File: | dev/pci/drm/i915/i915_perf.c |
Warning: | line 701, column 8 Although the value stored to 'taken' is used in the enclosing expression, the value is never actually read from 'taken' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | /* |
2 | * Copyright © 2015-2016 Intel Corporation |
3 | * |
4 | * Permission is hereby granted, free of charge, to any person obtaining a |
5 | * copy of this software and associated documentation files (the "Software"), |
6 | * to deal in the Software without restriction, including without limitation |
7 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, |
8 | * and/or sell copies of the Software, and to permit persons to whom the |
9 | * Software is furnished to do so, subject to the following conditions: |
10 | * |
11 | * The above copyright notice and this permission notice (including the next |
12 | * paragraph) shall be included in all copies or substantial portions of the |
13 | * Software. |
14 | * |
15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
16 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
17 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
18 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
19 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING |
20 | * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS |
21 | * IN THE SOFTWARE. |
22 | * |
23 | * Authors: |
24 | * Robert Bragg <robert@sixbynine.org> |
25 | */ |
26 | |
27 | |
28 | /** |
29 | * DOC: i915 Perf Overview |
30 | * |
31 | * Gen graphics supports a large number of performance counters that can help |
32 | * driver and application developers understand and optimize their use of the |
33 | * GPU. |
34 | * |
35 | * This i915 perf interface enables userspace to configure and open a file |
36 | * descriptor representing a stream of GPU metrics which can then be read() as |
37 | * a stream of sample records. |
38 | * |
39 | * The interface is particularly suited to exposing buffered metrics that are |
40 | * captured by DMA from the GPU, unsynchronized with and unrelated to the CPU. |
41 | * |
42 | * Streams representing a single context are accessible to applications with a |
43 | * corresponding drm file descriptor, such that OpenGL can use the interface |
44 | * without special privileges. Access to system-wide metrics requires root |
45 | * privileges by default, unless changed via the dev.i915.perf_event_paranoid |
46 | * sysctl option. |
47 | * |
48 | */ |
49 | |
50 | /** |
51 | * DOC: i915 Perf History and Comparison with Core Perf |
52 | * |
53 | * The interface was initially inspired by the core Perf infrastructure but |
54 | * some notable differences are: |
55 | * |
56 | * i915 perf file descriptors represent a "stream" instead of an "event"; where |
57 | * a perf event primarily corresponds to a single 64bit value, while a stream |
58 | * might sample sets of tightly-coupled counters, depending on the |
59 | * configuration. For example the Gen OA unit isn't designed to support |
60 | * orthogonal configurations of individual counters; it's configured for a set |
61 | * of related counters. Samples for an i915 perf stream capturing OA metrics |
62 | * will include a set of counter values packed in a compact HW specific format. |
63 | * The OA unit supports a number of different packing formats which can be |
64 | * selected by the user opening the stream. Perf has support for grouping |
65 | * events, but each event in the group is configured, validated and |
66 | * authenticated individually with separate system calls. |
67 | * |
68 | * i915 perf stream configurations are provided as an array of u64 (key,value) |
69 | * pairs, instead of a fixed struct with multiple miscellaneous config members, |
70 | * interleaved with event-type specific members. |
71 | * |
72 | * i915 perf doesn't support exposing metrics via an mmap'd circular buffer. |
73 | * The supported metrics are being written to memory by the GPU unsynchronized |
74 | * with the CPU, using HW specific packing formats for counter sets. Sometimes |
75 | * the constraints on HW configuration require reports to be filtered before it |
76 | * would be acceptable to expose them to unprivileged applications - to hide |
77 | * the metrics of other processes/contexts. For these use cases a read() based |
78 | * interface is a good fit, and provides an opportunity to filter data as it |
79 | * gets copied from the GPU mapped buffers to userspace buffers. |
80 | * |
81 | * |
82 | * Issues hit with first prototype based on Core Perf |
83 | * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
84 | * |
85 | * The first prototype of this driver was based on the core perf |
86 | * infrastructure, and while we did make that mostly work, with some changes to |
87 | * perf, we found we were breaking or working around too many assumptions baked |
88 | * into perf's currently cpu centric design. |
89 | * |
90 | * In the end we didn't see a clear benefit to making perf's implementation and |
91 | * interface more complex by changing design assumptions while we knew we still |
92 | * wouldn't be able to use any existing perf based userspace tools. |
93 | * |
94 | * Also considering the Gen specific nature of the Observability hardware and |
95 | * how userspace will sometimes need to combine i915 perf OA metrics with |
96 | * side-band OA data captured via MI_REPORT_PERF_COUNT commands; we're |
97 | * expecting the interface to be used by a platform specific userspace such as |
98 | * OpenGL or tools. This is to say; we aren't inherently missing out on having |
99 | * a standard vendor/architecture agnostic interface by not using perf. |
100 | * |
101 | * |
102 | * For posterity, in case we might re-visit trying to adapt core perf to be |
103 | * better suited to exposing i915 metrics these were the main pain points we |
104 | * hit: |
105 | * |
106 | * - The perf based OA PMU driver broke some significant design assumptions: |
107 | * |
108 | * Existing perf pmus are used for profiling work on a cpu and we were |
109 | * introducing the idea of _IS_DEVICE pmus with different security |
110 | * implications, the need to fake cpu-related data (such as user/kernel |
111 | * registers) to fit with perf's current design, and adding _DEVICE records |
112 | * as a way to forward device-specific status records. |
113 | * |
114 | * The OA unit writes reports of counters into a circular buffer, without |
115 | * involvement from the CPU, making our PMU driver the first of a kind. |
116 | * |
117 | * Given the way we were periodically forward data from the GPU-mapped, OA |
118 | * buffer to perf's buffer, those bursts of sample writes looked to perf like |
119 | * we were sampling too fast and so we had to subvert its throttling checks. |
120 | * |
121 | * Perf supports groups of counters and allows those to be read via |
122 | * transactions internally but transactions currently seem designed to be |
123 | * explicitly initiated from the cpu (say in response to a userspace read()) |
124 | * and while we could pull a report out of the OA buffer we can't |
125 | * trigger a report from the cpu on demand. |
126 | * |
127 | * Related to being report based; the OA counters are configured in HW as a |
128 | * set while perf generally expects counter configurations to be orthogonal. |
129 | * Although counters can be associated with a group leader as they are |
130 | * opened, there's no clear precedent for being able to provide group-wide |
131 | * configuration attributes (for example we want to let userspace choose the |
132 | * OA unit report format used to capture all counters in a set, or specify a |
133 | * GPU context to filter metrics on). We avoided using perf's grouping |
134 | * feature and forwarded OA reports to userspace via perf's 'raw' sample |
135 | * field. This suited our userspace well considering how coupled the counters |
136 | * are when dealing with normalizing. It would be inconvenient to split |
137 | * counters up into separate events, only to require userspace to recombine |
138 | * them. For Mesa it's also convenient to be forwarded raw, periodic reports |
139 | * for combining with the side-band raw reports it captures using |
140 | * MI_REPORT_PERF_COUNT commands. |
141 | * |
142 | * - As a side note on perf's grouping feature; there was also some concern |
143 | * that using PERF_FORMAT_GROUP as a way to pack together counter values |
144 | * would quite drastically inflate our sample sizes, which would likely |
145 | * lower the effective sampling resolutions we could use when the available |
146 | * memory bandwidth is limited. |
147 | * |
148 | * With the OA unit's report formats, counters are packed together as 32 |
149 | * or 40bit values, with the largest report size being 256 bytes. |
150 | * |
151 | * PERF_FORMAT_GROUP values are 64bit, but there doesn't appear to be a |
152 | * documented ordering to the values, implying PERF_FORMAT_ID must also be |
153 | * used to add a 64bit ID before each value; giving 16 bytes per counter. |
154 | * |
155 | * Related to counter orthogonality; we can't time share the OA unit, while |
156 | * event scheduling is a central design idea within perf for allowing |
157 | * userspace to open + enable more events than can be configured in HW at any |
158 | * one time. The OA unit is not designed to allow re-configuration while in |
159 | * use. We can't reconfigure the OA unit without losing internal OA unit |
160 | * state which we can't access explicitly to save and restore. Reconfiguring |
161 | * the OA unit is also relatively slow, involving ~100 register writes. From |
162 | * userspace Mesa also depends on a stable OA configuration when emitting |
163 | * MI_REPORT_PERF_COUNT commands and importantly the OA unit can't be |
164 | * disabled while there are outstanding MI_RPC commands lest we hang the |
165 | * command streamer. |
166 | * |
167 | * The contents of sample records aren't extensible by device drivers (i.e. |
168 | * the sample_type bits). As an example; Sourab Gupta had been looking to |
169 | * attach GPU timestamps to our OA samples. We were shoehorning OA reports |
170 | * into sample records by using the 'raw' field, but it's tricky to pack more |
171 | * than one thing into this field because events/core.c currently only lets a |
172 | * pmu give a single raw data pointer plus len which will be copied into the |
173 | * ring buffer. To include more than the OA report we'd have to copy the |
174 | * report into an intermediate larger buffer. I'd been considering allowing a |
175 | * vector of data+len values to be specified for copying the raw data, but |
176 | * it felt like a kludge to being using the raw field for this purpose. |
177 | * |
178 | * - It felt like our perf based PMU was making some technical compromises |
179 | * just for the sake of using perf: |
180 | * |
181 | * perf_event_open() requires events to either relate to a pid or a specific |
182 | * cpu core, while our device pmu related to neither. Events opened with a |
183 | * pid will be automatically enabled/disabled according to the scheduling of |
184 | * that process - so not appropriate for us. When an event is related to a |
185 | * cpu id, perf ensures pmu methods will be invoked via an inter process |
186 | * interrupt on that core. To avoid invasive changes our userspace opened OA |
187 | * perf events for a specific cpu. This was workable but it meant the |
188 | * majority of the OA driver ran in atomic context, including all OA report |
189 | * forwarding, which wasn't really necessary in our case and seems to make |
190 | * our locking requirements somewhat complex as we handled the interaction |
191 | * with the rest of the i915 driver. |
192 | */ |
193 | |
194 | #include <linux/anon_inodes.h> |
195 | #include <linux/sizes.h> |
196 | #include <linux/uuid.h> |
197 | |
198 | #include "gem/i915_gem_context.h" |
199 | #include "gt/intel_engine_pm.h" |
200 | #include "gt/intel_engine_user.h" |
201 | #include "gt/intel_gt.h" |
202 | #include "gt/intel_lrc_reg.h" |
203 | #include "gt/intel_ring.h" |
204 | |
205 | #include "i915_drv.h" |
206 | #include "i915_perf.h" |
207 | |
208 | /* HW requires this to be a power of two, between 128k and 16M, though driver |
209 | * is currently generally designed assuming the largest 16M size is used such |
210 | * that the overflow cases are unlikely in normal operation. |
211 | */ |
212 | #define OA_BUFFER_SIZE(1024 * 1024 * 16) SZ_16M(1024 * 1024 * 16) |
213 | |
214 | #define OA_TAKEN(tail, head)((tail - head) & ((1024 * 1024 * 16) - 1)) ((tail - head) & (OA_BUFFER_SIZE(1024 * 1024 * 16) - 1)) |
215 | |
216 | /** |
217 | * DOC: OA Tail Pointer Race |
218 | * |
219 | * There's a HW race condition between OA unit tail pointer register updates and |
220 | * writes to memory whereby the tail pointer can sometimes get ahead of what's |
221 | * been written out to the OA buffer so far (in terms of what's visible to the |
222 | * CPU). |
223 | * |
224 | * Although this can be observed explicitly while copying reports to userspace |
225 | * by checking for a zeroed report-id field in tail reports, we want to account |
226 | * for this earlier, as part of the oa_buffer_check_unlocked to avoid lots of |
227 | * redundant read() attempts. |
228 | * |
229 | * We workaround this issue in oa_buffer_check_unlocked() by reading the reports |
230 | * in the OA buffer, starting from the tail reported by the HW until we find a |
231 | * report with its first 2 dwords not 0 meaning its previous report is |
232 | * completely in memory and ready to be read. Those dwords are also set to 0 |
233 | * once read and the whole buffer is cleared upon OA buffer initialization. The |
234 | * first dword is the reason for this report while the second is the timestamp, |
235 | * making the chances of having those 2 fields at 0 fairly unlikely. A more |
236 | * detailed explanation is available in oa_buffer_check_unlocked(). |
237 | * |
238 | * Most of the implementation details for this workaround are in |
239 | * oa_buffer_check_unlocked() and _append_oa_reports() |
240 | * |
241 | * Note for posterity: previously the driver used to define an effective tail |
242 | * pointer that lagged the real pointer by a 'tail margin' measured in bytes |
243 | * derived from %OA_TAIL_MARGIN_NSEC and the configured sampling frequency. |
244 | * This was flawed considering that the OA unit may also automatically generate |
245 | * non-periodic reports (such as on context switch) or the OA unit may be |
246 | * enabled without any periodic sampling. |
247 | */ |
248 | #define OA_TAIL_MARGIN_NSEC100000ULL 100000ULL |
249 | #define INVALID_TAIL_PTR0xffffffff 0xffffffff |
250 | |
251 | /* The default frequency for checking whether the OA unit has written new |
252 | * reports to the circular OA buffer... |
253 | */ |
254 | #define DEFAULT_POLL_FREQUENCY_HZ200 200 |
255 | #define DEFAULT_POLL_PERIOD_NS(1000000000L / 200) (NSEC_PER_SEC1000000000L / DEFAULT_POLL_FREQUENCY_HZ200) |
256 | |
257 | /* for sysctl proc_dointvec_minmax of dev.i915.perf_stream_paranoid */ |
258 | static u32 i915_perf_stream_paranoid = true1; |
259 | |
260 | /* The maximum exponent the hardware accepts is 63 (essentially it selects one |
261 | * of the 64bit timestamp bits to trigger reports from) but there's currently |
262 | * no known use case for sampling as infrequently as once per 47 thousand years. |
263 | * |
264 | * Since the timestamps included in OA reports are only 32bits it seems |
265 | * reasonable to limit the OA exponent where it's still possible to account for |
266 | * overflow in OA report timestamps. |
267 | */ |
268 | #define OA_EXPONENT_MAX31 31 |
269 | |
270 | #define INVALID_CTX_ID0xffffffff 0xffffffff |
271 | |
272 | /* On Gen8+ automatically triggered OA reports include a 'reason' field... */ |
273 | #define OAREPORT_REASON_MASK0x3f 0x3f |
274 | #define OAREPORT_REASON_MASK_EXTENDED0x7f 0x7f |
275 | #define OAREPORT_REASON_SHIFT19 19 |
276 | #define OAREPORT_REASON_TIMER(1<<0) (1<<0) |
277 | #define OAREPORT_REASON_CTX_SWITCH(1<<3) (1<<3) |
278 | #define OAREPORT_REASON_CLK_RATIO(1<<5) (1<<5) |
279 | |
280 | |
281 | /* For sysctl proc_dointvec_minmax of i915_oa_max_sample_rate |
282 | * |
283 | * The highest sampling frequency we can theoretically program the OA unit |
284 | * with is always half the timestamp frequency: E.g. 6.25Mhz for Haswell. |
285 | * |
286 | * Initialized just before we register the sysctl parameter. |
287 | */ |
288 | static int oa_sample_rate_hard_limit; |
289 | |
290 | /* Theoretically we can program the OA unit to sample every 160ns but don't |
291 | * allow that by default unless root... |
292 | * |
293 | * The default threshold of 100000Hz is based on perf's similar |
294 | * kernel.perf_event_max_sample_rate sysctl parameter. |
295 | */ |
296 | static u32 i915_oa_max_sample_rate = 100000; |
297 | |
298 | /* XXX: beware if future OA HW adds new report formats that the current |
299 | * code assumes all reports have a power-of-two size and ~(size - 1) can |
300 | * be used as a mask to align the OA tail pointer. |
301 | */ |
302 | static const struct i915_oa_format hsw_oa_formats[I915_OA_FORMAT_MAX] = { |
303 | [I915_OA_FORMAT_A13] = { 0, 64 }, |
304 | [I915_OA_FORMAT_A29] = { 1, 128 }, |
305 | [I915_OA_FORMAT_A13_B8_C8] = { 2, 128 }, |
306 | /* A29_B8_C8 Disallowed as 192 bytes doesn't factor into buffer size */ |
307 | [I915_OA_FORMAT_B4_C8] = { 4, 64 }, |
308 | [I915_OA_FORMAT_A45_B8_C8] = { 5, 256 }, |
309 | [I915_OA_FORMAT_B4_C8_A16] = { 6, 128 }, |
310 | [I915_OA_FORMAT_C4_B8] = { 7, 64 }, |
311 | }; |
312 | |
313 | static const struct i915_oa_format gen8_plus_oa_formats[I915_OA_FORMAT_MAX] = { |
314 | [I915_OA_FORMAT_A12] = { 0, 64 }, |
315 | [I915_OA_FORMAT_A12_B8_C8] = { 2, 128 }, |
316 | [I915_OA_FORMAT_A32u40_A4u32_B8_C8] = { 5, 256 }, |
317 | [I915_OA_FORMAT_C4_B8] = { 7, 64 }, |
318 | }; |
319 | |
320 | static const struct i915_oa_format gen12_oa_formats[I915_OA_FORMAT_MAX] = { |
321 | [I915_OA_FORMAT_A32u40_A4u32_B8_C8] = { 5, 256 }, |
322 | }; |
323 | |
324 | #define SAMPLE_OA_REPORT(1<<0) (1<<0) |
325 | |
326 | /** |
327 | * struct perf_open_properties - for validated properties given to open a stream |
328 | * @sample_flags: `DRM_I915_PERF_PROP_SAMPLE_*` properties are tracked as flags |
329 | * @single_context: Whether a single or all gpu contexts should be monitored |
330 | * @hold_preemption: Whether the preemption is disabled for the filtered |
331 | * context |
332 | * @ctx_handle: A gem ctx handle for use with @single_context |
333 | * @metrics_set: An ID for an OA unit metric set advertised via sysfs |
334 | * @oa_format: An OA unit HW report format |
335 | * @oa_periodic: Whether to enable periodic OA unit sampling |
336 | * @oa_period_exponent: The OA unit sampling period is derived from this |
337 | * @engine: The engine (typically rcs0) being monitored by the OA unit |
338 | * @has_sseu: Whether @sseu was specified by userspace |
339 | * @sseu: internal SSEU configuration computed either from the userspace |
340 | * specified configuration in the opening parameters or a default value |
341 | * (see get_default_sseu_config()) |
342 | * @poll_oa_period: The period in nanoseconds at which the CPU will check for OA |
343 | * data availability |
344 | * |
345 | * As read_properties_unlocked() enumerates and validates the properties given |
346 | * to open a stream of metrics the configuration is built up in the structure |
347 | * which starts out zero initialized. |
348 | */ |
349 | struct perf_open_properties { |
350 | u32 sample_flags; |
351 | |
352 | u64 single_context:1; |
353 | u64 hold_preemption:1; |
354 | u64 ctx_handle; |
355 | |
356 | /* OA sampling state */ |
357 | int metrics_set; |
358 | int oa_format; |
359 | bool_Bool oa_periodic; |
360 | int oa_period_exponent; |
361 | |
362 | struct intel_engine_cs *engine; |
363 | |
364 | bool_Bool has_sseu; |
365 | struct intel_sseu sseu; |
366 | |
367 | u64 poll_oa_period; |
368 | }; |
369 | |
370 | struct i915_oa_config_bo { |
371 | struct llist_node node; |
372 | |
373 | struct i915_oa_config *oa_config; |
374 | struct i915_vma *vma; |
375 | }; |
376 | |
377 | static struct ctl_table_header *sysctl_header; |
378 | |
379 | static enum hrtimer_restart oa_poll_check_timer_cb(struct hrtimer *hrtimer); |
380 | |
381 | void i915_oa_config_release(struct kref *ref) |
382 | { |
383 | struct i915_oa_config *oa_config = |
384 | container_of(ref, typeof(*oa_config), ref)({ const __typeof( ((typeof(*oa_config) *)0)->ref ) *__mptr = (ref); (typeof(*oa_config) *)( (char *)__mptr - __builtin_offsetof (typeof(*oa_config), ref) );}); |
385 | |
386 | kfree(oa_config->flex_regs); |
387 | kfree(oa_config->b_counter_regs); |
388 | kfree(oa_config->mux_regs); |
389 | |
390 | kfree_rcu(oa_config, rcu)do { free((void *)oa_config, 145, 0); } while(0); |
391 | } |
392 | |
393 | struct i915_oa_config * |
394 | i915_perf_get_oa_config(struct i915_perf *perf, int metrics_set) |
395 | { |
396 | struct i915_oa_config *oa_config; |
397 | |
398 | rcu_read_lock(); |
399 | oa_config = idr_find(&perf->metrics_idr, metrics_set); |
400 | if (oa_config) |
401 | oa_config = i915_oa_config_get(oa_config); |
402 | rcu_read_unlock(); |
403 | |
404 | return oa_config; |
405 | } |
406 | |
407 | static void free_oa_config_bo(struct i915_oa_config_bo *oa_bo) |
408 | { |
409 | i915_oa_config_put(oa_bo->oa_config); |
410 | i915_vma_put(oa_bo->vma); |
411 | kfree(oa_bo); |
412 | } |
413 | |
414 | static u32 gen12_oa_hw_tail_read(struct i915_perf_stream *stream) |
415 | { |
416 | struct intel_uncore *uncore = stream->uncore; |
417 | |
418 | return intel_uncore_read(uncore, GEN12_OAG_OATAILPTR((const i915_reg_t){ .reg = (0xdb04) })) & |
419 | GEN12_OAG_OATAILPTR_MASK0xffffffc0; |
420 | } |
421 | |
422 | static u32 gen8_oa_hw_tail_read(struct i915_perf_stream *stream) |
423 | { |
424 | struct intel_uncore *uncore = stream->uncore; |
425 | |
426 | return intel_uncore_read(uncore, GEN8_OATAILPTR((const i915_reg_t){ .reg = (0x2B10) })) & GEN8_OATAILPTR_MASK0xffffffc0; |
427 | } |
428 | |
429 | static u32 gen7_oa_hw_tail_read(struct i915_perf_stream *stream) |
430 | { |
431 | struct intel_uncore *uncore = stream->uncore; |
432 | u32 oastatus1 = intel_uncore_read(uncore, GEN7_OASTATUS1((const i915_reg_t){ .reg = (0x2364) })); |
433 | |
434 | return oastatus1 & GEN7_OASTATUS1_TAIL_MASK0xffffffc0; |
435 | } |
436 | |
437 | /** |
438 | * oa_buffer_check_unlocked - check for data and update tail ptr state |
439 | * @stream: i915 stream instance |
440 | * |
441 | * This is either called via fops (for blocking reads in user ctx) or the poll |
442 | * check hrtimer (atomic ctx) to check the OA buffer tail pointer and check |
443 | * if there is data available for userspace to read. |
444 | * |
445 | * This function is central to providing a workaround for the OA unit tail |
446 | * pointer having a race with respect to what data is visible to the CPU. |
447 | * It is responsible for reading tail pointers from the hardware and giving |
448 | * the pointers time to 'age' before they are made available for reading. |
449 | * (See description of OA_TAIL_MARGIN_NSEC above for further details.) |
450 | * |
451 | * Besides returning true when there is data available to read() this function |
452 | * also updates the tail, aging_tail and aging_timestamp in the oa_buffer |
453 | * object. |
454 | * |
455 | * Note: It's safe to read OA config state here unlocked, assuming that this is |
456 | * only called while the stream is enabled, while the global OA configuration |
457 | * can't be modified. |
458 | * |
459 | * Returns: %true if the OA buffer contains data, else %false |
460 | */ |
461 | static bool_Bool oa_buffer_check_unlocked(struct i915_perf_stream *stream) |
462 | { |
463 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
464 | return false0; |
465 | #ifdef notyet |
466 | u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma); |
467 | int report_size = stream->oa_buffer.format_size; |
468 | unsigned long flags; |
469 | bool_Bool pollin; |
470 | u32 hw_tail; |
471 | u64 now; |
472 | |
473 | /* We have to consider the (unlikely) possibility that read() errors |
474 | * could result in an OA buffer reset which might reset the head and |
475 | * tail state. |
476 | */ |
477 | spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags)do { flags = 0; mtx_enter(&stream->oa_buffer.ptr_lock) ; } while (0); |
478 | |
479 | hw_tail = stream->perf->ops.oa_hw_tail_read(stream); |
480 | |
481 | /* The tail pointer increases in 64 byte increments, |
482 | * not in report_size steps... |
483 | */ |
484 | hw_tail &= ~(report_size - 1); |
485 | |
486 | now = ktime_get_mono_fast_ns(); |
487 | |
488 | if (hw_tail == stream->oa_buffer.aging_tail && |
489 | (now - stream->oa_buffer.aging_timestamp) > OA_TAIL_MARGIN_NSEC100000ULL) { |
490 | /* If the HW tail hasn't move since the last check and the HW |
491 | * tail has been aging for long enough, declare it the new |
492 | * tail. |
493 | */ |
494 | stream->oa_buffer.tail = stream->oa_buffer.aging_tail; |
495 | } else { |
496 | u32 head, tail, aged_tail; |
497 | |
498 | /* NB: The head we observe here might effectively be a little |
499 | * out of date. If a read() is in progress, the head could be |
500 | * anywhere between this head and stream->oa_buffer.tail. |
501 | */ |
502 | head = stream->oa_buffer.head - gtt_offset; |
503 | aged_tail = stream->oa_buffer.tail - gtt_offset; |
504 | |
505 | hw_tail -= gtt_offset; |
506 | tail = hw_tail; |
507 | |
508 | /* Walk the stream backward until we find a report with dword 0 |
509 | * & 1 not at 0. Since the circular buffer pointers progress by |
510 | * increments of 64 bytes and that reports can be up to 256 |
511 | * bytes long, we can't tell whether a report has fully landed |
512 | * in memory before the first 2 dwords of the following report |
513 | * have effectively landed. |
514 | * |
515 | * This is assuming that the writes of the OA unit land in |
516 | * memory in the order they were written to. |
517 | * If not : (╯°□°)╯︵ ┻━┻ |
518 | */ |
519 | while (OA_TAKEN(tail, aged_tail)((tail - aged_tail) & ((1024 * 1024 * 16) - 1)) >= report_size) { |
520 | u32 *report32 = (void *)(stream->oa_buffer.vaddr + tail); |
521 | |
522 | if (report32[0] != 0 || report32[1] != 0) |
523 | break; |
524 | |
525 | tail = (tail - report_size) & (OA_BUFFER_SIZE(1024 * 1024 * 16) - 1); |
526 | } |
527 | |
528 | if (OA_TAKEN(hw_tail, tail)((hw_tail - tail) & ((1024 * 1024 * 16) - 1)) > report_size && |
529 | __ratelimit(&stream->perf->tail_pointer_race)(1)) |
530 | DRM_NOTE("unlanded report(s) head=0x%x "printk("\0015" "[" "drm" "] " "unlanded report(s) head=0x%x " "tail=0x%x hw_tail=0x%x\n", head, tail, hw_tail) |
531 | "tail=0x%x hw_tail=0x%x\n",printk("\0015" "[" "drm" "] " "unlanded report(s) head=0x%x " "tail=0x%x hw_tail=0x%x\n", head, tail, hw_tail) |
532 | head, tail, hw_tail)printk("\0015" "[" "drm" "] " "unlanded report(s) head=0x%x " "tail=0x%x hw_tail=0x%x\n", head, tail, hw_tail); |
533 | |
534 | stream->oa_buffer.tail = gtt_offset + tail; |
535 | stream->oa_buffer.aging_tail = gtt_offset + hw_tail; |
536 | stream->oa_buffer.aging_timestamp = now; |
537 | } |
538 | |
539 | pollin = OA_TAKEN(stream->oa_buffer.tail - gtt_offset,((stream->oa_buffer.tail - gtt_offset - stream->oa_buffer .head - gtt_offset) & ((1024 * 1024 * 16) - 1)) |
540 | stream->oa_buffer.head - gtt_offset)((stream->oa_buffer.tail - gtt_offset - stream->oa_buffer .head - gtt_offset) & ((1024 * 1024 * 16) - 1)) >= report_size; |
541 | |
542 | spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags)do { (void)(flags); mtx_leave(&stream->oa_buffer.ptr_lock ); } while (0); |
543 | |
544 | return pollin; |
545 | #endif |
546 | } |
547 | |
548 | /** |
549 | * append_oa_status - Appends a status record to a userspace read() buffer. |
550 | * @stream: An i915-perf stream opened for OA metrics |
551 | * @buf: destination buffer given by userspace |
552 | * @count: the number of bytes userspace wants to read |
553 | * @offset: (inout): the current position for writing into @buf |
554 | * @type: The kind of status to report to userspace |
555 | * |
556 | * Writes a status record (such as `DRM_I915_PERF_RECORD_OA_REPORT_LOST`) |
557 | * into the userspace read() buffer. |
558 | * |
559 | * The @buf @offset will only be updated on success. |
560 | * |
561 | * Returns: 0 on success, negative error code on failure. |
562 | */ |
563 | static int append_oa_status(struct i915_perf_stream *stream, |
564 | char __user *buf, |
565 | size_t count, |
566 | size_t *offset, |
567 | enum drm_i915_perf_record_type type) |
568 | { |
569 | struct drm_i915_perf_record_header header = { type, 0, sizeof(header) }; |
570 | |
571 | if ((count - *offset) < header.size) |
572 | return -ENOSPC28; |
573 | |
574 | if (copy_to_user(buf + *offset, &header, sizeof(header))) |
575 | return -EFAULT14; |
576 | |
577 | (*offset) += header.size; |
578 | |
579 | return 0; |
580 | } |
581 | |
582 | /** |
583 | * append_oa_sample - Copies single OA report into userspace read() buffer. |
584 | * @stream: An i915-perf stream opened for OA metrics |
585 | * @buf: destination buffer given by userspace |
586 | * @count: the number of bytes userspace wants to read |
587 | * @offset: (inout): the current position for writing into @buf |
588 | * @report: A single OA report to (optionally) include as part of the sample |
589 | * |
590 | * The contents of a sample are configured through `DRM_I915_PERF_PROP_SAMPLE_*` |
591 | * properties when opening a stream, tracked as `stream->sample_flags`. This |
592 | * function copies the requested components of a single sample to the given |
593 | * read() @buf. |
594 | * |
595 | * The @buf @offset will only be updated on success. |
596 | * |
597 | * Returns: 0 on success, negative error code on failure. |
598 | */ |
599 | static int append_oa_sample(struct i915_perf_stream *stream, |
600 | char __user *buf, |
601 | size_t count, |
602 | size_t *offset, |
603 | const u8 *report) |
604 | { |
605 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
606 | return false0; |
607 | #ifdef notyet |
608 | int report_size = stream->oa_buffer.format_size; |
609 | struct drm_i915_perf_record_header header; |
610 | |
611 | header.type = DRM_I915_PERF_RECORD_SAMPLE; |
612 | header.pad = 0; |
613 | header.size = stream->sample_size; |
614 | |
615 | if ((count - *offset) < header.size) |
616 | return -ENOSPC28; |
617 | |
618 | buf += *offset; |
619 | if (copy_to_user(buf, &header, sizeof(header))) |
620 | return -EFAULT14; |
621 | buf += sizeof(header); |
622 | |
623 | if (copy_to_user(buf, report, report_size)) |
624 | return -EFAULT14; |
625 | |
626 | (*offset) += header.size; |
627 | |
628 | return 0; |
629 | #endif |
630 | } |
631 | |
632 | /** |
633 | * Copies all buffered OA reports into userspace read() buffer. |
634 | * @stream: An i915-perf stream opened for OA metrics |
635 | * @buf: destination buffer given by userspace |
636 | * @count: the number of bytes userspace wants to read |
637 | * @offset: (inout): the current position for writing into @buf |
638 | * |
639 | * Notably any error condition resulting in a short read (-%ENOSPC or |
640 | * -%EFAULT) will be returned even though one or more records may |
641 | * have been successfully copied. In this case it's up to the caller |
642 | * to decide if the error should be squashed before returning to |
643 | * userspace. |
644 | * |
645 | * Note: reports are consumed from the head, and appended to the |
646 | * tail, so the tail chases the head?... If you think that's mad |
647 | * and back-to-front you're not alone, but this follows the |
648 | * Gen PRM naming convention. |
649 | * |
650 | * Returns: 0 on success, negative error code on failure. |
651 | */ |
652 | static int gen8_append_oa_reports(struct i915_perf_stream *stream, |
653 | char __user *buf, |
654 | size_t count, |
655 | size_t *offset) |
656 | { |
657 | struct intel_uncore *uncore = stream->uncore; |
658 | int report_size = stream->oa_buffer.format_size; |
659 | u8 *oa_buf_base = stream->oa_buffer.vaddr; |
660 | u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma); |
661 | u32 mask = (OA_BUFFER_SIZE(1024 * 1024 * 16) - 1); |
662 | size_t start_offset = *offset; |
663 | unsigned long flags; |
664 | u32 head, tail; |
665 | u32 taken; |
666 | int ret = 0; |
667 | |
668 | if (drm_WARN_ON(&uncore->i915->drm, !stream->enabled)({ int __ret = !!((!stream->enabled)); if (__ret) printf("%s %s: " "%s", dev_driver_string(((&uncore->i915->drm))-> dev), "", "drm_WARN_ON(" "!stream->enabled" ")"); __builtin_expect (!!(__ret), 0); })) |
669 | return -EIO5; |
670 | |
671 | spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags)do { flags = 0; mtx_enter(&stream->oa_buffer.ptr_lock) ; } while (0); |
672 | |
673 | head = stream->oa_buffer.head; |
674 | tail = stream->oa_buffer.tail; |
675 | |
676 | spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags)do { (void)(flags); mtx_leave(&stream->oa_buffer.ptr_lock ); } while (0); |
677 | |
678 | /* |
679 | * NB: oa_buffer.head/tail include the gtt_offset which we don't want |
680 | * while indexing relative to oa_buf_base. |
681 | */ |
682 | head -= gtt_offset; |
683 | tail -= gtt_offset; |
684 | |
685 | /* |
686 | * An out of bounds or misaligned head or tail pointer implies a driver |
687 | * bug since we validate + align the tail pointers we read from the |
688 | * hardware and we are in full control of the head pointer which should |
689 | * only be incremented by multiples of the report size (notably also |
690 | * all a power of two). |
691 | */ |
692 | if (drm_WARN_ONCE(&uncore->i915->drm,({ static int __warned; int __ret = !!(head > (1024 * 1024 * 16) || head % report_size || tail > (1024 * 1024 * 16) || tail % report_size); if (__ret && !__warned) { printf ("%s %s: " "Inconsistent OA buffer pointers: head = %u, tail = %u\n" , dev_driver_string((&uncore->i915->drm)->dev), "" , head, tail); __warned = 1; } __builtin_expect(!!(__ret), 0) ; }) |
693 | head > OA_BUFFER_SIZE || head % report_size ||({ static int __warned; int __ret = !!(head > (1024 * 1024 * 16) || head % report_size || tail > (1024 * 1024 * 16) || tail % report_size); if (__ret && !__warned) { printf ("%s %s: " "Inconsistent OA buffer pointers: head = %u, tail = %u\n" , dev_driver_string((&uncore->i915->drm)->dev), "" , head, tail); __warned = 1; } __builtin_expect(!!(__ret), 0) ; }) |
694 | tail > OA_BUFFER_SIZE || tail % report_size,({ static int __warned; int __ret = !!(head > (1024 * 1024 * 16) || head % report_size || tail > (1024 * 1024 * 16) || tail % report_size); if (__ret && !__warned) { printf ("%s %s: " "Inconsistent OA buffer pointers: head = %u, tail = %u\n" , dev_driver_string((&uncore->i915->drm)->dev), "" , head, tail); __warned = 1; } __builtin_expect(!!(__ret), 0) ; }) |
695 | "Inconsistent OA buffer pointers: head = %u, tail = %u\n",({ static int __warned; int __ret = !!(head > (1024 * 1024 * 16) || head % report_size || tail > (1024 * 1024 * 16) || tail % report_size); if (__ret && !__warned) { printf ("%s %s: " "Inconsistent OA buffer pointers: head = %u, tail = %u\n" , dev_driver_string((&uncore->i915->drm)->dev), "" , head, tail); __warned = 1; } __builtin_expect(!!(__ret), 0) ; }) |
696 | head, tail)({ static int __warned; int __ret = !!(head > (1024 * 1024 * 16) || head % report_size || tail > (1024 * 1024 * 16) || tail % report_size); if (__ret && !__warned) { printf ("%s %s: " "Inconsistent OA buffer pointers: head = %u, tail = %u\n" , dev_driver_string((&uncore->i915->drm)->dev), "" , head, tail); __warned = 1; } __builtin_expect(!!(__ret), 0) ; })) |
697 | return -EIO5; |
698 | |
699 | |
700 | for (/* none */; |
701 | (taken = OA_TAKEN(tail, head)((tail - head) & ((1024 * 1024 * 16) - 1))); |
Although the value stored to 'taken' is used in the enclosing expression, the value is never actually read from 'taken' | |
702 | head = (head + report_size) & mask) { |
703 | u8 *report = oa_buf_base + head; |
704 | u32 *report32 = (void *)report; |
705 | u32 ctx_id; |
706 | u32 reason; |
707 | |
708 | /* |
709 | * All the report sizes factor neatly into the buffer |
710 | * size so we never expect to see a report split |
711 | * between the beginning and end of the buffer. |
712 | * |
713 | * Given the initial alignment check a misalignment |
714 | * here would imply a driver bug that would result |
715 | * in an overrun. |
716 | */ |
717 | if (drm_WARN_ON(&uncore->i915->drm,({ int __ret = !!((((1024 * 1024 * 16) - head) < report_size )); if (__ret) printf("%s %s: " "%s", dev_driver_string(((& uncore->i915->drm))->dev), "", "drm_WARN_ON(" "((1024 * 1024 * 16) - head) < report_size" ")"); __builtin_expect(!!(__ret), 0); }) |
718 | (OA_BUFFER_SIZE - head) < report_size)({ int __ret = !!((((1024 * 1024 * 16) - head) < report_size )); if (__ret) printf("%s %s: " "%s", dev_driver_string(((& uncore->i915->drm))->dev), "", "drm_WARN_ON(" "((1024 * 1024 * 16) - head) < report_size" ")"); __builtin_expect(!!(__ret), 0); })) { |
719 | drm_err(&uncore->i915->drm,printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "Spurious OA head ptr: non-integral report offset\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__) |
720 | "Spurious OA head ptr: non-integral report offset\n")printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "Spurious OA head ptr: non-integral report offset\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__); |
721 | break; |
722 | } |
723 | |
724 | /* |
725 | * The reason field includes flags identifying what |
726 | * triggered this specific report (mostly timer |
727 | * triggered or e.g. due to a context switch). |
728 | * |
729 | * This field is never expected to be zero so we can |
730 | * check that the report isn't invalid before copying |
731 | * it to userspace... |
732 | */ |
733 | reason = ((report32[0] >> OAREPORT_REASON_SHIFT19) & |
734 | (IS_GEN(stream->perf->i915, 12)(0 + (&(stream->perf->i915)->__info)->gen == ( 12)) ? |
735 | OAREPORT_REASON_MASK_EXTENDED0x7f : |
736 | OAREPORT_REASON_MASK0x3f)); |
737 | if (reason == 0) { |
738 | if (__ratelimit(&stream->perf->spurious_report_rs)(1)) |
739 | DRM_NOTE("Skipping spurious, invalid OA report\n")printk("\0015" "[" "drm" "] " "Skipping spurious, invalid OA report\n" ); |
740 | continue; |
741 | } |
742 | |
743 | ctx_id = report32[2] & stream->specific_ctx_id_mask; |
744 | |
745 | /* |
746 | * Squash whatever is in the CTX_ID field if it's marked as |
747 | * invalid to be sure we avoid false-positive, single-context |
748 | * filtering below... |
749 | * |
750 | * Note: that we don't clear the valid_ctx_bit so userspace can |
751 | * understand that the ID has been squashed by the kernel. |
752 | */ |
753 | if (!(report32[0] & stream->perf->gen8_valid_ctx_bit) && |
754 | INTEL_GEN(stream->perf->i915)((&(stream->perf->i915)->__info)->gen) <= 11) |
755 | ctx_id = report32[2] = INVALID_CTX_ID0xffffffff; |
756 | |
757 | /* |
758 | * NB: For Gen 8 the OA unit no longer supports clock gating |
759 | * off for a specific context and the kernel can't securely |
760 | * stop the counters from updating as system-wide / global |
761 | * values. |
762 | * |
763 | * Automatic reports now include a context ID so reports can be |
764 | * filtered on the cpu but it's not worth trying to |
765 | * automatically subtract/hide counter progress for other |
766 | * contexts while filtering since we can't stop userspace |
767 | * issuing MI_REPORT_PERF_COUNT commands which would still |
768 | * provide a side-band view of the real values. |
769 | * |
770 | * To allow userspace (such as Mesa/GL_INTEL_performance_query) |
771 | * to normalize counters for a single filtered context then it |
772 | * needs be forwarded bookend context-switch reports so that it |
773 | * can track switches in between MI_REPORT_PERF_COUNT commands |
774 | * and can itself subtract/ignore the progress of counters |
775 | * associated with other contexts. Note that the hardware |
776 | * automatically triggers reports when switching to a new |
777 | * context which are tagged with the ID of the newly active |
778 | * context. To avoid the complexity (and likely fragility) of |
779 | * reading ahead while parsing reports to try and minimize |
780 | * forwarding redundant context switch reports (i.e. between |
781 | * other, unrelated contexts) we simply elect to forward them |
782 | * all. |
783 | * |
784 | * We don't rely solely on the reason field to identify context |
785 | * switches since it's not-uncommon for periodic samples to |
786 | * identify a switch before any 'context switch' report. |
787 | */ |
788 | if (!stream->perf->exclusive_stream->ctx || |
789 | stream->specific_ctx_id == ctx_id || |
790 | stream->oa_buffer.last_ctx_id == stream->specific_ctx_id || |
791 | reason & OAREPORT_REASON_CTX_SWITCH(1<<3)) { |
792 | |
793 | /* |
794 | * While filtering for a single context we avoid |
795 | * leaking the IDs of other contexts. |
796 | */ |
797 | if (stream->perf->exclusive_stream->ctx && |
798 | stream->specific_ctx_id != ctx_id) { |
799 | report32[2] = INVALID_CTX_ID0xffffffff; |
800 | } |
801 | |
802 | ret = append_oa_sample(stream, buf, count, offset, |
803 | report); |
804 | if (ret) |
805 | break; |
806 | |
807 | stream->oa_buffer.last_ctx_id = ctx_id; |
808 | } |
809 | |
810 | /* |
811 | * Clear out the first 2 dword as a mean to detect unlanded |
812 | * reports. |
813 | */ |
814 | report32[0] = 0; |
815 | report32[1] = 0; |
816 | } |
817 | |
818 | if (start_offset != *offset) { |
819 | i915_reg_t oaheadptr; |
820 | |
821 | oaheadptr = IS_GEN(stream->perf->i915, 12)(0 + (&(stream->perf->i915)->__info)->gen == ( 12)) ? |
822 | GEN12_OAG_OAHEADPTR((const i915_reg_t){ .reg = (0xdb00) }) : GEN8_OAHEADPTR((const i915_reg_t){ .reg = (0x2B0C) }); |
823 | |
824 | spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags)do { flags = 0; mtx_enter(&stream->oa_buffer.ptr_lock) ; } while (0); |
825 | |
826 | /* |
827 | * We removed the gtt_offset for the copy loop above, indexing |
828 | * relative to oa_buf_base so put back here... |
829 | */ |
830 | head += gtt_offset; |
831 | intel_uncore_write(uncore, oaheadptr, |
832 | head & GEN12_OAG_OAHEADPTR_MASK0xffffffc0); |
833 | stream->oa_buffer.head = head; |
834 | |
835 | spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags)do { (void)(flags); mtx_leave(&stream->oa_buffer.ptr_lock ); } while (0); |
836 | } |
837 | |
838 | return ret; |
839 | } |
840 | |
841 | /** |
842 | * gen8_oa_read - copy status records then buffered OA reports |
843 | * @stream: An i915-perf stream opened for OA metrics |
844 | * @buf: destination buffer given by userspace |
845 | * @count: the number of bytes userspace wants to read |
846 | * @offset: (inout): the current position for writing into @buf |
847 | * |
848 | * Checks OA unit status registers and if necessary appends corresponding |
849 | * status records for userspace (such as for a buffer full condition) and then |
850 | * initiate appending any buffered OA reports. |
851 | * |
852 | * Updates @offset according to the number of bytes successfully copied into |
853 | * the userspace buffer. |
854 | * |
855 | * NB: some data may be successfully copied to the userspace buffer |
856 | * even if an error is returned, and this is reflected in the |
857 | * updated @offset. |
858 | * |
859 | * Returns: zero on success or a negative error code |
860 | */ |
861 | static int gen8_oa_read(struct i915_perf_stream *stream, |
862 | char __user *buf, |
863 | size_t count, |
864 | size_t *offset) |
865 | { |
866 | struct intel_uncore *uncore = stream->uncore; |
867 | u32 oastatus; |
868 | i915_reg_t oastatus_reg; |
869 | int ret; |
870 | |
871 | if (drm_WARN_ON(&uncore->i915->drm, !stream->oa_buffer.vaddr)({ int __ret = !!((!stream->oa_buffer.vaddr)); if (__ret) printf ("%s %s: " "%s", dev_driver_string(((&uncore->i915-> drm))->dev), "", "drm_WARN_ON(" "!stream->oa_buffer.vaddr" ")"); __builtin_expect(!!(__ret), 0); })) |
872 | return -EIO5; |
873 | |
874 | oastatus_reg = IS_GEN(stream->perf->i915, 12)(0 + (&(stream->perf->i915)->__info)->gen == ( 12)) ? |
875 | GEN12_OAG_OASTATUS((const i915_reg_t){ .reg = (0xdafc) }) : GEN8_OASTATUS((const i915_reg_t){ .reg = (0x2b08) }); |
876 | |
877 | oastatus = intel_uncore_read(uncore, oastatus_reg); |
878 | |
879 | /* |
880 | * We treat OABUFFER_OVERFLOW as a significant error: |
881 | * |
882 | * Although theoretically we could handle this more gracefully |
883 | * sometimes, some Gens don't correctly suppress certain |
884 | * automatically triggered reports in this condition and so we |
885 | * have to assume that old reports are now being trampled |
886 | * over. |
887 | * |
888 | * Considering how we don't currently give userspace control |
889 | * over the OA buffer size and always configure a large 16MB |
890 | * buffer, then a buffer overflow does anyway likely indicate |
891 | * that something has gone quite badly wrong. |
892 | */ |
893 | if (oastatus & GEN8_OASTATUS_OABUFFER_OVERFLOW(1 << 1)) { |
894 | ret = append_oa_status(stream, buf, count, offset, |
895 | DRM_I915_PERF_RECORD_OA_BUFFER_LOST); |
896 | if (ret) |
897 | return ret; |
898 | |
899 | DRM_DEBUG("OA buffer overflow (exponent = %d): force restart\n",__drm_dbg(DRM_UT_CORE, "OA buffer overflow (exponent = %d): force restart\n" , stream->period_exponent) |
900 | stream->period_exponent)__drm_dbg(DRM_UT_CORE, "OA buffer overflow (exponent = %d): force restart\n" , stream->period_exponent); |
901 | |
902 | stream->perf->ops.oa_disable(stream); |
903 | stream->perf->ops.oa_enable(stream); |
904 | |
905 | /* |
906 | * Note: .oa_enable() is expected to re-init the oabuffer and |
907 | * reset GEN8_OASTATUS for us |
908 | */ |
909 | oastatus = intel_uncore_read(uncore, oastatus_reg); |
910 | } |
911 | |
912 | if (oastatus & GEN8_OASTATUS_REPORT_LOST(1 << 0)) { |
913 | ret = append_oa_status(stream, buf, count, offset, |
914 | DRM_I915_PERF_RECORD_OA_REPORT_LOST); |
915 | if (ret) |
916 | return ret; |
917 | |
918 | intel_uncore_rmw(uncore, oastatus_reg, |
919 | GEN8_OASTATUS_COUNTER_OVERFLOW(1 << 2) | |
920 | GEN8_OASTATUS_REPORT_LOST(1 << 0), |
921 | IS_GEN_RANGE(uncore->i915, 8, 10)(!!((&(uncore->i915)->__info)->gen_mask & ( 0 + 0 + (((~0UL) >> (64 - (((10)) - 1) - 1)) & ((~0UL ) << (((8)) - 1)))))) ? |
922 | (GEN8_OASTATUS_HEAD_POINTER_WRAP(1 << 16) | |
923 | GEN8_OASTATUS_TAIL_POINTER_WRAP(1 << 17)) : 0); |
924 | } |
925 | |
926 | return gen8_append_oa_reports(stream, buf, count, offset); |
927 | } |
928 | |
929 | /** |
930 | * Copies all buffered OA reports into userspace read() buffer. |
931 | * @stream: An i915-perf stream opened for OA metrics |
932 | * @buf: destination buffer given by userspace |
933 | * @count: the number of bytes userspace wants to read |
934 | * @offset: (inout): the current position for writing into @buf |
935 | * |
936 | * Notably any error condition resulting in a short read (-%ENOSPC or |
937 | * -%EFAULT) will be returned even though one or more records may |
938 | * have been successfully copied. In this case it's up to the caller |
939 | * to decide if the error should be squashed before returning to |
940 | * userspace. |
941 | * |
942 | * Note: reports are consumed from the head, and appended to the |
943 | * tail, so the tail chases the head?... If you think that's mad |
944 | * and back-to-front you're not alone, but this follows the |
945 | * Gen PRM naming convention. |
946 | * |
947 | * Returns: 0 on success, negative error code on failure. |
948 | */ |
949 | static int gen7_append_oa_reports(struct i915_perf_stream *stream, |
950 | char __user *buf, |
951 | size_t count, |
952 | size_t *offset) |
953 | { |
954 | struct intel_uncore *uncore = stream->uncore; |
955 | int report_size = stream->oa_buffer.format_size; |
956 | u8 *oa_buf_base = stream->oa_buffer.vaddr; |
957 | u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma); |
958 | u32 mask = (OA_BUFFER_SIZE(1024 * 1024 * 16) - 1); |
959 | size_t start_offset = *offset; |
960 | unsigned long flags; |
961 | u32 head, tail; |
962 | u32 taken; |
963 | int ret = 0; |
964 | |
965 | if (drm_WARN_ON(&uncore->i915->drm, !stream->enabled)({ int __ret = !!((!stream->enabled)); if (__ret) printf("%s %s: " "%s", dev_driver_string(((&uncore->i915->drm))-> dev), "", "drm_WARN_ON(" "!stream->enabled" ")"); __builtin_expect (!!(__ret), 0); })) |
966 | return -EIO5; |
967 | |
968 | spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags)do { flags = 0; mtx_enter(&stream->oa_buffer.ptr_lock) ; } while (0); |
969 | |
970 | head = stream->oa_buffer.head; |
971 | tail = stream->oa_buffer.tail; |
972 | |
973 | spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags)do { (void)(flags); mtx_leave(&stream->oa_buffer.ptr_lock ); } while (0); |
974 | |
975 | /* NB: oa_buffer.head/tail include the gtt_offset which we don't want |
976 | * while indexing relative to oa_buf_base. |
977 | */ |
978 | head -= gtt_offset; |
979 | tail -= gtt_offset; |
980 | |
981 | /* An out of bounds or misaligned head or tail pointer implies a driver |
982 | * bug since we validate + align the tail pointers we read from the |
983 | * hardware and we are in full control of the head pointer which should |
984 | * only be incremented by multiples of the report size (notably also |
985 | * all a power of two). |
986 | */ |
987 | if (drm_WARN_ONCE(&uncore->i915->drm,({ static int __warned; int __ret = !!(head > (1024 * 1024 * 16) || head % report_size || tail > (1024 * 1024 * 16) || tail % report_size); if (__ret && !__warned) { printf ("%s %s: " "Inconsistent OA buffer pointers: head = %u, tail = %u\n" , dev_driver_string((&uncore->i915->drm)->dev), "" , head, tail); __warned = 1; } __builtin_expect(!!(__ret), 0) ; }) |
988 | head > OA_BUFFER_SIZE || head % report_size ||({ static int __warned; int __ret = !!(head > (1024 * 1024 * 16) || head % report_size || tail > (1024 * 1024 * 16) || tail % report_size); if (__ret && !__warned) { printf ("%s %s: " "Inconsistent OA buffer pointers: head = %u, tail = %u\n" , dev_driver_string((&uncore->i915->drm)->dev), "" , head, tail); __warned = 1; } __builtin_expect(!!(__ret), 0) ; }) |
989 | tail > OA_BUFFER_SIZE || tail % report_size,({ static int __warned; int __ret = !!(head > (1024 * 1024 * 16) || head % report_size || tail > (1024 * 1024 * 16) || tail % report_size); if (__ret && !__warned) { printf ("%s %s: " "Inconsistent OA buffer pointers: head = %u, tail = %u\n" , dev_driver_string((&uncore->i915->drm)->dev), "" , head, tail); __warned = 1; } __builtin_expect(!!(__ret), 0) ; }) |
990 | "Inconsistent OA buffer pointers: head = %u, tail = %u\n",({ static int __warned; int __ret = !!(head > (1024 * 1024 * 16) || head % report_size || tail > (1024 * 1024 * 16) || tail % report_size); if (__ret && !__warned) { printf ("%s %s: " "Inconsistent OA buffer pointers: head = %u, tail = %u\n" , dev_driver_string((&uncore->i915->drm)->dev), "" , head, tail); __warned = 1; } __builtin_expect(!!(__ret), 0) ; }) |
991 | head, tail)({ static int __warned; int __ret = !!(head > (1024 * 1024 * 16) || head % report_size || tail > (1024 * 1024 * 16) || tail % report_size); if (__ret && !__warned) { printf ("%s %s: " "Inconsistent OA buffer pointers: head = %u, tail = %u\n" , dev_driver_string((&uncore->i915->drm)->dev), "" , head, tail); __warned = 1; } __builtin_expect(!!(__ret), 0) ; })) |
992 | return -EIO5; |
993 | |
994 | |
995 | for (/* none */; |
996 | (taken = OA_TAKEN(tail, head)((tail - head) & ((1024 * 1024 * 16) - 1))); |
997 | head = (head + report_size) & mask) { |
998 | u8 *report = oa_buf_base + head; |
999 | u32 *report32 = (void *)report; |
1000 | |
1001 | /* All the report sizes factor neatly into the buffer |
1002 | * size so we never expect to see a report split |
1003 | * between the beginning and end of the buffer. |
1004 | * |
1005 | * Given the initial alignment check a misalignment |
1006 | * here would imply a driver bug that would result |
1007 | * in an overrun. |
1008 | */ |
1009 | if (drm_WARN_ON(&uncore->i915->drm,({ int __ret = !!((((1024 * 1024 * 16) - head) < report_size )); if (__ret) printf("%s %s: " "%s", dev_driver_string(((& uncore->i915->drm))->dev), "", "drm_WARN_ON(" "((1024 * 1024 * 16) - head) < report_size" ")"); __builtin_expect(!!(__ret), 0); }) |
1010 | (OA_BUFFER_SIZE - head) < report_size)({ int __ret = !!((((1024 * 1024 * 16) - head) < report_size )); if (__ret) printf("%s %s: " "%s", dev_driver_string(((& uncore->i915->drm))->dev), "", "drm_WARN_ON(" "((1024 * 1024 * 16) - head) < report_size" ")"); __builtin_expect(!!(__ret), 0); })) { |
1011 | drm_err(&uncore->i915->drm,printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "Spurious OA head ptr: non-integral report offset\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__) |
1012 | "Spurious OA head ptr: non-integral report offset\n")printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "Spurious OA head ptr: non-integral report offset\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__); |
1013 | break; |
1014 | } |
1015 | |
1016 | /* The report-ID field for periodic samples includes |
1017 | * some undocumented flags related to what triggered |
1018 | * the report and is never expected to be zero so we |
1019 | * can check that the report isn't invalid before |
1020 | * copying it to userspace... |
1021 | */ |
1022 | if (report32[0] == 0) { |
1023 | if (__ratelimit(&stream->perf->spurious_report_rs)(1)) |
1024 | DRM_NOTE("Skipping spurious, invalid OA report\n")printk("\0015" "[" "drm" "] " "Skipping spurious, invalid OA report\n" ); |
1025 | continue; |
1026 | } |
1027 | |
1028 | ret = append_oa_sample(stream, buf, count, offset, report); |
1029 | if (ret) |
1030 | break; |
1031 | |
1032 | /* Clear out the first 2 dwords as a mean to detect unlanded |
1033 | * reports. |
1034 | */ |
1035 | report32[0] = 0; |
1036 | report32[1] = 0; |
1037 | } |
1038 | |
1039 | if (start_offset != *offset) { |
1040 | spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags)do { flags = 0; mtx_enter(&stream->oa_buffer.ptr_lock) ; } while (0); |
1041 | |
1042 | /* We removed the gtt_offset for the copy loop above, indexing |
1043 | * relative to oa_buf_base so put back here... |
1044 | */ |
1045 | head += gtt_offset; |
1046 | |
1047 | intel_uncore_write(uncore, GEN7_OASTATUS2((const i915_reg_t){ .reg = (0x2368) }), |
1048 | (head & GEN7_OASTATUS2_HEAD_MASK0xffffffc0) | |
1049 | GEN7_OASTATUS2_MEM_SELECT_GGTT(1 << 0)); |
1050 | stream->oa_buffer.head = head; |
1051 | |
1052 | spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags)do { (void)(flags); mtx_leave(&stream->oa_buffer.ptr_lock ); } while (0); |
1053 | } |
1054 | |
1055 | return ret; |
1056 | } |
1057 | |
1058 | /** |
1059 | * gen7_oa_read - copy status records then buffered OA reports |
1060 | * @stream: An i915-perf stream opened for OA metrics |
1061 | * @buf: destination buffer given by userspace |
1062 | * @count: the number of bytes userspace wants to read |
1063 | * @offset: (inout): the current position for writing into @buf |
1064 | * |
1065 | * Checks Gen 7 specific OA unit status registers and if necessary appends |
1066 | * corresponding status records for userspace (such as for a buffer full |
1067 | * condition) and then initiate appending any buffered OA reports. |
1068 | * |
1069 | * Updates @offset according to the number of bytes successfully copied into |
1070 | * the userspace buffer. |
1071 | * |
1072 | * Returns: zero on success or a negative error code |
1073 | */ |
1074 | static int gen7_oa_read(struct i915_perf_stream *stream, |
1075 | char __user *buf, |
1076 | size_t count, |
1077 | size_t *offset) |
1078 | { |
1079 | struct intel_uncore *uncore = stream->uncore; |
1080 | u32 oastatus1; |
1081 | int ret; |
1082 | |
1083 | if (drm_WARN_ON(&uncore->i915->drm, !stream->oa_buffer.vaddr)({ int __ret = !!((!stream->oa_buffer.vaddr)); if (__ret) printf ("%s %s: " "%s", dev_driver_string(((&uncore->i915-> drm))->dev), "", "drm_WARN_ON(" "!stream->oa_buffer.vaddr" ")"); __builtin_expect(!!(__ret), 0); })) |
1084 | return -EIO5; |
1085 | |
1086 | oastatus1 = intel_uncore_read(uncore, GEN7_OASTATUS1((const i915_reg_t){ .reg = (0x2364) })); |
1087 | |
1088 | /* XXX: On Haswell we don't have a safe way to clear oastatus1 |
1089 | * bits while the OA unit is enabled (while the tail pointer |
1090 | * may be updated asynchronously) so we ignore status bits |
1091 | * that have already been reported to userspace. |
1092 | */ |
1093 | oastatus1 &= ~stream->perf->gen7_latched_oastatus1; |
1094 | |
1095 | /* We treat OABUFFER_OVERFLOW as a significant error: |
1096 | * |
1097 | * - The status can be interpreted to mean that the buffer is |
1098 | * currently full (with a higher precedence than OA_TAKEN() |
1099 | * which will start to report a near-empty buffer after an |
1100 | * overflow) but it's awkward that we can't clear the status |
1101 | * on Haswell, so without a reset we won't be able to catch |
1102 | * the state again. |
1103 | * |
1104 | * - Since it also implies the HW has started overwriting old |
1105 | * reports it may also affect our sanity checks for invalid |
1106 | * reports when copying to userspace that assume new reports |
1107 | * are being written to cleared memory. |
1108 | * |
1109 | * - In the future we may want to introduce a flight recorder |
1110 | * mode where the driver will automatically maintain a safe |
1111 | * guard band between head/tail, avoiding this overflow |
1112 | * condition, but we avoid the added driver complexity for |
1113 | * now. |
1114 | */ |
1115 | if (unlikely(oastatus1 & GEN7_OASTATUS1_OABUFFER_OVERFLOW)__builtin_expect(!!(oastatus1 & (1 << 1)), 0)) { |
1116 | ret = append_oa_status(stream, buf, count, offset, |
1117 | DRM_I915_PERF_RECORD_OA_BUFFER_LOST); |
1118 | if (ret) |
1119 | return ret; |
1120 | |
1121 | DRM_DEBUG("OA buffer overflow (exponent = %d): force restart\n",__drm_dbg(DRM_UT_CORE, "OA buffer overflow (exponent = %d): force restart\n" , stream->period_exponent) |
1122 | stream->period_exponent)__drm_dbg(DRM_UT_CORE, "OA buffer overflow (exponent = %d): force restart\n" , stream->period_exponent); |
1123 | |
1124 | stream->perf->ops.oa_disable(stream); |
1125 | stream->perf->ops.oa_enable(stream); |
1126 | |
1127 | oastatus1 = intel_uncore_read(uncore, GEN7_OASTATUS1((const i915_reg_t){ .reg = (0x2364) })); |
1128 | } |
1129 | |
1130 | if (unlikely(oastatus1 & GEN7_OASTATUS1_REPORT_LOST)__builtin_expect(!!(oastatus1 & (1 << 0)), 0)) { |
1131 | ret = append_oa_status(stream, buf, count, offset, |
1132 | DRM_I915_PERF_RECORD_OA_REPORT_LOST); |
1133 | if (ret) |
1134 | return ret; |
1135 | stream->perf->gen7_latched_oastatus1 |= |
1136 | GEN7_OASTATUS1_REPORT_LOST(1 << 0); |
1137 | } |
1138 | |
1139 | return gen7_append_oa_reports(stream, buf, count, offset); |
1140 | } |
1141 | |
1142 | /** |
1143 | * i915_oa_wait_unlocked - handles blocking IO until OA data available |
1144 | * @stream: An i915-perf stream opened for OA metrics |
1145 | * |
1146 | * Called when userspace tries to read() from a blocking stream FD opened |
1147 | * for OA metrics. It waits until the hrtimer callback finds a non-empty |
1148 | * OA buffer and wakes us. |
1149 | * |
1150 | * Note: it's acceptable to have this return with some false positives |
1151 | * since any subsequent read handling will return -EAGAIN if there isn't |
1152 | * really data ready for userspace yet. |
1153 | * |
1154 | * Returns: zero on success or a negative error code |
1155 | */ |
1156 | static int i915_oa_wait_unlocked(struct i915_perf_stream *stream) |
1157 | { |
1158 | /* We would wait indefinitely if periodic sampling is not enabled */ |
1159 | if (!stream->periodic) |
1160 | return -EIO5; |
1161 | |
1162 | return wait_event_interruptible(stream->poll_wq,({ int __ret = 0; if (!(oa_buffer_check_unlocked(stream))) __ret = ({ long ret = 0; do { int __error; unsigned long deadline; ((!cold) ? (void)0 : __assert("diagnostic ", "/usr/src/sys/dev/pci/drm/i915/i915_perf.c" , 1163, "!cold")); mtx_enter(&sch_mtx); deadline = jiffies + ret; __error = msleep(&stream->poll_wq, &sch_mtx , 0x100, "drmweti", ret); ret = deadline - jiffies; if (__error == -1 || __error == 4) { ret = -4; mtx_leave(&sch_mtx); break ; } if ((0) > 0 && (ret <= 0 || __error == 35)) { mtx_leave(&sch_mtx); ret = ((oa_buffer_check_unlocked( stream))) ? 1 : 0; break; } mtx_leave(&sch_mtx); } while ( ret > 0 && !(oa_buffer_check_unlocked(stream))); ret ; }); __ret; }) |
1163 | oa_buffer_check_unlocked(stream))({ int __ret = 0; if (!(oa_buffer_check_unlocked(stream))) __ret = ({ long ret = 0; do { int __error; unsigned long deadline; ((!cold) ? (void)0 : __assert("diagnostic ", "/usr/src/sys/dev/pci/drm/i915/i915_perf.c" , 1163, "!cold")); mtx_enter(&sch_mtx); deadline = jiffies + ret; __error = msleep(&stream->poll_wq, &sch_mtx , 0x100, "drmweti", ret); ret = deadline - jiffies; if (__error == -1 || __error == 4) { ret = -4; mtx_leave(&sch_mtx); break ; } if ((0) > 0 && (ret <= 0 || __error == 35)) { mtx_leave(&sch_mtx); ret = ((oa_buffer_check_unlocked( stream))) ? 1 : 0; break; } mtx_leave(&sch_mtx); } while ( ret > 0 && !(oa_buffer_check_unlocked(stream))); ret ; }); __ret; }); |
1164 | } |
1165 | |
1166 | #ifdef notyet |
1167 | /** |
1168 | * i915_oa_poll_wait - call poll_wait() for an OA stream poll() |
1169 | * @stream: An i915-perf stream opened for OA metrics |
1170 | * @file: An i915 perf stream file |
1171 | * @wait: poll() state table |
1172 | * |
1173 | * For handling userspace polling on an i915 perf stream opened for OA metrics, |
1174 | * this starts a poll_wait with the wait queue that our hrtimer callback wakes |
1175 | * when it sees data ready to read in the circular OA buffer. |
1176 | */ |
1177 | static void i915_oa_poll_wait(struct i915_perf_stream *stream, |
1178 | struct file *file, |
1179 | poll_table *wait) |
1180 | { |
1181 | poll_wait(file, &stream->poll_wq, wait); |
1182 | } |
1183 | #endif |
1184 | |
1185 | /** |
1186 | * i915_oa_read - just calls through to &i915_oa_ops->read |
1187 | * @stream: An i915-perf stream opened for OA metrics |
1188 | * @buf: destination buffer given by userspace |
1189 | * @count: the number of bytes userspace wants to read |
1190 | * @offset: (inout): the current position for writing into @buf |
1191 | * |
1192 | * Updates @offset according to the number of bytes successfully copied into |
1193 | * the userspace buffer. |
1194 | * |
1195 | * Returns: zero on success or a negative error code |
1196 | */ |
1197 | static int i915_oa_read(struct i915_perf_stream *stream, |
1198 | char __user *buf, |
1199 | size_t count, |
1200 | size_t *offset) |
1201 | { |
1202 | return stream->perf->ops.read(stream, buf, count, offset); |
1203 | } |
1204 | |
1205 | static struct intel_context *oa_pin_context(struct i915_perf_stream *stream) |
1206 | { |
1207 | struct i915_gem_engines_iter it; |
1208 | struct i915_gem_context *ctx = stream->ctx; |
1209 | struct intel_context *ce; |
1210 | struct i915_gem_ww_ctx ww; |
1211 | int err = -ENODEV19; |
1212 | |
1213 | for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it)for (i915_gem_engines_iter_init(&(it), (i915_gem_context_lock_engines (ctx))); ((ce) = i915_gem_engines_iter_next(&(it)));) { |
1214 | if (ce->engine != stream->engine) /* first match! */ |
1215 | continue; |
1216 | |
1217 | err = 0; |
1218 | break; |
1219 | } |
1220 | i915_gem_context_unlock_engines(ctx); |
1221 | |
1222 | if (err) |
1223 | return ERR_PTR(err); |
1224 | |
1225 | i915_gem_ww_ctx_init(&ww, true1); |
1226 | retry: |
1227 | /* |
1228 | * As the ID is the gtt offset of the context's vma we |
1229 | * pin the vma to ensure the ID remains fixed. |
1230 | */ |
1231 | err = intel_context_pin_ww(ce, &ww); |
1232 | if (err == -EDEADLK11) { |
1233 | err = i915_gem_ww_ctx_backoff(&ww); |
1234 | if (!err) |
1235 | goto retry; |
1236 | } |
1237 | i915_gem_ww_ctx_fini(&ww); |
1238 | |
1239 | if (err) |
1240 | return ERR_PTR(err); |
1241 | |
1242 | stream->pinned_ctx = ce; |
1243 | return stream->pinned_ctx; |
1244 | } |
1245 | |
1246 | /** |
1247 | * oa_get_render_ctx_id - determine and hold ctx hw id |
1248 | * @stream: An i915-perf stream opened for OA metrics |
1249 | * |
1250 | * Determine the render context hw id, and ensure it remains fixed for the |
1251 | * lifetime of the stream. This ensures that we don't have to worry about |
1252 | * updating the context ID in OACONTROL on the fly. |
1253 | * |
1254 | * Returns: zero on success or a negative error code |
1255 | */ |
1256 | static int oa_get_render_ctx_id(struct i915_perf_stream *stream) |
1257 | { |
1258 | struct intel_context *ce; |
1259 | |
1260 | ce = oa_pin_context(stream); |
1261 | if (IS_ERR(ce)) |
1262 | return PTR_ERR(ce); |
1263 | |
1264 | switch (INTEL_GEN(ce->engine->i915)((&(ce->engine->i915)->__info)->gen)) { |
1265 | case 7: { |
1266 | /* |
1267 | * On Haswell we don't do any post processing of the reports |
1268 | * and don't need to use the mask. |
1269 | */ |
1270 | stream->specific_ctx_id = i915_ggtt_offset(ce->state); |
1271 | stream->specific_ctx_id_mask = 0; |
1272 | break; |
1273 | } |
1274 | |
1275 | case 8: |
1276 | case 9: |
1277 | case 10: |
1278 | if (intel_engine_in_execlists_submission_mode(ce->engine)) { |
1279 | stream->specific_ctx_id_mask = |
1280 | (1U << GEN8_CTX_ID_WIDTH21) - 1; |
1281 | stream->specific_ctx_id = stream->specific_ctx_id_mask; |
1282 | } else { |
1283 | /* |
1284 | * When using GuC, the context descriptor we write in |
1285 | * i915 is read by GuC and rewritten before it's |
1286 | * actually written into the hardware. The LRCA is |
1287 | * what is put into the context id field of the |
1288 | * context descriptor by GuC. Because it's aligned to |
1289 | * a page, the lower 12bits are always at 0 and |
1290 | * dropped by GuC. They won't be part of the context |
1291 | * ID in the OA reports, so squash those lower bits. |
1292 | */ |
1293 | stream->specific_ctx_id = ce->lrc.lrca >> 12; |
1294 | |
1295 | /* |
1296 | * GuC uses the top bit to signal proxy submission, so |
1297 | * ignore that bit. |
1298 | */ |
1299 | stream->specific_ctx_id_mask = |
1300 | (1U << (GEN8_CTX_ID_WIDTH21 - 1)) - 1; |
1301 | } |
1302 | break; |
1303 | |
1304 | case 11: |
1305 | case 12: { |
1306 | stream->specific_ctx_id_mask = |
1307 | ((1U << GEN11_SW_CTX_ID_WIDTH11) - 1) << (GEN11_SW_CTX_ID_SHIFT37 - 32); |
1308 | /* |
1309 | * Pick an unused context id |
1310 | * 0 - BITS_PER_LONG are used by other contexts |
1311 | * GEN12_MAX_CONTEXT_HW_ID (0x7ff) is used by idle context |
1312 | */ |
1313 | stream->specific_ctx_id = (GEN12_MAX_CONTEXT_HW_ID((1<<11) - 1) - 1) << (GEN11_SW_CTX_ID_SHIFT37 - 32); |
1314 | break; |
1315 | } |
1316 | |
1317 | default: |
1318 | MISSING_CASE(INTEL_GEN(ce->engine->i915))({ int __ret = !!(1); if (__ret) printf("Missing case (%s == %ld)\n" , "((&(ce->engine->i915)->__info)->gen)", (long )(((&(ce->engine->i915)->__info)->gen))); __builtin_expect (!!(__ret), 0); }); |
1319 | } |
1320 | |
1321 | ce->tag = stream->specific_ctx_id; |
1322 | |
1323 | drm_dbg(&stream->perf->i915->drm,drm_dev_dbg((&stream->perf->i915->drm)->dev, DRM_UT_DRIVER , "filtering on ctx_id=0x%x ctx_id_mask=0x%x\n", stream->specific_ctx_id , stream->specific_ctx_id_mask) |
1324 | "filtering on ctx_id=0x%x ctx_id_mask=0x%x\n",drm_dev_dbg((&stream->perf->i915->drm)->dev, DRM_UT_DRIVER , "filtering on ctx_id=0x%x ctx_id_mask=0x%x\n", stream->specific_ctx_id , stream->specific_ctx_id_mask) |
1325 | stream->specific_ctx_id,drm_dev_dbg((&stream->perf->i915->drm)->dev, DRM_UT_DRIVER , "filtering on ctx_id=0x%x ctx_id_mask=0x%x\n", stream->specific_ctx_id , stream->specific_ctx_id_mask) |
1326 | stream->specific_ctx_id_mask)drm_dev_dbg((&stream->perf->i915->drm)->dev, DRM_UT_DRIVER , "filtering on ctx_id=0x%x ctx_id_mask=0x%x\n", stream->specific_ctx_id , stream->specific_ctx_id_mask); |
1327 | |
1328 | return 0; |
1329 | } |
1330 | |
1331 | /** |
1332 | * oa_put_render_ctx_id - counterpart to oa_get_render_ctx_id releases hold |
1333 | * @stream: An i915-perf stream opened for OA metrics |
1334 | * |
1335 | * In case anything needed doing to ensure the context HW ID would remain valid |
1336 | * for the lifetime of the stream, then that can be undone here. |
1337 | */ |
1338 | static void oa_put_render_ctx_id(struct i915_perf_stream *stream) |
1339 | { |
1340 | struct intel_context *ce; |
1341 | |
1342 | ce = fetch_and_zero(&stream->pinned_ctx)({ typeof(*&stream->pinned_ctx) __T = *(&stream-> pinned_ctx); *(&stream->pinned_ctx) = (typeof(*&stream ->pinned_ctx))0; __T; }); |
1343 | if (ce) { |
1344 | ce->tag = 0; /* recomputed on next submission after parking */ |
1345 | intel_context_unpin(ce); |
1346 | } |
1347 | |
1348 | stream->specific_ctx_id = INVALID_CTX_ID0xffffffff; |
1349 | stream->specific_ctx_id_mask = 0; |
1350 | } |
1351 | |
1352 | static void |
1353 | free_oa_buffer(struct i915_perf_stream *stream) |
1354 | { |
1355 | i915_vma_unpin_and_release(&stream->oa_buffer.vma, |
1356 | I915_VMA_RELEASE_MAP(1UL << (0))); |
1357 | |
1358 | stream->oa_buffer.vaddr = NULL((void *)0); |
1359 | } |
1360 | |
1361 | static void |
1362 | free_oa_configs(struct i915_perf_stream *stream) |
1363 | { |
1364 | struct i915_oa_config_bo *oa_bo, *tmp; |
1365 | |
1366 | i915_oa_config_put(stream->oa_config); |
1367 | llist_for_each_entry_safe(oa_bo, tmp, stream->oa_config_bos.first, node)for (oa_bo = (((stream->oa_config_bos.first)) ? ({ const __typeof ( ((__typeof(*oa_bo) *)0)->node ) *__mptr = ((stream->oa_config_bos .first)); (__typeof(*oa_bo) *)( (char *)__mptr - __builtin_offsetof (__typeof(*oa_bo), node) );}) : ((void *)0)); oa_bo != ((void *)0) && (tmp = ((oa_bo->node.next) ? ({ const __typeof ( ((__typeof(*oa_bo) *)0)->node ) *__mptr = (oa_bo->node .next); (__typeof(*oa_bo) *)( (char *)__mptr - __builtin_offsetof (__typeof(*oa_bo), node) );}) : ((void *)0)), oa_bo); oa_bo = tmp) |
1368 | free_oa_config_bo(oa_bo); |
1369 | } |
1370 | |
1371 | static void |
1372 | free_noa_wait(struct i915_perf_stream *stream) |
1373 | { |
1374 | i915_vma_unpin_and_release(&stream->noa_wait, 0); |
1375 | } |
1376 | |
1377 | static void i915_oa_stream_destroy(struct i915_perf_stream *stream) |
1378 | { |
1379 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
1380 | #ifdef notyet |
1381 | struct i915_perf *perf = stream->perf; |
1382 | |
1383 | BUG_ON(stream != perf->exclusive_stream)((!(stream != perf->exclusive_stream)) ? (void)0 : __assert ("diagnostic ", "/usr/src/sys/dev/pci/drm/i915/i915_perf.c", 1383 , "!(stream != perf->exclusive_stream)")); |
1384 | |
1385 | /* |
1386 | * Unset exclusive_stream first, it will be checked while disabling |
1387 | * the metric set on gen8+. |
1388 | * |
1389 | * See i915_oa_init_reg_state() and lrc_configure_all_contexts() |
1390 | */ |
1391 | WRITE_ONCE(perf->exclusive_stream, NULL)({ typeof(perf->exclusive_stream) __tmp = (((void *)0)); * (volatile typeof(perf->exclusive_stream) *)&(perf-> exclusive_stream) = __tmp; __tmp; }); |
1392 | perf->ops.disable_metric_set(stream); |
1393 | |
1394 | free_oa_buffer(stream); |
1395 | |
1396 | intel_uncore_forcewake_put(stream->uncore, FORCEWAKE_ALL); |
1397 | intel_engine_pm_put(stream->engine); |
1398 | |
1399 | if (stream->ctx) |
1400 | oa_put_render_ctx_id(stream); |
1401 | |
1402 | free_oa_configs(stream); |
1403 | free_noa_wait(stream); |
1404 | |
1405 | if (perf->spurious_report_rs.missed) { |
1406 | DRM_NOTE("%d spurious OA report notices suppressed due to ratelimiting\n",printk("\0015" "[" "drm" "] " "%d spurious OA report notices suppressed due to ratelimiting\n" , perf->spurious_report_rs.missed) |
1407 | perf->spurious_report_rs.missed)printk("\0015" "[" "drm" "] " "%d spurious OA report notices suppressed due to ratelimiting\n" , perf->spurious_report_rs.missed); |
1408 | } |
1409 | #endif |
1410 | } |
1411 | |
1412 | static void gen7_init_oa_buffer(struct i915_perf_stream *stream) |
1413 | { |
1414 | struct intel_uncore *uncore = stream->uncore; |
1415 | u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma); |
1416 | unsigned long flags; |
1417 | |
1418 | spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags)do { flags = 0; mtx_enter(&stream->oa_buffer.ptr_lock) ; } while (0); |
1419 | |
1420 | /* Pre-DevBDW: OABUFFER must be set with counters off, |
1421 | * before OASTATUS1, but after OASTATUS2 |
1422 | */ |
1423 | intel_uncore_write(uncore, GEN7_OASTATUS2((const i915_reg_t){ .reg = (0x2368) }), /* head */ |
1424 | gtt_offset | GEN7_OASTATUS2_MEM_SELECT_GGTT(1 << 0)); |
1425 | stream->oa_buffer.head = gtt_offset; |
1426 | |
1427 | intel_uncore_write(uncore, GEN7_OABUFFER((const i915_reg_t){ .reg = (0x23B0) }), gtt_offset); |
1428 | |
1429 | intel_uncore_write(uncore, GEN7_OASTATUS1((const i915_reg_t){ .reg = (0x2364) }), /* tail */ |
1430 | gtt_offset | OABUFFER_SIZE_16M(7 << 3)); |
1431 | |
1432 | /* Mark that we need updated tail pointers to read from... */ |
1433 | stream->oa_buffer.aging_tail = INVALID_TAIL_PTR0xffffffff; |
1434 | stream->oa_buffer.tail = gtt_offset; |
1435 | |
1436 | spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags)do { (void)(flags); mtx_leave(&stream->oa_buffer.ptr_lock ); } while (0); |
1437 | |
1438 | /* On Haswell we have to track which OASTATUS1 flags we've |
1439 | * already seen since they can't be cleared while periodic |
1440 | * sampling is enabled. |
1441 | */ |
1442 | stream->perf->gen7_latched_oastatus1 = 0; |
1443 | |
1444 | /* NB: although the OA buffer will initially be allocated |
1445 | * zeroed via shmfs (and so this memset is redundant when |
1446 | * first allocating), we may re-init the OA buffer, either |
1447 | * when re-enabling a stream or in error/reset paths. |
1448 | * |
1449 | * The reason we clear the buffer for each re-init is for the |
1450 | * sanity check in gen7_append_oa_reports() that looks at the |
1451 | * report-id field to make sure it's non-zero which relies on |
1452 | * the assumption that new reports are being written to zeroed |
1453 | * memory... |
1454 | */ |
1455 | memset(stream->oa_buffer.vaddr, 0, OA_BUFFER_SIZE)__builtin_memset((stream->oa_buffer.vaddr), (0), ((1024 * 1024 * 16))); |
1456 | } |
1457 | |
1458 | static void gen8_init_oa_buffer(struct i915_perf_stream *stream) |
1459 | { |
1460 | struct intel_uncore *uncore = stream->uncore; |
1461 | u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma); |
1462 | unsigned long flags; |
1463 | |
1464 | spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags)do { flags = 0; mtx_enter(&stream->oa_buffer.ptr_lock) ; } while (0); |
1465 | |
1466 | intel_uncore_write(uncore, GEN8_OASTATUS((const i915_reg_t){ .reg = (0x2b08) }), 0); |
1467 | intel_uncore_write(uncore, GEN8_OAHEADPTR((const i915_reg_t){ .reg = (0x2B0C) }), gtt_offset); |
1468 | stream->oa_buffer.head = gtt_offset; |
1469 | |
1470 | intel_uncore_write(uncore, GEN8_OABUFFER_UDW((const i915_reg_t){ .reg = (0x23b4) }), 0); |
1471 | |
1472 | /* |
1473 | * PRM says: |
1474 | * |
1475 | * "This MMIO must be set before the OATAILPTR |
1476 | * register and after the OAHEADPTR register. This is |
1477 | * to enable proper functionality of the overflow |
1478 | * bit." |
1479 | */ |
1480 | intel_uncore_write(uncore, GEN8_OABUFFER((const i915_reg_t){ .reg = (0x2b14) }), gtt_offset | |
1481 | OABUFFER_SIZE_16M(7 << 3) | GEN8_OABUFFER_MEM_SELECT_GGTT(1 << 0)); |
1482 | intel_uncore_write(uncore, GEN8_OATAILPTR((const i915_reg_t){ .reg = (0x2B10) }), gtt_offset & GEN8_OATAILPTR_MASK0xffffffc0); |
1483 | |
1484 | /* Mark that we need updated tail pointers to read from... */ |
1485 | stream->oa_buffer.aging_tail = INVALID_TAIL_PTR0xffffffff; |
1486 | stream->oa_buffer.tail = gtt_offset; |
1487 | |
1488 | /* |
1489 | * Reset state used to recognise context switches, affecting which |
1490 | * reports we will forward to userspace while filtering for a single |
1491 | * context. |
1492 | */ |
1493 | stream->oa_buffer.last_ctx_id = INVALID_CTX_ID0xffffffff; |
1494 | |
1495 | spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags)do { (void)(flags); mtx_leave(&stream->oa_buffer.ptr_lock ); } while (0); |
1496 | |
1497 | /* |
1498 | * NB: although the OA buffer will initially be allocated |
1499 | * zeroed via shmfs (and so this memset is redundant when |
1500 | * first allocating), we may re-init the OA buffer, either |
1501 | * when re-enabling a stream or in error/reset paths. |
1502 | * |
1503 | * The reason we clear the buffer for each re-init is for the |
1504 | * sanity check in gen8_append_oa_reports() that looks at the |
1505 | * reason field to make sure it's non-zero which relies on |
1506 | * the assumption that new reports are being written to zeroed |
1507 | * memory... |
1508 | */ |
1509 | memset(stream->oa_buffer.vaddr, 0, OA_BUFFER_SIZE)__builtin_memset((stream->oa_buffer.vaddr), (0), ((1024 * 1024 * 16))); |
1510 | } |
1511 | |
1512 | static void gen12_init_oa_buffer(struct i915_perf_stream *stream) |
1513 | { |
1514 | struct intel_uncore *uncore = stream->uncore; |
1515 | u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma); |
1516 | unsigned long flags; |
1517 | |
1518 | spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags)do { flags = 0; mtx_enter(&stream->oa_buffer.ptr_lock) ; } while (0); |
1519 | |
1520 | intel_uncore_write(uncore, GEN12_OAG_OASTATUS((const i915_reg_t){ .reg = (0xdafc) }), 0); |
1521 | intel_uncore_write(uncore, GEN12_OAG_OAHEADPTR((const i915_reg_t){ .reg = (0xdb00) }), |
1522 | gtt_offset & GEN12_OAG_OAHEADPTR_MASK0xffffffc0); |
1523 | stream->oa_buffer.head = gtt_offset; |
1524 | |
1525 | /* |
1526 | * PRM says: |
1527 | * |
1528 | * "This MMIO must be set before the OATAILPTR |
1529 | * register and after the OAHEADPTR register. This is |
1530 | * to enable proper functionality of the overflow |
1531 | * bit." |
1532 | */ |
1533 | intel_uncore_write(uncore, GEN12_OAG_OABUFFER((const i915_reg_t){ .reg = (0xdb08) }), gtt_offset | |
1534 | OABUFFER_SIZE_16M(7 << 3) | GEN8_OABUFFER_MEM_SELECT_GGTT(1 << 0)); |
1535 | intel_uncore_write(uncore, GEN12_OAG_OATAILPTR((const i915_reg_t){ .reg = (0xdb04) }), |
1536 | gtt_offset & GEN12_OAG_OATAILPTR_MASK0xffffffc0); |
1537 | |
1538 | /* Mark that we need updated tail pointers to read from... */ |
1539 | stream->oa_buffer.aging_tail = INVALID_TAIL_PTR0xffffffff; |
1540 | stream->oa_buffer.tail = gtt_offset; |
1541 | |
1542 | /* |
1543 | * Reset state used to recognise context switches, affecting which |
1544 | * reports we will forward to userspace while filtering for a single |
1545 | * context. |
1546 | */ |
1547 | stream->oa_buffer.last_ctx_id = INVALID_CTX_ID0xffffffff; |
1548 | |
1549 | spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags)do { (void)(flags); mtx_leave(&stream->oa_buffer.ptr_lock ); } while (0); |
1550 | |
1551 | /* |
1552 | * NB: although the OA buffer will initially be allocated |
1553 | * zeroed via shmfs (and so this memset is redundant when |
1554 | * first allocating), we may re-init the OA buffer, either |
1555 | * when re-enabling a stream or in error/reset paths. |
1556 | * |
1557 | * The reason we clear the buffer for each re-init is for the |
1558 | * sanity check in gen8_append_oa_reports() that looks at the |
1559 | * reason field to make sure it's non-zero which relies on |
1560 | * the assumption that new reports are being written to zeroed |
1561 | * memory... |
1562 | */ |
1563 | memset(stream->oa_buffer.vaddr, 0,__builtin_memset((stream->oa_buffer.vaddr), (0), (stream-> oa_buffer.vma->size)) |
1564 | stream->oa_buffer.vma->size)__builtin_memset((stream->oa_buffer.vaddr), (0), (stream-> oa_buffer.vma->size)); |
1565 | } |
1566 | |
1567 | static int alloc_oa_buffer(struct i915_perf_stream *stream) |
1568 | { |
1569 | struct drm_i915_privateinteldrm_softc *i915 = stream->perf->i915; |
1570 | struct drm_i915_gem_object *bo; |
1571 | struct i915_vma *vma; |
1572 | int ret; |
1573 | |
1574 | if (drm_WARN_ON(&i915->drm, stream->oa_buffer.vma)({ int __ret = !!((stream->oa_buffer.vma)); if (__ret) printf ("%s %s: " "%s", dev_driver_string(((&i915->drm))-> dev), "", "drm_WARN_ON(" "stream->oa_buffer.vma" ")"); __builtin_expect (!!(__ret), 0); })) |
1575 | return -ENODEV19; |
1576 | |
1577 | BUILD_BUG_ON_NOT_POWER_OF_2(OA_BUFFER_SIZE)0; |
1578 | BUILD_BUG_ON(OA_BUFFER_SIZE < SZ_128K || OA_BUFFER_SIZE > SZ_16M)extern char _ctassert[(!((1024 * 1024 * 16) < (1024 * 128) || (1024 * 1024 * 16) > (1024 * 1024 * 16))) ? 1 : -1 ] __attribute__ ((__unused__)); |
1579 | |
1580 | bo = i915_gem_object_create_shmem(stream->perf->i915, OA_BUFFER_SIZE(1024 * 1024 * 16)); |
1581 | if (IS_ERR(bo)) { |
1582 | drm_err(&i915->drm, "Failed to allocate OA buffer\n")printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "Failed to allocate OA buffer\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__); |
1583 | return PTR_ERR(bo); |
1584 | } |
1585 | |
1586 | i915_gem_object_set_cache_coherency(bo, I915_CACHE_LLC); |
1587 | |
1588 | /* PreHSW required 512K alignment, HSW requires 16M */ |
1589 | vma = i915_gem_object_ggtt_pin(bo, NULL((void *)0), 0, SZ_16M(1024 * 1024 * 16), 0); |
1590 | if (IS_ERR(vma)) { |
1591 | ret = PTR_ERR(vma); |
1592 | goto err_unref; |
1593 | } |
1594 | stream->oa_buffer.vma = vma; |
1595 | |
1596 | stream->oa_buffer.vaddr = |
1597 | i915_gem_object_pin_map(bo, I915_MAP_WB); |
1598 | if (IS_ERR(stream->oa_buffer.vaddr)) { |
1599 | ret = PTR_ERR(stream->oa_buffer.vaddr); |
1600 | goto err_unpin; |
1601 | } |
1602 | |
1603 | return 0; |
1604 | |
1605 | err_unpin: |
1606 | __i915_vma_unpin(vma); |
1607 | |
1608 | err_unref: |
1609 | i915_gem_object_put(bo); |
1610 | |
1611 | stream->oa_buffer.vaddr = NULL((void *)0); |
1612 | stream->oa_buffer.vma = NULL((void *)0); |
1613 | |
1614 | return ret; |
1615 | } |
1616 | |
1617 | static u32 *save_restore_register(struct i915_perf_stream *stream, u32 *cs, |
1618 | bool_Bool save, i915_reg_t reg, u32 offset, |
1619 | u32 dword_count) |
1620 | { |
1621 | u32 cmd; |
1622 | u32 d; |
1623 | |
1624 | cmd = save ? MI_STORE_REGISTER_MEM(((0x24) << 23) | (1)) : MI_LOAD_REGISTER_MEM(((0x29) << 23) | (1)); |
1625 | cmd |= MI_SRM_LRM_GLOBAL_GTT(1<<22); |
1626 | if (INTEL_GEN(stream->perf->i915)((&(stream->perf->i915)->__info)->gen) >= 8) |
1627 | cmd++; |
1628 | |
1629 | for (d = 0; d < dword_count; d++) { |
1630 | *cs++ = cmd; |
1631 | *cs++ = i915_mmio_reg_offset(reg) + 4 * d; |
1632 | *cs++ = intel_gt_scratch_offset(stream->engine->gt, |
1633 | offset) + 4 * d; |
1634 | *cs++ = 0; |
1635 | } |
1636 | |
1637 | return cs; |
1638 | } |
1639 | |
1640 | static int alloc_noa_wait(struct i915_perf_stream *stream) |
1641 | { |
1642 | struct drm_i915_privateinteldrm_softc *i915 = stream->perf->i915; |
1643 | struct drm_i915_gem_object *bo; |
1644 | struct i915_vma *vma; |
1645 | const u64 delay_ticks = 0xffffffffffffffff - |
1646 | i915_cs_timestamp_ns_to_ticks(i915, atomic64_read(&stream->perf->noa_programming_delay)({ typeof(*(&stream->perf->noa_programming_delay)) __tmp = *(volatile typeof(*(&stream->perf->noa_programming_delay )) *)&(*(&stream->perf->noa_programming_delay)) ; membar_datadep_consumer(); __tmp; })); |
1647 | const u32 base = stream->engine->mmio_base; |
1648 | #define CS_GPR(x)((const i915_reg_t){ .reg = ((base) + 0x600 + (x) * 8) }) GEN8_RING_CS_GPR(base, x)((const i915_reg_t){ .reg = ((base) + 0x600 + (x) * 8) }) |
1649 | u32 *batch, *ts0, *cs, *jump; |
1650 | int ret, i; |
1651 | enum { |
1652 | START_TS, |
1653 | NOW_TS, |
1654 | DELTA_TS, |
1655 | JUMP_PREDICATE, |
1656 | DELTA_TARGET, |
1657 | N_CS_GPR |
1658 | }; |
1659 | |
1660 | bo = i915_gem_object_create_internal(i915, 4096); |
1661 | if (IS_ERR(bo)) { |
1662 | drm_err(&i915->drm,printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "Failed to allocate NOA wait batchbuffer\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__) |
1663 | "Failed to allocate NOA wait batchbuffer\n")printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "Failed to allocate NOA wait batchbuffer\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__); |
1664 | return PTR_ERR(bo); |
1665 | } |
1666 | |
1667 | /* |
1668 | * We pin in GGTT because we jump into this buffer now because |
1669 | * multiple OA config BOs will have a jump to this address and it |
1670 | * needs to be fixed during the lifetime of the i915/perf stream. |
1671 | */ |
1672 | vma = i915_gem_object_ggtt_pin(bo, NULL((void *)0), 0, 0, PIN_HIGH(1ULL << (5))); |
1673 | if (IS_ERR(vma)) { |
1674 | ret = PTR_ERR(vma); |
1675 | goto err_unref; |
1676 | } |
1677 | |
1678 | batch = cs = i915_gem_object_pin_map(bo, I915_MAP_WB); |
1679 | if (IS_ERR(batch)) { |
1680 | ret = PTR_ERR(batch); |
1681 | goto err_unpin; |
1682 | } |
1683 | |
1684 | /* Save registers. */ |
1685 | for (i = 0; i < N_CS_GPR; i++) |
1686 | cs = save_restore_register( |
1687 | stream, cs, true1 /* save */, CS_GPR(i)((const i915_reg_t){ .reg = ((base) + 0x600 + (i) * 8) }), |
1688 | INTEL_GT_SCRATCH_FIELD_PERF_CS_GPR + 8 * i, 2); |
1689 | cs = save_restore_register( |
1690 | stream, cs, true1 /* save */, MI_PREDICATE_RESULT_1((const i915_reg_t){ .reg = (0x241c) }), |
1691 | INTEL_GT_SCRATCH_FIELD_PERF_PREDICATE_RESULT_1, 1); |
1692 | |
1693 | /* First timestamp snapshot location. */ |
1694 | ts0 = cs; |
1695 | |
1696 | /* |
1697 | * Initial snapshot of the timestamp register to implement the wait. |
1698 | * We work with 32b values, so clear out the top 32b bits of the |
1699 | * register because the ALU works 64bits. |
1700 | */ |
1701 | *cs++ = MI_LOAD_REGISTER_IMM(1)(((0x22) << 23) | (2*(1)-1)); |
1702 | *cs++ = i915_mmio_reg_offset(CS_GPR(START_TS)((const i915_reg_t){ .reg = ((base) + 0x600 + (START_TS) * 8) })) + 4; |
1703 | *cs++ = 0; |
1704 | *cs++ = MI_LOAD_REGISTER_REG(((0x2A) << 23) | (1)) | (3 - 2); |
1705 | *cs++ = i915_mmio_reg_offset(RING_TIMESTAMP(base)((const i915_reg_t){ .reg = ((base) + 0x358) })); |
1706 | *cs++ = i915_mmio_reg_offset(CS_GPR(START_TS)((const i915_reg_t){ .reg = ((base) + 0x600 + (START_TS) * 8) })); |
1707 | |
1708 | /* |
1709 | * This is the location we're going to jump back into until the |
1710 | * required amount of time has passed. |
1711 | */ |
1712 | jump = cs; |
1713 | |
1714 | /* |
1715 | * Take another snapshot of the timestamp register. Take care to clear |
1716 | * up the top 32bits of CS_GPR(1) as we're using it for other |
1717 | * operations below. |
1718 | */ |
1719 | *cs++ = MI_LOAD_REGISTER_IMM(1)(((0x22) << 23) | (2*(1)-1)); |
1720 | *cs++ = i915_mmio_reg_offset(CS_GPR(NOW_TS)((const i915_reg_t){ .reg = ((base) + 0x600 + (NOW_TS) * 8) } )) + 4; |
1721 | *cs++ = 0; |
1722 | *cs++ = MI_LOAD_REGISTER_REG(((0x2A) << 23) | (1)) | (3 - 2); |
1723 | *cs++ = i915_mmio_reg_offset(RING_TIMESTAMP(base)((const i915_reg_t){ .reg = ((base) + 0x358) })); |
1724 | *cs++ = i915_mmio_reg_offset(CS_GPR(NOW_TS)((const i915_reg_t){ .reg = ((base) + 0x600 + (NOW_TS) * 8) } )); |
1725 | |
1726 | /* |
1727 | * Do a diff between the 2 timestamps and store the result back into |
1728 | * CS_GPR(1). |
1729 | */ |
1730 | *cs++ = MI_MATH(5)(((0x1a) << 23) | ((5) - 1)); |
1731 | *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(NOW_TS))((0x080) << 20 | (0x20) << 10 | ((NOW_TS))); |
1732 | *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(START_TS))((0x080) << 20 | (0x21) << 10 | ((START_TS))); |
1733 | *cs++ = MI_MATH_SUB((0x101) << 20 | (0x0) << 10 | (0x0)); |
1734 | *cs++ = MI_MATH_STORE(MI_MATH_REG(DELTA_TS), MI_MATH_REG_ACCU)((0x180) << 20 | ((DELTA_TS)) << 10 | (0x31)); |
1735 | *cs++ = MI_MATH_STORE(MI_MATH_REG(JUMP_PREDICATE), MI_MATH_REG_CF)((0x180) << 20 | ((JUMP_PREDICATE)) << 10 | (0x33 )); |
1736 | |
1737 | /* |
1738 | * Transfer the carry flag (set to 1 if ts1 < ts0, meaning the |
1739 | * timestamp have rolled over the 32bits) into the predicate register |
1740 | * to be used for the predicated jump. |
1741 | */ |
1742 | *cs++ = MI_LOAD_REGISTER_REG(((0x2A) << 23) | (1)) | (3 - 2); |
1743 | *cs++ = i915_mmio_reg_offset(CS_GPR(JUMP_PREDICATE)((const i915_reg_t){ .reg = ((base) + 0x600 + (JUMP_PREDICATE ) * 8) })); |
1744 | *cs++ = i915_mmio_reg_offset(MI_PREDICATE_RESULT_1((const i915_reg_t){ .reg = (0x241c) })); |
1745 | |
1746 | /* Restart from the beginning if we had timestamps roll over. */ |
1747 | *cs++ = (INTEL_GEN(i915)((&(i915)->__info)->gen) < 8 ? |
1748 | MI_BATCH_BUFFER_START(((0x31) << 23) | (0)) : |
1749 | MI_BATCH_BUFFER_START_GEN8(((0x31) << 23) | (1))) | |
1750 | MI_BATCH_PREDICATE((u32)((1UL << (15)) + 0)); |
1751 | *cs++ = i915_ggtt_offset(vma) + (ts0 - batch) * 4; |
1752 | *cs++ = 0; |
1753 | |
1754 | /* |
1755 | * Now add the diff between to previous timestamps and add it to : |
1756 | * (((1 * << 64) - 1) - delay_ns) |
1757 | * |
1758 | * When the Carry Flag contains 1 this means the elapsed time is |
1759 | * longer than the expected delay, and we can exit the wait loop. |
1760 | */ |
1761 | *cs++ = MI_LOAD_REGISTER_IMM(2)(((0x22) << 23) | (2*(2)-1)); |
1762 | *cs++ = i915_mmio_reg_offset(CS_GPR(DELTA_TARGET)((const i915_reg_t){ .reg = ((base) + 0x600 + (DELTA_TARGET) * 8) })); |
1763 | *cs++ = lower_32_bits(delay_ticks)((u32)(delay_ticks)); |
1764 | *cs++ = i915_mmio_reg_offset(CS_GPR(DELTA_TARGET)((const i915_reg_t){ .reg = ((base) + 0x600 + (DELTA_TARGET) * 8) })) + 4; |
1765 | *cs++ = upper_32_bits(delay_ticks)((u32)(((delay_ticks) >> 16) >> 16)); |
1766 | |
1767 | *cs++ = MI_MATH(4)(((0x1a) << 23) | ((4) - 1)); |
1768 | *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(DELTA_TS))((0x080) << 20 | (0x20) << 10 | ((DELTA_TS))); |
1769 | *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(DELTA_TARGET))((0x080) << 20 | (0x21) << 10 | ((DELTA_TARGET))); |
1770 | *cs++ = MI_MATH_ADD((0x100) << 20 | (0x0) << 10 | (0x0)); |
1771 | *cs++ = MI_MATH_STOREINV(MI_MATH_REG(JUMP_PREDICATE), MI_MATH_REG_CF)((0x580) << 20 | ((JUMP_PREDICATE)) << 10 | (0x33 )); |
1772 | |
1773 | *cs++ = MI_ARB_CHECK(((0x05) << 23) | (0)); |
1774 | |
1775 | /* |
1776 | * Transfer the result into the predicate register to be used for the |
1777 | * predicated jump. |
1778 | */ |
1779 | *cs++ = MI_LOAD_REGISTER_REG(((0x2A) << 23) | (1)) | (3 - 2); |
1780 | *cs++ = i915_mmio_reg_offset(CS_GPR(JUMP_PREDICATE)((const i915_reg_t){ .reg = ((base) + 0x600 + (JUMP_PREDICATE ) * 8) })); |
1781 | *cs++ = i915_mmio_reg_offset(MI_PREDICATE_RESULT_1((const i915_reg_t){ .reg = (0x241c) })); |
1782 | |
1783 | /* Predicate the jump. */ |
1784 | *cs++ = (INTEL_GEN(i915)((&(i915)->__info)->gen) < 8 ? |
1785 | MI_BATCH_BUFFER_START(((0x31) << 23) | (0)) : |
1786 | MI_BATCH_BUFFER_START_GEN8(((0x31) << 23) | (1))) | |
1787 | MI_BATCH_PREDICATE((u32)((1UL << (15)) + 0)); |
1788 | *cs++ = i915_ggtt_offset(vma) + (jump - batch) * 4; |
1789 | *cs++ = 0; |
1790 | |
1791 | /* Restore registers. */ |
1792 | for (i = 0; i < N_CS_GPR; i++) |
1793 | cs = save_restore_register( |
1794 | stream, cs, false0 /* restore */, CS_GPR(i)((const i915_reg_t){ .reg = ((base) + 0x600 + (i) * 8) }), |
1795 | INTEL_GT_SCRATCH_FIELD_PERF_CS_GPR + 8 * i, 2); |
1796 | cs = save_restore_register( |
1797 | stream, cs, false0 /* restore */, MI_PREDICATE_RESULT_1((const i915_reg_t){ .reg = (0x241c) }), |
1798 | INTEL_GT_SCRATCH_FIELD_PERF_PREDICATE_RESULT_1, 1); |
1799 | |
1800 | /* And return to the ring. */ |
1801 | *cs++ = MI_BATCH_BUFFER_END(((0x0a) << 23) | (0)); |
1802 | |
1803 | GEM_BUG_ON(cs - batch > PAGE_SIZE / sizeof(*batch))((void)0); |
1804 | |
1805 | i915_gem_object_flush_map(bo); |
1806 | __i915_gem_object_release_map(bo); |
1807 | |
1808 | stream->noa_wait = vma; |
1809 | return 0; |
1810 | |
1811 | err_unpin: |
1812 | i915_vma_unpin_and_release(&vma, 0); |
1813 | err_unref: |
1814 | i915_gem_object_put(bo); |
1815 | return ret; |
1816 | } |
1817 | |
1818 | static u32 *write_cs_mi_lri(u32 *cs, |
1819 | const struct i915_oa_reg *reg_data, |
1820 | u32 n_regs) |
1821 | { |
1822 | u32 i; |
1823 | |
1824 | for (i = 0; i < n_regs; i++) { |
1825 | if ((i % MI_LOAD_REGISTER_IMM_MAX_REGS(126)) == 0) { |
1826 | u32 n_lri = min_t(u32,({ u32 __min_a = (n_regs - i); u32 __min_b = ((126)); __min_a < __min_b ? __min_a : __min_b; }) |
1827 | n_regs - i,({ u32 __min_a = (n_regs - i); u32 __min_b = ((126)); __min_a < __min_b ? __min_a : __min_b; }) |
1828 | MI_LOAD_REGISTER_IMM_MAX_REGS)({ u32 __min_a = (n_regs - i); u32 __min_b = ((126)); __min_a < __min_b ? __min_a : __min_b; }); |
1829 | |
1830 | *cs++ = MI_LOAD_REGISTER_IMM(n_lri)(((0x22) << 23) | (2*(n_lri)-1)); |
1831 | } |
1832 | *cs++ = i915_mmio_reg_offset(reg_data[i].addr); |
1833 | *cs++ = reg_data[i].value; |
1834 | } |
1835 | |
1836 | return cs; |
1837 | } |
1838 | |
1839 | static int num_lri_dwords(int num_regs) |
1840 | { |
1841 | int count = 0; |
1842 | |
1843 | if (num_regs > 0) { |
1844 | count += DIV_ROUND_UP(num_regs, MI_LOAD_REGISTER_IMM_MAX_REGS)(((num_regs) + (((126)) - 1)) / ((126))); |
1845 | count += num_regs * 2; |
1846 | } |
1847 | |
1848 | return count; |
1849 | } |
1850 | |
1851 | static struct i915_oa_config_bo * |
1852 | alloc_oa_config_buffer(struct i915_perf_stream *stream, |
1853 | struct i915_oa_config *oa_config) |
1854 | { |
1855 | struct drm_i915_gem_object *obj; |
1856 | struct i915_oa_config_bo *oa_bo; |
1857 | size_t config_length = 0; |
1858 | u32 *cs; |
1859 | int err; |
1860 | |
1861 | oa_bo = kzalloc(sizeof(*oa_bo), GFP_KERNEL(0x0001 | 0x0004)); |
1862 | if (!oa_bo) |
1863 | return ERR_PTR(-ENOMEM12); |
1864 | |
1865 | config_length += num_lri_dwords(oa_config->mux_regs_len); |
1866 | config_length += num_lri_dwords(oa_config->b_counter_regs_len); |
1867 | config_length += num_lri_dwords(oa_config->flex_regs_len); |
1868 | config_length += 3; /* MI_BATCH_BUFFER_START */ |
1869 | config_length = roundup2(sizeof(u32) * config_length, I915_GTT_PAGE_SIZE)(((sizeof(u32) * config_length) + (((1ULL << (12))) - 1 )) & (~((__typeof(sizeof(u32) * config_length))((1ULL << (12))) - 1))); |
1870 | |
1871 | obj = i915_gem_object_create_shmem(stream->perf->i915, config_length); |
1872 | if (IS_ERR(obj)) { |
1873 | err = PTR_ERR(obj); |
1874 | goto err_free; |
1875 | } |
1876 | |
1877 | cs = i915_gem_object_pin_map(obj, I915_MAP_WB); |
1878 | if (IS_ERR(cs)) { |
1879 | err = PTR_ERR(cs); |
1880 | goto err_oa_bo; |
1881 | } |
1882 | |
1883 | cs = write_cs_mi_lri(cs, |
1884 | oa_config->mux_regs, |
1885 | oa_config->mux_regs_len); |
1886 | cs = write_cs_mi_lri(cs, |
1887 | oa_config->b_counter_regs, |
1888 | oa_config->b_counter_regs_len); |
1889 | cs = write_cs_mi_lri(cs, |
1890 | oa_config->flex_regs, |
1891 | oa_config->flex_regs_len); |
1892 | |
1893 | /* Jump into the active wait. */ |
1894 | *cs++ = (INTEL_GEN(stream->perf->i915)((&(stream->perf->i915)->__info)->gen) < 8 ? |
1895 | MI_BATCH_BUFFER_START(((0x31) << 23) | (0)) : |
1896 | MI_BATCH_BUFFER_START_GEN8(((0x31) << 23) | (1))); |
1897 | *cs++ = i915_ggtt_offset(stream->noa_wait); |
1898 | *cs++ = 0; |
1899 | |
1900 | i915_gem_object_flush_map(obj); |
1901 | __i915_gem_object_release_map(obj); |
1902 | |
1903 | oa_bo->vma = i915_vma_instance(obj, |
1904 | &stream->engine->gt->ggtt->vm, |
1905 | NULL((void *)0)); |
1906 | if (IS_ERR(oa_bo->vma)) { |
1907 | err = PTR_ERR(oa_bo->vma); |
1908 | goto err_oa_bo; |
1909 | } |
1910 | |
1911 | oa_bo->oa_config = i915_oa_config_get(oa_config); |
1912 | llist_add(&oa_bo->node, &stream->oa_config_bos); |
1913 | |
1914 | return oa_bo; |
1915 | |
1916 | err_oa_bo: |
1917 | i915_gem_object_put(obj); |
1918 | err_free: |
1919 | kfree(oa_bo); |
1920 | return ERR_PTR(err); |
1921 | } |
1922 | |
1923 | static struct i915_vma * |
1924 | get_oa_vma(struct i915_perf_stream *stream, struct i915_oa_config *oa_config) |
1925 | { |
1926 | struct i915_oa_config_bo *oa_bo; |
1927 | |
1928 | /* |
1929 | * Look for the buffer in the already allocated BOs attached |
1930 | * to the stream. |
1931 | */ |
1932 | llist_for_each_entry(oa_bo, stream->oa_config_bos.first, node)for ((oa_bo) = (((stream->oa_config_bos.first)) ? ({ const __typeof( ((__typeof(*(oa_bo)) *)0)->node ) *__mptr = ((stream ->oa_config_bos.first)); (__typeof(*(oa_bo)) *)( (char *)__mptr - __builtin_offsetof(__typeof(*(oa_bo)), node) );}) : ((void *)0)); (oa_bo) != ((void *)0); (oa_bo) = (((oa_bo)->node. next) ? ({ const __typeof( ((__typeof(*(oa_bo)) *)0)->node ) *__mptr = ((oa_bo)->node.next); (__typeof(*(oa_bo)) *)( (char *)__mptr - __builtin_offsetof(__typeof(*(oa_bo)), node ) );}) : ((void *)0))) { |
1933 | if (oa_bo->oa_config == oa_config && |
1934 | memcmp(oa_bo->oa_config->uuid,__builtin_memcmp((oa_bo->oa_config->uuid), (oa_config-> uuid), (sizeof(oa_config->uuid))) |
1935 | oa_config->uuid,__builtin_memcmp((oa_bo->oa_config->uuid), (oa_config-> uuid), (sizeof(oa_config->uuid))) |
1936 | sizeof(oa_config->uuid))__builtin_memcmp((oa_bo->oa_config->uuid), (oa_config-> uuid), (sizeof(oa_config->uuid))) == 0) |
1937 | goto out; |
1938 | } |
1939 | |
1940 | oa_bo = alloc_oa_config_buffer(stream, oa_config); |
1941 | if (IS_ERR(oa_bo)) |
1942 | return ERR_CAST(oa_bo); |
1943 | |
1944 | out: |
1945 | return i915_vma_get(oa_bo->vma); |
1946 | } |
1947 | |
1948 | static int |
1949 | emit_oa_config(struct i915_perf_stream *stream, |
1950 | struct i915_oa_config *oa_config, |
1951 | struct intel_context *ce, |
1952 | struct i915_active *active) |
1953 | { |
1954 | struct i915_request *rq; |
1955 | struct i915_vma *vma; |
1956 | struct i915_gem_ww_ctx ww; |
1957 | int err; |
1958 | |
1959 | vma = get_oa_vma(stream, oa_config); |
1960 | if (IS_ERR(vma)) |
1961 | return PTR_ERR(vma); |
1962 | |
1963 | i915_gem_ww_ctx_init(&ww, true1); |
1964 | retry: |
1965 | err = i915_gem_object_lock(vma->obj, &ww); |
1966 | if (err) |
1967 | goto err; |
1968 | |
1969 | err = i915_vma_pin_ww(vma, &ww, 0, 0, PIN_GLOBAL(1ULL << (10)) | PIN_HIGH(1ULL << (5))); |
1970 | if (err) |
1971 | goto err; |
1972 | |
1973 | intel_engine_pm_get(ce->engine); |
1974 | rq = i915_request_create(ce); |
1975 | intel_engine_pm_put(ce->engine); |
1976 | if (IS_ERR(rq)) { |
1977 | err = PTR_ERR(rq); |
1978 | goto err_vma_unpin; |
1979 | } |
1980 | |
1981 | if (!IS_ERR_OR_NULL(active)) { |
1982 | /* After all individual context modifications */ |
1983 | err = i915_request_await_active(rq, active, |
1984 | I915_ACTIVE_AWAIT_ACTIVE(1UL << (1))); |
1985 | if (err) |
1986 | goto err_add_request; |
1987 | |
1988 | err = i915_active_add_request(active, rq); |
1989 | if (err) |
1990 | goto err_add_request; |
1991 | } |
1992 | |
1993 | err = i915_request_await_object(rq, vma->obj, 0); |
1994 | if (!err) |
1995 | err = i915_vma_move_to_active(vma, rq, 0); |
1996 | if (err) |
1997 | goto err_add_request; |
1998 | |
1999 | err = rq->engine->emit_bb_start(rq, |
2000 | vma->node.start, 0, |
2001 | I915_DISPATCH_SECURE(1UL << (0))); |
2002 | if (err) |
2003 | goto err_add_request; |
2004 | |
2005 | err_add_request: |
2006 | i915_request_add(rq); |
2007 | err_vma_unpin: |
2008 | i915_vma_unpin(vma); |
2009 | err: |
2010 | if (err == -EDEADLK11) { |
2011 | err = i915_gem_ww_ctx_backoff(&ww); |
2012 | if (!err) |
2013 | goto retry; |
2014 | } |
2015 | |
2016 | i915_gem_ww_ctx_fini(&ww); |
2017 | i915_vma_put(vma); |
2018 | return err; |
2019 | } |
2020 | |
2021 | static struct intel_context *oa_context(struct i915_perf_stream *stream) |
2022 | { |
2023 | return stream->pinned_ctx ?: stream->engine->kernel_context; |
2024 | } |
2025 | |
2026 | static int |
2027 | hsw_enable_metric_set(struct i915_perf_stream *stream, |
2028 | struct i915_active *active) |
2029 | { |
2030 | struct intel_uncore *uncore = stream->uncore; |
2031 | |
2032 | /* |
2033 | * PRM: |
2034 | * |
2035 | * OA unit is using “crclk” for its functionality. When trunk |
2036 | * level clock gating takes place, OA clock would be gated, |
2037 | * unable to count the events from non-render clock domain. |
2038 | * Render clock gating must be disabled when OA is enabled to |
2039 | * count the events from non-render domain. Unit level clock |
2040 | * gating for RCS should also be disabled. |
2041 | */ |
2042 | intel_uncore_rmw(uncore, GEN7_MISCCPCTL((const i915_reg_t){ .reg = (0x9424) }), |
2043 | GEN7_DOP_CLOCK_GATE_ENABLE(1 << 0), 0); |
2044 | intel_uncore_rmw(uncore, GEN6_UCGCTL1((const i915_reg_t){ .reg = (0x9400) }), |
2045 | 0, GEN6_CSUNIT_CLOCK_GATE_DISABLE(1 << 7)); |
2046 | |
2047 | return emit_oa_config(stream, |
2048 | stream->oa_config, oa_context(stream), |
2049 | active); |
2050 | } |
2051 | |
2052 | static void hsw_disable_metric_set(struct i915_perf_stream *stream) |
2053 | { |
2054 | struct intel_uncore *uncore = stream->uncore; |
2055 | |
2056 | intel_uncore_rmw(uncore, GEN6_UCGCTL1((const i915_reg_t){ .reg = (0x9400) }), |
2057 | GEN6_CSUNIT_CLOCK_GATE_DISABLE(1 << 7), 0); |
2058 | intel_uncore_rmw(uncore, GEN7_MISCCPCTL((const i915_reg_t){ .reg = (0x9424) }), |
2059 | 0, GEN7_DOP_CLOCK_GATE_ENABLE(1 << 0)); |
2060 | |
2061 | intel_uncore_rmw(uncore, GDT_CHICKEN_BITS((const i915_reg_t){ .reg = (0x9840) }), GT_NOA_ENABLE0x00000080, 0); |
2062 | } |
2063 | |
2064 | static u32 oa_config_flex_reg(const struct i915_oa_config *oa_config, |
2065 | i915_reg_t reg) |
2066 | { |
2067 | u32 mmio = i915_mmio_reg_offset(reg); |
2068 | int i; |
2069 | |
2070 | /* |
2071 | * This arbitrary default will select the 'EU FPU0 Pipeline |
2072 | * Active' event. In the future it's anticipated that there |
2073 | * will be an explicit 'No Event' we can select, but not yet... |
2074 | */ |
2075 | if (!oa_config) |
2076 | return 0; |
2077 | |
2078 | for (i = 0; i < oa_config->flex_regs_len; i++) { |
2079 | if (i915_mmio_reg_offset(oa_config->flex_regs[i].addr) == mmio) |
2080 | return oa_config->flex_regs[i].value; |
2081 | } |
2082 | |
2083 | return 0; |
2084 | } |
2085 | /* |
2086 | * NB: It must always remain pointer safe to run this even if the OA unit |
2087 | * has been disabled. |
2088 | * |
2089 | * It's fine to put out-of-date values into these per-context registers |
2090 | * in the case that the OA unit has been disabled. |
2091 | */ |
2092 | static void |
2093 | gen8_update_reg_state_unlocked(const struct intel_context *ce, |
2094 | const struct i915_perf_stream *stream) |
2095 | { |
2096 | u32 ctx_oactxctrl = stream->perf->ctx_oactxctrl_offset; |
2097 | u32 ctx_flexeu0 = stream->perf->ctx_flexeu0_offset; |
2098 | /* The MMIO offsets for Flex EU registers aren't contiguous */ |
2099 | i915_reg_t flex_regs[] = { |
2100 | EU_PERF_CNTL0((const i915_reg_t){ .reg = (0xe458) }), |
2101 | EU_PERF_CNTL1((const i915_reg_t){ .reg = (0xe558) }), |
2102 | EU_PERF_CNTL2((const i915_reg_t){ .reg = (0xe658) }), |
2103 | EU_PERF_CNTL3((const i915_reg_t){ .reg = (0xe758) }), |
2104 | EU_PERF_CNTL4((const i915_reg_t){ .reg = (0xe45c) }), |
2105 | EU_PERF_CNTL5((const i915_reg_t){ .reg = (0xe55c) }), |
2106 | EU_PERF_CNTL6((const i915_reg_t){ .reg = (0xe65c) }), |
2107 | }; |
2108 | u32 *reg_state = ce->lrc_reg_state; |
2109 | int i; |
2110 | |
2111 | reg_state[ctx_oactxctrl + 1] = |
2112 | (stream->period_exponent << GEN8_OA_TIMER_PERIOD_SHIFT2) | |
2113 | (stream->periodic ? GEN8_OA_TIMER_ENABLE(1 << 1) : 0) | |
2114 | GEN8_OA_COUNTER_RESUME(1 << 0); |
2115 | |
2116 | for (i = 0; i < ARRAY_SIZE(flex_regs)(sizeof((flex_regs)) / sizeof((flex_regs)[0])); i++) |
2117 | reg_state[ctx_flexeu0 + i * 2 + 1] = |
2118 | oa_config_flex_reg(stream->oa_config, flex_regs[i]); |
2119 | } |
2120 | |
2121 | struct flex { |
2122 | i915_reg_t reg; |
2123 | u32 offset; |
2124 | u32 value; |
2125 | }; |
2126 | |
2127 | static int |
2128 | gen8_store_flex(struct i915_request *rq, |
2129 | struct intel_context *ce, |
2130 | const struct flex *flex, unsigned int count) |
2131 | { |
2132 | u32 offset; |
2133 | u32 *cs; |
2134 | |
2135 | cs = intel_ring_begin(rq, 4 * count); |
2136 | if (IS_ERR(cs)) |
2137 | return PTR_ERR(cs); |
2138 | |
2139 | offset = i915_ggtt_offset(ce->state) + LRC_STATE_OFFSET(((0) + (1)) * (1 << 12)); |
2140 | do { |
2141 | *cs++ = MI_STORE_DWORD_IMM_GEN4(((0x20) << 23) | (2)) | MI_USE_GGTT(1 << 22); |
2142 | *cs++ = offset + flex->offset * sizeof(u32); |
2143 | *cs++ = 0; |
2144 | *cs++ = flex->value; |
2145 | } while (flex++, --count); |
2146 | |
2147 | intel_ring_advance(rq, cs); |
2148 | |
2149 | return 0; |
2150 | } |
2151 | |
2152 | static int |
2153 | gen8_load_flex(struct i915_request *rq, |
2154 | struct intel_context *ce, |
2155 | const struct flex *flex, unsigned int count) |
2156 | { |
2157 | u32 *cs; |
2158 | |
2159 | GEM_BUG_ON(!count || count > 63)((void)0); |
2160 | |
2161 | cs = intel_ring_begin(rq, 2 * count + 2); |
2162 | if (IS_ERR(cs)) |
2163 | return PTR_ERR(cs); |
2164 | |
2165 | *cs++ = MI_LOAD_REGISTER_IMM(count)(((0x22) << 23) | (2*(count)-1)); |
2166 | do { |
2167 | *cs++ = i915_mmio_reg_offset(flex->reg); |
2168 | *cs++ = flex->value; |
2169 | } while (flex++, --count); |
2170 | *cs++ = MI_NOOP(((0) << 23) | (0)); |
2171 | |
2172 | intel_ring_advance(rq, cs); |
2173 | |
2174 | return 0; |
2175 | } |
2176 | |
2177 | static int gen8_modify_context(struct intel_context *ce, |
2178 | const struct flex *flex, unsigned int count) |
2179 | { |
2180 | struct i915_request *rq; |
2181 | int err; |
2182 | |
2183 | rq = intel_engine_create_kernel_request(ce->engine); |
2184 | if (IS_ERR(rq)) |
2185 | return PTR_ERR(rq); |
2186 | |
2187 | /* Serialise with the remote context */ |
2188 | err = intel_context_prepare_remote_request(ce, rq); |
2189 | if (err == 0) |
2190 | err = gen8_store_flex(rq, ce, flex, count); |
2191 | |
2192 | i915_request_add(rq); |
2193 | return err; |
2194 | } |
2195 | |
2196 | static int |
2197 | gen8_modify_self(struct intel_context *ce, |
2198 | const struct flex *flex, unsigned int count, |
2199 | struct i915_active *active) |
2200 | { |
2201 | struct i915_request *rq; |
2202 | int err; |
2203 | |
2204 | intel_engine_pm_get(ce->engine); |
2205 | rq = i915_request_create(ce); |
2206 | intel_engine_pm_put(ce->engine); |
2207 | if (IS_ERR(rq)) |
2208 | return PTR_ERR(rq); |
2209 | |
2210 | if (!IS_ERR_OR_NULL(active)) { |
2211 | err = i915_active_add_request(active, rq); |
2212 | if (err) |
2213 | goto err_add_request; |
2214 | } |
2215 | |
2216 | err = gen8_load_flex(rq, ce, flex, count); |
2217 | if (err) |
2218 | goto err_add_request; |
2219 | |
2220 | err_add_request: |
2221 | i915_request_add(rq); |
2222 | return err; |
2223 | } |
2224 | |
2225 | static int gen8_configure_context(struct i915_gem_context *ctx, |
2226 | struct flex *flex, unsigned int count) |
2227 | { |
2228 | struct i915_gem_engines_iter it; |
2229 | struct intel_context *ce; |
2230 | int err = 0; |
2231 | |
2232 | for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it)for (i915_gem_engines_iter_init(&(it), (i915_gem_context_lock_engines (ctx))); ((ce) = i915_gem_engines_iter_next(&(it)));) { |
2233 | GEM_BUG_ON(ce == ce->engine->kernel_context)((void)0); |
2234 | |
2235 | if (ce->engine->class != RENDER_CLASS0) |
2236 | continue; |
2237 | |
2238 | /* Otherwise OA settings will be set upon first use */ |
2239 | if (!intel_context_pin_if_active(ce)) |
2240 | continue; |
2241 | |
2242 | flex->value = intel_sseu_make_rpcs(ce->engine->gt, &ce->sseu); |
2243 | err = gen8_modify_context(ce, flex, count); |
2244 | |
2245 | intel_context_unpin(ce); |
2246 | if (err) |
2247 | break; |
2248 | } |
2249 | i915_gem_context_unlock_engines(ctx); |
2250 | |
2251 | return err; |
2252 | } |
2253 | |
2254 | static int gen12_configure_oar_context(struct i915_perf_stream *stream, |
2255 | struct i915_active *active) |
2256 | { |
2257 | int err; |
2258 | struct intel_context *ce = stream->pinned_ctx; |
2259 | u32 format = stream->oa_buffer.format; |
2260 | struct flex regs_context[] = { |
2261 | { |
2262 | GEN8_OACTXCONTROL((const i915_reg_t){ .reg = (0x2360) }), |
2263 | stream->perf->ctx_oactxctrl_offset + 1, |
2264 | active ? GEN8_OA_COUNTER_RESUME(1 << 0) : 0, |
2265 | }, |
2266 | }; |
2267 | /* Offsets in regs_lri are not used since this configuration is only |
2268 | * applied using LRI. Initialize the correct offsets for posterity. |
2269 | */ |
2270 | #define GEN12_OAR_OACONTROL_OFFSET0x5B0 0x5B0 |
2271 | struct flex regs_lri[] = { |
2272 | { |
2273 | GEN12_OAR_OACONTROL((const i915_reg_t){ .reg = (0x2960) }), |
2274 | GEN12_OAR_OACONTROL_OFFSET0x5B0 + 1, |
2275 | (format << GEN12_OAR_OACONTROL_COUNTER_FORMAT_SHIFT1) | |
2276 | (active ? GEN12_OAR_OACONTROL_COUNTER_ENABLE(1 << 0) : 0) |
2277 | }, |
2278 | { |
2279 | RING_CONTEXT_CONTROL(ce->engine->mmio_base)((const i915_reg_t){ .reg = ((ce->engine->mmio_base) + 0x244 ) }), |
2280 | CTX_CONTEXT_CONTROL(0x02 + 1), |
2281 | _MASKED_FIELD(GEN12_CTX_CTRL_OAR_CONTEXT_ENABLE,({ if (__builtin_constant_p((1 << 8))) do { } while (0) ; if (__builtin_constant_p(active ? (1 << 8) : 0)) do { } while (0); if (__builtin_constant_p((1 << 8)) && __builtin_constant_p(active ? (1 << 8) : 0)) do { } while (0); (((1 << 8)) << 16 | (active ? (1 << 8 ) : 0)); }) |
2282 | active ?({ if (__builtin_constant_p((1 << 8))) do { } while (0) ; if (__builtin_constant_p(active ? (1 << 8) : 0)) do { } while (0); if (__builtin_constant_p((1 << 8)) && __builtin_constant_p(active ? (1 << 8) : 0)) do { } while (0); (((1 << 8)) << 16 | (active ? (1 << 8 ) : 0)); }) |
2283 | GEN12_CTX_CTRL_OAR_CONTEXT_ENABLE :({ if (__builtin_constant_p((1 << 8))) do { } while (0) ; if (__builtin_constant_p(active ? (1 << 8) : 0)) do { } while (0); if (__builtin_constant_p((1 << 8)) && __builtin_constant_p(active ? (1 << 8) : 0)) do { } while (0); (((1 << 8)) << 16 | (active ? (1 << 8 ) : 0)); }) |
2284 | 0)({ if (__builtin_constant_p((1 << 8))) do { } while (0) ; if (__builtin_constant_p(active ? (1 << 8) : 0)) do { } while (0); if (__builtin_constant_p((1 << 8)) && __builtin_constant_p(active ? (1 << 8) : 0)) do { } while (0); (((1 << 8)) << 16 | (active ? (1 << 8 ) : 0)); }) |
2285 | }, |
2286 | }; |
2287 | |
2288 | /* Modify the context image of pinned context with regs_context*/ |
2289 | err = intel_context_lock_pinned(ce); |
2290 | if (err) |
2291 | return err; |
2292 | |
2293 | err = gen8_modify_context(ce, regs_context, ARRAY_SIZE(regs_context)(sizeof((regs_context)) / sizeof((regs_context)[0]))); |
2294 | intel_context_unlock_pinned(ce); |
2295 | if (err) |
2296 | return err; |
2297 | |
2298 | /* Apply regs_lri using LRI with pinned context */ |
2299 | return gen8_modify_self(ce, regs_lri, ARRAY_SIZE(regs_lri)(sizeof((regs_lri)) / sizeof((regs_lri)[0])), active); |
2300 | } |
2301 | |
2302 | /* |
2303 | * Manages updating the per-context aspects of the OA stream |
2304 | * configuration across all contexts. |
2305 | * |
2306 | * The awkward consideration here is that OACTXCONTROL controls the |
2307 | * exponent for periodic sampling which is primarily used for system |
2308 | * wide profiling where we'd like a consistent sampling period even in |
2309 | * the face of context switches. |
2310 | * |
2311 | * Our approach of updating the register state context (as opposed to |
2312 | * say using a workaround batch buffer) ensures that the hardware |
2313 | * won't automatically reload an out-of-date timer exponent even |
2314 | * transiently before a WA BB could be parsed. |
2315 | * |
2316 | * This function needs to: |
2317 | * - Ensure the currently running context's per-context OA state is |
2318 | * updated |
2319 | * - Ensure that all existing contexts will have the correct per-context |
2320 | * OA state if they are scheduled for use. |
2321 | * - Ensure any new contexts will be initialized with the correct |
2322 | * per-context OA state. |
2323 | * |
2324 | * Note: it's only the RCS/Render context that has any OA state. |
2325 | * Note: the first flex register passed must always be R_PWR_CLK_STATE |
2326 | */ |
2327 | static int |
2328 | oa_configure_all_contexts(struct i915_perf_stream *stream, |
2329 | struct flex *regs, |
2330 | size_t num_regs, |
2331 | struct i915_active *active) |
2332 | { |
2333 | struct drm_i915_privateinteldrm_softc *i915 = stream->perf->i915; |
2334 | struct intel_engine_cs *engine; |
2335 | struct i915_gem_context *ctx, *cn; |
2336 | int err; |
2337 | |
2338 | lockdep_assert_held(&stream->perf->lock)do { (void)(&stream->perf->lock); } while(0); |
2339 | |
2340 | /* |
2341 | * The OA register config is setup through the context image. This image |
2342 | * might be written to by the GPU on context switch (in particular on |
2343 | * lite-restore). This means we can't safely update a context's image, |
2344 | * if this context is scheduled/submitted to run on the GPU. |
2345 | * |
2346 | * We could emit the OA register config through the batch buffer but |
2347 | * this might leave small interval of time where the OA unit is |
2348 | * configured at an invalid sampling period. |
2349 | * |
2350 | * Note that since we emit all requests from a single ring, there |
2351 | * is still an implicit global barrier here that may cause a high |
2352 | * priority context to wait for an otherwise independent low priority |
2353 | * context. Contexts idle at the time of reconfiguration are not |
2354 | * trapped behind the barrier. |
2355 | */ |
2356 | spin_lock(&i915->gem.contexts.lock)mtx_enter(&i915->gem.contexts.lock); |
2357 | list_for_each_entry_safe(ctx, cn, &i915->gem.contexts.list, link)for (ctx = ({ const __typeof( ((__typeof(*ctx) *)0)->link ) *__mptr = ((&i915->gem.contexts.list)->next); (__typeof (*ctx) *)( (char *)__mptr - __builtin_offsetof(__typeof(*ctx) , link) );}), cn = ({ const __typeof( ((__typeof(*ctx) *)0)-> link ) *__mptr = (ctx->link.next); (__typeof(*ctx) *)( (char *)__mptr - __builtin_offsetof(__typeof(*ctx), link) );}); & ctx->link != (&i915->gem.contexts.list); ctx = cn, cn = ({ const __typeof( ((__typeof(*cn) *)0)->link ) *__mptr = (cn->link.next); (__typeof(*cn) *)( (char *)__mptr - __builtin_offsetof (__typeof(*cn), link) );})) { |
2358 | if (!kref_get_unless_zero(&ctx->ref)) |
2359 | continue; |
2360 | |
2361 | spin_unlock(&i915->gem.contexts.lock)mtx_leave(&i915->gem.contexts.lock); |
2362 | |
2363 | err = gen8_configure_context(ctx, regs, num_regs); |
2364 | if (err) { |
2365 | i915_gem_context_put(ctx); |
2366 | return err; |
2367 | } |
2368 | |
2369 | spin_lock(&i915->gem.contexts.lock)mtx_enter(&i915->gem.contexts.lock); |
2370 | list_safe_reset_next(ctx, cn, link)cn = ({ const __typeof( ((typeof(*(ctx)) *)0)->link ) *__mptr = (((ctx)->link.next)); (typeof(*(ctx)) *)( (char *)__mptr - __builtin_offsetof(typeof(*(ctx)), link) );}); |
2371 | i915_gem_context_put(ctx); |
2372 | } |
2373 | spin_unlock(&i915->gem.contexts.lock)mtx_leave(&i915->gem.contexts.lock); |
2374 | |
2375 | /* |
2376 | * After updating all other contexts, we need to modify ourselves. |
2377 | * If we don't modify the kernel_context, we do not get events while |
2378 | * idle. |
2379 | */ |
2380 | for_each_uabi_engine(engine, i915)for ((engine) = (linux_root_RB_MINMAX((struct linux_root *)(& (i915)->uabi_engines), -1) ? ({ const __typeof( ((struct intel_engine_cs *)0)->uabi_node ) *__mptr = (linux_root_RB_MINMAX((struct linux_root *)(&(i915)->uabi_engines), -1)); (struct intel_engine_cs *)( (char *)__mptr - __builtin_offsetof(struct intel_engine_cs , uabi_node) );}) : ((void *)0)); (engine); (engine) = (linux_root_RB_NEXT ((&(engine)->uabi_node)) ? ({ const __typeof( ((struct intel_engine_cs *)0)->uabi_node ) *__mptr = (linux_root_RB_NEXT ((&(engine)->uabi_node))); (struct intel_engine_cs *)( (char *)__mptr - __builtin_offsetof(struct intel_engine_cs, uabi_node ) );}) : ((void *)0))) { |
2381 | struct intel_context *ce = engine->kernel_context; |
2382 | |
2383 | if (engine->class != RENDER_CLASS0) |
2384 | continue; |
2385 | |
2386 | regs[0].value = intel_sseu_make_rpcs(engine->gt, &ce->sseu); |
2387 | |
2388 | err = gen8_modify_self(ce, regs, num_regs, active); |
2389 | if (err) |
2390 | return err; |
2391 | } |
2392 | |
2393 | return 0; |
2394 | } |
2395 | |
2396 | static int |
2397 | gen12_configure_all_contexts(struct i915_perf_stream *stream, |
2398 | const struct i915_oa_config *oa_config, |
2399 | struct i915_active *active) |
2400 | { |
2401 | struct flex regs[] = { |
2402 | { |
2403 | GEN8_R_PWR_CLK_STATE((const i915_reg_t){ .reg = (0x20C8) }), |
2404 | CTX_R_PWR_CLK_STATE(0x42 + 1), |
2405 | }, |
2406 | }; |
2407 | |
2408 | return oa_configure_all_contexts(stream, |
2409 | regs, ARRAY_SIZE(regs)(sizeof((regs)) / sizeof((regs)[0])), |
2410 | active); |
2411 | } |
2412 | |
2413 | static int |
2414 | lrc_configure_all_contexts(struct i915_perf_stream *stream, |
2415 | const struct i915_oa_config *oa_config, |
2416 | struct i915_active *active) |
2417 | { |
2418 | /* The MMIO offsets for Flex EU registers aren't contiguous */ |
2419 | const u32 ctx_flexeu0 = stream->perf->ctx_flexeu0_offset; |
2420 | #define ctx_flexeuN(N) (ctx_flexeu0 + 2 * (N) + 1) |
2421 | struct flex regs[] = { |
2422 | { |
2423 | GEN8_R_PWR_CLK_STATE((const i915_reg_t){ .reg = (0x20C8) }), |
2424 | CTX_R_PWR_CLK_STATE(0x42 + 1), |
2425 | }, |
2426 | { |
2427 | GEN8_OACTXCONTROL((const i915_reg_t){ .reg = (0x2360) }), |
2428 | stream->perf->ctx_oactxctrl_offset + 1, |
2429 | }, |
2430 | { EU_PERF_CNTL0((const i915_reg_t){ .reg = (0xe458) }), ctx_flexeuN(0) }, |
2431 | { EU_PERF_CNTL1((const i915_reg_t){ .reg = (0xe558) }), ctx_flexeuN(1) }, |
2432 | { EU_PERF_CNTL2((const i915_reg_t){ .reg = (0xe658) }), ctx_flexeuN(2) }, |
2433 | { EU_PERF_CNTL3((const i915_reg_t){ .reg = (0xe758) }), ctx_flexeuN(3) }, |
2434 | { EU_PERF_CNTL4((const i915_reg_t){ .reg = (0xe45c) }), ctx_flexeuN(4) }, |
2435 | { EU_PERF_CNTL5((const i915_reg_t){ .reg = (0xe55c) }), ctx_flexeuN(5) }, |
2436 | { EU_PERF_CNTL6((const i915_reg_t){ .reg = (0xe65c) }), ctx_flexeuN(6) }, |
2437 | }; |
2438 | #undef ctx_flexeuN |
2439 | int i; |
2440 | |
2441 | regs[1].value = |
2442 | (stream->period_exponent << GEN8_OA_TIMER_PERIOD_SHIFT2) | |
2443 | (stream->periodic ? GEN8_OA_TIMER_ENABLE(1 << 1) : 0) | |
2444 | GEN8_OA_COUNTER_RESUME(1 << 0); |
2445 | |
2446 | for (i = 2; i < ARRAY_SIZE(regs)(sizeof((regs)) / sizeof((regs)[0])); i++) |
2447 | regs[i].value = oa_config_flex_reg(oa_config, regs[i].reg); |
2448 | |
2449 | return oa_configure_all_contexts(stream, |
2450 | regs, ARRAY_SIZE(regs)(sizeof((regs)) / sizeof((regs)[0])), |
2451 | active); |
2452 | } |
2453 | |
2454 | static int |
2455 | gen8_enable_metric_set(struct i915_perf_stream *stream, |
2456 | struct i915_active *active) |
2457 | { |
2458 | struct intel_uncore *uncore = stream->uncore; |
2459 | struct i915_oa_config *oa_config = stream->oa_config; |
2460 | int ret; |
2461 | |
2462 | /* |
2463 | * We disable slice/unslice clock ratio change reports on SKL since |
2464 | * they are too noisy. The HW generates a lot of redundant reports |
2465 | * where the ratio hasn't really changed causing a lot of redundant |
2466 | * work to processes and increasing the chances we'll hit buffer |
2467 | * overruns. |
2468 | * |
2469 | * Although we don't currently use the 'disable overrun' OABUFFER |
2470 | * feature it's worth noting that clock ratio reports have to be |
2471 | * disabled before considering to use that feature since the HW doesn't |
2472 | * correctly block these reports. |
2473 | * |
2474 | * Currently none of the high-level metrics we have depend on knowing |
2475 | * this ratio to normalize. |
2476 | * |
2477 | * Note: This register is not power context saved and restored, but |
2478 | * that's OK considering that we disable RC6 while the OA unit is |
2479 | * enabled. |
2480 | * |
2481 | * The _INCLUDE_CLK_RATIO bit allows the slice/unslice frequency to |
2482 | * be read back from automatically triggered reports, as part of the |
2483 | * RPT_ID field. |
2484 | */ |
2485 | if (IS_GEN_RANGE(stream->perf->i915, 9, 11)(!!((&(stream->perf->i915)->__info)->gen_mask & ( 0 + 0 + (((~0UL) >> (64 - (((11)) - 1) - 1)) & ((~0UL) << (((9)) - 1))))))) { |
2486 | intel_uncore_write(uncore, GEN8_OA_DEBUG((const i915_reg_t){ .reg = (0x2B04) }), |
2487 | _MASKED_BIT_ENABLE(GEN9_OA_DEBUG_DISABLE_CLK_RATIO_REPORTS |({ typeof((1 << 5) | (1 << 6)) _a = ((1 << 5 ) | (1 << 6)); ({ if (__builtin_constant_p(_a)) do { } while (0); if (__builtin_constant_p(_a)) do { } while (0); if (__builtin_constant_p (_a) && __builtin_constant_p(_a)) do { } while (0); ( (_a) << 16 | (_a)); }); }) |
2488 | GEN9_OA_DEBUG_INCLUDE_CLK_RATIO)({ typeof((1 << 5) | (1 << 6)) _a = ((1 << 5 ) | (1 << 6)); ({ if (__builtin_constant_p(_a)) do { } while (0); if (__builtin_constant_p(_a)) do { } while (0); if (__builtin_constant_p (_a) && __builtin_constant_p(_a)) do { } while (0); ( (_a) << 16 | (_a)); }); })); |
2489 | } |
2490 | |
2491 | /* |
2492 | * Update all contexts prior writing the mux configurations as we need |
2493 | * to make sure all slices/subslices are ON before writing to NOA |
2494 | * registers. |
2495 | */ |
2496 | ret = lrc_configure_all_contexts(stream, oa_config, active); |
2497 | if (ret) |
2498 | return ret; |
2499 | |
2500 | return emit_oa_config(stream, |
2501 | stream->oa_config, oa_context(stream), |
2502 | active); |
2503 | } |
2504 | |
2505 | static u32 oag_report_ctx_switches(const struct i915_perf_stream *stream) |
2506 | { |
2507 | return _MASKED_FIELD(GEN12_OAG_OA_DEBUG_DISABLE_CTX_SWITCH_REPORTS,({ if (__builtin_constant_p((1 << 1))) do { } while (0) ; if (__builtin_constant_p((stream->sample_flags & (1<< 0)) ? 0 : (1 << 1))) do { } while (0); if (__builtin_constant_p ((1 << 1)) && __builtin_constant_p((stream-> sample_flags & (1<<0)) ? 0 : (1 << 1))) do { } while (0); (((1 << 1)) << 16 | ((stream->sample_flags & (1<<0)) ? 0 : (1 << 1))); }) |
2508 | (stream->sample_flags & SAMPLE_OA_REPORT) ?({ if (__builtin_constant_p((1 << 1))) do { } while (0) ; if (__builtin_constant_p((stream->sample_flags & (1<< 0)) ? 0 : (1 << 1))) do { } while (0); if (__builtin_constant_p ((1 << 1)) && __builtin_constant_p((stream-> sample_flags & (1<<0)) ? 0 : (1 << 1))) do { } while (0); (((1 << 1)) << 16 | ((stream->sample_flags & (1<<0)) ? 0 : (1 << 1))); }) |
2509 | 0 : GEN12_OAG_OA_DEBUG_DISABLE_CTX_SWITCH_REPORTS)({ if (__builtin_constant_p((1 << 1))) do { } while (0) ; if (__builtin_constant_p((stream->sample_flags & (1<< 0)) ? 0 : (1 << 1))) do { } while (0); if (__builtin_constant_p ((1 << 1)) && __builtin_constant_p((stream-> sample_flags & (1<<0)) ? 0 : (1 << 1))) do { } while (0); (((1 << 1)) << 16 | ((stream->sample_flags & (1<<0)) ? 0 : (1 << 1))); }); |
2510 | } |
2511 | |
2512 | static int |
2513 | gen12_enable_metric_set(struct i915_perf_stream *stream, |
2514 | struct i915_active *active) |
2515 | { |
2516 | struct intel_uncore *uncore = stream->uncore; |
2517 | struct i915_oa_config *oa_config = stream->oa_config; |
2518 | bool_Bool periodic = stream->periodic; |
2519 | u32 period_exponent = stream->period_exponent; |
2520 | int ret; |
2521 | |
2522 | intel_uncore_write(uncore, GEN12_OAG_OA_DEBUG((const i915_reg_t){ .reg = (0xdaf8) }), |
2523 | /* Disable clk ratio reports, like previous Gens. */ |
2524 | _MASKED_BIT_ENABLE(GEN12_OAG_OA_DEBUG_DISABLE_CLK_RATIO_REPORTS |({ typeof((1 << 5) | (1 << 6)) _a = ((1 << 5 ) | (1 << 6)); ({ if (__builtin_constant_p(_a)) do { } while (0); if (__builtin_constant_p(_a)) do { } while (0); if (__builtin_constant_p (_a) && __builtin_constant_p(_a)) do { } while (0); ( (_a) << 16 | (_a)); }); }) |
2525 | GEN12_OAG_OA_DEBUG_INCLUDE_CLK_RATIO)({ typeof((1 << 5) | (1 << 6)) _a = ((1 << 5 ) | (1 << 6)); ({ if (__builtin_constant_p(_a)) do { } while (0); if (__builtin_constant_p(_a)) do { } while (0); if (__builtin_constant_p (_a) && __builtin_constant_p(_a)) do { } while (0); ( (_a) << 16 | (_a)); }); }) | |
2526 | /* |
2527 | * If the user didn't require OA reports, instruct |
2528 | * the hardware not to emit ctx switch reports. |
2529 | */ |
2530 | oag_report_ctx_switches(stream)); |
2531 | |
2532 | intel_uncore_write(uncore, GEN12_OAG_OAGLBCTXCTRL((const i915_reg_t){ .reg = (0x2b28) }), periodic ? |
2533 | (GEN12_OAG_OAGLBCTXCTRL_COUNTER_RESUME(1 << 0) | |
2534 | GEN12_OAG_OAGLBCTXCTRL_TIMER_ENABLE(1 << 1) | |
2535 | (period_exponent << GEN12_OAG_OAGLBCTXCTRL_TIMER_PERIOD_SHIFT2)) |
2536 | : 0); |
2537 | |
2538 | /* |
2539 | * Update all contexts prior writing the mux configurations as we need |
2540 | * to make sure all slices/subslices are ON before writing to NOA |
2541 | * registers. |
2542 | */ |
2543 | ret = gen12_configure_all_contexts(stream, oa_config, active); |
2544 | if (ret) |
2545 | return ret; |
2546 | |
2547 | /* |
2548 | * For Gen12, performance counters are context |
2549 | * saved/restored. Only enable it for the context that |
2550 | * requested this. |
2551 | */ |
2552 | if (stream->ctx) { |
2553 | ret = gen12_configure_oar_context(stream, active); |
2554 | if (ret) |
2555 | return ret; |
2556 | } |
2557 | |
2558 | return emit_oa_config(stream, |
2559 | stream->oa_config, oa_context(stream), |
2560 | active); |
2561 | } |
2562 | |
2563 | static void gen8_disable_metric_set(struct i915_perf_stream *stream) |
2564 | { |
2565 | struct intel_uncore *uncore = stream->uncore; |
2566 | |
2567 | /* Reset all contexts' slices/subslices configurations. */ |
2568 | lrc_configure_all_contexts(stream, NULL((void *)0), NULL((void *)0)); |
2569 | |
2570 | intel_uncore_rmw(uncore, GDT_CHICKEN_BITS((const i915_reg_t){ .reg = (0x9840) }), GT_NOA_ENABLE0x00000080, 0); |
2571 | } |
2572 | |
2573 | static void gen10_disable_metric_set(struct i915_perf_stream *stream) |
2574 | { |
2575 | struct intel_uncore *uncore = stream->uncore; |
2576 | |
2577 | /* Reset all contexts' slices/subslices configurations. */ |
2578 | lrc_configure_all_contexts(stream, NULL((void *)0), NULL((void *)0)); |
2579 | |
2580 | /* Make sure we disable noa to save power. */ |
2581 | intel_uncore_rmw(uncore, RPM_CONFIG1((const i915_reg_t){ .reg = (0x0D04) }), GEN10_GT_NOA_ENABLE(1 << 9), 0); |
2582 | } |
2583 | |
2584 | static void gen12_disable_metric_set(struct i915_perf_stream *stream) |
2585 | { |
2586 | struct intel_uncore *uncore = stream->uncore; |
2587 | |
2588 | /* Reset all contexts' slices/subslices configurations. */ |
2589 | gen12_configure_all_contexts(stream, NULL((void *)0), NULL((void *)0)); |
2590 | |
2591 | /* disable the context save/restore or OAR counters */ |
2592 | if (stream->ctx) |
2593 | gen12_configure_oar_context(stream, NULL((void *)0)); |
2594 | |
2595 | /* Make sure we disable noa to save power. */ |
2596 | intel_uncore_rmw(uncore, RPM_CONFIG1((const i915_reg_t){ .reg = (0x0D04) }), GEN10_GT_NOA_ENABLE(1 << 9), 0); |
2597 | } |
2598 | |
2599 | static void gen7_oa_enable(struct i915_perf_stream *stream) |
2600 | { |
2601 | struct intel_uncore *uncore = stream->uncore; |
2602 | struct i915_gem_context *ctx = stream->ctx; |
2603 | u32 ctx_id = stream->specific_ctx_id; |
2604 | bool_Bool periodic = stream->periodic; |
2605 | u32 period_exponent = stream->period_exponent; |
2606 | u32 report_format = stream->oa_buffer.format; |
2607 | |
2608 | /* |
2609 | * Reset buf pointers so we don't forward reports from before now. |
2610 | * |
2611 | * Think carefully if considering trying to avoid this, since it |
2612 | * also ensures status flags and the buffer itself are cleared |
2613 | * in error paths, and we have checks for invalid reports based |
2614 | * on the assumption that certain fields are written to zeroed |
2615 | * memory which this helps maintains. |
2616 | */ |
2617 | gen7_init_oa_buffer(stream); |
2618 | |
2619 | intel_uncore_write(uncore, GEN7_OACONTROL((const i915_reg_t){ .reg = (0x2360) }), |
2620 | (ctx_id & GEN7_OACONTROL_CTX_MASK0xFFFFF000) | |
2621 | (period_exponent << |
2622 | GEN7_OACONTROL_TIMER_PERIOD_SHIFT6) | |
2623 | (periodic ? GEN7_OACONTROL_TIMER_ENABLE(1 << 5) : 0) | |
2624 | (report_format << GEN7_OACONTROL_FORMAT_SHIFT2) | |
2625 | (ctx ? GEN7_OACONTROL_PER_CTX_ENABLE(1 << 1) : 0) | |
2626 | GEN7_OACONTROL_ENABLE(1 << 0)); |
2627 | } |
2628 | |
2629 | static void gen8_oa_enable(struct i915_perf_stream *stream) |
2630 | { |
2631 | struct intel_uncore *uncore = stream->uncore; |
2632 | u32 report_format = stream->oa_buffer.format; |
2633 | |
2634 | /* |
2635 | * Reset buf pointers so we don't forward reports from before now. |
2636 | * |
2637 | * Think carefully if considering trying to avoid this, since it |
2638 | * also ensures status flags and the buffer itself are cleared |
2639 | * in error paths, and we have checks for invalid reports based |
2640 | * on the assumption that certain fields are written to zeroed |
2641 | * memory which this helps maintains. |
2642 | */ |
2643 | gen8_init_oa_buffer(stream); |
2644 | |
2645 | /* |
2646 | * Note: we don't rely on the hardware to perform single context |
2647 | * filtering and instead filter on the cpu based on the context-id |
2648 | * field of reports |
2649 | */ |
2650 | intel_uncore_write(uncore, GEN8_OACONTROL((const i915_reg_t){ .reg = (0x2B00) }), |
2651 | (report_format << GEN8_OA_REPORT_FORMAT_SHIFT2) | |
2652 | GEN8_OA_COUNTER_ENABLE(1 << 0)); |
2653 | } |
2654 | |
2655 | static void gen12_oa_enable(struct i915_perf_stream *stream) |
2656 | { |
2657 | struct intel_uncore *uncore = stream->uncore; |
2658 | u32 report_format = stream->oa_buffer.format; |
2659 | |
2660 | /* |
2661 | * If we don't want OA reports from the OA buffer, then we don't even |
2662 | * need to program the OAG unit. |
2663 | */ |
2664 | if (!(stream->sample_flags & SAMPLE_OA_REPORT(1<<0))) |
2665 | return; |
2666 | |
2667 | gen12_init_oa_buffer(stream); |
2668 | |
2669 | intel_uncore_write(uncore, GEN12_OAG_OACONTROL((const i915_reg_t){ .reg = (0xdaf4) }), |
2670 | (report_format << GEN12_OAG_OACONTROL_OA_COUNTER_FORMAT_SHIFT2) | |
2671 | GEN12_OAG_OACONTROL_OA_COUNTER_ENABLE(1 << 0)); |
2672 | } |
2673 | |
2674 | /** |
2675 | * i915_oa_stream_enable - handle `I915_PERF_IOCTL_ENABLE` for OA stream |
2676 | * @stream: An i915 perf stream opened for OA metrics |
2677 | * |
2678 | * [Re]enables hardware periodic sampling according to the period configured |
2679 | * when opening the stream. This also starts a hrtimer that will periodically |
2680 | * check for data in the circular OA buffer for notifying userspace (e.g. |
2681 | * during a read() or poll()). |
2682 | */ |
2683 | static void i915_oa_stream_enable(struct i915_perf_stream *stream) |
2684 | { |
2685 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
2686 | #ifdef notyet |
2687 | stream->pollin = false0; |
2688 | |
2689 | stream->perf->ops.oa_enable(stream); |
2690 | |
2691 | if (stream->sample_flags & SAMPLE_OA_REPORT(1<<0)) |
2692 | hrtimer_start(&stream->poll_check_timer, |
2693 | ns_to_ktime(stream->poll_oa_period), |
2694 | HRTIMER_MODE_REL_PINNED); |
2695 | #endif |
2696 | } |
2697 | |
2698 | static void gen7_oa_disable(struct i915_perf_stream *stream) |
2699 | { |
2700 | struct intel_uncore *uncore = stream->uncore; |
2701 | |
2702 | intel_uncore_write(uncore, GEN7_OACONTROL((const i915_reg_t){ .reg = (0x2360) }), 0); |
2703 | if (intel_wait_for_register(uncore, |
2704 | GEN7_OACONTROL((const i915_reg_t){ .reg = (0x2360) }), GEN7_OACONTROL_ENABLE(1 << 0), 0, |
2705 | 50)) |
2706 | drm_err(&stream->perf->i915->drm,printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "wait for OA to be disabled timed out\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__) |
2707 | "wait for OA to be disabled timed out\n")printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "wait for OA to be disabled timed out\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__); |
2708 | } |
2709 | |
2710 | static void gen8_oa_disable(struct i915_perf_stream *stream) |
2711 | { |
2712 | struct intel_uncore *uncore = stream->uncore; |
2713 | |
2714 | intel_uncore_write(uncore, GEN8_OACONTROL((const i915_reg_t){ .reg = (0x2B00) }), 0); |
2715 | if (intel_wait_for_register(uncore, |
2716 | GEN8_OACONTROL((const i915_reg_t){ .reg = (0x2B00) }), GEN8_OA_COUNTER_ENABLE(1 << 0), 0, |
2717 | 50)) |
2718 | drm_err(&stream->perf->i915->drm,printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "wait for OA to be disabled timed out\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__) |
2719 | "wait for OA to be disabled timed out\n")printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "wait for OA to be disabled timed out\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__); |
2720 | } |
2721 | |
2722 | static void gen12_oa_disable(struct i915_perf_stream *stream) |
2723 | { |
2724 | struct intel_uncore *uncore = stream->uncore; |
2725 | |
2726 | intel_uncore_write(uncore, GEN12_OAG_OACONTROL((const i915_reg_t){ .reg = (0xdaf4) }), 0); |
2727 | if (intel_wait_for_register(uncore, |
2728 | GEN12_OAG_OACONTROL((const i915_reg_t){ .reg = (0xdaf4) }), |
2729 | GEN12_OAG_OACONTROL_OA_COUNTER_ENABLE(1 << 0), 0, |
2730 | 50)) |
2731 | drm_err(&stream->perf->i915->drm,printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "wait for OA to be disabled timed out\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__) |
2732 | "wait for OA to be disabled timed out\n")printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "wait for OA to be disabled timed out\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__); |
2733 | |
2734 | intel_uncore_write(uncore, GEN12_OA_TLB_INV_CR((const i915_reg_t){ .reg = (0xceec) }), 1); |
2735 | if (intel_wait_for_register(uncore, |
2736 | GEN12_OA_TLB_INV_CR((const i915_reg_t){ .reg = (0xceec) }), |
2737 | 1, 0, |
2738 | 50)) |
2739 | drm_err(&stream->perf->i915->drm,printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "wait for OA tlb invalidate timed out\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__) |
2740 | "wait for OA tlb invalidate timed out\n")printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "wait for OA tlb invalidate timed out\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__); |
2741 | } |
2742 | |
2743 | /** |
2744 | * i915_oa_stream_disable - handle `I915_PERF_IOCTL_DISABLE` for OA stream |
2745 | * @stream: An i915 perf stream opened for OA metrics |
2746 | * |
2747 | * Stops the OA unit from periodically writing counter reports into the |
2748 | * circular OA buffer. This also stops the hrtimer that periodically checks for |
2749 | * data in the circular OA buffer, for notifying userspace. |
2750 | */ |
2751 | static void i915_oa_stream_disable(struct i915_perf_stream *stream) |
2752 | { |
2753 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
2754 | #ifdef notyet |
2755 | stream->perf->ops.oa_disable(stream); |
2756 | |
2757 | if (stream->sample_flags & SAMPLE_OA_REPORT(1<<0)) |
2758 | hrtimer_cancel(&stream->poll_check_timer)timeout_del(&stream->poll_check_timer); |
2759 | #endif |
2760 | } |
2761 | |
2762 | #ifdef notyet |
2763 | static const struct i915_perf_stream_ops i915_oa_stream_ops = { |
2764 | .destroy = i915_oa_stream_destroy, |
2765 | .enable = i915_oa_stream_enable, |
2766 | .disable = i915_oa_stream_disable, |
2767 | .wait_unlocked = i915_oa_wait_unlocked, |
2768 | .poll_wait = i915_oa_poll_wait, |
2769 | .read = i915_oa_read, |
2770 | }; |
2771 | #endif |
2772 | |
2773 | static int i915_perf_stream_enable_sync(struct i915_perf_stream *stream) |
2774 | { |
2775 | struct i915_active *active; |
2776 | int err; |
2777 | |
2778 | active = i915_active_create(); |
2779 | if (!active) |
2780 | return -ENOMEM12; |
2781 | |
2782 | err = stream->perf->ops.enable_metric_set(stream, active); |
2783 | if (err == 0) |
2784 | __i915_active_wait(active, TASK_UNINTERRUPTIBLE0); |
2785 | |
2786 | i915_active_put(active); |
2787 | return err; |
2788 | } |
2789 | |
2790 | static void |
2791 | get_default_sseu_config(struct intel_sseu *out_sseu, |
2792 | struct intel_engine_cs *engine) |
2793 | { |
2794 | const struct sseu_dev_info *devinfo_sseu = &engine->gt->info.sseu; |
2795 | |
2796 | *out_sseu = intel_sseu_from_device_info(devinfo_sseu); |
2797 | |
2798 | if (IS_GEN(engine->i915, 11)(0 + (&(engine->i915)->__info)->gen == (11))) { |
2799 | /* |
2800 | * We only need subslice count so it doesn't matter which ones |
2801 | * we select - just turn off low bits in the amount of half of |
2802 | * all available subslices per slice. |
2803 | */ |
2804 | out_sseu->subslice_mask = |
2805 | ~(~0 << (hweight8(out_sseu->subslice_mask) / 2)); |
2806 | out_sseu->slice_mask = 0x1; |
2807 | } |
2808 | } |
2809 | |
2810 | static int |
2811 | get_sseu_config(struct intel_sseu *out_sseu, |
2812 | struct intel_engine_cs *engine, |
2813 | const struct drm_i915_gem_context_param_sseu *drm_sseu) |
2814 | { |
2815 | if (drm_sseu->engine.engine_class != engine->uabi_class || |
2816 | drm_sseu->engine.engine_instance != engine->uabi_instance) |
2817 | return -EINVAL22; |
2818 | |
2819 | return i915_gem_user_to_context_sseu(engine->gt, drm_sseu, out_sseu); |
2820 | } |
2821 | |
2822 | /** |
2823 | * i915_oa_stream_init - validate combined props for OA stream and init |
2824 | * @stream: An i915 perf stream |
2825 | * @param: The open parameters passed to `DRM_I915_PERF_OPEN` |
2826 | * @props: The property state that configures stream (individually validated) |
2827 | * |
2828 | * While read_properties_unlocked() validates properties in isolation it |
2829 | * doesn't ensure that the combination necessarily makes sense. |
2830 | * |
2831 | * At this point it has been determined that userspace wants a stream of |
2832 | * OA metrics, but still we need to further validate the combined |
2833 | * properties are OK. |
2834 | * |
2835 | * If the configuration makes sense then we can allocate memory for |
2836 | * a circular OA buffer and apply the requested metric set configuration. |
2837 | * |
2838 | * Returns: zero on success or a negative error code. |
2839 | */ |
2840 | static int i915_oa_stream_init(struct i915_perf_stream *stream, |
2841 | struct drm_i915_perf_open_param *param, |
2842 | struct perf_open_properties *props) |
2843 | { |
2844 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
2845 | return -ENOSYS78; |
2846 | #ifdef notyet |
2847 | struct drm_i915_privateinteldrm_softc *i915 = stream->perf->i915; |
2848 | struct i915_perf *perf = stream->perf; |
2849 | int format_size; |
2850 | int ret; |
2851 | |
2852 | if (!props->engine) { |
2853 | DRM_DEBUG("OA engine not specified\n")__drm_dbg(DRM_UT_CORE, "OA engine not specified\n"); |
2854 | return -EINVAL22; |
2855 | } |
2856 | |
2857 | /* |
2858 | * If the sysfs metrics/ directory wasn't registered for some |
2859 | * reason then don't let userspace try their luck with config |
2860 | * IDs |
2861 | */ |
2862 | if (!perf->metrics_kobj) { |
2863 | DRM_DEBUG("OA metrics weren't advertised via sysfs\n")__drm_dbg(DRM_UT_CORE, "OA metrics weren't advertised via sysfs\n" ); |
2864 | return -EINVAL22; |
2865 | } |
2866 | |
2867 | if (!(props->sample_flags & SAMPLE_OA_REPORT(1<<0)) && |
2868 | (INTEL_GEN(perf->i915)((&(perf->i915)->__info)->gen) < 12 || !stream->ctx)) { |
2869 | DRM_DEBUG("Only OA report sampling supported\n")__drm_dbg(DRM_UT_CORE, "Only OA report sampling supported\n"); |
2870 | return -EINVAL22; |
2871 | } |
2872 | |
2873 | if (!perf->ops.enable_metric_set) { |
2874 | DRM_DEBUG("OA unit not supported\n")__drm_dbg(DRM_UT_CORE, "OA unit not supported\n"); |
2875 | return -ENODEV19; |
2876 | } |
2877 | |
2878 | /* |
2879 | * To avoid the complexity of having to accurately filter |
2880 | * counter reports and marshal to the appropriate client |
2881 | * we currently only allow exclusive access |
2882 | */ |
2883 | if (perf->exclusive_stream) { |
2884 | DRM_DEBUG("OA unit already in use\n")__drm_dbg(DRM_UT_CORE, "OA unit already in use\n"); |
2885 | return -EBUSY16; |
2886 | } |
2887 | |
2888 | if (!props->oa_format) { |
2889 | DRM_DEBUG("OA report format not specified\n")__drm_dbg(DRM_UT_CORE, "OA report format not specified\n"); |
2890 | return -EINVAL22; |
2891 | } |
2892 | |
2893 | stream->engine = props->engine; |
2894 | stream->uncore = stream->engine->gt->uncore; |
2895 | |
2896 | stream->sample_size = sizeof(struct drm_i915_perf_record_header); |
2897 | |
2898 | format_size = perf->oa_formats[props->oa_format].size; |
2899 | |
2900 | stream->sample_flags = props->sample_flags; |
2901 | stream->sample_size += format_size; |
2902 | |
2903 | stream->oa_buffer.format_size = format_size; |
2904 | if (drm_WARN_ON(&i915->drm, stream->oa_buffer.format_size == 0)({ int __ret = !!((stream->oa_buffer.format_size == 0)); if (__ret) printf("%s %s: " "%s", dev_driver_string(((&i915 ->drm))->dev), "", "drm_WARN_ON(" "stream->oa_buffer.format_size == 0" ")"); __builtin_expect(!!(__ret), 0); })) |
2905 | return -EINVAL22; |
2906 | |
2907 | stream->hold_preemption = props->hold_preemption; |
2908 | |
2909 | stream->oa_buffer.format = |
2910 | perf->oa_formats[props->oa_format].format; |
2911 | |
2912 | stream->periodic = props->oa_periodic; |
2913 | if (stream->periodic) |
2914 | stream->period_exponent = props->oa_period_exponent; |
2915 | |
2916 | if (stream->ctx) { |
2917 | ret = oa_get_render_ctx_id(stream); |
2918 | if (ret) { |
2919 | DRM_DEBUG("Invalid context id to filter with\n")__drm_dbg(DRM_UT_CORE, "Invalid context id to filter with\n"); |
2920 | return ret; |
2921 | } |
2922 | } |
2923 | |
2924 | ret = alloc_noa_wait(stream); |
2925 | if (ret) { |
2926 | DRM_DEBUG("Unable to allocate NOA wait batch buffer\n")__drm_dbg(DRM_UT_CORE, "Unable to allocate NOA wait batch buffer\n" ); |
2927 | goto err_noa_wait_alloc; |
2928 | } |
2929 | |
2930 | stream->oa_config = i915_perf_get_oa_config(perf, props->metrics_set); |
2931 | if (!stream->oa_config) { |
2932 | DRM_DEBUG("Invalid OA config id=%i\n", props->metrics_set)__drm_dbg(DRM_UT_CORE, "Invalid OA config id=%i\n", props-> metrics_set); |
2933 | ret = -EINVAL22; |
2934 | goto err_config; |
2935 | } |
2936 | |
2937 | /* PRM - observability performance counters: |
2938 | * |
2939 | * OACONTROL, performance counter enable, note: |
2940 | * |
2941 | * "When this bit is set, in order to have coherent counts, |
2942 | * RC6 power state and trunk clock gating must be disabled. |
2943 | * This can be achieved by programming MMIO registers as |
2944 | * 0xA094=0 and 0xA090[31]=1" |
2945 | * |
2946 | * In our case we are expecting that taking pm + FORCEWAKE |
2947 | * references will effectively disable RC6. |
2948 | */ |
2949 | intel_engine_pm_get(stream->engine); |
2950 | intel_uncore_forcewake_get(stream->uncore, FORCEWAKE_ALL); |
2951 | |
2952 | ret = alloc_oa_buffer(stream); |
2953 | if (ret) |
2954 | goto err_oa_buf_alloc; |
2955 | |
2956 | stream->ops = &i915_oa_stream_ops; |
2957 | |
2958 | perf->sseu = props->sseu; |
2959 | WRITE_ONCE(perf->exclusive_stream, stream)({ typeof(perf->exclusive_stream) __tmp = (stream); *(volatile typeof(perf->exclusive_stream) *)&(perf->exclusive_stream ) = __tmp; __tmp; }); |
2960 | |
2961 | ret = i915_perf_stream_enable_sync(stream); |
2962 | if (ret) { |
2963 | DRM_DEBUG("Unable to enable metric set\n")__drm_dbg(DRM_UT_CORE, "Unable to enable metric set\n"); |
2964 | goto err_enable; |
2965 | } |
2966 | |
2967 | DRM_DEBUG("opening stream oa config uuid=%s\n",__drm_dbg(DRM_UT_CORE, "opening stream oa config uuid=%s\n", stream ->oa_config->uuid) |
2968 | stream->oa_config->uuid)__drm_dbg(DRM_UT_CORE, "opening stream oa config uuid=%s\n", stream ->oa_config->uuid); |
2969 | |
2970 | hrtimer_init(&stream->poll_check_timer, |
2971 | CLOCK_MONOTONIC3, HRTIMER_MODE_REL1); |
2972 | stream->poll_check_timer.function = oa_poll_check_timer_cb; |
2973 | init_waitqueue_head(&stream->poll_wq); |
2974 | mtx_init(&stream->oa_buffer.ptr_lock, IPL_TTY)do { (void)(((void *)0)); (void)(0); __mtx_init((&stream-> oa_buffer.ptr_lock), ((((0x9)) > 0x0 && ((0x9)) < 0x9) ? 0x9 : ((0x9)))); } while (0); |
2975 | |
2976 | return 0; |
2977 | |
2978 | err_enable: |
2979 | WRITE_ONCE(perf->exclusive_stream, NULL)({ typeof(perf->exclusive_stream) __tmp = (((void *)0)); * (volatile typeof(perf->exclusive_stream) *)&(perf-> exclusive_stream) = __tmp; __tmp; }); |
2980 | perf->ops.disable_metric_set(stream); |
2981 | |
2982 | free_oa_buffer(stream); |
2983 | |
2984 | err_oa_buf_alloc: |
2985 | free_oa_configs(stream); |
2986 | |
2987 | intel_uncore_forcewake_put(stream->uncore, FORCEWAKE_ALL); |
2988 | intel_engine_pm_put(stream->engine); |
2989 | |
2990 | err_config: |
2991 | free_noa_wait(stream); |
2992 | |
2993 | err_noa_wait_alloc: |
2994 | if (stream->ctx) |
2995 | oa_put_render_ctx_id(stream); |
2996 | |
2997 | return ret; |
2998 | #endif |
2999 | } |
3000 | |
3001 | void i915_oa_init_reg_state(const struct intel_context *ce, |
3002 | const struct intel_engine_cs *engine) |
3003 | { |
3004 | struct i915_perf_stream *stream; |
3005 | |
3006 | if (engine->class != RENDER_CLASS0) |
3007 | return; |
3008 | |
3009 | /* perf.exclusive_stream serialised by lrc_configure_all_contexts() */ |
3010 | stream = READ_ONCE(engine->i915->perf.exclusive_stream)({ typeof(engine->i915->perf.exclusive_stream) __tmp = * (volatile typeof(engine->i915->perf.exclusive_stream) * )&(engine->i915->perf.exclusive_stream); membar_datadep_consumer (); __tmp; }); |
3011 | if (stream && INTEL_GEN(stream->perf->i915)((&(stream->perf->i915)->__info)->gen) < 12) |
3012 | gen8_update_reg_state_unlocked(ce, stream); |
3013 | } |
3014 | |
3015 | /** |
3016 | * i915_perf_read - handles read() FOP for i915 perf stream FDs |
3017 | * @file: An i915 perf stream file |
3018 | * @buf: destination buffer given by userspace |
3019 | * @count: the number of bytes userspace wants to read |
3020 | * @ppos: (inout) file seek position (unused) |
3021 | * |
3022 | * The entry point for handling a read() on a stream file descriptor from |
3023 | * userspace. Most of the work is left to the i915_perf_read_locked() and |
3024 | * &i915_perf_stream_ops->read but to save having stream implementations (of |
3025 | * which we might have multiple later) we handle blocking read here. |
3026 | * |
3027 | * We can also consistently treat trying to read from a disabled stream |
3028 | * as an IO error so implementations can assume the stream is enabled |
3029 | * while reading. |
3030 | * |
3031 | * Returns: The number of bytes copied or a negative error code on failure. |
3032 | */ |
3033 | static ssize_t i915_perf_read(struct file *file, |
3034 | char __user *buf, |
3035 | size_t count, |
3036 | loff_t *ppos) |
3037 | { |
3038 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
3039 | return -ENOSYS78; |
3040 | #ifdef notyet |
3041 | struct i915_perf_stream *stream = file->private_data; |
3042 | struct i915_perf *perf = stream->perf; |
3043 | size_t offset = 0; |
3044 | int ret; |
3045 | |
3046 | /* To ensure it's handled consistently we simply treat all reads of a |
3047 | * disabled stream as an error. In particular it might otherwise lead |
3048 | * to a deadlock for blocking file descriptors... |
3049 | */ |
3050 | if (!stream->enabled || !(stream->sample_flags & SAMPLE_OA_REPORT(1<<0))) |
3051 | return -EIO5; |
3052 | |
3053 | if (!(file->f_flags & O_NONBLOCK0x0004)) { |
3054 | /* There's the small chance of false positives from |
3055 | * stream->ops->wait_unlocked. |
3056 | * |
3057 | * E.g. with single context filtering since we only wait until |
3058 | * oabuffer has >= 1 report we don't immediately know whether |
3059 | * any reports really belong to the current context |
3060 | */ |
3061 | do { |
3062 | ret = stream->ops->wait_unlocked(stream); |
3063 | if (ret) |
3064 | return ret; |
3065 | |
3066 | mutex_lock(&perf->lock)rw_enter_write(&perf->lock); |
3067 | ret = stream->ops->read(stream, buf, count, &offset); |
3068 | mutex_unlock(&perf->lock)rw_exit_write(&perf->lock); |
3069 | } while (!offset && !ret); |
3070 | } else { |
3071 | mutex_lock(&perf->lock)rw_enter_write(&perf->lock); |
3072 | ret = stream->ops->read(stream, buf, count, &offset); |
3073 | mutex_unlock(&perf->lock)rw_exit_write(&perf->lock); |
3074 | } |
3075 | |
3076 | /* We allow the poll checking to sometimes report false positive EPOLLIN |
3077 | * events where we might actually report EAGAIN on read() if there's |
3078 | * not really any data available. In this situation though we don't |
3079 | * want to enter a busy loop between poll() reporting a EPOLLIN event |
3080 | * and read() returning -EAGAIN. Clearing the oa.pollin state here |
3081 | * effectively ensures we back off until the next hrtimer callback |
3082 | * before reporting another EPOLLIN event. |
3083 | * The exception to this is if ops->read() returned -ENOSPC which means |
3084 | * that more OA data is available than could fit in the user provided |
3085 | * buffer. In this case we want the next poll() call to not block. |
3086 | */ |
3087 | if (ret != -ENOSPC28) |
3088 | stream->pollin = false0; |
3089 | |
3090 | /* Possible values for ret are 0, -EFAULT, -ENOSPC, -EIO, ... */ |
3091 | return offset ?: (ret ?: -EAGAIN35); |
3092 | #endif |
3093 | } |
3094 | |
3095 | static enum hrtimer_restart oa_poll_check_timer_cb(struct hrtimer *hrtimer) |
3096 | { |
3097 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
3098 | return 0; |
3099 | #ifdef notyet |
3100 | struct i915_perf_stream *stream = |
3101 | container_of(hrtimer, typeof(*stream), poll_check_timer)({ const __typeof( ((typeof(*stream) *)0)->poll_check_timer ) *__mptr = (hrtimer); (typeof(*stream) *)( (char *)__mptr - __builtin_offsetof(typeof(*stream), poll_check_timer) );}); |
3102 | |
3103 | if (oa_buffer_check_unlocked(stream)) { |
3104 | stream->pollin = true1; |
3105 | wake_up(&stream->poll_wq); |
3106 | } |
3107 | |
3108 | hrtimer_forward_now(hrtimer, |
3109 | ns_to_ktime(stream->poll_oa_period)); |
3110 | |
3111 | return HRTIMER_RESTART; |
3112 | #endif |
3113 | } |
3114 | |
3115 | #ifdef notyet |
3116 | |
3117 | /** |
3118 | * i915_perf_poll_locked - poll_wait() with a suitable wait queue for stream |
3119 | * @stream: An i915 perf stream |
3120 | * @file: An i915 perf stream file |
3121 | * @wait: poll() state table |
3122 | * |
3123 | * For handling userspace polling on an i915 perf stream, this calls through to |
3124 | * &i915_perf_stream_ops->poll_wait to call poll_wait() with a wait queue that |
3125 | * will be woken for new stream data. |
3126 | * |
3127 | * Note: The &perf->lock mutex has been taken to serialize |
3128 | * with any non-file-operation driver hooks. |
3129 | * |
3130 | * Returns: any poll events that are ready without sleeping |
3131 | */ |
3132 | static __poll_t i915_perf_poll_locked(struct i915_perf_stream *stream, |
3133 | struct file *file, |
3134 | poll_table *wait) |
3135 | { |
3136 | __poll_t events = 0; |
3137 | |
3138 | stream->ops->poll_wait(stream, file, wait); |
3139 | |
3140 | /* Note: we don't explicitly check whether there's something to read |
3141 | * here since this path may be very hot depending on what else |
3142 | * userspace is polling, or on the timeout in use. We rely solely on |
3143 | * the hrtimer/oa_poll_check_timer_cb to notify us when there are |
3144 | * samples to read. |
3145 | */ |
3146 | if (stream->pollin) |
3147 | events |= EPOLLIN; |
3148 | |
3149 | return events; |
3150 | } |
3151 | |
3152 | /** |
3153 | * i915_perf_poll - call poll_wait() with a suitable wait queue for stream |
3154 | * @file: An i915 perf stream file |
3155 | * @wait: poll() state table |
3156 | * |
3157 | * For handling userspace polling on an i915 perf stream, this ensures |
3158 | * poll_wait() gets called with a wait queue that will be woken for new stream |
3159 | * data. |
3160 | * |
3161 | * Note: Implementation deferred to i915_perf_poll_locked() |
3162 | * |
3163 | * Returns: any poll events that are ready without sleeping |
3164 | */ |
3165 | static __poll_t i915_perf_poll(struct file *file, poll_table *wait) |
3166 | { |
3167 | struct i915_perf_stream *stream = file->private_data; |
3168 | struct i915_perf *perf = stream->perf; |
3169 | __poll_t ret; |
3170 | |
3171 | mutex_lock(&perf->lock)rw_enter_write(&perf->lock); |
3172 | ret = i915_perf_poll_locked(stream, file, wait); |
3173 | mutex_unlock(&perf->lock)rw_exit_write(&perf->lock); |
3174 | |
3175 | return ret; |
3176 | } |
3177 | |
3178 | #endif /* notyet */ |
3179 | |
3180 | /** |
3181 | * i915_perf_enable_locked - handle `I915_PERF_IOCTL_ENABLE` ioctl |
3182 | * @stream: A disabled i915 perf stream |
3183 | * |
3184 | * [Re]enables the associated capture of data for this stream. |
3185 | * |
3186 | * If a stream was previously enabled then there's currently no intention |
3187 | * to provide userspace any guarantee about the preservation of previously |
3188 | * buffered data. |
3189 | */ |
3190 | static void i915_perf_enable_locked(struct i915_perf_stream *stream) |
3191 | { |
3192 | if (stream->enabled) |
3193 | return; |
3194 | |
3195 | /* Allow stream->ops->enable() to refer to this */ |
3196 | stream->enabled = true1; |
3197 | |
3198 | if (stream->ops->enable) |
3199 | stream->ops->enable(stream); |
3200 | |
3201 | if (stream->hold_preemption) |
3202 | intel_context_set_nopreempt(stream->pinned_ctx); |
3203 | } |
3204 | |
3205 | /** |
3206 | * i915_perf_disable_locked - handle `I915_PERF_IOCTL_DISABLE` ioctl |
3207 | * @stream: An enabled i915 perf stream |
3208 | * |
3209 | * Disables the associated capture of data for this stream. |
3210 | * |
3211 | * The intention is that disabling an re-enabling a stream will ideally be |
3212 | * cheaper than destroying and re-opening a stream with the same configuration, |
3213 | * though there are no formal guarantees about what state or buffered data |
3214 | * must be retained between disabling and re-enabling a stream. |
3215 | * |
3216 | * Note: while a stream is disabled it's considered an error for userspace |
3217 | * to attempt to read from the stream (-EIO). |
3218 | */ |
3219 | static void i915_perf_disable_locked(struct i915_perf_stream *stream) |
3220 | { |
3221 | if (!stream->enabled) |
3222 | return; |
3223 | |
3224 | /* Allow stream->ops->disable() to refer to this */ |
3225 | stream->enabled = false0; |
3226 | |
3227 | if (stream->hold_preemption) |
3228 | intel_context_clear_nopreempt(stream->pinned_ctx); |
3229 | |
3230 | if (stream->ops->disable) |
3231 | stream->ops->disable(stream); |
3232 | } |
3233 | |
3234 | static long i915_perf_config_locked(struct i915_perf_stream *stream, |
3235 | unsigned long metrics_set) |
3236 | { |
3237 | struct i915_oa_config *config; |
3238 | long ret = stream->oa_config->id; |
3239 | |
3240 | config = i915_perf_get_oa_config(stream->perf, metrics_set); |
3241 | if (!config) |
3242 | return -EINVAL22; |
3243 | |
3244 | if (config != stream->oa_config) { |
3245 | int err; |
3246 | |
3247 | /* |
3248 | * If OA is bound to a specific context, emit the |
3249 | * reconfiguration inline from that context. The update |
3250 | * will then be ordered with respect to submission on that |
3251 | * context. |
3252 | * |
3253 | * When set globally, we use a low priority kernel context, |
3254 | * so it will effectively take effect when idle. |
3255 | */ |
3256 | err = emit_oa_config(stream, config, oa_context(stream), NULL((void *)0)); |
3257 | if (!err) |
3258 | config = xchg(&stream->oa_config, config)__sync_lock_test_and_set(&stream->oa_config, config); |
3259 | else |
3260 | ret = err; |
3261 | } |
3262 | |
3263 | i915_oa_config_put(config); |
3264 | |
3265 | return ret; |
3266 | } |
3267 | |
3268 | /** |
3269 | * i915_perf_ioctl - support ioctl() usage with i915 perf stream FDs |
3270 | * @stream: An i915 perf stream |
3271 | * @cmd: the ioctl request |
3272 | * @arg: the ioctl data |
3273 | * |
3274 | * Note: The &perf->lock mutex has been taken to serialize |
3275 | * with any non-file-operation driver hooks. |
3276 | * |
3277 | * Returns: zero on success or a negative error code. Returns -EINVAL for |
3278 | * an unknown ioctl request. |
3279 | */ |
3280 | static long i915_perf_ioctl_locked(struct i915_perf_stream *stream, |
3281 | unsigned int cmd, |
3282 | unsigned long arg) |
3283 | { |
3284 | switch (cmd) { |
3285 | case I915_PERF_IOCTL_ENABLE((unsigned long)0x20000000 | ((0 & 0x1fff) << 16) | ((('i')) << 8) | ((0x0))): |
3286 | i915_perf_enable_locked(stream); |
3287 | return 0; |
3288 | case I915_PERF_IOCTL_DISABLE((unsigned long)0x20000000 | ((0 & 0x1fff) << 16) | ((('i')) << 8) | ((0x1))): |
3289 | i915_perf_disable_locked(stream); |
3290 | return 0; |
3291 | case I915_PERF_IOCTL_CONFIG((unsigned long)0x20000000 | ((0 & 0x1fff) << 16) | ((('i')) << 8) | ((0x2))): |
3292 | return i915_perf_config_locked(stream, arg); |
3293 | } |
3294 | |
3295 | return -EINVAL22; |
3296 | } |
3297 | |
3298 | /** |
3299 | * i915_perf_ioctl - support ioctl() usage with i915 perf stream FDs |
3300 | * @file: An i915 perf stream file |
3301 | * @cmd: the ioctl request |
3302 | * @arg: the ioctl data |
3303 | * |
3304 | * Implementation deferred to i915_perf_ioctl_locked(). |
3305 | * |
3306 | * Returns: zero on success or a negative error code. Returns -EINVAL for |
3307 | * an unknown ioctl request. |
3308 | */ |
3309 | static long i915_perf_ioctl(struct file *file, |
3310 | unsigned int cmd, |
3311 | unsigned long arg) |
3312 | { |
3313 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
3314 | return -ENOSYS78; |
3315 | #ifdef notyet |
3316 | struct i915_perf_stream *stream = file->private_data; |
3317 | struct i915_perf *perf = stream->perf; |
3318 | long ret; |
3319 | |
3320 | mutex_lock(&perf->lock)rw_enter_write(&perf->lock); |
3321 | ret = i915_perf_ioctl_locked(stream, cmd, arg); |
3322 | mutex_unlock(&perf->lock)rw_exit_write(&perf->lock); |
3323 | |
3324 | return ret; |
3325 | #endif |
3326 | } |
3327 | |
3328 | /** |
3329 | * i915_perf_destroy_locked - destroy an i915 perf stream |
3330 | * @stream: An i915 perf stream |
3331 | * |
3332 | * Frees all resources associated with the given i915 perf @stream, disabling |
3333 | * any associated data capture in the process. |
3334 | * |
3335 | * Note: The &perf->lock mutex has been taken to serialize |
3336 | * with any non-file-operation driver hooks. |
3337 | */ |
3338 | static void i915_perf_destroy_locked(struct i915_perf_stream *stream) |
3339 | { |
3340 | if (stream->enabled) |
3341 | i915_perf_disable_locked(stream); |
3342 | |
3343 | if (stream->ops->destroy) |
3344 | stream->ops->destroy(stream); |
3345 | |
3346 | if (stream->ctx) |
3347 | i915_gem_context_put(stream->ctx); |
3348 | |
3349 | kfree(stream); |
3350 | } |
3351 | |
3352 | #ifdef notyet |
3353 | |
3354 | /** |
3355 | * i915_perf_release - handles userspace close() of a stream file |
3356 | * @inode: anonymous inode associated with file |
3357 | * @file: An i915 perf stream file |
3358 | * |
3359 | * Cleans up any resources associated with an open i915 perf stream file. |
3360 | * |
3361 | * NB: close() can't really fail from the userspace point of view. |
3362 | * |
3363 | * Returns: zero on success or a negative error code. |
3364 | */ |
3365 | static int i915_perf_release(struct inode *inode, struct file *file) |
3366 | { |
3367 | struct i915_perf_stream *stream = file->private_data; |
3368 | struct i915_perf *perf = stream->perf; |
3369 | |
3370 | mutex_lock(&perf->lock)rw_enter_write(&perf->lock); |
3371 | i915_perf_destroy_locked(stream); |
3372 | mutex_unlock(&perf->lock)rw_exit_write(&perf->lock); |
3373 | |
3374 | /* Release the reference the perf stream kept on the driver. */ |
3375 | drm_dev_put(&perf->i915->drm); |
3376 | |
3377 | return 0; |
3378 | } |
3379 | |
3380 | |
3381 | static const struct file_operations fops = { |
3382 | .owner = THIS_MODULE((void *)0), |
3383 | .llseek = no_llseek, |
3384 | .release = i915_perf_release, |
3385 | .poll = i915_perf_poll, |
3386 | .read = i915_perf_read, |
3387 | .unlocked_ioctl = i915_perf_ioctl, |
3388 | /* Our ioctl have no arguments, so it's safe to use the same function |
3389 | * to handle 32bits compatibility. |
3390 | */ |
3391 | .compat_ioctl = i915_perf_ioctl, |
3392 | }; |
3393 | |
3394 | #endif /* notyet */ |
3395 | |
3396 | /** |
3397 | * i915_perf_open_ioctl_locked - DRM ioctl() for userspace to open a stream FD |
3398 | * @perf: i915 perf instance |
3399 | * @param: The open parameters passed to 'DRM_I915_PERF_OPEN` |
3400 | * @props: individually validated u64 property value pairs |
3401 | * @file: drm file |
3402 | * |
3403 | * See i915_perf_ioctl_open() for interface details. |
3404 | * |
3405 | * Implements further stream config validation and stream initialization on |
3406 | * behalf of i915_perf_open_ioctl() with the &perf->lock mutex |
3407 | * taken to serialize with any non-file-operation driver hooks. |
3408 | * |
3409 | * Note: at this point the @props have only been validated in isolation and |
3410 | * it's still necessary to validate that the combination of properties makes |
3411 | * sense. |
3412 | * |
3413 | * In the case where userspace is interested in OA unit metrics then further |
3414 | * config validation and stream initialization details will be handled by |
3415 | * i915_oa_stream_init(). The code here should only validate config state that |
3416 | * will be relevant to all stream types / backends. |
3417 | * |
3418 | * Returns: zero on success or a negative error code. |
3419 | */ |
3420 | static int |
3421 | i915_perf_open_ioctl_locked(struct i915_perf *perf, |
3422 | struct drm_i915_perf_open_param *param, |
3423 | struct perf_open_properties *props, |
3424 | struct drm_file *file) |
3425 | { |
3426 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
3427 | return -ENOSYS78; |
3428 | #ifdef notyet |
3429 | struct i915_gem_context *specific_ctx = NULL((void *)0); |
3430 | struct i915_perf_stream *stream = NULL((void *)0); |
3431 | unsigned long f_flags = 0; |
3432 | bool_Bool privileged_op = true1; |
3433 | int stream_fd; |
3434 | int ret; |
3435 | |
3436 | if (props->single_context) { |
3437 | u32 ctx_handle = props->ctx_handle; |
3438 | struct drm_i915_file_private *file_priv = file->driver_priv; |
3439 | |
3440 | specific_ctx = i915_gem_context_lookup(file_priv, ctx_handle); |
3441 | if (!specific_ctx) { |
3442 | DRM_DEBUG("Failed to look up context with ID %u for opening perf stream\n",__drm_dbg(DRM_UT_CORE, "Failed to look up context with ID %u for opening perf stream\n" , ctx_handle) |
3443 | ctx_handle)__drm_dbg(DRM_UT_CORE, "Failed to look up context with ID %u for opening perf stream\n" , ctx_handle); |
3444 | ret = -ENOENT2; |
3445 | goto err; |
3446 | } |
3447 | } |
3448 | |
3449 | /* |
3450 | * On Haswell the OA unit supports clock gating off for a specific |
3451 | * context and in this mode there's no visibility of metrics for the |
3452 | * rest of the system, which we consider acceptable for a |
3453 | * non-privileged client. |
3454 | * |
3455 | * For Gen8->11 the OA unit no longer supports clock gating off for a |
3456 | * specific context and the kernel can't securely stop the counters |
3457 | * from updating as system-wide / global values. Even though we can |
3458 | * filter reports based on the included context ID we can't block |
3459 | * clients from seeing the raw / global counter values via |
3460 | * MI_REPORT_PERF_COUNT commands and so consider it a privileged op to |
3461 | * enable the OA unit by default. |
3462 | * |
3463 | * For Gen12+ we gain a new OAR unit that only monitors the RCS on a |
3464 | * per context basis. So we can relax requirements there if the user |
3465 | * doesn't request global stream access (i.e. query based sampling |
3466 | * using MI_RECORD_PERF_COUNT. |
3467 | */ |
3468 | if (IS_HASWELL(perf->i915)IS_PLATFORM(perf->i915, INTEL_HASWELL) && specific_ctx) |
3469 | privileged_op = false0; |
3470 | else if (IS_GEN(perf->i915, 12)(0 + (&(perf->i915)->__info)->gen == (12)) && specific_ctx && |
3471 | (props->sample_flags & SAMPLE_OA_REPORT(1<<0)) == 0) |
3472 | privileged_op = false0; |
3473 | |
3474 | if (props->hold_preemption) { |
3475 | if (!props->single_context) { |
3476 | DRM_DEBUG("preemption disable with no context\n")__drm_dbg(DRM_UT_CORE, "preemption disable with no context\n" ); |
3477 | ret = -EINVAL22; |
3478 | goto err; |
3479 | } |
3480 | privileged_op = true1; |
3481 | } |
3482 | |
3483 | /* |
3484 | * Asking for SSEU configuration is a priviliged operation. |
3485 | */ |
3486 | if (props->has_sseu) |
3487 | privileged_op = true1; |
3488 | else |
3489 | get_default_sseu_config(&props->sseu, props->engine); |
3490 | |
3491 | /* Similar to perf's kernel.perf_paranoid_cpu sysctl option |
3492 | * we check a dev.i915.perf_stream_paranoid sysctl option |
3493 | * to determine if it's ok to access system wide OA counters |
3494 | * without CAP_PERFMON or CAP_SYS_ADMIN privileges. |
3495 | */ |
3496 | if (privileged_op && |
3497 | i915_perf_stream_paranoid && !perfmon_capable()) { |
3498 | DRM_DEBUG("Insufficient privileges to open i915 perf stream\n")__drm_dbg(DRM_UT_CORE, "Insufficient privileges to open i915 perf stream\n" ); |
3499 | ret = -EACCES13; |
3500 | goto err_ctx; |
3501 | } |
3502 | |
3503 | stream = kzalloc(sizeof(*stream), GFP_KERNEL(0x0001 | 0x0004)); |
3504 | if (!stream) { |
3505 | ret = -ENOMEM12; |
3506 | goto err_ctx; |
3507 | } |
3508 | |
3509 | stream->perf = perf; |
3510 | stream->ctx = specific_ctx; |
3511 | stream->poll_oa_period = props->poll_oa_period; |
3512 | |
3513 | ret = i915_oa_stream_init(stream, param, props); |
3514 | if (ret) |
3515 | goto err_alloc; |
3516 | |
3517 | /* we avoid simply assigning stream->sample_flags = props->sample_flags |
3518 | * to have _stream_init check the combination of sample flags more |
3519 | * thoroughly, but still this is the expected result at this point. |
3520 | */ |
3521 | if (WARN_ON(stream->sample_flags != props->sample_flags)({ int __ret = !!((stream->sample_flags != props->sample_flags )); if (__ret) printf("%s", "WARN_ON(" "stream->sample_flags != props->sample_flags" ")"); __builtin_expect(!!(__ret), 0); })) { |
3522 | ret = -ENODEV19; |
3523 | goto err_flags; |
3524 | } |
3525 | |
3526 | if (param->flags & I915_PERF_FLAG_FD_CLOEXEC(1<<0)) |
3527 | f_flags |= O_CLOEXEC0x10000; |
3528 | if (param->flags & I915_PERF_FLAG_FD_NONBLOCK(1<<1)) |
3529 | f_flags |= O_NONBLOCK0x0004; |
3530 | |
3531 | stream_fd = anon_inode_getfd("[i915_perf]", &fops, stream, f_flags); |
3532 | if (stream_fd < 0) { |
3533 | ret = stream_fd; |
3534 | goto err_flags; |
3535 | } |
3536 | |
3537 | if (!(param->flags & I915_PERF_FLAG_DISABLED(1<<2))) |
3538 | i915_perf_enable_locked(stream); |
3539 | |
3540 | /* Take a reference on the driver that will be kept with stream_fd |
3541 | * until its release. |
3542 | */ |
3543 | drm_dev_get(&perf->i915->drm); |
3544 | |
3545 | return stream_fd; |
3546 | |
3547 | err_flags: |
3548 | if (stream->ops->destroy) |
3549 | stream->ops->destroy(stream); |
3550 | err_alloc: |
3551 | kfree(stream); |
3552 | err_ctx: |
3553 | if (specific_ctx) |
3554 | i915_gem_context_put(specific_ctx); |
3555 | err: |
3556 | return ret; |
3557 | #endif |
3558 | } |
3559 | |
3560 | static u64 oa_exponent_to_ns(struct i915_perf *perf, int exponent) |
3561 | { |
3562 | return i915_cs_timestamp_ticks_to_ns(perf->i915, 2ULL << exponent); |
3563 | } |
3564 | |
3565 | /** |
3566 | * read_properties_unlocked - validate + copy userspace stream open properties |
3567 | * @perf: i915 perf instance |
3568 | * @uprops: The array of u64 key value pairs given by userspace |
3569 | * @n_props: The number of key value pairs expected in @uprops |
3570 | * @props: The stream configuration built up while validating properties |
3571 | * |
3572 | * Note this function only validates properties in isolation it doesn't |
3573 | * validate that the combination of properties makes sense or that all |
3574 | * properties necessary for a particular kind of stream have been set. |
3575 | * |
3576 | * Note that there currently aren't any ordering requirements for properties so |
3577 | * we shouldn't validate or assume anything about ordering here. This doesn't |
3578 | * rule out defining new properties with ordering requirements in the future. |
3579 | */ |
3580 | static int read_properties_unlocked(struct i915_perf *perf, |
3581 | u64 __user *uprops, |
3582 | u32 n_props, |
3583 | struct perf_open_properties *props) |
3584 | { |
3585 | u64 __user *uprop = uprops; |
3586 | u32 i; |
3587 | int ret; |
3588 | |
3589 | memset(props, 0, sizeof(struct perf_open_properties))__builtin_memset((props), (0), (sizeof(struct perf_open_properties ))); |
3590 | props->poll_oa_period = DEFAULT_POLL_PERIOD_NS(1000000000L / 200); |
3591 | |
3592 | if (!n_props) { |
3593 | DRM_DEBUG("No i915 perf properties given\n")__drm_dbg(DRM_UT_CORE, "No i915 perf properties given\n"); |
3594 | return -EINVAL22; |
3595 | } |
3596 | |
3597 | /* At the moment we only support using i915-perf on the RCS. */ |
3598 | props->engine = intel_engine_lookup_user(perf->i915, |
3599 | I915_ENGINE_CLASS_RENDER, |
3600 | 0); |
3601 | if (!props->engine) { |
3602 | DRM_DEBUG("No RENDER-capable engines\n")__drm_dbg(DRM_UT_CORE, "No RENDER-capable engines\n"); |
3603 | return -EINVAL22; |
3604 | } |
3605 | |
3606 | /* Considering that ID = 0 is reserved and assuming that we don't |
3607 | * (currently) expect any configurations to ever specify duplicate |
3608 | * values for a particular property ID then the last _PROP_MAX value is |
3609 | * one greater than the maximum number of properties we expect to get |
3610 | * from userspace. |
3611 | */ |
3612 | if (n_props >= DRM_I915_PERF_PROP_MAX) { |
3613 | DRM_DEBUG("More i915 perf properties specified than exist\n")__drm_dbg(DRM_UT_CORE, "More i915 perf properties specified than exist\n" ); |
3614 | return -EINVAL22; |
3615 | } |
3616 | |
3617 | for (i = 0; i < n_props; i++) { |
3618 | u64 oa_period, oa_freq_hz; |
3619 | u64 id, value; |
3620 | |
3621 | ret = get_user(id, uprop)-copyin(uprop, &(id), sizeof(id)); |
3622 | if (ret) |
3623 | return ret; |
3624 | |
3625 | ret = get_user(value, uprop + 1)-copyin(uprop + 1, &(value), sizeof(value)); |
3626 | if (ret) |
3627 | return ret; |
3628 | |
3629 | if (id == 0 || id >= DRM_I915_PERF_PROP_MAX) { |
3630 | DRM_DEBUG("Unknown i915 perf property ID\n")__drm_dbg(DRM_UT_CORE, "Unknown i915 perf property ID\n"); |
3631 | return -EINVAL22; |
3632 | } |
3633 | |
3634 | switch ((enum drm_i915_perf_property_id)id) { |
3635 | case DRM_I915_PERF_PROP_CTX_HANDLE: |
3636 | props->single_context = 1; |
3637 | props->ctx_handle = value; |
3638 | break; |
3639 | case DRM_I915_PERF_PROP_SAMPLE_OA: |
3640 | if (value) |
3641 | props->sample_flags |= SAMPLE_OA_REPORT(1<<0); |
3642 | break; |
3643 | case DRM_I915_PERF_PROP_OA_METRICS_SET: |
3644 | if (value == 0) { |
3645 | DRM_DEBUG("Unknown OA metric set ID\n")__drm_dbg(DRM_UT_CORE, "Unknown OA metric set ID\n"); |
3646 | return -EINVAL22; |
3647 | } |
3648 | props->metrics_set = value; |
3649 | break; |
3650 | case DRM_I915_PERF_PROP_OA_FORMAT: |
3651 | if (value == 0 || value >= I915_OA_FORMAT_MAX) { |
3652 | DRM_DEBUG("Out-of-range OA report format %llu\n",__drm_dbg(DRM_UT_CORE, "Out-of-range OA report format %llu\n" , value) |
3653 | value)__drm_dbg(DRM_UT_CORE, "Out-of-range OA report format %llu\n" , value); |
3654 | return -EINVAL22; |
3655 | } |
3656 | if (!perf->oa_formats[value].size) { |
3657 | DRM_DEBUG("Unsupported OA report format %llu\n",__drm_dbg(DRM_UT_CORE, "Unsupported OA report format %llu\n", value) |
3658 | value)__drm_dbg(DRM_UT_CORE, "Unsupported OA report format %llu\n", value); |
3659 | return -EINVAL22; |
3660 | } |
3661 | props->oa_format = value; |
3662 | break; |
3663 | case DRM_I915_PERF_PROP_OA_EXPONENT: |
3664 | if (value > OA_EXPONENT_MAX31) { |
3665 | DRM_DEBUG("OA timer exponent too high (> %u)\n",__drm_dbg(DRM_UT_CORE, "OA timer exponent too high (> %u)\n" , 31) |
3666 | OA_EXPONENT_MAX)__drm_dbg(DRM_UT_CORE, "OA timer exponent too high (> %u)\n" , 31); |
3667 | return -EINVAL22; |
3668 | } |
3669 | |
3670 | /* Theoretically we can program the OA unit to sample |
3671 | * e.g. every 160ns for HSW, 167ns for BDW/SKL or 104ns |
3672 | * for BXT. We don't allow such high sampling |
3673 | * frequencies by default unless root. |
3674 | */ |
3675 | |
3676 | BUILD_BUG_ON(sizeof(oa_period) != 8)extern char _ctassert[(!(sizeof(oa_period) != 8)) ? 1 : -1 ] __attribute__ ((__unused__)); |
3677 | oa_period = oa_exponent_to_ns(perf, value); |
3678 | |
3679 | /* This check is primarily to ensure that oa_period <= |
3680 | * UINT32_MAX (before passing to do_div which only |
3681 | * accepts a u32 denominator), but we can also skip |
3682 | * checking anything < 1Hz which implicitly can't be |
3683 | * limited via an integer oa_max_sample_rate. |
3684 | */ |
3685 | if (oa_period <= NSEC_PER_SEC1000000000L) { |
3686 | u64 tmp = NSEC_PER_SEC1000000000L; |
3687 | do_div(tmp, oa_period)({ uint32_t __base = (oa_period); uint32_t __rem = ((uint64_t )(tmp)) % __base; (tmp) = ((uint64_t)(tmp)) / __base; __rem; } ); |
3688 | oa_freq_hz = tmp; |
3689 | } else |
3690 | oa_freq_hz = 0; |
3691 | |
3692 | if (oa_freq_hz > i915_oa_max_sample_rate && !perfmon_capable()) { |
3693 | DRM_DEBUG("OA exponent would exceed the max sampling frequency (sysctl dev.i915.oa_max_sample_rate) %uHz without CAP_PERFMON or CAP_SYS_ADMIN privileges\n",__drm_dbg(DRM_UT_CORE, "OA exponent would exceed the max sampling frequency (sysctl dev.i915.oa_max_sample_rate) %uHz without CAP_PERFMON or CAP_SYS_ADMIN privileges\n" , i915_oa_max_sample_rate) |
3694 | i915_oa_max_sample_rate)__drm_dbg(DRM_UT_CORE, "OA exponent would exceed the max sampling frequency (sysctl dev.i915.oa_max_sample_rate) %uHz without CAP_PERFMON or CAP_SYS_ADMIN privileges\n" , i915_oa_max_sample_rate); |
3695 | return -EACCES13; |
3696 | } |
3697 | |
3698 | props->oa_periodic = true1; |
3699 | props->oa_period_exponent = value; |
3700 | break; |
3701 | case DRM_I915_PERF_PROP_HOLD_PREEMPTION: |
3702 | props->hold_preemption = !!value; |
3703 | break; |
3704 | case DRM_I915_PERF_PROP_GLOBAL_SSEU: { |
3705 | struct drm_i915_gem_context_param_sseu user_sseu; |
3706 | |
3707 | if (copy_from_user(&user_sseu, |
3708 | u64_to_user_ptr(value)((void *)(uintptr_t)(value)), |
3709 | sizeof(user_sseu))) { |
3710 | DRM_DEBUG("Unable to copy global sseu parameter\n")__drm_dbg(DRM_UT_CORE, "Unable to copy global sseu parameter\n" ); |
3711 | return -EFAULT14; |
3712 | } |
3713 | |
3714 | ret = get_sseu_config(&props->sseu, props->engine, &user_sseu); |
3715 | if (ret) { |
3716 | DRM_DEBUG("Invalid SSEU configuration\n")__drm_dbg(DRM_UT_CORE, "Invalid SSEU configuration\n"); |
3717 | return ret; |
3718 | } |
3719 | props->has_sseu = true1; |
3720 | break; |
3721 | } |
3722 | case DRM_I915_PERF_PROP_POLL_OA_PERIOD: |
3723 | if (value < 100000 /* 100us */) { |
3724 | DRM_DEBUG("OA availability timer too small (%lluns < 100us)\n",__drm_dbg(DRM_UT_CORE, "OA availability timer too small (%lluns < 100us)\n" , value) |
3725 | value)__drm_dbg(DRM_UT_CORE, "OA availability timer too small (%lluns < 100us)\n" , value); |
3726 | return -EINVAL22; |
3727 | } |
3728 | props->poll_oa_period = value; |
3729 | break; |
3730 | case DRM_I915_PERF_PROP_MAX: |
3731 | MISSING_CASE(id)({ int __ret = !!(1); if (__ret) printf("Missing case (%s == %ld)\n" , "id", (long)(id)); __builtin_expect(!!(__ret), 0); }); |
3732 | return -EINVAL22; |
3733 | } |
3734 | |
3735 | uprop += 2; |
3736 | } |
3737 | |
3738 | return 0; |
3739 | } |
3740 | |
3741 | /** |
3742 | * i915_perf_open_ioctl - DRM ioctl() for userspace to open a stream FD |
3743 | * @dev: drm device |
3744 | * @data: ioctl data copied from userspace (unvalidated) |
3745 | * @file: drm file |
3746 | * |
3747 | * Validates the stream open parameters given by userspace including flags |
3748 | * and an array of u64 key, value pair properties. |
3749 | * |
3750 | * Very little is assumed up front about the nature of the stream being |
3751 | * opened (for instance we don't assume it's for periodic OA unit metrics). An |
3752 | * i915-perf stream is expected to be a suitable interface for other forms of |
3753 | * buffered data written by the GPU besides periodic OA metrics. |
3754 | * |
3755 | * Note we copy the properties from userspace outside of the i915 perf |
3756 | * mutex to avoid an awkward lockdep with mmap_lock. |
3757 | * |
3758 | * Most of the implementation details are handled by |
3759 | * i915_perf_open_ioctl_locked() after taking the &perf->lock |
3760 | * mutex for serializing with any non-file-operation driver hooks. |
3761 | * |
3762 | * Return: A newly opened i915 Perf stream file descriptor or negative |
3763 | * error code on failure. |
3764 | */ |
3765 | int i915_perf_open_ioctl(struct drm_device *dev, void *data, |
3766 | struct drm_file *file) |
3767 | { |
3768 | struct i915_perf *perf = &to_i915(dev)->perf; |
3769 | struct drm_i915_perf_open_param *param = data; |
3770 | struct perf_open_properties props; |
3771 | u32 known_open_flags; |
3772 | int ret; |
3773 | |
3774 | if (!perf->i915) { |
3775 | DRM_DEBUG("i915 perf interface not available for this system\n")__drm_dbg(DRM_UT_CORE, "i915 perf interface not available for this system\n" ); |
3776 | return -ENOTSUPP91; |
3777 | } |
3778 | |
3779 | known_open_flags = I915_PERF_FLAG_FD_CLOEXEC(1<<0) | |
3780 | I915_PERF_FLAG_FD_NONBLOCK(1<<1) | |
3781 | I915_PERF_FLAG_DISABLED(1<<2); |
3782 | if (param->flags & ~known_open_flags) { |
3783 | DRM_DEBUG("Unknown drm_i915_perf_open_param flag\n")__drm_dbg(DRM_UT_CORE, "Unknown drm_i915_perf_open_param flag\n" ); |
3784 | return -EINVAL22; |
3785 | } |
3786 | |
3787 | ret = read_properties_unlocked(perf, |
3788 | u64_to_user_ptr(param->properties_ptr)((void *)(uintptr_t)(param->properties_ptr)), |
3789 | param->num_properties, |
3790 | &props); |
3791 | if (ret) |
3792 | return ret; |
3793 | |
3794 | mutex_lock(&perf->lock)rw_enter_write(&perf->lock); |
3795 | ret = i915_perf_open_ioctl_locked(perf, param, &props, file); |
3796 | mutex_unlock(&perf->lock)rw_exit_write(&perf->lock); |
3797 | |
3798 | return ret; |
3799 | } |
3800 | |
3801 | /** |
3802 | * i915_perf_register - exposes i915-perf to userspace |
3803 | * @i915: i915 device instance |
3804 | * |
3805 | * In particular OA metric sets are advertised under a sysfs metrics/ |
3806 | * directory allowing userspace to enumerate valid IDs that can be |
3807 | * used to open an i915-perf stream. |
3808 | */ |
3809 | void i915_perf_register(struct drm_i915_privateinteldrm_softc *i915) |
3810 | { |
3811 | #ifdef __linux__ |
3812 | struct i915_perf *perf = &i915->perf; |
3813 | |
3814 | if (!perf->i915) |
3815 | return; |
3816 | |
3817 | /* To be sure we're synchronized with an attempted |
3818 | * i915_perf_open_ioctl(); considering that we register after |
3819 | * being exposed to userspace. |
3820 | */ |
3821 | mutex_lock(&perf->lock)rw_enter_write(&perf->lock); |
3822 | |
3823 | perf->metrics_kobj = |
3824 | kobject_create_and_add("metrics", |
3825 | &i915->drm.primary->kdev->kobj); |
3826 | |
3827 | mutex_unlock(&perf->lock)rw_exit_write(&perf->lock); |
3828 | #endif |
3829 | } |
3830 | |
3831 | /** |
3832 | * i915_perf_unregister - hide i915-perf from userspace |
3833 | * @i915: i915 device instance |
3834 | * |
3835 | * i915-perf state cleanup is split up into an 'unregister' and |
3836 | * 'deinit' phase where the interface is first hidden from |
3837 | * userspace by i915_perf_unregister() before cleaning up |
3838 | * remaining state in i915_perf_fini(). |
3839 | */ |
3840 | void i915_perf_unregister(struct drm_i915_privateinteldrm_softc *i915) |
3841 | { |
3842 | struct i915_perf *perf = &i915->perf; |
3843 | |
3844 | if (!perf->metrics_kobj) |
3845 | return; |
3846 | |
3847 | kobject_put(perf->metrics_kobj); |
3848 | perf->metrics_kobj = NULL((void *)0); |
3849 | } |
3850 | |
3851 | static bool_Bool gen8_is_valid_flex_addr(struct i915_perf *perf, u32 addr) |
3852 | { |
3853 | static const i915_reg_t flex_eu_regs[] = { |
3854 | EU_PERF_CNTL0((const i915_reg_t){ .reg = (0xe458) }), |
3855 | EU_PERF_CNTL1((const i915_reg_t){ .reg = (0xe558) }), |
3856 | EU_PERF_CNTL2((const i915_reg_t){ .reg = (0xe658) }), |
3857 | EU_PERF_CNTL3((const i915_reg_t){ .reg = (0xe758) }), |
3858 | EU_PERF_CNTL4((const i915_reg_t){ .reg = (0xe45c) }), |
3859 | EU_PERF_CNTL5((const i915_reg_t){ .reg = (0xe55c) }), |
3860 | EU_PERF_CNTL6((const i915_reg_t){ .reg = (0xe65c) }), |
3861 | }; |
3862 | int i; |
3863 | |
3864 | for (i = 0; i < ARRAY_SIZE(flex_eu_regs)(sizeof((flex_eu_regs)) / sizeof((flex_eu_regs)[0])); i++) { |
3865 | if (i915_mmio_reg_offset(flex_eu_regs[i]) == addr) |
3866 | return true1; |
3867 | } |
3868 | return false0; |
3869 | } |
3870 | |
3871 | #define ADDR_IN_RANGE(addr, start, end)((addr) >= (start) && (addr) <= (end)) \ |
3872 | ((addr) >= (start) && \ |
3873 | (addr) <= (end)) |
3874 | |
3875 | #define REG_IN_RANGE(addr, start, end)((addr) >= i915_mmio_reg_offset(start) && (addr) <= i915_mmio_reg_offset(end)) \ |
3876 | ((addr) >= i915_mmio_reg_offset(start) && \ |
3877 | (addr) <= i915_mmio_reg_offset(end)) |
3878 | |
3879 | #define REG_EQUAL(addr, mmio)((addr) == i915_mmio_reg_offset(mmio)) \ |
3880 | ((addr) == i915_mmio_reg_offset(mmio)) |
3881 | |
3882 | static bool_Bool gen7_is_valid_b_counter_addr(struct i915_perf *perf, u32 addr) |
3883 | { |
3884 | return REG_IN_RANGE(addr, OASTARTTRIG1, OASTARTTRIG8)((addr) >= i915_mmio_reg_offset(((const i915_reg_t){ .reg = (0x2710) })) && (addr) <= i915_mmio_reg_offset((( const i915_reg_t){ .reg = (0x272c) }))) || |
3885 | REG_IN_RANGE(addr, OAREPORTTRIG1, OAREPORTTRIG8)((addr) >= i915_mmio_reg_offset(((const i915_reg_t){ .reg = (0x2740) })) && (addr) <= i915_mmio_reg_offset((( const i915_reg_t){ .reg = (0x275c) }))) || |
3886 | REG_IN_RANGE(addr, OACEC0_0, OACEC7_1)((addr) >= i915_mmio_reg_offset(((const i915_reg_t){ .reg = (0x2770) })) && (addr) <= i915_mmio_reg_offset((( const i915_reg_t){ .reg = (0x27ac) }))); |
3887 | } |
3888 | |
3889 | static bool_Bool gen7_is_valid_mux_addr(struct i915_perf *perf, u32 addr) |
3890 | { |
3891 | return REG_EQUAL(addr, HALF_SLICE_CHICKEN2)((addr) == i915_mmio_reg_offset(((const i915_reg_t){ .reg = ( 0xe180) }))) || |
3892 | REG_IN_RANGE(addr, MICRO_BP0_0, NOA_WRITE)((addr) >= i915_mmio_reg_offset(((const i915_reg_t){ .reg = (0x9800) })) && (addr) <= i915_mmio_reg_offset((( const i915_reg_t){ .reg = (0x9888) }))) || |
3893 | REG_IN_RANGE(addr, OA_PERFCNT1_LO, OA_PERFCNT2_HI)((addr) >= i915_mmio_reg_offset(((const i915_reg_t){ .reg = (0x91B8) })) && (addr) <= i915_mmio_reg_offset((( const i915_reg_t){ .reg = (0x91C4) }))) || |
3894 | REG_IN_RANGE(addr, OA_PERFMATRIX_LO, OA_PERFMATRIX_HI)((addr) >= i915_mmio_reg_offset(((const i915_reg_t){ .reg = (0x91C8) })) && (addr) <= i915_mmio_reg_offset((( const i915_reg_t){ .reg = (0x91CC) }))); |
3895 | } |
3896 | |
3897 | static bool_Bool gen8_is_valid_mux_addr(struct i915_perf *perf, u32 addr) |
3898 | { |
3899 | return gen7_is_valid_mux_addr(perf, addr) || |
3900 | REG_EQUAL(addr, WAIT_FOR_RC6_EXIT)((addr) == i915_mmio_reg_offset(((const i915_reg_t){ .reg = ( 0x20CC) }))) || |
3901 | REG_IN_RANGE(addr, RPM_CONFIG0, NOA_CONFIG(8))((addr) >= i915_mmio_reg_offset(((const i915_reg_t){ .reg = (0x0D00) })) && (addr) <= i915_mmio_reg_offset((( const i915_reg_t){ .reg = (0x0D0C + (8) * 4) }))); |
3902 | } |
3903 | |
3904 | static bool_Bool gen10_is_valid_mux_addr(struct i915_perf *perf, u32 addr) |
3905 | { |
3906 | return gen8_is_valid_mux_addr(perf, addr) || |
3907 | REG_EQUAL(addr, GEN10_NOA_WRITE_HIGH)((addr) == i915_mmio_reg_offset(((const i915_reg_t){ .reg = ( 0x9884) }))) || |
3908 | REG_IN_RANGE(addr, OA_PERFCNT3_LO, OA_PERFCNT4_HI)((addr) >= i915_mmio_reg_offset(((const i915_reg_t){ .reg = (0x91C8) })) && (addr) <= i915_mmio_reg_offset((( const i915_reg_t){ .reg = (0x91DC) }))); |
3909 | } |
3910 | |
3911 | static bool_Bool hsw_is_valid_mux_addr(struct i915_perf *perf, u32 addr) |
3912 | { |
3913 | return gen7_is_valid_mux_addr(perf, addr) || |
3914 | ADDR_IN_RANGE(addr, 0x25100, 0x2FF90)((addr) >= (0x25100) && (addr) <= (0x2FF90)) || |
3915 | REG_IN_RANGE(addr, HSW_MBVID2_NOA0, HSW_MBVID2_NOA9)((addr) >= i915_mmio_reg_offset(((const i915_reg_t){ .reg = (0x9E80) })) && (addr) <= i915_mmio_reg_offset((( const i915_reg_t){ .reg = (0x9EA4) }))) || |
3916 | REG_EQUAL(addr, HSW_MBVID2_MISR0)((addr) == i915_mmio_reg_offset(((const i915_reg_t){ .reg = ( 0x9EC0) }))); |
3917 | } |
3918 | |
3919 | static bool_Bool chv_is_valid_mux_addr(struct i915_perf *perf, u32 addr) |
3920 | { |
3921 | return gen7_is_valid_mux_addr(perf, addr) || |
3922 | ADDR_IN_RANGE(addr, 0x182300, 0x1823A4)((addr) >= (0x182300) && (addr) <= (0x1823A4)); |
3923 | } |
3924 | |
3925 | static bool_Bool gen12_is_valid_b_counter_addr(struct i915_perf *perf, u32 addr) |
3926 | { |
3927 | return REG_IN_RANGE(addr, GEN12_OAG_OASTARTTRIG1, GEN12_OAG_OASTARTTRIG8)((addr) >= i915_mmio_reg_offset(((const i915_reg_t){ .reg = (0xd900) })) && (addr) <= i915_mmio_reg_offset((( const i915_reg_t){ .reg = (0xd91c) }))) || |
3928 | REG_IN_RANGE(addr, GEN12_OAG_OAREPORTTRIG1, GEN12_OAG_OAREPORTTRIG8)((addr) >= i915_mmio_reg_offset(((const i915_reg_t){ .reg = (0xd920) })) && (addr) <= i915_mmio_reg_offset((( const i915_reg_t){ .reg = (0xd93c) }))) || |
3929 | REG_IN_RANGE(addr, GEN12_OAG_CEC0_0, GEN12_OAG_CEC7_1)((addr) >= i915_mmio_reg_offset(((const i915_reg_t){ .reg = (0xd940) })) && (addr) <= i915_mmio_reg_offset((( const i915_reg_t){ .reg = (0xd97c) }))) || |
3930 | REG_IN_RANGE(addr, GEN12_OAG_SCEC0_0, GEN12_OAG_SCEC7_1)((addr) >= i915_mmio_reg_offset(((const i915_reg_t){ .reg = (0xdc00) })) && (addr) <= i915_mmio_reg_offset((( const i915_reg_t){ .reg = (0xdc3c) }))) || |
3931 | REG_EQUAL(addr, GEN12_OAA_DBG_REG)((addr) == i915_mmio_reg_offset(((const i915_reg_t){ .reg = ( 0xdc44) }))) || |
3932 | REG_EQUAL(addr, GEN12_OAG_OA_PESS)((addr) == i915_mmio_reg_offset(((const i915_reg_t){ .reg = ( 0x2b2c) }))) || |
3933 | REG_EQUAL(addr, GEN12_OAG_SPCTR_CNF)((addr) == i915_mmio_reg_offset(((const i915_reg_t){ .reg = ( 0xdc40) }))); |
3934 | } |
3935 | |
3936 | static bool_Bool gen12_is_valid_mux_addr(struct i915_perf *perf, u32 addr) |
3937 | { |
3938 | return REG_EQUAL(addr, NOA_WRITE)((addr) == i915_mmio_reg_offset(((const i915_reg_t){ .reg = ( 0x9888) }))) || |
3939 | REG_EQUAL(addr, GEN10_NOA_WRITE_HIGH)((addr) == i915_mmio_reg_offset(((const i915_reg_t){ .reg = ( 0x9884) }))) || |
3940 | REG_EQUAL(addr, GDT_CHICKEN_BITS)((addr) == i915_mmio_reg_offset(((const i915_reg_t){ .reg = ( 0x9840) }))) || |
3941 | REG_EQUAL(addr, WAIT_FOR_RC6_EXIT)((addr) == i915_mmio_reg_offset(((const i915_reg_t){ .reg = ( 0x20CC) }))) || |
3942 | REG_EQUAL(addr, RPM_CONFIG0)((addr) == i915_mmio_reg_offset(((const i915_reg_t){ .reg = ( 0x0D00) }))) || |
3943 | REG_EQUAL(addr, RPM_CONFIG1)((addr) == i915_mmio_reg_offset(((const i915_reg_t){ .reg = ( 0x0D04) }))) || |
3944 | REG_IN_RANGE(addr, NOA_CONFIG(0), NOA_CONFIG(8))((addr) >= i915_mmio_reg_offset(((const i915_reg_t){ .reg = (0x0D0C + (0) * 4) })) && (addr) <= i915_mmio_reg_offset (((const i915_reg_t){ .reg = (0x0D0C + (8) * 4) }))); |
3945 | } |
3946 | |
3947 | static u32 mask_reg_value(u32 reg, u32 val) |
3948 | { |
3949 | /* HALF_SLICE_CHICKEN2 is programmed with a the |
3950 | * WaDisableSTUnitPowerOptimization workaround. Make sure the value |
3951 | * programmed by userspace doesn't change this. |
3952 | */ |
3953 | if (REG_EQUAL(reg, HALF_SLICE_CHICKEN2)((reg) == i915_mmio_reg_offset(((const i915_reg_t){ .reg = (0xe180 ) })))) |
3954 | val = val & ~_MASKED_BIT_ENABLE(GEN8_ST_PO_DISABLE)({ typeof((1 << 13)) _a = ((1 << 13)); ({ if (__builtin_constant_p (_a)) do { } while (0); if (__builtin_constant_p(_a)) do { } while (0); if (__builtin_constant_p(_a) && __builtin_constant_p (_a)) do { } while (0); ((_a) << 16 | (_a)); }); }); |
3955 | |
3956 | /* WAIT_FOR_RC6_EXIT has only one bit fullfilling the function |
3957 | * indicated by its name and a bunch of selection fields used by OA |
3958 | * configs. |
3959 | */ |
3960 | if (REG_EQUAL(reg, WAIT_FOR_RC6_EXIT)((reg) == i915_mmio_reg_offset(((const i915_reg_t){ .reg = (0x20CC ) })))) |
3961 | val = val & ~_MASKED_BIT_ENABLE(HSW_WAIT_FOR_RC6_EXIT_ENABLE)({ typeof((1 << 0)) _a = ((1 << 0)); ({ if (__builtin_constant_p (_a)) do { } while (0); if (__builtin_constant_p(_a)) do { } while (0); if (__builtin_constant_p(_a) && __builtin_constant_p (_a)) do { } while (0); ((_a) << 16 | (_a)); }); }); |
3962 | |
3963 | return val; |
3964 | } |
3965 | |
3966 | static struct i915_oa_reg *alloc_oa_regs(struct i915_perf *perf, |
3967 | bool_Bool (*is_valid)(struct i915_perf *perf, u32 addr), |
3968 | u32 __user *regs, |
3969 | u32 n_regs) |
3970 | { |
3971 | struct i915_oa_reg *oa_regs; |
3972 | int err; |
3973 | u32 i; |
3974 | |
3975 | if (!n_regs) |
3976 | return NULL((void *)0); |
3977 | |
3978 | /* No is_valid function means we're not allowing any register to be programmed. */ |
3979 | GEM_BUG_ON(!is_valid)((void)0); |
3980 | if (!is_valid) |
3981 | return ERR_PTR(-EINVAL22); |
3982 | |
3983 | oa_regs = kmalloc_array(n_regs, sizeof(*oa_regs), GFP_KERNEL(0x0001 | 0x0004)); |
3984 | if (!oa_regs) |
3985 | return ERR_PTR(-ENOMEM12); |
3986 | |
3987 | for (i = 0; i < n_regs; i++) { |
3988 | u32 addr, value; |
3989 | |
3990 | err = get_user(addr, regs)-copyin(regs, &(addr), sizeof(addr)); |
3991 | if (err) |
3992 | goto addr_err; |
3993 | |
3994 | if (!is_valid(perf, addr)) { |
3995 | DRM_DEBUG("Invalid oa_reg address: %X\n", addr)__drm_dbg(DRM_UT_CORE, "Invalid oa_reg address: %X\n", addr); |
3996 | err = -EINVAL22; |
3997 | goto addr_err; |
3998 | } |
3999 | |
4000 | err = get_user(value, regs + 1)-copyin(regs + 1, &(value), sizeof(value)); |
4001 | if (err) |
4002 | goto addr_err; |
4003 | |
4004 | oa_regs[i].addr = _MMIO(addr)((const i915_reg_t){ .reg = (addr) }); |
4005 | oa_regs[i].value = mask_reg_value(addr, value); |
4006 | |
4007 | regs += 2; |
4008 | } |
4009 | |
4010 | return oa_regs; |
4011 | |
4012 | addr_err: |
4013 | kfree(oa_regs); |
4014 | return ERR_PTR(err); |
4015 | } |
4016 | |
4017 | static ssize_t show_dynamic_id(struct device *dev, |
4018 | struct device_attribute *attr, |
4019 | char *buf) |
4020 | { |
4021 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
4022 | return -ENOSYS78; |
4023 | #ifdef notyet |
4024 | struct i915_oa_config *oa_config = |
4025 | container_of(attr, typeof(*oa_config), sysfs_metric_id)({ const __typeof( ((typeof(*oa_config) *)0)->sysfs_metric_id ) *__mptr = (attr); (typeof(*oa_config) *)( (char *)__mptr - __builtin_offsetof(typeof(*oa_config), sysfs_metric_id) );}); |
4026 | |
4027 | return sprintf(buf, "%d\n", oa_config->id); |
4028 | #endif |
4029 | } |
4030 | |
4031 | static int create_dynamic_oa_sysfs_entry(struct i915_perf *perf, |
4032 | struct i915_oa_config *oa_config) |
4033 | { |
4034 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
4035 | return -ENOSYS78; |
4036 | #ifdef notyet |
4037 | sysfs_attr_init(&oa_config->sysfs_metric_id.attr); |
4038 | oa_config->sysfs_metric_id.attr.name = "id"; |
4039 | oa_config->sysfs_metric_id.attr.mode = S_IRUGO; |
4040 | oa_config->sysfs_metric_id.show = show_dynamic_id; |
4041 | oa_config->sysfs_metric_id.store = NULL((void *)0); |
4042 | |
4043 | oa_config->attrs[0] = &oa_config->sysfs_metric_id.attr; |
4044 | oa_config->attrs[1] = NULL((void *)0); |
4045 | |
4046 | oa_config->sysfs_metric.name = oa_config->uuid; |
4047 | oa_config->sysfs_metric.attrs = oa_config->attrs; |
4048 | |
4049 | return sysfs_create_group(perf->metrics_kobj,0 |
4050 | &oa_config->sysfs_metric)0; |
4051 | #endif |
4052 | } |
4053 | |
4054 | /** |
4055 | * i915_perf_add_config_ioctl - DRM ioctl() for userspace to add a new OA config |
4056 | * @dev: drm device |
4057 | * @data: ioctl data (pointer to struct drm_i915_perf_oa_config) copied from |
4058 | * userspace (unvalidated) |
4059 | * @file: drm file |
4060 | * |
4061 | * Validates the submitted OA register to be saved into a new OA config that |
4062 | * can then be used for programming the OA unit and its NOA network. |
4063 | * |
4064 | * Returns: A new allocated config number to be used with the perf open ioctl |
4065 | * or a negative error code on failure. |
4066 | */ |
4067 | int i915_perf_add_config_ioctl(struct drm_device *dev, void *data, |
4068 | struct drm_file *file) |
4069 | { |
4070 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
4071 | return -ENOSYS78; |
4072 | #ifdef notyet |
4073 | struct i915_perf *perf = &to_i915(dev)->perf; |
4074 | struct drm_i915_perf_oa_config *args = data; |
4075 | struct i915_oa_config *oa_config, *tmp; |
4076 | struct i915_oa_reg *regs; |
4077 | int err, id; |
4078 | |
4079 | if (!perf->i915) { |
4080 | DRM_DEBUG("i915 perf interface not available for this system\n")__drm_dbg(DRM_UT_CORE, "i915 perf interface not available for this system\n" ); |
4081 | return -ENOTSUPP91; |
4082 | } |
4083 | |
4084 | if (!perf->metrics_kobj) { |
4085 | DRM_DEBUG("OA metrics weren't advertised via sysfs\n")__drm_dbg(DRM_UT_CORE, "OA metrics weren't advertised via sysfs\n" ); |
4086 | return -EINVAL22; |
4087 | } |
4088 | |
4089 | if (i915_perf_stream_paranoid && !perfmon_capable()) { |
4090 | DRM_DEBUG("Insufficient privileges to add i915 OA config\n")__drm_dbg(DRM_UT_CORE, "Insufficient privileges to add i915 OA config\n" ); |
4091 | return -EACCES13; |
4092 | } |
4093 | |
4094 | if ((!args->mux_regs_ptr || !args->n_mux_regs) && |
4095 | (!args->boolean_regs_ptr || !args->n_boolean_regs) && |
4096 | (!args->flex_regs_ptr || !args->n_flex_regs)) { |
4097 | DRM_DEBUG("No OA registers given\n")__drm_dbg(DRM_UT_CORE, "No OA registers given\n"); |
4098 | return -EINVAL22; |
4099 | } |
4100 | |
4101 | oa_config = kzalloc(sizeof(*oa_config), GFP_KERNEL(0x0001 | 0x0004)); |
4102 | if (!oa_config) { |
4103 | DRM_DEBUG("Failed to allocate memory for the OA config\n")__drm_dbg(DRM_UT_CORE, "Failed to allocate memory for the OA config\n" ); |
4104 | return -ENOMEM12; |
4105 | } |
4106 | |
4107 | oa_config->perf = perf; |
4108 | kref_init(&oa_config->ref); |
4109 | |
4110 | if (!uuid_is_valid(args->uuid)) { |
4111 | DRM_DEBUG("Invalid uuid format for OA config\n")__drm_dbg(DRM_UT_CORE, "Invalid uuid format for OA config\n"); |
4112 | err = -EINVAL22; |
4113 | goto reg_err; |
4114 | } |
4115 | |
4116 | /* Last character in oa_config->uuid will be 0 because oa_config is |
4117 | * kzalloc. |
4118 | */ |
4119 | memcpy(oa_config->uuid, args->uuid, sizeof(args->uuid))__builtin_memcpy((oa_config->uuid), (args->uuid), (sizeof (args->uuid))); |
4120 | |
4121 | oa_config->mux_regs_len = args->n_mux_regs; |
4122 | regs = alloc_oa_regs(perf, |
4123 | perf->ops.is_valid_mux_reg, |
4124 | u64_to_user_ptr(args->mux_regs_ptr)((void *)(uintptr_t)(args->mux_regs_ptr)), |
4125 | args->n_mux_regs); |
4126 | |
4127 | if (IS_ERR(regs)) { |
4128 | DRM_DEBUG("Failed to create OA config for mux_regs\n")__drm_dbg(DRM_UT_CORE, "Failed to create OA config for mux_regs\n" ); |
4129 | err = PTR_ERR(regs); |
4130 | goto reg_err; |
4131 | } |
4132 | oa_config->mux_regs = regs; |
4133 | |
4134 | oa_config->b_counter_regs_len = args->n_boolean_regs; |
4135 | regs = alloc_oa_regs(perf, |
4136 | perf->ops.is_valid_b_counter_reg, |
4137 | u64_to_user_ptr(args->boolean_regs_ptr)((void *)(uintptr_t)(args->boolean_regs_ptr)), |
4138 | args->n_boolean_regs); |
4139 | |
4140 | if (IS_ERR(regs)) { |
4141 | DRM_DEBUG("Failed to create OA config for b_counter_regs\n")__drm_dbg(DRM_UT_CORE, "Failed to create OA config for b_counter_regs\n" ); |
4142 | err = PTR_ERR(regs); |
4143 | goto reg_err; |
4144 | } |
4145 | oa_config->b_counter_regs = regs; |
4146 | |
4147 | if (INTEL_GEN(perf->i915)((&(perf->i915)->__info)->gen) < 8) { |
4148 | if (args->n_flex_regs != 0) { |
4149 | err = -EINVAL22; |
4150 | goto reg_err; |
4151 | } |
4152 | } else { |
4153 | oa_config->flex_regs_len = args->n_flex_regs; |
4154 | regs = alloc_oa_regs(perf, |
4155 | perf->ops.is_valid_flex_reg, |
4156 | u64_to_user_ptr(args->flex_regs_ptr)((void *)(uintptr_t)(args->flex_regs_ptr)), |
4157 | args->n_flex_regs); |
4158 | |
4159 | if (IS_ERR(regs)) { |
4160 | DRM_DEBUG("Failed to create OA config for flex_regs\n")__drm_dbg(DRM_UT_CORE, "Failed to create OA config for flex_regs\n" ); |
4161 | err = PTR_ERR(regs); |
4162 | goto reg_err; |
4163 | } |
4164 | oa_config->flex_regs = regs; |
4165 | } |
4166 | |
4167 | err = mutex_lock_interruptible(&perf->metrics_lock); |
4168 | if (err) |
4169 | goto reg_err; |
4170 | |
4171 | /* We shouldn't have too many configs, so this iteration shouldn't be |
4172 | * too costly. |
4173 | */ |
4174 | idr_for_each_entry(&perf->metrics_idr, tmp, id)for (id = 0; ((tmp) = idr_get_next(&perf->metrics_idr, &(id))) != ((void *)0); id++) { |
4175 | if (!strcmp(tmp->uuid, oa_config->uuid)) { |
4176 | DRM_DEBUG("OA config already exists with this uuid\n")__drm_dbg(DRM_UT_CORE, "OA config already exists with this uuid\n" ); |
4177 | err = -EADDRINUSE48; |
4178 | goto sysfs_err; |
4179 | } |
4180 | } |
4181 | |
4182 | err = create_dynamic_oa_sysfs_entry(perf, oa_config); |
4183 | if (err) { |
4184 | DRM_DEBUG("Failed to create sysfs entry for OA config\n")__drm_dbg(DRM_UT_CORE, "Failed to create sysfs entry for OA config\n" ); |
4185 | goto sysfs_err; |
4186 | } |
4187 | |
4188 | /* Config id 0 is invalid, id 1 for kernel stored test config. */ |
4189 | oa_config->id = idr_alloc(&perf->metrics_idr, |
4190 | oa_config, 2, |
4191 | 0, GFP_KERNEL(0x0001 | 0x0004)); |
4192 | if (oa_config->id < 0) { |
4193 | DRM_DEBUG("Failed to create sysfs entry for OA config\n")__drm_dbg(DRM_UT_CORE, "Failed to create sysfs entry for OA config\n" ); |
4194 | err = oa_config->id; |
4195 | goto sysfs_err; |
4196 | } |
4197 | |
4198 | mutex_unlock(&perf->metrics_lock)rw_exit_write(&perf->metrics_lock); |
4199 | |
4200 | DRM_DEBUG("Added config %s id=%i\n", oa_config->uuid, oa_config->id)__drm_dbg(DRM_UT_CORE, "Added config %s id=%i\n", oa_config-> uuid, oa_config->id); |
4201 | |
4202 | return oa_config->id; |
4203 | |
4204 | sysfs_err: |
4205 | mutex_unlock(&perf->metrics_lock)rw_exit_write(&perf->metrics_lock); |
4206 | reg_err: |
4207 | i915_oa_config_put(oa_config); |
4208 | DRM_DEBUG("Failed to add new OA config\n")__drm_dbg(DRM_UT_CORE, "Failed to add new OA config\n"); |
4209 | return err; |
4210 | #endif |
4211 | } |
4212 | |
4213 | /** |
4214 | * i915_perf_remove_config_ioctl - DRM ioctl() for userspace to remove an OA config |
4215 | * @dev: drm device |
4216 | * @data: ioctl data (pointer to u64 integer) copied from userspace |
4217 | * @file: drm file |
4218 | * |
4219 | * Configs can be removed while being used, the will stop appearing in sysfs |
4220 | * and their content will be freed when the stream using the config is closed. |
4221 | * |
4222 | * Returns: 0 on success or a negative error code on failure. |
4223 | */ |
4224 | int i915_perf_remove_config_ioctl(struct drm_device *dev, void *data, |
4225 | struct drm_file *file) |
4226 | { |
4227 | struct i915_perf *perf = &to_i915(dev)->perf; |
4228 | u64 *arg = data; |
4229 | struct i915_oa_config *oa_config; |
4230 | int ret; |
4231 | |
4232 | if (!perf->i915) { |
4233 | DRM_DEBUG("i915 perf interface not available for this system\n")__drm_dbg(DRM_UT_CORE, "i915 perf interface not available for this system\n" ); |
4234 | return -ENOTSUPP91; |
4235 | } |
4236 | |
4237 | if (i915_perf_stream_paranoid && !perfmon_capable()) { |
4238 | DRM_DEBUG("Insufficient privileges to remove i915 OA config\n")__drm_dbg(DRM_UT_CORE, "Insufficient privileges to remove i915 OA config\n" ); |
4239 | return -EACCES13; |
4240 | } |
4241 | |
4242 | ret = mutex_lock_interruptible(&perf->metrics_lock); |
4243 | if (ret) |
4244 | return ret; |
4245 | |
4246 | oa_config = idr_find(&perf->metrics_idr, *arg); |
4247 | if (!oa_config) { |
4248 | DRM_DEBUG("Failed to remove unknown OA config\n")__drm_dbg(DRM_UT_CORE, "Failed to remove unknown OA config\n" ); |
4249 | ret = -ENOENT2; |
4250 | goto err_unlock; |
4251 | } |
4252 | |
4253 | GEM_BUG_ON(*arg != oa_config->id)((void)0); |
4254 | |
4255 | sysfs_remove_group(perf->metrics_kobj, &oa_config->sysfs_metric); |
4256 | |
4257 | idr_remove(&perf->metrics_idr, *arg); |
4258 | |
4259 | mutex_unlock(&perf->metrics_lock)rw_exit_write(&perf->metrics_lock); |
4260 | |
4261 | DRM_DEBUG("Removed config %s id=%i\n", oa_config->uuid, oa_config->id)__drm_dbg(DRM_UT_CORE, "Removed config %s id=%i\n", oa_config ->uuid, oa_config->id); |
4262 | |
4263 | i915_oa_config_put(oa_config); |
4264 | |
4265 | return 0; |
4266 | |
4267 | err_unlock: |
4268 | mutex_unlock(&perf->metrics_lock)rw_exit_write(&perf->metrics_lock); |
4269 | return ret; |
4270 | } |
4271 | |
4272 | #ifdef notyet |
4273 | static struct ctl_table oa_table[] = { |
4274 | { |
4275 | .procname = "perf_stream_paranoid", |
4276 | .data = &i915_perf_stream_paranoid, |
4277 | .maxlen = sizeof(i915_perf_stream_paranoid), |
4278 | .mode = 0644, |
4279 | .proc_handler = proc_dointvec_minmax, |
4280 | .extra1 = SYSCTL_ZERO, |
4281 | .extra2 = SYSCTL_ONE, |
4282 | }, |
4283 | { |
4284 | .procname = "oa_max_sample_rate", |
4285 | .data = &i915_oa_max_sample_rate, |
4286 | .maxlen = sizeof(i915_oa_max_sample_rate), |
4287 | .mode = 0644, |
4288 | .proc_handler = proc_dointvec_minmax, |
4289 | .extra1 = SYSCTL_ZERO, |
4290 | .extra2 = &oa_sample_rate_hard_limit, |
4291 | }, |
4292 | {} |
4293 | }; |
4294 | |
4295 | static struct ctl_table i915_root[] = { |
4296 | { |
4297 | .procname = "i915", |
4298 | .maxlen = 0, |
4299 | .mode = 0555, |
4300 | .child = oa_table, |
4301 | }, |
4302 | {} |
4303 | }; |
4304 | |
4305 | static struct ctl_table dev_root[] = { |
4306 | { |
4307 | .procname = "dev", |
4308 | .maxlen = 0, |
4309 | .mode = 0555, |
4310 | .child = i915_root, |
4311 | }, |
4312 | {} |
4313 | }; |
4314 | #endif |
4315 | |
4316 | /** |
4317 | * i915_perf_init - initialize i915-perf state on module bind |
4318 | * @i915: i915 device instance |
4319 | * |
4320 | * Initializes i915-perf state without exposing anything to userspace. |
4321 | * |
4322 | * Note: i915-perf initialization is split into an 'init' and 'register' |
4323 | * phase with the i915_perf_register() exposing state to userspace. |
4324 | */ |
4325 | void i915_perf_init(struct drm_i915_privateinteldrm_softc *i915) |
4326 | { |
4327 | struct i915_perf *perf = &i915->perf; |
4328 | |
4329 | /* XXX const struct i915_perf_ops! */ |
4330 | |
4331 | if (IS_HASWELL(i915)IS_PLATFORM(i915, INTEL_HASWELL)) { |
4332 | perf->ops.is_valid_b_counter_reg = gen7_is_valid_b_counter_addr; |
4333 | perf->ops.is_valid_mux_reg = hsw_is_valid_mux_addr; |
4334 | perf->ops.is_valid_flex_reg = NULL((void *)0); |
4335 | perf->ops.enable_metric_set = hsw_enable_metric_set; |
4336 | perf->ops.disable_metric_set = hsw_disable_metric_set; |
4337 | perf->ops.oa_enable = gen7_oa_enable; |
4338 | perf->ops.oa_disable = gen7_oa_disable; |
4339 | perf->ops.read = gen7_oa_read; |
4340 | perf->ops.oa_hw_tail_read = gen7_oa_hw_tail_read; |
4341 | |
4342 | perf->oa_formats = hsw_oa_formats; |
4343 | } else if (HAS_LOGICAL_RING_CONTEXTS(i915)((&(i915)->__info)->has_logical_ring_contexts)) { |
4344 | /* Note: that although we could theoretically also support the |
4345 | * legacy ringbuffer mode on BDW (and earlier iterations of |
4346 | * this driver, before upstreaming did this) it didn't seem |
4347 | * worth the complexity to maintain now that BDW+ enable |
4348 | * execlist mode by default. |
4349 | */ |
4350 | perf->ops.read = gen8_oa_read; |
4351 | |
4352 | if (IS_GEN_RANGE(i915, 8, 9)(!!((&(i915)->__info)->gen_mask & ( 0 + 0 + ((( ~0UL) >> (64 - (((9)) - 1) - 1)) & ((~0UL) << (((8)) - 1))))))) { |
4353 | perf->oa_formats = gen8_plus_oa_formats; |
4354 | |
4355 | perf->ops.is_valid_b_counter_reg = |
4356 | gen7_is_valid_b_counter_addr; |
4357 | perf->ops.is_valid_mux_reg = |
4358 | gen8_is_valid_mux_addr; |
4359 | perf->ops.is_valid_flex_reg = |
4360 | gen8_is_valid_flex_addr; |
4361 | |
4362 | if (IS_CHERRYVIEW(i915)IS_PLATFORM(i915, INTEL_CHERRYVIEW)) { |
4363 | perf->ops.is_valid_mux_reg = |
4364 | chv_is_valid_mux_addr; |
4365 | } |
4366 | |
4367 | perf->ops.oa_enable = gen8_oa_enable; |
4368 | perf->ops.oa_disable = gen8_oa_disable; |
4369 | perf->ops.enable_metric_set = gen8_enable_metric_set; |
4370 | perf->ops.disable_metric_set = gen8_disable_metric_set; |
4371 | perf->ops.oa_hw_tail_read = gen8_oa_hw_tail_read; |
4372 | |
4373 | if (IS_GEN(i915, 8)(0 + (&(i915)->__info)->gen == (8))) { |
4374 | perf->ctx_oactxctrl_offset = 0x120; |
4375 | perf->ctx_flexeu0_offset = 0x2ce; |
4376 | |
4377 | perf->gen8_valid_ctx_bit = BIT(25)(1UL << (25)); |
4378 | } else { |
4379 | perf->ctx_oactxctrl_offset = 0x128; |
4380 | perf->ctx_flexeu0_offset = 0x3de; |
4381 | |
4382 | perf->gen8_valid_ctx_bit = BIT(16)(1UL << (16)); |
4383 | } |
4384 | } else if (IS_GEN_RANGE(i915, 10, 11)(!!((&(i915)->__info)->gen_mask & ( 0 + 0 + ((( ~0UL) >> (64 - (((11)) - 1) - 1)) & ((~0UL) << (((10)) - 1))))))) { |
4385 | perf->oa_formats = gen8_plus_oa_formats; |
4386 | |
4387 | perf->ops.is_valid_b_counter_reg = |
4388 | gen7_is_valid_b_counter_addr; |
4389 | perf->ops.is_valid_mux_reg = |
4390 | gen10_is_valid_mux_addr; |
4391 | perf->ops.is_valid_flex_reg = |
4392 | gen8_is_valid_flex_addr; |
4393 | |
4394 | perf->ops.oa_enable = gen8_oa_enable; |
4395 | perf->ops.oa_disable = gen8_oa_disable; |
4396 | perf->ops.enable_metric_set = gen8_enable_metric_set; |
4397 | perf->ops.disable_metric_set = gen10_disable_metric_set; |
4398 | perf->ops.oa_hw_tail_read = gen8_oa_hw_tail_read; |
4399 | |
4400 | if (IS_GEN(i915, 10)(0 + (&(i915)->__info)->gen == (10))) { |
4401 | perf->ctx_oactxctrl_offset = 0x128; |
4402 | perf->ctx_flexeu0_offset = 0x3de; |
4403 | } else { |
4404 | perf->ctx_oactxctrl_offset = 0x124; |
4405 | perf->ctx_flexeu0_offset = 0x78e; |
4406 | } |
4407 | perf->gen8_valid_ctx_bit = BIT(16)(1UL << (16)); |
4408 | } else if (IS_GEN(i915, 12)(0 + (&(i915)->__info)->gen == (12))) { |
4409 | perf->oa_formats = gen12_oa_formats; |
4410 | |
4411 | perf->ops.is_valid_b_counter_reg = |
4412 | gen12_is_valid_b_counter_addr; |
4413 | perf->ops.is_valid_mux_reg = |
4414 | gen12_is_valid_mux_addr; |
4415 | perf->ops.is_valid_flex_reg = |
4416 | gen8_is_valid_flex_addr; |
4417 | |
4418 | perf->ops.oa_enable = gen12_oa_enable; |
4419 | perf->ops.oa_disable = gen12_oa_disable; |
4420 | perf->ops.enable_metric_set = gen12_enable_metric_set; |
4421 | perf->ops.disable_metric_set = gen12_disable_metric_set; |
4422 | perf->ops.oa_hw_tail_read = gen12_oa_hw_tail_read; |
4423 | |
4424 | perf->ctx_flexeu0_offset = 0; |
4425 | perf->ctx_oactxctrl_offset = 0x144; |
4426 | } |
4427 | } |
4428 | |
4429 | if (perf->ops.enable_metric_set) { |
4430 | rw_init(&perf->lock, "perflk")_rw_init_flags(&perf->lock, "perflk", 0, ((void *)0)); |
4431 | |
4432 | oa_sample_rate_hard_limit = |
4433 | RUNTIME_INFO(i915)(&(i915)->__runtime)->cs_timestamp_frequency_hz / 2; |
4434 | |
4435 | rw_init(&perf->metrics_lock, "metricslk")_rw_init_flags(&perf->metrics_lock, "metricslk", 0, (( void *)0)); |
4436 | idr_init(&perf->metrics_idr); |
4437 | |
4438 | /* We set up some ratelimit state to potentially throttle any |
4439 | * _NOTES about spurious, invalid OA reports which we don't |
4440 | * forward to userspace. |
4441 | * |
4442 | * We print a _NOTE about any throttling when closing the |
4443 | * stream instead of waiting until driver _fini which no one |
4444 | * would ever see. |
4445 | * |
4446 | * Using the same limiting factors as printk_ratelimit() |
4447 | */ |
4448 | #ifdef notyet |
4449 | ratelimit_state_init(&perf->spurious_report_rs, 5 * HZ, 10); |
4450 | /* Since we use a DRM_NOTE for spurious reports it would be |
4451 | * inconsistent to let __ratelimit() automatically print a |
4452 | * warning for throttling. |
4453 | */ |
4454 | ratelimit_set_flags(&perf->spurious_report_rs, |
4455 | RATELIMIT_MSG_ON_RELEASE); |
4456 | #endif |
4457 | |
4458 | ratelimit_state_init(&perf->tail_pointer_race, |
4459 | 5 * HZ, 10); |
4460 | ratelimit_set_flags(&perf->tail_pointer_race, |
4461 | RATELIMIT_MSG_ON_RELEASE); |
4462 | |
4463 | atomic64_set(&perf->noa_programming_delay,({ typeof(*(&perf->noa_programming_delay)) __tmp = ((500 * 1000)); *(volatile typeof(*(&perf->noa_programming_delay )) *)&(*(&perf->noa_programming_delay)) = __tmp; __tmp ; }) |
4464 | 500 * 1000 /* 500us */)({ typeof(*(&perf->noa_programming_delay)) __tmp = ((500 * 1000)); *(volatile typeof(*(&perf->noa_programming_delay )) *)&(*(&perf->noa_programming_delay)) = __tmp; __tmp ; }); |
4465 | |
4466 | perf->i915 = i915; |
4467 | } |
4468 | } |
4469 | |
4470 | static int destroy_config(int id, void *p, void *data) |
4471 | { |
4472 | i915_oa_config_put(p); |
4473 | return 0; |
4474 | } |
4475 | |
4476 | void i915_perf_sysctl_register(void) |
4477 | { |
4478 | #ifdef notyet |
4479 | sysctl_header = register_sysctl_table(dev_root); |
4480 | #endif |
4481 | } |
4482 | |
4483 | void i915_perf_sysctl_unregister(void) |
4484 | { |
4485 | #ifdef notyet |
4486 | unregister_sysctl_table(sysctl_header); |
4487 | #endif |
4488 | } |
4489 | |
4490 | /** |
4491 | * i915_perf_fini - Counter part to i915_perf_init() |
4492 | * @i915: i915 device instance |
4493 | */ |
4494 | void i915_perf_fini(struct drm_i915_privateinteldrm_softc *i915) |
4495 | { |
4496 | struct i915_perf *perf = &i915->perf; |
4497 | |
4498 | if (!perf->i915) |
4499 | return; |
4500 | |
4501 | idr_for_each(&perf->metrics_idr, destroy_config, perf); |
4502 | idr_destroy(&perf->metrics_idr); |
4503 | |
4504 | memset(&perf->ops, 0, sizeof(perf->ops))__builtin_memset((&perf->ops), (0), (sizeof(perf->ops ))); |
4505 | perf->i915 = NULL((void *)0); |
4506 | } |
4507 | |
4508 | /** |
4509 | * i915_perf_ioctl_version - Version of the i915-perf subsystem |
4510 | * |
4511 | * This version number is used by userspace to detect available features. |
4512 | */ |
4513 | int i915_perf_ioctl_version(void) |
4514 | { |
4515 | /* |
4516 | * 1: Initial version |
4517 | * I915_PERF_IOCTL_ENABLE |
4518 | * I915_PERF_IOCTL_DISABLE |
4519 | * |
4520 | * 2: Added runtime modification of OA config. |
4521 | * I915_PERF_IOCTL_CONFIG |
4522 | * |
4523 | * 3: Add DRM_I915_PERF_PROP_HOLD_PREEMPTION parameter to hold |
4524 | * preemption on a particular context so that performance data is |
4525 | * accessible from a delta of MI_RPC reports without looking at the |
4526 | * OA buffer. |
4527 | * |
4528 | * 4: Add DRM_I915_PERF_PROP_ALLOWED_SSEU to limit what contexts can |
4529 | * be run for the duration of the performance recording based on |
4530 | * their SSEU configuration. |
4531 | * |
4532 | * 5: Add DRM_I915_PERF_PROP_POLL_OA_PERIOD parameter that controls the |
4533 | * interval for the hrtimer used to check for OA data. |
4534 | */ |
4535 | return 5; |
4536 | } |
4537 | |
4538 | #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)0 |
4539 | #include "selftests/i915_perf.c" |
4540 | #endif |