Commit | Line | Data |
---|---|---|
eec688e1 RB |
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 | ||
7abbd8d6 RB |
27 | |
28 | /** | |
16d98b31 | 29 | * DOC: i915 Perf Overview |
7abbd8d6 RB |
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 | * | |
16d98b31 RB |
48 | */ |
49 | ||
50 | /** | |
51 | * DOC: i915 Perf History and Comparison with Core Perf | |
7abbd8d6 RB |
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 | * | |
16d98b31 RB |
82 | * Issues hit with first prototype based on Core Perf |
83 | * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ | |
7abbd8d6 RB |
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 | * | |
16d98b31 | 142 | * - As a side note on perf's grouping feature; there was also some concern |
7abbd8d6 RB |
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 | ||
eec688e1 | 194 | #include <linux/anon_inodes.h> |
d7965152 | 195 | #include <linux/sizes.h> |
f89823c2 | 196 | #include <linux/uuid.h> |
eec688e1 RB |
197 | |
198 | #include "i915_drv.h" | |
d7965152 | 199 | #include "i915_oa_hsw.h" |
19f81df2 RB |
200 | #include "i915_oa_bdw.h" |
201 | #include "i915_oa_chv.h" | |
202 | #include "i915_oa_sklgt2.h" | |
203 | #include "i915_oa_sklgt3.h" | |
204 | #include "i915_oa_sklgt4.h" | |
205 | #include "i915_oa_bxt.h" | |
6c5c1d89 LL |
206 | #include "i915_oa_kblgt2.h" |
207 | #include "i915_oa_kblgt3.h" | |
28c7ef9e | 208 | #include "i915_oa_glk.h" |
22ea4f35 | 209 | #include "i915_oa_cflgt2.h" |
4407eaa9 | 210 | #include "i915_oa_cflgt3.h" |
95690a02 | 211 | #include "i915_oa_cnl.h" |
1de401c0 | 212 | #include "i915_oa_icl.h" |
35ab4fd2 | 213 | #include "intel_lrc_reg.h" |
d7965152 | 214 | |
fe841686 JL |
215 | /* HW requires this to be a power of two, between 128k and 16M, though driver |
216 | * is currently generally designed assuming the largest 16M size is used such | |
217 | * that the overflow cases are unlikely in normal operation. | |
218 | */ | |
219 | #define OA_BUFFER_SIZE SZ_16M | |
220 | ||
221 | #define OA_TAKEN(tail, head) ((tail - head) & (OA_BUFFER_SIZE - 1)) | |
d7965152 | 222 | |
0dd860cf RB |
223 | /** |
224 | * DOC: OA Tail Pointer Race | |
225 | * | |
226 | * There's a HW race condition between OA unit tail pointer register updates and | |
d7965152 | 227 | * writes to memory whereby the tail pointer can sometimes get ahead of what's |
0dd860cf RB |
228 | * been written out to the OA buffer so far (in terms of what's visible to the |
229 | * CPU). | |
230 | * | |
231 | * Although this can be observed explicitly while copying reports to userspace | |
232 | * by checking for a zeroed report-id field in tail reports, we want to account | |
19f81df2 | 233 | * for this earlier, as part of the oa_buffer_check to avoid lots of redundant |
0dd860cf RB |
234 | * read() attempts. |
235 | * | |
236 | * In effect we define a tail pointer for reading that lags the real tail | |
237 | * pointer by at least %OA_TAIL_MARGIN_NSEC nanoseconds, which gives enough | |
238 | * time for the corresponding reports to become visible to the CPU. | |
239 | * | |
240 | * To manage this we actually track two tail pointers: | |
241 | * 1) An 'aging' tail with an associated timestamp that is tracked until we | |
242 | * can trust the corresponding data is visible to the CPU; at which point | |
243 | * it is considered 'aged'. | |
244 | * 2) An 'aged' tail that can be used for read()ing. | |
d7965152 | 245 | * |
0dd860cf | 246 | * The two separate pointers let us decouple read()s from tail pointer aging. |
d7965152 | 247 | * |
0dd860cf | 248 | * The tail pointers are checked and updated at a limited rate within a hrtimer |
a9a08845 | 249 | * callback (the same callback that is used for delivering EPOLLIN events) |
d7965152 | 250 | * |
0dd860cf RB |
251 | * Initially the tails are marked invalid with %INVALID_TAIL_PTR which |
252 | * indicates that an updated tail pointer is needed. | |
253 | * | |
254 | * Most of the implementation details for this workaround are in | |
19f81df2 | 255 | * oa_buffer_check_unlocked() and _append_oa_reports() |
0dd860cf RB |
256 | * |
257 | * Note for posterity: previously the driver used to define an effective tail | |
258 | * pointer that lagged the real pointer by a 'tail margin' measured in bytes | |
259 | * derived from %OA_TAIL_MARGIN_NSEC and the configured sampling frequency. | |
260 | * This was flawed considering that the OA unit may also automatically generate | |
261 | * non-periodic reports (such as on context switch) or the OA unit may be | |
262 | * enabled without any periodic sampling. | |
d7965152 RB |
263 | */ |
264 | #define OA_TAIL_MARGIN_NSEC 100000ULL | |
0dd860cf | 265 | #define INVALID_TAIL_PTR 0xffffffff |
d7965152 RB |
266 | |
267 | /* frequency for checking whether the OA unit has written new reports to the | |
268 | * circular OA buffer... | |
269 | */ | |
270 | #define POLL_FREQUENCY 200 | |
271 | #define POLL_PERIOD (NSEC_PER_SEC / POLL_FREQUENCY) | |
272 | ||
ccdf6341 RB |
273 | /* for sysctl proc_dointvec_minmax of dev.i915.perf_stream_paranoid */ |
274 | static int zero; | |
275 | static int one = 1; | |
276 | static u32 i915_perf_stream_paranoid = true; | |
277 | ||
d7965152 RB |
278 | /* The maximum exponent the hardware accepts is 63 (essentially it selects one |
279 | * of the 64bit timestamp bits to trigger reports from) but there's currently | |
280 | * no known use case for sampling as infrequently as once per 47 thousand years. | |
281 | * | |
282 | * Since the timestamps included in OA reports are only 32bits it seems | |
283 | * reasonable to limit the OA exponent where it's still possible to account for | |
284 | * overflow in OA report timestamps. | |
285 | */ | |
286 | #define OA_EXPONENT_MAX 31 | |
287 | ||
288 | #define INVALID_CTX_ID 0xffffffff | |
289 | ||
19f81df2 RB |
290 | /* On Gen8+ automatically triggered OA reports include a 'reason' field... */ |
291 | #define OAREPORT_REASON_MASK 0x3f | |
292 | #define OAREPORT_REASON_SHIFT 19 | |
293 | #define OAREPORT_REASON_TIMER (1<<0) | |
294 | #define OAREPORT_REASON_CTX_SWITCH (1<<3) | |
295 | #define OAREPORT_REASON_CLK_RATIO (1<<5) | |
296 | ||
d7965152 | 297 | |
00319ba0 RB |
298 | /* For sysctl proc_dointvec_minmax of i915_oa_max_sample_rate |
299 | * | |
155e941f RB |
300 | * The highest sampling frequency we can theoretically program the OA unit |
301 | * with is always half the timestamp frequency: E.g. 6.25Mhz for Haswell. | |
302 | * | |
303 | * Initialized just before we register the sysctl parameter. | |
00319ba0 | 304 | */ |
155e941f | 305 | static int oa_sample_rate_hard_limit; |
00319ba0 RB |
306 | |
307 | /* Theoretically we can program the OA unit to sample every 160ns but don't | |
308 | * allow that by default unless root... | |
309 | * | |
310 | * The default threshold of 100000Hz is based on perf's similar | |
311 | * kernel.perf_event_max_sample_rate sysctl parameter. | |
312 | */ | |
313 | static u32 i915_oa_max_sample_rate = 100000; | |
314 | ||
d7965152 RB |
315 | /* XXX: beware if future OA HW adds new report formats that the current |
316 | * code assumes all reports have a power-of-two size and ~(size - 1) can | |
317 | * be used as a mask to align the OA tail pointer. | |
318 | */ | |
6ebb6d8e | 319 | static const struct i915_oa_format hsw_oa_formats[I915_OA_FORMAT_MAX] = { |
d7965152 RB |
320 | [I915_OA_FORMAT_A13] = { 0, 64 }, |
321 | [I915_OA_FORMAT_A29] = { 1, 128 }, | |
322 | [I915_OA_FORMAT_A13_B8_C8] = { 2, 128 }, | |
323 | /* A29_B8_C8 Disallowed as 192 bytes doesn't factor into buffer size */ | |
324 | [I915_OA_FORMAT_B4_C8] = { 4, 64 }, | |
325 | [I915_OA_FORMAT_A45_B8_C8] = { 5, 256 }, | |
326 | [I915_OA_FORMAT_B4_C8_A16] = { 6, 128 }, | |
327 | [I915_OA_FORMAT_C4_B8] = { 7, 64 }, | |
328 | }; | |
329 | ||
6ebb6d8e | 330 | static const struct i915_oa_format gen8_plus_oa_formats[I915_OA_FORMAT_MAX] = { |
19f81df2 RB |
331 | [I915_OA_FORMAT_A12] = { 0, 64 }, |
332 | [I915_OA_FORMAT_A12_B8_C8] = { 2, 128 }, | |
333 | [I915_OA_FORMAT_A32u40_A4u32_B8_C8] = { 5, 256 }, | |
334 | [I915_OA_FORMAT_C4_B8] = { 7, 64 }, | |
335 | }; | |
336 | ||
d7965152 | 337 | #define SAMPLE_OA_REPORT (1<<0) |
eec688e1 | 338 | |
16d98b31 RB |
339 | /** |
340 | * struct perf_open_properties - for validated properties given to open a stream | |
341 | * @sample_flags: `DRM_I915_PERF_PROP_SAMPLE_*` properties are tracked as flags | |
342 | * @single_context: Whether a single or all gpu contexts should be monitored | |
343 | * @ctx_handle: A gem ctx handle for use with @single_context | |
344 | * @metrics_set: An ID for an OA unit metric set advertised via sysfs | |
345 | * @oa_format: An OA unit HW report format | |
346 | * @oa_periodic: Whether to enable periodic OA unit sampling | |
347 | * @oa_period_exponent: The OA unit sampling period is derived from this | |
348 | * | |
349 | * As read_properties_unlocked() enumerates and validates the properties given | |
350 | * to open a stream of metrics the configuration is built up in the structure | |
351 | * which starts out zero initialized. | |
352 | */ | |
eec688e1 RB |
353 | struct perf_open_properties { |
354 | u32 sample_flags; | |
355 | ||
356 | u64 single_context:1; | |
357 | u64 ctx_handle; | |
d7965152 RB |
358 | |
359 | /* OA sampling state */ | |
360 | int metrics_set; | |
361 | int oa_format; | |
362 | bool oa_periodic; | |
363 | int oa_period_exponent; | |
364 | }; | |
365 | ||
f89823c2 LL |
366 | static void free_oa_config(struct drm_i915_private *dev_priv, |
367 | struct i915_oa_config *oa_config) | |
368 | { | |
369 | if (!PTR_ERR(oa_config->flex_regs)) | |
370 | kfree(oa_config->flex_regs); | |
371 | if (!PTR_ERR(oa_config->b_counter_regs)) | |
372 | kfree(oa_config->b_counter_regs); | |
373 | if (!PTR_ERR(oa_config->mux_regs)) | |
374 | kfree(oa_config->mux_regs); | |
375 | kfree(oa_config); | |
376 | } | |
377 | ||
378 | static void put_oa_config(struct drm_i915_private *dev_priv, | |
379 | struct i915_oa_config *oa_config) | |
380 | { | |
381 | if (!atomic_dec_and_test(&oa_config->ref_count)) | |
382 | return; | |
383 | ||
384 | free_oa_config(dev_priv, oa_config); | |
385 | } | |
386 | ||
387 | static int get_oa_config(struct drm_i915_private *dev_priv, | |
388 | int metrics_set, | |
389 | struct i915_oa_config **out_config) | |
390 | { | |
391 | int ret; | |
392 | ||
393 | if (metrics_set == 1) { | |
394 | *out_config = &dev_priv->perf.oa.test_config; | |
395 | atomic_inc(&dev_priv->perf.oa.test_config.ref_count); | |
396 | return 0; | |
397 | } | |
398 | ||
399 | ret = mutex_lock_interruptible(&dev_priv->perf.metrics_lock); | |
400 | if (ret) | |
401 | return ret; | |
402 | ||
403 | *out_config = idr_find(&dev_priv->perf.metrics_idr, metrics_set); | |
404 | if (!*out_config) | |
405 | ret = -EINVAL; | |
406 | else | |
407 | atomic_inc(&(*out_config)->ref_count); | |
408 | ||
409 | mutex_unlock(&dev_priv->perf.metrics_lock); | |
410 | ||
411 | return ret; | |
412 | } | |
413 | ||
19f81df2 RB |
414 | static u32 gen8_oa_hw_tail_read(struct drm_i915_private *dev_priv) |
415 | { | |
416 | return I915_READ(GEN8_OATAILPTR) & GEN8_OATAILPTR_MASK; | |
417 | } | |
418 | ||
419 | static u32 gen7_oa_hw_tail_read(struct drm_i915_private *dev_priv) | |
420 | { | |
421 | u32 oastatus1 = I915_READ(GEN7_OASTATUS1); | |
422 | ||
423 | return oastatus1 & GEN7_OASTATUS1_TAIL_MASK; | |
424 | } | |
425 | ||
0dd860cf | 426 | /** |
19f81df2 | 427 | * oa_buffer_check_unlocked - check for data and update tail ptr state |
0dd860cf | 428 | * @dev_priv: i915 device instance |
d7965152 | 429 | * |
0dd860cf RB |
430 | * This is either called via fops (for blocking reads in user ctx) or the poll |
431 | * check hrtimer (atomic ctx) to check the OA buffer tail pointer and check | |
432 | * if there is data available for userspace to read. | |
d7965152 | 433 | * |
0dd860cf RB |
434 | * This function is central to providing a workaround for the OA unit tail |
435 | * pointer having a race with respect to what data is visible to the CPU. | |
436 | * It is responsible for reading tail pointers from the hardware and giving | |
437 | * the pointers time to 'age' before they are made available for reading. | |
438 | * (See description of OA_TAIL_MARGIN_NSEC above for further details.) | |
439 | * | |
440 | * Besides returning true when there is data available to read() this function | |
441 | * also has the side effect of updating the oa_buffer.tails[], .aging_timestamp | |
442 | * and .aged_tail_idx state used for reading. | |
443 | * | |
444 | * Note: It's safe to read OA config state here unlocked, assuming that this is | |
445 | * only called while the stream is enabled, while the global OA configuration | |
446 | * can't be modified. | |
447 | * | |
448 | * Returns: %true if the OA buffer contains data, else %false | |
d7965152 | 449 | */ |
19f81df2 | 450 | static bool oa_buffer_check_unlocked(struct drm_i915_private *dev_priv) |
d7965152 RB |
451 | { |
452 | int report_size = dev_priv->perf.oa.oa_buffer.format_size; | |
0dd860cf RB |
453 | unsigned long flags; |
454 | unsigned int aged_idx; | |
0dd860cf RB |
455 | u32 head, hw_tail, aged_tail, aging_tail; |
456 | u64 now; | |
457 | ||
458 | /* We have to consider the (unlikely) possibility that read() errors | |
459 | * could result in an OA buffer reset which might reset the head, | |
460 | * tails[] and aged_tail state. | |
461 | */ | |
462 | spin_lock_irqsave(&dev_priv->perf.oa.oa_buffer.ptr_lock, flags); | |
463 | ||
464 | /* NB: The head we observe here might effectively be a little out of | |
465 | * date (between head and tails[aged_idx].offset if there is currently | |
466 | * a read() in progress. | |
467 | */ | |
468 | head = dev_priv->perf.oa.oa_buffer.head; | |
469 | ||
470 | aged_idx = dev_priv->perf.oa.oa_buffer.aged_tail_idx; | |
471 | aged_tail = dev_priv->perf.oa.oa_buffer.tails[aged_idx].offset; | |
472 | aging_tail = dev_priv->perf.oa.oa_buffer.tails[!aged_idx].offset; | |
473 | ||
19f81df2 | 474 | hw_tail = dev_priv->perf.oa.ops.oa_hw_tail_read(dev_priv); |
0dd860cf RB |
475 | |
476 | /* The tail pointer increases in 64 byte increments, | |
477 | * not in report_size steps... | |
478 | */ | |
479 | hw_tail &= ~(report_size - 1); | |
480 | ||
481 | now = ktime_get_mono_fast_ns(); | |
482 | ||
4117ebc7 RB |
483 | /* Update the aged tail |
484 | * | |
485 | * Flip the tail pointer available for read()s once the aging tail is | |
486 | * old enough to trust that the corresponding data will be visible to | |
487 | * the CPU... | |
488 | * | |
489 | * Do this before updating the aging pointer in case we may be able to | |
490 | * immediately start aging a new pointer too (if new data has become | |
491 | * available) without needing to wait for a later hrtimer callback. | |
492 | */ | |
493 | if (aging_tail != INVALID_TAIL_PTR && | |
494 | ((now - dev_priv->perf.oa.oa_buffer.aging_timestamp) > | |
495 | OA_TAIL_MARGIN_NSEC)) { | |
19f81df2 | 496 | |
4117ebc7 RB |
497 | aged_idx ^= 1; |
498 | dev_priv->perf.oa.oa_buffer.aged_tail_idx = aged_idx; | |
499 | ||
500 | aged_tail = aging_tail; | |
501 | ||
502 | /* Mark that we need a new pointer to start aging... */ | |
503 | dev_priv->perf.oa.oa_buffer.tails[!aged_idx].offset = INVALID_TAIL_PTR; | |
504 | aging_tail = INVALID_TAIL_PTR; | |
505 | } | |
506 | ||
0dd860cf RB |
507 | /* Update the aging tail |
508 | * | |
509 | * We throttle aging tail updates until we have a new tail that | |
510 | * represents >= one report more data than is already available for | |
511 | * reading. This ensures there will be enough data for a successful | |
512 | * read once this new pointer has aged and ensures we will give the new | |
513 | * pointer time to age. | |
514 | */ | |
515 | if (aging_tail == INVALID_TAIL_PTR && | |
516 | (aged_tail == INVALID_TAIL_PTR || | |
517 | OA_TAKEN(hw_tail, aged_tail) >= report_size)) { | |
518 | struct i915_vma *vma = dev_priv->perf.oa.oa_buffer.vma; | |
519 | u32 gtt_offset = i915_ggtt_offset(vma); | |
520 | ||
521 | /* Be paranoid and do a bounds check on the pointer read back | |
522 | * from hardware, just in case some spurious hardware condition | |
523 | * could put the tail out of bounds... | |
524 | */ | |
525 | if (hw_tail >= gtt_offset && | |
fe841686 | 526 | hw_tail < (gtt_offset + OA_BUFFER_SIZE)) { |
0dd860cf RB |
527 | dev_priv->perf.oa.oa_buffer.tails[!aged_idx].offset = |
528 | aging_tail = hw_tail; | |
529 | dev_priv->perf.oa.oa_buffer.aging_timestamp = now; | |
530 | } else { | |
531 | DRM_ERROR("Ignoring spurious out of range OA buffer tail pointer = %u\n", | |
532 | hw_tail); | |
533 | } | |
534 | } | |
535 | ||
0dd860cf RB |
536 | spin_unlock_irqrestore(&dev_priv->perf.oa.oa_buffer.ptr_lock, flags); |
537 | ||
538 | return aged_tail == INVALID_TAIL_PTR ? | |
539 | false : OA_TAKEN(aged_tail, head) >= report_size; | |
d7965152 RB |
540 | } |
541 | ||
542 | /** | |
16d98b31 RB |
543 | * append_oa_status - Appends a status record to a userspace read() buffer. |
544 | * @stream: An i915-perf stream opened for OA metrics | |
545 | * @buf: destination buffer given by userspace | |
546 | * @count: the number of bytes userspace wants to read | |
547 | * @offset: (inout): the current position for writing into @buf | |
548 | * @type: The kind of status to report to userspace | |
549 | * | |
550 | * Writes a status record (such as `DRM_I915_PERF_RECORD_OA_REPORT_LOST`) | |
551 | * into the userspace read() buffer. | |
552 | * | |
553 | * The @buf @offset will only be updated on success. | |
554 | * | |
555 | * Returns: 0 on success, negative error code on failure. | |
d7965152 RB |
556 | */ |
557 | static int append_oa_status(struct i915_perf_stream *stream, | |
558 | char __user *buf, | |
559 | size_t count, | |
560 | size_t *offset, | |
561 | enum drm_i915_perf_record_type type) | |
562 | { | |
563 | struct drm_i915_perf_record_header header = { type, 0, sizeof(header) }; | |
564 | ||
565 | if ((count - *offset) < header.size) | |
566 | return -ENOSPC; | |
567 | ||
568 | if (copy_to_user(buf + *offset, &header, sizeof(header))) | |
569 | return -EFAULT; | |
570 | ||
571 | (*offset) += header.size; | |
572 | ||
573 | return 0; | |
574 | } | |
575 | ||
576 | /** | |
16d98b31 RB |
577 | * append_oa_sample - Copies single OA report into userspace read() buffer. |
578 | * @stream: An i915-perf stream opened for OA metrics | |
579 | * @buf: destination buffer given by userspace | |
580 | * @count: the number of bytes userspace wants to read | |
581 | * @offset: (inout): the current position for writing into @buf | |
582 | * @report: A single OA report to (optionally) include as part of the sample | |
583 | * | |
584 | * The contents of a sample are configured through `DRM_I915_PERF_PROP_SAMPLE_*` | |
585 | * properties when opening a stream, tracked as `stream->sample_flags`. This | |
586 | * function copies the requested components of a single sample to the given | |
587 | * read() @buf. | |
588 | * | |
589 | * The @buf @offset will only be updated on success. | |
590 | * | |
591 | * Returns: 0 on success, negative error code on failure. | |
d7965152 RB |
592 | */ |
593 | static int append_oa_sample(struct i915_perf_stream *stream, | |
594 | char __user *buf, | |
595 | size_t count, | |
596 | size_t *offset, | |
597 | const u8 *report) | |
598 | { | |
599 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
600 | int report_size = dev_priv->perf.oa.oa_buffer.format_size; | |
601 | struct drm_i915_perf_record_header header; | |
602 | u32 sample_flags = stream->sample_flags; | |
603 | ||
604 | header.type = DRM_I915_PERF_RECORD_SAMPLE; | |
605 | header.pad = 0; | |
606 | header.size = stream->sample_size; | |
607 | ||
608 | if ((count - *offset) < header.size) | |
609 | return -ENOSPC; | |
610 | ||
611 | buf += *offset; | |
612 | if (copy_to_user(buf, &header, sizeof(header))) | |
613 | return -EFAULT; | |
614 | buf += sizeof(header); | |
615 | ||
616 | if (sample_flags & SAMPLE_OA_REPORT) { | |
617 | if (copy_to_user(buf, report, report_size)) | |
618 | return -EFAULT; | |
619 | } | |
620 | ||
621 | (*offset) += header.size; | |
622 | ||
623 | return 0; | |
624 | } | |
625 | ||
19f81df2 RB |
626 | /** |
627 | * Copies all buffered OA reports into userspace read() buffer. | |
628 | * @stream: An i915-perf stream opened for OA metrics | |
629 | * @buf: destination buffer given by userspace | |
630 | * @count: the number of bytes userspace wants to read | |
631 | * @offset: (inout): the current position for writing into @buf | |
632 | * | |
633 | * Notably any error condition resulting in a short read (-%ENOSPC or | |
634 | * -%EFAULT) will be returned even though one or more records may | |
635 | * have been successfully copied. In this case it's up to the caller | |
636 | * to decide if the error should be squashed before returning to | |
637 | * userspace. | |
638 | * | |
639 | * Note: reports are consumed from the head, and appended to the | |
640 | * tail, so the tail chases the head?... If you think that's mad | |
641 | * and back-to-front you're not alone, but this follows the | |
642 | * Gen PRM naming convention. | |
643 | * | |
644 | * Returns: 0 on success, negative error code on failure. | |
645 | */ | |
646 | static int gen8_append_oa_reports(struct i915_perf_stream *stream, | |
647 | char __user *buf, | |
648 | size_t count, | |
649 | size_t *offset) | |
650 | { | |
651 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
652 | int report_size = dev_priv->perf.oa.oa_buffer.format_size; | |
653 | u8 *oa_buf_base = dev_priv->perf.oa.oa_buffer.vaddr; | |
654 | u32 gtt_offset = i915_ggtt_offset(dev_priv->perf.oa.oa_buffer.vma); | |
fe841686 | 655 | u32 mask = (OA_BUFFER_SIZE - 1); |
19f81df2 RB |
656 | size_t start_offset = *offset; |
657 | unsigned long flags; | |
658 | unsigned int aged_tail_idx; | |
659 | u32 head, tail; | |
660 | u32 taken; | |
661 | int ret = 0; | |
662 | ||
663 | if (WARN_ON(!stream->enabled)) | |
664 | return -EIO; | |
665 | ||
666 | spin_lock_irqsave(&dev_priv->perf.oa.oa_buffer.ptr_lock, flags); | |
667 | ||
668 | head = dev_priv->perf.oa.oa_buffer.head; | |
669 | aged_tail_idx = dev_priv->perf.oa.oa_buffer.aged_tail_idx; | |
670 | tail = dev_priv->perf.oa.oa_buffer.tails[aged_tail_idx].offset; | |
671 | ||
672 | spin_unlock_irqrestore(&dev_priv->perf.oa.oa_buffer.ptr_lock, flags); | |
673 | ||
674 | /* | |
675 | * An invalid tail pointer here means we're still waiting for the poll | |
676 | * hrtimer callback to give us a pointer | |
677 | */ | |
678 | if (tail == INVALID_TAIL_PTR) | |
679 | return -EAGAIN; | |
680 | ||
681 | /* | |
682 | * NB: oa_buffer.head/tail include the gtt_offset which we don't want | |
683 | * while indexing relative to oa_buf_base. | |
684 | */ | |
685 | head -= gtt_offset; | |
686 | tail -= gtt_offset; | |
687 | ||
688 | /* | |
689 | * An out of bounds or misaligned head or tail pointer implies a driver | |
690 | * bug since we validate + align the tail pointers we read from the | |
691 | * hardware and we are in full control of the head pointer which should | |
692 | * only be incremented by multiples of the report size (notably also | |
693 | * all a power of two). | |
694 | */ | |
fe841686 JL |
695 | if (WARN_ONCE(head > OA_BUFFER_SIZE || head % report_size || |
696 | tail > OA_BUFFER_SIZE || tail % report_size, | |
19f81df2 RB |
697 | "Inconsistent OA buffer pointers: head = %u, tail = %u\n", |
698 | head, tail)) | |
699 | return -EIO; | |
700 | ||
701 | ||
702 | for (/* none */; | |
703 | (taken = OA_TAKEN(tail, head)); | |
704 | head = (head + report_size) & mask) { | |
705 | u8 *report = oa_buf_base + head; | |
706 | u32 *report32 = (void *)report; | |
707 | u32 ctx_id; | |
708 | u32 reason; | |
709 | ||
710 | /* | |
711 | * All the report sizes factor neatly into the buffer | |
712 | * size so we never expect to see a report split | |
713 | * between the beginning and end of the buffer. | |
714 | * | |
715 | * Given the initial alignment check a misalignment | |
716 | * here would imply a driver bug that would result | |
717 | * in an overrun. | |
718 | */ | |
fe841686 | 719 | if (WARN_ON((OA_BUFFER_SIZE - head) < report_size)) { |
19f81df2 RB |
720 | DRM_ERROR("Spurious OA head ptr: non-integral report offset\n"); |
721 | break; | |
722 | } | |
723 | ||
724 | /* | |
725 | * The reason field includes flags identifying what | |
726 | * triggered this specific report (mostly timer | |
727 | * triggered or e.g. due to a context switch). | |
728 | * | |
729 | * This field is never expected to be zero so we can | |
730 | * check that the report isn't invalid before copying | |
731 | * it to userspace... | |
732 | */ | |
733 | reason = ((report32[0] >> OAREPORT_REASON_SHIFT) & | |
734 | OAREPORT_REASON_MASK); | |
735 | if (reason == 0) { | |
736 | if (__ratelimit(&dev_priv->perf.oa.spurious_report_rs)) | |
737 | DRM_NOTE("Skipping spurious, invalid OA report\n"); | |
738 | continue; | |
739 | } | |
740 | ||
61d5676b | 741 | ctx_id = report32[2] & dev_priv->perf.oa.specific_ctx_id_mask; |
19f81df2 RB |
742 | |
743 | /* | |
744 | * Squash whatever is in the CTX_ID field if it's marked as | |
745 | * invalid to be sure we avoid false-positive, single-context | |
746 | * filtering below... | |
747 | * | |
748 | * Note: that we don't clear the valid_ctx_bit so userspace can | |
749 | * understand that the ID has been squashed by the kernel. | |
750 | */ | |
751 | if (!(report32[0] & dev_priv->perf.oa.gen8_valid_ctx_bit)) | |
752 | ctx_id = report32[2] = INVALID_CTX_ID; | |
753 | ||
754 | /* | |
755 | * NB: For Gen 8 the OA unit no longer supports clock gating | |
756 | * off for a specific context and the kernel can't securely | |
757 | * stop the counters from updating as system-wide / global | |
758 | * values. | |
759 | * | |
760 | * Automatic reports now include a context ID so reports can be | |
761 | * filtered on the cpu but it's not worth trying to | |
762 | * automatically subtract/hide counter progress for other | |
763 | * contexts while filtering since we can't stop userspace | |
764 | * issuing MI_REPORT_PERF_COUNT commands which would still | |
765 | * provide a side-band view of the real values. | |
766 | * | |
767 | * To allow userspace (such as Mesa/GL_INTEL_performance_query) | |
768 | * to normalize counters for a single filtered context then it | |
769 | * needs be forwarded bookend context-switch reports so that it | |
770 | * can track switches in between MI_REPORT_PERF_COUNT commands | |
771 | * and can itself subtract/ignore the progress of counters | |
772 | * associated with other contexts. Note that the hardware | |
773 | * automatically triggers reports when switching to a new | |
774 | * context which are tagged with the ID of the newly active | |
775 | * context. To avoid the complexity (and likely fragility) of | |
776 | * reading ahead while parsing reports to try and minimize | |
777 | * forwarding redundant context switch reports (i.e. between | |
778 | * other, unrelated contexts) we simply elect to forward them | |
779 | * all. | |
780 | * | |
781 | * We don't rely solely on the reason field to identify context | |
782 | * switches since it's not-uncommon for periodic samples to | |
783 | * identify a switch before any 'context switch' report. | |
784 | */ | |
785 | if (!dev_priv->perf.oa.exclusive_stream->ctx || | |
786 | dev_priv->perf.oa.specific_ctx_id == ctx_id || | |
787 | (dev_priv->perf.oa.oa_buffer.last_ctx_id == | |
788 | dev_priv->perf.oa.specific_ctx_id) || | |
789 | reason & OAREPORT_REASON_CTX_SWITCH) { | |
790 | ||
791 | /* | |
792 | * While filtering for a single context we avoid | |
793 | * leaking the IDs of other contexts. | |
794 | */ | |
795 | if (dev_priv->perf.oa.exclusive_stream->ctx && | |
796 | dev_priv->perf.oa.specific_ctx_id != ctx_id) { | |
797 | report32[2] = INVALID_CTX_ID; | |
798 | } | |
799 | ||
800 | ret = append_oa_sample(stream, buf, count, offset, | |
801 | report); | |
802 | if (ret) | |
803 | break; | |
804 | ||
805 | dev_priv->perf.oa.oa_buffer.last_ctx_id = ctx_id; | |
806 | } | |
807 | ||
808 | /* | |
809 | * The above reason field sanity check is based on | |
810 | * the assumption that the OA buffer is initially | |
811 | * zeroed and we reset the field after copying so the | |
812 | * check is still meaningful once old reports start | |
813 | * being overwritten. | |
814 | */ | |
815 | report32[0] = 0; | |
816 | } | |
817 | ||
818 | if (start_offset != *offset) { | |
819 | spin_lock_irqsave(&dev_priv->perf.oa.oa_buffer.ptr_lock, flags); | |
820 | ||
821 | /* | |
822 | * We removed the gtt_offset for the copy loop above, indexing | |
823 | * relative to oa_buf_base so put back here... | |
824 | */ | |
825 | head += gtt_offset; | |
826 | ||
827 | I915_WRITE(GEN8_OAHEADPTR, head & GEN8_OAHEADPTR_MASK); | |
828 | dev_priv->perf.oa.oa_buffer.head = head; | |
829 | ||
830 | spin_unlock_irqrestore(&dev_priv->perf.oa.oa_buffer.ptr_lock, flags); | |
831 | } | |
832 | ||
833 | return ret; | |
834 | } | |
835 | ||
836 | /** | |
837 | * gen8_oa_read - copy status records then buffered OA reports | |
838 | * @stream: An i915-perf stream opened for OA metrics | |
839 | * @buf: destination buffer given by userspace | |
840 | * @count: the number of bytes userspace wants to read | |
841 | * @offset: (inout): the current position for writing into @buf | |
842 | * | |
843 | * Checks OA unit status registers and if necessary appends corresponding | |
844 | * status records for userspace (such as for a buffer full condition) and then | |
845 | * initiate appending any buffered OA reports. | |
846 | * | |
847 | * Updates @offset according to the number of bytes successfully copied into | |
848 | * the userspace buffer. | |
849 | * | |
850 | * NB: some data may be successfully copied to the userspace buffer | |
851 | * even if an error is returned, and this is reflected in the | |
852 | * updated @offset. | |
853 | * | |
854 | * Returns: zero on success or a negative error code | |
855 | */ | |
856 | static int gen8_oa_read(struct i915_perf_stream *stream, | |
857 | char __user *buf, | |
858 | size_t count, | |
859 | size_t *offset) | |
860 | { | |
861 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
862 | u32 oastatus; | |
863 | int ret; | |
864 | ||
865 | if (WARN_ON(!dev_priv->perf.oa.oa_buffer.vaddr)) | |
866 | return -EIO; | |
867 | ||
868 | oastatus = I915_READ(GEN8_OASTATUS); | |
869 | ||
870 | /* | |
871 | * We treat OABUFFER_OVERFLOW as a significant error: | |
872 | * | |
873 | * Although theoretically we could handle this more gracefully | |
874 | * sometimes, some Gens don't correctly suppress certain | |
875 | * automatically triggered reports in this condition and so we | |
876 | * have to assume that old reports are now being trampled | |
877 | * over. | |
fe841686 JL |
878 | * |
879 | * Considering how we don't currently give userspace control | |
880 | * over the OA buffer size and always configure a large 16MB | |
881 | * buffer, then a buffer overflow does anyway likely indicate | |
882 | * that something has gone quite badly wrong. | |
19f81df2 RB |
883 | */ |
884 | if (oastatus & GEN8_OASTATUS_OABUFFER_OVERFLOW) { | |
885 | ret = append_oa_status(stream, buf, count, offset, | |
886 | DRM_I915_PERF_RECORD_OA_BUFFER_LOST); | |
887 | if (ret) | |
888 | return ret; | |
889 | ||
890 | DRM_DEBUG("OA buffer overflow (exponent = %d): force restart\n", | |
891 | dev_priv->perf.oa.period_exponent); | |
892 | ||
5728de2f LL |
893 | dev_priv->perf.oa.ops.oa_disable(stream); |
894 | dev_priv->perf.oa.ops.oa_enable(stream); | |
19f81df2 RB |
895 | |
896 | /* | |
897 | * Note: .oa_enable() is expected to re-init the oabuffer and | |
898 | * reset GEN8_OASTATUS for us | |
899 | */ | |
900 | oastatus = I915_READ(GEN8_OASTATUS); | |
901 | } | |
902 | ||
903 | if (oastatus & GEN8_OASTATUS_REPORT_LOST) { | |
904 | ret = append_oa_status(stream, buf, count, offset, | |
905 | DRM_I915_PERF_RECORD_OA_REPORT_LOST); | |
906 | if (ret) | |
907 | return ret; | |
908 | I915_WRITE(GEN8_OASTATUS, | |
909 | oastatus & ~GEN8_OASTATUS_REPORT_LOST); | |
910 | } | |
911 | ||
912 | return gen8_append_oa_reports(stream, buf, count, offset); | |
913 | } | |
914 | ||
d7965152 RB |
915 | /** |
916 | * Copies all buffered OA reports into userspace read() buffer. | |
917 | * @stream: An i915-perf stream opened for OA metrics | |
918 | * @buf: destination buffer given by userspace | |
919 | * @count: the number of bytes userspace wants to read | |
920 | * @offset: (inout): the current position for writing into @buf | |
d7965152 | 921 | * |
16d98b31 RB |
922 | * Notably any error condition resulting in a short read (-%ENOSPC or |
923 | * -%EFAULT) will be returned even though one or more records may | |
d7965152 RB |
924 | * have been successfully copied. In this case it's up to the caller |
925 | * to decide if the error should be squashed before returning to | |
926 | * userspace. | |
927 | * | |
928 | * Note: reports are consumed from the head, and appended to the | |
e81b3a55 | 929 | * tail, so the tail chases the head?... If you think that's mad |
d7965152 RB |
930 | * and back-to-front you're not alone, but this follows the |
931 | * Gen PRM naming convention. | |
16d98b31 RB |
932 | * |
933 | * Returns: 0 on success, negative error code on failure. | |
d7965152 RB |
934 | */ |
935 | static int gen7_append_oa_reports(struct i915_perf_stream *stream, | |
936 | char __user *buf, | |
937 | size_t count, | |
3bb335c1 | 938 | size_t *offset) |
d7965152 RB |
939 | { |
940 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
941 | int report_size = dev_priv->perf.oa.oa_buffer.format_size; | |
942 | u8 *oa_buf_base = dev_priv->perf.oa.oa_buffer.vaddr; | |
d7965152 | 943 | u32 gtt_offset = i915_ggtt_offset(dev_priv->perf.oa.oa_buffer.vma); |
fe841686 | 944 | u32 mask = (OA_BUFFER_SIZE - 1); |
3bb335c1 | 945 | size_t start_offset = *offset; |
0dd860cf RB |
946 | unsigned long flags; |
947 | unsigned int aged_tail_idx; | |
948 | u32 head, tail; | |
d7965152 RB |
949 | u32 taken; |
950 | int ret = 0; | |
951 | ||
952 | if (WARN_ON(!stream->enabled)) | |
953 | return -EIO; | |
954 | ||
0dd860cf | 955 | spin_lock_irqsave(&dev_priv->perf.oa.oa_buffer.ptr_lock, flags); |
f279020a | 956 | |
0dd860cf RB |
957 | head = dev_priv->perf.oa.oa_buffer.head; |
958 | aged_tail_idx = dev_priv->perf.oa.oa_buffer.aged_tail_idx; | |
959 | tail = dev_priv->perf.oa.oa_buffer.tails[aged_tail_idx].offset; | |
f279020a | 960 | |
0dd860cf | 961 | spin_unlock_irqrestore(&dev_priv->perf.oa.oa_buffer.ptr_lock, flags); |
d7965152 | 962 | |
0dd860cf RB |
963 | /* An invalid tail pointer here means we're still waiting for the poll |
964 | * hrtimer callback to give us a pointer | |
d7965152 | 965 | */ |
0dd860cf RB |
966 | if (tail == INVALID_TAIL_PTR) |
967 | return -EAGAIN; | |
d7965152 | 968 | |
0dd860cf RB |
969 | /* NB: oa_buffer.head/tail include the gtt_offset which we don't want |
970 | * while indexing relative to oa_buf_base. | |
d7965152 | 971 | */ |
0dd860cf RB |
972 | head -= gtt_offset; |
973 | tail -= gtt_offset; | |
d7965152 | 974 | |
0dd860cf RB |
975 | /* An out of bounds or misaligned head or tail pointer implies a driver |
976 | * bug since we validate + align the tail pointers we read from the | |
977 | * hardware and we are in full control of the head pointer which should | |
978 | * only be incremented by multiples of the report size (notably also | |
979 | * all a power of two). | |
d7965152 | 980 | */ |
fe841686 JL |
981 | if (WARN_ONCE(head > OA_BUFFER_SIZE || head % report_size || |
982 | tail > OA_BUFFER_SIZE || tail % report_size, | |
0dd860cf RB |
983 | "Inconsistent OA buffer pointers: head = %u, tail = %u\n", |
984 | head, tail)) | |
985 | return -EIO; | |
d7965152 | 986 | |
d7965152 RB |
987 | |
988 | for (/* none */; | |
989 | (taken = OA_TAKEN(tail, head)); | |
990 | head = (head + report_size) & mask) { | |
991 | u8 *report = oa_buf_base + head; | |
992 | u32 *report32 = (void *)report; | |
993 | ||
994 | /* All the report sizes factor neatly into the buffer | |
995 | * size so we never expect to see a report split | |
996 | * between the beginning and end of the buffer. | |
997 | * | |
998 | * Given the initial alignment check a misalignment | |
999 | * here would imply a driver bug that would result | |
1000 | * in an overrun. | |
1001 | */ | |
fe841686 | 1002 | if (WARN_ON((OA_BUFFER_SIZE - head) < report_size)) { |
d7965152 RB |
1003 | DRM_ERROR("Spurious OA head ptr: non-integral report offset\n"); |
1004 | break; | |
1005 | } | |
1006 | ||
1007 | /* The report-ID field for periodic samples includes | |
1008 | * some undocumented flags related to what triggered | |
1009 | * the report and is never expected to be zero so we | |
1010 | * can check that the report isn't invalid before | |
1011 | * copying it to userspace... | |
1012 | */ | |
1013 | if (report32[0] == 0) { | |
712122ea RB |
1014 | if (__ratelimit(&dev_priv->perf.oa.spurious_report_rs)) |
1015 | DRM_NOTE("Skipping spurious, invalid OA report\n"); | |
d7965152 RB |
1016 | continue; |
1017 | } | |
1018 | ||
1019 | ret = append_oa_sample(stream, buf, count, offset, report); | |
1020 | if (ret) | |
1021 | break; | |
1022 | ||
1023 | /* The above report-id field sanity check is based on | |
1024 | * the assumption that the OA buffer is initially | |
1025 | * zeroed and we reset the field after copying so the | |
1026 | * check is still meaningful once old reports start | |
1027 | * being overwritten. | |
1028 | */ | |
1029 | report32[0] = 0; | |
1030 | } | |
1031 | ||
3bb335c1 | 1032 | if (start_offset != *offset) { |
0dd860cf RB |
1033 | spin_lock_irqsave(&dev_priv->perf.oa.oa_buffer.ptr_lock, flags); |
1034 | ||
3bb335c1 RB |
1035 | /* We removed the gtt_offset for the copy loop above, indexing |
1036 | * relative to oa_buf_base so put back here... | |
1037 | */ | |
1038 | head += gtt_offset; | |
1039 | ||
1040 | I915_WRITE(GEN7_OASTATUS2, | |
1041 | ((head & GEN7_OASTATUS2_HEAD_MASK) | | |
b82ed43d | 1042 | GEN7_OASTATUS2_MEM_SELECT_GGTT)); |
3bb335c1 | 1043 | dev_priv->perf.oa.oa_buffer.head = head; |
0dd860cf RB |
1044 | |
1045 | spin_unlock_irqrestore(&dev_priv->perf.oa.oa_buffer.ptr_lock, flags); | |
3bb335c1 | 1046 | } |
d7965152 RB |
1047 | |
1048 | return ret; | |
1049 | } | |
1050 | ||
16d98b31 RB |
1051 | /** |
1052 | * gen7_oa_read - copy status records then buffered OA reports | |
1053 | * @stream: An i915-perf stream opened for OA metrics | |
1054 | * @buf: destination buffer given by userspace | |
1055 | * @count: the number of bytes userspace wants to read | |
1056 | * @offset: (inout): the current position for writing into @buf | |
1057 | * | |
1058 | * Checks Gen 7 specific OA unit status registers and if necessary appends | |
1059 | * corresponding status records for userspace (such as for a buffer full | |
1060 | * condition) and then initiate appending any buffered OA reports. | |
1061 | * | |
1062 | * Updates @offset according to the number of bytes successfully copied into | |
1063 | * the userspace buffer. | |
1064 | * | |
1065 | * Returns: zero on success or a negative error code | |
1066 | */ | |
d7965152 RB |
1067 | static int gen7_oa_read(struct i915_perf_stream *stream, |
1068 | char __user *buf, | |
1069 | size_t count, | |
1070 | size_t *offset) | |
1071 | { | |
1072 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
d7965152 | 1073 | u32 oastatus1; |
d7965152 RB |
1074 | int ret; |
1075 | ||
1076 | if (WARN_ON(!dev_priv->perf.oa.oa_buffer.vaddr)) | |
1077 | return -EIO; | |
1078 | ||
d7965152 RB |
1079 | oastatus1 = I915_READ(GEN7_OASTATUS1); |
1080 | ||
d7965152 RB |
1081 | /* XXX: On Haswell we don't have a safe way to clear oastatus1 |
1082 | * bits while the OA unit is enabled (while the tail pointer | |
1083 | * may be updated asynchronously) so we ignore status bits | |
1084 | * that have already been reported to userspace. | |
1085 | */ | |
1086 | oastatus1 &= ~dev_priv->perf.oa.gen7_latched_oastatus1; | |
1087 | ||
1088 | /* We treat OABUFFER_OVERFLOW as a significant error: | |
1089 | * | |
1090 | * - The status can be interpreted to mean that the buffer is | |
1091 | * currently full (with a higher precedence than OA_TAKEN() | |
1092 | * which will start to report a near-empty buffer after an | |
1093 | * overflow) but it's awkward that we can't clear the status | |
1094 | * on Haswell, so without a reset we won't be able to catch | |
1095 | * the state again. | |
1096 | * | |
1097 | * - Since it also implies the HW has started overwriting old | |
1098 | * reports it may also affect our sanity checks for invalid | |
1099 | * reports when copying to userspace that assume new reports | |
1100 | * are being written to cleared memory. | |
1101 | * | |
1102 | * - In the future we may want to introduce a flight recorder | |
1103 | * mode where the driver will automatically maintain a safe | |
1104 | * guard band between head/tail, avoiding this overflow | |
1105 | * condition, but we avoid the added driver complexity for | |
1106 | * now. | |
1107 | */ | |
1108 | if (unlikely(oastatus1 & GEN7_OASTATUS1_OABUFFER_OVERFLOW)) { | |
1109 | ret = append_oa_status(stream, buf, count, offset, | |
1110 | DRM_I915_PERF_RECORD_OA_BUFFER_LOST); | |
1111 | if (ret) | |
1112 | return ret; | |
1113 | ||
19f81df2 RB |
1114 | DRM_DEBUG("OA buffer overflow (exponent = %d): force restart\n", |
1115 | dev_priv->perf.oa.period_exponent); | |
d7965152 | 1116 | |
5728de2f LL |
1117 | dev_priv->perf.oa.ops.oa_disable(stream); |
1118 | dev_priv->perf.oa.ops.oa_enable(stream); | |
d7965152 | 1119 | |
d7965152 | 1120 | oastatus1 = I915_READ(GEN7_OASTATUS1); |
d7965152 RB |
1121 | } |
1122 | ||
1123 | if (unlikely(oastatus1 & GEN7_OASTATUS1_REPORT_LOST)) { | |
1124 | ret = append_oa_status(stream, buf, count, offset, | |
1125 | DRM_I915_PERF_RECORD_OA_REPORT_LOST); | |
1126 | if (ret) | |
1127 | return ret; | |
1128 | dev_priv->perf.oa.gen7_latched_oastatus1 |= | |
1129 | GEN7_OASTATUS1_REPORT_LOST; | |
1130 | } | |
1131 | ||
3bb335c1 | 1132 | return gen7_append_oa_reports(stream, buf, count, offset); |
d7965152 RB |
1133 | } |
1134 | ||
16d98b31 RB |
1135 | /** |
1136 | * i915_oa_wait_unlocked - handles blocking IO until OA data available | |
1137 | * @stream: An i915-perf stream opened for OA metrics | |
1138 | * | |
1139 | * Called when userspace tries to read() from a blocking stream FD opened | |
1140 | * for OA metrics. It waits until the hrtimer callback finds a non-empty | |
1141 | * OA buffer and wakes us. | |
1142 | * | |
1143 | * Note: it's acceptable to have this return with some false positives | |
1144 | * since any subsequent read handling will return -EAGAIN if there isn't | |
1145 | * really data ready for userspace yet. | |
1146 | * | |
1147 | * Returns: zero on success or a negative error code | |
1148 | */ | |
d7965152 RB |
1149 | static int i915_oa_wait_unlocked(struct i915_perf_stream *stream) |
1150 | { | |
1151 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
1152 | ||
1153 | /* We would wait indefinitely if periodic sampling is not enabled */ | |
1154 | if (!dev_priv->perf.oa.periodic) | |
1155 | return -EIO; | |
1156 | ||
d7965152 | 1157 | return wait_event_interruptible(dev_priv->perf.oa.poll_wq, |
19f81df2 | 1158 | oa_buffer_check_unlocked(dev_priv)); |
d7965152 RB |
1159 | } |
1160 | ||
16d98b31 RB |
1161 | /** |
1162 | * i915_oa_poll_wait - call poll_wait() for an OA stream poll() | |
1163 | * @stream: An i915-perf stream opened for OA metrics | |
1164 | * @file: An i915 perf stream file | |
1165 | * @wait: poll() state table | |
1166 | * | |
1167 | * For handling userspace polling on an i915 perf stream opened for OA metrics, | |
1168 | * this starts a poll_wait with the wait queue that our hrtimer callback wakes | |
1169 | * when it sees data ready to read in the circular OA buffer. | |
1170 | */ | |
d7965152 RB |
1171 | static void i915_oa_poll_wait(struct i915_perf_stream *stream, |
1172 | struct file *file, | |
1173 | poll_table *wait) | |
1174 | { | |
1175 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
1176 | ||
1177 | poll_wait(file, &dev_priv->perf.oa.poll_wq, wait); | |
1178 | } | |
1179 | ||
16d98b31 RB |
1180 | /** |
1181 | * i915_oa_read - just calls through to &i915_oa_ops->read | |
1182 | * @stream: An i915-perf stream opened for OA metrics | |
1183 | * @buf: destination buffer given by userspace | |
1184 | * @count: the number of bytes userspace wants to read | |
1185 | * @offset: (inout): the current position for writing into @buf | |
1186 | * | |
1187 | * Updates @offset according to the number of bytes successfully copied into | |
1188 | * the userspace buffer. | |
1189 | * | |
1190 | * Returns: zero on success or a negative error code | |
1191 | */ | |
d7965152 RB |
1192 | static int i915_oa_read(struct i915_perf_stream *stream, |
1193 | char __user *buf, | |
1194 | size_t count, | |
1195 | size_t *offset) | |
1196 | { | |
1197 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
1198 | ||
1199 | return dev_priv->perf.oa.ops.read(stream, buf, count, offset); | |
1200 | } | |
1201 | ||
61d5676b LL |
1202 | static struct intel_context *oa_pin_context(struct drm_i915_private *i915, |
1203 | struct i915_gem_context *ctx) | |
1204 | { | |
8a68d464 | 1205 | struct intel_engine_cs *engine = i915->engine[RCS0]; |
61d5676b LL |
1206 | struct intel_context *ce; |
1207 | int ret; | |
1208 | ||
1209 | ret = i915_mutex_lock_interruptible(&i915->drm); | |
1210 | if (ret) | |
1211 | return ERR_PTR(ret); | |
1212 | ||
1213 | /* | |
1214 | * As the ID is the gtt offset of the context's vma we | |
1215 | * pin the vma to ensure the ID remains fixed. | |
1216 | * | |
1217 | * NB: implied RCS engine... | |
1218 | */ | |
1219 | ce = intel_context_pin(ctx, engine); | |
1220 | mutex_unlock(&i915->drm.struct_mutex); | |
1221 | if (IS_ERR(ce)) | |
1222 | return ce; | |
1223 | ||
1224 | i915->perf.oa.pinned_ctx = ce; | |
1225 | ||
1226 | return ce; | |
1227 | } | |
1228 | ||
16d98b31 RB |
1229 | /** |
1230 | * oa_get_render_ctx_id - determine and hold ctx hw id | |
1231 | * @stream: An i915-perf stream opened for OA metrics | |
1232 | * | |
1233 | * Determine the render context hw id, and ensure it remains fixed for the | |
d7965152 RB |
1234 | * lifetime of the stream. This ensures that we don't have to worry about |
1235 | * updating the context ID in OACONTROL on the fly. | |
16d98b31 RB |
1236 | * |
1237 | * Returns: zero on success or a negative error code | |
d7965152 RB |
1238 | */ |
1239 | static int oa_get_render_ctx_id(struct i915_perf_stream *stream) | |
1240 | { | |
61d5676b LL |
1241 | struct drm_i915_private *i915 = stream->dev_priv; |
1242 | struct intel_context *ce; | |
d7965152 | 1243 | |
61d5676b LL |
1244 | ce = oa_pin_context(i915, stream->ctx); |
1245 | if (IS_ERR(ce)) | |
1246 | return PTR_ERR(ce); | |
19f81df2 | 1247 | |
61d5676b LL |
1248 | switch (INTEL_GEN(i915)) { |
1249 | case 7: { | |
19f81df2 | 1250 | /* |
61d5676b LL |
1251 | * On Haswell we don't do any post processing of the reports |
1252 | * and don't need to use the mask. | |
19f81df2 | 1253 | */ |
61d5676b LL |
1254 | i915->perf.oa.specific_ctx_id = i915_ggtt_offset(ce->state); |
1255 | i915->perf.oa.specific_ctx_id_mask = 0; | |
1256 | break; | |
1257 | } | |
d7965152 | 1258 | |
61d5676b LL |
1259 | case 8: |
1260 | case 9: | |
1261 | case 10: | |
1262 | if (USES_GUC_SUBMISSION(i915)) { | |
1263 | /* | |
1264 | * When using GuC, the context descriptor we write in | |
1265 | * i915 is read by GuC and rewritten before it's | |
1266 | * actually written into the hardware. The LRCA is | |
1267 | * what is put into the context id field of the | |
1268 | * context descriptor by GuC. Because it's aligned to | |
1269 | * a page, the lower 12bits are always at 0 and | |
1270 | * dropped by GuC. They won't be part of the context | |
1271 | * ID in the OA reports, so squash those lower bits. | |
1272 | */ | |
1273 | i915->perf.oa.specific_ctx_id = | |
1274 | lower_32_bits(ce->lrc_desc) >> 12; | |
19f81df2 | 1275 | |
61d5676b LL |
1276 | /* |
1277 | * GuC uses the top bit to signal proxy submission, so | |
1278 | * ignore that bit. | |
1279 | */ | |
1280 | i915->perf.oa.specific_ctx_id_mask = | |
1281 | (1U << (GEN8_CTX_ID_WIDTH - 1)) - 1; | |
1282 | } else { | |
61d5676b LL |
1283 | i915->perf.oa.specific_ctx_id_mask = |
1284 | (1U << GEN8_CTX_ID_WIDTH) - 1; | |
9904b156 MT |
1285 | i915->perf.oa.specific_ctx_id = |
1286 | upper_32_bits(ce->lrc_desc); | |
1287 | i915->perf.oa.specific_ctx_id &= | |
1288 | i915->perf.oa.specific_ctx_id_mask; | |
61d5676b LL |
1289 | } |
1290 | break; | |
1291 | ||
1292 | case 11: { | |
61d5676b LL |
1293 | i915->perf.oa.specific_ctx_id_mask = |
1294 | ((1U << GEN11_SW_CTX_ID_WIDTH) - 1) << (GEN11_SW_CTX_ID_SHIFT - 32) | | |
1295 | ((1U << GEN11_ENGINE_INSTANCE_WIDTH) - 1) << (GEN11_ENGINE_INSTANCE_SHIFT - 32) | | |
1296 | ((1 << GEN11_ENGINE_CLASS_WIDTH) - 1) << (GEN11_ENGINE_CLASS_SHIFT - 32); | |
2b9a8203 MT |
1297 | i915->perf.oa.specific_ctx_id = upper_32_bits(ce->lrc_desc); |
1298 | i915->perf.oa.specific_ctx_id &= | |
1299 | i915->perf.oa.specific_ctx_id_mask; | |
61d5676b LL |
1300 | break; |
1301 | } | |
1302 | ||
1303 | default: | |
1304 | MISSING_CASE(INTEL_GEN(i915)); | |
19f81df2 | 1305 | } |
d7965152 | 1306 | |
61d5676b LL |
1307 | DRM_DEBUG_DRIVER("filtering on ctx_id=0x%x ctx_id_mask=0x%x\n", |
1308 | i915->perf.oa.specific_ctx_id, | |
1309 | i915->perf.oa.specific_ctx_id_mask); | |
1310 | ||
266a240b | 1311 | return 0; |
d7965152 RB |
1312 | } |
1313 | ||
16d98b31 RB |
1314 | /** |
1315 | * oa_put_render_ctx_id - counterpart to oa_get_render_ctx_id releases hold | |
1316 | * @stream: An i915-perf stream opened for OA metrics | |
1317 | * | |
1318 | * In case anything needed doing to ensure the context HW ID would remain valid | |
1319 | * for the lifetime of the stream, then that can be undone here. | |
1320 | */ | |
d7965152 RB |
1321 | static void oa_put_render_ctx_id(struct i915_perf_stream *stream) |
1322 | { | |
1323 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
1fc44d9b | 1324 | struct intel_context *ce; |
d7965152 | 1325 | |
1fc44d9b | 1326 | dev_priv->perf.oa.specific_ctx_id = INVALID_CTX_ID; |
61d5676b | 1327 | dev_priv->perf.oa.specific_ctx_id_mask = 0; |
d7965152 | 1328 | |
1fc44d9b CW |
1329 | ce = fetch_and_zero(&dev_priv->perf.oa.pinned_ctx); |
1330 | if (ce) { | |
19f81df2 | 1331 | mutex_lock(&dev_priv->drm.struct_mutex); |
1fc44d9b | 1332 | intel_context_unpin(ce); |
19f81df2 RB |
1333 | mutex_unlock(&dev_priv->drm.struct_mutex); |
1334 | } | |
d7965152 RB |
1335 | } |
1336 | ||
1337 | static void | |
1338 | free_oa_buffer(struct drm_i915_private *i915) | |
1339 | { | |
1340 | mutex_lock(&i915->drm.struct_mutex); | |
1341 | ||
6a2f59e4 CW |
1342 | i915_vma_unpin_and_release(&i915->perf.oa.oa_buffer.vma, |
1343 | I915_VMA_RELEASE_MAP); | |
d7965152 RB |
1344 | |
1345 | mutex_unlock(&i915->drm.struct_mutex); | |
6a2f59e4 CW |
1346 | |
1347 | i915->perf.oa.oa_buffer.vaddr = NULL; | |
d7965152 RB |
1348 | } |
1349 | ||
1350 | static void i915_oa_stream_destroy(struct i915_perf_stream *stream) | |
1351 | { | |
1352 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
1353 | ||
1354 | BUG_ON(stream != dev_priv->perf.oa.exclusive_stream); | |
1355 | ||
19f81df2 | 1356 | /* |
f89823c2 LL |
1357 | * Unset exclusive_stream first, it will be checked while disabling |
1358 | * the metric set on gen8+. | |
19f81df2 | 1359 | */ |
701f8231 | 1360 | mutex_lock(&dev_priv->drm.struct_mutex); |
19f81df2 | 1361 | dev_priv->perf.oa.exclusive_stream = NULL; |
d7965152 | 1362 | dev_priv->perf.oa.ops.disable_metric_set(dev_priv); |
41d3fdcd | 1363 | mutex_unlock(&dev_priv->drm.struct_mutex); |
d7965152 RB |
1364 | |
1365 | free_oa_buffer(dev_priv); | |
1366 | ||
3ceea6a1 | 1367 | intel_uncore_forcewake_put(&dev_priv->uncore, FORCEWAKE_ALL); |
6619c007 | 1368 | intel_runtime_pm_put(dev_priv, stream->wakeref); |
d7965152 RB |
1369 | |
1370 | if (stream->ctx) | |
1371 | oa_put_render_ctx_id(stream); | |
1372 | ||
f89823c2 LL |
1373 | put_oa_config(dev_priv, stream->oa_config); |
1374 | ||
712122ea RB |
1375 | if (dev_priv->perf.oa.spurious_report_rs.missed) { |
1376 | DRM_NOTE("%d spurious OA report notices suppressed due to ratelimiting\n", | |
1377 | dev_priv->perf.oa.spurious_report_rs.missed); | |
1378 | } | |
d7965152 RB |
1379 | } |
1380 | ||
1381 | static void gen7_init_oa_buffer(struct drm_i915_private *dev_priv) | |
1382 | { | |
1383 | u32 gtt_offset = i915_ggtt_offset(dev_priv->perf.oa.oa_buffer.vma); | |
0dd860cf RB |
1384 | unsigned long flags; |
1385 | ||
1386 | spin_lock_irqsave(&dev_priv->perf.oa.oa_buffer.ptr_lock, flags); | |
d7965152 RB |
1387 | |
1388 | /* Pre-DevBDW: OABUFFER must be set with counters off, | |
1389 | * before OASTATUS1, but after OASTATUS2 | |
1390 | */ | |
b82ed43d LL |
1391 | I915_WRITE(GEN7_OASTATUS2, |
1392 | gtt_offset | GEN7_OASTATUS2_MEM_SELECT_GGTT); /* head */ | |
f279020a RB |
1393 | dev_priv->perf.oa.oa_buffer.head = gtt_offset; |
1394 | ||
d7965152 | 1395 | I915_WRITE(GEN7_OABUFFER, gtt_offset); |
f279020a | 1396 | |
fe841686 | 1397 | I915_WRITE(GEN7_OASTATUS1, gtt_offset | OABUFFER_SIZE_16M); /* tail */ |
d7965152 | 1398 | |
0dd860cf RB |
1399 | /* Mark that we need updated tail pointers to read from... */ |
1400 | dev_priv->perf.oa.oa_buffer.tails[0].offset = INVALID_TAIL_PTR; | |
1401 | dev_priv->perf.oa.oa_buffer.tails[1].offset = INVALID_TAIL_PTR; | |
1402 | ||
1403 | spin_unlock_irqrestore(&dev_priv->perf.oa.oa_buffer.ptr_lock, flags); | |
1404 | ||
d7965152 RB |
1405 | /* On Haswell we have to track which OASTATUS1 flags we've |
1406 | * already seen since they can't be cleared while periodic | |
1407 | * sampling is enabled. | |
1408 | */ | |
1409 | dev_priv->perf.oa.gen7_latched_oastatus1 = 0; | |
1410 | ||
1411 | /* NB: although the OA buffer will initially be allocated | |
1412 | * zeroed via shmfs (and so this memset is redundant when | |
1413 | * first allocating), we may re-init the OA buffer, either | |
1414 | * when re-enabling a stream or in error/reset paths. | |
1415 | * | |
1416 | * The reason we clear the buffer for each re-init is for the | |
1417 | * sanity check in gen7_append_oa_reports() that looks at the | |
1418 | * report-id field to make sure it's non-zero which relies on | |
1419 | * the assumption that new reports are being written to zeroed | |
1420 | * memory... | |
1421 | */ | |
fe841686 | 1422 | memset(dev_priv->perf.oa.oa_buffer.vaddr, 0, OA_BUFFER_SIZE); |
d7965152 RB |
1423 | |
1424 | /* Maybe make ->pollin per-stream state if we support multiple | |
1425 | * concurrent streams in the future. | |
1426 | */ | |
1427 | dev_priv->perf.oa.pollin = false; | |
1428 | } | |
1429 | ||
19f81df2 RB |
1430 | static void gen8_init_oa_buffer(struct drm_i915_private *dev_priv) |
1431 | { | |
1432 | u32 gtt_offset = i915_ggtt_offset(dev_priv->perf.oa.oa_buffer.vma); | |
1433 | unsigned long flags; | |
1434 | ||
1435 | spin_lock_irqsave(&dev_priv->perf.oa.oa_buffer.ptr_lock, flags); | |
1436 | ||
1437 | I915_WRITE(GEN8_OASTATUS, 0); | |
1438 | I915_WRITE(GEN8_OAHEADPTR, gtt_offset); | |
1439 | dev_priv->perf.oa.oa_buffer.head = gtt_offset; | |
1440 | ||
1441 | I915_WRITE(GEN8_OABUFFER_UDW, 0); | |
1442 | ||
1443 | /* | |
1444 | * PRM says: | |
1445 | * | |
1446 | * "This MMIO must be set before the OATAILPTR | |
1447 | * register and after the OAHEADPTR register. This is | |
1448 | * to enable proper functionality of the overflow | |
1449 | * bit." | |
1450 | */ | |
1451 | I915_WRITE(GEN8_OABUFFER, gtt_offset | | |
fe841686 | 1452 | OABUFFER_SIZE_16M | GEN8_OABUFFER_MEM_SELECT_GGTT); |
19f81df2 RB |
1453 | I915_WRITE(GEN8_OATAILPTR, gtt_offset & GEN8_OATAILPTR_MASK); |
1454 | ||
1455 | /* Mark that we need updated tail pointers to read from... */ | |
1456 | dev_priv->perf.oa.oa_buffer.tails[0].offset = INVALID_TAIL_PTR; | |
1457 | dev_priv->perf.oa.oa_buffer.tails[1].offset = INVALID_TAIL_PTR; | |
1458 | ||
1459 | /* | |
1460 | * Reset state used to recognise context switches, affecting which | |
1461 | * reports we will forward to userspace while filtering for a single | |
1462 | * context. | |
1463 | */ | |
1464 | dev_priv->perf.oa.oa_buffer.last_ctx_id = INVALID_CTX_ID; | |
1465 | ||
1466 | spin_unlock_irqrestore(&dev_priv->perf.oa.oa_buffer.ptr_lock, flags); | |
1467 | ||
1468 | /* | |
1469 | * NB: although the OA buffer will initially be allocated | |
1470 | * zeroed via shmfs (and so this memset is redundant when | |
1471 | * first allocating), we may re-init the OA buffer, either | |
1472 | * when re-enabling a stream or in error/reset paths. | |
1473 | * | |
1474 | * The reason we clear the buffer for each re-init is for the | |
1475 | * sanity check in gen8_append_oa_reports() that looks at the | |
1476 | * reason field to make sure it's non-zero which relies on | |
1477 | * the assumption that new reports are being written to zeroed | |
1478 | * memory... | |
1479 | */ | |
fe841686 | 1480 | memset(dev_priv->perf.oa.oa_buffer.vaddr, 0, OA_BUFFER_SIZE); |
19f81df2 RB |
1481 | |
1482 | /* | |
1483 | * Maybe make ->pollin per-stream state if we support multiple | |
1484 | * concurrent streams in the future. | |
1485 | */ | |
1486 | dev_priv->perf.oa.pollin = false; | |
1487 | } | |
1488 | ||
fe841686 | 1489 | static int alloc_oa_buffer(struct drm_i915_private *dev_priv) |
d7965152 RB |
1490 | { |
1491 | struct drm_i915_gem_object *bo; | |
1492 | struct i915_vma *vma; | |
1493 | int ret; | |
1494 | ||
1495 | if (WARN_ON(dev_priv->perf.oa.oa_buffer.vma)) | |
1496 | return -ENODEV; | |
1497 | ||
1498 | ret = i915_mutex_lock_interruptible(&dev_priv->drm); | |
1499 | if (ret) | |
1500 | return ret; | |
1501 | ||
fe841686 JL |
1502 | BUILD_BUG_ON_NOT_POWER_OF_2(OA_BUFFER_SIZE); |
1503 | BUILD_BUG_ON(OA_BUFFER_SIZE < SZ_128K || OA_BUFFER_SIZE > SZ_16M); | |
1504 | ||
1505 | bo = i915_gem_object_create(dev_priv, OA_BUFFER_SIZE); | |
d7965152 RB |
1506 | if (IS_ERR(bo)) { |
1507 | DRM_ERROR("Failed to allocate OA buffer\n"); | |
1508 | ret = PTR_ERR(bo); | |
1509 | goto unlock; | |
1510 | } | |
1511 | ||
1512 | ret = i915_gem_object_set_cache_level(bo, I915_CACHE_LLC); | |
1513 | if (ret) | |
1514 | goto err_unref; | |
1515 | ||
1516 | /* PreHSW required 512K alignment, HSW requires 16M */ | |
1517 | vma = i915_gem_object_ggtt_pin(bo, NULL, 0, SZ_16M, 0); | |
1518 | if (IS_ERR(vma)) { | |
1519 | ret = PTR_ERR(vma); | |
1520 | goto err_unref; | |
1521 | } | |
1522 | dev_priv->perf.oa.oa_buffer.vma = vma; | |
1523 | ||
1524 | dev_priv->perf.oa.oa_buffer.vaddr = | |
1525 | i915_gem_object_pin_map(bo, I915_MAP_WB); | |
1526 | if (IS_ERR(dev_priv->perf.oa.oa_buffer.vaddr)) { | |
1527 | ret = PTR_ERR(dev_priv->perf.oa.oa_buffer.vaddr); | |
1528 | goto err_unpin; | |
1529 | } | |
1530 | ||
fe841686 | 1531 | DRM_DEBUG_DRIVER("OA Buffer initialized, gtt offset = 0x%x, vaddr = %p\n", |
d7965152 | 1532 | i915_ggtt_offset(dev_priv->perf.oa.oa_buffer.vma), |
fe841686 | 1533 | dev_priv->perf.oa.oa_buffer.vaddr); |
d7965152 RB |
1534 | |
1535 | goto unlock; | |
1536 | ||
1537 | err_unpin: | |
1538 | __i915_vma_unpin(vma); | |
1539 | ||
1540 | err_unref: | |
1541 | i915_gem_object_put(bo); | |
1542 | ||
1543 | dev_priv->perf.oa.oa_buffer.vaddr = NULL; | |
1544 | dev_priv->perf.oa.oa_buffer.vma = NULL; | |
1545 | ||
1546 | unlock: | |
1547 | mutex_unlock(&dev_priv->drm.struct_mutex); | |
1548 | return ret; | |
1549 | } | |
1550 | ||
1551 | static void config_oa_regs(struct drm_i915_private *dev_priv, | |
1552 | const struct i915_oa_reg *regs, | |
701f8231 | 1553 | u32 n_regs) |
d7965152 | 1554 | { |
701f8231 | 1555 | u32 i; |
d7965152 RB |
1556 | |
1557 | for (i = 0; i < n_regs; i++) { | |
1558 | const struct i915_oa_reg *reg = regs + i; | |
1559 | ||
1560 | I915_WRITE(reg->addr, reg->value); | |
1561 | } | |
1562 | } | |
1563 | ||
5728de2f | 1564 | static int hsw_enable_metric_set(struct i915_perf_stream *stream) |
d7965152 | 1565 | { |
5728de2f LL |
1566 | struct drm_i915_private *dev_priv = stream->dev_priv; |
1567 | const struct i915_oa_config *oa_config = stream->oa_config; | |
1568 | ||
d7965152 RB |
1569 | /* PRM: |
1570 | * | |
1571 | * OA unit is using “crclk” for its functionality. When trunk | |
1572 | * level clock gating takes place, OA clock would be gated, | |
1573 | * unable to count the events from non-render clock domain. | |
1574 | * Render clock gating must be disabled when OA is enabled to | |
1575 | * count the events from non-render domain. Unit level clock | |
1576 | * gating for RCS should also be disabled. | |
1577 | */ | |
1578 | I915_WRITE(GEN7_MISCCPCTL, (I915_READ(GEN7_MISCCPCTL) & | |
1579 | ~GEN7_DOP_CLOCK_GATE_ENABLE)); | |
1580 | I915_WRITE(GEN6_UCGCTL1, (I915_READ(GEN6_UCGCTL1) | | |
1581 | GEN6_CSUNIT_CLOCK_GATE_DISABLE)); | |
1582 | ||
701f8231 | 1583 | config_oa_regs(dev_priv, oa_config->mux_regs, oa_config->mux_regs_len); |
d7965152 RB |
1584 | |
1585 | /* It apparently takes a fairly long time for a new MUX | |
1586 | * configuration to be be applied after these register writes. | |
1587 | * This delay duration was derived empirically based on the | |
1588 | * render_basic config but hopefully it covers the maximum | |
1589 | * configuration latency. | |
1590 | * | |
1591 | * As a fallback, the checks in _append_oa_reports() to skip | |
1592 | * invalid OA reports do also seem to work to discard reports | |
1593 | * generated before this config has completed - albeit not | |
1594 | * silently. | |
1595 | * | |
1596 | * Unfortunately this is essentially a magic number, since we | |
1597 | * don't currently know of a reliable mechanism for predicting | |
1598 | * how long the MUX config will take to apply and besides | |
1599 | * seeing invalid reports we don't know of a reliable way to | |
1600 | * explicitly check that the MUX config has landed. | |
1601 | * | |
1602 | * It's even possible we've miss characterized the underlying | |
1603 | * problem - it just seems like the simplest explanation why | |
1604 | * a delay at this location would mitigate any invalid reports. | |
1605 | */ | |
1606 | usleep_range(15000, 20000); | |
1607 | ||
701f8231 LL |
1608 | config_oa_regs(dev_priv, oa_config->b_counter_regs, |
1609 | oa_config->b_counter_regs_len); | |
d7965152 RB |
1610 | |
1611 | return 0; | |
1612 | } | |
1613 | ||
1614 | static void hsw_disable_metric_set(struct drm_i915_private *dev_priv) | |
1615 | { | |
1616 | I915_WRITE(GEN6_UCGCTL1, (I915_READ(GEN6_UCGCTL1) & | |
1617 | ~GEN6_CSUNIT_CLOCK_GATE_DISABLE)); | |
1618 | I915_WRITE(GEN7_MISCCPCTL, (I915_READ(GEN7_MISCCPCTL) | | |
1619 | GEN7_DOP_CLOCK_GATE_ENABLE)); | |
1620 | ||
1621 | I915_WRITE(GDT_CHICKEN_BITS, (I915_READ(GDT_CHICKEN_BITS) & | |
1622 | ~GT_NOA_ENABLE)); | |
1623 | } | |
1624 | ||
19f81df2 RB |
1625 | /* |
1626 | * NB: It must always remain pointer safe to run this even if the OA unit | |
1627 | * has been disabled. | |
1628 | * | |
1629 | * It's fine to put out-of-date values into these per-context registers | |
1630 | * in the case that the OA unit has been disabled. | |
1631 | */ | |
b146e5ef CW |
1632 | static void |
1633 | gen8_update_reg_state_unlocked(struct intel_context *ce, | |
1634 | u32 *reg_state, | |
1635 | const struct i915_oa_config *oa_config) | |
19f81df2 | 1636 | { |
b146e5ef CW |
1637 | struct drm_i915_private *i915 = ce->gem_context->i915; |
1638 | u32 ctx_oactxctrl = i915->perf.oa.ctx_oactxctrl_offset; | |
1639 | u32 ctx_flexeu0 = i915->perf.oa.ctx_flexeu0_offset; | |
19f81df2 | 1640 | /* The MMIO offsets for Flex EU registers aren't contiguous */ |
35ab4fd2 LL |
1641 | i915_reg_t flex_regs[] = { |
1642 | EU_PERF_CNTL0, | |
1643 | EU_PERF_CNTL1, | |
1644 | EU_PERF_CNTL2, | |
1645 | EU_PERF_CNTL3, | |
1646 | EU_PERF_CNTL4, | |
1647 | EU_PERF_CNTL5, | |
1648 | EU_PERF_CNTL6, | |
19f81df2 RB |
1649 | }; |
1650 | int i; | |
1651 | ||
35ab4fd2 | 1652 | CTX_REG(reg_state, ctx_oactxctrl, GEN8_OACTXCONTROL, |
b146e5ef CW |
1653 | (i915->perf.oa.period_exponent << GEN8_OA_TIMER_PERIOD_SHIFT) | |
1654 | (i915->perf.oa.periodic ? GEN8_OA_TIMER_ENABLE : 0) | | |
35ab4fd2 | 1655 | GEN8_OA_COUNTER_RESUME); |
19f81df2 | 1656 | |
35ab4fd2 | 1657 | for (i = 0; i < ARRAY_SIZE(flex_regs); i++) { |
19f81df2 | 1658 | u32 state_offset = ctx_flexeu0 + i * 2; |
35ab4fd2 | 1659 | u32 mmio = i915_mmio_reg_offset(flex_regs[i]); |
19f81df2 RB |
1660 | |
1661 | /* | |
1662 | * This arbitrary default will select the 'EU FPU0 Pipeline | |
1663 | * Active' event. In the future it's anticipated that there | |
1664 | * will be an explicit 'No Event' we can select, but not yet... | |
1665 | */ | |
1666 | u32 value = 0; | |
19f81df2 | 1667 | |
701f8231 LL |
1668 | if (oa_config) { |
1669 | u32 j; | |
1670 | ||
1671 | for (j = 0; j < oa_config->flex_regs_len; j++) { | |
1672 | if (i915_mmio_reg_offset(oa_config->flex_regs[j].addr) == mmio) { | |
1673 | value = oa_config->flex_regs[j].value; | |
1674 | break; | |
1675 | } | |
19f81df2 RB |
1676 | } |
1677 | } | |
1678 | ||
35ab4fd2 | 1679 | CTX_REG(reg_state, state_offset, flex_regs[i], value); |
19f81df2 | 1680 | } |
ec431eae | 1681 | |
b146e5ef CW |
1682 | CTX_REG(reg_state, |
1683 | CTX_R_PWR_CLK_STATE, GEN8_R_PWR_CLK_STATE, | |
1684 | gen8_make_rpcs(i915, &ce->sseu)); | |
19f81df2 RB |
1685 | } |
1686 | ||
19f81df2 RB |
1687 | /* |
1688 | * Manages updating the per-context aspects of the OA stream | |
1689 | * configuration across all contexts. | |
1690 | * | |
1691 | * The awkward consideration here is that OACTXCONTROL controls the | |
1692 | * exponent for periodic sampling which is primarily used for system | |
1693 | * wide profiling where we'd like a consistent sampling period even in | |
1694 | * the face of context switches. | |
1695 | * | |
1696 | * Our approach of updating the register state context (as opposed to | |
1697 | * say using a workaround batch buffer) ensures that the hardware | |
1698 | * won't automatically reload an out-of-date timer exponent even | |
1699 | * transiently before a WA BB could be parsed. | |
1700 | * | |
1701 | * This function needs to: | |
1702 | * - Ensure the currently running context's per-context OA state is | |
1703 | * updated | |
1704 | * - Ensure that all existing contexts will have the correct per-context | |
1705 | * OA state if they are scheduled for use. | |
1706 | * - Ensure any new contexts will be initialized with the correct | |
1707 | * per-context OA state. | |
1708 | * | |
1709 | * Note: it's only the RCS/Render context that has any OA state. | |
1710 | */ | |
1711 | static int gen8_configure_all_contexts(struct drm_i915_private *dev_priv, | |
41d3fdcd | 1712 | const struct i915_oa_config *oa_config) |
19f81df2 | 1713 | { |
8a68d464 | 1714 | struct intel_engine_cs *engine = dev_priv->engine[RCS0]; |
666424ab | 1715 | unsigned int map_type = i915_coherent_map_type(dev_priv); |
19f81df2 | 1716 | struct i915_gem_context *ctx; |
722f3de3 | 1717 | struct i915_request *rq; |
19f81df2 | 1718 | int ret; |
19f81df2 | 1719 | |
41d3fdcd | 1720 | lockdep_assert_held(&dev_priv->drm.struct_mutex); |
19f81df2 | 1721 | |
19f81df2 RB |
1722 | /* |
1723 | * The OA register config is setup through the context image. This image | |
1724 | * might be written to by the GPU on context switch (in particular on | |
1725 | * lite-restore). This means we can't safely update a context's image, | |
1726 | * if this context is scheduled/submitted to run on the GPU. | |
1727 | * | |
1728 | * We could emit the OA register config through the batch buffer but | |
1729 | * this might leave small interval of time where the OA unit is | |
1730 | * configured at an invalid sampling period. | |
1731 | * | |
1732 | * So far the best way to work around this issue seems to be draining | |
1733 | * the GPU from any submitted work. | |
1734 | */ | |
ec625fb9 | 1735 | ret = i915_gem_wait_for_idle(dev_priv, |
722f3de3 | 1736 | I915_WAIT_LOCKED, |
ec625fb9 | 1737 | MAX_SCHEDULE_TIMEOUT); |
19f81df2 | 1738 | if (ret) |
1c71bc56 | 1739 | return ret; |
19f81df2 RB |
1740 | |
1741 | /* Update all contexts now that we've stalled the submission. */ | |
829a0af2 | 1742 | list_for_each_entry(ctx, &dev_priv->contexts.list, link) { |
c4d52feb | 1743 | struct intel_context *ce = intel_context_lookup(ctx, engine); |
19f81df2 RB |
1744 | u32 *regs; |
1745 | ||
1746 | /* OA settings will be set upon first use */ | |
c4d52feb | 1747 | if (!ce || !ce->state) |
19f81df2 RB |
1748 | continue; |
1749 | ||
666424ab | 1750 | regs = i915_gem_object_pin_map(ce->state->obj, map_type); |
1c71bc56 LL |
1751 | if (IS_ERR(regs)) |
1752 | return PTR_ERR(regs); | |
19f81df2 RB |
1753 | |
1754 | ce->state->obj->mm.dirty = true; | |
1755 | regs += LRC_STATE_PN * PAGE_SIZE / sizeof(*regs); | |
1756 | ||
b146e5ef | 1757 | gen8_update_reg_state_unlocked(ce, regs, oa_config); |
19f81df2 RB |
1758 | |
1759 | i915_gem_object_unpin_map(ce->state->obj); | |
1760 | } | |
1761 | ||
722f3de3 TU |
1762 | /* |
1763 | * Apply the configuration by doing one context restore of the edited | |
1764 | * context image. | |
1765 | */ | |
1766 | rq = i915_request_alloc(engine, dev_priv->kernel_context); | |
1767 | if (IS_ERR(rq)) | |
1768 | return PTR_ERR(rq); | |
1769 | ||
1770 | i915_request_add(rq); | |
1771 | ||
1772 | return 0; | |
19f81df2 RB |
1773 | } |
1774 | ||
5728de2f | 1775 | static int gen8_enable_metric_set(struct i915_perf_stream *stream) |
19f81df2 | 1776 | { |
5728de2f LL |
1777 | struct drm_i915_private *dev_priv = stream->dev_priv; |
1778 | const struct i915_oa_config *oa_config = stream->oa_config; | |
701f8231 | 1779 | int ret; |
19f81df2 RB |
1780 | |
1781 | /* | |
1782 | * We disable slice/unslice clock ratio change reports on SKL since | |
1783 | * they are too noisy. The HW generates a lot of redundant reports | |
1784 | * where the ratio hasn't really changed causing a lot of redundant | |
1785 | * work to processes and increasing the chances we'll hit buffer | |
1786 | * overruns. | |
1787 | * | |
1788 | * Although we don't currently use the 'disable overrun' OABUFFER | |
1789 | * feature it's worth noting that clock ratio reports have to be | |
1790 | * disabled before considering to use that feature since the HW doesn't | |
1791 | * correctly block these reports. | |
1792 | * | |
1793 | * Currently none of the high-level metrics we have depend on knowing | |
1794 | * this ratio to normalize. | |
1795 | * | |
1796 | * Note: This register is not power context saved and restored, but | |
1797 | * that's OK considering that we disable RC6 while the OA unit is | |
1798 | * enabled. | |
1799 | * | |
1800 | * The _INCLUDE_CLK_RATIO bit allows the slice/unslice frequency to | |
1801 | * be read back from automatically triggered reports, as part of the | |
1802 | * RPT_ID field. | |
1803 | */ | |
00690008 | 1804 | if (IS_GEN_RANGE(dev_priv, 9, 11)) { |
19f81df2 RB |
1805 | I915_WRITE(GEN8_OA_DEBUG, |
1806 | _MASKED_BIT_ENABLE(GEN9_OA_DEBUG_DISABLE_CLK_RATIO_REPORTS | | |
1807 | GEN9_OA_DEBUG_INCLUDE_CLK_RATIO)); | |
1808 | } | |
1809 | ||
1810 | /* | |
1811 | * Update all contexts prior writing the mux configurations as we need | |
1812 | * to make sure all slices/subslices are ON before writing to NOA | |
1813 | * registers. | |
1814 | */ | |
41d3fdcd | 1815 | ret = gen8_configure_all_contexts(dev_priv, oa_config); |
19f81df2 RB |
1816 | if (ret) |
1817 | return ret; | |
1818 | ||
701f8231 LL |
1819 | config_oa_regs(dev_priv, oa_config->mux_regs, oa_config->mux_regs_len); |
1820 | ||
701f8231 LL |
1821 | config_oa_regs(dev_priv, oa_config->b_counter_regs, |
1822 | oa_config->b_counter_regs_len); | |
19f81df2 RB |
1823 | |
1824 | return 0; | |
1825 | } | |
1826 | ||
1827 | static void gen8_disable_metric_set(struct drm_i915_private *dev_priv) | |
1828 | { | |
1829 | /* Reset all contexts' slices/subslices configurations. */ | |
41d3fdcd | 1830 | gen8_configure_all_contexts(dev_priv, NULL); |
28964cf2 LL |
1831 | |
1832 | I915_WRITE(GDT_CHICKEN_BITS, (I915_READ(GDT_CHICKEN_BITS) & | |
1833 | ~GT_NOA_ENABLE)); | |
19f81df2 RB |
1834 | } |
1835 | ||
95690a02 LL |
1836 | static void gen10_disable_metric_set(struct drm_i915_private *dev_priv) |
1837 | { | |
1838 | /* Reset all contexts' slices/subslices configurations. */ | |
41d3fdcd | 1839 | gen8_configure_all_contexts(dev_priv, NULL); |
95690a02 LL |
1840 | |
1841 | /* Make sure we disable noa to save power. */ | |
1842 | I915_WRITE(RPM_CONFIG1, | |
1843 | I915_READ(RPM_CONFIG1) & ~GEN10_GT_NOA_ENABLE); | |
1844 | } | |
1845 | ||
5728de2f | 1846 | static void gen7_oa_enable(struct i915_perf_stream *stream) |
d7965152 | 1847 | { |
5728de2f LL |
1848 | struct drm_i915_private *dev_priv = stream->dev_priv; |
1849 | struct i915_gem_context *ctx = stream->ctx; | |
11051303 LL |
1850 | u32 ctx_id = dev_priv->perf.oa.specific_ctx_id; |
1851 | bool periodic = dev_priv->perf.oa.periodic; | |
1852 | u32 period_exponent = dev_priv->perf.oa.period_exponent; | |
1853 | u32 report_format = dev_priv->perf.oa.oa_buffer.format; | |
1854 | ||
1bef3409 RB |
1855 | /* |
1856 | * Reset buf pointers so we don't forward reports from before now. | |
1857 | * | |
1858 | * Think carefully if considering trying to avoid this, since it | |
1859 | * also ensures status flags and the buffer itself are cleared | |
1860 | * in error paths, and we have checks for invalid reports based | |
1861 | * on the assumption that certain fields are written to zeroed | |
1862 | * memory which this helps maintains. | |
1863 | */ | |
1864 | gen7_init_oa_buffer(dev_priv); | |
d7965152 | 1865 | |
11051303 LL |
1866 | I915_WRITE(GEN7_OACONTROL, |
1867 | (ctx_id & GEN7_OACONTROL_CTX_MASK) | | |
1868 | (period_exponent << | |
1869 | GEN7_OACONTROL_TIMER_PERIOD_SHIFT) | | |
1870 | (periodic ? GEN7_OACONTROL_TIMER_ENABLE : 0) | | |
1871 | (report_format << GEN7_OACONTROL_FORMAT_SHIFT) | | |
1872 | (ctx ? GEN7_OACONTROL_PER_CTX_ENABLE : 0) | | |
1873 | GEN7_OACONTROL_ENABLE); | |
d7965152 RB |
1874 | } |
1875 | ||
5728de2f | 1876 | static void gen8_oa_enable(struct i915_perf_stream *stream) |
19f81df2 | 1877 | { |
5728de2f | 1878 | struct drm_i915_private *dev_priv = stream->dev_priv; |
19f81df2 RB |
1879 | u32 report_format = dev_priv->perf.oa.oa_buffer.format; |
1880 | ||
1881 | /* | |
1882 | * Reset buf pointers so we don't forward reports from before now. | |
1883 | * | |
1884 | * Think carefully if considering trying to avoid this, since it | |
1885 | * also ensures status flags and the buffer itself are cleared | |
1886 | * in error paths, and we have checks for invalid reports based | |
1887 | * on the assumption that certain fields are written to zeroed | |
1888 | * memory which this helps maintains. | |
1889 | */ | |
1890 | gen8_init_oa_buffer(dev_priv); | |
1891 | ||
1892 | /* | |
1893 | * Note: we don't rely on the hardware to perform single context | |
1894 | * filtering and instead filter on the cpu based on the context-id | |
1895 | * field of reports | |
1896 | */ | |
1897 | I915_WRITE(GEN8_OACONTROL, (report_format << | |
1898 | GEN8_OA_REPORT_FORMAT_SHIFT) | | |
1899 | GEN8_OA_COUNTER_ENABLE); | |
1900 | } | |
1901 | ||
16d98b31 RB |
1902 | /** |
1903 | * i915_oa_stream_enable - handle `I915_PERF_IOCTL_ENABLE` for OA stream | |
1904 | * @stream: An i915 perf stream opened for OA metrics | |
1905 | * | |
1906 | * [Re]enables hardware periodic sampling according to the period configured | |
1907 | * when opening the stream. This also starts a hrtimer that will periodically | |
1908 | * check for data in the circular OA buffer for notifying userspace (e.g. | |
1909 | * during a read() or poll()). | |
1910 | */ | |
d7965152 RB |
1911 | static void i915_oa_stream_enable(struct i915_perf_stream *stream) |
1912 | { | |
1913 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
1914 | ||
5728de2f | 1915 | dev_priv->perf.oa.ops.oa_enable(stream); |
d7965152 RB |
1916 | |
1917 | if (dev_priv->perf.oa.periodic) | |
1918 | hrtimer_start(&dev_priv->perf.oa.poll_check_timer, | |
1919 | ns_to_ktime(POLL_PERIOD), | |
1920 | HRTIMER_MODE_REL_PINNED); | |
1921 | } | |
1922 | ||
5728de2f | 1923 | static void gen7_oa_disable(struct i915_perf_stream *stream) |
d7965152 | 1924 | { |
5728de2f LL |
1925 | struct drm_i915_private *dev_priv = stream->dev_priv; |
1926 | ||
d7965152 | 1927 | I915_WRITE(GEN7_OACONTROL, 0); |
e896d29a CW |
1928 | if (intel_wait_for_register(dev_priv, |
1929 | GEN7_OACONTROL, GEN7_OACONTROL_ENABLE, 0, | |
1930 | 50)) | |
1931 | DRM_ERROR("wait for OA to be disabled timed out\n"); | |
d7965152 RB |
1932 | } |
1933 | ||
5728de2f | 1934 | static void gen8_oa_disable(struct i915_perf_stream *stream) |
19f81df2 | 1935 | { |
5728de2f LL |
1936 | struct drm_i915_private *dev_priv = stream->dev_priv; |
1937 | ||
19f81df2 | 1938 | I915_WRITE(GEN8_OACONTROL, 0); |
e896d29a CW |
1939 | if (intel_wait_for_register(dev_priv, |
1940 | GEN8_OACONTROL, GEN8_OA_COUNTER_ENABLE, 0, | |
1941 | 50)) | |
1942 | DRM_ERROR("wait for OA to be disabled timed out\n"); | |
19f81df2 RB |
1943 | } |
1944 | ||
16d98b31 RB |
1945 | /** |
1946 | * i915_oa_stream_disable - handle `I915_PERF_IOCTL_DISABLE` for OA stream | |
1947 | * @stream: An i915 perf stream opened for OA metrics | |
1948 | * | |
1949 | * Stops the OA unit from periodically writing counter reports into the | |
1950 | * circular OA buffer. This also stops the hrtimer that periodically checks for | |
1951 | * data in the circular OA buffer, for notifying userspace. | |
1952 | */ | |
d7965152 RB |
1953 | static void i915_oa_stream_disable(struct i915_perf_stream *stream) |
1954 | { | |
1955 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
1956 | ||
5728de2f | 1957 | dev_priv->perf.oa.ops.oa_disable(stream); |
d7965152 RB |
1958 | |
1959 | if (dev_priv->perf.oa.periodic) | |
1960 | hrtimer_cancel(&dev_priv->perf.oa.poll_check_timer); | |
1961 | } | |
1962 | ||
d7965152 RB |
1963 | static const struct i915_perf_stream_ops i915_oa_stream_ops = { |
1964 | .destroy = i915_oa_stream_destroy, | |
1965 | .enable = i915_oa_stream_enable, | |
1966 | .disable = i915_oa_stream_disable, | |
1967 | .wait_unlocked = i915_oa_wait_unlocked, | |
1968 | .poll_wait = i915_oa_poll_wait, | |
1969 | .read = i915_oa_read, | |
eec688e1 RB |
1970 | }; |
1971 | ||
16d98b31 RB |
1972 | /** |
1973 | * i915_oa_stream_init - validate combined props for OA stream and init | |
1974 | * @stream: An i915 perf stream | |
1975 | * @param: The open parameters passed to `DRM_I915_PERF_OPEN` | |
1976 | * @props: The property state that configures stream (individually validated) | |
1977 | * | |
1978 | * While read_properties_unlocked() validates properties in isolation it | |
1979 | * doesn't ensure that the combination necessarily makes sense. | |
1980 | * | |
1981 | * At this point it has been determined that userspace wants a stream of | |
1982 | * OA metrics, but still we need to further validate the combined | |
1983 | * properties are OK. | |
1984 | * | |
1985 | * If the configuration makes sense then we can allocate memory for | |
1986 | * a circular OA buffer and apply the requested metric set configuration. | |
1987 | * | |
1988 | * Returns: zero on success or a negative error code. | |
1989 | */ | |
d7965152 RB |
1990 | static int i915_oa_stream_init(struct i915_perf_stream *stream, |
1991 | struct drm_i915_perf_open_param *param, | |
1992 | struct perf_open_properties *props) | |
1993 | { | |
1994 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
1995 | int format_size; | |
1996 | int ret; | |
1997 | ||
442b8c06 RB |
1998 | /* If the sysfs metrics/ directory wasn't registered for some |
1999 | * reason then don't let userspace try their luck with config | |
2000 | * IDs | |
2001 | */ | |
2002 | if (!dev_priv->perf.metrics_kobj) { | |
7708550c | 2003 | DRM_DEBUG("OA metrics weren't advertised via sysfs\n"); |
442b8c06 RB |
2004 | return -EINVAL; |
2005 | } | |
2006 | ||
d7965152 | 2007 | if (!(props->sample_flags & SAMPLE_OA_REPORT)) { |
7708550c | 2008 | DRM_DEBUG("Only OA report sampling supported\n"); |
d7965152 RB |
2009 | return -EINVAL; |
2010 | } | |
2011 | ||
784b1a84 | 2012 | if (!dev_priv->perf.oa.ops.enable_metric_set) { |
7708550c | 2013 | DRM_DEBUG("OA unit not supported\n"); |
d7965152 RB |
2014 | return -ENODEV; |
2015 | } | |
2016 | ||
2017 | /* To avoid the complexity of having to accurately filter | |
2018 | * counter reports and marshal to the appropriate client | |
2019 | * we currently only allow exclusive access | |
2020 | */ | |
2021 | if (dev_priv->perf.oa.exclusive_stream) { | |
7708550c | 2022 | DRM_DEBUG("OA unit already in use\n"); |
d7965152 RB |
2023 | return -EBUSY; |
2024 | } | |
2025 | ||
d7965152 | 2026 | if (!props->oa_format) { |
7708550c | 2027 | DRM_DEBUG("OA report format not specified\n"); |
d7965152 RB |
2028 | return -EINVAL; |
2029 | } | |
2030 | ||
712122ea RB |
2031 | /* We set up some ratelimit state to potentially throttle any _NOTES |
2032 | * about spurious, invalid OA reports which we don't forward to | |
2033 | * userspace. | |
2034 | * | |
2035 | * The initialization is associated with opening the stream (not driver | |
2036 | * init) considering we print a _NOTE about any throttling when closing | |
2037 | * the stream instead of waiting until driver _fini which no one would | |
2038 | * ever see. | |
2039 | * | |
2040 | * Using the same limiting factors as printk_ratelimit() | |
2041 | */ | |
2042 | ratelimit_state_init(&dev_priv->perf.oa.spurious_report_rs, | |
2043 | 5 * HZ, 10); | |
2044 | /* Since we use a DRM_NOTE for spurious reports it would be | |
2045 | * inconsistent to let __ratelimit() automatically print a warning for | |
2046 | * throttling. | |
2047 | */ | |
2048 | ratelimit_set_flags(&dev_priv->perf.oa.spurious_report_rs, | |
2049 | RATELIMIT_MSG_ON_RELEASE); | |
2050 | ||
d7965152 RB |
2051 | stream->sample_size = sizeof(struct drm_i915_perf_record_header); |
2052 | ||
2053 | format_size = dev_priv->perf.oa.oa_formats[props->oa_format].size; | |
2054 | ||
2055 | stream->sample_flags |= SAMPLE_OA_REPORT; | |
2056 | stream->sample_size += format_size; | |
2057 | ||
2058 | dev_priv->perf.oa.oa_buffer.format_size = format_size; | |
2059 | if (WARN_ON(dev_priv->perf.oa.oa_buffer.format_size == 0)) | |
2060 | return -EINVAL; | |
2061 | ||
2062 | dev_priv->perf.oa.oa_buffer.format = | |
2063 | dev_priv->perf.oa.oa_formats[props->oa_format].format; | |
2064 | ||
d7965152 | 2065 | dev_priv->perf.oa.periodic = props->oa_periodic; |
0dd860cf | 2066 | if (dev_priv->perf.oa.periodic) |
d7965152 RB |
2067 | dev_priv->perf.oa.period_exponent = props->oa_period_exponent; |
2068 | ||
d7965152 RB |
2069 | if (stream->ctx) { |
2070 | ret = oa_get_render_ctx_id(stream); | |
9bd9be66 LL |
2071 | if (ret) { |
2072 | DRM_DEBUG("Invalid context id to filter with\n"); | |
d7965152 | 2073 | return ret; |
9bd9be66 | 2074 | } |
d7965152 RB |
2075 | } |
2076 | ||
f89823c2 | 2077 | ret = get_oa_config(dev_priv, props->metrics_set, &stream->oa_config); |
9bd9be66 LL |
2078 | if (ret) { |
2079 | DRM_DEBUG("Invalid OA config id=%i\n", props->metrics_set); | |
f89823c2 | 2080 | goto err_config; |
9bd9be66 | 2081 | } |
701f8231 | 2082 | |
d7965152 RB |
2083 | /* PRM - observability performance counters: |
2084 | * | |
2085 | * OACONTROL, performance counter enable, note: | |
2086 | * | |
2087 | * "When this bit is set, in order to have coherent counts, | |
2088 | * RC6 power state and trunk clock gating must be disabled. | |
2089 | * This can be achieved by programming MMIO registers as | |
2090 | * 0xA094=0 and 0xA090[31]=1" | |
2091 | * | |
2092 | * In our case we are expecting that taking pm + FORCEWAKE | |
2093 | * references will effectively disable RC6. | |
2094 | */ | |
6619c007 | 2095 | stream->wakeref = intel_runtime_pm_get(dev_priv); |
3ceea6a1 | 2096 | intel_uncore_forcewake_get(&dev_priv->uncore, FORCEWAKE_ALL); |
d7965152 | 2097 | |
fe841686 | 2098 | ret = alloc_oa_buffer(dev_priv); |
987f8c44 | 2099 | if (ret) |
2100 | goto err_oa_buf_alloc; | |
2101 | ||
41d3fdcd LL |
2102 | ret = i915_mutex_lock_interruptible(&dev_priv->drm); |
2103 | if (ret) | |
2104 | goto err_lock; | |
2105 | ||
ec431eae LL |
2106 | stream->ops = &i915_oa_stream_ops; |
2107 | dev_priv->perf.oa.exclusive_stream = stream; | |
2108 | ||
5728de2f | 2109 | ret = dev_priv->perf.oa.ops.enable_metric_set(stream); |
9bd9be66 LL |
2110 | if (ret) { |
2111 | DRM_DEBUG("Unable to enable metric set\n"); | |
d7965152 | 2112 | goto err_enable; |
9bd9be66 | 2113 | } |
d7965152 | 2114 | |
701f8231 LL |
2115 | mutex_unlock(&dev_priv->drm.struct_mutex); |
2116 | ||
d7965152 RB |
2117 | return 0; |
2118 | ||
41d3fdcd | 2119 | err_enable: |
ec431eae | 2120 | dev_priv->perf.oa.exclusive_stream = NULL; |
701f8231 | 2121 | dev_priv->perf.oa.ops.disable_metric_set(dev_priv); |
41d3fdcd | 2122 | mutex_unlock(&dev_priv->drm.struct_mutex); |
701f8231 | 2123 | |
41d3fdcd | 2124 | err_lock: |
d7965152 RB |
2125 | free_oa_buffer(dev_priv); |
2126 | ||
2127 | err_oa_buf_alloc: | |
f89823c2 LL |
2128 | put_oa_config(dev_priv, stream->oa_config); |
2129 | ||
3ceea6a1 | 2130 | intel_uncore_forcewake_put(&dev_priv->uncore, FORCEWAKE_ALL); |
6619c007 | 2131 | intel_runtime_pm_put(dev_priv, stream->wakeref); |
f89823c2 LL |
2132 | |
2133 | err_config: | |
d7965152 RB |
2134 | if (stream->ctx) |
2135 | oa_put_render_ctx_id(stream); | |
2136 | ||
2137 | return ret; | |
2138 | } | |
2139 | ||
19f81df2 | 2140 | void i915_oa_init_reg_state(struct intel_engine_cs *engine, |
b146e5ef CW |
2141 | struct intel_context *ce, |
2142 | u32 *regs) | |
19f81df2 | 2143 | { |
28b6cb08 | 2144 | struct i915_perf_stream *stream; |
19f81df2 | 2145 | |
8a68d464 | 2146 | if (engine->class != RENDER_CLASS) |
19f81df2 RB |
2147 | return; |
2148 | ||
28b6cb08 | 2149 | stream = engine->i915->perf.oa.exclusive_stream; |
701f8231 | 2150 | if (stream) |
b146e5ef | 2151 | gen8_update_reg_state_unlocked(ce, regs, stream->oa_config); |
19f81df2 RB |
2152 | } |
2153 | ||
16d98b31 RB |
2154 | /** |
2155 | * i915_perf_read_locked - &i915_perf_stream_ops->read with error normalisation | |
2156 | * @stream: An i915 perf stream | |
2157 | * @file: An i915 perf stream file | |
2158 | * @buf: destination buffer given by userspace | |
2159 | * @count: the number of bytes userspace wants to read | |
2160 | * @ppos: (inout) file seek position (unused) | |
2161 | * | |
2162 | * Besides wrapping &i915_perf_stream_ops->read this provides a common place to | |
2163 | * ensure that if we've successfully copied any data then reporting that takes | |
2164 | * precedence over any internal error status, so the data isn't lost. | |
2165 | * | |
2166 | * For example ret will be -ENOSPC whenever there is more buffered data than | |
2167 | * can be copied to userspace, but that's only interesting if we weren't able | |
2168 | * to copy some data because it implies the userspace buffer is too small to | |
2169 | * receive a single record (and we never split records). | |
2170 | * | |
2171 | * Another case with ret == -EFAULT is more of a grey area since it would seem | |
2172 | * like bad form for userspace to ask us to overrun its buffer, but the user | |
2173 | * knows best: | |
2174 | * | |
2175 | * http://yarchive.net/comp/linux/partial_reads_writes.html | |
2176 | * | |
2177 | * Returns: The number of bytes copied or a negative error code on failure. | |
2178 | */ | |
eec688e1 RB |
2179 | static ssize_t i915_perf_read_locked(struct i915_perf_stream *stream, |
2180 | struct file *file, | |
2181 | char __user *buf, | |
2182 | size_t count, | |
2183 | loff_t *ppos) | |
2184 | { | |
2185 | /* Note we keep the offset (aka bytes read) separate from any | |
2186 | * error status so that the final check for whether we return | |
2187 | * the bytes read with a higher precedence than any error (see | |
2188 | * comment below) doesn't need to be handled/duplicated in | |
2189 | * stream->ops->read() implementations. | |
2190 | */ | |
2191 | size_t offset = 0; | |
2192 | int ret = stream->ops->read(stream, buf, count, &offset); | |
2193 | ||
eec688e1 RB |
2194 | return offset ?: (ret ?: -EAGAIN); |
2195 | } | |
2196 | ||
16d98b31 RB |
2197 | /** |
2198 | * i915_perf_read - handles read() FOP for i915 perf stream FDs | |
2199 | * @file: An i915 perf stream file | |
2200 | * @buf: destination buffer given by userspace | |
2201 | * @count: the number of bytes userspace wants to read | |
2202 | * @ppos: (inout) file seek position (unused) | |
2203 | * | |
2204 | * The entry point for handling a read() on a stream file descriptor from | |
2205 | * userspace. Most of the work is left to the i915_perf_read_locked() and | |
2206 | * &i915_perf_stream_ops->read but to save having stream implementations (of | |
2207 | * which we might have multiple later) we handle blocking read here. | |
2208 | * | |
2209 | * We can also consistently treat trying to read from a disabled stream | |
2210 | * as an IO error so implementations can assume the stream is enabled | |
2211 | * while reading. | |
2212 | * | |
2213 | * Returns: The number of bytes copied or a negative error code on failure. | |
2214 | */ | |
eec688e1 RB |
2215 | static ssize_t i915_perf_read(struct file *file, |
2216 | char __user *buf, | |
2217 | size_t count, | |
2218 | loff_t *ppos) | |
2219 | { | |
2220 | struct i915_perf_stream *stream = file->private_data; | |
2221 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
2222 | ssize_t ret; | |
2223 | ||
d7965152 RB |
2224 | /* To ensure it's handled consistently we simply treat all reads of a |
2225 | * disabled stream as an error. In particular it might otherwise lead | |
2226 | * to a deadlock for blocking file descriptors... | |
2227 | */ | |
2228 | if (!stream->enabled) | |
2229 | return -EIO; | |
2230 | ||
eec688e1 | 2231 | if (!(file->f_flags & O_NONBLOCK)) { |
d7965152 RB |
2232 | /* There's the small chance of false positives from |
2233 | * stream->ops->wait_unlocked. | |
2234 | * | |
2235 | * E.g. with single context filtering since we only wait until | |
2236 | * oabuffer has >= 1 report we don't immediately know whether | |
2237 | * any reports really belong to the current context | |
eec688e1 RB |
2238 | */ |
2239 | do { | |
2240 | ret = stream->ops->wait_unlocked(stream); | |
2241 | if (ret) | |
2242 | return ret; | |
2243 | ||
2244 | mutex_lock(&dev_priv->perf.lock); | |
2245 | ret = i915_perf_read_locked(stream, file, | |
2246 | buf, count, ppos); | |
2247 | mutex_unlock(&dev_priv->perf.lock); | |
2248 | } while (ret == -EAGAIN); | |
2249 | } else { | |
2250 | mutex_lock(&dev_priv->perf.lock); | |
2251 | ret = i915_perf_read_locked(stream, file, buf, count, ppos); | |
2252 | mutex_unlock(&dev_priv->perf.lock); | |
2253 | } | |
2254 | ||
a9a08845 | 2255 | /* We allow the poll checking to sometimes report false positive EPOLLIN |
26ebd9c7 RB |
2256 | * events where we might actually report EAGAIN on read() if there's |
2257 | * not really any data available. In this situation though we don't | |
a9a08845 | 2258 | * want to enter a busy loop between poll() reporting a EPOLLIN event |
26ebd9c7 RB |
2259 | * and read() returning -EAGAIN. Clearing the oa.pollin state here |
2260 | * effectively ensures we back off until the next hrtimer callback | |
a9a08845 | 2261 | * before reporting another EPOLLIN event. |
26ebd9c7 RB |
2262 | */ |
2263 | if (ret >= 0 || ret == -EAGAIN) { | |
d7965152 RB |
2264 | /* Maybe make ->pollin per-stream state if we support multiple |
2265 | * concurrent streams in the future. | |
2266 | */ | |
2267 | dev_priv->perf.oa.pollin = false; | |
2268 | } | |
2269 | ||
eec688e1 RB |
2270 | return ret; |
2271 | } | |
2272 | ||
d7965152 RB |
2273 | static enum hrtimer_restart oa_poll_check_timer_cb(struct hrtimer *hrtimer) |
2274 | { | |
2275 | struct drm_i915_private *dev_priv = | |
2276 | container_of(hrtimer, typeof(*dev_priv), | |
2277 | perf.oa.poll_check_timer); | |
2278 | ||
19f81df2 | 2279 | if (oa_buffer_check_unlocked(dev_priv)) { |
d7965152 RB |
2280 | dev_priv->perf.oa.pollin = true; |
2281 | wake_up(&dev_priv->perf.oa.poll_wq); | |
2282 | } | |
2283 | ||
2284 | hrtimer_forward_now(hrtimer, ns_to_ktime(POLL_PERIOD)); | |
2285 | ||
2286 | return HRTIMER_RESTART; | |
2287 | } | |
2288 | ||
16d98b31 RB |
2289 | /** |
2290 | * i915_perf_poll_locked - poll_wait() with a suitable wait queue for stream | |
2291 | * @dev_priv: i915 device instance | |
2292 | * @stream: An i915 perf stream | |
2293 | * @file: An i915 perf stream file | |
2294 | * @wait: poll() state table | |
2295 | * | |
2296 | * For handling userspace polling on an i915 perf stream, this calls through to | |
2297 | * &i915_perf_stream_ops->poll_wait to call poll_wait() with a wait queue that | |
2298 | * will be woken for new stream data. | |
2299 | * | |
2300 | * Note: The &drm_i915_private->perf.lock mutex has been taken to serialize | |
2301 | * with any non-file-operation driver hooks. | |
2302 | * | |
2303 | * Returns: any poll events that are ready without sleeping | |
2304 | */ | |
afc9a42b | 2305 | static __poll_t i915_perf_poll_locked(struct drm_i915_private *dev_priv, |
d7965152 | 2306 | struct i915_perf_stream *stream, |
eec688e1 RB |
2307 | struct file *file, |
2308 | poll_table *wait) | |
2309 | { | |
afc9a42b | 2310 | __poll_t events = 0; |
eec688e1 RB |
2311 | |
2312 | stream->ops->poll_wait(stream, file, wait); | |
2313 | ||
d7965152 RB |
2314 | /* Note: we don't explicitly check whether there's something to read |
2315 | * here since this path may be very hot depending on what else | |
2316 | * userspace is polling, or on the timeout in use. We rely solely on | |
2317 | * the hrtimer/oa_poll_check_timer_cb to notify us when there are | |
2318 | * samples to read. | |
2319 | */ | |
2320 | if (dev_priv->perf.oa.pollin) | |
a9a08845 | 2321 | events |= EPOLLIN; |
eec688e1 | 2322 | |
d7965152 | 2323 | return events; |
eec688e1 RB |
2324 | } |
2325 | ||
16d98b31 RB |
2326 | /** |
2327 | * i915_perf_poll - call poll_wait() with a suitable wait queue for stream | |
2328 | * @file: An i915 perf stream file | |
2329 | * @wait: poll() state table | |
2330 | * | |
2331 | * For handling userspace polling on an i915 perf stream, this ensures | |
2332 | * poll_wait() gets called with a wait queue that will be woken for new stream | |
2333 | * data. | |
2334 | * | |
2335 | * Note: Implementation deferred to i915_perf_poll_locked() | |
2336 | * | |
2337 | * Returns: any poll events that are ready without sleeping | |
2338 | */ | |
afc9a42b | 2339 | static __poll_t i915_perf_poll(struct file *file, poll_table *wait) |
eec688e1 RB |
2340 | { |
2341 | struct i915_perf_stream *stream = file->private_data; | |
2342 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
afc9a42b | 2343 | __poll_t ret; |
eec688e1 RB |
2344 | |
2345 | mutex_lock(&dev_priv->perf.lock); | |
d7965152 | 2346 | ret = i915_perf_poll_locked(dev_priv, stream, file, wait); |
eec688e1 RB |
2347 | mutex_unlock(&dev_priv->perf.lock); |
2348 | ||
2349 | return ret; | |
2350 | } | |
2351 | ||
16d98b31 RB |
2352 | /** |
2353 | * i915_perf_enable_locked - handle `I915_PERF_IOCTL_ENABLE` ioctl | |
2354 | * @stream: A disabled i915 perf stream | |
2355 | * | |
2356 | * [Re]enables the associated capture of data for this stream. | |
2357 | * | |
2358 | * If a stream was previously enabled then there's currently no intention | |
2359 | * to provide userspace any guarantee about the preservation of previously | |
2360 | * buffered data. | |
2361 | */ | |
eec688e1 RB |
2362 | static void i915_perf_enable_locked(struct i915_perf_stream *stream) |
2363 | { | |
2364 | if (stream->enabled) | |
2365 | return; | |
2366 | ||
2367 | /* Allow stream->ops->enable() to refer to this */ | |
2368 | stream->enabled = true; | |
2369 | ||
2370 | if (stream->ops->enable) | |
2371 | stream->ops->enable(stream); | |
2372 | } | |
2373 | ||
16d98b31 RB |
2374 | /** |
2375 | * i915_perf_disable_locked - handle `I915_PERF_IOCTL_DISABLE` ioctl | |
2376 | * @stream: An enabled i915 perf stream | |
2377 | * | |
2378 | * Disables the associated capture of data for this stream. | |
2379 | * | |
2380 | * The intention is that disabling an re-enabling a stream will ideally be | |
2381 | * cheaper than destroying and re-opening a stream with the same configuration, | |
2382 | * though there are no formal guarantees about what state or buffered data | |
2383 | * must be retained between disabling and re-enabling a stream. | |
2384 | * | |
2385 | * Note: while a stream is disabled it's considered an error for userspace | |
2386 | * to attempt to read from the stream (-EIO). | |
2387 | */ | |
eec688e1 RB |
2388 | static void i915_perf_disable_locked(struct i915_perf_stream *stream) |
2389 | { | |
2390 | if (!stream->enabled) | |
2391 | return; | |
2392 | ||
2393 | /* Allow stream->ops->disable() to refer to this */ | |
2394 | stream->enabled = false; | |
2395 | ||
2396 | if (stream->ops->disable) | |
2397 | stream->ops->disable(stream); | |
2398 | } | |
2399 | ||
16d98b31 RB |
2400 | /** |
2401 | * i915_perf_ioctl - support ioctl() usage with i915 perf stream FDs | |
2402 | * @stream: An i915 perf stream | |
2403 | * @cmd: the ioctl request | |
2404 | * @arg: the ioctl data | |
2405 | * | |
2406 | * Note: The &drm_i915_private->perf.lock mutex has been taken to serialize | |
2407 | * with any non-file-operation driver hooks. | |
2408 | * | |
2409 | * Returns: zero on success or a negative error code. Returns -EINVAL for | |
2410 | * an unknown ioctl request. | |
2411 | */ | |
eec688e1 RB |
2412 | static long i915_perf_ioctl_locked(struct i915_perf_stream *stream, |
2413 | unsigned int cmd, | |
2414 | unsigned long arg) | |
2415 | { | |
2416 | switch (cmd) { | |
2417 | case I915_PERF_IOCTL_ENABLE: | |
2418 | i915_perf_enable_locked(stream); | |
2419 | return 0; | |
2420 | case I915_PERF_IOCTL_DISABLE: | |
2421 | i915_perf_disable_locked(stream); | |
2422 | return 0; | |
2423 | } | |
2424 | ||
2425 | return -EINVAL; | |
2426 | } | |
2427 | ||
16d98b31 RB |
2428 | /** |
2429 | * i915_perf_ioctl - support ioctl() usage with i915 perf stream FDs | |
2430 | * @file: An i915 perf stream file | |
2431 | * @cmd: the ioctl request | |
2432 | * @arg: the ioctl data | |
2433 | * | |
2434 | * Implementation deferred to i915_perf_ioctl_locked(). | |
2435 | * | |
2436 | * Returns: zero on success or a negative error code. Returns -EINVAL for | |
2437 | * an unknown ioctl request. | |
2438 | */ | |
eec688e1 RB |
2439 | static long i915_perf_ioctl(struct file *file, |
2440 | unsigned int cmd, | |
2441 | unsigned long arg) | |
2442 | { | |
2443 | struct i915_perf_stream *stream = file->private_data; | |
2444 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
2445 | long ret; | |
2446 | ||
2447 | mutex_lock(&dev_priv->perf.lock); | |
2448 | ret = i915_perf_ioctl_locked(stream, cmd, arg); | |
2449 | mutex_unlock(&dev_priv->perf.lock); | |
2450 | ||
2451 | return ret; | |
2452 | } | |
2453 | ||
16d98b31 RB |
2454 | /** |
2455 | * i915_perf_destroy_locked - destroy an i915 perf stream | |
2456 | * @stream: An i915 perf stream | |
2457 | * | |
2458 | * Frees all resources associated with the given i915 perf @stream, disabling | |
2459 | * any associated data capture in the process. | |
2460 | * | |
2461 | * Note: The &drm_i915_private->perf.lock mutex has been taken to serialize | |
2462 | * with any non-file-operation driver hooks. | |
2463 | */ | |
eec688e1 RB |
2464 | static void i915_perf_destroy_locked(struct i915_perf_stream *stream) |
2465 | { | |
eec688e1 RB |
2466 | if (stream->enabled) |
2467 | i915_perf_disable_locked(stream); | |
2468 | ||
2469 | if (stream->ops->destroy) | |
2470 | stream->ops->destroy(stream); | |
2471 | ||
2472 | list_del(&stream->link); | |
2473 | ||
69df05e1 | 2474 | if (stream->ctx) |
5f09a9c8 | 2475 | i915_gem_context_put(stream->ctx); |
eec688e1 RB |
2476 | |
2477 | kfree(stream); | |
2478 | } | |
2479 | ||
16d98b31 RB |
2480 | /** |
2481 | * i915_perf_release - handles userspace close() of a stream file | |
2482 | * @inode: anonymous inode associated with file | |
2483 | * @file: An i915 perf stream file | |
2484 | * | |
2485 | * Cleans up any resources associated with an open i915 perf stream file. | |
2486 | * | |
2487 | * NB: close() can't really fail from the userspace point of view. | |
2488 | * | |
2489 | * Returns: zero on success or a negative error code. | |
2490 | */ | |
eec688e1 RB |
2491 | static int i915_perf_release(struct inode *inode, struct file *file) |
2492 | { | |
2493 | struct i915_perf_stream *stream = file->private_data; | |
2494 | struct drm_i915_private *dev_priv = stream->dev_priv; | |
2495 | ||
2496 | mutex_lock(&dev_priv->perf.lock); | |
2497 | i915_perf_destroy_locked(stream); | |
2498 | mutex_unlock(&dev_priv->perf.lock); | |
2499 | ||
2500 | return 0; | |
2501 | } | |
2502 | ||
2503 | ||
2504 | static const struct file_operations fops = { | |
2505 | .owner = THIS_MODULE, | |
2506 | .llseek = no_llseek, | |
2507 | .release = i915_perf_release, | |
2508 | .poll = i915_perf_poll, | |
2509 | .read = i915_perf_read, | |
2510 | .unlocked_ioctl = i915_perf_ioctl, | |
191f8960 LL |
2511 | /* Our ioctl have no arguments, so it's safe to use the same function |
2512 | * to handle 32bits compatibility. | |
2513 | */ | |
2514 | .compat_ioctl = i915_perf_ioctl, | |
eec688e1 RB |
2515 | }; |
2516 | ||
2517 | ||
16d98b31 RB |
2518 | /** |
2519 | * i915_perf_open_ioctl_locked - DRM ioctl() for userspace to open a stream FD | |
2520 | * @dev_priv: i915 device instance | |
2521 | * @param: The open parameters passed to 'DRM_I915_PERF_OPEN` | |
2522 | * @props: individually validated u64 property value pairs | |
2523 | * @file: drm file | |
2524 | * | |
2525 | * See i915_perf_ioctl_open() for interface details. | |
2526 | * | |
2527 | * Implements further stream config validation and stream initialization on | |
2528 | * behalf of i915_perf_open_ioctl() with the &drm_i915_private->perf.lock mutex | |
2529 | * taken to serialize with any non-file-operation driver hooks. | |
2530 | * | |
2531 | * Note: at this point the @props have only been validated in isolation and | |
2532 | * it's still necessary to validate that the combination of properties makes | |
2533 | * sense. | |
2534 | * | |
2535 | * In the case where userspace is interested in OA unit metrics then further | |
2536 | * config validation and stream initialization details will be handled by | |
2537 | * i915_oa_stream_init(). The code here should only validate config state that | |
2538 | * will be relevant to all stream types / backends. | |
2539 | * | |
2540 | * Returns: zero on success or a negative error code. | |
2541 | */ | |
eec688e1 RB |
2542 | static int |
2543 | i915_perf_open_ioctl_locked(struct drm_i915_private *dev_priv, | |
2544 | struct drm_i915_perf_open_param *param, | |
2545 | struct perf_open_properties *props, | |
2546 | struct drm_file *file) | |
2547 | { | |
2548 | struct i915_gem_context *specific_ctx = NULL; | |
2549 | struct i915_perf_stream *stream = NULL; | |
2550 | unsigned long f_flags = 0; | |
19f81df2 | 2551 | bool privileged_op = true; |
eec688e1 RB |
2552 | int stream_fd; |
2553 | int ret; | |
2554 | ||
2555 | if (props->single_context) { | |
2556 | u32 ctx_handle = props->ctx_handle; | |
2557 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
2558 | ||
635f56c3 ID |
2559 | specific_ctx = i915_gem_context_lookup(file_priv, ctx_handle); |
2560 | if (!specific_ctx) { | |
2561 | DRM_DEBUG("Failed to look up context with ID %u for opening perf stream\n", | |
2562 | ctx_handle); | |
2563 | ret = -ENOENT; | |
eec688e1 RB |
2564 | goto err; |
2565 | } | |
2566 | } | |
2567 | ||
19f81df2 RB |
2568 | /* |
2569 | * On Haswell the OA unit supports clock gating off for a specific | |
2570 | * context and in this mode there's no visibility of metrics for the | |
2571 | * rest of the system, which we consider acceptable for a | |
2572 | * non-privileged client. | |
2573 | * | |
2574 | * For Gen8+ the OA unit no longer supports clock gating off for a | |
2575 | * specific context and the kernel can't securely stop the counters | |
2576 | * from updating as system-wide / global values. Even though we can | |
2577 | * filter reports based on the included context ID we can't block | |
2578 | * clients from seeing the raw / global counter values via | |
2579 | * MI_REPORT_PERF_COUNT commands and so consider it a privileged op to | |
2580 | * enable the OA unit by default. | |
2581 | */ | |
2582 | if (IS_HASWELL(dev_priv) && specific_ctx) | |
2583 | privileged_op = false; | |
2584 | ||
ccdf6341 RB |
2585 | /* Similar to perf's kernel.perf_paranoid_cpu sysctl option |
2586 | * we check a dev.i915.perf_stream_paranoid sysctl option | |
2587 | * to determine if it's ok to access system wide OA counters | |
2588 | * without CAP_SYS_ADMIN privileges. | |
2589 | */ | |
19f81df2 | 2590 | if (privileged_op && |
ccdf6341 | 2591 | i915_perf_stream_paranoid && !capable(CAP_SYS_ADMIN)) { |
7708550c | 2592 | DRM_DEBUG("Insufficient privileges to open system-wide i915 perf stream\n"); |
eec688e1 RB |
2593 | ret = -EACCES; |
2594 | goto err_ctx; | |
2595 | } | |
2596 | ||
2597 | stream = kzalloc(sizeof(*stream), GFP_KERNEL); | |
2598 | if (!stream) { | |
2599 | ret = -ENOMEM; | |
2600 | goto err_ctx; | |
2601 | } | |
2602 | ||
eec688e1 RB |
2603 | stream->dev_priv = dev_priv; |
2604 | stream->ctx = specific_ctx; | |
2605 | ||
d7965152 RB |
2606 | ret = i915_oa_stream_init(stream, param, props); |
2607 | if (ret) | |
2608 | goto err_alloc; | |
2609 | ||
2610 | /* we avoid simply assigning stream->sample_flags = props->sample_flags | |
2611 | * to have _stream_init check the combination of sample flags more | |
2612 | * thoroughly, but still this is the expected result at this point. | |
eec688e1 | 2613 | */ |
d7965152 RB |
2614 | if (WARN_ON(stream->sample_flags != props->sample_flags)) { |
2615 | ret = -ENODEV; | |
22f880ca | 2616 | goto err_flags; |
d7965152 | 2617 | } |
eec688e1 RB |
2618 | |
2619 | list_add(&stream->link, &dev_priv->perf.streams); | |
2620 | ||
2621 | if (param->flags & I915_PERF_FLAG_FD_CLOEXEC) | |
2622 | f_flags |= O_CLOEXEC; | |
2623 | if (param->flags & I915_PERF_FLAG_FD_NONBLOCK) | |
2624 | f_flags |= O_NONBLOCK; | |
2625 | ||
2626 | stream_fd = anon_inode_getfd("[i915_perf]", &fops, stream, f_flags); | |
2627 | if (stream_fd < 0) { | |
2628 | ret = stream_fd; | |
2629 | goto err_open; | |
2630 | } | |
2631 | ||
2632 | if (!(param->flags & I915_PERF_FLAG_DISABLED)) | |
2633 | i915_perf_enable_locked(stream); | |
2634 | ||
2635 | return stream_fd; | |
2636 | ||
2637 | err_open: | |
2638 | list_del(&stream->link); | |
22f880ca | 2639 | err_flags: |
eec688e1 RB |
2640 | if (stream->ops->destroy) |
2641 | stream->ops->destroy(stream); | |
2642 | err_alloc: | |
2643 | kfree(stream); | |
2644 | err_ctx: | |
69df05e1 | 2645 | if (specific_ctx) |
5f09a9c8 | 2646 | i915_gem_context_put(specific_ctx); |
eec688e1 RB |
2647 | err: |
2648 | return ret; | |
2649 | } | |
2650 | ||
155e941f RB |
2651 | static u64 oa_exponent_to_ns(struct drm_i915_private *dev_priv, int exponent) |
2652 | { | |
9f9b2792 | 2653 | return div64_u64(1000000000ULL * (2ULL << exponent), |
0258404f | 2654 | 1000ULL * RUNTIME_INFO(dev_priv)->cs_timestamp_frequency_khz); |
155e941f RB |
2655 | } |
2656 | ||
16d98b31 RB |
2657 | /** |
2658 | * read_properties_unlocked - validate + copy userspace stream open properties | |
2659 | * @dev_priv: i915 device instance | |
2660 | * @uprops: The array of u64 key value pairs given by userspace | |
2661 | * @n_props: The number of key value pairs expected in @uprops | |
2662 | * @props: The stream configuration built up while validating properties | |
eec688e1 RB |
2663 | * |
2664 | * Note this function only validates properties in isolation it doesn't | |
2665 | * validate that the combination of properties makes sense or that all | |
2666 | * properties necessary for a particular kind of stream have been set. | |
16d98b31 RB |
2667 | * |
2668 | * Note that there currently aren't any ordering requirements for properties so | |
2669 | * we shouldn't validate or assume anything about ordering here. This doesn't | |
2670 | * rule out defining new properties with ordering requirements in the future. | |
eec688e1 RB |
2671 | */ |
2672 | static int read_properties_unlocked(struct drm_i915_private *dev_priv, | |
2673 | u64 __user *uprops, | |
2674 | u32 n_props, | |
2675 | struct perf_open_properties *props) | |
2676 | { | |
2677 | u64 __user *uprop = uprops; | |
701f8231 | 2678 | u32 i; |
eec688e1 RB |
2679 | |
2680 | memset(props, 0, sizeof(struct perf_open_properties)); | |
2681 | ||
2682 | if (!n_props) { | |
7708550c | 2683 | DRM_DEBUG("No i915 perf properties given\n"); |
eec688e1 RB |
2684 | return -EINVAL; |
2685 | } | |
2686 | ||
2687 | /* Considering that ID = 0 is reserved and assuming that we don't | |
2688 | * (currently) expect any configurations to ever specify duplicate | |
2689 | * values for a particular property ID then the last _PROP_MAX value is | |
2690 | * one greater than the maximum number of properties we expect to get | |
2691 | * from userspace. | |
2692 | */ | |
2693 | if (n_props >= DRM_I915_PERF_PROP_MAX) { | |
7708550c | 2694 | DRM_DEBUG("More i915 perf properties specified than exist\n"); |
eec688e1 RB |
2695 | return -EINVAL; |
2696 | } | |
2697 | ||
2698 | for (i = 0; i < n_props; i++) { | |
00319ba0 | 2699 | u64 oa_period, oa_freq_hz; |
eec688e1 RB |
2700 | u64 id, value; |
2701 | int ret; | |
2702 | ||
2703 | ret = get_user(id, uprop); | |
2704 | if (ret) | |
2705 | return ret; | |
2706 | ||
2707 | ret = get_user(value, uprop + 1); | |
2708 | if (ret) | |
2709 | return ret; | |
2710 | ||
0a309f9e MA |
2711 | if (id == 0 || id >= DRM_I915_PERF_PROP_MAX) { |
2712 | DRM_DEBUG("Unknown i915 perf property ID\n"); | |
2713 | return -EINVAL; | |
2714 | } | |
2715 | ||
eec688e1 RB |
2716 | switch ((enum drm_i915_perf_property_id)id) { |
2717 | case DRM_I915_PERF_PROP_CTX_HANDLE: | |
2718 | props->single_context = 1; | |
2719 | props->ctx_handle = value; | |
2720 | break; | |
d7965152 | 2721 | case DRM_I915_PERF_PROP_SAMPLE_OA: |
b6dd47b9 LL |
2722 | if (value) |
2723 | props->sample_flags |= SAMPLE_OA_REPORT; | |
d7965152 RB |
2724 | break; |
2725 | case DRM_I915_PERF_PROP_OA_METRICS_SET: | |
701f8231 | 2726 | if (value == 0) { |
7708550c | 2727 | DRM_DEBUG("Unknown OA metric set ID\n"); |
d7965152 RB |
2728 | return -EINVAL; |
2729 | } | |
2730 | props->metrics_set = value; | |
2731 | break; | |
2732 | case DRM_I915_PERF_PROP_OA_FORMAT: | |
2733 | if (value == 0 || value >= I915_OA_FORMAT_MAX) { | |
52c57c26 RB |
2734 | DRM_DEBUG("Out-of-range OA report format %llu\n", |
2735 | value); | |
d7965152 RB |
2736 | return -EINVAL; |
2737 | } | |
2738 | if (!dev_priv->perf.oa.oa_formats[value].size) { | |
52c57c26 RB |
2739 | DRM_DEBUG("Unsupported OA report format %llu\n", |
2740 | value); | |
d7965152 RB |
2741 | return -EINVAL; |
2742 | } | |
2743 | props->oa_format = value; | |
2744 | break; | |
2745 | case DRM_I915_PERF_PROP_OA_EXPONENT: | |
2746 | if (value > OA_EXPONENT_MAX) { | |
7708550c RB |
2747 | DRM_DEBUG("OA timer exponent too high (> %u)\n", |
2748 | OA_EXPONENT_MAX); | |
d7965152 RB |
2749 | return -EINVAL; |
2750 | } | |
2751 | ||
00319ba0 | 2752 | /* Theoretically we can program the OA unit to sample |
155e941f RB |
2753 | * e.g. every 160ns for HSW, 167ns for BDW/SKL or 104ns |
2754 | * for BXT. We don't allow such high sampling | |
2755 | * frequencies by default unless root. | |
00319ba0 | 2756 | */ |
155e941f | 2757 | |
00319ba0 | 2758 | BUILD_BUG_ON(sizeof(oa_period) != 8); |
155e941f | 2759 | oa_period = oa_exponent_to_ns(dev_priv, value); |
00319ba0 RB |
2760 | |
2761 | /* This check is primarily to ensure that oa_period <= | |
2762 | * UINT32_MAX (before passing to do_div which only | |
2763 | * accepts a u32 denominator), but we can also skip | |
2764 | * checking anything < 1Hz which implicitly can't be | |
2765 | * limited via an integer oa_max_sample_rate. | |
d7965152 | 2766 | */ |
00319ba0 RB |
2767 | if (oa_period <= NSEC_PER_SEC) { |
2768 | u64 tmp = NSEC_PER_SEC; | |
2769 | do_div(tmp, oa_period); | |
2770 | oa_freq_hz = tmp; | |
2771 | } else | |
2772 | oa_freq_hz = 0; | |
2773 | ||
2774 | if (oa_freq_hz > i915_oa_max_sample_rate && | |
2775 | !capable(CAP_SYS_ADMIN)) { | |
7708550c | 2776 | DRM_DEBUG("OA exponent would exceed the max sampling frequency (sysctl dev.i915.oa_max_sample_rate) %uHz without root privileges\n", |
00319ba0 | 2777 | i915_oa_max_sample_rate); |
d7965152 RB |
2778 | return -EACCES; |
2779 | } | |
2780 | ||
2781 | props->oa_periodic = true; | |
2782 | props->oa_period_exponent = value; | |
2783 | break; | |
0a309f9e | 2784 | case DRM_I915_PERF_PROP_MAX: |
eec688e1 | 2785 | MISSING_CASE(id); |
eec688e1 RB |
2786 | return -EINVAL; |
2787 | } | |
2788 | ||
2789 | uprop += 2; | |
2790 | } | |
2791 | ||
2792 | return 0; | |
2793 | } | |
2794 | ||
16d98b31 RB |
2795 | /** |
2796 | * i915_perf_open_ioctl - DRM ioctl() for userspace to open a stream FD | |
2797 | * @dev: drm device | |
2798 | * @data: ioctl data copied from userspace (unvalidated) | |
2799 | * @file: drm file | |
2800 | * | |
2801 | * Validates the stream open parameters given by userspace including flags | |
2802 | * and an array of u64 key, value pair properties. | |
2803 | * | |
2804 | * Very little is assumed up front about the nature of the stream being | |
2805 | * opened (for instance we don't assume it's for periodic OA unit metrics). An | |
2806 | * i915-perf stream is expected to be a suitable interface for other forms of | |
2807 | * buffered data written by the GPU besides periodic OA metrics. | |
2808 | * | |
2809 | * Note we copy the properties from userspace outside of the i915 perf | |
2810 | * mutex to avoid an awkward lockdep with mmap_sem. | |
2811 | * | |
2812 | * Most of the implementation details are handled by | |
2813 | * i915_perf_open_ioctl_locked() after taking the &drm_i915_private->perf.lock | |
2814 | * mutex for serializing with any non-file-operation driver hooks. | |
2815 | * | |
2816 | * Return: A newly opened i915 Perf stream file descriptor or negative | |
2817 | * error code on failure. | |
2818 | */ | |
eec688e1 RB |
2819 | int i915_perf_open_ioctl(struct drm_device *dev, void *data, |
2820 | struct drm_file *file) | |
2821 | { | |
2822 | struct drm_i915_private *dev_priv = dev->dev_private; | |
2823 | struct drm_i915_perf_open_param *param = data; | |
2824 | struct perf_open_properties props; | |
2825 | u32 known_open_flags; | |
2826 | int ret; | |
2827 | ||
2828 | if (!dev_priv->perf.initialized) { | |
7708550c | 2829 | DRM_DEBUG("i915 perf interface not available for this system\n"); |
eec688e1 RB |
2830 | return -ENOTSUPP; |
2831 | } | |
2832 | ||
2833 | known_open_flags = I915_PERF_FLAG_FD_CLOEXEC | | |
2834 | I915_PERF_FLAG_FD_NONBLOCK | | |
2835 | I915_PERF_FLAG_DISABLED; | |
2836 | if (param->flags & ~known_open_flags) { | |
7708550c | 2837 | DRM_DEBUG("Unknown drm_i915_perf_open_param flag\n"); |
eec688e1 RB |
2838 | return -EINVAL; |
2839 | } | |
2840 | ||
2841 | ret = read_properties_unlocked(dev_priv, | |
2842 | u64_to_user_ptr(param->properties_ptr), | |
2843 | param->num_properties, | |
2844 | &props); | |
2845 | if (ret) | |
2846 | return ret; | |
2847 | ||
2848 | mutex_lock(&dev_priv->perf.lock); | |
2849 | ret = i915_perf_open_ioctl_locked(dev_priv, param, &props, file); | |
2850 | mutex_unlock(&dev_priv->perf.lock); | |
2851 | ||
2852 | return ret; | |
2853 | } | |
2854 | ||
16d98b31 RB |
2855 | /** |
2856 | * i915_perf_register - exposes i915-perf to userspace | |
2857 | * @dev_priv: i915 device instance | |
2858 | * | |
2859 | * In particular OA metric sets are advertised under a sysfs metrics/ | |
2860 | * directory allowing userspace to enumerate valid IDs that can be | |
2861 | * used to open an i915-perf stream. | |
2862 | */ | |
442b8c06 RB |
2863 | void i915_perf_register(struct drm_i915_private *dev_priv) |
2864 | { | |
701f8231 LL |
2865 | int ret; |
2866 | ||
442b8c06 RB |
2867 | if (!dev_priv->perf.initialized) |
2868 | return; | |
2869 | ||
2870 | /* To be sure we're synchronized with an attempted | |
2871 | * i915_perf_open_ioctl(); considering that we register after | |
2872 | * being exposed to userspace. | |
2873 | */ | |
2874 | mutex_lock(&dev_priv->perf.lock); | |
2875 | ||
2876 | dev_priv->perf.metrics_kobj = | |
2877 | kobject_create_and_add("metrics", | |
2878 | &dev_priv->drm.primary->kdev->kobj); | |
2879 | if (!dev_priv->perf.metrics_kobj) | |
2880 | goto exit; | |
2881 | ||
40f75ea4 | 2882 | sysfs_attr_init(&dev_priv->perf.oa.test_config.sysfs_metric_id.attr); |
701f8231 | 2883 | |
2dd24a9c | 2884 | if (INTEL_GEN(dev_priv) >= 11) { |
993298af RV |
2885 | i915_perf_load_test_config_icl(dev_priv); |
2886 | } else if (IS_CANNONLAKE(dev_priv)) { | |
2887 | i915_perf_load_test_config_cnl(dev_priv); | |
2888 | } else if (IS_COFFEELAKE(dev_priv)) { | |
2889 | if (IS_CFL_GT2(dev_priv)) | |
2890 | i915_perf_load_test_config_cflgt2(dev_priv); | |
2891 | if (IS_CFL_GT3(dev_priv)) | |
2892 | i915_perf_load_test_config_cflgt3(dev_priv); | |
2893 | } else if (IS_GEMINILAKE(dev_priv)) { | |
2894 | i915_perf_load_test_config_glk(dev_priv); | |
2895 | } else if (IS_KABYLAKE(dev_priv)) { | |
2896 | if (IS_KBL_GT2(dev_priv)) | |
2897 | i915_perf_load_test_config_kblgt2(dev_priv); | |
2898 | else if (IS_KBL_GT3(dev_priv)) | |
2899 | i915_perf_load_test_config_kblgt3(dev_priv); | |
2900 | } else if (IS_BROXTON(dev_priv)) { | |
2901 | i915_perf_load_test_config_bxt(dev_priv); | |
19f81df2 | 2902 | } else if (IS_SKYLAKE(dev_priv)) { |
701f8231 LL |
2903 | if (IS_SKL_GT2(dev_priv)) |
2904 | i915_perf_load_test_config_sklgt2(dev_priv); | |
2905 | else if (IS_SKL_GT3(dev_priv)) | |
2906 | i915_perf_load_test_config_sklgt3(dev_priv); | |
2907 | else if (IS_SKL_GT4(dev_priv)) | |
2908 | i915_perf_load_test_config_sklgt4(dev_priv); | |
993298af RV |
2909 | } else if (IS_CHERRYVIEW(dev_priv)) { |
2910 | i915_perf_load_test_config_chv(dev_priv); | |
2911 | } else if (IS_BROADWELL(dev_priv)) { | |
2912 | i915_perf_load_test_config_bdw(dev_priv); | |
2913 | } else if (IS_HASWELL(dev_priv)) { | |
2914 | i915_perf_load_test_config_hsw(dev_priv); | |
2915 | } | |
442b8c06 | 2916 | |
701f8231 LL |
2917 | if (dev_priv->perf.oa.test_config.id == 0) |
2918 | goto sysfs_error; | |
2919 | ||
2920 | ret = sysfs_create_group(dev_priv->perf.metrics_kobj, | |
2921 | &dev_priv->perf.oa.test_config.sysfs_metric); | |
2922 | if (ret) | |
2923 | goto sysfs_error; | |
f89823c2 LL |
2924 | |
2925 | atomic_set(&dev_priv->perf.oa.test_config.ref_count, 1); | |
2926 | ||
19f81df2 RB |
2927 | goto exit; |
2928 | ||
2929 | sysfs_error: | |
2930 | kobject_put(dev_priv->perf.metrics_kobj); | |
2931 | dev_priv->perf.metrics_kobj = NULL; | |
2932 | ||
442b8c06 RB |
2933 | exit: |
2934 | mutex_unlock(&dev_priv->perf.lock); | |
2935 | } | |
2936 | ||
16d98b31 RB |
2937 | /** |
2938 | * i915_perf_unregister - hide i915-perf from userspace | |
2939 | * @dev_priv: i915 device instance | |
2940 | * | |
2941 | * i915-perf state cleanup is split up into an 'unregister' and | |
2942 | * 'deinit' phase where the interface is first hidden from | |
2943 | * userspace by i915_perf_unregister() before cleaning up | |
2944 | * remaining state in i915_perf_fini(). | |
2945 | */ | |
442b8c06 RB |
2946 | void i915_perf_unregister(struct drm_i915_private *dev_priv) |
2947 | { | |
442b8c06 RB |
2948 | if (!dev_priv->perf.metrics_kobj) |
2949 | return; | |
2950 | ||
701f8231 LL |
2951 | sysfs_remove_group(dev_priv->perf.metrics_kobj, |
2952 | &dev_priv->perf.oa.test_config.sysfs_metric); | |
442b8c06 RB |
2953 | |
2954 | kobject_put(dev_priv->perf.metrics_kobj); | |
2955 | dev_priv->perf.metrics_kobj = NULL; | |
2956 | } | |
2957 | ||
f89823c2 LL |
2958 | static bool gen8_is_valid_flex_addr(struct drm_i915_private *dev_priv, u32 addr) |
2959 | { | |
2960 | static const i915_reg_t flex_eu_regs[] = { | |
2961 | EU_PERF_CNTL0, | |
2962 | EU_PERF_CNTL1, | |
2963 | EU_PERF_CNTL2, | |
2964 | EU_PERF_CNTL3, | |
2965 | EU_PERF_CNTL4, | |
2966 | EU_PERF_CNTL5, | |
2967 | EU_PERF_CNTL6, | |
2968 | }; | |
2969 | int i; | |
2970 | ||
2971 | for (i = 0; i < ARRAY_SIZE(flex_eu_regs); i++) { | |
7c52a221 | 2972 | if (i915_mmio_reg_offset(flex_eu_regs[i]) == addr) |
f89823c2 LL |
2973 | return true; |
2974 | } | |
2975 | return false; | |
2976 | } | |
2977 | ||
2978 | static bool gen7_is_valid_b_counter_addr(struct drm_i915_private *dev_priv, u32 addr) | |
2979 | { | |
7c52a221 LL |
2980 | return (addr >= i915_mmio_reg_offset(OASTARTTRIG1) && |
2981 | addr <= i915_mmio_reg_offset(OASTARTTRIG8)) || | |
2982 | (addr >= i915_mmio_reg_offset(OAREPORTTRIG1) && | |
2983 | addr <= i915_mmio_reg_offset(OAREPORTTRIG8)) || | |
2984 | (addr >= i915_mmio_reg_offset(OACEC0_0) && | |
2985 | addr <= i915_mmio_reg_offset(OACEC7_1)); | |
f89823c2 LL |
2986 | } |
2987 | ||
2988 | static bool gen7_is_valid_mux_addr(struct drm_i915_private *dev_priv, u32 addr) | |
2989 | { | |
7c52a221 LL |
2990 | return addr == i915_mmio_reg_offset(HALF_SLICE_CHICKEN2) || |
2991 | (addr >= i915_mmio_reg_offset(MICRO_BP0_0) && | |
2992 | addr <= i915_mmio_reg_offset(NOA_WRITE)) || | |
2993 | (addr >= i915_mmio_reg_offset(OA_PERFCNT1_LO) && | |
2994 | addr <= i915_mmio_reg_offset(OA_PERFCNT2_HI)) || | |
2995 | (addr >= i915_mmio_reg_offset(OA_PERFMATRIX_LO) && | |
2996 | addr <= i915_mmio_reg_offset(OA_PERFMATRIX_HI)); | |
f89823c2 LL |
2997 | } |
2998 | ||
2999 | static bool gen8_is_valid_mux_addr(struct drm_i915_private *dev_priv, u32 addr) | |
3000 | { | |
3001 | return gen7_is_valid_mux_addr(dev_priv, addr) || | |
7c52a221 LL |
3002 | addr == i915_mmio_reg_offset(WAIT_FOR_RC6_EXIT) || |
3003 | (addr >= i915_mmio_reg_offset(RPM_CONFIG0) && | |
3004 | addr <= i915_mmio_reg_offset(NOA_CONFIG(8))); | |
f89823c2 LL |
3005 | } |
3006 | ||
95690a02 LL |
3007 | static bool gen10_is_valid_mux_addr(struct drm_i915_private *dev_priv, u32 addr) |
3008 | { | |
3009 | return gen8_is_valid_mux_addr(dev_priv, addr) || | |
7c52a221 LL |
3010 | (addr >= i915_mmio_reg_offset(OA_PERFCNT3_LO) && |
3011 | addr <= i915_mmio_reg_offset(OA_PERFCNT4_HI)); | |
95690a02 LL |
3012 | } |
3013 | ||
f89823c2 LL |
3014 | static bool hsw_is_valid_mux_addr(struct drm_i915_private *dev_priv, u32 addr) |
3015 | { | |
3016 | return gen7_is_valid_mux_addr(dev_priv, addr) || | |
3017 | (addr >= 0x25100 && addr <= 0x2FF90) || | |
7c52a221 LL |
3018 | (addr >= i915_mmio_reg_offset(HSW_MBVID2_NOA0) && |
3019 | addr <= i915_mmio_reg_offset(HSW_MBVID2_NOA9)) || | |
3020 | addr == i915_mmio_reg_offset(HSW_MBVID2_MISR0); | |
f89823c2 LL |
3021 | } |
3022 | ||
3023 | static bool chv_is_valid_mux_addr(struct drm_i915_private *dev_priv, u32 addr) | |
3024 | { | |
3025 | return gen7_is_valid_mux_addr(dev_priv, addr) || | |
3026 | (addr >= 0x182300 && addr <= 0x1823A4); | |
3027 | } | |
3028 | ||
739f3abd | 3029 | static u32 mask_reg_value(u32 reg, u32 val) |
f89823c2 LL |
3030 | { |
3031 | /* HALF_SLICE_CHICKEN2 is programmed with a the | |
3032 | * WaDisableSTUnitPowerOptimization workaround. Make sure the value | |
3033 | * programmed by userspace doesn't change this. | |
3034 | */ | |
7c52a221 | 3035 | if (i915_mmio_reg_offset(HALF_SLICE_CHICKEN2) == reg) |
f89823c2 LL |
3036 | val = val & ~_MASKED_BIT_ENABLE(GEN8_ST_PO_DISABLE); |
3037 | ||
3038 | /* WAIT_FOR_RC6_EXIT has only one bit fullfilling the function | |
3039 | * indicated by its name and a bunch of selection fields used by OA | |
3040 | * configs. | |
3041 | */ | |
7c52a221 | 3042 | if (i915_mmio_reg_offset(WAIT_FOR_RC6_EXIT) == reg) |
f89823c2 LL |
3043 | val = val & ~_MASKED_BIT_ENABLE(HSW_WAIT_FOR_RC6_EXIT_ENABLE); |
3044 | ||
3045 | return val; | |
3046 | } | |
3047 | ||
3048 | static struct i915_oa_reg *alloc_oa_regs(struct drm_i915_private *dev_priv, | |
3049 | bool (*is_valid)(struct drm_i915_private *dev_priv, u32 addr), | |
3050 | u32 __user *regs, | |
3051 | u32 n_regs) | |
3052 | { | |
3053 | struct i915_oa_reg *oa_regs; | |
3054 | int err; | |
3055 | u32 i; | |
3056 | ||
3057 | if (!n_regs) | |
3058 | return NULL; | |
3059 | ||
96d4f267 | 3060 | if (!access_ok(regs, n_regs * sizeof(u32) * 2)) |
f89823c2 LL |
3061 | return ERR_PTR(-EFAULT); |
3062 | ||
3063 | /* No is_valid function means we're not allowing any register to be programmed. */ | |
3064 | GEM_BUG_ON(!is_valid); | |
3065 | if (!is_valid) | |
3066 | return ERR_PTR(-EINVAL); | |
3067 | ||
3068 | oa_regs = kmalloc_array(n_regs, sizeof(*oa_regs), GFP_KERNEL); | |
3069 | if (!oa_regs) | |
3070 | return ERR_PTR(-ENOMEM); | |
3071 | ||
3072 | for (i = 0; i < n_regs; i++) { | |
3073 | u32 addr, value; | |
3074 | ||
3075 | err = get_user(addr, regs); | |
3076 | if (err) | |
3077 | goto addr_err; | |
3078 | ||
3079 | if (!is_valid(dev_priv, addr)) { | |
3080 | DRM_DEBUG("Invalid oa_reg address: %X\n", addr); | |
3081 | err = -EINVAL; | |
3082 | goto addr_err; | |
3083 | } | |
3084 | ||
3085 | err = get_user(value, regs + 1); | |
3086 | if (err) | |
3087 | goto addr_err; | |
3088 | ||
3089 | oa_regs[i].addr = _MMIO(addr); | |
3090 | oa_regs[i].value = mask_reg_value(addr, value); | |
3091 | ||
3092 | regs += 2; | |
3093 | } | |
3094 | ||
3095 | return oa_regs; | |
3096 | ||
3097 | addr_err: | |
3098 | kfree(oa_regs); | |
3099 | return ERR_PTR(err); | |
3100 | } | |
3101 | ||
3102 | static ssize_t show_dynamic_id(struct device *dev, | |
3103 | struct device_attribute *attr, | |
3104 | char *buf) | |
3105 | { | |
3106 | struct i915_oa_config *oa_config = | |
3107 | container_of(attr, typeof(*oa_config), sysfs_metric_id); | |
3108 | ||
3109 | return sprintf(buf, "%d\n", oa_config->id); | |
3110 | } | |
3111 | ||
3112 | static int create_dynamic_oa_sysfs_entry(struct drm_i915_private *dev_priv, | |
3113 | struct i915_oa_config *oa_config) | |
3114 | { | |
28152a23 | 3115 | sysfs_attr_init(&oa_config->sysfs_metric_id.attr); |
f89823c2 LL |
3116 | oa_config->sysfs_metric_id.attr.name = "id"; |
3117 | oa_config->sysfs_metric_id.attr.mode = S_IRUGO; | |
3118 | oa_config->sysfs_metric_id.show = show_dynamic_id; | |
3119 | oa_config->sysfs_metric_id.store = NULL; | |
3120 | ||
3121 | oa_config->attrs[0] = &oa_config->sysfs_metric_id.attr; | |
3122 | oa_config->attrs[1] = NULL; | |
3123 | ||
3124 | oa_config->sysfs_metric.name = oa_config->uuid; | |
3125 | oa_config->sysfs_metric.attrs = oa_config->attrs; | |
3126 | ||
3127 | return sysfs_create_group(dev_priv->perf.metrics_kobj, | |
3128 | &oa_config->sysfs_metric); | |
3129 | } | |
3130 | ||
3131 | /** | |
3132 | * i915_perf_add_config_ioctl - DRM ioctl() for userspace to add a new OA config | |
3133 | * @dev: drm device | |
3134 | * @data: ioctl data (pointer to struct drm_i915_perf_oa_config) copied from | |
3135 | * userspace (unvalidated) | |
3136 | * @file: drm file | |
3137 | * | |
3138 | * Validates the submitted OA register to be saved into a new OA config that | |
3139 | * can then be used for programming the OA unit and its NOA network. | |
3140 | * | |
3141 | * Returns: A new allocated config number to be used with the perf open ioctl | |
3142 | * or a negative error code on failure. | |
3143 | */ | |
3144 | int i915_perf_add_config_ioctl(struct drm_device *dev, void *data, | |
3145 | struct drm_file *file) | |
3146 | { | |
3147 | struct drm_i915_private *dev_priv = dev->dev_private; | |
3148 | struct drm_i915_perf_oa_config *args = data; | |
3149 | struct i915_oa_config *oa_config, *tmp; | |
3150 | int err, id; | |
3151 | ||
3152 | if (!dev_priv->perf.initialized) { | |
3153 | DRM_DEBUG("i915 perf interface not available for this system\n"); | |
3154 | return -ENOTSUPP; | |
3155 | } | |
3156 | ||
3157 | if (!dev_priv->perf.metrics_kobj) { | |
3158 | DRM_DEBUG("OA metrics weren't advertised via sysfs\n"); | |
3159 | return -EINVAL; | |
3160 | } | |
3161 | ||
3162 | if (i915_perf_stream_paranoid && !capable(CAP_SYS_ADMIN)) { | |
3163 | DRM_DEBUG("Insufficient privileges to add i915 OA config\n"); | |
3164 | return -EACCES; | |
3165 | } | |
3166 | ||
3167 | if ((!args->mux_regs_ptr || !args->n_mux_regs) && | |
3168 | (!args->boolean_regs_ptr || !args->n_boolean_regs) && | |
3169 | (!args->flex_regs_ptr || !args->n_flex_regs)) { | |
3170 | DRM_DEBUG("No OA registers given\n"); | |
3171 | return -EINVAL; | |
3172 | } | |
3173 | ||
3174 | oa_config = kzalloc(sizeof(*oa_config), GFP_KERNEL); | |
3175 | if (!oa_config) { | |
3176 | DRM_DEBUG("Failed to allocate memory for the OA config\n"); | |
3177 | return -ENOMEM; | |
3178 | } | |
3179 | ||
3180 | atomic_set(&oa_config->ref_count, 1); | |
3181 | ||
3182 | if (!uuid_is_valid(args->uuid)) { | |
3183 | DRM_DEBUG("Invalid uuid format for OA config\n"); | |
3184 | err = -EINVAL; | |
3185 | goto reg_err; | |
3186 | } | |
3187 | ||
3188 | /* Last character in oa_config->uuid will be 0 because oa_config is | |
3189 | * kzalloc. | |
3190 | */ | |
3191 | memcpy(oa_config->uuid, args->uuid, sizeof(args->uuid)); | |
3192 | ||
3193 | oa_config->mux_regs_len = args->n_mux_regs; | |
3194 | oa_config->mux_regs = | |
3195 | alloc_oa_regs(dev_priv, | |
3196 | dev_priv->perf.oa.ops.is_valid_mux_reg, | |
3197 | u64_to_user_ptr(args->mux_regs_ptr), | |
3198 | args->n_mux_regs); | |
3199 | ||
3200 | if (IS_ERR(oa_config->mux_regs)) { | |
3201 | DRM_DEBUG("Failed to create OA config for mux_regs\n"); | |
3202 | err = PTR_ERR(oa_config->mux_regs); | |
3203 | goto reg_err; | |
3204 | } | |
3205 | ||
3206 | oa_config->b_counter_regs_len = args->n_boolean_regs; | |
3207 | oa_config->b_counter_regs = | |
3208 | alloc_oa_regs(dev_priv, | |
3209 | dev_priv->perf.oa.ops.is_valid_b_counter_reg, | |
3210 | u64_to_user_ptr(args->boolean_regs_ptr), | |
3211 | args->n_boolean_regs); | |
3212 | ||
3213 | if (IS_ERR(oa_config->b_counter_regs)) { | |
3214 | DRM_DEBUG("Failed to create OA config for b_counter_regs\n"); | |
3215 | err = PTR_ERR(oa_config->b_counter_regs); | |
3216 | goto reg_err; | |
3217 | } | |
3218 | ||
3219 | if (INTEL_GEN(dev_priv) < 8) { | |
3220 | if (args->n_flex_regs != 0) { | |
3221 | err = -EINVAL; | |
3222 | goto reg_err; | |
3223 | } | |
3224 | } else { | |
3225 | oa_config->flex_regs_len = args->n_flex_regs; | |
3226 | oa_config->flex_regs = | |
3227 | alloc_oa_regs(dev_priv, | |
3228 | dev_priv->perf.oa.ops.is_valid_flex_reg, | |
3229 | u64_to_user_ptr(args->flex_regs_ptr), | |
3230 | args->n_flex_regs); | |
3231 | ||
3232 | if (IS_ERR(oa_config->flex_regs)) { | |
3233 | DRM_DEBUG("Failed to create OA config for flex_regs\n"); | |
3234 | err = PTR_ERR(oa_config->flex_regs); | |
3235 | goto reg_err; | |
3236 | } | |
3237 | } | |
3238 | ||
3239 | err = mutex_lock_interruptible(&dev_priv->perf.metrics_lock); | |
3240 | if (err) | |
3241 | goto reg_err; | |
3242 | ||
3243 | /* We shouldn't have too many configs, so this iteration shouldn't be | |
3244 | * too costly. | |
3245 | */ | |
3246 | idr_for_each_entry(&dev_priv->perf.metrics_idr, tmp, id) { | |
3247 | if (!strcmp(tmp->uuid, oa_config->uuid)) { | |
3248 | DRM_DEBUG("OA config already exists with this uuid\n"); | |
3249 | err = -EADDRINUSE; | |
3250 | goto sysfs_err; | |
3251 | } | |
3252 | } | |
3253 | ||
3254 | err = create_dynamic_oa_sysfs_entry(dev_priv, oa_config); | |
3255 | if (err) { | |
3256 | DRM_DEBUG("Failed to create sysfs entry for OA config\n"); | |
3257 | goto sysfs_err; | |
3258 | } | |
3259 | ||
3260 | /* Config id 0 is invalid, id 1 for kernel stored test config. */ | |
3261 | oa_config->id = idr_alloc(&dev_priv->perf.metrics_idr, | |
3262 | oa_config, 2, | |
3263 | 0, GFP_KERNEL); | |
3264 | if (oa_config->id < 0) { | |
3265 | DRM_DEBUG("Failed to create sysfs entry for OA config\n"); | |
3266 | err = oa_config->id; | |
3267 | goto sysfs_err; | |
3268 | } | |
3269 | ||
3270 | mutex_unlock(&dev_priv->perf.metrics_lock); | |
3271 | ||
9bd9be66 LL |
3272 | DRM_DEBUG("Added config %s id=%i\n", oa_config->uuid, oa_config->id); |
3273 | ||
f89823c2 LL |
3274 | return oa_config->id; |
3275 | ||
3276 | sysfs_err: | |
3277 | mutex_unlock(&dev_priv->perf.metrics_lock); | |
3278 | reg_err: | |
3279 | put_oa_config(dev_priv, oa_config); | |
3280 | DRM_DEBUG("Failed to add new OA config\n"); | |
3281 | return err; | |
3282 | } | |
3283 | ||
3284 | /** | |
3285 | * i915_perf_remove_config_ioctl - DRM ioctl() for userspace to remove an OA config | |
3286 | * @dev: drm device | |
3287 | * @data: ioctl data (pointer to u64 integer) copied from userspace | |
3288 | * @file: drm file | |
3289 | * | |
3290 | * Configs can be removed while being used, the will stop appearing in sysfs | |
3291 | * and their content will be freed when the stream using the config is closed. | |
3292 | * | |
3293 | * Returns: 0 on success or a negative error code on failure. | |
3294 | */ | |
3295 | int i915_perf_remove_config_ioctl(struct drm_device *dev, void *data, | |
3296 | struct drm_file *file) | |
3297 | { | |
3298 | struct drm_i915_private *dev_priv = dev->dev_private; | |
3299 | u64 *arg = data; | |
3300 | struct i915_oa_config *oa_config; | |
3301 | int ret; | |
3302 | ||
3303 | if (!dev_priv->perf.initialized) { | |
3304 | DRM_DEBUG("i915 perf interface not available for this system\n"); | |
3305 | return -ENOTSUPP; | |
3306 | } | |
3307 | ||
3308 | if (i915_perf_stream_paranoid && !capable(CAP_SYS_ADMIN)) { | |
3309 | DRM_DEBUG("Insufficient privileges to remove i915 OA config\n"); | |
3310 | return -EACCES; | |
3311 | } | |
3312 | ||
3313 | ret = mutex_lock_interruptible(&dev_priv->perf.metrics_lock); | |
3314 | if (ret) | |
3315 | goto lock_err; | |
3316 | ||
3317 | oa_config = idr_find(&dev_priv->perf.metrics_idr, *arg); | |
3318 | if (!oa_config) { | |
3319 | DRM_DEBUG("Failed to remove unknown OA config\n"); | |
3320 | ret = -ENOENT; | |
3321 | goto config_err; | |
3322 | } | |
3323 | ||
3324 | GEM_BUG_ON(*arg != oa_config->id); | |
3325 | ||
3326 | sysfs_remove_group(dev_priv->perf.metrics_kobj, | |
3327 | &oa_config->sysfs_metric); | |
3328 | ||
3329 | idr_remove(&dev_priv->perf.metrics_idr, *arg); | |
9bd9be66 LL |
3330 | |
3331 | DRM_DEBUG("Removed config %s id=%i\n", oa_config->uuid, oa_config->id); | |
3332 | ||
f89823c2 LL |
3333 | put_oa_config(dev_priv, oa_config); |
3334 | ||
3335 | config_err: | |
3336 | mutex_unlock(&dev_priv->perf.metrics_lock); | |
3337 | lock_err: | |
3338 | return ret; | |
3339 | } | |
3340 | ||
ccdf6341 RB |
3341 | static struct ctl_table oa_table[] = { |
3342 | { | |
3343 | .procname = "perf_stream_paranoid", | |
3344 | .data = &i915_perf_stream_paranoid, | |
3345 | .maxlen = sizeof(i915_perf_stream_paranoid), | |
3346 | .mode = 0644, | |
3347 | .proc_handler = proc_dointvec_minmax, | |
3348 | .extra1 = &zero, | |
3349 | .extra2 = &one, | |
3350 | }, | |
00319ba0 RB |
3351 | { |
3352 | .procname = "oa_max_sample_rate", | |
3353 | .data = &i915_oa_max_sample_rate, | |
3354 | .maxlen = sizeof(i915_oa_max_sample_rate), | |
3355 | .mode = 0644, | |
3356 | .proc_handler = proc_dointvec_minmax, | |
3357 | .extra1 = &zero, | |
3358 | .extra2 = &oa_sample_rate_hard_limit, | |
3359 | }, | |
ccdf6341 RB |
3360 | {} |
3361 | }; | |
3362 | ||
3363 | static struct ctl_table i915_root[] = { | |
3364 | { | |
3365 | .procname = "i915", | |
3366 | .maxlen = 0, | |
3367 | .mode = 0555, | |
3368 | .child = oa_table, | |
3369 | }, | |
3370 | {} | |
3371 | }; | |
3372 | ||
3373 | static struct ctl_table dev_root[] = { | |
3374 | { | |
3375 | .procname = "dev", | |
3376 | .maxlen = 0, | |
3377 | .mode = 0555, | |
3378 | .child = i915_root, | |
3379 | }, | |
3380 | {} | |
3381 | }; | |
3382 | ||
16d98b31 RB |
3383 | /** |
3384 | * i915_perf_init - initialize i915-perf state on module load | |
3385 | * @dev_priv: i915 device instance | |
3386 | * | |
3387 | * Initializes i915-perf state without exposing anything to userspace. | |
3388 | * | |
3389 | * Note: i915-perf initialization is split into an 'init' and 'register' | |
3390 | * phase with the i915_perf_register() exposing state to userspace. | |
3391 | */ | |
eec688e1 RB |
3392 | void i915_perf_init(struct drm_i915_private *dev_priv) |
3393 | { | |
19f81df2 | 3394 | if (IS_HASWELL(dev_priv)) { |
f89823c2 LL |
3395 | dev_priv->perf.oa.ops.is_valid_b_counter_reg = |
3396 | gen7_is_valid_b_counter_addr; | |
3397 | dev_priv->perf.oa.ops.is_valid_mux_reg = | |
3398 | hsw_is_valid_mux_addr; | |
3399 | dev_priv->perf.oa.ops.is_valid_flex_reg = NULL; | |
19f81df2 RB |
3400 | dev_priv->perf.oa.ops.enable_metric_set = hsw_enable_metric_set; |
3401 | dev_priv->perf.oa.ops.disable_metric_set = hsw_disable_metric_set; | |
3402 | dev_priv->perf.oa.ops.oa_enable = gen7_oa_enable; | |
3403 | dev_priv->perf.oa.ops.oa_disable = gen7_oa_disable; | |
3404 | dev_priv->perf.oa.ops.read = gen7_oa_read; | |
3405 | dev_priv->perf.oa.ops.oa_hw_tail_read = | |
3406 | gen7_oa_hw_tail_read; | |
3407 | ||
3408 | dev_priv->perf.oa.oa_formats = hsw_oa_formats; | |
fb5c551a | 3409 | } else if (HAS_LOGICAL_RING_CONTEXTS(dev_priv)) { |
19f81df2 RB |
3410 | /* Note: that although we could theoretically also support the |
3411 | * legacy ringbuffer mode on BDW (and earlier iterations of | |
3412 | * this driver, before upstreaming did this) it didn't seem | |
3413 | * worth the complexity to maintain now that BDW+ enable | |
3414 | * execlist mode by default. | |
3415 | */ | |
ba6b7c1a | 3416 | dev_priv->perf.oa.oa_formats = gen8_plus_oa_formats; |
d7965152 | 3417 | |
701f8231 LL |
3418 | dev_priv->perf.oa.ops.oa_enable = gen8_oa_enable; |
3419 | dev_priv->perf.oa.ops.oa_disable = gen8_oa_disable; | |
3420 | dev_priv->perf.oa.ops.read = gen8_oa_read; | |
3421 | dev_priv->perf.oa.ops.oa_hw_tail_read = gen8_oa_hw_tail_read; | |
3422 | ||
f3ce44a0 | 3423 | if (IS_GEN_RANGE(dev_priv, 8, 9)) { |
ba6b7c1a LL |
3424 | dev_priv->perf.oa.ops.is_valid_b_counter_reg = |
3425 | gen7_is_valid_b_counter_addr; | |
3426 | dev_priv->perf.oa.ops.is_valid_mux_reg = | |
3427 | gen8_is_valid_mux_addr; | |
3428 | dev_priv->perf.oa.ops.is_valid_flex_reg = | |
3429 | gen8_is_valid_flex_addr; | |
155e941f | 3430 | |
f89823c2 LL |
3431 | if (IS_CHERRYVIEW(dev_priv)) { |
3432 | dev_priv->perf.oa.ops.is_valid_mux_reg = | |
3433 | chv_is_valid_mux_addr; | |
3434 | } | |
155e941f | 3435 | |
ba6b7c1a LL |
3436 | dev_priv->perf.oa.ops.enable_metric_set = gen8_enable_metric_set; |
3437 | dev_priv->perf.oa.ops.disable_metric_set = gen8_disable_metric_set; | |
3438 | ||
cf819eff | 3439 | if (IS_GEN(dev_priv, 8)) { |
ba6b7c1a LL |
3440 | dev_priv->perf.oa.ctx_oactxctrl_offset = 0x120; |
3441 | dev_priv->perf.oa.ctx_flexeu0_offset = 0x2ce; | |
3442 | ||
3443 | dev_priv->perf.oa.gen8_valid_ctx_bit = (1<<25); | |
3444 | } else { | |
3445 | dev_priv->perf.oa.ctx_oactxctrl_offset = 0x128; | |
3446 | dev_priv->perf.oa.ctx_flexeu0_offset = 0x3de; | |
3447 | ||
3448 | dev_priv->perf.oa.gen8_valid_ctx_bit = (1<<16); | |
3449 | } | |
00690008 | 3450 | } else if (IS_GEN_RANGE(dev_priv, 10, 11)) { |
95690a02 LL |
3451 | dev_priv->perf.oa.ops.is_valid_b_counter_reg = |
3452 | gen7_is_valid_b_counter_addr; | |
3453 | dev_priv->perf.oa.ops.is_valid_mux_reg = | |
3454 | gen10_is_valid_mux_addr; | |
3455 | dev_priv->perf.oa.ops.is_valid_flex_reg = | |
3456 | gen8_is_valid_flex_addr; | |
3457 | ||
3458 | dev_priv->perf.oa.ops.enable_metric_set = gen8_enable_metric_set; | |
3459 | dev_priv->perf.oa.ops.disable_metric_set = gen10_disable_metric_set; | |
3460 | ||
3461 | dev_priv->perf.oa.ctx_oactxctrl_offset = 0x128; | |
3462 | dev_priv->perf.oa.ctx_flexeu0_offset = 0x3de; | |
3463 | ||
3464 | dev_priv->perf.oa.gen8_valid_ctx_bit = (1<<16); | |
19f81df2 | 3465 | } |
19f81df2 | 3466 | } |
d7965152 | 3467 | |
9f9b2792 | 3468 | if (dev_priv->perf.oa.ops.enable_metric_set) { |
19f81df2 RB |
3469 | hrtimer_init(&dev_priv->perf.oa.poll_check_timer, |
3470 | CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
3471 | dev_priv->perf.oa.poll_check_timer.function = oa_poll_check_timer_cb; | |
3472 | init_waitqueue_head(&dev_priv->perf.oa.poll_wq); | |
d7965152 | 3473 | |
19f81df2 RB |
3474 | INIT_LIST_HEAD(&dev_priv->perf.streams); |
3475 | mutex_init(&dev_priv->perf.lock); | |
19f81df2 | 3476 | spin_lock_init(&dev_priv->perf.oa.oa_buffer.ptr_lock); |
eec688e1 | 3477 | |
9f9b2792 | 3478 | oa_sample_rate_hard_limit = 1000 * |
0258404f | 3479 | (RUNTIME_INFO(dev_priv)->cs_timestamp_frequency_khz / 2); |
19f81df2 | 3480 | dev_priv->perf.sysctl_header = register_sysctl_table(dev_root); |
ccdf6341 | 3481 | |
f89823c2 LL |
3482 | mutex_init(&dev_priv->perf.metrics_lock); |
3483 | idr_init(&dev_priv->perf.metrics_idr); | |
3484 | ||
19f81df2 RB |
3485 | dev_priv->perf.initialized = true; |
3486 | } | |
eec688e1 RB |
3487 | } |
3488 | ||
f89823c2 LL |
3489 | static int destroy_config(int id, void *p, void *data) |
3490 | { | |
3491 | struct drm_i915_private *dev_priv = data; | |
3492 | struct i915_oa_config *oa_config = p; | |
3493 | ||
3494 | put_oa_config(dev_priv, oa_config); | |
3495 | ||
3496 | return 0; | |
3497 | } | |
3498 | ||
16d98b31 RB |
3499 | /** |
3500 | * i915_perf_fini - Counter part to i915_perf_init() | |
3501 | * @dev_priv: i915 device instance | |
3502 | */ | |
eec688e1 RB |
3503 | void i915_perf_fini(struct drm_i915_private *dev_priv) |
3504 | { | |
3505 | if (!dev_priv->perf.initialized) | |
3506 | return; | |
3507 | ||
f89823c2 LL |
3508 | idr_for_each(&dev_priv->perf.metrics_idr, destroy_config, dev_priv); |
3509 | idr_destroy(&dev_priv->perf.metrics_idr); | |
3510 | ||
ccdf6341 RB |
3511 | unregister_sysctl_table(dev_priv->perf.sysctl_header); |
3512 | ||
d7965152 | 3513 | memset(&dev_priv->perf.oa.ops, 0, sizeof(dev_priv->perf.oa.ops)); |
19f81df2 | 3514 | |
eec688e1 RB |
3515 | dev_priv->perf.initialized = false; |
3516 | } |