Loading...
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/nospec.h>
196#include <linux/sizes.h>
197#include <linux/uuid.h>
198
199#include "gem/i915_gem_context.h"
200#include "gem/i915_gem_internal.h"
201#include "gt/intel_engine_pm.h"
202#include "gt/intel_engine_regs.h"
203#include "gt/intel_engine_user.h"
204#include "gt/intel_execlists_submission.h"
205#include "gt/intel_gpu_commands.h"
206#include "gt/intel_gt.h"
207#include "gt/intel_gt_clock_utils.h"
208#include "gt/intel_gt_mcr.h"
209#include "gt/intel_gt_print.h"
210#include "gt/intel_gt_regs.h"
211#include "gt/intel_lrc.h"
212#include "gt/intel_lrc_reg.h"
213#include "gt/intel_rc6.h"
214#include "gt/intel_ring.h"
215#include "gt/uc/intel_guc_slpc.h"
216
217#include "i915_drv.h"
218#include "i915_file_private.h"
219#include "i915_perf.h"
220#include "i915_perf_oa_regs.h"
221#include "i915_reg.h"
222
223/* HW requires this to be a power of two, between 128k and 16M, though driver
224 * is currently generally designed assuming the largest 16M size is used such
225 * that the overflow cases are unlikely in normal operation.
226 */
227#define OA_BUFFER_SIZE SZ_16M
228
229#define OA_TAKEN(tail, head) ((tail - head) & (OA_BUFFER_SIZE - 1))
230
231/**
232 * DOC: OA Tail Pointer Race
233 *
234 * There's a HW race condition between OA unit tail pointer register updates and
235 * writes to memory whereby the tail pointer can sometimes get ahead of what's
236 * been written out to the OA buffer so far (in terms of what's visible to the
237 * CPU).
238 *
239 * Although this can be observed explicitly while copying reports to userspace
240 * by checking for a zeroed report-id field in tail reports, we want to account
241 * for this earlier, as part of the oa_buffer_check_unlocked to avoid lots of
242 * redundant read() attempts.
243 *
244 * We workaround this issue in oa_buffer_check_unlocked() by reading the reports
245 * in the OA buffer, starting from the tail reported by the HW until we find a
246 * report with its first 2 dwords not 0 meaning its previous report is
247 * completely in memory and ready to be read. Those dwords are also set to 0
248 * once read and the whole buffer is cleared upon OA buffer initialization. The
249 * first dword is the reason for this report while the second is the timestamp,
250 * making the chances of having those 2 fields at 0 fairly unlikely. A more
251 * detailed explanation is available in oa_buffer_check_unlocked().
252 *
253 * Most of the implementation details for this workaround are in
254 * oa_buffer_check_unlocked() and _append_oa_reports()
255 *
256 * Note for posterity: previously the driver used to define an effective tail
257 * pointer that lagged the real pointer by a 'tail margin' measured in bytes
258 * derived from %OA_TAIL_MARGIN_NSEC and the configured sampling frequency.
259 * This was flawed considering that the OA unit may also automatically generate
260 * non-periodic reports (such as on context switch) or the OA unit may be
261 * enabled without any periodic sampling.
262 */
263#define OA_TAIL_MARGIN_NSEC 100000ULL
264#define INVALID_TAIL_PTR 0xffffffff
265
266/* The default frequency for checking whether the OA unit has written new
267 * reports to the circular OA buffer...
268 */
269#define DEFAULT_POLL_FREQUENCY_HZ 200
270#define DEFAULT_POLL_PERIOD_NS (NSEC_PER_SEC / DEFAULT_POLL_FREQUENCY_HZ)
271
272/* for sysctl proc_dointvec_minmax of dev.i915.perf_stream_paranoid */
273static u32 i915_perf_stream_paranoid = true;
274
275/* The maximum exponent the hardware accepts is 63 (essentially it selects one
276 * of the 64bit timestamp bits to trigger reports from) but there's currently
277 * no known use case for sampling as infrequently as once per 47 thousand years.
278 *
279 * Since the timestamps included in OA reports are only 32bits it seems
280 * reasonable to limit the OA exponent where it's still possible to account for
281 * overflow in OA report timestamps.
282 */
283#define OA_EXPONENT_MAX 31
284
285#define INVALID_CTX_ID 0xffffffff
286
287/* On Gen8+ automatically triggered OA reports include a 'reason' field... */
288#define OAREPORT_REASON_MASK 0x3f
289#define OAREPORT_REASON_MASK_EXTENDED 0x7f
290#define OAREPORT_REASON_SHIFT 19
291#define OAREPORT_REASON_TIMER (1<<0)
292#define OAREPORT_REASON_CTX_SWITCH (1<<3)
293#define OAREPORT_REASON_CLK_RATIO (1<<5)
294
295#define HAS_MI_SET_PREDICATE(i915) (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 55))
296
297/* For sysctl proc_dointvec_minmax of i915_oa_max_sample_rate
298 *
299 * The highest sampling frequency we can theoretically program the OA unit
300 * with is always half the timestamp frequency: E.g. 6.25Mhz for Haswell.
301 *
302 * Initialized just before we register the sysctl parameter.
303 */
304static int oa_sample_rate_hard_limit;
305
306/* Theoretically we can program the OA unit to sample every 160ns but don't
307 * allow that by default unless root...
308 *
309 * The default threshold of 100000Hz is based on perf's similar
310 * kernel.perf_event_max_sample_rate sysctl parameter.
311 */
312static u32 i915_oa_max_sample_rate = 100000;
313
314/* XXX: beware if future OA HW adds new report formats that the current
315 * code assumes all reports have a power-of-two size and ~(size - 1) can
316 * be used as a mask to align the OA tail pointer.
317 */
318static const struct i915_oa_format oa_formats[I915_OA_FORMAT_MAX] = {
319 [I915_OA_FORMAT_A13] = { 0, 64 },
320 [I915_OA_FORMAT_A29] = { 1, 128 },
321 [I915_OA_FORMAT_A13_B8_C8] = { 2, 128 },
322 /* A29_B8_C8 Disallowed as 192 bytes doesn't factor into buffer size */
323 [I915_OA_FORMAT_B4_C8] = { 4, 64 },
324 [I915_OA_FORMAT_A45_B8_C8] = { 5, 256 },
325 [I915_OA_FORMAT_B4_C8_A16] = { 6, 128 },
326 [I915_OA_FORMAT_C4_B8] = { 7, 64 },
327 [I915_OA_FORMAT_A12] = { 0, 64 },
328 [I915_OA_FORMAT_A12_B8_C8] = { 2, 128 },
329 [I915_OA_FORMAT_A32u40_A4u32_B8_C8] = { 5, 256 },
330 [I915_OAR_FORMAT_A32u40_A4u32_B8_C8] = { 5, 256 },
331 [I915_OA_FORMAT_A24u40_A14u32_B8_C8] = { 5, 256 },
332 [I915_OAM_FORMAT_MPEC8u64_B8_C8] = { 1, 192, TYPE_OAM, HDR_64_BIT },
333 [I915_OAM_FORMAT_MPEC8u32_B8_C8] = { 2, 128, TYPE_OAM, HDR_64_BIT },
334};
335
336static const u32 mtl_oa_base[] = {
337 [PERF_GROUP_OAM_SAMEDIA_0] = 0x393000,
338};
339
340#define SAMPLE_OA_REPORT (1<<0)
341
342/**
343 * struct perf_open_properties - for validated properties given to open a stream
344 * @sample_flags: `DRM_I915_PERF_PROP_SAMPLE_*` properties are tracked as flags
345 * @single_context: Whether a single or all gpu contexts should be monitored
346 * @hold_preemption: Whether the preemption is disabled for the filtered
347 * context
348 * @ctx_handle: A gem ctx handle for use with @single_context
349 * @metrics_set: An ID for an OA unit metric set advertised via sysfs
350 * @oa_format: An OA unit HW report format
351 * @oa_periodic: Whether to enable periodic OA unit sampling
352 * @oa_period_exponent: The OA unit sampling period is derived from this
353 * @engine: The engine (typically rcs0) being monitored by the OA unit
354 * @has_sseu: Whether @sseu was specified by userspace
355 * @sseu: internal SSEU configuration computed either from the userspace
356 * specified configuration in the opening parameters or a default value
357 * (see get_default_sseu_config())
358 * @poll_oa_period: The period in nanoseconds at which the CPU will check for OA
359 * data availability
360 *
361 * As read_properties_unlocked() enumerates and validates the properties given
362 * to open a stream of metrics the configuration is built up in the structure
363 * which starts out zero initialized.
364 */
365struct perf_open_properties {
366 u32 sample_flags;
367
368 u64 single_context:1;
369 u64 hold_preemption:1;
370 u64 ctx_handle;
371
372 /* OA sampling state */
373 int metrics_set;
374 int oa_format;
375 bool oa_periodic;
376 int oa_period_exponent;
377
378 struct intel_engine_cs *engine;
379
380 bool has_sseu;
381 struct intel_sseu sseu;
382
383 u64 poll_oa_period;
384};
385
386struct i915_oa_config_bo {
387 struct llist_node node;
388
389 struct i915_oa_config *oa_config;
390 struct i915_vma *vma;
391};
392
393static struct ctl_table_header *sysctl_header;
394
395static enum hrtimer_restart oa_poll_check_timer_cb(struct hrtimer *hrtimer);
396
397void i915_oa_config_release(struct kref *ref)
398{
399 struct i915_oa_config *oa_config =
400 container_of(ref, typeof(*oa_config), ref);
401
402 kfree(oa_config->flex_regs);
403 kfree(oa_config->b_counter_regs);
404 kfree(oa_config->mux_regs);
405
406 kfree_rcu(oa_config, rcu);
407}
408
409struct i915_oa_config *
410i915_perf_get_oa_config(struct i915_perf *perf, int metrics_set)
411{
412 struct i915_oa_config *oa_config;
413
414 rcu_read_lock();
415 oa_config = idr_find(&perf->metrics_idr, metrics_set);
416 if (oa_config)
417 oa_config = i915_oa_config_get(oa_config);
418 rcu_read_unlock();
419
420 return oa_config;
421}
422
423static void free_oa_config_bo(struct i915_oa_config_bo *oa_bo)
424{
425 i915_oa_config_put(oa_bo->oa_config);
426 i915_vma_put(oa_bo->vma);
427 kfree(oa_bo);
428}
429
430static inline const
431struct i915_perf_regs *__oa_regs(struct i915_perf_stream *stream)
432{
433 return &stream->engine->oa_group->regs;
434}
435
436static u32 gen12_oa_hw_tail_read(struct i915_perf_stream *stream)
437{
438 struct intel_uncore *uncore = stream->uncore;
439
440 return intel_uncore_read(uncore, __oa_regs(stream)->oa_tail_ptr) &
441 GEN12_OAG_OATAILPTR_MASK;
442}
443
444static u32 gen8_oa_hw_tail_read(struct i915_perf_stream *stream)
445{
446 struct intel_uncore *uncore = stream->uncore;
447
448 return intel_uncore_read(uncore, GEN8_OATAILPTR) & GEN8_OATAILPTR_MASK;
449}
450
451static u32 gen7_oa_hw_tail_read(struct i915_perf_stream *stream)
452{
453 struct intel_uncore *uncore = stream->uncore;
454 u32 oastatus1 = intel_uncore_read(uncore, GEN7_OASTATUS1);
455
456 return oastatus1 & GEN7_OASTATUS1_TAIL_MASK;
457}
458
459#define oa_report_header_64bit(__s) \
460 ((__s)->oa_buffer.format->header == HDR_64_BIT)
461
462static u64 oa_report_id(struct i915_perf_stream *stream, void *report)
463{
464 return oa_report_header_64bit(stream) ? *(u64 *)report : *(u32 *)report;
465}
466
467static u64 oa_report_reason(struct i915_perf_stream *stream, void *report)
468{
469 return (oa_report_id(stream, report) >> OAREPORT_REASON_SHIFT) &
470 (GRAPHICS_VER(stream->perf->i915) == 12 ?
471 OAREPORT_REASON_MASK_EXTENDED :
472 OAREPORT_REASON_MASK);
473}
474
475static void oa_report_id_clear(struct i915_perf_stream *stream, u32 *report)
476{
477 if (oa_report_header_64bit(stream))
478 *(u64 *)report = 0;
479 else
480 *report = 0;
481}
482
483static bool oa_report_ctx_invalid(struct i915_perf_stream *stream, void *report)
484{
485 return !(oa_report_id(stream, report) &
486 stream->perf->gen8_valid_ctx_bit);
487}
488
489static u64 oa_timestamp(struct i915_perf_stream *stream, void *report)
490{
491 return oa_report_header_64bit(stream) ?
492 *((u64 *)report + 1) :
493 *((u32 *)report + 1);
494}
495
496static void oa_timestamp_clear(struct i915_perf_stream *stream, u32 *report)
497{
498 if (oa_report_header_64bit(stream))
499 *(u64 *)&report[2] = 0;
500 else
501 report[1] = 0;
502}
503
504static u32 oa_context_id(struct i915_perf_stream *stream, u32 *report)
505{
506 u32 ctx_id = oa_report_header_64bit(stream) ? report[4] : report[2];
507
508 return ctx_id & stream->specific_ctx_id_mask;
509}
510
511static void oa_context_id_squash(struct i915_perf_stream *stream, u32 *report)
512{
513 if (oa_report_header_64bit(stream))
514 report[4] = INVALID_CTX_ID;
515 else
516 report[2] = INVALID_CTX_ID;
517}
518
519/**
520 * oa_buffer_check_unlocked - check for data and update tail ptr state
521 * @stream: i915 stream instance
522 *
523 * This is either called via fops (for blocking reads in user ctx) or the poll
524 * check hrtimer (atomic ctx) to check the OA buffer tail pointer and check
525 * if there is data available for userspace to read.
526 *
527 * This function is central to providing a workaround for the OA unit tail
528 * pointer having a race with respect to what data is visible to the CPU.
529 * It is responsible for reading tail pointers from the hardware and giving
530 * the pointers time to 'age' before they are made available for reading.
531 * (See description of OA_TAIL_MARGIN_NSEC above for further details.)
532 *
533 * Besides returning true when there is data available to read() this function
534 * also updates the tail in the oa_buffer object.
535 *
536 * Note: It's safe to read OA config state here unlocked, assuming that this is
537 * only called while the stream is enabled, while the global OA configuration
538 * can't be modified.
539 *
540 * Returns: %true if the OA buffer contains data, else %false
541 */
542static bool oa_buffer_check_unlocked(struct i915_perf_stream *stream)
543{
544 u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma);
545 int report_size = stream->oa_buffer.format->size;
546 u32 tail, hw_tail;
547 unsigned long flags;
548 bool pollin;
549 u32 partial_report_size;
550
551 /* We have to consider the (unlikely) possibility that read() errors
552 * could result in an OA buffer reset which might reset the head and
553 * tail state.
554 */
555 spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
556
557 hw_tail = stream->perf->ops.oa_hw_tail_read(stream);
558 hw_tail -= gtt_offset;
559
560 /* The tail pointer increases in 64 byte increments, not in report_size
561 * steps. Also the report size may not be a power of 2. Compute
562 * potentially partially landed report in the OA buffer
563 */
564 partial_report_size = OA_TAKEN(hw_tail, stream->oa_buffer.tail);
565 partial_report_size %= report_size;
566
567 /* Subtract partial amount off the tail */
568 hw_tail = OA_TAKEN(hw_tail, partial_report_size);
569
570 tail = hw_tail;
571
572 /* Walk the stream backward until we find a report with report
573 * id and timestmap not at 0. Since the circular buffer pointers
574 * progress by increments of 64 bytes and that reports can be up
575 * to 256 bytes long, we can't tell whether a report has fully
576 * landed in memory before the report id and timestamp of the
577 * following report have effectively landed.
578 *
579 * This is assuming that the writes of the OA unit land in
580 * memory in the order they were written to.
581 * If not : (╯°□°）╯︵ ┻━┻
582 */
583 while (OA_TAKEN(tail, stream->oa_buffer.tail) >= report_size) {
584 void *report = stream->oa_buffer.vaddr + tail;
585
586 if (oa_report_id(stream, report) ||
587 oa_timestamp(stream, report))
588 break;
589
590 tail = (tail - report_size) & (OA_BUFFER_SIZE - 1);
591 }
592
593 if (OA_TAKEN(hw_tail, tail) > report_size &&
594 __ratelimit(&stream->perf->tail_pointer_race))
595 drm_notice(&stream->uncore->i915->drm,
596 "unlanded report(s) head=0x%x tail=0x%x hw_tail=0x%x\n",
597 stream->oa_buffer.head, tail, hw_tail);
598
599 stream->oa_buffer.tail = tail;
600
601 pollin = OA_TAKEN(stream->oa_buffer.tail,
602 stream->oa_buffer.head) >= report_size;
603
604 spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
605
606 return pollin;
607}
608
609/**
610 * append_oa_status - Appends a status record to a userspace read() buffer.
611 * @stream: An i915-perf stream opened for OA metrics
612 * @buf: destination buffer given by userspace
613 * @count: the number of bytes userspace wants to read
614 * @offset: (inout): the current position for writing into @buf
615 * @type: The kind of status to report to userspace
616 *
617 * Writes a status record (such as `DRM_I915_PERF_RECORD_OA_REPORT_LOST`)
618 * into the userspace read() buffer.
619 *
620 * The @buf @offset will only be updated on success.
621 *
622 * Returns: 0 on success, negative error code on failure.
623 */
624static int append_oa_status(struct i915_perf_stream *stream,
625 char __user *buf,
626 size_t count,
627 size_t *offset,
628 enum drm_i915_perf_record_type type)
629{
630 struct drm_i915_perf_record_header header = { type, 0, sizeof(header) };
631
632 if ((count - *offset) < header.size)
633 return -ENOSPC;
634
635 if (copy_to_user(buf + *offset, &header, sizeof(header)))
636 return -EFAULT;
637
638 (*offset) += header.size;
639
640 return 0;
641}
642
643/**
644 * append_oa_sample - Copies single OA report into userspace read() buffer.
645 * @stream: An i915-perf stream opened for OA metrics
646 * @buf: destination buffer given by userspace
647 * @count: the number of bytes userspace wants to read
648 * @offset: (inout): the current position for writing into @buf
649 * @report: A single OA report to (optionally) include as part of the sample
650 *
651 * The contents of a sample are configured through `DRM_I915_PERF_PROP_SAMPLE_*`
652 * properties when opening a stream, tracked as `stream->sample_flags`. This
653 * function copies the requested components of a single sample to the given
654 * read() @buf.
655 *
656 * The @buf @offset will only be updated on success.
657 *
658 * Returns: 0 on success, negative error code on failure.
659 */
660static int append_oa_sample(struct i915_perf_stream *stream,
661 char __user *buf,
662 size_t count,
663 size_t *offset,
664 const u8 *report)
665{
666 int report_size = stream->oa_buffer.format->size;
667 struct drm_i915_perf_record_header header;
668 int report_size_partial;
669 u8 *oa_buf_end;
670
671 header.type = DRM_I915_PERF_RECORD_SAMPLE;
672 header.pad = 0;
673 header.size = stream->sample_size;
674
675 if ((count - *offset) < header.size)
676 return -ENOSPC;
677
678 buf += *offset;
679 if (copy_to_user(buf, &header, sizeof(header)))
680 return -EFAULT;
681 buf += sizeof(header);
682
683 oa_buf_end = stream->oa_buffer.vaddr + OA_BUFFER_SIZE;
684 report_size_partial = oa_buf_end - report;
685
686 if (report_size_partial < report_size) {
687 if (copy_to_user(buf, report, report_size_partial))
688 return -EFAULT;
689 buf += report_size_partial;
690
691 if (copy_to_user(buf, stream->oa_buffer.vaddr,
692 report_size - report_size_partial))
693 return -EFAULT;
694 } else if (copy_to_user(buf, report, report_size)) {
695 return -EFAULT;
696 }
697
698 (*offset) += header.size;
699
700 return 0;
701}
702
703/**
704 * gen8_append_oa_reports - Copies all buffered OA reports into
705 * userspace read() buffer.
706 * @stream: An i915-perf stream opened for OA metrics
707 * @buf: destination buffer given by userspace
708 * @count: the number of bytes userspace wants to read
709 * @offset: (inout): the current position for writing into @buf
710 *
711 * Notably any error condition resulting in a short read (-%ENOSPC or
712 * -%EFAULT) will be returned even though one or more records may
713 * have been successfully copied. In this case it's up to the caller
714 * to decide if the error should be squashed before returning to
715 * userspace.
716 *
717 * Note: reports are consumed from the head, and appended to the
718 * tail, so the tail chases the head?... If you think that's mad
719 * and back-to-front you're not alone, but this follows the
720 * Gen PRM naming convention.
721 *
722 * Returns: 0 on success, negative error code on failure.
723 */
724static int gen8_append_oa_reports(struct i915_perf_stream *stream,
725 char __user *buf,
726 size_t count,
727 size_t *offset)
728{
729 struct intel_uncore *uncore = stream->uncore;
730 int report_size = stream->oa_buffer.format->size;
731 u8 *oa_buf_base = stream->oa_buffer.vaddr;
732 u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma);
733 u32 mask = (OA_BUFFER_SIZE - 1);
734 size_t start_offset = *offset;
735 unsigned long flags;
736 u32 head, tail;
737 int ret = 0;
738
739 if (drm_WARN_ON(&uncore->i915->drm, !stream->enabled))
740 return -EIO;
741
742 spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
743
744 head = stream->oa_buffer.head;
745 tail = stream->oa_buffer.tail;
746
747 spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
748
749 /*
750 * An out of bounds or misaligned head or tail pointer implies a driver
751 * bug since we validate + align the tail pointers we read from the
752 * hardware and we are in full control of the head pointer which should
753 * only be incremented by multiples of the report size.
754 */
755 if (drm_WARN_ONCE(&uncore->i915->drm,
756 head > OA_BUFFER_SIZE ||
757 tail > OA_BUFFER_SIZE,
758 "Inconsistent OA buffer pointers: head = %u, tail = %u\n",
759 head, tail))
760 return -EIO;
761
762
763 for (/* none */;
764 OA_TAKEN(tail, head);
765 head = (head + report_size) & mask) {
766 u8 *report = oa_buf_base + head;
767 u32 *report32 = (void *)report;
768 u32 ctx_id;
769 u64 reason;
770
771 /*
772 * The reason field includes flags identifying what
773 * triggered this specific report (mostly timer
774 * triggered or e.g. due to a context switch).
775 */
776 reason = oa_report_reason(stream, report);
777 ctx_id = oa_context_id(stream, report32);
778
779 /*
780 * Squash whatever is in the CTX_ID field if it's marked as
781 * invalid to be sure we avoid false-positive, single-context
782 * filtering below...
783 *
784 * Note: that we don't clear the valid_ctx_bit so userspace can
785 * understand that the ID has been squashed by the kernel.
786 *
787 * Update:
788 *
789 * On XEHP platforms the behavior of context id valid bit has
790 * changed compared to prior platforms. To describe this, we
791 * define a few terms:
792 *
793 * context-switch-report: This is a report with the reason type
794 * being context-switch. It is generated when a context switches
795 * out.
796 *
797 * context-valid-bit: A bit that is set in the report ID field
798 * to indicate that a valid context has been loaded.
799 *
800 * gpu-idle: A condition characterized by a
801 * context-switch-report with context-valid-bit set to 0.
802 *
803 * On prior platforms, context-id-valid bit is set to 0 only
804 * when GPU goes idle. In all other reports, it is set to 1.
805 *
806 * On XEHP platforms, context-valid-bit is set to 1 in a context
807 * switch report if a new context switched in. For all other
808 * reports it is set to 0.
809 *
810 * This change in behavior causes an issue with MMIO triggered
811 * reports. MMIO triggered reports have the markers in the
812 * context ID field and the context-valid-bit is 0. The logic
813 * below to squash the context ID would render the report
814 * useless since the user will not be able to find it in the OA
815 * buffer. Since MMIO triggered reports exist only on XEHP,
816 * we should avoid squashing these for XEHP platforms.
817 */
818
819 if (oa_report_ctx_invalid(stream, report) &&
820 GRAPHICS_VER_FULL(stream->engine->i915) < IP_VER(12, 55)) {
821 ctx_id = INVALID_CTX_ID;
822 oa_context_id_squash(stream, report32);
823 }
824
825 /*
826 * NB: For Gen 8 the OA unit no longer supports clock gating
827 * off for a specific context and the kernel can't securely
828 * stop the counters from updating as system-wide / global
829 * values.
830 *
831 * Automatic reports now include a context ID so reports can be
832 * filtered on the cpu but it's not worth trying to
833 * automatically subtract/hide counter progress for other
834 * contexts while filtering since we can't stop userspace
835 * issuing MI_REPORT_PERF_COUNT commands which would still
836 * provide a side-band view of the real values.
837 *
838 * To allow userspace (such as Mesa/GL_INTEL_performance_query)
839 * to normalize counters for a single filtered context then it
840 * needs be forwarded bookend context-switch reports so that it
841 * can track switches in between MI_REPORT_PERF_COUNT commands
842 * and can itself subtract/ignore the progress of counters
843 * associated with other contexts. Note that the hardware
844 * automatically triggers reports when switching to a new
845 * context which are tagged with the ID of the newly active
846 * context. To avoid the complexity (and likely fragility) of
847 * reading ahead while parsing reports to try and minimize
848 * forwarding redundant context switch reports (i.e. between
849 * other, unrelated contexts) we simply elect to forward them
850 * all.
851 *
852 * We don't rely solely on the reason field to identify context
853 * switches since it's not-uncommon for periodic samples to
854 * identify a switch before any 'context switch' report.
855 */
856 if (!stream->ctx ||
857 stream->specific_ctx_id == ctx_id ||
858 stream->oa_buffer.last_ctx_id == stream->specific_ctx_id ||
859 reason & OAREPORT_REASON_CTX_SWITCH) {
860
861 /*
862 * While filtering for a single context we avoid
863 * leaking the IDs of other contexts.
864 */
865 if (stream->ctx &&
866 stream->specific_ctx_id != ctx_id) {
867 oa_context_id_squash(stream, report32);
868 }
869
870 ret = append_oa_sample(stream, buf, count, offset,
871 report);
872 if (ret)
873 break;
874
875 stream->oa_buffer.last_ctx_id = ctx_id;
876 }
877
878 if (is_power_of_2(report_size)) {
879 /*
880 * Clear out the report id and timestamp as a means
881 * to detect unlanded reports.
882 */
883 oa_report_id_clear(stream, report32);
884 oa_timestamp_clear(stream, report32);
885 } else {
886 u8 *oa_buf_end = stream->oa_buffer.vaddr +
887 OA_BUFFER_SIZE;
888 u32 part = oa_buf_end - (u8 *)report32;
889
890 /* Zero out the entire report */
891 if (report_size <= part) {
892 memset(report32, 0, report_size);
893 } else {
894 memset(report32, 0, part);
895 memset(oa_buf_base, 0, report_size - part);
896 }
897 }
898 }
899
900 if (start_offset != *offset) {
901 i915_reg_t oaheadptr;
902
903 oaheadptr = GRAPHICS_VER(stream->perf->i915) == 12 ?
904 __oa_regs(stream)->oa_head_ptr :
905 GEN8_OAHEADPTR;
906
907 spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
908
909 /*
910 * We removed the gtt_offset for the copy loop above, indexing
911 * relative to oa_buf_base so put back here...
912 */
913 intel_uncore_write(uncore, oaheadptr,
914 (head + gtt_offset) & GEN12_OAG_OAHEADPTR_MASK);
915 stream->oa_buffer.head = head;
916
917 spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
918 }
919
920 return ret;
921}
922
923/**
924 * gen8_oa_read - copy status records then buffered OA reports
925 * @stream: An i915-perf stream opened for OA metrics
926 * @buf: destination buffer given by userspace
927 * @count: the number of bytes userspace wants to read
928 * @offset: (inout): the current position for writing into @buf
929 *
930 * Checks OA unit status registers and if necessary appends corresponding
931 * status records for userspace (such as for a buffer full condition) and then
932 * initiate appending any buffered OA reports.
933 *
934 * Updates @offset according to the number of bytes successfully copied into
935 * the userspace buffer.
936 *
937 * NB: some data may be successfully copied to the userspace buffer
938 * even if an error is returned, and this is reflected in the
939 * updated @offset.
940 *
941 * Returns: zero on success or a negative error code
942 */
943static int gen8_oa_read(struct i915_perf_stream *stream,
944 char __user *buf,
945 size_t count,
946 size_t *offset)
947{
948 struct intel_uncore *uncore = stream->uncore;
949 u32 oastatus;
950 i915_reg_t oastatus_reg;
951 int ret;
952
953 if (drm_WARN_ON(&uncore->i915->drm, !stream->oa_buffer.vaddr))
954 return -EIO;
955
956 oastatus_reg = GRAPHICS_VER(stream->perf->i915) == 12 ?
957 __oa_regs(stream)->oa_status :
958 GEN8_OASTATUS;
959
960 oastatus = intel_uncore_read(uncore, oastatus_reg);
961
962 /*
963 * We treat OABUFFER_OVERFLOW as a significant error:
964 *
965 * Although theoretically we could handle this more gracefully
966 * sometimes, some Gens don't correctly suppress certain
967 * automatically triggered reports in this condition and so we
968 * have to assume that old reports are now being trampled
969 * over.
970 *
971 * Considering how we don't currently give userspace control
972 * over the OA buffer size and always configure a large 16MB
973 * buffer, then a buffer overflow does anyway likely indicate
974 * that something has gone quite badly wrong.
975 */
976 if (oastatus & GEN8_OASTATUS_OABUFFER_OVERFLOW) {
977 ret = append_oa_status(stream, buf, count, offset,
978 DRM_I915_PERF_RECORD_OA_BUFFER_LOST);
979 if (ret)
980 return ret;
981
982 drm_dbg(&stream->perf->i915->drm,
983 "OA buffer overflow (exponent = %d): force restart\n",
984 stream->period_exponent);
985
986 stream->perf->ops.oa_disable(stream);
987 stream->perf->ops.oa_enable(stream);
988
989 /*
990 * Note: .oa_enable() is expected to re-init the oabuffer and
991 * reset GEN8_OASTATUS for us
992 */
993 oastatus = intel_uncore_read(uncore, oastatus_reg);
994 }
995
996 if (oastatus & GEN8_OASTATUS_REPORT_LOST) {
997 ret = append_oa_status(stream, buf, count, offset,
998 DRM_I915_PERF_RECORD_OA_REPORT_LOST);
999 if (ret)
1000 return ret;
1001
1002 intel_uncore_rmw(uncore, oastatus_reg,
1003 GEN8_OASTATUS_COUNTER_OVERFLOW |
1004 GEN8_OASTATUS_REPORT_LOST,
1005 IS_GRAPHICS_VER(uncore->i915, 8, 11) ?
1006 (GEN8_OASTATUS_HEAD_POINTER_WRAP |
1007 GEN8_OASTATUS_TAIL_POINTER_WRAP) : 0);
1008 }
1009
1010 return gen8_append_oa_reports(stream, buf, count, offset);
1011}
1012
1013/**
1014 * gen7_append_oa_reports - Copies all buffered OA reports into
1015 * userspace read() buffer.
1016 * @stream: An i915-perf stream opened for OA metrics
1017 * @buf: destination buffer given by userspace
1018 * @count: the number of bytes userspace wants to read
1019 * @offset: (inout): the current position for writing into @buf
1020 *
1021 * Notably any error condition resulting in a short read (-%ENOSPC or
1022 * -%EFAULT) will be returned even though one or more records may
1023 * have been successfully copied. In this case it's up to the caller
1024 * to decide if the error should be squashed before returning to
1025 * userspace.
1026 *
1027 * Note: reports are consumed from the head, and appended to the
1028 * tail, so the tail chases the head?... If you think that's mad
1029 * and back-to-front you're not alone, but this follows the
1030 * Gen PRM naming convention.
1031 *
1032 * Returns: 0 on success, negative error code on failure.
1033 */
1034static int gen7_append_oa_reports(struct i915_perf_stream *stream,
1035 char __user *buf,
1036 size_t count,
1037 size_t *offset)
1038{
1039 struct intel_uncore *uncore = stream->uncore;
1040 int report_size = stream->oa_buffer.format->size;
1041 u8 *oa_buf_base = stream->oa_buffer.vaddr;
1042 u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma);
1043 u32 mask = (OA_BUFFER_SIZE - 1);
1044 size_t start_offset = *offset;
1045 unsigned long flags;
1046 u32 head, tail;
1047 int ret = 0;
1048
1049 if (drm_WARN_ON(&uncore->i915->drm, !stream->enabled))
1050 return -EIO;
1051
1052 spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
1053
1054 head = stream->oa_buffer.head;
1055 tail = stream->oa_buffer.tail;
1056
1057 spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
1058
1059 /* An out of bounds or misaligned head or tail pointer implies a driver
1060 * bug since we validate + align the tail pointers we read from the
1061 * hardware and we are in full control of the head pointer which should
1062 * only be incremented by multiples of the report size (notably also
1063 * all a power of two).
1064 */
1065 if (drm_WARN_ONCE(&uncore->i915->drm,
1066 head > OA_BUFFER_SIZE || head % report_size ||
1067 tail > OA_BUFFER_SIZE || tail % report_size,
1068 "Inconsistent OA buffer pointers: head = %u, tail = %u\n",
1069 head, tail))
1070 return -EIO;
1071
1072
1073 for (/* none */;
1074 OA_TAKEN(tail, head);
1075 head = (head + report_size) & mask) {
1076 u8 *report = oa_buf_base + head;
1077 u32 *report32 = (void *)report;
1078
1079 /* All the report sizes factor neatly into the buffer
1080 * size so we never expect to see a report split
1081 * between the beginning and end of the buffer.
1082 *
1083 * Given the initial alignment check a misalignment
1084 * here would imply a driver bug that would result
1085 * in an overrun.
1086 */
1087 if (drm_WARN_ON(&uncore->i915->drm,
1088 (OA_BUFFER_SIZE - head) < report_size)) {
1089 drm_err(&uncore->i915->drm,
1090 "Spurious OA head ptr: non-integral report offset\n");
1091 break;
1092 }
1093
1094 /* The report-ID field for periodic samples includes
1095 * some undocumented flags related to what triggered
1096 * the report and is never expected to be zero so we
1097 * can check that the report isn't invalid before
1098 * copying it to userspace...
1099 */
1100 if (report32[0] == 0) {
1101 if (__ratelimit(&stream->perf->spurious_report_rs))
1102 drm_notice(&uncore->i915->drm,
1103 "Skipping spurious, invalid OA report\n");
1104 continue;
1105 }
1106
1107 ret = append_oa_sample(stream, buf, count, offset, report);
1108 if (ret)
1109 break;
1110
1111 /* Clear out the first 2 dwords as a mean to detect unlanded
1112 * reports.
1113 */
1114 report32[0] = 0;
1115 report32[1] = 0;
1116 }
1117
1118 if (start_offset != *offset) {
1119 spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
1120
1121 intel_uncore_write(uncore, GEN7_OASTATUS2,
1122 ((head + gtt_offset) & GEN7_OASTATUS2_HEAD_MASK) |
1123 GEN7_OASTATUS2_MEM_SELECT_GGTT);
1124 stream->oa_buffer.head = head;
1125
1126 spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
1127 }
1128
1129 return ret;
1130}
1131
1132/**
1133 * gen7_oa_read - copy status records then buffered OA reports
1134 * @stream: An i915-perf stream opened for OA metrics
1135 * @buf: destination buffer given by userspace
1136 * @count: the number of bytes userspace wants to read
1137 * @offset: (inout): the current position for writing into @buf
1138 *
1139 * Checks Gen 7 specific OA unit status registers and if necessary appends
1140 * corresponding status records for userspace (such as for a buffer full
1141 * condition) and then initiate appending any buffered OA reports.
1142 *
1143 * Updates @offset according to the number of bytes successfully copied into
1144 * the userspace buffer.
1145 *
1146 * Returns: zero on success or a negative error code
1147 */
1148static int gen7_oa_read(struct i915_perf_stream *stream,
1149 char __user *buf,
1150 size_t count,
1151 size_t *offset)
1152{
1153 struct intel_uncore *uncore = stream->uncore;
1154 u32 oastatus1;
1155 int ret;
1156
1157 if (drm_WARN_ON(&uncore->i915->drm, !stream->oa_buffer.vaddr))
1158 return -EIO;
1159
1160 oastatus1 = intel_uncore_read(uncore, GEN7_OASTATUS1);
1161
1162 /* XXX: On Haswell we don't have a safe way to clear oastatus1
1163 * bits while the OA unit is enabled (while the tail pointer
1164 * may be updated asynchronously) so we ignore status bits
1165 * that have already been reported to userspace.
1166 */
1167 oastatus1 &= ~stream->perf->gen7_latched_oastatus1;
1168
1169 /* We treat OABUFFER_OVERFLOW as a significant error:
1170 *
1171 * - The status can be interpreted to mean that the buffer is
1172 * currently full (with a higher precedence than OA_TAKEN()
1173 * which will start to report a near-empty buffer after an
1174 * overflow) but it's awkward that we can't clear the status
1175 * on Haswell, so without a reset we won't be able to catch
1176 * the state again.
1177 *
1178 * - Since it also implies the HW has started overwriting old
1179 * reports it may also affect our sanity checks for invalid
1180 * reports when copying to userspace that assume new reports
1181 * are being written to cleared memory.
1182 *
1183 * - In the future we may want to introduce a flight recorder
1184 * mode where the driver will automatically maintain a safe
1185 * guard band between head/tail, avoiding this overflow
1186 * condition, but we avoid the added driver complexity for
1187 * now.
1188 */
1189 if (unlikely(oastatus1 & GEN7_OASTATUS1_OABUFFER_OVERFLOW)) {
1190 ret = append_oa_status(stream, buf, count, offset,
1191 DRM_I915_PERF_RECORD_OA_BUFFER_LOST);
1192 if (ret)
1193 return ret;
1194
1195 drm_dbg(&stream->perf->i915->drm,
1196 "OA buffer overflow (exponent = %d): force restart\n",
1197 stream->period_exponent);
1198
1199 stream->perf->ops.oa_disable(stream);
1200 stream->perf->ops.oa_enable(stream);
1201
1202 oastatus1 = intel_uncore_read(uncore, GEN7_OASTATUS1);
1203 }
1204
1205 if (unlikely(oastatus1 & GEN7_OASTATUS1_REPORT_LOST)) {
1206 ret = append_oa_status(stream, buf, count, offset,
1207 DRM_I915_PERF_RECORD_OA_REPORT_LOST);
1208 if (ret)
1209 return ret;
1210 stream->perf->gen7_latched_oastatus1 |=
1211 GEN7_OASTATUS1_REPORT_LOST;
1212 }
1213
1214 return gen7_append_oa_reports(stream, buf, count, offset);
1215}
1216
1217/**
1218 * i915_oa_wait_unlocked - handles blocking IO until OA data available
1219 * @stream: An i915-perf stream opened for OA metrics
1220 *
1221 * Called when userspace tries to read() from a blocking stream FD opened
1222 * for OA metrics. It waits until the hrtimer callback finds a non-empty
1223 * OA buffer and wakes us.
1224 *
1225 * Note: it's acceptable to have this return with some false positives
1226 * since any subsequent read handling will return -EAGAIN if there isn't
1227 * really data ready for userspace yet.
1228 *
1229 * Returns: zero on success or a negative error code
1230 */
1231static int i915_oa_wait_unlocked(struct i915_perf_stream *stream)
1232{
1233 /* We would wait indefinitely if periodic sampling is not enabled */
1234 if (!stream->periodic)
1235 return -EIO;
1236
1237 return wait_event_interruptible(stream->poll_wq,
1238 oa_buffer_check_unlocked(stream));
1239}
1240
1241/**
1242 * i915_oa_poll_wait - call poll_wait() for an OA stream poll()
1243 * @stream: An i915-perf stream opened for OA metrics
1244 * @file: An i915 perf stream file
1245 * @wait: poll() state table
1246 *
1247 * For handling userspace polling on an i915 perf stream opened for OA metrics,
1248 * this starts a poll_wait with the wait queue that our hrtimer callback wakes
1249 * when it sees data ready to read in the circular OA buffer.
1250 */
1251static void i915_oa_poll_wait(struct i915_perf_stream *stream,
1252 struct file *file,
1253 poll_table *wait)
1254{
1255 poll_wait(file, &stream->poll_wq, wait);
1256}
1257
1258/**
1259 * i915_oa_read - just calls through to &i915_oa_ops->read
1260 * @stream: An i915-perf stream opened for OA metrics
1261 * @buf: destination buffer given by userspace
1262 * @count: the number of bytes userspace wants to read
1263 * @offset: (inout): the current position for writing into @buf
1264 *
1265 * Updates @offset according to the number of bytes successfully copied into
1266 * the userspace buffer.
1267 *
1268 * Returns: zero on success or a negative error code
1269 */
1270static int i915_oa_read(struct i915_perf_stream *stream,
1271 char __user *buf,
1272 size_t count,
1273 size_t *offset)
1274{
1275 return stream->perf->ops.read(stream, buf, count, offset);
1276}
1277
1278static struct intel_context *oa_pin_context(struct i915_perf_stream *stream)
1279{
1280 struct i915_gem_engines_iter it;
1281 struct i915_gem_context *ctx = stream->ctx;
1282 struct intel_context *ce;
1283 struct i915_gem_ww_ctx ww;
1284 int err = -ENODEV;
1285
1286 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
1287 if (ce->engine != stream->engine) /* first match! */
1288 continue;
1289
1290 err = 0;
1291 break;
1292 }
1293 i915_gem_context_unlock_engines(ctx);
1294
1295 if (err)
1296 return ERR_PTR(err);
1297
1298 i915_gem_ww_ctx_init(&ww, true);
1299retry:
1300 /*
1301 * As the ID is the gtt offset of the context's vma we
1302 * pin the vma to ensure the ID remains fixed.
1303 */
1304 err = intel_context_pin_ww(ce, &ww);
1305 if (err == -EDEADLK) {
1306 err = i915_gem_ww_ctx_backoff(&ww);
1307 if (!err)
1308 goto retry;
1309 }
1310 i915_gem_ww_ctx_fini(&ww);
1311
1312 if (err)
1313 return ERR_PTR(err);
1314
1315 stream->pinned_ctx = ce;
1316 return stream->pinned_ctx;
1317}
1318
1319static int
1320__store_reg_to_mem(struct i915_request *rq, i915_reg_t reg, u32 ggtt_offset)
1321{
1322 u32 *cs, cmd;
1323
1324 cmd = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
1325 if (GRAPHICS_VER(rq->i915) >= 8)
1326 cmd++;
1327
1328 cs = intel_ring_begin(rq, 4);
1329 if (IS_ERR(cs))
1330 return PTR_ERR(cs);
1331
1332 *cs++ = cmd;
1333 *cs++ = i915_mmio_reg_offset(reg);
1334 *cs++ = ggtt_offset;
1335 *cs++ = 0;
1336
1337 intel_ring_advance(rq, cs);
1338
1339 return 0;
1340}
1341
1342static int
1343__read_reg(struct intel_context *ce, i915_reg_t reg, u32 ggtt_offset)
1344{
1345 struct i915_request *rq;
1346 int err;
1347
1348 rq = i915_request_create(ce);
1349 if (IS_ERR(rq))
1350 return PTR_ERR(rq);
1351
1352 i915_request_get(rq);
1353
1354 err = __store_reg_to_mem(rq, reg, ggtt_offset);
1355
1356 i915_request_add(rq);
1357 if (!err && i915_request_wait(rq, 0, HZ / 2) < 0)
1358 err = -ETIME;
1359
1360 i915_request_put(rq);
1361
1362 return err;
1363}
1364
1365static int
1366gen12_guc_sw_ctx_id(struct intel_context *ce, u32 *ctx_id)
1367{
1368 struct i915_vma *scratch;
1369 u32 *val;
1370 int err;
1371
1372 scratch = __vm_create_scratch_for_read_pinned(&ce->engine->gt->ggtt->vm, 4);
1373 if (IS_ERR(scratch))
1374 return PTR_ERR(scratch);
1375
1376 err = i915_vma_sync(scratch);
1377 if (err)
1378 goto err_scratch;
1379
1380 err = __read_reg(ce, RING_EXECLIST_STATUS_HI(ce->engine->mmio_base),
1381 i915_ggtt_offset(scratch));
1382 if (err)
1383 goto err_scratch;
1384
1385 val = i915_gem_object_pin_map_unlocked(scratch->obj, I915_MAP_WB);
1386 if (IS_ERR(val)) {
1387 err = PTR_ERR(val);
1388 goto err_scratch;
1389 }
1390
1391 *ctx_id = *val;
1392 i915_gem_object_unpin_map(scratch->obj);
1393
1394err_scratch:
1395 i915_vma_unpin_and_release(&scratch, 0);
1396 return err;
1397}
1398
1399/*
1400 * For execlist mode of submission, pick an unused context id
1401 * 0 - (NUM_CONTEXT_TAG -1) are used by other contexts
1402 * XXX_MAX_CONTEXT_HW_ID is used by idle context
1403 *
1404 * For GuC mode of submission read context id from the upper dword of the
1405 * EXECLIST_STATUS register. Note that we read this value only once and expect
1406 * that the value stays fixed for the entire OA use case. There are cases where
1407 * GuC KMD implementation may deregister a context to reuse it's context id, but
1408 * we prevent that from happening to the OA context by pinning it.
1409 */
1410static int gen12_get_render_context_id(struct i915_perf_stream *stream)
1411{
1412 u32 ctx_id, mask;
1413 int ret;
1414
1415 if (intel_engine_uses_guc(stream->engine)) {
1416 ret = gen12_guc_sw_ctx_id(stream->pinned_ctx, &ctx_id);
1417 if (ret)
1418 return ret;
1419
1420 mask = ((1U << GEN12_GUC_SW_CTX_ID_WIDTH) - 1) <<
1421 (GEN12_GUC_SW_CTX_ID_SHIFT - 32);
1422 } else if (GRAPHICS_VER_FULL(stream->engine->i915) >= IP_VER(12, 55)) {
1423 ctx_id = (XEHP_MAX_CONTEXT_HW_ID - 1) <<
1424 (XEHP_SW_CTX_ID_SHIFT - 32);
1425
1426 mask = ((1U << XEHP_SW_CTX_ID_WIDTH) - 1) <<
1427 (XEHP_SW_CTX_ID_SHIFT - 32);
1428 } else {
1429 ctx_id = (GEN12_MAX_CONTEXT_HW_ID - 1) <<
1430 (GEN11_SW_CTX_ID_SHIFT - 32);
1431
1432 mask = ((1U << GEN11_SW_CTX_ID_WIDTH) - 1) <<
1433 (GEN11_SW_CTX_ID_SHIFT - 32);
1434 }
1435 stream->specific_ctx_id = ctx_id & mask;
1436 stream->specific_ctx_id_mask = mask;
1437
1438 return 0;
1439}
1440
1441static bool oa_find_reg_in_lri(u32 *state, u32 reg, u32 *offset, u32 end)
1442{
1443 u32 idx = *offset;
1444 u32 len = min(MI_LRI_LEN(state[idx]) + idx, end);
1445 bool found = false;
1446
1447 idx++;
1448 for (; idx < len; idx += 2) {
1449 if (state[idx] == reg) {
1450 found = true;
1451 break;
1452 }
1453 }
1454
1455 *offset = idx;
1456 return found;
1457}
1458
1459static u32 oa_context_image_offset(struct intel_context *ce, u32 reg)
1460{
1461 u32 offset, len = (ce->engine->context_size - PAGE_SIZE) / 4;
1462 u32 *state = ce->lrc_reg_state;
1463
1464 if (drm_WARN_ON(&ce->engine->i915->drm, !state))
1465 return U32_MAX;
1466
1467 for (offset = 0; offset < len; ) {
1468 if (IS_MI_LRI_CMD(state[offset])) {
1469 /*
1470 * We expect reg-value pairs in MI_LRI command, so
1471 * MI_LRI_LEN() should be even, if not, issue a warning.
1472 */
1473 drm_WARN_ON(&ce->engine->i915->drm,
1474 MI_LRI_LEN(state[offset]) & 0x1);
1475
1476 if (oa_find_reg_in_lri(state, reg, &offset, len))
1477 break;
1478 } else {
1479 offset++;
1480 }
1481 }
1482
1483 return offset < len ? offset : U32_MAX;
1484}
1485
1486static int set_oa_ctx_ctrl_offset(struct intel_context *ce)
1487{
1488 i915_reg_t reg = GEN12_OACTXCONTROL(ce->engine->mmio_base);
1489 struct i915_perf *perf = &ce->engine->i915->perf;
1490 u32 offset = perf->ctx_oactxctrl_offset;
1491
1492 /* Do this only once. Failure is stored as offset of U32_MAX */
1493 if (offset)
1494 goto exit;
1495
1496 offset = oa_context_image_offset(ce, i915_mmio_reg_offset(reg));
1497 perf->ctx_oactxctrl_offset = offset;
1498
1499 drm_dbg(&ce->engine->i915->drm,
1500 "%s oa ctx control at 0x%08x dword offset\n",
1501 ce->engine->name, offset);
1502
1503exit:
1504 return offset && offset != U32_MAX ? 0 : -ENODEV;
1505}
1506
1507static bool engine_supports_mi_query(struct intel_engine_cs *engine)
1508{
1509 return engine->class == RENDER_CLASS;
1510}
1511
1512/**
1513 * oa_get_render_ctx_id - determine and hold ctx hw id
1514 * @stream: An i915-perf stream opened for OA metrics
1515 *
1516 * Determine the render context hw id, and ensure it remains fixed for the
1517 * lifetime of the stream. This ensures that we don't have to worry about
1518 * updating the context ID in OACONTROL on the fly.
1519 *
1520 * Returns: zero on success or a negative error code
1521 */
1522static int oa_get_render_ctx_id(struct i915_perf_stream *stream)
1523{
1524 struct intel_context *ce;
1525 int ret = 0;
1526
1527 ce = oa_pin_context(stream);
1528 if (IS_ERR(ce))
1529 return PTR_ERR(ce);
1530
1531 if (engine_supports_mi_query(stream->engine) &&
1532 HAS_LOGICAL_RING_CONTEXTS(stream->perf->i915)) {
1533 /*
1534 * We are enabling perf query here. If we don't find the context
1535 * offset here, just return an error.
1536 */
1537 ret = set_oa_ctx_ctrl_offset(ce);
1538 if (ret) {
1539 intel_context_unpin(ce);
1540 drm_err(&stream->perf->i915->drm,
1541 "Enabling perf query failed for %s\n",
1542 stream->engine->name);
1543 return ret;
1544 }
1545 }
1546
1547 switch (GRAPHICS_VER(ce->engine->i915)) {
1548 case 7: {
1549 /*
1550 * On Haswell we don't do any post processing of the reports
1551 * and don't need to use the mask.
1552 */
1553 stream->specific_ctx_id = i915_ggtt_offset(ce->state);
1554 stream->specific_ctx_id_mask = 0;
1555 break;
1556 }
1557
1558 case 8:
1559 case 9:
1560 if (intel_engine_uses_guc(ce->engine)) {
1561 /*
1562 * When using GuC, the context descriptor we write in
1563 * i915 is read by GuC and rewritten before it's
1564 * actually written into the hardware. The LRCA is
1565 * what is put into the context id field of the
1566 * context descriptor by GuC. Because it's aligned to
1567 * a page, the lower 12bits are always at 0 and
1568 * dropped by GuC. They won't be part of the context
1569 * ID in the OA reports, so squash those lower bits.
1570 */
1571 stream->specific_ctx_id = ce->lrc.lrca >> 12;
1572
1573 /*
1574 * GuC uses the top bit to signal proxy submission, so
1575 * ignore that bit.
1576 */
1577 stream->specific_ctx_id_mask =
1578 (1U << (GEN8_CTX_ID_WIDTH - 1)) - 1;
1579 } else {
1580 stream->specific_ctx_id_mask =
1581 (1U << GEN8_CTX_ID_WIDTH) - 1;
1582 stream->specific_ctx_id = stream->specific_ctx_id_mask;
1583 }
1584 break;
1585
1586 case 11:
1587 case 12:
1588 ret = gen12_get_render_context_id(stream);
1589 break;
1590
1591 default:
1592 MISSING_CASE(GRAPHICS_VER(ce->engine->i915));
1593 }
1594
1595 ce->tag = stream->specific_ctx_id;
1596
1597 drm_dbg(&stream->perf->i915->drm,
1598 "filtering on ctx_id=0x%x ctx_id_mask=0x%x\n",
1599 stream->specific_ctx_id,
1600 stream->specific_ctx_id_mask);
1601
1602 return ret;
1603}
1604
1605/**
1606 * oa_put_render_ctx_id - counterpart to oa_get_render_ctx_id releases hold
1607 * @stream: An i915-perf stream opened for OA metrics
1608 *
1609 * In case anything needed doing to ensure the context HW ID would remain valid
1610 * for the lifetime of the stream, then that can be undone here.
1611 */
1612static void oa_put_render_ctx_id(struct i915_perf_stream *stream)
1613{
1614 struct intel_context *ce;
1615
1616 ce = fetch_and_zero(&stream->pinned_ctx);
1617 if (ce) {
1618 ce->tag = 0; /* recomputed on next submission after parking */
1619 intel_context_unpin(ce);
1620 }
1621
1622 stream->specific_ctx_id = INVALID_CTX_ID;
1623 stream->specific_ctx_id_mask = 0;
1624}
1625
1626static void
1627free_oa_buffer(struct i915_perf_stream *stream)
1628{
1629 i915_vma_unpin_and_release(&stream->oa_buffer.vma,
1630 I915_VMA_RELEASE_MAP);
1631
1632 stream->oa_buffer.vaddr = NULL;
1633}
1634
1635static void
1636free_oa_configs(struct i915_perf_stream *stream)
1637{
1638 struct i915_oa_config_bo *oa_bo, *tmp;
1639
1640 i915_oa_config_put(stream->oa_config);
1641 llist_for_each_entry_safe(oa_bo, tmp, stream->oa_config_bos.first, node)
1642 free_oa_config_bo(oa_bo);
1643}
1644
1645static void
1646free_noa_wait(struct i915_perf_stream *stream)
1647{
1648 i915_vma_unpin_and_release(&stream->noa_wait, 0);
1649}
1650
1651static bool engine_supports_oa(const struct intel_engine_cs *engine)
1652{
1653 return engine->oa_group;
1654}
1655
1656static bool engine_supports_oa_format(struct intel_engine_cs *engine, int type)
1657{
1658 return engine->oa_group && engine->oa_group->type == type;
1659}
1660
1661static void i915_oa_stream_destroy(struct i915_perf_stream *stream)
1662{
1663 struct i915_perf *perf = stream->perf;
1664 struct intel_gt *gt = stream->engine->gt;
1665 struct i915_perf_group *g = stream->engine->oa_group;
1666
1667 if (WARN_ON(stream != g->exclusive_stream))
1668 return;
1669
1670 /*
1671 * Unset exclusive_stream first, it will be checked while disabling
1672 * the metric set on gen8+.
1673 *
1674 * See i915_oa_init_reg_state() and lrc_configure_all_contexts()
1675 */
1676 WRITE_ONCE(g->exclusive_stream, NULL);
1677 perf->ops.disable_metric_set(stream);
1678
1679 free_oa_buffer(stream);
1680
1681 intel_uncore_forcewake_put(stream->uncore, FORCEWAKE_ALL);
1682 intel_engine_pm_put(stream->engine);
1683
1684 if (stream->ctx)
1685 oa_put_render_ctx_id(stream);
1686
1687 free_oa_configs(stream);
1688 free_noa_wait(stream);
1689
1690 if (perf->spurious_report_rs.missed) {
1691 gt_notice(gt, "%d spurious OA report notices suppressed due to ratelimiting\n",
1692 perf->spurious_report_rs.missed);
1693 }
1694}
1695
1696static void gen7_init_oa_buffer(struct i915_perf_stream *stream)
1697{
1698 struct intel_uncore *uncore = stream->uncore;
1699 u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma);
1700 unsigned long flags;
1701
1702 spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
1703
1704 /* Pre-DevBDW: OABUFFER must be set with counters off,
1705 * before OASTATUS1, but after OASTATUS2
1706 */
1707 intel_uncore_write(uncore, GEN7_OASTATUS2, /* head */
1708 gtt_offset | GEN7_OASTATUS2_MEM_SELECT_GGTT);
1709 stream->oa_buffer.head = 0;
1710
1711 intel_uncore_write(uncore, GEN7_OABUFFER, gtt_offset);
1712
1713 intel_uncore_write(uncore, GEN7_OASTATUS1, /* tail */
1714 gtt_offset | OABUFFER_SIZE_16M);
1715
1716 /* Mark that we need updated tail pointers to read from... */
1717 stream->oa_buffer.tail = 0;
1718
1719 spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
1720
1721 /* On Haswell we have to track which OASTATUS1 flags we've
1722 * already seen since they can't be cleared while periodic
1723 * sampling is enabled.
1724 */
1725 stream->perf->gen7_latched_oastatus1 = 0;
1726
1727 /* NB: although the OA buffer will initially be allocated
1728 * zeroed via shmfs (and so this memset is redundant when
1729 * first allocating), we may re-init the OA buffer, either
1730 * when re-enabling a stream or in error/reset paths.
1731 *
1732 * The reason we clear the buffer for each re-init is for the
1733 * sanity check in gen7_append_oa_reports() that looks at the
1734 * report-id field to make sure it's non-zero which relies on
1735 * the assumption that new reports are being written to zeroed
1736 * memory...
1737 */
1738 memset(stream->oa_buffer.vaddr, 0, OA_BUFFER_SIZE);
1739}
1740
1741static void gen8_init_oa_buffer(struct i915_perf_stream *stream)
1742{
1743 struct intel_uncore *uncore = stream->uncore;
1744 u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma);
1745 unsigned long flags;
1746
1747 spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
1748
1749 intel_uncore_write(uncore, GEN8_OASTATUS, 0);
1750 intel_uncore_write(uncore, GEN8_OAHEADPTR, gtt_offset);
1751 stream->oa_buffer.head = 0;
1752
1753 intel_uncore_write(uncore, GEN8_OABUFFER_UDW, 0);
1754
1755 /*
1756 * PRM says:
1757 *
1758 * "This MMIO must be set before the OATAILPTR
1759 * register and after the OAHEADPTR register. This is
1760 * to enable proper functionality of the overflow
1761 * bit."
1762 */
1763 intel_uncore_write(uncore, GEN8_OABUFFER, gtt_offset |
1764 OABUFFER_SIZE_16M | GEN8_OABUFFER_MEM_SELECT_GGTT);
1765 intel_uncore_write(uncore, GEN8_OATAILPTR, gtt_offset & GEN8_OATAILPTR_MASK);
1766
1767 /* Mark that we need updated tail pointers to read from... */
1768 stream->oa_buffer.tail = 0;
1769
1770 /*
1771 * Reset state used to recognise context switches, affecting which
1772 * reports we will forward to userspace while filtering for a single
1773 * context.
1774 */
1775 stream->oa_buffer.last_ctx_id = INVALID_CTX_ID;
1776
1777 spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
1778
1779 /*
1780 * NB: although the OA buffer will initially be allocated
1781 * zeroed via shmfs (and so this memset is redundant when
1782 * first allocating), we may re-init the OA buffer, either
1783 * when re-enabling a stream or in error/reset paths.
1784 *
1785 * The reason we clear the buffer for each re-init is for the
1786 * sanity check in gen8_append_oa_reports() that looks at the
1787 * reason field to make sure it's non-zero which relies on
1788 * the assumption that new reports are being written to zeroed
1789 * memory...
1790 */
1791 memset(stream->oa_buffer.vaddr, 0, OA_BUFFER_SIZE);
1792}
1793
1794static void gen12_init_oa_buffer(struct i915_perf_stream *stream)
1795{
1796 struct intel_uncore *uncore = stream->uncore;
1797 u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma);
1798 unsigned long flags;
1799
1800 spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
1801
1802 intel_uncore_write(uncore, __oa_regs(stream)->oa_status, 0);
1803 intel_uncore_write(uncore, __oa_regs(stream)->oa_head_ptr,
1804 gtt_offset & GEN12_OAG_OAHEADPTR_MASK);
1805 stream->oa_buffer.head = 0;
1806
1807 /*
1808 * PRM says:
1809 *
1810 * "This MMIO must be set before the OATAILPTR
1811 * register and after the OAHEADPTR register. This is
1812 * to enable proper functionality of the overflow
1813 * bit."
1814 */
1815 intel_uncore_write(uncore, __oa_regs(stream)->oa_buffer, gtt_offset |
1816 OABUFFER_SIZE_16M | GEN8_OABUFFER_MEM_SELECT_GGTT);
1817 intel_uncore_write(uncore, __oa_regs(stream)->oa_tail_ptr,
1818 gtt_offset & GEN12_OAG_OATAILPTR_MASK);
1819
1820 /* Mark that we need updated tail pointers to read from... */
1821 stream->oa_buffer.tail = 0;
1822
1823 /*
1824 * Reset state used to recognise context switches, affecting which
1825 * reports we will forward to userspace while filtering for a single
1826 * context.
1827 */
1828 stream->oa_buffer.last_ctx_id = INVALID_CTX_ID;
1829
1830 spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
1831
1832 /*
1833 * NB: although the OA buffer will initially be allocated
1834 * zeroed via shmfs (and so this memset is redundant when
1835 * first allocating), we may re-init the OA buffer, either
1836 * when re-enabling a stream or in error/reset paths.
1837 *
1838 * The reason we clear the buffer for each re-init is for the
1839 * sanity check in gen8_append_oa_reports() that looks at the
1840 * reason field to make sure it's non-zero which relies on
1841 * the assumption that new reports are being written to zeroed
1842 * memory...
1843 */
1844 memset(stream->oa_buffer.vaddr, 0,
1845 stream->oa_buffer.vma->size);
1846}
1847
1848static int alloc_oa_buffer(struct i915_perf_stream *stream)
1849{
1850 struct drm_i915_private *i915 = stream->perf->i915;
1851 struct intel_gt *gt = stream->engine->gt;
1852 struct drm_i915_gem_object *bo;
1853 struct i915_vma *vma;
1854 int ret;
1855
1856 if (drm_WARN_ON(&i915->drm, stream->oa_buffer.vma))
1857 return -ENODEV;
1858
1859 BUILD_BUG_ON_NOT_POWER_OF_2(OA_BUFFER_SIZE);
1860 BUILD_BUG_ON(OA_BUFFER_SIZE < SZ_128K || OA_BUFFER_SIZE > SZ_16M);
1861
1862 bo = i915_gem_object_create_shmem(stream->perf->i915, OA_BUFFER_SIZE);
1863 if (IS_ERR(bo)) {
1864 drm_err(&i915->drm, "Failed to allocate OA buffer\n");
1865 return PTR_ERR(bo);
1866 }
1867
1868 i915_gem_object_set_cache_coherency(bo, I915_CACHE_LLC);
1869
1870 /* PreHSW required 512K alignment, HSW requires 16M */
1871 vma = i915_vma_instance(bo, >->ggtt->vm, NULL);
1872 if (IS_ERR(vma)) {
1873 ret = PTR_ERR(vma);
1874 goto err_unref;
1875 }
1876
1877 /*
1878 * PreHSW required 512K alignment.
1879 * HSW and onwards, align to requested size of OA buffer.
1880 */
1881 ret = i915_vma_pin(vma, 0, SZ_16M, PIN_GLOBAL | PIN_HIGH);
1882 if (ret) {
1883 gt_err(gt, "Failed to pin OA buffer %d\n", ret);
1884 goto err_unref;
1885 }
1886
1887 stream->oa_buffer.vma = vma;
1888
1889 stream->oa_buffer.vaddr =
1890 i915_gem_object_pin_map_unlocked(bo, I915_MAP_WB);
1891 if (IS_ERR(stream->oa_buffer.vaddr)) {
1892 ret = PTR_ERR(stream->oa_buffer.vaddr);
1893 goto err_unpin;
1894 }
1895
1896 return 0;
1897
1898err_unpin:
1899 __i915_vma_unpin(vma);
1900
1901err_unref:
1902 i915_gem_object_put(bo);
1903
1904 stream->oa_buffer.vaddr = NULL;
1905 stream->oa_buffer.vma = NULL;
1906
1907 return ret;
1908}
1909
1910static u32 *save_restore_register(struct i915_perf_stream *stream, u32 *cs,
1911 bool save, i915_reg_t reg, u32 offset,
1912 u32 dword_count)
1913{
1914 u32 cmd;
1915 u32 d;
1916
1917 cmd = save ? MI_STORE_REGISTER_MEM : MI_LOAD_REGISTER_MEM;
1918 cmd |= MI_SRM_LRM_GLOBAL_GTT;
1919 if (GRAPHICS_VER(stream->perf->i915) >= 8)
1920 cmd++;
1921
1922 for (d = 0; d < dword_count; d++) {
1923 *cs++ = cmd;
1924 *cs++ = i915_mmio_reg_offset(reg) + 4 * d;
1925 *cs++ = i915_ggtt_offset(stream->noa_wait) + offset + 4 * d;
1926 *cs++ = 0;
1927 }
1928
1929 return cs;
1930}
1931
1932static int alloc_noa_wait(struct i915_perf_stream *stream)
1933{
1934 struct drm_i915_private *i915 = stream->perf->i915;
1935 struct intel_gt *gt = stream->engine->gt;
1936 struct drm_i915_gem_object *bo;
1937 struct i915_vma *vma;
1938 const u64 delay_ticks = 0xffffffffffffffff -
1939 intel_gt_ns_to_clock_interval(to_gt(stream->perf->i915),
1940 atomic64_read(&stream->perf->noa_programming_delay));
1941 const u32 base = stream->engine->mmio_base;
1942#define CS_GPR(x) GEN8_RING_CS_GPR(base, x)
1943 u32 *batch, *ts0, *cs, *jump;
1944 struct i915_gem_ww_ctx ww;
1945 int ret, i;
1946 enum {
1947 START_TS,
1948 NOW_TS,
1949 DELTA_TS,
1950 JUMP_PREDICATE,
1951 DELTA_TARGET,
1952 N_CS_GPR
1953 };
1954 i915_reg_t mi_predicate_result = HAS_MI_SET_PREDICATE(i915) ?
1955 MI_PREDICATE_RESULT_2_ENGINE(base) :
1956 MI_PREDICATE_RESULT_1(RENDER_RING_BASE);
1957
1958 /*
1959 * gt->scratch was being used to save/restore the GPR registers, but on
1960 * MTL the scratch uses stolen lmem. An MI_SRM to this memory region
1961 * causes an engine hang. Instead allocate an additional page here to
1962 * save/restore GPR registers
1963 */
1964 bo = i915_gem_object_create_internal(i915, 8192);
1965 if (IS_ERR(bo)) {
1966 drm_err(&i915->drm,
1967 "Failed to allocate NOA wait batchbuffer\n");
1968 return PTR_ERR(bo);
1969 }
1970
1971 i915_gem_ww_ctx_init(&ww, true);
1972retry:
1973 ret = i915_gem_object_lock(bo, &ww);
1974 if (ret)
1975 goto out_ww;
1976
1977 /*
1978 * We pin in GGTT because we jump into this buffer now because
1979 * multiple OA config BOs will have a jump to this address and it
1980 * needs to be fixed during the lifetime of the i915/perf stream.
1981 */
1982 vma = i915_vma_instance(bo, >->ggtt->vm, NULL);
1983 if (IS_ERR(vma)) {
1984 ret = PTR_ERR(vma);
1985 goto out_ww;
1986 }
1987
1988 ret = i915_vma_pin_ww(vma, &ww, 0, 0, PIN_GLOBAL | PIN_HIGH);
1989 if (ret)
1990 goto out_ww;
1991
1992 batch = cs = i915_gem_object_pin_map(bo, I915_MAP_WB);
1993 if (IS_ERR(batch)) {
1994 ret = PTR_ERR(batch);
1995 goto err_unpin;
1996 }
1997
1998 stream->noa_wait = vma;
1999
2000#define GPR_SAVE_OFFSET 4096
2001#define PREDICATE_SAVE_OFFSET 4160
2002
2003 /* Save registers. */
2004 for (i = 0; i < N_CS_GPR; i++)
2005 cs = save_restore_register(
2006 stream, cs, true /* save */, CS_GPR(i),
2007 GPR_SAVE_OFFSET + 8 * i, 2);
2008 cs = save_restore_register(
2009 stream, cs, true /* save */, mi_predicate_result,
2010 PREDICATE_SAVE_OFFSET, 1);
2011
2012 /* First timestamp snapshot location. */
2013 ts0 = cs;
2014
2015 /*
2016 * Initial snapshot of the timestamp register to implement the wait.
2017 * We work with 32b values, so clear out the top 32b bits of the
2018 * register because the ALU works 64bits.
2019 */
2020 *cs++ = MI_LOAD_REGISTER_IMM(1);
2021 *cs++ = i915_mmio_reg_offset(CS_GPR(START_TS)) + 4;
2022 *cs++ = 0;
2023 *cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
2024 *cs++ = i915_mmio_reg_offset(RING_TIMESTAMP(base));
2025 *cs++ = i915_mmio_reg_offset(CS_GPR(START_TS));
2026
2027 /*
2028 * This is the location we're going to jump back into until the
2029 * required amount of time has passed.
2030 */
2031 jump = cs;
2032
2033 /*
2034 * Take another snapshot of the timestamp register. Take care to clear
2035 * up the top 32bits of CS_GPR(1) as we're using it for other
2036 * operations below.
2037 */
2038 *cs++ = MI_LOAD_REGISTER_IMM(1);
2039 *cs++ = i915_mmio_reg_offset(CS_GPR(NOW_TS)) + 4;
2040 *cs++ = 0;
2041 *cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
2042 *cs++ = i915_mmio_reg_offset(RING_TIMESTAMP(base));
2043 *cs++ = i915_mmio_reg_offset(CS_GPR(NOW_TS));
2044
2045 /*
2046 * Do a diff between the 2 timestamps and store the result back into
2047 * CS_GPR(1).
2048 */
2049 *cs++ = MI_MATH(5);
2050 *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(NOW_TS));
2051 *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(START_TS));
2052 *cs++ = MI_MATH_SUB;
2053 *cs++ = MI_MATH_STORE(MI_MATH_REG(DELTA_TS), MI_MATH_REG_ACCU);
2054 *cs++ = MI_MATH_STORE(MI_MATH_REG(JUMP_PREDICATE), MI_MATH_REG_CF);
2055
2056 /*
2057 * Transfer the carry flag (set to 1 if ts1 < ts0, meaning the
2058 * timestamp have rolled over the 32bits) into the predicate register
2059 * to be used for the predicated jump.
2060 */
2061 *cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
2062 *cs++ = i915_mmio_reg_offset(CS_GPR(JUMP_PREDICATE));
2063 *cs++ = i915_mmio_reg_offset(mi_predicate_result);
2064
2065 if (HAS_MI_SET_PREDICATE(i915))
2066 *cs++ = MI_SET_PREDICATE | 1;
2067
2068 /* Restart from the beginning if we had timestamps roll over. */
2069 *cs++ = (GRAPHICS_VER(i915) < 8 ?
2070 MI_BATCH_BUFFER_START :
2071 MI_BATCH_BUFFER_START_GEN8) |
2072 MI_BATCH_PREDICATE;
2073 *cs++ = i915_ggtt_offset(vma) + (ts0 - batch) * 4;
2074 *cs++ = 0;
2075
2076 if (HAS_MI_SET_PREDICATE(i915))
2077 *cs++ = MI_SET_PREDICATE;
2078
2079 /*
2080 * Now add the diff between to previous timestamps and add it to :
2081 * (((1 * << 64) - 1) - delay_ns)
2082 *
2083 * When the Carry Flag contains 1 this means the elapsed time is
2084 * longer than the expected delay, and we can exit the wait loop.
2085 */
2086 *cs++ = MI_LOAD_REGISTER_IMM(2);
2087 *cs++ = i915_mmio_reg_offset(CS_GPR(DELTA_TARGET));
2088 *cs++ = lower_32_bits(delay_ticks);
2089 *cs++ = i915_mmio_reg_offset(CS_GPR(DELTA_TARGET)) + 4;
2090 *cs++ = upper_32_bits(delay_ticks);
2091
2092 *cs++ = MI_MATH(4);
2093 *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(DELTA_TS));
2094 *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(DELTA_TARGET));
2095 *cs++ = MI_MATH_ADD;
2096 *cs++ = MI_MATH_STOREINV(MI_MATH_REG(JUMP_PREDICATE), MI_MATH_REG_CF);
2097
2098 *cs++ = MI_ARB_CHECK;
2099
2100 /*
2101 * Transfer the result into the predicate register to be used for the
2102 * predicated jump.
2103 */
2104 *cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
2105 *cs++ = i915_mmio_reg_offset(CS_GPR(JUMP_PREDICATE));
2106 *cs++ = i915_mmio_reg_offset(mi_predicate_result);
2107
2108 if (HAS_MI_SET_PREDICATE(i915))
2109 *cs++ = MI_SET_PREDICATE | 1;
2110
2111 /* Predicate the jump. */
2112 *cs++ = (GRAPHICS_VER(i915) < 8 ?
2113 MI_BATCH_BUFFER_START :
2114 MI_BATCH_BUFFER_START_GEN8) |
2115 MI_BATCH_PREDICATE;
2116 *cs++ = i915_ggtt_offset(vma) + (jump - batch) * 4;
2117 *cs++ = 0;
2118
2119 if (HAS_MI_SET_PREDICATE(i915))
2120 *cs++ = MI_SET_PREDICATE;
2121
2122 /* Restore registers. */
2123 for (i = 0; i < N_CS_GPR; i++)
2124 cs = save_restore_register(
2125 stream, cs, false /* restore */, CS_GPR(i),
2126 GPR_SAVE_OFFSET + 8 * i, 2);
2127 cs = save_restore_register(
2128 stream, cs, false /* restore */, mi_predicate_result,
2129 PREDICATE_SAVE_OFFSET, 1);
2130
2131 /* And return to the ring. */
2132 *cs++ = MI_BATCH_BUFFER_END;
2133
2134 GEM_BUG_ON(cs - batch > PAGE_SIZE / sizeof(*batch));
2135
2136 i915_gem_object_flush_map(bo);
2137 __i915_gem_object_release_map(bo);
2138
2139 goto out_ww;
2140
2141err_unpin:
2142 i915_vma_unpin_and_release(&vma, 0);
2143out_ww:
2144 if (ret == -EDEADLK) {
2145 ret = i915_gem_ww_ctx_backoff(&ww);
2146 if (!ret)
2147 goto retry;
2148 }
2149 i915_gem_ww_ctx_fini(&ww);
2150 if (ret)
2151 i915_gem_object_put(bo);
2152 return ret;
2153}
2154
2155static u32 *write_cs_mi_lri(u32 *cs,
2156 const struct i915_oa_reg *reg_data,
2157 u32 n_regs)
2158{
2159 u32 i;
2160
2161 for (i = 0; i < n_regs; i++) {
2162 if ((i % MI_LOAD_REGISTER_IMM_MAX_REGS) == 0) {
2163 u32 n_lri = min_t(u32,
2164 n_regs - i,
2165 MI_LOAD_REGISTER_IMM_MAX_REGS);
2166
2167 *cs++ = MI_LOAD_REGISTER_IMM(n_lri);
2168 }
2169 *cs++ = i915_mmio_reg_offset(reg_data[i].addr);
2170 *cs++ = reg_data[i].value;
2171 }
2172
2173 return cs;
2174}
2175
2176static int num_lri_dwords(int num_regs)
2177{
2178 int count = 0;
2179
2180 if (num_regs > 0) {
2181 count += DIV_ROUND_UP(num_regs, MI_LOAD_REGISTER_IMM_MAX_REGS);
2182 count += num_regs * 2;
2183 }
2184
2185 return count;
2186}
2187
2188static struct i915_oa_config_bo *
2189alloc_oa_config_buffer(struct i915_perf_stream *stream,
2190 struct i915_oa_config *oa_config)
2191{
2192 struct drm_i915_gem_object *obj;
2193 struct i915_oa_config_bo *oa_bo;
2194 struct i915_gem_ww_ctx ww;
2195 size_t config_length = 0;
2196 u32 *cs;
2197 int err;
2198
2199 oa_bo = kzalloc(sizeof(*oa_bo), GFP_KERNEL);
2200 if (!oa_bo)
2201 return ERR_PTR(-ENOMEM);
2202
2203 config_length += num_lri_dwords(oa_config->mux_regs_len);
2204 config_length += num_lri_dwords(oa_config->b_counter_regs_len);
2205 config_length += num_lri_dwords(oa_config->flex_regs_len);
2206 config_length += 3; /* MI_BATCH_BUFFER_START */
2207 config_length = ALIGN(sizeof(u32) * config_length, I915_GTT_PAGE_SIZE);
2208
2209 obj = i915_gem_object_create_shmem(stream->perf->i915, config_length);
2210 if (IS_ERR(obj)) {
2211 err = PTR_ERR(obj);
2212 goto err_free;
2213 }
2214
2215 i915_gem_ww_ctx_init(&ww, true);
2216retry:
2217 err = i915_gem_object_lock(obj, &ww);
2218 if (err)
2219 goto out_ww;
2220
2221 cs = i915_gem_object_pin_map(obj, I915_MAP_WB);
2222 if (IS_ERR(cs)) {
2223 err = PTR_ERR(cs);
2224 goto out_ww;
2225 }
2226
2227 cs = write_cs_mi_lri(cs,
2228 oa_config->mux_regs,
2229 oa_config->mux_regs_len);
2230 cs = write_cs_mi_lri(cs,
2231 oa_config->b_counter_regs,
2232 oa_config->b_counter_regs_len);
2233 cs = write_cs_mi_lri(cs,
2234 oa_config->flex_regs,
2235 oa_config->flex_regs_len);
2236
2237 /* Jump into the active wait. */
2238 *cs++ = (GRAPHICS_VER(stream->perf->i915) < 8 ?
2239 MI_BATCH_BUFFER_START :
2240 MI_BATCH_BUFFER_START_GEN8);
2241 *cs++ = i915_ggtt_offset(stream->noa_wait);
2242 *cs++ = 0;
2243
2244 i915_gem_object_flush_map(obj);
2245 __i915_gem_object_release_map(obj);
2246
2247 oa_bo->vma = i915_vma_instance(obj,
2248 &stream->engine->gt->ggtt->vm,
2249 NULL);
2250 if (IS_ERR(oa_bo->vma)) {
2251 err = PTR_ERR(oa_bo->vma);
2252 goto out_ww;
2253 }
2254
2255 oa_bo->oa_config = i915_oa_config_get(oa_config);
2256 llist_add(&oa_bo->node, &stream->oa_config_bos);
2257
2258out_ww:
2259 if (err == -EDEADLK) {
2260 err = i915_gem_ww_ctx_backoff(&ww);
2261 if (!err)
2262 goto retry;
2263 }
2264 i915_gem_ww_ctx_fini(&ww);
2265
2266 if (err)
2267 i915_gem_object_put(obj);
2268err_free:
2269 if (err) {
2270 kfree(oa_bo);
2271 return ERR_PTR(err);
2272 }
2273 return oa_bo;
2274}
2275
2276static struct i915_vma *
2277get_oa_vma(struct i915_perf_stream *stream, struct i915_oa_config *oa_config)
2278{
2279 struct i915_oa_config_bo *oa_bo;
2280
2281 /*
2282 * Look for the buffer in the already allocated BOs attached
2283 * to the stream.
2284 */
2285 llist_for_each_entry(oa_bo, stream->oa_config_bos.first, node) {
2286 if (oa_bo->oa_config == oa_config &&
2287 memcmp(oa_bo->oa_config->uuid,
2288 oa_config->uuid,
2289 sizeof(oa_config->uuid)) == 0)
2290 goto out;
2291 }
2292
2293 oa_bo = alloc_oa_config_buffer(stream, oa_config);
2294 if (IS_ERR(oa_bo))
2295 return ERR_CAST(oa_bo);
2296
2297out:
2298 return i915_vma_get(oa_bo->vma);
2299}
2300
2301static int
2302emit_oa_config(struct i915_perf_stream *stream,
2303 struct i915_oa_config *oa_config,
2304 struct intel_context *ce,
2305 struct i915_active *active)
2306{
2307 struct i915_request *rq;
2308 struct i915_vma *vma;
2309 struct i915_gem_ww_ctx ww;
2310 int err;
2311
2312 vma = get_oa_vma(stream, oa_config);
2313 if (IS_ERR(vma))
2314 return PTR_ERR(vma);
2315
2316 i915_gem_ww_ctx_init(&ww, true);
2317retry:
2318 err = i915_gem_object_lock(vma->obj, &ww);
2319 if (err)
2320 goto err;
2321
2322 err = i915_vma_pin_ww(vma, &ww, 0, 0, PIN_GLOBAL | PIN_HIGH);
2323 if (err)
2324 goto err;
2325
2326 intel_engine_pm_get(ce->engine);
2327 rq = i915_request_create(ce);
2328 intel_engine_pm_put(ce->engine);
2329 if (IS_ERR(rq)) {
2330 err = PTR_ERR(rq);
2331 goto err_vma_unpin;
2332 }
2333
2334 if (!IS_ERR_OR_NULL(active)) {
2335 /* After all individual context modifications */
2336 err = i915_request_await_active(rq, active,
2337 I915_ACTIVE_AWAIT_ACTIVE);
2338 if (err)
2339 goto err_add_request;
2340
2341 err = i915_active_add_request(active, rq);
2342 if (err)
2343 goto err_add_request;
2344 }
2345
2346 err = i915_vma_move_to_active(vma, rq, 0);
2347 if (err)
2348 goto err_add_request;
2349
2350 err = rq->engine->emit_bb_start(rq,
2351 i915_vma_offset(vma), 0,
2352 I915_DISPATCH_SECURE);
2353 if (err)
2354 goto err_add_request;
2355
2356err_add_request:
2357 i915_request_add(rq);
2358err_vma_unpin:
2359 i915_vma_unpin(vma);
2360err:
2361 if (err == -EDEADLK) {
2362 err = i915_gem_ww_ctx_backoff(&ww);
2363 if (!err)
2364 goto retry;
2365 }
2366
2367 i915_gem_ww_ctx_fini(&ww);
2368 i915_vma_put(vma);
2369 return err;
2370}
2371
2372static struct intel_context *oa_context(struct i915_perf_stream *stream)
2373{
2374 return stream->pinned_ctx ?: stream->engine->kernel_context;
2375}
2376
2377static int
2378hsw_enable_metric_set(struct i915_perf_stream *stream,
2379 struct i915_active *active)
2380{
2381 struct intel_uncore *uncore = stream->uncore;
2382
2383 /*
2384 * PRM:
2385 *
2386 * OA unit is using “crclk” for its functionality. When trunk
2387 * level clock gating takes place, OA clock would be gated,
2388 * unable to count the events from non-render clock domain.
2389 * Render clock gating must be disabled when OA is enabled to
2390 * count the events from non-render domain. Unit level clock
2391 * gating for RCS should also be disabled.
2392 */
2393 intel_uncore_rmw(uncore, GEN7_MISCCPCTL,
2394 GEN7_DOP_CLOCK_GATE_ENABLE, 0);
2395 intel_uncore_rmw(uncore, GEN6_UCGCTL1,
2396 0, GEN6_CSUNIT_CLOCK_GATE_DISABLE);
2397
2398 return emit_oa_config(stream,
2399 stream->oa_config, oa_context(stream),
2400 active);
2401}
2402
2403static void hsw_disable_metric_set(struct i915_perf_stream *stream)
2404{
2405 struct intel_uncore *uncore = stream->uncore;
2406
2407 intel_uncore_rmw(uncore, GEN6_UCGCTL1,
2408 GEN6_CSUNIT_CLOCK_GATE_DISABLE, 0);
2409 intel_uncore_rmw(uncore, GEN7_MISCCPCTL,
2410 0, GEN7_DOP_CLOCK_GATE_ENABLE);
2411
2412 intel_uncore_rmw(uncore, GDT_CHICKEN_BITS, GT_NOA_ENABLE, 0);
2413}
2414
2415static u32 oa_config_flex_reg(const struct i915_oa_config *oa_config,
2416 i915_reg_t reg)
2417{
2418 u32 mmio = i915_mmio_reg_offset(reg);
2419 int i;
2420
2421 /*
2422 * This arbitrary default will select the 'EU FPU0 Pipeline
2423 * Active' event. In the future it's anticipated that there
2424 * will be an explicit 'No Event' we can select, but not yet...
2425 */
2426 if (!oa_config)
2427 return 0;
2428
2429 for (i = 0; i < oa_config->flex_regs_len; i++) {
2430 if (i915_mmio_reg_offset(oa_config->flex_regs[i].addr) == mmio)
2431 return oa_config->flex_regs[i].value;
2432 }
2433
2434 return 0;
2435}
2436/*
2437 * NB: It must always remain pointer safe to run this even if the OA unit
2438 * has been disabled.
2439 *
2440 * It's fine to put out-of-date values into these per-context registers
2441 * in the case that the OA unit has been disabled.
2442 */
2443static void
2444gen8_update_reg_state_unlocked(const struct intel_context *ce,
2445 const struct i915_perf_stream *stream)
2446{
2447 u32 ctx_oactxctrl = stream->perf->ctx_oactxctrl_offset;
2448 u32 ctx_flexeu0 = stream->perf->ctx_flexeu0_offset;
2449 /* The MMIO offsets for Flex EU registers aren't contiguous */
2450 static const i915_reg_t flex_regs[] = {
2451 EU_PERF_CNTL0,
2452 EU_PERF_CNTL1,
2453 EU_PERF_CNTL2,
2454 EU_PERF_CNTL3,
2455 EU_PERF_CNTL4,
2456 EU_PERF_CNTL5,
2457 EU_PERF_CNTL6,
2458 };
2459 u32 *reg_state = ce->lrc_reg_state;
2460 int i;
2461
2462 reg_state[ctx_oactxctrl + 1] =
2463 (stream->period_exponent << GEN8_OA_TIMER_PERIOD_SHIFT) |
2464 (stream->periodic ? GEN8_OA_TIMER_ENABLE : 0) |
2465 GEN8_OA_COUNTER_RESUME;
2466
2467 for (i = 0; i < ARRAY_SIZE(flex_regs); i++)
2468 reg_state[ctx_flexeu0 + i * 2 + 1] =
2469 oa_config_flex_reg(stream->oa_config, flex_regs[i]);
2470}
2471
2472struct flex {
2473 i915_reg_t reg;
2474 u32 offset;
2475 u32 value;
2476};
2477
2478static int
2479gen8_store_flex(struct i915_request *rq,
2480 struct intel_context *ce,
2481 const struct flex *flex, unsigned int count)
2482{
2483 u32 offset;
2484 u32 *cs;
2485
2486 cs = intel_ring_begin(rq, 4 * count);
2487 if (IS_ERR(cs))
2488 return PTR_ERR(cs);
2489
2490 offset = i915_ggtt_offset(ce->state) + LRC_STATE_OFFSET;
2491 do {
2492 *cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
2493 *cs++ = offset + flex->offset * sizeof(u32);
2494 *cs++ = 0;
2495 *cs++ = flex->value;
2496 } while (flex++, --count);
2497
2498 intel_ring_advance(rq, cs);
2499
2500 return 0;
2501}
2502
2503static int
2504gen8_load_flex(struct i915_request *rq,
2505 struct intel_context *ce,
2506 const struct flex *flex, unsigned int count)
2507{
2508 u32 *cs;
2509
2510 GEM_BUG_ON(!count || count > 63);
2511
2512 cs = intel_ring_begin(rq, 2 * count + 2);
2513 if (IS_ERR(cs))
2514 return PTR_ERR(cs);
2515
2516 *cs++ = MI_LOAD_REGISTER_IMM(count);
2517 do {
2518 *cs++ = i915_mmio_reg_offset(flex->reg);
2519 *cs++ = flex->value;
2520 } while (flex++, --count);
2521 *cs++ = MI_NOOP;
2522
2523 intel_ring_advance(rq, cs);
2524
2525 return 0;
2526}
2527
2528static int gen8_modify_context(struct intel_context *ce,
2529 const struct flex *flex, unsigned int count)
2530{
2531 struct i915_request *rq;
2532 int err;
2533
2534 rq = intel_engine_create_kernel_request(ce->engine);
2535 if (IS_ERR(rq))
2536 return PTR_ERR(rq);
2537
2538 /* Serialise with the remote context */
2539 err = intel_context_prepare_remote_request(ce, rq);
2540 if (err == 0)
2541 err = gen8_store_flex(rq, ce, flex, count);
2542
2543 i915_request_add(rq);
2544 return err;
2545}
2546
2547static int
2548gen8_modify_self(struct intel_context *ce,
2549 const struct flex *flex, unsigned int count,
2550 struct i915_active *active)
2551{
2552 struct i915_request *rq;
2553 int err;
2554
2555 intel_engine_pm_get(ce->engine);
2556 rq = i915_request_create(ce);
2557 intel_engine_pm_put(ce->engine);
2558 if (IS_ERR(rq))
2559 return PTR_ERR(rq);
2560
2561 if (!IS_ERR_OR_NULL(active)) {
2562 err = i915_active_add_request(active, rq);
2563 if (err)
2564 goto err_add_request;
2565 }
2566
2567 err = gen8_load_flex(rq, ce, flex, count);
2568 if (err)
2569 goto err_add_request;
2570
2571err_add_request:
2572 i915_request_add(rq);
2573 return err;
2574}
2575
2576static int gen8_configure_context(struct i915_perf_stream *stream,
2577 struct i915_gem_context *ctx,
2578 struct flex *flex, unsigned int count)
2579{
2580 struct i915_gem_engines_iter it;
2581 struct intel_context *ce;
2582 int err = 0;
2583
2584 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
2585 GEM_BUG_ON(ce == ce->engine->kernel_context);
2586
2587 if (ce->engine->class != RENDER_CLASS)
2588 continue;
2589
2590 /* Otherwise OA settings will be set upon first use */
2591 if (!intel_context_pin_if_active(ce))
2592 continue;
2593
2594 flex->value = intel_sseu_make_rpcs(ce->engine->gt, &ce->sseu);
2595 err = gen8_modify_context(ce, flex, count);
2596
2597 intel_context_unpin(ce);
2598 if (err)
2599 break;
2600 }
2601 i915_gem_context_unlock_engines(ctx);
2602
2603 return err;
2604}
2605
2606static int gen12_configure_oar_context(struct i915_perf_stream *stream,
2607 struct i915_active *active)
2608{
2609 int err;
2610 struct intel_context *ce = stream->pinned_ctx;
2611 u32 format = stream->oa_buffer.format->format;
2612 u32 offset = stream->perf->ctx_oactxctrl_offset;
2613 struct flex regs_context[] = {
2614 {
2615 GEN8_OACTXCONTROL,
2616 offset + 1,
2617 active ? GEN8_OA_COUNTER_RESUME : 0,
2618 },
2619 };
2620 /* Offsets in regs_lri are not used since this configuration is only
2621 * applied using LRI. Initialize the correct offsets for posterity.
2622 */
2623#define GEN12_OAR_OACONTROL_OFFSET 0x5B0
2624 struct flex regs_lri[] = {
2625 {
2626 GEN12_OAR_OACONTROL,
2627 GEN12_OAR_OACONTROL_OFFSET + 1,
2628 (format << GEN12_OAR_OACONTROL_COUNTER_FORMAT_SHIFT) |
2629 (active ? GEN12_OAR_OACONTROL_COUNTER_ENABLE : 0)
2630 },
2631 {
2632 RING_CONTEXT_CONTROL(ce->engine->mmio_base),
2633 CTX_CONTEXT_CONTROL,
2634 _MASKED_FIELD(GEN12_CTX_CTRL_OAR_CONTEXT_ENABLE,
2635 active ?
2636 GEN12_CTX_CTRL_OAR_CONTEXT_ENABLE :
2637 0)
2638 },
2639 };
2640
2641 /* Modify the context image of pinned context with regs_context */
2642 err = intel_context_lock_pinned(ce);
2643 if (err)
2644 return err;
2645
2646 err = gen8_modify_context(ce, regs_context,
2647 ARRAY_SIZE(regs_context));
2648 intel_context_unlock_pinned(ce);
2649 if (err)
2650 return err;
2651
2652 /* Apply regs_lri using LRI with pinned context */
2653 return gen8_modify_self(ce, regs_lri, ARRAY_SIZE(regs_lri), active);
2654}
2655
2656/*
2657 * Manages updating the per-context aspects of the OA stream
2658 * configuration across all contexts.
2659 *
2660 * The awkward consideration here is that OACTXCONTROL controls the
2661 * exponent for periodic sampling which is primarily used for system
2662 * wide profiling where we'd like a consistent sampling period even in
2663 * the face of context switches.
2664 *
2665 * Our approach of updating the register state context (as opposed to
2666 * say using a workaround batch buffer) ensures that the hardware
2667 * won't automatically reload an out-of-date timer exponent even
2668 * transiently before a WA BB could be parsed.
2669 *
2670 * This function needs to:
2671 * - Ensure the currently running context's per-context OA state is
2672 * updated
2673 * - Ensure that all existing contexts will have the correct per-context
2674 * OA state if they are scheduled for use.
2675 * - Ensure any new contexts will be initialized with the correct
2676 * per-context OA state.
2677 *
2678 * Note: it's only the RCS/Render context that has any OA state.
2679 * Note: the first flex register passed must always be R_PWR_CLK_STATE
2680 */
2681static int
2682oa_configure_all_contexts(struct i915_perf_stream *stream,
2683 struct flex *regs,
2684 size_t num_regs,
2685 struct i915_active *active)
2686{
2687 struct drm_i915_private *i915 = stream->perf->i915;
2688 struct intel_engine_cs *engine;
2689 struct intel_gt *gt = stream->engine->gt;
2690 struct i915_gem_context *ctx, *cn;
2691 int err;
2692
2693 lockdep_assert_held(>->perf.lock);
2694
2695 /*
2696 * The OA register config is setup through the context image. This image
2697 * might be written to by the GPU on context switch (in particular on
2698 * lite-restore). This means we can't safely update a context's image,
2699 * if this context is scheduled/submitted to run on the GPU.
2700 *
2701 * We could emit the OA register config through the batch buffer but
2702 * this might leave small interval of time where the OA unit is
2703 * configured at an invalid sampling period.
2704 *
2705 * Note that since we emit all requests from a single ring, there
2706 * is still an implicit global barrier here that may cause a high
2707 * priority context to wait for an otherwise independent low priority
2708 * context. Contexts idle at the time of reconfiguration are not
2709 * trapped behind the barrier.
2710 */
2711 spin_lock(&i915->gem.contexts.lock);
2712 list_for_each_entry_safe(ctx, cn, &i915->gem.contexts.list, link) {
2713 if (!kref_get_unless_zero(&ctx->ref))
2714 continue;
2715
2716 spin_unlock(&i915->gem.contexts.lock);
2717
2718 err = gen8_configure_context(stream, ctx, regs, num_regs);
2719 if (err) {
2720 i915_gem_context_put(ctx);
2721 return err;
2722 }
2723
2724 spin_lock(&i915->gem.contexts.lock);
2725 list_safe_reset_next(ctx, cn, link);
2726 i915_gem_context_put(ctx);
2727 }
2728 spin_unlock(&i915->gem.contexts.lock);
2729
2730 /*
2731 * After updating all other contexts, we need to modify ourselves.
2732 * If we don't modify the kernel_context, we do not get events while
2733 * idle.
2734 */
2735 for_each_uabi_engine(engine, i915) {
2736 struct intel_context *ce = engine->kernel_context;
2737
2738 if (engine->class != RENDER_CLASS)
2739 continue;
2740
2741 regs[0].value = intel_sseu_make_rpcs(engine->gt, &ce->sseu);
2742
2743 err = gen8_modify_self(ce, regs, num_regs, active);
2744 if (err)
2745 return err;
2746 }
2747
2748 return 0;
2749}
2750
2751static int
2752lrc_configure_all_contexts(struct i915_perf_stream *stream,
2753 const struct i915_oa_config *oa_config,
2754 struct i915_active *active)
2755{
2756 u32 ctx_oactxctrl = stream->perf->ctx_oactxctrl_offset;
2757 /* The MMIO offsets for Flex EU registers aren't contiguous */
2758 const u32 ctx_flexeu0 = stream->perf->ctx_flexeu0_offset;
2759#define ctx_flexeuN(N) (ctx_flexeu0 + 2 * (N) + 1)
2760 struct flex regs[] = {
2761 {
2762 GEN8_R_PWR_CLK_STATE(RENDER_RING_BASE),
2763 CTX_R_PWR_CLK_STATE,
2764 },
2765 {
2766 GEN8_OACTXCONTROL,
2767 ctx_oactxctrl + 1,
2768 },
2769 { EU_PERF_CNTL0, ctx_flexeuN(0) },
2770 { EU_PERF_CNTL1, ctx_flexeuN(1) },
2771 { EU_PERF_CNTL2, ctx_flexeuN(2) },
2772 { EU_PERF_CNTL3, ctx_flexeuN(3) },
2773 { EU_PERF_CNTL4, ctx_flexeuN(4) },
2774 { EU_PERF_CNTL5, ctx_flexeuN(5) },
2775 { EU_PERF_CNTL6, ctx_flexeuN(6) },
2776 };
2777#undef ctx_flexeuN
2778 int i;
2779
2780 regs[1].value =
2781 (stream->period_exponent << GEN8_OA_TIMER_PERIOD_SHIFT) |
2782 (stream->periodic ? GEN8_OA_TIMER_ENABLE : 0) |
2783 GEN8_OA_COUNTER_RESUME;
2784
2785 for (i = 2; i < ARRAY_SIZE(regs); i++)
2786 regs[i].value = oa_config_flex_reg(oa_config, regs[i].reg);
2787
2788 return oa_configure_all_contexts(stream,
2789 regs, ARRAY_SIZE(regs),
2790 active);
2791}
2792
2793static int
2794gen8_enable_metric_set(struct i915_perf_stream *stream,
2795 struct i915_active *active)
2796{
2797 struct intel_uncore *uncore = stream->uncore;
2798 struct i915_oa_config *oa_config = stream->oa_config;
2799 int ret;
2800
2801 /*
2802 * We disable slice/unslice clock ratio change reports on SKL since
2803 * they are too noisy. The HW generates a lot of redundant reports
2804 * where the ratio hasn't really changed causing a lot of redundant
2805 * work to processes and increasing the chances we'll hit buffer
2806 * overruns.
2807 *
2808 * Although we don't currently use the 'disable overrun' OABUFFER
2809 * feature it's worth noting that clock ratio reports have to be
2810 * disabled before considering to use that feature since the HW doesn't
2811 * correctly block these reports.
2812 *
2813 * Currently none of the high-level metrics we have depend on knowing
2814 * this ratio to normalize.
2815 *
2816 * Note: This register is not power context saved and restored, but
2817 * that's OK considering that we disable RC6 while the OA unit is
2818 * enabled.
2819 *
2820 * The _INCLUDE_CLK_RATIO bit allows the slice/unslice frequency to
2821 * be read back from automatically triggered reports, as part of the
2822 * RPT_ID field.
2823 */
2824 if (IS_GRAPHICS_VER(stream->perf->i915, 9, 11)) {
2825 intel_uncore_write(uncore, GEN8_OA_DEBUG,
2826 _MASKED_BIT_ENABLE(GEN9_OA_DEBUG_DISABLE_CLK_RATIO_REPORTS |
2827 GEN9_OA_DEBUG_INCLUDE_CLK_RATIO));
2828 }
2829
2830 /*
2831 * Update all contexts prior writing the mux configurations as we need
2832 * to make sure all slices/subslices are ON before writing to NOA
2833 * registers.
2834 */
2835 ret = lrc_configure_all_contexts(stream, oa_config, active);
2836 if (ret)
2837 return ret;
2838
2839 return emit_oa_config(stream,
2840 stream->oa_config, oa_context(stream),
2841 active);
2842}
2843
2844static u32 oag_report_ctx_switches(const struct i915_perf_stream *stream)
2845{
2846 return _MASKED_FIELD(GEN12_OAG_OA_DEBUG_DISABLE_CTX_SWITCH_REPORTS,
2847 (stream->sample_flags & SAMPLE_OA_REPORT) ?
2848 0 : GEN12_OAG_OA_DEBUG_DISABLE_CTX_SWITCH_REPORTS);
2849}
2850
2851static int
2852gen12_enable_metric_set(struct i915_perf_stream *stream,
2853 struct i915_active *active)
2854{
2855 struct drm_i915_private *i915 = stream->perf->i915;
2856 struct intel_uncore *uncore = stream->uncore;
2857 bool periodic = stream->periodic;
2858 u32 period_exponent = stream->period_exponent;
2859 u32 sqcnt1;
2860 int ret;
2861
2862 /*
2863 * Wa_1508761755
2864 * EU NOA signals behave incorrectly if EU clock gating is enabled.
2865 * Disable thread stall DOP gating and EU DOP gating.
2866 */
2867 if (IS_DG2(i915)) {
2868 intel_gt_mcr_multicast_write(uncore->gt, GEN8_ROW_CHICKEN,
2869 _MASKED_BIT_ENABLE(STALL_DOP_GATING_DISABLE));
2870 intel_uncore_write(uncore, GEN7_ROW_CHICKEN2,
2871 _MASKED_BIT_ENABLE(GEN12_DISABLE_DOP_GATING));
2872 }
2873
2874 intel_uncore_write(uncore, __oa_regs(stream)->oa_debug,
2875 /* Disable clk ratio reports, like previous Gens. */
2876 _MASKED_BIT_ENABLE(GEN12_OAG_OA_DEBUG_DISABLE_CLK_RATIO_REPORTS |
2877 GEN12_OAG_OA_DEBUG_INCLUDE_CLK_RATIO) |
2878 /*
2879 * If the user didn't require OA reports, instruct
2880 * the hardware not to emit ctx switch reports.
2881 */
2882 oag_report_ctx_switches(stream));
2883
2884 intel_uncore_write(uncore, __oa_regs(stream)->oa_ctx_ctrl, periodic ?
2885 (GEN12_OAG_OAGLBCTXCTRL_COUNTER_RESUME |
2886 GEN12_OAG_OAGLBCTXCTRL_TIMER_ENABLE |
2887 (period_exponent << GEN12_OAG_OAGLBCTXCTRL_TIMER_PERIOD_SHIFT))
2888 : 0);
2889
2890 /*
2891 * Initialize Super Queue Internal Cnt Register
2892 * Set PMON Enable in order to collect valid metrics.
2893 * Enable bytes per clock reporting in OA.
2894 */
2895 sqcnt1 = GEN12_SQCNT1_PMON_ENABLE |
2896 (HAS_OA_BPC_REPORTING(i915) ? GEN12_SQCNT1_OABPC : 0);
2897
2898 intel_uncore_rmw(uncore, GEN12_SQCNT1, 0, sqcnt1);
2899
2900 /*
2901 * For Gen12, performance counters are context
2902 * saved/restored. Only enable it for the context that
2903 * requested this.
2904 */
2905 if (stream->ctx) {
2906 ret = gen12_configure_oar_context(stream, active);
2907 if (ret)
2908 return ret;
2909 }
2910
2911 return emit_oa_config(stream,
2912 stream->oa_config, oa_context(stream),
2913 active);
2914}
2915
2916static void gen8_disable_metric_set(struct i915_perf_stream *stream)
2917{
2918 struct intel_uncore *uncore = stream->uncore;
2919
2920 /* Reset all contexts' slices/subslices configurations. */
2921 lrc_configure_all_contexts(stream, NULL, NULL);
2922
2923 intel_uncore_rmw(uncore, GDT_CHICKEN_BITS, GT_NOA_ENABLE, 0);
2924}
2925
2926static void gen11_disable_metric_set(struct i915_perf_stream *stream)
2927{
2928 struct intel_uncore *uncore = stream->uncore;
2929
2930 /* Reset all contexts' slices/subslices configurations. */
2931 lrc_configure_all_contexts(stream, NULL, NULL);
2932
2933 /* Make sure we disable noa to save power. */
2934 intel_uncore_rmw(uncore, RPM_CONFIG1, GEN10_GT_NOA_ENABLE, 0);
2935}
2936
2937static void gen12_disable_metric_set(struct i915_perf_stream *stream)
2938{
2939 struct intel_uncore *uncore = stream->uncore;
2940 struct drm_i915_private *i915 = stream->perf->i915;
2941 u32 sqcnt1;
2942
2943 /*
2944 * Wa_1508761755: Enable thread stall DOP gating and EU DOP gating.
2945 */
2946 if (IS_DG2(i915)) {
2947 intel_gt_mcr_multicast_write(uncore->gt, GEN8_ROW_CHICKEN,
2948 _MASKED_BIT_DISABLE(STALL_DOP_GATING_DISABLE));
2949 intel_uncore_write(uncore, GEN7_ROW_CHICKEN2,
2950 _MASKED_BIT_DISABLE(GEN12_DISABLE_DOP_GATING));
2951 }
2952
2953 /* disable the context save/restore or OAR counters */
2954 if (stream->ctx)
2955 gen12_configure_oar_context(stream, NULL);
2956
2957 /* Make sure we disable noa to save power. */
2958 intel_uncore_rmw(uncore, RPM_CONFIG1, GEN10_GT_NOA_ENABLE, 0);
2959
2960 sqcnt1 = GEN12_SQCNT1_PMON_ENABLE |
2961 (HAS_OA_BPC_REPORTING(i915) ? GEN12_SQCNT1_OABPC : 0);
2962
2963 /* Reset PMON Enable to save power. */
2964 intel_uncore_rmw(uncore, GEN12_SQCNT1, sqcnt1, 0);
2965}
2966
2967static void gen7_oa_enable(struct i915_perf_stream *stream)
2968{
2969 struct intel_uncore *uncore = stream->uncore;
2970 struct i915_gem_context *ctx = stream->ctx;
2971 u32 ctx_id = stream->specific_ctx_id;
2972 bool periodic = stream->periodic;
2973 u32 period_exponent = stream->period_exponent;
2974 u32 report_format = stream->oa_buffer.format->format;
2975
2976 /*
2977 * Reset buf pointers so we don't forward reports from before now.
2978 *
2979 * Think carefully if considering trying to avoid this, since it
2980 * also ensures status flags and the buffer itself are cleared
2981 * in error paths, and we have checks for invalid reports based
2982 * on the assumption that certain fields are written to zeroed
2983 * memory which this helps maintains.
2984 */
2985 gen7_init_oa_buffer(stream);
2986
2987 intel_uncore_write(uncore, GEN7_OACONTROL,
2988 (ctx_id & GEN7_OACONTROL_CTX_MASK) |
2989 (period_exponent <<
2990 GEN7_OACONTROL_TIMER_PERIOD_SHIFT) |
2991 (periodic ? GEN7_OACONTROL_TIMER_ENABLE : 0) |
2992 (report_format << GEN7_OACONTROL_FORMAT_SHIFT) |
2993 (ctx ? GEN7_OACONTROL_PER_CTX_ENABLE : 0) |
2994 GEN7_OACONTROL_ENABLE);
2995}
2996
2997static void gen8_oa_enable(struct i915_perf_stream *stream)
2998{
2999 struct intel_uncore *uncore = stream->uncore;
3000 u32 report_format = stream->oa_buffer.format->format;
3001
3002 /*
3003 * Reset buf pointers so we don't forward reports from before now.
3004 *
3005 * Think carefully if considering trying to avoid this, since it
3006 * also ensures status flags and the buffer itself are cleared
3007 * in error paths, and we have checks for invalid reports based
3008 * on the assumption that certain fields are written to zeroed
3009 * memory which this helps maintains.
3010 */
3011 gen8_init_oa_buffer(stream);
3012
3013 /*
3014 * Note: we don't rely on the hardware to perform single context
3015 * filtering and instead filter on the cpu based on the context-id
3016 * field of reports
3017 */
3018 intel_uncore_write(uncore, GEN8_OACONTROL,
3019 (report_format << GEN8_OA_REPORT_FORMAT_SHIFT) |
3020 GEN8_OA_COUNTER_ENABLE);
3021}
3022
3023static void gen12_oa_enable(struct i915_perf_stream *stream)
3024{
3025 const struct i915_perf_regs *regs;
3026 u32 val;
3027
3028 /*
3029 * If we don't want OA reports from the OA buffer, then we don't even
3030 * need to program the OAG unit.
3031 */
3032 if (!(stream->sample_flags & SAMPLE_OA_REPORT))
3033 return;
3034
3035 gen12_init_oa_buffer(stream);
3036
3037 regs = __oa_regs(stream);
3038 val = (stream->oa_buffer.format->format << regs->oa_ctrl_counter_format_shift) |
3039 GEN12_OAG_OACONTROL_OA_COUNTER_ENABLE;
3040
3041 intel_uncore_write(stream->uncore, regs->oa_ctrl, val);
3042}
3043
3044/**
3045 * i915_oa_stream_enable - handle `I915_PERF_IOCTL_ENABLE` for OA stream
3046 * @stream: An i915 perf stream opened for OA metrics
3047 *
3048 * [Re]enables hardware periodic sampling according to the period configured
3049 * when opening the stream. This also starts a hrtimer that will periodically
3050 * check for data in the circular OA buffer for notifying userspace (e.g.
3051 * during a read() or poll()).
3052 */
3053static void i915_oa_stream_enable(struct i915_perf_stream *stream)
3054{
3055 stream->pollin = false;
3056
3057 stream->perf->ops.oa_enable(stream);
3058
3059 if (stream->sample_flags & SAMPLE_OA_REPORT)
3060 hrtimer_start(&stream->poll_check_timer,
3061 ns_to_ktime(stream->poll_oa_period),
3062 HRTIMER_MODE_REL_PINNED);
3063}
3064
3065static void gen7_oa_disable(struct i915_perf_stream *stream)
3066{
3067 struct intel_uncore *uncore = stream->uncore;
3068
3069 intel_uncore_write(uncore, GEN7_OACONTROL, 0);
3070 if (intel_wait_for_register(uncore,
3071 GEN7_OACONTROL, GEN7_OACONTROL_ENABLE, 0,
3072 50))
3073 drm_err(&stream->perf->i915->drm,
3074 "wait for OA to be disabled timed out\n");
3075}
3076
3077static void gen8_oa_disable(struct i915_perf_stream *stream)
3078{
3079 struct intel_uncore *uncore = stream->uncore;
3080
3081 intel_uncore_write(uncore, GEN8_OACONTROL, 0);
3082 if (intel_wait_for_register(uncore,
3083 GEN8_OACONTROL, GEN8_OA_COUNTER_ENABLE, 0,
3084 50))
3085 drm_err(&stream->perf->i915->drm,
3086 "wait for OA to be disabled timed out\n");
3087}
3088
3089static void gen12_oa_disable(struct i915_perf_stream *stream)
3090{
3091 struct intel_uncore *uncore = stream->uncore;
3092
3093 intel_uncore_write(uncore, __oa_regs(stream)->oa_ctrl, 0);
3094 if (intel_wait_for_register(uncore,
3095 __oa_regs(stream)->oa_ctrl,
3096 GEN12_OAG_OACONTROL_OA_COUNTER_ENABLE, 0,
3097 50))
3098 drm_err(&stream->perf->i915->drm,
3099 "wait for OA to be disabled timed out\n");
3100
3101 intel_uncore_write(uncore, GEN12_OA_TLB_INV_CR, 1);
3102 if (intel_wait_for_register(uncore,
3103 GEN12_OA_TLB_INV_CR,
3104 1, 0,
3105 50))
3106 drm_err(&stream->perf->i915->drm,
3107 "wait for OA tlb invalidate timed out\n");
3108}
3109
3110/**
3111 * i915_oa_stream_disable - handle `I915_PERF_IOCTL_DISABLE` for OA stream
3112 * @stream: An i915 perf stream opened for OA metrics
3113 *
3114 * Stops the OA unit from periodically writing counter reports into the
3115 * circular OA buffer. This also stops the hrtimer that periodically checks for
3116 * data in the circular OA buffer, for notifying userspace.
3117 */
3118static void i915_oa_stream_disable(struct i915_perf_stream *stream)
3119{
3120 stream->perf->ops.oa_disable(stream);
3121
3122 if (stream->sample_flags & SAMPLE_OA_REPORT)
3123 hrtimer_cancel(&stream->poll_check_timer);
3124}
3125
3126static const struct i915_perf_stream_ops i915_oa_stream_ops = {
3127 .destroy = i915_oa_stream_destroy,
3128 .enable = i915_oa_stream_enable,
3129 .disable = i915_oa_stream_disable,
3130 .wait_unlocked = i915_oa_wait_unlocked,
3131 .poll_wait = i915_oa_poll_wait,
3132 .read = i915_oa_read,
3133};
3134
3135static int i915_perf_stream_enable_sync(struct i915_perf_stream *stream)
3136{
3137 struct i915_active *active;
3138 int err;
3139
3140 active = i915_active_create();
3141 if (!active)
3142 return -ENOMEM;
3143
3144 err = stream->perf->ops.enable_metric_set(stream, active);
3145 if (err == 0)
3146 __i915_active_wait(active, TASK_UNINTERRUPTIBLE);
3147
3148 i915_active_put(active);
3149 return err;
3150}
3151
3152static void
3153get_default_sseu_config(struct intel_sseu *out_sseu,
3154 struct intel_engine_cs *engine)
3155{
3156 const struct sseu_dev_info *devinfo_sseu = &engine->gt->info.sseu;
3157
3158 *out_sseu = intel_sseu_from_device_info(devinfo_sseu);
3159
3160 if (GRAPHICS_VER(engine->i915) == 11) {
3161 /*
3162 * We only need subslice count so it doesn't matter which ones
3163 * we select - just turn off low bits in the amount of half of
3164 * all available subslices per slice.
3165 */
3166 out_sseu->subslice_mask =
3167 ~(~0 << (hweight8(out_sseu->subslice_mask) / 2));
3168 out_sseu->slice_mask = 0x1;
3169 }
3170}
3171
3172static int
3173get_sseu_config(struct intel_sseu *out_sseu,
3174 struct intel_engine_cs *engine,
3175 const struct drm_i915_gem_context_param_sseu *drm_sseu)
3176{
3177 if (drm_sseu->engine.engine_class != engine->uabi_class ||
3178 drm_sseu->engine.engine_instance != engine->uabi_instance)
3179 return -EINVAL;
3180
3181 return i915_gem_user_to_context_sseu(engine->gt, drm_sseu, out_sseu);
3182}
3183
3184/*
3185 * OA timestamp frequency = CS timestamp frequency in most platforms. On some
3186 * platforms OA unit ignores the CTC_SHIFT and the 2 timestamps differ. In such
3187 * cases, return the adjusted CS timestamp frequency to the user.
3188 */
3189u32 i915_perf_oa_timestamp_frequency(struct drm_i915_private *i915)
3190{
3191 struct intel_gt *gt = to_gt(i915);
3192
3193 /* Wa_18013179988 */
3194 if (IS_DG2(i915) || IS_GFX_GT_IP_RANGE(gt, IP_VER(12, 70), IP_VER(12, 74))) {
3195 intel_wakeref_t wakeref;
3196 u32 reg, shift;
3197
3198 with_intel_runtime_pm(to_gt(i915)->uncore->rpm, wakeref)
3199 reg = intel_uncore_read(to_gt(i915)->uncore, RPM_CONFIG0);
3200
3201 shift = REG_FIELD_GET(GEN10_RPM_CONFIG0_CTC_SHIFT_PARAMETER_MASK,
3202 reg);
3203
3204 return to_gt(i915)->clock_frequency << (3 - shift);
3205 }
3206
3207 return to_gt(i915)->clock_frequency;
3208}
3209
3210/**
3211 * i915_oa_stream_init - validate combined props for OA stream and init
3212 * @stream: An i915 perf stream
3213 * @param: The open parameters passed to `DRM_I915_PERF_OPEN`
3214 * @props: The property state that configures stream (individually validated)
3215 *
3216 * While read_properties_unlocked() validates properties in isolation it
3217 * doesn't ensure that the combination necessarily makes sense.
3218 *
3219 * At this point it has been determined that userspace wants a stream of
3220 * OA metrics, but still we need to further validate the combined
3221 * properties are OK.
3222 *
3223 * If the configuration makes sense then we can allocate memory for
3224 * a circular OA buffer and apply the requested metric set configuration.
3225 *
3226 * Returns: zero on success or a negative error code.
3227 */
3228static int i915_oa_stream_init(struct i915_perf_stream *stream,
3229 struct drm_i915_perf_open_param *param,
3230 struct perf_open_properties *props)
3231{
3232 struct drm_i915_private *i915 = stream->perf->i915;
3233 struct i915_perf *perf = stream->perf;
3234 struct i915_perf_group *g;
3235 int ret;
3236
3237 if (!props->engine) {
3238 drm_dbg(&stream->perf->i915->drm,
3239 "OA engine not specified\n");
3240 return -EINVAL;
3241 }
3242 g = props->engine->oa_group;
3243
3244 /*
3245 * If the sysfs metrics/ directory wasn't registered for some
3246 * reason then don't let userspace try their luck with config
3247 * IDs
3248 */
3249 if (!perf->metrics_kobj) {
3250 drm_dbg(&stream->perf->i915->drm,
3251 "OA metrics weren't advertised via sysfs\n");
3252 return -EINVAL;
3253 }
3254
3255 if (!(props->sample_flags & SAMPLE_OA_REPORT) &&
3256 (GRAPHICS_VER(perf->i915) < 12 || !stream->ctx)) {
3257 drm_dbg(&stream->perf->i915->drm,
3258 "Only OA report sampling supported\n");
3259 return -EINVAL;
3260 }
3261
3262 if (!perf->ops.enable_metric_set) {
3263 drm_dbg(&stream->perf->i915->drm,
3264 "OA unit not supported\n");
3265 return -ENODEV;
3266 }
3267
3268 /*
3269 * To avoid the complexity of having to accurately filter
3270 * counter reports and marshal to the appropriate client
3271 * we currently only allow exclusive access
3272 */
3273 if (g->exclusive_stream) {
3274 drm_dbg(&stream->perf->i915->drm,
3275 "OA unit already in use\n");
3276 return -EBUSY;
3277 }
3278
3279 if (!props->oa_format) {
3280 drm_dbg(&stream->perf->i915->drm,
3281 "OA report format not specified\n");
3282 return -EINVAL;
3283 }
3284
3285 stream->engine = props->engine;
3286 stream->uncore = stream->engine->gt->uncore;
3287
3288 stream->sample_size = sizeof(struct drm_i915_perf_record_header);
3289
3290 stream->oa_buffer.format = &perf->oa_formats[props->oa_format];
3291 if (drm_WARN_ON(&i915->drm, stream->oa_buffer.format->size == 0))
3292 return -EINVAL;
3293
3294 stream->sample_flags = props->sample_flags;
3295 stream->sample_size += stream->oa_buffer.format->size;
3296
3297 stream->hold_preemption = props->hold_preemption;
3298
3299 stream->periodic = props->oa_periodic;
3300 if (stream->periodic)
3301 stream->period_exponent = props->oa_period_exponent;
3302
3303 if (stream->ctx) {
3304 ret = oa_get_render_ctx_id(stream);
3305 if (ret) {
3306 drm_dbg(&stream->perf->i915->drm,
3307 "Invalid context id to filter with\n");
3308 return ret;
3309 }
3310 }
3311
3312 ret = alloc_noa_wait(stream);
3313 if (ret) {
3314 drm_dbg(&stream->perf->i915->drm,
3315 "Unable to allocate NOA wait batch buffer\n");
3316 goto err_noa_wait_alloc;
3317 }
3318
3319 stream->oa_config = i915_perf_get_oa_config(perf, props->metrics_set);
3320 if (!stream->oa_config) {
3321 drm_dbg(&stream->perf->i915->drm,
3322 "Invalid OA config id=%i\n", props->metrics_set);
3323 ret = -EINVAL;
3324 goto err_config;
3325 }
3326
3327 /* PRM - observability performance counters:
3328 *
3329 * OACONTROL, performance counter enable, note:
3330 *
3331 * "When this bit is set, in order to have coherent counts,
3332 * RC6 power state and trunk clock gating must be disabled.
3333 * This can be achieved by programming MMIO registers as
3334 * 0xA094=0 and 0xA090[31]=1"
3335 *
3336 * In our case we are expecting that taking pm + FORCEWAKE
3337 * references will effectively disable RC6.
3338 */
3339 intel_engine_pm_get(stream->engine);
3340 intel_uncore_forcewake_get(stream->uncore, FORCEWAKE_ALL);
3341
3342 ret = alloc_oa_buffer(stream);
3343 if (ret)
3344 goto err_oa_buf_alloc;
3345
3346 stream->ops = &i915_oa_stream_ops;
3347
3348 stream->engine->gt->perf.sseu = props->sseu;
3349 WRITE_ONCE(g->exclusive_stream, stream);
3350
3351 ret = i915_perf_stream_enable_sync(stream);
3352 if (ret) {
3353 drm_dbg(&stream->perf->i915->drm,
3354 "Unable to enable metric set\n");
3355 goto err_enable;
3356 }
3357
3358 drm_dbg(&stream->perf->i915->drm,
3359 "opening stream oa config uuid=%s\n",
3360 stream->oa_config->uuid);
3361
3362 hrtimer_init(&stream->poll_check_timer,
3363 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3364 stream->poll_check_timer.function = oa_poll_check_timer_cb;
3365 init_waitqueue_head(&stream->poll_wq);
3366 spin_lock_init(&stream->oa_buffer.ptr_lock);
3367 mutex_init(&stream->lock);
3368
3369 return 0;
3370
3371err_enable:
3372 WRITE_ONCE(g->exclusive_stream, NULL);
3373 perf->ops.disable_metric_set(stream);
3374
3375 free_oa_buffer(stream);
3376
3377err_oa_buf_alloc:
3378 intel_uncore_forcewake_put(stream->uncore, FORCEWAKE_ALL);
3379 intel_engine_pm_put(stream->engine);
3380
3381 free_oa_configs(stream);
3382
3383err_config:
3384 free_noa_wait(stream);
3385
3386err_noa_wait_alloc:
3387 if (stream->ctx)
3388 oa_put_render_ctx_id(stream);
3389
3390 return ret;
3391}
3392
3393void i915_oa_init_reg_state(const struct intel_context *ce,
3394 const struct intel_engine_cs *engine)
3395{
3396 struct i915_perf_stream *stream;
3397
3398 if (engine->class != RENDER_CLASS)
3399 return;
3400
3401 /* perf.exclusive_stream serialised by lrc_configure_all_contexts() */
3402 stream = READ_ONCE(engine->oa_group->exclusive_stream);
3403 if (stream && GRAPHICS_VER(stream->perf->i915) < 12)
3404 gen8_update_reg_state_unlocked(ce, stream);
3405}
3406
3407/**
3408 * i915_perf_read - handles read() FOP for i915 perf stream FDs
3409 * @file: An i915 perf stream file
3410 * @buf: destination buffer given by userspace
3411 * @count: the number of bytes userspace wants to read
3412 * @ppos: (inout) file seek position (unused)
3413 *
3414 * The entry point for handling a read() on a stream file descriptor from
3415 * userspace. Most of the work is left to the i915_perf_read_locked() and
3416 * &i915_perf_stream_ops->read but to save having stream implementations (of
3417 * which we might have multiple later) we handle blocking read here.
3418 *
3419 * We can also consistently treat trying to read from a disabled stream
3420 * as an IO error so implementations can assume the stream is enabled
3421 * while reading.
3422 *
3423 * Returns: The number of bytes copied or a negative error code on failure.
3424 */
3425static ssize_t i915_perf_read(struct file *file,
3426 char __user *buf,
3427 size_t count,
3428 loff_t *ppos)
3429{
3430 struct i915_perf_stream *stream = file->private_data;
3431 size_t offset = 0;
3432 int ret;
3433
3434 /* To ensure it's handled consistently we simply treat all reads of a
3435 * disabled stream as an error. In particular it might otherwise lead
3436 * to a deadlock for blocking file descriptors...
3437 */
3438 if (!stream->enabled || !(stream->sample_flags & SAMPLE_OA_REPORT))
3439 return -EIO;
3440
3441 if (!(file->f_flags & O_NONBLOCK)) {
3442 /* There's the small chance of false positives from
3443 * stream->ops->wait_unlocked.
3444 *
3445 * E.g. with single context filtering since we only wait until
3446 * oabuffer has >= 1 report we don't immediately know whether
3447 * any reports really belong to the current context
3448 */
3449 do {
3450 ret = stream->ops->wait_unlocked(stream);
3451 if (ret)
3452 return ret;
3453
3454 mutex_lock(&stream->lock);
3455 ret = stream->ops->read(stream, buf, count, &offset);
3456 mutex_unlock(&stream->lock);
3457 } while (!offset && !ret);
3458 } else {
3459 mutex_lock(&stream->lock);
3460 ret = stream->ops->read(stream, buf, count, &offset);
3461 mutex_unlock(&stream->lock);
3462 }
3463
3464 /* We allow the poll checking to sometimes report false positive EPOLLIN
3465 * events where we might actually report EAGAIN on read() if there's
3466 * not really any data available. In this situation though we don't
3467 * want to enter a busy loop between poll() reporting a EPOLLIN event
3468 * and read() returning -EAGAIN. Clearing the oa.pollin state here
3469 * effectively ensures we back off until the next hrtimer callback
3470 * before reporting another EPOLLIN event.
3471 * The exception to this is if ops->read() returned -ENOSPC which means
3472 * that more OA data is available than could fit in the user provided
3473 * buffer. In this case we want the next poll() call to not block.
3474 */
3475 if (ret != -ENOSPC)
3476 stream->pollin = false;
3477
3478 /* Possible values for ret are 0, -EFAULT, -ENOSPC, -EIO, ... */
3479 return offset ?: (ret ?: -EAGAIN);
3480}
3481
3482static enum hrtimer_restart oa_poll_check_timer_cb(struct hrtimer *hrtimer)
3483{
3484 struct i915_perf_stream *stream =
3485 container_of(hrtimer, typeof(*stream), poll_check_timer);
3486
3487 if (oa_buffer_check_unlocked(stream)) {
3488 stream->pollin = true;
3489 wake_up(&stream->poll_wq);
3490 }
3491
3492 hrtimer_forward_now(hrtimer,
3493 ns_to_ktime(stream->poll_oa_period));
3494
3495 return HRTIMER_RESTART;
3496}
3497
3498/**
3499 * i915_perf_poll_locked - poll_wait() with a suitable wait queue for stream
3500 * @stream: An i915 perf stream
3501 * @file: An i915 perf stream file
3502 * @wait: poll() state table
3503 *
3504 * For handling userspace polling on an i915 perf stream, this calls through to
3505 * &i915_perf_stream_ops->poll_wait to call poll_wait() with a wait queue that
3506 * will be woken for new stream data.
3507 *
3508 * Returns: any poll events that are ready without sleeping
3509 */
3510static __poll_t i915_perf_poll_locked(struct i915_perf_stream *stream,
3511 struct file *file,
3512 poll_table *wait)
3513{
3514 __poll_t events = 0;
3515
3516 stream->ops->poll_wait(stream, file, wait);
3517
3518 /* Note: we don't explicitly check whether there's something to read
3519 * here since this path may be very hot depending on what else
3520 * userspace is polling, or on the timeout in use. We rely solely on
3521 * the hrtimer/oa_poll_check_timer_cb to notify us when there are
3522 * samples to read.
3523 */
3524 if (stream->pollin)
3525 events |= EPOLLIN;
3526
3527 return events;
3528}
3529
3530/**
3531 * i915_perf_poll - call poll_wait() with a suitable wait queue for stream
3532 * @file: An i915 perf stream file
3533 * @wait: poll() state table
3534 *
3535 * For handling userspace polling on an i915 perf stream, this ensures
3536 * poll_wait() gets called with a wait queue that will be woken for new stream
3537 * data.
3538 *
3539 * Note: Implementation deferred to i915_perf_poll_locked()
3540 *
3541 * Returns: any poll events that are ready without sleeping
3542 */
3543static __poll_t i915_perf_poll(struct file *file, poll_table *wait)
3544{
3545 struct i915_perf_stream *stream = file->private_data;
3546 __poll_t ret;
3547
3548 mutex_lock(&stream->lock);
3549 ret = i915_perf_poll_locked(stream, file, wait);
3550 mutex_unlock(&stream->lock);
3551
3552 return ret;
3553}
3554
3555/**
3556 * i915_perf_enable_locked - handle `I915_PERF_IOCTL_ENABLE` ioctl
3557 * @stream: A disabled i915 perf stream
3558 *
3559 * [Re]enables the associated capture of data for this stream.
3560 *
3561 * If a stream was previously enabled then there's currently no intention
3562 * to provide userspace any guarantee about the preservation of previously
3563 * buffered data.
3564 */
3565static void i915_perf_enable_locked(struct i915_perf_stream *stream)
3566{
3567 if (stream->enabled)
3568 return;
3569
3570 /* Allow stream->ops->enable() to refer to this */
3571 stream->enabled = true;
3572
3573 if (stream->ops->enable)
3574 stream->ops->enable(stream);
3575
3576 if (stream->hold_preemption)
3577 intel_context_set_nopreempt(stream->pinned_ctx);
3578}
3579
3580/**
3581 * i915_perf_disable_locked - handle `I915_PERF_IOCTL_DISABLE` ioctl
3582 * @stream: An enabled i915 perf stream
3583 *
3584 * Disables the associated capture of data for this stream.
3585 *
3586 * The intention is that disabling an re-enabling a stream will ideally be
3587 * cheaper than destroying and re-opening a stream with the same configuration,
3588 * though there are no formal guarantees about what state or buffered data
3589 * must be retained between disabling and re-enabling a stream.
3590 *
3591 * Note: while a stream is disabled it's considered an error for userspace
3592 * to attempt to read from the stream (-EIO).
3593 */
3594static void i915_perf_disable_locked(struct i915_perf_stream *stream)
3595{
3596 if (!stream->enabled)
3597 return;
3598
3599 /* Allow stream->ops->disable() to refer to this */
3600 stream->enabled = false;
3601
3602 if (stream->hold_preemption)
3603 intel_context_clear_nopreempt(stream->pinned_ctx);
3604
3605 if (stream->ops->disable)
3606 stream->ops->disable(stream);
3607}
3608
3609static long i915_perf_config_locked(struct i915_perf_stream *stream,
3610 unsigned long metrics_set)
3611{
3612 struct i915_oa_config *config;
3613 long ret = stream->oa_config->id;
3614
3615 config = i915_perf_get_oa_config(stream->perf, metrics_set);
3616 if (!config)
3617 return -EINVAL;
3618
3619 if (config != stream->oa_config) {
3620 int err;
3621
3622 /*
3623 * If OA is bound to a specific context, emit the
3624 * reconfiguration inline from that context. The update
3625 * will then be ordered with respect to submission on that
3626 * context.
3627 *
3628 * When set globally, we use a low priority kernel context,
3629 * so it will effectively take effect when idle.
3630 */
3631 err = emit_oa_config(stream, config, oa_context(stream), NULL);
3632 if (!err)
3633 config = xchg(&stream->oa_config, config);
3634 else
3635 ret = err;
3636 }
3637
3638 i915_oa_config_put(config);
3639
3640 return ret;
3641}
3642
3643/**
3644 * i915_perf_ioctl_locked - support ioctl() usage with i915 perf stream FDs
3645 * @stream: An i915 perf stream
3646 * @cmd: the ioctl request
3647 * @arg: the ioctl data
3648 *
3649 * Returns: zero on success or a negative error code. Returns -EINVAL for
3650 * an unknown ioctl request.
3651 */
3652static long i915_perf_ioctl_locked(struct i915_perf_stream *stream,
3653 unsigned int cmd,
3654 unsigned long arg)
3655{
3656 switch (cmd) {
3657 case I915_PERF_IOCTL_ENABLE:
3658 i915_perf_enable_locked(stream);
3659 return 0;
3660 case I915_PERF_IOCTL_DISABLE:
3661 i915_perf_disable_locked(stream);
3662 return 0;
3663 case I915_PERF_IOCTL_CONFIG:
3664 return i915_perf_config_locked(stream, arg);
3665 }
3666
3667 return -EINVAL;
3668}
3669
3670/**
3671 * i915_perf_ioctl - support ioctl() usage with i915 perf stream FDs
3672 * @file: An i915 perf stream file
3673 * @cmd: the ioctl request
3674 * @arg: the ioctl data
3675 *
3676 * Implementation deferred to i915_perf_ioctl_locked().
3677 *
3678 * Returns: zero on success or a negative error code. Returns -EINVAL for
3679 * an unknown ioctl request.
3680 */
3681static long i915_perf_ioctl(struct file *file,
3682 unsigned int cmd,
3683 unsigned long arg)
3684{
3685 struct i915_perf_stream *stream = file->private_data;
3686 long ret;
3687
3688 mutex_lock(&stream->lock);
3689 ret = i915_perf_ioctl_locked(stream, cmd, arg);
3690 mutex_unlock(&stream->lock);
3691
3692 return ret;
3693}
3694
3695/**
3696 * i915_perf_destroy_locked - destroy an i915 perf stream
3697 * @stream: An i915 perf stream
3698 *
3699 * Frees all resources associated with the given i915 perf @stream, disabling
3700 * any associated data capture in the process.
3701 *
3702 * Note: The >->perf.lock mutex has been taken to serialize
3703 * with any non-file-operation driver hooks.
3704 */
3705static void i915_perf_destroy_locked(struct i915_perf_stream *stream)
3706{
3707 if (stream->enabled)
3708 i915_perf_disable_locked(stream);
3709
3710 if (stream->ops->destroy)
3711 stream->ops->destroy(stream);
3712
3713 if (stream->ctx)
3714 i915_gem_context_put(stream->ctx);
3715
3716 kfree(stream);
3717}
3718
3719/**
3720 * i915_perf_release - handles userspace close() of a stream file
3721 * @inode: anonymous inode associated with file
3722 * @file: An i915 perf stream file
3723 *
3724 * Cleans up any resources associated with an open i915 perf stream file.
3725 *
3726 * NB: close() can't really fail from the userspace point of view.
3727 *
3728 * Returns: zero on success or a negative error code.
3729 */
3730static int i915_perf_release(struct inode *inode, struct file *file)
3731{
3732 struct i915_perf_stream *stream = file->private_data;
3733 struct i915_perf *perf = stream->perf;
3734 struct intel_gt *gt = stream->engine->gt;
3735
3736 /*
3737 * Within this call, we know that the fd is being closed and we have no
3738 * other user of stream->lock. Use the perf lock to destroy the stream
3739 * here.
3740 */
3741 mutex_lock(>->perf.lock);
3742 i915_perf_destroy_locked(stream);
3743 mutex_unlock(>->perf.lock);
3744
3745 /* Release the reference the perf stream kept on the driver. */
3746 drm_dev_put(&perf->i915->drm);
3747
3748 return 0;
3749}
3750
3751
3752static const struct file_operations fops = {
3753 .owner = THIS_MODULE,
3754 .release = i915_perf_release,
3755 .poll = i915_perf_poll,
3756 .read = i915_perf_read,
3757 .unlocked_ioctl = i915_perf_ioctl,
3758 /* Our ioctl have no arguments, so it's safe to use the same function
3759 * to handle 32bits compatibility.
3760 */
3761 .compat_ioctl = i915_perf_ioctl,
3762};
3763
3764
3765/**
3766 * i915_perf_open_ioctl_locked - DRM ioctl() for userspace to open a stream FD
3767 * @perf: i915 perf instance
3768 * @param: The open parameters passed to 'DRM_I915_PERF_OPEN`
3769 * @props: individually validated u64 property value pairs
3770 * @file: drm file
3771 *
3772 * See i915_perf_ioctl_open() for interface details.
3773 *
3774 * Implements further stream config validation and stream initialization on
3775 * behalf of i915_perf_open_ioctl() with the >->perf.lock mutex
3776 * taken to serialize with any non-file-operation driver hooks.
3777 *
3778 * Note: at this point the @props have only been validated in isolation and
3779 * it's still necessary to validate that the combination of properties makes
3780 * sense.
3781 *
3782 * In the case where userspace is interested in OA unit metrics then further
3783 * config validation and stream initialization details will be handled by
3784 * i915_oa_stream_init(). The code here should only validate config state that
3785 * will be relevant to all stream types / backends.
3786 *
3787 * Returns: zero on success or a negative error code.
3788 */
3789static int
3790i915_perf_open_ioctl_locked(struct i915_perf *perf,
3791 struct drm_i915_perf_open_param *param,
3792 struct perf_open_properties *props,
3793 struct drm_file *file)
3794{
3795 struct i915_gem_context *specific_ctx = NULL;
3796 struct i915_perf_stream *stream = NULL;
3797 unsigned long f_flags = 0;
3798 bool privileged_op = true;
3799 int stream_fd;
3800 int ret;
3801
3802 if (props->single_context) {
3803 u32 ctx_handle = props->ctx_handle;
3804 struct drm_i915_file_private *file_priv = file->driver_priv;
3805
3806 specific_ctx = i915_gem_context_lookup(file_priv, ctx_handle);
3807 if (IS_ERR(specific_ctx)) {
3808 drm_dbg(&perf->i915->drm,
3809 "Failed to look up context with ID %u for opening perf stream\n",
3810 ctx_handle);
3811 ret = PTR_ERR(specific_ctx);
3812 goto err;
3813 }
3814 }
3815
3816 /*
3817 * On Haswell the OA unit supports clock gating off for a specific
3818 * context and in this mode there's no visibility of metrics for the
3819 * rest of the system, which we consider acceptable for a
3820 * non-privileged client.
3821 *
3822 * For Gen8->11 the OA unit no longer supports clock gating off for a
3823 * specific context and the kernel can't securely stop the counters
3824 * from updating as system-wide / global values. Even though we can
3825 * filter reports based on the included context ID we can't block
3826 * clients from seeing the raw / global counter values via
3827 * MI_REPORT_PERF_COUNT commands and so consider it a privileged op to
3828 * enable the OA unit by default.
3829 *
3830 * For Gen12+ we gain a new OAR unit that only monitors the RCS on a
3831 * per context basis. So we can relax requirements there if the user
3832 * doesn't request global stream access (i.e. query based sampling
3833 * using MI_RECORD_PERF_COUNT.
3834 */
3835 if (IS_HASWELL(perf->i915) && specific_ctx)
3836 privileged_op = false;
3837 else if (GRAPHICS_VER(perf->i915) == 12 && specific_ctx &&
3838 (props->sample_flags & SAMPLE_OA_REPORT) == 0)
3839 privileged_op = false;
3840
3841 if (props->hold_preemption) {
3842 if (!props->single_context) {
3843 drm_dbg(&perf->i915->drm,
3844 "preemption disable with no context\n");
3845 ret = -EINVAL;
3846 goto err;
3847 }
3848 privileged_op = true;
3849 }
3850
3851 /*
3852 * Asking for SSEU configuration is a priviliged operation.
3853 */
3854 if (props->has_sseu)
3855 privileged_op = true;
3856 else
3857 get_default_sseu_config(&props->sseu, props->engine);
3858
3859 /* Similar to perf's kernel.perf_paranoid_cpu sysctl option
3860 * we check a dev.i915.perf_stream_paranoid sysctl option
3861 * to determine if it's ok to access system wide OA counters
3862 * without CAP_PERFMON or CAP_SYS_ADMIN privileges.
3863 */
3864 if (privileged_op &&
3865 i915_perf_stream_paranoid && !perfmon_capable()) {
3866 drm_dbg(&perf->i915->drm,
3867 "Insufficient privileges to open i915 perf stream\n");
3868 ret = -EACCES;
3869 goto err_ctx;
3870 }
3871
3872 stream = kzalloc(sizeof(*stream), GFP_KERNEL);
3873 if (!stream) {
3874 ret = -ENOMEM;
3875 goto err_ctx;
3876 }
3877
3878 stream->perf = perf;
3879 stream->ctx = specific_ctx;
3880 stream->poll_oa_period = props->poll_oa_period;
3881
3882 ret = i915_oa_stream_init(stream, param, props);
3883 if (ret)
3884 goto err_alloc;
3885
3886 /* we avoid simply assigning stream->sample_flags = props->sample_flags
3887 * to have _stream_init check the combination of sample flags more
3888 * thoroughly, but still this is the expected result at this point.
3889 */
3890 if (WARN_ON(stream->sample_flags != props->sample_flags)) {
3891 ret = -ENODEV;
3892 goto err_flags;
3893 }
3894
3895 if (param->flags & I915_PERF_FLAG_FD_CLOEXEC)
3896 f_flags |= O_CLOEXEC;
3897 if (param->flags & I915_PERF_FLAG_FD_NONBLOCK)
3898 f_flags |= O_NONBLOCK;
3899
3900 stream_fd = anon_inode_getfd("[i915_perf]", &fops, stream, f_flags);
3901 if (stream_fd < 0) {
3902 ret = stream_fd;
3903 goto err_flags;
3904 }
3905
3906 if (!(param->flags & I915_PERF_FLAG_DISABLED))
3907 i915_perf_enable_locked(stream);
3908
3909 /* Take a reference on the driver that will be kept with stream_fd
3910 * until its release.
3911 */
3912 drm_dev_get(&perf->i915->drm);
3913
3914 return stream_fd;
3915
3916err_flags:
3917 if (stream->ops->destroy)
3918 stream->ops->destroy(stream);
3919err_alloc:
3920 kfree(stream);
3921err_ctx:
3922 if (specific_ctx)
3923 i915_gem_context_put(specific_ctx);
3924err:
3925 return ret;
3926}
3927
3928static u64 oa_exponent_to_ns(struct i915_perf *perf, int exponent)
3929{
3930 u64 nom = (2ULL << exponent) * NSEC_PER_SEC;
3931 u32 den = i915_perf_oa_timestamp_frequency(perf->i915);
3932
3933 return div_u64(nom + den - 1, den);
3934}
3935
3936static __always_inline bool
3937oa_format_valid(struct i915_perf *perf, enum drm_i915_oa_format format)
3938{
3939 return test_bit(format, perf->format_mask);
3940}
3941
3942static __always_inline void
3943oa_format_add(struct i915_perf *perf, enum drm_i915_oa_format format)
3944{
3945 __set_bit(format, perf->format_mask);
3946}
3947
3948/**
3949 * read_properties_unlocked - validate + copy userspace stream open properties
3950 * @perf: i915 perf instance
3951 * @uprops: The array of u64 key value pairs given by userspace
3952 * @n_props: The number of key value pairs expected in @uprops
3953 * @props: The stream configuration built up while validating properties
3954 *
3955 * Note this function only validates properties in isolation it doesn't
3956 * validate that the combination of properties makes sense or that all
3957 * properties necessary for a particular kind of stream have been set.
3958 *
3959 * Note that there currently aren't any ordering requirements for properties so
3960 * we shouldn't validate or assume anything about ordering here. This doesn't
3961 * rule out defining new properties with ordering requirements in the future.
3962 */
3963static int read_properties_unlocked(struct i915_perf *perf,
3964 u64 __user *uprops,
3965 u32 n_props,
3966 struct perf_open_properties *props)
3967{
3968 struct drm_i915_gem_context_param_sseu user_sseu;
3969 const struct i915_oa_format *f;
3970 u64 __user *uprop = uprops;
3971 bool config_instance = false;
3972 bool config_class = false;
3973 bool config_sseu = false;
3974 u8 class, instance;
3975 u32 i;
3976 int ret;
3977
3978 memset(props, 0, sizeof(struct perf_open_properties));
3979 props->poll_oa_period = DEFAULT_POLL_PERIOD_NS;
3980
3981 /* Considering that ID = 0 is reserved and assuming that we don't
3982 * (currently) expect any configurations to ever specify duplicate
3983 * values for a particular property ID then the last _PROP_MAX value is
3984 * one greater than the maximum number of properties we expect to get
3985 * from userspace.
3986 */
3987 if (!n_props || n_props >= DRM_I915_PERF_PROP_MAX) {
3988 drm_dbg(&perf->i915->drm,
3989 "Invalid number of i915 perf properties given\n");
3990 return -EINVAL;
3991 }
3992
3993 /* Defaults when class:instance is not passed */
3994 class = I915_ENGINE_CLASS_RENDER;
3995 instance = 0;
3996
3997 for (i = 0; i < n_props; i++) {
3998 u64 oa_period, oa_freq_hz;
3999 u64 id, value;
4000
4001 ret = get_user(id, uprop);
4002 if (ret)
4003 return ret;
4004
4005 ret = get_user(value, uprop + 1);
4006 if (ret)
4007 return ret;
4008
4009 if (id == 0 || id >= DRM_I915_PERF_PROP_MAX) {
4010 drm_dbg(&perf->i915->drm,
4011 "Unknown i915 perf property ID\n");
4012 return -EINVAL;
4013 }
4014
4015 switch ((enum drm_i915_perf_property_id)id) {
4016 case DRM_I915_PERF_PROP_CTX_HANDLE:
4017 props->single_context = 1;
4018 props->ctx_handle = value;
4019 break;
4020 case DRM_I915_PERF_PROP_SAMPLE_OA:
4021 if (value)
4022 props->sample_flags |= SAMPLE_OA_REPORT;
4023 break;
4024 case DRM_I915_PERF_PROP_OA_METRICS_SET:
4025 if (value == 0) {
4026 drm_dbg(&perf->i915->drm,
4027 "Unknown OA metric set ID\n");
4028 return -EINVAL;
4029 }
4030 props->metrics_set = value;
4031 break;
4032 case DRM_I915_PERF_PROP_OA_FORMAT:
4033 if (value == 0 || value >= I915_OA_FORMAT_MAX) {
4034 drm_dbg(&perf->i915->drm,
4035 "Out-of-range OA report format %llu\n",
4036 value);
4037 return -EINVAL;
4038 }
4039 if (!oa_format_valid(perf, value)) {
4040 drm_dbg(&perf->i915->drm,
4041 "Unsupported OA report format %llu\n",
4042 value);
4043 return -EINVAL;
4044 }
4045 props->oa_format = value;
4046 break;
4047 case DRM_I915_PERF_PROP_OA_EXPONENT:
4048 if (value > OA_EXPONENT_MAX) {
4049 drm_dbg(&perf->i915->drm,
4050 "OA timer exponent too high (> %u)\n",
4051 OA_EXPONENT_MAX);
4052 return -EINVAL;
4053 }
4054
4055 /* Theoretically we can program the OA unit to sample
4056 * e.g. every 160ns for HSW, 167ns for BDW/SKL or 104ns
4057 * for BXT. We don't allow such high sampling
4058 * frequencies by default unless root.
4059 */
4060
4061 BUILD_BUG_ON(sizeof(oa_period) != 8);
4062 oa_period = oa_exponent_to_ns(perf, value);
4063
4064 /* This check is primarily to ensure that oa_period <=
4065 * UINT32_MAX (before passing to do_div which only
4066 * accepts a u32 denominator), but we can also skip
4067 * checking anything < 1Hz which implicitly can't be
4068 * limited via an integer oa_max_sample_rate.
4069 */
4070 if (oa_period <= NSEC_PER_SEC) {
4071 u64 tmp = NSEC_PER_SEC;
4072 do_div(tmp, oa_period);
4073 oa_freq_hz = tmp;
4074 } else
4075 oa_freq_hz = 0;
4076
4077 if (oa_freq_hz > i915_oa_max_sample_rate && !perfmon_capable()) {
4078 drm_dbg(&perf->i915->drm,
4079 "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",
4080 i915_oa_max_sample_rate);
4081 return -EACCES;
4082 }
4083
4084 props->oa_periodic = true;
4085 props->oa_period_exponent = value;
4086 break;
4087 case DRM_I915_PERF_PROP_HOLD_PREEMPTION:
4088 props->hold_preemption = !!value;
4089 break;
4090 case DRM_I915_PERF_PROP_GLOBAL_SSEU: {
4091 if (GRAPHICS_VER_FULL(perf->i915) >= IP_VER(12, 55)) {
4092 drm_dbg(&perf->i915->drm,
4093 "SSEU config not supported on gfx %x\n",
4094 GRAPHICS_VER_FULL(perf->i915));
4095 return -ENODEV;
4096 }
4097
4098 if (copy_from_user(&user_sseu,
4099 u64_to_user_ptr(value),
4100 sizeof(user_sseu))) {
4101 drm_dbg(&perf->i915->drm,
4102 "Unable to copy global sseu parameter\n");
4103 return -EFAULT;
4104 }
4105 config_sseu = true;
4106 break;
4107 }
4108 case DRM_I915_PERF_PROP_POLL_OA_PERIOD:
4109 if (value < 100000 /* 100us */) {
4110 drm_dbg(&perf->i915->drm,
4111 "OA availability timer too small (%lluns < 100us)\n",
4112 value);
4113 return -EINVAL;
4114 }
4115 props->poll_oa_period = value;
4116 break;
4117 case DRM_I915_PERF_PROP_OA_ENGINE_CLASS:
4118 class = (u8)value;
4119 config_class = true;
4120 break;
4121 case DRM_I915_PERF_PROP_OA_ENGINE_INSTANCE:
4122 instance = (u8)value;
4123 config_instance = true;
4124 break;
4125 default:
4126 MISSING_CASE(id);
4127 return -EINVAL;
4128 }
4129
4130 uprop += 2;
4131 }
4132
4133 if ((config_class && !config_instance) ||
4134 (config_instance && !config_class)) {
4135 drm_dbg(&perf->i915->drm,
4136 "OA engine-class and engine-instance parameters must be passed together\n");
4137 return -EINVAL;
4138 }
4139
4140 props->engine = intel_engine_lookup_user(perf->i915, class, instance);
4141 if (!props->engine) {
4142 drm_dbg(&perf->i915->drm,
4143 "OA engine class and instance invalid %d:%d\n",
4144 class, instance);
4145 return -EINVAL;
4146 }
4147
4148 if (!engine_supports_oa(props->engine)) {
4149 drm_dbg(&perf->i915->drm,
4150 "Engine not supported by OA %d:%d\n",
4151 class, instance);
4152 return -EINVAL;
4153 }
4154
4155 /*
4156 * Wa_14017512683: mtl[a0..c0): Use of OAM must be preceded with Media
4157 * C6 disable in BIOS. Fail if Media C6 is enabled on steppings where OAM
4158 * does not work as expected.
4159 */
4160 if (IS_MEDIA_GT_IP_STEP(props->engine->gt, IP_VER(13, 0), STEP_A0, STEP_C0) &&
4161 props->engine->oa_group->type == TYPE_OAM &&
4162 intel_check_bios_c6_setup(&props->engine->gt->rc6)) {
4163 drm_dbg(&perf->i915->drm,
4164 "OAM requires media C6 to be disabled in BIOS\n");
4165 return -EINVAL;
4166 }
4167
4168 i = array_index_nospec(props->oa_format, I915_OA_FORMAT_MAX);
4169 f = &perf->oa_formats[i];
4170 if (!engine_supports_oa_format(props->engine, f->type)) {
4171 drm_dbg(&perf->i915->drm,
4172 "Invalid OA format %d for class %d\n",
4173 f->type, props->engine->class);
4174 return -EINVAL;
4175 }
4176
4177 if (config_sseu) {
4178 ret = get_sseu_config(&props->sseu, props->engine, &user_sseu);
4179 if (ret) {
4180 drm_dbg(&perf->i915->drm,
4181 "Invalid SSEU configuration\n");
4182 return ret;
4183 }
4184 props->has_sseu = true;
4185 }
4186
4187 return 0;
4188}
4189
4190/**
4191 * i915_perf_open_ioctl - DRM ioctl() for userspace to open a stream FD
4192 * @dev: drm device
4193 * @data: ioctl data copied from userspace (unvalidated)
4194 * @file: drm file
4195 *
4196 * Validates the stream open parameters given by userspace including flags
4197 * and an array of u64 key, value pair properties.
4198 *
4199 * Very little is assumed up front about the nature of the stream being
4200 * opened (for instance we don't assume it's for periodic OA unit metrics). An
4201 * i915-perf stream is expected to be a suitable interface for other forms of
4202 * buffered data written by the GPU besides periodic OA metrics.
4203 *
4204 * Note we copy the properties from userspace outside of the i915 perf
4205 * mutex to avoid an awkward lockdep with mmap_lock.
4206 *
4207 * Most of the implementation details are handled by
4208 * i915_perf_open_ioctl_locked() after taking the >->perf.lock
4209 * mutex for serializing with any non-file-operation driver hooks.
4210 *
4211 * Return: A newly opened i915 Perf stream file descriptor or negative
4212 * error code on failure.
4213 */
4214int i915_perf_open_ioctl(struct drm_device *dev, void *data,
4215 struct drm_file *file)
4216{
4217 struct i915_perf *perf = &to_i915(dev)->perf;
4218 struct drm_i915_perf_open_param *param = data;
4219 struct intel_gt *gt;
4220 struct perf_open_properties props;
4221 u32 known_open_flags;
4222 int ret;
4223
4224 if (!perf->i915)
4225 return -ENOTSUPP;
4226
4227 known_open_flags = I915_PERF_FLAG_FD_CLOEXEC |
4228 I915_PERF_FLAG_FD_NONBLOCK |
4229 I915_PERF_FLAG_DISABLED;
4230 if (param->flags & ~known_open_flags) {
4231 drm_dbg(&perf->i915->drm,
4232 "Unknown drm_i915_perf_open_param flag\n");
4233 return -EINVAL;
4234 }
4235
4236 ret = read_properties_unlocked(perf,
4237 u64_to_user_ptr(param->properties_ptr),
4238 param->num_properties,
4239 &props);
4240 if (ret)
4241 return ret;
4242
4243 gt = props.engine->gt;
4244
4245 mutex_lock(>->perf.lock);
4246 ret = i915_perf_open_ioctl_locked(perf, param, &props, file);
4247 mutex_unlock(>->perf.lock);
4248
4249 return ret;
4250}
4251
4252/**
4253 * i915_perf_register - exposes i915-perf to userspace
4254 * @i915: i915 device instance
4255 *
4256 * In particular OA metric sets are advertised under a sysfs metrics/
4257 * directory allowing userspace to enumerate valid IDs that can be
4258 * used to open an i915-perf stream.
4259 */
4260void i915_perf_register(struct drm_i915_private *i915)
4261{
4262 struct i915_perf *perf = &i915->perf;
4263 struct intel_gt *gt = to_gt(i915);
4264
4265 if (!perf->i915)
4266 return;
4267
4268 /* To be sure we're synchronized with an attempted
4269 * i915_perf_open_ioctl(); considering that we register after
4270 * being exposed to userspace.
4271 */
4272 mutex_lock(>->perf.lock);
4273
4274 perf->metrics_kobj =
4275 kobject_create_and_add("metrics",
4276 &i915->drm.primary->kdev->kobj);
4277
4278 mutex_unlock(>->perf.lock);
4279}
4280
4281/**
4282 * i915_perf_unregister - hide i915-perf from userspace
4283 * @i915: i915 device instance
4284 *
4285 * i915-perf state cleanup is split up into an 'unregister' and
4286 * 'deinit' phase where the interface is first hidden from
4287 * userspace by i915_perf_unregister() before cleaning up
4288 * remaining state in i915_perf_fini().
4289 */
4290void i915_perf_unregister(struct drm_i915_private *i915)
4291{
4292 struct i915_perf *perf = &i915->perf;
4293
4294 if (!perf->metrics_kobj)
4295 return;
4296
4297 kobject_put(perf->metrics_kobj);
4298 perf->metrics_kobj = NULL;
4299}
4300
4301static bool gen8_is_valid_flex_addr(struct i915_perf *perf, u32 addr)
4302{
4303 static const i915_reg_t flex_eu_regs[] = {
4304 EU_PERF_CNTL0,
4305 EU_PERF_CNTL1,
4306 EU_PERF_CNTL2,
4307 EU_PERF_CNTL3,
4308 EU_PERF_CNTL4,
4309 EU_PERF_CNTL5,
4310 EU_PERF_CNTL6,
4311 };
4312 int i;
4313
4314 for (i = 0; i < ARRAY_SIZE(flex_eu_regs); i++) {
4315 if (i915_mmio_reg_offset(flex_eu_regs[i]) == addr)
4316 return true;
4317 }
4318 return false;
4319}
4320
4321static bool reg_in_range_table(u32 addr, const struct i915_range *table)
4322{
4323 while (table->start || table->end) {
4324 if (addr >= table->start && addr <= table->end)
4325 return true;
4326
4327 table++;
4328 }
4329
4330 return false;
4331}
4332
4333#define REG_EQUAL(addr, mmio) \
4334 ((addr) == i915_mmio_reg_offset(mmio))
4335
4336static const struct i915_range gen7_oa_b_counters[] = {
4337 { .start = 0x2710, .end = 0x272c }, /* OASTARTTRIG[1-8] */
4338 { .start = 0x2740, .end = 0x275c }, /* OAREPORTTRIG[1-8] */
4339 { .start = 0x2770, .end = 0x27ac }, /* OACEC[0-7][0-1] */
4340 {}
4341};
4342
4343static const struct i915_range gen12_oa_b_counters[] = {
4344 { .start = 0x2b2c, .end = 0x2b2c }, /* GEN12_OAG_OA_PESS */
4345 { .start = 0xd900, .end = 0xd91c }, /* GEN12_OAG_OASTARTTRIG[1-8] */
4346 { .start = 0xd920, .end = 0xd93c }, /* GEN12_OAG_OAREPORTTRIG1[1-8] */
4347 { .start = 0xd940, .end = 0xd97c }, /* GEN12_OAG_CEC[0-7][0-1] */
4348 { .start = 0xdc00, .end = 0xdc3c }, /* GEN12_OAG_SCEC[0-7][0-1] */
4349 { .start = 0xdc40, .end = 0xdc40 }, /* GEN12_OAG_SPCTR_CNF */
4350 { .start = 0xdc44, .end = 0xdc44 }, /* GEN12_OAA_DBG_REG */
4351 {}
4352};
4353
4354static const struct i915_range mtl_oam_b_counters[] = {
4355 { .start = 0x393000, .end = 0x39301c }, /* GEN12_OAM_STARTTRIG1[1-8] */
4356 { .start = 0x393020, .end = 0x39303c }, /* GEN12_OAM_REPORTTRIG1[1-8] */
4357 { .start = 0x393040, .end = 0x39307c }, /* GEN12_OAM_CEC[0-7][0-1] */
4358 { .start = 0x393200, .end = 0x39323C }, /* MPES[0-7] */
4359 {}
4360};
4361
4362static const struct i915_range xehp_oa_b_counters[] = {
4363 { .start = 0xdc48, .end = 0xdc48 }, /* OAA_ENABLE_REG */
4364 { .start = 0xdd00, .end = 0xdd48 }, /* OAG_LCE0_0 - OAA_LENABLE_REG */
4365 {}
4366};
4367
4368static const struct i915_range gen7_oa_mux_regs[] = {
4369 { .start = 0x91b8, .end = 0x91cc }, /* OA_PERFCNT[1-2], OA_PERFMATRIX */
4370 { .start = 0x9800, .end = 0x9888 }, /* MICRO_BP0_0 - NOA_WRITE */
4371 { .start = 0xe180, .end = 0xe180 }, /* HALF_SLICE_CHICKEN2 */
4372 {}
4373};
4374
4375static const struct i915_range hsw_oa_mux_regs[] = {
4376 { .start = 0x09e80, .end = 0x09ea4 }, /* HSW_MBVID2_NOA[0-9] */
4377 { .start = 0x09ec0, .end = 0x09ec0 }, /* HSW_MBVID2_MISR0 */
4378 { .start = 0x25100, .end = 0x2ff90 },
4379 {}
4380};
4381
4382static const struct i915_range chv_oa_mux_regs[] = {
4383 { .start = 0x182300, .end = 0x1823a4 },
4384 {}
4385};
4386
4387static const struct i915_range gen8_oa_mux_regs[] = {
4388 { .start = 0x0d00, .end = 0x0d2c }, /* RPM_CONFIG[0-1], NOA_CONFIG[0-8] */
4389 { .start = 0x20cc, .end = 0x20cc }, /* WAIT_FOR_RC6_EXIT */
4390 {}
4391};
4392
4393static const struct i915_range gen11_oa_mux_regs[] = {
4394 { .start = 0x91c8, .end = 0x91dc }, /* OA_PERFCNT[3-4] */
4395 {}
4396};
4397
4398static const struct i915_range gen12_oa_mux_regs[] = {
4399 { .start = 0x0d00, .end = 0x0d04 }, /* RPM_CONFIG[0-1] */
4400 { .start = 0x0d0c, .end = 0x0d2c }, /* NOA_CONFIG[0-8] */
4401 { .start = 0x9840, .end = 0x9840 }, /* GDT_CHICKEN_BITS */
4402 { .start = 0x9884, .end = 0x9888 }, /* NOA_WRITE */
4403 { .start = 0x20cc, .end = 0x20cc }, /* WAIT_FOR_RC6_EXIT */
4404 {}
4405};
4406
4407/*
4408 * Ref: 14010536224:
4409 * 0x20cc is repurposed on MTL, so use a separate array for MTL.
4410 */
4411static const struct i915_range mtl_oa_mux_regs[] = {
4412 { .start = 0x0d00, .end = 0x0d04 }, /* RPM_CONFIG[0-1] */
4413 { .start = 0x0d0c, .end = 0x0d2c }, /* NOA_CONFIG[0-8] */
4414 { .start = 0x9840, .end = 0x9840 }, /* GDT_CHICKEN_BITS */
4415 { .start = 0x9884, .end = 0x9888 }, /* NOA_WRITE */
4416 { .start = 0x38d100, .end = 0x38d114}, /* VISACTL */
4417 {}
4418};
4419
4420static bool gen7_is_valid_b_counter_addr(struct i915_perf *perf, u32 addr)
4421{
4422 return reg_in_range_table(addr, gen7_oa_b_counters);
4423}
4424
4425static bool gen8_is_valid_mux_addr(struct i915_perf *perf, u32 addr)
4426{
4427 return reg_in_range_table(addr, gen7_oa_mux_regs) ||
4428 reg_in_range_table(addr, gen8_oa_mux_regs);
4429}
4430
4431static bool gen11_is_valid_mux_addr(struct i915_perf *perf, u32 addr)
4432{
4433 return reg_in_range_table(addr, gen7_oa_mux_regs) ||
4434 reg_in_range_table(addr, gen8_oa_mux_regs) ||
4435 reg_in_range_table(addr, gen11_oa_mux_regs);
4436}
4437
4438static bool hsw_is_valid_mux_addr(struct i915_perf *perf, u32 addr)
4439{
4440 return reg_in_range_table(addr, gen7_oa_mux_regs) ||
4441 reg_in_range_table(addr, hsw_oa_mux_regs);
4442}
4443
4444static bool chv_is_valid_mux_addr(struct i915_perf *perf, u32 addr)
4445{
4446 return reg_in_range_table(addr, gen7_oa_mux_regs) ||
4447 reg_in_range_table(addr, chv_oa_mux_regs);
4448}
4449
4450static bool gen12_is_valid_b_counter_addr(struct i915_perf *perf, u32 addr)
4451{
4452 return reg_in_range_table(addr, gen12_oa_b_counters);
4453}
4454
4455static bool mtl_is_valid_oam_b_counter_addr(struct i915_perf *perf, u32 addr)
4456{
4457 if (HAS_OAM(perf->i915) &&
4458 GRAPHICS_VER_FULL(perf->i915) >= IP_VER(12, 70))
4459 return reg_in_range_table(addr, mtl_oam_b_counters);
4460
4461 return false;
4462}
4463
4464static bool xehp_is_valid_b_counter_addr(struct i915_perf *perf, u32 addr)
4465{
4466 return reg_in_range_table(addr, xehp_oa_b_counters) ||
4467 reg_in_range_table(addr, gen12_oa_b_counters) ||
4468 mtl_is_valid_oam_b_counter_addr(perf, addr);
4469}
4470
4471static bool gen12_is_valid_mux_addr(struct i915_perf *perf, u32 addr)
4472{
4473 if (GRAPHICS_VER_FULL(perf->i915) >= IP_VER(12, 70))
4474 return reg_in_range_table(addr, mtl_oa_mux_regs);
4475 else
4476 return reg_in_range_table(addr, gen12_oa_mux_regs);
4477}
4478
4479static u32 mask_reg_value(u32 reg, u32 val)
4480{
4481 /* HALF_SLICE_CHICKEN2 is programmed with a the
4482 * WaDisableSTUnitPowerOptimization workaround. Make sure the value
4483 * programmed by userspace doesn't change this.
4484 */
4485 if (REG_EQUAL(reg, HALF_SLICE_CHICKEN2))
4486 val = val & ~_MASKED_BIT_ENABLE(GEN8_ST_PO_DISABLE);
4487
4488 /* WAIT_FOR_RC6_EXIT has only one bit fullfilling the function
4489 * indicated by its name and a bunch of selection fields used by OA
4490 * configs.
4491 */
4492 if (REG_EQUAL(reg, WAIT_FOR_RC6_EXIT))
4493 val = val & ~_MASKED_BIT_ENABLE(HSW_WAIT_FOR_RC6_EXIT_ENABLE);
4494
4495 return val;
4496}
4497
4498static struct i915_oa_reg *alloc_oa_regs(struct i915_perf *perf,
4499 bool (*is_valid)(struct i915_perf *perf, u32 addr),
4500 u32 __user *regs,
4501 u32 n_regs)
4502{
4503 struct i915_oa_reg *oa_regs;
4504 int err;
4505 u32 i;
4506
4507 if (!n_regs)
4508 return NULL;
4509
4510 /* No is_valid function means we're not allowing any register to be programmed. */
4511 GEM_BUG_ON(!is_valid);
4512 if (!is_valid)
4513 return ERR_PTR(-EINVAL);
4514
4515 oa_regs = kmalloc_array(n_regs, sizeof(*oa_regs), GFP_KERNEL);
4516 if (!oa_regs)
4517 return ERR_PTR(-ENOMEM);
4518
4519 for (i = 0; i < n_regs; i++) {
4520 u32 addr, value;
4521
4522 err = get_user(addr, regs);
4523 if (err)
4524 goto addr_err;
4525
4526 if (!is_valid(perf, addr)) {
4527 drm_dbg(&perf->i915->drm,
4528 "Invalid oa_reg address: %X\n", addr);
4529 err = -EINVAL;
4530 goto addr_err;
4531 }
4532
4533 err = get_user(value, regs + 1);
4534 if (err)
4535 goto addr_err;
4536
4537 oa_regs[i].addr = _MMIO(addr);
4538 oa_regs[i].value = mask_reg_value(addr, value);
4539
4540 regs += 2;
4541 }
4542
4543 return oa_regs;
4544
4545addr_err:
4546 kfree(oa_regs);
4547 return ERR_PTR(err);
4548}
4549
4550static ssize_t show_dynamic_id(struct kobject *kobj,
4551 struct kobj_attribute *attr,
4552 char *buf)
4553{
4554 struct i915_oa_config *oa_config =
4555 container_of(attr, typeof(*oa_config), sysfs_metric_id);
4556
4557 return sprintf(buf, "%d\n", oa_config->id);
4558}
4559
4560static int create_dynamic_oa_sysfs_entry(struct i915_perf *perf,
4561 struct i915_oa_config *oa_config)
4562{
4563 sysfs_attr_init(&oa_config->sysfs_metric_id.attr);
4564 oa_config->sysfs_metric_id.attr.name = "id";
4565 oa_config->sysfs_metric_id.attr.mode = S_IRUGO;
4566 oa_config->sysfs_metric_id.show = show_dynamic_id;
4567 oa_config->sysfs_metric_id.store = NULL;
4568
4569 oa_config->attrs[0] = &oa_config->sysfs_metric_id.attr;
4570 oa_config->attrs[1] = NULL;
4571
4572 oa_config->sysfs_metric.name = oa_config->uuid;
4573 oa_config->sysfs_metric.attrs = oa_config->attrs;
4574
4575 return sysfs_create_group(perf->metrics_kobj,
4576 &oa_config->sysfs_metric);
4577}
4578
4579/**
4580 * i915_perf_add_config_ioctl - DRM ioctl() for userspace to add a new OA config
4581 * @dev: drm device
4582 * @data: ioctl data (pointer to struct drm_i915_perf_oa_config) copied from
4583 * userspace (unvalidated)
4584 * @file: drm file
4585 *
4586 * Validates the submitted OA register to be saved into a new OA config that
4587 * can then be used for programming the OA unit and its NOA network.
4588 *
4589 * Returns: A new allocated config number to be used with the perf open ioctl
4590 * or a negative error code on failure.
4591 */
4592int i915_perf_add_config_ioctl(struct drm_device *dev, void *data,
4593 struct drm_file *file)
4594{
4595 struct i915_perf *perf = &to_i915(dev)->perf;
4596 struct drm_i915_perf_oa_config *args = data;
4597 struct i915_oa_config *oa_config, *tmp;
4598 struct i915_oa_reg *regs;
4599 int err, id;
4600
4601 if (!perf->i915)
4602 return -ENOTSUPP;
4603
4604 if (!perf->metrics_kobj) {
4605 drm_dbg(&perf->i915->drm,
4606 "OA metrics weren't advertised via sysfs\n");
4607 return -EINVAL;
4608 }
4609
4610 if (i915_perf_stream_paranoid && !perfmon_capable()) {
4611 drm_dbg(&perf->i915->drm,
4612 "Insufficient privileges to add i915 OA config\n");
4613 return -EACCES;
4614 }
4615
4616 if ((!args->mux_regs_ptr || !args->n_mux_regs) &&
4617 (!args->boolean_regs_ptr || !args->n_boolean_regs) &&
4618 (!args->flex_regs_ptr || !args->n_flex_regs)) {
4619 drm_dbg(&perf->i915->drm,
4620 "No OA registers given\n");
4621 return -EINVAL;
4622 }
4623
4624 oa_config = kzalloc(sizeof(*oa_config), GFP_KERNEL);
4625 if (!oa_config) {
4626 drm_dbg(&perf->i915->drm,
4627 "Failed to allocate memory for the OA config\n");
4628 return -ENOMEM;
4629 }
4630
4631 oa_config->perf = perf;
4632 kref_init(&oa_config->ref);
4633
4634 if (!uuid_is_valid(args->uuid)) {
4635 drm_dbg(&perf->i915->drm,
4636 "Invalid uuid format for OA config\n");
4637 err = -EINVAL;
4638 goto reg_err;
4639 }
4640
4641 /* Last character in oa_config->uuid will be 0 because oa_config is
4642 * kzalloc.
4643 */
4644 memcpy(oa_config->uuid, args->uuid, sizeof(args->uuid));
4645
4646 oa_config->mux_regs_len = args->n_mux_regs;
4647 regs = alloc_oa_regs(perf,
4648 perf->ops.is_valid_mux_reg,
4649 u64_to_user_ptr(args->mux_regs_ptr),
4650 args->n_mux_regs);
4651
4652 if (IS_ERR(regs)) {
4653 drm_dbg(&perf->i915->drm,
4654 "Failed to create OA config for mux_regs\n");
4655 err = PTR_ERR(regs);
4656 goto reg_err;
4657 }
4658 oa_config->mux_regs = regs;
4659
4660 oa_config->b_counter_regs_len = args->n_boolean_regs;
4661 regs = alloc_oa_regs(perf,
4662 perf->ops.is_valid_b_counter_reg,
4663 u64_to_user_ptr(args->boolean_regs_ptr),
4664 args->n_boolean_regs);
4665
4666 if (IS_ERR(regs)) {
4667 drm_dbg(&perf->i915->drm,
4668 "Failed to create OA config for b_counter_regs\n");
4669 err = PTR_ERR(regs);
4670 goto reg_err;
4671 }
4672 oa_config->b_counter_regs = regs;
4673
4674 if (GRAPHICS_VER(perf->i915) < 8) {
4675 if (args->n_flex_regs != 0) {
4676 err = -EINVAL;
4677 goto reg_err;
4678 }
4679 } else {
4680 oa_config->flex_regs_len = args->n_flex_regs;
4681 regs = alloc_oa_regs(perf,
4682 perf->ops.is_valid_flex_reg,
4683 u64_to_user_ptr(args->flex_regs_ptr),
4684 args->n_flex_regs);
4685
4686 if (IS_ERR(regs)) {
4687 drm_dbg(&perf->i915->drm,
4688 "Failed to create OA config for flex_regs\n");
4689 err = PTR_ERR(regs);
4690 goto reg_err;
4691 }
4692 oa_config->flex_regs = regs;
4693 }
4694
4695 err = mutex_lock_interruptible(&perf->metrics_lock);
4696 if (err)
4697 goto reg_err;
4698
4699 /* We shouldn't have too many configs, so this iteration shouldn't be
4700 * too costly.
4701 */
4702 idr_for_each_entry(&perf->metrics_idr, tmp, id) {
4703 if (!strcmp(tmp->uuid, oa_config->uuid)) {
4704 drm_dbg(&perf->i915->drm,
4705 "OA config already exists with this uuid\n");
4706 err = -EADDRINUSE;
4707 goto sysfs_err;
4708 }
4709 }
4710
4711 err = create_dynamic_oa_sysfs_entry(perf, oa_config);
4712 if (err) {
4713 drm_dbg(&perf->i915->drm,
4714 "Failed to create sysfs entry for OA config\n");
4715 goto sysfs_err;
4716 }
4717
4718 /* Config id 0 is invalid, id 1 for kernel stored test config. */
4719 oa_config->id = idr_alloc(&perf->metrics_idr,
4720 oa_config, 2,
4721 0, GFP_KERNEL);
4722 if (oa_config->id < 0) {
4723 drm_dbg(&perf->i915->drm,
4724 "Failed to create sysfs entry for OA config\n");
4725 err = oa_config->id;
4726 goto sysfs_err;
4727 }
4728 id = oa_config->id;
4729
4730 drm_dbg(&perf->i915->drm,
4731 "Added config %s id=%i\n", oa_config->uuid, oa_config->id);
4732 mutex_unlock(&perf->metrics_lock);
4733
4734 return id;
4735
4736sysfs_err:
4737 mutex_unlock(&perf->metrics_lock);
4738reg_err:
4739 i915_oa_config_put(oa_config);
4740 drm_dbg(&perf->i915->drm,
4741 "Failed to add new OA config\n");
4742 return err;
4743}
4744
4745/**
4746 * i915_perf_remove_config_ioctl - DRM ioctl() for userspace to remove an OA config
4747 * @dev: drm device
4748 * @data: ioctl data (pointer to u64 integer) copied from userspace
4749 * @file: drm file
4750 *
4751 * Configs can be removed while being used, the will stop appearing in sysfs
4752 * and their content will be freed when the stream using the config is closed.
4753 *
4754 * Returns: 0 on success or a negative error code on failure.
4755 */
4756int i915_perf_remove_config_ioctl(struct drm_device *dev, void *data,
4757 struct drm_file *file)
4758{
4759 struct i915_perf *perf = &to_i915(dev)->perf;
4760 u64 *arg = data;
4761 struct i915_oa_config *oa_config;
4762 int ret;
4763
4764 if (!perf->i915)
4765 return -ENOTSUPP;
4766
4767 if (i915_perf_stream_paranoid && !perfmon_capable()) {
4768 drm_dbg(&perf->i915->drm,
4769 "Insufficient privileges to remove i915 OA config\n");
4770 return -EACCES;
4771 }
4772
4773 ret = mutex_lock_interruptible(&perf->metrics_lock);
4774 if (ret)
4775 return ret;
4776
4777 oa_config = idr_find(&perf->metrics_idr, *arg);
4778 if (!oa_config) {
4779 drm_dbg(&perf->i915->drm,
4780 "Failed to remove unknown OA config\n");
4781 ret = -ENOENT;
4782 goto err_unlock;
4783 }
4784
4785 GEM_BUG_ON(*arg != oa_config->id);
4786
4787 sysfs_remove_group(perf->metrics_kobj, &oa_config->sysfs_metric);
4788
4789 idr_remove(&perf->metrics_idr, *arg);
4790
4791 mutex_unlock(&perf->metrics_lock);
4792
4793 drm_dbg(&perf->i915->drm,
4794 "Removed config %s id=%i\n", oa_config->uuid, oa_config->id);
4795
4796 i915_oa_config_put(oa_config);
4797
4798 return 0;
4799
4800err_unlock:
4801 mutex_unlock(&perf->metrics_lock);
4802 return ret;
4803}
4804
4805static struct ctl_table oa_table[] = {
4806 {
4807 .procname = "perf_stream_paranoid",
4808 .data = &i915_perf_stream_paranoid,
4809 .maxlen = sizeof(i915_perf_stream_paranoid),
4810 .mode = 0644,
4811 .proc_handler = proc_dointvec_minmax,
4812 .extra1 = SYSCTL_ZERO,
4813 .extra2 = SYSCTL_ONE,
4814 },
4815 {
4816 .procname = "oa_max_sample_rate",
4817 .data = &i915_oa_max_sample_rate,
4818 .maxlen = sizeof(i915_oa_max_sample_rate),
4819 .mode = 0644,
4820 .proc_handler = proc_dointvec_minmax,
4821 .extra1 = SYSCTL_ZERO,
4822 .extra2 = &oa_sample_rate_hard_limit,
4823 },
4824};
4825
4826static u32 num_perf_groups_per_gt(struct intel_gt *gt)
4827{
4828 return 1;
4829}
4830
4831static u32 __oam_engine_group(struct intel_engine_cs *engine)
4832{
4833 if (GRAPHICS_VER_FULL(engine->i915) >= IP_VER(12, 70)) {
4834 /*
4835 * There's 1 SAMEDIA gt and 1 OAM per SAMEDIA gt. All media slices
4836 * within the gt use the same OAM. All MTL SKUs list 1 SA MEDIA.
4837 */
4838 drm_WARN_ON(&engine->i915->drm,
4839 engine->gt->type != GT_MEDIA);
4840
4841 return PERF_GROUP_OAM_SAMEDIA_0;
4842 }
4843
4844 return PERF_GROUP_INVALID;
4845}
4846
4847static u32 __oa_engine_group(struct intel_engine_cs *engine)
4848{
4849 switch (engine->class) {
4850 case RENDER_CLASS:
4851 return PERF_GROUP_OAG;
4852
4853 case VIDEO_DECODE_CLASS:
4854 case VIDEO_ENHANCEMENT_CLASS:
4855 return __oam_engine_group(engine);
4856
4857 default:
4858 return PERF_GROUP_INVALID;
4859 }
4860}
4861
4862static struct i915_perf_regs __oam_regs(u32 base)
4863{
4864 return (struct i915_perf_regs) {
4865 base,
4866 GEN12_OAM_HEAD_POINTER(base),
4867 GEN12_OAM_TAIL_POINTER(base),
4868 GEN12_OAM_BUFFER(base),
4869 GEN12_OAM_CONTEXT_CONTROL(base),
4870 GEN12_OAM_CONTROL(base),
4871 GEN12_OAM_DEBUG(base),
4872 GEN12_OAM_STATUS(base),
4873 GEN12_OAM_CONTROL_COUNTER_FORMAT_SHIFT,
4874 };
4875}
4876
4877static struct i915_perf_regs __oag_regs(void)
4878{
4879 return (struct i915_perf_regs) {
4880 0,
4881 GEN12_OAG_OAHEADPTR,
4882 GEN12_OAG_OATAILPTR,
4883 GEN12_OAG_OABUFFER,
4884 GEN12_OAG_OAGLBCTXCTRL,
4885 GEN12_OAG_OACONTROL,
4886 GEN12_OAG_OA_DEBUG,
4887 GEN12_OAG_OASTATUS,
4888 GEN12_OAG_OACONTROL_OA_COUNTER_FORMAT_SHIFT,
4889 };
4890}
4891
4892static void oa_init_groups(struct intel_gt *gt)
4893{
4894 int i, num_groups = gt->perf.num_perf_groups;
4895
4896 for (i = 0; i < num_groups; i++) {
4897 struct i915_perf_group *g = >->perf.group[i];
4898
4899 /* Fused off engines can result in a group with num_engines == 0 */
4900 if (g->num_engines == 0)
4901 continue;
4902
4903 if (i == PERF_GROUP_OAG && gt->type != GT_MEDIA) {
4904 g->regs = __oag_regs();
4905 g->type = TYPE_OAG;
4906 } else if (GRAPHICS_VER_FULL(gt->i915) >= IP_VER(12, 70)) {
4907 g->regs = __oam_regs(mtl_oa_base[i]);
4908 g->type = TYPE_OAM;
4909 }
4910 }
4911}
4912
4913static int oa_init_gt(struct intel_gt *gt)
4914{
4915 u32 num_groups = num_perf_groups_per_gt(gt);
4916 struct intel_engine_cs *engine;
4917 struct i915_perf_group *g;
4918 intel_engine_mask_t tmp;
4919
4920 g = kcalloc(num_groups, sizeof(*g), GFP_KERNEL);
4921 if (!g)
4922 return -ENOMEM;
4923
4924 for_each_engine_masked(engine, gt, ALL_ENGINES, tmp) {
4925 u32 index = __oa_engine_group(engine);
4926
4927 engine->oa_group = NULL;
4928 if (index < num_groups) {
4929 g[index].num_engines++;
4930 engine->oa_group = &g[index];
4931 }
4932 }
4933
4934 gt->perf.num_perf_groups = num_groups;
4935 gt->perf.group = g;
4936
4937 oa_init_groups(gt);
4938
4939 return 0;
4940}
4941
4942static int oa_init_engine_groups(struct i915_perf *perf)
4943{
4944 struct intel_gt *gt;
4945 int i, ret;
4946
4947 for_each_gt(gt, perf->i915, i) {
4948 ret = oa_init_gt(gt);
4949 if (ret)
4950 return ret;
4951 }
4952
4953 return 0;
4954}
4955
4956static void oa_init_supported_formats(struct i915_perf *perf)
4957{
4958 struct drm_i915_private *i915 = perf->i915;
4959 enum intel_platform platform = INTEL_INFO(i915)->platform;
4960
4961 switch (platform) {
4962 case INTEL_HASWELL:
4963 oa_format_add(perf, I915_OA_FORMAT_A13);
4964 oa_format_add(perf, I915_OA_FORMAT_A13);
4965 oa_format_add(perf, I915_OA_FORMAT_A29);
4966 oa_format_add(perf, I915_OA_FORMAT_A13_B8_C8);
4967 oa_format_add(perf, I915_OA_FORMAT_B4_C8);
4968 oa_format_add(perf, I915_OA_FORMAT_A45_B8_C8);
4969 oa_format_add(perf, I915_OA_FORMAT_B4_C8_A16);
4970 oa_format_add(perf, I915_OA_FORMAT_C4_B8);
4971 break;
4972
4973 case INTEL_BROADWELL:
4974 case INTEL_CHERRYVIEW:
4975 case INTEL_SKYLAKE:
4976 case INTEL_BROXTON:
4977 case INTEL_KABYLAKE:
4978 case INTEL_GEMINILAKE:
4979 case INTEL_COFFEELAKE:
4980 case INTEL_COMETLAKE:
4981 case INTEL_ICELAKE:
4982 case INTEL_ELKHARTLAKE:
4983 case INTEL_JASPERLAKE:
4984 case INTEL_TIGERLAKE:
4985 case INTEL_ROCKETLAKE:
4986 case INTEL_DG1:
4987 case INTEL_ALDERLAKE_S:
4988 case INTEL_ALDERLAKE_P:
4989 oa_format_add(perf, I915_OA_FORMAT_A12);
4990 oa_format_add(perf, I915_OA_FORMAT_A12_B8_C8);
4991 oa_format_add(perf, I915_OA_FORMAT_A32u40_A4u32_B8_C8);
4992 oa_format_add(perf, I915_OA_FORMAT_C4_B8);
4993 break;
4994
4995 case INTEL_DG2:
4996 oa_format_add(perf, I915_OAR_FORMAT_A32u40_A4u32_B8_C8);
4997 oa_format_add(perf, I915_OA_FORMAT_A24u40_A14u32_B8_C8);
4998 break;
4999
5000 case INTEL_METEORLAKE:
5001 oa_format_add(perf, I915_OAR_FORMAT_A32u40_A4u32_B8_C8);
5002 oa_format_add(perf, I915_OA_FORMAT_A24u40_A14u32_B8_C8);
5003 oa_format_add(perf, I915_OAM_FORMAT_MPEC8u64_B8_C8);
5004 oa_format_add(perf, I915_OAM_FORMAT_MPEC8u32_B8_C8);
5005 break;
5006
5007 default:
5008 MISSING_CASE(platform);
5009 }
5010}
5011
5012static void i915_perf_init_info(struct drm_i915_private *i915)
5013{
5014 struct i915_perf *perf = &i915->perf;
5015
5016 switch (GRAPHICS_VER(i915)) {
5017 case 8:
5018 perf->ctx_oactxctrl_offset = 0x120;
5019 perf->ctx_flexeu0_offset = 0x2ce;
5020 perf->gen8_valid_ctx_bit = BIT(25);
5021 break;
5022 case 9:
5023 perf->ctx_oactxctrl_offset = 0x128;
5024 perf->ctx_flexeu0_offset = 0x3de;
5025 perf->gen8_valid_ctx_bit = BIT(16);
5026 break;
5027 case 11:
5028 perf->ctx_oactxctrl_offset = 0x124;
5029 perf->ctx_flexeu0_offset = 0x78e;
5030 perf->gen8_valid_ctx_bit = BIT(16);
5031 break;
5032 case 12:
5033 perf->gen8_valid_ctx_bit = BIT(16);
5034 /*
5035 * Calculate offset at runtime in oa_pin_context for gen12 and
5036 * cache the value in perf->ctx_oactxctrl_offset.
5037 */
5038 break;
5039 default:
5040 MISSING_CASE(GRAPHICS_VER(i915));
5041 }
5042}
5043
5044/**
5045 * i915_perf_init - initialize i915-perf state on module bind
5046 * @i915: i915 device instance
5047 *
5048 * Initializes i915-perf state without exposing anything to userspace.
5049 *
5050 * Note: i915-perf initialization is split into an 'init' and 'register'
5051 * phase with the i915_perf_register() exposing state to userspace.
5052 */
5053int i915_perf_init(struct drm_i915_private *i915)
5054{
5055 struct i915_perf *perf = &i915->perf;
5056
5057 perf->oa_formats = oa_formats;
5058 if (IS_HASWELL(i915)) {
5059 perf->ops.is_valid_b_counter_reg = gen7_is_valid_b_counter_addr;
5060 perf->ops.is_valid_mux_reg = hsw_is_valid_mux_addr;
5061 perf->ops.is_valid_flex_reg = NULL;
5062 perf->ops.enable_metric_set = hsw_enable_metric_set;
5063 perf->ops.disable_metric_set = hsw_disable_metric_set;
5064 perf->ops.oa_enable = gen7_oa_enable;
5065 perf->ops.oa_disable = gen7_oa_disable;
5066 perf->ops.read = gen7_oa_read;
5067 perf->ops.oa_hw_tail_read = gen7_oa_hw_tail_read;
5068 } else if (HAS_LOGICAL_RING_CONTEXTS(i915)) {
5069 /* Note: that although we could theoretically also support the
5070 * legacy ringbuffer mode on BDW (and earlier iterations of
5071 * this driver, before upstreaming did this) it didn't seem
5072 * worth the complexity to maintain now that BDW+ enable
5073 * execlist mode by default.
5074 */
5075 perf->ops.read = gen8_oa_read;
5076 i915_perf_init_info(i915);
5077
5078 if (IS_GRAPHICS_VER(i915, 8, 9)) {
5079 perf->ops.is_valid_b_counter_reg =
5080 gen7_is_valid_b_counter_addr;
5081 perf->ops.is_valid_mux_reg =
5082 gen8_is_valid_mux_addr;
5083 perf->ops.is_valid_flex_reg =
5084 gen8_is_valid_flex_addr;
5085
5086 if (IS_CHERRYVIEW(i915)) {
5087 perf->ops.is_valid_mux_reg =
5088 chv_is_valid_mux_addr;
5089 }
5090
5091 perf->ops.oa_enable = gen8_oa_enable;
5092 perf->ops.oa_disable = gen8_oa_disable;
5093 perf->ops.enable_metric_set = gen8_enable_metric_set;
5094 perf->ops.disable_metric_set = gen8_disable_metric_set;
5095 perf->ops.oa_hw_tail_read = gen8_oa_hw_tail_read;
5096 } else if (GRAPHICS_VER(i915) == 11) {
5097 perf->ops.is_valid_b_counter_reg =
5098 gen7_is_valid_b_counter_addr;
5099 perf->ops.is_valid_mux_reg =
5100 gen11_is_valid_mux_addr;
5101 perf->ops.is_valid_flex_reg =
5102 gen8_is_valid_flex_addr;
5103
5104 perf->ops.oa_enable = gen8_oa_enable;
5105 perf->ops.oa_disable = gen8_oa_disable;
5106 perf->ops.enable_metric_set = gen8_enable_metric_set;
5107 perf->ops.disable_metric_set = gen11_disable_metric_set;
5108 perf->ops.oa_hw_tail_read = gen8_oa_hw_tail_read;
5109 } else if (GRAPHICS_VER(i915) == 12) {
5110 perf->ops.is_valid_b_counter_reg =
5111 HAS_OA_SLICE_CONTRIB_LIMITS(i915) ?
5112 xehp_is_valid_b_counter_addr :
5113 gen12_is_valid_b_counter_addr;
5114 perf->ops.is_valid_mux_reg =
5115 gen12_is_valid_mux_addr;
5116 perf->ops.is_valid_flex_reg =
5117 gen8_is_valid_flex_addr;
5118
5119 perf->ops.oa_enable = gen12_oa_enable;
5120 perf->ops.oa_disable = gen12_oa_disable;
5121 perf->ops.enable_metric_set = gen12_enable_metric_set;
5122 perf->ops.disable_metric_set = gen12_disable_metric_set;
5123 perf->ops.oa_hw_tail_read = gen12_oa_hw_tail_read;
5124 }
5125 }
5126
5127 if (perf->ops.enable_metric_set) {
5128 struct intel_gt *gt;
5129 int i, ret;
5130
5131 for_each_gt(gt, i915, i)
5132 mutex_init(>->perf.lock);
5133
5134 /* Choose a representative limit */
5135 oa_sample_rate_hard_limit = to_gt(i915)->clock_frequency / 2;
5136
5137 mutex_init(&perf->metrics_lock);
5138 idr_init_base(&perf->metrics_idr, 1);
5139
5140 /* We set up some ratelimit state to potentially throttle any
5141 * _NOTES about spurious, invalid OA reports which we don't
5142 * forward to userspace.
5143 *
5144 * We print a _NOTE about any throttling when closing the
5145 * stream instead of waiting until driver _fini which no one
5146 * would ever see.
5147 *
5148 * Using the same limiting factors as printk_ratelimit()
5149 */
5150 ratelimit_state_init(&perf->spurious_report_rs, 5 * HZ, 10);
5151 /* Since we use a DRM_NOTE for spurious reports it would be
5152 * inconsistent to let __ratelimit() automatically print a
5153 * warning for throttling.
5154 */
5155 ratelimit_set_flags(&perf->spurious_report_rs,
5156 RATELIMIT_MSG_ON_RELEASE);
5157
5158 ratelimit_state_init(&perf->tail_pointer_race,
5159 5 * HZ, 10);
5160 ratelimit_set_flags(&perf->tail_pointer_race,
5161 RATELIMIT_MSG_ON_RELEASE);
5162
5163 atomic64_set(&perf->noa_programming_delay,
5164 500 * 1000 /* 500us */);
5165
5166 perf->i915 = i915;
5167
5168 ret = oa_init_engine_groups(perf);
5169 if (ret) {
5170 drm_err(&i915->drm,
5171 "OA initialization failed %d\n", ret);
5172 return ret;
5173 }
5174
5175 oa_init_supported_formats(perf);
5176 }
5177
5178 return 0;
5179}
5180
5181static int destroy_config(int id, void *p, void *data)
5182{
5183 i915_oa_config_put(p);
5184 return 0;
5185}
5186
5187int i915_perf_sysctl_register(void)
5188{
5189 sysctl_header = register_sysctl("dev/i915", oa_table);
5190 return 0;
5191}
5192
5193void i915_perf_sysctl_unregister(void)
5194{
5195 unregister_sysctl_table(sysctl_header);
5196}
5197
5198/**
5199 * i915_perf_fini - Counter part to i915_perf_init()
5200 * @i915: i915 device instance
5201 */
5202void i915_perf_fini(struct drm_i915_private *i915)
5203{
5204 struct i915_perf *perf = &i915->perf;
5205 struct intel_gt *gt;
5206 int i;
5207
5208 if (!perf->i915)
5209 return;
5210
5211 for_each_gt(gt, perf->i915, i)
5212 kfree(gt->perf.group);
5213
5214 idr_for_each(&perf->metrics_idr, destroy_config, perf);
5215 idr_destroy(&perf->metrics_idr);
5216
5217 memset(&perf->ops, 0, sizeof(perf->ops));
5218 perf->i915 = NULL;
5219}
5220
5221/**
5222 * i915_perf_ioctl_version - Version of the i915-perf subsystem
5223 * @i915: The i915 device
5224 *
5225 * This version number is used by userspace to detect available features.
5226 */
5227int i915_perf_ioctl_version(struct drm_i915_private *i915)
5228{
5229 /*
5230 * 1: Initial version
5231 * I915_PERF_IOCTL_ENABLE
5232 * I915_PERF_IOCTL_DISABLE
5233 *
5234 * 2: Added runtime modification of OA config.
5235 * I915_PERF_IOCTL_CONFIG
5236 *
5237 * 3: Add DRM_I915_PERF_PROP_HOLD_PREEMPTION parameter to hold
5238 * preemption on a particular context so that performance data is
5239 * accessible from a delta of MI_RPC reports without looking at the
5240 * OA buffer.
5241 *
5242 * 4: Add DRM_I915_PERF_PROP_ALLOWED_SSEU to limit what contexts can
5243 * be run for the duration of the performance recording based on
5244 * their SSEU configuration.
5245 *
5246 * 5: Add DRM_I915_PERF_PROP_POLL_OA_PERIOD parameter that controls the
5247 * interval for the hrtimer used to check for OA data.
5248 *
5249 * 6: Add DRM_I915_PERF_PROP_OA_ENGINE_CLASS and
5250 * DRM_I915_PERF_PROP_OA_ENGINE_INSTANCE
5251 *
5252 * 7: Add support for video decode and enhancement classes.
5253 */
5254
5255 /*
5256 * Wa_14017512683: mtl[a0..c0): Use of OAM must be preceded with Media
5257 * C6 disable in BIOS. If Media C6 is enabled in BIOS, return version 6
5258 * to indicate that OA media is not supported.
5259 */
5260 if (IS_MEDIA_GT_IP_STEP(i915->media_gt, IP_VER(13, 0), STEP_A0, STEP_C0) &&
5261 intel_check_bios_c6_setup(&i915->media_gt->rc6))
5262 return 6;
5263
5264 return 7;
5265}
5266
5267#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
5268#include "selftests/i915_perf.c"
5269#endif
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_execlists_submission.h"
202#include "gt/intel_gpu_commands.h"
203#include "gt/intel_gt.h"
204#include "gt/intel_gt_clock_utils.h"
205#include "gt/intel_lrc.h"
206#include "gt/intel_ring.h"
207
208#include "i915_drv.h"
209#include "i915_perf.h"
210
211/* HW requires this to be a power of two, between 128k and 16M, though driver
212 * is currently generally designed assuming the largest 16M size is used such
213 * that the overflow cases are unlikely in normal operation.
214 */
215#define OA_BUFFER_SIZE SZ_16M
216
217#define OA_TAKEN(tail, head) ((tail - head) & (OA_BUFFER_SIZE - 1))
218
219/**
220 * DOC: OA Tail Pointer Race
221 *
222 * There's a HW race condition between OA unit tail pointer register updates and
223 * writes to memory whereby the tail pointer can sometimes get ahead of what's
224 * been written out to the OA buffer so far (in terms of what's visible to the
225 * CPU).
226 *
227 * Although this can be observed explicitly while copying reports to userspace
228 * by checking for a zeroed report-id field in tail reports, we want to account
229 * for this earlier, as part of the oa_buffer_check_unlocked to avoid lots of
230 * redundant read() attempts.
231 *
232 * We workaround this issue in oa_buffer_check_unlocked() by reading the reports
233 * in the OA buffer, starting from the tail reported by the HW until we find a
234 * report with its first 2 dwords not 0 meaning its previous report is
235 * completely in memory and ready to be read. Those dwords are also set to 0
236 * once read and the whole buffer is cleared upon OA buffer initialization. The
237 * first dword is the reason for this report while the second is the timestamp,
238 * making the chances of having those 2 fields at 0 fairly unlikely. A more
239 * detailed explanation is available in oa_buffer_check_unlocked().
240 *
241 * Most of the implementation details for this workaround are in
242 * oa_buffer_check_unlocked() and _append_oa_reports()
243 *
244 * Note for posterity: previously the driver used to define an effective tail
245 * pointer that lagged the real pointer by a 'tail margin' measured in bytes
246 * derived from %OA_TAIL_MARGIN_NSEC and the configured sampling frequency.
247 * This was flawed considering that the OA unit may also automatically generate
248 * non-periodic reports (such as on context switch) or the OA unit may be
249 * enabled without any periodic sampling.
250 */
251#define OA_TAIL_MARGIN_NSEC 100000ULL
252#define INVALID_TAIL_PTR 0xffffffff
253
254/* The default frequency for checking whether the OA unit has written new
255 * reports to the circular OA buffer...
256 */
257#define DEFAULT_POLL_FREQUENCY_HZ 200
258#define DEFAULT_POLL_PERIOD_NS (NSEC_PER_SEC / DEFAULT_POLL_FREQUENCY_HZ)
259
260/* for sysctl proc_dointvec_minmax of dev.i915.perf_stream_paranoid */
261static u32 i915_perf_stream_paranoid = true;
262
263/* The maximum exponent the hardware accepts is 63 (essentially it selects one
264 * of the 64bit timestamp bits to trigger reports from) but there's currently
265 * no known use case for sampling as infrequently as once per 47 thousand years.
266 *
267 * Since the timestamps included in OA reports are only 32bits it seems
268 * reasonable to limit the OA exponent where it's still possible to account for
269 * overflow in OA report timestamps.
270 */
271#define OA_EXPONENT_MAX 31
272
273#define INVALID_CTX_ID 0xffffffff
274
275/* On Gen8+ automatically triggered OA reports include a 'reason' field... */
276#define OAREPORT_REASON_MASK 0x3f
277#define OAREPORT_REASON_MASK_EXTENDED 0x7f
278#define OAREPORT_REASON_SHIFT 19
279#define OAREPORT_REASON_TIMER (1<<0)
280#define OAREPORT_REASON_CTX_SWITCH (1<<3)
281#define OAREPORT_REASON_CLK_RATIO (1<<5)
282
283
284/* For sysctl proc_dointvec_minmax of i915_oa_max_sample_rate
285 *
286 * The highest sampling frequency we can theoretically program the OA unit
287 * with is always half the timestamp frequency: E.g. 6.25Mhz for Haswell.
288 *
289 * Initialized just before we register the sysctl parameter.
290 */
291static int oa_sample_rate_hard_limit;
292
293/* Theoretically we can program the OA unit to sample every 160ns but don't
294 * allow that by default unless root...
295 *
296 * The default threshold of 100000Hz is based on perf's similar
297 * kernel.perf_event_max_sample_rate sysctl parameter.
298 */
299static u32 i915_oa_max_sample_rate = 100000;
300
301/* XXX: beware if future OA HW adds new report formats that the current
302 * code assumes all reports have a power-of-two size and ~(size - 1) can
303 * be used as a mask to align the OA tail pointer.
304 */
305static const struct i915_oa_format oa_formats[I915_OA_FORMAT_MAX] = {
306 [I915_OA_FORMAT_A13] = { 0, 64 },
307 [I915_OA_FORMAT_A29] = { 1, 128 },
308 [I915_OA_FORMAT_A13_B8_C8] = { 2, 128 },
309 /* A29_B8_C8 Disallowed as 192 bytes doesn't factor into buffer size */
310 [I915_OA_FORMAT_B4_C8] = { 4, 64 },
311 [I915_OA_FORMAT_A45_B8_C8] = { 5, 256 },
312 [I915_OA_FORMAT_B4_C8_A16] = { 6, 128 },
313 [I915_OA_FORMAT_C4_B8] = { 7, 64 },
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};
318
319#define SAMPLE_OA_REPORT (1<<0)
320
321/**
322 * struct perf_open_properties - for validated properties given to open a stream
323 * @sample_flags: `DRM_I915_PERF_PROP_SAMPLE_*` properties are tracked as flags
324 * @single_context: Whether a single or all gpu contexts should be monitored
325 * @hold_preemption: Whether the preemption is disabled for the filtered
326 * context
327 * @ctx_handle: A gem ctx handle for use with @single_context
328 * @metrics_set: An ID for an OA unit metric set advertised via sysfs
329 * @oa_format: An OA unit HW report format
330 * @oa_periodic: Whether to enable periodic OA unit sampling
331 * @oa_period_exponent: The OA unit sampling period is derived from this
332 * @engine: The engine (typically rcs0) being monitored by the OA unit
333 * @has_sseu: Whether @sseu was specified by userspace
334 * @sseu: internal SSEU configuration computed either from the userspace
335 * specified configuration in the opening parameters or a default value
336 * (see get_default_sseu_config())
337 * @poll_oa_period: The period in nanoseconds at which the CPU will check for OA
338 * data availability
339 *
340 * As read_properties_unlocked() enumerates and validates the properties given
341 * to open a stream of metrics the configuration is built up in the structure
342 * which starts out zero initialized.
343 */
344struct perf_open_properties {
345 u32 sample_flags;
346
347 u64 single_context:1;
348 u64 hold_preemption:1;
349 u64 ctx_handle;
350
351 /* OA sampling state */
352 int metrics_set;
353 int oa_format;
354 bool oa_periodic;
355 int oa_period_exponent;
356
357 struct intel_engine_cs *engine;
358
359 bool has_sseu;
360 struct intel_sseu sseu;
361
362 u64 poll_oa_period;
363};
364
365struct i915_oa_config_bo {
366 struct llist_node node;
367
368 struct i915_oa_config *oa_config;
369 struct i915_vma *vma;
370};
371
372static struct ctl_table_header *sysctl_header;
373
374static enum hrtimer_restart oa_poll_check_timer_cb(struct hrtimer *hrtimer);
375
376void i915_oa_config_release(struct kref *ref)
377{
378 struct i915_oa_config *oa_config =
379 container_of(ref, typeof(*oa_config), ref);
380
381 kfree(oa_config->flex_regs);
382 kfree(oa_config->b_counter_regs);
383 kfree(oa_config->mux_regs);
384
385 kfree_rcu(oa_config, rcu);
386}
387
388struct i915_oa_config *
389i915_perf_get_oa_config(struct i915_perf *perf, int metrics_set)
390{
391 struct i915_oa_config *oa_config;
392
393 rcu_read_lock();
394 oa_config = idr_find(&perf->metrics_idr, metrics_set);
395 if (oa_config)
396 oa_config = i915_oa_config_get(oa_config);
397 rcu_read_unlock();
398
399 return oa_config;
400}
401
402static void free_oa_config_bo(struct i915_oa_config_bo *oa_bo)
403{
404 i915_oa_config_put(oa_bo->oa_config);
405 i915_vma_put(oa_bo->vma);
406 kfree(oa_bo);
407}
408
409static u32 gen12_oa_hw_tail_read(struct i915_perf_stream *stream)
410{
411 struct intel_uncore *uncore = stream->uncore;
412
413 return intel_uncore_read(uncore, GEN12_OAG_OATAILPTR) &
414 GEN12_OAG_OATAILPTR_MASK;
415}
416
417static u32 gen8_oa_hw_tail_read(struct i915_perf_stream *stream)
418{
419 struct intel_uncore *uncore = stream->uncore;
420
421 return intel_uncore_read(uncore, GEN8_OATAILPTR) & GEN8_OATAILPTR_MASK;
422}
423
424static u32 gen7_oa_hw_tail_read(struct i915_perf_stream *stream)
425{
426 struct intel_uncore *uncore = stream->uncore;
427 u32 oastatus1 = intel_uncore_read(uncore, GEN7_OASTATUS1);
428
429 return oastatus1 & GEN7_OASTATUS1_TAIL_MASK;
430}
431
432/**
433 * oa_buffer_check_unlocked - check for data and update tail ptr state
434 * @stream: i915 stream instance
435 *
436 * This is either called via fops (for blocking reads in user ctx) or the poll
437 * check hrtimer (atomic ctx) to check the OA buffer tail pointer and check
438 * if there is data available for userspace to read.
439 *
440 * This function is central to providing a workaround for the OA unit tail
441 * pointer having a race with respect to what data is visible to the CPU.
442 * It is responsible for reading tail pointers from the hardware and giving
443 * the pointers time to 'age' before they are made available for reading.
444 * (See description of OA_TAIL_MARGIN_NSEC above for further details.)
445 *
446 * Besides returning true when there is data available to read() this function
447 * also updates the tail, aging_tail and aging_timestamp in the oa_buffer
448 * object.
449 *
450 * Note: It's safe to read OA config state here unlocked, assuming that this is
451 * only called while the stream is enabled, while the global OA configuration
452 * can't be modified.
453 *
454 * Returns: %true if the OA buffer contains data, else %false
455 */
456static bool oa_buffer_check_unlocked(struct i915_perf_stream *stream)
457{
458 u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma);
459 int report_size = stream->oa_buffer.format_size;
460 unsigned long flags;
461 bool pollin;
462 u32 hw_tail;
463 u64 now;
464
465 /* We have to consider the (unlikely) possibility that read() errors
466 * could result in an OA buffer reset which might reset the head and
467 * tail state.
468 */
469 spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
470
471 hw_tail = stream->perf->ops.oa_hw_tail_read(stream);
472
473 /* The tail pointer increases in 64 byte increments,
474 * not in report_size steps...
475 */
476 hw_tail &= ~(report_size - 1);
477
478 now = ktime_get_mono_fast_ns();
479
480 if (hw_tail == stream->oa_buffer.aging_tail &&
481 (now - stream->oa_buffer.aging_timestamp) > OA_TAIL_MARGIN_NSEC) {
482 /* If the HW tail hasn't move since the last check and the HW
483 * tail has been aging for long enough, declare it the new
484 * tail.
485 */
486 stream->oa_buffer.tail = stream->oa_buffer.aging_tail;
487 } else {
488 u32 head, tail, aged_tail;
489
490 /* NB: The head we observe here might effectively be a little
491 * out of date. If a read() is in progress, the head could be
492 * anywhere between this head and stream->oa_buffer.tail.
493 */
494 head = stream->oa_buffer.head - gtt_offset;
495 aged_tail = stream->oa_buffer.tail - gtt_offset;
496
497 hw_tail -= gtt_offset;
498 tail = hw_tail;
499
500 /* Walk the stream backward until we find a report with dword 0
501 * & 1 not at 0. Since the circular buffer pointers progress by
502 * increments of 64 bytes and that reports can be up to 256
503 * bytes long, we can't tell whether a report has fully landed
504 * in memory before the first 2 dwords of the following report
505 * have effectively landed.
506 *
507 * This is assuming that the writes of the OA unit land in
508 * memory in the order they were written to.
509 * If not : (╯°□°）╯︵ ┻━┻
510 */
511 while (OA_TAKEN(tail, aged_tail) >= report_size) {
512 u32 *report32 = (void *)(stream->oa_buffer.vaddr + tail);
513
514 if (report32[0] != 0 || report32[1] != 0)
515 break;
516
517 tail = (tail - report_size) & (OA_BUFFER_SIZE - 1);
518 }
519
520 if (OA_TAKEN(hw_tail, tail) > report_size &&
521 __ratelimit(&stream->perf->tail_pointer_race))
522 DRM_NOTE("unlanded report(s) head=0x%x "
523 "tail=0x%x hw_tail=0x%x\n",
524 head, tail, hw_tail);
525
526 stream->oa_buffer.tail = gtt_offset + tail;
527 stream->oa_buffer.aging_tail = gtt_offset + hw_tail;
528 stream->oa_buffer.aging_timestamp = now;
529 }
530
531 pollin = OA_TAKEN(stream->oa_buffer.tail - gtt_offset,
532 stream->oa_buffer.head - gtt_offset) >= report_size;
533
534 spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
535
536 return pollin;
537}
538
539/**
540 * append_oa_status - Appends a status record to a userspace read() buffer.
541 * @stream: An i915-perf stream opened for OA metrics
542 * @buf: destination buffer given by userspace
543 * @count: the number of bytes userspace wants to read
544 * @offset: (inout): the current position for writing into @buf
545 * @type: The kind of status to report to userspace
546 *
547 * Writes a status record (such as `DRM_I915_PERF_RECORD_OA_REPORT_LOST`)
548 * into the userspace read() buffer.
549 *
550 * The @buf @offset will only be updated on success.
551 *
552 * Returns: 0 on success, negative error code on failure.
553 */
554static int append_oa_status(struct i915_perf_stream *stream,
555 char __user *buf,
556 size_t count,
557 size_t *offset,
558 enum drm_i915_perf_record_type type)
559{
560 struct drm_i915_perf_record_header header = { type, 0, sizeof(header) };
561
562 if ((count - *offset) < header.size)
563 return -ENOSPC;
564
565 if (copy_to_user(buf + *offset, &header, sizeof(header)))
566 return -EFAULT;
567
568 (*offset) += header.size;
569
570 return 0;
571}
572
573/**
574 * append_oa_sample - Copies single OA report into userspace read() buffer.
575 * @stream: An i915-perf stream opened for OA metrics
576 * @buf: destination buffer given by userspace
577 * @count: the number of bytes userspace wants to read
578 * @offset: (inout): the current position for writing into @buf
579 * @report: A single OA report to (optionally) include as part of the sample
580 *
581 * The contents of a sample are configured through `DRM_I915_PERF_PROP_SAMPLE_*`
582 * properties when opening a stream, tracked as `stream->sample_flags`. This
583 * function copies the requested components of a single sample to the given
584 * read() @buf.
585 *
586 * The @buf @offset will only be updated on success.
587 *
588 * Returns: 0 on success, negative error code on failure.
589 */
590static int append_oa_sample(struct i915_perf_stream *stream,
591 char __user *buf,
592 size_t count,
593 size_t *offset,
594 const u8 *report)
595{
596 int report_size = stream->oa_buffer.format_size;
597 struct drm_i915_perf_record_header header;
598
599 header.type = DRM_I915_PERF_RECORD_SAMPLE;
600 header.pad = 0;
601 header.size = stream->sample_size;
602
603 if ((count - *offset) < header.size)
604 return -ENOSPC;
605
606 buf += *offset;
607 if (copy_to_user(buf, &header, sizeof(header)))
608 return -EFAULT;
609 buf += sizeof(header);
610
611 if (copy_to_user(buf, report, report_size))
612 return -EFAULT;
613
614 (*offset) += header.size;
615
616 return 0;
617}
618
619/**
620 * gen8_append_oa_reports - Copies all buffered OA reports into
621 * userspace read() buffer.
622 * @stream: An i915-perf stream opened for OA metrics
623 * @buf: destination buffer given by userspace
624 * @count: the number of bytes userspace wants to read
625 * @offset: (inout): the current position for writing into @buf
626 *
627 * Notably any error condition resulting in a short read (-%ENOSPC or
628 * -%EFAULT) will be returned even though one or more records may
629 * have been successfully copied. In this case it's up to the caller
630 * to decide if the error should be squashed before returning to
631 * userspace.
632 *
633 * Note: reports are consumed from the head, and appended to the
634 * tail, so the tail chases the head?... If you think that's mad
635 * and back-to-front you're not alone, but this follows the
636 * Gen PRM naming convention.
637 *
638 * Returns: 0 on success, negative error code on failure.
639 */
640static int gen8_append_oa_reports(struct i915_perf_stream *stream,
641 char __user *buf,
642 size_t count,
643 size_t *offset)
644{
645 struct intel_uncore *uncore = stream->uncore;
646 int report_size = stream->oa_buffer.format_size;
647 u8 *oa_buf_base = stream->oa_buffer.vaddr;
648 u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma);
649 u32 mask = (OA_BUFFER_SIZE - 1);
650 size_t start_offset = *offset;
651 unsigned long flags;
652 u32 head, tail;
653 u32 taken;
654 int ret = 0;
655
656 if (drm_WARN_ON(&uncore->i915->drm, !stream->enabled))
657 return -EIO;
658
659 spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
660
661 head = stream->oa_buffer.head;
662 tail = stream->oa_buffer.tail;
663
664 spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
665
666 /*
667 * NB: oa_buffer.head/tail include the gtt_offset which we don't want
668 * while indexing relative to oa_buf_base.
669 */
670 head -= gtt_offset;
671 tail -= gtt_offset;
672
673 /*
674 * An out of bounds or misaligned head or tail pointer implies a driver
675 * bug since we validate + align the tail pointers we read from the
676 * hardware and we are in full control of the head pointer which should
677 * only be incremented by multiples of the report size (notably also
678 * all a power of two).
679 */
680 if (drm_WARN_ONCE(&uncore->i915->drm,
681 head > OA_BUFFER_SIZE || head % report_size ||
682 tail > OA_BUFFER_SIZE || tail % report_size,
683 "Inconsistent OA buffer pointers: head = %u, tail = %u\n",
684 head, tail))
685 return -EIO;
686
687
688 for (/* none */;
689 (taken = OA_TAKEN(tail, head));
690 head = (head + report_size) & mask) {
691 u8 *report = oa_buf_base + head;
692 u32 *report32 = (void *)report;
693 u32 ctx_id;
694 u32 reason;
695
696 /*
697 * All the report sizes factor neatly into the buffer
698 * size so we never expect to see a report split
699 * between the beginning and end of the buffer.
700 *
701 * Given the initial alignment check a misalignment
702 * here would imply a driver bug that would result
703 * in an overrun.
704 */
705 if (drm_WARN_ON(&uncore->i915->drm,
706 (OA_BUFFER_SIZE - head) < report_size)) {
707 drm_err(&uncore->i915->drm,
708 "Spurious OA head ptr: non-integral report offset\n");
709 break;
710 }
711
712 /*
713 * The reason field includes flags identifying what
714 * triggered this specific report (mostly timer
715 * triggered or e.g. due to a context switch).
716 *
717 * This field is never expected to be zero so we can
718 * check that the report isn't invalid before copying
719 * it to userspace...
720 */
721 reason = ((report32[0] >> OAREPORT_REASON_SHIFT) &
722 (GRAPHICS_VER(stream->perf->i915) == 12 ?
723 OAREPORT_REASON_MASK_EXTENDED :
724 OAREPORT_REASON_MASK));
725
726 ctx_id = report32[2] & stream->specific_ctx_id_mask;
727
728 /*
729 * Squash whatever is in the CTX_ID field if it's marked as
730 * invalid to be sure we avoid false-positive, single-context
731 * filtering below...
732 *
733 * Note: that we don't clear the valid_ctx_bit so userspace can
734 * understand that the ID has been squashed by the kernel.
735 */
736 if (!(report32[0] & stream->perf->gen8_valid_ctx_bit) &&
737 GRAPHICS_VER(stream->perf->i915) <= 11)
738 ctx_id = report32[2] = INVALID_CTX_ID;
739
740 /*
741 * NB: For Gen 8 the OA unit no longer supports clock gating
742 * off for a specific context and the kernel can't securely
743 * stop the counters from updating as system-wide / global
744 * values.
745 *
746 * Automatic reports now include a context ID so reports can be
747 * filtered on the cpu but it's not worth trying to
748 * automatically subtract/hide counter progress for other
749 * contexts while filtering since we can't stop userspace
750 * issuing MI_REPORT_PERF_COUNT commands which would still
751 * provide a side-band view of the real values.
752 *
753 * To allow userspace (such as Mesa/GL_INTEL_performance_query)
754 * to normalize counters for a single filtered context then it
755 * needs be forwarded bookend context-switch reports so that it
756 * can track switches in between MI_REPORT_PERF_COUNT commands
757 * and can itself subtract/ignore the progress of counters
758 * associated with other contexts. Note that the hardware
759 * automatically triggers reports when switching to a new
760 * context which are tagged with the ID of the newly active
761 * context. To avoid the complexity (and likely fragility) of
762 * reading ahead while parsing reports to try and minimize
763 * forwarding redundant context switch reports (i.e. between
764 * other, unrelated contexts) we simply elect to forward them
765 * all.
766 *
767 * We don't rely solely on the reason field to identify context
768 * switches since it's not-uncommon for periodic samples to
769 * identify a switch before any 'context switch' report.
770 */
771 if (!stream->perf->exclusive_stream->ctx ||
772 stream->specific_ctx_id == ctx_id ||
773 stream->oa_buffer.last_ctx_id == stream->specific_ctx_id ||
774 reason & OAREPORT_REASON_CTX_SWITCH) {
775
776 /*
777 * While filtering for a single context we avoid
778 * leaking the IDs of other contexts.
779 */
780 if (stream->perf->exclusive_stream->ctx &&
781 stream->specific_ctx_id != ctx_id) {
782 report32[2] = INVALID_CTX_ID;
783 }
784
785 ret = append_oa_sample(stream, buf, count, offset,
786 report);
787 if (ret)
788 break;
789
790 stream->oa_buffer.last_ctx_id = ctx_id;
791 }
792
793 /*
794 * Clear out the first 2 dword as a mean to detect unlanded
795 * reports.
796 */
797 report32[0] = 0;
798 report32[1] = 0;
799 }
800
801 if (start_offset != *offset) {
802 i915_reg_t oaheadptr;
803
804 oaheadptr = GRAPHICS_VER(stream->perf->i915) == 12 ?
805 GEN12_OAG_OAHEADPTR : GEN8_OAHEADPTR;
806
807 spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
808
809 /*
810 * We removed the gtt_offset for the copy loop above, indexing
811 * relative to oa_buf_base so put back here...
812 */
813 head += gtt_offset;
814 intel_uncore_write(uncore, oaheadptr,
815 head & GEN12_OAG_OAHEADPTR_MASK);
816 stream->oa_buffer.head = head;
817
818 spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
819 }
820
821 return ret;
822}
823
824/**
825 * gen8_oa_read - copy status records then buffered OA reports
826 * @stream: An i915-perf stream opened for OA metrics
827 * @buf: destination buffer given by userspace
828 * @count: the number of bytes userspace wants to read
829 * @offset: (inout): the current position for writing into @buf
830 *
831 * Checks OA unit status registers and if necessary appends corresponding
832 * status records for userspace (such as for a buffer full condition) and then
833 * initiate appending any buffered OA reports.
834 *
835 * Updates @offset according to the number of bytes successfully copied into
836 * the userspace buffer.
837 *
838 * NB: some data may be successfully copied to the userspace buffer
839 * even if an error is returned, and this is reflected in the
840 * updated @offset.
841 *
842 * Returns: zero on success or a negative error code
843 */
844static int gen8_oa_read(struct i915_perf_stream *stream,
845 char __user *buf,
846 size_t count,
847 size_t *offset)
848{
849 struct intel_uncore *uncore = stream->uncore;
850 u32 oastatus;
851 i915_reg_t oastatus_reg;
852 int ret;
853
854 if (drm_WARN_ON(&uncore->i915->drm, !stream->oa_buffer.vaddr))
855 return -EIO;
856
857 oastatus_reg = GRAPHICS_VER(stream->perf->i915) == 12 ?
858 GEN12_OAG_OASTATUS : GEN8_OASTATUS;
859
860 oastatus = intel_uncore_read(uncore, oastatus_reg);
861
862 /*
863 * We treat OABUFFER_OVERFLOW as a significant error:
864 *
865 * Although theoretically we could handle this more gracefully
866 * sometimes, some Gens don't correctly suppress certain
867 * automatically triggered reports in this condition and so we
868 * have to assume that old reports are now being trampled
869 * over.
870 *
871 * Considering how we don't currently give userspace control
872 * over the OA buffer size and always configure a large 16MB
873 * buffer, then a buffer overflow does anyway likely indicate
874 * that something has gone quite badly wrong.
875 */
876 if (oastatus & GEN8_OASTATUS_OABUFFER_OVERFLOW) {
877 ret = append_oa_status(stream, buf, count, offset,
878 DRM_I915_PERF_RECORD_OA_BUFFER_LOST);
879 if (ret)
880 return ret;
881
882 DRM_DEBUG("OA buffer overflow (exponent = %d): force restart\n",
883 stream->period_exponent);
884
885 stream->perf->ops.oa_disable(stream);
886 stream->perf->ops.oa_enable(stream);
887
888 /*
889 * Note: .oa_enable() is expected to re-init the oabuffer and
890 * reset GEN8_OASTATUS for us
891 */
892 oastatus = intel_uncore_read(uncore, oastatus_reg);
893 }
894
895 if (oastatus & GEN8_OASTATUS_REPORT_LOST) {
896 ret = append_oa_status(stream, buf, count, offset,
897 DRM_I915_PERF_RECORD_OA_REPORT_LOST);
898 if (ret)
899 return ret;
900
901 intel_uncore_rmw(uncore, oastatus_reg,
902 GEN8_OASTATUS_COUNTER_OVERFLOW |
903 GEN8_OASTATUS_REPORT_LOST,
904 IS_GRAPHICS_VER(uncore->i915, 8, 11) ?
905 (GEN8_OASTATUS_HEAD_POINTER_WRAP |
906 GEN8_OASTATUS_TAIL_POINTER_WRAP) : 0);
907 }
908
909 return gen8_append_oa_reports(stream, buf, count, offset);
910}
911
912/**
913 * gen7_append_oa_reports - Copies all buffered OA reports into
914 * userspace read() buffer.
915 * @stream: An i915-perf stream opened for OA metrics
916 * @buf: destination buffer given by userspace
917 * @count: the number of bytes userspace wants to read
918 * @offset: (inout): the current position for writing into @buf
919 *
920 * Notably any error condition resulting in a short read (-%ENOSPC or
921 * -%EFAULT) will be returned even though one or more records may
922 * have been successfully copied. In this case it's up to the caller
923 * to decide if the error should be squashed before returning to
924 * userspace.
925 *
926 * Note: reports are consumed from the head, and appended to the
927 * tail, so the tail chases the head?... If you think that's mad
928 * and back-to-front you're not alone, but this follows the
929 * Gen PRM naming convention.
930 *
931 * Returns: 0 on success, negative error code on failure.
932 */
933static int gen7_append_oa_reports(struct i915_perf_stream *stream,
934 char __user *buf,
935 size_t count,
936 size_t *offset)
937{
938 struct intel_uncore *uncore = stream->uncore;
939 int report_size = stream->oa_buffer.format_size;
940 u8 *oa_buf_base = stream->oa_buffer.vaddr;
941 u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma);
942 u32 mask = (OA_BUFFER_SIZE - 1);
943 size_t start_offset = *offset;
944 unsigned long flags;
945 u32 head, tail;
946 u32 taken;
947 int ret = 0;
948
949 if (drm_WARN_ON(&uncore->i915->drm, !stream->enabled))
950 return -EIO;
951
952 spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
953
954 head = stream->oa_buffer.head;
955 tail = stream->oa_buffer.tail;
956
957 spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
958
959 /* NB: oa_buffer.head/tail include the gtt_offset which we don't want
960 * while indexing relative to oa_buf_base.
961 */
962 head -= gtt_offset;
963 tail -= gtt_offset;
964
965 /* An out of bounds or misaligned head or tail pointer implies a driver
966 * bug since we validate + align the tail pointers we read from the
967 * hardware and we are in full control of the head pointer which should
968 * only be incremented by multiples of the report size (notably also
969 * all a power of two).
970 */
971 if (drm_WARN_ONCE(&uncore->i915->drm,
972 head > OA_BUFFER_SIZE || head % report_size ||
973 tail > OA_BUFFER_SIZE || tail % report_size,
974 "Inconsistent OA buffer pointers: head = %u, tail = %u\n",
975 head, tail))
976 return -EIO;
977
978
979 for (/* none */;
980 (taken = OA_TAKEN(tail, head));
981 head = (head + report_size) & mask) {
982 u8 *report = oa_buf_base + head;
983 u32 *report32 = (void *)report;
984
985 /* All the report sizes factor neatly into the buffer
986 * size so we never expect to see a report split
987 * between the beginning and end of the buffer.
988 *
989 * Given the initial alignment check a misalignment
990 * here would imply a driver bug that would result
991 * in an overrun.
992 */
993 if (drm_WARN_ON(&uncore->i915->drm,
994 (OA_BUFFER_SIZE - head) < report_size)) {
995 drm_err(&uncore->i915->drm,
996 "Spurious OA head ptr: non-integral report offset\n");
997 break;
998 }
999
1000 /* The report-ID field for periodic samples includes
1001 * some undocumented flags related to what triggered
1002 * the report and is never expected to be zero so we
1003 * can check that the report isn't invalid before
1004 * copying it to userspace...
1005 */
1006 if (report32[0] == 0) {
1007 if (__ratelimit(&stream->perf->spurious_report_rs))
1008 DRM_NOTE("Skipping spurious, invalid OA report\n");
1009 continue;
1010 }
1011
1012 ret = append_oa_sample(stream, buf, count, offset, report);
1013 if (ret)
1014 break;
1015
1016 /* Clear out the first 2 dwords as a mean to detect unlanded
1017 * reports.
1018 */
1019 report32[0] = 0;
1020 report32[1] = 0;
1021 }
1022
1023 if (start_offset != *offset) {
1024 spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
1025
1026 /* We removed the gtt_offset for the copy loop above, indexing
1027 * relative to oa_buf_base so put back here...
1028 */
1029 head += gtt_offset;
1030
1031 intel_uncore_write(uncore, GEN7_OASTATUS2,
1032 (head & GEN7_OASTATUS2_HEAD_MASK) |
1033 GEN7_OASTATUS2_MEM_SELECT_GGTT);
1034 stream->oa_buffer.head = head;
1035
1036 spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
1037 }
1038
1039 return ret;
1040}
1041
1042/**
1043 * gen7_oa_read - copy status records then buffered OA reports
1044 * @stream: An i915-perf stream opened for OA metrics
1045 * @buf: destination buffer given by userspace
1046 * @count: the number of bytes userspace wants to read
1047 * @offset: (inout): the current position for writing into @buf
1048 *
1049 * Checks Gen 7 specific OA unit status registers and if necessary appends
1050 * corresponding status records for userspace (such as for a buffer full
1051 * condition) and then initiate appending any buffered OA reports.
1052 *
1053 * Updates @offset according to the number of bytes successfully copied into
1054 * the userspace buffer.
1055 *
1056 * Returns: zero on success or a negative error code
1057 */
1058static int gen7_oa_read(struct i915_perf_stream *stream,
1059 char __user *buf,
1060 size_t count,
1061 size_t *offset)
1062{
1063 struct intel_uncore *uncore = stream->uncore;
1064 u32 oastatus1;
1065 int ret;
1066
1067 if (drm_WARN_ON(&uncore->i915->drm, !stream->oa_buffer.vaddr))
1068 return -EIO;
1069
1070 oastatus1 = intel_uncore_read(uncore, GEN7_OASTATUS1);
1071
1072 /* XXX: On Haswell we don't have a safe way to clear oastatus1
1073 * bits while the OA unit is enabled (while the tail pointer
1074 * may be updated asynchronously) so we ignore status bits
1075 * that have already been reported to userspace.
1076 */
1077 oastatus1 &= ~stream->perf->gen7_latched_oastatus1;
1078
1079 /* We treat OABUFFER_OVERFLOW as a significant error:
1080 *
1081 * - The status can be interpreted to mean that the buffer is
1082 * currently full (with a higher precedence than OA_TAKEN()
1083 * which will start to report a near-empty buffer after an
1084 * overflow) but it's awkward that we can't clear the status
1085 * on Haswell, so without a reset we won't be able to catch
1086 * the state again.
1087 *
1088 * - Since it also implies the HW has started overwriting old
1089 * reports it may also affect our sanity checks for invalid
1090 * reports when copying to userspace that assume new reports
1091 * are being written to cleared memory.
1092 *
1093 * - In the future we may want to introduce a flight recorder
1094 * mode where the driver will automatically maintain a safe
1095 * guard band between head/tail, avoiding this overflow
1096 * condition, but we avoid the added driver complexity for
1097 * now.
1098 */
1099 if (unlikely(oastatus1 & GEN7_OASTATUS1_OABUFFER_OVERFLOW)) {
1100 ret = append_oa_status(stream, buf, count, offset,
1101 DRM_I915_PERF_RECORD_OA_BUFFER_LOST);
1102 if (ret)
1103 return ret;
1104
1105 DRM_DEBUG("OA buffer overflow (exponent = %d): force restart\n",
1106 stream->period_exponent);
1107
1108 stream->perf->ops.oa_disable(stream);
1109 stream->perf->ops.oa_enable(stream);
1110
1111 oastatus1 = intel_uncore_read(uncore, GEN7_OASTATUS1);
1112 }
1113
1114 if (unlikely(oastatus1 & GEN7_OASTATUS1_REPORT_LOST)) {
1115 ret = append_oa_status(stream, buf, count, offset,
1116 DRM_I915_PERF_RECORD_OA_REPORT_LOST);
1117 if (ret)
1118 return ret;
1119 stream->perf->gen7_latched_oastatus1 |=
1120 GEN7_OASTATUS1_REPORT_LOST;
1121 }
1122
1123 return gen7_append_oa_reports(stream, buf, count, offset);
1124}
1125
1126/**
1127 * i915_oa_wait_unlocked - handles blocking IO until OA data available
1128 * @stream: An i915-perf stream opened for OA metrics
1129 *
1130 * Called when userspace tries to read() from a blocking stream FD opened
1131 * for OA metrics. It waits until the hrtimer callback finds a non-empty
1132 * OA buffer and wakes us.
1133 *
1134 * Note: it's acceptable to have this return with some false positives
1135 * since any subsequent read handling will return -EAGAIN if there isn't
1136 * really data ready for userspace yet.
1137 *
1138 * Returns: zero on success or a negative error code
1139 */
1140static int i915_oa_wait_unlocked(struct i915_perf_stream *stream)
1141{
1142 /* We would wait indefinitely if periodic sampling is not enabled */
1143 if (!stream->periodic)
1144 return -EIO;
1145
1146 return wait_event_interruptible(stream->poll_wq,
1147 oa_buffer_check_unlocked(stream));
1148}
1149
1150/**
1151 * i915_oa_poll_wait - call poll_wait() for an OA stream poll()
1152 * @stream: An i915-perf stream opened for OA metrics
1153 * @file: An i915 perf stream file
1154 * @wait: poll() state table
1155 *
1156 * For handling userspace polling on an i915 perf stream opened for OA metrics,
1157 * this starts a poll_wait with the wait queue that our hrtimer callback wakes
1158 * when it sees data ready to read in the circular OA buffer.
1159 */
1160static void i915_oa_poll_wait(struct i915_perf_stream *stream,
1161 struct file *file,
1162 poll_table *wait)
1163{
1164 poll_wait(file, &stream->poll_wq, wait);
1165}
1166
1167/**
1168 * i915_oa_read - just calls through to &i915_oa_ops->read
1169 * @stream: An i915-perf stream opened for OA metrics
1170 * @buf: destination buffer given by userspace
1171 * @count: the number of bytes userspace wants to read
1172 * @offset: (inout): the current position for writing into @buf
1173 *
1174 * Updates @offset according to the number of bytes successfully copied into
1175 * the userspace buffer.
1176 *
1177 * Returns: zero on success or a negative error code
1178 */
1179static int i915_oa_read(struct i915_perf_stream *stream,
1180 char __user *buf,
1181 size_t count,
1182 size_t *offset)
1183{
1184 return stream->perf->ops.read(stream, buf, count, offset);
1185}
1186
1187static struct intel_context *oa_pin_context(struct i915_perf_stream *stream)
1188{
1189 struct i915_gem_engines_iter it;
1190 struct i915_gem_context *ctx = stream->ctx;
1191 struct intel_context *ce;
1192 struct i915_gem_ww_ctx ww;
1193 int err = -ENODEV;
1194
1195 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
1196 if (ce->engine != stream->engine) /* first match! */
1197 continue;
1198
1199 err = 0;
1200 break;
1201 }
1202 i915_gem_context_unlock_engines(ctx);
1203
1204 if (err)
1205 return ERR_PTR(err);
1206
1207 i915_gem_ww_ctx_init(&ww, true);
1208retry:
1209 /*
1210 * As the ID is the gtt offset of the context's vma we
1211 * pin the vma to ensure the ID remains fixed.
1212 */
1213 err = intel_context_pin_ww(ce, &ww);
1214 if (err == -EDEADLK) {
1215 err = i915_gem_ww_ctx_backoff(&ww);
1216 if (!err)
1217 goto retry;
1218 }
1219 i915_gem_ww_ctx_fini(&ww);
1220
1221 if (err)
1222 return ERR_PTR(err);
1223
1224 stream->pinned_ctx = ce;
1225 return stream->pinned_ctx;
1226}
1227
1228/**
1229 * oa_get_render_ctx_id - determine and hold ctx hw id
1230 * @stream: An i915-perf stream opened for OA metrics
1231 *
1232 * Determine the render context hw id, and ensure it remains fixed for the
1233 * lifetime of the stream. This ensures that we don't have to worry about
1234 * updating the context ID in OACONTROL on the fly.
1235 *
1236 * Returns: zero on success or a negative error code
1237 */
1238static int oa_get_render_ctx_id(struct i915_perf_stream *stream)
1239{
1240 struct intel_context *ce;
1241
1242 ce = oa_pin_context(stream);
1243 if (IS_ERR(ce))
1244 return PTR_ERR(ce);
1245
1246 switch (GRAPHICS_VER(ce->engine->i915)) {
1247 case 7: {
1248 /*
1249 * On Haswell we don't do any post processing of the reports
1250 * and don't need to use the mask.
1251 */
1252 stream->specific_ctx_id = i915_ggtt_offset(ce->state);
1253 stream->specific_ctx_id_mask = 0;
1254 break;
1255 }
1256
1257 case 8:
1258 case 9:
1259 case 10:
1260 if (intel_engine_uses_guc(ce->engine)) {
1261 /*
1262 * When using GuC, the context descriptor we write in
1263 * i915 is read by GuC and rewritten before it's
1264 * actually written into the hardware. The LRCA is
1265 * what is put into the context id field of the
1266 * context descriptor by GuC. Because it's aligned to
1267 * a page, the lower 12bits are always at 0 and
1268 * dropped by GuC. They won't be part of the context
1269 * ID in the OA reports, so squash those lower bits.
1270 */
1271 stream->specific_ctx_id = ce->lrc.lrca >> 12;
1272
1273 /*
1274 * GuC uses the top bit to signal proxy submission, so
1275 * ignore that bit.
1276 */
1277 stream->specific_ctx_id_mask =
1278 (1U << (GEN8_CTX_ID_WIDTH - 1)) - 1;
1279 } else {
1280 stream->specific_ctx_id_mask =
1281 (1U << GEN8_CTX_ID_WIDTH) - 1;
1282 stream->specific_ctx_id = stream->specific_ctx_id_mask;
1283 }
1284 break;
1285
1286 case 11:
1287 case 12: {
1288 stream->specific_ctx_id_mask =
1289 ((1U << GEN11_SW_CTX_ID_WIDTH) - 1) << (GEN11_SW_CTX_ID_SHIFT - 32);
1290 /*
1291 * Pick an unused context id
1292 * 0 - BITS_PER_LONG are used by other contexts
1293 * GEN12_MAX_CONTEXT_HW_ID (0x7ff) is used by idle context
1294 */
1295 stream->specific_ctx_id = (GEN12_MAX_CONTEXT_HW_ID - 1) << (GEN11_SW_CTX_ID_SHIFT - 32);
1296 break;
1297 }
1298
1299 default:
1300 MISSING_CASE(GRAPHICS_VER(ce->engine->i915));
1301 }
1302
1303 ce->tag = stream->specific_ctx_id;
1304
1305 drm_dbg(&stream->perf->i915->drm,
1306 "filtering on ctx_id=0x%x ctx_id_mask=0x%x\n",
1307 stream->specific_ctx_id,
1308 stream->specific_ctx_id_mask);
1309
1310 return 0;
1311}
1312
1313/**
1314 * oa_put_render_ctx_id - counterpart to oa_get_render_ctx_id releases hold
1315 * @stream: An i915-perf stream opened for OA metrics
1316 *
1317 * In case anything needed doing to ensure the context HW ID would remain valid
1318 * for the lifetime of the stream, then that can be undone here.
1319 */
1320static void oa_put_render_ctx_id(struct i915_perf_stream *stream)
1321{
1322 struct intel_context *ce;
1323
1324 ce = fetch_and_zero(&stream->pinned_ctx);
1325 if (ce) {
1326 ce->tag = 0; /* recomputed on next submission after parking */
1327 intel_context_unpin(ce);
1328 }
1329
1330 stream->specific_ctx_id = INVALID_CTX_ID;
1331 stream->specific_ctx_id_mask = 0;
1332}
1333
1334static void
1335free_oa_buffer(struct i915_perf_stream *stream)
1336{
1337 i915_vma_unpin_and_release(&stream->oa_buffer.vma,
1338 I915_VMA_RELEASE_MAP);
1339
1340 stream->oa_buffer.vaddr = NULL;
1341}
1342
1343static void
1344free_oa_configs(struct i915_perf_stream *stream)
1345{
1346 struct i915_oa_config_bo *oa_bo, *tmp;
1347
1348 i915_oa_config_put(stream->oa_config);
1349 llist_for_each_entry_safe(oa_bo, tmp, stream->oa_config_bos.first, node)
1350 free_oa_config_bo(oa_bo);
1351}
1352
1353static void
1354free_noa_wait(struct i915_perf_stream *stream)
1355{
1356 i915_vma_unpin_and_release(&stream->noa_wait, 0);
1357}
1358
1359static void i915_oa_stream_destroy(struct i915_perf_stream *stream)
1360{
1361 struct i915_perf *perf = stream->perf;
1362
1363 BUG_ON(stream != perf->exclusive_stream);
1364
1365 /*
1366 * Unset exclusive_stream first, it will be checked while disabling
1367 * the metric set on gen8+.
1368 *
1369 * See i915_oa_init_reg_state() and lrc_configure_all_contexts()
1370 */
1371 WRITE_ONCE(perf->exclusive_stream, NULL);
1372 perf->ops.disable_metric_set(stream);
1373
1374 free_oa_buffer(stream);
1375
1376 intel_uncore_forcewake_put(stream->uncore, FORCEWAKE_ALL);
1377 intel_engine_pm_put(stream->engine);
1378
1379 if (stream->ctx)
1380 oa_put_render_ctx_id(stream);
1381
1382 free_oa_configs(stream);
1383 free_noa_wait(stream);
1384
1385 if (perf->spurious_report_rs.missed) {
1386 DRM_NOTE("%d spurious OA report notices suppressed due to ratelimiting\n",
1387 perf->spurious_report_rs.missed);
1388 }
1389}
1390
1391static void gen7_init_oa_buffer(struct i915_perf_stream *stream)
1392{
1393 struct intel_uncore *uncore = stream->uncore;
1394 u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma);
1395 unsigned long flags;
1396
1397 spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
1398
1399 /* Pre-DevBDW: OABUFFER must be set with counters off,
1400 * before OASTATUS1, but after OASTATUS2
1401 */
1402 intel_uncore_write(uncore, GEN7_OASTATUS2, /* head */
1403 gtt_offset | GEN7_OASTATUS2_MEM_SELECT_GGTT);
1404 stream->oa_buffer.head = gtt_offset;
1405
1406 intel_uncore_write(uncore, GEN7_OABUFFER, gtt_offset);
1407
1408 intel_uncore_write(uncore, GEN7_OASTATUS1, /* tail */
1409 gtt_offset | OABUFFER_SIZE_16M);
1410
1411 /* Mark that we need updated tail pointers to read from... */
1412 stream->oa_buffer.aging_tail = INVALID_TAIL_PTR;
1413 stream->oa_buffer.tail = gtt_offset;
1414
1415 spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
1416
1417 /* On Haswell we have to track which OASTATUS1 flags we've
1418 * already seen since they can't be cleared while periodic
1419 * sampling is enabled.
1420 */
1421 stream->perf->gen7_latched_oastatus1 = 0;
1422
1423 /* NB: although the OA buffer will initially be allocated
1424 * zeroed via shmfs (and so this memset is redundant when
1425 * first allocating), we may re-init the OA buffer, either
1426 * when re-enabling a stream or in error/reset paths.
1427 *
1428 * The reason we clear the buffer for each re-init is for the
1429 * sanity check in gen7_append_oa_reports() that looks at the
1430 * report-id field to make sure it's non-zero which relies on
1431 * the assumption that new reports are being written to zeroed
1432 * memory...
1433 */
1434 memset(stream->oa_buffer.vaddr, 0, OA_BUFFER_SIZE);
1435}
1436
1437static void gen8_init_oa_buffer(struct i915_perf_stream *stream)
1438{
1439 struct intel_uncore *uncore = stream->uncore;
1440 u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma);
1441 unsigned long flags;
1442
1443 spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
1444
1445 intel_uncore_write(uncore, GEN8_OASTATUS, 0);
1446 intel_uncore_write(uncore, GEN8_OAHEADPTR, gtt_offset);
1447 stream->oa_buffer.head = gtt_offset;
1448
1449 intel_uncore_write(uncore, GEN8_OABUFFER_UDW, 0);
1450
1451 /*
1452 * PRM says:
1453 *
1454 * "This MMIO must be set before the OATAILPTR
1455 * register and after the OAHEADPTR register. This is
1456 * to enable proper functionality of the overflow
1457 * bit."
1458 */
1459 intel_uncore_write(uncore, GEN8_OABUFFER, gtt_offset |
1460 OABUFFER_SIZE_16M | GEN8_OABUFFER_MEM_SELECT_GGTT);
1461 intel_uncore_write(uncore, GEN8_OATAILPTR, gtt_offset & GEN8_OATAILPTR_MASK);
1462
1463 /* Mark that we need updated tail pointers to read from... */
1464 stream->oa_buffer.aging_tail = INVALID_TAIL_PTR;
1465 stream->oa_buffer.tail = gtt_offset;
1466
1467 /*
1468 * Reset state used to recognise context switches, affecting which
1469 * reports we will forward to userspace while filtering for a single
1470 * context.
1471 */
1472 stream->oa_buffer.last_ctx_id = INVALID_CTX_ID;
1473
1474 spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
1475
1476 /*
1477 * NB: although the OA buffer will initially be allocated
1478 * zeroed via shmfs (and so this memset is redundant when
1479 * first allocating), we may re-init the OA buffer, either
1480 * when re-enabling a stream or in error/reset paths.
1481 *
1482 * The reason we clear the buffer for each re-init is for the
1483 * sanity check in gen8_append_oa_reports() that looks at the
1484 * reason field to make sure it's non-zero which relies on
1485 * the assumption that new reports are being written to zeroed
1486 * memory...
1487 */
1488 memset(stream->oa_buffer.vaddr, 0, OA_BUFFER_SIZE);
1489}
1490
1491static void gen12_init_oa_buffer(struct i915_perf_stream *stream)
1492{
1493 struct intel_uncore *uncore = stream->uncore;
1494 u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma);
1495 unsigned long flags;
1496
1497 spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags);
1498
1499 intel_uncore_write(uncore, GEN12_OAG_OASTATUS, 0);
1500 intel_uncore_write(uncore, GEN12_OAG_OAHEADPTR,
1501 gtt_offset & GEN12_OAG_OAHEADPTR_MASK);
1502 stream->oa_buffer.head = gtt_offset;
1503
1504 /*
1505 * PRM says:
1506 *
1507 * "This MMIO must be set before the OATAILPTR
1508 * register and after the OAHEADPTR register. This is
1509 * to enable proper functionality of the overflow
1510 * bit."
1511 */
1512 intel_uncore_write(uncore, GEN12_OAG_OABUFFER, gtt_offset |
1513 OABUFFER_SIZE_16M | GEN8_OABUFFER_MEM_SELECT_GGTT);
1514 intel_uncore_write(uncore, GEN12_OAG_OATAILPTR,
1515 gtt_offset & GEN12_OAG_OATAILPTR_MASK);
1516
1517 /* Mark that we need updated tail pointers to read from... */
1518 stream->oa_buffer.aging_tail = INVALID_TAIL_PTR;
1519 stream->oa_buffer.tail = gtt_offset;
1520
1521 /*
1522 * Reset state used to recognise context switches, affecting which
1523 * reports we will forward to userspace while filtering for a single
1524 * context.
1525 */
1526 stream->oa_buffer.last_ctx_id = INVALID_CTX_ID;
1527
1528 spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags);
1529
1530 /*
1531 * NB: although the OA buffer will initially be allocated
1532 * zeroed via shmfs (and so this memset is redundant when
1533 * first allocating), we may re-init the OA buffer, either
1534 * when re-enabling a stream or in error/reset paths.
1535 *
1536 * The reason we clear the buffer for each re-init is for the
1537 * sanity check in gen8_append_oa_reports() that looks at the
1538 * reason field to make sure it's non-zero which relies on
1539 * the assumption that new reports are being written to zeroed
1540 * memory...
1541 */
1542 memset(stream->oa_buffer.vaddr, 0,
1543 stream->oa_buffer.vma->size);
1544}
1545
1546static int alloc_oa_buffer(struct i915_perf_stream *stream)
1547{
1548 struct drm_i915_private *i915 = stream->perf->i915;
1549 struct drm_i915_gem_object *bo;
1550 struct i915_vma *vma;
1551 int ret;
1552
1553 if (drm_WARN_ON(&i915->drm, stream->oa_buffer.vma))
1554 return -ENODEV;
1555
1556 BUILD_BUG_ON_NOT_POWER_OF_2(OA_BUFFER_SIZE);
1557 BUILD_BUG_ON(OA_BUFFER_SIZE < SZ_128K || OA_BUFFER_SIZE > SZ_16M);
1558
1559 bo = i915_gem_object_create_shmem(stream->perf->i915, OA_BUFFER_SIZE);
1560 if (IS_ERR(bo)) {
1561 drm_err(&i915->drm, "Failed to allocate OA buffer\n");
1562 return PTR_ERR(bo);
1563 }
1564
1565 i915_gem_object_set_cache_coherency(bo, I915_CACHE_LLC);
1566
1567 /* PreHSW required 512K alignment, HSW requires 16M */
1568 vma = i915_gem_object_ggtt_pin(bo, NULL, 0, SZ_16M, 0);
1569 if (IS_ERR(vma)) {
1570 ret = PTR_ERR(vma);
1571 goto err_unref;
1572 }
1573 stream->oa_buffer.vma = vma;
1574
1575 stream->oa_buffer.vaddr =
1576 i915_gem_object_pin_map_unlocked(bo, I915_MAP_WB);
1577 if (IS_ERR(stream->oa_buffer.vaddr)) {
1578 ret = PTR_ERR(stream->oa_buffer.vaddr);
1579 goto err_unpin;
1580 }
1581
1582 return 0;
1583
1584err_unpin:
1585 __i915_vma_unpin(vma);
1586
1587err_unref:
1588 i915_gem_object_put(bo);
1589
1590 stream->oa_buffer.vaddr = NULL;
1591 stream->oa_buffer.vma = NULL;
1592
1593 return ret;
1594}
1595
1596static u32 *save_restore_register(struct i915_perf_stream *stream, u32 *cs,
1597 bool save, i915_reg_t reg, u32 offset,
1598 u32 dword_count)
1599{
1600 u32 cmd;
1601 u32 d;
1602
1603 cmd = save ? MI_STORE_REGISTER_MEM : MI_LOAD_REGISTER_MEM;
1604 cmd |= MI_SRM_LRM_GLOBAL_GTT;
1605 if (GRAPHICS_VER(stream->perf->i915) >= 8)
1606 cmd++;
1607
1608 for (d = 0; d < dword_count; d++) {
1609 *cs++ = cmd;
1610 *cs++ = i915_mmio_reg_offset(reg) + 4 * d;
1611 *cs++ = intel_gt_scratch_offset(stream->engine->gt,
1612 offset) + 4 * d;
1613 *cs++ = 0;
1614 }
1615
1616 return cs;
1617}
1618
1619static int alloc_noa_wait(struct i915_perf_stream *stream)
1620{
1621 struct drm_i915_private *i915 = stream->perf->i915;
1622 struct drm_i915_gem_object *bo;
1623 struct i915_vma *vma;
1624 const u64 delay_ticks = 0xffffffffffffffff -
1625 intel_gt_ns_to_clock_interval(stream->perf->i915->ggtt.vm.gt,
1626 atomic64_read(&stream->perf->noa_programming_delay));
1627 const u32 base = stream->engine->mmio_base;
1628#define CS_GPR(x) GEN8_RING_CS_GPR(base, x)
1629 u32 *batch, *ts0, *cs, *jump;
1630 struct i915_gem_ww_ctx ww;
1631 int ret, i;
1632 enum {
1633 START_TS,
1634 NOW_TS,
1635 DELTA_TS,
1636 JUMP_PREDICATE,
1637 DELTA_TARGET,
1638 N_CS_GPR
1639 };
1640
1641 bo = i915_gem_object_create_internal(i915, 4096);
1642 if (IS_ERR(bo)) {
1643 drm_err(&i915->drm,
1644 "Failed to allocate NOA wait batchbuffer\n");
1645 return PTR_ERR(bo);
1646 }
1647
1648 i915_gem_ww_ctx_init(&ww, true);
1649retry:
1650 ret = i915_gem_object_lock(bo, &ww);
1651 if (ret)
1652 goto out_ww;
1653
1654 /*
1655 * We pin in GGTT because we jump into this buffer now because
1656 * multiple OA config BOs will have a jump to this address and it
1657 * needs to be fixed during the lifetime of the i915/perf stream.
1658 */
1659 vma = i915_gem_object_ggtt_pin_ww(bo, &ww, NULL, 0, 0, PIN_HIGH);
1660 if (IS_ERR(vma)) {
1661 ret = PTR_ERR(vma);
1662 goto out_ww;
1663 }
1664
1665 batch = cs = i915_gem_object_pin_map(bo, I915_MAP_WB);
1666 if (IS_ERR(batch)) {
1667 ret = PTR_ERR(batch);
1668 goto err_unpin;
1669 }
1670
1671 /* Save registers. */
1672 for (i = 0; i < N_CS_GPR; i++)
1673 cs = save_restore_register(
1674 stream, cs, true /* save */, CS_GPR(i),
1675 INTEL_GT_SCRATCH_FIELD_PERF_CS_GPR + 8 * i, 2);
1676 cs = save_restore_register(
1677 stream, cs, true /* save */, MI_PREDICATE_RESULT_1,
1678 INTEL_GT_SCRATCH_FIELD_PERF_PREDICATE_RESULT_1, 1);
1679
1680 /* First timestamp snapshot location. */
1681 ts0 = cs;
1682
1683 /*
1684 * Initial snapshot of the timestamp register to implement the wait.
1685 * We work with 32b values, so clear out the top 32b bits of the
1686 * register because the ALU works 64bits.
1687 */
1688 *cs++ = MI_LOAD_REGISTER_IMM(1);
1689 *cs++ = i915_mmio_reg_offset(CS_GPR(START_TS)) + 4;
1690 *cs++ = 0;
1691 *cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
1692 *cs++ = i915_mmio_reg_offset(RING_TIMESTAMP(base));
1693 *cs++ = i915_mmio_reg_offset(CS_GPR(START_TS));
1694
1695 /*
1696 * This is the location we're going to jump back into until the
1697 * required amount of time has passed.
1698 */
1699 jump = cs;
1700
1701 /*
1702 * Take another snapshot of the timestamp register. Take care to clear
1703 * up the top 32bits of CS_GPR(1) as we're using it for other
1704 * operations below.
1705 */
1706 *cs++ = MI_LOAD_REGISTER_IMM(1);
1707 *cs++ = i915_mmio_reg_offset(CS_GPR(NOW_TS)) + 4;
1708 *cs++ = 0;
1709 *cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
1710 *cs++ = i915_mmio_reg_offset(RING_TIMESTAMP(base));
1711 *cs++ = i915_mmio_reg_offset(CS_GPR(NOW_TS));
1712
1713 /*
1714 * Do a diff between the 2 timestamps and store the result back into
1715 * CS_GPR(1).
1716 */
1717 *cs++ = MI_MATH(5);
1718 *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(NOW_TS));
1719 *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(START_TS));
1720 *cs++ = MI_MATH_SUB;
1721 *cs++ = MI_MATH_STORE(MI_MATH_REG(DELTA_TS), MI_MATH_REG_ACCU);
1722 *cs++ = MI_MATH_STORE(MI_MATH_REG(JUMP_PREDICATE), MI_MATH_REG_CF);
1723
1724 /*
1725 * Transfer the carry flag (set to 1 if ts1 < ts0, meaning the
1726 * timestamp have rolled over the 32bits) into the predicate register
1727 * to be used for the predicated jump.
1728 */
1729 *cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
1730 *cs++ = i915_mmio_reg_offset(CS_GPR(JUMP_PREDICATE));
1731 *cs++ = i915_mmio_reg_offset(MI_PREDICATE_RESULT_1);
1732
1733 /* Restart from the beginning if we had timestamps roll over. */
1734 *cs++ = (GRAPHICS_VER(i915) < 8 ?
1735 MI_BATCH_BUFFER_START :
1736 MI_BATCH_BUFFER_START_GEN8) |
1737 MI_BATCH_PREDICATE;
1738 *cs++ = i915_ggtt_offset(vma) + (ts0 - batch) * 4;
1739 *cs++ = 0;
1740
1741 /*
1742 * Now add the diff between to previous timestamps and add it to :
1743 * (((1 * << 64) - 1) - delay_ns)
1744 *
1745 * When the Carry Flag contains 1 this means the elapsed time is
1746 * longer than the expected delay, and we can exit the wait loop.
1747 */
1748 *cs++ = MI_LOAD_REGISTER_IMM(2);
1749 *cs++ = i915_mmio_reg_offset(CS_GPR(DELTA_TARGET));
1750 *cs++ = lower_32_bits(delay_ticks);
1751 *cs++ = i915_mmio_reg_offset(CS_GPR(DELTA_TARGET)) + 4;
1752 *cs++ = upper_32_bits(delay_ticks);
1753
1754 *cs++ = MI_MATH(4);
1755 *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(DELTA_TS));
1756 *cs++ = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(DELTA_TARGET));
1757 *cs++ = MI_MATH_ADD;
1758 *cs++ = MI_MATH_STOREINV(MI_MATH_REG(JUMP_PREDICATE), MI_MATH_REG_CF);
1759
1760 *cs++ = MI_ARB_CHECK;
1761
1762 /*
1763 * Transfer the result into the predicate register to be used for the
1764 * predicated jump.
1765 */
1766 *cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
1767 *cs++ = i915_mmio_reg_offset(CS_GPR(JUMP_PREDICATE));
1768 *cs++ = i915_mmio_reg_offset(MI_PREDICATE_RESULT_1);
1769
1770 /* Predicate the jump. */
1771 *cs++ = (GRAPHICS_VER(i915) < 8 ?
1772 MI_BATCH_BUFFER_START :
1773 MI_BATCH_BUFFER_START_GEN8) |
1774 MI_BATCH_PREDICATE;
1775 *cs++ = i915_ggtt_offset(vma) + (jump - batch) * 4;
1776 *cs++ = 0;
1777
1778 /* Restore registers. */
1779 for (i = 0; i < N_CS_GPR; i++)
1780 cs = save_restore_register(
1781 stream, cs, false /* restore */, CS_GPR(i),
1782 INTEL_GT_SCRATCH_FIELD_PERF_CS_GPR + 8 * i, 2);
1783 cs = save_restore_register(
1784 stream, cs, false /* restore */, MI_PREDICATE_RESULT_1,
1785 INTEL_GT_SCRATCH_FIELD_PERF_PREDICATE_RESULT_1, 1);
1786
1787 /* And return to the ring. */
1788 *cs++ = MI_BATCH_BUFFER_END;
1789
1790 GEM_BUG_ON(cs - batch > PAGE_SIZE / sizeof(*batch));
1791
1792 i915_gem_object_flush_map(bo);
1793 __i915_gem_object_release_map(bo);
1794
1795 stream->noa_wait = vma;
1796 goto out_ww;
1797
1798err_unpin:
1799 i915_vma_unpin_and_release(&vma, 0);
1800out_ww:
1801 if (ret == -EDEADLK) {
1802 ret = i915_gem_ww_ctx_backoff(&ww);
1803 if (!ret)
1804 goto retry;
1805 }
1806 i915_gem_ww_ctx_fini(&ww);
1807 if (ret)
1808 i915_gem_object_put(bo);
1809 return ret;
1810}
1811
1812static u32 *write_cs_mi_lri(u32 *cs,
1813 const struct i915_oa_reg *reg_data,
1814 u32 n_regs)
1815{
1816 u32 i;
1817
1818 for (i = 0; i < n_regs; i++) {
1819 if ((i % MI_LOAD_REGISTER_IMM_MAX_REGS) == 0) {
1820 u32 n_lri = min_t(u32,
1821 n_regs - i,
1822 MI_LOAD_REGISTER_IMM_MAX_REGS);
1823
1824 *cs++ = MI_LOAD_REGISTER_IMM(n_lri);
1825 }
1826 *cs++ = i915_mmio_reg_offset(reg_data[i].addr);
1827 *cs++ = reg_data[i].value;
1828 }
1829
1830 return cs;
1831}
1832
1833static int num_lri_dwords(int num_regs)
1834{
1835 int count = 0;
1836
1837 if (num_regs > 0) {
1838 count += DIV_ROUND_UP(num_regs, MI_LOAD_REGISTER_IMM_MAX_REGS);
1839 count += num_regs * 2;
1840 }
1841
1842 return count;
1843}
1844
1845static struct i915_oa_config_bo *
1846alloc_oa_config_buffer(struct i915_perf_stream *stream,
1847 struct i915_oa_config *oa_config)
1848{
1849 struct drm_i915_gem_object *obj;
1850 struct i915_oa_config_bo *oa_bo;
1851 struct i915_gem_ww_ctx ww;
1852 size_t config_length = 0;
1853 u32 *cs;
1854 int err;
1855
1856 oa_bo = kzalloc(sizeof(*oa_bo), GFP_KERNEL);
1857 if (!oa_bo)
1858 return ERR_PTR(-ENOMEM);
1859
1860 config_length += num_lri_dwords(oa_config->mux_regs_len);
1861 config_length += num_lri_dwords(oa_config->b_counter_regs_len);
1862 config_length += num_lri_dwords(oa_config->flex_regs_len);
1863 config_length += 3; /* MI_BATCH_BUFFER_START */
1864 config_length = ALIGN(sizeof(u32) * config_length, I915_GTT_PAGE_SIZE);
1865
1866 obj = i915_gem_object_create_shmem(stream->perf->i915, config_length);
1867 if (IS_ERR(obj)) {
1868 err = PTR_ERR(obj);
1869 goto err_free;
1870 }
1871
1872 i915_gem_ww_ctx_init(&ww, true);
1873retry:
1874 err = i915_gem_object_lock(obj, &ww);
1875 if (err)
1876 goto out_ww;
1877
1878 cs = i915_gem_object_pin_map(obj, I915_MAP_WB);
1879 if (IS_ERR(cs)) {
1880 err = PTR_ERR(cs);
1881 goto out_ww;
1882 }
1883
1884 cs = write_cs_mi_lri(cs,
1885 oa_config->mux_regs,
1886 oa_config->mux_regs_len);
1887 cs = write_cs_mi_lri(cs,
1888 oa_config->b_counter_regs,
1889 oa_config->b_counter_regs_len);
1890 cs = write_cs_mi_lri(cs,
1891 oa_config->flex_regs,
1892 oa_config->flex_regs_len);
1893
1894 /* Jump into the active wait. */
1895 *cs++ = (GRAPHICS_VER(stream->perf->i915) < 8 ?
1896 MI_BATCH_BUFFER_START :
1897 MI_BATCH_BUFFER_START_GEN8);
1898 *cs++ = i915_ggtt_offset(stream->noa_wait);
1899 *cs++ = 0;
1900
1901 i915_gem_object_flush_map(obj);
1902 __i915_gem_object_release_map(obj);
1903
1904 oa_bo->vma = i915_vma_instance(obj,
1905 &stream->engine->gt->ggtt->vm,
1906 NULL);
1907 if (IS_ERR(oa_bo->vma)) {
1908 err = PTR_ERR(oa_bo->vma);
1909 goto out_ww;
1910 }
1911
1912 oa_bo->oa_config = i915_oa_config_get(oa_config);
1913 llist_add(&oa_bo->node, &stream->oa_config_bos);
1914
1915out_ww:
1916 if (err == -EDEADLK) {
1917 err = i915_gem_ww_ctx_backoff(&ww);
1918 if (!err)
1919 goto retry;
1920 }
1921 i915_gem_ww_ctx_fini(&ww);
1922
1923 if (err)
1924 i915_gem_object_put(obj);
1925err_free:
1926 if (err) {
1927 kfree(oa_bo);
1928 return ERR_PTR(err);
1929 }
1930 return oa_bo;
1931}
1932
1933static struct i915_vma *
1934get_oa_vma(struct i915_perf_stream *stream, struct i915_oa_config *oa_config)
1935{
1936 struct i915_oa_config_bo *oa_bo;
1937
1938 /*
1939 * Look for the buffer in the already allocated BOs attached
1940 * to the stream.
1941 */
1942 llist_for_each_entry(oa_bo, stream->oa_config_bos.first, node) {
1943 if (oa_bo->oa_config == oa_config &&
1944 memcmp(oa_bo->oa_config->uuid,
1945 oa_config->uuid,
1946 sizeof(oa_config->uuid)) == 0)
1947 goto out;
1948 }
1949
1950 oa_bo = alloc_oa_config_buffer(stream, oa_config);
1951 if (IS_ERR(oa_bo))
1952 return ERR_CAST(oa_bo);
1953
1954out:
1955 return i915_vma_get(oa_bo->vma);
1956}
1957
1958static int
1959emit_oa_config(struct i915_perf_stream *stream,
1960 struct i915_oa_config *oa_config,
1961 struct intel_context *ce,
1962 struct i915_active *active)
1963{
1964 struct i915_request *rq;
1965 struct i915_vma *vma;
1966 struct i915_gem_ww_ctx ww;
1967 int err;
1968
1969 vma = get_oa_vma(stream, oa_config);
1970 if (IS_ERR(vma))
1971 return PTR_ERR(vma);
1972
1973 i915_gem_ww_ctx_init(&ww, true);
1974retry:
1975 err = i915_gem_object_lock(vma->obj, &ww);
1976 if (err)
1977 goto err;
1978
1979 err = i915_vma_pin_ww(vma, &ww, 0, 0, PIN_GLOBAL | PIN_HIGH);
1980 if (err)
1981 goto err;
1982
1983 intel_engine_pm_get(ce->engine);
1984 rq = i915_request_create(ce);
1985 intel_engine_pm_put(ce->engine);
1986 if (IS_ERR(rq)) {
1987 err = PTR_ERR(rq);
1988 goto err_vma_unpin;
1989 }
1990
1991 if (!IS_ERR_OR_NULL(active)) {
1992 /* After all individual context modifications */
1993 err = i915_request_await_active(rq, active,
1994 I915_ACTIVE_AWAIT_ACTIVE);
1995 if (err)
1996 goto err_add_request;
1997
1998 err = i915_active_add_request(active, rq);
1999 if (err)
2000 goto err_add_request;
2001 }
2002
2003 err = i915_request_await_object(rq, vma->obj, 0);
2004 if (!err)
2005 err = i915_vma_move_to_active(vma, rq, 0);
2006 if (err)
2007 goto err_add_request;
2008
2009 err = rq->engine->emit_bb_start(rq,
2010 vma->node.start, 0,
2011 I915_DISPATCH_SECURE);
2012 if (err)
2013 goto err_add_request;
2014
2015err_add_request:
2016 i915_request_add(rq);
2017err_vma_unpin:
2018 i915_vma_unpin(vma);
2019err:
2020 if (err == -EDEADLK) {
2021 err = i915_gem_ww_ctx_backoff(&ww);
2022 if (!err)
2023 goto retry;
2024 }
2025
2026 i915_gem_ww_ctx_fini(&ww);
2027 i915_vma_put(vma);
2028 return err;
2029}
2030
2031static struct intel_context *oa_context(struct i915_perf_stream *stream)
2032{
2033 return stream->pinned_ctx ?: stream->engine->kernel_context;
2034}
2035
2036static int
2037hsw_enable_metric_set(struct i915_perf_stream *stream,
2038 struct i915_active *active)
2039{
2040 struct intel_uncore *uncore = stream->uncore;
2041
2042 /*
2043 * PRM:
2044 *
2045 * OA unit is using “crclk” for its functionality. When trunk
2046 * level clock gating takes place, OA clock would be gated,
2047 * unable to count the events from non-render clock domain.
2048 * Render clock gating must be disabled when OA is enabled to
2049 * count the events from non-render domain. Unit level clock
2050 * gating for RCS should also be disabled.
2051 */
2052 intel_uncore_rmw(uncore, GEN7_MISCCPCTL,
2053 GEN7_DOP_CLOCK_GATE_ENABLE, 0);
2054 intel_uncore_rmw(uncore, GEN6_UCGCTL1,
2055 0, GEN6_CSUNIT_CLOCK_GATE_DISABLE);
2056
2057 return emit_oa_config(stream,
2058 stream->oa_config, oa_context(stream),
2059 active);
2060}
2061
2062static void hsw_disable_metric_set(struct i915_perf_stream *stream)
2063{
2064 struct intel_uncore *uncore = stream->uncore;
2065
2066 intel_uncore_rmw(uncore, GEN6_UCGCTL1,
2067 GEN6_CSUNIT_CLOCK_GATE_DISABLE, 0);
2068 intel_uncore_rmw(uncore, GEN7_MISCCPCTL,
2069 0, GEN7_DOP_CLOCK_GATE_ENABLE);
2070
2071 intel_uncore_rmw(uncore, GDT_CHICKEN_BITS, GT_NOA_ENABLE, 0);
2072}
2073
2074static u32 oa_config_flex_reg(const struct i915_oa_config *oa_config,
2075 i915_reg_t reg)
2076{
2077 u32 mmio = i915_mmio_reg_offset(reg);
2078 int i;
2079
2080 /*
2081 * This arbitrary default will select the 'EU FPU0 Pipeline
2082 * Active' event. In the future it's anticipated that there
2083 * will be an explicit 'No Event' we can select, but not yet...
2084 */
2085 if (!oa_config)
2086 return 0;
2087
2088 for (i = 0; i < oa_config->flex_regs_len; i++) {
2089 if (i915_mmio_reg_offset(oa_config->flex_regs[i].addr) == mmio)
2090 return oa_config->flex_regs[i].value;
2091 }
2092
2093 return 0;
2094}
2095/*
2096 * NB: It must always remain pointer safe to run this even if the OA unit
2097 * has been disabled.
2098 *
2099 * It's fine to put out-of-date values into these per-context registers
2100 * in the case that the OA unit has been disabled.
2101 */
2102static void
2103gen8_update_reg_state_unlocked(const struct intel_context *ce,
2104 const struct i915_perf_stream *stream)
2105{
2106 u32 ctx_oactxctrl = stream->perf->ctx_oactxctrl_offset;
2107 u32 ctx_flexeu0 = stream->perf->ctx_flexeu0_offset;
2108 /* The MMIO offsets for Flex EU registers aren't contiguous */
2109 i915_reg_t flex_regs[] = {
2110 EU_PERF_CNTL0,
2111 EU_PERF_CNTL1,
2112 EU_PERF_CNTL2,
2113 EU_PERF_CNTL3,
2114 EU_PERF_CNTL4,
2115 EU_PERF_CNTL5,
2116 EU_PERF_CNTL6,
2117 };
2118 u32 *reg_state = ce->lrc_reg_state;
2119 int i;
2120
2121 reg_state[ctx_oactxctrl + 1] =
2122 (stream->period_exponent << GEN8_OA_TIMER_PERIOD_SHIFT) |
2123 (stream->periodic ? GEN8_OA_TIMER_ENABLE : 0) |
2124 GEN8_OA_COUNTER_RESUME;
2125
2126 for (i = 0; i < ARRAY_SIZE(flex_regs); i++)
2127 reg_state[ctx_flexeu0 + i * 2 + 1] =
2128 oa_config_flex_reg(stream->oa_config, flex_regs[i]);
2129}
2130
2131struct flex {
2132 i915_reg_t reg;
2133 u32 offset;
2134 u32 value;
2135};
2136
2137static int
2138gen8_store_flex(struct i915_request *rq,
2139 struct intel_context *ce,
2140 const struct flex *flex, unsigned int count)
2141{
2142 u32 offset;
2143 u32 *cs;
2144
2145 cs = intel_ring_begin(rq, 4 * count);
2146 if (IS_ERR(cs))
2147 return PTR_ERR(cs);
2148
2149 offset = i915_ggtt_offset(ce->state) + LRC_STATE_OFFSET;
2150 do {
2151 *cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
2152 *cs++ = offset + flex->offset * sizeof(u32);
2153 *cs++ = 0;
2154 *cs++ = flex->value;
2155 } while (flex++, --count);
2156
2157 intel_ring_advance(rq, cs);
2158
2159 return 0;
2160}
2161
2162static int
2163gen8_load_flex(struct i915_request *rq,
2164 struct intel_context *ce,
2165 const struct flex *flex, unsigned int count)
2166{
2167 u32 *cs;
2168
2169 GEM_BUG_ON(!count || count > 63);
2170
2171 cs = intel_ring_begin(rq, 2 * count + 2);
2172 if (IS_ERR(cs))
2173 return PTR_ERR(cs);
2174
2175 *cs++ = MI_LOAD_REGISTER_IMM(count);
2176 do {
2177 *cs++ = i915_mmio_reg_offset(flex->reg);
2178 *cs++ = flex->value;
2179 } while (flex++, --count);
2180 *cs++ = MI_NOOP;
2181
2182 intel_ring_advance(rq, cs);
2183
2184 return 0;
2185}
2186
2187static int gen8_modify_context(struct intel_context *ce,
2188 const struct flex *flex, unsigned int count)
2189{
2190 struct i915_request *rq;
2191 int err;
2192
2193 rq = intel_engine_create_kernel_request(ce->engine);
2194 if (IS_ERR(rq))
2195 return PTR_ERR(rq);
2196
2197 /* Serialise with the remote context */
2198 err = intel_context_prepare_remote_request(ce, rq);
2199 if (err == 0)
2200 err = gen8_store_flex(rq, ce, flex, count);
2201
2202 i915_request_add(rq);
2203 return err;
2204}
2205
2206static int
2207gen8_modify_self(struct intel_context *ce,
2208 const struct flex *flex, unsigned int count,
2209 struct i915_active *active)
2210{
2211 struct i915_request *rq;
2212 int err;
2213
2214 intel_engine_pm_get(ce->engine);
2215 rq = i915_request_create(ce);
2216 intel_engine_pm_put(ce->engine);
2217 if (IS_ERR(rq))
2218 return PTR_ERR(rq);
2219
2220 if (!IS_ERR_OR_NULL(active)) {
2221 err = i915_active_add_request(active, rq);
2222 if (err)
2223 goto err_add_request;
2224 }
2225
2226 err = gen8_load_flex(rq, ce, flex, count);
2227 if (err)
2228 goto err_add_request;
2229
2230err_add_request:
2231 i915_request_add(rq);
2232 return err;
2233}
2234
2235static int gen8_configure_context(struct i915_gem_context *ctx,
2236 struct flex *flex, unsigned int count)
2237{
2238 struct i915_gem_engines_iter it;
2239 struct intel_context *ce;
2240 int err = 0;
2241
2242 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
2243 GEM_BUG_ON(ce == ce->engine->kernel_context);
2244
2245 if (ce->engine->class != RENDER_CLASS)
2246 continue;
2247
2248 /* Otherwise OA settings will be set upon first use */
2249 if (!intel_context_pin_if_active(ce))
2250 continue;
2251
2252 flex->value = intel_sseu_make_rpcs(ce->engine->gt, &ce->sseu);
2253 err = gen8_modify_context(ce, flex, count);
2254
2255 intel_context_unpin(ce);
2256 if (err)
2257 break;
2258 }
2259 i915_gem_context_unlock_engines(ctx);
2260
2261 return err;
2262}
2263
2264static int gen12_configure_oar_context(struct i915_perf_stream *stream,
2265 struct i915_active *active)
2266{
2267 int err;
2268 struct intel_context *ce = stream->pinned_ctx;
2269 u32 format = stream->oa_buffer.format;
2270 struct flex regs_context[] = {
2271 {
2272 GEN8_OACTXCONTROL,
2273 stream->perf->ctx_oactxctrl_offset + 1,
2274 active ? GEN8_OA_COUNTER_RESUME : 0,
2275 },
2276 };
2277 /* Offsets in regs_lri are not used since this configuration is only
2278 * applied using LRI. Initialize the correct offsets for posterity.
2279 */
2280#define GEN12_OAR_OACONTROL_OFFSET 0x5B0
2281 struct flex regs_lri[] = {
2282 {
2283 GEN12_OAR_OACONTROL,
2284 GEN12_OAR_OACONTROL_OFFSET + 1,
2285 (format << GEN12_OAR_OACONTROL_COUNTER_FORMAT_SHIFT) |
2286 (active ? GEN12_OAR_OACONTROL_COUNTER_ENABLE : 0)
2287 },
2288 {
2289 RING_CONTEXT_CONTROL(ce->engine->mmio_base),
2290 CTX_CONTEXT_CONTROL,
2291 _MASKED_FIELD(GEN12_CTX_CTRL_OAR_CONTEXT_ENABLE,
2292 active ?
2293 GEN12_CTX_CTRL_OAR_CONTEXT_ENABLE :
2294 0)
2295 },
2296 };
2297
2298 /* Modify the context image of pinned context with regs_context*/
2299 err = intel_context_lock_pinned(ce);
2300 if (err)
2301 return err;
2302
2303 err = gen8_modify_context(ce, regs_context, ARRAY_SIZE(regs_context));
2304 intel_context_unlock_pinned(ce);
2305 if (err)
2306 return err;
2307
2308 /* Apply regs_lri using LRI with pinned context */
2309 return gen8_modify_self(ce, regs_lri, ARRAY_SIZE(regs_lri), active);
2310}
2311
2312/*
2313 * Manages updating the per-context aspects of the OA stream
2314 * configuration across all contexts.
2315 *
2316 * The awkward consideration here is that OACTXCONTROL controls the
2317 * exponent for periodic sampling which is primarily used for system
2318 * wide profiling where we'd like a consistent sampling period even in
2319 * the face of context switches.
2320 *
2321 * Our approach of updating the register state context (as opposed to
2322 * say using a workaround batch buffer) ensures that the hardware
2323 * won't automatically reload an out-of-date timer exponent even
2324 * transiently before a WA BB could be parsed.
2325 *
2326 * This function needs to:
2327 * - Ensure the currently running context's per-context OA state is
2328 * updated
2329 * - Ensure that all existing contexts will have the correct per-context
2330 * OA state if they are scheduled for use.
2331 * - Ensure any new contexts will be initialized with the correct
2332 * per-context OA state.
2333 *
2334 * Note: it's only the RCS/Render context that has any OA state.
2335 * Note: the first flex register passed must always be R_PWR_CLK_STATE
2336 */
2337static int
2338oa_configure_all_contexts(struct i915_perf_stream *stream,
2339 struct flex *regs,
2340 size_t num_regs,
2341 struct i915_active *active)
2342{
2343 struct drm_i915_private *i915 = stream->perf->i915;
2344 struct intel_engine_cs *engine;
2345 struct i915_gem_context *ctx, *cn;
2346 int err;
2347
2348 lockdep_assert_held(&stream->perf->lock);
2349
2350 /*
2351 * The OA register config is setup through the context image. This image
2352 * might be written to by the GPU on context switch (in particular on
2353 * lite-restore). This means we can't safely update a context's image,
2354 * if this context is scheduled/submitted to run on the GPU.
2355 *
2356 * We could emit the OA register config through the batch buffer but
2357 * this might leave small interval of time where the OA unit is
2358 * configured at an invalid sampling period.
2359 *
2360 * Note that since we emit all requests from a single ring, there
2361 * is still an implicit global barrier here that may cause a high
2362 * priority context to wait for an otherwise independent low priority
2363 * context. Contexts idle at the time of reconfiguration are not
2364 * trapped behind the barrier.
2365 */
2366 spin_lock(&i915->gem.contexts.lock);
2367 list_for_each_entry_safe(ctx, cn, &i915->gem.contexts.list, link) {
2368 if (!kref_get_unless_zero(&ctx->ref))
2369 continue;
2370
2371 spin_unlock(&i915->gem.contexts.lock);
2372
2373 err = gen8_configure_context(ctx, regs, num_regs);
2374 if (err) {
2375 i915_gem_context_put(ctx);
2376 return err;
2377 }
2378
2379 spin_lock(&i915->gem.contexts.lock);
2380 list_safe_reset_next(ctx, cn, link);
2381 i915_gem_context_put(ctx);
2382 }
2383 spin_unlock(&i915->gem.contexts.lock);
2384
2385 /*
2386 * After updating all other contexts, we need to modify ourselves.
2387 * If we don't modify the kernel_context, we do not get events while
2388 * idle.
2389 */
2390 for_each_uabi_engine(engine, i915) {
2391 struct intel_context *ce = engine->kernel_context;
2392
2393 if (engine->class != RENDER_CLASS)
2394 continue;
2395
2396 regs[0].value = intel_sseu_make_rpcs(engine->gt, &ce->sseu);
2397
2398 err = gen8_modify_self(ce, regs, num_regs, active);
2399 if (err)
2400 return err;
2401 }
2402
2403 return 0;
2404}
2405
2406static int
2407gen12_configure_all_contexts(struct i915_perf_stream *stream,
2408 const struct i915_oa_config *oa_config,
2409 struct i915_active *active)
2410{
2411 struct flex regs[] = {
2412 {
2413 GEN8_R_PWR_CLK_STATE,
2414 CTX_R_PWR_CLK_STATE,
2415 },
2416 };
2417
2418 return oa_configure_all_contexts(stream,
2419 regs, ARRAY_SIZE(regs),
2420 active);
2421}
2422
2423static int
2424lrc_configure_all_contexts(struct i915_perf_stream *stream,
2425 const struct i915_oa_config *oa_config,
2426 struct i915_active *active)
2427{
2428 /* The MMIO offsets for Flex EU registers aren't contiguous */
2429 const u32 ctx_flexeu0 = stream->perf->ctx_flexeu0_offset;
2430#define ctx_flexeuN(N) (ctx_flexeu0 + 2 * (N) + 1)
2431 struct flex regs[] = {
2432 {
2433 GEN8_R_PWR_CLK_STATE,
2434 CTX_R_PWR_CLK_STATE,
2435 },
2436 {
2437 GEN8_OACTXCONTROL,
2438 stream->perf->ctx_oactxctrl_offset + 1,
2439 },
2440 { EU_PERF_CNTL0, ctx_flexeuN(0) },
2441 { EU_PERF_CNTL1, ctx_flexeuN(1) },
2442 { EU_PERF_CNTL2, ctx_flexeuN(2) },
2443 { EU_PERF_CNTL3, ctx_flexeuN(3) },
2444 { EU_PERF_CNTL4, ctx_flexeuN(4) },
2445 { EU_PERF_CNTL5, ctx_flexeuN(5) },
2446 { EU_PERF_CNTL6, ctx_flexeuN(6) },
2447 };
2448#undef ctx_flexeuN
2449 int i;
2450
2451 regs[1].value =
2452 (stream->period_exponent << GEN8_OA_TIMER_PERIOD_SHIFT) |
2453 (stream->periodic ? GEN8_OA_TIMER_ENABLE : 0) |
2454 GEN8_OA_COUNTER_RESUME;
2455
2456 for (i = 2; i < ARRAY_SIZE(regs); i++)
2457 regs[i].value = oa_config_flex_reg(oa_config, regs[i].reg);
2458
2459 return oa_configure_all_contexts(stream,
2460 regs, ARRAY_SIZE(regs),
2461 active);
2462}
2463
2464static int
2465gen8_enable_metric_set(struct i915_perf_stream *stream,
2466 struct i915_active *active)
2467{
2468 struct intel_uncore *uncore = stream->uncore;
2469 struct i915_oa_config *oa_config = stream->oa_config;
2470 int ret;
2471
2472 /*
2473 * We disable slice/unslice clock ratio change reports on SKL since
2474 * they are too noisy. The HW generates a lot of redundant reports
2475 * where the ratio hasn't really changed causing a lot of redundant
2476 * work to processes and increasing the chances we'll hit buffer
2477 * overruns.
2478 *
2479 * Although we don't currently use the 'disable overrun' OABUFFER
2480 * feature it's worth noting that clock ratio reports have to be
2481 * disabled before considering to use that feature since the HW doesn't
2482 * correctly block these reports.
2483 *
2484 * Currently none of the high-level metrics we have depend on knowing
2485 * this ratio to normalize.
2486 *
2487 * Note: This register is not power context saved and restored, but
2488 * that's OK considering that we disable RC6 while the OA unit is
2489 * enabled.
2490 *
2491 * The _INCLUDE_CLK_RATIO bit allows the slice/unslice frequency to
2492 * be read back from automatically triggered reports, as part of the
2493 * RPT_ID field.
2494 */
2495 if (IS_GRAPHICS_VER(stream->perf->i915, 9, 11)) {
2496 intel_uncore_write(uncore, GEN8_OA_DEBUG,
2497 _MASKED_BIT_ENABLE(GEN9_OA_DEBUG_DISABLE_CLK_RATIO_REPORTS |
2498 GEN9_OA_DEBUG_INCLUDE_CLK_RATIO));
2499 }
2500
2501 /*
2502 * Update all contexts prior writing the mux configurations as we need
2503 * to make sure all slices/subslices are ON before writing to NOA
2504 * registers.
2505 */
2506 ret = lrc_configure_all_contexts(stream, oa_config, active);
2507 if (ret)
2508 return ret;
2509
2510 return emit_oa_config(stream,
2511 stream->oa_config, oa_context(stream),
2512 active);
2513}
2514
2515static u32 oag_report_ctx_switches(const struct i915_perf_stream *stream)
2516{
2517 return _MASKED_FIELD(GEN12_OAG_OA_DEBUG_DISABLE_CTX_SWITCH_REPORTS,
2518 (stream->sample_flags & SAMPLE_OA_REPORT) ?
2519 0 : GEN12_OAG_OA_DEBUG_DISABLE_CTX_SWITCH_REPORTS);
2520}
2521
2522static int
2523gen12_enable_metric_set(struct i915_perf_stream *stream,
2524 struct i915_active *active)
2525{
2526 struct intel_uncore *uncore = stream->uncore;
2527 struct i915_oa_config *oa_config = stream->oa_config;
2528 bool periodic = stream->periodic;
2529 u32 period_exponent = stream->period_exponent;
2530 int ret;
2531
2532 intel_uncore_write(uncore, GEN12_OAG_OA_DEBUG,
2533 /* Disable clk ratio reports, like previous Gens. */
2534 _MASKED_BIT_ENABLE(GEN12_OAG_OA_DEBUG_DISABLE_CLK_RATIO_REPORTS |
2535 GEN12_OAG_OA_DEBUG_INCLUDE_CLK_RATIO) |
2536 /*
2537 * If the user didn't require OA reports, instruct
2538 * the hardware not to emit ctx switch reports.
2539 */
2540 oag_report_ctx_switches(stream));
2541
2542 intel_uncore_write(uncore, GEN12_OAG_OAGLBCTXCTRL, periodic ?
2543 (GEN12_OAG_OAGLBCTXCTRL_COUNTER_RESUME |
2544 GEN12_OAG_OAGLBCTXCTRL_TIMER_ENABLE |
2545 (period_exponent << GEN12_OAG_OAGLBCTXCTRL_TIMER_PERIOD_SHIFT))
2546 : 0);
2547
2548 /*
2549 * Update all contexts prior writing the mux configurations as we need
2550 * to make sure all slices/subslices are ON before writing to NOA
2551 * registers.
2552 */
2553 ret = gen12_configure_all_contexts(stream, oa_config, active);
2554 if (ret)
2555 return ret;
2556
2557 /*
2558 * For Gen12, performance counters are context
2559 * saved/restored. Only enable it for the context that
2560 * requested this.
2561 */
2562 if (stream->ctx) {
2563 ret = gen12_configure_oar_context(stream, active);
2564 if (ret)
2565 return ret;
2566 }
2567
2568 return emit_oa_config(stream,
2569 stream->oa_config, oa_context(stream),
2570 active);
2571}
2572
2573static void gen8_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, NULL);
2579
2580 intel_uncore_rmw(uncore, GDT_CHICKEN_BITS, GT_NOA_ENABLE, 0);
2581}
2582
2583static void gen10_disable_metric_set(struct i915_perf_stream *stream)
2584{
2585 struct intel_uncore *uncore = stream->uncore;
2586
2587 /* Reset all contexts' slices/subslices configurations. */
2588 lrc_configure_all_contexts(stream, NULL, NULL);
2589
2590 /* Make sure we disable noa to save power. */
2591 intel_uncore_rmw(uncore, RPM_CONFIG1, GEN10_GT_NOA_ENABLE, 0);
2592}
2593
2594static void gen12_disable_metric_set(struct i915_perf_stream *stream)
2595{
2596 struct intel_uncore *uncore = stream->uncore;
2597
2598 /* Reset all contexts' slices/subslices configurations. */
2599 gen12_configure_all_contexts(stream, NULL, NULL);
2600
2601 /* disable the context save/restore or OAR counters */
2602 if (stream->ctx)
2603 gen12_configure_oar_context(stream, NULL);
2604
2605 /* Make sure we disable noa to save power. */
2606 intel_uncore_rmw(uncore, RPM_CONFIG1, GEN10_GT_NOA_ENABLE, 0);
2607}
2608
2609static void gen7_oa_enable(struct i915_perf_stream *stream)
2610{
2611 struct intel_uncore *uncore = stream->uncore;
2612 struct i915_gem_context *ctx = stream->ctx;
2613 u32 ctx_id = stream->specific_ctx_id;
2614 bool periodic = stream->periodic;
2615 u32 period_exponent = stream->period_exponent;
2616 u32 report_format = stream->oa_buffer.format;
2617
2618 /*
2619 * Reset buf pointers so we don't forward reports from before now.
2620 *
2621 * Think carefully if considering trying to avoid this, since it
2622 * also ensures status flags and the buffer itself are cleared
2623 * in error paths, and we have checks for invalid reports based
2624 * on the assumption that certain fields are written to zeroed
2625 * memory which this helps maintains.
2626 */
2627 gen7_init_oa_buffer(stream);
2628
2629 intel_uncore_write(uncore, GEN7_OACONTROL,
2630 (ctx_id & GEN7_OACONTROL_CTX_MASK) |
2631 (period_exponent <<
2632 GEN7_OACONTROL_TIMER_PERIOD_SHIFT) |
2633 (periodic ? GEN7_OACONTROL_TIMER_ENABLE : 0) |
2634 (report_format << GEN7_OACONTROL_FORMAT_SHIFT) |
2635 (ctx ? GEN7_OACONTROL_PER_CTX_ENABLE : 0) |
2636 GEN7_OACONTROL_ENABLE);
2637}
2638
2639static void gen8_oa_enable(struct i915_perf_stream *stream)
2640{
2641 struct intel_uncore *uncore = stream->uncore;
2642 u32 report_format = stream->oa_buffer.format;
2643
2644 /*
2645 * Reset buf pointers so we don't forward reports from before now.
2646 *
2647 * Think carefully if considering trying to avoid this, since it
2648 * also ensures status flags and the buffer itself are cleared
2649 * in error paths, and we have checks for invalid reports based
2650 * on the assumption that certain fields are written to zeroed
2651 * memory which this helps maintains.
2652 */
2653 gen8_init_oa_buffer(stream);
2654
2655 /*
2656 * Note: we don't rely on the hardware to perform single context
2657 * filtering and instead filter on the cpu based on the context-id
2658 * field of reports
2659 */
2660 intel_uncore_write(uncore, GEN8_OACONTROL,
2661 (report_format << GEN8_OA_REPORT_FORMAT_SHIFT) |
2662 GEN8_OA_COUNTER_ENABLE);
2663}
2664
2665static void gen12_oa_enable(struct i915_perf_stream *stream)
2666{
2667 struct intel_uncore *uncore = stream->uncore;
2668 u32 report_format = stream->oa_buffer.format;
2669
2670 /*
2671 * If we don't want OA reports from the OA buffer, then we don't even
2672 * need to program the OAG unit.
2673 */
2674 if (!(stream->sample_flags & SAMPLE_OA_REPORT))
2675 return;
2676
2677 gen12_init_oa_buffer(stream);
2678
2679 intel_uncore_write(uncore, GEN12_OAG_OACONTROL,
2680 (report_format << GEN12_OAG_OACONTROL_OA_COUNTER_FORMAT_SHIFT) |
2681 GEN12_OAG_OACONTROL_OA_COUNTER_ENABLE);
2682}
2683
2684/**
2685 * i915_oa_stream_enable - handle `I915_PERF_IOCTL_ENABLE` for OA stream
2686 * @stream: An i915 perf stream opened for OA metrics
2687 *
2688 * [Re]enables hardware periodic sampling according to the period configured
2689 * when opening the stream. This also starts a hrtimer that will periodically
2690 * check for data in the circular OA buffer for notifying userspace (e.g.
2691 * during a read() or poll()).
2692 */
2693static void i915_oa_stream_enable(struct i915_perf_stream *stream)
2694{
2695 stream->pollin = false;
2696
2697 stream->perf->ops.oa_enable(stream);
2698
2699 if (stream->sample_flags & SAMPLE_OA_REPORT)
2700 hrtimer_start(&stream->poll_check_timer,
2701 ns_to_ktime(stream->poll_oa_period),
2702 HRTIMER_MODE_REL_PINNED);
2703}
2704
2705static void gen7_oa_disable(struct i915_perf_stream *stream)
2706{
2707 struct intel_uncore *uncore = stream->uncore;
2708
2709 intel_uncore_write(uncore, GEN7_OACONTROL, 0);
2710 if (intel_wait_for_register(uncore,
2711 GEN7_OACONTROL, GEN7_OACONTROL_ENABLE, 0,
2712 50))
2713 drm_err(&stream->perf->i915->drm,
2714 "wait for OA to be disabled timed out\n");
2715}
2716
2717static void gen8_oa_disable(struct i915_perf_stream *stream)
2718{
2719 struct intel_uncore *uncore = stream->uncore;
2720
2721 intel_uncore_write(uncore, GEN8_OACONTROL, 0);
2722 if (intel_wait_for_register(uncore,
2723 GEN8_OACONTROL, GEN8_OA_COUNTER_ENABLE, 0,
2724 50))
2725 drm_err(&stream->perf->i915->drm,
2726 "wait for OA to be disabled timed out\n");
2727}
2728
2729static void gen12_oa_disable(struct i915_perf_stream *stream)
2730{
2731 struct intel_uncore *uncore = stream->uncore;
2732
2733 intel_uncore_write(uncore, GEN12_OAG_OACONTROL, 0);
2734 if (intel_wait_for_register(uncore,
2735 GEN12_OAG_OACONTROL,
2736 GEN12_OAG_OACONTROL_OA_COUNTER_ENABLE, 0,
2737 50))
2738 drm_err(&stream->perf->i915->drm,
2739 "wait for OA to be disabled timed out\n");
2740
2741 intel_uncore_write(uncore, GEN12_OA_TLB_INV_CR, 1);
2742 if (intel_wait_for_register(uncore,
2743 GEN12_OA_TLB_INV_CR,
2744 1, 0,
2745 50))
2746 drm_err(&stream->perf->i915->drm,
2747 "wait for OA tlb invalidate timed out\n");
2748}
2749
2750/**
2751 * i915_oa_stream_disable - handle `I915_PERF_IOCTL_DISABLE` for OA stream
2752 * @stream: An i915 perf stream opened for OA metrics
2753 *
2754 * Stops the OA unit from periodically writing counter reports into the
2755 * circular OA buffer. This also stops the hrtimer that periodically checks for
2756 * data in the circular OA buffer, for notifying userspace.
2757 */
2758static void i915_oa_stream_disable(struct i915_perf_stream *stream)
2759{
2760 stream->perf->ops.oa_disable(stream);
2761
2762 if (stream->sample_flags & SAMPLE_OA_REPORT)
2763 hrtimer_cancel(&stream->poll_check_timer);
2764}
2765
2766static const struct i915_perf_stream_ops i915_oa_stream_ops = {
2767 .destroy = i915_oa_stream_destroy,
2768 .enable = i915_oa_stream_enable,
2769 .disable = i915_oa_stream_disable,
2770 .wait_unlocked = i915_oa_wait_unlocked,
2771 .poll_wait = i915_oa_poll_wait,
2772 .read = i915_oa_read,
2773};
2774
2775static int i915_perf_stream_enable_sync(struct i915_perf_stream *stream)
2776{
2777 struct i915_active *active;
2778 int err;
2779
2780 active = i915_active_create();
2781 if (!active)
2782 return -ENOMEM;
2783
2784 err = stream->perf->ops.enable_metric_set(stream, active);
2785 if (err == 0)
2786 __i915_active_wait(active, TASK_UNINTERRUPTIBLE);
2787
2788 i915_active_put(active);
2789 return err;
2790}
2791
2792static void
2793get_default_sseu_config(struct intel_sseu *out_sseu,
2794 struct intel_engine_cs *engine)
2795{
2796 const struct sseu_dev_info *devinfo_sseu = &engine->gt->info.sseu;
2797
2798 *out_sseu = intel_sseu_from_device_info(devinfo_sseu);
2799
2800 if (GRAPHICS_VER(engine->i915) == 11) {
2801 /*
2802 * We only need subslice count so it doesn't matter which ones
2803 * we select - just turn off low bits in the amount of half of
2804 * all available subslices per slice.
2805 */
2806 out_sseu->subslice_mask =
2807 ~(~0 << (hweight8(out_sseu->subslice_mask) / 2));
2808 out_sseu->slice_mask = 0x1;
2809 }
2810}
2811
2812static int
2813get_sseu_config(struct intel_sseu *out_sseu,
2814 struct intel_engine_cs *engine,
2815 const struct drm_i915_gem_context_param_sseu *drm_sseu)
2816{
2817 if (drm_sseu->engine.engine_class != engine->uabi_class ||
2818 drm_sseu->engine.engine_instance != engine->uabi_instance)
2819 return -EINVAL;
2820
2821 return i915_gem_user_to_context_sseu(engine->gt, drm_sseu, out_sseu);
2822}
2823
2824/**
2825 * i915_oa_stream_init - validate combined props for OA stream and init
2826 * @stream: An i915 perf stream
2827 * @param: The open parameters passed to `DRM_I915_PERF_OPEN`
2828 * @props: The property state that configures stream (individually validated)
2829 *
2830 * While read_properties_unlocked() validates properties in isolation it
2831 * doesn't ensure that the combination necessarily makes sense.
2832 *
2833 * At this point it has been determined that userspace wants a stream of
2834 * OA metrics, but still we need to further validate the combined
2835 * properties are OK.
2836 *
2837 * If the configuration makes sense then we can allocate memory for
2838 * a circular OA buffer and apply the requested metric set configuration.
2839 *
2840 * Returns: zero on success or a negative error code.
2841 */
2842static int i915_oa_stream_init(struct i915_perf_stream *stream,
2843 struct drm_i915_perf_open_param *param,
2844 struct perf_open_properties *props)
2845{
2846 struct drm_i915_private *i915 = stream->perf->i915;
2847 struct i915_perf *perf = stream->perf;
2848 int format_size;
2849 int ret;
2850
2851 if (!props->engine) {
2852 DRM_DEBUG("OA engine not specified\n");
2853 return -EINVAL;
2854 }
2855
2856 /*
2857 * If the sysfs metrics/ directory wasn't registered for some
2858 * reason then don't let userspace try their luck with config
2859 * IDs
2860 */
2861 if (!perf->metrics_kobj) {
2862 DRM_DEBUG("OA metrics weren't advertised via sysfs\n");
2863 return -EINVAL;
2864 }
2865
2866 if (!(props->sample_flags & SAMPLE_OA_REPORT) &&
2867 (GRAPHICS_VER(perf->i915) < 12 || !stream->ctx)) {
2868 DRM_DEBUG("Only OA report sampling supported\n");
2869 return -EINVAL;
2870 }
2871
2872 if (!perf->ops.enable_metric_set) {
2873 DRM_DEBUG("OA unit not supported\n");
2874 return -ENODEV;
2875 }
2876
2877 /*
2878 * To avoid the complexity of having to accurately filter
2879 * counter reports and marshal to the appropriate client
2880 * we currently only allow exclusive access
2881 */
2882 if (perf->exclusive_stream) {
2883 DRM_DEBUG("OA unit already in use\n");
2884 return -EBUSY;
2885 }
2886
2887 if (!props->oa_format) {
2888 DRM_DEBUG("OA report format not specified\n");
2889 return -EINVAL;
2890 }
2891
2892 stream->engine = props->engine;
2893 stream->uncore = stream->engine->gt->uncore;
2894
2895 stream->sample_size = sizeof(struct drm_i915_perf_record_header);
2896
2897 format_size = perf->oa_formats[props->oa_format].size;
2898
2899 stream->sample_flags = props->sample_flags;
2900 stream->sample_size += format_size;
2901
2902 stream->oa_buffer.format_size = format_size;
2903 if (drm_WARN_ON(&i915->drm, stream->oa_buffer.format_size == 0))
2904 return -EINVAL;
2905
2906 stream->hold_preemption = props->hold_preemption;
2907
2908 stream->oa_buffer.format =
2909 perf->oa_formats[props->oa_format].format;
2910
2911 stream->periodic = props->oa_periodic;
2912 if (stream->periodic)
2913 stream->period_exponent = props->oa_period_exponent;
2914
2915 if (stream->ctx) {
2916 ret = oa_get_render_ctx_id(stream);
2917 if (ret) {
2918 DRM_DEBUG("Invalid context id to filter with\n");
2919 return ret;
2920 }
2921 }
2922
2923 ret = alloc_noa_wait(stream);
2924 if (ret) {
2925 DRM_DEBUG("Unable to allocate NOA wait batch buffer\n");
2926 goto err_noa_wait_alloc;
2927 }
2928
2929 stream->oa_config = i915_perf_get_oa_config(perf, props->metrics_set);
2930 if (!stream->oa_config) {
2931 DRM_DEBUG("Invalid OA config id=%i\n", props->metrics_set);
2932 ret = -EINVAL;
2933 goto err_config;
2934 }
2935
2936 /* PRM - observability performance counters:
2937 *
2938 * OACONTROL, performance counter enable, note:
2939 *
2940 * "When this bit is set, in order to have coherent counts,
2941 * RC6 power state and trunk clock gating must be disabled.
2942 * This can be achieved by programming MMIO registers as
2943 * 0xA094=0 and 0xA090[31]=1"
2944 *
2945 * In our case we are expecting that taking pm + FORCEWAKE
2946 * references will effectively disable RC6.
2947 */
2948 intel_engine_pm_get(stream->engine);
2949 intel_uncore_forcewake_get(stream->uncore, FORCEWAKE_ALL);
2950
2951 ret = alloc_oa_buffer(stream);
2952 if (ret)
2953 goto err_oa_buf_alloc;
2954
2955 stream->ops = &i915_oa_stream_ops;
2956
2957 perf->sseu = props->sseu;
2958 WRITE_ONCE(perf->exclusive_stream, stream);
2959
2960 ret = i915_perf_stream_enable_sync(stream);
2961 if (ret) {
2962 DRM_DEBUG("Unable to enable metric set\n");
2963 goto err_enable;
2964 }
2965
2966 DRM_DEBUG("opening stream oa config uuid=%s\n",
2967 stream->oa_config->uuid);
2968
2969 hrtimer_init(&stream->poll_check_timer,
2970 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2971 stream->poll_check_timer.function = oa_poll_check_timer_cb;
2972 init_waitqueue_head(&stream->poll_wq);
2973 spin_lock_init(&stream->oa_buffer.ptr_lock);
2974
2975 return 0;
2976
2977err_enable:
2978 WRITE_ONCE(perf->exclusive_stream, NULL);
2979 perf->ops.disable_metric_set(stream);
2980
2981 free_oa_buffer(stream);
2982
2983err_oa_buf_alloc:
2984 free_oa_configs(stream);
2985
2986 intel_uncore_forcewake_put(stream->uncore, FORCEWAKE_ALL);
2987 intel_engine_pm_put(stream->engine);
2988
2989err_config:
2990 free_noa_wait(stream);
2991
2992err_noa_wait_alloc:
2993 if (stream->ctx)
2994 oa_put_render_ctx_id(stream);
2995
2996 return ret;
2997}
2998
2999void i915_oa_init_reg_state(const struct intel_context *ce,
3000 const struct intel_engine_cs *engine)
3001{
3002 struct i915_perf_stream *stream;
3003
3004 if (engine->class != RENDER_CLASS)
3005 return;
3006
3007 /* perf.exclusive_stream serialised by lrc_configure_all_contexts() */
3008 stream = READ_ONCE(engine->i915->perf.exclusive_stream);
3009 if (stream && GRAPHICS_VER(stream->perf->i915) < 12)
3010 gen8_update_reg_state_unlocked(ce, stream);
3011}
3012
3013/**
3014 * i915_perf_read - handles read() FOP for i915 perf stream FDs
3015 * @file: An i915 perf stream file
3016 * @buf: destination buffer given by userspace
3017 * @count: the number of bytes userspace wants to read
3018 * @ppos: (inout) file seek position (unused)
3019 *
3020 * The entry point for handling a read() on a stream file descriptor from
3021 * userspace. Most of the work is left to the i915_perf_read_locked() and
3022 * &i915_perf_stream_ops->read but to save having stream implementations (of
3023 * which we might have multiple later) we handle blocking read here.
3024 *
3025 * We can also consistently treat trying to read from a disabled stream
3026 * as an IO error so implementations can assume the stream is enabled
3027 * while reading.
3028 *
3029 * Returns: The number of bytes copied or a negative error code on failure.
3030 */
3031static ssize_t i915_perf_read(struct file *file,
3032 char __user *buf,
3033 size_t count,
3034 loff_t *ppos)
3035{
3036 struct i915_perf_stream *stream = file->private_data;
3037 struct i915_perf *perf = stream->perf;
3038 size_t offset = 0;
3039 int ret;
3040
3041 /* To ensure it's handled consistently we simply treat all reads of a
3042 * disabled stream as an error. In particular it might otherwise lead
3043 * to a deadlock for blocking file descriptors...
3044 */
3045 if (!stream->enabled || !(stream->sample_flags & SAMPLE_OA_REPORT))
3046 return -EIO;
3047
3048 if (!(file->f_flags & O_NONBLOCK)) {
3049 /* There's the small chance of false positives from
3050 * stream->ops->wait_unlocked.
3051 *
3052 * E.g. with single context filtering since we only wait until
3053 * oabuffer has >= 1 report we don't immediately know whether
3054 * any reports really belong to the current context
3055 */
3056 do {
3057 ret = stream->ops->wait_unlocked(stream);
3058 if (ret)
3059 return ret;
3060
3061 mutex_lock(&perf->lock);
3062 ret = stream->ops->read(stream, buf, count, &offset);
3063 mutex_unlock(&perf->lock);
3064 } while (!offset && !ret);
3065 } else {
3066 mutex_lock(&perf->lock);
3067 ret = stream->ops->read(stream, buf, count, &offset);
3068 mutex_unlock(&perf->lock);
3069 }
3070
3071 /* We allow the poll checking to sometimes report false positive EPOLLIN
3072 * events where we might actually report EAGAIN on read() if there's
3073 * not really any data available. In this situation though we don't
3074 * want to enter a busy loop between poll() reporting a EPOLLIN event
3075 * and read() returning -EAGAIN. Clearing the oa.pollin state here
3076 * effectively ensures we back off until the next hrtimer callback
3077 * before reporting another EPOLLIN event.
3078 * The exception to this is if ops->read() returned -ENOSPC which means
3079 * that more OA data is available than could fit in the user provided
3080 * buffer. In this case we want the next poll() call to not block.
3081 */
3082 if (ret != -ENOSPC)
3083 stream->pollin = false;
3084
3085 /* Possible values for ret are 0, -EFAULT, -ENOSPC, -EIO, ... */
3086 return offset ?: (ret ?: -EAGAIN);
3087}
3088
3089static enum hrtimer_restart oa_poll_check_timer_cb(struct hrtimer *hrtimer)
3090{
3091 struct i915_perf_stream *stream =
3092 container_of(hrtimer, typeof(*stream), poll_check_timer);
3093
3094 if (oa_buffer_check_unlocked(stream)) {
3095 stream->pollin = true;
3096 wake_up(&stream->poll_wq);
3097 }
3098
3099 hrtimer_forward_now(hrtimer,
3100 ns_to_ktime(stream->poll_oa_period));
3101
3102 return HRTIMER_RESTART;
3103}
3104
3105/**
3106 * i915_perf_poll_locked - poll_wait() with a suitable wait queue for stream
3107 * @stream: An i915 perf stream
3108 * @file: An i915 perf stream file
3109 * @wait: poll() state table
3110 *
3111 * For handling userspace polling on an i915 perf stream, this calls through to
3112 * &i915_perf_stream_ops->poll_wait to call poll_wait() with a wait queue that
3113 * will be woken for new stream data.
3114 *
3115 * Note: The &perf->lock mutex has been taken to serialize
3116 * with any non-file-operation driver hooks.
3117 *
3118 * Returns: any poll events that are ready without sleeping
3119 */
3120static __poll_t i915_perf_poll_locked(struct i915_perf_stream *stream,
3121 struct file *file,
3122 poll_table *wait)
3123{
3124 __poll_t events = 0;
3125
3126 stream->ops->poll_wait(stream, file, wait);
3127
3128 /* Note: we don't explicitly check whether there's something to read
3129 * here since this path may be very hot depending on what else
3130 * userspace is polling, or on the timeout in use. We rely solely on
3131 * the hrtimer/oa_poll_check_timer_cb to notify us when there are
3132 * samples to read.
3133 */
3134 if (stream->pollin)
3135 events |= EPOLLIN;
3136
3137 return events;
3138}
3139
3140/**
3141 * i915_perf_poll - call poll_wait() with a suitable wait queue for stream
3142 * @file: An i915 perf stream file
3143 * @wait: poll() state table
3144 *
3145 * For handling userspace polling on an i915 perf stream, this ensures
3146 * poll_wait() gets called with a wait queue that will be woken for new stream
3147 * data.
3148 *
3149 * Note: Implementation deferred to i915_perf_poll_locked()
3150 *
3151 * Returns: any poll events that are ready without sleeping
3152 */
3153static __poll_t i915_perf_poll(struct file *file, poll_table *wait)
3154{
3155 struct i915_perf_stream *stream = file->private_data;
3156 struct i915_perf *perf = stream->perf;
3157 __poll_t ret;
3158
3159 mutex_lock(&perf->lock);
3160 ret = i915_perf_poll_locked(stream, file, wait);
3161 mutex_unlock(&perf->lock);
3162
3163 return ret;
3164}
3165
3166/**
3167 * i915_perf_enable_locked - handle `I915_PERF_IOCTL_ENABLE` ioctl
3168 * @stream: A disabled i915 perf stream
3169 *
3170 * [Re]enables the associated capture of data for this stream.
3171 *
3172 * If a stream was previously enabled then there's currently no intention
3173 * to provide userspace any guarantee about the preservation of previously
3174 * buffered data.
3175 */
3176static void i915_perf_enable_locked(struct i915_perf_stream *stream)
3177{
3178 if (stream->enabled)
3179 return;
3180
3181 /* Allow stream->ops->enable() to refer to this */
3182 stream->enabled = true;
3183
3184 if (stream->ops->enable)
3185 stream->ops->enable(stream);
3186
3187 if (stream->hold_preemption)
3188 intel_context_set_nopreempt(stream->pinned_ctx);
3189}
3190
3191/**
3192 * i915_perf_disable_locked - handle `I915_PERF_IOCTL_DISABLE` ioctl
3193 * @stream: An enabled i915 perf stream
3194 *
3195 * Disables the associated capture of data for this stream.
3196 *
3197 * The intention is that disabling an re-enabling a stream will ideally be
3198 * cheaper than destroying and re-opening a stream with the same configuration,
3199 * though there are no formal guarantees about what state or buffered data
3200 * must be retained between disabling and re-enabling a stream.
3201 *
3202 * Note: while a stream is disabled it's considered an error for userspace
3203 * to attempt to read from the stream (-EIO).
3204 */
3205static void i915_perf_disable_locked(struct i915_perf_stream *stream)
3206{
3207 if (!stream->enabled)
3208 return;
3209
3210 /* Allow stream->ops->disable() to refer to this */
3211 stream->enabled = false;
3212
3213 if (stream->hold_preemption)
3214 intel_context_clear_nopreempt(stream->pinned_ctx);
3215
3216 if (stream->ops->disable)
3217 stream->ops->disable(stream);
3218}
3219
3220static long i915_perf_config_locked(struct i915_perf_stream *stream,
3221 unsigned long metrics_set)
3222{
3223 struct i915_oa_config *config;
3224 long ret = stream->oa_config->id;
3225
3226 config = i915_perf_get_oa_config(stream->perf, metrics_set);
3227 if (!config)
3228 return -EINVAL;
3229
3230 if (config != stream->oa_config) {
3231 int err;
3232
3233 /*
3234 * If OA is bound to a specific context, emit the
3235 * reconfiguration inline from that context. The update
3236 * will then be ordered with respect to submission on that
3237 * context.
3238 *
3239 * When set globally, we use a low priority kernel context,
3240 * so it will effectively take effect when idle.
3241 */
3242 err = emit_oa_config(stream, config, oa_context(stream), NULL);
3243 if (!err)
3244 config = xchg(&stream->oa_config, config);
3245 else
3246 ret = err;
3247 }
3248
3249 i915_oa_config_put(config);
3250
3251 return ret;
3252}
3253
3254/**
3255 * i915_perf_ioctl_locked - support ioctl() usage with i915 perf stream FDs
3256 * @stream: An i915 perf stream
3257 * @cmd: the ioctl request
3258 * @arg: the ioctl data
3259 *
3260 * Note: The &perf->lock mutex has been taken to serialize
3261 * with any non-file-operation driver hooks.
3262 *
3263 * Returns: zero on success or a negative error code. Returns -EINVAL for
3264 * an unknown ioctl request.
3265 */
3266static long i915_perf_ioctl_locked(struct i915_perf_stream *stream,
3267 unsigned int cmd,
3268 unsigned long arg)
3269{
3270 switch (cmd) {
3271 case I915_PERF_IOCTL_ENABLE:
3272 i915_perf_enable_locked(stream);
3273 return 0;
3274 case I915_PERF_IOCTL_DISABLE:
3275 i915_perf_disable_locked(stream);
3276 return 0;
3277 case I915_PERF_IOCTL_CONFIG:
3278 return i915_perf_config_locked(stream, arg);
3279 }
3280
3281 return -EINVAL;
3282}
3283
3284/**
3285 * i915_perf_ioctl - support ioctl() usage with i915 perf stream FDs
3286 * @file: An i915 perf stream file
3287 * @cmd: the ioctl request
3288 * @arg: the ioctl data
3289 *
3290 * Implementation deferred to i915_perf_ioctl_locked().
3291 *
3292 * Returns: zero on success or a negative error code. Returns -EINVAL for
3293 * an unknown ioctl request.
3294 */
3295static long i915_perf_ioctl(struct file *file,
3296 unsigned int cmd,
3297 unsigned long arg)
3298{
3299 struct i915_perf_stream *stream = file->private_data;
3300 struct i915_perf *perf = stream->perf;
3301 long ret;
3302
3303 mutex_lock(&perf->lock);
3304 ret = i915_perf_ioctl_locked(stream, cmd, arg);
3305 mutex_unlock(&perf->lock);
3306
3307 return ret;
3308}
3309
3310/**
3311 * i915_perf_destroy_locked - destroy an i915 perf stream
3312 * @stream: An i915 perf stream
3313 *
3314 * Frees all resources associated with the given i915 perf @stream, disabling
3315 * any associated data capture in the process.
3316 *
3317 * Note: The &perf->lock mutex has been taken to serialize
3318 * with any non-file-operation driver hooks.
3319 */
3320static void i915_perf_destroy_locked(struct i915_perf_stream *stream)
3321{
3322 if (stream->enabled)
3323 i915_perf_disable_locked(stream);
3324
3325 if (stream->ops->destroy)
3326 stream->ops->destroy(stream);
3327
3328 if (stream->ctx)
3329 i915_gem_context_put(stream->ctx);
3330
3331 kfree(stream);
3332}
3333
3334/**
3335 * i915_perf_release - handles userspace close() of a stream file
3336 * @inode: anonymous inode associated with file
3337 * @file: An i915 perf stream file
3338 *
3339 * Cleans up any resources associated with an open i915 perf stream file.
3340 *
3341 * NB: close() can't really fail from the userspace point of view.
3342 *
3343 * Returns: zero on success or a negative error code.
3344 */
3345static int i915_perf_release(struct inode *inode, struct file *file)
3346{
3347 struct i915_perf_stream *stream = file->private_data;
3348 struct i915_perf *perf = stream->perf;
3349
3350 mutex_lock(&perf->lock);
3351 i915_perf_destroy_locked(stream);
3352 mutex_unlock(&perf->lock);
3353
3354 /* Release the reference the perf stream kept on the driver. */
3355 drm_dev_put(&perf->i915->drm);
3356
3357 return 0;
3358}
3359
3360
3361static const struct file_operations fops = {
3362 .owner = THIS_MODULE,
3363 .llseek = no_llseek,
3364 .release = i915_perf_release,
3365 .poll = i915_perf_poll,
3366 .read = i915_perf_read,
3367 .unlocked_ioctl = i915_perf_ioctl,
3368 /* Our ioctl have no arguments, so it's safe to use the same function
3369 * to handle 32bits compatibility.
3370 */
3371 .compat_ioctl = i915_perf_ioctl,
3372};
3373
3374
3375/**
3376 * i915_perf_open_ioctl_locked - DRM ioctl() for userspace to open a stream FD
3377 * @perf: i915 perf instance
3378 * @param: The open parameters passed to 'DRM_I915_PERF_OPEN`
3379 * @props: individually validated u64 property value pairs
3380 * @file: drm file
3381 *
3382 * See i915_perf_ioctl_open() for interface details.
3383 *
3384 * Implements further stream config validation and stream initialization on
3385 * behalf of i915_perf_open_ioctl() with the &perf->lock mutex
3386 * taken to serialize with any non-file-operation driver hooks.
3387 *
3388 * Note: at this point the @props have only been validated in isolation and
3389 * it's still necessary to validate that the combination of properties makes
3390 * sense.
3391 *
3392 * In the case where userspace is interested in OA unit metrics then further
3393 * config validation and stream initialization details will be handled by
3394 * i915_oa_stream_init(). The code here should only validate config state that
3395 * will be relevant to all stream types / backends.
3396 *
3397 * Returns: zero on success or a negative error code.
3398 */
3399static int
3400i915_perf_open_ioctl_locked(struct i915_perf *perf,
3401 struct drm_i915_perf_open_param *param,
3402 struct perf_open_properties *props,
3403 struct drm_file *file)
3404{
3405 struct i915_gem_context *specific_ctx = NULL;
3406 struct i915_perf_stream *stream = NULL;
3407 unsigned long f_flags = 0;
3408 bool privileged_op = true;
3409 int stream_fd;
3410 int ret;
3411
3412 if (props->single_context) {
3413 u32 ctx_handle = props->ctx_handle;
3414 struct drm_i915_file_private *file_priv = file->driver_priv;
3415
3416 specific_ctx = i915_gem_context_lookup(file_priv, ctx_handle);
3417 if (!specific_ctx) {
3418 DRM_DEBUG("Failed to look up context with ID %u for opening perf stream\n",
3419 ctx_handle);
3420 ret = -ENOENT;
3421 goto err;
3422 }
3423 }
3424
3425 /*
3426 * On Haswell the OA unit supports clock gating off for a specific
3427 * context and in this mode there's no visibility of metrics for the
3428 * rest of the system, which we consider acceptable for a
3429 * non-privileged client.
3430 *
3431 * For Gen8->11 the OA unit no longer supports clock gating off for a
3432 * specific context and the kernel can't securely stop the counters
3433 * from updating as system-wide / global values. Even though we can
3434 * filter reports based on the included context ID we can't block
3435 * clients from seeing the raw / global counter values via
3436 * MI_REPORT_PERF_COUNT commands and so consider it a privileged op to
3437 * enable the OA unit by default.
3438 *
3439 * For Gen12+ we gain a new OAR unit that only monitors the RCS on a
3440 * per context basis. So we can relax requirements there if the user
3441 * doesn't request global stream access (i.e. query based sampling
3442 * using MI_RECORD_PERF_COUNT.
3443 */
3444 if (IS_HASWELL(perf->i915) && specific_ctx)
3445 privileged_op = false;
3446 else if (GRAPHICS_VER(perf->i915) == 12 && specific_ctx &&
3447 (props->sample_flags & SAMPLE_OA_REPORT) == 0)
3448 privileged_op = false;
3449
3450 if (props->hold_preemption) {
3451 if (!props->single_context) {
3452 DRM_DEBUG("preemption disable with no context\n");
3453 ret = -EINVAL;
3454 goto err;
3455 }
3456 privileged_op = true;
3457 }
3458
3459 /*
3460 * Asking for SSEU configuration is a priviliged operation.
3461 */
3462 if (props->has_sseu)
3463 privileged_op = true;
3464 else
3465 get_default_sseu_config(&props->sseu, props->engine);
3466
3467 /* Similar to perf's kernel.perf_paranoid_cpu sysctl option
3468 * we check a dev.i915.perf_stream_paranoid sysctl option
3469 * to determine if it's ok to access system wide OA counters
3470 * without CAP_PERFMON or CAP_SYS_ADMIN privileges.
3471 */
3472 if (privileged_op &&
3473 i915_perf_stream_paranoid && !perfmon_capable()) {
3474 DRM_DEBUG("Insufficient privileges to open i915 perf stream\n");
3475 ret = -EACCES;
3476 goto err_ctx;
3477 }
3478
3479 stream = kzalloc(sizeof(*stream), GFP_KERNEL);
3480 if (!stream) {
3481 ret = -ENOMEM;
3482 goto err_ctx;
3483 }
3484
3485 stream->perf = perf;
3486 stream->ctx = specific_ctx;
3487 stream->poll_oa_period = props->poll_oa_period;
3488
3489 ret = i915_oa_stream_init(stream, param, props);
3490 if (ret)
3491 goto err_alloc;
3492
3493 /* we avoid simply assigning stream->sample_flags = props->sample_flags
3494 * to have _stream_init check the combination of sample flags more
3495 * thoroughly, but still this is the expected result at this point.
3496 */
3497 if (WARN_ON(stream->sample_flags != props->sample_flags)) {
3498 ret = -ENODEV;
3499 goto err_flags;
3500 }
3501
3502 if (param->flags & I915_PERF_FLAG_FD_CLOEXEC)
3503 f_flags |= O_CLOEXEC;
3504 if (param->flags & I915_PERF_FLAG_FD_NONBLOCK)
3505 f_flags |= O_NONBLOCK;
3506
3507 stream_fd = anon_inode_getfd("[i915_perf]", &fops, stream, f_flags);
3508 if (stream_fd < 0) {
3509 ret = stream_fd;
3510 goto err_flags;
3511 }
3512
3513 if (!(param->flags & I915_PERF_FLAG_DISABLED))
3514 i915_perf_enable_locked(stream);
3515
3516 /* Take a reference on the driver that will be kept with stream_fd
3517 * until its release.
3518 */
3519 drm_dev_get(&perf->i915->drm);
3520
3521 return stream_fd;
3522
3523err_flags:
3524 if (stream->ops->destroy)
3525 stream->ops->destroy(stream);
3526err_alloc:
3527 kfree(stream);
3528err_ctx:
3529 if (specific_ctx)
3530 i915_gem_context_put(specific_ctx);
3531err:
3532 return ret;
3533}
3534
3535static u64 oa_exponent_to_ns(struct i915_perf *perf, int exponent)
3536{
3537 return intel_gt_clock_interval_to_ns(perf->i915->ggtt.vm.gt,
3538 2ULL << exponent);
3539}
3540
3541static __always_inline bool
3542oa_format_valid(struct i915_perf *perf, enum drm_i915_oa_format format)
3543{
3544 return test_bit(format, perf->format_mask);
3545}
3546
3547static __always_inline void
3548oa_format_add(struct i915_perf *perf, enum drm_i915_oa_format format)
3549{
3550 __set_bit(format, perf->format_mask);
3551}
3552
3553/**
3554 * read_properties_unlocked - validate + copy userspace stream open properties
3555 * @perf: i915 perf instance
3556 * @uprops: The array of u64 key value pairs given by userspace
3557 * @n_props: The number of key value pairs expected in @uprops
3558 * @props: The stream configuration built up while validating properties
3559 *
3560 * Note this function only validates properties in isolation it doesn't
3561 * validate that the combination of properties makes sense or that all
3562 * properties necessary for a particular kind of stream have been set.
3563 *
3564 * Note that there currently aren't any ordering requirements for properties so
3565 * we shouldn't validate or assume anything about ordering here. This doesn't
3566 * rule out defining new properties with ordering requirements in the future.
3567 */
3568static int read_properties_unlocked(struct i915_perf *perf,
3569 u64 __user *uprops,
3570 u32 n_props,
3571 struct perf_open_properties *props)
3572{
3573 u64 __user *uprop = uprops;
3574 u32 i;
3575 int ret;
3576
3577 memset(props, 0, sizeof(struct perf_open_properties));
3578 props->poll_oa_period = DEFAULT_POLL_PERIOD_NS;
3579
3580 if (!n_props) {
3581 DRM_DEBUG("No i915 perf properties given\n");
3582 return -EINVAL;
3583 }
3584
3585 /* At the moment we only support using i915-perf on the RCS. */
3586 props->engine = intel_engine_lookup_user(perf->i915,
3587 I915_ENGINE_CLASS_RENDER,
3588 0);
3589 if (!props->engine) {
3590 DRM_DEBUG("No RENDER-capable engines\n");
3591 return -EINVAL;
3592 }
3593
3594 /* Considering that ID = 0 is reserved and assuming that we don't
3595 * (currently) expect any configurations to ever specify duplicate
3596 * values for a particular property ID then the last _PROP_MAX value is
3597 * one greater than the maximum number of properties we expect to get
3598 * from userspace.
3599 */
3600 if (n_props >= DRM_I915_PERF_PROP_MAX) {
3601 DRM_DEBUG("More i915 perf properties specified than exist\n");
3602 return -EINVAL;
3603 }
3604
3605 for (i = 0; i < n_props; i++) {
3606 u64 oa_period, oa_freq_hz;
3607 u64 id, value;
3608
3609 ret = get_user(id, uprop);
3610 if (ret)
3611 return ret;
3612
3613 ret = get_user(value, uprop + 1);
3614 if (ret)
3615 return ret;
3616
3617 if (id == 0 || id >= DRM_I915_PERF_PROP_MAX) {
3618 DRM_DEBUG("Unknown i915 perf property ID\n");
3619 return -EINVAL;
3620 }
3621
3622 switch ((enum drm_i915_perf_property_id)id) {
3623 case DRM_I915_PERF_PROP_CTX_HANDLE:
3624 props->single_context = 1;
3625 props->ctx_handle = value;
3626 break;
3627 case DRM_I915_PERF_PROP_SAMPLE_OA:
3628 if (value)
3629 props->sample_flags |= SAMPLE_OA_REPORT;
3630 break;
3631 case DRM_I915_PERF_PROP_OA_METRICS_SET:
3632 if (value == 0) {
3633 DRM_DEBUG("Unknown OA metric set ID\n");
3634 return -EINVAL;
3635 }
3636 props->metrics_set = value;
3637 break;
3638 case DRM_I915_PERF_PROP_OA_FORMAT:
3639 if (value == 0 || value >= I915_OA_FORMAT_MAX) {
3640 DRM_DEBUG("Out-of-range OA report format %llu\n",
3641 value);
3642 return -EINVAL;
3643 }
3644 if (!oa_format_valid(perf, value)) {
3645 DRM_DEBUG("Unsupported OA report format %llu\n",
3646 value);
3647 return -EINVAL;
3648 }
3649 props->oa_format = value;
3650 break;
3651 case DRM_I915_PERF_PROP_OA_EXPONENT:
3652 if (value > OA_EXPONENT_MAX) {
3653 DRM_DEBUG("OA timer exponent too high (> %u)\n",
3654 OA_EXPONENT_MAX);
3655 return -EINVAL;
3656 }
3657
3658 /* Theoretically we can program the OA unit to sample
3659 * e.g. every 160ns for HSW, 167ns for BDW/SKL or 104ns
3660 * for BXT. We don't allow such high sampling
3661 * frequencies by default unless root.
3662 */
3663
3664 BUILD_BUG_ON(sizeof(oa_period) != 8);
3665 oa_period = oa_exponent_to_ns(perf, value);
3666
3667 /* This check is primarily to ensure that oa_period <=
3668 * UINT32_MAX (before passing to do_div which only
3669 * accepts a u32 denominator), but we can also skip
3670 * checking anything < 1Hz which implicitly can't be
3671 * limited via an integer oa_max_sample_rate.
3672 */
3673 if (oa_period <= NSEC_PER_SEC) {
3674 u64 tmp = NSEC_PER_SEC;
3675 do_div(tmp, oa_period);
3676 oa_freq_hz = tmp;
3677 } else
3678 oa_freq_hz = 0;
3679
3680 if (oa_freq_hz > i915_oa_max_sample_rate && !perfmon_capable()) {
3681 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",
3682 i915_oa_max_sample_rate);
3683 return -EACCES;
3684 }
3685
3686 props->oa_periodic = true;
3687 props->oa_period_exponent = value;
3688 break;
3689 case DRM_I915_PERF_PROP_HOLD_PREEMPTION:
3690 props->hold_preemption = !!value;
3691 break;
3692 case DRM_I915_PERF_PROP_GLOBAL_SSEU: {
3693 struct drm_i915_gem_context_param_sseu user_sseu;
3694
3695 if (copy_from_user(&user_sseu,
3696 u64_to_user_ptr(value),
3697 sizeof(user_sseu))) {
3698 DRM_DEBUG("Unable to copy global sseu parameter\n");
3699 return -EFAULT;
3700 }
3701
3702 ret = get_sseu_config(&props->sseu, props->engine, &user_sseu);
3703 if (ret) {
3704 DRM_DEBUG("Invalid SSEU configuration\n");
3705 return ret;
3706 }
3707 props->has_sseu = true;
3708 break;
3709 }
3710 case DRM_I915_PERF_PROP_POLL_OA_PERIOD:
3711 if (value < 100000 /* 100us */) {
3712 DRM_DEBUG("OA availability timer too small (%lluns < 100us)\n",
3713 value);
3714 return -EINVAL;
3715 }
3716 props->poll_oa_period = value;
3717 break;
3718 case DRM_I915_PERF_PROP_MAX:
3719 MISSING_CASE(id);
3720 return -EINVAL;
3721 }
3722
3723 uprop += 2;
3724 }
3725
3726 return 0;
3727}
3728
3729/**
3730 * i915_perf_open_ioctl - DRM ioctl() for userspace to open a stream FD
3731 * @dev: drm device
3732 * @data: ioctl data copied from userspace (unvalidated)
3733 * @file: drm file
3734 *
3735 * Validates the stream open parameters given by userspace including flags
3736 * and an array of u64 key, value pair properties.
3737 *
3738 * Very little is assumed up front about the nature of the stream being
3739 * opened (for instance we don't assume it's for periodic OA unit metrics). An
3740 * i915-perf stream is expected to be a suitable interface for other forms of
3741 * buffered data written by the GPU besides periodic OA metrics.
3742 *
3743 * Note we copy the properties from userspace outside of the i915 perf
3744 * mutex to avoid an awkward lockdep with mmap_lock.
3745 *
3746 * Most of the implementation details are handled by
3747 * i915_perf_open_ioctl_locked() after taking the &perf->lock
3748 * mutex for serializing with any non-file-operation driver hooks.
3749 *
3750 * Return: A newly opened i915 Perf stream file descriptor or negative
3751 * error code on failure.
3752 */
3753int i915_perf_open_ioctl(struct drm_device *dev, void *data,
3754 struct drm_file *file)
3755{
3756 struct i915_perf *perf = &to_i915(dev)->perf;
3757 struct drm_i915_perf_open_param *param = data;
3758 struct perf_open_properties props;
3759 u32 known_open_flags;
3760 int ret;
3761
3762 if (!perf->i915) {
3763 DRM_DEBUG("i915 perf interface not available for this system\n");
3764 return -ENOTSUPP;
3765 }
3766
3767 known_open_flags = I915_PERF_FLAG_FD_CLOEXEC |
3768 I915_PERF_FLAG_FD_NONBLOCK |
3769 I915_PERF_FLAG_DISABLED;
3770 if (param->flags & ~known_open_flags) {
3771 DRM_DEBUG("Unknown drm_i915_perf_open_param flag\n");
3772 return -EINVAL;
3773 }
3774
3775 ret = read_properties_unlocked(perf,
3776 u64_to_user_ptr(param->properties_ptr),
3777 param->num_properties,
3778 &props);
3779 if (ret)
3780 return ret;
3781
3782 mutex_lock(&perf->lock);
3783 ret = i915_perf_open_ioctl_locked(perf, param, &props, file);
3784 mutex_unlock(&perf->lock);
3785
3786 return ret;
3787}
3788
3789/**
3790 * i915_perf_register - exposes i915-perf to userspace
3791 * @i915: i915 device instance
3792 *
3793 * In particular OA metric sets are advertised under a sysfs metrics/
3794 * directory allowing userspace to enumerate valid IDs that can be
3795 * used to open an i915-perf stream.
3796 */
3797void i915_perf_register(struct drm_i915_private *i915)
3798{
3799 struct i915_perf *perf = &i915->perf;
3800
3801 if (!perf->i915)
3802 return;
3803
3804 /* To be sure we're synchronized with an attempted
3805 * i915_perf_open_ioctl(); considering that we register after
3806 * being exposed to userspace.
3807 */
3808 mutex_lock(&perf->lock);
3809
3810 perf->metrics_kobj =
3811 kobject_create_and_add("metrics",
3812 &i915->drm.primary->kdev->kobj);
3813
3814 mutex_unlock(&perf->lock);
3815}
3816
3817/**
3818 * i915_perf_unregister - hide i915-perf from userspace
3819 * @i915: i915 device instance
3820 *
3821 * i915-perf state cleanup is split up into an 'unregister' and
3822 * 'deinit' phase where the interface is first hidden from
3823 * userspace by i915_perf_unregister() before cleaning up
3824 * remaining state in i915_perf_fini().
3825 */
3826void i915_perf_unregister(struct drm_i915_private *i915)
3827{
3828 struct i915_perf *perf = &i915->perf;
3829
3830 if (!perf->metrics_kobj)
3831 return;
3832
3833 kobject_put(perf->metrics_kobj);
3834 perf->metrics_kobj = NULL;
3835}
3836
3837static bool gen8_is_valid_flex_addr(struct i915_perf *perf, u32 addr)
3838{
3839 static const i915_reg_t flex_eu_regs[] = {
3840 EU_PERF_CNTL0,
3841 EU_PERF_CNTL1,
3842 EU_PERF_CNTL2,
3843 EU_PERF_CNTL3,
3844 EU_PERF_CNTL4,
3845 EU_PERF_CNTL5,
3846 EU_PERF_CNTL6,
3847 };
3848 int i;
3849
3850 for (i = 0; i < ARRAY_SIZE(flex_eu_regs); i++) {
3851 if (i915_mmio_reg_offset(flex_eu_regs[i]) == addr)
3852 return true;
3853 }
3854 return false;
3855}
3856
3857#define ADDR_IN_RANGE(addr, start, end) \
3858 ((addr) >= (start) && \
3859 (addr) <= (end))
3860
3861#define REG_IN_RANGE(addr, start, end) \
3862 ((addr) >= i915_mmio_reg_offset(start) && \
3863 (addr) <= i915_mmio_reg_offset(end))
3864
3865#define REG_EQUAL(addr, mmio) \
3866 ((addr) == i915_mmio_reg_offset(mmio))
3867
3868static bool gen7_is_valid_b_counter_addr(struct i915_perf *perf, u32 addr)
3869{
3870 return REG_IN_RANGE(addr, OASTARTTRIG1, OASTARTTRIG8) ||
3871 REG_IN_RANGE(addr, OAREPORTTRIG1, OAREPORTTRIG8) ||
3872 REG_IN_RANGE(addr, OACEC0_0, OACEC7_1);
3873}
3874
3875static bool gen7_is_valid_mux_addr(struct i915_perf *perf, u32 addr)
3876{
3877 return REG_EQUAL(addr, HALF_SLICE_CHICKEN2) ||
3878 REG_IN_RANGE(addr, MICRO_BP0_0, NOA_WRITE) ||
3879 REG_IN_RANGE(addr, OA_PERFCNT1_LO, OA_PERFCNT2_HI) ||
3880 REG_IN_RANGE(addr, OA_PERFMATRIX_LO, OA_PERFMATRIX_HI);
3881}
3882
3883static bool gen8_is_valid_mux_addr(struct i915_perf *perf, u32 addr)
3884{
3885 return gen7_is_valid_mux_addr(perf, addr) ||
3886 REG_EQUAL(addr, WAIT_FOR_RC6_EXIT) ||
3887 REG_IN_RANGE(addr, RPM_CONFIG0, NOA_CONFIG(8));
3888}
3889
3890static bool gen10_is_valid_mux_addr(struct i915_perf *perf, u32 addr)
3891{
3892 return gen8_is_valid_mux_addr(perf, addr) ||
3893 REG_EQUAL(addr, GEN10_NOA_WRITE_HIGH) ||
3894 REG_IN_RANGE(addr, OA_PERFCNT3_LO, OA_PERFCNT4_HI);
3895}
3896
3897static bool hsw_is_valid_mux_addr(struct i915_perf *perf, u32 addr)
3898{
3899 return gen7_is_valid_mux_addr(perf, addr) ||
3900 ADDR_IN_RANGE(addr, 0x25100, 0x2FF90) ||
3901 REG_IN_RANGE(addr, HSW_MBVID2_NOA0, HSW_MBVID2_NOA9) ||
3902 REG_EQUAL(addr, HSW_MBVID2_MISR0);
3903}
3904
3905static bool chv_is_valid_mux_addr(struct i915_perf *perf, u32 addr)
3906{
3907 return gen7_is_valid_mux_addr(perf, addr) ||
3908 ADDR_IN_RANGE(addr, 0x182300, 0x1823A4);
3909}
3910
3911static bool gen12_is_valid_b_counter_addr(struct i915_perf *perf, u32 addr)
3912{
3913 return REG_IN_RANGE(addr, GEN12_OAG_OASTARTTRIG1, GEN12_OAG_OASTARTTRIG8) ||
3914 REG_IN_RANGE(addr, GEN12_OAG_OAREPORTTRIG1, GEN12_OAG_OAREPORTTRIG8) ||
3915 REG_IN_RANGE(addr, GEN12_OAG_CEC0_0, GEN12_OAG_CEC7_1) ||
3916 REG_IN_RANGE(addr, GEN12_OAG_SCEC0_0, GEN12_OAG_SCEC7_1) ||
3917 REG_EQUAL(addr, GEN12_OAA_DBG_REG) ||
3918 REG_EQUAL(addr, GEN12_OAG_OA_PESS) ||
3919 REG_EQUAL(addr, GEN12_OAG_SPCTR_CNF);
3920}
3921
3922static bool gen12_is_valid_mux_addr(struct i915_perf *perf, u32 addr)
3923{
3924 return REG_EQUAL(addr, NOA_WRITE) ||
3925 REG_EQUAL(addr, GEN10_NOA_WRITE_HIGH) ||
3926 REG_EQUAL(addr, GDT_CHICKEN_BITS) ||
3927 REG_EQUAL(addr, WAIT_FOR_RC6_EXIT) ||
3928 REG_EQUAL(addr, RPM_CONFIG0) ||
3929 REG_EQUAL(addr, RPM_CONFIG1) ||
3930 REG_IN_RANGE(addr, NOA_CONFIG(0), NOA_CONFIG(8));
3931}
3932
3933static u32 mask_reg_value(u32 reg, u32 val)
3934{
3935 /* HALF_SLICE_CHICKEN2 is programmed with a the
3936 * WaDisableSTUnitPowerOptimization workaround. Make sure the value
3937 * programmed by userspace doesn't change this.
3938 */
3939 if (REG_EQUAL(reg, HALF_SLICE_CHICKEN2))
3940 val = val & ~_MASKED_BIT_ENABLE(GEN8_ST_PO_DISABLE);
3941
3942 /* WAIT_FOR_RC6_EXIT has only one bit fullfilling the function
3943 * indicated by its name and a bunch of selection fields used by OA
3944 * configs.
3945 */
3946 if (REG_EQUAL(reg, WAIT_FOR_RC6_EXIT))
3947 val = val & ~_MASKED_BIT_ENABLE(HSW_WAIT_FOR_RC6_EXIT_ENABLE);
3948
3949 return val;
3950}
3951
3952static struct i915_oa_reg *alloc_oa_regs(struct i915_perf *perf,
3953 bool (*is_valid)(struct i915_perf *perf, u32 addr),
3954 u32 __user *regs,
3955 u32 n_regs)
3956{
3957 struct i915_oa_reg *oa_regs;
3958 int err;
3959 u32 i;
3960
3961 if (!n_regs)
3962 return NULL;
3963
3964 /* No is_valid function means we're not allowing any register to be programmed. */
3965 GEM_BUG_ON(!is_valid);
3966 if (!is_valid)
3967 return ERR_PTR(-EINVAL);
3968
3969 oa_regs = kmalloc_array(n_regs, sizeof(*oa_regs), GFP_KERNEL);
3970 if (!oa_regs)
3971 return ERR_PTR(-ENOMEM);
3972
3973 for (i = 0; i < n_regs; i++) {
3974 u32 addr, value;
3975
3976 err = get_user(addr, regs);
3977 if (err)
3978 goto addr_err;
3979
3980 if (!is_valid(perf, addr)) {
3981 DRM_DEBUG("Invalid oa_reg address: %X\n", addr);
3982 err = -EINVAL;
3983 goto addr_err;
3984 }
3985
3986 err = get_user(value, regs + 1);
3987 if (err)
3988 goto addr_err;
3989
3990 oa_regs[i].addr = _MMIO(addr);
3991 oa_regs[i].value = mask_reg_value(addr, value);
3992
3993 regs += 2;
3994 }
3995
3996 return oa_regs;
3997
3998addr_err:
3999 kfree(oa_regs);
4000 return ERR_PTR(err);
4001}
4002
4003static ssize_t show_dynamic_id(struct device *dev,
4004 struct device_attribute *attr,
4005 char *buf)
4006{
4007 struct i915_oa_config *oa_config =
4008 container_of(attr, typeof(*oa_config), sysfs_metric_id);
4009
4010 return sprintf(buf, "%d\n", oa_config->id);
4011}
4012
4013static int create_dynamic_oa_sysfs_entry(struct i915_perf *perf,
4014 struct i915_oa_config *oa_config)
4015{
4016 sysfs_attr_init(&oa_config->sysfs_metric_id.attr);
4017 oa_config->sysfs_metric_id.attr.name = "id";
4018 oa_config->sysfs_metric_id.attr.mode = S_IRUGO;
4019 oa_config->sysfs_metric_id.show = show_dynamic_id;
4020 oa_config->sysfs_metric_id.store = NULL;
4021
4022 oa_config->attrs[0] = &oa_config->sysfs_metric_id.attr;
4023 oa_config->attrs[1] = NULL;
4024
4025 oa_config->sysfs_metric.name = oa_config->uuid;
4026 oa_config->sysfs_metric.attrs = oa_config->attrs;
4027
4028 return sysfs_create_group(perf->metrics_kobj,
4029 &oa_config->sysfs_metric);
4030}
4031
4032/**
4033 * i915_perf_add_config_ioctl - DRM ioctl() for userspace to add a new OA config
4034 * @dev: drm device
4035 * @data: ioctl data (pointer to struct drm_i915_perf_oa_config) copied from
4036 * userspace (unvalidated)
4037 * @file: drm file
4038 *
4039 * Validates the submitted OA register to be saved into a new OA config that
4040 * can then be used for programming the OA unit and its NOA network.
4041 *
4042 * Returns: A new allocated config number to be used with the perf open ioctl
4043 * or a negative error code on failure.
4044 */
4045int i915_perf_add_config_ioctl(struct drm_device *dev, void *data,
4046 struct drm_file *file)
4047{
4048 struct i915_perf *perf = &to_i915(dev)->perf;
4049 struct drm_i915_perf_oa_config *args = data;
4050 struct i915_oa_config *oa_config, *tmp;
4051 struct i915_oa_reg *regs;
4052 int err, id;
4053
4054 if (!perf->i915) {
4055 DRM_DEBUG("i915 perf interface not available for this system\n");
4056 return -ENOTSUPP;
4057 }
4058
4059 if (!perf->metrics_kobj) {
4060 DRM_DEBUG("OA metrics weren't advertised via sysfs\n");
4061 return -EINVAL;
4062 }
4063
4064 if (i915_perf_stream_paranoid && !perfmon_capable()) {
4065 DRM_DEBUG("Insufficient privileges to add i915 OA config\n");
4066 return -EACCES;
4067 }
4068
4069 if ((!args->mux_regs_ptr || !args->n_mux_regs) &&
4070 (!args->boolean_regs_ptr || !args->n_boolean_regs) &&
4071 (!args->flex_regs_ptr || !args->n_flex_regs)) {
4072 DRM_DEBUG("No OA registers given\n");
4073 return -EINVAL;
4074 }
4075
4076 oa_config = kzalloc(sizeof(*oa_config), GFP_KERNEL);
4077 if (!oa_config) {
4078 DRM_DEBUG("Failed to allocate memory for the OA config\n");
4079 return -ENOMEM;
4080 }
4081
4082 oa_config->perf = perf;
4083 kref_init(&oa_config->ref);
4084
4085 if (!uuid_is_valid(args->uuid)) {
4086 DRM_DEBUG("Invalid uuid format for OA config\n");
4087 err = -EINVAL;
4088 goto reg_err;
4089 }
4090
4091 /* Last character in oa_config->uuid will be 0 because oa_config is
4092 * kzalloc.
4093 */
4094 memcpy(oa_config->uuid, args->uuid, sizeof(args->uuid));
4095
4096 oa_config->mux_regs_len = args->n_mux_regs;
4097 regs = alloc_oa_regs(perf,
4098 perf->ops.is_valid_mux_reg,
4099 u64_to_user_ptr(args->mux_regs_ptr),
4100 args->n_mux_regs);
4101
4102 if (IS_ERR(regs)) {
4103 DRM_DEBUG("Failed to create OA config for mux_regs\n");
4104 err = PTR_ERR(regs);
4105 goto reg_err;
4106 }
4107 oa_config->mux_regs = regs;
4108
4109 oa_config->b_counter_regs_len = args->n_boolean_regs;
4110 regs = alloc_oa_regs(perf,
4111 perf->ops.is_valid_b_counter_reg,
4112 u64_to_user_ptr(args->boolean_regs_ptr),
4113 args->n_boolean_regs);
4114
4115 if (IS_ERR(regs)) {
4116 DRM_DEBUG("Failed to create OA config for b_counter_regs\n");
4117 err = PTR_ERR(regs);
4118 goto reg_err;
4119 }
4120 oa_config->b_counter_regs = regs;
4121
4122 if (GRAPHICS_VER(perf->i915) < 8) {
4123 if (args->n_flex_regs != 0) {
4124 err = -EINVAL;
4125 goto reg_err;
4126 }
4127 } else {
4128 oa_config->flex_regs_len = args->n_flex_regs;
4129 regs = alloc_oa_regs(perf,
4130 perf->ops.is_valid_flex_reg,
4131 u64_to_user_ptr(args->flex_regs_ptr),
4132 args->n_flex_regs);
4133
4134 if (IS_ERR(regs)) {
4135 DRM_DEBUG("Failed to create OA config for flex_regs\n");
4136 err = PTR_ERR(regs);
4137 goto reg_err;
4138 }
4139 oa_config->flex_regs = regs;
4140 }
4141
4142 err = mutex_lock_interruptible(&perf->metrics_lock);
4143 if (err)
4144 goto reg_err;
4145
4146 /* We shouldn't have too many configs, so this iteration shouldn't be
4147 * too costly.
4148 */
4149 idr_for_each_entry(&perf->metrics_idr, tmp, id) {
4150 if (!strcmp(tmp->uuid, oa_config->uuid)) {
4151 DRM_DEBUG("OA config already exists with this uuid\n");
4152 err = -EADDRINUSE;
4153 goto sysfs_err;
4154 }
4155 }
4156
4157 err = create_dynamic_oa_sysfs_entry(perf, oa_config);
4158 if (err) {
4159 DRM_DEBUG("Failed to create sysfs entry for OA config\n");
4160 goto sysfs_err;
4161 }
4162
4163 /* Config id 0 is invalid, id 1 for kernel stored test config. */
4164 oa_config->id = idr_alloc(&perf->metrics_idr,
4165 oa_config, 2,
4166 0, GFP_KERNEL);
4167 if (oa_config->id < 0) {
4168 DRM_DEBUG("Failed to create sysfs entry for OA config\n");
4169 err = oa_config->id;
4170 goto sysfs_err;
4171 }
4172
4173 mutex_unlock(&perf->metrics_lock);
4174
4175 DRM_DEBUG("Added config %s id=%i\n", oa_config->uuid, oa_config->id);
4176
4177 return oa_config->id;
4178
4179sysfs_err:
4180 mutex_unlock(&perf->metrics_lock);
4181reg_err:
4182 i915_oa_config_put(oa_config);
4183 DRM_DEBUG("Failed to add new OA config\n");
4184 return err;
4185}
4186
4187/**
4188 * i915_perf_remove_config_ioctl - DRM ioctl() for userspace to remove an OA config
4189 * @dev: drm device
4190 * @data: ioctl data (pointer to u64 integer) copied from userspace
4191 * @file: drm file
4192 *
4193 * Configs can be removed while being used, the will stop appearing in sysfs
4194 * and their content will be freed when the stream using the config is closed.
4195 *
4196 * Returns: 0 on success or a negative error code on failure.
4197 */
4198int i915_perf_remove_config_ioctl(struct drm_device *dev, void *data,
4199 struct drm_file *file)
4200{
4201 struct i915_perf *perf = &to_i915(dev)->perf;
4202 u64 *arg = data;
4203 struct i915_oa_config *oa_config;
4204 int ret;
4205
4206 if (!perf->i915) {
4207 DRM_DEBUG("i915 perf interface not available for this system\n");
4208 return -ENOTSUPP;
4209 }
4210
4211 if (i915_perf_stream_paranoid && !perfmon_capable()) {
4212 DRM_DEBUG("Insufficient privileges to remove i915 OA config\n");
4213 return -EACCES;
4214 }
4215
4216 ret = mutex_lock_interruptible(&perf->metrics_lock);
4217 if (ret)
4218 return ret;
4219
4220 oa_config = idr_find(&perf->metrics_idr, *arg);
4221 if (!oa_config) {
4222 DRM_DEBUG("Failed to remove unknown OA config\n");
4223 ret = -ENOENT;
4224 goto err_unlock;
4225 }
4226
4227 GEM_BUG_ON(*arg != oa_config->id);
4228
4229 sysfs_remove_group(perf->metrics_kobj, &oa_config->sysfs_metric);
4230
4231 idr_remove(&perf->metrics_idr, *arg);
4232
4233 mutex_unlock(&perf->metrics_lock);
4234
4235 DRM_DEBUG("Removed config %s id=%i\n", oa_config->uuid, oa_config->id);
4236
4237 i915_oa_config_put(oa_config);
4238
4239 return 0;
4240
4241err_unlock:
4242 mutex_unlock(&perf->metrics_lock);
4243 return ret;
4244}
4245
4246static struct ctl_table oa_table[] = {
4247 {
4248 .procname = "perf_stream_paranoid",
4249 .data = &i915_perf_stream_paranoid,
4250 .maxlen = sizeof(i915_perf_stream_paranoid),
4251 .mode = 0644,
4252 .proc_handler = proc_dointvec_minmax,
4253 .extra1 = SYSCTL_ZERO,
4254 .extra2 = SYSCTL_ONE,
4255 },
4256 {
4257 .procname = "oa_max_sample_rate",
4258 .data = &i915_oa_max_sample_rate,
4259 .maxlen = sizeof(i915_oa_max_sample_rate),
4260 .mode = 0644,
4261 .proc_handler = proc_dointvec_minmax,
4262 .extra1 = SYSCTL_ZERO,
4263 .extra2 = &oa_sample_rate_hard_limit,
4264 },
4265 {}
4266};
4267
4268static struct ctl_table i915_root[] = {
4269 {
4270 .procname = "i915",
4271 .maxlen = 0,
4272 .mode = 0555,
4273 .child = oa_table,
4274 },
4275 {}
4276};
4277
4278static struct ctl_table dev_root[] = {
4279 {
4280 .procname = "dev",
4281 .maxlen = 0,
4282 .mode = 0555,
4283 .child = i915_root,
4284 },
4285 {}
4286};
4287
4288static void oa_init_supported_formats(struct i915_perf *perf)
4289{
4290 struct drm_i915_private *i915 = perf->i915;
4291 enum intel_platform platform = INTEL_INFO(i915)->platform;
4292
4293 switch (platform) {
4294 case INTEL_HASWELL:
4295 oa_format_add(perf, I915_OA_FORMAT_A13);
4296 oa_format_add(perf, I915_OA_FORMAT_A13);
4297 oa_format_add(perf, I915_OA_FORMAT_A29);
4298 oa_format_add(perf, I915_OA_FORMAT_A13_B8_C8);
4299 oa_format_add(perf, I915_OA_FORMAT_B4_C8);
4300 oa_format_add(perf, I915_OA_FORMAT_A45_B8_C8);
4301 oa_format_add(perf, I915_OA_FORMAT_B4_C8_A16);
4302 oa_format_add(perf, I915_OA_FORMAT_C4_B8);
4303 break;
4304
4305 case INTEL_BROADWELL:
4306 case INTEL_CHERRYVIEW:
4307 case INTEL_SKYLAKE:
4308 case INTEL_BROXTON:
4309 case INTEL_KABYLAKE:
4310 case INTEL_GEMINILAKE:
4311 case INTEL_COFFEELAKE:
4312 case INTEL_COMETLAKE:
4313 case INTEL_CANNONLAKE:
4314 case INTEL_ICELAKE:
4315 case INTEL_ELKHARTLAKE:
4316 case INTEL_JASPERLAKE:
4317 case INTEL_TIGERLAKE:
4318 case INTEL_ROCKETLAKE:
4319 case INTEL_DG1:
4320 case INTEL_ALDERLAKE_S:
4321 case INTEL_ALDERLAKE_P:
4322 oa_format_add(perf, I915_OA_FORMAT_A12);
4323 oa_format_add(perf, I915_OA_FORMAT_A12_B8_C8);
4324 oa_format_add(perf, I915_OA_FORMAT_A32u40_A4u32_B8_C8);
4325 oa_format_add(perf, I915_OA_FORMAT_C4_B8);
4326 break;
4327
4328 default:
4329 MISSING_CASE(platform);
4330 }
4331}
4332
4333/**
4334 * i915_perf_init - initialize i915-perf state on module bind
4335 * @i915: i915 device instance
4336 *
4337 * Initializes i915-perf state without exposing anything to userspace.
4338 *
4339 * Note: i915-perf initialization is split into an 'init' and 'register'
4340 * phase with the i915_perf_register() exposing state to userspace.
4341 */
4342void i915_perf_init(struct drm_i915_private *i915)
4343{
4344 struct i915_perf *perf = &i915->perf;
4345
4346 /* XXX const struct i915_perf_ops! */
4347
4348 perf->oa_formats = oa_formats;
4349 if (IS_HASWELL(i915)) {
4350 perf->ops.is_valid_b_counter_reg = gen7_is_valid_b_counter_addr;
4351 perf->ops.is_valid_mux_reg = hsw_is_valid_mux_addr;
4352 perf->ops.is_valid_flex_reg = NULL;
4353 perf->ops.enable_metric_set = hsw_enable_metric_set;
4354 perf->ops.disable_metric_set = hsw_disable_metric_set;
4355 perf->ops.oa_enable = gen7_oa_enable;
4356 perf->ops.oa_disable = gen7_oa_disable;
4357 perf->ops.read = gen7_oa_read;
4358 perf->ops.oa_hw_tail_read = gen7_oa_hw_tail_read;
4359 } else if (HAS_LOGICAL_RING_CONTEXTS(i915)) {
4360 /* Note: that although we could theoretically also support the
4361 * legacy ringbuffer mode on BDW (and earlier iterations of
4362 * this driver, before upstreaming did this) it didn't seem
4363 * worth the complexity to maintain now that BDW+ enable
4364 * execlist mode by default.
4365 */
4366 perf->ops.read = gen8_oa_read;
4367
4368 if (IS_GRAPHICS_VER(i915, 8, 9)) {
4369 perf->ops.is_valid_b_counter_reg =
4370 gen7_is_valid_b_counter_addr;
4371 perf->ops.is_valid_mux_reg =
4372 gen8_is_valid_mux_addr;
4373 perf->ops.is_valid_flex_reg =
4374 gen8_is_valid_flex_addr;
4375
4376 if (IS_CHERRYVIEW(i915)) {
4377 perf->ops.is_valid_mux_reg =
4378 chv_is_valid_mux_addr;
4379 }
4380
4381 perf->ops.oa_enable = gen8_oa_enable;
4382 perf->ops.oa_disable = gen8_oa_disable;
4383 perf->ops.enable_metric_set = gen8_enable_metric_set;
4384 perf->ops.disable_metric_set = gen8_disable_metric_set;
4385 perf->ops.oa_hw_tail_read = gen8_oa_hw_tail_read;
4386
4387 if (GRAPHICS_VER(i915) == 8) {
4388 perf->ctx_oactxctrl_offset = 0x120;
4389 perf->ctx_flexeu0_offset = 0x2ce;
4390
4391 perf->gen8_valid_ctx_bit = BIT(25);
4392 } else {
4393 perf->ctx_oactxctrl_offset = 0x128;
4394 perf->ctx_flexeu0_offset = 0x3de;
4395
4396 perf->gen8_valid_ctx_bit = BIT(16);
4397 }
4398 } else if (IS_GRAPHICS_VER(i915, 10, 11)) {
4399 perf->ops.is_valid_b_counter_reg =
4400 gen7_is_valid_b_counter_addr;
4401 perf->ops.is_valid_mux_reg =
4402 gen10_is_valid_mux_addr;
4403 perf->ops.is_valid_flex_reg =
4404 gen8_is_valid_flex_addr;
4405
4406 perf->ops.oa_enable = gen8_oa_enable;
4407 perf->ops.oa_disable = gen8_oa_disable;
4408 perf->ops.enable_metric_set = gen8_enable_metric_set;
4409 perf->ops.disable_metric_set = gen10_disable_metric_set;
4410 perf->ops.oa_hw_tail_read = gen8_oa_hw_tail_read;
4411
4412 if (GRAPHICS_VER(i915) == 10) {
4413 perf->ctx_oactxctrl_offset = 0x128;
4414 perf->ctx_flexeu0_offset = 0x3de;
4415 } else {
4416 perf->ctx_oactxctrl_offset = 0x124;
4417 perf->ctx_flexeu0_offset = 0x78e;
4418 }
4419 perf->gen8_valid_ctx_bit = BIT(16);
4420 } else if (GRAPHICS_VER(i915) == 12) {
4421 perf->ops.is_valid_b_counter_reg =
4422 gen12_is_valid_b_counter_addr;
4423 perf->ops.is_valid_mux_reg =
4424 gen12_is_valid_mux_addr;
4425 perf->ops.is_valid_flex_reg =
4426 gen8_is_valid_flex_addr;
4427
4428 perf->ops.oa_enable = gen12_oa_enable;
4429 perf->ops.oa_disable = gen12_oa_disable;
4430 perf->ops.enable_metric_set = gen12_enable_metric_set;
4431 perf->ops.disable_metric_set = gen12_disable_metric_set;
4432 perf->ops.oa_hw_tail_read = gen12_oa_hw_tail_read;
4433
4434 perf->ctx_flexeu0_offset = 0;
4435 perf->ctx_oactxctrl_offset = 0x144;
4436 }
4437 }
4438
4439 if (perf->ops.enable_metric_set) {
4440 mutex_init(&perf->lock);
4441
4442 /* Choose a representative limit */
4443 oa_sample_rate_hard_limit = i915->gt.clock_frequency / 2;
4444
4445 mutex_init(&perf->metrics_lock);
4446 idr_init_base(&perf->metrics_idr, 1);
4447
4448 /* We set up some ratelimit state to potentially throttle any
4449 * _NOTES about spurious, invalid OA reports which we don't
4450 * forward to userspace.
4451 *
4452 * We print a _NOTE about any throttling when closing the
4453 * stream instead of waiting until driver _fini which no one
4454 * would ever see.
4455 *
4456 * Using the same limiting factors as printk_ratelimit()
4457 */
4458 ratelimit_state_init(&perf->spurious_report_rs, 5 * HZ, 10);
4459 /* Since we use a DRM_NOTE for spurious reports it would be
4460 * inconsistent to let __ratelimit() automatically print a
4461 * warning for throttling.
4462 */
4463 ratelimit_set_flags(&perf->spurious_report_rs,
4464 RATELIMIT_MSG_ON_RELEASE);
4465
4466 ratelimit_state_init(&perf->tail_pointer_race,
4467 5 * HZ, 10);
4468 ratelimit_set_flags(&perf->tail_pointer_race,
4469 RATELIMIT_MSG_ON_RELEASE);
4470
4471 atomic64_set(&perf->noa_programming_delay,
4472 500 * 1000 /* 500us */);
4473
4474 perf->i915 = i915;
4475
4476 oa_init_supported_formats(perf);
4477 }
4478}
4479
4480static int destroy_config(int id, void *p, void *data)
4481{
4482 i915_oa_config_put(p);
4483 return 0;
4484}
4485
4486void i915_perf_sysctl_register(void)
4487{
4488 sysctl_header = register_sysctl_table(dev_root);
4489}
4490
4491void i915_perf_sysctl_unregister(void)
4492{
4493 unregister_sysctl_table(sysctl_header);
4494}
4495
4496/**
4497 * i915_perf_fini - Counter part to i915_perf_init()
4498 * @i915: i915 device instance
4499 */
4500void i915_perf_fini(struct drm_i915_private *i915)
4501{
4502 struct i915_perf *perf = &i915->perf;
4503
4504 if (!perf->i915)
4505 return;
4506
4507 idr_for_each(&perf->metrics_idr, destroy_config, perf);
4508 idr_destroy(&perf->metrics_idr);
4509
4510 memset(&perf->ops, 0, sizeof(perf->ops));
4511 perf->i915 = NULL;
4512}
4513
4514/**
4515 * i915_perf_ioctl_version - Version of the i915-perf subsystem
4516 *
4517 * This version number is used by userspace to detect available features.
4518 */
4519int i915_perf_ioctl_version(void)
4520{
4521 /*
4522 * 1: Initial version
4523 * I915_PERF_IOCTL_ENABLE
4524 * I915_PERF_IOCTL_DISABLE
4525 *
4526 * 2: Added runtime modification of OA config.
4527 * I915_PERF_IOCTL_CONFIG
4528 *
4529 * 3: Add DRM_I915_PERF_PROP_HOLD_PREEMPTION parameter to hold
4530 * preemption on a particular context so that performance data is
4531 * accessible from a delta of MI_RPC reports without looking at the
4532 * OA buffer.
4533 *
4534 * 4: Add DRM_I915_PERF_PROP_ALLOWED_SSEU to limit what contexts can
4535 * be run for the duration of the performance recording based on
4536 * their SSEU configuration.
4537 *
4538 * 5: Add DRM_I915_PERF_PROP_POLL_OA_PERIOD parameter that controls the
4539 * interval for the hrtimer used to check for OA data.
4540 */
4541 return 5;
4542}
4543
4544#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
4545#include "selftests/i915_perf.c"
4546#endif