1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2019 Intel Corporation
4  */
5 
6 #include <linux/kobject.h>
7 #include <linux/sysfs.h>
8 
9 #include "i915_drv.h"
10 #include "intel_engine.h"
11 #include "intel_engine_heartbeat.h"
12 #include "sysfs_engines.h"
13 
14 struct kobj_engine {
15 	struct kobject base;
16 	struct intel_engine_cs *engine;
17 };
18 
kobj_to_engine(struct kobject * kobj)19 static struct intel_engine_cs *kobj_to_engine(struct kobject *kobj)
20 {
21 	return container_of(kobj, struct kobj_engine, base)->engine;
22 }
23 
24 static ssize_t
name_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)25 name_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
26 {
27 	return sprintf(buf, "%s\n", kobj_to_engine(kobj)->name);
28 }
29 
30 static struct kobj_attribute name_attr =
31 __ATTR(name, 0444, name_show, NULL);
32 
33 static ssize_t
class_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)34 class_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
35 {
36 	return sprintf(buf, "%d\n", kobj_to_engine(kobj)->uabi_class);
37 }
38 
39 static struct kobj_attribute class_attr =
40 __ATTR(class, 0444, class_show, NULL);
41 
42 static ssize_t
inst_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)43 inst_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
44 {
45 	return sprintf(buf, "%d\n", kobj_to_engine(kobj)->uabi_instance);
46 }
47 
48 static struct kobj_attribute inst_attr =
49 __ATTR(instance, 0444, inst_show, NULL);
50 
51 static ssize_t
mmio_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)52 mmio_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
53 {
54 	return sprintf(buf, "0x%x\n", kobj_to_engine(kobj)->mmio_base);
55 }
56 
57 static struct kobj_attribute mmio_attr =
58 __ATTR(mmio_base, 0444, mmio_show, NULL);
59 
60 static const char * const vcs_caps[] = {
61 	[ilog2(I915_VIDEO_CLASS_CAPABILITY_HEVC)] = "hevc",
62 	[ilog2(I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC)] = "sfc",
63 };
64 
65 static const char * const vecs_caps[] = {
66 	[ilog2(I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC)] = "sfc",
67 };
68 
repr_trim(char * buf,ssize_t len)69 static ssize_t repr_trim(char *buf, ssize_t len)
70 {
71 	/* Trim off the trailing space and replace with a newline */
72 	if (len > PAGE_SIZE)
73 		len = PAGE_SIZE;
74 	if (len > 0)
75 		buf[len - 1] = '\n';
76 
77 	return len;
78 }
79 
80 static ssize_t
__caps_show(struct intel_engine_cs * engine,unsigned long caps,char * buf,bool show_unknown)81 __caps_show(struct intel_engine_cs *engine,
82 	    unsigned long caps, char *buf, bool show_unknown)
83 {
84 	const char * const *repr;
85 	int count, n;
86 	ssize_t len;
87 
88 	switch (engine->class) {
89 	case VIDEO_DECODE_CLASS:
90 		repr = vcs_caps;
91 		count = ARRAY_SIZE(vcs_caps);
92 		break;
93 
94 	case VIDEO_ENHANCEMENT_CLASS:
95 		repr = vecs_caps;
96 		count = ARRAY_SIZE(vecs_caps);
97 		break;
98 
99 	default:
100 		repr = NULL;
101 		count = 0;
102 		break;
103 	}
104 	GEM_BUG_ON(count > BITS_PER_LONG);
105 
106 	len = 0;
107 	for_each_set_bit(n, &caps, show_unknown ? BITS_PER_LONG : count) {
108 		if (n >= count || !repr[n]) {
109 			if (GEM_WARN_ON(show_unknown))
110 				len += snprintf(buf + len, PAGE_SIZE - len,
111 						"[%x] ", n);
112 		} else {
113 			len += snprintf(buf + len, PAGE_SIZE - len,
114 					"%s ", repr[n]);
115 		}
116 		if (GEM_WARN_ON(len >= PAGE_SIZE))
117 			break;
118 	}
119 	return repr_trim(buf, len);
120 }
121 
122 static ssize_t
caps_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)123 caps_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
124 {
125 	struct intel_engine_cs *engine = kobj_to_engine(kobj);
126 
127 	return __caps_show(engine, engine->uabi_capabilities, buf, true);
128 }
129 
130 static struct kobj_attribute caps_attr =
131 __ATTR(capabilities, 0444, caps_show, NULL);
132 
133 static ssize_t
all_caps_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)134 all_caps_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
135 {
136 	return __caps_show(kobj_to_engine(kobj), -1, buf, false);
137 }
138 
139 static struct kobj_attribute all_caps_attr =
140 __ATTR(known_capabilities, 0444, all_caps_show, NULL);
141 
142 static ssize_t
max_spin_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)143 max_spin_store(struct kobject *kobj, struct kobj_attribute *attr,
144 	       const char *buf, size_t count)
145 {
146 	struct intel_engine_cs *engine = kobj_to_engine(kobj);
147 	unsigned long long duration;
148 	int err;
149 
150 	/*
151 	 * When waiting for a request, if is it currently being executed
152 	 * on the GPU, we busywait for a short while before sleeping. The
153 	 * premise is that most requests are short, and if it is already
154 	 * executing then there is a good chance that it will complete
155 	 * before we can setup the interrupt handler and go to sleep.
156 	 * We try to offset the cost of going to sleep, by first spinning
157 	 * on the request -- if it completed in less time than it would take
158 	 * to go sleep, process the interrupt and return back to the client,
159 	 * then we have saved the client some latency, albeit at the cost
160 	 * of spinning on an expensive CPU core.
161 	 *
162 	 * While we try to avoid waiting at all for a request that is unlikely
163 	 * to complete, deciding how long it is worth spinning is for is an
164 	 * arbitrary decision: trading off power vs latency.
165 	 */
166 
167 	err = kstrtoull(buf, 0, &duration);
168 	if (err)
169 		return err;
170 
171 	if (duration > jiffies_to_nsecs(2))
172 		return -EINVAL;
173 
174 	WRITE_ONCE(engine->props.max_busywait_duration_ns, duration);
175 
176 	return count;
177 }
178 
179 static ssize_t
max_spin_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)180 max_spin_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
181 {
182 	struct intel_engine_cs *engine = kobj_to_engine(kobj);
183 
184 	return sprintf(buf, "%lu\n", engine->props.max_busywait_duration_ns);
185 }
186 
187 static struct kobj_attribute max_spin_attr =
188 __ATTR(max_busywait_duration_ns, 0644, max_spin_show, max_spin_store);
189 
190 static ssize_t
max_spin_default(struct kobject * kobj,struct kobj_attribute * attr,char * buf)191 max_spin_default(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
192 {
193 	struct intel_engine_cs *engine = kobj_to_engine(kobj);
194 
195 	return sprintf(buf, "%lu\n", engine->defaults.max_busywait_duration_ns);
196 }
197 
198 static struct kobj_attribute max_spin_def =
199 __ATTR(max_busywait_duration_ns, 0444, max_spin_default, NULL);
200 
201 static ssize_t
timeslice_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)202 timeslice_store(struct kobject *kobj, struct kobj_attribute *attr,
203 		const char *buf, size_t count)
204 {
205 	struct intel_engine_cs *engine = kobj_to_engine(kobj);
206 	unsigned long long duration;
207 	int err;
208 
209 	/*
210 	 * Execlists uses a scheduling quantum (a timeslice) to alternate
211 	 * execution between ready-to-run contexts of equal priority. This
212 	 * ensures that all users (though only if they of equal importance)
213 	 * have the opportunity to run and prevents livelocks where contexts
214 	 * may have implicit ordering due to userspace semaphores.
215 	 */
216 
217 	err = kstrtoull(buf, 0, &duration);
218 	if (err)
219 		return err;
220 
221 	if (duration > jiffies_to_msecs(MAX_SCHEDULE_TIMEOUT))
222 		return -EINVAL;
223 
224 	WRITE_ONCE(engine->props.timeslice_duration_ms, duration);
225 
226 	if (execlists_active(&engine->execlists))
227 		set_timer_ms(&engine->execlists.timer, duration);
228 
229 	return count;
230 }
231 
232 static ssize_t
timeslice_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)233 timeslice_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
234 {
235 	struct intel_engine_cs *engine = kobj_to_engine(kobj);
236 
237 	return sprintf(buf, "%lu\n", engine->props.timeslice_duration_ms);
238 }
239 
240 static struct kobj_attribute timeslice_duration_attr =
241 __ATTR(timeslice_duration_ms, 0644, timeslice_show, timeslice_store);
242 
243 static ssize_t
timeslice_default(struct kobject * kobj,struct kobj_attribute * attr,char * buf)244 timeslice_default(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
245 {
246 	struct intel_engine_cs *engine = kobj_to_engine(kobj);
247 
248 	return sprintf(buf, "%lu\n", engine->defaults.timeslice_duration_ms);
249 }
250 
251 static struct kobj_attribute timeslice_duration_def =
252 __ATTR(timeslice_duration_ms, 0444, timeslice_default, NULL);
253 
254 static ssize_t
stop_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)255 stop_store(struct kobject *kobj, struct kobj_attribute *attr,
256 	   const char *buf, size_t count)
257 {
258 	struct intel_engine_cs *engine = kobj_to_engine(kobj);
259 	unsigned long long duration;
260 	int err;
261 
262 	/*
263 	 * When we allow ourselves to sleep before a GPU reset after disabling
264 	 * submission, even for a few milliseconds, gives an innocent context
265 	 * the opportunity to clear the GPU before the reset occurs. However,
266 	 * how long to sleep depends on the typical non-preemptible duration
267 	 * (a similar problem to determining the ideal preempt-reset timeout
268 	 * or even the heartbeat interval).
269 	 */
270 
271 	err = kstrtoull(buf, 0, &duration);
272 	if (err)
273 		return err;
274 
275 	if (duration > jiffies_to_msecs(MAX_SCHEDULE_TIMEOUT))
276 		return -EINVAL;
277 
278 	WRITE_ONCE(engine->props.stop_timeout_ms, duration);
279 	return count;
280 }
281 
282 static ssize_t
stop_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)283 stop_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
284 {
285 	struct intel_engine_cs *engine = kobj_to_engine(kobj);
286 
287 	return sprintf(buf, "%lu\n", engine->props.stop_timeout_ms);
288 }
289 
290 static struct kobj_attribute stop_timeout_attr =
291 __ATTR(stop_timeout_ms, 0644, stop_show, stop_store);
292 
293 static ssize_t
stop_default(struct kobject * kobj,struct kobj_attribute * attr,char * buf)294 stop_default(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
295 {
296 	struct intel_engine_cs *engine = kobj_to_engine(kobj);
297 
298 	return sprintf(buf, "%lu\n", engine->defaults.stop_timeout_ms);
299 }
300 
301 static struct kobj_attribute stop_timeout_def =
302 __ATTR(stop_timeout_ms, 0444, stop_default, NULL);
303 
304 static ssize_t
preempt_timeout_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)305 preempt_timeout_store(struct kobject *kobj, struct kobj_attribute *attr,
306 		      const char *buf, size_t count)
307 {
308 	struct intel_engine_cs *engine = kobj_to_engine(kobj);
309 	unsigned long long timeout;
310 	int err;
311 
312 	/*
313 	 * After initialising a preemption request, we give the current
314 	 * resident a small amount of time to vacate the GPU. The preemption
315 	 * request is for a higher priority context and should be immediate to
316 	 * maintain high quality of service (and avoid priority inversion).
317 	 * However, the preemption granularity of the GPU can be quite coarse
318 	 * and so we need a compromise.
319 	 */
320 
321 	err = kstrtoull(buf, 0, &timeout);
322 	if (err)
323 		return err;
324 
325 	if (timeout > jiffies_to_msecs(MAX_SCHEDULE_TIMEOUT))
326 		return -EINVAL;
327 
328 	WRITE_ONCE(engine->props.preempt_timeout_ms, timeout);
329 
330 	if (READ_ONCE(engine->execlists.pending[0]))
331 		set_timer_ms(&engine->execlists.preempt, timeout);
332 
333 	return count;
334 }
335 
336 static ssize_t
preempt_timeout_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)337 preempt_timeout_show(struct kobject *kobj, struct kobj_attribute *attr,
338 		     char *buf)
339 {
340 	struct intel_engine_cs *engine = kobj_to_engine(kobj);
341 
342 	return sprintf(buf, "%lu\n", engine->props.preempt_timeout_ms);
343 }
344 
345 static struct kobj_attribute preempt_timeout_attr =
346 __ATTR(preempt_timeout_ms, 0644, preempt_timeout_show, preempt_timeout_store);
347 
348 static ssize_t
preempt_timeout_default(struct kobject * kobj,struct kobj_attribute * attr,char * buf)349 preempt_timeout_default(struct kobject *kobj, struct kobj_attribute *attr,
350 			char *buf)
351 {
352 	struct intel_engine_cs *engine = kobj_to_engine(kobj);
353 
354 	return sprintf(buf, "%lu\n", engine->defaults.preempt_timeout_ms);
355 }
356 
357 static struct kobj_attribute preempt_timeout_def =
358 __ATTR(preempt_timeout_ms, 0444, preempt_timeout_default, NULL);
359 
360 static ssize_t
heartbeat_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)361 heartbeat_store(struct kobject *kobj, struct kobj_attribute *attr,
362 		const char *buf, size_t count)
363 {
364 	struct intel_engine_cs *engine = kobj_to_engine(kobj);
365 	unsigned long long delay;
366 	int err;
367 
368 	/*
369 	 * We monitor the health of the system via periodic heartbeat pulses.
370 	 * The pulses also provide the opportunity to perform garbage
371 	 * collection.  However, we interpret an incomplete pulse (a missed
372 	 * heartbeat) as an indication that the system is no longer responsive,
373 	 * i.e. hung, and perform an engine or full GPU reset. Given that the
374 	 * preemption granularity can be very coarse on a system, the optimal
375 	 * value for any workload is unknowable!
376 	 */
377 
378 	err = kstrtoull(buf, 0, &delay);
379 	if (err)
380 		return err;
381 
382 	if (delay >= jiffies_to_msecs(MAX_SCHEDULE_TIMEOUT))
383 		return -EINVAL;
384 
385 	err = intel_engine_set_heartbeat(engine, delay);
386 	if (err)
387 		return err;
388 
389 	return count;
390 }
391 
392 static ssize_t
heartbeat_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)393 heartbeat_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
394 {
395 	struct intel_engine_cs *engine = kobj_to_engine(kobj);
396 
397 	return sprintf(buf, "%lu\n", engine->props.heartbeat_interval_ms);
398 }
399 
400 static struct kobj_attribute heartbeat_interval_attr =
401 __ATTR(heartbeat_interval_ms, 0644, heartbeat_show, heartbeat_store);
402 
403 static ssize_t
heartbeat_default(struct kobject * kobj,struct kobj_attribute * attr,char * buf)404 heartbeat_default(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
405 {
406 	struct intel_engine_cs *engine = kobj_to_engine(kobj);
407 
408 	return sprintf(buf, "%lu\n", engine->defaults.heartbeat_interval_ms);
409 }
410 
411 static struct kobj_attribute heartbeat_interval_def =
412 __ATTR(heartbeat_interval_ms, 0444, heartbeat_default, NULL);
413 
kobj_engine_release(struct kobject * kobj)414 static void kobj_engine_release(struct kobject *kobj)
415 {
416 	kfree(kobj);
417 }
418 
419 static struct kobj_type kobj_engine_type = {
420 	.release = kobj_engine_release,
421 	.sysfs_ops = &kobj_sysfs_ops
422 };
423 
424 static struct kobject *
kobj_engine(struct kobject * dir,struct intel_engine_cs * engine)425 kobj_engine(struct kobject *dir, struct intel_engine_cs *engine)
426 {
427 	struct kobj_engine *ke;
428 
429 	ke = kzalloc(sizeof(*ke), GFP_KERNEL);
430 	if (!ke)
431 		return NULL;
432 
433 	kobject_init(&ke->base, &kobj_engine_type);
434 	ke->engine = engine;
435 
436 	if (kobject_add(&ke->base, dir, "%s", engine->name)) {
437 		kobject_put(&ke->base);
438 		return NULL;
439 	}
440 
441 	/* xfer ownership to sysfs tree */
442 	return &ke->base;
443 }
444 
add_defaults(struct kobj_engine * parent)445 static void add_defaults(struct kobj_engine *parent)
446 {
447 	static const struct attribute *files[] = {
448 		&max_spin_def.attr,
449 		&stop_timeout_def.attr,
450 #if CONFIG_DRM_I915_HEARTBEAT_INTERVAL
451 		&heartbeat_interval_def.attr,
452 #endif
453 		NULL
454 	};
455 	struct kobj_engine *ke;
456 
457 	ke = kzalloc(sizeof(*ke), GFP_KERNEL);
458 	if (!ke)
459 		return;
460 
461 	kobject_init(&ke->base, &kobj_engine_type);
462 	ke->engine = parent->engine;
463 
464 	if (kobject_add(&ke->base, &parent->base, "%s", ".defaults")) {
465 		kobject_put(&ke->base);
466 		return;
467 	}
468 
469 	if (sysfs_create_files(&ke->base, files))
470 		return;
471 
472 	if (intel_engine_has_timeslices(ke->engine) &&
473 	    sysfs_create_file(&ke->base, &timeslice_duration_def.attr))
474 		return;
475 
476 	if (intel_engine_has_preempt_reset(ke->engine) &&
477 	    sysfs_create_file(&ke->base, &preempt_timeout_def.attr))
478 		return;
479 }
480 
intel_engines_add_sysfs(struct drm_i915_private * i915)481 void intel_engines_add_sysfs(struct drm_i915_private *i915)
482 {
483 	static const struct attribute *files[] = {
484 		&name_attr.attr,
485 		&class_attr.attr,
486 		&inst_attr.attr,
487 		&mmio_attr.attr,
488 		&caps_attr.attr,
489 		&all_caps_attr.attr,
490 		&max_spin_attr.attr,
491 		&stop_timeout_attr.attr,
492 #if CONFIG_DRM_I915_HEARTBEAT_INTERVAL
493 		&heartbeat_interval_attr.attr,
494 #endif
495 		NULL
496 	};
497 
498 	struct device *kdev = i915->drm.primary->kdev;
499 	struct intel_engine_cs *engine;
500 	struct kobject *dir;
501 
502 	dir = kobject_create_and_add("engine", &kdev->kobj);
503 	if (!dir)
504 		return;
505 
506 	for_each_uabi_engine(engine, i915) {
507 		struct kobject *kobj;
508 
509 		kobj = kobj_engine(dir, engine);
510 		if (!kobj)
511 			goto err_engine;
512 
513 		if (sysfs_create_files(kobj, files))
514 			goto err_object;
515 
516 		if (intel_engine_has_timeslices(engine) &&
517 		    sysfs_create_file(kobj, &timeslice_duration_attr.attr))
518 			goto err_engine;
519 
520 		if (intel_engine_has_preempt_reset(engine) &&
521 		    sysfs_create_file(kobj, &preempt_timeout_attr.attr))
522 			goto err_engine;
523 
524 		add_defaults(container_of(kobj, struct kobj_engine, base));
525 
526 		if (0) {
527 err_object:
528 			kobject_put(kobj);
529 err_engine:
530 			dev_err(kdev, "Failed to add sysfs engine '%s'\n",
531 				engine->name);
532 			break;
533 		}
534 	}
535 }
536