File: | dev/pci/drm/amd/amdgpu/amdgpu_device.c |
Warning: | line 5498, column 3 Access to field 'next' results in a dereference of a null pointer |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | /* | |||
2 | * Copyright 2008 Advanced Micro Devices, Inc. | |||
3 | * Copyright 2008 Red Hat Inc. | |||
4 | * Copyright 2009 Jerome Glisse. | |||
5 | * | |||
6 | * Permission is hereby granted, free of charge, to any person obtaining a | |||
7 | * copy of this software and associated documentation files (the "Software"), | |||
8 | * to deal in the Software without restriction, including without limitation | |||
9 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, | |||
10 | * and/or sell copies of the Software, and to permit persons to whom the | |||
11 | * Software is furnished to do so, subject to the following conditions: | |||
12 | * | |||
13 | * The above copyright notice and this permission notice shall be included in | |||
14 | * all copies or substantial portions of the Software. | |||
15 | * | |||
16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | |||
17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | |||
18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL | |||
19 | * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR | |||
20 | * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, | |||
21 | * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR | |||
22 | * OTHER DEALINGS IN THE SOFTWARE. | |||
23 | * | |||
24 | * Authors: Dave Airlie | |||
25 | * Alex Deucher | |||
26 | * Jerome Glisse | |||
27 | */ | |||
28 | #include <linux/power_supply.h> | |||
29 | #include <linux/kthread.h> | |||
30 | #include <linux/module.h> | |||
31 | #include <linux/console.h> | |||
32 | #include <linux/slab.h> | |||
33 | #include <linux/iommu.h> | |||
34 | #include <linux/pci.h> | |||
35 | #include <linux/devcoredump.h> | |||
36 | #include <generated/utsrelease.h> | |||
37 | #include <linux/pci-p2pdma.h> | |||
38 | #include <linux/apple-gmux.h> | |||
39 | ||||
40 | #include <drm/drm_aperture.h> | |||
41 | #include <drm/drm_atomic_helper.h> | |||
42 | #include <drm/drm_probe_helper.h> | |||
43 | #include <drm/amdgpu_drm.h> | |||
44 | #include <linux/device.h> | |||
45 | #include <linux/vgaarb.h> | |||
46 | #include <linux/vga_switcheroo.h> | |||
47 | #include <linux/efi.h> | |||
48 | #include "amdgpu.h" | |||
49 | #include "amdgpu_trace.h" | |||
50 | #include "amdgpu_i2c.h" | |||
51 | #include "atom.h" | |||
52 | #include "amdgpu_atombios.h" | |||
53 | #include "amdgpu_atomfirmware.h" | |||
54 | #include "amd_pcie.h" | |||
55 | #ifdef CONFIG_DRM_AMDGPU_SI | |||
56 | #include "si.h" | |||
57 | #endif | |||
58 | #ifdef CONFIG_DRM_AMDGPU_CIK | |||
59 | #include "cik.h" | |||
60 | #endif | |||
61 | #include "vi.h" | |||
62 | #include "soc15.h" | |||
63 | #include "nv.h" | |||
64 | #include "bif/bif_4_1_d.h" | |||
65 | #include <linux/firmware.h> | |||
66 | #include "amdgpu_vf_error.h" | |||
67 | ||||
68 | #include "amdgpu_amdkfd.h" | |||
69 | #include "amdgpu_pm.h" | |||
70 | ||||
71 | #include "amdgpu_xgmi.h" | |||
72 | #include "amdgpu_ras.h" | |||
73 | #include "amdgpu_pmu.h" | |||
74 | #include "amdgpu_fru_eeprom.h" | |||
75 | #include "amdgpu_reset.h" | |||
76 | ||||
77 | #include <linux/suspend.h> | |||
78 | #include <drm/task_barrier.h> | |||
79 | #include <linux/pm_runtime.h> | |||
80 | ||||
81 | #include <drm/drm_drv.h> | |||
82 | ||||
83 | #if IS_ENABLED(CONFIG_X86)1 && defined(__linux__) | |||
84 | #include <asm/intel-family.h> | |||
85 | #endif | |||
86 | ||||
87 | MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin"); | |||
88 | MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin"); | |||
89 | MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin"); | |||
90 | MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin"); | |||
91 | MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin"); | |||
92 | MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin"); | |||
93 | MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin"); | |||
94 | ||||
95 | #define AMDGPU_RESUME_MS2000 2000 | |||
96 | #define AMDGPU_MAX_RETRY_LIMIT2 2 | |||
97 | #define AMDGPU_RETRY_SRIOV_RESET(r)((r) == -16 || (r) == -60 || (r) == -22) ((r) == -EBUSY16 || (r) == -ETIMEDOUT60 || (r) == -EINVAL22) | |||
98 | ||||
99 | static const struct drm_driver amdgpu_kms_driver; | |||
100 | ||||
101 | const char *amdgpu_asic_name[] = { | |||
102 | "TAHITI", | |||
103 | "PITCAIRN", | |||
104 | "VERDE", | |||
105 | "OLAND", | |||
106 | "HAINAN", | |||
107 | "BONAIRE", | |||
108 | "KAVERI", | |||
109 | "KABINI", | |||
110 | "HAWAII", | |||
111 | "MULLINS", | |||
112 | "TOPAZ", | |||
113 | "TONGA", | |||
114 | "FIJI", | |||
115 | "CARRIZO", | |||
116 | "STONEY", | |||
117 | "POLARIS10", | |||
118 | "POLARIS11", | |||
119 | "POLARIS12", | |||
120 | "VEGAM", | |||
121 | "VEGA10", | |||
122 | "VEGA12", | |||
123 | "VEGA20", | |||
124 | "RAVEN", | |||
125 | "ARCTURUS", | |||
126 | "RENOIR", | |||
127 | "ALDEBARAN", | |||
128 | "NAVI10", | |||
129 | "CYAN_SKILLFISH", | |||
130 | "NAVI14", | |||
131 | "NAVI12", | |||
132 | "SIENNA_CICHLID", | |||
133 | "NAVY_FLOUNDER", | |||
134 | "VANGOGH", | |||
135 | "DIMGREY_CAVEFISH", | |||
136 | "BEIGE_GOBY", | |||
137 | "YELLOW_CARP", | |||
138 | "IP DISCOVERY", | |||
139 | "LAST", | |||
140 | }; | |||
141 | ||||
142 | /** | |||
143 | * DOC: pcie_replay_count | |||
144 | * | |||
145 | * The amdgpu driver provides a sysfs API for reporting the total number | |||
146 | * of PCIe replays (NAKs) | |||
147 | * The file pcie_replay_count is used for this and returns the total | |||
148 | * number of replays as a sum of the NAKs generated and NAKs received | |||
149 | */ | |||
150 | ||||
151 | static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev, | |||
152 | struct device_attribute *attr, char *buf) | |||
153 | { | |||
154 | struct drm_device *ddev = dev_get_drvdata(dev); | |||
155 | struct amdgpu_device *adev = drm_to_adev(ddev); | |||
156 | uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev)((adev)->asic_funcs->get_pcie_replay_count((adev))); | |||
157 | ||||
158 | return sysfs_emit(buf, "%llu\n", cnt); | |||
159 | } | |||
160 | ||||
161 | static DEVICE_ATTR(pcie_replay_count, S_IRUGO,struct device_attribute dev_attr_pcie_replay_count | |||
162 | amdgpu_device_get_pcie_replay_count, NULL)struct device_attribute dev_attr_pcie_replay_count; | |||
163 | ||||
164 | static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev); | |||
165 | ||||
166 | /** | |||
167 | * DOC: product_name | |||
168 | * | |||
169 | * The amdgpu driver provides a sysfs API for reporting the product name | |||
170 | * for the device | |||
171 | * The file serial_number is used for this and returns the product name | |||
172 | * as returned from the FRU. | |||
173 | * NOTE: This is only available for certain server cards | |||
174 | */ | |||
175 | ||||
176 | static ssize_t amdgpu_device_get_product_name(struct device *dev, | |||
177 | struct device_attribute *attr, char *buf) | |||
178 | { | |||
179 | struct drm_device *ddev = dev_get_drvdata(dev); | |||
180 | struct amdgpu_device *adev = drm_to_adev(ddev); | |||
181 | ||||
182 | return sysfs_emit(buf, "%s\n", adev->product_name); | |||
183 | } | |||
184 | ||||
185 | static DEVICE_ATTR(product_name, S_IRUGO,struct device_attribute dev_attr_product_name | |||
186 | amdgpu_device_get_product_name, NULL)struct device_attribute dev_attr_product_name; | |||
187 | ||||
188 | /** | |||
189 | * DOC: product_number | |||
190 | * | |||
191 | * The amdgpu driver provides a sysfs API for reporting the part number | |||
192 | * for the device | |||
193 | * The file serial_number is used for this and returns the part number | |||
194 | * as returned from the FRU. | |||
195 | * NOTE: This is only available for certain server cards | |||
196 | */ | |||
197 | ||||
198 | static ssize_t amdgpu_device_get_product_number(struct device *dev, | |||
199 | struct device_attribute *attr, char *buf) | |||
200 | { | |||
201 | struct drm_device *ddev = dev_get_drvdata(dev); | |||
202 | struct amdgpu_device *adev = drm_to_adev(ddev); | |||
203 | ||||
204 | return sysfs_emit(buf, "%s\n", adev->product_number); | |||
205 | } | |||
206 | ||||
207 | static DEVICE_ATTR(product_number, S_IRUGO,struct device_attribute dev_attr_product_number | |||
208 | amdgpu_device_get_product_number, NULL)struct device_attribute dev_attr_product_number; | |||
209 | ||||
210 | /** | |||
211 | * DOC: serial_number | |||
212 | * | |||
213 | * The amdgpu driver provides a sysfs API for reporting the serial number | |||
214 | * for the device | |||
215 | * The file serial_number is used for this and returns the serial number | |||
216 | * as returned from the FRU. | |||
217 | * NOTE: This is only available for certain server cards | |||
218 | */ | |||
219 | ||||
220 | static ssize_t amdgpu_device_get_serial_number(struct device *dev, | |||
221 | struct device_attribute *attr, char *buf) | |||
222 | { | |||
223 | struct drm_device *ddev = dev_get_drvdata(dev); | |||
224 | struct amdgpu_device *adev = drm_to_adev(ddev); | |||
225 | ||||
226 | return sysfs_emit(buf, "%s\n", adev->serial); | |||
227 | } | |||
228 | ||||
229 | static DEVICE_ATTR(serial_number, S_IRUGO,struct device_attribute dev_attr_serial_number | |||
230 | amdgpu_device_get_serial_number, NULL)struct device_attribute dev_attr_serial_number; | |||
231 | ||||
232 | /** | |||
233 | * amdgpu_device_supports_px - Is the device a dGPU with ATPX power control | |||
234 | * | |||
235 | * @dev: drm_device pointer | |||
236 | * | |||
237 | * Returns true if the device is a dGPU with ATPX power control, | |||
238 | * otherwise return false. | |||
239 | */ | |||
240 | bool_Bool amdgpu_device_supports_px(struct drm_device *dev) | |||
241 | { | |||
242 | struct amdgpu_device *adev = drm_to_adev(dev); | |||
243 | ||||
244 | if ((adev->flags & AMD_IS_PX) && !amdgpu_is_atpx_hybrid()) | |||
245 | return true1; | |||
246 | return false0; | |||
247 | } | |||
248 | ||||
249 | /** | |||
250 | * amdgpu_device_supports_boco - Is the device a dGPU with ACPI power resources | |||
251 | * | |||
252 | * @dev: drm_device pointer | |||
253 | * | |||
254 | * Returns true if the device is a dGPU with ACPI power control, | |||
255 | * otherwise return false. | |||
256 | */ | |||
257 | bool_Bool amdgpu_device_supports_boco(struct drm_device *dev) | |||
258 | { | |||
259 | struct amdgpu_device *adev = drm_to_adev(dev); | |||
260 | ||||
261 | if (adev->has_pr3 || | |||
262 | ((adev->flags & AMD_IS_PX) && amdgpu_is_atpx_hybrid())) | |||
263 | return true1; | |||
264 | return false0; | |||
265 | } | |||
266 | ||||
267 | /** | |||
268 | * amdgpu_device_supports_baco - Does the device support BACO | |||
269 | * | |||
270 | * @dev: drm_device pointer | |||
271 | * | |||
272 | * Returns true if the device supporte BACO, | |||
273 | * otherwise return false. | |||
274 | */ | |||
275 | bool_Bool amdgpu_device_supports_baco(struct drm_device *dev) | |||
276 | { | |||
277 | struct amdgpu_device *adev = drm_to_adev(dev); | |||
278 | ||||
279 | return amdgpu_asic_supports_baco(adev)(adev)->asic_funcs->supports_baco((adev)); | |||
280 | } | |||
281 | ||||
282 | /** | |||
283 | * amdgpu_device_supports_smart_shift - Is the device dGPU with | |||
284 | * smart shift support | |||
285 | * | |||
286 | * @dev: drm_device pointer | |||
287 | * | |||
288 | * Returns true if the device is a dGPU with Smart Shift support, | |||
289 | * otherwise returns false. | |||
290 | */ | |||
291 | bool_Bool amdgpu_device_supports_smart_shift(struct drm_device *dev) | |||
292 | { | |||
293 | return (amdgpu_device_supports_boco(dev) && | |||
294 | amdgpu_acpi_is_power_shift_control_supported()); | |||
295 | } | |||
296 | ||||
297 | /* | |||
298 | * VRAM access helper functions | |||
299 | */ | |||
300 | ||||
301 | /** | |||
302 | * amdgpu_device_mm_access - access vram by MM_INDEX/MM_DATA | |||
303 | * | |||
304 | * @adev: amdgpu_device pointer | |||
305 | * @pos: offset of the buffer in vram | |||
306 | * @buf: virtual address of the buffer in system memory | |||
307 | * @size: read/write size, sizeof(@buf) must > @size | |||
308 | * @write: true - write to vram, otherwise - read from vram | |||
309 | */ | |||
310 | void amdgpu_device_mm_access(struct amdgpu_device *adev, loff_t pos, | |||
311 | void *buf, size_t size, bool_Bool write) | |||
312 | { | |||
313 | unsigned long flags; | |||
314 | uint32_t hi = ~0, tmp = 0; | |||
315 | uint32_t *data = buf; | |||
316 | uint64_t last; | |||
317 | int idx; | |||
318 | ||||
319 | if (!drm_dev_enter(adev_to_drm(adev), &idx)) | |||
320 | return; | |||
321 | ||||
322 | BUG_ON(!IS_ALIGNED(pos, 4) || !IS_ALIGNED(size, 4))((!(!(((pos) & ((4) - 1)) == 0) || !(((size) & ((4) - 1)) == 0))) ? (void)0 : __assert("diagnostic ", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_device.c" , 322, "!(!(((pos) & ((4) - 1)) == 0) || !(((size) & ((4) - 1)) == 0))" )); | |||
323 | ||||
324 | spin_lock_irqsave(&adev->mmio_idx_lock, flags)do { flags = 0; mtx_enter(&adev->mmio_idx_lock); } while (0); | |||
325 | for (last = pos + size; pos < last; pos += 4) { | |||
326 | tmp = pos >> 31; | |||
327 | ||||
328 | WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000)amdgpu_device_wreg(adev, (0x0), (((uint32_t)pos) | 0x80000000 ), (1<<1)); | |||
329 | if (tmp != hi) { | |||
330 | WREG32_NO_KIQ(mmMM_INDEX_HI, tmp)amdgpu_device_wreg(adev, (0x6), (tmp), (1<<1)); | |||
331 | hi = tmp; | |||
332 | } | |||
333 | if (write) | |||
334 | WREG32_NO_KIQ(mmMM_DATA, *data++)amdgpu_device_wreg(adev, (0x1), (*data++), (1<<1)); | |||
335 | else | |||
336 | *data++ = RREG32_NO_KIQ(mmMM_DATA)amdgpu_device_rreg(adev, (0x1), (1<<1)); | |||
337 | } | |||
338 | ||||
339 | spin_unlock_irqrestore(&adev->mmio_idx_lock, flags)do { (void)(flags); mtx_leave(&adev->mmio_idx_lock); } while (0); | |||
340 | drm_dev_exit(idx); | |||
341 | } | |||
342 | ||||
343 | /** | |||
344 | * amdgpu_device_aper_access - access vram by vram aperature | |||
345 | * | |||
346 | * @adev: amdgpu_device pointer | |||
347 | * @pos: offset of the buffer in vram | |||
348 | * @buf: virtual address of the buffer in system memory | |||
349 | * @size: read/write size, sizeof(@buf) must > @size | |||
350 | * @write: true - write to vram, otherwise - read from vram | |||
351 | * | |||
352 | * The return value means how many bytes have been transferred. | |||
353 | */ | |||
354 | size_t amdgpu_device_aper_access(struct amdgpu_device *adev, loff_t pos, | |||
355 | void *buf, size_t size, bool_Bool write) | |||
356 | { | |||
357 | #ifdef CONFIG_64BIT1 | |||
358 | void __iomem *addr; | |||
359 | size_t count = 0; | |||
360 | uint64_t last; | |||
361 | ||||
362 | if (!adev->mman.aper_base_kaddr) | |||
363 | return 0; | |||
364 | ||||
365 | last = min(pos + size, adev->gmc.visible_vram_size)(((pos + size)<(adev->gmc.visible_vram_size))?(pos + size ):(adev->gmc.visible_vram_size)); | |||
366 | if (last > pos) { | |||
367 | addr = adev->mman.aper_base_kaddr + pos; | |||
368 | count = last - pos; | |||
369 | ||||
370 | if (write) { | |||
371 | memcpy_toio(addr, buf, count)__builtin_memcpy((addr), (buf), (count)); | |||
372 | mb()do { __asm volatile("mfence" ::: "memory"); } while (0); | |||
373 | amdgpu_device_flush_hdp(adev, NULL((void *)0)); | |||
374 | } else { | |||
375 | amdgpu_device_invalidate_hdp(adev, NULL((void *)0)); | |||
376 | mb()do { __asm volatile("mfence" ::: "memory"); } while (0); | |||
377 | memcpy_fromio(buf, addr, count)__builtin_memcpy((buf), (addr), (count)); | |||
378 | } | |||
379 | ||||
380 | } | |||
381 | ||||
382 | return count; | |||
383 | #else | |||
384 | return 0; | |||
385 | #endif | |||
386 | } | |||
387 | ||||
388 | /** | |||
389 | * amdgpu_device_vram_access - read/write a buffer in vram | |||
390 | * | |||
391 | * @adev: amdgpu_device pointer | |||
392 | * @pos: offset of the buffer in vram | |||
393 | * @buf: virtual address of the buffer in system memory | |||
394 | * @size: read/write size, sizeof(@buf) must > @size | |||
395 | * @write: true - write to vram, otherwise - read from vram | |||
396 | */ | |||
397 | void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, | |||
398 | void *buf, size_t size, bool_Bool write) | |||
399 | { | |||
400 | size_t count; | |||
401 | ||||
402 | /* try to using vram apreature to access vram first */ | |||
403 | count = amdgpu_device_aper_access(adev, pos, buf, size, write); | |||
404 | size -= count; | |||
405 | if (size) { | |||
406 | /* using MM to access rest vram */ | |||
407 | pos += count; | |||
408 | buf += count; | |||
409 | amdgpu_device_mm_access(adev, pos, buf, size, write); | |||
410 | } | |||
411 | } | |||
412 | ||||
413 | /* | |||
414 | * register access helper functions. | |||
415 | */ | |||
416 | ||||
417 | /* Check if hw access should be skipped because of hotplug or device error */ | |||
418 | bool_Bool amdgpu_device_skip_hw_access(struct amdgpu_device *adev) | |||
419 | { | |||
420 | if (adev->no_hw_access) | |||
421 | return true1; | |||
422 | ||||
423 | #ifdef CONFIG_LOCKDEP | |||
424 | /* | |||
425 | * This is a bit complicated to understand, so worth a comment. What we assert | |||
426 | * here is that the GPU reset is not running on another thread in parallel. | |||
427 | * | |||
428 | * For this we trylock the read side of the reset semaphore, if that succeeds | |||
429 | * we know that the reset is not running in paralell. | |||
430 | * | |||
431 | * If the trylock fails we assert that we are either already holding the read | |||
432 | * side of the lock or are the reset thread itself and hold the write side of | |||
433 | * the lock. | |||
434 | */ | |||
435 | if (in_task()) { | |||
436 | if (down_read_trylock(&adev->reset_domain->sem)(rw_enter(&adev->reset_domain->sem, 0x0002UL | 0x0040UL ) == 0)) | |||
437 | up_read(&adev->reset_domain->sem)rw_exit_read(&adev->reset_domain->sem); | |||
438 | else | |||
439 | lockdep_assert_held(&adev->reset_domain->sem)do { (void)(&adev->reset_domain->sem); } while(0); | |||
440 | } | |||
441 | #endif | |||
442 | return false0; | |||
443 | } | |||
444 | ||||
445 | /** | |||
446 | * amdgpu_device_rreg - read a memory mapped IO or indirect register | |||
447 | * | |||
448 | * @adev: amdgpu_device pointer | |||
449 | * @reg: dword aligned register offset | |||
450 | * @acc_flags: access flags which require special behavior | |||
451 | * | |||
452 | * Returns the 32 bit value from the offset specified. | |||
453 | */ | |||
454 | uint32_t amdgpu_device_rreg(struct amdgpu_device *adev, | |||
455 | uint32_t reg, uint32_t acc_flags) | |||
456 | { | |||
457 | uint32_t ret; | |||
458 | ||||
459 | if (amdgpu_device_skip_hw_access(adev)) | |||
460 | return 0; | |||
461 | ||||
462 | if ((reg * 4) < adev->rmmio_size) { | |||
463 | if (!(acc_flags & AMDGPU_REGS_NO_KIQ(1<<1)) && | |||
464 | amdgpu_sriov_runtime(adev)((adev)->virt.caps & (1 << 4)) && | |||
465 | down_read_trylock(&adev->reset_domain->sem)(rw_enter(&adev->reset_domain->sem, 0x0002UL | 0x0040UL ) == 0)) { | |||
466 | ret = amdgpu_kiq_rreg(adev, reg); | |||
467 | up_read(&adev->reset_domain->sem)rw_exit_read(&adev->reset_domain->sem); | |||
468 | } else { | |||
469 | ret = readl(((void __iomem *)adev->rmmio) + (reg * 4))ioread32(((void *)adev->rmmio) + (reg * 4)); | |||
470 | } | |||
471 | } else { | |||
472 | ret = adev->pcie_rreg(adev, reg * 4); | |||
473 | } | |||
474 | ||||
475 | trace_amdgpu_device_rreg(adev->pdev->device, reg, ret); | |||
476 | ||||
477 | return ret; | |||
478 | } | |||
479 | ||||
480 | /* | |||
481 | * MMIO register read with bytes helper functions | |||
482 | * @offset:bytes offset from MMIO start | |||
483 | * | |||
484 | */ | |||
485 | ||||
486 | /** | |||
487 | * amdgpu_mm_rreg8 - read a memory mapped IO register | |||
488 | * | |||
489 | * @adev: amdgpu_device pointer | |||
490 | * @offset: byte aligned register offset | |||
491 | * | |||
492 | * Returns the 8 bit value from the offset specified. | |||
493 | */ | |||
494 | uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) | |||
495 | { | |||
496 | if (amdgpu_device_skip_hw_access(adev)) | |||
497 | return 0; | |||
498 | ||||
499 | if (offset < adev->rmmio_size) | |||
500 | return (readb(adev->rmmio + offset)ioread8(adev->rmmio + offset)); | |||
501 | BUG()do { panic("BUG at %s:%d", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_device.c" , 501); } while (0); | |||
502 | } | |||
503 | ||||
504 | /* | |||
505 | * MMIO register write with bytes helper functions | |||
506 | * @offset:bytes offset from MMIO start | |||
507 | * @value: the value want to be written to the register | |||
508 | * | |||
509 | */ | |||
510 | /** | |||
511 | * amdgpu_mm_wreg8 - read a memory mapped IO register | |||
512 | * | |||
513 | * @adev: amdgpu_device pointer | |||
514 | * @offset: byte aligned register offset | |||
515 | * @value: 8 bit value to write | |||
516 | * | |||
517 | * Writes the value specified to the offset specified. | |||
518 | */ | |||
519 | void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) | |||
520 | { | |||
521 | if (amdgpu_device_skip_hw_access(adev)) | |||
522 | return; | |||
523 | ||||
524 | if (offset < adev->rmmio_size) | |||
525 | writeb(value, adev->rmmio + offset)iowrite8(value, adev->rmmio + offset); | |||
526 | else | |||
527 | BUG()do { panic("BUG at %s:%d", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_device.c" , 527); } while (0); | |||
528 | } | |||
529 | ||||
530 | /** | |||
531 | * amdgpu_device_wreg - write to a memory mapped IO or indirect register | |||
532 | * | |||
533 | * @adev: amdgpu_device pointer | |||
534 | * @reg: dword aligned register offset | |||
535 | * @v: 32 bit value to write to the register | |||
536 | * @acc_flags: access flags which require special behavior | |||
537 | * | |||
538 | * Writes the value specified to the offset specified. | |||
539 | */ | |||
540 | void amdgpu_device_wreg(struct amdgpu_device *adev, | |||
541 | uint32_t reg, uint32_t v, | |||
542 | uint32_t acc_flags) | |||
543 | { | |||
544 | if (amdgpu_device_skip_hw_access(adev)) | |||
545 | return; | |||
546 | ||||
547 | if ((reg * 4) < adev->rmmio_size) { | |||
548 | if (!(acc_flags & AMDGPU_REGS_NO_KIQ(1<<1)) && | |||
549 | amdgpu_sriov_runtime(adev)((adev)->virt.caps & (1 << 4)) && | |||
550 | down_read_trylock(&adev->reset_domain->sem)(rw_enter(&adev->reset_domain->sem, 0x0002UL | 0x0040UL ) == 0)) { | |||
551 | amdgpu_kiq_wreg(adev, reg, v); | |||
552 | up_read(&adev->reset_domain->sem)rw_exit_read(&adev->reset_domain->sem); | |||
553 | } else { | |||
554 | writel(v, ((void __iomem *)adev->rmmio) + (reg * 4))iowrite32(v, ((void *)adev->rmmio) + (reg * 4)); | |||
555 | } | |||
556 | } else { | |||
557 | adev->pcie_wreg(adev, reg * 4, v); | |||
558 | } | |||
559 | ||||
560 | trace_amdgpu_device_wreg(adev->pdev->device, reg, v); | |||
561 | } | |||
562 | ||||
563 | /** | |||
564 | * amdgpu_mm_wreg_mmio_rlc - write register either with direct/indirect mmio or with RLC path if in range | |||
565 | * | |||
566 | * @adev: amdgpu_device pointer | |||
567 | * @reg: mmio/rlc register | |||
568 | * @v: value to write | |||
569 | * | |||
570 | * this function is invoked only for the debugfs register access | |||
571 | */ | |||
572 | void amdgpu_mm_wreg_mmio_rlc(struct amdgpu_device *adev, | |||
573 | uint32_t reg, uint32_t v) | |||
574 | { | |||
575 | if (amdgpu_device_skip_hw_access(adev)) | |||
576 | return; | |||
577 | ||||
578 | if (amdgpu_sriov_fullaccess(adev)((((adev))->virt.caps & (1 << 2)) && !(( (adev))->virt.caps & (1 << 4))) && | |||
579 | adev->gfx.rlc.funcs && | |||
580 | adev->gfx.rlc.funcs->is_rlcg_access_range) { | |||
581 | if (adev->gfx.rlc.funcs->is_rlcg_access_range(adev, reg)) | |||
582 | return amdgpu_sriov_wreg(adev, reg, v, 0, 0); | |||
583 | } else if ((reg * 4) >= adev->rmmio_size) { | |||
584 | adev->pcie_wreg(adev, reg * 4, v); | |||
585 | } else { | |||
586 | writel(v, ((void __iomem *)adev->rmmio) + (reg * 4))iowrite32(v, ((void *)adev->rmmio) + (reg * 4)); | |||
587 | } | |||
588 | } | |||
589 | ||||
590 | /** | |||
591 | * amdgpu_mm_rdoorbell - read a doorbell dword | |||
592 | * | |||
593 | * @adev: amdgpu_device pointer | |||
594 | * @index: doorbell index | |||
595 | * | |||
596 | * Returns the value in the doorbell aperture at the | |||
597 | * requested doorbell index (CIK). | |||
598 | */ | |||
599 | u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index) | |||
600 | { | |||
601 | if (amdgpu_device_skip_hw_access(adev)) | |||
602 | return 0; | |||
603 | ||||
604 | if (index < adev->doorbell.num_doorbells) { | |||
605 | return readl(adev->doorbell.ptr + index)ioread32(adev->doorbell.ptr + index); | |||
606 | } else { | |||
607 | DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index)__drm_err("reading beyond doorbell aperture: 0x%08x!\n", index ); | |||
608 | return 0; | |||
609 | } | |||
610 | } | |||
611 | ||||
612 | /** | |||
613 | * amdgpu_mm_wdoorbell - write a doorbell dword | |||
614 | * | |||
615 | * @adev: amdgpu_device pointer | |||
616 | * @index: doorbell index | |||
617 | * @v: value to write | |||
618 | * | |||
619 | * Writes @v to the doorbell aperture at the | |||
620 | * requested doorbell index (CIK). | |||
621 | */ | |||
622 | void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v) | |||
623 | { | |||
624 | if (amdgpu_device_skip_hw_access(adev)) | |||
625 | return; | |||
626 | ||||
627 | if (index < adev->doorbell.num_doorbells) { | |||
628 | writel(v, adev->doorbell.ptr + index)iowrite32(v, adev->doorbell.ptr + index); | |||
629 | } else { | |||
630 | DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index)__drm_err("writing beyond doorbell aperture: 0x%08x!\n", index ); | |||
631 | } | |||
632 | } | |||
633 | ||||
634 | /** | |||
635 | * amdgpu_mm_rdoorbell64 - read a doorbell Qword | |||
636 | * | |||
637 | * @adev: amdgpu_device pointer | |||
638 | * @index: doorbell index | |||
639 | * | |||
640 | * Returns the value in the doorbell aperture at the | |||
641 | * requested doorbell index (VEGA10+). | |||
642 | */ | |||
643 | u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index) | |||
644 | { | |||
645 | if (amdgpu_device_skip_hw_access(adev)) | |||
646 | return 0; | |||
647 | ||||
648 | if (index < adev->doorbell.num_doorbells) { | |||
649 | return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index))({ typeof(*((atomic64_t *)(adev->doorbell.ptr + index))) __tmp = *(volatile typeof(*((atomic64_t *)(adev->doorbell.ptr + index))) *)&(*((atomic64_t *)(adev->doorbell.ptr + index ))); membar_datadep_consumer(); __tmp; }); | |||
650 | } else { | |||
651 | DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index)__drm_err("reading beyond doorbell aperture: 0x%08x!\n", index ); | |||
652 | return 0; | |||
653 | } | |||
654 | } | |||
655 | ||||
656 | /** | |||
657 | * amdgpu_mm_wdoorbell64 - write a doorbell Qword | |||
658 | * | |||
659 | * @adev: amdgpu_device pointer | |||
660 | * @index: doorbell index | |||
661 | * @v: value to write | |||
662 | * | |||
663 | * Writes @v to the doorbell aperture at the | |||
664 | * requested doorbell index (VEGA10+). | |||
665 | */ | |||
666 | void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v) | |||
667 | { | |||
668 | if (amdgpu_device_skip_hw_access(adev)) | |||
669 | return; | |||
670 | ||||
671 | if (index < adev->doorbell.num_doorbells) { | |||
672 | atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v)({ typeof(*((atomic64_t *)(adev->doorbell.ptr + index))) __tmp = ((v)); *(volatile typeof(*((atomic64_t *)(adev->doorbell .ptr + index))) *)&(*((atomic64_t *)(adev->doorbell.ptr + index))) = __tmp; __tmp; }); | |||
673 | } else { | |||
674 | DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index)__drm_err("writing beyond doorbell aperture: 0x%08x!\n", index ); | |||
675 | } | |||
676 | } | |||
677 | ||||
678 | /** | |||
679 | * amdgpu_device_indirect_rreg - read an indirect register | |||
680 | * | |||
681 | * @adev: amdgpu_device pointer | |||
682 | * @pcie_index: mmio register offset | |||
683 | * @pcie_data: mmio register offset | |||
684 | * @reg_addr: indirect register address to read from | |||
685 | * | |||
686 | * Returns the value of indirect register @reg_addr | |||
687 | */ | |||
688 | u32 amdgpu_device_indirect_rreg(struct amdgpu_device *adev, | |||
689 | u32 pcie_index, u32 pcie_data, | |||
690 | u32 reg_addr) | |||
691 | { | |||
692 | unsigned long flags; | |||
693 | u32 r; | |||
694 | void __iomem *pcie_index_offset; | |||
695 | void __iomem *pcie_data_offset; | |||
696 | ||||
697 | spin_lock_irqsave(&adev->pcie_idx_lock, flags)do { flags = 0; mtx_enter(&adev->pcie_idx_lock); } while (0); | |||
698 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |||
699 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |||
700 | ||||
701 | writel(reg_addr, pcie_index_offset)iowrite32(reg_addr, pcie_index_offset); | |||
702 | readl(pcie_index_offset)ioread32(pcie_index_offset); | |||
703 | r = readl(pcie_data_offset)ioread32(pcie_data_offset); | |||
704 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags)do { (void)(flags); mtx_leave(&adev->pcie_idx_lock); } while (0); | |||
705 | ||||
706 | return r; | |||
707 | } | |||
708 | ||||
709 | /** | |||
710 | * amdgpu_device_indirect_rreg64 - read a 64bits indirect register | |||
711 | * | |||
712 | * @adev: amdgpu_device pointer | |||
713 | * @pcie_index: mmio register offset | |||
714 | * @pcie_data: mmio register offset | |||
715 | * @reg_addr: indirect register address to read from | |||
716 | * | |||
717 | * Returns the value of indirect register @reg_addr | |||
718 | */ | |||
719 | u64 amdgpu_device_indirect_rreg64(struct amdgpu_device *adev, | |||
720 | u32 pcie_index, u32 pcie_data, | |||
721 | u32 reg_addr) | |||
722 | { | |||
723 | unsigned long flags; | |||
724 | u64 r; | |||
725 | void __iomem *pcie_index_offset; | |||
726 | void __iomem *pcie_data_offset; | |||
727 | ||||
728 | spin_lock_irqsave(&adev->pcie_idx_lock, flags)do { flags = 0; mtx_enter(&adev->pcie_idx_lock); } while (0); | |||
729 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |||
730 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |||
731 | ||||
732 | /* read low 32 bits */ | |||
733 | writel(reg_addr, pcie_index_offset)iowrite32(reg_addr, pcie_index_offset); | |||
734 | readl(pcie_index_offset)ioread32(pcie_index_offset); | |||
735 | r = readl(pcie_data_offset)ioread32(pcie_data_offset); | |||
736 | /* read high 32 bits */ | |||
737 | writel(reg_addr + 4, pcie_index_offset)iowrite32(reg_addr + 4, pcie_index_offset); | |||
738 | readl(pcie_index_offset)ioread32(pcie_index_offset); | |||
739 | r |= ((u64)readl(pcie_data_offset)ioread32(pcie_data_offset) << 32); | |||
740 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags)do { (void)(flags); mtx_leave(&adev->pcie_idx_lock); } while (0); | |||
741 | ||||
742 | return r; | |||
743 | } | |||
744 | ||||
745 | /** | |||
746 | * amdgpu_device_indirect_wreg - write an indirect register address | |||
747 | * | |||
748 | * @adev: amdgpu_device pointer | |||
749 | * @pcie_index: mmio register offset | |||
750 | * @pcie_data: mmio register offset | |||
751 | * @reg_addr: indirect register offset | |||
752 | * @reg_data: indirect register data | |||
753 | * | |||
754 | */ | |||
755 | void amdgpu_device_indirect_wreg(struct amdgpu_device *adev, | |||
756 | u32 pcie_index, u32 pcie_data, | |||
757 | u32 reg_addr, u32 reg_data) | |||
758 | { | |||
759 | unsigned long flags; | |||
760 | void __iomem *pcie_index_offset; | |||
761 | void __iomem *pcie_data_offset; | |||
762 | ||||
763 | spin_lock_irqsave(&adev->pcie_idx_lock, flags)do { flags = 0; mtx_enter(&adev->pcie_idx_lock); } while (0); | |||
764 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |||
765 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |||
766 | ||||
767 | writel(reg_addr, pcie_index_offset)iowrite32(reg_addr, pcie_index_offset); | |||
768 | readl(pcie_index_offset)ioread32(pcie_index_offset); | |||
769 | writel(reg_data, pcie_data_offset)iowrite32(reg_data, pcie_data_offset); | |||
770 | readl(pcie_data_offset)ioread32(pcie_data_offset); | |||
771 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags)do { (void)(flags); mtx_leave(&adev->pcie_idx_lock); } while (0); | |||
772 | } | |||
773 | ||||
774 | /** | |||
775 | * amdgpu_device_indirect_wreg64 - write a 64bits indirect register address | |||
776 | * | |||
777 | * @adev: amdgpu_device pointer | |||
778 | * @pcie_index: mmio register offset | |||
779 | * @pcie_data: mmio register offset | |||
780 | * @reg_addr: indirect register offset | |||
781 | * @reg_data: indirect register data | |||
782 | * | |||
783 | */ | |||
784 | void amdgpu_device_indirect_wreg64(struct amdgpu_device *adev, | |||
785 | u32 pcie_index, u32 pcie_data, | |||
786 | u32 reg_addr, u64 reg_data) | |||
787 | { | |||
788 | unsigned long flags; | |||
789 | void __iomem *pcie_index_offset; | |||
790 | void __iomem *pcie_data_offset; | |||
791 | ||||
792 | spin_lock_irqsave(&adev->pcie_idx_lock, flags)do { flags = 0; mtx_enter(&adev->pcie_idx_lock); } while (0); | |||
793 | pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4; | |||
794 | pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4; | |||
795 | ||||
796 | /* write low 32 bits */ | |||
797 | writel(reg_addr, pcie_index_offset)iowrite32(reg_addr, pcie_index_offset); | |||
798 | readl(pcie_index_offset)ioread32(pcie_index_offset); | |||
799 | writel((u32)(reg_data & 0xffffffffULL), pcie_data_offset)iowrite32((u32)(reg_data & 0xffffffffULL), pcie_data_offset ); | |||
800 | readl(pcie_data_offset)ioread32(pcie_data_offset); | |||
801 | /* write high 32 bits */ | |||
802 | writel(reg_addr + 4, pcie_index_offset)iowrite32(reg_addr + 4, pcie_index_offset); | |||
803 | readl(pcie_index_offset)ioread32(pcie_index_offset); | |||
804 | writel((u32)(reg_data >> 32), pcie_data_offset)iowrite32((u32)(reg_data >> 32), pcie_data_offset); | |||
805 | readl(pcie_data_offset)ioread32(pcie_data_offset); | |||
806 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags)do { (void)(flags); mtx_leave(&adev->pcie_idx_lock); } while (0); | |||
807 | } | |||
808 | ||||
809 | /** | |||
810 | * amdgpu_invalid_rreg - dummy reg read function | |||
811 | * | |||
812 | * @adev: amdgpu_device pointer | |||
813 | * @reg: offset of register | |||
814 | * | |||
815 | * Dummy register read function. Used for register blocks | |||
816 | * that certain asics don't have (all asics). | |||
817 | * Returns the value in the register. | |||
818 | */ | |||
819 | static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg) | |||
820 | { | |||
821 | DRM_ERROR("Invalid callback to read register 0x%04X\n", reg)__drm_err("Invalid callback to read register 0x%04X\n", reg); | |||
822 | BUG()do { panic("BUG at %s:%d", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_device.c" , 822); } while (0); | |||
823 | return 0; | |||
824 | } | |||
825 | ||||
826 | /** | |||
827 | * amdgpu_invalid_wreg - dummy reg write function | |||
828 | * | |||
829 | * @adev: amdgpu_device pointer | |||
830 | * @reg: offset of register | |||
831 | * @v: value to write to the register | |||
832 | * | |||
833 | * Dummy register read function. Used for register blocks | |||
834 | * that certain asics don't have (all asics). | |||
835 | */ | |||
836 | static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v) | |||
837 | { | |||
838 | DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",__drm_err("Invalid callback to write register 0x%04X with 0x%08X\n" , reg, v) | |||
839 | reg, v)__drm_err("Invalid callback to write register 0x%04X with 0x%08X\n" , reg, v); | |||
840 | BUG()do { panic("BUG at %s:%d", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_device.c" , 840); } while (0); | |||
841 | } | |||
842 | ||||
843 | /** | |||
844 | * amdgpu_invalid_rreg64 - dummy 64 bit reg read function | |||
845 | * | |||
846 | * @adev: amdgpu_device pointer | |||
847 | * @reg: offset of register | |||
848 | * | |||
849 | * Dummy register read function. Used for register blocks | |||
850 | * that certain asics don't have (all asics). | |||
851 | * Returns the value in the register. | |||
852 | */ | |||
853 | static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg) | |||
854 | { | |||
855 | DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg)__drm_err("Invalid callback to read 64 bit register 0x%04X\n" , reg); | |||
856 | BUG()do { panic("BUG at %s:%d", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_device.c" , 856); } while (0); | |||
857 | return 0; | |||
858 | } | |||
859 | ||||
860 | /** | |||
861 | * amdgpu_invalid_wreg64 - dummy reg write function | |||
862 | * | |||
863 | * @adev: amdgpu_device pointer | |||
864 | * @reg: offset of register | |||
865 | * @v: value to write to the register | |||
866 | * | |||
867 | * Dummy register read function. Used for register blocks | |||
868 | * that certain asics don't have (all asics). | |||
869 | */ | |||
870 | static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v) | |||
871 | { | |||
872 | DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n",__drm_err("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n" , reg, v) | |||
873 | reg, v)__drm_err("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n" , reg, v); | |||
874 | BUG()do { panic("BUG at %s:%d", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_device.c" , 874); } while (0); | |||
875 | } | |||
876 | ||||
877 | /** | |||
878 | * amdgpu_block_invalid_rreg - dummy reg read function | |||
879 | * | |||
880 | * @adev: amdgpu_device pointer | |||
881 | * @block: offset of instance | |||
882 | * @reg: offset of register | |||
883 | * | |||
884 | * Dummy register read function. Used for register blocks | |||
885 | * that certain asics don't have (all asics). | |||
886 | * Returns the value in the register. | |||
887 | */ | |||
888 | static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev, | |||
889 | uint32_t block, uint32_t reg) | |||
890 | { | |||
891 | DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",__drm_err("Invalid callback to read register 0x%04X in block 0x%04X\n" , reg, block) | |||
892 | reg, block)__drm_err("Invalid callback to read register 0x%04X in block 0x%04X\n" , reg, block); | |||
893 | BUG()do { panic("BUG at %s:%d", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_device.c" , 893); } while (0); | |||
894 | return 0; | |||
895 | } | |||
896 | ||||
897 | /** | |||
898 | * amdgpu_block_invalid_wreg - dummy reg write function | |||
899 | * | |||
900 | * @adev: amdgpu_device pointer | |||
901 | * @block: offset of instance | |||
902 | * @reg: offset of register | |||
903 | * @v: value to write to the register | |||
904 | * | |||
905 | * Dummy register read function. Used for register blocks | |||
906 | * that certain asics don't have (all asics). | |||
907 | */ | |||
908 | static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev, | |||
909 | uint32_t block, | |||
910 | uint32_t reg, uint32_t v) | |||
911 | { | |||
912 | DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",__drm_err("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n" , reg, block, v) | |||
913 | reg, block, v)__drm_err("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n" , reg, block, v); | |||
914 | BUG()do { panic("BUG at %s:%d", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_device.c" , 914); } while (0); | |||
915 | } | |||
916 | ||||
917 | /** | |||
918 | * amdgpu_device_asic_init - Wrapper for atom asic_init | |||
919 | * | |||
920 | * @adev: amdgpu_device pointer | |||
921 | * | |||
922 | * Does any asic specific work and then calls atom asic init. | |||
923 | */ | |||
924 | static int amdgpu_device_asic_init(struct amdgpu_device *adev) | |||
925 | { | |||
926 | amdgpu_asic_pre_asic_init(adev)(adev)->asic_funcs->pre_asic_init((adev)); | |||
927 | ||||
928 | if (adev->ip_versions[GC_HWIP][0] >= IP_VERSION(11, 0, 0)(((11) << 16) | ((0) << 8) | (0))) | |||
929 | return amdgpu_atomfirmware_asic_init(adev, true1); | |||
930 | else | |||
931 | return amdgpu_atom_asic_init(adev->mode_info.atom_context); | |||
932 | } | |||
933 | ||||
934 | /** | |||
935 | * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page | |||
936 | * | |||
937 | * @adev: amdgpu_device pointer | |||
938 | * | |||
939 | * Allocates a scratch page of VRAM for use by various things in the | |||
940 | * driver. | |||
941 | */ | |||
942 | static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev) | |||
943 | { | |||
944 | return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE4096, | |||
945 | PAGE_SIZE(1 << 12), AMDGPU_GEM_DOMAIN_VRAM0x4, | |||
946 | &adev->vram_scratch.robj, | |||
947 | &adev->vram_scratch.gpu_addr, | |||
948 | (void **)&adev->vram_scratch.ptr); | |||
949 | } | |||
950 | ||||
951 | /** | |||
952 | * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page | |||
953 | * | |||
954 | * @adev: amdgpu_device pointer | |||
955 | * | |||
956 | * Frees the VRAM scratch page. | |||
957 | */ | |||
958 | static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev) | |||
959 | { | |||
960 | amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL((void *)0), NULL((void *)0)); | |||
961 | } | |||
962 | ||||
963 | /** | |||
964 | * amdgpu_device_program_register_sequence - program an array of registers. | |||
965 | * | |||
966 | * @adev: amdgpu_device pointer | |||
967 | * @registers: pointer to the register array | |||
968 | * @array_size: size of the register array | |||
969 | * | |||
970 | * Programs an array or registers with and and or masks. | |||
971 | * This is a helper for setting golden registers. | |||
972 | */ | |||
973 | void amdgpu_device_program_register_sequence(struct amdgpu_device *adev, | |||
974 | const u32 *registers, | |||
975 | const u32 array_size) | |||
976 | { | |||
977 | u32 tmp, reg, and_mask, or_mask; | |||
978 | int i; | |||
979 | ||||
980 | if (array_size % 3) | |||
981 | return; | |||
982 | ||||
983 | for (i = 0; i < array_size; i +=3) { | |||
984 | reg = registers[i + 0]; | |||
985 | and_mask = registers[i + 1]; | |||
986 | or_mask = registers[i + 2]; | |||
987 | ||||
988 | if (and_mask == 0xffffffff) { | |||
989 | tmp = or_mask; | |||
990 | } else { | |||
991 | tmp = RREG32(reg)amdgpu_device_rreg(adev, (reg), 0); | |||
992 | tmp &= ~and_mask; | |||
993 | if (adev->family >= AMDGPU_FAMILY_AI141) | |||
994 | tmp |= (or_mask & and_mask); | |||
995 | else | |||
996 | tmp |= or_mask; | |||
997 | } | |||
998 | WREG32(reg, tmp)amdgpu_device_wreg(adev, (reg), (tmp), 0); | |||
999 | } | |||
1000 | } | |||
1001 | ||||
1002 | /** | |||
1003 | * amdgpu_device_pci_config_reset - reset the GPU | |||
1004 | * | |||
1005 | * @adev: amdgpu_device pointer | |||
1006 | * | |||
1007 | * Resets the GPU using the pci config reset sequence. | |||
1008 | * Only applicable to asics prior to vega10. | |||
1009 | */ | |||
1010 | void amdgpu_device_pci_config_reset(struct amdgpu_device *adev) | |||
1011 | { | |||
1012 | pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA0x39d5e86b); | |||
1013 | } | |||
1014 | ||||
1015 | /** | |||
1016 | * amdgpu_device_pci_reset - reset the GPU using generic PCI means | |||
1017 | * | |||
1018 | * @adev: amdgpu_device pointer | |||
1019 | * | |||
1020 | * Resets the GPU using generic pci reset interfaces (FLR, SBR, etc.). | |||
1021 | */ | |||
1022 | int amdgpu_device_pci_reset(struct amdgpu_device *adev) | |||
1023 | { | |||
1024 | STUB()do { printf("%s: stub\n", __func__); } while(0); | |||
1025 | return -ENOSYS78; | |||
1026 | #ifdef notyet | |||
1027 | return pci_reset_function(adev->pdev); | |||
1028 | #endif | |||
1029 | } | |||
1030 | ||||
1031 | /* | |||
1032 | * GPU doorbell aperture helpers function. | |||
1033 | */ | |||
1034 | /** | |||
1035 | * amdgpu_device_doorbell_init - Init doorbell driver information. | |||
1036 | * | |||
1037 | * @adev: amdgpu_device pointer | |||
1038 | * | |||
1039 | * Init doorbell driver information (CIK) | |||
1040 | * Returns 0 on success, error on failure. | |||
1041 | */ | |||
1042 | static int amdgpu_device_doorbell_init(struct amdgpu_device *adev) | |||
1043 | { | |||
1044 | ||||
1045 | /* No doorbell on SI hardware generation */ | |||
1046 | if (adev->asic_type < CHIP_BONAIRE) { | |||
1047 | adev->doorbell.base = 0; | |||
1048 | adev->doorbell.size = 0; | |||
1049 | adev->doorbell.num_doorbells = 0; | |||
1050 | adev->doorbell.ptr = NULL((void *)0); | |||
1051 | return 0; | |||
1052 | } | |||
1053 | ||||
1054 | #ifdef __linux__ | |||
1055 | if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET) | |||
1056 | return -EINVAL22; | |||
1057 | #endif | |||
1058 | ||||
1059 | amdgpu_asic_init_doorbell_index(adev)(adev)->asic_funcs->init_doorbell_index((adev)); | |||
1060 | ||||
1061 | /* doorbell bar mapping */ | |||
1062 | #ifdef __linux__ | |||
1063 | adev->doorbell.base = pci_resource_start(adev->pdev, 2); | |||
1064 | adev->doorbell.size = pci_resource_len(adev->pdev, 2); | |||
1065 | #endif | |||
1066 | ||||
1067 | if (adev->enable_mes) { | |||
1068 | adev->doorbell.num_doorbells = | |||
1069 | adev->doorbell.size / sizeof(u32); | |||
1070 | } else { | |||
1071 | adev->doorbell.num_doorbells = | |||
1072 | min_t(u32, adev->doorbell.size / sizeof(u32),({ u32 __min_a = (adev->doorbell.size / sizeof(u32)); u32 __min_b = (adev->doorbell_index.max_assignment+1); __min_a < __min_b ? __min_a : __min_b; }) | |||
1073 | adev->doorbell_index.max_assignment+1)({ u32 __min_a = (adev->doorbell.size / sizeof(u32)); u32 __min_b = (adev->doorbell_index.max_assignment+1); __min_a < __min_b ? __min_a : __min_b; }); | |||
1074 | if (adev->doorbell.num_doorbells == 0) | |||
1075 | return -EINVAL22; | |||
1076 | ||||
1077 | /* For Vega, reserve and map two pages on doorbell BAR since SDMA | |||
1078 | * paging queue doorbell use the second page. The | |||
1079 | * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the | |||
1080 | * doorbells are in the first page. So with paging queue enabled, | |||
1081 | * the max num_doorbells should + 1 page (0x400 in dword) | |||
1082 | */ | |||
1083 | if (adev->asic_type >= CHIP_VEGA10) | |||
1084 | adev->doorbell.num_doorbells += 0x400; | |||
1085 | } | |||
1086 | ||||
1087 | #ifdef __linux__ | |||
1088 | adev->doorbell.ptr = ioremap(adev->doorbell.base, | |||
1089 | adev->doorbell.num_doorbells * | |||
1090 | sizeof(u32)); | |||
1091 | if (adev->doorbell.ptr == NULL((void *)0)) | |||
1092 | return -ENOMEM12; | |||
1093 | #endif | |||
1094 | ||||
1095 | return 0; | |||
1096 | } | |||
1097 | ||||
1098 | /** | |||
1099 | * amdgpu_device_doorbell_fini - Tear down doorbell driver information. | |||
1100 | * | |||
1101 | * @adev: amdgpu_device pointer | |||
1102 | * | |||
1103 | * Tear down doorbell driver information (CIK) | |||
1104 | */ | |||
1105 | static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev) | |||
1106 | { | |||
1107 | #ifdef __linux__ | |||
1108 | iounmap(adev->doorbell.ptr); | |||
1109 | #else | |||
1110 | if (adev->doorbell.size > 0) | |||
1111 | bus_space_unmap(adev->doorbell.bst, adev->doorbell.bsh, | |||
1112 | adev->doorbell.size); | |||
1113 | #endif | |||
1114 | adev->doorbell.ptr = NULL((void *)0); | |||
1115 | } | |||
1116 | ||||
1117 | ||||
1118 | ||||
1119 | /* | |||
1120 | * amdgpu_device_wb_*() | |||
1121 | * Writeback is the method by which the GPU updates special pages in memory | |||
1122 | * with the status of certain GPU events (fences, ring pointers,etc.). | |||
1123 | */ | |||
1124 | ||||
1125 | /** | |||
1126 | * amdgpu_device_wb_fini - Disable Writeback and free memory | |||
1127 | * | |||
1128 | * @adev: amdgpu_device pointer | |||
1129 | * | |||
1130 | * Disables Writeback and frees the Writeback memory (all asics). | |||
1131 | * Used at driver shutdown. | |||
1132 | */ | |||
1133 | static void amdgpu_device_wb_fini(struct amdgpu_device *adev) | |||
1134 | { | |||
1135 | if (adev->wb.wb_obj) { | |||
1136 | amdgpu_bo_free_kernel(&adev->wb.wb_obj, | |||
1137 | &adev->wb.gpu_addr, | |||
1138 | (void **)&adev->wb.wb); | |||
1139 | adev->wb.wb_obj = NULL((void *)0); | |||
1140 | } | |||
1141 | } | |||
1142 | ||||
1143 | /** | |||
1144 | * amdgpu_device_wb_init - Init Writeback driver info and allocate memory | |||
1145 | * | |||
1146 | * @adev: amdgpu_device pointer | |||
1147 | * | |||
1148 | * Initializes writeback and allocates writeback memory (all asics). | |||
1149 | * Used at driver startup. | |||
1150 | * Returns 0 on success or an -error on failure. | |||
1151 | */ | |||
1152 | static int amdgpu_device_wb_init(struct amdgpu_device *adev) | |||
1153 | { | |||
1154 | int r; | |||
1155 | ||||
1156 | if (adev->wb.wb_obj == NULL((void *)0)) { | |||
1157 | /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */ | |||
1158 | r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB256 * sizeof(uint32_t) * 8, | |||
1159 | PAGE_SIZE(1 << 12), AMDGPU_GEM_DOMAIN_GTT0x2, | |||
1160 | &adev->wb.wb_obj, &adev->wb.gpu_addr, | |||
1161 | (void **)&adev->wb.wb); | |||
1162 | if (r) { | |||
1163 | dev_warn(adev->dev, "(%d) create WB bo failed\n", r)printf("drm:pid%d:%s *WARNING* " "(%d) create WB bo failed\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , r); | |||
1164 | return r; | |||
1165 | } | |||
1166 | ||||
1167 | adev->wb.num_wb = AMDGPU_MAX_WB256; | |||
1168 | memset(&adev->wb.used, 0, sizeof(adev->wb.used))__builtin_memset((&adev->wb.used), (0), (sizeof(adev-> wb.used))); | |||
1169 | ||||
1170 | /* clear wb memory */ | |||
1171 | memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8)__builtin_memset(((char *)adev->wb.wb), (0), (256 * sizeof (uint32_t) * 8)); | |||
1172 | } | |||
1173 | ||||
1174 | return 0; | |||
1175 | } | |||
1176 | ||||
1177 | /** | |||
1178 | * amdgpu_device_wb_get - Allocate a wb entry | |||
1179 | * | |||
1180 | * @adev: amdgpu_device pointer | |||
1181 | * @wb: wb index | |||
1182 | * | |||
1183 | * Allocate a wb slot for use by the driver (all asics). | |||
1184 | * Returns 0 on success or -EINVAL on failure. | |||
1185 | */ | |||
1186 | int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb) | |||
1187 | { | |||
1188 | unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb); | |||
1189 | ||||
1190 | if (offset < adev->wb.num_wb) { | |||
1191 | __set_bit(offset, adev->wb.used); | |||
1192 | *wb = offset << 3; /* convert to dw offset */ | |||
1193 | return 0; | |||
1194 | } else { | |||
1195 | return -EINVAL22; | |||
1196 | } | |||
1197 | } | |||
1198 | ||||
1199 | /** | |||
1200 | * amdgpu_device_wb_free - Free a wb entry | |||
1201 | * | |||
1202 | * @adev: amdgpu_device pointer | |||
1203 | * @wb: wb index | |||
1204 | * | |||
1205 | * Free a wb slot allocated for use by the driver (all asics) | |||
1206 | */ | |||
1207 | void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb) | |||
1208 | { | |||
1209 | wb >>= 3; | |||
1210 | if (wb < adev->wb.num_wb) | |||
1211 | __clear_bit(wb, adev->wb.used); | |||
1212 | } | |||
1213 | ||||
1214 | /** | |||
1215 | * amdgpu_device_resize_fb_bar - try to resize FB BAR | |||
1216 | * | |||
1217 | * @adev: amdgpu_device pointer | |||
1218 | * | |||
1219 | * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not | |||
1220 | * to fail, but if any of the BARs is not accessible after the size we abort | |||
1221 | * driver loading by returning -ENODEV. | |||
1222 | */ | |||
1223 | int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev) | |||
1224 | { | |||
1225 | #ifdef __linux__ | |||
1226 | int rbar_size = pci_rebar_bytes_to_size(adev->gmc.real_vram_size); | |||
1227 | struct pci_bus *root; | |||
1228 | struct resource *res; | |||
1229 | unsigned i; | |||
1230 | u16 cmd; | |||
1231 | int r; | |||
1232 | ||||
1233 | if (!IS_ENABLED(CONFIG_PHYS_ADDR_T_64BIT)0) | |||
1234 | return 0; | |||
1235 | ||||
1236 | /* Bypass for VF */ | |||
1237 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) | |||
1238 | return 0; | |||
1239 | ||||
1240 | /* skip if the bios has already enabled large BAR */ | |||
1241 | if (adev->gmc.real_vram_size && | |||
1242 | (pci_resource_len(adev->pdev, 0) >= adev->gmc.real_vram_size)) | |||
1243 | return 0; | |||
1244 | ||||
1245 | /* Check if the root BUS has 64bit memory resources */ | |||
1246 | root = adev->pdev->bus; | |||
1247 | while (root->parent) | |||
1248 | root = root->parent; | |||
1249 | ||||
1250 | pci_bus_for_each_resource(root, res, i) { | |||
1251 | if (res && res->flags & (IORESOURCE_MEM0x0001 | IORESOURCE_MEM_64) && | |||
1252 | res->start > 0x100000000ull) | |||
1253 | break; | |||
1254 | } | |||
1255 | ||||
1256 | /* Trying to resize is pointless without a root hub window above 4GB */ | |||
1257 | if (!res) | |||
1258 | return 0; | |||
1259 | ||||
1260 | /* Limit the BAR size to what is available */ | |||
1261 | rbar_size = min(fls(pci_rebar_get_possible_sizes(adev->pdev, 0)) - 1,(((fls(pci_rebar_get_possible_sizes(adev->pdev, 0)) - 1)< (rbar_size))?(fls(pci_rebar_get_possible_sizes(adev->pdev, 0)) - 1):(rbar_size)) | |||
1262 | rbar_size)(((fls(pci_rebar_get_possible_sizes(adev->pdev, 0)) - 1)< (rbar_size))?(fls(pci_rebar_get_possible_sizes(adev->pdev, 0)) - 1):(rbar_size)); | |||
1263 | ||||
1264 | /* Disable memory decoding while we change the BAR addresses and size */ | |||
1265 | pci_read_config_word(adev->pdev, PCI_COMMAND0x04, &cmd); | |||
1266 | pci_write_config_word(adev->pdev, PCI_COMMAND0x04, | |||
1267 | cmd & ~PCI_COMMAND_MEMORY0x00000002); | |||
1268 | ||||
1269 | /* Free the VRAM and doorbell BAR, we most likely need to move both. */ | |||
1270 | amdgpu_device_doorbell_fini(adev); | |||
1271 | if (adev->asic_type >= CHIP_BONAIRE) | |||
1272 | pci_release_resource(adev->pdev, 2); | |||
1273 | ||||
1274 | pci_release_resource(adev->pdev, 0); | |||
1275 | ||||
1276 | r = pci_resize_resource(adev->pdev, 0, rbar_size); | |||
1277 | if (r == -ENOSPC28) | |||
1278 | DRM_INFO("Not enough PCI address space for a large BAR.")printk("\0016" "[" "drm" "] " "Not enough PCI address space for a large BAR." ); | |||
1279 | else if (r && r != -ENOTSUPP91) | |||
1280 | DRM_ERROR("Problem resizing BAR0 (%d).", r)__drm_err("Problem resizing BAR0 (%d).", r); | |||
1281 | ||||
1282 | pci_assign_unassigned_bus_resources(adev->pdev->bus); | |||
1283 | ||||
1284 | /* When the doorbell or fb BAR isn't available we have no chance of | |||
1285 | * using the device. | |||
1286 | */ | |||
1287 | r = amdgpu_device_doorbell_init(adev); | |||
1288 | if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET)) | |||
1289 | return -ENODEV19; | |||
1290 | ||||
1291 | pci_write_config_word(adev->pdev, PCI_COMMAND0x04, cmd); | |||
1292 | #endif /* __linux__ */ | |||
1293 | ||||
1294 | return 0; | |||
1295 | } | |||
1296 | ||||
1297 | /* | |||
1298 | * GPU helpers function. | |||
1299 | */ | |||
1300 | /** | |||
1301 | * amdgpu_device_need_post - check if the hw need post or not | |||
1302 | * | |||
1303 | * @adev: amdgpu_device pointer | |||
1304 | * | |||
1305 | * Check if the asic has been initialized (all asics) at driver startup | |||
1306 | * or post is needed if hw reset is performed. | |||
1307 | * Returns true if need or false if not. | |||
1308 | */ | |||
1309 | bool_Bool amdgpu_device_need_post(struct amdgpu_device *adev) | |||
1310 | { | |||
1311 | uint32_t reg; | |||
1312 | ||||
1313 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) | |||
1314 | return false0; | |||
1315 | ||||
1316 | if (amdgpu_passthrough(adev)((adev)->virt.caps & (1 << 3))) { | |||
1317 | /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot | |||
1318 | * some old smc fw still need driver do vPost otherwise gpu hang, while | |||
1319 | * those smc fw version above 22.15 doesn't have this flaw, so we force | |||
1320 | * vpost executed for smc version below 22.15 | |||
1321 | */ | |||
1322 | if (adev->asic_type == CHIP_FIJI) { | |||
1323 | int err; | |||
1324 | uint32_t fw_ver; | |||
1325 | err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev); | |||
1326 | /* force vPost if error occured */ | |||
1327 | if (err) | |||
1328 | return true1; | |||
1329 | ||||
1330 | fw_ver = *((uint32_t *)adev->pm.fw->data + 69); | |||
1331 | if (fw_ver < 0x00160e00) | |||
1332 | return true1; | |||
1333 | } | |||
1334 | } | |||
1335 | ||||
1336 | /* Don't post if we need to reset whole hive on init */ | |||
1337 | if (adev->gmc.xgmi.pending_reset) | |||
1338 | return false0; | |||
1339 | ||||
1340 | if (adev->has_hw_reset) { | |||
1341 | adev->has_hw_reset = false0; | |||
1342 | return true1; | |||
1343 | } | |||
1344 | ||||
1345 | /* bios scratch used on CIK+ */ | |||
1346 | if (adev->asic_type >= CHIP_BONAIRE) | |||
1347 | return amdgpu_atombios_scratch_need_asic_init(adev); | |||
1348 | ||||
1349 | /* check MEM_SIZE for older asics */ | |||
1350 | reg = amdgpu_asic_get_config_memsize(adev)(adev)->asic_funcs->get_config_memsize((adev)); | |||
1351 | ||||
1352 | if ((reg != 0) && (reg != 0xffffffff)) | |||
1353 | return false0; | |||
1354 | ||||
1355 | return true1; | |||
1356 | } | |||
1357 | ||||
1358 | /* | |||
1359 | * Intel hosts such as Raptor Lake and Sapphire Rapids don't support dynamic | |||
1360 | * speed switching. Until we have confirmation from Intel that a specific host | |||
1361 | * supports it, it's safer that we keep it disabled for all. | |||
1362 | * | |||
1363 | * https://edc.intel.com/content/www/us/en/design/products/platforms/details/raptor-lake-s/13th-generation-core-processors-datasheet-volume-1-of-2/005/pci-express-support/ | |||
1364 | * https://gitlab.freedesktop.org/drm/amd/-/issues/2663 | |||
1365 | */ | |||
1366 | bool_Bool amdgpu_device_pcie_dynamic_switching_supported(void) | |||
1367 | { | |||
1368 | #if IS_ENABLED(CONFIG_X86)1 | |||
1369 | #ifdef __linux__ | |||
1370 | struct cpuinfo_x86 *c = &cpu_data(0); | |||
1371 | ||||
1372 | if (c->x86_vendor == X86_VENDOR_INTEL) | |||
1373 | #else | |||
1374 | if (strcmp(cpu_vendor, "GenuineIntel") == 0) | |||
1375 | #endif | |||
1376 | return false0; | |||
1377 | #endif | |||
1378 | return true1; | |||
1379 | } | |||
1380 | ||||
1381 | /** | |||
1382 | * amdgpu_device_should_use_aspm - check if the device should program ASPM | |||
1383 | * | |||
1384 | * @adev: amdgpu_device pointer | |||
1385 | * | |||
1386 | * Confirm whether the module parameter and pcie bridge agree that ASPM should | |||
1387 | * be set for this device. | |||
1388 | * | |||
1389 | * Returns true if it should be used or false if not. | |||
1390 | */ | |||
1391 | bool_Bool amdgpu_device_should_use_aspm(struct amdgpu_device *adev) | |||
1392 | { | |||
1393 | switch (amdgpu_aspm) { | |||
1394 | case -1: | |||
1395 | break; | |||
1396 | case 0: | |||
1397 | return false0; | |||
1398 | case 1: | |||
1399 | return true1; | |||
1400 | default: | |||
1401 | return false0; | |||
1402 | } | |||
1403 | return pcie_aspm_enabled(adev->pdev); | |||
1404 | } | |||
1405 | ||||
1406 | bool_Bool amdgpu_device_aspm_support_quirk(void) | |||
1407 | { | |||
1408 | #if IS_ENABLED(CONFIG_X86)1 | |||
1409 | struct cpu_info *ci = curcpu()({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;}); | |||
1410 | ||||
1411 | return !(ci->ci_family == 6 && ci->ci_model == 0x97); | |||
1412 | #else | |||
1413 | return true1; | |||
1414 | #endif | |||
1415 | } | |||
1416 | ||||
1417 | /* if we get transitioned to only one device, take VGA back */ | |||
1418 | /** | |||
1419 | * amdgpu_device_vga_set_decode - enable/disable vga decode | |||
1420 | * | |||
1421 | * @pdev: PCI device pointer | |||
1422 | * @state: enable/disable vga decode | |||
1423 | * | |||
1424 | * Enable/disable vga decode (all asics). | |||
1425 | * Returns VGA resource flags. | |||
1426 | */ | |||
1427 | #ifdef notyet | |||
1428 | static unsigned int amdgpu_device_vga_set_decode(struct pci_dev *pdev, | |||
1429 | bool_Bool state) | |||
1430 | { | |||
1431 | struct amdgpu_device *adev = drm_to_adev(pci_get_drvdata(pdev)); | |||
1432 | amdgpu_asic_set_vga_state(adev, state)(adev)->asic_funcs->set_vga_state((adev), (state)); | |||
1433 | if (state) | |||
1434 | return VGA_RSRC_LEGACY_IO0x01 | VGA_RSRC_LEGACY_MEM | | |||
1435 | VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM; | |||
1436 | else | |||
1437 | return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM; | |||
1438 | } | |||
1439 | #endif | |||
1440 | ||||
1441 | /** | |||
1442 | * amdgpu_device_check_block_size - validate the vm block size | |||
1443 | * | |||
1444 | * @adev: amdgpu_device pointer | |||
1445 | * | |||
1446 | * Validates the vm block size specified via module parameter. | |||
1447 | * The vm block size defines number of bits in page table versus page directory, | |||
1448 | * a page is 4KB so we have 12 bits offset, minimum 9 bits in the | |||
1449 | * page table and the remaining bits are in the page directory. | |||
1450 | */ | |||
1451 | static void amdgpu_device_check_block_size(struct amdgpu_device *adev) | |||
1452 | { | |||
1453 | /* defines number of bits in page table versus page directory, | |||
1454 | * a page is 4KB so we have 12 bits offset, minimum 9 bits in the | |||
1455 | * page table and the remaining bits are in the page directory */ | |||
1456 | if (amdgpu_vm_block_size == -1) | |||
1457 | return; | |||
1458 | ||||
1459 | if (amdgpu_vm_block_size < 9) { | |||
1460 | dev_warn(adev->dev, "VM page table size (%d) too small\n",printf("drm:pid%d:%s *WARNING* " "VM page table size (%d) too small\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , amdgpu_vm_block_size ) | |||
1461 | amdgpu_vm_block_size)printf("drm:pid%d:%s *WARNING* " "VM page table size (%d) too small\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , amdgpu_vm_block_size ); | |||
1462 | amdgpu_vm_block_size = -1; | |||
1463 | } | |||
1464 | } | |||
1465 | ||||
1466 | /** | |||
1467 | * amdgpu_device_check_vm_size - validate the vm size | |||
1468 | * | |||
1469 | * @adev: amdgpu_device pointer | |||
1470 | * | |||
1471 | * Validates the vm size in GB specified via module parameter. | |||
1472 | * The VM size is the size of the GPU virtual memory space in GB. | |||
1473 | */ | |||
1474 | static void amdgpu_device_check_vm_size(struct amdgpu_device *adev) | |||
1475 | { | |||
1476 | /* no need to check the default value */ | |||
1477 | if (amdgpu_vm_size == -1) | |||
1478 | return; | |||
1479 | ||||
1480 | if (amdgpu_vm_size < 1) { | |||
1481 | dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",printf("drm:pid%d:%s *WARNING* " "VM size (%d) too small, min is 1GB\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , amdgpu_vm_size ) | |||
1482 | amdgpu_vm_size)printf("drm:pid%d:%s *WARNING* " "VM size (%d) too small, min is 1GB\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , amdgpu_vm_size ); | |||
1483 | amdgpu_vm_size = -1; | |||
1484 | } | |||
1485 | } | |||
1486 | ||||
1487 | static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev) | |||
1488 | { | |||
1489 | #ifdef __linux__ | |||
1490 | struct sysinfo si; | |||
1491 | #endif | |||
1492 | bool_Bool is_os_64 = (sizeof(void *) == 8); | |||
1493 | uint64_t total_memory; | |||
1494 | uint64_t dram_size_seven_GB = 0x1B8000000; | |||
1495 | uint64_t dram_size_three_GB = 0xB8000000; | |||
1496 | ||||
1497 | if (amdgpu_smu_memory_pool_size == 0) | |||
1498 | return; | |||
1499 | ||||
1500 | if (!is_os_64) { | |||
1501 | DRM_WARN("Not 64-bit OS, feature not supported\n")printk("\0014" "[" "drm" "] " "Not 64-bit OS, feature not supported\n" ); | |||
1502 | goto def_value; | |||
1503 | } | |||
1504 | #ifdef __linux__ | |||
1505 | si_meminfo(&si); | |||
1506 | total_memory = (uint64_t)si.totalram * si.mem_unit; | |||
1507 | #else | |||
1508 | total_memory = ptoa(physmem)((paddr_t)(physmem) << 12); | |||
1509 | #endif | |||
1510 | ||||
1511 | if ((amdgpu_smu_memory_pool_size == 1) || | |||
1512 | (amdgpu_smu_memory_pool_size == 2)) { | |||
1513 | if (total_memory < dram_size_three_GB) | |||
1514 | goto def_value1; | |||
1515 | } else if ((amdgpu_smu_memory_pool_size == 4) || | |||
1516 | (amdgpu_smu_memory_pool_size == 8)) { | |||
1517 | if (total_memory < dram_size_seven_GB) | |||
1518 | goto def_value1; | |||
1519 | } else { | |||
1520 | DRM_WARN("Smu memory pool size not supported\n")printk("\0014" "[" "drm" "] " "Smu memory pool size not supported\n" ); | |||
1521 | goto def_value; | |||
1522 | } | |||
1523 | adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28; | |||
1524 | ||||
1525 | return; | |||
1526 | ||||
1527 | def_value1: | |||
1528 | DRM_WARN("No enough system memory\n")printk("\0014" "[" "drm" "] " "No enough system memory\n"); | |||
1529 | def_value: | |||
1530 | adev->pm.smu_prv_buffer_size = 0; | |||
1531 | } | |||
1532 | ||||
1533 | static int amdgpu_device_init_apu_flags(struct amdgpu_device *adev) | |||
1534 | { | |||
1535 | if (!(adev->flags & AMD_IS_APU) || | |||
1536 | adev->asic_type < CHIP_RAVEN) | |||
1537 | return 0; | |||
1538 | ||||
1539 | switch (adev->asic_type) { | |||
1540 | case CHIP_RAVEN: | |||
1541 | if (adev->pdev->device == 0x15dd) | |||
1542 | adev->apu_flags |= AMD_APU_IS_RAVEN; | |||
1543 | if (adev->pdev->device == 0x15d8) | |||
1544 | adev->apu_flags |= AMD_APU_IS_PICASSO; | |||
1545 | break; | |||
1546 | case CHIP_RENOIR: | |||
1547 | if ((adev->pdev->device == 0x1636) || | |||
1548 | (adev->pdev->device == 0x164c)) | |||
1549 | adev->apu_flags |= AMD_APU_IS_RENOIR; | |||
1550 | else | |||
1551 | adev->apu_flags |= AMD_APU_IS_GREEN_SARDINE; | |||
1552 | break; | |||
1553 | case CHIP_VANGOGH: | |||
1554 | adev->apu_flags |= AMD_APU_IS_VANGOGH; | |||
1555 | break; | |||
1556 | case CHIP_YELLOW_CARP: | |||
1557 | break; | |||
1558 | case CHIP_CYAN_SKILLFISH: | |||
1559 | if ((adev->pdev->device == 0x13FE) || | |||
1560 | (adev->pdev->device == 0x143F)) | |||
1561 | adev->apu_flags |= AMD_APU_IS_CYAN_SKILLFISH2; | |||
1562 | break; | |||
1563 | default: | |||
1564 | break; | |||
1565 | } | |||
1566 | ||||
1567 | return 0; | |||
1568 | } | |||
1569 | ||||
1570 | /** | |||
1571 | * amdgpu_device_check_arguments - validate module params | |||
1572 | * | |||
1573 | * @adev: amdgpu_device pointer | |||
1574 | * | |||
1575 | * Validates certain module parameters and updates | |||
1576 | * the associated values used by the driver (all asics). | |||
1577 | */ | |||
1578 | static int amdgpu_device_check_arguments(struct amdgpu_device *adev) | |||
1579 | { | |||
1580 | if (amdgpu_sched_jobs < 4) { | |||
1581 | dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",printf("drm:pid%d:%s *WARNING* " "sched jobs (%d) must be at least 4\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , amdgpu_sched_jobs ) | |||
1582 | amdgpu_sched_jobs)printf("drm:pid%d:%s *WARNING* " "sched jobs (%d) must be at least 4\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , amdgpu_sched_jobs ); | |||
1583 | amdgpu_sched_jobs = 4; | |||
1584 | } else if (!is_power_of_2(amdgpu_sched_jobs)(((amdgpu_sched_jobs) != 0) && (((amdgpu_sched_jobs) - 1) & (amdgpu_sched_jobs)) == 0)){ | |||
1585 | dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",printf("drm:pid%d:%s *WARNING* " "sched jobs (%d) must be a power of 2\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , amdgpu_sched_jobs ) | |||
1586 | amdgpu_sched_jobs)printf("drm:pid%d:%s *WARNING* " "sched jobs (%d) must be a power of 2\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , amdgpu_sched_jobs ); | |||
1587 | amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs); | |||
1588 | } | |||
1589 | ||||
1590 | if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) { | |||
1591 | /* gart size must be greater or equal to 32M */ | |||
1592 | dev_warn(adev->dev, "gart size (%d) too small\n",printf("drm:pid%d:%s *WARNING* " "gart size (%d) too small\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , amdgpu_gart_size ) | |||
1593 | amdgpu_gart_size)printf("drm:pid%d:%s *WARNING* " "gart size (%d) too small\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , amdgpu_gart_size ); | |||
1594 | amdgpu_gart_size = -1; | |||
1595 | } | |||
1596 | ||||
1597 | if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) { | |||
1598 | /* gtt size must be greater or equal to 32M */ | |||
1599 | dev_warn(adev->dev, "gtt size (%d) too small\n",printf("drm:pid%d:%s *WARNING* " "gtt size (%d) too small\n", ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , amdgpu_gtt_size ) | |||
1600 | amdgpu_gtt_size)printf("drm:pid%d:%s *WARNING* " "gtt size (%d) too small\n", ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , amdgpu_gtt_size ); | |||
1601 | amdgpu_gtt_size = -1; | |||
1602 | } | |||
1603 | ||||
1604 | /* valid range is between 4 and 9 inclusive */ | |||
1605 | if (amdgpu_vm_fragment_size != -1 && | |||
1606 | (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) { | |||
1607 | dev_warn(adev->dev, "valid range is between 4 and 9\n")printf("drm:pid%d:%s *WARNING* " "valid range is between 4 and 9\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
1608 | amdgpu_vm_fragment_size = -1; | |||
1609 | } | |||
1610 | ||||
1611 | if (amdgpu_sched_hw_submission < 2) { | |||
1612 | dev_warn(adev->dev, "sched hw submission jobs (%d) must be at least 2\n",printf("drm:pid%d:%s *WARNING* " "sched hw submission jobs (%d) must be at least 2\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , amdgpu_sched_hw_submission ) | |||
1613 | amdgpu_sched_hw_submission)printf("drm:pid%d:%s *WARNING* " "sched hw submission jobs (%d) must be at least 2\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , amdgpu_sched_hw_submission ); | |||
1614 | amdgpu_sched_hw_submission = 2; | |||
1615 | } else if (!is_power_of_2(amdgpu_sched_hw_submission)(((amdgpu_sched_hw_submission) != 0) && (((amdgpu_sched_hw_submission ) - 1) & (amdgpu_sched_hw_submission)) == 0)) { | |||
1616 | dev_warn(adev->dev, "sched hw submission jobs (%d) must be a power of 2\n",printf("drm:pid%d:%s *WARNING* " "sched hw submission jobs (%d) must be a power of 2\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , amdgpu_sched_hw_submission ) | |||
1617 | amdgpu_sched_hw_submission)printf("drm:pid%d:%s *WARNING* " "sched hw submission jobs (%d) must be a power of 2\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , amdgpu_sched_hw_submission ); | |||
1618 | amdgpu_sched_hw_submission = roundup_pow_of_two(amdgpu_sched_hw_submission); | |||
1619 | } | |||
1620 | ||||
1621 | if (amdgpu_reset_method < -1 || amdgpu_reset_method > 4) { | |||
1622 | dev_warn(adev->dev, "invalid option for reset method, reverting to default\n")printf("drm:pid%d:%s *WARNING* " "invalid option for reset method, reverting to default\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
1623 | amdgpu_reset_method = -1; | |||
1624 | } | |||
1625 | ||||
1626 | amdgpu_device_check_smu_prv_buffer_size(adev); | |||
1627 | ||||
1628 | amdgpu_device_check_vm_size(adev); | |||
1629 | ||||
1630 | amdgpu_device_check_block_size(adev); | |||
1631 | ||||
1632 | adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type); | |||
1633 | ||||
1634 | return 0; | |||
1635 | } | |||
1636 | ||||
1637 | #ifdef __linux__ | |||
1638 | /** | |||
1639 | * amdgpu_switcheroo_set_state - set switcheroo state | |||
1640 | * | |||
1641 | * @pdev: pci dev pointer | |||
1642 | * @state: vga_switcheroo state | |||
1643 | * | |||
1644 | * Callback for the switcheroo driver. Suspends or resumes the | |||
1645 | * the asics before or after it is powered up using ACPI methods. | |||
1646 | */ | |||
1647 | static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, | |||
1648 | enum vga_switcheroo_state state) | |||
1649 | { | |||
1650 | struct drm_device *dev = pci_get_drvdata(pdev); | |||
1651 | int r; | |||
1652 | ||||
1653 | if (amdgpu_device_supports_px(dev) && state == VGA_SWITCHEROO_OFF) | |||
1654 | return; | |||
1655 | ||||
1656 | if (state == VGA_SWITCHEROO_ON) { | |||
1657 | pr_info("switched on\n")do { } while(0); | |||
1658 | /* don't suspend or resume card normally */ | |||
1659 | dev->switch_power_state = DRM_SWITCH_POWER_CHANGING; | |||
1660 | ||||
1661 | pci_set_power_state(pdev, PCI_D0); | |||
1662 | amdgpu_device_load_pci_state(pdev); | |||
1663 | r = pci_enable_device(pdev); | |||
1664 | if (r) | |||
1665 | DRM_WARN("pci_enable_device failed (%d)\n", r)printk("\0014" "[" "drm" "] " "pci_enable_device failed (%d)\n" , r); | |||
1666 | amdgpu_device_resume(dev, true1); | |||
1667 | ||||
1668 | dev->switch_power_state = DRM_SWITCH_POWER_ON; | |||
1669 | } else { | |||
1670 | pr_info("switched off\n")do { } while(0); | |||
1671 | dev->switch_power_state = DRM_SWITCH_POWER_CHANGING; | |||
1672 | amdgpu_device_suspend(dev, true1); | |||
1673 | amdgpu_device_cache_pci_state(pdev); | |||
1674 | /* Shut down the device */ | |||
1675 | pci_disable_device(pdev); | |||
1676 | pci_set_power_state(pdev, PCI_D3cold); | |||
1677 | dev->switch_power_state = DRM_SWITCH_POWER_OFF; | |||
1678 | } | |||
1679 | } | |||
1680 | ||||
1681 | /** | |||
1682 | * amdgpu_switcheroo_can_switch - see if switcheroo state can change | |||
1683 | * | |||
1684 | * @pdev: pci dev pointer | |||
1685 | * | |||
1686 | * Callback for the switcheroo driver. Check of the switcheroo | |||
1687 | * state can be changed. | |||
1688 | * Returns true if the state can be changed, false if not. | |||
1689 | */ | |||
1690 | static bool_Bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev) | |||
1691 | { | |||
1692 | struct drm_device *dev = pci_get_drvdata(pdev); | |||
1693 | ||||
1694 | /* | |||
1695 | * FIXME: open_count is protected by drm_global_mutex but that would lead to | |||
1696 | * locking inversion with the driver load path. And the access here is | |||
1697 | * completely racy anyway. So don't bother with locking for now. | |||
1698 | */ | |||
1699 | return atomic_read(&dev->open_count)({ typeof(*(&dev->open_count)) __tmp = *(volatile typeof (*(&dev->open_count)) *)&(*(&dev->open_count )); membar_datadep_consumer(); __tmp; }) == 0; | |||
1700 | } | |||
1701 | #endif /* __linux__ */ | |||
1702 | ||||
1703 | static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = { | |||
1704 | #ifdef notyet | |||
1705 | .set_gpu_state = amdgpu_switcheroo_set_state, | |||
1706 | .reprobe = NULL((void *)0), | |||
1707 | .can_switch = amdgpu_switcheroo_can_switch, | |||
1708 | #endif | |||
1709 | }; | |||
1710 | ||||
1711 | /** | |||
1712 | * amdgpu_device_ip_set_clockgating_state - set the CG state | |||
1713 | * | |||
1714 | * @dev: amdgpu_device pointer | |||
1715 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) | |||
1716 | * @state: clockgating state (gate or ungate) | |||
1717 | * | |||
1718 | * Sets the requested clockgating state for all instances of | |||
1719 | * the hardware IP specified. | |||
1720 | * Returns the error code from the last instance. | |||
1721 | */ | |||
1722 | int amdgpu_device_ip_set_clockgating_state(void *dev, | |||
1723 | enum amd_ip_block_type block_type, | |||
1724 | enum amd_clockgating_state state) | |||
1725 | { | |||
1726 | struct amdgpu_device *adev = dev; | |||
1727 | int i, r = 0; | |||
1728 | ||||
1729 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
1730 | if (!adev->ip_blocks[i].status.valid) | |||
1731 | continue; | |||
1732 | if (adev->ip_blocks[i].version->type != block_type) | |||
1733 | continue; | |||
1734 | if (!adev->ip_blocks[i].version->funcs->set_clockgating_state) | |||
1735 | continue; | |||
1736 | r = adev->ip_blocks[i].version->funcs->set_clockgating_state( | |||
1737 | (void *)adev, state); | |||
1738 | if (r) | |||
1739 | DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",__drm_err("set_clockgating_state of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r) | |||
1740 | adev->ip_blocks[i].version->funcs->name, r)__drm_err("set_clockgating_state of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r); | |||
1741 | } | |||
1742 | return r; | |||
1743 | } | |||
1744 | ||||
1745 | /** | |||
1746 | * amdgpu_device_ip_set_powergating_state - set the PG state | |||
1747 | * | |||
1748 | * @dev: amdgpu_device pointer | |||
1749 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) | |||
1750 | * @state: powergating state (gate or ungate) | |||
1751 | * | |||
1752 | * Sets the requested powergating state for all instances of | |||
1753 | * the hardware IP specified. | |||
1754 | * Returns the error code from the last instance. | |||
1755 | */ | |||
1756 | int amdgpu_device_ip_set_powergating_state(void *dev, | |||
1757 | enum amd_ip_block_type block_type, | |||
1758 | enum amd_powergating_state state) | |||
1759 | { | |||
1760 | struct amdgpu_device *adev = dev; | |||
1761 | int i, r = 0; | |||
1762 | ||||
1763 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
1764 | if (!adev->ip_blocks[i].status.valid) | |||
1765 | continue; | |||
1766 | if (adev->ip_blocks[i].version->type != block_type) | |||
1767 | continue; | |||
1768 | if (!adev->ip_blocks[i].version->funcs->set_powergating_state) | |||
1769 | continue; | |||
1770 | r = adev->ip_blocks[i].version->funcs->set_powergating_state( | |||
1771 | (void *)adev, state); | |||
1772 | if (r) | |||
1773 | DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",__drm_err("set_powergating_state of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r) | |||
1774 | adev->ip_blocks[i].version->funcs->name, r)__drm_err("set_powergating_state of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r); | |||
1775 | } | |||
1776 | return r; | |||
1777 | } | |||
1778 | ||||
1779 | /** | |||
1780 | * amdgpu_device_ip_get_clockgating_state - get the CG state | |||
1781 | * | |||
1782 | * @adev: amdgpu_device pointer | |||
1783 | * @flags: clockgating feature flags | |||
1784 | * | |||
1785 | * Walks the list of IPs on the device and updates the clockgating | |||
1786 | * flags for each IP. | |||
1787 | * Updates @flags with the feature flags for each hardware IP where | |||
1788 | * clockgating is enabled. | |||
1789 | */ | |||
1790 | void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev, | |||
1791 | u64 *flags) | |||
1792 | { | |||
1793 | int i; | |||
1794 | ||||
1795 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
1796 | if (!adev->ip_blocks[i].status.valid) | |||
1797 | continue; | |||
1798 | if (adev->ip_blocks[i].version->funcs->get_clockgating_state) | |||
1799 | adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags); | |||
1800 | } | |||
1801 | } | |||
1802 | ||||
1803 | /** | |||
1804 | * amdgpu_device_ip_wait_for_idle - wait for idle | |||
1805 | * | |||
1806 | * @adev: amdgpu_device pointer | |||
1807 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) | |||
1808 | * | |||
1809 | * Waits for the request hardware IP to be idle. | |||
1810 | * Returns 0 for success or a negative error code on failure. | |||
1811 | */ | |||
1812 | int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev, | |||
1813 | enum amd_ip_block_type block_type) | |||
1814 | { | |||
1815 | int i, r; | |||
1816 | ||||
1817 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
1818 | if (!adev->ip_blocks[i].status.valid) | |||
1819 | continue; | |||
1820 | if (adev->ip_blocks[i].version->type == block_type) { | |||
1821 | r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev); | |||
1822 | if (r) | |||
1823 | return r; | |||
1824 | break; | |||
1825 | } | |||
1826 | } | |||
1827 | return 0; | |||
1828 | ||||
1829 | } | |||
1830 | ||||
1831 | /** | |||
1832 | * amdgpu_device_ip_is_idle - is the hardware IP idle | |||
1833 | * | |||
1834 | * @adev: amdgpu_device pointer | |||
1835 | * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) | |||
1836 | * | |||
1837 | * Check if the hardware IP is idle or not. | |||
1838 | * Returns true if it the IP is idle, false if not. | |||
1839 | */ | |||
1840 | bool_Bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev, | |||
1841 | enum amd_ip_block_type block_type) | |||
1842 | { | |||
1843 | int i; | |||
1844 | ||||
1845 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
1846 | if (!adev->ip_blocks[i].status.valid) | |||
1847 | continue; | |||
1848 | if (adev->ip_blocks[i].version->type == block_type) | |||
1849 | return adev->ip_blocks[i].version->funcs->is_idle((void *)adev); | |||
1850 | } | |||
1851 | return true1; | |||
1852 | ||||
1853 | } | |||
1854 | ||||
1855 | /** | |||
1856 | * amdgpu_device_ip_get_ip_block - get a hw IP pointer | |||
1857 | * | |||
1858 | * @adev: amdgpu_device pointer | |||
1859 | * @type: Type of hardware IP (SMU, GFX, UVD, etc.) | |||
1860 | * | |||
1861 | * Returns a pointer to the hardware IP block structure | |||
1862 | * if it exists for the asic, otherwise NULL. | |||
1863 | */ | |||
1864 | struct amdgpu_ip_block * | |||
1865 | amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev, | |||
1866 | enum amd_ip_block_type type) | |||
1867 | { | |||
1868 | int i; | |||
1869 | ||||
1870 | for (i = 0; i < adev->num_ip_blocks; i++) | |||
1871 | if (adev->ip_blocks[i].version->type == type) | |||
1872 | return &adev->ip_blocks[i]; | |||
1873 | ||||
1874 | return NULL((void *)0); | |||
1875 | } | |||
1876 | ||||
1877 | /** | |||
1878 | * amdgpu_device_ip_block_version_cmp | |||
1879 | * | |||
1880 | * @adev: amdgpu_device pointer | |||
1881 | * @type: enum amd_ip_block_type | |||
1882 | * @major: major version | |||
1883 | * @minor: minor version | |||
1884 | * | |||
1885 | * return 0 if equal or greater | |||
1886 | * return 1 if smaller or the ip_block doesn't exist | |||
1887 | */ | |||
1888 | int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev, | |||
1889 | enum amd_ip_block_type type, | |||
1890 | u32 major, u32 minor) | |||
1891 | { | |||
1892 | struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type); | |||
1893 | ||||
1894 | if (ip_block && ((ip_block->version->major > major) || | |||
1895 | ((ip_block->version->major == major) && | |||
1896 | (ip_block->version->minor >= minor)))) | |||
1897 | return 0; | |||
1898 | ||||
1899 | return 1; | |||
1900 | } | |||
1901 | ||||
1902 | /** | |||
1903 | * amdgpu_device_ip_block_add | |||
1904 | * | |||
1905 | * @adev: amdgpu_device pointer | |||
1906 | * @ip_block_version: pointer to the IP to add | |||
1907 | * | |||
1908 | * Adds the IP block driver information to the collection of IPs | |||
1909 | * on the asic. | |||
1910 | */ | |||
1911 | int amdgpu_device_ip_block_add(struct amdgpu_device *adev, | |||
1912 | const struct amdgpu_ip_block_version *ip_block_version) | |||
1913 | { | |||
1914 | if (!ip_block_version) | |||
1915 | return -EINVAL22; | |||
1916 | ||||
1917 | switch (ip_block_version->type) { | |||
1918 | case AMD_IP_BLOCK_TYPE_VCN: | |||
1919 | if (adev->harvest_ip_mask & AMD_HARVEST_IP_VCN_MASK) | |||
1920 | return 0; | |||
1921 | break; | |||
1922 | case AMD_IP_BLOCK_TYPE_JPEG: | |||
1923 | if (adev->harvest_ip_mask & AMD_HARVEST_IP_JPEG_MASK) | |||
1924 | return 0; | |||
1925 | break; | |||
1926 | default: | |||
1927 | break; | |||
1928 | } | |||
1929 | ||||
1930 | DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,printk("\0016" "[" "drm" "] " "add ip block number %d <%s>\n" , adev->num_ip_blocks, ip_block_version->funcs->name ) | |||
1931 | ip_block_version->funcs->name)printk("\0016" "[" "drm" "] " "add ip block number %d <%s>\n" , adev->num_ip_blocks, ip_block_version->funcs->name ); | |||
1932 | ||||
1933 | adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version; | |||
1934 | ||||
1935 | return 0; | |||
1936 | } | |||
1937 | ||||
1938 | /** | |||
1939 | * amdgpu_device_enable_virtual_display - enable virtual display feature | |||
1940 | * | |||
1941 | * @adev: amdgpu_device pointer | |||
1942 | * | |||
1943 | * Enabled the virtual display feature if the user has enabled it via | |||
1944 | * the module parameter virtual_display. This feature provides a virtual | |||
1945 | * display hardware on headless boards or in virtualized environments. | |||
1946 | * This function parses and validates the configuration string specified by | |||
1947 | * the user and configues the virtual display configuration (number of | |||
1948 | * virtual connectors, crtcs, etc.) specified. | |||
1949 | */ | |||
1950 | static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev) | |||
1951 | { | |||
1952 | adev->enable_virtual_display = false0; | |||
1953 | ||||
1954 | #ifdef notyet | |||
1955 | if (amdgpu_virtual_display) { | |||
1956 | const char *pci_address_name = pci_name(adev->pdev); | |||
1957 | char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname; | |||
1958 | ||||
1959 | pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL(0x0001 | 0x0004)); | |||
1960 | pciaddstr_tmp = pciaddstr; | |||
1961 | while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) { | |||
1962 | pciaddname = strsep(&pciaddname_tmp, ","); | |||
1963 | if (!strcmp("all", pciaddname) | |||
1964 | || !strcmp(pci_address_name, pciaddname)) { | |||
1965 | long num_crtc; | |||
1966 | int res = -1; | |||
1967 | ||||
1968 | adev->enable_virtual_display = true1; | |||
1969 | ||||
1970 | if (pciaddname_tmp) | |||
1971 | res = kstrtol(pciaddname_tmp, 10, | |||
1972 | &num_crtc); | |||
1973 | ||||
1974 | if (!res) { | |||
1975 | if (num_crtc < 1) | |||
1976 | num_crtc = 1; | |||
1977 | if (num_crtc > 6) | |||
1978 | num_crtc = 6; | |||
1979 | adev->mode_info.num_crtc = num_crtc; | |||
1980 | } else { | |||
1981 | adev->mode_info.num_crtc = 1; | |||
1982 | } | |||
1983 | break; | |||
1984 | } | |||
1985 | } | |||
1986 | ||||
1987 | DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",printk("\0016" "[" "drm" "] " "virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n" , amdgpu_virtual_display, pci_address_name, adev->enable_virtual_display , adev->mode_info.num_crtc) | |||
1988 | amdgpu_virtual_display, pci_address_name,printk("\0016" "[" "drm" "] " "virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n" , amdgpu_virtual_display, pci_address_name, adev->enable_virtual_display , adev->mode_info.num_crtc) | |||
1989 | adev->enable_virtual_display, adev->mode_info.num_crtc)printk("\0016" "[" "drm" "] " "virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n" , amdgpu_virtual_display, pci_address_name, adev->enable_virtual_display , adev->mode_info.num_crtc); | |||
1990 | ||||
1991 | kfree(pciaddstr); | |||
1992 | } | |||
1993 | #endif | |||
1994 | } | |||
1995 | ||||
1996 | /** | |||
1997 | * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware | |||
1998 | * | |||
1999 | * @adev: amdgpu_device pointer | |||
2000 | * | |||
2001 | * Parses the asic configuration parameters specified in the gpu info | |||
2002 | * firmware and makes them availale to the driver for use in configuring | |||
2003 | * the asic. | |||
2004 | * Returns 0 on success, -EINVAL on failure. | |||
2005 | */ | |||
2006 | static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev) | |||
2007 | { | |||
2008 | const char *chip_name; | |||
2009 | char fw_name[40]; | |||
2010 | int err; | |||
2011 | const struct gpu_info_firmware_header_v1_0 *hdr; | |||
2012 | ||||
2013 | adev->firmware.gpu_info_fw = NULL((void *)0); | |||
2014 | ||||
2015 | if (adev->mman.discovery_bin) { | |||
2016 | /* | |||
2017 | * FIXME: The bounding box is still needed by Navi12, so | |||
2018 | * temporarily read it from gpu_info firmware. Should be dropped | |||
2019 | * when DAL no longer needs it. | |||
2020 | */ | |||
2021 | if (adev->asic_type != CHIP_NAVI12) | |||
2022 | return 0; | |||
2023 | } | |||
2024 | ||||
2025 | switch (adev->asic_type) { | |||
2026 | default: | |||
2027 | return 0; | |||
2028 | case CHIP_VEGA10: | |||
2029 | chip_name = "vega10"; | |||
2030 | break; | |||
2031 | case CHIP_VEGA12: | |||
2032 | chip_name = "vega12"; | |||
2033 | break; | |||
2034 | case CHIP_RAVEN: | |||
2035 | if (adev->apu_flags & AMD_APU_IS_RAVEN2) | |||
2036 | chip_name = "raven2"; | |||
2037 | else if (adev->apu_flags & AMD_APU_IS_PICASSO) | |||
2038 | chip_name = "picasso"; | |||
2039 | else | |||
2040 | chip_name = "raven"; | |||
2041 | break; | |||
2042 | case CHIP_ARCTURUS: | |||
2043 | chip_name = "arcturus"; | |||
2044 | break; | |||
2045 | case CHIP_NAVI12: | |||
2046 | chip_name = "navi12"; | |||
2047 | break; | |||
2048 | } | |||
2049 | ||||
2050 | snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name); | |||
2051 | err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev); | |||
2052 | if (err) { | |||
2053 | dev_err(adev->dev,printf("drm:pid%d:%s *ERROR* " "Failed to load gpu_info firmware \"%s\"\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , fw_name ) | |||
2054 | "Failed to load gpu_info firmware \"%s\"\n",printf("drm:pid%d:%s *ERROR* " "Failed to load gpu_info firmware \"%s\"\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , fw_name ) | |||
2055 | fw_name)printf("drm:pid%d:%s *ERROR* " "Failed to load gpu_info firmware \"%s\"\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , fw_name ); | |||
2056 | goto out; | |||
2057 | } | |||
2058 | err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw); | |||
2059 | if (err) { | |||
2060 | dev_err(adev->dev,printf("drm:pid%d:%s *ERROR* " "Failed to validate gpu_info firmware \"%s\"\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , fw_name ) | |||
2061 | "Failed to validate gpu_info firmware \"%s\"\n",printf("drm:pid%d:%s *ERROR* " "Failed to validate gpu_info firmware \"%s\"\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , fw_name ) | |||
2062 | fw_name)printf("drm:pid%d:%s *ERROR* " "Failed to validate gpu_info firmware \"%s\"\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , fw_name ); | |||
2063 | goto out; | |||
2064 | } | |||
2065 | ||||
2066 | hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data; | |||
2067 | amdgpu_ucode_print_gpu_info_hdr(&hdr->header); | |||
2068 | ||||
2069 | switch (hdr->version_major) { | |||
2070 | case 1: | |||
2071 | { | |||
2072 | const struct gpu_info_firmware_v1_0 *gpu_info_fw = | |||
2073 | (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data + | |||
2074 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)((__uint32_t)(hdr->header.ucode_array_offset_bytes))); | |||
2075 | ||||
2076 | /* | |||
2077 | * Should be droped when DAL no longer needs it. | |||
2078 | */ | |||
2079 | if (adev->asic_type == CHIP_NAVI12) | |||
2080 | goto parse_soc_bounding_box; | |||
2081 | ||||
2082 | adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se)((__uint32_t)(gpu_info_fw->gc_num_se)); | |||
2083 | adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh)((__uint32_t)(gpu_info_fw->gc_num_cu_per_sh)); | |||
2084 | adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se)((__uint32_t)(gpu_info_fw->gc_num_sh_per_se)); | |||
2085 | adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se)((__uint32_t)(gpu_info_fw->gc_num_rb_per_se)); | |||
2086 | adev->gfx.config.max_texture_channel_caches = | |||
2087 | le32_to_cpu(gpu_info_fw->gc_num_tccs)((__uint32_t)(gpu_info_fw->gc_num_tccs)); | |||
2088 | adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs)((__uint32_t)(gpu_info_fw->gc_num_gprs)); | |||
2089 | adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds)((__uint32_t)(gpu_info_fw->gc_num_max_gs_thds)); | |||
2090 | adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth)((__uint32_t)(gpu_info_fw->gc_gs_table_depth)); | |||
2091 | adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth)((__uint32_t)(gpu_info_fw->gc_gsprim_buff_depth)); | |||
2092 | adev->gfx.config.double_offchip_lds_buf = | |||
2093 | le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer)((__uint32_t)(gpu_info_fw->gc_double_offchip_lds_buffer)); | |||
2094 | adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size)((__uint32_t)(gpu_info_fw->gc_wave_size)); | |||
2095 | adev->gfx.cu_info.max_waves_per_simd = | |||
2096 | le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd)((__uint32_t)(gpu_info_fw->gc_max_waves_per_simd)); | |||
2097 | adev->gfx.cu_info.max_scratch_slots_per_cu = | |||
2098 | le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu)((__uint32_t)(gpu_info_fw->gc_max_scratch_slots_per_cu)); | |||
2099 | adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size)((__uint32_t)(gpu_info_fw->gc_lds_size)); | |||
2100 | if (hdr->version_minor >= 1) { | |||
2101 | const struct gpu_info_firmware_v1_1 *gpu_info_fw = | |||
2102 | (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data + | |||
2103 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)((__uint32_t)(hdr->header.ucode_array_offset_bytes))); | |||
2104 | adev->gfx.config.num_sc_per_sh = | |||
2105 | le32_to_cpu(gpu_info_fw->num_sc_per_sh)((__uint32_t)(gpu_info_fw->num_sc_per_sh)); | |||
2106 | adev->gfx.config.num_packer_per_sc = | |||
2107 | le32_to_cpu(gpu_info_fw->num_packer_per_sc)((__uint32_t)(gpu_info_fw->num_packer_per_sc)); | |||
2108 | } | |||
2109 | ||||
2110 | parse_soc_bounding_box: | |||
2111 | /* | |||
2112 | * soc bounding box info is not integrated in disocovery table, | |||
2113 | * we always need to parse it from gpu info firmware if needed. | |||
2114 | */ | |||
2115 | if (hdr->version_minor == 2) { | |||
2116 | const struct gpu_info_firmware_v1_2 *gpu_info_fw = | |||
2117 | (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data + | |||
2118 | le32_to_cpu(hdr->header.ucode_array_offset_bytes)((__uint32_t)(hdr->header.ucode_array_offset_bytes))); | |||
2119 | adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box; | |||
2120 | } | |||
2121 | break; | |||
2122 | } | |||
2123 | default: | |||
2124 | dev_err(adev->dev,printf("drm:pid%d:%s *ERROR* " "Unsupported gpu_info table %d\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , hdr-> header.ucode_version) | |||
2125 | "Unsupported gpu_info table %d\n", hdr->header.ucode_version)printf("drm:pid%d:%s *ERROR* " "Unsupported gpu_info table %d\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , hdr-> header.ucode_version); | |||
2126 | err = -EINVAL22; | |||
2127 | goto out; | |||
2128 | } | |||
2129 | out: | |||
2130 | return err; | |||
2131 | } | |||
2132 | ||||
2133 | /** | |||
2134 | * amdgpu_device_ip_early_init - run early init for hardware IPs | |||
2135 | * | |||
2136 | * @adev: amdgpu_device pointer | |||
2137 | * | |||
2138 | * Early initialization pass for hardware IPs. The hardware IPs that make | |||
2139 | * up each asic are discovered each IP's early_init callback is run. This | |||
2140 | * is the first stage in initializing the asic. | |||
2141 | * Returns 0 on success, negative error code on failure. | |||
2142 | */ | |||
2143 | static int amdgpu_device_ip_early_init(struct amdgpu_device *adev) | |||
2144 | { | |||
2145 | struct pci_dev *parent; | |||
2146 | int i, r; | |||
2147 | ||||
2148 | amdgpu_device_enable_virtual_display(adev); | |||
2149 | ||||
2150 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) { | |||
2151 | r = amdgpu_virt_request_full_gpu(adev, true1); | |||
2152 | if (r) | |||
2153 | return r; | |||
2154 | } | |||
2155 | ||||
2156 | switch (adev->asic_type) { | |||
2157 | #ifdef CONFIG_DRM_AMDGPU_SI | |||
2158 | case CHIP_VERDE: | |||
2159 | case CHIP_TAHITI: | |||
2160 | case CHIP_PITCAIRN: | |||
2161 | case CHIP_OLAND: | |||
2162 | case CHIP_HAINAN: | |||
2163 | adev->family = AMDGPU_FAMILY_SI110; | |||
2164 | r = si_set_ip_blocks(adev); | |||
2165 | if (r) | |||
2166 | return r; | |||
2167 | break; | |||
2168 | #endif | |||
2169 | #ifdef CONFIG_DRM_AMDGPU_CIK | |||
2170 | case CHIP_BONAIRE: | |||
2171 | case CHIP_HAWAII: | |||
2172 | case CHIP_KAVERI: | |||
2173 | case CHIP_KABINI: | |||
2174 | case CHIP_MULLINS: | |||
2175 | if (adev->flags & AMD_IS_APU) | |||
2176 | adev->family = AMDGPU_FAMILY_KV125; | |||
2177 | else | |||
2178 | adev->family = AMDGPU_FAMILY_CI120; | |||
2179 | ||||
2180 | r = cik_set_ip_blocks(adev); | |||
2181 | if (r) | |||
2182 | return r; | |||
2183 | break; | |||
2184 | #endif | |||
2185 | case CHIP_TOPAZ: | |||
2186 | case CHIP_TONGA: | |||
2187 | case CHIP_FIJI: | |||
2188 | case CHIP_POLARIS10: | |||
2189 | case CHIP_POLARIS11: | |||
2190 | case CHIP_POLARIS12: | |||
2191 | case CHIP_VEGAM: | |||
2192 | case CHIP_CARRIZO: | |||
2193 | case CHIP_STONEY: | |||
2194 | if (adev->flags & AMD_IS_APU) | |||
2195 | adev->family = AMDGPU_FAMILY_CZ135; | |||
2196 | else | |||
2197 | adev->family = AMDGPU_FAMILY_VI130; | |||
2198 | ||||
2199 | r = vi_set_ip_blocks(adev); | |||
2200 | if (r) | |||
2201 | return r; | |||
2202 | break; | |||
2203 | default: | |||
2204 | r = amdgpu_discovery_set_ip_blocks(adev); | |||
2205 | if (r) | |||
2206 | return r; | |||
2207 | break; | |||
2208 | } | |||
2209 | ||||
2210 | if (amdgpu_has_atpx() && | |||
2211 | (amdgpu_is_atpx_hybrid() || | |||
2212 | amdgpu_has_atpx_dgpu_power_cntl()) && | |||
2213 | ((adev->flags & AMD_IS_APU) == 0) && | |||
2214 | !dev_is_removable(&adev->pdev->dev)0) | |||
2215 | adev->flags |= AMD_IS_PX; | |||
2216 | ||||
2217 | if (!(adev->flags & AMD_IS_APU)) { | |||
2218 | #ifdef notyet | |||
2219 | parent = pcie_find_root_port(adev->pdev); | |||
2220 | adev->has_pr3 = parent ? pci_pr3_present(parent) : false0; | |||
2221 | #else | |||
2222 | adev->has_pr3 = false0; | |||
2223 | #endif | |||
2224 | } | |||
2225 | ||||
2226 | amdgpu_amdkfd_device_probe(adev); | |||
2227 | ||||
2228 | adev->pm.pp_feature = amdgpu_pp_feature_mask; | |||
2229 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2)) || sched_policy == KFD_SCHED_POLICY_NO_HWS) | |||
2230 | adev->pm.pp_feature &= ~PP_GFXOFF_MASK; | |||
2231 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2)) && adev->asic_type == CHIP_SIENNA_CICHLID) | |||
2232 | adev->pm.pp_feature &= ~PP_OVERDRIVE_MASK; | |||
2233 | ||||
2234 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
2235 | if ((amdgpu_ip_block_mask & (1 << i)) == 0) { | |||
2236 | DRM_ERROR("disabled ip block: %d <%s>\n",__drm_err("disabled ip block: %d <%s>\n", i, adev->ip_blocks [i].version->funcs->name) | |||
2237 | i, adev->ip_blocks[i].version->funcs->name)__drm_err("disabled ip block: %d <%s>\n", i, adev->ip_blocks [i].version->funcs->name); | |||
2238 | adev->ip_blocks[i].status.valid = false0; | |||
2239 | } else { | |||
2240 | if (adev->ip_blocks[i].version->funcs->early_init) { | |||
2241 | r = adev->ip_blocks[i].version->funcs->early_init((void *)adev); | |||
2242 | if (r == -ENOENT2) { | |||
2243 | adev->ip_blocks[i].status.valid = false0; | |||
2244 | } else if (r) { | |||
2245 | DRM_ERROR("early_init of IP block <%s> failed %d\n",__drm_err("early_init of IP block <%s> failed %d\n", adev ->ip_blocks[i].version->funcs->name, r) | |||
2246 | adev->ip_blocks[i].version->funcs->name, r)__drm_err("early_init of IP block <%s> failed %d\n", adev ->ip_blocks[i].version->funcs->name, r); | |||
2247 | return r; | |||
2248 | } else { | |||
2249 | adev->ip_blocks[i].status.valid = true1; | |||
2250 | } | |||
2251 | } else { | |||
2252 | adev->ip_blocks[i].status.valid = true1; | |||
2253 | } | |||
2254 | } | |||
2255 | /* get the vbios after the asic_funcs are set up */ | |||
2256 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) { | |||
2257 | r = amdgpu_device_parse_gpu_info_fw(adev); | |||
2258 | if (r) | |||
2259 | return r; | |||
2260 | ||||
2261 | /* Read BIOS */ | |||
2262 | if (!amdgpu_get_bios(adev)) | |||
2263 | return -EINVAL22; | |||
2264 | ||||
2265 | r = amdgpu_atombios_init(adev); | |||
2266 | if (r) { | |||
2267 | dev_err(adev->dev, "amdgpu_atombios_init failed\n")printf("drm:pid%d:%s *ERROR* " "amdgpu_atombios_init failed\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
2268 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0); | |||
2269 | return r; | |||
2270 | } | |||
2271 | ||||
2272 | /*get pf2vf msg info at it's earliest time*/ | |||
2273 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) | |||
2274 | amdgpu_virt_init_data_exchange(adev); | |||
2275 | ||||
2276 | } | |||
2277 | } | |||
2278 | ||||
2279 | adev->cg_flags &= amdgpu_cg_mask; | |||
2280 | adev->pg_flags &= amdgpu_pg_mask; | |||
2281 | ||||
2282 | return 0; | |||
2283 | } | |||
2284 | ||||
2285 | static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev) | |||
2286 | { | |||
2287 | int i, r; | |||
2288 | ||||
2289 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
2290 | if (!adev->ip_blocks[i].status.sw) | |||
2291 | continue; | |||
2292 | if (adev->ip_blocks[i].status.hw) | |||
2293 | continue; | |||
2294 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || | |||
2295 | (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2)) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) || | |||
2296 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) { | |||
2297 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); | |||
2298 | if (r) { | |||
2299 | DRM_ERROR("hw_init of IP block <%s> failed %d\n",__drm_err("hw_init of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r) | |||
2300 | adev->ip_blocks[i].version->funcs->name, r)__drm_err("hw_init of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r); | |||
2301 | return r; | |||
2302 | } | |||
2303 | adev->ip_blocks[i].status.hw = true1; | |||
2304 | } | |||
2305 | } | |||
2306 | ||||
2307 | return 0; | |||
2308 | } | |||
2309 | ||||
2310 | static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev) | |||
2311 | { | |||
2312 | int i, r; | |||
2313 | ||||
2314 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
2315 | if (!adev->ip_blocks[i].status.sw) | |||
2316 | continue; | |||
2317 | if (adev->ip_blocks[i].status.hw) | |||
2318 | continue; | |||
2319 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); | |||
2320 | if (r) { | |||
2321 | DRM_ERROR("hw_init of IP block <%s> failed %d\n",__drm_err("hw_init of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r) | |||
2322 | adev->ip_blocks[i].version->funcs->name, r)__drm_err("hw_init of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r); | |||
2323 | return r; | |||
2324 | } | |||
2325 | adev->ip_blocks[i].status.hw = true1; | |||
2326 | } | |||
2327 | ||||
2328 | return 0; | |||
2329 | } | |||
2330 | ||||
2331 | static int amdgpu_device_fw_loading(struct amdgpu_device *adev) | |||
2332 | { | |||
2333 | int r = 0; | |||
2334 | int i; | |||
2335 | uint32_t smu_version; | |||
2336 | ||||
2337 | if (adev->asic_type >= CHIP_VEGA10) { | |||
2338 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
2339 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP) | |||
2340 | continue; | |||
2341 | ||||
2342 | if (!adev->ip_blocks[i].status.sw) | |||
2343 | continue; | |||
2344 | ||||
2345 | /* no need to do the fw loading again if already done*/ | |||
2346 | if (adev->ip_blocks[i].status.hw == true1) | |||
2347 | break; | |||
2348 | ||||
2349 | if (amdgpu_in_reset(adev) || adev->in_suspend) { | |||
2350 | r = adev->ip_blocks[i].version->funcs->resume(adev); | |||
2351 | if (r) { | |||
2352 | DRM_ERROR("resume of IP block <%s> failed %d\n",__drm_err("resume of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r) | |||
2353 | adev->ip_blocks[i].version->funcs->name, r)__drm_err("resume of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r); | |||
2354 | return r; | |||
2355 | } | |||
2356 | } else { | |||
2357 | r = adev->ip_blocks[i].version->funcs->hw_init(adev); | |||
2358 | if (r) { | |||
2359 | DRM_ERROR("hw_init of IP block <%s> failed %d\n",__drm_err("hw_init of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r) | |||
2360 | adev->ip_blocks[i].version->funcs->name, r)__drm_err("hw_init of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r); | |||
2361 | return r; | |||
2362 | } | |||
2363 | } | |||
2364 | ||||
2365 | adev->ip_blocks[i].status.hw = true1; | |||
2366 | break; | |||
2367 | } | |||
2368 | } | |||
2369 | ||||
2370 | if (!amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2)) || adev->asic_type == CHIP_TONGA) | |||
2371 | r = amdgpu_pm_load_smu_firmware(adev, &smu_version); | |||
2372 | ||||
2373 | return r; | |||
2374 | } | |||
2375 | ||||
2376 | static int amdgpu_device_init_schedulers(struct amdgpu_device *adev) | |||
2377 | { | |||
2378 | long timeout; | |||
2379 | int r, i; | |||
2380 | ||||
2381 | for (i = 0; i < AMDGPU_MAX_RINGS28; ++i) { | |||
2382 | struct amdgpu_ring *ring = adev->rings[i]; | |||
2383 | ||||
2384 | /* No need to setup the GPU scheduler for rings that don't need it */ | |||
2385 | if (!ring || ring->no_scheduler) | |||
2386 | continue; | |||
2387 | ||||
2388 | switch (ring->funcs->type) { | |||
2389 | case AMDGPU_RING_TYPE_GFX: | |||
2390 | timeout = adev->gfx_timeout; | |||
2391 | break; | |||
2392 | case AMDGPU_RING_TYPE_COMPUTE: | |||
2393 | timeout = adev->compute_timeout; | |||
2394 | break; | |||
2395 | case AMDGPU_RING_TYPE_SDMA: | |||
2396 | timeout = adev->sdma_timeout; | |||
2397 | break; | |||
2398 | default: | |||
2399 | timeout = adev->video_timeout; | |||
2400 | break; | |||
2401 | } | |||
2402 | ||||
2403 | r = drm_sched_init(&ring->sched, &amdgpu_sched_ops, | |||
2404 | ring->num_hw_submission, amdgpu_job_hang_limit, | |||
2405 | timeout, adev->reset_domain->wq, | |||
2406 | ring->sched_score, ring->name, | |||
2407 | adev->dev); | |||
2408 | if (r) { | |||
2409 | DRM_ERROR("Failed to create scheduler on ring %s.\n",__drm_err("Failed to create scheduler on ring %s.\n", ring-> name) | |||
2410 | ring->name)__drm_err("Failed to create scheduler on ring %s.\n", ring-> name); | |||
2411 | return r; | |||
2412 | } | |||
2413 | } | |||
2414 | ||||
2415 | return 0; | |||
2416 | } | |||
2417 | ||||
2418 | ||||
2419 | /** | |||
2420 | * amdgpu_device_ip_init - run init for hardware IPs | |||
2421 | * | |||
2422 | * @adev: amdgpu_device pointer | |||
2423 | * | |||
2424 | * Main initialization pass for hardware IPs. The list of all the hardware | |||
2425 | * IPs that make up the asic is walked and the sw_init and hw_init callbacks | |||
2426 | * are run. sw_init initializes the software state associated with each IP | |||
2427 | * and hw_init initializes the hardware associated with each IP. | |||
2428 | * Returns 0 on success, negative error code on failure. | |||
2429 | */ | |||
2430 | static int amdgpu_device_ip_init(struct amdgpu_device *adev) | |||
2431 | { | |||
2432 | int i, r; | |||
2433 | ||||
2434 | r = amdgpu_ras_init(adev); | |||
2435 | if (r) | |||
2436 | return r; | |||
2437 | ||||
2438 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
2439 | if (!adev->ip_blocks[i].status.valid) | |||
2440 | continue; | |||
2441 | r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev); | |||
2442 | if (r) { | |||
2443 | DRM_ERROR("sw_init of IP block <%s> failed %d\n",__drm_err("sw_init of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r) | |||
2444 | adev->ip_blocks[i].version->funcs->name, r)__drm_err("sw_init of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r); | |||
2445 | goto init_failed; | |||
2446 | } | |||
2447 | adev->ip_blocks[i].status.sw = true1; | |||
2448 | ||||
2449 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) { | |||
2450 | /* need to do common hw init early so everything is set up for gmc */ | |||
2451 | r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev); | |||
2452 | if (r) { | |||
2453 | DRM_ERROR("hw_init %d failed %d\n", i, r)__drm_err("hw_init %d failed %d\n", i, r); | |||
2454 | goto init_failed; | |||
2455 | } | |||
2456 | adev->ip_blocks[i].status.hw = true1; | |||
2457 | } else if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) { | |||
2458 | /* need to do gmc hw init early so we can allocate gpu mem */ | |||
2459 | /* Try to reserve bad pages early */ | |||
2460 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) | |||
2461 | amdgpu_virt_exchange_data(adev); | |||
2462 | ||||
2463 | r = amdgpu_device_vram_scratch_init(adev); | |||
2464 | if (r) { | |||
2465 | DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r)__drm_err("amdgpu_vram_scratch_init failed %d\n", r); | |||
2466 | goto init_failed; | |||
2467 | } | |||
2468 | r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev); | |||
2469 | if (r) { | |||
2470 | DRM_ERROR("hw_init %d failed %d\n", i, r)__drm_err("hw_init %d failed %d\n", i, r); | |||
2471 | goto init_failed; | |||
2472 | } | |||
2473 | r = amdgpu_device_wb_init(adev); | |||
2474 | if (r) { | |||
2475 | DRM_ERROR("amdgpu_device_wb_init failed %d\n", r)__drm_err("amdgpu_device_wb_init failed %d\n", r); | |||
2476 | goto init_failed; | |||
2477 | } | |||
2478 | adev->ip_blocks[i].status.hw = true1; | |||
2479 | ||||
2480 | /* right after GMC hw init, we create CSA */ | |||
2481 | if (amdgpu_mcbp || amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) { | |||
2482 | r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj, | |||
2483 | AMDGPU_GEM_DOMAIN_VRAM0x4, | |||
2484 | AMDGPU_CSA_SIZE(128 * 1024)); | |||
2485 | if (r) { | |||
2486 | DRM_ERROR("allocate CSA failed %d\n", r)__drm_err("allocate CSA failed %d\n", r); | |||
2487 | goto init_failed; | |||
2488 | } | |||
2489 | } | |||
2490 | } | |||
2491 | } | |||
2492 | ||||
2493 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) | |||
2494 | amdgpu_virt_init_data_exchange(adev); | |||
2495 | ||||
2496 | r = amdgpu_ib_pool_init(adev); | |||
2497 | if (r) { | |||
2498 | dev_err(adev->dev, "IB initialization failed (%d).\n", r)printf("drm:pid%d:%s *ERROR* " "IB initialization failed (%d).\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , r); | |||
2499 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r); | |||
2500 | goto init_failed; | |||
2501 | } | |||
2502 | ||||
2503 | r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/ | |||
2504 | if (r) | |||
2505 | goto init_failed; | |||
2506 | ||||
2507 | r = amdgpu_device_ip_hw_init_phase1(adev); | |||
2508 | if (r) | |||
2509 | goto init_failed; | |||
2510 | ||||
2511 | r = amdgpu_device_fw_loading(adev); | |||
2512 | if (r) | |||
2513 | goto init_failed; | |||
2514 | ||||
2515 | r = amdgpu_device_ip_hw_init_phase2(adev); | |||
2516 | if (r) | |||
2517 | goto init_failed; | |||
2518 | ||||
2519 | /* | |||
2520 | * retired pages will be loaded from eeprom and reserved here, | |||
2521 | * it should be called after amdgpu_device_ip_hw_init_phase2 since | |||
2522 | * for some ASICs the RAS EEPROM code relies on SMU fully functioning | |||
2523 | * for I2C communication which only true at this point. | |||
2524 | * | |||
2525 | * amdgpu_ras_recovery_init may fail, but the upper only cares the | |||
2526 | * failure from bad gpu situation and stop amdgpu init process | |||
2527 | * accordingly. For other failed cases, it will still release all | |||
2528 | * the resource and print error message, rather than returning one | |||
2529 | * negative value to upper level. | |||
2530 | * | |||
2531 | * Note: theoretically, this should be called before all vram allocations | |||
2532 | * to protect retired page from abusing | |||
2533 | */ | |||
2534 | r = amdgpu_ras_recovery_init(adev); | |||
2535 | if (r) | |||
2536 | goto init_failed; | |||
2537 | ||||
2538 | /** | |||
2539 | * In case of XGMI grab extra reference for reset domain for this device | |||
2540 | */ | |||
2541 | if (adev->gmc.xgmi.num_physical_nodes > 1) { | |||
2542 | if (amdgpu_xgmi_add_device(adev) == 0) { | |||
2543 | if (!amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) { | |||
2544 | struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev); | |||
2545 | ||||
2546 | if (WARN_ON(!hive)({ int __ret = !!(!hive); if (__ret) printf("WARNING %s failed at %s:%d\n" , "!hive", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_device.c" , 2546); __builtin_expect(!!(__ret), 0); })) { | |||
2547 | r = -ENOENT2; | |||
2548 | goto init_failed; | |||
2549 | } | |||
2550 | ||||
2551 | if (!hive->reset_domain || | |||
2552 | !amdgpu_reset_get_reset_domain(hive->reset_domain)) { | |||
2553 | r = -ENOENT2; | |||
2554 | amdgpu_put_xgmi_hive(hive); | |||
2555 | goto init_failed; | |||
2556 | } | |||
2557 | ||||
2558 | /* Drop the early temporary reset domain we created for device */ | |||
2559 | amdgpu_reset_put_reset_domain(adev->reset_domain); | |||
2560 | adev->reset_domain = hive->reset_domain; | |||
2561 | amdgpu_put_xgmi_hive(hive); | |||
2562 | } | |||
2563 | } | |||
2564 | } | |||
2565 | ||||
2566 | r = amdgpu_device_init_schedulers(adev); | |||
2567 | if (r) | |||
2568 | goto init_failed; | |||
2569 | ||||
2570 | /* Don't init kfd if whole hive need to be reset during init */ | |||
2571 | if (!adev->gmc.xgmi.pending_reset) | |||
2572 | amdgpu_amdkfd_device_init(adev); | |||
2573 | ||||
2574 | amdgpu_fru_get_product_info(adev); | |||
2575 | ||||
2576 | init_failed: | |||
2577 | ||||
2578 | return r; | |||
2579 | } | |||
2580 | ||||
2581 | /** | |||
2582 | * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer | |||
2583 | * | |||
2584 | * @adev: amdgpu_device pointer | |||
2585 | * | |||
2586 | * Writes a reset magic value to the gart pointer in VRAM. The driver calls | |||
2587 | * this function before a GPU reset. If the value is retained after a | |||
2588 | * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents. | |||
2589 | */ | |||
2590 | static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev) | |||
2591 | { | |||
2592 | memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM)__builtin_memcpy((adev->reset_magic), (adev->gart.ptr), (64)); | |||
2593 | } | |||
2594 | ||||
2595 | /** | |||
2596 | * amdgpu_device_check_vram_lost - check if vram is valid | |||
2597 | * | |||
2598 | * @adev: amdgpu_device pointer | |||
2599 | * | |||
2600 | * Checks the reset magic value written to the gart pointer in VRAM. | |||
2601 | * The driver calls this after a GPU reset to see if the contents of | |||
2602 | * VRAM is lost or now. | |||
2603 | * returns true if vram is lost, false if not. | |||
2604 | */ | |||
2605 | static bool_Bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev) | |||
2606 | { | |||
2607 | if (memcmp(adev->gart.ptr, adev->reset_magic,__builtin_memcmp((adev->gart.ptr), (adev->reset_magic), (64)) | |||
2608 | AMDGPU_RESET_MAGIC_NUM)__builtin_memcmp((adev->gart.ptr), (adev->reset_magic), (64))) | |||
2609 | return true1; | |||
2610 | ||||
2611 | if (!amdgpu_in_reset(adev)) | |||
2612 | return false0; | |||
2613 | ||||
2614 | /* | |||
2615 | * For all ASICs with baco/mode1 reset, the VRAM is | |||
2616 | * always assumed to be lost. | |||
2617 | */ | |||
2618 | switch (amdgpu_asic_reset_method(adev)(adev)->asic_funcs->reset_method((adev))) { | |||
2619 | case AMD_RESET_METHOD_BACO: | |||
2620 | case AMD_RESET_METHOD_MODE1: | |||
2621 | return true1; | |||
2622 | default: | |||
2623 | return false0; | |||
2624 | } | |||
2625 | } | |||
2626 | ||||
2627 | /** | |||
2628 | * amdgpu_device_set_cg_state - set clockgating for amdgpu device | |||
2629 | * | |||
2630 | * @adev: amdgpu_device pointer | |||
2631 | * @state: clockgating state (gate or ungate) | |||
2632 | * | |||
2633 | * The list of all the hardware IPs that make up the asic is walked and the | |||
2634 | * set_clockgating_state callbacks are run. | |||
2635 | * Late initialization pass enabling clockgating for hardware IPs. | |||
2636 | * Fini or suspend, pass disabling clockgating for hardware IPs. | |||
2637 | * Returns 0 on success, negative error code on failure. | |||
2638 | */ | |||
2639 | ||||
2640 | int amdgpu_device_set_cg_state(struct amdgpu_device *adev, | |||
2641 | enum amd_clockgating_state state) | |||
2642 | { | |||
2643 | int i, j, r; | |||
2644 | ||||
2645 | if (amdgpu_emu_mode == 1) | |||
2646 | return 0; | |||
2647 | ||||
2648 | for (j = 0; j < adev->num_ip_blocks; j++) { | |||
2649 | i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1; | |||
2650 | if (!adev->ip_blocks[i].status.late_initialized) | |||
2651 | continue; | |||
2652 | /* skip CG for GFX on S0ix */ | |||
2653 | if (adev->in_s0ix && | |||
2654 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX) | |||
2655 | continue; | |||
2656 | /* skip CG for VCE/UVD, it's handled specially */ | |||
2657 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && | |||
2658 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && | |||
2659 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && | |||
2660 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG && | |||
2661 | adev->ip_blocks[i].version->funcs->set_clockgating_state) { | |||
2662 | /* enable clockgating to save power */ | |||
2663 | r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev, | |||
2664 | state); | |||
2665 | if (r) { | |||
2666 | DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",__drm_err("set_clockgating_state(gate) of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r) | |||
2667 | adev->ip_blocks[i].version->funcs->name, r)__drm_err("set_clockgating_state(gate) of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r); | |||
2668 | return r; | |||
2669 | } | |||
2670 | } | |||
2671 | } | |||
2672 | ||||
2673 | return 0; | |||
2674 | } | |||
2675 | ||||
2676 | int amdgpu_device_set_pg_state(struct amdgpu_device *adev, | |||
2677 | enum amd_powergating_state state) | |||
2678 | { | |||
2679 | int i, j, r; | |||
2680 | ||||
2681 | if (amdgpu_emu_mode == 1) | |||
2682 | return 0; | |||
2683 | ||||
2684 | for (j = 0; j < adev->num_ip_blocks; j++) { | |||
2685 | i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1; | |||
2686 | if (!adev->ip_blocks[i].status.late_initialized) | |||
2687 | continue; | |||
2688 | /* skip PG for GFX on S0ix */ | |||
2689 | if (adev->in_s0ix && | |||
2690 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX) | |||
2691 | continue; | |||
2692 | /* skip CG for VCE/UVD, it's handled specially */ | |||
2693 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && | |||
2694 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && | |||
2695 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && | |||
2696 | adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG && | |||
2697 | adev->ip_blocks[i].version->funcs->set_powergating_state) { | |||
2698 | /* enable powergating to save power */ | |||
2699 | r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev, | |||
2700 | state); | |||
2701 | if (r) { | |||
2702 | DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",__drm_err("set_powergating_state(gate) of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r) | |||
2703 | adev->ip_blocks[i].version->funcs->name, r)__drm_err("set_powergating_state(gate) of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r); | |||
2704 | return r; | |||
2705 | } | |||
2706 | } | |||
2707 | } | |||
2708 | return 0; | |||
2709 | } | |||
2710 | ||||
2711 | static int amdgpu_device_enable_mgpu_fan_boost(void) | |||
2712 | { | |||
2713 | struct amdgpu_gpu_instance *gpu_ins; | |||
2714 | struct amdgpu_device *adev; | |||
2715 | int i, ret = 0; | |||
2716 | ||||
2717 | mutex_lock(&mgpu_info.mutex)rw_enter_write(&mgpu_info.mutex); | |||
2718 | ||||
2719 | /* | |||
2720 | * MGPU fan boost feature should be enabled | |||
2721 | * only when there are two or more dGPUs in | |||
2722 | * the system | |||
2723 | */ | |||
2724 | if (mgpu_info.num_dgpu < 2) | |||
2725 | goto out; | |||
2726 | ||||
2727 | for (i = 0; i < mgpu_info.num_dgpu; i++) { | |||
2728 | gpu_ins = &(mgpu_info.gpu_ins[i]); | |||
2729 | adev = gpu_ins->adev; | |||
2730 | if (!(adev->flags & AMD_IS_APU) && | |||
2731 | !gpu_ins->mgpu_fan_enabled) { | |||
2732 | ret = amdgpu_dpm_enable_mgpu_fan_boost(adev); | |||
2733 | if (ret) | |||
2734 | break; | |||
2735 | ||||
2736 | gpu_ins->mgpu_fan_enabled = 1; | |||
2737 | } | |||
2738 | } | |||
2739 | ||||
2740 | out: | |||
2741 | mutex_unlock(&mgpu_info.mutex)rw_exit_write(&mgpu_info.mutex); | |||
2742 | ||||
2743 | return ret; | |||
2744 | } | |||
2745 | ||||
2746 | /** | |||
2747 | * amdgpu_device_ip_late_init - run late init for hardware IPs | |||
2748 | * | |||
2749 | * @adev: amdgpu_device pointer | |||
2750 | * | |||
2751 | * Late initialization pass for hardware IPs. The list of all the hardware | |||
2752 | * IPs that make up the asic is walked and the late_init callbacks are run. | |||
2753 | * late_init covers any special initialization that an IP requires | |||
2754 | * after all of the have been initialized or something that needs to happen | |||
2755 | * late in the init process. | |||
2756 | * Returns 0 on success, negative error code on failure. | |||
2757 | */ | |||
2758 | static int amdgpu_device_ip_late_init(struct amdgpu_device *adev) | |||
2759 | { | |||
2760 | struct amdgpu_gpu_instance *gpu_instance; | |||
2761 | int i = 0, r; | |||
2762 | ||||
2763 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
2764 | if (!adev->ip_blocks[i].status.hw) | |||
2765 | continue; | |||
2766 | if (adev->ip_blocks[i].version->funcs->late_init) { | |||
2767 | r = adev->ip_blocks[i].version->funcs->late_init((void *)adev); | |||
2768 | if (r) { | |||
2769 | DRM_ERROR("late_init of IP block <%s> failed %d\n",__drm_err("late_init of IP block <%s> failed %d\n", adev ->ip_blocks[i].version->funcs->name, r) | |||
2770 | adev->ip_blocks[i].version->funcs->name, r)__drm_err("late_init of IP block <%s> failed %d\n", adev ->ip_blocks[i].version->funcs->name, r); | |||
2771 | return r; | |||
2772 | } | |||
2773 | } | |||
2774 | adev->ip_blocks[i].status.late_initialized = true1; | |||
2775 | } | |||
2776 | ||||
2777 | r = amdgpu_ras_late_init(adev); | |||
2778 | if (r) { | |||
2779 | DRM_ERROR("amdgpu_ras_late_init failed %d", r)__drm_err("amdgpu_ras_late_init failed %d", r); | |||
2780 | return r; | |||
2781 | } | |||
2782 | ||||
2783 | amdgpu_ras_set_error_query_ready(adev, true1); | |||
2784 | ||||
2785 | amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE); | |||
2786 | amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE); | |||
2787 | ||||
2788 | amdgpu_device_fill_reset_magic(adev); | |||
2789 | ||||
2790 | r = amdgpu_device_enable_mgpu_fan_boost(); | |||
2791 | if (r) | |||
2792 | DRM_ERROR("enable mgpu fan boost failed (%d).\n", r)__drm_err("enable mgpu fan boost failed (%d).\n", r); | |||
2793 | ||||
2794 | /* For passthrough configuration on arcturus and aldebaran, enable special handling SBR */ | |||
2795 | if (amdgpu_passthrough(adev)((adev)->virt.caps & (1 << 3)) && ((adev->asic_type == CHIP_ARCTURUS && adev->gmc.xgmi.num_physical_nodes > 1)|| | |||
2796 | adev->asic_type == CHIP_ALDEBARAN )) | |||
2797 | amdgpu_dpm_handle_passthrough_sbr(adev, true1); | |||
2798 | ||||
2799 | if (adev->gmc.xgmi.num_physical_nodes > 1) { | |||
2800 | mutex_lock(&mgpu_info.mutex)rw_enter_write(&mgpu_info.mutex); | |||
2801 | ||||
2802 | /* | |||
2803 | * Reset device p-state to low as this was booted with high. | |||
2804 | * | |||
2805 | * This should be performed only after all devices from the same | |||
2806 | * hive get initialized. | |||
2807 | * | |||
2808 | * However, it's unknown how many device in the hive in advance. | |||
2809 | * As this is counted one by one during devices initializations. | |||
2810 | * | |||
2811 | * So, we wait for all XGMI interlinked devices initialized. | |||
2812 | * This may bring some delays as those devices may come from | |||
2813 | * different hives. But that should be OK. | |||
2814 | */ | |||
2815 | if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) { | |||
2816 | for (i = 0; i < mgpu_info.num_gpu; i++) { | |||
2817 | gpu_instance = &(mgpu_info.gpu_ins[i]); | |||
2818 | if (gpu_instance->adev->flags & AMD_IS_APU) | |||
2819 | continue; | |||
2820 | ||||
2821 | r = amdgpu_xgmi_set_pstate(gpu_instance->adev, | |||
2822 | AMDGPU_XGMI_PSTATE_MIN); | |||
2823 | if (r) { | |||
2824 | DRM_ERROR("pstate setting failed (%d).\n", r)__drm_err("pstate setting failed (%d).\n", r); | |||
2825 | break; | |||
2826 | } | |||
2827 | } | |||
2828 | } | |||
2829 | ||||
2830 | mutex_unlock(&mgpu_info.mutex)rw_exit_write(&mgpu_info.mutex); | |||
2831 | } | |||
2832 | ||||
2833 | return 0; | |||
2834 | } | |||
2835 | ||||
2836 | /** | |||
2837 | * amdgpu_device_smu_fini_early - smu hw_fini wrapper | |||
2838 | * | |||
2839 | * @adev: amdgpu_device pointer | |||
2840 | * | |||
2841 | * For ASICs need to disable SMC first | |||
2842 | */ | |||
2843 | static void amdgpu_device_smu_fini_early(struct amdgpu_device *adev) | |||
2844 | { | |||
2845 | int i, r; | |||
2846 | ||||
2847 | if (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0)(((9) << 16) | ((0) << 8) | (0))) | |||
2848 | return; | |||
2849 | ||||
2850 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
2851 | if (!adev->ip_blocks[i].status.hw) | |||
2852 | continue; | |||
2853 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) { | |||
2854 | r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev); | |||
2855 | /* XXX handle errors */ | |||
2856 | if (r) { | |||
2857 | DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",___drm_dbg(((void *)0), DRM_UT_CORE, "hw_fini of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r) | |||
2858 | adev->ip_blocks[i].version->funcs->name, r)___drm_dbg(((void *)0), DRM_UT_CORE, "hw_fini of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r); | |||
2859 | } | |||
2860 | adev->ip_blocks[i].status.hw = false0; | |||
2861 | break; | |||
2862 | } | |||
2863 | } | |||
2864 | } | |||
2865 | ||||
2866 | static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev) | |||
2867 | { | |||
2868 | int i, r; | |||
2869 | ||||
2870 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
2871 | if (!adev->ip_blocks[i].version->funcs->early_fini) | |||
2872 | continue; | |||
2873 | ||||
2874 | r = adev->ip_blocks[i].version->funcs->early_fini((void *)adev); | |||
2875 | if (r) { | |||
2876 | DRM_DEBUG("early_fini of IP block <%s> failed %d\n",___drm_dbg(((void *)0), DRM_UT_CORE, "early_fini of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r) | |||
2877 | adev->ip_blocks[i].version->funcs->name, r)___drm_dbg(((void *)0), DRM_UT_CORE, "early_fini of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r); | |||
2878 | } | |||
2879 | } | |||
2880 | ||||
2881 | amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE); | |||
2882 | amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE); | |||
2883 | ||||
2884 | amdgpu_amdkfd_suspend(adev, false0); | |||
2885 | ||||
2886 | /* Workaroud for ASICs need to disable SMC first */ | |||
2887 | amdgpu_device_smu_fini_early(adev); | |||
2888 | ||||
2889 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { | |||
2890 | if (!adev->ip_blocks[i].status.hw) | |||
2891 | continue; | |||
2892 | ||||
2893 | r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev); | |||
2894 | /* XXX handle errors */ | |||
2895 | if (r) { | |||
2896 | DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",___drm_dbg(((void *)0), DRM_UT_CORE, "hw_fini of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r) | |||
2897 | adev->ip_blocks[i].version->funcs->name, r)___drm_dbg(((void *)0), DRM_UT_CORE, "hw_fini of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r); | |||
2898 | } | |||
2899 | ||||
2900 | adev->ip_blocks[i].status.hw = false0; | |||
2901 | } | |||
2902 | ||||
2903 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) { | |||
2904 | if (amdgpu_virt_release_full_gpu(adev, false0)) | |||
2905 | DRM_ERROR("failed to release exclusive mode on fini\n")__drm_err("failed to release exclusive mode on fini\n"); | |||
2906 | } | |||
2907 | ||||
2908 | return 0; | |||
2909 | } | |||
2910 | ||||
2911 | /** | |||
2912 | * amdgpu_device_ip_fini - run fini for hardware IPs | |||
2913 | * | |||
2914 | * @adev: amdgpu_device pointer | |||
2915 | * | |||
2916 | * Main teardown pass for hardware IPs. The list of all the hardware | |||
2917 | * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks | |||
2918 | * are run. hw_fini tears down the hardware associated with each IP | |||
2919 | * and sw_fini tears down any software state associated with each IP. | |||
2920 | * Returns 0 on success, negative error code on failure. | |||
2921 | */ | |||
2922 | static int amdgpu_device_ip_fini(struct amdgpu_device *adev) | |||
2923 | { | |||
2924 | int i, r; | |||
2925 | ||||
2926 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2)) && adev->virt.ras_init_done) | |||
2927 | amdgpu_virt_release_ras_err_handler_data(adev); | |||
2928 | ||||
2929 | if (adev->gmc.xgmi.num_physical_nodes > 1) | |||
2930 | amdgpu_xgmi_remove_device(adev); | |||
2931 | ||||
2932 | amdgpu_amdkfd_device_fini_sw(adev); | |||
2933 | ||||
2934 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { | |||
2935 | if (!adev->ip_blocks[i].status.sw) | |||
2936 | continue; | |||
2937 | ||||
2938 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) { | |||
2939 | amdgpu_ucode_free_bo(adev); | |||
2940 | amdgpu_free_static_csa(&adev->virt.csa_obj); | |||
2941 | amdgpu_device_wb_fini(adev); | |||
2942 | amdgpu_device_vram_scratch_fini(adev); | |||
2943 | amdgpu_ib_pool_fini(adev); | |||
2944 | } | |||
2945 | ||||
2946 | r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev); | |||
2947 | /* XXX handle errors */ | |||
2948 | if (r) { | |||
2949 | DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",___drm_dbg(((void *)0), DRM_UT_CORE, "sw_fini of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r) | |||
2950 | adev->ip_blocks[i].version->funcs->name, r)___drm_dbg(((void *)0), DRM_UT_CORE, "sw_fini of IP block <%s> failed %d\n" , adev->ip_blocks[i].version->funcs->name, r); | |||
2951 | } | |||
2952 | adev->ip_blocks[i].status.sw = false0; | |||
2953 | adev->ip_blocks[i].status.valid = false0; | |||
2954 | } | |||
2955 | ||||
2956 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { | |||
2957 | if (!adev->ip_blocks[i].status.late_initialized) | |||
2958 | continue; | |||
2959 | if (adev->ip_blocks[i].version->funcs->late_fini) | |||
2960 | adev->ip_blocks[i].version->funcs->late_fini((void *)adev); | |||
2961 | adev->ip_blocks[i].status.late_initialized = false0; | |||
2962 | } | |||
2963 | ||||
2964 | amdgpu_ras_fini(adev); | |||
2965 | ||||
2966 | return 0; | |||
2967 | } | |||
2968 | ||||
2969 | /** | |||
2970 | * amdgpu_device_delayed_init_work_handler - work handler for IB tests | |||
2971 | * | |||
2972 | * @work: work_struct. | |||
2973 | */ | |||
2974 | static void amdgpu_device_delayed_init_work_handler(struct work_struct *work) | |||
2975 | { | |||
2976 | struct amdgpu_device *adev = | |||
2977 | container_of(work, struct amdgpu_device, delayed_init_work.work)({ const __typeof( ((struct amdgpu_device *)0)->delayed_init_work .work ) *__mptr = (work); (struct amdgpu_device *)( (char *)__mptr - __builtin_offsetof(struct amdgpu_device, delayed_init_work .work) );}); | |||
2978 | int r; | |||
2979 | ||||
2980 | r = amdgpu_ib_ring_tests(adev); | |||
2981 | if (r) | |||
2982 | DRM_ERROR("ib ring test failed (%d).\n", r)__drm_err("ib ring test failed (%d).\n", r); | |||
2983 | } | |||
2984 | ||||
2985 | static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work) | |||
2986 | { | |||
2987 | struct amdgpu_device *adev = | |||
2988 | container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work)({ const __typeof( ((struct amdgpu_device *)0)->gfx.gfx_off_delay_work .work ) *__mptr = (work); (struct amdgpu_device *)( (char *)__mptr - __builtin_offsetof(struct amdgpu_device, gfx.gfx_off_delay_work .work) );}); | |||
2989 | ||||
2990 | WARN_ON_ONCE(adev->gfx.gfx_off_state)({ static int __warned; int __ret = !!(adev->gfx.gfx_off_state ); if (__ret && !__warned) { printf("WARNING %s failed at %s:%d\n" , "adev->gfx.gfx_off_state", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_device.c" , 2990); __warned = 1; } __builtin_expect(!!(__ret), 0); }); | |||
2991 | WARN_ON_ONCE(adev->gfx.gfx_off_req_count)({ static int __warned; int __ret = !!(adev->gfx.gfx_off_req_count ); if (__ret && !__warned) { printf("WARNING %s failed at %s:%d\n" , "adev->gfx.gfx_off_req_count", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_device.c" , 2991); __warned = 1; } __builtin_expect(!!(__ret), 0); }); | |||
2992 | ||||
2993 | if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true1)) | |||
2994 | adev->gfx.gfx_off_state = true1; | |||
2995 | } | |||
2996 | ||||
2997 | /** | |||
2998 | * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1) | |||
2999 | * | |||
3000 | * @adev: amdgpu_device pointer | |||
3001 | * | |||
3002 | * Main suspend function for hardware IPs. The list of all the hardware | |||
3003 | * IPs that make up the asic is walked, clockgating is disabled and the | |||
3004 | * suspend callbacks are run. suspend puts the hardware and software state | |||
3005 | * in each IP into a state suitable for suspend. | |||
3006 | * Returns 0 on success, negative error code on failure. | |||
3007 | */ | |||
3008 | static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev) | |||
3009 | { | |||
3010 | int i, r; | |||
3011 | ||||
3012 | amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE); | |||
3013 | amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE); | |||
3014 | ||||
3015 | /* | |||
3016 | * Per PMFW team's suggestion, driver needs to handle gfxoff | |||
3017 | * and df cstate features disablement for gpu reset(e.g. Mode1Reset) | |||
3018 | * scenario. Add the missing df cstate disablement here. | |||
3019 | */ | |||
3020 | if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW)) | |||
3021 | dev_warn(adev->dev, "Failed to disallow df cstate")printf("drm:pid%d:%s *WARNING* " "Failed to disallow df cstate" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
3022 | ||||
3023 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { | |||
3024 | if (!adev->ip_blocks[i].status.valid) | |||
3025 | continue; | |||
3026 | ||||
3027 | /* displays are handled separately */ | |||
3028 | if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_DCE) | |||
3029 | continue; | |||
3030 | ||||
3031 | /* XXX handle errors */ | |||
3032 | r = adev->ip_blocks[i].version->funcs->suspend(adev); | |||
3033 | /* XXX handle errors */ | |||
3034 | if (r) { | |||
3035 | DRM_ERROR("suspend of IP block <%s> failed %d\n",__drm_err("suspend of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r) | |||
3036 | adev->ip_blocks[i].version->funcs->name, r)__drm_err("suspend of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r); | |||
3037 | return r; | |||
3038 | } | |||
3039 | ||||
3040 | adev->ip_blocks[i].status.hw = false0; | |||
3041 | } | |||
3042 | ||||
3043 | return 0; | |||
3044 | } | |||
3045 | ||||
3046 | /** | |||
3047 | * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2) | |||
3048 | * | |||
3049 | * @adev: amdgpu_device pointer | |||
3050 | * | |||
3051 | * Main suspend function for hardware IPs. The list of all the hardware | |||
3052 | * IPs that make up the asic is walked, clockgating is disabled and the | |||
3053 | * suspend callbacks are run. suspend puts the hardware and software state | |||
3054 | * in each IP into a state suitable for suspend. | |||
3055 | * Returns 0 on success, negative error code on failure. | |||
3056 | */ | |||
3057 | static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev) | |||
3058 | { | |||
3059 | int i, r; | |||
3060 | ||||
3061 | if (adev->in_s0ix) | |||
3062 | amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D3Entry); | |||
3063 | ||||
3064 | for (i = adev->num_ip_blocks - 1; i >= 0; i--) { | |||
3065 | if (!adev->ip_blocks[i].status.valid) | |||
3066 | continue; | |||
3067 | /* displays are handled in phase1 */ | |||
3068 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) | |||
3069 | continue; | |||
3070 | /* PSP lost connection when err_event_athub occurs */ | |||
3071 | if (amdgpu_ras_intr_triggered() && | |||
3072 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) { | |||
3073 | adev->ip_blocks[i].status.hw = false0; | |||
3074 | continue; | |||
3075 | } | |||
3076 | ||||
3077 | /* skip unnecessary suspend if we do not initialize them yet */ | |||
3078 | if (adev->gmc.xgmi.pending_reset && | |||
3079 | !(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || | |||
3080 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC || | |||
3081 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || | |||
3082 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)) { | |||
3083 | adev->ip_blocks[i].status.hw = false0; | |||
3084 | continue; | |||
3085 | } | |||
3086 | ||||
3087 | /* skip suspend of gfx/mes and psp for S0ix | |||
3088 | * gfx is in gfxoff state, so on resume it will exit gfxoff just | |||
3089 | * like at runtime. PSP is also part of the always on hardware | |||
3090 | * so no need to suspend it. | |||
3091 | */ | |||
3092 | if (adev->in_s0ix && | |||
3093 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP || | |||
3094 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX || | |||
3095 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_MES)) | |||
3096 | continue; | |||
3097 | ||||
3098 | /* SDMA 5.x+ is part of GFX power domain so it's covered by GFXOFF */ | |||
3099 | if (adev->in_s0ix && | |||
3100 | (adev->ip_versions[SDMA0_HWIP][0] >= IP_VERSION(5, 0, 0)(((5) << 16) | ((0) << 8) | (0))) && | |||
3101 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA)) | |||
3102 | continue; | |||
3103 | ||||
3104 | /* Once swPSP provides the IMU, RLC FW binaries to TOS during cold-boot. | |||
3105 | * These are in TMR, hence are expected to be reused by PSP-TOS to reload | |||
3106 | * from this location and RLC Autoload automatically also gets loaded | |||
3107 | * from here based on PMFW -> PSP message during re-init sequence. | |||
3108 | * Therefore, the psp suspend & resume should be skipped to avoid destroy | |||
3109 | * the TMR and reload FWs again for IMU enabled APU ASICs. | |||
3110 | */ | |||
3111 | if (amdgpu_in_reset(adev) && | |||
3112 | (adev->flags & AMD_IS_APU) && adev->gfx.imu.funcs && | |||
3113 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) | |||
3114 | continue; | |||
3115 | ||||
3116 | /* XXX handle errors */ | |||
3117 | r = adev->ip_blocks[i].version->funcs->suspend(adev); | |||
3118 | /* XXX handle errors */ | |||
3119 | if (r) { | |||
3120 | DRM_ERROR("suspend of IP block <%s> failed %d\n",__drm_err("suspend of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r) | |||
3121 | adev->ip_blocks[i].version->funcs->name, r)__drm_err("suspend of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r); | |||
3122 | } | |||
3123 | adev->ip_blocks[i].status.hw = false0; | |||
3124 | /* handle putting the SMC in the appropriate state */ | |||
3125 | if(!amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))){ | |||
3126 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) { | |||
3127 | r = amdgpu_dpm_set_mp1_state(adev, adev->mp1_state); | |||
3128 | if (r) { | |||
3129 | DRM_ERROR("SMC failed to set mp1 state %d, %d\n",__drm_err("SMC failed to set mp1 state %d, %d\n", adev->mp1_state , r) | |||
3130 | adev->mp1_state, r)__drm_err("SMC failed to set mp1 state %d, %d\n", adev->mp1_state , r); | |||
3131 | return r; | |||
3132 | } | |||
3133 | } | |||
3134 | } | |||
3135 | } | |||
3136 | ||||
3137 | return 0; | |||
3138 | } | |||
3139 | ||||
3140 | /** | |||
3141 | * amdgpu_device_ip_suspend - run suspend for hardware IPs | |||
3142 | * | |||
3143 | * @adev: amdgpu_device pointer | |||
3144 | * | |||
3145 | * Main suspend function for hardware IPs. The list of all the hardware | |||
3146 | * IPs that make up the asic is walked, clockgating is disabled and the | |||
3147 | * suspend callbacks are run. suspend puts the hardware and software state | |||
3148 | * in each IP into a state suitable for suspend. | |||
3149 | * Returns 0 on success, negative error code on failure. | |||
3150 | */ | |||
3151 | int amdgpu_device_ip_suspend(struct amdgpu_device *adev) | |||
3152 | { | |||
3153 | int r; | |||
3154 | ||||
3155 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) { | |||
3156 | amdgpu_virt_fini_data_exchange(adev); | |||
3157 | amdgpu_virt_request_full_gpu(adev, false0); | |||
3158 | } | |||
3159 | ||||
3160 | r = amdgpu_device_ip_suspend_phase1(adev); | |||
3161 | if (r) | |||
3162 | return r; | |||
3163 | r = amdgpu_device_ip_suspend_phase2(adev); | |||
3164 | ||||
3165 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) | |||
3166 | amdgpu_virt_release_full_gpu(adev, false0); | |||
3167 | ||||
3168 | return r; | |||
3169 | } | |||
3170 | ||||
3171 | static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev) | |||
3172 | { | |||
3173 | int i, r; | |||
3174 | ||||
3175 | static enum amd_ip_block_type ip_order[] = { | |||
3176 | AMD_IP_BLOCK_TYPE_COMMON, | |||
3177 | AMD_IP_BLOCK_TYPE_GMC, | |||
3178 | AMD_IP_BLOCK_TYPE_PSP, | |||
3179 | AMD_IP_BLOCK_TYPE_IH, | |||
3180 | }; | |||
3181 | ||||
3182 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
3183 | int j; | |||
3184 | struct amdgpu_ip_block *block; | |||
3185 | ||||
3186 | block = &adev->ip_blocks[i]; | |||
3187 | block->status.hw = false0; | |||
3188 | ||||
3189 | for (j = 0; j < ARRAY_SIZE(ip_order)(sizeof((ip_order)) / sizeof((ip_order)[0])); j++) { | |||
3190 | ||||
3191 | if (block->version->type != ip_order[j] || | |||
3192 | !block->status.valid) | |||
3193 | continue; | |||
3194 | ||||
3195 | r = block->version->funcs->hw_init(adev); | |||
3196 | DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded")printk("\0016" "[" "drm" "] " "RE-INIT-early: %s %s\n", block ->version->funcs->name, r?"failed":"succeeded"); | |||
3197 | if (r) | |||
3198 | return r; | |||
3199 | block->status.hw = true1; | |||
3200 | } | |||
3201 | } | |||
3202 | ||||
3203 | return 0; | |||
3204 | } | |||
3205 | ||||
3206 | static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev) | |||
3207 | { | |||
3208 | int i, r; | |||
3209 | ||||
3210 | static enum amd_ip_block_type ip_order[] = { | |||
3211 | AMD_IP_BLOCK_TYPE_SMC, | |||
3212 | AMD_IP_BLOCK_TYPE_DCE, | |||
3213 | AMD_IP_BLOCK_TYPE_GFX, | |||
3214 | AMD_IP_BLOCK_TYPE_SDMA, | |||
3215 | AMD_IP_BLOCK_TYPE_UVD, | |||
3216 | AMD_IP_BLOCK_TYPE_VCE, | |||
3217 | AMD_IP_BLOCK_TYPE_VCN | |||
3218 | }; | |||
3219 | ||||
3220 | for (i = 0; i < ARRAY_SIZE(ip_order)(sizeof((ip_order)) / sizeof((ip_order)[0])); i++) { | |||
3221 | int j; | |||
3222 | struct amdgpu_ip_block *block; | |||
3223 | ||||
3224 | for (j = 0; j < adev->num_ip_blocks; j++) { | |||
3225 | block = &adev->ip_blocks[j]; | |||
3226 | ||||
3227 | if (block->version->type != ip_order[i] || | |||
3228 | !block->status.valid || | |||
3229 | block->status.hw) | |||
3230 | continue; | |||
3231 | ||||
3232 | if (block->version->type == AMD_IP_BLOCK_TYPE_SMC) | |||
3233 | r = block->version->funcs->resume(adev); | |||
3234 | else | |||
3235 | r = block->version->funcs->hw_init(adev); | |||
3236 | ||||
3237 | DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded")printk("\0016" "[" "drm" "] " "RE-INIT-late: %s %s\n", block-> version->funcs->name, r?"failed":"succeeded"); | |||
3238 | if (r) | |||
3239 | return r; | |||
3240 | block->status.hw = true1; | |||
3241 | } | |||
3242 | } | |||
3243 | ||||
3244 | return 0; | |||
3245 | } | |||
3246 | ||||
3247 | /** | |||
3248 | * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs | |||
3249 | * | |||
3250 | * @adev: amdgpu_device pointer | |||
3251 | * | |||
3252 | * First resume function for hardware IPs. The list of all the hardware | |||
3253 | * IPs that make up the asic is walked and the resume callbacks are run for | |||
3254 | * COMMON, GMC, and IH. resume puts the hardware into a functional state | |||
3255 | * after a suspend and updates the software state as necessary. This | |||
3256 | * function is also used for restoring the GPU after a GPU reset. | |||
3257 | * Returns 0 on success, negative error code on failure. | |||
3258 | */ | |||
3259 | static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev) | |||
3260 | { | |||
3261 | int i, r; | |||
3262 | ||||
3263 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
3264 | if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) | |||
3265 | continue; | |||
3266 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || | |||
3267 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || | |||
3268 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH || | |||
3269 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP && amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2)))) { | |||
3270 | ||||
3271 | r = adev->ip_blocks[i].version->funcs->resume(adev); | |||
3272 | if (r) { | |||
3273 | DRM_ERROR("resume of IP block <%s> failed %d\n",__drm_err("resume of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r) | |||
3274 | adev->ip_blocks[i].version->funcs->name, r)__drm_err("resume of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r); | |||
3275 | return r; | |||
3276 | } | |||
3277 | adev->ip_blocks[i].status.hw = true1; | |||
3278 | } | |||
3279 | } | |||
3280 | ||||
3281 | return 0; | |||
3282 | } | |||
3283 | ||||
3284 | /** | |||
3285 | * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs | |||
3286 | * | |||
3287 | * @adev: amdgpu_device pointer | |||
3288 | * | |||
3289 | * First resume function for hardware IPs. The list of all the hardware | |||
3290 | * IPs that make up the asic is walked and the resume callbacks are run for | |||
3291 | * all blocks except COMMON, GMC, and IH. resume puts the hardware into a | |||
3292 | * functional state after a suspend and updates the software state as | |||
3293 | * necessary. This function is also used for restoring the GPU after a GPU | |||
3294 | * reset. | |||
3295 | * Returns 0 on success, negative error code on failure. | |||
3296 | */ | |||
3297 | static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev) | |||
3298 | { | |||
3299 | int i, r; | |||
3300 | ||||
3301 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
3302 | if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) | |||
3303 | continue; | |||
3304 | if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || | |||
3305 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || | |||
3306 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH || | |||
3307 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) | |||
3308 | continue; | |||
3309 | r = adev->ip_blocks[i].version->funcs->resume(adev); | |||
3310 | if (r) { | |||
3311 | DRM_ERROR("resume of IP block <%s> failed %d\n",__drm_err("resume of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r) | |||
3312 | adev->ip_blocks[i].version->funcs->name, r)__drm_err("resume of IP block <%s> failed %d\n", adev-> ip_blocks[i].version->funcs->name, r); | |||
3313 | return r; | |||
3314 | } | |||
3315 | adev->ip_blocks[i].status.hw = true1; | |||
3316 | ||||
3317 | if (adev->in_s0ix && adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) { | |||
3318 | /* disable gfxoff for IP resume. The gfxoff will be re-enabled in | |||
3319 | * amdgpu_device_resume() after IP resume. | |||
3320 | */ | |||
3321 | amdgpu_gfx_off_ctrl(adev, false0); | |||
3322 | DRM_DEBUG("will disable gfxoff for re-initializing other blocks\n")___drm_dbg(((void *)0), DRM_UT_CORE, "will disable gfxoff for re-initializing other blocks\n" ); | |||
3323 | } | |||
3324 | ||||
3325 | } | |||
3326 | ||||
3327 | return 0; | |||
3328 | } | |||
3329 | ||||
3330 | /** | |||
3331 | * amdgpu_device_ip_resume - run resume for hardware IPs | |||
3332 | * | |||
3333 | * @adev: amdgpu_device pointer | |||
3334 | * | |||
3335 | * Main resume function for hardware IPs. The hardware IPs | |||
3336 | * are split into two resume functions because they are | |||
3337 | * are also used in in recovering from a GPU reset and some additional | |||
3338 | * steps need to be take between them. In this case (S3/S4) they are | |||
3339 | * run sequentially. | |||
3340 | * Returns 0 on success, negative error code on failure. | |||
3341 | */ | |||
3342 | static int amdgpu_device_ip_resume(struct amdgpu_device *adev) | |||
3343 | { | |||
3344 | int r; | |||
3345 | ||||
3346 | r = amdgpu_amdkfd_resume_iommu(adev); | |||
3347 | if (r) | |||
3348 | return r; | |||
3349 | ||||
3350 | r = amdgpu_device_ip_resume_phase1(adev); | |||
3351 | if (r) | |||
3352 | return r; | |||
3353 | ||||
3354 | r = amdgpu_device_fw_loading(adev); | |||
3355 | if (r) | |||
3356 | return r; | |||
3357 | ||||
3358 | r = amdgpu_device_ip_resume_phase2(adev); | |||
3359 | ||||
3360 | return r; | |||
3361 | } | |||
3362 | ||||
3363 | /** | |||
3364 | * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV | |||
3365 | * | |||
3366 | * @adev: amdgpu_device pointer | |||
3367 | * | |||
3368 | * Query the VBIOS data tables to determine if the board supports SR-IOV. | |||
3369 | */ | |||
3370 | static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev) | |||
3371 | { | |||
3372 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) { | |||
3373 | if (adev->is_atom_fw) { | |||
3374 | if (amdgpu_atomfirmware_gpu_virtualization_supported(adev)) | |||
3375 | adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS(1 << 0); | |||
3376 | } else { | |||
3377 | if (amdgpu_atombios_has_gpu_virtualization_table(adev)) | |||
3378 | adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS(1 << 0); | |||
3379 | } | |||
3380 | ||||
3381 | if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS(1 << 0))) | |||
3382 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0); | |||
3383 | } | |||
3384 | } | |||
3385 | ||||
3386 | /** | |||
3387 | * amdgpu_device_asic_has_dc_support - determine if DC supports the asic | |||
3388 | * | |||
3389 | * @asic_type: AMD asic type | |||
3390 | * | |||
3391 | * Check if there is DC (new modesetting infrastructre) support for an asic. | |||
3392 | * returns true if DC has support, false if not. | |||
3393 | */ | |||
3394 | bool_Bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type) | |||
3395 | { | |||
3396 | switch (asic_type) { | |||
3397 | #ifdef CONFIG_DRM_AMDGPU_SI | |||
3398 | case CHIP_HAINAN: | |||
3399 | #endif | |||
3400 | case CHIP_TOPAZ: | |||
3401 | /* chips with no display hardware */ | |||
3402 | return false0; | |||
3403 | #if defined(CONFIG_DRM_AMD_DC1) | |||
3404 | case CHIP_TAHITI: | |||
3405 | case CHIP_PITCAIRN: | |||
3406 | case CHIP_VERDE: | |||
3407 | case CHIP_OLAND: | |||
3408 | /* | |||
3409 | * We have systems in the wild with these ASICs that require | |||
3410 | * LVDS and VGA support which is not supported with DC. | |||
3411 | * | |||
3412 | * Fallback to the non-DC driver here by default so as not to | |||
3413 | * cause regressions. | |||
3414 | */ | |||
3415 | #if defined(CONFIG_DRM_AMD_DC_SI) | |||
3416 | return amdgpu_dc > 0; | |||
3417 | #else | |||
3418 | return false0; | |||
3419 | #endif | |||
3420 | case CHIP_BONAIRE: | |||
3421 | case CHIP_KAVERI: | |||
3422 | case CHIP_KABINI: | |||
3423 | case CHIP_MULLINS: | |||
3424 | /* | |||
3425 | * We have systems in the wild with these ASICs that require | |||
3426 | * VGA support which is not supported with DC. | |||
3427 | * | |||
3428 | * Fallback to the non-DC driver here by default so as not to | |||
3429 | * cause regressions. | |||
3430 | */ | |||
3431 | return amdgpu_dc > 0; | |||
3432 | default: | |||
3433 | return amdgpu_dc != 0; | |||
3434 | #else | |||
3435 | default: | |||
3436 | if (amdgpu_dc > 0) | |||
3437 | DRM_INFO_ONCE("Display Core has been requested via kernel parameter "({ static int __warned; if (!__warned) { printk("\0016" "[" "drm" "] " "Display Core has been requested via kernel parameter " "but isn't supported by ASIC, ignoring\n"); __warned = 1; } } ) | |||
3438 | "but isn't supported by ASIC, ignoring\n")({ static int __warned; if (!__warned) { printk("\0016" "[" "drm" "] " "Display Core has been requested via kernel parameter " "but isn't supported by ASIC, ignoring\n"); __warned = 1; } } ); | |||
3439 | return false0; | |||
3440 | #endif | |||
3441 | } | |||
3442 | } | |||
3443 | ||||
3444 | /** | |||
3445 | * amdgpu_device_has_dc_support - check if dc is supported | |||
3446 | * | |||
3447 | * @adev: amdgpu_device pointer | |||
3448 | * | |||
3449 | * Returns true for supported, false for not supported | |||
3450 | */ | |||
3451 | bool_Bool amdgpu_device_has_dc_support(struct amdgpu_device *adev) | |||
3452 | { | |||
3453 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2)) || | |||
3454 | adev->enable_virtual_display || | |||
3455 | (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK)) | |||
3456 | return false0; | |||
3457 | ||||
3458 | return amdgpu_device_asic_has_dc_support(adev->asic_type); | |||
3459 | } | |||
3460 | ||||
3461 | static void amdgpu_device_xgmi_reset_func(struct work_struct *__work) | |||
3462 | { | |||
3463 | struct amdgpu_device *adev = | |||
3464 | container_of(__work, struct amdgpu_device, xgmi_reset_work)({ const __typeof( ((struct amdgpu_device *)0)->xgmi_reset_work ) *__mptr = (__work); (struct amdgpu_device *)( (char *)__mptr - __builtin_offsetof(struct amdgpu_device, xgmi_reset_work) ) ;}); | |||
3465 | struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev); | |||
3466 | ||||
3467 | /* It's a bug to not have a hive within this function */ | |||
3468 | if (WARN_ON(!hive)({ int __ret = !!(!hive); if (__ret) printf("WARNING %s failed at %s:%d\n" , "!hive", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_device.c" , 3468); __builtin_expect(!!(__ret), 0); })) | |||
3469 | return; | |||
3470 | ||||
3471 | /* | |||
3472 | * Use task barrier to synchronize all xgmi reset works across the | |||
3473 | * hive. task_barrier_enter and task_barrier_exit will block | |||
3474 | * until all the threads running the xgmi reset works reach | |||
3475 | * those points. task_barrier_full will do both blocks. | |||
3476 | */ | |||
3477 | if (amdgpu_asic_reset_method(adev)(adev)->asic_funcs->reset_method((adev)) == AMD_RESET_METHOD_BACO) { | |||
3478 | ||||
3479 | task_barrier_enter(&hive->tb); | |||
3480 | adev->asic_reset_res = amdgpu_device_baco_enter(adev_to_drm(adev)); | |||
3481 | ||||
3482 | if (adev->asic_reset_res) | |||
3483 | goto fail; | |||
3484 | ||||
3485 | task_barrier_exit(&hive->tb); | |||
3486 | adev->asic_reset_res = amdgpu_device_baco_exit(adev_to_drm(adev)); | |||
3487 | ||||
3488 | if (adev->asic_reset_res) | |||
3489 | goto fail; | |||
3490 | ||||
3491 | if (adev->mmhub.ras && adev->mmhub.ras->ras_block.hw_ops && | |||
3492 | adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count) | |||
3493 | adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count(adev); | |||
3494 | } else { | |||
3495 | ||||
3496 | task_barrier_full(&hive->tb); | |||
3497 | adev->asic_reset_res = amdgpu_asic_reset(adev)(adev)->asic_funcs->reset((adev)); | |||
3498 | } | |||
3499 | ||||
3500 | fail: | |||
3501 | if (adev->asic_reset_res) | |||
3502 | DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",printk("\0014" "[" "drm" "] " "ASIC reset failed with error, %d for drm dev, %s" , adev->asic_reset_res, adev_to_drm(adev)->unique) | |||
3503 | adev->asic_reset_res, adev_to_drm(adev)->unique)printk("\0014" "[" "drm" "] " "ASIC reset failed with error, %d for drm dev, %s" , adev->asic_reset_res, adev_to_drm(adev)->unique); | |||
3504 | amdgpu_put_xgmi_hive(hive); | |||
3505 | } | |||
3506 | ||||
3507 | static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev) | |||
3508 | { | |||
3509 | char *input = amdgpu_lockup_timeout; | |||
3510 | char *timeout_setting = NULL((void *)0); | |||
3511 | int index = 0; | |||
3512 | long timeout; | |||
3513 | int ret = 0; | |||
3514 | ||||
3515 | /* | |||
3516 | * By default timeout for non compute jobs is 10000 | |||
3517 | * and 60000 for compute jobs. | |||
3518 | * In SR-IOV or passthrough mode, timeout for compute | |||
3519 | * jobs are 60000 by default. | |||
3520 | */ | |||
3521 | adev->gfx_timeout = msecs_to_jiffies(10000)(((uint64_t)(10000)) * hz / 1000); | |||
3522 | adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout; | |||
3523 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) | |||
3524 | adev->compute_timeout = amdgpu_sriov_is_pp_one_vf(adev)((adev)->virt.gim_feature & AMDGIM_FEATURE_PP_ONE_VF) ? | |||
3525 | msecs_to_jiffies(60000)(((uint64_t)(60000)) * hz / 1000) : msecs_to_jiffies(10000)(((uint64_t)(10000)) * hz / 1000); | |||
3526 | else | |||
3527 | adev->compute_timeout = msecs_to_jiffies(60000)(((uint64_t)(60000)) * hz / 1000); | |||
3528 | ||||
3529 | #ifdef notyet | |||
3530 | if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH256)) { | |||
3531 | while ((timeout_setting = strsep(&input, ",")) && | |||
3532 | strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH256)) { | |||
3533 | ret = kstrtol(timeout_setting, 0, &timeout); | |||
3534 | if (ret) | |||
3535 | return ret; | |||
3536 | ||||
3537 | if (timeout == 0) { | |||
3538 | index++; | |||
3539 | continue; | |||
3540 | } else if (timeout < 0) { | |||
3541 | timeout = MAX_SCHEDULE_TIMEOUT(0x7fffffff); | |||
3542 | dev_warn(adev->dev, "lockup timeout disabled")printf("drm:pid%d:%s *WARNING* " "lockup timeout disabled", ( {struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
3543 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); | |||
3544 | } else { | |||
3545 | timeout = msecs_to_jiffies(timeout)(((uint64_t)(timeout)) * hz / 1000); | |||
3546 | } | |||
3547 | ||||
3548 | switch (index++) { | |||
3549 | case 0: | |||
3550 | adev->gfx_timeout = timeout; | |||
3551 | break; | |||
3552 | case 1: | |||
3553 | adev->compute_timeout = timeout; | |||
3554 | break; | |||
3555 | case 2: | |||
3556 | adev->sdma_timeout = timeout; | |||
3557 | break; | |||
3558 | case 3: | |||
3559 | adev->video_timeout = timeout; | |||
3560 | break; | |||
3561 | default: | |||
3562 | break; | |||
3563 | } | |||
3564 | } | |||
3565 | /* | |||
3566 | * There is only one value specified and | |||
3567 | * it should apply to all non-compute jobs. | |||
3568 | */ | |||
3569 | if (index == 1) { | |||
3570 | adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout; | |||
3571 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2)) || amdgpu_passthrough(adev)((adev)->virt.caps & (1 << 3))) | |||
3572 | adev->compute_timeout = adev->gfx_timeout; | |||
3573 | } | |||
3574 | } | |||
3575 | #endif | |||
3576 | ||||
3577 | return ret; | |||
3578 | } | |||
3579 | ||||
3580 | /** | |||
3581 | * amdgpu_device_check_iommu_direct_map - check if RAM direct mapped to GPU | |||
3582 | * | |||
3583 | * @adev: amdgpu_device pointer | |||
3584 | * | |||
3585 | * RAM direct mapped to GPU if IOMMU is not enabled or is pass through mode | |||
3586 | */ | |||
3587 | static void amdgpu_device_check_iommu_direct_map(struct amdgpu_device *adev) | |||
3588 | { | |||
3589 | #ifdef notyet | |||
3590 | struct iommu_domain *domain; | |||
3591 | ||||
3592 | domain = iommu_get_domain_for_dev(adev->dev); | |||
3593 | if (!domain || domain->type == IOMMU_DOMAIN_IDENTITY) | |||
3594 | #endif | |||
3595 | adev->ram_is_direct_mapped = true1; | |||
3596 | } | |||
3597 | ||||
3598 | static const struct attribute *amdgpu_dev_attributes[] = { | |||
3599 | &dev_attr_product_name.attr, | |||
3600 | &dev_attr_product_number.attr, | |||
3601 | &dev_attr_serial_number.attr, | |||
3602 | &dev_attr_pcie_replay_count.attr, | |||
3603 | NULL((void *)0) | |||
3604 | }; | |||
3605 | ||||
3606 | /** | |||
3607 | * amdgpu_device_init - initialize the driver | |||
3608 | * | |||
3609 | * @adev: amdgpu_device pointer | |||
3610 | * @flags: driver flags | |||
3611 | * | |||
3612 | * Initializes the driver info and hw (all asics). | |||
3613 | * Returns 0 for success or an error on failure. | |||
3614 | * Called at driver startup. | |||
3615 | */ | |||
3616 | int amdgpu_device_init(struct amdgpu_device *adev, | |||
3617 | uint32_t flags) | |||
3618 | { | |||
3619 | struct drm_device *ddev = adev_to_drm(adev); | |||
3620 | struct pci_dev *pdev = adev->pdev; | |||
3621 | int r, i; | |||
3622 | bool_Bool px = false0; | |||
3623 | u32 max_MBps; | |||
3624 | int tmp; | |||
3625 | ||||
3626 | adev->shutdown = false0; | |||
3627 | adev->flags = flags; | |||
3628 | ||||
3629 | if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST) | |||
3630 | adev->asic_type = amdgpu_force_asic_type; | |||
3631 | else | |||
3632 | adev->asic_type = flags & AMD_ASIC_MASK; | |||
3633 | ||||
3634 | adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT100000; | |||
3635 | if (amdgpu_emu_mode == 1) | |||
3636 | adev->usec_timeout *= 10; | |||
3637 | adev->gmc.gart_size = 512 * 1024 * 1024; | |||
3638 | adev->accel_working = false0; | |||
3639 | adev->num_rings = 0; | |||
3640 | RCU_INIT_POINTER(adev->gang_submit, dma_fence_get_stub())do { (adev->gang_submit) = (dma_fence_get_stub()); } while (0); | |||
3641 | adev->mman.buffer_funcs = NULL((void *)0); | |||
3642 | adev->mman.buffer_funcs_ring = NULL((void *)0); | |||
3643 | adev->vm_manager.vm_pte_funcs = NULL((void *)0); | |||
3644 | adev->vm_manager.vm_pte_num_scheds = 0; | |||
3645 | adev->gmc.gmc_funcs = NULL((void *)0); | |||
3646 | adev->harvest_ip_mask = 0x0; | |||
3647 | adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS28); | |||
3648 | bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES128); | |||
3649 | ||||
3650 | adev->smc_rreg = &amdgpu_invalid_rreg; | |||
3651 | adev->smc_wreg = &amdgpu_invalid_wreg; | |||
3652 | adev->pcie_rreg = &amdgpu_invalid_rreg; | |||
3653 | adev->pcie_wreg = &amdgpu_invalid_wreg; | |||
3654 | adev->pciep_rreg = &amdgpu_invalid_rreg; | |||
3655 | adev->pciep_wreg = &amdgpu_invalid_wreg; | |||
3656 | adev->pcie_rreg64 = &amdgpu_invalid_rreg64; | |||
3657 | adev->pcie_wreg64 = &amdgpu_invalid_wreg64; | |||
3658 | adev->uvd_ctx_rreg = &amdgpu_invalid_rreg; | |||
3659 | adev->uvd_ctx_wreg = &amdgpu_invalid_wreg; | |||
3660 | adev->didt_rreg = &amdgpu_invalid_rreg; | |||
3661 | adev->didt_wreg = &amdgpu_invalid_wreg; | |||
3662 | adev->gc_cac_rreg = &amdgpu_invalid_rreg; | |||
3663 | adev->gc_cac_wreg = &amdgpu_invalid_wreg; | |||
3664 | adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg; | |||
3665 | adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg; | |||
3666 | ||||
3667 | DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",printk("\0016" "[" "drm" "] " "initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n" , amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev ->device, pdev->subsystem_vendor, pdev->subsystem_device , pdev->revision) | |||
3668 | amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,printk("\0016" "[" "drm" "] " "initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n" , amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev ->device, pdev->subsystem_vendor, pdev->subsystem_device , pdev->revision) | |||
3669 | pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision)printk("\0016" "[" "drm" "] " "initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n" , amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev ->device, pdev->subsystem_vendor, pdev->subsystem_device , pdev->revision); | |||
3670 | ||||
3671 | /* mutex initialization are all done here so we | |||
3672 | * can recall function without having locking issues */ | |||
3673 | rw_init(&adev->firmware.mutex, "agfw")_rw_init_flags(&adev->firmware.mutex, "agfw", 0, ((void *)0)); | |||
3674 | rw_init(&adev->pm.mutex, "agpm")_rw_init_flags(&adev->pm.mutex, "agpm", 0, ((void *)0) ); | |||
3675 | rw_init(&adev->gfx.gpu_clock_mutex, "gfxclk")_rw_init_flags(&adev->gfx.gpu_clock_mutex, "gfxclk", 0 , ((void *)0)); | |||
3676 | rw_init(&adev->srbm_mutex, "srbm")_rw_init_flags(&adev->srbm_mutex, "srbm", 0, ((void *) 0)); | |||
3677 | rw_init(&adev->gfx.pipe_reserve_mutex, "pipers")_rw_init_flags(&adev->gfx.pipe_reserve_mutex, "pipers" , 0, ((void *)0)); | |||
3678 | rw_init(&adev->gfx.gfx_off_mutex, "gfxoff")_rw_init_flags(&adev->gfx.gfx_off_mutex, "gfxoff", 0, ( (void *)0)); | |||
3679 | rw_init(&adev->grbm_idx_mutex, "grbmidx")_rw_init_flags(&adev->grbm_idx_mutex, "grbmidx", 0, (( void *)0)); | |||
3680 | rw_init(&adev->mn_lock, "agpumn")_rw_init_flags(&adev->mn_lock, "agpumn", 0, ((void *)0 )); | |||
3681 | rw_init(&adev->virt.vf_errors.lock, "vferr")_rw_init_flags(&adev->virt.vf_errors.lock, "vferr", 0, ((void *)0)); | |||
3682 | hash_init(adev->mn_hash)__hash_init(adev->mn_hash, (sizeof((adev->mn_hash)) / sizeof ((adev->mn_hash)[0]))); | |||
3683 | rw_init(&adev->psp.mutex, "agpsp")_rw_init_flags(&adev->psp.mutex, "agpsp", 0, ((void *) 0)); | |||
3684 | rw_init(&adev->notifier_lock, "agnf")_rw_init_flags(&adev->notifier_lock, "agnf", 0, ((void *)0)); | |||
3685 | rw_init(&adev->pm.stable_pstate_ctx_lock, "agps")_rw_init_flags(&adev->pm.stable_pstate_ctx_lock, "agps" , 0, ((void *)0)); | |||
3686 | rw_init(&adev->benchmark_mutex, "agbm")_rw_init_flags(&adev->benchmark_mutex, "agbm", 0, ((void *)0)); | |||
3687 | ||||
3688 | amdgpu_device_init_apu_flags(adev); | |||
3689 | ||||
3690 | r = amdgpu_device_check_arguments(adev); | |||
3691 | if (r) | |||
3692 | return r; | |||
3693 | ||||
3694 | mtx_init(&adev->mmio_idx_lock, IPL_TTY)do { (void)(((void *)0)); (void)(0); __mtx_init((&adev-> mmio_idx_lock), ((((0x9)) > 0x0 && ((0x9)) < 0x9 ) ? 0x9 : ((0x9)))); } while (0); | |||
3695 | mtx_init(&adev->smc_idx_lock, IPL_TTY)do { (void)(((void *)0)); (void)(0); __mtx_init((&adev-> smc_idx_lock), ((((0x9)) > 0x0 && ((0x9)) < 0x9 ) ? 0x9 : ((0x9)))); } while (0); | |||
3696 | mtx_init(&adev->pcie_idx_lock, IPL_TTY)do { (void)(((void *)0)); (void)(0); __mtx_init((&adev-> pcie_idx_lock), ((((0x9)) > 0x0 && ((0x9)) < 0x9 ) ? 0x9 : ((0x9)))); } while (0); | |||
3697 | mtx_init(&adev->uvd_ctx_idx_lock, IPL_TTY)do { (void)(((void *)0)); (void)(0); __mtx_init((&adev-> uvd_ctx_idx_lock), ((((0x9)) > 0x0 && ((0x9)) < 0x9) ? 0x9 : ((0x9)))); } while (0); | |||
3698 | mtx_init(&adev->didt_idx_lock, IPL_TTY)do { (void)(((void *)0)); (void)(0); __mtx_init((&adev-> didt_idx_lock), ((((0x9)) > 0x0 && ((0x9)) < 0x9 ) ? 0x9 : ((0x9)))); } while (0); | |||
3699 | mtx_init(&adev->gc_cac_idx_lock, IPL_TTY)do { (void)(((void *)0)); (void)(0); __mtx_init((&adev-> gc_cac_idx_lock), ((((0x9)) > 0x0 && ((0x9)) < 0x9 ) ? 0x9 : ((0x9)))); } while (0); | |||
3700 | mtx_init(&adev->se_cac_idx_lock, IPL_TTY)do { (void)(((void *)0)); (void)(0); __mtx_init((&adev-> se_cac_idx_lock), ((((0x9)) > 0x0 && ((0x9)) < 0x9 ) ? 0x9 : ((0x9)))); } while (0); | |||
3701 | mtx_init(&adev->audio_endpt_idx_lock, IPL_TTY)do { (void)(((void *)0)); (void)(0); __mtx_init((&adev-> audio_endpt_idx_lock), ((((0x9)) > 0x0 && ((0x9)) < 0x9) ? 0x9 : ((0x9)))); } while (0); | |||
3702 | mtx_init(&adev->mm_stats.lock, IPL_NONE)do { (void)(((void *)0)); (void)(0); __mtx_init((&adev-> mm_stats.lock), ((((0x0)) > 0x0 && ((0x0)) < 0x9 ) ? 0x9 : ((0x0)))); } while (0); | |||
3703 | ||||
3704 | INIT_LIST_HEAD(&adev->shadow_list); | |||
3705 | rw_init(&adev->shadow_list_lock, "sdwlst")_rw_init_flags(&adev->shadow_list_lock, "sdwlst", 0, ( (void *)0)); | |||
3706 | ||||
3707 | INIT_LIST_HEAD(&adev->reset_list); | |||
3708 | ||||
3709 | INIT_LIST_HEAD(&adev->ras_list); | |||
3710 | ||||
3711 | INIT_DELAYED_WORK(&adev->delayed_init_work, | |||
3712 | amdgpu_device_delayed_init_work_handler); | |||
3713 | INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work, | |||
3714 | amdgpu_device_delay_enable_gfx_off); | |||
3715 | ||||
3716 | INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func); | |||
3717 | ||||
3718 | adev->gfx.gfx_off_req_count = 1; | |||
3719 | adev->gfx.gfx_off_residency = 0; | |||
3720 | adev->gfx.gfx_off_entrycount = 0; | |||
3721 | adev->pm.ac_power = power_supply_is_system_supplied() > 0; | |||
3722 | ||||
3723 | atomic_set(&adev->throttling_logging_enabled, 1)({ typeof(*(&adev->throttling_logging_enabled)) __tmp = ((1)); *(volatile typeof(*(&adev->throttling_logging_enabled )) *)&(*(&adev->throttling_logging_enabled)) = __tmp ; __tmp; }); | |||
3724 | /* | |||
3725 | * If throttling continues, logging will be performed every minute | |||
3726 | * to avoid log flooding. "-1" is subtracted since the thermal | |||
3727 | * throttling interrupt comes every second. Thus, the total logging | |||
3728 | * interval is 59 seconds(retelimited printk interval) + 1(waiting | |||
3729 | * for throttling interrupt) = 60 seconds. | |||
3730 | */ | |||
3731 | ratelimit_state_init(&adev->throttling_logging_rs, (60 - 1) * HZhz, 1); | |||
3732 | ratelimit_set_flags(&adev->throttling_logging_rs, RATELIMIT_MSG_ON_RELEASE(1 << 0)); | |||
3733 | ||||
3734 | #ifdef __linux__ | |||
3735 | /* Registers mapping */ | |||
3736 | /* TODO: block userspace mapping of io register */ | |||
3737 | if (adev->asic_type >= CHIP_BONAIRE) { | |||
3738 | adev->rmmio_base = pci_resource_start(adev->pdev, 5); | |||
3739 | adev->rmmio_size = pci_resource_len(adev->pdev, 5); | |||
3740 | } else { | |||
3741 | adev->rmmio_base = pci_resource_start(adev->pdev, 2); | |||
3742 | adev->rmmio_size = pci_resource_len(adev->pdev, 2); | |||
3743 | } | |||
3744 | ||||
3745 | for (i = 0; i < AMD_IP_BLOCK_TYPE_NUM; i++) | |||
3746 | atomic_set(&adev->pm.pwr_state[i], POWER_STATE_UNKNOWN)({ typeof(*(&adev->pm.pwr_state[i])) __tmp = ((POWER_STATE_UNKNOWN )); *(volatile typeof(*(&adev->pm.pwr_state[i])) *)& (*(&adev->pm.pwr_state[i])) = __tmp; __tmp; }); | |||
3747 | ||||
3748 | adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size); | |||
3749 | if (adev->rmmio == NULL((void *)0)) { | |||
3750 | return -ENOMEM12; | |||
3751 | } | |||
3752 | #endif | |||
3753 | DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base)printk("\0016" "[" "drm" "] " "register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base); | |||
3754 | DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size)printk("\0016" "[" "drm" "] " "register mmio size: %u\n", (unsigned )adev->rmmio_size); | |||
3755 | ||||
3756 | amdgpu_device_get_pcie_info(adev); | |||
3757 | ||||
3758 | if (amdgpu_mcbp) | |||
3759 | DRM_INFO("MCBP is enabled\n")printk("\0016" "[" "drm" "] " "MCBP is enabled\n"); | |||
3760 | ||||
3761 | /* | |||
3762 | * Reset domain needs to be present early, before XGMI hive discovered | |||
3763 | * (if any) and intitialized to use reset sem and in_gpu reset flag | |||
3764 | * early on during init and before calling to RREG32. | |||
3765 | */ | |||
3766 | adev->reset_domain = amdgpu_reset_create_reset_domain(SINGLE_DEVICE, "amdgpu-reset-dev"); | |||
3767 | if (!adev->reset_domain) | |||
3768 | return -ENOMEM12; | |||
3769 | ||||
3770 | /* detect hw virtualization here */ | |||
3771 | amdgpu_detect_virtualization(adev); | |||
3772 | ||||
3773 | r = amdgpu_device_get_job_timeout_settings(adev); | |||
3774 | if (r) { | |||
3775 | dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n")printf("drm:pid%d:%s *ERROR* " "invalid lockup_timeout parameter syntax\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
3776 | return r; | |||
3777 | } | |||
3778 | ||||
3779 | /* early init functions */ | |||
3780 | r = amdgpu_device_ip_early_init(adev); | |||
3781 | if (r) | |||
3782 | return r; | |||
3783 | ||||
3784 | /* Get rid of things like offb */ | |||
3785 | r = drm_aperture_remove_conflicting_pci_framebuffers(adev->pdev, &amdgpu_kms_driver); | |||
3786 | if (r) | |||
3787 | return r; | |||
3788 | ||||
3789 | /* Enable TMZ based on IP_VERSION */ | |||
3790 | amdgpu_gmc_tmz_set(adev); | |||
3791 | ||||
3792 | amdgpu_gmc_noretry_set(adev); | |||
3793 | /* Need to get xgmi info early to decide the reset behavior*/ | |||
3794 | if (adev->gmc.xgmi.supported) { | |||
3795 | r = adev->gfxhub.funcs->get_xgmi_info(adev); | |||
3796 | if (r) | |||
3797 | return r; | |||
3798 | } | |||
3799 | ||||
3800 | /* enable PCIE atomic ops */ | |||
3801 | #ifdef notyet | |||
3802 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) | |||
3803 | adev->have_atomics_support = ((struct amd_sriov_msg_pf2vf_info *) | |||
3804 | adev->virt.fw_reserve.p_pf2vf)->pcie_atomic_ops_support_flags == | |||
3805 | (PCI_EXP_DEVCAP2_ATOMIC_COMP32 | PCI_EXP_DEVCAP2_ATOMIC_COMP64); | |||
3806 | /* APUs w/ gfx9 onwards doesn't reply on PCIe atomics, rather it is a | |||
3807 | * internal path natively support atomics, set have_atomics_support to true. | |||
3808 | */ | |||
3809 | else if ((adev->flags & AMD_IS_APU) && | |||
3810 | (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0)(((9) << 16) | ((0) << 8) | (0)))) | |||
3811 | adev->have_atomics_support = true1; | |||
3812 | else | |||
3813 | adev->have_atomics_support = | |||
3814 | !pci_enable_atomic_ops_to_root(adev->pdev, | |||
3815 | PCI_EXP_DEVCAP2_ATOMIC_COMP32 | | |||
3816 | PCI_EXP_DEVCAP2_ATOMIC_COMP64); | |||
3817 | if (!adev->have_atomics_support) | |||
3818 | dev_info(adev->dev, "PCIE atomic ops is not supported\n")do { } while(0); | |||
3819 | #else | |||
3820 | /* APUs w/ gfx9 onwards doesn't reply on PCIe atomics, rather it is a | |||
3821 | * internal path natively support atomics, set have_atomics_support to true. | |||
3822 | */ | |||
3823 | if ((adev->flags & AMD_IS_APU) && | |||
3824 | (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0)(((9) << 16) | ((0) << 8) | (0)))) | |||
3825 | adev->have_atomics_support = true1; | |||
3826 | else | |||
3827 | adev->have_atomics_support = false0; | |||
3828 | #endif | |||
3829 | ||||
3830 | /* doorbell bar mapping and doorbell index init*/ | |||
3831 | amdgpu_device_doorbell_init(adev); | |||
3832 | ||||
3833 | if (amdgpu_emu_mode == 1) { | |||
3834 | /* post the asic on emulation mode */ | |||
3835 | emu_soc_asic_init(adev); | |||
3836 | goto fence_driver_init; | |||
3837 | } | |||
3838 | ||||
3839 | amdgpu_reset_init(adev); | |||
3840 | ||||
3841 | /* detect if we are with an SRIOV vbios */ | |||
3842 | amdgpu_device_detect_sriov_bios(adev); | |||
3843 | ||||
3844 | /* check if we need to reset the asic | |||
3845 | * E.g., driver was not cleanly unloaded previously, etc. | |||
3846 | */ | |||
3847 | if (!amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2)) && amdgpu_asic_need_reset_on_init(adev)(adev)->asic_funcs->need_reset_on_init((adev))) { | |||
3848 | if (adev->gmc.xgmi.num_physical_nodes) { | |||
3849 | dev_info(adev->dev, "Pending hive reset.\n")do { } while(0); | |||
3850 | adev->gmc.xgmi.pending_reset = true1; | |||
3851 | /* Only need to init necessary block for SMU to handle the reset */ | |||
3852 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
3853 | if (!adev->ip_blocks[i].status.valid) | |||
3854 | continue; | |||
3855 | if (!(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || | |||
3856 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || | |||
3857 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH || | |||
3858 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC)) { | |||
3859 | DRM_DEBUG("IP %s disabled for hw_init.\n",___drm_dbg(((void *)0), DRM_UT_CORE, "IP %s disabled for hw_init.\n" , adev->ip_blocks[i].version->funcs->name) | |||
3860 | adev->ip_blocks[i].version->funcs->name)___drm_dbg(((void *)0), DRM_UT_CORE, "IP %s disabled for hw_init.\n" , adev->ip_blocks[i].version->funcs->name); | |||
3861 | adev->ip_blocks[i].status.hw = true1; | |||
3862 | } | |||
3863 | } | |||
3864 | } else { | |||
3865 | tmp = amdgpu_reset_method; | |||
3866 | /* It should do a default reset when loading or reloading the driver, | |||
3867 | * regardless of the module parameter reset_method. | |||
3868 | */ | |||
3869 | amdgpu_reset_method = AMD_RESET_METHOD_NONE; | |||
3870 | r = amdgpu_asic_reset(adev)(adev)->asic_funcs->reset((adev)); | |||
3871 | amdgpu_reset_method = tmp; | |||
3872 | if (r) { | |||
3873 | dev_err(adev->dev, "asic reset on init failed\n")printf("drm:pid%d:%s *ERROR* " "asic reset on init failed\n", ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
3874 | goto failed; | |||
3875 | } | |||
3876 | } | |||
3877 | } | |||
3878 | ||||
3879 | pci_enable_pcie_error_reporting(adev->pdev); | |||
3880 | ||||
3881 | /* Post card if necessary */ | |||
3882 | if (amdgpu_device_need_post(adev)) { | |||
3883 | if (!adev->bios) { | |||
3884 | dev_err(adev->dev, "no vBIOS found\n")printf("drm:pid%d:%s *ERROR* " "no vBIOS found\n", ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof (struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p-> ps_pid, __func__); | |||
3885 | r = -EINVAL22; | |||
3886 | goto failed; | |||
3887 | } | |||
3888 | DRM_INFO("GPU posting now...\n")printk("\0016" "[" "drm" "] " "GPU posting now...\n"); | |||
3889 | r = amdgpu_device_asic_init(adev); | |||
3890 | if (r) { | |||
3891 | dev_err(adev->dev, "gpu post error!\n")printf("drm:pid%d:%s *ERROR* " "gpu post error!\n", ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof (struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p-> ps_pid, __func__); | |||
3892 | goto failed; | |||
3893 | } | |||
3894 | } | |||
3895 | ||||
3896 | if (adev->is_atom_fw) { | |||
3897 | /* Initialize clocks */ | |||
3898 | r = amdgpu_atomfirmware_get_clock_info(adev); | |||
3899 | if (r) { | |||
3900 | dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n")printf("drm:pid%d:%s *ERROR* " "amdgpu_atomfirmware_get_clock_info failed\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
3901 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0); | |||
3902 | goto failed; | |||
3903 | } | |||
3904 | } else { | |||
3905 | /* Initialize clocks */ | |||
3906 | r = amdgpu_atombios_get_clock_info(adev); | |||
3907 | if (r) { | |||
3908 | dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n")printf("drm:pid%d:%s *ERROR* " "amdgpu_atombios_get_clock_info failed\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
3909 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0); | |||
3910 | goto failed; | |||
3911 | } | |||
3912 | /* init i2c buses */ | |||
3913 | if (!amdgpu_device_has_dc_support(adev)) | |||
3914 | amdgpu_atombios_i2c_init(adev); | |||
3915 | } | |||
3916 | ||||
3917 | fence_driver_init: | |||
3918 | /* Fence driver */ | |||
3919 | r = amdgpu_fence_driver_sw_init(adev); | |||
3920 | if (r) { | |||
3921 | dev_err(adev->dev, "amdgpu_fence_driver_sw_init failed\n")printf("drm:pid%d:%s *ERROR* " "amdgpu_fence_driver_sw_init failed\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
3922 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0); | |||
3923 | goto failed; | |||
3924 | } | |||
3925 | ||||
3926 | /* init the mode config */ | |||
3927 | drm_mode_config_init(adev_to_drm(adev)); | |||
3928 | ||||
3929 | r = amdgpu_device_ip_init(adev); | |||
3930 | if (r) { | |||
3931 | dev_err(adev->dev, "amdgpu_device_ip_init failed\n")printf("drm:pid%d:%s *ERROR* " "amdgpu_device_ip_init failed\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
3932 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0); | |||
3933 | goto release_ras_con; | |||
3934 | } | |||
3935 | ||||
3936 | amdgpu_fence_driver_hw_init(adev); | |||
3937 | ||||
3938 | dev_info(adev->dev,do { } while(0) | |||
3939 | "SE %d, SH per SE %d, CU per SH %d, active_cu_number %d\n",do { } while(0) | |||
3940 | adev->gfx.config.max_shader_engines,do { } while(0) | |||
3941 | adev->gfx.config.max_sh_per_se,do { } while(0) | |||
3942 | adev->gfx.config.max_cu_per_sh,do { } while(0) | |||
3943 | adev->gfx.cu_info.number)do { } while(0); | |||
3944 | ||||
3945 | #ifdef __OpenBSD__1 | |||
3946 | { | |||
3947 | const char *chip_name; | |||
3948 | uint32_t version = adev->ip_versions[GC_HWIP][0]; | |||
3949 | int maj, min, rev; | |||
3950 | ||||
3951 | switch (adev->asic_type) { | |||
3952 | case CHIP_RAVEN: | |||
3953 | if (adev->apu_flags & AMD_APU_IS_RAVEN2) | |||
3954 | chip_name = "RAVEN2"; | |||
3955 | else if (adev->apu_flags & AMD_APU_IS_PICASSO) | |||
3956 | chip_name = "PICASSO"; | |||
3957 | else | |||
3958 | chip_name = "RAVEN"; | |||
3959 | break; | |||
3960 | case CHIP_RENOIR: | |||
3961 | if (adev->apu_flags & AMD_APU_IS_RENOIR) | |||
3962 | chip_name = "RENOIR"; | |||
3963 | else | |||
3964 | chip_name = "GREEN_SARDINE"; | |||
3965 | break; | |||
3966 | default: | |||
3967 | chip_name = amdgpu_asic_name[adev->asic_type]; | |||
3968 | } | |||
3969 | ||||
3970 | printf("%s: %s", adev->self.dv_xname, chip_name); | |||
3971 | /* show graphics/compute ip block version, not set on < GFX9 */ | |||
3972 | if (version) { | |||
3973 | maj = IP_VERSION_MAJ(version)((version) >> 16); | |||
3974 | min = IP_VERSION_MIN(version)(((version) >> 8) & 0xFF); | |||
3975 | rev = IP_VERSION_REV(version)((version) & 0xFF); | |||
3976 | printf(" GC %d.%d.%d", maj, min, rev); | |||
3977 | } | |||
3978 | printf(" %d CU rev 0x%02x\n", adev->gfx.cu_info.number, adev->rev_id); | |||
3979 | } | |||
3980 | #endif | |||
3981 | ||||
3982 | adev->accel_working = true1; | |||
3983 | ||||
3984 | amdgpu_vm_check_compute_bug(adev); | |||
3985 | ||||
3986 | /* Initialize the buffer migration limit. */ | |||
3987 | if (amdgpu_moverate >= 0) | |||
3988 | max_MBps = amdgpu_moverate; | |||
3989 | else | |||
3990 | max_MBps = 8; /* Allow 8 MB/s. */ | |||
3991 | /* Get a log2 for easy divisions. */ | |||
3992 | adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps))((sizeof((((1u)>(max_MBps))?(1u):(max_MBps))) <= 4) ? ( fls((((1u)>(max_MBps))?(1u):(max_MBps))) - 1) : (flsl((((1u )>(max_MBps))?(1u):(max_MBps))) - 1)); | |||
3993 | ||||
3994 | r = amdgpu_pm_sysfs_init(adev); | |||
3995 | if (r) { | |||
3996 | adev->pm_sysfs_en = false0; | |||
3997 | DRM_ERROR("registering pm debugfs failed (%d).\n", r)__drm_err("registering pm debugfs failed (%d).\n", r); | |||
3998 | } else | |||
3999 | adev->pm_sysfs_en = true1; | |||
4000 | ||||
4001 | r = amdgpu_ucode_sysfs_init(adev); | |||
4002 | if (r) { | |||
4003 | adev->ucode_sysfs_en = false0; | |||
4004 | DRM_ERROR("Creating firmware sysfs failed (%d).\n", r)__drm_err("Creating firmware sysfs failed (%d).\n", r); | |||
4005 | } else | |||
4006 | adev->ucode_sysfs_en = true1; | |||
4007 | ||||
4008 | r = amdgpu_psp_sysfs_init(adev); | |||
4009 | if (r) { | |||
4010 | adev->psp_sysfs_en = false0; | |||
4011 | if (!amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) | |||
4012 | DRM_ERROR("Creating psp sysfs failed\n")__drm_err("Creating psp sysfs failed\n"); | |||
4013 | } else | |||
4014 | adev->psp_sysfs_en = true1; | |||
4015 | ||||
4016 | /* | |||
4017 | * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost. | |||
4018 | * Otherwise the mgpu fan boost feature will be skipped due to the | |||
4019 | * gpu instance is counted less. | |||
4020 | */ | |||
4021 | amdgpu_register_gpu_instance(adev); | |||
4022 | ||||
4023 | /* enable clockgating, etc. after ib tests, etc. since some blocks require | |||
4024 | * explicit gating rather than handling it automatically. | |||
4025 | */ | |||
4026 | if (!adev->gmc.xgmi.pending_reset) { | |||
4027 | r = amdgpu_device_ip_late_init(adev); | |||
4028 | if (r) { | |||
4029 | dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n")printf("drm:pid%d:%s *ERROR* " "amdgpu_device_ip_late_init failed\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
4030 | amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r); | |||
4031 | goto release_ras_con; | |||
4032 | } | |||
4033 | /* must succeed. */ | |||
4034 | amdgpu_ras_resume(adev); | |||
4035 | queue_delayed_work(system_wq, &adev->delayed_init_work, | |||
4036 | msecs_to_jiffies(AMDGPU_RESUME_MS)(((uint64_t)(2000)) * hz / 1000)); | |||
4037 | } | |||
4038 | ||||
4039 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) { | |||
4040 | amdgpu_virt_release_full_gpu(adev, true1); | |||
4041 | flush_delayed_work(&adev->delayed_init_work); | |||
4042 | } | |||
4043 | ||||
4044 | r = sysfs_create_files(&adev->dev->kobj, amdgpu_dev_attributes)0; | |||
4045 | if (r) | |||
4046 | dev_err(adev->dev, "Could not create amdgpu device attr\n")printf("drm:pid%d:%s *ERROR* " "Could not create amdgpu device attr\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
4047 | ||||
4048 | if (IS_ENABLED(CONFIG_PERF_EVENTS)0) | |||
4049 | r = amdgpu_pmu_init(adev); | |||
4050 | if (r) | |||
4051 | dev_err(adev->dev, "amdgpu_pmu_init failed\n")printf("drm:pid%d:%s *ERROR* " "amdgpu_pmu_init failed\n", ({ struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
4052 | ||||
4053 | /* Have stored pci confspace at hand for restore in sudden PCI error */ | |||
4054 | if (amdgpu_device_cache_pci_state(adev->pdev)) | |||
4055 | pci_restore_state(pdev); | |||
4056 | ||||
4057 | /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */ | |||
4058 | /* this will fail for cards that aren't VGA class devices, just | |||
4059 | * ignore it */ | |||
4060 | #ifdef notyet | |||
4061 | if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA((0x03 << 8) | 0x00)) | |||
4062 | vga_client_register(adev->pdev, amdgpu_device_vga_set_decode); | |||
4063 | #endif | |||
4064 | ||||
4065 | px = amdgpu_device_supports_px(ddev); | |||
4066 | ||||
4067 | if (px || (!dev_is_removable(&adev->pdev->dev)0 && | |||
4068 | apple_gmux_detect(NULL((void *)0), NULL((void *)0)))) | |||
4069 | vga_switcheroo_register_client(adev->pdev, | |||
4070 | &amdgpu_switcheroo_ops, px); | |||
4071 | ||||
4072 | if (px) | |||
4073 | vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain); | |||
4074 | ||||
4075 | if (adev->gmc.xgmi.pending_reset) | |||
4076 | queue_delayed_work(system_wq, &mgpu_info.delayed_reset_work, | |||
4077 | msecs_to_jiffies(AMDGPU_RESUME_MS)(((uint64_t)(2000)) * hz / 1000)); | |||
4078 | ||||
4079 | amdgpu_device_check_iommu_direct_map(adev); | |||
4080 | ||||
4081 | return 0; | |||
4082 | ||||
4083 | release_ras_con: | |||
4084 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) | |||
4085 | amdgpu_virt_release_full_gpu(adev, true1); | |||
4086 | ||||
4087 | /* failed in exclusive mode due to timeout */ | |||
4088 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2)) && | |||
4089 | !amdgpu_sriov_runtime(adev)((adev)->virt.caps & (1 << 4)) && | |||
4090 | amdgpu_virt_mmio_blocked(adev) && | |||
4091 | !amdgpu_virt_wait_reset(adev)) { | |||
4092 | dev_err(adev->dev, "VF exclusive mode timeout\n")printf("drm:pid%d:%s *ERROR* " "VF exclusive mode timeout\n", ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
4093 | /* Don't send request since VF is inactive. */ | |||
4094 | adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME(1 << 4); | |||
4095 | adev->virt.ops = NULL((void *)0); | |||
4096 | r = -EAGAIN35; | |||
4097 | } | |||
4098 | amdgpu_release_ras_context(adev); | |||
4099 | ||||
4100 | failed: | |||
4101 | amdgpu_vf_error_trans_all(adev); | |||
4102 | ||||
4103 | return r; | |||
4104 | } | |||
4105 | ||||
4106 | static void amdgpu_device_unmap_mmio(struct amdgpu_device *adev) | |||
4107 | { | |||
4108 | STUB()do { printf("%s: stub\n", __func__); } while(0); | |||
4109 | #ifdef notyet | |||
4110 | /* Clear all CPU mappings pointing to this device */ | |||
4111 | unmap_mapping_range(adev->ddev.anon_inode->i_mapping, 0, 0, 1); | |||
4112 | #endif | |||
4113 | ||||
4114 | /* Unmap all mapped bars - Doorbell, registers and VRAM */ | |||
4115 | amdgpu_device_doorbell_fini(adev); | |||
4116 | ||||
4117 | #ifdef __linux__ | |||
4118 | iounmap(adev->rmmio); | |||
4119 | adev->rmmio = NULL((void *)0); | |||
4120 | if (adev->mman.aper_base_kaddr) | |||
4121 | iounmap(adev->mman.aper_base_kaddr); | |||
4122 | adev->mman.aper_base_kaddr = NULL((void *)0); | |||
4123 | #else | |||
4124 | if (adev->rmmio_size > 0) | |||
4125 | bus_space_unmap(adev->rmmio_bst, adev->rmmio_bsh, | |||
4126 | adev->rmmio_size); | |||
4127 | adev->rmmio_size = 0; | |||
4128 | adev->rmmio = NULL((void *)0); | |||
4129 | if (adev->mman.aper_base_kaddr) | |||
4130 | bus_space_unmap(adev->memt, adev->mman.aper_bsh, | |||
4131 | adev->gmc.visible_vram_size); | |||
4132 | adev->mman.aper_base_kaddr = NULL((void *)0); | |||
4133 | #endif | |||
4134 | ||||
4135 | /* Memory manager related */ | |||
4136 | if (!adev->gmc.xgmi.connected_to_cpu) { | |||
4137 | #ifdef __linux__ | |||
4138 | arch_phys_wc_del(adev->gmc.vram_mtrr); | |||
4139 | arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size); | |||
4140 | #else | |||
4141 | drm_mtrr_del(0, adev->gmc.aper_base, adev->gmc.aper_size, DRM_MTRR_WC(1<<1)); | |||
4142 | #endif | |||
4143 | } | |||
4144 | } | |||
4145 | ||||
4146 | /** | |||
4147 | * amdgpu_device_fini_hw - tear down the driver | |||
4148 | * | |||
4149 | * @adev: amdgpu_device pointer | |||
4150 | * | |||
4151 | * Tear down the driver info (all asics). | |||
4152 | * Called at driver shutdown. | |||
4153 | */ | |||
4154 | void amdgpu_device_fini_hw(struct amdgpu_device *adev) | |||
4155 | { | |||
4156 | dev_info(adev->dev, "amdgpu: finishing device.\n")do { } while(0); | |||
4157 | flush_delayed_work(&adev->delayed_init_work); | |||
4158 | adev->shutdown = true1; | |||
4159 | ||||
4160 | /* make sure IB test finished before entering exclusive mode | |||
4161 | * to avoid preemption on IB test | |||
4162 | * */ | |||
4163 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) { | |||
4164 | amdgpu_virt_request_full_gpu(adev, false0); | |||
4165 | amdgpu_virt_fini_data_exchange(adev); | |||
4166 | } | |||
4167 | ||||
4168 | /* disable all interrupts */ | |||
4169 | amdgpu_irq_disable_all(adev); | |||
4170 | if (adev->mode_info.mode_config_initialized){ | |||
4171 | if (!drm_drv_uses_atomic_modeset(adev_to_drm(adev))) | |||
4172 | drm_helper_force_disable_all(adev_to_drm(adev)); | |||
4173 | else | |||
4174 | drm_atomic_helper_shutdown(adev_to_drm(adev)); | |||
4175 | } | |||
4176 | amdgpu_fence_driver_hw_fini(adev); | |||
4177 | ||||
4178 | if (adev->mman.initialized) { | |||
4179 | flush_delayed_work(&adev->mman.bdev.wq); | |||
4180 | ttm_bo_lock_delayed_workqueue(&adev->mman.bdev); | |||
4181 | } | |||
4182 | ||||
4183 | if (adev->pm_sysfs_en) | |||
4184 | amdgpu_pm_sysfs_fini(adev); | |||
4185 | if (adev->ucode_sysfs_en) | |||
4186 | amdgpu_ucode_sysfs_fini(adev); | |||
4187 | if (adev->psp_sysfs_en) | |||
4188 | amdgpu_psp_sysfs_fini(adev); | |||
4189 | sysfs_remove_files(&adev->dev->kobj, amdgpu_dev_attributes); | |||
4190 | ||||
4191 | /* disable ras feature must before hw fini */ | |||
4192 | amdgpu_ras_pre_fini(adev); | |||
4193 | ||||
4194 | amdgpu_device_ip_fini_early(adev); | |||
4195 | ||||
4196 | amdgpu_irq_fini_hw(adev); | |||
4197 | ||||
4198 | if (adev->mman.initialized) | |||
4199 | ttm_device_clear_dma_mappings(&adev->mman.bdev); | |||
4200 | ||||
4201 | amdgpu_gart_dummy_page_fini(adev); | |||
4202 | ||||
4203 | if (drm_dev_is_unplugged(adev_to_drm(adev))) | |||
4204 | amdgpu_device_unmap_mmio(adev); | |||
4205 | ||||
4206 | } | |||
4207 | ||||
4208 | void amdgpu_device_fini_sw(struct amdgpu_device *adev) | |||
4209 | { | |||
4210 | int idx; | |||
4211 | bool_Bool px; | |||
4212 | ||||
4213 | amdgpu_fence_driver_sw_fini(adev); | |||
4214 | amdgpu_device_ip_fini(adev); | |||
4215 | release_firmware(adev->firmware.gpu_info_fw); | |||
4216 | adev->firmware.gpu_info_fw = NULL((void *)0); | |||
4217 | adev->accel_working = false0; | |||
4218 | dma_fence_put(rcu_dereference_protected(adev->gang_submit, true)(adev->gang_submit)); | |||
4219 | ||||
4220 | amdgpu_reset_fini(adev); | |||
4221 | ||||
4222 | /* free i2c buses */ | |||
4223 | if (!amdgpu_device_has_dc_support(adev)) | |||
4224 | amdgpu_i2c_fini(adev); | |||
4225 | ||||
4226 | if (amdgpu_emu_mode != 1) | |||
4227 | amdgpu_atombios_fini(adev); | |||
4228 | ||||
4229 | kfree(adev->bios); | |||
4230 | adev->bios = NULL((void *)0); | |||
4231 | ||||
4232 | px = amdgpu_device_supports_px(adev_to_drm(adev)); | |||
4233 | ||||
4234 | if (px || (!dev_is_removable(&adev->pdev->dev)0 && | |||
4235 | apple_gmux_detect(NULL((void *)0), NULL((void *)0)))) | |||
4236 | vga_switcheroo_unregister_client(adev->pdev); | |||
4237 | ||||
4238 | if (px) | |||
4239 | vga_switcheroo_fini_domain_pm_ops(adev->dev); | |||
4240 | ||||
4241 | if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA((0x03 << 8) | 0x00)) | |||
4242 | vga_client_unregister(adev->pdev); | |||
4243 | ||||
4244 | if (drm_dev_enter(adev_to_drm(adev), &idx)) { | |||
4245 | #ifdef __linux__ | |||
4246 | iounmap(adev->rmmio); | |||
4247 | adev->rmmio = NULL((void *)0); | |||
4248 | #else | |||
4249 | if (adev->rmmio_size > 0) | |||
4250 | bus_space_unmap(adev->rmmio_bst, adev->rmmio_bsh, | |||
4251 | adev->rmmio_size); | |||
4252 | adev->rmmio_size = 0; | |||
4253 | adev->rmmio = NULL((void *)0); | |||
4254 | #endif | |||
4255 | amdgpu_device_doorbell_fini(adev); | |||
4256 | drm_dev_exit(idx); | |||
4257 | } | |||
4258 | ||||
4259 | if (IS_ENABLED(CONFIG_PERF_EVENTS)0) | |||
4260 | amdgpu_pmu_fini(adev); | |||
4261 | if (adev->mman.discovery_bin) | |||
4262 | amdgpu_discovery_fini(adev); | |||
4263 | ||||
4264 | amdgpu_reset_put_reset_domain(adev->reset_domain); | |||
4265 | adev->reset_domain = NULL((void *)0); | |||
4266 | ||||
4267 | kfree(adev->pci_state); | |||
4268 | ||||
4269 | } | |||
4270 | ||||
4271 | /** | |||
4272 | * amdgpu_device_evict_resources - evict device resources | |||
4273 | * @adev: amdgpu device object | |||
4274 | * | |||
4275 | * Evicts all ttm device resources(vram BOs, gart table) from the lru list | |||
4276 | * of the vram memory type. Mainly used for evicting device resources | |||
4277 | * at suspend time. | |||
4278 | * | |||
4279 | */ | |||
4280 | static int amdgpu_device_evict_resources(struct amdgpu_device *adev) | |||
4281 | { | |||
4282 | int ret; | |||
4283 | ||||
4284 | /* No need to evict vram on APUs for suspend to ram or s2idle */ | |||
4285 | if ((adev->in_s3 || adev->in_s0ix) && (adev->flags & AMD_IS_APU)) | |||
4286 | return 0; | |||
4287 | ||||
4288 | ret = amdgpu_ttm_evict_resources(adev, TTM_PL_VRAM2); | |||
4289 | if (ret) | |||
4290 | DRM_WARN("evicting device resources failed\n")printk("\0014" "[" "drm" "] " "evicting device resources failed\n" ); | |||
4291 | return ret; | |||
4292 | } | |||
4293 | ||||
4294 | /* | |||
4295 | * Suspend & resume. | |||
4296 | */ | |||
4297 | /** | |||
4298 | * amdgpu_device_suspend - initiate device suspend | |||
4299 | * | |||
4300 | * @dev: drm dev pointer | |||
4301 | * @fbcon : notify the fbdev of suspend | |||
4302 | * | |||
4303 | * Puts the hw in the suspend state (all asics). | |||
4304 | * Returns 0 for success or an error on failure. | |||
4305 | * Called at driver suspend. | |||
4306 | */ | |||
4307 | int amdgpu_device_suspend(struct drm_device *dev, bool_Bool fbcon) | |||
4308 | { | |||
4309 | struct amdgpu_device *adev = drm_to_adev(dev); | |||
4310 | int r = 0; | |||
4311 | ||||
4312 | if (adev->shutdown) | |||
4313 | return 0; | |||
4314 | ||||
4315 | #ifdef notyet | |||
4316 | if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) | |||
4317 | return 0; | |||
4318 | #endif | |||
4319 | ||||
4320 | adev->in_suspend = true1; | |||
4321 | ||||
4322 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) { | |||
4323 | amdgpu_virt_fini_data_exchange(adev); | |||
4324 | r = amdgpu_virt_request_full_gpu(adev, false0); | |||
4325 | if (r) | |||
4326 | return r; | |||
4327 | } | |||
4328 | ||||
4329 | if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D3)) | |||
4330 | DRM_WARN("smart shift update failed\n")printk("\0014" "[" "drm" "] " "smart shift update failed\n"); | |||
4331 | ||||
4332 | drm_kms_helper_poll_disable(dev); | |||
4333 | ||||
4334 | if (fbcon) | |||
4335 | drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, true1); | |||
4336 | ||||
4337 | cancel_delayed_work_sync(&adev->delayed_init_work); | |||
4338 | flush_delayed_work(&adev->gfx.gfx_off_delay_work); | |||
4339 | ||||
4340 | amdgpu_ras_suspend(adev); | |||
4341 | ||||
4342 | amdgpu_device_ip_suspend_phase1(adev); | |||
4343 | ||||
4344 | if (!adev->in_s0ix) | |||
4345 | amdgpu_amdkfd_suspend(adev, adev->in_runpm); | |||
4346 | ||||
4347 | r = amdgpu_device_evict_resources(adev); | |||
4348 | if (r) | |||
4349 | return r; | |||
4350 | ||||
4351 | amdgpu_fence_driver_hw_fini(adev); | |||
4352 | ||||
4353 | amdgpu_device_ip_suspend_phase2(adev); | |||
4354 | ||||
4355 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) | |||
4356 | amdgpu_virt_release_full_gpu(adev, false0); | |||
4357 | ||||
4358 | return 0; | |||
4359 | } | |||
4360 | ||||
4361 | /** | |||
4362 | * amdgpu_device_resume - initiate device resume | |||
4363 | * | |||
4364 | * @dev: drm dev pointer | |||
4365 | * @fbcon : notify the fbdev of resume | |||
4366 | * | |||
4367 | * Bring the hw back to operating state (all asics). | |||
4368 | * Returns 0 for success or an error on failure. | |||
4369 | * Called at driver resume. | |||
4370 | */ | |||
4371 | int amdgpu_device_resume(struct drm_device *dev, bool_Bool fbcon) | |||
4372 | { | |||
4373 | struct amdgpu_device *adev = drm_to_adev(dev); | |||
4374 | int r = 0; | |||
4375 | ||||
4376 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) { | |||
4377 | r = amdgpu_virt_request_full_gpu(adev, true1); | |||
4378 | if (r) | |||
4379 | return r; | |||
4380 | } | |||
4381 | ||||
4382 | #ifdef notyet | |||
4383 | if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) | |||
4384 | return 0; | |||
4385 | #endif | |||
4386 | ||||
4387 | if (adev->in_s0ix) | |||
4388 | amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D0Entry); | |||
4389 | ||||
4390 | /* post card */ | |||
4391 | if (amdgpu_device_need_post(adev)) { | |||
4392 | r = amdgpu_device_asic_init(adev); | |||
4393 | if (r) | |||
4394 | dev_err(adev->dev, "amdgpu asic init failed\n")printf("drm:pid%d:%s *ERROR* " "amdgpu asic init failed\n", ( {struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
4395 | } | |||
4396 | ||||
4397 | r = amdgpu_device_ip_resume(adev); | |||
4398 | ||||
4399 | /* no matter what r is, always need to properly release full GPU */ | |||
4400 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) { | |||
4401 | amdgpu_virt_init_data_exchange(adev); | |||
4402 | amdgpu_virt_release_full_gpu(adev, true1); | |||
4403 | } | |||
4404 | ||||
4405 | if (r) { | |||
4406 | dev_err(adev->dev, "amdgpu_device_ip_resume failed (%d).\n", r)printf("drm:pid%d:%s *ERROR* " "amdgpu_device_ip_resume failed (%d).\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , r); | |||
4407 | return r; | |||
4408 | } | |||
4409 | amdgpu_fence_driver_hw_init(adev); | |||
4410 | ||||
4411 | r = amdgpu_device_ip_late_init(adev); | |||
4412 | if (r) | |||
4413 | return r; | |||
4414 | ||||
4415 | queue_delayed_work(system_wq, &adev->delayed_init_work, | |||
4416 | msecs_to_jiffies(AMDGPU_RESUME_MS)(((uint64_t)(2000)) * hz / 1000)); | |||
4417 | ||||
4418 | if (!adev->in_s0ix) { | |||
4419 | r = amdgpu_amdkfd_resume(adev, adev->in_runpm); | |||
4420 | if (r) | |||
4421 | return r; | |||
4422 | } | |||
4423 | ||||
4424 | /* Make sure IB tests flushed */ | |||
4425 | flush_delayed_work(&adev->delayed_init_work); | |||
4426 | ||||
4427 | if (adev->in_s0ix) { | |||
4428 | /* re-enable gfxoff after IP resume. This re-enables gfxoff after | |||
4429 | * it was disabled for IP resume in amdgpu_device_ip_resume_phase2(). | |||
4430 | */ | |||
4431 | amdgpu_gfx_off_ctrl(adev, true1); | |||
4432 | DRM_DEBUG("will enable gfxoff for the mission mode\n")___drm_dbg(((void *)0), DRM_UT_CORE, "will enable gfxoff for the mission mode\n" ); | |||
4433 | } | |||
4434 | if (fbcon) | |||
4435 | drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, false0); | |||
4436 | ||||
4437 | drm_kms_helper_poll_enable(dev); | |||
4438 | ||||
4439 | amdgpu_ras_resume(adev); | |||
4440 | ||||
4441 | /* | |||
4442 | * Most of the connector probing functions try to acquire runtime pm | |||
4443 | * refs to ensure that the GPU is powered on when connector polling is | |||
4444 | * performed. Since we're calling this from a runtime PM callback, | |||
4445 | * trying to acquire rpm refs will cause us to deadlock. | |||
4446 | * | |||
4447 | * Since we're guaranteed to be holding the rpm lock, it's safe to | |||
4448 | * temporarily disable the rpm helpers so this doesn't deadlock us. | |||
4449 | */ | |||
4450 | #if defined(CONFIG_PM) && defined(__linux__) | |||
4451 | dev->dev->power.disable_depth++; | |||
4452 | #endif | |||
4453 | if (!amdgpu_device_has_dc_support(adev)) | |||
4454 | drm_helper_hpd_irq_event(dev); | |||
4455 | else | |||
4456 | drm_kms_helper_hotplug_event(dev); | |||
4457 | #if defined(CONFIG_PM) && defined(__linux__) | |||
4458 | dev->dev->power.disable_depth--; | |||
4459 | #endif | |||
4460 | adev->in_suspend = false0; | |||
4461 | ||||
4462 | if (adev->enable_mes) | |||
4463 | amdgpu_mes_self_test(adev); | |||
4464 | ||||
4465 | if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D0)) | |||
4466 | DRM_WARN("smart shift update failed\n")printk("\0014" "[" "drm" "] " "smart shift update failed\n"); | |||
4467 | ||||
4468 | return 0; | |||
4469 | } | |||
4470 | ||||
4471 | /** | |||
4472 | * amdgpu_device_ip_check_soft_reset - did soft reset succeed | |||
4473 | * | |||
4474 | * @adev: amdgpu_device pointer | |||
4475 | * | |||
4476 | * The list of all the hardware IPs that make up the asic is walked and | |||
4477 | * the check_soft_reset callbacks are run. check_soft_reset determines | |||
4478 | * if the asic is still hung or not. | |||
4479 | * Returns true if any of the IPs are still in a hung state, false if not. | |||
4480 | */ | |||
4481 | static bool_Bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev) | |||
4482 | { | |||
4483 | int i; | |||
4484 | bool_Bool asic_hang = false0; | |||
4485 | ||||
4486 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) | |||
4487 | return true1; | |||
4488 | ||||
4489 | if (amdgpu_asic_need_full_reset(adev)(adev)->asic_funcs->need_full_reset((adev))) | |||
4490 | return true1; | |||
4491 | ||||
4492 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
4493 | if (!adev->ip_blocks[i].status.valid) | |||
4494 | continue; | |||
4495 | if (adev->ip_blocks[i].version->funcs->check_soft_reset) | |||
4496 | adev->ip_blocks[i].status.hang = | |||
4497 | adev->ip_blocks[i].version->funcs->check_soft_reset(adev); | |||
4498 | if (adev->ip_blocks[i].status.hang) { | |||
4499 | dev_info(adev->dev, "IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name)do { } while(0); | |||
4500 | asic_hang = true1; | |||
4501 | } | |||
4502 | } | |||
4503 | return asic_hang; | |||
4504 | } | |||
4505 | ||||
4506 | /** | |||
4507 | * amdgpu_device_ip_pre_soft_reset - prepare for soft reset | |||
4508 | * | |||
4509 | * @adev: amdgpu_device pointer | |||
4510 | * | |||
4511 | * The list of all the hardware IPs that make up the asic is walked and the | |||
4512 | * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset | |||
4513 | * handles any IP specific hardware or software state changes that are | |||
4514 | * necessary for a soft reset to succeed. | |||
4515 | * Returns 0 on success, negative error code on failure. | |||
4516 | */ | |||
4517 | static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev) | |||
4518 | { | |||
4519 | int i, r = 0; | |||
4520 | ||||
4521 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
4522 | if (!adev->ip_blocks[i].status.valid) | |||
4523 | continue; | |||
4524 | if (adev->ip_blocks[i].status.hang && | |||
4525 | adev->ip_blocks[i].version->funcs->pre_soft_reset) { | |||
4526 | r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev); | |||
4527 | if (r) | |||
4528 | return r; | |||
4529 | } | |||
4530 | } | |||
4531 | ||||
4532 | return 0; | |||
4533 | } | |||
4534 | ||||
4535 | /** | |||
4536 | * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed | |||
4537 | * | |||
4538 | * @adev: amdgpu_device pointer | |||
4539 | * | |||
4540 | * Some hardware IPs cannot be soft reset. If they are hung, a full gpu | |||
4541 | * reset is necessary to recover. | |||
4542 | * Returns true if a full asic reset is required, false if not. | |||
4543 | */ | |||
4544 | static bool_Bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev) | |||
4545 | { | |||
4546 | int i; | |||
4547 | ||||
4548 | if (amdgpu_asic_need_full_reset(adev)(adev)->asic_funcs->need_full_reset((adev))) | |||
4549 | return true1; | |||
4550 | ||||
4551 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
4552 | if (!adev->ip_blocks[i].status.valid) | |||
4553 | continue; | |||
4554 | if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) || | |||
4555 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) || | |||
4556 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) || | |||
4557 | (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) || | |||
4558 | adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) { | |||
4559 | if (adev->ip_blocks[i].status.hang) { | |||
4560 | dev_info(adev->dev, "Some block need full reset!\n")do { } while(0); | |||
4561 | return true1; | |||
4562 | } | |||
4563 | } | |||
4564 | } | |||
4565 | return false0; | |||
4566 | } | |||
4567 | ||||
4568 | /** | |||
4569 | * amdgpu_device_ip_soft_reset - do a soft reset | |||
4570 | * | |||
4571 | * @adev: amdgpu_device pointer | |||
4572 | * | |||
4573 | * The list of all the hardware IPs that make up the asic is walked and the | |||
4574 | * soft_reset callbacks are run if the block is hung. soft_reset handles any | |||
4575 | * IP specific hardware or software state changes that are necessary to soft | |||
4576 | * reset the IP. | |||
4577 | * Returns 0 on success, negative error code on failure. | |||
4578 | */ | |||
4579 | static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev) | |||
4580 | { | |||
4581 | int i, r = 0; | |||
4582 | ||||
4583 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
4584 | if (!adev->ip_blocks[i].status.valid) | |||
4585 | continue; | |||
4586 | if (adev->ip_blocks[i].status.hang && | |||
4587 | adev->ip_blocks[i].version->funcs->soft_reset) { | |||
4588 | r = adev->ip_blocks[i].version->funcs->soft_reset(adev); | |||
4589 | if (r) | |||
4590 | return r; | |||
4591 | } | |||
4592 | } | |||
4593 | ||||
4594 | return 0; | |||
4595 | } | |||
4596 | ||||
4597 | /** | |||
4598 | * amdgpu_device_ip_post_soft_reset - clean up from soft reset | |||
4599 | * | |||
4600 | * @adev: amdgpu_device pointer | |||
4601 | * | |||
4602 | * The list of all the hardware IPs that make up the asic is walked and the | |||
4603 | * post_soft_reset callbacks are run if the asic was hung. post_soft_reset | |||
4604 | * handles any IP specific hardware or software state changes that are | |||
4605 | * necessary after the IP has been soft reset. | |||
4606 | * Returns 0 on success, negative error code on failure. | |||
4607 | */ | |||
4608 | static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev) | |||
4609 | { | |||
4610 | int i, r = 0; | |||
4611 | ||||
4612 | for (i = 0; i < adev->num_ip_blocks; i++) { | |||
4613 | if (!adev->ip_blocks[i].status.valid) | |||
4614 | continue; | |||
4615 | if (adev->ip_blocks[i].status.hang && | |||
4616 | adev->ip_blocks[i].version->funcs->post_soft_reset) | |||
4617 | r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev); | |||
4618 | if (r) | |||
4619 | return r; | |||
4620 | } | |||
4621 | ||||
4622 | return 0; | |||
4623 | } | |||
4624 | ||||
4625 | /** | |||
4626 | * amdgpu_device_recover_vram - Recover some VRAM contents | |||
4627 | * | |||
4628 | * @adev: amdgpu_device pointer | |||
4629 | * | |||
4630 | * Restores the contents of VRAM buffers from the shadows in GTT. Used to | |||
4631 | * restore things like GPUVM page tables after a GPU reset where | |||
4632 | * the contents of VRAM might be lost. | |||
4633 | * | |||
4634 | * Returns: | |||
4635 | * 0 on success, negative error code on failure. | |||
4636 | */ | |||
4637 | static int amdgpu_device_recover_vram(struct amdgpu_device *adev) | |||
4638 | { | |||
4639 | struct dma_fence *fence = NULL((void *)0), *next = NULL((void *)0); | |||
4640 | struct amdgpu_bo *shadow; | |||
4641 | struct amdgpu_bo_vm *vmbo; | |||
4642 | long r = 1, tmo; | |||
4643 | ||||
4644 | if (amdgpu_sriov_runtime(adev)((adev)->virt.caps & (1 << 4))) | |||
4645 | tmo = msecs_to_jiffies(8000)(((uint64_t)(8000)) * hz / 1000); | |||
4646 | else | |||
4647 | tmo = msecs_to_jiffies(100)(((uint64_t)(100)) * hz / 1000); | |||
4648 | ||||
4649 | dev_info(adev->dev, "recover vram bo from shadow start\n")do { } while(0); | |||
4650 | mutex_lock(&adev->shadow_list_lock)rw_enter_write(&adev->shadow_list_lock); | |||
4651 | list_for_each_entry(vmbo, &adev->shadow_list, shadow_list)for (vmbo = ({ const __typeof( ((__typeof(*vmbo) *)0)->shadow_list ) *__mptr = ((&adev->shadow_list)->next); (__typeof (*vmbo) *)( (char *)__mptr - __builtin_offsetof(__typeof(*vmbo ), shadow_list) );}); &vmbo->shadow_list != (&adev ->shadow_list); vmbo = ({ const __typeof( ((__typeof(*vmbo ) *)0)->shadow_list ) *__mptr = (vmbo->shadow_list.next ); (__typeof(*vmbo) *)( (char *)__mptr - __builtin_offsetof(__typeof (*vmbo), shadow_list) );})) { | |||
4652 | /* If vm is compute context or adev is APU, shadow will be NULL */ | |||
4653 | if (!vmbo->shadow) | |||
4654 | continue; | |||
4655 | shadow = vmbo->shadow; | |||
4656 | ||||
4657 | /* No need to recover an evicted BO */ | |||
4658 | if (shadow->tbo.resource->mem_type != TTM_PL_TT1 || | |||
4659 | shadow->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET0x7fffffffffffffffL || | |||
4660 | shadow->parent->tbo.resource->mem_type != TTM_PL_VRAM2) | |||
4661 | continue; | |||
4662 | ||||
4663 | r = amdgpu_bo_restore_shadow(shadow, &next); | |||
4664 | if (r) | |||
4665 | break; | |||
4666 | ||||
4667 | if (fence) { | |||
4668 | tmo = dma_fence_wait_timeout(fence, false0, tmo); | |||
4669 | dma_fence_put(fence); | |||
4670 | fence = next; | |||
4671 | if (tmo == 0) { | |||
4672 | r = -ETIMEDOUT60; | |||
4673 | break; | |||
4674 | } else if (tmo < 0) { | |||
4675 | r = tmo; | |||
4676 | break; | |||
4677 | } | |||
4678 | } else { | |||
4679 | fence = next; | |||
4680 | } | |||
4681 | } | |||
4682 | mutex_unlock(&adev->shadow_list_lock)rw_exit_write(&adev->shadow_list_lock); | |||
4683 | ||||
4684 | if (fence) | |||
4685 | tmo = dma_fence_wait_timeout(fence, false0, tmo); | |||
4686 | dma_fence_put(fence); | |||
4687 | ||||
4688 | if (r < 0 || tmo <= 0) { | |||
4689 | dev_err(adev->dev, "recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo)printf("drm:pid%d:%s *ERROR* " "recover vram bo from shadow failed, r is %ld, tmo is %ld\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , r, tmo ); | |||
4690 | return -EIO5; | |||
4691 | } | |||
4692 | ||||
4693 | dev_info(adev->dev, "recover vram bo from shadow done\n")do { } while(0); | |||
4694 | return 0; | |||
4695 | } | |||
4696 | ||||
4697 | ||||
4698 | /** | |||
4699 | * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf | |||
4700 | * | |||
4701 | * @adev: amdgpu_device pointer | |||
4702 | * @from_hypervisor: request from hypervisor | |||
4703 | * | |||
4704 | * do VF FLR and reinitialize Asic | |||
4705 | * return 0 means succeeded otherwise failed | |||
4706 | */ | |||
4707 | static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, | |||
4708 | bool_Bool from_hypervisor) | |||
4709 | { | |||
4710 | int r; | |||
4711 | struct amdgpu_hive_info *hive = NULL((void *)0); | |||
4712 | int retry_limit = 0; | |||
4713 | ||||
4714 | retry: | |||
4715 | amdgpu_amdkfd_pre_reset(adev); | |||
4716 | ||||
4717 | if (from_hypervisor) | |||
4718 | r = amdgpu_virt_request_full_gpu(adev, true1); | |||
4719 | else | |||
4720 | r = amdgpu_virt_reset_gpu(adev); | |||
4721 | if (r) | |||
4722 | return r; | |||
4723 | ||||
4724 | /* Resume IP prior to SMC */ | |||
4725 | r = amdgpu_device_ip_reinit_early_sriov(adev); | |||
4726 | if (r) | |||
4727 | goto error; | |||
4728 | ||||
4729 | amdgpu_virt_init_data_exchange(adev); | |||
4730 | ||||
4731 | r = amdgpu_device_fw_loading(adev); | |||
4732 | if (r) | |||
4733 | return r; | |||
4734 | ||||
4735 | /* now we are okay to resume SMC/CP/SDMA */ | |||
4736 | r = amdgpu_device_ip_reinit_late_sriov(adev); | |||
4737 | if (r) | |||
4738 | goto error; | |||
4739 | ||||
4740 | hive = amdgpu_get_xgmi_hive(adev); | |||
4741 | /* Update PSP FW topology after reset */ | |||
4742 | if (hive && adev->gmc.xgmi.num_physical_nodes > 1) | |||
4743 | r = amdgpu_xgmi_update_topology(hive, adev); | |||
4744 | ||||
4745 | if (hive) | |||
4746 | amdgpu_put_xgmi_hive(hive); | |||
4747 | ||||
4748 | if (!r) { | |||
4749 | amdgpu_irq_gpu_reset_resume_helper(adev); | |||
4750 | r = amdgpu_ib_ring_tests(adev); | |||
4751 | ||||
4752 | amdgpu_amdkfd_post_reset(adev); | |||
4753 | } | |||
4754 | ||||
4755 | error: | |||
4756 | if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) { | |||
4757 | amdgpu_inc_vram_lost(adev)__sync_fetch_and_add(&((adev)->vram_lost_counter), 1);; | |||
4758 | r = amdgpu_device_recover_vram(adev); | |||
4759 | } | |||
4760 | amdgpu_virt_release_full_gpu(adev, true1); | |||
4761 | ||||
4762 | if (AMDGPU_RETRY_SRIOV_RESET(r)((r) == -16 || (r) == -60 || (r) == -22)) { | |||
4763 | if (retry_limit < AMDGPU_MAX_RETRY_LIMIT2) { | |||
4764 | retry_limit++; | |||
4765 | goto retry; | |||
4766 | } else | |||
4767 | DRM_ERROR("GPU reset retry is beyond the retry limit\n")__drm_err("GPU reset retry is beyond the retry limit\n"); | |||
4768 | } | |||
4769 | ||||
4770 | return r; | |||
4771 | } | |||
4772 | ||||
4773 | /** | |||
4774 | * amdgpu_device_has_job_running - check if there is any job in mirror list | |||
4775 | * | |||
4776 | * @adev: amdgpu_device pointer | |||
4777 | * | |||
4778 | * check if there is any job in mirror list | |||
4779 | */ | |||
4780 | bool_Bool amdgpu_device_has_job_running(struct amdgpu_device *adev) | |||
4781 | { | |||
4782 | int i; | |||
4783 | struct drm_sched_job *job; | |||
4784 | ||||
4785 | for (i = 0; i < AMDGPU_MAX_RINGS28; ++i) { | |||
4786 | struct amdgpu_ring *ring = adev->rings[i]; | |||
4787 | ||||
4788 | if (!ring || !ring->sched.thread) | |||
4789 | continue; | |||
4790 | ||||
4791 | spin_lock(&ring->sched.job_list_lock)mtx_enter(&ring->sched.job_list_lock); | |||
4792 | job = list_first_entry_or_null(&ring->sched.pending_list,(list_empty(&ring->sched.pending_list) ? ((void *)0) : ({ const __typeof( ((struct drm_sched_job *)0)->list ) *__mptr = ((&ring->sched.pending_list)->next); (struct drm_sched_job *)( (char *)__mptr - __builtin_offsetof(struct drm_sched_job , list) );})) | |||
4793 | struct drm_sched_job, list)(list_empty(&ring->sched.pending_list) ? ((void *)0) : ({ const __typeof( ((struct drm_sched_job *)0)->list ) *__mptr = ((&ring->sched.pending_list)->next); (struct drm_sched_job *)( (char *)__mptr - __builtin_offsetof(struct drm_sched_job , list) );})); | |||
4794 | spin_unlock(&ring->sched.job_list_lock)mtx_leave(&ring->sched.job_list_lock); | |||
4795 | if (job) | |||
4796 | return true1; | |||
4797 | } | |||
4798 | return false0; | |||
4799 | } | |||
4800 | ||||
4801 | /** | |||
4802 | * amdgpu_device_should_recover_gpu - check if we should try GPU recovery | |||
4803 | * | |||
4804 | * @adev: amdgpu_device pointer | |||
4805 | * | |||
4806 | * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover | |||
4807 | * a hung GPU. | |||
4808 | */ | |||
4809 | bool_Bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev) | |||
4810 | { | |||
4811 | ||||
4812 | if (amdgpu_gpu_recovery == 0) | |||
4813 | goto disabled; | |||
4814 | ||||
4815 | if (!amdgpu_device_ip_check_soft_reset(adev)) { | |||
4816 | dev_info(adev->dev,"Timeout, but no hardware hang detected.\n")do { } while(0); | |||
4817 | return false0; | |||
4818 | } | |||
4819 | ||||
4820 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) | |||
4821 | return true1; | |||
4822 | ||||
4823 | if (amdgpu_gpu_recovery == -1) { | |||
4824 | switch (adev->asic_type) { | |||
4825 | #ifdef CONFIG_DRM_AMDGPU_SI | |||
4826 | case CHIP_VERDE: | |||
4827 | case CHIP_TAHITI: | |||
4828 | case CHIP_PITCAIRN: | |||
4829 | case CHIP_OLAND: | |||
4830 | case CHIP_HAINAN: | |||
4831 | #endif | |||
4832 | #ifdef CONFIG_DRM_AMDGPU_CIK | |||
4833 | case CHIP_KAVERI: | |||
4834 | case CHIP_KABINI: | |||
4835 | case CHIP_MULLINS: | |||
4836 | #endif | |||
4837 | case CHIP_CARRIZO: | |||
4838 | case CHIP_STONEY: | |||
4839 | case CHIP_CYAN_SKILLFISH: | |||
4840 | goto disabled; | |||
4841 | default: | |||
4842 | break; | |||
4843 | } | |||
4844 | } | |||
4845 | ||||
4846 | return true1; | |||
4847 | ||||
4848 | disabled: | |||
4849 | dev_info(adev->dev, "GPU recovery disabled.\n")do { } while(0); | |||
4850 | return false0; | |||
4851 | } | |||
4852 | ||||
4853 | int amdgpu_device_mode1_reset(struct amdgpu_device *adev) | |||
4854 | { | |||
4855 | u32 i; | |||
4856 | int ret = 0; | |||
4857 | ||||
4858 | amdgpu_atombios_scratch_regs_engine_hung(adev, true1); | |||
4859 | ||||
4860 | dev_info(adev->dev, "GPU mode1 reset\n")do { } while(0); | |||
4861 | ||||
4862 | /* disable BM */ | |||
4863 | pci_clear_master(adev->pdev); | |||
4864 | ||||
4865 | amdgpu_device_cache_pci_state(adev->pdev); | |||
4866 | ||||
4867 | if (amdgpu_dpm_is_mode1_reset_supported(adev)) { | |||
4868 | dev_info(adev->dev, "GPU smu mode1 reset\n")do { } while(0); | |||
4869 | ret = amdgpu_dpm_mode1_reset(adev); | |||
4870 | } else { | |||
4871 | dev_info(adev->dev, "GPU psp mode1 reset\n")do { } while(0); | |||
4872 | ret = psp_gpu_reset(adev); | |||
4873 | } | |||
4874 | ||||
4875 | if (ret) | |||
4876 | dev_err(adev->dev, "GPU mode1 reset failed\n")printf("drm:pid%d:%s *ERROR* " "GPU mode1 reset failed\n", ({ struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
4877 | ||||
4878 | amdgpu_device_load_pci_state(adev->pdev); | |||
4879 | ||||
4880 | /* wait for asic to come out of reset */ | |||
4881 | for (i = 0; i < adev->usec_timeout; i++) { | |||
4882 | u32 memsize = adev->nbio.funcs->get_memsize(adev); | |||
4883 | ||||
4884 | if (memsize != 0xffffffff) | |||
4885 | break; | |||
4886 | udelay(1); | |||
4887 | } | |||
4888 | ||||
4889 | amdgpu_atombios_scratch_regs_engine_hung(adev, false0); | |||
4890 | return ret; | |||
4891 | } | |||
4892 | ||||
4893 | int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, | |||
4894 | struct amdgpu_reset_context *reset_context) | |||
4895 | { | |||
4896 | int i, r = 0; | |||
4897 | struct amdgpu_job *job = NULL((void *)0); | |||
4898 | bool_Bool need_full_reset = | |||
4899 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |||
4900 | ||||
4901 | if (reset_context->reset_req_dev == adev) | |||
4902 | job = reset_context->job; | |||
4903 | ||||
4904 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) { | |||
4905 | /* stop the data exchange thread */ | |||
4906 | amdgpu_virt_fini_data_exchange(adev); | |||
4907 | } | |||
4908 | ||||
4909 | amdgpu_fence_driver_isr_toggle(adev, true1); | |||
4910 | ||||
4911 | /* block all schedulers and reset given job's ring */ | |||
4912 | for (i = 0; i < AMDGPU_MAX_RINGS28; ++i) { | |||
4913 | struct amdgpu_ring *ring = adev->rings[i]; | |||
4914 | ||||
4915 | if (!ring || !ring->sched.thread) | |||
4916 | continue; | |||
4917 | ||||
4918 | /*clear job fence from fence drv to avoid force_completion | |||
4919 | *leave NULL and vm flush fence in fence drv */ | |||
4920 | amdgpu_fence_driver_clear_job_fences(ring); | |||
4921 | ||||
4922 | /* after all hw jobs are reset, hw fence is meaningless, so force_completion */ | |||
4923 | amdgpu_fence_driver_force_completion(ring); | |||
4924 | } | |||
4925 | ||||
4926 | amdgpu_fence_driver_isr_toggle(adev, false0); | |||
4927 | ||||
4928 | if (job && job->vm) | |||
4929 | drm_sched_increase_karma(&job->base); | |||
4930 | ||||
4931 | r = amdgpu_reset_prepare_hwcontext(adev, reset_context); | |||
4932 | /* If reset handler not implemented, continue; otherwise return */ | |||
4933 | if (r == -ENOSYS78) | |||
4934 | r = 0; | |||
4935 | else | |||
4936 | return r; | |||
4937 | ||||
4938 | /* Don't suspend on bare metal if we are not going to HW reset the ASIC */ | |||
4939 | if (!amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) { | |||
4940 | ||||
4941 | if (!need_full_reset) | |||
4942 | need_full_reset = amdgpu_device_ip_need_full_reset(adev); | |||
4943 | ||||
4944 | if (!need_full_reset && amdgpu_gpu_recovery) { | |||
4945 | amdgpu_device_ip_pre_soft_reset(adev); | |||
4946 | r = amdgpu_device_ip_soft_reset(adev); | |||
4947 | amdgpu_device_ip_post_soft_reset(adev); | |||
4948 | if (r || amdgpu_device_ip_check_soft_reset(adev)) { | |||
4949 | dev_info(adev->dev, "soft reset failed, will fallback to full reset!\n")do { } while(0); | |||
4950 | need_full_reset = true1; | |||
4951 | } | |||
4952 | } | |||
4953 | ||||
4954 | if (need_full_reset) | |||
4955 | r = amdgpu_device_ip_suspend(adev); | |||
4956 | if (need_full_reset) | |||
4957 | set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |||
4958 | else | |||
4959 | clear_bit(AMDGPU_NEED_FULL_RESET, | |||
4960 | &reset_context->flags); | |||
4961 | } | |||
4962 | ||||
4963 | return r; | |||
4964 | } | |||
4965 | ||||
4966 | static int amdgpu_reset_reg_dumps(struct amdgpu_device *adev) | |||
4967 | { | |||
4968 | int i; | |||
4969 | ||||
4970 | lockdep_assert_held(&adev->reset_domain->sem)do { (void)(&adev->reset_domain->sem); } while(0); | |||
4971 | ||||
4972 | for (i = 0; i < adev->num_regs; i++) { | |||
4973 | adev->reset_dump_reg_value[i] = RREG32(adev->reset_dump_reg_list[i])amdgpu_device_rreg(adev, (adev->reset_dump_reg_list[i]), 0 ); | |||
4974 | trace_amdgpu_reset_reg_dumps(adev->reset_dump_reg_list[i], | |||
4975 | adev->reset_dump_reg_value[i]); | |||
4976 | } | |||
4977 | ||||
4978 | return 0; | |||
4979 | } | |||
4980 | ||||
4981 | #ifdef CONFIG_DEV_COREDUMP | |||
4982 | static ssize_t amdgpu_devcoredump_read(char *buffer, loff_t offset, | |||
4983 | size_t count, void *data, size_t datalen) | |||
4984 | { | |||
4985 | struct drm_printer p; | |||
4986 | struct amdgpu_device *adev = data; | |||
4987 | struct drm_print_iterator iter; | |||
4988 | int i; | |||
4989 | ||||
4990 | iter.data = buffer; | |||
4991 | iter.offset = 0; | |||
4992 | iter.start = offset; | |||
4993 | iter.remain = count; | |||
4994 | ||||
4995 | p = drm_coredump_printer(&iter); | |||
4996 | ||||
4997 | drm_printf(&p, "**** AMDGPU Device Coredump ****\n"); | |||
4998 | drm_printf(&p, "kernel: " UTS_RELEASE"" "\n"); | |||
4999 | drm_printf(&p, "module: " KBUILD_MODNAME "\n"); | |||
5000 | drm_printf(&p, "time: %lld.%09ld\n", adev->reset_time.tv_sec, adev->reset_time.tv_nsec); | |||
5001 | if (adev->reset_task_info.pid) | |||
5002 | drm_printf(&p, "process_name: %s PID: %d\n", | |||
5003 | adev->reset_task_info.process_name, | |||
5004 | adev->reset_task_info.pid); | |||
5005 | ||||
5006 | if (adev->reset_vram_lost) | |||
5007 | drm_printf(&p, "VRAM is lost due to GPU reset!\n"); | |||
5008 | if (adev->num_regs) { | |||
5009 | drm_printf(&p, "AMDGPU register dumps:\nOffset: Value:\n"); | |||
5010 | ||||
5011 | for (i = 0; i < adev->num_regs; i++) | |||
5012 | drm_printf(&p, "0x%08x: 0x%08x\n", | |||
5013 | adev->reset_dump_reg_list[i], | |||
5014 | adev->reset_dump_reg_value[i]); | |||
5015 | } | |||
5016 | ||||
5017 | return count - iter.remain; | |||
5018 | } | |||
5019 | ||||
5020 | static void amdgpu_devcoredump_free(void *data) | |||
5021 | { | |||
5022 | } | |||
5023 | ||||
5024 | static void amdgpu_reset_capture_coredumpm(struct amdgpu_device *adev) | |||
5025 | { | |||
5026 | struct drm_device *dev = adev_to_drm(adev); | |||
5027 | ||||
5028 | ktime_get_ts64(&adev->reset_time); | |||
5029 | dev_coredumpm(dev->dev, THIS_MODULE((void *)0), adev, 0, GFP_KERNEL(0x0001 | 0x0004), | |||
5030 | amdgpu_devcoredump_read, amdgpu_devcoredump_free); | |||
5031 | } | |||
5032 | #endif | |||
5033 | ||||
5034 | int amdgpu_do_asic_reset(struct list_head *device_list_handle, | |||
5035 | struct amdgpu_reset_context *reset_context) | |||
5036 | { | |||
5037 | struct amdgpu_device *tmp_adev = NULL((void *)0); | |||
5038 | bool_Bool need_full_reset, skip_hw_reset, vram_lost = false0; | |||
5039 | int r = 0; | |||
5040 | bool_Bool gpu_reset_for_dev_remove = 0; | |||
5041 | ||||
5042 | /* Try reset handler method first */ | |||
5043 | tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,({ const __typeof( ((struct amdgpu_device *)0)->reset_list ) *__mptr = ((device_list_handle)->next); (struct amdgpu_device *)( (char *)__mptr - __builtin_offsetof(struct amdgpu_device , reset_list) );}) | |||
5044 | reset_list)({ const __typeof( ((struct amdgpu_device *)0)->reset_list ) *__mptr = ((device_list_handle)->next); (struct amdgpu_device *)( (char *)__mptr - __builtin_offsetof(struct amdgpu_device , reset_list) );}); | |||
5045 | amdgpu_reset_reg_dumps(tmp_adev); | |||
5046 | ||||
5047 | reset_context->reset_device_list = device_list_handle; | |||
5048 | r = amdgpu_reset_perform_reset(tmp_adev, reset_context); | |||
5049 | /* If reset handler not implemented, continue; otherwise return */ | |||
5050 | if (r == -ENOSYS78) | |||
5051 | r = 0; | |||
5052 | else | |||
5053 | return r; | |||
5054 | ||||
5055 | /* Reset handler not implemented, use the default method */ | |||
5056 | need_full_reset = | |||
5057 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |||
5058 | skip_hw_reset = test_bit(AMDGPU_SKIP_HW_RESET, &reset_context->flags); | |||
5059 | ||||
5060 | gpu_reset_for_dev_remove = | |||
5061 | test_bit(AMDGPU_RESET_FOR_DEVICE_REMOVE, &reset_context->flags) && | |||
5062 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |||
5063 | ||||
5064 | /* | |||
5065 | * ASIC reset has to be done on all XGMI hive nodes ASAP | |||
5066 | * to allow proper links negotiation in FW (within 1 sec) | |||
5067 | */ | |||
5068 | if (!skip_hw_reset && need_full_reset) { | |||
5069 | list_for_each_entry(tmp_adev, device_list_handle, reset_list)for (tmp_adev = ({ const __typeof( ((__typeof(*tmp_adev) *)0) ->reset_list ) *__mptr = ((device_list_handle)->next); ( __typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof(__typeof (*tmp_adev), reset_list) );}); &tmp_adev->reset_list != (device_list_handle); tmp_adev = ({ const __typeof( ((__typeof (*tmp_adev) *)0)->reset_list ) *__mptr = (tmp_adev->reset_list .next); (__typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof (__typeof(*tmp_adev), reset_list) );})) { | |||
5070 | /* For XGMI run all resets in parallel to speed up the process */ | |||
5071 | if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) { | |||
5072 | tmp_adev->gmc.xgmi.pending_reset = false0; | |||
5073 | if (!queue_work(system_unbound_wq, &tmp_adev->xgmi_reset_work)) | |||
5074 | r = -EALREADY37; | |||
5075 | } else | |||
5076 | r = amdgpu_asic_reset(tmp_adev)(tmp_adev)->asic_funcs->reset((tmp_adev)); | |||
5077 | ||||
5078 | if (r) { | |||
5079 | dev_err(tmp_adev->dev, "ASIC reset failed with error, %d for drm dev, %s",printf("drm:pid%d:%s *ERROR* " "ASIC reset failed with error, %d for drm dev, %s" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , r, adev_to_drm (tmp_adev)->unique) | |||
5080 | r, adev_to_drm(tmp_adev)->unique)printf("drm:pid%d:%s *ERROR* " "ASIC reset failed with error, %d for drm dev, %s" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , r, adev_to_drm (tmp_adev)->unique); | |||
5081 | break; | |||
5082 | } | |||
5083 | } | |||
5084 | ||||
5085 | /* For XGMI wait for all resets to complete before proceed */ | |||
5086 | if (!r) { | |||
5087 | list_for_each_entry(tmp_adev, device_list_handle, reset_list)for (tmp_adev = ({ const __typeof( ((__typeof(*tmp_adev) *)0) ->reset_list ) *__mptr = ((device_list_handle)->next); ( __typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof(__typeof (*tmp_adev), reset_list) );}); &tmp_adev->reset_list != (device_list_handle); tmp_adev = ({ const __typeof( ((__typeof (*tmp_adev) *)0)->reset_list ) *__mptr = (tmp_adev->reset_list .next); (__typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof (__typeof(*tmp_adev), reset_list) );})) { | |||
5088 | if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) { | |||
5089 | flush_work(&tmp_adev->xgmi_reset_work); | |||
5090 | r = tmp_adev->asic_reset_res; | |||
5091 | if (r) | |||
5092 | break; | |||
5093 | } | |||
5094 | } | |||
5095 | } | |||
5096 | } | |||
5097 | ||||
5098 | if (!r && amdgpu_ras_intr_triggered()) { | |||
5099 | list_for_each_entry(tmp_adev, device_list_handle, reset_list)for (tmp_adev = ({ const __typeof( ((__typeof(*tmp_adev) *)0) ->reset_list ) *__mptr = ((device_list_handle)->next); ( __typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof(__typeof (*tmp_adev), reset_list) );}); &tmp_adev->reset_list != (device_list_handle); tmp_adev = ({ const __typeof( ((__typeof (*tmp_adev) *)0)->reset_list ) *__mptr = (tmp_adev->reset_list .next); (__typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof (__typeof(*tmp_adev), reset_list) );})) { | |||
5100 | if (tmp_adev->mmhub.ras && tmp_adev->mmhub.ras->ras_block.hw_ops && | |||
5101 | tmp_adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count) | |||
5102 | tmp_adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count(tmp_adev); | |||
5103 | } | |||
5104 | ||||
5105 | amdgpu_ras_intr_cleared(); | |||
5106 | } | |||
5107 | ||||
5108 | /* Since the mode1 reset affects base ip blocks, the | |||
5109 | * phase1 ip blocks need to be resumed. Otherwise there | |||
5110 | * will be a BIOS signature error and the psp bootloader | |||
5111 | * can't load kdb on the next amdgpu install. | |||
5112 | */ | |||
5113 | if (gpu_reset_for_dev_remove) { | |||
5114 | list_for_each_entry(tmp_adev, device_list_handle, reset_list)for (tmp_adev = ({ const __typeof( ((__typeof(*tmp_adev) *)0) ->reset_list ) *__mptr = ((device_list_handle)->next); ( __typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof(__typeof (*tmp_adev), reset_list) );}); &tmp_adev->reset_list != (device_list_handle); tmp_adev = ({ const __typeof( ((__typeof (*tmp_adev) *)0)->reset_list ) *__mptr = (tmp_adev->reset_list .next); (__typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof (__typeof(*tmp_adev), reset_list) );})) | |||
5115 | amdgpu_device_ip_resume_phase1(tmp_adev); | |||
5116 | ||||
5117 | goto end; | |||
5118 | } | |||
5119 | ||||
5120 | list_for_each_entry(tmp_adev, device_list_handle, reset_list)for (tmp_adev = ({ const __typeof( ((__typeof(*tmp_adev) *)0) ->reset_list ) *__mptr = ((device_list_handle)->next); ( __typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof(__typeof (*tmp_adev), reset_list) );}); &tmp_adev->reset_list != (device_list_handle); tmp_adev = ({ const __typeof( ((__typeof (*tmp_adev) *)0)->reset_list ) *__mptr = (tmp_adev->reset_list .next); (__typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof (__typeof(*tmp_adev), reset_list) );})) { | |||
5121 | if (need_full_reset) { | |||
5122 | /* post card */ | |||
5123 | r = amdgpu_device_asic_init(tmp_adev); | |||
5124 | if (r) { | |||
5125 | dev_warn(tmp_adev->dev, "asic atom init failed!")printf("drm:pid%d:%s *WARNING* " "asic atom init failed!", ({ struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
5126 | } else { | |||
5127 | dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n")do { } while(0); | |||
5128 | r = amdgpu_amdkfd_resume_iommu(tmp_adev); | |||
5129 | if (r) | |||
5130 | goto out; | |||
5131 | ||||
5132 | r = amdgpu_device_ip_resume_phase1(tmp_adev); | |||
5133 | if (r) | |||
5134 | goto out; | |||
5135 | ||||
5136 | vram_lost = amdgpu_device_check_vram_lost(tmp_adev); | |||
5137 | #ifdef CONFIG_DEV_COREDUMP | |||
5138 | tmp_adev->reset_vram_lost = vram_lost; | |||
5139 | memset(&tmp_adev->reset_task_info, 0,__builtin_memset((&tmp_adev->reset_task_info), (0), (sizeof (tmp_adev->reset_task_info))) | |||
5140 | sizeof(tmp_adev->reset_task_info))__builtin_memset((&tmp_adev->reset_task_info), (0), (sizeof (tmp_adev->reset_task_info))); | |||
5141 | if (reset_context->job && reset_context->job->vm) | |||
5142 | tmp_adev->reset_task_info = | |||
5143 | reset_context->job->vm->task_info; | |||
5144 | amdgpu_reset_capture_coredumpm(tmp_adev); | |||
5145 | #endif | |||
5146 | if (vram_lost) { | |||
5147 | DRM_INFO("VRAM is lost due to GPU reset!\n")printk("\0016" "[" "drm" "] " "VRAM is lost due to GPU reset!\n" ); | |||
5148 | amdgpu_inc_vram_lost(tmp_adev)__sync_fetch_and_add(&((tmp_adev)->vram_lost_counter), 1);; | |||
5149 | } | |||
5150 | ||||
5151 | r = amdgpu_device_fw_loading(tmp_adev); | |||
5152 | if (r) | |||
5153 | return r; | |||
5154 | ||||
5155 | r = amdgpu_device_ip_resume_phase2(tmp_adev); | |||
5156 | if (r) | |||
5157 | goto out; | |||
5158 | ||||
5159 | if (vram_lost) | |||
5160 | amdgpu_device_fill_reset_magic(tmp_adev); | |||
5161 | ||||
5162 | /* | |||
5163 | * Add this ASIC as tracked as reset was already | |||
5164 | * complete successfully. | |||
5165 | */ | |||
5166 | amdgpu_register_gpu_instance(tmp_adev); | |||
5167 | ||||
5168 | if (!reset_context->hive && | |||
5169 | tmp_adev->gmc.xgmi.num_physical_nodes > 1) | |||
5170 | amdgpu_xgmi_add_device(tmp_adev); | |||
5171 | ||||
5172 | r = amdgpu_device_ip_late_init(tmp_adev); | |||
5173 | if (r) | |||
5174 | goto out; | |||
5175 | ||||
5176 | drm_fb_helper_set_suspend_unlocked(adev_to_drm(tmp_adev)->fb_helper, false0); | |||
5177 | ||||
5178 | /* | |||
5179 | * The GPU enters bad state once faulty pages | |||
5180 | * by ECC has reached the threshold, and ras | |||
5181 | * recovery is scheduled next. So add one check | |||
5182 | * here to break recovery if it indeed exceeds | |||
5183 | * bad page threshold, and remind user to | |||
5184 | * retire this GPU or setting one bigger | |||
5185 | * bad_page_threshold value to fix this once | |||
5186 | * probing driver again. | |||
5187 | */ | |||
5188 | if (!amdgpu_ras_eeprom_check_err_threshold(tmp_adev)) { | |||
5189 | /* must succeed. */ | |||
5190 | amdgpu_ras_resume(tmp_adev); | |||
5191 | } else { | |||
5192 | r = -EINVAL22; | |||
5193 | goto out; | |||
5194 | } | |||
5195 | ||||
5196 | /* Update PSP FW topology after reset */ | |||
5197 | if (reset_context->hive && | |||
5198 | tmp_adev->gmc.xgmi.num_physical_nodes > 1) | |||
5199 | r = amdgpu_xgmi_update_topology( | |||
5200 | reset_context->hive, tmp_adev); | |||
5201 | } | |||
5202 | } | |||
5203 | ||||
5204 | out: | |||
5205 | if (!r) { | |||
5206 | amdgpu_irq_gpu_reset_resume_helper(tmp_adev); | |||
5207 | r = amdgpu_ib_ring_tests(tmp_adev); | |||
5208 | if (r) { | |||
5209 | dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r)printf("drm:pid%d:%s *ERROR* " "ib ring test failed (%d).\n", ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , r); | |||
5210 | need_full_reset = true1; | |||
5211 | r = -EAGAIN35; | |||
5212 | goto end; | |||
5213 | } | |||
5214 | } | |||
5215 | ||||
5216 | if (!r) | |||
5217 | r = amdgpu_device_recover_vram(tmp_adev); | |||
5218 | else | |||
5219 | tmp_adev->asic_reset_res = r; | |||
5220 | } | |||
5221 | ||||
5222 | end: | |||
5223 | if (need_full_reset) | |||
5224 | set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |||
5225 | else | |||
5226 | clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |||
5227 | return r; | |||
5228 | } | |||
5229 | ||||
5230 | static void amdgpu_device_set_mp1_state(struct amdgpu_device *adev) | |||
5231 | { | |||
5232 | ||||
5233 | switch (amdgpu_asic_reset_method(adev)(adev)->asic_funcs->reset_method((adev))) { | |||
5234 | case AMD_RESET_METHOD_MODE1: | |||
5235 | adev->mp1_state = PP_MP1_STATE_SHUTDOWN; | |||
5236 | break; | |||
5237 | case AMD_RESET_METHOD_MODE2: | |||
5238 | adev->mp1_state = PP_MP1_STATE_RESET; | |||
5239 | break; | |||
5240 | default: | |||
5241 | adev->mp1_state = PP_MP1_STATE_NONE; | |||
5242 | break; | |||
5243 | } | |||
5244 | ||||
5245 | pci_dev_put(p); | |||
5246 | } | |||
5247 | ||||
5248 | static void amdgpu_device_unset_mp1_state(struct amdgpu_device *adev) | |||
5249 | { | |||
5250 | amdgpu_vf_error_trans_all(adev); | |||
5251 | adev->mp1_state = PP_MP1_STATE_NONE; | |||
5252 | } | |||
5253 | ||||
5254 | static void amdgpu_device_resume_display_audio(struct amdgpu_device *adev) | |||
5255 | { | |||
5256 | STUB()do { printf("%s: stub\n", __func__); } while(0); | |||
5257 | #ifdef notyet | |||
5258 | struct pci_dev *p = NULL((void *)0); | |||
5259 | ||||
5260 | p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus), | |||
5261 | adev->pdev->bus->number, 1); | |||
5262 | if (p) { | |||
5263 | pm_runtime_enable(&(p->dev)); | |||
5264 | pm_runtime_resume(&(p->dev)); | |||
5265 | } | |||
5266 | #endif | |||
5267 | } | |||
5268 | ||||
5269 | static int amdgpu_device_suspend_display_audio(struct amdgpu_device *adev) | |||
5270 | { | |||
5271 | enum amd_reset_method reset_method; | |||
5272 | struct pci_dev *p = NULL((void *)0); | |||
5273 | u64 expires; | |||
5274 | ||||
5275 | /* | |||
5276 | * For now, only BACO and mode1 reset are confirmed | |||
5277 | * to suffer the audio issue without proper suspended. | |||
5278 | */ | |||
5279 | reset_method = amdgpu_asic_reset_method(adev)(adev)->asic_funcs->reset_method((adev)); | |||
5280 | if ((reset_method != AMD_RESET_METHOD_BACO) && | |||
5281 | (reset_method != AMD_RESET_METHOD_MODE1)) | |||
5282 | return -EINVAL22; | |||
5283 | ||||
5284 | STUB()do { printf("%s: stub\n", __func__); } while(0); | |||
5285 | return -ENOSYS78; | |||
5286 | #ifdef notyet | |||
5287 | ||||
5288 | p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus), | |||
5289 | adev->pdev->bus->number, 1); | |||
5290 | if (!p) | |||
5291 | return -ENODEV19; | |||
5292 | ||||
5293 | expires = pm_runtime_autosuspend_expiration(&(p->dev)); | |||
5294 | if (!expires) | |||
5295 | /* | |||
5296 | * If we cannot get the audio device autosuspend delay, | |||
5297 | * a fixed 4S interval will be used. Considering 3S is | |||
5298 | * the audio controller default autosuspend delay setting. | |||
5299 | * 4S used here is guaranteed to cover that. | |||
5300 | */ | |||
5301 | expires = ktime_get_mono_fast_ns() + NSEC_PER_SEC1000000000L * 4ULL; | |||
5302 | ||||
5303 | while (!pm_runtime_status_suspended(&(p->dev))) { | |||
5304 | if (!pm_runtime_suspend(&(p->dev))) | |||
5305 | break; | |||
5306 | ||||
5307 | if (expires < ktime_get_mono_fast_ns()) { | |||
5308 | dev_warn(adev->dev, "failed to suspend display audio\n")printf("drm:pid%d:%s *WARNING* " "failed to suspend display audio\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); | |||
5309 | pci_dev_put(p); | |||
5310 | /* TODO: abort the succeeding gpu reset? */ | |||
5311 | return -ETIMEDOUT60; | |||
5312 | } | |||
5313 | } | |||
5314 | ||||
5315 | pm_runtime_disable(&(p->dev)); | |||
5316 | ||||
5317 | pci_dev_put(p); | |||
5318 | return 0; | |||
5319 | #endif | |||
5320 | } | |||
5321 | ||||
5322 | static void amdgpu_device_recheck_guilty_jobs( | |||
5323 | struct amdgpu_device *adev, struct list_head *device_list_handle, | |||
5324 | struct amdgpu_reset_context *reset_context) | |||
5325 | { | |||
5326 | int i, r = 0; | |||
5327 | ||||
5328 | for (i = 0; i < AMDGPU_MAX_RINGS28; ++i) { | |||
5329 | struct amdgpu_ring *ring = adev->rings[i]; | |||
5330 | int ret = 0; | |||
5331 | struct drm_sched_job *s_job; | |||
5332 | ||||
5333 | if (!ring || !ring->sched.thread) | |||
5334 | continue; | |||
5335 | ||||
5336 | s_job = list_first_entry_or_null(&ring->sched.pending_list,(list_empty(&ring->sched.pending_list) ? ((void *)0) : ({ const __typeof( ((struct drm_sched_job *)0)->list ) *__mptr = ((&ring->sched.pending_list)->next); (struct drm_sched_job *)( (char *)__mptr - __builtin_offsetof(struct drm_sched_job , list) );})) | |||
5337 | struct drm_sched_job, list)(list_empty(&ring->sched.pending_list) ? ((void *)0) : ({ const __typeof( ((struct drm_sched_job *)0)->list ) *__mptr = ((&ring->sched.pending_list)->next); (struct drm_sched_job *)( (char *)__mptr - __builtin_offsetof(struct drm_sched_job , list) );})); | |||
5338 | if (s_job == NULL((void *)0)) | |||
5339 | continue; | |||
5340 | ||||
5341 | /* clear job's guilty and depend the folowing step to decide the real one */ | |||
5342 | drm_sched_reset_karma(s_job); | |||
5343 | drm_sched_resubmit_jobs_ext(&ring->sched, 1); | |||
5344 | ||||
5345 | if (!s_job->s_fence->parent) { | |||
5346 | DRM_WARN("Failed to get a HW fence for job!")printk("\0014" "[" "drm" "] " "Failed to get a HW fence for job!" ); | |||
5347 | continue; | |||
5348 | } | |||
5349 | ||||
5350 | ret = dma_fence_wait_timeout(s_job->s_fence->parent, false0, ring->sched.timeout); | |||
5351 | if (ret == 0) { /* timeout */ | |||
5352 | DRM_ERROR("Found the real bad job! ring:%s, job_id:%llx\n",__drm_err("Found the real bad job! ring:%s, job_id:%llx\n", ring ->sched.name, s_job->id) | |||
5353 | ring->sched.name, s_job->id)__drm_err("Found the real bad job! ring:%s, job_id:%llx\n", ring ->sched.name, s_job->id); | |||
5354 | ||||
5355 | ||||
5356 | amdgpu_fence_driver_isr_toggle(adev, true1); | |||
5357 | ||||
5358 | /* Clear this failed job from fence array */ | |||
5359 | amdgpu_fence_driver_clear_job_fences(ring); | |||
5360 | ||||
5361 | amdgpu_fence_driver_isr_toggle(adev, false0); | |||
5362 | ||||
5363 | /* Since the job won't signal and we go for | |||
5364 | * another resubmit drop this parent pointer | |||
5365 | */ | |||
5366 | dma_fence_put(s_job->s_fence->parent); | |||
5367 | s_job->s_fence->parent = NULL((void *)0); | |||
5368 | ||||
5369 | /* set guilty */ | |||
5370 | drm_sched_increase_karma(s_job); | |||
5371 | amdgpu_reset_prepare_hwcontext(adev, reset_context); | |||
5372 | retry: | |||
5373 | /* do hw reset */ | |||
5374 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) { | |||
5375 | amdgpu_virt_fini_data_exchange(adev); | |||
5376 | r = amdgpu_device_reset_sriov(adev, false0); | |||
5377 | if (r) | |||
5378 | adev->asic_reset_res = r; | |||
5379 | } else { | |||
5380 | clear_bit(AMDGPU_SKIP_HW_RESET, | |||
5381 | &reset_context->flags); | |||
5382 | r = amdgpu_do_asic_reset(device_list_handle, | |||
5383 | reset_context); | |||
5384 | if (r && r == -EAGAIN35) | |||
5385 | goto retry; | |||
5386 | } | |||
5387 | ||||
5388 | /* | |||
5389 | * add reset counter so that the following | |||
5390 | * resubmitted job could flush vmid | |||
5391 | */ | |||
5392 | atomic_inc(&adev->gpu_reset_counter)__sync_fetch_and_add(&adev->gpu_reset_counter, 1); | |||
5393 | continue; | |||
5394 | } | |||
5395 | ||||
5396 | /* got the hw fence, signal finished fence */ | |||
5397 | atomic_dec(ring->sched.score)__sync_fetch_and_sub(ring->sched.score, 1); | |||
5398 | dma_fence_get(&s_job->s_fence->finished); | |||
5399 | dma_fence_signal(&s_job->s_fence->finished); | |||
5400 | dma_fence_put(&s_job->s_fence->finished); | |||
5401 | ||||
5402 | /* remove node from list and free the job */ | |||
5403 | spin_lock(&ring->sched.job_list_lock)mtx_enter(&ring->sched.job_list_lock); | |||
5404 | list_del_init(&s_job->list); | |||
5405 | spin_unlock(&ring->sched.job_list_lock)mtx_leave(&ring->sched.job_list_lock); | |||
5406 | ring->sched.ops->free_job(s_job); | |||
5407 | } | |||
5408 | } | |||
5409 | ||||
5410 | static inline void amdgpu_device_stop_pending_resets(struct amdgpu_device *adev) | |||
5411 | { | |||
5412 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |||
5413 | ||||
5414 | #if defined(CONFIG_DEBUG_FS) | |||
5415 | if (!amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) | |||
5416 | cancel_work(&adev->reset_work); | |||
5417 | #endif | |||
5418 | ||||
5419 | if (adev->kfd.dev) | |||
5420 | cancel_work(&adev->kfd.reset_work); | |||
5421 | ||||
5422 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) | |||
5423 | cancel_work(&adev->virt.flr_work); | |||
5424 | ||||
5425 | if (con && adev->ras_enabled) | |||
5426 | cancel_work(&con->recovery_work); | |||
5427 | ||||
5428 | } | |||
5429 | ||||
5430 | ||||
5431 | /** | |||
5432 | * amdgpu_device_gpu_recover - reset the asic and recover scheduler | |||
5433 | * | |||
5434 | * @adev: amdgpu_device pointer | |||
5435 | * @job: which job trigger hang | |||
5436 | * | |||
5437 | * Attempt to reset the GPU if it has hung (all asics). | |||
5438 | * Attempt to do soft-reset or full-reset and reinitialize Asic | |||
5439 | * Returns 0 for success or an error on failure. | |||
5440 | */ | |||
5441 | ||||
5442 | int amdgpu_device_gpu_recover(struct amdgpu_device *adev, | |||
5443 | struct amdgpu_job *job, | |||
5444 | struct amdgpu_reset_context *reset_context) | |||
5445 | { | |||
5446 | struct list_head device_list, *device_list_handle = NULL((void *)0); | |||
5447 | bool_Bool job_signaled = false0; | |||
5448 | struct amdgpu_hive_info *hive = NULL((void *)0); | |||
5449 | struct amdgpu_device *tmp_adev = NULL((void *)0); | |||
5450 | int i, r = 0; | |||
5451 | bool_Bool need_emergency_restart = false0; | |||
5452 | bool_Bool audio_suspended = false0; | |||
5453 | int tmp_vram_lost_counter; | |||
5454 | bool_Bool gpu_reset_for_dev_remove = false0; | |||
5455 | ||||
5456 | gpu_reset_for_dev_remove = | |||
5457 | test_bit(AMDGPU_RESET_FOR_DEVICE_REMOVE, &reset_context->flags) && | |||
5458 | test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags); | |||
5459 | ||||
5460 | /* | |||
5461 | * Special case: RAS triggered and full reset isn't supported | |||
5462 | */ | |||
5463 | need_emergency_restart = amdgpu_ras_need_emergency_restart(adev); | |||
5464 | ||||
5465 | /* | |||
5466 | * Flush RAM to disk so that after reboot | |||
5467 | * the user can read log and see why the system rebooted. | |||
5468 | */ | |||
5469 | if (need_emergency_restart && amdgpu_ras_get_context(adev) && | |||
| ||||
5470 | amdgpu_ras_get_context(adev)->reboot) { | |||
5471 | DRM_WARN("Emergency reboot.")printk("\0014" "[" "drm" "] " "Emergency reboot."); | |||
5472 | ||||
5473 | #ifdef notyet | |||
5474 | ksys_sync_helper(); | |||
5475 | emergency_restart(); | |||
5476 | #else | |||
5477 | panic("emergency_restart"); | |||
5478 | #endif | |||
5479 | } | |||
5480 | ||||
5481 | dev_info(adev->dev, "GPU %s begin!\n",do { } while(0) | |||
5482 | need_emergency_restart ? "jobs stop":"reset")do { } while(0); | |||
5483 | ||||
5484 | if (!amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) | |||
5485 | hive = amdgpu_get_xgmi_hive(adev); | |||
5486 | if (hive) | |||
5487 | mutex_lock(&hive->hive_lock)rw_enter_write(&hive->hive_lock); | |||
5488 | ||||
5489 | reset_context->job = job; | |||
5490 | reset_context->hive = hive; | |||
5491 | /* | |||
5492 | * Build list of devices to reset. | |||
5493 | * In case we are in XGMI hive mode, resort the device list | |||
5494 | * to put adev in the 1st position. | |||
5495 | */ | |||
5496 | INIT_LIST_HEAD(&device_list); | |||
5497 | if (!amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2)) && (adev->gmc.xgmi.num_physical_nodes > 1)) { | |||
5498 | list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head)for (tmp_adev = ({ const __typeof( ((__typeof(*tmp_adev) *)0) ->gmc.xgmi.head ) *__mptr = ((&hive->device_list)-> next); (__typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof (__typeof(*tmp_adev), gmc.xgmi.head) );}); &tmp_adev-> gmc.xgmi.head != (&hive->device_list); tmp_adev = ({ const __typeof( ((__typeof(*tmp_adev) *)0)->gmc.xgmi.head ) *__mptr = (tmp_adev->gmc.xgmi.head.next); (__typeof(*tmp_adev) *) ( (char *)__mptr - __builtin_offsetof(__typeof(*tmp_adev), gmc .xgmi.head) );})) { | |||
| ||||
5499 | list_add_tail(&tmp_adev->reset_list, &device_list); | |||
5500 | if (gpu_reset_for_dev_remove && adev->shutdown) | |||
5501 | tmp_adev->shutdown = true1; | |||
5502 | } | |||
5503 | if (!list_is_first(&adev->reset_list, &device_list)) | |||
5504 | list_rotate_to_front(&adev->reset_list, &device_list); | |||
5505 | device_list_handle = &device_list; | |||
5506 | } else { | |||
5507 | list_add_tail(&adev->reset_list, &device_list); | |||
5508 | device_list_handle = &device_list; | |||
5509 | } | |||
5510 | ||||
5511 | /* We need to lock reset domain only once both for XGMI and single device */ | |||
5512 | tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,({ const __typeof( ((struct amdgpu_device *)0)->reset_list ) *__mptr = ((device_list_handle)->next); (struct amdgpu_device *)( (char *)__mptr - __builtin_offsetof(struct amdgpu_device , reset_list) );}) | |||
5513 | reset_list)({ const __typeof( ((struct amdgpu_device *)0)->reset_list ) *__mptr = ((device_list_handle)->next); (struct amdgpu_device *)( (char *)__mptr - __builtin_offsetof(struct amdgpu_device , reset_list) );}); | |||
5514 | amdgpu_device_lock_reset_domain(tmp_adev->reset_domain); | |||
5515 | ||||
5516 | /* block all schedulers and reset given job's ring */ | |||
5517 | list_for_each_entry(tmp_adev, device_list_handle, reset_list)for (tmp_adev = ({ const __typeof( ((__typeof(*tmp_adev) *)0) ->reset_list ) *__mptr = ((device_list_handle)->next); ( __typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof(__typeof (*tmp_adev), reset_list) );}); &tmp_adev->reset_list != (device_list_handle); tmp_adev = ({ const __typeof( ((__typeof (*tmp_adev) *)0)->reset_list ) *__mptr = (tmp_adev->reset_list .next); (__typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof (__typeof(*tmp_adev), reset_list) );})) { | |||
5518 | ||||
5519 | amdgpu_device_set_mp1_state(tmp_adev); | |||
5520 | ||||
5521 | /* | |||
5522 | * Try to put the audio codec into suspend state | |||
5523 | * before gpu reset started. | |||
5524 | * | |||
5525 | * Due to the power domain of the graphics device | |||
5526 | * is shared with AZ power domain. Without this, | |||
5527 | * we may change the audio hardware from behind | |||
5528 | * the audio driver's back. That will trigger | |||
5529 | * some audio codec errors. | |||
5530 | */ | |||
5531 | if (!amdgpu_device_suspend_display_audio(tmp_adev)) | |||
5532 | audio_suspended = true1; | |||
5533 | ||||
5534 | amdgpu_ras_set_error_query_ready(tmp_adev, false0); | |||
5535 | ||||
5536 | cancel_delayed_work_sync(&tmp_adev->delayed_init_work); | |||
5537 | ||||
5538 | if (!amdgpu_sriov_vf(tmp_adev)((tmp_adev)->virt.caps & (1 << 2))) | |||
5539 | amdgpu_amdkfd_pre_reset(tmp_adev); | |||
5540 | ||||
5541 | /* | |||
5542 | * Mark these ASICs to be reseted as untracked first | |||
5543 | * And add them back after reset completed | |||
5544 | */ | |||
5545 | amdgpu_unregister_gpu_instance(tmp_adev); | |||
5546 | ||||
5547 | drm_fb_helper_set_suspend_unlocked(adev_to_drm(tmp_adev)->fb_helper, true1); | |||
5548 | ||||
5549 | /* disable ras on ALL IPs */ | |||
5550 | if (!need_emergency_restart && | |||
5551 | amdgpu_device_ip_need_full_reset(tmp_adev)) | |||
5552 | amdgpu_ras_suspend(tmp_adev); | |||
5553 | ||||
5554 | for (i = 0; i < AMDGPU_MAX_RINGS28; ++i) { | |||
5555 | struct amdgpu_ring *ring = tmp_adev->rings[i]; | |||
5556 | ||||
5557 | if (!ring || !ring->sched.thread) | |||
5558 | continue; | |||
5559 | ||||
5560 | drm_sched_stop(&ring->sched, job ? &job->base : NULL((void *)0)); | |||
5561 | ||||
5562 | if (need_emergency_restart) | |||
5563 | amdgpu_job_stop_all_jobs_on_sched(&ring->sched); | |||
5564 | } | |||
5565 | atomic_inc(&tmp_adev->gpu_reset_counter)__sync_fetch_and_add(&tmp_adev->gpu_reset_counter, 1); | |||
5566 | } | |||
5567 | ||||
5568 | if (need_emergency_restart) | |||
5569 | goto skip_sched_resume; | |||
5570 | ||||
5571 | /* | |||
5572 | * Must check guilty signal here since after this point all old | |||
5573 | * HW fences are force signaled. | |||
5574 | * | |||
5575 | * job->base holds a reference to parent fence | |||
5576 | */ | |||
5577 | if (job && dma_fence_is_signaled(&job->hw_fence)) { | |||
5578 | job_signaled = true1; | |||
5579 | dev_info(adev->dev, "Guilty job already signaled, skipping HW reset")do { } while(0); | |||
5580 | goto skip_hw_reset; | |||
5581 | } | |||
5582 | ||||
5583 | retry: /* Rest of adevs pre asic reset from XGMI hive. */ | |||
5584 | list_for_each_entry(tmp_adev, device_list_handle, reset_list)for (tmp_adev = ({ const __typeof( ((__typeof(*tmp_adev) *)0) ->reset_list ) *__mptr = ((device_list_handle)->next); ( __typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof(__typeof (*tmp_adev), reset_list) );}); &tmp_adev->reset_list != (device_list_handle); tmp_adev = ({ const __typeof( ((__typeof (*tmp_adev) *)0)->reset_list ) *__mptr = (tmp_adev->reset_list .next); (__typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof (__typeof(*tmp_adev), reset_list) );})) { | |||
5585 | if (gpu_reset_for_dev_remove) { | |||
5586 | /* Workaroud for ASICs need to disable SMC first */ | |||
5587 | amdgpu_device_smu_fini_early(tmp_adev); | |||
5588 | } | |||
5589 | r = amdgpu_device_pre_asic_reset(tmp_adev, reset_context); | |||
5590 | /*TODO Should we stop ?*/ | |||
5591 | if (r) { | |||
5592 | dev_err(tmp_adev->dev, "GPU pre asic reset failed with err, %d for drm dev, %s ",printf("drm:pid%d:%s *ERROR* " "GPU pre asic reset failed with err, %d for drm dev, %s " , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , r, adev_to_drm (tmp_adev)->unique) | |||
5593 | r, adev_to_drm(tmp_adev)->unique)printf("drm:pid%d:%s *ERROR* " "GPU pre asic reset failed with err, %d for drm dev, %s " , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , r, adev_to_drm (tmp_adev)->unique); | |||
5594 | tmp_adev->asic_reset_res = r; | |||
5595 | } | |||
5596 | ||||
5597 | /* | |||
5598 | * Drop all pending non scheduler resets. Scheduler resets | |||
5599 | * were already dropped during drm_sched_stop | |||
5600 | */ | |||
5601 | amdgpu_device_stop_pending_resets(tmp_adev); | |||
5602 | } | |||
5603 | ||||
5604 | tmp_vram_lost_counter = atomic_read(&((adev)->vram_lost_counter))({ typeof(*(&((adev)->vram_lost_counter))) __tmp = *(volatile typeof(*(&((adev)->vram_lost_counter))) *)&(*(& ((adev)->vram_lost_counter))); membar_datadep_consumer(); __tmp ; }); | |||
5605 | /* Actual ASIC resets if needed.*/ | |||
5606 | /* Host driver will handle XGMI hive reset for SRIOV */ | |||
5607 | if (amdgpu_sriov_vf(adev)((adev)->virt.caps & (1 << 2))) { | |||
5608 | r = amdgpu_device_reset_sriov(adev, job ? false0 : true1); | |||
5609 | if (r) | |||
5610 | adev->asic_reset_res = r; | |||
5611 | ||||
5612 | /* Aldebaran supports ras in SRIOV, so need resume ras during reset */ | |||
5613 | if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2)(((9) << 16) | ((4) << 8) | (2))) | |||
5614 | amdgpu_ras_resume(adev); | |||
5615 | } else { | |||
5616 | r = amdgpu_do_asic_reset(device_list_handle, reset_context); | |||
5617 | if (r && r == -EAGAIN35) | |||
5618 | goto retry; | |||
5619 | ||||
5620 | if (!r && gpu_reset_for_dev_remove) | |||
5621 | goto recover_end; | |||
5622 | } | |||
5623 | ||||
5624 | skip_hw_reset: | |||
5625 | ||||
5626 | /* Post ASIC reset for all devs .*/ | |||
5627 | list_for_each_entry(tmp_adev, device_list_handle, reset_list)for (tmp_adev = ({ const __typeof( ((__typeof(*tmp_adev) *)0) ->reset_list ) *__mptr = ((device_list_handle)->next); ( __typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof(__typeof (*tmp_adev), reset_list) );}); &tmp_adev->reset_list != (device_list_handle); tmp_adev = ({ const __typeof( ((__typeof (*tmp_adev) *)0)->reset_list ) *__mptr = (tmp_adev->reset_list .next); (__typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof (__typeof(*tmp_adev), reset_list) );})) { | |||
5628 | ||||
5629 | /* | |||
5630 | * Sometimes a later bad compute job can block a good gfx job as gfx | |||
5631 | * and compute ring share internal GC HW mutually. We add an additional | |||
5632 | * guilty jobs recheck step to find the real guilty job, it synchronously | |||
5633 | * submits and pends for the first job being signaled. If it gets timeout, | |||
5634 | * we identify it as a real guilty job. | |||
5635 | */ | |||
5636 | if (amdgpu_gpu_recovery == 2 && | |||
5637 | !(tmp_vram_lost_counter < atomic_read(&adev->vram_lost_counter)({ typeof(*(&adev->vram_lost_counter)) __tmp = *(volatile typeof(*(&adev->vram_lost_counter)) *)&(*(&adev ->vram_lost_counter)); membar_datadep_consumer(); __tmp; } ))) | |||
5638 | amdgpu_device_recheck_guilty_jobs( | |||
5639 | tmp_adev, device_list_handle, reset_context); | |||
5640 | ||||
5641 | for (i = 0; i < AMDGPU_MAX_RINGS28; ++i) { | |||
5642 | struct amdgpu_ring *ring = tmp_adev->rings[i]; | |||
5643 | ||||
5644 | if (!ring || !ring->sched.thread) | |||
5645 | continue; | |||
5646 | ||||
5647 | /* No point to resubmit jobs if we didn't HW reset*/ | |||
5648 | if (!tmp_adev->asic_reset_res && !job_signaled) | |||
5649 | drm_sched_resubmit_jobs(&ring->sched); | |||
5650 | ||||
5651 | drm_sched_start(&ring->sched, !tmp_adev->asic_reset_res); | |||
5652 | } | |||
5653 | ||||
5654 | if (adev->enable_mes && adev->ip_versions[GC_HWIP][0] != IP_VERSION(11, 0, 3)(((11) << 16) | ((0) << 8) | (3))) | |||
5655 | amdgpu_mes_self_test(tmp_adev); | |||
5656 | ||||
5657 | if (!drm_drv_uses_atomic_modeset(adev_to_drm(tmp_adev)) && !job_signaled) { | |||
5658 | drm_helper_resume_force_mode(adev_to_drm(tmp_adev)); | |||
5659 | } | |||
5660 | ||||
5661 | if (tmp_adev->asic_reset_res) | |||
5662 | r = tmp_adev->asic_reset_res; | |||
5663 | ||||
5664 | tmp_adev->asic_reset_res = 0; | |||
5665 | ||||
5666 | if (r) { | |||
5667 | /* bad news, how to tell it to userspace ? */ | |||
5668 | dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter))do { } while(0); | |||
5669 | amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r); | |||
5670 | } else { | |||
5671 | dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter))do { } while(0); | |||
5672 | if (amdgpu_acpi_smart_shift_update(adev_to_drm(tmp_adev), AMDGPU_SS_DEV_D0)) | |||
5673 | DRM_WARN("smart shift update failed\n")printk("\0014" "[" "drm" "] " "smart shift update failed\n"); | |||
5674 | } | |||
5675 | } | |||
5676 | ||||
5677 | skip_sched_resume: | |||
5678 | list_for_each_entry(tmp_adev, device_list_handle, reset_list)for (tmp_adev = ({ const __typeof( ((__typeof(*tmp_adev) *)0) ->reset_list ) *__mptr = ((device_list_handle)->next); ( __typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof(__typeof (*tmp_adev), reset_list) );}); &tmp_adev->reset_list != (device_list_handle); tmp_adev = ({ const __typeof( ((__typeof (*tmp_adev) *)0)->reset_list ) *__mptr = (tmp_adev->reset_list .next); (__typeof(*tmp_adev) *)( (char *)__mptr - __builtin_offsetof (__typeof(*tmp_adev), reset_list) );})) { | |||
5679 | /* unlock kfd: SRIOV would do it separately */ | |||
5680 | if (!need_emergency_restart && !amdgpu_sriov_vf(tmp_adev)((tmp_adev)->virt.caps & (1 << 2))) | |||
5681 | amdgpu_amdkfd_post_reset(tmp_adev); | |||
5682 | ||||
5683 | /* kfd_post_reset will do nothing if kfd device is not initialized, | |||
5684 | * need to bring up kfd here if it's not be initialized before | |||
5685 | */ | |||
5686 | if (!adev->kfd.init_complete) | |||
5687 | amdgpu_amdkfd_device_init(adev); | |||
5688 | ||||
5689 | if (audio_suspended) | |||
5690 | amdgpu_device_resume_display_audio(tmp_adev); | |||
5691 | ||||
5692 | amdgpu_device_unset_mp1_state(tmp_adev); | |||
5693 | } | |||
5694 | ||||
5695 | recover_end: | |||
5696 | tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,({ const __typeof( ((struct amdgpu_device *)0)->reset_list ) *__mptr = ((device_list_handle)->next); (struct amdgpu_device *)( (char *)__mptr - __builtin_offsetof(struct amdgpu_device , reset_list) );}) | |||
5697 | reset_list)({ const __typeof( ((struct amdgpu_device *)0)->reset_list ) *__mptr = ((device_list_handle)->next); (struct amdgpu_device *)( (char *)__mptr - __builtin_offsetof(struct amdgpu_device , reset_list) );}); | |||
5698 | amdgpu_device_unlock_reset_domain(tmp_adev->reset_domain); | |||
5699 | ||||
5700 | if (hive) { | |||
5701 | mutex_unlock(&hive->hive_lock)rw_exit_write(&hive->hive_lock); | |||
5702 | amdgpu_put_xgmi_hive(hive); | |||
5703 | } | |||
5704 | ||||
5705 | if (r) | |||
5706 | dev_info(adev->dev, "GPU reset end with ret = %d\n", r)do { } while(0); | |||
5707 | ||||
5708 | atomic_set(&adev->reset_domain->reset_res, r)({ typeof(*(&adev->reset_domain->reset_res)) __tmp = ((r)); *(volatile typeof(*(&adev->reset_domain->reset_res )) *)&(*(&adev->reset_domain->reset_res)) = __tmp ; __tmp; }); | |||
5709 | return r; | |||
5710 | } | |||
5711 | ||||
5712 | /** | |||
5713 | * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot | |||
5714 | * | |||
5715 | * @adev: amdgpu_device pointer | |||
5716 | * | |||
5717 | * Fetchs and stores in the driver the PCIE capabilities (gen speed | |||
5718 | * and lanes) of the slot the device is in. Handles APUs and | |||
5719 | * virtualized environments where PCIE config space may not be available. | |||
5720 | */ | |||
5721 | static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev) | |||
5722 | { | |||
5723 | struct pci_dev *pdev; | |||
5724 | enum pci_bus_speed speed_cap, platform_speed_cap; | |||
5725 | enum pcie_link_width platform_link_width; | |||
5726 | ||||
5727 | if (amdgpu_pcie_gen_cap) | |||
5728 | adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap; | |||
5729 | ||||
5730 | if (amdgpu_pcie_lane_cap) | |||
5731 | adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap; | |||
5732 | ||||
5733 | /* covers APUs as well */ | |||
5734 | if (pci_is_root_bus(adev->pdev->bus)) { | |||
5735 | if (adev->pm.pcie_gen_mask == 0) | |||
5736 | adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK(0x00010000 | 0x00020000 | 0x00000001 | 0x00000002 | 0x00000004 ); | |||
5737 | if (adev->pm.pcie_mlw_mask == 0) | |||
5738 | adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK(0x00010000 | 0x00020000 | 0x00040000 | 0x00080000 | 0x00200000 ); | |||
5739 | return; | |||
5740 | } | |||
5741 | ||||
5742 | if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask) | |||
5743 | return; | |||
5744 | ||||
5745 | pcie_bandwidth_available(adev->pdev, NULL((void *)0), | |||
5746 | &platform_speed_cap, &platform_link_width); | |||
5747 | ||||
5748 | if (adev->pm.pcie_gen_mask == 0) { | |||
5749 | /* asic caps */ | |||
5750 | pdev = adev->pdev; | |||
5751 | speed_cap = pcie_get_speed_cap(pdev); | |||
5752 | if (speed_cap == PCI_SPEED_UNKNOWN) { | |||
5753 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN10x00000001 | | |||
5754 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN20x00000002 | | |||
5755 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN30x00000004); | |||
5756 | } else { | |||
5757 | if (speed_cap == PCIE_SPEED_32_0GT) | |||
5758 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN10x00000001 | | |||
5759 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN20x00000002 | | |||
5760 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN30x00000004 | | |||
5761 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN40x00000008 | | |||
5762 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN50x00000010); | |||
5763 | else if (speed_cap == PCIE_SPEED_16_0GT) | |||
5764 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN10x00000001 | | |||
5765 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN20x00000002 | | |||
5766 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN30x00000004 | | |||
5767 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN40x00000008); | |||
5768 | else if (speed_cap == PCIE_SPEED_8_0GT) | |||
5769 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN10x00000001 | | |||
5770 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN20x00000002 | | |||
5771 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN30x00000004); | |||
5772 | else if (speed_cap == PCIE_SPEED_5_0GT) | |||
5773 | adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN10x00000001 | | |||
5774 | CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN20x00000002); | |||
5775 | else | |||
5776 | adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN10x00000001; | |||
5777 | } | |||
5778 | /* platform caps */ | |||
5779 | if (platform_speed_cap == PCI_SPEED_UNKNOWN) { | |||
5780 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN10x00010000 | | |||
5781 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN20x00020000); | |||
5782 | } else { | |||
5783 | if (platform_speed_cap == PCIE_SPEED_32_0GT) | |||
5784 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN10x00010000 | | |||
5785 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN20x00020000 | | |||
5786 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN30x00040000 | | |||
5787 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN40x00080000 | | |||
5788 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN50x00100000); | |||
5789 | else if (platform_speed_cap == PCIE_SPEED_16_0GT) | |||
5790 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN10x00010000 | | |||
5791 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN20x00020000 | | |||
5792 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN30x00040000 | | |||
5793 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN40x00080000); | |||
5794 | else if (platform_speed_cap == PCIE_SPEED_8_0GT) | |||
5795 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN10x00010000 | | |||
5796 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN20x00020000 | | |||
5797 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN30x00040000); | |||
5798 | else if (platform_speed_cap == PCIE_SPEED_5_0GT) | |||
5799 | adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN10x00010000 | | |||
5800 | CAIL_PCIE_LINK_SPEED_SUPPORT_GEN20x00020000); | |||
5801 | else | |||
5802 | adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN10x00010000; | |||
5803 | ||||
5804 | } | |||
5805 | } | |||
5806 | if (adev->pm.pcie_mlw_mask == 0) { | |||
5807 | if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) { | |||
5808 | adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK(0x00010000 | 0x00020000 | 0x00040000 | 0x00080000 | 0x00200000 ); | |||
5809 | } else { | |||
5810 | switch (platform_link_width) { | |||
5811 | case PCIE_LNK_X32: | |||
5812 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X320x00400000 | | |||
5813 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X160x00200000 | | |||
5814 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X120x00100000 | | |||
5815 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X80x00080000 | | |||
5816 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X40x00040000 | | |||
5817 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X20x00020000 | | |||
5818 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X10x00010000); | |||
5819 | break; | |||
5820 | case PCIE_LNK_X16: | |||
5821 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X160x00200000 | | |||
5822 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X120x00100000 | | |||
5823 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X80x00080000 | | |||
5824 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X40x00040000 | | |||
5825 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X20x00020000 | | |||
5826 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X10x00010000); | |||
5827 | break; | |||
5828 | case PCIE_LNK_X12: | |||
5829 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X120x00100000 | | |||
5830 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X80x00080000 | | |||
5831 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X40x00040000 | | |||
5832 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X20x00020000 | | |||
5833 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X10x00010000); | |||
5834 | break; | |||
5835 | case PCIE_LNK_X8: | |||
5836 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X80x00080000 | | |||
5837 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X40x00040000 | | |||
5838 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X20x00020000 | | |||
5839 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X10x00010000); | |||
5840 | break; | |||
5841 | case PCIE_LNK_X4: | |||
5842 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X40x00040000 | | |||
5843 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X20x00020000 | | |||
5844 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X10x00010000); | |||
5845 | break; | |||
5846 | case PCIE_LNK_X2: | |||
5847 | adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X20x00020000 | | |||
5848 | CAIL_PCIE_LINK_WIDTH_SUPPORT_X10x00010000); | |||
5849 | break; | |||
5850 | case PCIE_LNK_X1: | |||
5851 | adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X10x00010000; | |||
5852 | break; | |||
5853 | default: | |||
5854 | break; | |||
5855 | } | |||
5856 | } | |||
5857 | } | |||
5858 | } | |||
5859 | ||||
5860 | /** | |||
5861 | * amdgpu_device_is_peer_accessible - Check peer access through PCIe BAR | |||
5862 | * | |||
5863 | * @adev: amdgpu_device pointer | |||
5864 | * @peer_adev: amdgpu_device pointer for peer device trying to access @adev | |||
5865 | * | |||
5866 | * Return true if @peer_adev can access (DMA) @adev through the PCIe | |||
5867 | * BAR, i.e. @adev is "large BAR" and the BAR matches the DMA mask of | |||
5868 | * @peer_adev. | |||
5869 | */ | |||
5870 | bool_Bool amdgpu_device_is_peer_accessible(struct amdgpu_device *adev, | |||
5871 | struct amdgpu_device *peer_adev) | |||
5872 | { | |||
5873 | #ifdef CONFIG_HSA_AMD_P2P | |||
5874 | uint64_t address_mask = peer_adev->dev->dma_mask ? | |||
5875 | ~*peer_adev->dev->dma_mask : ~((1ULL << 32) - 1); | |||
5876 | resource_size_t aper_limit = | |||
5877 | adev->gmc.aper_base + adev->gmc.aper_size - 1; | |||
5878 | bool_Bool p2p_access = | |||
5879 | !adev->gmc.xgmi.connected_to_cpu && | |||
5880 | !(pci_p2pdma_distance(adev->pdev, peer_adev->dev, false0) < 0); | |||
5881 | ||||
5882 | return pcie_p2p && p2p_access && (adev->gmc.visible_vram_size && | |||
5883 | adev->gmc.real_vram_size == adev->gmc.visible_vram_size && | |||
5884 | !(adev->gmc.aper_base & address_mask || | |||
5885 | aper_limit & address_mask)); | |||
5886 | #else | |||
5887 | return false0; | |||
5888 | #endif | |||
5889 | } | |||
5890 | ||||
5891 | int amdgpu_device_baco_enter(struct drm_device *dev) | |||
5892 | { | |||
5893 | struct amdgpu_device *adev = drm_to_adev(dev); | |||
5894 | struct amdgpu_ras *ras = amdgpu_ras_get_context(adev); | |||
5895 | ||||
5896 | if (!amdgpu_device_supports_baco(adev_to_drm(adev))) | |||
5897 | return -ENOTSUPP91; | |||
5898 | ||||
5899 | if (ras && adev->ras_enabled && | |||
5900 | adev->nbio.funcs->enable_doorbell_interrupt) | |||
5901 | adev->nbio.funcs->enable_doorbell_interrupt(adev, false0); | |||
5902 | ||||
5903 | return amdgpu_dpm_baco_enter(adev); | |||
5904 | } | |||
5905 | ||||
5906 | int amdgpu_device_baco_exit(struct drm_device *dev) | |||
5907 | { | |||
5908 | struct amdgpu_device *adev = drm_to_adev(dev); | |||
5909 | struct amdgpu_ras *ras = amdgpu_ras_get_context(adev); | |||
5910 | int ret = 0; | |||
5911 | ||||
5912 | if (!amdgpu_device_supports_baco(adev_to_drm(adev))) | |||
5913 | return -ENOTSUPP91; | |||
5914 | ||||
5915 | ret = amdgpu_dpm_baco_exit(adev); | |||
5916 | if (ret) | |||
5917 | return ret; | |||
5918 | ||||
5919 | if (ras && adev->ras_enabled && | |||
5920 | adev->nbio.funcs->enable_doorbell_interrupt) | |||
5921 | adev->nbio.funcs->enable_doorbell_interrupt(adev, true1); | |||
5922 | ||||
5923 | if (amdgpu_passthrough(adev)((adev)->virt.caps & (1 << 3)) && | |||
5924 | adev->nbio.funcs->clear_doorbell_interrupt) | |||
5925 | adev->nbio.funcs->clear_doorbell_interrupt(adev); | |||
5926 | ||||
5927 | return 0; | |||
5928 | } | |||
5929 | ||||
5930 | /** | |||
5931 | * amdgpu_pci_error_detected - Called when a PCI error is detected. | |||
5932 | * @pdev: PCI device struct | |||
5933 | * @state: PCI channel state | |||
5934 | * | |||
5935 | * Description: Called when a PCI error is detected. | |||
5936 | * | |||
5937 | * Return: PCI_ERS_RESULT_NEED_RESET or PCI_ERS_RESULT_DISCONNECT. | |||
5938 | */ | |||
5939 | pci_ers_result_t amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state) | |||
5940 | { | |||
5941 | STUB()do { printf("%s: stub\n", __func__); } while(0); | |||
5942 | return 0; | |||
5943 | #ifdef notyet | |||
5944 | struct drm_device *dev = pci_get_drvdata(pdev); | |||
5945 | struct amdgpu_device *adev = drm_to_adev(dev); | |||
5946 | int i; | |||
5947 | ||||
5948 | DRM_INFO("PCI error: detected callback, state(%d)!!\n", state)printk("\0016" "[" "drm" "] " "PCI error: detected callback, state(%d)!!\n" , state); | |||
5949 | ||||
5950 | if (adev->gmc.xgmi.num_physical_nodes > 1) { | |||
5951 | DRM_WARN("No support for XGMI hive yet...")printk("\0014" "[" "drm" "] " "No support for XGMI hive yet..." ); | |||
5952 | return PCI_ERS_RESULT_DISCONNECT0; | |||
5953 | } | |||
5954 | ||||
5955 | adev->pci_channel_state = state; | |||
5956 | ||||
5957 | switch (state) { | |||
5958 | case pci_channel_io_normal: | |||
5959 | return PCI_ERS_RESULT_CAN_RECOVER; | |||
5960 | /* Fatal error, prepare for slot reset */ | |||
5961 | case pci_channel_io_frozen: | |||
5962 | /* | |||
5963 | * Locking adev->reset_domain->sem will prevent any external access | |||
5964 | * to GPU during PCI error recovery | |||
5965 | */ | |||
5966 | amdgpu_device_lock_reset_domain(adev->reset_domain); | |||
5967 | amdgpu_device_set_mp1_state(adev); | |||
5968 | ||||
5969 | /* | |||
5970 | * Block any work scheduling as we do for regular GPU reset | |||
5971 | * for the duration of the recovery | |||
5972 | */ | |||
5973 | for (i = 0; i < AMDGPU_MAX_RINGS28; ++i) { | |||
5974 | struct amdgpu_ring *ring = adev->rings[i]; | |||
5975 | ||||
5976 | if (!ring || !ring->sched.thread) | |||
5977 | continue; | |||
5978 | ||||
5979 | drm_sched_stop(&ring->sched, NULL((void *)0)); | |||
5980 | } | |||
5981 | atomic_inc(&adev->gpu_reset_counter)__sync_fetch_and_add(&adev->gpu_reset_counter, 1); | |||
5982 | return PCI_ERS_RESULT_NEED_RESET; | |||
5983 | case pci_channel_io_perm_failure: | |||
5984 | /* Permanent error, prepare for device removal */ | |||
5985 | return PCI_ERS_RESULT_DISCONNECT0; | |||
5986 | } | |||
5987 | ||||
5988 | return PCI_ERS_RESULT_NEED_RESET; | |||
5989 | #endif | |||
5990 | } | |||
5991 | ||||
5992 | /** | |||
5993 | * amdgpu_pci_mmio_enabled - Enable MMIO and dump debug registers | |||
5994 | * @pdev: pointer to PCI device | |||
5995 | */ | |||
5996 | pci_ers_result_t amdgpu_pci_mmio_enabled(struct pci_dev *pdev) | |||
5997 | { | |||
5998 | ||||
5999 | DRM_INFO("PCI error: mmio enabled callback!!\n")printk("\0016" "[" "drm" "] " "PCI error: mmio enabled callback!!\n" ); | |||
6000 | ||||
6001 | /* TODO - dump whatever for debugging purposes */ | |||
6002 | ||||
6003 | /* This called only if amdgpu_pci_error_detected returns | |||
6004 | * PCI_ERS_RESULT_CAN_RECOVER. Read/write to the device still | |||
6005 | * works, no need to reset slot. | |||
6006 | */ | |||
6007 | ||||
6008 | return PCI_ERS_RESULT_RECOVERED1; | |||
6009 | } | |||
6010 | ||||
6011 | /** | |||
6012 | * amdgpu_pci_slot_reset - Called when PCI slot has been reset. | |||
6013 | * @pdev: PCI device struct | |||
6014 | * | |||
6015 | * Description: This routine is called by the pci error recovery | |||
6016 | * code after the PCI slot has been reset, just before we | |||
6017 | * should resume normal operations. | |||
6018 | */ | |||
6019 | pci_ers_result_t amdgpu_pci_slot_reset(struct pci_dev *pdev) | |||
6020 | { | |||
6021 | STUB()do { printf("%s: stub\n", __func__); } while(0); | |||
6022 | return PCI_ERS_RESULT_RECOVERED1; | |||
6023 | #ifdef notyet | |||
6024 | struct drm_device *dev = pci_get_drvdata(pdev); | |||
6025 | struct amdgpu_device *adev = drm_to_adev(dev); | |||
6026 | int r, i; | |||
6027 | struct amdgpu_reset_context reset_context; | |||
6028 | u32 memsize; | |||
6029 | struct list_head device_list; | |||
6030 | ||||
6031 | DRM_INFO("PCI error: slot reset callback!!\n")printk("\0016" "[" "drm" "] " "PCI error: slot reset callback!!\n" ); | |||
6032 | ||||
6033 | memset(&reset_context, 0, sizeof(reset_context))__builtin_memset((&reset_context), (0), (sizeof(reset_context ))); | |||
6034 | ||||
6035 | INIT_LIST_HEAD(&device_list); | |||
6036 | list_add_tail(&adev->reset_list, &device_list); | |||
6037 | ||||
6038 | /* wait for asic to come out of reset */ | |||
6039 | drm_msleep(500)mdelay(500); | |||
6040 | ||||
6041 | /* Restore PCI confspace */ | |||
6042 | amdgpu_device_load_pci_state(pdev); | |||
6043 | ||||
6044 | /* confirm ASIC came out of reset */ | |||
6045 | for (i = 0; i < adev->usec_timeout; i++) { | |||
6046 | memsize = amdgpu_asic_get_config_memsize(adev)(adev)->asic_funcs->get_config_memsize((adev)); | |||
6047 | ||||
6048 | if (memsize != 0xffffffff) | |||
6049 | break; | |||
6050 | udelay(1); | |||
6051 | } | |||
6052 | if (memsize == 0xffffffff) { | |||
6053 | r = -ETIME60; | |||
6054 | goto out; | |||
6055 | } | |||
6056 | ||||
6057 | reset_context.method = AMD_RESET_METHOD_NONE; | |||
6058 | reset_context.reset_req_dev = adev; | |||
6059 | set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags); | |||
6060 | set_bit(AMDGPU_SKIP_HW_RESET, &reset_context.flags); | |||
6061 | ||||
6062 | adev->no_hw_access = true1; | |||
6063 | r = amdgpu_device_pre_asic_reset(adev, &reset_context); | |||
6064 | adev->no_hw_access = false0; | |||
6065 | if (r) | |||
6066 | goto out; | |||
6067 | ||||
6068 | r = amdgpu_do_asic_reset(&device_list, &reset_context); | |||
6069 | ||||
6070 | out: | |||
6071 | if (!r) { | |||
6072 | if (amdgpu_device_cache_pci_state(adev->pdev)) | |||
6073 | pci_restore_state(adev->pdev); | |||
6074 | ||||
6075 | DRM_INFO("PCIe error recovery succeeded\n")printk("\0016" "[" "drm" "] " "PCIe error recovery succeeded\n" ); | |||
6076 | } else { | |||
6077 | DRM_ERROR("PCIe error recovery failed, err:%d", r)__drm_err("PCIe error recovery failed, err:%d", r); | |||
6078 | amdgpu_device_unset_mp1_state(adev); | |||
6079 | amdgpu_device_unlock_reset_domain(adev->reset_domain); | |||
6080 | } | |||
6081 | ||||
6082 | return r ? PCI_ERS_RESULT_DISCONNECT0 : PCI_ERS_RESULT_RECOVERED1; | |||
6083 | #endif | |||
6084 | } | |||
6085 | ||||
6086 | /** | |||
6087 | * amdgpu_pci_resume() - resume normal ops after PCI reset | |||
6088 | * @pdev: pointer to PCI device | |||
6089 | * | |||
6090 | * Called when the error recovery driver tells us that its | |||
6091 | * OK to resume normal operation. | |||
6092 | */ | |||
6093 | void amdgpu_pci_resume(struct pci_dev *pdev) | |||
6094 | { | |||
6095 | STUB()do { printf("%s: stub\n", __func__); } while(0); | |||
6096 | #ifdef notyet | |||
6097 | struct drm_device *dev = pci_get_drvdata(pdev); | |||
6098 | struct amdgpu_device *adev = drm_to_adev(dev); | |||
6099 | int i; | |||
6100 | ||||
6101 | ||||
6102 | DRM_INFO("PCI error: resume callback!!\n")printk("\0016" "[" "drm" "] " "PCI error: resume callback!!\n" ); | |||
6103 | ||||
6104 | /* Only continue execution for the case of pci_channel_io_frozen */ | |||
6105 | if (adev->pci_channel_state != pci_channel_io_frozen) | |||
6106 | return; | |||
6107 | ||||
6108 | for (i = 0; i < AMDGPU_MAX_RINGS28; ++i) { | |||
6109 | struct amdgpu_ring *ring = adev->rings[i]; | |||
6110 | ||||
6111 | if (!ring || !ring->sched.thread) | |||
6112 | continue; | |||
6113 | ||||
6114 | ||||
6115 | drm_sched_resubmit_jobs(&ring->sched); | |||
6116 | drm_sched_start(&ring->sched, true1); | |||
6117 | } | |||
6118 | ||||
6119 | amdgpu_device_unset_mp1_state(adev); | |||
6120 | amdgpu_device_unlock_reset_domain(adev->reset_domain); | |||
6121 | #endif | |||
6122 | } | |||
6123 | ||||
6124 | bool_Bool amdgpu_device_cache_pci_state(struct pci_dev *pdev) | |||
6125 | { | |||
6126 | return false0; | |||
6127 | #ifdef notyet | |||
6128 | struct drm_device *dev = pci_get_drvdata(pdev); | |||
6129 | struct amdgpu_device *adev = drm_to_adev(dev); | |||
6130 | int r; | |||
6131 | ||||
6132 | r = pci_save_state(pdev); | |||
6133 | if (!r) { | |||
6134 | kfree(adev->pci_state); | |||
6135 | ||||
6136 | adev->pci_state = pci_store_saved_state(pdev); | |||
6137 | ||||
6138 | if (!adev->pci_state) { | |||
6139 | DRM_ERROR("Failed to store PCI saved state")__drm_err("Failed to store PCI saved state"); | |||
6140 | return false0; | |||
6141 | } | |||
6142 | } else { | |||
6143 | DRM_WARN("Failed to save PCI state, err:%d\n", r)printk("\0014" "[" "drm" "] " "Failed to save PCI state, err:%d\n" , r); | |||
6144 | return false0; | |||
6145 | } | |||
6146 | ||||
6147 | return true1; | |||
6148 | #endif | |||
6149 | } | |||
6150 | ||||
6151 | bool_Bool amdgpu_device_load_pci_state(struct pci_dev *pdev) | |||
6152 | { | |||
6153 | STUB()do { printf("%s: stub\n", __func__); } while(0); | |||
6154 | return false0; | |||
6155 | #ifdef notyet | |||
6156 | struct drm_device *dev = pci_get_drvdata(pdev); | |||
6157 | struct amdgpu_device *adev = drm_to_adev(dev); | |||
6158 | int r; | |||
6159 | ||||
6160 | if (!adev->pci_state) | |||
6161 | return false0; | |||
6162 | ||||
6163 | r = pci_load_saved_state(pdev, adev->pci_state); | |||
6164 | ||||
6165 | if (!r) { | |||
6166 | pci_restore_state(pdev); | |||
6167 | } else { | |||
6168 | DRM_WARN("Failed to load PCI state, err:%d\n", r)printk("\0014" "[" "drm" "] " "Failed to load PCI state, err:%d\n" , r); | |||
6169 | return false0; | |||
6170 | } | |||
6171 | ||||
6172 | return true1; | |||
6173 | #endif | |||
6174 | } | |||
6175 | ||||
6176 | void amdgpu_device_flush_hdp(struct amdgpu_device *adev, | |||
6177 | struct amdgpu_ring *ring) | |||
6178 | { | |||
6179 | #ifdef CONFIG_X86_641 | |||
6180 | if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)((adev)->virt.caps & (1 << 3))) | |||
6181 | return; | |||
6182 | #endif | |||
6183 | if (adev->gmc.xgmi.connected_to_cpu) | |||
6184 | return; | |||
6185 | ||||
6186 | if (ring && ring->funcs->emit_hdp_flush) | |||
6187 | amdgpu_ring_emit_hdp_flush(ring)(ring)->funcs->emit_hdp_flush((ring)); | |||
6188 | else | |||
6189 | amdgpu_asic_flush_hdp(adev, ring)((adev)->asic_funcs->flush_hdp ? (adev)->asic_funcs-> flush_hdp((adev), (ring)) : (adev)->hdp.funcs->flush_hdp ((adev), (ring))); | |||
6190 | } | |||
6191 | ||||
6192 | void amdgpu_device_invalidate_hdp(struct amdgpu_device *adev, | |||
6193 | struct amdgpu_ring *ring) | |||
6194 | { | |||
6195 | #ifdef CONFIG_X86_641 | |||
6196 | if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)((adev)->virt.caps & (1 << 3))) | |||
6197 | return; | |||
6198 | #endif | |||
6199 | if (adev->gmc.xgmi.connected_to_cpu) | |||
6200 | return; | |||
6201 | ||||
6202 | amdgpu_asic_invalidate_hdp(adev, ring)((adev)->asic_funcs->invalidate_hdp ? (adev)->asic_funcs ->invalidate_hdp((adev), (ring)) : ((adev)->hdp.funcs-> invalidate_hdp ? (adev)->hdp.funcs->invalidate_hdp((adev ), (ring)) : 0)); | |||
6203 | } | |||
6204 | ||||
6205 | int amdgpu_in_reset(struct amdgpu_device *adev) | |||
6206 | { | |||
6207 | return atomic_read(&adev->reset_domain->in_gpu_reset)({ typeof(*(&adev->reset_domain->in_gpu_reset)) __tmp = *(volatile typeof(*(&adev->reset_domain->in_gpu_reset )) *)&(*(&adev->reset_domain->in_gpu_reset)); membar_datadep_consumer (); __tmp; }); | |||
6208 | } | |||
6209 | ||||
6210 | /** | |||
6211 | * amdgpu_device_halt() - bring hardware to some kind of halt state | |||
6212 | * | |||
6213 | * @adev: amdgpu_device pointer | |||
6214 | * | |||
6215 | * Bring hardware to some kind of halt state so that no one can touch it | |||
6216 | * any more. It will help to maintain error context when error occurred. | |||
6217 | * Compare to a simple hang, the system will keep stable at least for SSH | |||
6218 | * access. Then it should be trivial to inspect the hardware state and | |||
6219 | * see what's going on. Implemented as following: | |||
6220 | * | |||
6221 | * 1. drm_dev_unplug() makes device inaccessible to user space(IOCTLs, etc), | |||
6222 | * clears all CPU mappings to device, disallows remappings through page faults | |||
6223 | * 2. amdgpu_irq_disable_all() disables all interrupts | |||
6224 | * 3. amdgpu_fence_driver_hw_fini() signals all HW fences | |||
6225 | * 4. set adev->no_hw_access to avoid potential crashes after setp 5 | |||
6226 | * 5. amdgpu_device_unmap_mmio() clears all MMIO mappings | |||
6227 | * 6. pci_disable_device() and pci_wait_for_pending_transaction() | |||
6228 | * flush any in flight DMA operations | |||
6229 | */ | |||
6230 | void amdgpu_device_halt(struct amdgpu_device *adev) | |||
6231 | { | |||
6232 | struct pci_dev *pdev = adev->pdev; | |||
6233 | struct drm_device *ddev = adev_to_drm(adev); | |||
6234 | ||||
6235 | drm_dev_unplug(ddev); | |||
6236 | ||||
6237 | amdgpu_irq_disable_all(adev); | |||
6238 | ||||
6239 | amdgpu_fence_driver_hw_fini(adev); | |||
6240 | ||||
6241 | adev->no_hw_access = true1; | |||
6242 | ||||
6243 | amdgpu_device_unmap_mmio(adev); | |||
6244 | ||||
6245 | pci_disable_device(pdev); | |||
6246 | pci_wait_for_pending_transaction(pdev); | |||
6247 | } | |||
6248 | ||||
6249 | u32 amdgpu_device_pcie_port_rreg(struct amdgpu_device *adev, | |||
6250 | u32 reg) | |||
6251 | { | |||
6252 | unsigned long flags, address, data; | |||
6253 | u32 r; | |||
6254 | ||||
6255 | address = adev->nbio.funcs->get_pcie_port_index_offset(adev); | |||
6256 | data = adev->nbio.funcs->get_pcie_port_data_offset(adev); | |||
6257 | ||||
6258 | spin_lock_irqsave(&adev->pcie_idx_lock, flags)do { flags = 0; mtx_enter(&adev->pcie_idx_lock); } while (0); | |||
6259 | WREG32(address, reg * 4)amdgpu_device_wreg(adev, (address), (reg * 4), 0); | |||
6260 | (void)RREG32(address)amdgpu_device_rreg(adev, (address), 0); | |||
6261 | r = RREG32(data)amdgpu_device_rreg(adev, (data), 0); | |||
6262 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags)do { (void)(flags); mtx_leave(&adev->pcie_idx_lock); } while (0); | |||
6263 | return r; | |||
6264 | } | |||
6265 | ||||
6266 | void amdgpu_device_pcie_port_wreg(struct amdgpu_device *adev, | |||
6267 | u32 reg, u32 v) | |||
6268 | { | |||
6269 | unsigned long flags, address, data; | |||
6270 | ||||
6271 | address = adev->nbio.funcs->get_pcie_port_index_offset(adev); | |||
6272 | data = adev->nbio.funcs->get_pcie_port_data_offset(adev); | |||
6273 | ||||
6274 | spin_lock_irqsave(&adev->pcie_idx_lock, flags)do { flags = 0; mtx_enter(&adev->pcie_idx_lock); } while (0); | |||
6275 | WREG32(address, reg * 4)amdgpu_device_wreg(adev, (address), (reg * 4), 0); | |||
6276 | (void)RREG32(address)amdgpu_device_rreg(adev, (address), 0); | |||
6277 | WREG32(data, v)amdgpu_device_wreg(adev, (data), (v), 0); | |||
6278 | (void)RREG32(data)amdgpu_device_rreg(adev, (data), 0); | |||
6279 | spin_unlock_irqrestore(&adev->pcie_idx_lock, flags)do { (void)(flags); mtx_leave(&adev->pcie_idx_lock); } while (0); | |||
6280 | } | |||
6281 | ||||
6282 | /** | |||
6283 | * amdgpu_device_switch_gang - switch to a new gang | |||
6284 | * @adev: amdgpu_device pointer | |||
6285 | * @gang: the gang to switch to | |||
6286 | * | |||
6287 | * Try to switch to a new gang. | |||
6288 | * Returns: NULL if we switched to the new gang or a reference to the current | |||
6289 | * gang leader. | |||
6290 | */ | |||
6291 | struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev, | |||
6292 | struct dma_fence *gang) | |||
6293 | { | |||
6294 | struct dma_fence *old = NULL((void *)0); | |||
6295 | ||||
6296 | do { | |||
6297 | dma_fence_put(old); | |||
6298 | rcu_read_lock(); | |||
6299 | old = dma_fence_get_rcu_safe(&adev->gang_submit); | |||
6300 | rcu_read_unlock(); | |||
6301 | ||||
6302 | if (old == gang) | |||
6303 | break; | |||
6304 | ||||
6305 | if (!dma_fence_is_signaled(old)) | |||
6306 | return old; | |||
6307 | ||||
6308 | } while (cmpxchg((struct dma_fence __force **)&adev->gang_submit,__sync_val_compare_and_swap((struct dma_fence **)&adev-> gang_submit, old, gang) | |||
6309 | old, gang)__sync_val_compare_and_swap((struct dma_fence **)&adev-> gang_submit, old, gang) != old); | |||
6310 | ||||
6311 | dma_fence_put(old); | |||
6312 | return NULL((void *)0); | |||
6313 | } | |||
6314 | ||||
6315 | bool_Bool amdgpu_device_has_display_hardware(struct amdgpu_device *adev) | |||
6316 | { | |||
6317 | switch (adev->asic_type) { | |||
6318 | #ifdef CONFIG_DRM_AMDGPU_SI | |||
6319 | case CHIP_HAINAN: | |||
6320 | #endif | |||
6321 | case CHIP_TOPAZ: | |||
6322 | /* chips with no display hardware */ | |||
6323 | return false0; | |||
6324 | #ifdef CONFIG_DRM_AMDGPU_SI | |||
6325 | case CHIP_TAHITI: | |||
6326 | case CHIP_PITCAIRN: | |||
6327 | case CHIP_VERDE: | |||
6328 | case CHIP_OLAND: | |||
6329 | #endif | |||
6330 | #ifdef CONFIG_DRM_AMDGPU_CIK | |||
6331 | case CHIP_BONAIRE: | |||
6332 | case CHIP_HAWAII: | |||
6333 | case CHIP_KAVERI: | |||
6334 | case CHIP_KABINI: | |||
6335 | case CHIP_MULLINS: | |||
6336 | #endif | |||
6337 | case CHIP_TONGA: | |||
6338 | case CHIP_FIJI: | |||
6339 | case CHIP_POLARIS10: | |||
6340 | case CHIP_POLARIS11: | |||
6341 | case CHIP_POLARIS12: | |||
6342 | case CHIP_VEGAM: | |||
6343 | case CHIP_CARRIZO: | |||
6344 | case CHIP_STONEY: | |||
6345 | /* chips with display hardware */ | |||
6346 | return true1; | |||
6347 | default: | |||
6348 | /* IP discovery */ | |||
6349 | if (!adev->ip_versions[DCE_HWIP][0] || | |||
6350 | (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK)) | |||
6351 | return false0; | |||
6352 | return true1; | |||
6353 | } | |||
6354 | } |