1/*
2 * Copyright © 2019 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
24#include <unistd.h>
25#include <poll.h>
26
27#include "common/intel_gem.h"
28
29#include "dev/intel_debug.h"
30#include "dev/intel_device_info.h"
31
32#include "perf/intel_perf.h"
33#include "perf/intel_perf_mdapi.h"
34#include "perf/intel_perf_private.h"
35#include "perf/intel_perf_query.h"
36#include "perf/intel_perf_regs.h"
37
38#include "drm-uapi/i915_drm.h"
39
40#include "util/compiler.h"
41#include "util/u_math.h"
42
43#define FILE_DEBUG_FLAG DEBUG_PERFMON
44
45#define MI_RPC_BO_SIZE                (4096)
46#define MI_FREQ_OFFSET_BYTES          (256)
47#define MI_PERF_COUNTERS_OFFSET_BYTES (260)
48
49#define ALIGN(x, y) (((x) + (y)-1) & ~((y)-1))
50
51/* Align to 64bytes, requirement for OA report write address. */
52#define TOTAL_QUERY_DATA_SIZE            \
53   ALIGN(256 /* OA report */ +           \
54         4  /* freq register */ +        \
55         8 + 8 /* perf counter 1 & 2 */, \
56         64)
57
58
59static uint32_t field_offset(bool end, uint32_t offset)
60{
61   return (end ? TOTAL_QUERY_DATA_SIZE : 0) + offset;
62}
63
64#define MAP_READ  (1 << 0)
65#define MAP_WRITE (1 << 1)
66
67/**
68 * Periodic OA samples are read() into these buffer structures via the
69 * i915 perf kernel interface and appended to the
70 * perf_ctx->sample_buffers linked list. When we process the
71 * results of an OA metrics query we need to consider all the periodic
72 * samples between the Begin and End MI_REPORT_PERF_COUNT command
73 * markers.
74 *
75 * 'Periodic' is a simplification as there are other automatic reports
76 * written by the hardware also buffered here.
77 *
78 * Considering three queries, A, B and C:
79 *
80 *  Time ---->
81 *                ________________A_________________
82 *                |                                |
83 *                | ________B_________ _____C___________
84 *                | |                | |           |   |
85 *
86 * And an illustration of sample buffers read over this time frame:
87 * [HEAD ][     ][     ][     ][     ][     ][     ][     ][TAIL ]
88 *
89 * These nodes may hold samples for query A:
90 * [     ][     ][  A  ][  A  ][  A  ][  A  ][  A  ][     ][     ]
91 *
92 * These nodes may hold samples for query B:
93 * [     ][     ][  B  ][  B  ][  B  ][     ][     ][     ][     ]
94 *
95 * These nodes may hold samples for query C:
96 * [     ][     ][     ][     ][     ][  C  ][  C  ][  C  ][     ]
97 *
98 * The illustration assumes we have an even distribution of periodic
99 * samples so all nodes have the same size plotted against time:
100 *
101 * Note, to simplify code, the list is never empty.
102 *
103 * With overlapping queries we can see that periodic OA reports may
104 * relate to multiple queries and care needs to be take to keep
105 * track of sample buffers until there are no queries that might
106 * depend on their contents.
107 *
108 * We use a node ref counting system where a reference ensures that a
109 * node and all following nodes can't be freed/recycled until the
110 * reference drops to zero.
111 *
112 * E.g. with a ref of one here:
113 * [  0  ][  0  ][  1  ][  0  ][  0  ][  0  ][  0  ][  0  ][  0  ]
114 *
115 * These nodes could be freed or recycled ("reaped"):
116 * [  0  ][  0  ]
117 *
118 * These must be preserved until the leading ref drops to zero:
119 *               [  1  ][  0  ][  0  ][  0  ][  0  ][  0  ][  0  ]
120 *
121 * When a query starts we take a reference on the current tail of
122 * the list, knowing that no already-buffered samples can possibly
123 * relate to the newly-started query. A pointer to this node is
124 * also saved in the query object's ->oa.samples_head.
125 *
126 * E.g. starting query A while there are two nodes in .sample_buffers:
127 *                ________________A________
128 *                |
129 *
130 * [  0  ][  1  ]
131 *           ^_______ Add a reference and store pointer to node in
132 *                    A->oa.samples_head
133 *
134 * Moving forward to when the B query starts with no new buffer nodes:
135 * (for reference, i915 perf reads() are only done when queries finish)
136 *                ________________A_______
137 *                | ________B___
138 *                | |
139 *
140 * [  0  ][  2  ]
141 *           ^_______ Add a reference and store pointer to
142 *                    node in B->oa.samples_head
143 *
144 * Once a query is finished, after an OA query has become 'Ready',
145 * once the End OA report has landed and after we we have processed
146 * all the intermediate periodic samples then we drop the
147 * ->oa.samples_head reference we took at the start.
148 *
149 * So when the B query has finished we have:
150 *                ________________A________
151 *                | ______B___________
152 *                | |                |
153 * [  0  ][  1  ][  0  ][  0  ][  0  ]
154 *           ^_______ Drop B->oa.samples_head reference
155 *
156 * We still can't free these due to the A->oa.samples_head ref:
157 *        [  1  ][  0  ][  0  ][  0  ]
158 *
159 * When the A query finishes: (note there's a new ref for C's samples_head)
160 *                ________________A_________________
161 *                |                                |
162 *                |                    _____C_________
163 *                |                    |           |
164 * [  0  ][  0  ][  0  ][  0  ][  1  ][  0  ][  0  ]
165 *           ^_______ Drop A->oa.samples_head reference
166 *
167 * And we can now reap these nodes up to the C->oa.samples_head:
168 * [  X  ][  X  ][  X  ][  X  ]
169 *                  keeping -> [  1  ][  0  ][  0  ]
170 *
171 * We reap old sample buffers each time we finish processing an OA
172 * query by iterating the sample_buffers list from the head until we
173 * find a referenced node and stop.
174 *
175 * Reaped buffers move to a perfquery.free_sample_buffers list and
176 * when we come to read() we first look to recycle a buffer from the
177 * free_sample_buffers list before allocating a new buffer.
178 */
179struct oa_sample_buf {
180   struct exec_node link;
181   int refcount;
182   int len;
183   uint8_t buf[I915_PERF_OA_SAMPLE_SIZE * 10];
184   uint32_t last_timestamp;
185};
186
187/**
188 * gen representation of a performance query object.
189 *
190 * NB: We want to keep this structure relatively lean considering that
191 * applications may expect to allocate enough objects to be able to
192 * query around all draw calls in a frame.
193 */
194struct intel_perf_query_object
195{
196   const struct intel_perf_query_info *queryinfo;
197
198   /* See query->kind to know which state below is in use... */
199   union {
200      struct {
201
202         /**
203          * BO containing OA counter snapshots at query Begin/End time.
204          */
205         void *bo;
206
207         /**
208          * Address of mapped of @bo
209          */
210         void *map;
211
212         /**
213          * The MI_REPORT_PERF_COUNT command lets us specify a unique
214          * ID that will be reflected in the resulting OA report
215          * that's written by the GPU. This is the ID we're expecting
216          * in the begin report and the the end report should be
217          * @begin_report_id + 1.
218          */
219         int begin_report_id;
220
221         /**
222          * Reference the head of the brw->perfquery.sample_buffers
223          * list at the time that the query started (so we only need
224          * to look at nodes after this point when looking for samples
225          * related to this query)
226          *
227          * (See struct brw_oa_sample_buf description for more details)
228          */
229         struct exec_node *samples_head;
230
231         /**
232          * false while in the unaccumulated_elements list, and set to
233          * true when the final, end MI_RPC snapshot has been
234          * accumulated.
235          */
236         bool results_accumulated;
237
238         /**
239          * Accumulated OA results between begin and end of the query.
240          */
241         struct intel_perf_query_result result;
242      } oa;
243
244      struct {
245         /**
246          * BO containing starting and ending snapshots for the
247          * statistics counters.
248          */
249         void *bo;
250      } pipeline_stats;
251   };
252};
253
254struct intel_perf_context {
255   struct intel_perf_config *perf;
256
257   void * mem_ctx; /* ralloc context */
258   void * ctx;  /* driver context (eg, brw_context) */
259   void * bufmgr;
260   const struct intel_device_info *devinfo;
261
262   uint32_t hw_ctx;
263   int drm_fd;
264
265   /* The i915 perf stream we open to setup + enable the OA counters */
266   int oa_stream_fd;
267
268   /* An i915 perf stream fd gives exclusive access to the OA unit that will
269    * report counter snapshots for a specific counter set/profile in a
270    * specific layout/format so we can only start OA queries that are
271    * compatible with the currently open fd...
272    */
273   int current_oa_metrics_set_id;
274   int current_oa_format;
275
276   /* List of buffers containing OA reports */
277   struct exec_list sample_buffers;
278
279   /* Cached list of empty sample buffers */
280   struct exec_list free_sample_buffers;
281
282   int n_active_oa_queries;
283   int n_active_pipeline_stats_queries;
284
285   /* The number of queries depending on running OA counters which
286    * extends beyond brw_end_perf_query() since we need to wait until
287    * the last MI_RPC command has parsed by the GPU.
288    *
289    * Accurate accounting is important here as emitting an
290    * MI_REPORT_PERF_COUNT command while the OA unit is disabled will
291    * effectively hang the gpu.
292    */
293   int n_oa_users;
294
295   /* To help catch an spurious problem with the hardware or perf
296    * forwarding samples, we emit each MI_REPORT_PERF_COUNT command
297    * with a unique ID that we can explicitly check for...
298    */
299   int next_query_start_report_id;
300
301   /**
302    * An array of queries whose results haven't yet been assembled
303    * based on the data in buffer objects.
304    *
305    * These may be active, or have already ended.  However, the
306    * results have not been requested.
307    */
308   struct intel_perf_query_object **unaccumulated;
309   int unaccumulated_elements;
310   int unaccumulated_array_size;
311
312   /* The total number of query objects so we can relinquish
313    * our exclusive access to perf if the application deletes
314    * all of its objects. (NB: We only disable perf while
315    * there are no active queries)
316    */
317   int n_query_instances;
318
319   int period_exponent;
320};
321
322static bool
323inc_n_users(struct intel_perf_context *perf_ctx)
324{
325   if (perf_ctx->n_oa_users == 0 &&
326       intel_ioctl(perf_ctx->oa_stream_fd, I915_PERF_IOCTL_ENABLE, 0) < 0)
327   {
328      return false;
329   }
330   ++perf_ctx->n_oa_users;
331
332   return true;
333}
334
335static void
336dec_n_users(struct intel_perf_context *perf_ctx)
337{
338   /* Disabling the i915 perf stream will effectively disable the OA
339    * counters.  Note it's important to be sure there are no outstanding
340    * MI_RPC commands at this point since they could stall the CS
341    * indefinitely once OACONTROL is disabled.
342    */
343   --perf_ctx->n_oa_users;
344   if (perf_ctx->n_oa_users == 0 &&
345       intel_ioctl(perf_ctx->oa_stream_fd, I915_PERF_IOCTL_DISABLE, 0) < 0)
346   {
347      DBG("WARNING: Error disabling gen perf stream: %m\n");
348   }
349}
350
351void
352intel_perf_close(struct intel_perf_context *perfquery,
353                 const struct intel_perf_query_info *query)
354{
355   if (perfquery->oa_stream_fd != -1) {
356      close(perfquery->oa_stream_fd);
357      perfquery->oa_stream_fd = -1;
358   }
359   if (query && query->kind == INTEL_PERF_QUERY_TYPE_RAW) {
360      struct intel_perf_query_info *raw_query =
361         (struct intel_perf_query_info *) query;
362      raw_query->oa_metrics_set_id = 0;
363   }
364}
365
366bool
367intel_perf_open(struct intel_perf_context *perf_ctx,
368                int metrics_set_id,
369                int report_format,
370                int period_exponent,
371                int drm_fd,
372                uint32_t ctx_id,
373                bool enable)
374{
375   uint64_t properties[DRM_I915_PERF_PROP_MAX * 2];
376   uint32_t p = 0;
377
378   /* Single context sampling if valid context id. */
379   if (ctx_id != INTEL_PERF_INVALID_CTX_ID) {
380      properties[p++] = DRM_I915_PERF_PROP_CTX_HANDLE;
381      properties[p++] = ctx_id;
382   }
383
384   /* Include OA reports in samples */
385   properties[p++] = DRM_I915_PERF_PROP_SAMPLE_OA;
386   properties[p++] = true;
387
388   /* OA unit configuration */
389   properties[p++] = DRM_I915_PERF_PROP_OA_METRICS_SET;
390   properties[p++] = metrics_set_id;
391
392   properties[p++] = DRM_I915_PERF_PROP_OA_FORMAT;
393   properties[p++] = report_format;
394
395   properties[p++] = DRM_I915_PERF_PROP_OA_EXPONENT;
396   properties[p++] = period_exponent;
397
398   /* SSEU configuration */
399   if (intel_perf_has_global_sseu(perf_ctx->perf)) {
400      properties[p++] = DRM_I915_PERF_PROP_GLOBAL_SSEU;
401      properties[p++] = to_user_pointer(&perf_ctx->perf->sseu);
402   }
403
404   assert(p <= ARRAY_SIZE(properties));
405
406   struct drm_i915_perf_open_param param = {
407      .flags = I915_PERF_FLAG_FD_CLOEXEC |
408               I915_PERF_FLAG_FD_NONBLOCK |
409               (enable ? 0 : I915_PERF_FLAG_DISABLED),
410      .num_properties = p / 2,
411      .properties_ptr = (uintptr_t) properties,
412   };
413   int fd = intel_ioctl(drm_fd, DRM_IOCTL_I915_PERF_OPEN, &param);
414   if (fd == -1) {
415      DBG("Error opening gen perf OA stream: %m\n");
416      return false;
417   }
418
419   perf_ctx->oa_stream_fd = fd;
420
421   perf_ctx->current_oa_metrics_set_id = metrics_set_id;
422   perf_ctx->current_oa_format = report_format;
423
424   if (enable)
425      ++perf_ctx->n_oa_users;
426
427   return true;
428}
429
430static uint64_t
431get_metric_id(struct intel_perf_config *perf,
432              const struct intel_perf_query_info *query)
433{
434   /* These queries are know not to ever change, their config ID has been
435    * loaded upon the first query creation. No need to look them up again.
436    */
437   if (query->kind == INTEL_PERF_QUERY_TYPE_OA)
438      return query->oa_metrics_set_id;
439
440   assert(query->kind == INTEL_PERF_QUERY_TYPE_RAW);
441
442   /* Raw queries can be reprogrammed up by an external application/library.
443    * When a raw query is used for the first time it's id is set to a value !=
444    * 0. When it stops being used the id returns to 0. No need to reload the
445    * ID when it's already loaded.
446    */
447   if (query->oa_metrics_set_id != 0) {
448      DBG("Raw query '%s' guid=%s using cached ID: %"PRIu64"\n",
449          query->name, query->guid, query->oa_metrics_set_id);
450      return query->oa_metrics_set_id;
451   }
452
453   struct intel_perf_query_info *raw_query = (struct intel_perf_query_info *)query;
454   if (!intel_perf_load_metric_id(perf, query->guid,
455                                &raw_query->oa_metrics_set_id)) {
456      DBG("Unable to read query guid=%s ID, falling back to test config\n", query->guid);
457      raw_query->oa_metrics_set_id = perf->fallback_raw_oa_metric;
458   } else {
459      DBG("Raw query '%s'guid=%s loaded ID: %"PRIu64"\n",
460          query->name, query->guid, query->oa_metrics_set_id);
461   }
462   return query->oa_metrics_set_id;
463}
464
465static struct oa_sample_buf *
466get_free_sample_buf(struct intel_perf_context *perf_ctx)
467{
468   struct exec_node *node = exec_list_pop_head(&perf_ctx->free_sample_buffers);
469   struct oa_sample_buf *buf;
470
471   if (node)
472      buf = exec_node_data(struct oa_sample_buf, node, link);
473   else {
474      buf = ralloc_size(perf_ctx->perf, sizeof(*buf));
475
476      exec_node_init(&buf->link);
477      buf->refcount = 0;
478   }
479   buf->len = 0;
480
481   return buf;
482}
483
484static void
485reap_old_sample_buffers(struct intel_perf_context *perf_ctx)
486{
487   struct exec_node *tail_node =
488      exec_list_get_tail(&perf_ctx->sample_buffers);
489   struct oa_sample_buf *tail_buf =
490      exec_node_data(struct oa_sample_buf, tail_node, link);
491
492   /* Remove all old, unreferenced sample buffers walking forward from
493    * the head of the list, except always leave at least one node in
494    * the list so we always have a node to reference when we Begin
495    * a new query.
496    */
497   foreach_list_typed_safe(struct oa_sample_buf, buf, link,
498                           &perf_ctx->sample_buffers)
499   {
500      if (buf->refcount == 0 && buf != tail_buf) {
501         exec_node_remove(&buf->link);
502         exec_list_push_head(&perf_ctx->free_sample_buffers, &buf->link);
503      } else
504         return;
505   }
506}
507
508static void
509free_sample_bufs(struct intel_perf_context *perf_ctx)
510{
511   foreach_list_typed_safe(struct oa_sample_buf, buf, link,
512                           &perf_ctx->free_sample_buffers)
513      ralloc_free(buf);
514
515   exec_list_make_empty(&perf_ctx->free_sample_buffers);
516}
517
518
519struct intel_perf_query_object *
520intel_perf_new_query(struct intel_perf_context *perf_ctx, unsigned query_index)
521{
522   const struct intel_perf_query_info *query =
523      &perf_ctx->perf->queries[query_index];
524
525   switch (query->kind) {
526   case INTEL_PERF_QUERY_TYPE_OA:
527   case INTEL_PERF_QUERY_TYPE_RAW:
528      if (perf_ctx->period_exponent == 0)
529         return NULL;
530      break;
531   case INTEL_PERF_QUERY_TYPE_PIPELINE:
532      break;
533   }
534
535   struct intel_perf_query_object *obj =
536      calloc(1, sizeof(struct intel_perf_query_object));
537
538   if (!obj)
539      return NULL;
540
541   obj->queryinfo = query;
542
543   perf_ctx->n_query_instances++;
544   return obj;
545}
546
547int
548intel_perf_active_queries(struct intel_perf_context *perf_ctx,
549                          const struct intel_perf_query_info *query)
550{
551   assert(perf_ctx->n_active_oa_queries == 0 || perf_ctx->n_active_pipeline_stats_queries == 0);
552
553   switch (query->kind) {
554   case INTEL_PERF_QUERY_TYPE_OA:
555   case INTEL_PERF_QUERY_TYPE_RAW:
556      return perf_ctx->n_active_oa_queries;
557      break;
558
559   case INTEL_PERF_QUERY_TYPE_PIPELINE:
560      return perf_ctx->n_active_pipeline_stats_queries;
561      break;
562
563   default:
564      unreachable("Unknown query type");
565      break;
566   }
567}
568
569const struct intel_perf_query_info*
570intel_perf_query_info(const struct intel_perf_query_object *query)
571{
572   return query->queryinfo;
573}
574
575struct intel_perf_context *
576intel_perf_new_context(void *parent)
577{
578   struct intel_perf_context *ctx = rzalloc(parent, struct intel_perf_context);
579   if (! ctx)
580      fprintf(stderr, "%s: failed to alloc context\n", __func__);
581   return ctx;
582}
583
584struct intel_perf_config *
585intel_perf_config(struct intel_perf_context *ctx)
586{
587   return ctx->perf;
588}
589
590void
591intel_perf_init_context(struct intel_perf_context *perf_ctx,
592                        struct intel_perf_config *perf_cfg,
593                        void * mem_ctx, /* ralloc context */
594                        void * ctx,  /* driver context (eg, brw_context) */
595                        void * bufmgr,  /* eg brw_bufmgr */
596                        const struct intel_device_info *devinfo,
597                        uint32_t hw_ctx,
598                        int drm_fd)
599{
600   perf_ctx->perf = perf_cfg;
601   perf_ctx->mem_ctx = mem_ctx;
602   perf_ctx->ctx = ctx;
603   perf_ctx->bufmgr = bufmgr;
604   perf_ctx->drm_fd = drm_fd;
605   perf_ctx->hw_ctx = hw_ctx;
606   perf_ctx->devinfo = devinfo;
607
608   perf_ctx->unaccumulated =
609      ralloc_array(mem_ctx, struct intel_perf_query_object *, 2);
610   perf_ctx->unaccumulated_elements = 0;
611   perf_ctx->unaccumulated_array_size = 2;
612
613   exec_list_make_empty(&perf_ctx->sample_buffers);
614   exec_list_make_empty(&perf_ctx->free_sample_buffers);
615
616   /* It's convenient to guarantee that this linked list of sample
617    * buffers is never empty so we add an empty head so when we
618    * Begin an OA query we can always take a reference on a buffer
619    * in this list.
620    */
621   struct oa_sample_buf *buf = get_free_sample_buf(perf_ctx);
622   exec_list_push_head(&perf_ctx->sample_buffers, &buf->link);
623
624   perf_ctx->oa_stream_fd = -1;
625   perf_ctx->next_query_start_report_id = 1000;
626
627   /* The period_exponent gives a sampling period as follows:
628    *   sample_period = timestamp_period * 2^(period_exponent + 1)
629    *
630    * The timestamps increments every 80ns (HSW), ~52ns (GFX9LP) or
631    * ~83ns (GFX8/9).
632    *
633    * The counter overflow period is derived from the EuActive counter
634    * which reads a counter that increments by the number of clock
635    * cycles multiplied by the number of EUs. It can be calculated as:
636    *
637    * 2^(number of bits in A counter) / (n_eus * max_intel_freq * 2)
638    *
639    * (E.g. 40 EUs @ 1GHz = ~53ms)
640    *
641    * We select a sampling period inferior to that overflow period to
642    * ensure we cannot see more than 1 counter overflow, otherwise we
643    * could loose information.
644    */
645
646   int a_counter_in_bits = 32;
647   if (devinfo->ver >= 8)
648      a_counter_in_bits = 40;
649
650   uint64_t overflow_period = pow(2, a_counter_in_bits) / (perf_cfg->sys_vars.n_eus *
651       /* drop 1GHz freq to have units in nanoseconds */
652       2);
653
654   DBG("A counter overflow period: %"PRIu64"ns, %"PRIu64"ms (n_eus=%"PRIu64")\n",
655       overflow_period, overflow_period / 1000000ul, perf_cfg->sys_vars.n_eus);
656
657   int period_exponent = 0;
658   uint64_t prev_sample_period, next_sample_period;
659   for (int e = 0; e < 30; e++) {
660      prev_sample_period = 1000000000ull * pow(2, e + 1) / devinfo->timestamp_frequency;
661      next_sample_period = 1000000000ull * pow(2, e + 2) / devinfo->timestamp_frequency;
662
663      /* Take the previous sampling period, lower than the overflow
664       * period.
665       */
666      if (prev_sample_period < overflow_period &&
667          next_sample_period > overflow_period)
668         period_exponent = e + 1;
669   }
670
671   perf_ctx->period_exponent = period_exponent;
672
673   if (period_exponent == 0) {
674      DBG("WARNING: enable to find a sampling exponent\n");
675   } else {
676      DBG("OA sampling exponent: %i ~= %"PRIu64"ms\n", period_exponent,
677            prev_sample_period / 1000000ul);
678   }
679}
680
681/**
682 * Add a query to the global list of "unaccumulated queries."
683 *
684 * Queries are tracked here until all the associated OA reports have
685 * been accumulated via accumulate_oa_reports() after the end
686 * MI_REPORT_PERF_COUNT has landed in query->oa.bo.
687 */
688static void
689add_to_unaccumulated_query_list(struct intel_perf_context *perf_ctx,
690                                struct intel_perf_query_object *obj)
691{
692   if (perf_ctx->unaccumulated_elements >=
693       perf_ctx->unaccumulated_array_size)
694   {
695      perf_ctx->unaccumulated_array_size *= 1.5;
696      perf_ctx->unaccumulated =
697         reralloc(perf_ctx->mem_ctx, perf_ctx->unaccumulated,
698                  struct intel_perf_query_object *,
699                  perf_ctx->unaccumulated_array_size);
700   }
701
702   perf_ctx->unaccumulated[perf_ctx->unaccumulated_elements++] = obj;
703}
704
705/**
706 * Emit MI_STORE_REGISTER_MEM commands to capture all of the
707 * pipeline statistics for the performance query object.
708 */
709static void
710snapshot_statistics_registers(struct intel_perf_context *ctx,
711                              struct intel_perf_query_object *obj,
712                              uint32_t offset_in_bytes)
713{
714   struct intel_perf_config *perf = ctx->perf;
715   const struct intel_perf_query_info *query = obj->queryinfo;
716   const int n_counters = query->n_counters;
717
718   for (int i = 0; i < n_counters; i++) {
719      const struct intel_perf_query_counter *counter = &query->counters[i];
720
721      assert(counter->data_type == INTEL_PERF_COUNTER_DATA_TYPE_UINT64);
722
723      perf->vtbl.store_register_mem(ctx->ctx, obj->pipeline_stats.bo,
724                                    counter->pipeline_stat.reg, 8,
725                                    offset_in_bytes + counter->offset);
726   }
727}
728
729static void
730snapshot_query_layout(struct intel_perf_context *perf_ctx,
731                      struct intel_perf_query_object *query,
732                      bool end_snapshot)
733{
734   struct intel_perf_config *perf_cfg = perf_ctx->perf;
735   const struct intel_perf_query_field_layout *layout = &perf_cfg->query_layout;
736   uint32_t offset = end_snapshot ? align(layout->size, layout->alignment) : 0;
737
738   for (uint32_t f = 0; f < layout->n_fields; f++) {
739      const struct intel_perf_query_field *field =
740         &layout->fields[end_snapshot ? f : (layout->n_fields - 1 - f)];
741
742      switch (field->type) {
743      case INTEL_PERF_QUERY_FIELD_TYPE_MI_RPC:
744         perf_cfg->vtbl.emit_mi_report_perf_count(perf_ctx->ctx, query->oa.bo,
745                                                  offset + field->location,
746                                                  query->oa.begin_report_id +
747                                                  (end_snapshot ? 1 : 0));
748         break;
749      case INTEL_PERF_QUERY_FIELD_TYPE_SRM_PERFCNT:
750      case INTEL_PERF_QUERY_FIELD_TYPE_SRM_RPSTAT:
751      case INTEL_PERF_QUERY_FIELD_TYPE_SRM_OA_B:
752      case INTEL_PERF_QUERY_FIELD_TYPE_SRM_OA_C:
753         perf_cfg->vtbl.store_register_mem(perf_ctx->ctx, query->oa.bo,
754                                           field->mmio_offset, field->size,
755                                           offset + field->location);
756         break;
757      default:
758         unreachable("Invalid field type");
759      }
760   }
761}
762
763bool
764intel_perf_begin_query(struct intel_perf_context *perf_ctx,
765                       struct intel_perf_query_object *query)
766{
767   struct intel_perf_config *perf_cfg = perf_ctx->perf;
768   const struct intel_perf_query_info *queryinfo = query->queryinfo;
769
770   /* XXX: We have to consider that the command parser unit that parses batch
771    * buffer commands and is used to capture begin/end counter snapshots isn't
772    * implicitly synchronized with what's currently running across other GPU
773    * units (such as the EUs running shaders) that the performance counters are
774    * associated with.
775    *
776    * The intention of performance queries is to measure the work associated
777    * with commands between the begin/end delimiters and so for that to be the
778    * case we need to explicitly synchronize the parsing of commands to capture
779    * Begin/End counter snapshots with what's running across other parts of the
780    * GPU.
781    *
782    * When the command parser reaches a Begin marker it effectively needs to
783    * drain everything currently running on the GPU until the hardware is idle
784    * before capturing the first snapshot of counters - otherwise the results
785    * would also be measuring the effects of earlier commands.
786    *
787    * When the command parser reaches an End marker it needs to stall until
788    * everything currently running on the GPU has finished before capturing the
789    * end snapshot - otherwise the results won't be a complete representation
790    * of the work.
791    *
792    * To achieve this, we stall the pipeline at pixel scoreboard (prevent any
793    * additional work to be processed by the pipeline until all pixels of the
794    * previous draw has be completed).
795    *
796    * N.B. The final results are based on deltas of counters between (inside)
797    * Begin/End markers so even though the total wall clock time of the
798    * workload is stretched by larger pipeline bubbles the bubbles themselves
799    * are generally invisible to the query results. Whether that's a good or a
800    * bad thing depends on the use case. For a lower real-time impact while
801    * capturing metrics then periodic sampling may be a better choice than
802    * INTEL_performance_query.
803    *
804    *
805    * This is our Begin synchronization point to drain current work on the
806    * GPU before we capture our first counter snapshot...
807    */
808   perf_cfg->vtbl.emit_stall_at_pixel_scoreboard(perf_ctx->ctx);
809
810   switch (queryinfo->kind) {
811   case INTEL_PERF_QUERY_TYPE_OA:
812   case INTEL_PERF_QUERY_TYPE_RAW: {
813
814      /* Opening an i915 perf stream implies exclusive access to the OA unit
815       * which will generate counter reports for a specific counter set with a
816       * specific layout/format so we can't begin any OA based queries that
817       * require a different counter set or format unless we get an opportunity
818       * to close the stream and open a new one...
819       */
820      uint64_t metric_id = get_metric_id(perf_ctx->perf, queryinfo);
821
822      if (perf_ctx->oa_stream_fd != -1 &&
823          perf_ctx->current_oa_metrics_set_id != metric_id) {
824
825         if (perf_ctx->n_oa_users != 0) {
826            DBG("WARNING: Begin failed already using perf config=%i/%"PRIu64"\n",
827                perf_ctx->current_oa_metrics_set_id, metric_id);
828            return false;
829         } else
830            intel_perf_close(perf_ctx, queryinfo);
831      }
832
833      /* If the OA counters aren't already on, enable them. */
834      if (perf_ctx->oa_stream_fd == -1) {
835         assert(perf_ctx->period_exponent != 0);
836
837         if (!intel_perf_open(perf_ctx, metric_id, queryinfo->oa_format,
838                            perf_ctx->period_exponent, perf_ctx->drm_fd,
839                            perf_ctx->hw_ctx, false))
840            return false;
841      } else {
842         assert(perf_ctx->current_oa_metrics_set_id == metric_id &&
843                perf_ctx->current_oa_format == queryinfo->oa_format);
844      }
845
846      if (!inc_n_users(perf_ctx)) {
847         DBG("WARNING: Error enabling i915 perf stream: %m\n");
848         return false;
849      }
850
851      if (query->oa.bo) {
852         perf_cfg->vtbl.bo_unreference(query->oa.bo);
853         query->oa.bo = NULL;
854      }
855
856      query->oa.bo = perf_cfg->vtbl.bo_alloc(perf_ctx->bufmgr,
857                                             "perf. query OA MI_RPC bo",
858                                             MI_RPC_BO_SIZE);
859#ifdef DEBUG
860      /* Pre-filling the BO helps debug whether writes landed. */
861      void *map = perf_cfg->vtbl.bo_map(perf_ctx->ctx, query->oa.bo, MAP_WRITE);
862      memset(map, 0x80, MI_RPC_BO_SIZE);
863      perf_cfg->vtbl.bo_unmap(query->oa.bo);
864#endif
865
866      query->oa.begin_report_id = perf_ctx->next_query_start_report_id;
867      perf_ctx->next_query_start_report_id += 2;
868
869      snapshot_query_layout(perf_ctx, query, false /* end_snapshot */);
870
871      ++perf_ctx->n_active_oa_queries;
872
873      /* No already-buffered samples can possibly be associated with this query
874       * so create a marker within the list of sample buffers enabling us to
875       * easily ignore earlier samples when processing this query after
876       * completion.
877       */
878      assert(!exec_list_is_empty(&perf_ctx->sample_buffers));
879      query->oa.samples_head = exec_list_get_tail(&perf_ctx->sample_buffers);
880
881      struct oa_sample_buf *buf =
882         exec_node_data(struct oa_sample_buf, query->oa.samples_head, link);
883
884      /* This reference will ensure that future/following sample
885       * buffers (that may relate to this query) can't be freed until
886       * this drops to zero.
887       */
888      buf->refcount++;
889
890      intel_perf_query_result_clear(&query->oa.result);
891      query->oa.results_accumulated = false;
892
893      add_to_unaccumulated_query_list(perf_ctx, query);
894      break;
895   }
896
897   case INTEL_PERF_QUERY_TYPE_PIPELINE:
898      if (query->pipeline_stats.bo) {
899         perf_cfg->vtbl.bo_unreference(query->pipeline_stats.bo);
900         query->pipeline_stats.bo = NULL;
901      }
902
903      query->pipeline_stats.bo =
904         perf_cfg->vtbl.bo_alloc(perf_ctx->bufmgr,
905                                 "perf. query pipeline stats bo",
906                                 STATS_BO_SIZE);
907
908      /* Take starting snapshots. */
909      snapshot_statistics_registers(perf_ctx, query, 0);
910
911      ++perf_ctx->n_active_pipeline_stats_queries;
912      break;
913
914   default:
915      unreachable("Unknown query type");
916      break;
917   }
918
919   return true;
920}
921
922void
923intel_perf_end_query(struct intel_perf_context *perf_ctx,
924                     struct intel_perf_query_object *query)
925{
926   struct intel_perf_config *perf_cfg = perf_ctx->perf;
927
928   /* Ensure that the work associated with the queried commands will have
929    * finished before taking our query end counter readings.
930    *
931    * For more details see comment in brw_begin_perf_query for
932    * corresponding flush.
933    */
934   perf_cfg->vtbl.emit_stall_at_pixel_scoreboard(perf_ctx->ctx);
935
936   switch (query->queryinfo->kind) {
937   case INTEL_PERF_QUERY_TYPE_OA:
938   case INTEL_PERF_QUERY_TYPE_RAW:
939
940      /* NB: It's possible that the query will have already been marked
941       * as 'accumulated' if an error was seen while reading samples
942       * from perf. In this case we mustn't try and emit a closing
943       * MI_RPC command in case the OA unit has already been disabled
944       */
945      if (!query->oa.results_accumulated)
946         snapshot_query_layout(perf_ctx, query, true /* end_snapshot */);
947
948      --perf_ctx->n_active_oa_queries;
949
950      /* NB: even though the query has now ended, it can't be accumulated
951       * until the end MI_REPORT_PERF_COUNT snapshot has been written
952       * to query->oa.bo
953       */
954      break;
955
956   case INTEL_PERF_QUERY_TYPE_PIPELINE:
957      snapshot_statistics_registers(perf_ctx, query,
958                                    STATS_BO_END_OFFSET_BYTES);
959      --perf_ctx->n_active_pipeline_stats_queries;
960      break;
961
962   default:
963      unreachable("Unknown query type");
964      break;
965   }
966}
967
968bool intel_perf_oa_stream_ready(struct intel_perf_context *perf_ctx)
969{
970   struct pollfd pfd;
971
972   pfd.fd = perf_ctx->oa_stream_fd;
973   pfd.events = POLLIN;
974   pfd.revents = 0;
975
976   if (poll(&pfd, 1, 0) < 0) {
977      DBG("Error polling OA stream\n");
978      return false;
979   }
980
981   if (!(pfd.revents & POLLIN))
982      return false;
983
984   return true;
985}
986
987ssize_t
988intel_perf_read_oa_stream(struct intel_perf_context *perf_ctx,
989                          void* buf,
990                          size_t nbytes)
991{
992   return read(perf_ctx->oa_stream_fd, buf, nbytes);
993}
994
995enum OaReadStatus {
996   OA_READ_STATUS_ERROR,
997   OA_READ_STATUS_UNFINISHED,
998   OA_READ_STATUS_FINISHED,
999};
1000
1001static enum OaReadStatus
1002read_oa_samples_until(struct intel_perf_context *perf_ctx,
1003                      uint32_t start_timestamp,
1004                      uint32_t end_timestamp)
1005{
1006   struct exec_node *tail_node =
1007      exec_list_get_tail(&perf_ctx->sample_buffers);
1008   struct oa_sample_buf *tail_buf =
1009      exec_node_data(struct oa_sample_buf, tail_node, link);
1010   uint32_t last_timestamp =
1011      tail_buf->len == 0 ? start_timestamp : tail_buf->last_timestamp;
1012
1013   while (1) {
1014      struct oa_sample_buf *buf = get_free_sample_buf(perf_ctx);
1015      uint32_t offset;
1016      int len;
1017
1018      while ((len = read(perf_ctx->oa_stream_fd, buf->buf,
1019                         sizeof(buf->buf))) < 0 && errno == EINTR)
1020         ;
1021
1022      if (len <= 0) {
1023         exec_list_push_tail(&perf_ctx->free_sample_buffers, &buf->link);
1024
1025         if (len == 0) {
1026            DBG("Spurious EOF reading i915 perf samples\n");
1027            return OA_READ_STATUS_ERROR;
1028         }
1029
1030         if (errno != EAGAIN) {
1031            DBG("Error reading i915 perf samples: %m\n");
1032            return OA_READ_STATUS_ERROR;
1033         }
1034
1035         if ((last_timestamp - start_timestamp) >= INT32_MAX)
1036            return OA_READ_STATUS_UNFINISHED;
1037
1038         if ((last_timestamp - start_timestamp) <
1039              (end_timestamp - start_timestamp))
1040            return OA_READ_STATUS_UNFINISHED;
1041
1042         return OA_READ_STATUS_FINISHED;
1043      }
1044
1045      buf->len = len;
1046      exec_list_push_tail(&perf_ctx->sample_buffers, &buf->link);
1047
1048      /* Go through the reports and update the last timestamp. */
1049      offset = 0;
1050      while (offset < buf->len) {
1051         const struct drm_i915_perf_record_header *header =
1052            (const struct drm_i915_perf_record_header *) &buf->buf[offset];
1053         uint32_t *report = (uint32_t *) (header + 1);
1054
1055         if (header->type == DRM_I915_PERF_RECORD_SAMPLE)
1056            last_timestamp = report[1];
1057
1058         offset += header->size;
1059      }
1060
1061      buf->last_timestamp = last_timestamp;
1062   }
1063
1064   unreachable("not reached");
1065   return OA_READ_STATUS_ERROR;
1066}
1067
1068/**
1069 * Try to read all the reports until either the delimiting timestamp
1070 * or an error arises.
1071 */
1072static bool
1073read_oa_samples_for_query(struct intel_perf_context *perf_ctx,
1074                          struct intel_perf_query_object *query,
1075                          void *current_batch)
1076{
1077   uint32_t *start;
1078   uint32_t *last;
1079   uint32_t *end;
1080   struct intel_perf_config *perf_cfg = perf_ctx->perf;
1081
1082   /* We need the MI_REPORT_PERF_COUNT to land before we can start
1083    * accumulate. */
1084   assert(!perf_cfg->vtbl.batch_references(current_batch, query->oa.bo) &&
1085          !perf_cfg->vtbl.bo_busy(query->oa.bo));
1086
1087   /* Map the BO once here and let accumulate_oa_reports() unmap
1088    * it. */
1089   if (query->oa.map == NULL)
1090      query->oa.map = perf_cfg->vtbl.bo_map(perf_ctx->ctx, query->oa.bo, MAP_READ);
1091
1092   start = last = query->oa.map + field_offset(false, 0);
1093   end = query->oa.map + field_offset(true, 0);
1094
1095   if (start[0] != query->oa.begin_report_id) {
1096      DBG("Spurious start report id=%"PRIu32"\n", start[0]);
1097      return true;
1098   }
1099   if (end[0] != (query->oa.begin_report_id + 1)) {
1100      DBG("Spurious end report id=%"PRIu32"\n", end[0]);
1101      return true;
1102   }
1103
1104   /* Read the reports until the end timestamp. */
1105   switch (read_oa_samples_until(perf_ctx, start[1], end[1])) {
1106   case OA_READ_STATUS_ERROR:
1107      FALLTHROUGH; /* Let accumulate_oa_reports() deal with the error. */
1108   case OA_READ_STATUS_FINISHED:
1109      return true;
1110   case OA_READ_STATUS_UNFINISHED:
1111      return false;
1112   }
1113
1114   unreachable("invalid read status");
1115   return false;
1116}
1117
1118void
1119intel_perf_wait_query(struct intel_perf_context *perf_ctx,
1120                      struct intel_perf_query_object *query,
1121                      void *current_batch)
1122{
1123   struct intel_perf_config *perf_cfg = perf_ctx->perf;
1124   struct brw_bo *bo = NULL;
1125
1126   switch (query->queryinfo->kind) {
1127   case INTEL_PERF_QUERY_TYPE_OA:
1128   case INTEL_PERF_QUERY_TYPE_RAW:
1129      bo = query->oa.bo;
1130      break;
1131
1132   case INTEL_PERF_QUERY_TYPE_PIPELINE:
1133      bo = query->pipeline_stats.bo;
1134      break;
1135
1136   default:
1137      unreachable("Unknown query type");
1138      break;
1139   }
1140
1141   if (bo == NULL)
1142      return;
1143
1144   /* If the current batch references our results bo then we need to
1145    * flush first...
1146    */
1147   if (perf_cfg->vtbl.batch_references(current_batch, bo))
1148      perf_cfg->vtbl.batchbuffer_flush(perf_ctx->ctx, __FILE__, __LINE__);
1149
1150   perf_cfg->vtbl.bo_wait_rendering(bo);
1151}
1152
1153bool
1154intel_perf_is_query_ready(struct intel_perf_context *perf_ctx,
1155                          struct intel_perf_query_object *query,
1156                          void *current_batch)
1157{
1158   struct intel_perf_config *perf_cfg = perf_ctx->perf;
1159
1160   switch (query->queryinfo->kind) {
1161   case INTEL_PERF_QUERY_TYPE_OA:
1162   case INTEL_PERF_QUERY_TYPE_RAW:
1163      return (query->oa.results_accumulated ||
1164              (query->oa.bo &&
1165               !perf_cfg->vtbl.batch_references(current_batch, query->oa.bo) &&
1166               !perf_cfg->vtbl.bo_busy(query->oa.bo)));
1167
1168   case INTEL_PERF_QUERY_TYPE_PIPELINE:
1169      return (query->pipeline_stats.bo &&
1170              !perf_cfg->vtbl.batch_references(current_batch, query->pipeline_stats.bo) &&
1171              !perf_cfg->vtbl.bo_busy(query->pipeline_stats.bo));
1172
1173   default:
1174      unreachable("Unknown query type");
1175      break;
1176   }
1177
1178   return false;
1179}
1180
1181/**
1182 * Remove a query from the global list of unaccumulated queries once
1183 * after successfully accumulating the OA reports associated with the
1184 * query in accumulate_oa_reports() or when discarding unwanted query
1185 * results.
1186 */
1187static void
1188drop_from_unaccumulated_query_list(struct intel_perf_context *perf_ctx,
1189                                   struct intel_perf_query_object *query)
1190{
1191   for (int i = 0; i < perf_ctx->unaccumulated_elements; i++) {
1192      if (perf_ctx->unaccumulated[i] == query) {
1193         int last_elt = --perf_ctx->unaccumulated_elements;
1194
1195         if (i == last_elt)
1196            perf_ctx->unaccumulated[i] = NULL;
1197         else {
1198            perf_ctx->unaccumulated[i] =
1199               perf_ctx->unaccumulated[last_elt];
1200         }
1201
1202         break;
1203      }
1204   }
1205
1206   /* Drop our samples_head reference so that associated periodic
1207    * sample data buffers can potentially be reaped if they aren't
1208    * referenced by any other queries...
1209    */
1210
1211   struct oa_sample_buf *buf =
1212      exec_node_data(struct oa_sample_buf, query->oa.samples_head, link);
1213
1214   assert(buf->refcount > 0);
1215   buf->refcount--;
1216
1217   query->oa.samples_head = NULL;
1218
1219   reap_old_sample_buffers(perf_ctx);
1220}
1221
1222/* In general if we see anything spurious while accumulating results,
1223 * we don't try and continue accumulating the current query, hoping
1224 * for the best, we scrap anything outstanding, and then hope for the
1225 * best with new queries.
1226 */
1227static void
1228discard_all_queries(struct intel_perf_context *perf_ctx)
1229{
1230   while (perf_ctx->unaccumulated_elements) {
1231      struct intel_perf_query_object *query = perf_ctx->unaccumulated[0];
1232
1233      query->oa.results_accumulated = true;
1234      drop_from_unaccumulated_query_list(perf_ctx, query);
1235
1236      dec_n_users(perf_ctx);
1237   }
1238}
1239
1240/* Looks for the validity bit of context ID (dword 2) of an OA report. */
1241static bool
1242oa_report_ctx_id_valid(const struct intel_device_info *devinfo,
1243                       const uint32_t *report)
1244{
1245   assert(devinfo->ver >= 8);
1246   if (devinfo->ver == 8)
1247      return (report[0] & (1 << 25)) != 0;
1248   return (report[0] & (1 << 16)) != 0;
1249}
1250
1251/**
1252 * Accumulate raw OA counter values based on deltas between pairs of
1253 * OA reports.
1254 *
1255 * Accumulation starts from the first report captured via
1256 * MI_REPORT_PERF_COUNT (MI_RPC) by brw_begin_perf_query() until the
1257 * last MI_RPC report requested by brw_end_perf_query(). Between these
1258 * two reports there may also some number of periodically sampled OA
1259 * reports collected via the i915 perf interface - depending on the
1260 * duration of the query.
1261 *
1262 * These periodic snapshots help to ensure we handle counter overflow
1263 * correctly by being frequent enough to ensure we don't miss multiple
1264 * overflows of a counter between snapshots. For Gfx8+ the i915 perf
1265 * snapshots provide the extra context-switch reports that let us
1266 * subtract out the progress of counters associated with other
1267 * contexts running on the system.
1268 */
1269static void
1270accumulate_oa_reports(struct intel_perf_context *perf_ctx,
1271                      struct intel_perf_query_object *query)
1272{
1273   const struct intel_device_info *devinfo = perf_ctx->devinfo;
1274   uint32_t *start;
1275   uint32_t *last;
1276   uint32_t *end;
1277   struct exec_node *first_samples_node;
1278   bool last_report_ctx_match = true;
1279   int out_duration = 0;
1280
1281   assert(query->oa.map != NULL);
1282
1283   start = last = query->oa.map + field_offset(false, 0);
1284   end = query->oa.map + field_offset(true, 0);
1285
1286   if (start[0] != query->oa.begin_report_id) {
1287      DBG("Spurious start report id=%"PRIu32"\n", start[0]);
1288      goto error;
1289   }
1290   if (end[0] != (query->oa.begin_report_id + 1)) {
1291      DBG("Spurious end report id=%"PRIu32"\n", end[0]);
1292      goto error;
1293   }
1294
1295   /* On Gfx12+ OA reports are sourced from per context counters, so we don't
1296    * ever have to look at the global OA buffer. Yey \o/
1297    */
1298   if (perf_ctx->devinfo->ver >= 12) {
1299      last = start;
1300      goto end;
1301   }
1302
1303   /* See if we have any periodic reports to accumulate too... */
1304
1305   /* N.B. The oa.samples_head was set when the query began and
1306    * pointed to the tail of the perf_ctx->sample_buffers list at
1307    * the time the query started. Since the buffer existed before the
1308    * first MI_REPORT_PERF_COUNT command was emitted we therefore know
1309    * that no data in this particular node's buffer can possibly be
1310    * associated with the query - so skip ahead one...
1311    */
1312   first_samples_node = query->oa.samples_head->next;
1313
1314   foreach_list_typed_from(struct oa_sample_buf, buf, link,
1315                           &perf_ctx->sample_buffers,
1316                           first_samples_node)
1317   {
1318      int offset = 0;
1319
1320      while (offset < buf->len) {
1321         const struct drm_i915_perf_record_header *header =
1322            (const struct drm_i915_perf_record_header *)(buf->buf + offset);
1323
1324         assert(header->size != 0);
1325         assert(header->size <= buf->len);
1326
1327         offset += header->size;
1328
1329         switch (header->type) {
1330         case DRM_I915_PERF_RECORD_SAMPLE: {
1331            uint32_t *report = (uint32_t *)(header + 1);
1332            bool report_ctx_match = true;
1333            bool add = true;
1334
1335            /* Ignore reports that come before the start marker.
1336             * (Note: takes care to allow overflow of 32bit timestamps)
1337             */
1338            if (intel_device_info_timebase_scale(devinfo,
1339                                               report[1] - start[1]) > 5000000000) {
1340               continue;
1341            }
1342
1343            /* Ignore reports that come after the end marker.
1344             * (Note: takes care to allow overflow of 32bit timestamps)
1345             */
1346            if (intel_device_info_timebase_scale(devinfo,
1347                                               report[1] - end[1]) <= 5000000000) {
1348               goto end;
1349            }
1350
1351            /* For Gfx8+ since the counters continue while other
1352             * contexts are running we need to discount any unrelated
1353             * deltas. The hardware automatically generates a report
1354             * on context switch which gives us a new reference point
1355             * to continuing adding deltas from.
1356             *
1357             * For Haswell we can rely on the HW to stop the progress
1358             * of OA counters while any other context is acctive.
1359             */
1360            if (devinfo->ver >= 8) {
1361               /* Consider that the current report matches our context only if
1362                * the report says the report ID is valid.
1363                */
1364               report_ctx_match = oa_report_ctx_id_valid(devinfo, report) &&
1365                  report[2] == start[2];
1366               if (report_ctx_match)
1367                  out_duration = 0;
1368               else
1369                  out_duration++;
1370
1371               /* Only add the delta between <last, report> if the last report
1372                * was clearly identified as our context, or if we have at most
1373                * 1 report without a matching ID.
1374                *
1375                * The OA unit will sometimes label reports with an invalid
1376                * context ID when i915 rewrites the execlist submit register
1377                * with the same context as the one currently running. This
1378                * happens when i915 wants to notify the HW of ringbuffer tail
1379                * register update. We have to consider this report as part of
1380                * our context as the 3d pipeline behind the OACS unit is still
1381                * processing the operations started at the previous execlist
1382                * submission.
1383                */
1384               add = last_report_ctx_match && out_duration < 2;
1385            }
1386
1387            if (add) {
1388               intel_perf_query_result_accumulate(&query->oa.result,
1389                                                query->queryinfo,
1390                                                devinfo,
1391                                                last, report);
1392            } else {
1393               /* We're not adding the delta because we've identified it's not
1394                * for the context we filter for. We can consider that the
1395                * query was split.
1396                */
1397               query->oa.result.query_disjoint = true;
1398            }
1399
1400            last = report;
1401            last_report_ctx_match = report_ctx_match;
1402
1403            break;
1404         }
1405
1406         case DRM_I915_PERF_RECORD_OA_BUFFER_LOST:
1407             DBG("i915 perf: OA error: all reports lost\n");
1408             goto error;
1409         case DRM_I915_PERF_RECORD_OA_REPORT_LOST:
1410             DBG("i915 perf: OA report lost\n");
1411             break;
1412         }
1413      }
1414   }
1415
1416end:
1417
1418   intel_perf_query_result_accumulate(&query->oa.result, query->queryinfo,
1419                                    devinfo, last, end);
1420
1421   query->oa.results_accumulated = true;
1422   drop_from_unaccumulated_query_list(perf_ctx, query);
1423   dec_n_users(perf_ctx);
1424
1425   return;
1426
1427error:
1428
1429   discard_all_queries(perf_ctx);
1430}
1431
1432void
1433intel_perf_delete_query(struct intel_perf_context *perf_ctx,
1434                        struct intel_perf_query_object *query)
1435{
1436   struct intel_perf_config *perf_cfg = perf_ctx->perf;
1437
1438   /* We can assume that the frontend waits for a query to complete
1439    * before ever calling into here, so we don't have to worry about
1440    * deleting an in-flight query object.
1441    */
1442   switch (query->queryinfo->kind) {
1443   case INTEL_PERF_QUERY_TYPE_OA:
1444   case INTEL_PERF_QUERY_TYPE_RAW:
1445      if (query->oa.bo) {
1446         if (!query->oa.results_accumulated) {
1447            drop_from_unaccumulated_query_list(perf_ctx, query);
1448            dec_n_users(perf_ctx);
1449         }
1450
1451         perf_cfg->vtbl.bo_unreference(query->oa.bo);
1452         query->oa.bo = NULL;
1453      }
1454
1455      query->oa.results_accumulated = false;
1456      break;
1457
1458   case INTEL_PERF_QUERY_TYPE_PIPELINE:
1459      if (query->pipeline_stats.bo) {
1460         perf_cfg->vtbl.bo_unreference(query->pipeline_stats.bo);
1461         query->pipeline_stats.bo = NULL;
1462      }
1463      break;
1464
1465   default:
1466      unreachable("Unknown query type");
1467      break;
1468   }
1469
1470   /* As an indication that the INTEL_performance_query extension is no
1471    * longer in use, it's a good time to free our cache of sample
1472    * buffers and close any current i915-perf stream.
1473    */
1474   if (--perf_ctx->n_query_instances == 0) {
1475      free_sample_bufs(perf_ctx);
1476      intel_perf_close(perf_ctx, query->queryinfo);
1477   }
1478
1479   free(query);
1480}
1481
1482static int
1483get_oa_counter_data(struct intel_perf_context *perf_ctx,
1484                    struct intel_perf_query_object *query,
1485                    size_t data_size,
1486                    uint8_t *data)
1487{
1488   struct intel_perf_config *perf_cfg = perf_ctx->perf;
1489   const struct intel_perf_query_info *queryinfo = query->queryinfo;
1490   int n_counters = queryinfo->n_counters;
1491   int written = 0;
1492
1493   for (int i = 0; i < n_counters; i++) {
1494      const struct intel_perf_query_counter *counter = &queryinfo->counters[i];
1495      uint64_t *out_uint64;
1496      float *out_float;
1497      size_t counter_size = intel_perf_query_counter_get_size(counter);
1498
1499      if (counter_size) {
1500         switch (counter->data_type) {
1501         case INTEL_PERF_COUNTER_DATA_TYPE_UINT64:
1502            out_uint64 = (uint64_t *)(data + counter->offset);
1503            *out_uint64 =
1504               counter->oa_counter_read_uint64(perf_cfg, queryinfo,
1505                                               &query->oa.result);
1506            break;
1507         case INTEL_PERF_COUNTER_DATA_TYPE_FLOAT:
1508            out_float = (float *)(data + counter->offset);
1509            *out_float =
1510               counter->oa_counter_read_float(perf_cfg, queryinfo,
1511                                              &query->oa.result);
1512            break;
1513         default:
1514            /* So far we aren't using uint32, double or bool32... */
1515            unreachable("unexpected counter data type");
1516         }
1517
1518         if (counter->offset + counter_size > written)
1519            written = counter->offset + counter_size;
1520      }
1521   }
1522
1523   return written;
1524}
1525
1526static int
1527get_pipeline_stats_data(struct intel_perf_context *perf_ctx,
1528                        struct intel_perf_query_object *query,
1529                        size_t data_size,
1530                        uint8_t *data)
1531
1532{
1533   struct intel_perf_config *perf_cfg = perf_ctx->perf;
1534   const struct intel_perf_query_info *queryinfo = query->queryinfo;
1535   int n_counters = queryinfo->n_counters;
1536   uint8_t *p = data;
1537
1538   uint64_t *start = perf_cfg->vtbl.bo_map(perf_ctx->ctx, query->pipeline_stats.bo, MAP_READ);
1539   uint64_t *end = start + (STATS_BO_END_OFFSET_BYTES / sizeof(uint64_t));
1540
1541   for (int i = 0; i < n_counters; i++) {
1542      const struct intel_perf_query_counter *counter = &queryinfo->counters[i];
1543      uint64_t value = end[i] - start[i];
1544
1545      if (counter->pipeline_stat.numerator !=
1546          counter->pipeline_stat.denominator) {
1547         value *= counter->pipeline_stat.numerator;
1548         value /= counter->pipeline_stat.denominator;
1549      }
1550
1551      *((uint64_t *)p) = value;
1552      p += 8;
1553   }
1554
1555   perf_cfg->vtbl.bo_unmap(query->pipeline_stats.bo);
1556
1557   return p - data;
1558}
1559
1560void
1561intel_perf_get_query_data(struct intel_perf_context *perf_ctx,
1562                          struct intel_perf_query_object *query,
1563                          void *current_batch,
1564                          int data_size,
1565                          unsigned *data,
1566                          unsigned *bytes_written)
1567{
1568   struct intel_perf_config *perf_cfg = perf_ctx->perf;
1569   int written = 0;
1570
1571   switch (query->queryinfo->kind) {
1572   case INTEL_PERF_QUERY_TYPE_OA:
1573   case INTEL_PERF_QUERY_TYPE_RAW:
1574      if (!query->oa.results_accumulated) {
1575         /* Due to the sampling frequency of the OA buffer by the i915-perf
1576          * driver, there can be a 5ms delay between the Mesa seeing the query
1577          * complete and i915 making all the OA buffer reports available to us.
1578          * We need to wait for all the reports to come in before we can do
1579          * the post processing removing unrelated deltas.
1580          * There is a i915-perf series to address this issue, but it's
1581          * not been merged upstream yet.
1582          */
1583         while (!read_oa_samples_for_query(perf_ctx, query, current_batch))
1584            ;
1585
1586         uint32_t *begin_report = query->oa.map;
1587         uint32_t *end_report = query->oa.map + perf_cfg->query_layout.size;
1588         intel_perf_query_result_accumulate_fields(&query->oa.result,
1589                                                 query->queryinfo,
1590                                                 perf_ctx->devinfo,
1591                                                 begin_report,
1592                                                 end_report,
1593                                                 true /* no_oa_accumulate */);
1594         accumulate_oa_reports(perf_ctx, query);
1595         assert(query->oa.results_accumulated);
1596
1597         perf_cfg->vtbl.bo_unmap(query->oa.bo);
1598         query->oa.map = NULL;
1599      }
1600      if (query->queryinfo->kind == INTEL_PERF_QUERY_TYPE_OA) {
1601         written = get_oa_counter_data(perf_ctx, query, data_size, (uint8_t *)data);
1602      } else {
1603         const struct intel_device_info *devinfo = perf_ctx->devinfo;
1604
1605         written = intel_perf_query_result_write_mdapi((uint8_t *)data, data_size,
1606                                                     devinfo, query->queryinfo,
1607                                                     &query->oa.result);
1608      }
1609      break;
1610
1611   case INTEL_PERF_QUERY_TYPE_PIPELINE:
1612      written = get_pipeline_stats_data(perf_ctx, query, data_size, (uint8_t *)data);
1613      break;
1614
1615   default:
1616      unreachable("Unknown query type");
1617      break;
1618   }
1619
1620   if (bytes_written)
1621      *bytes_written = written;
1622}
1623
1624void
1625intel_perf_dump_query_count(struct intel_perf_context *perf_ctx)
1626{
1627   DBG("Queries: (Open queries = %d, OA users = %d)\n",
1628       perf_ctx->n_active_oa_queries, perf_ctx->n_oa_users);
1629}
1630
1631void
1632intel_perf_dump_query(struct intel_perf_context *ctx,
1633                      struct intel_perf_query_object *obj,
1634                      void *current_batch)
1635{
1636   switch (obj->queryinfo->kind) {
1637   case INTEL_PERF_QUERY_TYPE_OA:
1638   case INTEL_PERF_QUERY_TYPE_RAW:
1639      DBG("BO: %-4s OA data: %-10s %-15s\n",
1640          obj->oa.bo ? "yes," : "no,",
1641          intel_perf_is_query_ready(ctx, obj, current_batch) ? "ready," : "not ready,",
1642          obj->oa.results_accumulated ? "accumulated" : "not accumulated");
1643      break;
1644   case INTEL_PERF_QUERY_TYPE_PIPELINE:
1645      DBG("BO: %-4s\n",
1646          obj->pipeline_stats.bo ? "yes" : "no");
1647      break;
1648   default:
1649      unreachable("Unknown query type");
1650      break;
1651   }
1652}
1653