Bug Summary

File:dev/pci/drm/amd/amdgpu/amdgpu_vm.c
Warning:line 1577, column 41
The result of the left shift is undefined due to shifting by '4294967295', which is greater or equal to the width of type 'uint64_t'

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple amd64-unknown-openbsd7.0 -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name amdgpu_vm.c -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -mrelocation-model static -mframe-pointer=all -relaxed-aliasing -fno-rounding-math -mconstructor-aliases -ffreestanding -mcmodel=kernel -target-cpu x86-64 -target-feature +retpoline-indirect-calls -target-feature +retpoline-indirect-branches -target-feature -sse2 -target-feature -sse -target-feature -3dnow -target-feature -mmx -target-feature +save-args -disable-red-zone -no-implicit-float -tune-cpu generic -debugger-tuning=gdb -fcoverage-compilation-dir=/usr/src/sys/arch/amd64/compile/GENERIC.MP/obj -nostdsysteminc -nobuiltininc -resource-dir /usr/local/lib/clang/13.0.0 -I /usr/src/sys -I /usr/src/sys/arch/amd64/compile/GENERIC.MP/obj -I /usr/src/sys/arch -I /usr/src/sys/dev/pci/drm/include -I /usr/src/sys/dev/pci/drm/include/uapi -I /usr/src/sys/dev/pci/drm/amd/include/asic_reg -I /usr/src/sys/dev/pci/drm/amd/include -I /usr/src/sys/dev/pci/drm/amd/amdgpu -I /usr/src/sys/dev/pci/drm/amd/display -I /usr/src/sys/dev/pci/drm/amd/display/include -I /usr/src/sys/dev/pci/drm/amd/display/dc -I /usr/src/sys/dev/pci/drm/amd/display/amdgpu_dm -I /usr/src/sys/dev/pci/drm/amd/pm/inc -I /usr/src/sys/dev/pci/drm/amd/pm/swsmu -I /usr/src/sys/dev/pci/drm/amd/pm/swsmu/smu11 -I /usr/src/sys/dev/pci/drm/amd/pm/swsmu/smu12 -I /usr/src/sys/dev/pci/drm/amd/pm/powerplay -I /usr/src/sys/dev/pci/drm/amd/pm/powerplay/hwmgr -I /usr/src/sys/dev/pci/drm/amd/pm/powerplay/smumgr -I /usr/src/sys/dev/pci/drm/amd/display/dc/inc -I /usr/src/sys/dev/pci/drm/amd/display/dc/inc/hw -I /usr/src/sys/dev/pci/drm/amd/display/dc/clk_mgr -I /usr/src/sys/dev/pci/drm/amd/display/modules/inc -I /usr/src/sys/dev/pci/drm/amd/display/modules/hdcp -I /usr/src/sys/dev/pci/drm/amd/display/dmub/inc -I /usr/src/sys/dev/pci/drm/i915 -D DDB -D DIAGNOSTIC -D KTRACE -D ACCOUNTING -D KMEMSTATS -D PTRACE -D POOL_DEBUG -D CRYPTO -D SYSVMSG -D SYSVSEM -D SYSVSHM -D UVM_SWAP_ENCRYPT -D FFS -D FFS2 -D FFS_SOFTUPDATES -D UFS_DIRHASH -D QUOTA -D EXT2FS -D MFS -D NFSCLIENT -D NFSSERVER -D CD9660 -D UDF -D MSDOSFS -D FIFO -D FUSE -D SOCKET_SPLICE -D TCP_ECN -D TCP_SIGNATURE -D INET6 -D IPSEC -D PPP_BSDCOMP -D PPP_DEFLATE -D PIPEX -D MROUTING -D MPLS -D BOOT_CONFIG -D USER_PCICONF -D APERTURE -D MTRR -D NTFS -D HIBERNATE -D PCIVERBOSE -D USBVERBOSE -D WSDISPLAY_COMPAT_USL -D WSDISPLAY_COMPAT_RAWKBD -D WSDISPLAY_DEFAULTSCREENS=6 -D X86EMU -D ONEWIREVERBOSE -D MULTIPROCESSOR -D MAXUSERS=80 -D _KERNEL -D CONFIG_DRM_AMD_DC_DCN3_0 -O2 -Wno-pointer-sign -Wno-address-of-packed-member -Wno-constant-conversion -Wno-unused-but-set-variable -Wno-gnu-folding-constant -fdebug-compilation-dir=/usr/src/sys/arch/amd64/compile/GENERIC.MP/obj -ferror-limit 19 -fwrapv -D_RET_PROTECTOR -ret-protector -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -fno-builtin-malloc -fno-builtin-calloc -fno-builtin-realloc -fno-builtin-valloc -fno-builtin-free -fno-builtin-strdup -fno-builtin-strndup -analyzer-output=html -faddrsig -o /usr/obj/sys/arch/amd64/compile/GENERIC.MP/scan-build/2022-01-12-131800-47421-1 -x c /usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_vm.c
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/dma-fence-array.h>
29#include <linux/interval_tree_generic.h>
30#include <linux/idr.h>
31#include <linux/dma-buf.h>
32
33#include <drm/amdgpu_drm.h>
34#include "amdgpu.h"
35#include "amdgpu_trace.h"
36#include "amdgpu_amdkfd.h"
37#include "amdgpu_gmc.h"
38#include "amdgpu_xgmi.h"
39#include "amdgpu_dma_buf.h"
40
41/**
42 * DOC: GPUVM
43 *
44 * GPUVM is similar to the legacy gart on older asics, however
45 * rather than there being a single global gart table
46 * for the entire GPU, there are multiple VM page tables active
47 * at any given time. The VM page tables can contain a mix
48 * vram pages and system memory pages and system memory pages
49 * can be mapped as snooped (cached system pages) or unsnooped
50 * (uncached system pages).
51 * Each VM has an ID associated with it and there is a page table
52 * associated with each VMID. When execting a command buffer,
53 * the kernel tells the the ring what VMID to use for that command
54 * buffer. VMIDs are allocated dynamically as commands are submitted.
55 * The userspace drivers maintain their own address space and the kernel
56 * sets up their pages tables accordingly when they submit their
57 * command buffers and a VMID is assigned.
58 * Cayman/Trinity support up to 8 active VMs at any given time;
59 * SI supports 16.
60 */
61
62#define START(node) ((node)->start)
63#define LAST(node) ((node)->last)
64
65#ifdef __linux__
66INTERVAL_TREE_DEFINE(struct amdgpu_bo_va_mapping, rb, uint64_t, __subtree_last,
67 START, LAST, static, amdgpu_vm_it)
68#else
69static struct amdgpu_bo_va_mapping *
70amdgpu_vm_it_iter_first(struct rb_root_cached *root, uint64_t start,
71 uint64_t last)
72{
73 struct amdgpu_bo_va_mapping *node;
74 struct rb_node *rb;
75
76 for (rb = rb_first_cached(root)linux_root_RB_MINMAX((struct linux_root *)(&(root)->rb_root
), -1)
; rb; rb = rb_next(rb)linux_root_RB_NEXT((rb))) {
77 node = rb_entry(rb, typeof(*node), rb)({ const __typeof( ((typeof(*node) *)0)->rb ) *__mptr = (rb
); (typeof(*node) *)( (char *)__mptr - __builtin_offsetof(typeof
(*node), rb) );})
;
78 if (LAST(node) >= start && START(node) <= last)
79 return node;
80 }
81 return NULL((void *)0);
82}
83
84static struct amdgpu_bo_va_mapping *
85amdgpu_vm_it_iter_next(struct amdgpu_bo_va_mapping *node, uint64_t start,
86 uint64_t last)
87{
88 STUB()do { printf("%s: stub\n", __func__); } while(0);
89 struct rb_node *rb = &node->rb;
90
91 for (rb = rb_next(rb)linux_root_RB_NEXT((rb)); rb; rb = rb_next(rb)linux_root_RB_NEXT((rb))) {
92 node = rb_entry(rb, typeof(*node), rb)({ const __typeof( ((typeof(*node) *)0)->rb ) *__mptr = (rb
); (typeof(*node) *)( (char *)__mptr - __builtin_offsetof(typeof
(*node), rb) );})
;
93 if (LAST(node) >= start && START(node) <= last)
94 return node;
95 }
96 return NULL((void *)0);
97}
98
99static void
100amdgpu_vm_it_remove(struct amdgpu_bo_va_mapping *node,
101 struct rb_root_cached *root)
102{
103 rb_erase_cached(&node->rb, root)linux_root_RB_REMOVE((struct linux_root *)(&(root)->rb_root
), (&node->rb))
;
104}
105
106static void
107amdgpu_vm_it_insert(struct amdgpu_bo_va_mapping *node,
108 struct rb_root_cached *root)
109{
110 struct rb_node **iter = &root->rb_root.rb_node;
111 struct rb_node *parent = NULL((void *)0);
112 struct amdgpu_bo_va_mapping *iter_node;
113
114 while (*iter) {
115 parent = *iter;
116 iter_node = rb_entry(*iter, struct amdgpu_bo_va_mapping, rb)({ const __typeof( ((struct amdgpu_bo_va_mapping *)0)->rb )
*__mptr = (*iter); (struct amdgpu_bo_va_mapping *)( (char *)
__mptr - __builtin_offsetof(struct amdgpu_bo_va_mapping, rb) )
;})
;
117
118 if (node->start < iter_node->start)
119 iter = &(*iter)->rb_left__entry.rbe_left;
120 else
121 iter = &(*iter)->rb_right__entry.rbe_right;
122 }
123
124 rb_link_node(&node->rb, parent, iter);
125 rb_insert_color_cached(&node->rb, root, false)linux_root_RB_INSERT_COLOR((struct linux_root *)(&(root)->
rb_root), (&node->rb))
;
126}
127#endif
128
129#undef START
130#undef LAST
131
132/**
133 * struct amdgpu_prt_cb - Helper to disable partial resident texture feature from a fence callback
134 */
135struct amdgpu_prt_cb {
136
137 /**
138 * @adev: amdgpu device
139 */
140 struct amdgpu_device *adev;
141
142 /**
143 * @cb: callback
144 */
145 struct dma_fence_cb cb;
146};
147
148/*
149 * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS
150 * happens while holding this lock anywhere to prevent deadlocks when
151 * an MMU notifier runs in reclaim-FS context.
152 */
153static inline void amdgpu_vm_eviction_lock(struct amdgpu_vm *vm)
154{
155 mutex_lock(&vm->eviction_lock)rw_enter_write(&vm->eviction_lock);
156#ifdef notyet
157 vm->saved_flags = memalloc_nofs_save();
158#endif
159}
160
161static inline int amdgpu_vm_eviction_trylock(struct amdgpu_vm *vm)
162{
163 if (mutex_trylock(&vm->eviction_lock)(rw_enter(&vm->eviction_lock, 0x0001UL | 0x0040UL) == 0
)
) {
164#ifdef notyet
165 vm->saved_flags = memalloc_nofs_save();
166#endif
167 return 1;
168 }
169 return 0;
170}
171
172static inline void amdgpu_vm_eviction_unlock(struct amdgpu_vm *vm)
173{
174#ifdef notyet
175 memalloc_nofs_restore(vm->saved_flags);
176#endif
177 mutex_unlock(&vm->eviction_lock)rw_exit_write(&vm->eviction_lock);
178}
179
180/**
181 * amdgpu_vm_level_shift - return the addr shift for each level
182 *
183 * @adev: amdgpu_device pointer
184 * @level: VMPT level
185 *
186 * Returns:
187 * The number of bits the pfn needs to be right shifted for a level.
188 */
189static unsigned amdgpu_vm_level_shift(struct amdgpu_device *adev,
190 unsigned level)
191{
192 switch (level) {
39
Control jumps to the 'default' case at line 200
193 case AMDGPU_VM_PDB2:
194 case AMDGPU_VM_PDB1:
195 case AMDGPU_VM_PDB0:
196 return 9 * (AMDGPU_VM_PDB0 - level) +
197 adev->vm_manager.block_size;
198 case AMDGPU_VM_PTB:
199 return 0;
200 default:
201 return ~0;
40
Returning the value 4294967295
202 }
203}
204
205/**
206 * amdgpu_vm_num_entries - return the number of entries in a PD/PT
207 *
208 * @adev: amdgpu_device pointer
209 * @level: VMPT level
210 *
211 * Returns:
212 * The number of entries in a page directory or page table.
213 */
214static unsigned amdgpu_vm_num_entries(struct amdgpu_device *adev,
215 unsigned level)
216{
217 unsigned shift = amdgpu_vm_level_shift(adev,
218 adev->vm_manager.root_level);
219
220 if (level == adev->vm_manager.root_level)
221 /* For the root directory */
222 return round_up(adev->vm_manager.max_pfn, 1ULL << shift)((((adev->vm_manager.max_pfn) + ((1ULL << shift) - 1
)) / (1ULL << shift)) * (1ULL << shift))
223 >> shift;
224 else if (level != AMDGPU_VM_PTB)
225 /* Everything in between */
226 return 512;
227 else
228 /* For the page tables on the leaves */
229 return AMDGPU_VM_PTE_COUNT(adev)(1 << (adev)->vm_manager.block_size);
230}
231
232/**
233 * amdgpu_vm_num_ats_entries - return the number of ATS entries in the root PD
234 *
235 * @adev: amdgpu_device pointer
236 *
237 * Returns:
238 * The number of entries in the root page directory which needs the ATS setting.
239 */
240static unsigned amdgpu_vm_num_ats_entries(struct amdgpu_device *adev)
241{
242 unsigned shift;
243
244 shift = amdgpu_vm_level_shift(adev, adev->vm_manager.root_level);
245 return AMDGPU_GMC_HOLE_START0x0000800000000000ULL >> (shift + AMDGPU_GPU_PAGE_SHIFT12);
246}
247
248/**
249 * amdgpu_vm_entries_mask - the mask to get the entry number of a PD/PT
250 *
251 * @adev: amdgpu_device pointer
252 * @level: VMPT level
253 *
254 * Returns:
255 * The mask to extract the entry number of a PD/PT from an address.
256 */
257static uint32_t amdgpu_vm_entries_mask(struct amdgpu_device *adev,
258 unsigned int level)
259{
260 if (level <= adev->vm_manager.root_level)
261 return 0xffffffff;
262 else if (level != AMDGPU_VM_PTB)
263 return 0x1ff;
264 else
265 return AMDGPU_VM_PTE_COUNT(adev)(1 << (adev)->vm_manager.block_size) - 1;
266}
267
268/**
269 * amdgpu_vm_bo_size - returns the size of the BOs in bytes
270 *
271 * @adev: amdgpu_device pointer
272 * @level: VMPT level
273 *
274 * Returns:
275 * The size of the BO for a page directory or page table in bytes.
276 */
277static unsigned amdgpu_vm_bo_size(struct amdgpu_device *adev, unsigned level)
278{
279 return AMDGPU_GPU_PAGE_ALIGN(amdgpu_vm_num_entries(adev, level) * 8)(((amdgpu_vm_num_entries(adev, level) * 8) + (4096 - 1)) &
~(4096 - 1))
;
280}
281
282/**
283 * amdgpu_vm_bo_evicted - vm_bo is evicted
284 *
285 * @vm_bo: vm_bo which is evicted
286 *
287 * State for PDs/PTs and per VM BOs which are not at the location they should
288 * be.
289 */
290static void amdgpu_vm_bo_evicted(struct amdgpu_vm_bo_base *vm_bo)
291{
292 struct amdgpu_vm *vm = vm_bo->vm;
293 struct amdgpu_bo *bo = vm_bo->bo;
294
295 vm_bo->moved = true1;
296 if (bo->tbo.type == ttm_bo_type_kernel)
297 list_move(&vm_bo->vm_status, &vm->evicted);
298 else
299 list_move_tail(&vm_bo->vm_status, &vm->evicted);
300}
301/**
302 * amdgpu_vm_bo_moved - vm_bo is moved
303 *
304 * @vm_bo: vm_bo which is moved
305 *
306 * State for per VM BOs which are moved, but that change is not yet reflected
307 * in the page tables.
308 */
309static void amdgpu_vm_bo_moved(struct amdgpu_vm_bo_base *vm_bo)
310{
311 list_move(&vm_bo->vm_status, &vm_bo->vm->moved);
312}
313
314/**
315 * amdgpu_vm_bo_idle - vm_bo is idle
316 *
317 * @vm_bo: vm_bo which is now idle
318 *
319 * State for PDs/PTs and per VM BOs which have gone through the state machine
320 * and are now idle.
321 */
322static void amdgpu_vm_bo_idle(struct amdgpu_vm_bo_base *vm_bo)
323{
324 list_move(&vm_bo->vm_status, &vm_bo->vm->idle);
325 vm_bo->moved = false0;
326}
327
328/**
329 * amdgpu_vm_bo_invalidated - vm_bo is invalidated
330 *
331 * @vm_bo: vm_bo which is now invalidated
332 *
333 * State for normal BOs which are invalidated and that change not yet reflected
334 * in the PTs.
335 */
336static void amdgpu_vm_bo_invalidated(struct amdgpu_vm_bo_base *vm_bo)
337{
338 spin_lock(&vm_bo->vm->invalidated_lock)mtx_enter(&vm_bo->vm->invalidated_lock);
339 list_move(&vm_bo->vm_status, &vm_bo->vm->invalidated);
340 spin_unlock(&vm_bo->vm->invalidated_lock)mtx_leave(&vm_bo->vm->invalidated_lock);
341}
342
343/**
344 * amdgpu_vm_bo_relocated - vm_bo is reloacted
345 *
346 * @vm_bo: vm_bo which is relocated
347 *
348 * State for PDs/PTs which needs to update their parent PD.
349 * For the root PD, just move to idle state.
350 */
351static void amdgpu_vm_bo_relocated(struct amdgpu_vm_bo_base *vm_bo)
352{
353 if (vm_bo->bo->parent)
354 list_move(&vm_bo->vm_status, &vm_bo->vm->relocated);
355 else
356 amdgpu_vm_bo_idle(vm_bo);
357}
358
359/**
360 * amdgpu_vm_bo_done - vm_bo is done
361 *
362 * @vm_bo: vm_bo which is now done
363 *
364 * State for normal BOs which are invalidated and that change has been updated
365 * in the PTs.
366 */
367static void amdgpu_vm_bo_done(struct amdgpu_vm_bo_base *vm_bo)
368{
369 spin_lock(&vm_bo->vm->invalidated_lock)mtx_enter(&vm_bo->vm->invalidated_lock);
370 list_del_init(&vm_bo->vm_status);
371 spin_unlock(&vm_bo->vm->invalidated_lock)mtx_leave(&vm_bo->vm->invalidated_lock);
372}
373
374/**
375 * amdgpu_vm_bo_base_init - Adds bo to the list of bos associated with the vm
376 *
377 * @base: base structure for tracking BO usage in a VM
378 * @vm: vm to which bo is to be added
379 * @bo: amdgpu buffer object
380 *
381 * Initialize a bo_va_base structure and add it to the appropriate lists
382 *
383 */
384static void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
385 struct amdgpu_vm *vm,
386 struct amdgpu_bo *bo)
387{
388 base->vm = vm;
389 base->bo = bo;
390 base->next = NULL((void *)0);
391 INIT_LIST_HEAD(&base->vm_status);
392
393 if (!bo)
394 return;
395 base->next = bo->vm_bo;
396 bo->vm_bo = base;
397
398 if (bo->tbo.base.resv != vm->root.base.bo->tbo.base.resv)
399 return;
400
401 vm->bulk_moveable = false0;
402 if (bo->tbo.type == ttm_bo_type_kernel && bo->parent)
403 amdgpu_vm_bo_relocated(base);
404 else
405 amdgpu_vm_bo_idle(base);
406
407 if (bo->preferred_domains &
408 amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type))
409 return;
410
411 /*
412 * we checked all the prerequisites, but it looks like this per vm bo
413 * is currently evicted. add the bo to the evicted list to make sure it
414 * is validated on next vm use to avoid fault.
415 * */
416 amdgpu_vm_bo_evicted(base);
417}
418
419/**
420 * amdgpu_vm_pt_parent - get the parent page directory
421 *
422 * @pt: child page table
423 *
424 * Helper to get the parent entry for the child page table. NULL if we are at
425 * the root page directory.
426 */
427static struct amdgpu_vm_pt *amdgpu_vm_pt_parent(struct amdgpu_vm_pt *pt)
428{
429 struct amdgpu_bo *parent = pt->base.bo->parent;
430
431 if (!parent)
432 return NULL((void *)0);
433
434 return container_of(parent->vm_bo, struct amdgpu_vm_pt, base)({ const __typeof( ((struct amdgpu_vm_pt *)0)->base ) *__mptr
= (parent->vm_bo); (struct amdgpu_vm_pt *)( (char *)__mptr
- __builtin_offsetof(struct amdgpu_vm_pt, base) );})
;
435}
436
437/*
438 * amdgpu_vm_pt_cursor - state for for_each_amdgpu_vm_pt
439 */
440struct amdgpu_vm_pt_cursor {
441 uint64_t pfn;
442 struct amdgpu_vm_pt *parent;
443 struct amdgpu_vm_pt *entry;
444 unsigned level;
445};
446
447/**
448 * amdgpu_vm_pt_start - start PD/PT walk
449 *
450 * @adev: amdgpu_device pointer
451 * @vm: amdgpu_vm structure
452 * @start: start address of the walk
453 * @cursor: state to initialize
454 *
455 * Initialize a amdgpu_vm_pt_cursor to start a walk.
456 */
457static void amdgpu_vm_pt_start(struct amdgpu_device *adev,
458 struct amdgpu_vm *vm, uint64_t start,
459 struct amdgpu_vm_pt_cursor *cursor)
460{
461 cursor->pfn = start;
462 cursor->parent = NULL((void *)0);
463 cursor->entry = &vm->root;
464 cursor->level = adev->vm_manager.root_level;
465}
31
Returning without writing to 'vm->root.entries', which participates in a condition later
32
Returning without writing to 'vm->root.base.bo', which participates in a condition later
466
467/**
468 * amdgpu_vm_pt_descendant - go to child node
469 *
470 * @adev: amdgpu_device pointer
471 * @cursor: current state
472 *
473 * Walk to the child node of the current node.
474 * Returns:
475 * True if the walk was possible, false otherwise.
476 */
477static bool_Bool amdgpu_vm_pt_descendant(struct amdgpu_device *adev,
478 struct amdgpu_vm_pt_cursor *cursor)
479{
480 unsigned mask, shift, idx;
481
482 if (!cursor->entry->entries)
45
Assuming field 'entries' is null, which participates in a condition later
46
Taking true branch
483 return false0;
47
Returning zero, which participates in a condition later
484
485 BUG_ON(!cursor->entry->base.bo)((!(!cursor->entry->base.bo)) ? (void)0 : __assert("diagnostic "
, "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_vm.c", 485, "!(!cursor->entry->base.bo)"
))
;
486 mask = amdgpu_vm_entries_mask(adev, cursor->level);
487 shift = amdgpu_vm_level_shift(adev, cursor->level);
488
489 ++cursor->level;
490 idx = (cursor->pfn >> shift) & mask;
491 cursor->parent = cursor->entry;
492 cursor->entry = &cursor->entry->entries[idx];
493 return true1;
494}
495
496/**
497 * amdgpu_vm_pt_sibling - go to sibling node
498 *
499 * @adev: amdgpu_device pointer
500 * @cursor: current state
501 *
502 * Walk to the sibling node of the current node.
503 * Returns:
504 * True if the walk was possible, false otherwise.
505 */
506static bool_Bool amdgpu_vm_pt_sibling(struct amdgpu_device *adev,
507 struct amdgpu_vm_pt_cursor *cursor)
508{
509 unsigned shift, num_entries;
510
511 /* Root doesn't have a sibling */
512 if (!cursor->parent)
513 return false0;
514
515 /* Go to our parents and see if we got a sibling */
516 shift = amdgpu_vm_level_shift(adev, cursor->level - 1);
517 num_entries = amdgpu_vm_num_entries(adev, cursor->level - 1);
518
519 if (cursor->entry == &cursor->parent->entries[num_entries - 1])
520 return false0;
521
522 cursor->pfn += 1ULL << shift;
523 cursor->pfn &= ~((1ULL << shift) - 1);
524 ++cursor->entry;
525 return true1;
526}
527
528/**
529 * amdgpu_vm_pt_ancestor - go to parent node
530 *
531 * @cursor: current state
532 *
533 * Walk to the parent node of the current node.
534 * Returns:
535 * True if the walk was possible, false otherwise.
536 */
537static bool_Bool amdgpu_vm_pt_ancestor(struct amdgpu_vm_pt_cursor *cursor)
538{
539 if (!cursor->parent)
540 return false0;
541
542 --cursor->level;
543 cursor->entry = cursor->parent;
544 cursor->parent = amdgpu_vm_pt_parent(cursor->parent);
545 return true1;
546}
547
548/**
549 * amdgpu_vm_pt_next - get next PD/PT in hieratchy
550 *
551 * @adev: amdgpu_device pointer
552 * @cursor: current state
553 *
554 * Walk the PD/PT tree to the next node.
555 */
556static void amdgpu_vm_pt_next(struct amdgpu_device *adev,
557 struct amdgpu_vm_pt_cursor *cursor)
558{
559 /* First try a newborn child */
560 if (amdgpu_vm_pt_descendant(adev, cursor))
561 return;
562
563 /* If that didn't worked try to find a sibling */
564 while (!amdgpu_vm_pt_sibling(adev, cursor)) {
565 /* No sibling, go to our parents and grandparents */
566 if (!amdgpu_vm_pt_ancestor(cursor)) {
567 cursor->pfn = ~0ll;
568 return;
569 }
570 }
571}
572
573/**
574 * amdgpu_vm_pt_first_dfs - start a deep first search
575 *
576 * @adev: amdgpu_device structure
577 * @vm: amdgpu_vm structure
578 * @start: optional cursor to start with
579 * @cursor: state to initialize
580 *
581 * Starts a deep first traversal of the PD/PT tree.
582 */
583static void amdgpu_vm_pt_first_dfs(struct amdgpu_device *adev,
584 struct amdgpu_vm *vm,
585 struct amdgpu_vm_pt_cursor *start,
586 struct amdgpu_vm_pt_cursor *cursor)
587{
588 if (start)
589 *cursor = *start;
590 else
591 amdgpu_vm_pt_start(adev, vm, 0, cursor);
592 while (amdgpu_vm_pt_descendant(adev, cursor));
593}
594
595/**
596 * amdgpu_vm_pt_continue_dfs - check if the deep first search should continue
597 *
598 * @start: starting point for the search
599 * @entry: current entry
600 *
601 * Returns:
602 * True when the search should continue, false otherwise.
603 */
604static bool_Bool amdgpu_vm_pt_continue_dfs(struct amdgpu_vm_pt_cursor *start,
605 struct amdgpu_vm_pt *entry)
606{
607 return entry && (!start || entry != start->entry);
608}
609
610/**
611 * amdgpu_vm_pt_next_dfs - get the next node for a deep first search
612 *
613 * @adev: amdgpu_device structure
614 * @cursor: current state
615 *
616 * Move the cursor to the next node in a deep first search.
617 */
618static void amdgpu_vm_pt_next_dfs(struct amdgpu_device *adev,
619 struct amdgpu_vm_pt_cursor *cursor)
620{
621 if (!cursor->entry)
622 return;
623
624 if (!cursor->parent)
625 cursor->entry = NULL((void *)0);
626 else if (amdgpu_vm_pt_sibling(adev, cursor))
627 while (amdgpu_vm_pt_descendant(adev, cursor));
628 else
629 amdgpu_vm_pt_ancestor(cursor);
630}
631
632/*
633 * for_each_amdgpu_vm_pt_dfs_safe - safe deep first search of all PDs/PTs
634 */
635#define for_each_amdgpu_vm_pt_dfs_safe(adev, vm, start, cursor, entry)for (amdgpu_vm_pt_first_dfs((adev), (vm), (start), &(cursor
)), (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &
(cursor)); amdgpu_vm_pt_continue_dfs((start), (entry)); (entry
) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &(cursor
)))
\
636 for (amdgpu_vm_pt_first_dfs((adev), (vm), (start), &(cursor)), \
637 (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &(cursor));\
638 amdgpu_vm_pt_continue_dfs((start), (entry)); \
639 (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &(cursor)))
640
641/**
642 * amdgpu_vm_get_pd_bo - add the VM PD to a validation list
643 *
644 * @vm: vm providing the BOs
645 * @validated: head of validation list
646 * @entry: entry to add
647 *
648 * Add the page directory to the list of BOs to
649 * validate for command submission.
650 */
651void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
652 struct list_head *validated,
653 struct amdgpu_bo_list_entry *entry)
654{
655 entry->priority = 0;
656 entry->tv.bo = &vm->root.base.bo->tbo;
657 /* Two for VM updates, one for TTM and one for the CS job */
658 entry->tv.num_shared = 4;
659 entry->user_pages = NULL((void *)0);
660 list_add(&entry->tv.head, validated);
661}
662
663/**
664 * amdgpu_vm_del_from_lru_notify - update bulk_moveable flag
665 *
666 * @bo: BO which was removed from the LRU
667 *
668 * Make sure the bulk_moveable flag is updated when a BO is removed from the
669 * LRU.
670 */
671void amdgpu_vm_del_from_lru_notify(struct ttm_buffer_object *bo)
672{
673 struct amdgpu_bo *abo;
674 struct amdgpu_vm_bo_base *bo_base;
675
676 if (!amdgpu_bo_is_amdgpu_bo(bo))
677 return;
678
679 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT(1 << 21))
680 return;
681
682 abo = ttm_to_amdgpu_bo(bo);
683 if (!abo->parent)
684 return;
685 for (bo_base = abo->vm_bo; bo_base; bo_base = bo_base->next) {
686 struct amdgpu_vm *vm = bo_base->vm;
687
688 if (abo->tbo.base.resv == vm->root.base.bo->tbo.base.resv)
689 vm->bulk_moveable = false0;
690 }
691
692}
693/**
694 * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
695 *
696 * @adev: amdgpu device pointer
697 * @vm: vm providing the BOs
698 *
699 * Move all BOs to the end of LRU and remember their positions to put them
700 * together.
701 */
702void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
703 struct amdgpu_vm *vm)
704{
705 struct amdgpu_vm_bo_base *bo_base;
706
707 if (vm->bulk_moveable) {
708 spin_lock(&ttm_bo_glob.lru_lock)mtx_enter(&ttm_bo_glob.lru_lock);
709 ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move);
710 spin_unlock(&ttm_bo_glob.lru_lock)mtx_leave(&ttm_bo_glob.lru_lock);
711 return;
712 }
713
714 memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move))__builtin_memset((&vm->lru_bulk_move), (0), (sizeof(vm
->lru_bulk_move)))
;
715
716 spin_lock(&ttm_bo_glob.lru_lock)mtx_enter(&ttm_bo_glob.lru_lock);
717 list_for_each_entry(bo_base, &vm->idle, vm_status)for (bo_base = ({ const __typeof( ((__typeof(*bo_base) *)0)->
vm_status ) *__mptr = ((&vm->idle)->next); (__typeof
(*bo_base) *)( (char *)__mptr - __builtin_offsetof(__typeof(*
bo_base), vm_status) );}); &bo_base->vm_status != (&
vm->idle); bo_base = ({ const __typeof( ((__typeof(*bo_base
) *)0)->vm_status ) *__mptr = (bo_base->vm_status.next)
; (__typeof(*bo_base) *)( (char *)__mptr - __builtin_offsetof
(__typeof(*bo_base), vm_status) );}))
{
718 struct amdgpu_bo *bo = bo_base->bo;
719
720 if (!bo->parent)
721 continue;
722
723 ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
724 if (bo->shadow)
725 ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
726 &vm->lru_bulk_move);
727 }
728 spin_unlock(&ttm_bo_glob.lru_lock)mtx_leave(&ttm_bo_glob.lru_lock);
729
730 vm->bulk_moveable = true1;
731}
732
733/**
734 * amdgpu_vm_validate_pt_bos - validate the page table BOs
735 *
736 * @adev: amdgpu device pointer
737 * @vm: vm providing the BOs
738 * @validate: callback to do the validation
739 * @param: parameter for the validation callback
740 *
741 * Validate the page table BOs on command submission if neccessary.
742 *
743 * Returns:
744 * Validation result.
745 */
746int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
747 int (*validate)(void *p, struct amdgpu_bo *bo),
748 void *param)
749{
750 struct amdgpu_vm_bo_base *bo_base, *tmp;
751 int r;
752
753 vm->bulk_moveable &= list_empty(&vm->evicted);
754
755 list_for_each_entry_safe(bo_base, tmp, &vm->evicted, vm_status)for (bo_base = ({ const __typeof( ((__typeof(*bo_base) *)0)->
vm_status ) *__mptr = ((&vm->evicted)->next); (__typeof
(*bo_base) *)( (char *)__mptr - __builtin_offsetof(__typeof(*
bo_base), vm_status) );}), tmp = ({ const __typeof( ((__typeof
(*bo_base) *)0)->vm_status ) *__mptr = (bo_base->vm_status
.next); (__typeof(*bo_base) *)( (char *)__mptr - __builtin_offsetof
(__typeof(*bo_base), vm_status) );}); &bo_base->vm_status
!= (&vm->evicted); bo_base = tmp, tmp = ({ const __typeof
( ((__typeof(*tmp) *)0)->vm_status ) *__mptr = (tmp->vm_status
.next); (__typeof(*tmp) *)( (char *)__mptr - __builtin_offsetof
(__typeof(*tmp), vm_status) );}))
{
756 struct amdgpu_bo *bo = bo_base->bo;
757
758 r = validate(param, bo);
759 if (r)
760 return r;
761
762 if (bo->tbo.type != ttm_bo_type_kernel) {
763 amdgpu_vm_bo_moved(bo_base);
764 } else {
765 vm->update_funcs->map_table(bo);
766 amdgpu_vm_bo_relocated(bo_base);
767 }
768 }
769
770 amdgpu_vm_eviction_lock(vm);
771 vm->evicting = false0;
772 amdgpu_vm_eviction_unlock(vm);
773
774 return 0;
775}
776
777/**
778 * amdgpu_vm_ready - check VM is ready for updates
779 *
780 * @vm: VM to check
781 *
782 * Check if all VM PDs/PTs are ready for updates
783 *
784 * Returns:
785 * True if eviction list is empty.
786 */
787bool_Bool amdgpu_vm_ready(struct amdgpu_vm *vm)
788{
789 return list_empty(&vm->evicted);
790}
791
792/**
793 * amdgpu_vm_clear_bo - initially clear the PDs/PTs
794 *
795 * @adev: amdgpu_device pointer
796 * @vm: VM to clear BO from
797 * @bo: BO to clear
798 * @immediate: use an immediate update
799 *
800 * Root PD needs to be reserved when calling this.
801 *
802 * Returns:
803 * 0 on success, errno otherwise.
804 */
805static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
806 struct amdgpu_vm *vm,
807 struct amdgpu_bo *bo,
808 bool_Bool immediate)
809{
810 struct ttm_operation_ctx ctx = { true1, false0 };
811 unsigned level = adev->vm_manager.root_level;
812 struct amdgpu_vm_update_params params;
813 struct amdgpu_bo *ancestor = bo;
814 unsigned entries, ats_entries;
815 uint64_t addr;
816 int r;
817
818 /* Figure out our place in the hierarchy */
819 if (ancestor->parent) {
820 ++level;
821 while (ancestor->parent->parent) {
822 ++level;
823 ancestor = ancestor->parent;
824 }
825 }
826
827 entries = amdgpu_bo_size(bo) / 8;
828 if (!vm->pte_support_ats) {
829 ats_entries = 0;
830
831 } else if (!bo->parent) {
832 ats_entries = amdgpu_vm_num_ats_entries(adev);
833 ats_entries = min(ats_entries, entries)(((ats_entries)<(entries))?(ats_entries):(entries));
834 entries -= ats_entries;
835
836 } else {
837 struct amdgpu_vm_pt *pt;
838
839 pt = container_of(ancestor->vm_bo, struct amdgpu_vm_pt, base)({ const __typeof( ((struct amdgpu_vm_pt *)0)->base ) *__mptr
= (ancestor->vm_bo); (struct amdgpu_vm_pt *)( (char *)__mptr
- __builtin_offsetof(struct amdgpu_vm_pt, base) );})
;
840 ats_entries = amdgpu_vm_num_ats_entries(adev);
841 if ((pt - vm->root.entries) >= ats_entries) {
842 ats_entries = 0;
843 } else {
844 ats_entries = entries;
845 entries = 0;
846 }
847 }
848
849 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
850 if (r)
851 return r;
852
853 if (bo->shadow) {
854 r = ttm_bo_validate(&bo->shadow->tbo, &bo->shadow->placement,
855 &ctx);
856 if (r)
857 return r;
858 }
859
860 r = vm->update_funcs->map_table(bo);
861 if (r)
862 return r;
863
864 memset(&params, 0, sizeof(params))__builtin_memset((&params), (0), (sizeof(params)));
865 params.adev = adev;
866 params.vm = vm;
867 params.immediate = immediate;
868
869 r = vm->update_funcs->prepare(&params, NULL((void *)0), AMDGPU_SYNC_EXPLICIT);
870 if (r)
871 return r;
872
873 addr = 0;
874 if (ats_entries) {
875 uint64_t value = 0, flags;
876
877 flags = AMDGPU_PTE_DEFAULT_ATC((1ULL << 1) | (1ULL << 2) | (1ULL << 4) | (
1ULL << 5) | (1ULL << 6) | ((uint64_t)(2) <<
57))
;
878 if (level != AMDGPU_VM_PTB) {
879 /* Handle leaf PDEs as PTEs */
880 flags |= AMDGPU_PDE_PTE(1ULL << 54);
881 amdgpu_gmc_get_vm_pde(adev, level, &value, &flags)(adev)->gmc.gmc_funcs->get_vm_pde((adev), (level), (&
value), (&flags))
;
882 }
883
884 r = vm->update_funcs->update(&params, bo, addr, 0, ats_entries,
885 value, flags);
886 if (r)
887 return r;
888
889 addr += ats_entries * 8;
890 }
891
892 if (entries) {
893 uint64_t value = 0, flags = 0;
894
895 if (adev->asic_type >= CHIP_VEGA10) {
896 if (level != AMDGPU_VM_PTB) {
897 /* Handle leaf PDEs as PTEs */
898 flags |= AMDGPU_PDE_PTE(1ULL << 54);
899 amdgpu_gmc_get_vm_pde(adev, level,(adev)->gmc.gmc_funcs->get_vm_pde((adev), (level), (&
value), (&flags))
900 &value, &flags)(adev)->gmc.gmc_funcs->get_vm_pde((adev), (level), (&
value), (&flags))
;
901 } else {
902 /* Workaround for fault priority problem on GMC9 */
903 flags = AMDGPU_PTE_EXECUTABLE(1ULL << 4);
904 }
905 }
906
907 r = vm->update_funcs->update(&params, bo, addr, 0, entries,
908 value, flags);
909 if (r)
910 return r;
911 }
912
913 return vm->update_funcs->commit(&params, NULL((void *)0));
914}
915
916/**
917 * amdgpu_vm_bo_param - fill in parameters for PD/PT allocation
918 *
919 * @adev: amdgpu_device pointer
920 * @vm: requesting vm
921 * @level: the page table level
922 * @immediate: use a immediate update
923 * @bp: resulting BO allocation parameters
924 */
925static void amdgpu_vm_bo_param(struct amdgpu_device *adev, struct amdgpu_vm *vm,
926 int level, bool_Bool immediate,
927 struct amdgpu_bo_param *bp)
928{
929 memset(bp, 0, sizeof(*bp))__builtin_memset((bp), (0), (sizeof(*bp)));
930
931 bp->size = amdgpu_vm_bo_size(adev, level);
932 bp->byte_align = AMDGPU_GPU_PAGE_SIZE4096;
933 bp->domain = AMDGPU_GEM_DOMAIN_VRAM0x4;
934 bp->domain = amdgpu_bo_get_preferred_pin_domain(adev, bp->domain);
935 bp->flags = AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS(1 << 5) |
936 AMDGPU_GEM_CREATE_CPU_GTT_USWC(1 << 2);
937 if (vm->use_cpu_for_update)
938 bp->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED(1 << 0);
939 else if (!vm->root.base.bo || vm->root.base.bo->shadow)
940 bp->flags |= AMDGPU_GEM_CREATE_SHADOW(1 << 4);
941 bp->type = ttm_bo_type_kernel;
942 bp->no_wait_gpu = immediate;
943 if (vm->root.base.bo)
944 bp->resv = vm->root.base.bo->tbo.base.resv;
945}
946
947/**
948 * amdgpu_vm_alloc_pts - Allocate a specific page table
949 *
950 * @adev: amdgpu_device pointer
951 * @vm: VM to allocate page tables for
952 * @cursor: Which page table to allocate
953 * @immediate: use an immediate update
954 *
955 * Make sure a specific page table or directory is allocated.
956 *
957 * Returns:
958 * 1 if page table needed to be allocated, 0 if page table was already
959 * allocated, negative errno if an error occurred.
960 */
961static int amdgpu_vm_alloc_pts(struct amdgpu_device *adev,
962 struct amdgpu_vm *vm,
963 struct amdgpu_vm_pt_cursor *cursor,
964 bool_Bool immediate)
965{
966 struct amdgpu_vm_pt *entry = cursor->entry;
967 struct amdgpu_bo_param bp;
968 struct amdgpu_bo *pt;
969 int r;
970
971 if (cursor->level < AMDGPU_VM_PTB && !entry->entries) {
972 unsigned num_entries;
973
974 num_entries = amdgpu_vm_num_entries(adev, cursor->level);
975 entry->entries = kvmalloc_array(num_entries,
976 sizeof(*entry->entries),
977 GFP_KERNEL(0x0001 | 0x0004) | __GFP_ZERO0x0008);
978 if (!entry->entries)
979 return -ENOMEM12;
980 }
981
982 if (entry->base.bo)
983 return 0;
984
985 amdgpu_vm_bo_param(adev, vm, cursor->level, immediate, &bp);
986
987 r = amdgpu_bo_create(adev, &bp, &pt);
988 if (r)
989 return r;
990
991 /* Keep a reference to the root directory to avoid
992 * freeing them up in the wrong order.
993 */
994 pt->parent = amdgpu_bo_ref(cursor->parent->base.bo);
995 amdgpu_vm_bo_base_init(&entry->base, vm, pt);
996
997 r = amdgpu_vm_clear_bo(adev, vm, pt, immediate);
998 if (r)
999 goto error_free_pt;
1000
1001 return 0;
1002
1003error_free_pt:
1004 amdgpu_bo_unref(&pt->shadow);
1005 amdgpu_bo_unref(&pt);
1006 return r;
1007}
1008
1009/**
1010 * amdgpu_vm_free_table - fre one PD/PT
1011 *
1012 * @entry: PDE to free
1013 */
1014static void amdgpu_vm_free_table(struct amdgpu_vm_pt *entry)
1015{
1016 if (entry->base.bo) {
1017 entry->base.bo->vm_bo = NULL((void *)0);
1018 list_del(&entry->base.vm_status);
1019 amdgpu_bo_unref(&entry->base.bo->shadow);
1020 amdgpu_bo_unref(&entry->base.bo);
1021 }
1022 kvfree(entry->entries);
1023 entry->entries = NULL((void *)0);
1024}
1025
1026/**
1027 * amdgpu_vm_free_pts - free PD/PT levels
1028 *
1029 * @adev: amdgpu device structure
1030 * @vm: amdgpu vm structure
1031 * @start: optional cursor where to start freeing PDs/PTs
1032 *
1033 * Free the page directory or page table level and all sub levels.
1034 */
1035static void amdgpu_vm_free_pts(struct amdgpu_device *adev,
1036 struct amdgpu_vm *vm,
1037 struct amdgpu_vm_pt_cursor *start)
1038{
1039 struct amdgpu_vm_pt_cursor cursor;
1040 struct amdgpu_vm_pt *entry;
1041
1042 vm->bulk_moveable = false0;
1043
1044 for_each_amdgpu_vm_pt_dfs_safe(adev, vm, start, cursor, entry)for (amdgpu_vm_pt_first_dfs((adev), (vm), (start), &(cursor
)), (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &
(cursor)); amdgpu_vm_pt_continue_dfs((start), (entry)); (entry
) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &(cursor
)))
1045 amdgpu_vm_free_table(entry);
1046
1047 if (start)
1048 amdgpu_vm_free_table(start->entry);
1049}
1050
1051/**
1052 * amdgpu_vm_check_compute_bug - check whether asic has compute vm bug
1053 *
1054 * @adev: amdgpu_device pointer
1055 */
1056void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev)
1057{
1058 const struct amdgpu_ip_block *ip_block;
1059 bool_Bool has_compute_vm_bug;
1060 struct amdgpu_ring *ring;
1061 int i;
1062
1063 has_compute_vm_bug = false0;
1064
1065 ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_GFX);
1066 if (ip_block) {
1067 /* Compute has a VM bug for GFX version < 7.
1068 Compute has a VM bug for GFX 8 MEC firmware version < 673.*/
1069 if (ip_block->version->major <= 7)
1070 has_compute_vm_bug = true1;
1071 else if (ip_block->version->major == 8)
1072 if (adev->gfx.mec_fw_version < 673)
1073 has_compute_vm_bug = true1;
1074 }
1075
1076 for (i = 0; i < adev->num_rings; i++) {
1077 ring = adev->rings[i];
1078 if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
1079 /* only compute rings */
1080 ring->has_compute_vm_bug = has_compute_vm_bug;
1081 else
1082 ring->has_compute_vm_bug = false0;
1083 }
1084}
1085
1086/**
1087 * amdgpu_vm_need_pipeline_sync - Check if pipe sync is needed for job.
1088 *
1089 * @ring: ring on which the job will be submitted
1090 * @job: job to submit
1091 *
1092 * Returns:
1093 * True if sync is needed.
1094 */
1095bool_Bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
1096 struct amdgpu_job *job)
1097{
1098 struct amdgpu_device *adev = ring->adev;
1099 unsigned vmhub = ring->funcs->vmhub;
1100 struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
1101 struct amdgpu_vmid *id;
1102 bool_Bool gds_switch_needed;
1103 bool_Bool vm_flush_needed = job->vm_needs_flush || ring->has_compute_vm_bug;
1104
1105 if (job->vmid == 0)
1106 return false0;
1107 id = &id_mgr->ids[job->vmid];
1108 gds_switch_needed = ring->funcs->emit_gds_switch && (
1109 id->gds_base != job->gds_base ||
1110 id->gds_size != job->gds_size ||
1111 id->gws_base != job->gws_base ||
1112 id->gws_size != job->gws_size ||
1113 id->oa_base != job->oa_base ||
1114 id->oa_size != job->oa_size);
1115
1116 if (amdgpu_vmid_had_gpu_reset(adev, id))
1117 return true1;
1118
1119 return vm_flush_needed || gds_switch_needed;
1120}
1121
1122/**
1123 * amdgpu_vm_flush - hardware flush the vm
1124 *
1125 * @ring: ring to use for flush
1126 * @job: related job
1127 * @need_pipe_sync: is pipe sync needed
1128 *
1129 * Emit a VM flush when it is necessary.
1130 *
1131 * Returns:
1132 * 0 on success, errno otherwise.
1133 */
1134int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job,
1135 bool_Bool need_pipe_sync)
1136{
1137 struct amdgpu_device *adev = ring->adev;
1138 unsigned vmhub = ring->funcs->vmhub;
1139 struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
1140 struct amdgpu_vmid *id = &id_mgr->ids[job->vmid];
1141 bool_Bool gds_switch_needed = ring->funcs->emit_gds_switch && (
1142 id->gds_base != job->gds_base ||
1143 id->gds_size != job->gds_size ||
1144 id->gws_base != job->gws_base ||
1145 id->gws_size != job->gws_size ||
1146 id->oa_base != job->oa_base ||
1147 id->oa_size != job->oa_size);
1148 bool_Bool vm_flush_needed = job->vm_needs_flush;
1149 struct dma_fence *fence = NULL((void *)0);
1150 bool_Bool pasid_mapping_needed = false0;
1151 unsigned patch_offset = 0;
1152 bool_Bool update_spm_vmid_needed = (job->vm && (job->vm->reserved_vmid[vmhub] != NULL((void *)0)));
1153 int r;
1154
1155 if (update_spm_vmid_needed && adev->gfx.rlc.funcs->update_spm_vmid)
1156 adev->gfx.rlc.funcs->update_spm_vmid(adev, job->vmid);
1157
1158 if (amdgpu_vmid_had_gpu_reset(adev, id)) {
1159 gds_switch_needed = true1;
1160 vm_flush_needed = true1;
1161 pasid_mapping_needed = true1;
1162 }
1163
1164 mutex_lock(&id_mgr->lock)rw_enter_write(&id_mgr->lock);
1165 if (id->pasid != job->pasid || !id->pasid_mapping ||
1166 !dma_fence_is_signaled(id->pasid_mapping))
1167 pasid_mapping_needed = true1;
1168 mutex_unlock(&id_mgr->lock)rw_exit_write(&id_mgr->lock);
1169
1170 gds_switch_needed &= !!ring->funcs->emit_gds_switch;
1171 vm_flush_needed &= !!ring->funcs->emit_vm_flush &&
1172 job->vm_pd_addr != AMDGPU_BO_INVALID_OFFSET0x7fffffffffffffffL;
1173 pasid_mapping_needed &= adev->gmc.gmc_funcs->emit_pasid_mapping &&
1174 ring->funcs->emit_wreg;
1175
1176 if (!vm_flush_needed && !gds_switch_needed && !need_pipe_sync)
1177 return 0;
1178
1179 if (ring->funcs->init_cond_exec)
1180 patch_offset = amdgpu_ring_init_cond_exec(ring)(ring)->funcs->init_cond_exec((ring));
1181
1182 if (need_pipe_sync)
1183 amdgpu_ring_emit_pipeline_sync(ring)(ring)->funcs->emit_pipeline_sync((ring));
1184
1185 if (vm_flush_needed) {
1186 trace_amdgpu_vm_flush(ring, job->vmid, job->vm_pd_addr);
1187 amdgpu_ring_emit_vm_flush(ring, job->vmid, job->vm_pd_addr)(ring)->funcs->emit_vm_flush((ring), (job->vmid), (job
->vm_pd_addr))
;
1188 }
1189
1190 if (pasid_mapping_needed)
1191 amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid)(ring)->adev->gmc.gmc_funcs->emit_pasid_mapping((ring
), (job->vmid), (job->pasid))
;
1192
1193 if (vm_flush_needed || pasid_mapping_needed) {
1194 r = amdgpu_fence_emit(ring, &fence, 0);
1195 if (r)
1196 return r;
1197 }
1198
1199 if (vm_flush_needed) {
1200 mutex_lock(&id_mgr->lock)rw_enter_write(&id_mgr->lock);
1201 dma_fence_put(id->last_flush);
1202 id->last_flush = dma_fence_get(fence);
1203 id->current_gpu_reset_count =
1204 atomic_read(&adev->gpu_reset_counter)({ typeof(*(&adev->gpu_reset_counter)) __tmp = *(volatile
typeof(*(&adev->gpu_reset_counter)) *)&(*(&adev
->gpu_reset_counter)); membar_datadep_consumer(); __tmp; }
)
;
1205 mutex_unlock(&id_mgr->lock)rw_exit_write(&id_mgr->lock);
1206 }
1207
1208 if (pasid_mapping_needed) {
1209 mutex_lock(&id_mgr->lock)rw_enter_write(&id_mgr->lock);
1210 id->pasid = job->pasid;
1211 dma_fence_put(id->pasid_mapping);
1212 id->pasid_mapping = dma_fence_get(fence);
1213 mutex_unlock(&id_mgr->lock)rw_exit_write(&id_mgr->lock);
1214 }
1215 dma_fence_put(fence);
1216
1217 if (ring->funcs->emit_gds_switch && gds_switch_needed) {
1218 id->gds_base = job->gds_base;
1219 id->gds_size = job->gds_size;
1220 id->gws_base = job->gws_base;
1221 id->gws_size = job->gws_size;
1222 id->oa_base = job->oa_base;
1223 id->oa_size = job->oa_size;
1224 amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base,(ring)->funcs->emit_gds_switch((ring), (job->vmid), (
job->gds_base), (job->gds_size), (job->gws_base), (job
->gws_size), (job->oa_base), (job->oa_size))
1225 job->gds_size, job->gws_base,(ring)->funcs->emit_gds_switch((ring), (job->vmid), (
job->gds_base), (job->gds_size), (job->gws_base), (job
->gws_size), (job->oa_base), (job->oa_size))
1226 job->gws_size, job->oa_base,(ring)->funcs->emit_gds_switch((ring), (job->vmid), (
job->gds_base), (job->gds_size), (job->gws_base), (job
->gws_size), (job->oa_base), (job->oa_size))
1227 job->oa_size)(ring)->funcs->emit_gds_switch((ring), (job->vmid), (
job->gds_base), (job->gds_size), (job->gws_base), (job
->gws_size), (job->oa_base), (job->oa_size))
;
1228 }
1229
1230 if (ring->funcs->patch_cond_exec)
1231 amdgpu_ring_patch_cond_exec(ring, patch_offset)(ring)->funcs->patch_cond_exec((ring),(patch_offset));
1232
1233 /* the double SWITCH_BUFFER here *cannot* be skipped by COND_EXEC */
1234 if (ring->funcs->emit_switch_buffer) {
1235 amdgpu_ring_emit_switch_buffer(ring)(ring)->funcs->emit_switch_buffer((ring));
1236 amdgpu_ring_emit_switch_buffer(ring)(ring)->funcs->emit_switch_buffer((ring));
1237 }
1238 return 0;
1239}
1240
1241/**
1242 * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
1243 *
1244 * @vm: requested vm
1245 * @bo: requested buffer object
1246 *
1247 * Find @bo inside the requested vm.
1248 * Search inside the @bos vm list for the requested vm
1249 * Returns the found bo_va or NULL if none is found
1250 *
1251 * Object has to be reserved!
1252 *
1253 * Returns:
1254 * Found bo_va or NULL.
1255 */
1256struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
1257 struct amdgpu_bo *bo)
1258{
1259 struct amdgpu_vm_bo_base *base;
1260
1261 for (base = bo->vm_bo; base; base = base->next) {
1262 if (base->vm != vm)
1263 continue;
1264
1265 return container_of(base, struct amdgpu_bo_va, base)({ const __typeof( ((struct amdgpu_bo_va *)0)->base ) *__mptr
= (base); (struct amdgpu_bo_va *)( (char *)__mptr - __builtin_offsetof
(struct amdgpu_bo_va, base) );})
;
1266 }
1267 return NULL((void *)0);
1268}
1269
1270/**
1271 * amdgpu_vm_map_gart - Resolve gart mapping of addr
1272 *
1273 * @pages_addr: optional DMA address to use for lookup
1274 * @addr: the unmapped addr
1275 *
1276 * Look up the physical address of the page that the pte resolves
1277 * to.
1278 *
1279 * Returns:
1280 * The pointer for the page table entry.
1281 */
1282uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr)
1283{
1284 uint64_t result;
1285
1286 /* page table offset */
1287 result = pages_addr[addr >> PAGE_SHIFT12];
1288
1289 /* in case cpu page size != gpu page size*/
1290 result |= addr & (~LINUX_PAGE_MASK(~((1 << 12) - 1)));
1291
1292 result &= 0xFFFFFFFFFFFFF000ULL;
1293
1294 return result;
1295}
1296
1297/**
1298 * amdgpu_vm_update_pde - update a single level in the hierarchy
1299 *
1300 * @params: parameters for the update
1301 * @vm: requested vm
1302 * @entry: entry to update
1303 *
1304 * Makes sure the requested entry in parent is up to date.
1305 */
1306static int amdgpu_vm_update_pde(struct amdgpu_vm_update_params *params,
1307 struct amdgpu_vm *vm,
1308 struct amdgpu_vm_pt *entry)
1309{
1310 struct amdgpu_vm_pt *parent = amdgpu_vm_pt_parent(entry);
1311 struct amdgpu_bo *bo = parent->base.bo, *pbo;
1312 uint64_t pde, pt, flags;
1313 unsigned level;
1314
1315 for (level = 0, pbo = bo->parent; pbo; ++level)
1316 pbo = pbo->parent;
1317
1318 level += params->adev->vm_manager.root_level;
1319 amdgpu_gmc_get_pde_for_bo(entry->base.bo, level, &pt, &flags);
1320 pde = (entry - parent->entries) * 8;
1321 return vm->update_funcs->update(params, bo, pde, pt, 1, 0, flags);
1322}
1323
1324/**
1325 * amdgpu_vm_invalidate_pds - mark all PDs as invalid
1326 *
1327 * @adev: amdgpu_device pointer
1328 * @vm: related vm
1329 *
1330 * Mark all PD level as invalid after an error.
1331 */
1332static void amdgpu_vm_invalidate_pds(struct amdgpu_device *adev,
1333 struct amdgpu_vm *vm)
1334{
1335 struct amdgpu_vm_pt_cursor cursor;
1336 struct amdgpu_vm_pt *entry;
1337
1338 for_each_amdgpu_vm_pt_dfs_safe(adev, vm, NULL, cursor, entry)for (amdgpu_vm_pt_first_dfs((adev), (vm), (((void *)0)), &
(cursor)), (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev
), &(cursor)); amdgpu_vm_pt_continue_dfs((((void *)0)), (
entry)); (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev
), &(cursor)))
1339 if (entry->base.bo && !entry->base.moved)
1340 amdgpu_vm_bo_relocated(&entry->base);
1341}
1342
1343/**
1344 * amdgpu_vm_update_pdes - make sure that all directories are valid
1345 *
1346 * @adev: amdgpu_device pointer
1347 * @vm: requested vm
1348 * @immediate: submit immediately to the paging queue
1349 *
1350 * Makes sure all directories are up to date.
1351 *
1352 * Returns:
1353 * 0 for success, error for failure.
1354 */
1355int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
1356 struct amdgpu_vm *vm, bool_Bool immediate)
1357{
1358 struct amdgpu_vm_update_params params;
1359 int r;
1360
1361 if (list_empty(&vm->relocated))
1362 return 0;
1363
1364 memset(&params, 0, sizeof(params))__builtin_memset((&params), (0), (sizeof(params)));
1365 params.adev = adev;
1366 params.vm = vm;
1367 params.immediate = immediate;
1368
1369 r = vm->update_funcs->prepare(&params, NULL((void *)0), AMDGPU_SYNC_EXPLICIT);
1370 if (r)
1371 return r;
1372
1373 while (!list_empty(&vm->relocated)) {
1374 struct amdgpu_vm_pt *entry;
1375
1376 entry = list_first_entry(&vm->relocated, struct amdgpu_vm_pt,({ const __typeof( ((struct amdgpu_vm_pt *)0)->base.vm_status
) *__mptr = ((&vm->relocated)->next); (struct amdgpu_vm_pt
*)( (char *)__mptr - __builtin_offsetof(struct amdgpu_vm_pt,
base.vm_status) );})
1377 base.vm_status)({ const __typeof( ((struct amdgpu_vm_pt *)0)->base.vm_status
) *__mptr = ((&vm->relocated)->next); (struct amdgpu_vm_pt
*)( (char *)__mptr - __builtin_offsetof(struct amdgpu_vm_pt,
base.vm_status) );})
;
1378 amdgpu_vm_bo_idle(&entry->base);
1379
1380 r = amdgpu_vm_update_pde(&params, vm, entry);
1381 if (r)
1382 goto error;
1383 }
1384
1385 r = vm->update_funcs->commit(&params, &vm->last_update);
1386 if (r)
1387 goto error;
1388 return 0;
1389
1390error:
1391 amdgpu_vm_invalidate_pds(adev, vm);
1392 return r;
1393}
1394
1395/*
1396 * amdgpu_vm_update_flags - figure out flags for PTE updates
1397 *
1398 * Make sure to set the right flags for the PTEs at the desired level.
1399 */
1400static void amdgpu_vm_update_flags(struct amdgpu_vm_update_params *params,
1401 struct amdgpu_bo *bo, unsigned level,
1402 uint64_t pe, uint64_t addr,
1403 unsigned count, uint32_t incr,
1404 uint64_t flags)
1405
1406{
1407 if (level != AMDGPU_VM_PTB) {
1408 flags |= AMDGPU_PDE_PTE(1ULL << 54);
1409 amdgpu_gmc_get_vm_pde(params->adev, level, &addr, &flags)(params->adev)->gmc.gmc_funcs->get_vm_pde((params->
adev), (level), (&addr), (&flags))
;
1410
1411 } else if (params->adev->asic_type >= CHIP_VEGA10 &&
1412 !(flags & AMDGPU_PTE_VALID(1ULL << 0)) &&
1413 !(flags & AMDGPU_PTE_PRT(1ULL << 51))) {
1414
1415 /* Workaround for fault priority problem on GMC9 */
1416 flags |= AMDGPU_PTE_EXECUTABLE(1ULL << 4);
1417 }
1418
1419 params->vm->update_funcs->update(params, bo, pe, addr, count, incr,
1420 flags);
1421}
1422
1423/**
1424 * amdgpu_vm_fragment - get fragment for PTEs
1425 *
1426 * @params: see amdgpu_vm_update_params definition
1427 * @start: first PTE to handle
1428 * @end: last PTE to handle
1429 * @flags: hw mapping flags
1430 * @frag: resulting fragment size
1431 * @frag_end: end of this fragment
1432 *
1433 * Returns the first possible fragment for the start and end address.
1434 */
1435static void amdgpu_vm_fragment(struct amdgpu_vm_update_params *params,
1436 uint64_t start, uint64_t end, uint64_t flags,
1437 unsigned int *frag, uint64_t *frag_end)
1438{
1439 /**
1440 * The MC L1 TLB supports variable sized pages, based on a fragment
1441 * field in the PTE. When this field is set to a non-zero value, page
1442 * granularity is increased from 4KB to (1 << (12 + frag)). The PTE
1443 * flags are considered valid for all PTEs within the fragment range
1444 * and corresponding mappings are assumed to be physically contiguous.
1445 *
1446 * The L1 TLB can store a single PTE for the whole fragment,
1447 * significantly increasing the space available for translation
1448 * caching. This leads to large improvements in throughput when the
1449 * TLB is under pressure.
1450 *
1451 * The L2 TLB distributes small and large fragments into two
1452 * asymmetric partitions. The large fragment cache is significantly
1453 * larger. Thus, we try to use large fragments wherever possible.
1454 * Userspace can support this by aligning virtual base address and
1455 * allocation size to the fragment size.
1456 *
1457 * Starting with Vega10 the fragment size only controls the L1. The L2
1458 * is now directly feed with small/huge/giant pages from the walker.
1459 */
1460 unsigned max_frag;
1461
1462 if (params->adev->asic_type < CHIP_VEGA10)
24
Assuming field 'asic_type' is >= CHIP_VEGA10
25
Taking false branch
1463 max_frag = params->adev->vm_manager.fragment_size;
1464 else
1465 max_frag = 31;
1466
1467 /* system pages are non continuously */
1468 if (params->pages_addr) {
26
Assuming field 'pages_addr' is non-null
27
Taking true branch
1469 *frag = 0;
1470 *frag_end = end;
1471 return;
28
Returning without writing to 'params->unlocked', which participates in a condition later
1472 }
1473
1474 /* This intentionally wraps around if no bit is set */
1475 *frag = min((unsigned)ffs(start) - 1, (unsigned)fls64(end - start) - 1)((((unsigned)ffs(start) - 1)<((unsigned)fls64(end - start)
- 1))?((unsigned)ffs(start) - 1):((unsigned)fls64(end - start
) - 1))
;
1476 if (*frag >= max_frag) {
1477 *frag = max_frag;
1478 *frag_end = end & ~((1ULL << max_frag) - 1);
1479 } else {
1480 *frag_end = start + (1 << *frag);
1481 }
1482}
1483
1484/**
1485 * amdgpu_vm_update_ptes - make sure that page tables are valid
1486 *
1487 * @params: see amdgpu_vm_update_params definition
1488 * @start: start of GPU address range
1489 * @end: end of GPU address range
1490 * @dst: destination address to map to, the next dst inside the function
1491 * @flags: mapping flags
1492 *
1493 * Update the page tables in the range @start - @end.
1494 *
1495 * Returns:
1496 * 0 for success, -EINVAL for failure.
1497 */
1498static int amdgpu_vm_update_ptes(struct amdgpu_vm_update_params *params,
1499 uint64_t start, uint64_t end,
1500 uint64_t dst, uint64_t flags)
1501{
1502 struct amdgpu_device *adev = params->adev;
1503 struct amdgpu_vm_pt_cursor cursor;
1504 uint64_t frag_start = start, frag_end;
1505 unsigned int frag;
1506 int r;
1507
1508 /* figure out the initial fragment */
1509 amdgpu_vm_fragment(params, frag_start, end, flags, &frag, &frag_end);
23
Calling 'amdgpu_vm_fragment'
29
Returning from 'amdgpu_vm_fragment'
1510
1511 /* walk over the address space and update the PTs */
1512 amdgpu_vm_pt_start(adev, params->vm, start, &cursor);
30
Calling 'amdgpu_vm_pt_start'
33
Returning from 'amdgpu_vm_pt_start'
1513 while (cursor.pfn < end) {
34
Assuming 'end' is > field 'pfn'
35
Loop condition is true. Entering loop body
1514 unsigned shift, parent_shift, mask;
1515 uint64_t incr, entry_end, pe_start;
1516 struct amdgpu_bo *pt;
1517
1518 if (!params->unlocked) {
36
Assuming field 'unlocked' is true
37
Taking false branch
1519 /* make sure that the page tables covering the
1520 * address range are actually allocated
1521 */
1522 r = amdgpu_vm_alloc_pts(params->adev, params->vm,
1523 &cursor, params->immediate);
1524 if (r)
1525 return r;
1526 }
1527
1528 shift = amdgpu_vm_level_shift(adev, cursor.level);
38
Calling 'amdgpu_vm_level_shift'
41
Returning from 'amdgpu_vm_level_shift'
42
The value 4294967295 is assigned to 'shift'
1529 parent_shift = amdgpu_vm_level_shift(adev, cursor.level - 1);
1530 if (params->unlocked
42.1
Field 'unlocked' is true
) {
43
Taking true branch
1531 /* Unlocked updates are only allowed on the leaves */
1532 if (amdgpu_vm_pt_descendant(adev, &cursor))
44
Calling 'amdgpu_vm_pt_descendant'
48
Returning from 'amdgpu_vm_pt_descendant'
49
Taking false branch
1533 continue;
1534 } else if (adev->asic_type < CHIP_VEGA10 &&
1535 (flags & AMDGPU_PTE_VALID(1ULL << 0))) {
1536 /* No huge page support before GMC v9 */
1537 if (cursor.level != AMDGPU_VM_PTB) {
1538 if (!amdgpu_vm_pt_descendant(adev, &cursor))
1539 return -ENOENT2;
1540 continue;
1541 }
1542 } else if (frag < shift) {
1543 /* We can't use this level when the fragment size is
1544 * smaller than the address shift. Go to the next
1545 * child entry and try again.
1546 */
1547 if (amdgpu_vm_pt_descendant(adev, &cursor))
1548 continue;
1549 } else if (frag >= parent_shift) {
1550 /* If the fragment size is even larger than the parent
1551 * shift we should go up one level and check it again.
1552 */
1553 if (!amdgpu_vm_pt_ancestor(&cursor))
1554 return -EINVAL22;
1555 continue;
1556 }
1557
1558 pt = cursor.entry->base.bo;
1559 if (!pt) {
50
Assuming 'pt' is non-null
51
Taking false branch
1560 /* We need all PDs and PTs for mapping something, */
1561 if (flags & AMDGPU_PTE_VALID(1ULL << 0))
1562 return -ENOENT2;
1563
1564 /* but unmapping something can happen at a higher
1565 * level.
1566 */
1567 if (!amdgpu_vm_pt_ancestor(&cursor))
1568 return -EINVAL22;
1569
1570 pt = cursor.entry->base.bo;
1571 shift = parent_shift;
1572 frag_end = max(frag_end, roundup2(frag_start + 1,(((frag_end)>((((frag_start + 1) + ((1ULL << shift) -
1)) & (~((__typeof(frag_start + 1))(1ULL << shift)
- 1)))))?(frag_end):((((frag_start + 1) + ((1ULL << shift
) - 1)) & (~((__typeof(frag_start + 1))(1ULL << shift
) - 1)))))
1573 1ULL << shift))(((frag_end)>((((frag_start + 1) + ((1ULL << shift) -
1)) & (~((__typeof(frag_start + 1))(1ULL << shift)
- 1)))))?(frag_end):((((frag_start + 1) + ((1ULL << shift
) - 1)) & (~((__typeof(frag_start + 1))(1ULL << shift
) - 1)))))
;
1574 }
1575
1576 /* Looks good so far, calculate parameters for the update */
1577 incr = (uint64_t)AMDGPU_GPU_PAGE_SIZE4096 << shift;
52
The result of the left shift is undefined due to shifting by '4294967295', which is greater or equal to the width of type 'uint64_t'
1578 mask = amdgpu_vm_entries_mask(adev, cursor.level);
1579 pe_start = ((cursor.pfn >> shift) & mask) * 8;
1580 entry_end = ((uint64_t)mask + 1) << shift;
1581 entry_end += cursor.pfn & ~(entry_end - 1);
1582 entry_end = min(entry_end, end)(((entry_end)<(end))?(entry_end):(end));
1583
1584 do {
1585 struct amdgpu_vm *vm = params->vm;
1586 uint64_t upd_end = min(entry_end, frag_end)(((entry_end)<(frag_end))?(entry_end):(frag_end));
1587 unsigned nptes = (upd_end - frag_start) >> shift;
1588 uint64_t upd_flags = flags | AMDGPU_PTE_FRAG(frag)((frag & 0x1fULL) << 7);
1589
1590 /* This can happen when we set higher level PDs to
1591 * silent to stop fault floods.
1592 */
1593 nptes = max(nptes, 1u)(((nptes)>(1u))?(nptes):(1u));
1594
1595 trace_amdgpu_vm_update_ptes(params, frag_start, upd_end,
1596 nptes, dst, incr, upd_flags,
1597 vm->task_info.pid,
1598 vm->immediate.fence_context);
1599 amdgpu_vm_update_flags(params, pt, cursor.level,
1600 pe_start, dst, nptes, incr,
1601 upd_flags);
1602
1603 pe_start += nptes * 8;
1604 dst += nptes * incr;
1605
1606 frag_start = upd_end;
1607 if (frag_start >= frag_end) {
1608 /* figure out the next fragment */
1609 amdgpu_vm_fragment(params, frag_start, end,
1610 flags, &frag, &frag_end);
1611 if (frag < shift)
1612 break;
1613 }
1614 } while (frag_start < entry_end);
1615
1616 if (amdgpu_vm_pt_descendant(adev, &cursor)) {
1617 /* Free all child entries.
1618 * Update the tables with the flags and addresses and free up subsequent
1619 * tables in the case of huge pages or freed up areas.
1620 * This is the maximum you can free, because all other page tables are not
1621 * completely covered by the range and so potentially still in use.
1622 */
1623 while (cursor.pfn < frag_start) {
1624 amdgpu_vm_free_pts(adev, params->vm, &cursor);
1625 amdgpu_vm_pt_next(adev, &cursor);
1626 }
1627
1628 } else if (frag >= shift) {
1629 /* or just move on to the next on the same level. */
1630 amdgpu_vm_pt_next(adev, &cursor);
1631 }
1632 }
1633
1634 return 0;
1635}
1636
1637/**
1638 * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table
1639 *
1640 * @adev: amdgpu_device pointer
1641 * @vm: requested vm
1642 * @immediate: immediate submission in a page fault
1643 * @unlocked: unlocked invalidation during MM callback
1644 * @resv: fences we need to sync to
1645 * @start: start of mapped range
1646 * @last: last mapped entry
1647 * @flags: flags for the entries
1648 * @addr: addr to set the area to
1649 * @pages_addr: DMA addresses to use for mapping
1650 * @fence: optional resulting fence
1651 *
1652 * Fill in the page table entries between @start and @last.
1653 *
1654 * Returns:
1655 * 0 for success, -EINVAL for failure.
1656 */
1657static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
1658 struct amdgpu_vm *vm, bool_Bool immediate,
1659 bool_Bool unlocked, struct dma_resv *resv,
1660 uint64_t start, uint64_t last,
1661 uint64_t flags, uint64_t addr,
1662 dma_addr_t *pages_addr,
1663 struct dma_fence **fence)
1664{
1665 struct amdgpu_vm_update_params params;
1666 enum amdgpu_sync_mode sync_mode;
1667 int r;
1668
1669 memset(&params, 0, sizeof(params))__builtin_memset((&params), (0), (sizeof(params)));
1670 params.adev = adev;
1671 params.vm = vm;
1672 params.immediate = immediate;
1673 params.pages_addr = pages_addr;
1674 params.unlocked = unlocked;
1675
1676 /* Implicitly sync to command submissions in the same VM before
1677 * unmapping. Sync to moving fences before mapping.
1678 */
1679 if (!(flags & AMDGPU_PTE_VALID(1ULL << 0)))
15
Taking false branch
1680 sync_mode = AMDGPU_SYNC_EQ_OWNER;
1681 else
1682 sync_mode = AMDGPU_SYNC_EXPLICIT;
1683
1684 amdgpu_vm_eviction_lock(vm);
1685 if (vm->evicting) {
16
Assuming field 'evicting' is false
17
Taking false branch
1686 r = -EBUSY16;
1687 goto error_unlock;
1688 }
1689
1690 if (!unlocked
17.1
'unlocked' is false
&& !dma_fence_is_signaled(vm->last_unlocked)) {
18
Assuming the condition is false
19
Taking false branch
1691 struct dma_fence *tmp = dma_fence_get_stub();
1692
1693 amdgpu_bo_fence(vm->root.base.bo, vm->last_unlocked, true1);
1694 swap(vm->last_unlocked, tmp)do { __typeof(vm->last_unlocked) __tmp = (vm->last_unlocked
); (vm->last_unlocked) = (tmp); (tmp) = __tmp; } while(0)
;
1695 dma_fence_put(tmp);
1696 }
1697
1698 r = vm->update_funcs->prepare(&params, resv, sync_mode);
1699 if (r)
20
Assuming 'r' is 0
21
Taking false branch
1700 goto error_unlock;
1701
1702 r = amdgpu_vm_update_ptes(&params, start, last + 1, addr, flags);
22
Calling 'amdgpu_vm_update_ptes'
1703 if (r)
1704 goto error_unlock;
1705
1706 r = vm->update_funcs->commit(&params, fence);
1707
1708error_unlock:
1709 amdgpu_vm_eviction_unlock(vm);
1710 return r;
1711}
1712
1713/**
1714 * amdgpu_vm_bo_split_mapping - split a mapping into smaller chunks
1715 *
1716 * @adev: amdgpu_device pointer
1717 * @resv: fences we need to sync to
1718 * @pages_addr: DMA addresses to use for mapping
1719 * @vm: requested vm
1720 * @mapping: mapped range and flags to use for the update
1721 * @flags: HW flags for the mapping
1722 * @bo_adev: amdgpu_device pointer that bo actually been allocated
1723 * @nodes: array of drm_mm_nodes with the MC addresses
1724 * @fence: optional resulting fence
1725 *
1726 * Split the mapping into smaller chunks so that each update fits
1727 * into a SDMA IB.
1728 *
1729 * Returns:
1730 * 0 for success, -EINVAL for failure.
1731 */
1732static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev,
1733 struct dma_resv *resv,
1734 dma_addr_t *pages_addr,
1735 struct amdgpu_vm *vm,
1736 struct amdgpu_bo_va_mapping *mapping,
1737 uint64_t flags,
1738 struct amdgpu_device *bo_adev,
1739 struct drm_mm_node *nodes,
1740 struct dma_fence **fence)
1741{
1742 unsigned min_linear_pages = 1 << adev->vm_manager.fragment_size;
1743 uint64_t pfn, start = mapping->start;
1744 int r;
1745
1746 /* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
1747 * but in case of something, we filter the flags in first place
1748 */
1749 if (!(mapping->flags & AMDGPU_PTE_READABLE(1ULL << 5)))
1750 flags &= ~AMDGPU_PTE_READABLE(1ULL << 5);
1751 if (!(mapping->flags & AMDGPU_PTE_WRITEABLE(1ULL << 6)))
1752 flags &= ~AMDGPU_PTE_WRITEABLE(1ULL << 6);
1753
1754 /* Apply ASIC specific mapping flags */
1755 amdgpu_gmc_get_vm_pte(adev, mapping, &flags)(adev)->gmc.gmc_funcs->get_vm_pte((adev), (mapping), (&
flags))
;
1756
1757 trace_amdgpu_vm_bo_update(mapping);
1758
1759 pfn = mapping->offset >> PAGE_SHIFT12;
1760 if (nodes) {
1761 while (pfn >= nodes->size) {
1762 pfn -= nodes->size;
1763 ++nodes;
1764 }
1765 }
1766
1767 do {
1768 dma_addr_t *dma_addr = NULL((void *)0);
1769 uint64_t max_entries;
1770 uint64_t addr, last;
1771
1772 max_entries = mapping->last - start + 1;
1773 if (nodes) {
1774 addr = nodes->start << PAGE_SHIFT12;
1775 max_entries = min((nodes->size - pfn) *((((nodes->size - pfn) * ((1 << 12) / 4096))<(max_entries
))?((nodes->size - pfn) * ((1 << 12) / 4096)):(max_entries
))
1776 AMDGPU_GPU_PAGES_IN_CPU_PAGE, max_entries)((((nodes->size - pfn) * ((1 << 12) / 4096))<(max_entries
))?((nodes->size - pfn) * ((1 << 12) / 4096)):(max_entries
))
;
1777 } else {
1778 addr = 0;
1779 }
1780
1781 if (pages_addr) {
1782 uint64_t count;
1783
1784 for (count = 1;
1785 count < max_entries / AMDGPU_GPU_PAGES_IN_CPU_PAGE((1 << 12) / 4096);
1786 ++count) {
1787 uint64_t idx = pfn + count;
1788
1789 if (pages_addr[idx] !=
1790 (pages_addr[idx - 1] + PAGE_SIZE(1 << 12)))
1791 break;
1792 }
1793
1794 if (count < min_linear_pages) {
1795 addr = pfn << PAGE_SHIFT12;
1796 dma_addr = pages_addr;
1797 } else {
1798 addr = pages_addr[pfn];
1799 max_entries = count *
1800 AMDGPU_GPU_PAGES_IN_CPU_PAGE((1 << 12) / 4096);
1801 }
1802
1803 } else if (flags & (AMDGPU_PTE_VALID(1ULL << 0) | AMDGPU_PTE_PRT(1ULL << 51))) {
1804 addr += bo_adev->vm_manager.vram_base_offset;
1805 addr += pfn << PAGE_SHIFT12;
1806 }
1807
1808 last = start + max_entries - 1;
1809 r = amdgpu_vm_bo_update_mapping(adev, vm, false0, false0, resv,
1810 start, last, flags, addr,
1811 dma_addr, fence);
1812 if (r)
1813 return r;
1814
1815 pfn += (last - start + 1) / AMDGPU_GPU_PAGES_IN_CPU_PAGE((1 << 12) / 4096);
1816 if (nodes && nodes->size == pfn) {
1817 pfn = 0;
1818 ++nodes;
1819 }
1820 start = last + 1;
1821
1822 } while (unlikely(start != mapping->last + 1)__builtin_expect(!!(start != mapping->last + 1), 0));
1823
1824 return 0;
1825}
1826
1827/**
1828 * amdgpu_vm_bo_update - update all BO mappings in the vm page table
1829 *
1830 * @adev: amdgpu_device pointer
1831 * @bo_va: requested BO and VM object
1832 * @clear: if true clear the entries
1833 *
1834 * Fill in the page table entries for @bo_va.
1835 *
1836 * Returns:
1837 * 0 for success, -EINVAL for failure.
1838 */
1839int amdgpu_vm_bo_update(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va,
1840 bool_Bool clear)
1841{
1842 struct amdgpu_bo *bo = bo_va->base.bo;
1843 struct amdgpu_vm *vm = bo_va->base.vm;
1844 struct amdgpu_bo_va_mapping *mapping;
1845 dma_addr_t *pages_addr = NULL((void *)0);
1846 struct ttm_resource *mem;
1847 struct drm_mm_node *nodes;
1848 struct dma_fence **last_update;
1849 struct dma_resv *resv;
1850 uint64_t flags;
1851 struct amdgpu_device *bo_adev = adev;
1852 int r;
1853
1854 if (clear || !bo) {
1855 mem = NULL((void *)0);
1856 nodes = NULL((void *)0);
1857 resv = vm->root.base.bo->tbo.base.resv;
1858 } else {
1859 struct drm_gem_object *obj = &bo->tbo.base;
1860 struct ttm_dma_tt *ttm;
1861
1862 resv = bo->tbo.base.resv;
1863#ifdef notyet
1864 if (obj->import_attach && bo_va->is_xgmi) {
1865 struct dma_buf *dma_buf = obj->import_attach->dmabuf;
1866 struct drm_gem_object *gobj = dma_buf->priv;
1867 struct amdgpu_bo *abo = gem_to_amdgpu_bo(gobj)({ const __typeof( ((struct amdgpu_bo *)0)->tbo.base ) *__mptr
= ((gobj)); (struct amdgpu_bo *)( (char *)__mptr - __builtin_offsetof
(struct amdgpu_bo, tbo.base) );})
;
1868
1869 if (abo->tbo.mem.mem_type == TTM_PL_VRAM2)
1870 bo = gem_to_amdgpu_bo(gobj)({ const __typeof( ((struct amdgpu_bo *)0)->tbo.base ) *__mptr
= ((gobj)); (struct amdgpu_bo *)( (char *)__mptr - __builtin_offsetof
(struct amdgpu_bo, tbo.base) );})
;
1871 }
1872#endif
1873 mem = &bo->tbo.mem;
1874 nodes = mem->mm_node;
1875 if (mem->mem_type == TTM_PL_TT1) {
1876 ttm = container_of(bo->tbo.ttm, struct ttm_dma_tt, ttm)({ const __typeof( ((struct ttm_dma_tt *)0)->ttm ) *__mptr
= (bo->tbo.ttm); (struct ttm_dma_tt *)( (char *)__mptr - __builtin_offsetof
(struct ttm_dma_tt, ttm) );})
;
1877 pages_addr = ttm->dma_address;
1878 }
1879 }
1880
1881 if (bo) {
1882 flags = amdgpu_ttm_tt_pte_flags(adev, bo->tbo.ttm, mem);
1883
1884 if (amdgpu_bo_encrypted(bo))
1885 flags |= AMDGPU_PTE_TMZ(1ULL << 3);
1886
1887 bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
1888 } else {
1889 flags = 0x0;
1890 }
1891
1892 if (clear || (bo && bo->tbo.base.resv ==
1893 vm->root.base.bo->tbo.base.resv))
1894 last_update = &vm->last_update;
1895 else
1896 last_update = &bo_va->last_pt_update;
1897
1898 if (!clear && bo_va->base.moved) {
1899 bo_va->base.moved = false0;
1900 list_splice_init(&bo_va->valids, &bo_va->invalids);
1901
1902 } else if (bo_va->cleared != clear) {
1903 list_splice_init(&bo_va->valids, &bo_va->invalids);
1904 }
1905
1906 list_for_each_entry(mapping, &bo_va->invalids, list)for (mapping = ({ const __typeof( ((__typeof(*mapping) *)0)->
list ) *__mptr = ((&bo_va->invalids)->next); (__typeof
(*mapping) *)( (char *)__mptr - __builtin_offsetof(__typeof(*
mapping), list) );}); &mapping->list != (&bo_va->
invalids); mapping = ({ const __typeof( ((__typeof(*mapping) *
)0)->list ) *__mptr = (mapping->list.next); (__typeof(*
mapping) *)( (char *)__mptr - __builtin_offsetof(__typeof(*mapping
), list) );}))
{
1907 r = amdgpu_vm_bo_split_mapping(adev, resv, pages_addr, vm,
1908 mapping, flags, bo_adev, nodes,
1909 last_update);
1910 if (r)
1911 return r;
1912 }
1913
1914 /* If the BO is not in its preferred location add it back to
1915 * the evicted list so that it gets validated again on the
1916 * next command submission.
1917 */
1918 if (bo && bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv) {
1919 uint32_t mem_type = bo->tbo.mem.mem_type;
1920
1921 if (!(bo->preferred_domains &
1922 amdgpu_mem_type_to_domain(mem_type)))
1923 amdgpu_vm_bo_evicted(&bo_va->base);
1924 else
1925 amdgpu_vm_bo_idle(&bo_va->base);
1926 } else {
1927 amdgpu_vm_bo_done(&bo_va->base);
1928 }
1929
1930 list_splice_init(&bo_va->invalids, &bo_va->valids);
1931 bo_va->cleared = clear;
1932
1933 if (trace_amdgpu_vm_bo_mapping_enabled()) {
1934 list_for_each_entry(mapping, &bo_va->valids, list)for (mapping = ({ const __typeof( ((__typeof(*mapping) *)0)->
list ) *__mptr = ((&bo_va->valids)->next); (__typeof
(*mapping) *)( (char *)__mptr - __builtin_offsetof(__typeof(*
mapping), list) );}); &mapping->list != (&bo_va->
valids); mapping = ({ const __typeof( ((__typeof(*mapping) *)
0)->list ) *__mptr = (mapping->list.next); (__typeof(*mapping
) *)( (char *)__mptr - __builtin_offsetof(__typeof(*mapping),
list) );}))
1935 trace_amdgpu_vm_bo_mapping(mapping);
1936 }
1937
1938 return 0;
1939}
1940
1941/**
1942 * amdgpu_vm_update_prt_state - update the global PRT state
1943 *
1944 * @adev: amdgpu_device pointer
1945 */
1946static void amdgpu_vm_update_prt_state(struct amdgpu_device *adev)
1947{
1948 unsigned long flags;
1949 bool_Bool enable;
1950
1951 spin_lock_irqsave(&adev->vm_manager.prt_lock, flags)do { flags = 0; mtx_enter(&adev->vm_manager.prt_lock);
} while (0)
;
1952 enable = !!atomic_read(&adev->vm_manager.num_prt_users)({ typeof(*(&adev->vm_manager.num_prt_users)) __tmp = *
(volatile typeof(*(&adev->vm_manager.num_prt_users)) *
)&(*(&adev->vm_manager.num_prt_users)); membar_datadep_consumer
(); __tmp; })
;
1953 adev->gmc.gmc_funcs->set_prt(adev, enable);
1954 spin_unlock_irqrestore(&adev->vm_manager.prt_lock, flags)do { (void)(flags); mtx_leave(&adev->vm_manager.prt_lock
); } while (0)
;
1955}
1956
1957/**
1958 * amdgpu_vm_prt_get - add a PRT user
1959 *
1960 * @adev: amdgpu_device pointer
1961 */
1962static void amdgpu_vm_prt_get(struct amdgpu_device *adev)
1963{
1964 if (!adev->gmc.gmc_funcs->set_prt)
1965 return;
1966
1967 if (atomic_inc_return(&adev->vm_manager.num_prt_users)__sync_add_and_fetch((&adev->vm_manager.num_prt_users)
, 1)
== 1)
1968 amdgpu_vm_update_prt_state(adev);
1969}
1970
1971/**
1972 * amdgpu_vm_prt_put - drop a PRT user
1973 *
1974 * @adev: amdgpu_device pointer
1975 */
1976static void amdgpu_vm_prt_put(struct amdgpu_device *adev)
1977{
1978 if (atomic_dec_return(&adev->vm_manager.num_prt_users)__sync_sub_and_fetch((&adev->vm_manager.num_prt_users)
, 1)
== 0)
1979 amdgpu_vm_update_prt_state(adev);
1980}
1981
1982/**
1983 * amdgpu_vm_prt_cb - callback for updating the PRT status
1984 *
1985 * @fence: fence for the callback
1986 * @_cb: the callback function
1987 */
1988static void amdgpu_vm_prt_cb(struct dma_fence *fence, struct dma_fence_cb *_cb)
1989{
1990 struct amdgpu_prt_cb *cb = container_of(_cb, struct amdgpu_prt_cb, cb)({ const __typeof( ((struct amdgpu_prt_cb *)0)->cb ) *__mptr
= (_cb); (struct amdgpu_prt_cb *)( (char *)__mptr - __builtin_offsetof
(struct amdgpu_prt_cb, cb) );})
;
1991
1992 amdgpu_vm_prt_put(cb->adev);
1993 kfree(cb);
1994}
1995
1996/**
1997 * amdgpu_vm_add_prt_cb - add callback for updating the PRT status
1998 *
1999 * @adev: amdgpu_device pointer
2000 * @fence: fence for the callback
2001 */
2002static void amdgpu_vm_add_prt_cb(struct amdgpu_device *adev,
2003 struct dma_fence *fence)
2004{
2005 struct amdgpu_prt_cb *cb;
2006
2007 if (!adev->gmc.gmc_funcs->set_prt)
2008 return;
2009
2010 cb = kmalloc(sizeof(struct amdgpu_prt_cb), GFP_KERNEL(0x0001 | 0x0004));
2011 if (!cb) {
2012 /* Last resort when we are OOM */
2013 if (fence)
2014 dma_fence_wait(fence, false0);
2015
2016 amdgpu_vm_prt_put(adev);
2017 } else {
2018 cb->adev = adev;
2019 if (!fence || dma_fence_add_callback(fence, &cb->cb,
2020 amdgpu_vm_prt_cb))
2021 amdgpu_vm_prt_cb(fence, &cb->cb);
2022 }
2023}
2024
2025/**
2026 * amdgpu_vm_free_mapping - free a mapping
2027 *
2028 * @adev: amdgpu_device pointer
2029 * @vm: requested vm
2030 * @mapping: mapping to be freed
2031 * @fence: fence of the unmap operation
2032 *
2033 * Free a mapping and make sure we decrease the PRT usage count if applicable.
2034 */
2035static void amdgpu_vm_free_mapping(struct amdgpu_device *adev,
2036 struct amdgpu_vm *vm,
2037 struct amdgpu_bo_va_mapping *mapping,
2038 struct dma_fence *fence)
2039{
2040 if (mapping->flags & AMDGPU_PTE_PRT(1ULL << 51))
2041 amdgpu_vm_add_prt_cb(adev, fence);
2042 kfree(mapping);
2043}
2044
2045/**
2046 * amdgpu_vm_prt_fini - finish all prt mappings
2047 *
2048 * @adev: amdgpu_device pointer
2049 * @vm: requested vm
2050 *
2051 * Register a cleanup callback to disable PRT support after VM dies.
2052 */
2053static void amdgpu_vm_prt_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2054{
2055 struct dma_resv *resv = vm->root.base.bo->tbo.base.resv;
2056 struct dma_fence *excl, **shared;
2057 unsigned i, shared_count;
2058 int r;
2059
2060 r = dma_resv_get_fences_rcu(resv, &excl,
2061 &shared_count, &shared);
2062 if (r) {
2063 /* Not enough memory to grab the fence list, as last resort
2064 * block for all the fences to complete.
2065 */
2066 dma_resv_wait_timeout_rcu(resv, true1, false0,
2067 MAX_SCHEDULE_TIMEOUT(0x7fffffff));
2068 return;
2069 }
2070
2071 /* Add a callback for each fence in the reservation object */
2072 amdgpu_vm_prt_get(adev);
2073 amdgpu_vm_add_prt_cb(adev, excl);
2074
2075 for (i = 0; i < shared_count; ++i) {
2076 amdgpu_vm_prt_get(adev);
2077 amdgpu_vm_add_prt_cb(adev, shared[i]);
2078 }
2079
2080 kfree(shared);
2081}
2082
2083/**
2084 * amdgpu_vm_clear_freed - clear freed BOs in the PT
2085 *
2086 * @adev: amdgpu_device pointer
2087 * @vm: requested vm
2088 * @fence: optional resulting fence (unchanged if no work needed to be done
2089 * or if an error occurred)
2090 *
2091 * Make sure all freed BOs are cleared in the PT.
2092 * PTs have to be reserved and mutex must be locked!
2093 *
2094 * Returns:
2095 * 0 for success.
2096 *
2097 */
2098int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
2099 struct amdgpu_vm *vm,
2100 struct dma_fence **fence)
2101{
2102 struct dma_resv *resv = vm->root.base.bo->tbo.base.resv;
2103 struct amdgpu_bo_va_mapping *mapping;
2104 uint64_t init_pte_value = 0;
2105 struct dma_fence *f = NULL((void *)0);
2106 int r;
2107
2108 while (!list_empty(&vm->freed)) {
2109 mapping = list_first_entry(&vm->freed,({ const __typeof( ((struct amdgpu_bo_va_mapping *)0)->list
) *__mptr = ((&vm->freed)->next); (struct amdgpu_bo_va_mapping
*)( (char *)__mptr - __builtin_offsetof(struct amdgpu_bo_va_mapping
, list) );})
2110 struct amdgpu_bo_va_mapping, list)({ const __typeof( ((struct amdgpu_bo_va_mapping *)0)->list
) *__mptr = ((&vm->freed)->next); (struct amdgpu_bo_va_mapping
*)( (char *)__mptr - __builtin_offsetof(struct amdgpu_bo_va_mapping
, list) );})
;
2111 list_del(&mapping->list);
2112
2113 if (vm->pte_support_ats &&
2114 mapping->start < AMDGPU_GMC_HOLE_START0x0000800000000000ULL)
2115 init_pte_value = AMDGPU_PTE_DEFAULT_ATC((1ULL << 1) | (1ULL << 2) | (1ULL << 4) | (
1ULL << 5) | (1ULL << 6) | ((uint64_t)(2) <<
57))
;
2116
2117 r = amdgpu_vm_bo_update_mapping(adev, vm, false0, false0, resv,
2118 mapping->start, mapping->last,
2119 init_pte_value, 0, NULL((void *)0), &f);
2120 amdgpu_vm_free_mapping(adev, vm, mapping, f);
2121 if (r) {
2122 dma_fence_put(f);
2123 return r;
2124 }
2125 }
2126
2127 if (fence && f) {
2128 dma_fence_put(*fence);
2129 *fence = f;
2130 } else {
2131 dma_fence_put(f);
2132 }
2133
2134 return 0;
2135
2136}
2137
2138/**
2139 * amdgpu_vm_handle_moved - handle moved BOs in the PT
2140 *
2141 * @adev: amdgpu_device pointer
2142 * @vm: requested vm
2143 *
2144 * Make sure all BOs which are moved are updated in the PTs.
2145 *
2146 * Returns:
2147 * 0 for success.
2148 *
2149 * PTs have to be reserved!
2150 */
2151int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
2152 struct amdgpu_vm *vm)
2153{
2154 struct amdgpu_bo_va *bo_va, *tmp;
2155 struct dma_resv *resv;
2156 bool_Bool clear;
2157 int r;
2158
2159 list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status)for (bo_va = ({ const __typeof( ((__typeof(*bo_va) *)0)->base
.vm_status ) *__mptr = ((&vm->moved)->next); (__typeof
(*bo_va) *)( (char *)__mptr - __builtin_offsetof(__typeof(*bo_va
), base.vm_status) );}), tmp = ({ const __typeof( ((__typeof(
*bo_va) *)0)->base.vm_status ) *__mptr = (bo_va->base.vm_status
.next); (__typeof(*bo_va) *)( (char *)__mptr - __builtin_offsetof
(__typeof(*bo_va), base.vm_status) );}); &bo_va->base.
vm_status != (&vm->moved); bo_va = tmp, tmp = ({ const
__typeof( ((__typeof(*tmp) *)0)->base.vm_status ) *__mptr
= (tmp->base.vm_status.next); (__typeof(*tmp) *)( (char *
)__mptr - __builtin_offsetof(__typeof(*tmp), base.vm_status) )
;}))
{
2160 /* Per VM BOs never need to bo cleared in the page tables */
2161 r = amdgpu_vm_bo_update(adev, bo_va, false0);
2162 if (r)
2163 return r;
2164 }
2165
2166 spin_lock(&vm->invalidated_lock)mtx_enter(&vm->invalidated_lock);
2167 while (!list_empty(&vm->invalidated)) {
2168 bo_va = list_first_entry(&vm->invalidated, struct amdgpu_bo_va,({ const __typeof( ((struct amdgpu_bo_va *)0)->base.vm_status
) *__mptr = ((&vm->invalidated)->next); (struct amdgpu_bo_va
*)( (char *)__mptr - __builtin_offsetof(struct amdgpu_bo_va,
base.vm_status) );})
2169 base.vm_status)({ const __typeof( ((struct amdgpu_bo_va *)0)->base.vm_status
) *__mptr = ((&vm->invalidated)->next); (struct amdgpu_bo_va
*)( (char *)__mptr - __builtin_offsetof(struct amdgpu_bo_va,
base.vm_status) );})
;
2170 resv = bo_va->base.bo->tbo.base.resv;
2171 spin_unlock(&vm->invalidated_lock)mtx_leave(&vm->invalidated_lock);
2172
2173 /* Try to reserve the BO to avoid clearing its ptes */
2174 if (!amdgpu_vm_debug && dma_resv_trylock(resv))
2175 clear = false0;
2176 /* Somebody else is using the BO right now */
2177 else
2178 clear = true1;
2179
2180 r = amdgpu_vm_bo_update(adev, bo_va, clear);
2181 if (r)
2182 return r;
2183
2184 if (!clear)
2185 dma_resv_unlock(resv);
2186 spin_lock(&vm->invalidated_lock)mtx_enter(&vm->invalidated_lock);
2187 }
2188 spin_unlock(&vm->invalidated_lock)mtx_leave(&vm->invalidated_lock);
2189
2190 return 0;
2191}
2192
2193/**
2194 * amdgpu_vm_bo_add - add a bo to a specific vm
2195 *
2196 * @adev: amdgpu_device pointer
2197 * @vm: requested vm
2198 * @bo: amdgpu buffer object
2199 *
2200 * Add @bo into the requested vm.
2201 * Add @bo to the list of bos associated with the vm
2202 *
2203 * Returns:
2204 * Newly added bo_va or NULL for failure
2205 *
2206 * Object has to be reserved!
2207 */
2208struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
2209 struct amdgpu_vm *vm,
2210 struct amdgpu_bo *bo)
2211{
2212 struct amdgpu_bo_va *bo_va;
2213
2214 bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL(0x0001 | 0x0004));
2215 if (bo_va == NULL((void *)0)) {
2216 return NULL((void *)0);
2217 }
2218 amdgpu_vm_bo_base_init(&bo_va->base, vm, bo);
2219
2220 bo_va->ref_count = 1;
2221 INIT_LIST_HEAD(&bo_va->valids);
2222 INIT_LIST_HEAD(&bo_va->invalids);
2223
2224 if (!bo)
2225 return bo_va;
2226
2227 if (amdgpu_dmabuf_is_xgmi_accessible(adev, bo)) {
2228 bo_va->is_xgmi = true1;
2229 /* Power up XGMI if it can be potentially used */
2230 amdgpu_xgmi_set_pstate(adev, AMDGPU_XGMI_PSTATE_MAX_VEGA20);
2231 }
2232
2233 return bo_va;
2234}
2235
2236
2237/**
2238 * amdgpu_vm_bo_insert_mapping - insert a new mapping
2239 *
2240 * @adev: amdgpu_device pointer
2241 * @bo_va: bo_va to store the address
2242 * @mapping: the mapping to insert
2243 *
2244 * Insert a new mapping into all structures.
2245 */
2246static void amdgpu_vm_bo_insert_map(struct amdgpu_device *adev,
2247 struct amdgpu_bo_va *bo_va,
2248 struct amdgpu_bo_va_mapping *mapping)
2249{
2250 struct amdgpu_vm *vm = bo_va->base.vm;
2251 struct amdgpu_bo *bo = bo_va->base.bo;
2252
2253 mapping->bo_va = bo_va;
2254 list_add(&mapping->list, &bo_va->invalids);
2255 amdgpu_vm_it_insert(mapping, &vm->va);
2256
2257 if (mapping->flags & AMDGPU_PTE_PRT(1ULL << 51))
2258 amdgpu_vm_prt_get(adev);
2259
2260 if (bo && bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv &&
2261 !bo_va->base.moved) {
2262 list_move(&bo_va->base.vm_status, &vm->moved);
2263 }
2264 trace_amdgpu_vm_bo_map(bo_va, mapping);
2265}
2266
2267/**
2268 * amdgpu_vm_bo_map - map bo inside a vm
2269 *
2270 * @adev: amdgpu_device pointer
2271 * @bo_va: bo_va to store the address
2272 * @saddr: where to map the BO
2273 * @offset: requested offset in the BO
2274 * @size: BO size in bytes
2275 * @flags: attributes of pages (read/write/valid/etc.)
2276 *
2277 * Add a mapping of the BO at the specefied addr into the VM.
2278 *
2279 * Returns:
2280 * 0 for success, error for failure.
2281 *
2282 * Object has to be reserved and unreserved outside!
2283 */
2284int amdgpu_vm_bo_map(struct amdgpu_device *adev,
2285 struct amdgpu_bo_va *bo_va,
2286 uint64_t saddr, uint64_t offset,
2287 uint64_t size, uint64_t flags)
2288{
2289 struct amdgpu_bo_va_mapping *mapping, *tmp;
2290 struct amdgpu_bo *bo = bo_va->base.bo;
2291 struct amdgpu_vm *vm = bo_va->base.vm;
2292 uint64_t eaddr;
2293
2294 /* validate the parameters */
2295 if (saddr & ~LINUX_PAGE_MASK(~((1 << 12) - 1)) || offset & ~LINUX_PAGE_MASK(~((1 << 12) - 1)) ||
2296 size == 0 || size & ~LINUX_PAGE_MASK(~((1 << 12) - 1)))
2297 return -EINVAL22;
2298
2299 /* make sure object fit at this offset */
2300 eaddr = saddr + size - 1;
2301 if (saddr >= eaddr ||
2302 (bo && offset + size > amdgpu_bo_size(bo)) ||
2303 (eaddr >= adev->vm_manager.max_pfn << AMDGPU_GPU_PAGE_SHIFT12))
2304 return -EINVAL22;
2305
2306 saddr /= AMDGPU_GPU_PAGE_SIZE4096;
2307 eaddr /= AMDGPU_GPU_PAGE_SIZE4096;
2308
2309 tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
2310 if (tmp) {
2311 /* bo and tmp overlap, invalid addr */
2312 dev_err(adev->dev, "bo %p va 0x%010llx-0x%010llx conflict with "printf("drm:pid%d:%s *ERROR* " "bo %p va 0x%010llx-0x%010llx conflict with "
"0x%010llx-0x%010llx\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__ , bo, saddr, eaddr, tmp->start, tmp->last + 1
)
2313 "0x%010llx-0x%010llx\n", bo, saddr, eaddr,printf("drm:pid%d:%s *ERROR* " "bo %p va 0x%010llx-0x%010llx conflict with "
"0x%010llx-0x%010llx\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__ , bo, saddr, eaddr, tmp->start, tmp->last + 1
)
2314 tmp->start, tmp->last + 1)printf("drm:pid%d:%s *ERROR* " "bo %p va 0x%010llx-0x%010llx conflict with "
"0x%010llx-0x%010llx\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__ , bo, saddr, eaddr, tmp->start, tmp->last + 1
)
;
2315 return -EINVAL22;
2316 }
2317
2318 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL(0x0001 | 0x0004));
2319 if (!mapping)
2320 return -ENOMEM12;
2321
2322 mapping->start = saddr;
2323 mapping->last = eaddr;
2324 mapping->offset = offset;
2325 mapping->flags = flags;
2326
2327 amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
2328
2329 return 0;
2330}
2331
2332/**
2333 * amdgpu_vm_bo_replace_map - map bo inside a vm, replacing existing mappings
2334 *
2335 * @adev: amdgpu_device pointer
2336 * @bo_va: bo_va to store the address
2337 * @saddr: where to map the BO
2338 * @offset: requested offset in the BO
2339 * @size: BO size in bytes
2340 * @flags: attributes of pages (read/write/valid/etc.)
2341 *
2342 * Add a mapping of the BO at the specefied addr into the VM. Replace existing
2343 * mappings as we do so.
2344 *
2345 * Returns:
2346 * 0 for success, error for failure.
2347 *
2348 * Object has to be reserved and unreserved outside!
2349 */
2350int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
2351 struct amdgpu_bo_va *bo_va,
2352 uint64_t saddr, uint64_t offset,
2353 uint64_t size, uint64_t flags)
2354{
2355 struct amdgpu_bo_va_mapping *mapping;
2356 struct amdgpu_bo *bo = bo_va->base.bo;
2357 uint64_t eaddr;
2358 int r;
2359
2360 /* validate the parameters */
2361 if (saddr & ~LINUX_PAGE_MASK(~((1 << 12) - 1)) || offset & ~LINUX_PAGE_MASK(~((1 << 12) - 1)) ||
2362 size == 0 || size & ~LINUX_PAGE_MASK(~((1 << 12) - 1)))
2363 return -EINVAL22;
2364
2365 /* make sure object fit at this offset */
2366 eaddr = saddr + size - 1;
2367 if (saddr >= eaddr ||
2368 (bo && offset + size > amdgpu_bo_size(bo)) ||
2369 (eaddr >= adev->vm_manager.max_pfn << AMDGPU_GPU_PAGE_SHIFT12))
2370 return -EINVAL22;
2371
2372 /* Allocate all the needed memory */
2373 mapping = kmalloc(sizeof(*mapping), GFP_KERNEL(0x0001 | 0x0004));
2374 if (!mapping)
2375 return -ENOMEM12;
2376
2377 r = amdgpu_vm_bo_clear_mappings(adev, bo_va->base.vm, saddr, size);
2378 if (r) {
2379 kfree(mapping);
2380 return r;
2381 }
2382
2383 saddr /= AMDGPU_GPU_PAGE_SIZE4096;
2384 eaddr /= AMDGPU_GPU_PAGE_SIZE4096;
2385
2386 mapping->start = saddr;
2387 mapping->last = eaddr;
2388 mapping->offset = offset;
2389 mapping->flags = flags;
2390
2391 amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
2392
2393 return 0;
2394}
2395
2396/**
2397 * amdgpu_vm_bo_unmap - remove bo mapping from vm
2398 *
2399 * @adev: amdgpu_device pointer
2400 * @bo_va: bo_va to remove the address from
2401 * @saddr: where to the BO is mapped
2402 *
2403 * Remove a mapping of the BO at the specefied addr from the VM.
2404 *
2405 * Returns:
2406 * 0 for success, error for failure.
2407 *
2408 * Object has to be reserved and unreserved outside!
2409 */
2410int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
2411 struct amdgpu_bo_va *bo_va,
2412 uint64_t saddr)
2413{
2414 struct amdgpu_bo_va_mapping *mapping;
2415 struct amdgpu_vm *vm = bo_va->base.vm;
2416 bool_Bool valid = true1;
2417
2418 saddr /= AMDGPU_GPU_PAGE_SIZE4096;
2419
2420 list_for_each_entry(mapping, &bo_va->valids, list)for (mapping = ({ const __typeof( ((__typeof(*mapping) *)0)->
list ) *__mptr = ((&bo_va->valids)->next); (__typeof
(*mapping) *)( (char *)__mptr - __builtin_offsetof(__typeof(*
mapping), list) );}); &mapping->list != (&bo_va->
valids); mapping = ({ const __typeof( ((__typeof(*mapping) *)
0)->list ) *__mptr = (mapping->list.next); (__typeof(*mapping
) *)( (char *)__mptr - __builtin_offsetof(__typeof(*mapping),
list) );}))
{
2421 if (mapping->start == saddr)
2422 break;
2423 }
2424
2425 if (&mapping->list == &bo_va->valids) {
2426 valid = false0;
2427
2428 list_for_each_entry(mapping, &bo_va->invalids, list)for (mapping = ({ const __typeof( ((__typeof(*mapping) *)0)->
list ) *__mptr = ((&bo_va->invalids)->next); (__typeof
(*mapping) *)( (char *)__mptr - __builtin_offsetof(__typeof(*
mapping), list) );}); &mapping->list != (&bo_va->
invalids); mapping = ({ const __typeof( ((__typeof(*mapping) *
)0)->list ) *__mptr = (mapping->list.next); (__typeof(*
mapping) *)( (char *)__mptr - __builtin_offsetof(__typeof(*mapping
), list) );}))
{
2429 if (mapping->start == saddr)
2430 break;
2431 }
2432
2433 if (&mapping->list == &bo_va->invalids)
2434 return -ENOENT2;
2435 }
2436
2437 list_del(&mapping->list);
2438 amdgpu_vm_it_remove(mapping, &vm->va);
2439 mapping->bo_va = NULL((void *)0);
2440 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2441
2442 if (valid)
2443 list_add(&mapping->list, &vm->freed);
2444 else
2445 amdgpu_vm_free_mapping(adev, vm, mapping,
2446 bo_va->last_pt_update);
2447
2448 return 0;
2449}
2450
2451/**
2452 * amdgpu_vm_bo_clear_mappings - remove all mappings in a specific range
2453 *
2454 * @adev: amdgpu_device pointer
2455 * @vm: VM structure to use
2456 * @saddr: start of the range
2457 * @size: size of the range
2458 *
2459 * Remove all mappings in a range, split them as appropriate.
2460 *
2461 * Returns:
2462 * 0 for success, error for failure.
2463 */
2464int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
2465 struct amdgpu_vm *vm,
2466 uint64_t saddr, uint64_t size)
2467{
2468 struct amdgpu_bo_va_mapping *before, *after, *tmp, *next;
2469 DRM_LIST_HEAD(removed)struct list_head removed = { &(removed), &(removed) };
2470 uint64_t eaddr;
2471
2472 eaddr = saddr + size - 1;
2473 saddr /= AMDGPU_GPU_PAGE_SIZE4096;
2474 eaddr /= AMDGPU_GPU_PAGE_SIZE4096;
2475
2476 /* Allocate all the needed memory */
2477 before = kzalloc(sizeof(*before), GFP_KERNEL(0x0001 | 0x0004));
2478 if (!before)
2479 return -ENOMEM12;
2480 INIT_LIST_HEAD(&before->list);
2481
2482 after = kzalloc(sizeof(*after), GFP_KERNEL(0x0001 | 0x0004));
2483 if (!after) {
2484 kfree(before);
2485 return -ENOMEM12;
2486 }
2487 INIT_LIST_HEAD(&after->list);
2488
2489 /* Now gather all removed mappings */
2490 tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
2491 while (tmp) {
2492 /* Remember mapping split at the start */
2493 if (tmp->start < saddr) {
2494 before->start = tmp->start;
2495 before->last = saddr - 1;
2496 before->offset = tmp->offset;
2497 before->flags = tmp->flags;
2498 before->bo_va = tmp->bo_va;
2499 list_add(&before->list, &tmp->bo_va->invalids);
2500 }
2501
2502 /* Remember mapping split at the end */
2503 if (tmp->last > eaddr) {
2504 after->start = eaddr + 1;
2505 after->last = tmp->last;
2506 after->offset = tmp->offset;
2507 after->offset += (after->start - tmp->start) << PAGE_SHIFT12;
2508 after->flags = tmp->flags;
2509 after->bo_va = tmp->bo_va;
2510 list_add(&after->list, &tmp->bo_va->invalids);
2511 }
2512
2513 list_del(&tmp->list);
2514 list_add(&tmp->list, &removed);
2515
2516 tmp = amdgpu_vm_it_iter_next(tmp, saddr, eaddr);
2517 }
2518
2519 /* And free them up */
2520 list_for_each_entry_safe(tmp, next, &removed, list)for (tmp = ({ const __typeof( ((__typeof(*tmp) *)0)->list )
*__mptr = ((&removed)->next); (__typeof(*tmp) *)( (char
*)__mptr - __builtin_offsetof(__typeof(*tmp), list) );}), next
= ({ const __typeof( ((__typeof(*tmp) *)0)->list ) *__mptr
= (tmp->list.next); (__typeof(*tmp) *)( (char *)__mptr - __builtin_offsetof
(__typeof(*tmp), list) );}); &tmp->list != (&removed
); tmp = next, next = ({ const __typeof( ((__typeof(*next) *)
0)->list ) *__mptr = (next->list.next); (__typeof(*next
) *)( (char *)__mptr - __builtin_offsetof(__typeof(*next), list
) );}))
{
2521 amdgpu_vm_it_remove(tmp, &vm->va);
2522 list_del(&tmp->list);
2523
2524 if (tmp->start < saddr)
2525 tmp->start = saddr;
2526 if (tmp->last > eaddr)
2527 tmp->last = eaddr;
2528
2529 tmp->bo_va = NULL((void *)0);
2530 list_add(&tmp->list, &vm->freed);
2531 trace_amdgpu_vm_bo_unmap(NULL((void *)0), tmp);
2532 }
2533
2534 /* Insert partial mapping before the range */
2535 if (!list_empty(&before->list)) {
2536 amdgpu_vm_it_insert(before, &vm->va);
2537 if (before->flags & AMDGPU_PTE_PRT(1ULL << 51))
2538 amdgpu_vm_prt_get(adev);
2539 } else {
2540 kfree(before);
2541 }
2542
2543 /* Insert partial mapping after the range */
2544 if (!list_empty(&after->list)) {
2545 amdgpu_vm_it_insert(after, &vm->va);
2546 if (after->flags & AMDGPU_PTE_PRT(1ULL << 51))
2547 amdgpu_vm_prt_get(adev);
2548 } else {
2549 kfree(after);
2550 }
2551
2552 return 0;
2553}
2554
2555/**
2556 * amdgpu_vm_bo_lookup_mapping - find mapping by address
2557 *
2558 * @vm: the requested VM
2559 * @addr: the address
2560 *
2561 * Find a mapping by it's address.
2562 *
2563 * Returns:
2564 * The amdgpu_bo_va_mapping matching for addr or NULL
2565 *
2566 */
2567struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
2568 uint64_t addr)
2569{
2570 return amdgpu_vm_it_iter_first(&vm->va, addr, addr);
2571}
2572
2573/**
2574 * amdgpu_vm_bo_trace_cs - trace all reserved mappings
2575 *
2576 * @vm: the requested vm
2577 * @ticket: CS ticket
2578 *
2579 * Trace all mappings of BOs reserved during a command submission.
2580 */
2581void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket)
2582{
2583 struct amdgpu_bo_va_mapping *mapping;
2584
2585 if (!trace_amdgpu_vm_bo_cs_enabled())
2586 return;
2587
2588 for (mapping = amdgpu_vm_it_iter_first(&vm->va, 0, U64_MAX0xffffffffffffffffULL); mapping;
2589 mapping = amdgpu_vm_it_iter_next(mapping, 0, U64_MAX0xffffffffffffffffULL)) {
2590 if (mapping->bo_va && mapping->bo_va->base.bo) {
2591 struct amdgpu_bo *bo;
2592
2593 bo = mapping->bo_va->base.bo;
2594 if (dma_resv_locking_ctx(bo->tbo.base.resv) !=
2595 ticket)
2596 continue;
2597 }
2598
2599 trace_amdgpu_vm_bo_cs(mapping);
2600 }
2601}
2602
2603/**
2604 * amdgpu_vm_bo_rmv - remove a bo to a specific vm
2605 *
2606 * @adev: amdgpu_device pointer
2607 * @bo_va: requested bo_va
2608 *
2609 * Remove @bo_va->bo from the requested vm.
2610 *
2611 * Object have to be reserved!
2612 */
2613void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
2614 struct amdgpu_bo_va *bo_va)
2615{
2616 struct amdgpu_bo_va_mapping *mapping, *next;
2617 struct amdgpu_bo *bo = bo_va->base.bo;
2618 struct amdgpu_vm *vm = bo_va->base.vm;
2619 struct amdgpu_vm_bo_base **base;
2620
2621 if (bo) {
2622 if (bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv)
2623 vm->bulk_moveable = false0;
2624
2625 for (base = &bo_va->base.bo->vm_bo; *base;
2626 base = &(*base)->next) {
2627 if (*base != &bo_va->base)
2628 continue;
2629
2630 *base = bo_va->base.next;
2631 break;
2632 }
2633 }
2634
2635 spin_lock(&vm->invalidated_lock)mtx_enter(&vm->invalidated_lock);
2636 list_del(&bo_va->base.vm_status);
2637 spin_unlock(&vm->invalidated_lock)mtx_leave(&vm->invalidated_lock);
2638
2639 list_for_each_entry_safe(mapping, next, &bo_va->valids, list)for (mapping = ({ const __typeof( ((__typeof(*mapping) *)0)->
list ) *__mptr = ((&bo_va->valids)->next); (__typeof
(*mapping) *)( (char *)__mptr - __builtin_offsetof(__typeof(*
mapping), list) );}), next = ({ const __typeof( ((__typeof(*mapping
) *)0)->list ) *__mptr = (mapping->list.next); (__typeof
(*mapping) *)( (char *)__mptr - __builtin_offsetof(__typeof(*
mapping), list) );}); &mapping->list != (&bo_va->
valids); mapping = next, next = ({ const __typeof( ((__typeof
(*next) *)0)->list ) *__mptr = (next->list.next); (__typeof
(*next) *)( (char *)__mptr - __builtin_offsetof(__typeof(*next
), list) );}))
{
2640 list_del(&mapping->list);
2641 amdgpu_vm_it_remove(mapping, &vm->va);
2642 mapping->bo_va = NULL((void *)0);
2643 trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2644 list_add(&mapping->list, &vm->freed);
2645 }
2646 list_for_each_entry_safe(mapping, next, &bo_va->invalids, list)for (mapping = ({ const __typeof( ((__typeof(*mapping) *)0)->
list ) *__mptr = ((&bo_va->invalids)->next); (__typeof
(*mapping) *)( (char *)__mptr - __builtin_offsetof(__typeof(*
mapping), list) );}), next = ({ const __typeof( ((__typeof(*mapping
) *)0)->list ) *__mptr = (mapping->list.next); (__typeof
(*mapping) *)( (char *)__mptr - __builtin_offsetof(__typeof(*
mapping), list) );}); &mapping->list != (&bo_va->
invalids); mapping = next, next = ({ const __typeof( ((__typeof
(*next) *)0)->list ) *__mptr = (next->list.next); (__typeof
(*next) *)( (char *)__mptr - __builtin_offsetof(__typeof(*next
), list) );}))
{
2647 list_del(&mapping->list);
2648 amdgpu_vm_it_remove(mapping, &vm->va);
2649 amdgpu_vm_free_mapping(adev, vm, mapping,
2650 bo_va->last_pt_update);
2651 }
2652
2653 dma_fence_put(bo_va->last_pt_update);
2654
2655 if (bo && bo_va->is_xgmi)
2656 amdgpu_xgmi_set_pstate(adev, AMDGPU_XGMI_PSTATE_MIN);
2657
2658 kfree(bo_va);
2659}
2660
2661/**
2662 * amdgpu_vm_evictable - check if we can evict a VM
2663 *
2664 * @bo: A page table of the VM.
2665 *
2666 * Check if it is possible to evict a VM.
2667 */
2668bool_Bool amdgpu_vm_evictable(struct amdgpu_bo *bo)
2669{
2670 struct amdgpu_vm_bo_base *bo_base = bo->vm_bo;
2671
2672 /* Page tables of a destroyed VM can go away immediately */
2673 if (!bo_base || !bo_base->vm)
2674 return true1;
2675
2676 /* Don't evict VM page tables while they are busy */
2677 if (!dma_resv_test_signaled_rcu(bo->tbo.base.resv, true1))
2678 return false0;
2679
2680 /* Try to block ongoing updates */
2681 if (!amdgpu_vm_eviction_trylock(bo_base->vm))
2682 return false0;
2683
2684 /* Don't evict VM page tables while they are updated */
2685 if (!dma_fence_is_signaled(bo_base->vm->last_unlocked)) {
2686 amdgpu_vm_eviction_unlock(bo_base->vm);
2687 return false0;
2688 }
2689
2690 bo_base->vm->evicting = true1;
2691 amdgpu_vm_eviction_unlock(bo_base->vm);
2692 return true1;
2693}
2694
2695/**
2696 * amdgpu_vm_bo_invalidate - mark the bo as invalid
2697 *
2698 * @adev: amdgpu_device pointer
2699 * @bo: amdgpu buffer object
2700 * @evicted: is the BO evicted
2701 *
2702 * Mark @bo as invalid.
2703 */
2704void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
2705 struct amdgpu_bo *bo, bool_Bool evicted)
2706{
2707 struct amdgpu_vm_bo_base *bo_base;
2708
2709 /* shadow bo doesn't have bo base, its validation needs its parent */
2710 if (bo->parent && bo->parent->shadow == bo)
2711 bo = bo->parent;
2712
2713 for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
2714 struct amdgpu_vm *vm = bo_base->vm;
2715
2716 if (evicted && bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv) {
2717 amdgpu_vm_bo_evicted(bo_base);
2718 continue;
2719 }
2720
2721 if (bo_base->moved)
2722 continue;
2723 bo_base->moved = true1;
2724
2725 if (bo->tbo.type == ttm_bo_type_kernel)
2726 amdgpu_vm_bo_relocated(bo_base);
2727 else if (bo->tbo.base.resv == vm->root.base.bo->tbo.base.resv)
2728 amdgpu_vm_bo_moved(bo_base);
2729 else
2730 amdgpu_vm_bo_invalidated(bo_base);
2731 }
2732}
2733
2734/**
2735 * amdgpu_vm_get_block_size - calculate VM page table size as power of two
2736 *
2737 * @vm_size: VM size
2738 *
2739 * Returns:
2740 * VM page table as power of two
2741 */
2742static uint32_t amdgpu_vm_get_block_size(uint64_t vm_size)
2743{
2744 /* Total bits covered by PD + PTs */
2745 unsigned bits = ilog2(vm_size)((sizeof(vm_size) <= 4) ? (fls(vm_size) - 1) : (flsl(vm_size
) - 1))
+ 18;
2746
2747 /* Make sure the PD is 4K in size up to 8GB address space.
2748 Above that split equal between PD and PTs */
2749 if (vm_size <= 8)
2750 return (bits - 9);
2751 else
2752 return ((bits + 3) / 2);
2753}
2754
2755/**
2756 * amdgpu_vm_adjust_size - adjust vm size, block size and fragment size
2757 *
2758 * @adev: amdgpu_device pointer
2759 * @min_vm_size: the minimum vm size in GB if it's set auto
2760 * @fragment_size_default: Default PTE fragment size
2761 * @max_level: max VMPT level
2762 * @max_bits: max address space size in bits
2763 *
2764 */
2765void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
2766 uint32_t fragment_size_default, unsigned max_level,
2767 unsigned max_bits)
2768{
2769 unsigned int max_size = 1 << (max_bits - 30);
2770 unsigned int vm_size;
2771 uint64_t tmp;
2772
2773 /* adjust vm size first */
2774 if (amdgpu_vm_size != -1) {
2775 vm_size = amdgpu_vm_size;
2776 if (vm_size > max_size) {
2777 dev_warn(adev->dev, "VM size (%d) too large, max is %u GB\n",printf("drm:pid%d:%s *WARNING* " "VM size (%d) too large, max is %u GB\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
, max_size)
2778 amdgpu_vm_size, max_size)printf("drm:pid%d:%s *WARNING* " "VM size (%d) too large, max is %u GB\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
, max_size)
;
2779 vm_size = max_size;
2780 }
2781 } else {
2782#ifdef __linux__
2783 struct sysinfo si;
2784#endif
2785 unsigned int phys_ram_gb;
2786
2787 /* Optimal VM size depends on the amount of physical
2788 * RAM available. Underlying requirements and
2789 * assumptions:
2790 *
2791 * - Need to map system memory and VRAM from all GPUs
2792 * - VRAM from other GPUs not known here
2793 * - Assume VRAM <= system memory
2794 * - On GFX8 and older, VM space can be segmented for
2795 * different MTYPEs
2796 * - Need to allow room for fragmentation, guard pages etc.
2797 *
2798 * This adds up to a rough guess of system memory x3.
2799 * Round up to power of two to maximize the available
2800 * VM size with the given page table size.
2801 */
2802#ifdef __linux__
2803 si_meminfo(&si);
2804 phys_ram_gb = ((uint64_t)si.totalram * si.mem_unit +
2805 (1 << 30) - 1) >> 30;
2806#else
2807 phys_ram_gb = ((uint64_t)ptoa(physmem)((paddr_t)(physmem) << 12) +
2808 (1 << 30) - 1) >> 30;
2809#endif
2810 vm_size = roundup_pow_of_two(
2811 min(max(phys_ram_gb * 3, min_vm_size), max_size)((((((phys_ram_gb * 3)>(min_vm_size))?(phys_ram_gb * 3):(min_vm_size
)))<(max_size))?((((phys_ram_gb * 3)>(min_vm_size))?(phys_ram_gb
* 3):(min_vm_size))):(max_size))
);
2812 }
2813
2814 adev->vm_manager.max_pfn = (uint64_t)vm_size << 18;
2815
2816 tmp = roundup_pow_of_two(adev->vm_manager.max_pfn);
2817 if (amdgpu_vm_block_size != -1)
2818 tmp >>= amdgpu_vm_block_size - 9;
2819 tmp = DIV_ROUND_UP(fls64(tmp) - 1, 9)(((fls64(tmp) - 1) + ((9) - 1)) / (9)) - 1;
2820 adev->vm_manager.num_level = min(max_level, (unsigned)tmp)(((max_level)<((unsigned)tmp))?(max_level):((unsigned)tmp)
)
;
2821 switch (adev->vm_manager.num_level) {
2822 case 3:
2823 adev->vm_manager.root_level = AMDGPU_VM_PDB2;
2824 break;
2825 case 2:
2826 adev->vm_manager.root_level = AMDGPU_VM_PDB1;
2827 break;
2828 case 1:
2829 adev->vm_manager.root_level = AMDGPU_VM_PDB0;
2830 break;
2831 default:
2832 dev_err(adev->dev, "VMPT only supports 2~4+1 levels\n")printf("drm:pid%d:%s *ERROR* " "VMPT only supports 2~4+1 levels\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__)
;
2833 }
2834 /* block size depends on vm size and hw setup*/
2835 if (amdgpu_vm_block_size != -1)
2836 adev->vm_manager.block_size =
2837 min((unsigned)amdgpu_vm_block_size, max_bits((((unsigned)amdgpu_vm_block_size)<(max_bits - 12 - 9 * adev
->vm_manager.num_level))?((unsigned)amdgpu_vm_block_size):
(max_bits - 12 - 9 * adev->vm_manager.num_level))
2838 - AMDGPU_GPU_PAGE_SHIFT((((unsigned)amdgpu_vm_block_size)<(max_bits - 12 - 9 * adev
->vm_manager.num_level))?((unsigned)amdgpu_vm_block_size):
(max_bits - 12 - 9 * adev->vm_manager.num_level))
2839 - 9 * adev->vm_manager.num_level)((((unsigned)amdgpu_vm_block_size)<(max_bits - 12 - 9 * adev
->vm_manager.num_level))?((unsigned)amdgpu_vm_block_size):
(max_bits - 12 - 9 * adev->vm_manager.num_level))
;
2840 else if (adev->vm_manager.num_level > 1)
2841 adev->vm_manager.block_size = 9;
2842 else
2843 adev->vm_manager.block_size = amdgpu_vm_get_block_size(tmp);
2844
2845 if (amdgpu_vm_fragment_size == -1)
2846 adev->vm_manager.fragment_size = fragment_size_default;
2847 else
2848 adev->vm_manager.fragment_size = amdgpu_vm_fragment_size;
2849
2850 DRM_INFO("vm size is %u GB, %u levels, block size is %u-bit, fragment size is %u-bit\n",printk("\0016" "[" "drm" "] " "vm size is %u GB, %u levels, block size is %u-bit, fragment size is %u-bit\n"
, vm_size, adev->vm_manager.num_level + 1, adev->vm_manager
.block_size, adev->vm_manager.fragment_size)
2851 vm_size, adev->vm_manager.num_level + 1,printk("\0016" "[" "drm" "] " "vm size is %u GB, %u levels, block size is %u-bit, fragment size is %u-bit\n"
, vm_size, adev->vm_manager.num_level + 1, adev->vm_manager
.block_size, adev->vm_manager.fragment_size)
2852 adev->vm_manager.block_size,printk("\0016" "[" "drm" "] " "vm size is %u GB, %u levels, block size is %u-bit, fragment size is %u-bit\n"
, vm_size, adev->vm_manager.num_level + 1, adev->vm_manager
.block_size, adev->vm_manager.fragment_size)
2853 adev->vm_manager.fragment_size)printk("\0016" "[" "drm" "] " "vm size is %u GB, %u levels, block size is %u-bit, fragment size is %u-bit\n"
, vm_size, adev->vm_manager.num_level + 1, adev->vm_manager
.block_size, adev->vm_manager.fragment_size)
;
2854}
2855
2856/**
2857 * amdgpu_vm_wait_idle - wait for the VM to become idle
2858 *
2859 * @vm: VM object to wait for
2860 * @timeout: timeout to wait for VM to become idle
2861 */
2862long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout)
2863{
2864 timeout = dma_resv_wait_timeout_rcu(vm->root.base.bo->tbo.base.resv,
2865 true1, true1, timeout);
2866 if (timeout <= 0)
2867 return timeout;
2868
2869 return dma_fence_wait_timeout(vm->last_unlocked, true1, timeout);
2870}
2871
2872/**
2873 * amdgpu_vm_init - initialize a vm instance
2874 *
2875 * @adev: amdgpu_device pointer
2876 * @vm: requested vm
2877 * @vm_context: Indicates if it GFX or Compute context
2878 * @pasid: Process address space identifier
2879 *
2880 * Init @vm fields.
2881 *
2882 * Returns:
2883 * 0 for success, error for failure.
2884 */
2885int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
2886 int vm_context, u32 pasid)
2887{
2888 struct amdgpu_bo_param bp;
2889 struct amdgpu_bo *root;
2890 int r, i;
2891
2892 vm->va = RB_ROOT_CACHED(struct rb_root_cached) { ((void *)0) };
2893 for (i = 0; i < AMDGPU_MAX_VMHUBS3; i++)
2894 vm->reserved_vmid[i] = NULL((void *)0);
2895 INIT_LIST_HEAD(&vm->evicted);
2896 INIT_LIST_HEAD(&vm->relocated);
2897 INIT_LIST_HEAD(&vm->moved);
2898 INIT_LIST_HEAD(&vm->idle);
2899 INIT_LIST_HEAD(&vm->invalidated);
2900 mtx_init(&vm->invalidated_lock, IPL_NONE)do { (void)(((void *)0)); (void)(0); __mtx_init((&vm->
invalidated_lock), ((((0x0)) > 0x0 && ((0x0)) <
0x9) ? 0x9 : ((0x0)))); } while (0)
;
2901 INIT_LIST_HEAD(&vm->freed);
2902
2903
2904 /* create scheduler entities for page table updates */
2905 r = drm_sched_entity_init(&vm->immediate, DRM_SCHED_PRIORITY_NORMAL,
2906 adev->vm_manager.vm_pte_scheds,
2907 adev->vm_manager.vm_pte_num_scheds, NULL((void *)0));
2908 if (r)
2909 return r;
2910
2911 r = drm_sched_entity_init(&vm->delayed, DRM_SCHED_PRIORITY_NORMAL,
2912 adev->vm_manager.vm_pte_scheds,
2913 adev->vm_manager.vm_pte_num_scheds, NULL((void *)0));
2914 if (r)
2915 goto error_free_immediate;
2916
2917 vm->pte_support_ats = false0;
2918 vm->is_compute_context = false0;
2919
2920 if (vm_context == AMDGPU_VM_CONTEXT_COMPUTE1) {
2921 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2922 AMDGPU_VM_USE_CPU_FOR_COMPUTE(1 << 1));
2923
2924 if (adev->asic_type == CHIP_RAVEN)
2925 vm->pte_support_ats = true1;
2926 } else {
2927 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2928 AMDGPU_VM_USE_CPU_FOR_GFX(1 << 0));
2929 }
2930 DRM_DEBUG_DRIVER("VM update mode is %s\n",__drm_dbg(DRM_UT_DRIVER, "VM update mode is %s\n", vm->use_cpu_for_update
? "CPU" : "SDMA")
2931 vm->use_cpu_for_update ? "CPU" : "SDMA")__drm_dbg(DRM_UT_DRIVER, "VM update mode is %s\n", vm->use_cpu_for_update
? "CPU" : "SDMA")
;
2932 WARN_ONCE((vm->use_cpu_for_update &&({ static int __warned; int __ret = !!((vm->use_cpu_for_update
&& !amdgpu_gmc_vram_full_visible(&adev->gmc))
); if (__ret && !__warned) { printf("CPU update of VM recommended only for large BAR system\n"
); __warned = 1; } __builtin_expect(!!(__ret), 0); })
2933 !amdgpu_gmc_vram_full_visible(&adev->gmc)),({ static int __warned; int __ret = !!((vm->use_cpu_for_update
&& !amdgpu_gmc_vram_full_visible(&adev->gmc))
); if (__ret && !__warned) { printf("CPU update of VM recommended only for large BAR system\n"
); __warned = 1; } __builtin_expect(!!(__ret), 0); })
2934 "CPU update of VM recommended only for large BAR system\n")({ static int __warned; int __ret = !!((vm->use_cpu_for_update
&& !amdgpu_gmc_vram_full_visible(&adev->gmc))
); if (__ret && !__warned) { printf("CPU update of VM recommended only for large BAR system\n"
); __warned = 1; } __builtin_expect(!!(__ret), 0); })
;
2935
2936 if (vm->use_cpu_for_update)
2937 vm->update_funcs = &amdgpu_vm_cpu_funcs;
2938 else
2939 vm->update_funcs = &amdgpu_vm_sdma_funcs;
2940 vm->last_update = NULL((void *)0);
2941 vm->last_unlocked = dma_fence_get_stub();
2942
2943 rw_init(&vm->eviction_lock, "avmev")_rw_init_flags(&vm->eviction_lock, "avmev", 0, ((void *
)0))
;
2944 vm->evicting = false0;
2945
2946 amdgpu_vm_bo_param(adev, vm, adev->vm_manager.root_level, false0, &bp);
2947 if (vm_context == AMDGPU_VM_CONTEXT_COMPUTE1)
2948 bp.flags &= ~AMDGPU_GEM_CREATE_SHADOW(1 << 4);
2949 r = amdgpu_bo_create(adev, &bp, &root);
2950 if (r)
2951 goto error_free_delayed;
2952
2953 r = amdgpu_bo_reserve(root, true1);
2954 if (r)
2955 goto error_free_root;
2956
2957 r = dma_resv_reserve_shared(root->tbo.base.resv, 1);
2958 if (r)
2959 goto error_unreserve;
2960
2961 amdgpu_vm_bo_base_init(&vm->root.base, vm, root);
2962
2963 r = amdgpu_vm_clear_bo(adev, vm, root, false0);
2964 if (r)
2965 goto error_unreserve;
2966
2967 amdgpu_bo_unreserve(vm->root.base.bo);
2968
2969 if (pasid) {
2970 unsigned long flags;
2971
2972 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags)do { flags = 0; mtx_enter(&adev->vm_manager.pasid_lock
); } while (0)
;
2973 r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1,
2974 GFP_ATOMIC0x0002);
2975 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags)do { (void)(flags); mtx_leave(&adev->vm_manager.pasid_lock
); } while (0)
;
2976 if (r < 0)
2977 goto error_free_root;
2978
2979 vm->pasid = pasid;
2980 }
2981
2982#ifdef __linux__
2983 INIT_KFIFO(vm->faults);
2984#else
2985 SIMPLEQ_INIT(&vm->faults)do { (&vm->faults)->sqh_first = ((void *)0); (&
vm->faults)->sqh_last = &(&vm->faults)->sqh_first
; } while (0)
;
2986#endif
2987
2988 return 0;
2989
2990error_unreserve:
2991 amdgpu_bo_unreserve(vm->root.base.bo);
2992
2993error_free_root:
2994 amdgpu_bo_unref(&vm->root.base.bo->shadow);
2995 amdgpu_bo_unref(&vm->root.base.bo);
2996 vm->root.base.bo = NULL((void *)0);
2997
2998error_free_delayed:
2999 dma_fence_put(vm->last_unlocked);
3000 drm_sched_entity_destroy(&vm->delayed);
3001
3002error_free_immediate:
3003 drm_sched_entity_destroy(&vm->immediate);
3004
3005 return r;
3006}
3007
3008/**
3009 * amdgpu_vm_check_clean_reserved - check if a VM is clean
3010 *
3011 * @adev: amdgpu_device pointer
3012 * @vm: the VM to check
3013 *
3014 * check all entries of the root PD, if any subsequent PDs are allocated,
3015 * it means there are page table creating and filling, and is no a clean
3016 * VM
3017 *
3018 * Returns:
3019 * 0 if this VM is clean
3020 */
3021static int amdgpu_vm_check_clean_reserved(struct amdgpu_device *adev,
3022 struct amdgpu_vm *vm)
3023{
3024 enum amdgpu_vm_level root = adev->vm_manager.root_level;
3025 unsigned int entries = amdgpu_vm_num_entries(adev, root);
3026 unsigned int i = 0;
3027
3028 if (!(vm->root.entries))
3029 return 0;
3030
3031 for (i = 0; i < entries; i++) {
3032 if (vm->root.entries[i].base.bo)
3033 return -EINVAL22;
3034 }
3035
3036 return 0;
3037}
3038
3039/**
3040 * amdgpu_vm_make_compute - Turn a GFX VM into a compute VM
3041 *
3042 * @adev: amdgpu_device pointer
3043 * @vm: requested vm
3044 * @pasid: pasid to use
3045 *
3046 * This only works on GFX VMs that don't have any BOs added and no
3047 * page tables allocated yet.
3048 *
3049 * Changes the following VM parameters:
3050 * - use_cpu_for_update
3051 * - pte_supports_ats
3052 * - pasid (old PASID is released, because compute manages its own PASIDs)
3053 *
3054 * Reinitializes the page directory to reflect the changed ATS
3055 * setting.
3056 *
3057 * Returns:
3058 * 0 for success, -errno for errors.
3059 */
3060int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm,
3061 u32 pasid)
3062{
3063 bool_Bool pte_support_ats = (adev->asic_type == CHIP_RAVEN);
3064 int r;
3065
3066 r = amdgpu_bo_reserve(vm->root.base.bo, true1);
3067 if (r)
3068 return r;
3069
3070 /* Sanity checks */
3071 r = amdgpu_vm_check_clean_reserved(adev, vm);
3072 if (r)
3073 goto unreserve_bo;
3074
3075 if (pasid) {
3076 unsigned long flags;
3077
3078 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags)do { flags = 0; mtx_enter(&adev->vm_manager.pasid_lock
); } while (0)
;
3079 r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1,
3080 GFP_ATOMIC0x0002);
3081 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags)do { (void)(flags); mtx_leave(&adev->vm_manager.pasid_lock
); } while (0)
;
3082
3083 if (r == -ENOSPC28)
3084 goto unreserve_bo;
3085 r = 0;
3086 }
3087
3088 /* Check if PD needs to be reinitialized and do it before
3089 * changing any other state, in case it fails.
3090 */
3091 if (pte_support_ats != vm->pte_support_ats) {
3092 vm->pte_support_ats = pte_support_ats;
3093 r = amdgpu_vm_clear_bo(adev, vm, vm->root.base.bo, false0);
3094 if (r)
3095 goto free_idr;
3096 }
3097
3098 /* Update VM state */
3099 vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
3100 AMDGPU_VM_USE_CPU_FOR_COMPUTE(1 << 1));
3101 DRM_DEBUG_DRIVER("VM update mode is %s\n",__drm_dbg(DRM_UT_DRIVER, "VM update mode is %s\n", vm->use_cpu_for_update
? "CPU" : "SDMA")
3102 vm->use_cpu_for_update ? "CPU" : "SDMA")__drm_dbg(DRM_UT_DRIVER, "VM update mode is %s\n", vm->use_cpu_for_update
? "CPU" : "SDMA")
;
3103 WARN_ONCE((vm->use_cpu_for_update &&({ static int __warned; int __ret = !!((vm->use_cpu_for_update
&& !amdgpu_gmc_vram_full_visible(&adev->gmc))
); if (__ret && !__warned) { printf("CPU update of VM recommended only for large BAR system\n"
); __warned = 1; } __builtin_expect(!!(__ret), 0); })
3104 !amdgpu_gmc_vram_full_visible(&adev->gmc)),({ static int __warned; int __ret = !!((vm->use_cpu_for_update
&& !amdgpu_gmc_vram_full_visible(&adev->gmc))
); if (__ret && !__warned) { printf("CPU update of VM recommended only for large BAR system\n"
); __warned = 1; } __builtin_expect(!!(__ret), 0); })
3105 "CPU update of VM recommended only for large BAR system\n")({ static int __warned; int __ret = !!((vm->use_cpu_for_update
&& !amdgpu_gmc_vram_full_visible(&adev->gmc))
); if (__ret && !__warned) { printf("CPU update of VM recommended only for large BAR system\n"
); __warned = 1; } __builtin_expect(!!(__ret), 0); })
;
3106
3107 if (vm->use_cpu_for_update) {
3108 /* Sync with last SDMA update/clear before switching to CPU */
3109 r = amdgpu_bo_sync_wait(vm->root.base.bo,
3110 AMDGPU_FENCE_OWNER_UNDEFINED((void *)0ul), true1);
3111 if (r)
3112 goto free_idr;
3113
3114 vm->update_funcs = &amdgpu_vm_cpu_funcs;
3115 } else {
3116 vm->update_funcs = &amdgpu_vm_sdma_funcs;
3117 }
3118 dma_fence_put(vm->last_update);
3119 vm->last_update = NULL((void *)0);
3120 vm->is_compute_context = true1;
3121
3122 if (vm->pasid) {
3123 unsigned long flags;
3124
3125 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags)do { flags = 0; mtx_enter(&adev->vm_manager.pasid_lock
); } while (0)
;
3126 idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
3127 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags)do { (void)(flags); mtx_leave(&adev->vm_manager.pasid_lock
); } while (0)
;
3128
3129 /* Free the original amdgpu allocated pasid
3130 * Will be replaced with kfd allocated pasid
3131 */
3132 amdgpu_pasid_free(vm->pasid);
3133 vm->pasid = 0;
3134 }
3135
3136 /* Free the shadow bo for compute VM */
3137 amdgpu_bo_unref(&vm->root.base.bo->shadow);
3138
3139 if (pasid)
3140 vm->pasid = pasid;
3141
3142 goto unreserve_bo;
3143
3144free_idr:
3145 if (pasid) {
3146 unsigned long flags;
3147
3148 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags)do { flags = 0; mtx_enter(&adev->vm_manager.pasid_lock
); } while (0)
;
3149 idr_remove(&adev->vm_manager.pasid_idr, pasid);
3150 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags)do { (void)(flags); mtx_leave(&adev->vm_manager.pasid_lock
); } while (0)
;
3151 }
3152unreserve_bo:
3153 amdgpu_bo_unreserve(vm->root.base.bo);
3154 return r;
3155}
3156
3157/**
3158 * amdgpu_vm_release_compute - release a compute vm
3159 * @adev: amdgpu_device pointer
3160 * @vm: a vm turned into compute vm by calling amdgpu_vm_make_compute
3161 *
3162 * This is a correspondant of amdgpu_vm_make_compute. It decouples compute
3163 * pasid from vm. Compute should stop use of vm after this call.
3164 */
3165void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
3166{
3167 if (vm->pasid) {
3168 unsigned long flags;
3169
3170 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags)do { flags = 0; mtx_enter(&adev->vm_manager.pasid_lock
); } while (0)
;
3171 idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
3172 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags)do { (void)(flags); mtx_leave(&adev->vm_manager.pasid_lock
); } while (0)
;
3173 }
3174 vm->pasid = 0;
3175 vm->is_compute_context = false0;
3176}
3177
3178/**
3179 * amdgpu_vm_fini - tear down a vm instance
3180 *
3181 * @adev: amdgpu_device pointer
3182 * @vm: requested vm
3183 *
3184 * Tear down @vm.
3185 * Unbind the VM and remove all bos from the vm bo list
3186 */
3187void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
3188{
3189 struct amdgpu_bo_va_mapping *mapping, *tmp;
3190 bool_Bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
3191 struct amdgpu_bo *root;
3192 int i;
3193
3194 amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
3195
3196 root = amdgpu_bo_ref(vm->root.base.bo);
3197 amdgpu_bo_reserve(root, true1);
3198 if (vm->pasid) {
3199 unsigned long flags;
3200
3201 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags)do { flags = 0; mtx_enter(&adev->vm_manager.pasid_lock
); } while (0)
;
3202 idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
3203 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags)do { (void)(flags); mtx_leave(&adev->vm_manager.pasid_lock
); } while (0)
;
3204 vm->pasid = 0;
3205 }
3206
3207 dma_fence_wait(vm->last_unlocked, false0);
3208 dma_fence_put(vm->last_unlocked);
3209
3210 list_for_each_entry_safe(mapping, tmp, &vm->freed, list)for (mapping = ({ const __typeof( ((__typeof(*mapping) *)0)->
list ) *__mptr = ((&vm->freed)->next); (__typeof(*mapping
) *)( (char *)__mptr - __builtin_offsetof(__typeof(*mapping),
list) );}), tmp = ({ const __typeof( ((__typeof(*mapping) *)
0)->list ) *__mptr = (mapping->list.next); (__typeof(*mapping
) *)( (char *)__mptr - __builtin_offsetof(__typeof(*mapping),
list) );}); &mapping->list != (&vm->freed); mapping
= tmp, tmp = ({ const __typeof( ((__typeof(*tmp) *)0)->list
) *__mptr = (tmp->list.next); (__typeof(*tmp) *)( (char *
)__mptr - __builtin_offsetof(__typeof(*tmp), list) );}))
{
3211 if (mapping->flags & AMDGPU_PTE_PRT(1ULL << 51) && prt_fini_needed) {
3212 amdgpu_vm_prt_fini(adev, vm);
3213 prt_fini_needed = false0;
3214 }
3215
3216 list_del(&mapping->list);
3217 amdgpu_vm_free_mapping(adev, vm, mapping, NULL((void *)0));
3218 }
3219
3220 amdgpu_vm_free_pts(adev, vm, NULL((void *)0));
3221 amdgpu_bo_unreserve(root);
3222 amdgpu_bo_unref(&root);
3223 WARN_ON(vm->root.base.bo)({ int __ret = !!(vm->root.base.bo); if (__ret) printf("WARNING %s failed at %s:%d\n"
, "vm->root.base.bo", "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_vm.c"
, 3223); __builtin_expect(!!(__ret), 0); })
;
3224
3225 drm_sched_entity_destroy(&vm->immediate);
3226 drm_sched_entity_destroy(&vm->delayed);
3227
3228 if (!RB_EMPTY_ROOT(&vm->va.rb_root)((&vm->va.rb_root)->rb_node == ((void *)0))) {
3229 dev_err(adev->dev, "still active bo inside vm\n")printf("drm:pid%d:%s *ERROR* " "still active bo inside vm\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__)
;
3230 }
3231 rbtree_postorder_for_each_entry_safe(mapping, tmp,for ((mapping) = (__rb_deepest_left((&vm->va.rb_root)->
rb_node) ? ({ const __typeof( ((__typeof(*mapping) *)0)->rb
) *__mptr = (__rb_deepest_left((&vm->va.rb_root)->
rb_node)); (__typeof(*mapping) *)( (char *)__mptr - __builtin_offsetof
(__typeof(*mapping), rb) );}) : ((void *)0)); ((mapping) != (
(void *)0)) && ((tmp) = (rb_next_postorder(&mapping
->rb) ? ({ const __typeof( ((typeof(*mapping) *)0)->rb )
*__mptr = (rb_next_postorder(&mapping->rb)); (typeof(
*mapping) *)( (char *)__mptr - __builtin_offsetof(typeof(*mapping
), rb) );}) : ((void *)0)), 1); (mapping) = (tmp))
3232 &vm->va.rb_root, rb)for ((mapping) = (__rb_deepest_left((&vm->va.rb_root)->
rb_node) ? ({ const __typeof( ((__typeof(*mapping) *)0)->rb
) *__mptr = (__rb_deepest_left((&vm->va.rb_root)->
rb_node)); (__typeof(*mapping) *)( (char *)__mptr - __builtin_offsetof
(__typeof(*mapping), rb) );}) : ((void *)0)); ((mapping) != (
(void *)0)) && ((tmp) = (rb_next_postorder(&mapping
->rb) ? ({ const __typeof( ((typeof(*mapping) *)0)->rb )
*__mptr = (rb_next_postorder(&mapping->rb)); (typeof(
*mapping) *)( (char *)__mptr - __builtin_offsetof(typeof(*mapping
), rb) );}) : ((void *)0)), 1); (mapping) = (tmp))
{
3233 /* Don't remove the mapping here, we don't want to trigger a
3234 * rebalance and the tree is about to be destroyed anyway.
3235 */
3236 list_del(&mapping->list);
3237 kfree(mapping);
3238 }
3239
3240 dma_fence_put(vm->last_update);
3241 for (i = 0; i < AMDGPU_MAX_VMHUBS3; i++)
3242 amdgpu_vmid_free_reserved(adev, vm, i);
3243}
3244
3245/**
3246 * amdgpu_vm_manager_init - init the VM manager
3247 *
3248 * @adev: amdgpu_device pointer
3249 *
3250 * Initialize the VM manager structures
3251 */
3252void amdgpu_vm_manager_init(struct amdgpu_device *adev)
3253{
3254 unsigned i;
3255
3256 /* Concurrent flushes are only possible starting with Vega10 and
3257 * are broken on Navi10 and Navi14.
3258 */
3259 adev->vm_manager.concurrent_flush = !(adev->asic_type < CHIP_VEGA10 ||
3260 adev->asic_type == CHIP_NAVI10 ||
3261 adev->asic_type == CHIP_NAVI14);
3262 amdgpu_vmid_mgr_init(adev);
3263
3264 adev->vm_manager.fence_context =
3265 dma_fence_context_alloc(AMDGPU_MAX_RINGS28);
3266 for (i = 0; i < AMDGPU_MAX_RINGS28; ++i)
3267 adev->vm_manager.seqno[i] = 0;
3268
3269 mtx_init(&adev->vm_manager.prt_lock, IPL_TTY)do { (void)(((void *)0)); (void)(0); __mtx_init((&adev->
vm_manager.prt_lock), ((((0x9)) > 0x0 && ((0x9)) <
0x9) ? 0x9 : ((0x9)))); } while (0)
;
3270 atomic_set(&adev->vm_manager.num_prt_users, 0)({ typeof(*(&adev->vm_manager.num_prt_users)) __tmp = (
(0)); *(volatile typeof(*(&adev->vm_manager.num_prt_users
)) *)&(*(&adev->vm_manager.num_prt_users)) = __tmp
; __tmp; })
;
3271
3272 /* If not overridden by the user, by default, only in large BAR systems
3273 * Compute VM tables will be updated by CPU
3274 */
3275#ifdef CONFIG_X86_641
3276 if (amdgpu_vm_update_mode == -1) {
3277 if (amdgpu_gmc_vram_full_visible(&adev->gmc))
3278 adev->vm_manager.vm_update_mode =
3279 AMDGPU_VM_USE_CPU_FOR_COMPUTE(1 << 1);
3280 else
3281 adev->vm_manager.vm_update_mode = 0;
3282 } else
3283 adev->vm_manager.vm_update_mode = amdgpu_vm_update_mode;
3284#else
3285 adev->vm_manager.vm_update_mode = 0;
3286#endif
3287
3288 idr_init(&adev->vm_manager.pasid_idr);
3289 mtx_init(&adev->vm_manager.pasid_lock, IPL_TTY)do { (void)(((void *)0)); (void)(0); __mtx_init((&adev->
vm_manager.pasid_lock), ((((0x9)) > 0x0 && ((0x9))
< 0x9) ? 0x9 : ((0x9)))); } while (0)
;
3290}
3291
3292/**
3293 * amdgpu_vm_manager_fini - cleanup VM manager
3294 *
3295 * @adev: amdgpu_device pointer
3296 *
3297 * Cleanup the VM manager and free resources.
3298 */
3299void amdgpu_vm_manager_fini(struct amdgpu_device *adev)
3300{
3301 WARN_ON(!idr_is_empty(&adev->vm_manager.pasid_idr))({ int __ret = !!(!idr_is_empty(&adev->vm_manager.pasid_idr
)); if (__ret) printf("WARNING %s failed at %s:%d\n", "!idr_is_empty(&adev->vm_manager.pasid_idr)"
, "/usr/src/sys/dev/pci/drm/amd/amdgpu/amdgpu_vm.c", 3301); __builtin_expect
(!!(__ret), 0); })
;
3302 idr_destroy(&adev->vm_manager.pasid_idr);
3303
3304 amdgpu_vmid_mgr_fini(adev);
3305}
3306
3307/**
3308 * amdgpu_vm_ioctl - Manages VMID reservation for vm hubs.
3309 *
3310 * @dev: drm device pointer
3311 * @data: drm_amdgpu_vm
3312 * @filp: drm file pointer
3313 *
3314 * Returns:
3315 * 0 for success, -errno for errors.
3316 */
3317int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
3318{
3319 union drm_amdgpu_vm *args = data;
3320 struct amdgpu_device *adev = drm_to_adev(dev);
3321 struct amdgpu_fpriv *fpriv = filp->driver_priv;
3322 long timeout = msecs_to_jiffies(2000)(((uint64_t)(2000)) * hz / 1000);
3323 int r;
3324
3325 switch (args->in.op) {
3326 case AMDGPU_VM_OP_RESERVE_VMID1:
3327 /* We only have requirement to reserve vmid from gfxhub */
3328 r = amdgpu_vmid_alloc_reserved(adev, &fpriv->vm,
3329 AMDGPU_GFXHUB_00);
3330 if (r)
3331 return r;
3332 break;
3333 case AMDGPU_VM_OP_UNRESERVE_VMID2:
3334 if (amdgpu_sriov_runtime(adev)((adev)->virt.caps & (1 << 4)))
3335 timeout = 8 * timeout;
3336
3337 /* Wait vm idle to make sure the vmid set in SPM_VMID is
3338 * not referenced anymore.
3339 */
3340 r = amdgpu_bo_reserve(fpriv->vm.root.base.bo, true1);
3341 if (r)
3342 return r;
3343
3344 r = amdgpu_vm_wait_idle(&fpriv->vm, timeout);
3345 if (r < 0)
3346 return r;
3347
3348 amdgpu_bo_unreserve(fpriv->vm.root.base.bo);
3349 amdgpu_vmid_free_reserved(adev, &fpriv->vm, AMDGPU_GFXHUB_00);
3350 break;
3351 default:
3352 return -EINVAL22;
3353 }
3354
3355 return 0;
3356}
3357
3358/**
3359 * amdgpu_vm_get_task_info - Extracts task info for a PASID.
3360 *
3361 * @adev: drm device pointer
3362 * @pasid: PASID identifier for VM
3363 * @task_info: task_info to fill.
3364 */
3365void amdgpu_vm_get_task_info(struct amdgpu_device *adev, u32 pasid,
3366 struct amdgpu_task_info *task_info)
3367{
3368 struct amdgpu_vm *vm;
3369 unsigned long flags;
3370
3371 spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags)do { flags = 0; mtx_enter(&adev->vm_manager.pasid_lock
); } while (0)
;
3372
3373 vm = idr_find(&adev->vm_manager.pasid_idr, pasid);
3374 if (vm)
3375 *task_info = vm->task_info;
3376
3377 spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags)do { (void)(flags); mtx_leave(&adev->vm_manager.pasid_lock
); } while (0)
;
3378}
3379
3380/**
3381 * amdgpu_vm_set_task_info - Sets VMs task info.
3382 *
3383 * @vm: vm for which to set the info
3384 */
3385void amdgpu_vm_set_task_info(struct amdgpu_vm *vm)
3386{
3387 if (vm->task_info.pid)
3388 return;
3389
3390#ifdef __linux__
3391 vm->task_info.pid = current->pid;
3392 get_task_comm(vm->task_info.task_name, current);
3393
3394 if (current->group_leader->mm != current->mm)
3395 return;
3396
3397 vm->task_info.tgid = current->group_leader->pid;
3398 get_task_comm(vm->task_info.process_name, current->group_leader);
3399#else
3400 vm->task_info.pid = curproc({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;
3401 strlcpy(vm->task_info.task_name, curproc({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_comm,
3402 sizeof(vm->task_info.task_name));
3403#endif
3404}
3405
3406/**
3407 * amdgpu_vm_handle_fault - graceful handling of VM faults.
3408 * @adev: amdgpu device pointer
3409 * @pasid: PASID of the VM
3410 * @addr: Address of the fault
3411 *
3412 * Try to gracefully handle a VM fault. Return true if the fault was handled and
3413 * shouldn't be reported any more.
3414 */
3415bool_Bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
3416 uint64_t addr)
3417{
3418 struct amdgpu_bo *root;
3419 uint64_t value, flags;
3420 struct amdgpu_vm *vm;
3421 long r;
3422
3423 spin_lock(&adev->vm_manager.pasid_lock)mtx_enter(&adev->vm_manager.pasid_lock);
3424 vm = idr_find(&adev->vm_manager.pasid_idr, pasid);
3425 if (vm)
1
Assuming 'vm' is non-null
2
Taking true branch
3426 root = amdgpu_bo_ref(vm->root.base.bo);
3427 else
3428 root = NULL((void *)0);
3429 spin_unlock(&adev->vm_manager.pasid_lock)mtx_leave(&adev->vm_manager.pasid_lock);
3430
3431 if (!root)
3
Assuming 'root' is non-null
4
Taking false branch
3432 return false0;
3433
3434 r = amdgpu_bo_reserve(root, true1);
3435 if (r
4.1
'r' is 0
)
5
Taking false branch
3436 goto error_unref;
3437
3438 /* Double check that the VM still exists */
3439 spin_lock(&adev->vm_manager.pasid_lock)mtx_enter(&adev->vm_manager.pasid_lock);
3440 vm = idr_find(&adev->vm_manager.pasid_idr, pasid);
3441 if (vm && vm->root.base.bo != root)
6
Assuming 'vm' is non-null
7
Assuming 'root' is equal to field 'bo'
8
Taking false branch
3442 vm = NULL((void *)0);
3443 spin_unlock(&adev->vm_manager.pasid_lock)mtx_leave(&adev->vm_manager.pasid_lock);
3444 if (!vm
8.1
'vm' is non-null
)
9
Taking false branch
3445 goto error_unlock;
3446
3447 addr /= AMDGPU_GPU_PAGE_SIZE4096;
3448 flags = AMDGPU_PTE_VALID(1ULL << 0) | AMDGPU_PTE_SNOOPED(1ULL << 2) |
3449 AMDGPU_PTE_SYSTEM(1ULL << 1);
3450
3451 if (vm->is_compute_context) {
10
Assuming field 'is_compute_context' is false
11
Taking false branch
3452 /* Intentionally setting invalid PTE flag
3453 * combination to force a no-retry-fault
3454 */
3455 flags = AMDGPU_PTE_EXECUTABLE(1ULL << 4) | AMDGPU_PDE_PTE(1ULL << 54) |
3456 AMDGPU_PTE_TF(1ULL << 56);
3457 value = 0;
3458
3459 } else if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_NEVER0) {
12
Assuming 'amdgpu_vm_fault_stop' is not equal to AMDGPU_VM_FAULT_STOP_NEVER
13
Taking false branch
3460 /* Redirect the access to the dummy page */
3461 value = adev->dummy_page_addr;
3462 flags |= AMDGPU_PTE_EXECUTABLE(1ULL << 4) | AMDGPU_PTE_READABLE(1ULL << 5) |
3463 AMDGPU_PTE_WRITEABLE(1ULL << 6);
3464
3465 } else {
3466 /* Let the hw retry silently on the PTE */
3467 value = 0;
3468 }
3469
3470 r = amdgpu_vm_bo_update_mapping(adev, vm, true1, false0, NULL((void *)0), addr,
14
Calling 'amdgpu_vm_bo_update_mapping'
3471 addr + 1, flags, value, NULL((void *)0), NULL((void *)0));
3472 if (r)
3473 goto error_unlock;
3474
3475 r = amdgpu_vm_update_pdes(adev, vm, true1);
3476
3477error_unlock:
3478 amdgpu_bo_unreserve(root);
3479 if (r < 0)
3480 DRM_ERROR("Can't handle page fault (%ld)\n", r)__drm_err("Can't handle page fault (%ld)\n", r);
3481
3482error_unref:
3483 amdgpu_bo_unref(&root);
3484
3485 return false0;
3486}