1 /*
2  * Copyright 2009 Jerome Glisse.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the
7  * "Software"), to deal in the Software without restriction, including
8  * without limitation the rights to use, copy, modify, merge, publish,
9  * distribute, sub license, and/or sell copies of the Software, and to
10  * permit persons to whom the Software is furnished to do so, subject to
11  * the following conditions:
12  *
13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
20  *
21  * The above copyright notice and this permission notice (including the
22  * next paragraph) shall be included in all copies or substantial portions
23  * of the Software.
24  *
25  */
26 /*
27  * Authors:
28  *    Jerome Glisse <glisse@freedesktop.org>
29  *    Dave Airlie
30  */
31 
32 #include <linux/atomic.h>
33 #include <linux/firmware.h>
34 #include <linux/kref.h>
35 #include <linux/sched/signal.h>
36 #include <linux/seq_file.h>
37 #include <linux/slab.h>
38 #include <linux/wait.h>
39 
40 #include <drm/drm_device.h>
41 #include <drm/drm_file.h>
42 
43 #include "radeon.h"
44 #include "radeon_reg.h"
45 #include "radeon_trace.h"
46 
47 /*
48  * Fences
49  * Fences mark an event in the GPUs pipeline and are used
50  * for GPU/CPU synchronization.  When the fence is written,
51  * it is expected that all buffers associated with that fence
52  * are no longer in use by the associated ring on the GPU and
53  * that the relevant GPU caches have been flushed.  Whether
54  * we use a scratch register or memory location depends on the asic
55  * and whether writeback is enabled.
56  */
57 
58 /**
59  * radeon_fence_write - write a fence value
60  *
61  * @rdev: radeon_device pointer
62  * @seq: sequence number to write
63  * @ring: ring index the fence is associated with
64  *
65  * Writes a fence value to memory or a scratch register (all asics).
66  */
radeon_fence_write(struct radeon_device * rdev,u32 seq,int ring)67 static void radeon_fence_write(struct radeon_device *rdev, u32 seq, int ring)
68 {
69 	struct radeon_fence_driver *drv = &rdev->fence_drv[ring];
70 	if (likely(rdev->wb.enabled || !drv->scratch_reg)) {
71 		if (drv->cpu_addr) {
72 			*drv->cpu_addr = cpu_to_le32(seq);
73 		}
74 	} else {
75 		WREG32(drv->scratch_reg, seq);
76 	}
77 }
78 
79 /**
80  * radeon_fence_read - read a fence value
81  *
82  * @rdev: radeon_device pointer
83  * @ring: ring index the fence is associated with
84  *
85  * Reads a fence value from memory or a scratch register (all asics).
86  * Returns the value of the fence read from memory or register.
87  */
radeon_fence_read(struct radeon_device * rdev,int ring)88 static u32 radeon_fence_read(struct radeon_device *rdev, int ring)
89 {
90 	struct radeon_fence_driver *drv = &rdev->fence_drv[ring];
91 	u32 seq = 0;
92 
93 	if (likely(rdev->wb.enabled || !drv->scratch_reg)) {
94 		if (drv->cpu_addr) {
95 			seq = le32_to_cpu(*drv->cpu_addr);
96 		} else {
97 			seq = lower_32_bits(atomic64_read(&drv->last_seq));
98 		}
99 	} else {
100 		seq = RREG32(drv->scratch_reg);
101 	}
102 	return seq;
103 }
104 
105 /**
106  * radeon_fence_schedule_check - schedule lockup check
107  *
108  * @rdev: radeon_device pointer
109  * @ring: ring index we should work with
110  *
111  * Queues a delayed work item to check for lockups.
112  */
radeon_fence_schedule_check(struct radeon_device * rdev,int ring)113 static void radeon_fence_schedule_check(struct radeon_device *rdev, int ring)
114 {
115 	/*
116 	 * Do not reset the timer here with mod_delayed_work,
117 	 * this can livelock in an interaction with TTM delayed destroy.
118 	 */
119 	queue_delayed_work(system_power_efficient_wq,
120 			   &rdev->fence_drv[ring].lockup_work,
121 			   RADEON_FENCE_JIFFIES_TIMEOUT);
122 }
123 
124 /**
125  * radeon_fence_emit - emit a fence on the requested ring
126  *
127  * @rdev: radeon_device pointer
128  * @fence: radeon fence object
129  * @ring: ring index the fence is associated with
130  *
131  * Emits a fence command on the requested ring (all asics).
132  * Returns 0 on success, -ENOMEM on failure.
133  */
radeon_fence_emit(struct radeon_device * rdev,struct radeon_fence ** fence,int ring)134 int radeon_fence_emit(struct radeon_device *rdev,
135 		      struct radeon_fence **fence,
136 		      int ring)
137 {
138 	u64 seq;
139 
140 	/* we are protected by the ring emission mutex */
141 	*fence = kmalloc(sizeof(struct radeon_fence), GFP_KERNEL);
142 	if ((*fence) == NULL) {
143 		return -ENOMEM;
144 	}
145 	(*fence)->rdev = rdev;
146 	(*fence)->seq = seq = ++rdev->fence_drv[ring].sync_seq[ring];
147 	(*fence)->ring = ring;
148 	(*fence)->is_vm_update = false;
149 	dma_fence_init(&(*fence)->base, &radeon_fence_ops,
150 		       &rdev->fence_queue.lock,
151 		       rdev->fence_context + ring,
152 		       seq);
153 	radeon_fence_ring_emit(rdev, ring, *fence);
154 	trace_radeon_fence_emit(rdev->ddev, ring, (*fence)->seq);
155 	radeon_fence_schedule_check(rdev, ring);
156 	return 0;
157 }
158 
159 /*
160  * radeon_fence_check_signaled - callback from fence_queue
161  *
162  * this function is called with fence_queue lock held, which is also used
163  * for the fence locking itself, so unlocked variants are used for
164  * fence_signal, and remove_wait_queue.
165  */
radeon_fence_check_signaled(wait_queue_entry_t * wait,unsigned mode,int flags,void * key)166 static int radeon_fence_check_signaled(wait_queue_entry_t *wait, unsigned mode, int flags, void *key)
167 {
168 	struct radeon_fence *fence;
169 	u64 seq;
170 
171 	fence = container_of(wait, struct radeon_fence, fence_wake);
172 
173 	/*
174 	 * We cannot use radeon_fence_process here because we're already
175 	 * in the waitqueue, in a call from wake_up_all.
176 	 */
177 	seq = atomic64_read(&fence->rdev->fence_drv[fence->ring].last_seq);
178 	if (seq >= fence->seq) {
179 		dma_fence_signal_locked(&fence->base);
180 		radeon_irq_kms_sw_irq_put(fence->rdev, fence->ring);
181 		__remove_wait_queue(&fence->rdev->fence_queue, &fence->fence_wake);
182 		dma_fence_put(&fence->base);
183 	}
184 	return 0;
185 }
186 
187 /**
188  * radeon_fence_activity - check for fence activity
189  *
190  * @rdev: radeon_device pointer
191  * @ring: ring index the fence is associated with
192  *
193  * Checks the current fence value and calculates the last
194  * signalled fence value. Returns true if activity occured
195  * on the ring, and the fence_queue should be waken up.
196  */
radeon_fence_activity(struct radeon_device * rdev,int ring)197 static bool radeon_fence_activity(struct radeon_device *rdev, int ring)
198 {
199 	uint64_t seq, last_seq, last_emitted;
200 	unsigned count_loop = 0;
201 	bool wake = false;
202 
203 	/* Note there is a scenario here for an infinite loop but it's
204 	 * very unlikely to happen. For it to happen, the current polling
205 	 * process need to be interrupted by another process and another
206 	 * process needs to update the last_seq btw the atomic read and
207 	 * xchg of the current process.
208 	 *
209 	 * More over for this to go in infinite loop there need to be
210 	 * continuously new fence signaled ie radeon_fence_read needs
211 	 * to return a different value each time for both the currently
212 	 * polling process and the other process that xchg the last_seq
213 	 * btw atomic read and xchg of the current process. And the
214 	 * value the other process set as last seq must be higher than
215 	 * the seq value we just read. Which means that current process
216 	 * need to be interrupted after radeon_fence_read and before
217 	 * atomic xchg.
218 	 *
219 	 * To be even more safe we count the number of time we loop and
220 	 * we bail after 10 loop just accepting the fact that we might
221 	 * have temporarly set the last_seq not to the true real last
222 	 * seq but to an older one.
223 	 */
224 	last_seq = atomic64_read(&rdev->fence_drv[ring].last_seq);
225 	do {
226 		last_emitted = rdev->fence_drv[ring].sync_seq[ring];
227 		seq = radeon_fence_read(rdev, ring);
228 		seq |= last_seq & 0xffffffff00000000LL;
229 		if (seq < last_seq) {
230 			seq &= 0xffffffff;
231 			seq |= last_emitted & 0xffffffff00000000LL;
232 		}
233 
234 		if (seq <= last_seq || seq > last_emitted) {
235 			break;
236 		}
237 		/* If we loop over we don't want to return without
238 		 * checking if a fence is signaled as it means that the
239 		 * seq we just read is different from the previous on.
240 		 */
241 		wake = true;
242 		last_seq = seq;
243 		if ((count_loop++) > 10) {
244 			/* We looped over too many time leave with the
245 			 * fact that we might have set an older fence
246 			 * seq then the current real last seq as signaled
247 			 * by the hw.
248 			 */
249 			break;
250 		}
251 	} while (atomic64_xchg(&rdev->fence_drv[ring].last_seq, seq) > seq);
252 
253 	if (seq < last_emitted)
254 		radeon_fence_schedule_check(rdev, ring);
255 
256 	return wake;
257 }
258 
259 /**
260  * radeon_fence_check_lockup - check for hardware lockup
261  *
262  * @work: delayed work item
263  *
264  * Checks for fence activity and if there is none probe
265  * the hardware if a lockup occured.
266  */
radeon_fence_check_lockup(struct work_struct * work)267 static void radeon_fence_check_lockup(struct work_struct *work)
268 {
269 	struct radeon_fence_driver *fence_drv;
270 	struct radeon_device *rdev;
271 	int ring;
272 
273 	fence_drv = container_of(work, struct radeon_fence_driver,
274 				 lockup_work.work);
275 	rdev = fence_drv->rdev;
276 	ring = fence_drv - &rdev->fence_drv[0];
277 
278 	if (!down_read_trylock(&rdev->exclusive_lock)) {
279 		/* just reschedule the check if a reset is going on */
280 		radeon_fence_schedule_check(rdev, ring);
281 		return;
282 	}
283 
284 	if (fence_drv->delayed_irq && rdev->irq.installed) {
285 		unsigned long irqflags;
286 
287 		fence_drv->delayed_irq = false;
288 		spin_lock_irqsave(&rdev->irq.lock, irqflags);
289 		radeon_irq_set(rdev);
290 		spin_unlock_irqrestore(&rdev->irq.lock, irqflags);
291 	}
292 
293 	if (radeon_fence_activity(rdev, ring))
294 		wake_up_all(&rdev->fence_queue);
295 
296 	else if (radeon_ring_is_lockup(rdev, ring, &rdev->ring[ring])) {
297 
298 		/* good news we believe it's a lockup */
299 		dev_warn(rdev->dev, "GPU lockup (current fence id "
300 			 "0x%016llx last fence id 0x%016llx on ring %d)\n",
301 			 (uint64_t)atomic64_read(&fence_drv->last_seq),
302 			 fence_drv->sync_seq[ring], ring);
303 
304 		/* remember that we need an reset */
305 		rdev->needs_reset = true;
306 		wake_up_all(&rdev->fence_queue);
307 	}
308 	up_read(&rdev->exclusive_lock);
309 }
310 
311 /**
312  * radeon_fence_process - process a fence
313  *
314  * @rdev: radeon_device pointer
315  * @ring: ring index the fence is associated with
316  *
317  * Checks the current fence value and wakes the fence queue
318  * if the sequence number has increased (all asics).
319  */
radeon_fence_process(struct radeon_device * rdev,int ring)320 void radeon_fence_process(struct radeon_device *rdev, int ring)
321 {
322 	if (radeon_fence_activity(rdev, ring))
323 		wake_up_all(&rdev->fence_queue);
324 }
325 
326 /**
327  * radeon_fence_seq_signaled - check if a fence sequence number has signaled
328  *
329  * @rdev: radeon device pointer
330  * @seq: sequence number
331  * @ring: ring index the fence is associated with
332  *
333  * Check if the last signaled fence sequnce number is >= the requested
334  * sequence number (all asics).
335  * Returns true if the fence has signaled (current fence value
336  * is >= requested value) or false if it has not (current fence
337  * value is < the requested value.  Helper function for
338  * radeon_fence_signaled().
339  */
radeon_fence_seq_signaled(struct radeon_device * rdev,u64 seq,unsigned ring)340 static bool radeon_fence_seq_signaled(struct radeon_device *rdev,
341 				      u64 seq, unsigned ring)
342 {
343 	if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
344 		return true;
345 	}
346 	/* poll new last sequence at least once */
347 	radeon_fence_process(rdev, ring);
348 	if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
349 		return true;
350 	}
351 	return false;
352 }
353 
radeon_fence_is_signaled(struct dma_fence * f)354 static bool radeon_fence_is_signaled(struct dma_fence *f)
355 {
356 	struct radeon_fence *fence = to_radeon_fence(f);
357 	struct radeon_device *rdev = fence->rdev;
358 	unsigned ring = fence->ring;
359 	u64 seq = fence->seq;
360 
361 	if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
362 		return true;
363 	}
364 
365 	if (down_read_trylock(&rdev->exclusive_lock)) {
366 		radeon_fence_process(rdev, ring);
367 		up_read(&rdev->exclusive_lock);
368 
369 		if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
370 			return true;
371 		}
372 	}
373 	return false;
374 }
375 
376 /**
377  * radeon_fence_enable_signaling - enable signalling on fence
378  * @f: fence
379  *
380  * This function is called with fence_queue lock held, and adds a callback
381  * to fence_queue that checks if this fence is signaled, and if so it
382  * signals the fence and removes itself.
383  */
radeon_fence_enable_signaling(struct dma_fence * f)384 static bool radeon_fence_enable_signaling(struct dma_fence *f)
385 {
386 	struct radeon_fence *fence = to_radeon_fence(f);
387 	struct radeon_device *rdev = fence->rdev;
388 
389 	if (atomic64_read(&rdev->fence_drv[fence->ring].last_seq) >= fence->seq)
390 		return false;
391 
392 	if (down_read_trylock(&rdev->exclusive_lock)) {
393 		radeon_irq_kms_sw_irq_get(rdev, fence->ring);
394 
395 		if (radeon_fence_activity(rdev, fence->ring))
396 			wake_up_all_locked(&rdev->fence_queue);
397 
398 		/* did fence get signaled after we enabled the sw irq? */
399 		if (atomic64_read(&rdev->fence_drv[fence->ring].last_seq) >= fence->seq) {
400 			radeon_irq_kms_sw_irq_put(rdev, fence->ring);
401 			up_read(&rdev->exclusive_lock);
402 			return false;
403 		}
404 
405 		up_read(&rdev->exclusive_lock);
406 	} else {
407 		/* we're probably in a lockup, lets not fiddle too much */
408 		if (radeon_irq_kms_sw_irq_get_delayed(rdev, fence->ring))
409 			rdev->fence_drv[fence->ring].delayed_irq = true;
410 		radeon_fence_schedule_check(rdev, fence->ring);
411 	}
412 
413 	fence->fence_wake.flags = 0;
414 	fence->fence_wake.private = NULL;
415 	fence->fence_wake.func = radeon_fence_check_signaled;
416 	__add_wait_queue(&rdev->fence_queue, &fence->fence_wake);
417 	dma_fence_get(f);
418 	return true;
419 }
420 
421 /**
422  * radeon_fence_signaled - check if a fence has signaled
423  *
424  * @fence: radeon fence object
425  *
426  * Check if the requested fence has signaled (all asics).
427  * Returns true if the fence has signaled or false if it has not.
428  */
radeon_fence_signaled(struct radeon_fence * fence)429 bool radeon_fence_signaled(struct radeon_fence *fence)
430 {
431 	if (!fence)
432 		return true;
433 
434 	if (radeon_fence_seq_signaled(fence->rdev, fence->seq, fence->ring)) {
435 		dma_fence_signal(&fence->base);
436 		return true;
437 	}
438 	return false;
439 }
440 
441 /**
442  * radeon_fence_any_seq_signaled - check if any sequence number is signaled
443  *
444  * @rdev: radeon device pointer
445  * @seq: sequence numbers
446  *
447  * Check if the last signaled fence sequnce number is >= the requested
448  * sequence number (all asics).
449  * Returns true if any has signaled (current value is >= requested value)
450  * or false if it has not. Helper function for radeon_fence_wait_seq.
451  */
radeon_fence_any_seq_signaled(struct radeon_device * rdev,u64 * seq)452 static bool radeon_fence_any_seq_signaled(struct radeon_device *rdev, u64 *seq)
453 {
454 	unsigned i;
455 
456 	for (i = 0; i < RADEON_NUM_RINGS; ++i) {
457 		if (seq[i] && radeon_fence_seq_signaled(rdev, seq[i], i))
458 			return true;
459 	}
460 	return false;
461 }
462 
463 /**
464  * radeon_fence_wait_seq_timeout - wait for a specific sequence numbers
465  *
466  * @rdev: radeon device pointer
467  * @target_seq: sequence number(s) we want to wait for
468  * @intr: use interruptable sleep
469  * @timeout: maximum time to wait, or MAX_SCHEDULE_TIMEOUT for infinite wait
470  *
471  * Wait for the requested sequence number(s) to be written by any ring
472  * (all asics).  Sequnce number array is indexed by ring id.
473  * @intr selects whether to use interruptable (true) or non-interruptable
474  * (false) sleep when waiting for the sequence number.  Helper function
475  * for radeon_fence_wait_*().
476  * Returns remaining time if the sequence number has passed, 0 when
477  * the wait timeout, or an error for all other cases.
478  * -EDEADLK is returned when a GPU lockup has been detected.
479  */
radeon_fence_wait_seq_timeout(struct radeon_device * rdev,u64 * target_seq,bool intr,long timeout)480 static long radeon_fence_wait_seq_timeout(struct radeon_device *rdev,
481 					  u64 *target_seq, bool intr,
482 					  long timeout)
483 {
484 	long r;
485 	int i;
486 
487 	if (radeon_fence_any_seq_signaled(rdev, target_seq))
488 		return timeout;
489 
490 	/* enable IRQs and tracing */
491 	for (i = 0; i < RADEON_NUM_RINGS; ++i) {
492 		if (!target_seq[i])
493 			continue;
494 
495 		trace_radeon_fence_wait_begin(rdev->ddev, i, target_seq[i]);
496 		radeon_irq_kms_sw_irq_get(rdev, i);
497 	}
498 
499 	if (intr) {
500 		r = wait_event_interruptible_timeout(rdev->fence_queue, (
501 			radeon_fence_any_seq_signaled(rdev, target_seq)
502 			 || rdev->needs_reset), timeout);
503 	} else {
504 		r = wait_event_timeout(rdev->fence_queue, (
505 			radeon_fence_any_seq_signaled(rdev, target_seq)
506 			 || rdev->needs_reset), timeout);
507 	}
508 
509 	if (rdev->needs_reset)
510 		r = -EDEADLK;
511 
512 	for (i = 0; i < RADEON_NUM_RINGS; ++i) {
513 		if (!target_seq[i])
514 			continue;
515 
516 		radeon_irq_kms_sw_irq_put(rdev, i);
517 		trace_radeon_fence_wait_end(rdev->ddev, i, target_seq[i]);
518 	}
519 
520 	return r;
521 }
522 
523 /**
524  * radeon_fence_wait_timeout - wait for a fence to signal with timeout
525  *
526  * @fence: radeon fence object
527  * @intr: use interruptible sleep
528  *
529  * Wait for the requested fence to signal (all asics).
530  * @intr selects whether to use interruptable (true) or non-interruptable
531  * (false) sleep when waiting for the fence.
532  * @timeout: maximum time to wait, or MAX_SCHEDULE_TIMEOUT for infinite wait
533  * Returns remaining time if the sequence number has passed, 0 when
534  * the wait timeout, or an error for all other cases.
535  */
radeon_fence_wait_timeout(struct radeon_fence * fence,bool intr,long timeout)536 long radeon_fence_wait_timeout(struct radeon_fence *fence, bool intr, long timeout)
537 {
538 	uint64_t seq[RADEON_NUM_RINGS] = {};
539 	long r;
540 
541 	/*
542 	 * This function should not be called on !radeon fences.
543 	 * If this is the case, it would mean this function can
544 	 * also be called on radeon fences belonging to another card.
545 	 * exclusive_lock is not held in that case.
546 	 */
547 	if (WARN_ON_ONCE(!to_radeon_fence(&fence->base)))
548 		return dma_fence_wait(&fence->base, intr);
549 
550 	seq[fence->ring] = fence->seq;
551 	r = radeon_fence_wait_seq_timeout(fence->rdev, seq, intr, timeout);
552 	if (r <= 0) {
553 		return r;
554 	}
555 
556 	dma_fence_signal(&fence->base);
557 	return r;
558 }
559 
560 /**
561  * radeon_fence_wait - wait for a fence to signal
562  *
563  * @fence: radeon fence object
564  * @intr: use interruptible sleep
565  *
566  * Wait for the requested fence to signal (all asics).
567  * @intr selects whether to use interruptable (true) or non-interruptable
568  * (false) sleep when waiting for the fence.
569  * Returns 0 if the fence has passed, error for all other cases.
570  */
radeon_fence_wait(struct radeon_fence * fence,bool intr)571 int radeon_fence_wait(struct radeon_fence *fence, bool intr)
572 {
573 	long r = radeon_fence_wait_timeout(fence, intr, MAX_SCHEDULE_TIMEOUT);
574 	if (r > 0) {
575 		return 0;
576 	} else {
577 		return r;
578 	}
579 }
580 
581 /**
582  * radeon_fence_wait_any - wait for a fence to signal on any ring
583  *
584  * @rdev: radeon device pointer
585  * @fences: radeon fence object(s)
586  * @intr: use interruptable sleep
587  *
588  * Wait for any requested fence to signal (all asics).  Fence
589  * array is indexed by ring id.  @intr selects whether to use
590  * interruptable (true) or non-interruptable (false) sleep when
591  * waiting for the fences. Used by the suballocator.
592  * Returns 0 if any fence has passed, error for all other cases.
593  */
radeon_fence_wait_any(struct radeon_device * rdev,struct radeon_fence ** fences,bool intr)594 int radeon_fence_wait_any(struct radeon_device *rdev,
595 			  struct radeon_fence **fences,
596 			  bool intr)
597 {
598 	uint64_t seq[RADEON_NUM_RINGS];
599 	unsigned i, num_rings = 0;
600 	long r;
601 
602 	for (i = 0; i < RADEON_NUM_RINGS; ++i) {
603 		seq[i] = 0;
604 
605 		if (!fences[i]) {
606 			continue;
607 		}
608 
609 		seq[i] = fences[i]->seq;
610 		++num_rings;
611 	}
612 
613 	/* nothing to wait for ? */
614 	if (num_rings == 0)
615 		return -ENOENT;
616 
617 	r = radeon_fence_wait_seq_timeout(rdev, seq, intr, MAX_SCHEDULE_TIMEOUT);
618 	if (r < 0) {
619 		return r;
620 	}
621 	return 0;
622 }
623 
624 /**
625  * radeon_fence_wait_next - wait for the next fence to signal
626  *
627  * @rdev: radeon device pointer
628  * @ring: ring index the fence is associated with
629  *
630  * Wait for the next fence on the requested ring to signal (all asics).
631  * Returns 0 if the next fence has passed, error for all other cases.
632  * Caller must hold ring lock.
633  */
radeon_fence_wait_next(struct radeon_device * rdev,int ring)634 int radeon_fence_wait_next(struct radeon_device *rdev, int ring)
635 {
636 	uint64_t seq[RADEON_NUM_RINGS] = {};
637 	long r;
638 
639 	seq[ring] = atomic64_read(&rdev->fence_drv[ring].last_seq) + 1ULL;
640 	if (seq[ring] >= rdev->fence_drv[ring].sync_seq[ring]) {
641 		/* nothing to wait for, last_seq is
642 		   already the last emited fence */
643 		return -ENOENT;
644 	}
645 	r = radeon_fence_wait_seq_timeout(rdev, seq, false, MAX_SCHEDULE_TIMEOUT);
646 	if (r < 0)
647 		return r;
648 	return 0;
649 }
650 
651 /**
652  * radeon_fence_wait_empty - wait for all fences to signal
653  *
654  * @rdev: radeon device pointer
655  * @ring: ring index the fence is associated with
656  *
657  * Wait for all fences on the requested ring to signal (all asics).
658  * Returns 0 if the fences have passed, error for all other cases.
659  * Caller must hold ring lock.
660  */
radeon_fence_wait_empty(struct radeon_device * rdev,int ring)661 int radeon_fence_wait_empty(struct radeon_device *rdev, int ring)
662 {
663 	uint64_t seq[RADEON_NUM_RINGS] = {};
664 	long r;
665 
666 	seq[ring] = rdev->fence_drv[ring].sync_seq[ring];
667 	if (!seq[ring])
668 		return 0;
669 
670 	r = radeon_fence_wait_seq_timeout(rdev, seq, false, MAX_SCHEDULE_TIMEOUT);
671 	if (r < 0) {
672 		if (r == -EDEADLK)
673 			return -EDEADLK;
674 
675 		dev_err(rdev->dev, "error waiting for ring[%d] to become idle (%ld)\n",
676 			ring, r);
677 	}
678 	return 0;
679 }
680 
681 /**
682  * radeon_fence_ref - take a ref on a fence
683  *
684  * @fence: radeon fence object
685  *
686  * Take a reference on a fence (all asics).
687  * Returns the fence.
688  */
radeon_fence_ref(struct radeon_fence * fence)689 struct radeon_fence *radeon_fence_ref(struct radeon_fence *fence)
690 {
691 	dma_fence_get(&fence->base);
692 	return fence;
693 }
694 
695 /**
696  * radeon_fence_unref - remove a ref on a fence
697  *
698  * @fence: radeon fence object
699  *
700  * Remove a reference on a fence (all asics).
701  */
radeon_fence_unref(struct radeon_fence ** fence)702 void radeon_fence_unref(struct radeon_fence **fence)
703 {
704 	struct radeon_fence *tmp = *fence;
705 
706 	*fence = NULL;
707 	if (tmp) {
708 		dma_fence_put(&tmp->base);
709 	}
710 }
711 
712 /**
713  * radeon_fence_count_emitted - get the count of emitted fences
714  *
715  * @rdev: radeon device pointer
716  * @ring: ring index the fence is associated with
717  *
718  * Get the number of fences emitted on the requested ring (all asics).
719  * Returns the number of emitted fences on the ring.  Used by the
720  * dynpm code to ring track activity.
721  */
radeon_fence_count_emitted(struct radeon_device * rdev,int ring)722 unsigned radeon_fence_count_emitted(struct radeon_device *rdev, int ring)
723 {
724 	uint64_t emitted;
725 
726 	/* We are not protected by ring lock when reading the last sequence
727 	 * but it's ok to report slightly wrong fence count here.
728 	 */
729 	radeon_fence_process(rdev, ring);
730 	emitted = rdev->fence_drv[ring].sync_seq[ring]
731 		- atomic64_read(&rdev->fence_drv[ring].last_seq);
732 	/* to avoid 32bits warp around */
733 	if (emitted > 0x10000000) {
734 		emitted = 0x10000000;
735 	}
736 	return (unsigned)emitted;
737 }
738 
739 /**
740  * radeon_fence_need_sync - do we need a semaphore
741  *
742  * @fence: radeon fence object
743  * @dst_ring: which ring to check against
744  *
745  * Check if the fence needs to be synced against another ring
746  * (all asics).  If so, we need to emit a semaphore.
747  * Returns true if we need to sync with another ring, false if
748  * not.
749  */
radeon_fence_need_sync(struct radeon_fence * fence,int dst_ring)750 bool radeon_fence_need_sync(struct radeon_fence *fence, int dst_ring)
751 {
752 	struct radeon_fence_driver *fdrv;
753 
754 	if (!fence) {
755 		return false;
756 	}
757 
758 	if (fence->ring == dst_ring) {
759 		return false;
760 	}
761 
762 	/* we are protected by the ring mutex */
763 	fdrv = &fence->rdev->fence_drv[dst_ring];
764 	if (fence->seq <= fdrv->sync_seq[fence->ring]) {
765 		return false;
766 	}
767 
768 	return true;
769 }
770 
771 /**
772  * radeon_fence_note_sync - record the sync point
773  *
774  * @fence: radeon fence object
775  * @dst_ring: which ring to check against
776  *
777  * Note the sequence number at which point the fence will
778  * be synced with the requested ring (all asics).
779  */
radeon_fence_note_sync(struct radeon_fence * fence,int dst_ring)780 void radeon_fence_note_sync(struct radeon_fence *fence, int dst_ring)
781 {
782 	struct radeon_fence_driver *dst, *src;
783 	unsigned i;
784 
785 	if (!fence) {
786 		return;
787 	}
788 
789 	if (fence->ring == dst_ring) {
790 		return;
791 	}
792 
793 	/* we are protected by the ring mutex */
794 	src = &fence->rdev->fence_drv[fence->ring];
795 	dst = &fence->rdev->fence_drv[dst_ring];
796 	for (i = 0; i < RADEON_NUM_RINGS; ++i) {
797 		if (i == dst_ring) {
798 			continue;
799 		}
800 		dst->sync_seq[i] = max(dst->sync_seq[i], src->sync_seq[i]);
801 	}
802 }
803 
804 /**
805  * radeon_fence_driver_start_ring - make the fence driver
806  * ready for use on the requested ring.
807  *
808  * @rdev: radeon device pointer
809  * @ring: ring index to start the fence driver on
810  *
811  * Make the fence driver ready for processing (all asics).
812  * Not all asics have all rings, so each asic will only
813  * start the fence driver on the rings it has.
814  * Returns 0 for success, errors for failure.
815  */
radeon_fence_driver_start_ring(struct radeon_device * rdev,int ring)816 int radeon_fence_driver_start_ring(struct radeon_device *rdev, int ring)
817 {
818 	uint64_t index;
819 	int r;
820 
821 	radeon_scratch_free(rdev, rdev->fence_drv[ring].scratch_reg);
822 	if (rdev->wb.use_event || !radeon_ring_supports_scratch_reg(rdev, &rdev->ring[ring])) {
823 		rdev->fence_drv[ring].scratch_reg = 0;
824 		if (ring != R600_RING_TYPE_UVD_INDEX) {
825 			index = R600_WB_EVENT_OFFSET + ring * 4;
826 			rdev->fence_drv[ring].cpu_addr = &rdev->wb.wb[index/4];
827 			rdev->fence_drv[ring].gpu_addr = rdev->wb.gpu_addr +
828 							 index;
829 
830 		} else {
831 			/* put fence directly behind firmware */
832 			index = ALIGN(rdev->uvd_fw->size, 8);
833 			rdev->fence_drv[ring].cpu_addr = rdev->uvd.cpu_addr + index;
834 			rdev->fence_drv[ring].gpu_addr = rdev->uvd.gpu_addr + index;
835 		}
836 
837 	} else {
838 		r = radeon_scratch_get(rdev, &rdev->fence_drv[ring].scratch_reg);
839 		if (r) {
840 			dev_err(rdev->dev, "fence failed to get scratch register\n");
841 			return r;
842 		}
843 		index = RADEON_WB_SCRATCH_OFFSET +
844 			rdev->fence_drv[ring].scratch_reg -
845 			rdev->scratch.reg_base;
846 		rdev->fence_drv[ring].cpu_addr = &rdev->wb.wb[index/4];
847 		rdev->fence_drv[ring].gpu_addr = rdev->wb.gpu_addr + index;
848 	}
849 	radeon_fence_write(rdev, atomic64_read(&rdev->fence_drv[ring].last_seq), ring);
850 	rdev->fence_drv[ring].initialized = true;
851 	dev_info(rdev->dev, "fence driver on ring %d use gpu addr 0x%016llx\n",
852 		 ring, rdev->fence_drv[ring].gpu_addr);
853 	return 0;
854 }
855 
856 /**
857  * radeon_fence_driver_init_ring - init the fence driver
858  * for the requested ring.
859  *
860  * @rdev: radeon device pointer
861  * @ring: ring index to start the fence driver on
862  *
863  * Init the fence driver for the requested ring (all asics).
864  * Helper function for radeon_fence_driver_init().
865  */
radeon_fence_driver_init_ring(struct radeon_device * rdev,int ring)866 static void radeon_fence_driver_init_ring(struct radeon_device *rdev, int ring)
867 {
868 	int i;
869 
870 	rdev->fence_drv[ring].scratch_reg = -1;
871 	rdev->fence_drv[ring].cpu_addr = NULL;
872 	rdev->fence_drv[ring].gpu_addr = 0;
873 	for (i = 0; i < RADEON_NUM_RINGS; ++i)
874 		rdev->fence_drv[ring].sync_seq[i] = 0;
875 	atomic64_set(&rdev->fence_drv[ring].last_seq, 0);
876 	rdev->fence_drv[ring].initialized = false;
877 	INIT_DELAYED_WORK(&rdev->fence_drv[ring].lockup_work,
878 			  radeon_fence_check_lockup);
879 	rdev->fence_drv[ring].rdev = rdev;
880 }
881 
882 /**
883  * radeon_fence_driver_init - init the fence driver
884  * for all possible rings.
885  *
886  * @rdev: radeon device pointer
887  *
888  * Init the fence driver for all possible rings (all asics).
889  * Not all asics have all rings, so each asic will only
890  * start the fence driver on the rings it has using
891  * radeon_fence_driver_start_ring().
892  */
radeon_fence_driver_init(struct radeon_device * rdev)893 void radeon_fence_driver_init(struct radeon_device *rdev)
894 {
895 	int ring;
896 
897 	init_waitqueue_head(&rdev->fence_queue);
898 	for (ring = 0; ring < RADEON_NUM_RINGS; ring++) {
899 		radeon_fence_driver_init_ring(rdev, ring);
900 	}
901 
902 	radeon_debugfs_fence_init(rdev);
903 }
904 
905 /**
906  * radeon_fence_driver_fini - tear down the fence driver
907  * for all possible rings.
908  *
909  * @rdev: radeon device pointer
910  *
911  * Tear down the fence driver for all possible rings (all asics).
912  */
radeon_fence_driver_fini(struct radeon_device * rdev)913 void radeon_fence_driver_fini(struct radeon_device *rdev)
914 {
915 	int ring, r;
916 
917 	mutex_lock(&rdev->ring_lock);
918 	for (ring = 0; ring < RADEON_NUM_RINGS; ring++) {
919 		if (!rdev->fence_drv[ring].initialized)
920 			continue;
921 		r = radeon_fence_wait_empty(rdev, ring);
922 		if (r) {
923 			/* no need to trigger GPU reset as we are unloading */
924 			radeon_fence_driver_force_completion(rdev, ring);
925 		}
926 		cancel_delayed_work_sync(&rdev->fence_drv[ring].lockup_work);
927 		wake_up_all(&rdev->fence_queue);
928 		radeon_scratch_free(rdev, rdev->fence_drv[ring].scratch_reg);
929 		rdev->fence_drv[ring].initialized = false;
930 	}
931 	mutex_unlock(&rdev->ring_lock);
932 }
933 
934 /**
935  * radeon_fence_driver_force_completion - force all fence waiter to complete
936  *
937  * @rdev: radeon device pointer
938  * @ring: the ring to complete
939  *
940  * In case of GPU reset failure make sure no process keep waiting on fence
941  * that will never complete.
942  */
radeon_fence_driver_force_completion(struct radeon_device * rdev,int ring)943 void radeon_fence_driver_force_completion(struct radeon_device *rdev, int ring)
944 {
945 	if (rdev->fence_drv[ring].initialized) {
946 		radeon_fence_write(rdev, rdev->fence_drv[ring].sync_seq[ring], ring);
947 		cancel_delayed_work_sync(&rdev->fence_drv[ring].lockup_work);
948 	}
949 }
950 
951 
952 /*
953  * Fence debugfs
954  */
955 #if defined(CONFIG_DEBUG_FS)
radeon_debugfs_fence_info_show(struct seq_file * m,void * data)956 static int radeon_debugfs_fence_info_show(struct seq_file *m, void *data)
957 {
958 	struct radeon_device *rdev = m->private;
959 	int i, j;
960 
961 	for (i = 0; i < RADEON_NUM_RINGS; ++i) {
962 		if (!rdev->fence_drv[i].initialized)
963 			continue;
964 
965 		radeon_fence_process(rdev, i);
966 
967 		seq_printf(m, "--- ring %d ---\n", i);
968 		seq_printf(m, "Last signaled fence 0x%016llx\n",
969 			   (unsigned long long)atomic64_read(&rdev->fence_drv[i].last_seq));
970 		seq_printf(m, "Last emitted        0x%016llx\n",
971 			   rdev->fence_drv[i].sync_seq[i]);
972 
973 		for (j = 0; j < RADEON_NUM_RINGS; ++j) {
974 			if (i != j && rdev->fence_drv[j].initialized)
975 				seq_printf(m, "Last sync to ring %d 0x%016llx\n",
976 					   j, rdev->fence_drv[i].sync_seq[j]);
977 		}
978 	}
979 	return 0;
980 }
981 
982 /*
983  * radeon_debugfs_gpu_reset - manually trigger a gpu reset
984  *
985  * Manually trigger a gpu reset at the next fence wait.
986  */
radeon_debugfs_gpu_reset(void * data,u64 * val)987 static int radeon_debugfs_gpu_reset(void *data, u64 *val)
988 {
989 	struct radeon_device *rdev = (struct radeon_device *)data;
990 
991 	down_read(&rdev->exclusive_lock);
992 	*val = rdev->needs_reset;
993 	rdev->needs_reset = true;
994 	wake_up_all(&rdev->fence_queue);
995 	up_read(&rdev->exclusive_lock);
996 
997 	return 0;
998 }
999 DEFINE_SHOW_ATTRIBUTE(radeon_debugfs_fence_info);
1000 DEFINE_DEBUGFS_ATTRIBUTE(radeon_debugfs_gpu_reset_fops,
1001 			 radeon_debugfs_gpu_reset, NULL, "%lld\n");
1002 #endif
1003 
radeon_debugfs_fence_init(struct radeon_device * rdev)1004 void radeon_debugfs_fence_init(struct radeon_device *rdev)
1005 {
1006 #if defined(CONFIG_DEBUG_FS)
1007 	struct dentry *root = rdev->ddev->primary->debugfs_root;
1008 
1009 	debugfs_create_file("radeon_gpu_reset", 0444, root, rdev,
1010 			    &radeon_debugfs_gpu_reset_fops);
1011 	debugfs_create_file("radeon_fence_info", 0444, root, rdev,
1012 			    &radeon_debugfs_fence_info_fops);
1013 
1014 
1015 #endif
1016 }
1017 
radeon_fence_get_driver_name(struct dma_fence * fence)1018 static const char *radeon_fence_get_driver_name(struct dma_fence *fence)
1019 {
1020 	return "radeon";
1021 }
1022 
radeon_fence_get_timeline_name(struct dma_fence * f)1023 static const char *radeon_fence_get_timeline_name(struct dma_fence *f)
1024 {
1025 	struct radeon_fence *fence = to_radeon_fence(f);
1026 	switch (fence->ring) {
1027 	case RADEON_RING_TYPE_GFX_INDEX: return "radeon.gfx";
1028 	case CAYMAN_RING_TYPE_CP1_INDEX: return "radeon.cp1";
1029 	case CAYMAN_RING_TYPE_CP2_INDEX: return "radeon.cp2";
1030 	case R600_RING_TYPE_DMA_INDEX: return "radeon.dma";
1031 	case CAYMAN_RING_TYPE_DMA1_INDEX: return "radeon.dma1";
1032 	case R600_RING_TYPE_UVD_INDEX: return "radeon.uvd";
1033 	case TN_RING_TYPE_VCE1_INDEX: return "radeon.vce1";
1034 	case TN_RING_TYPE_VCE2_INDEX: return "radeon.vce2";
1035 	default: WARN_ON_ONCE(1); return "radeon.unk";
1036 	}
1037 }
1038 
radeon_test_signaled(struct radeon_fence * fence)1039 static inline bool radeon_test_signaled(struct radeon_fence *fence)
1040 {
1041 	return test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->base.flags);
1042 }
1043 
1044 struct radeon_wait_cb {
1045 	struct dma_fence_cb base;
1046 	struct task_struct *task;
1047 };
1048 
1049 static void
radeon_fence_wait_cb(struct dma_fence * fence,struct dma_fence_cb * cb)1050 radeon_fence_wait_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
1051 {
1052 	struct radeon_wait_cb *wait =
1053 		container_of(cb, struct radeon_wait_cb, base);
1054 
1055 	wake_up_process(wait->task);
1056 }
1057 
radeon_fence_default_wait(struct dma_fence * f,bool intr,signed long t)1058 static signed long radeon_fence_default_wait(struct dma_fence *f, bool intr,
1059 					     signed long t)
1060 {
1061 	struct radeon_fence *fence = to_radeon_fence(f);
1062 	struct radeon_device *rdev = fence->rdev;
1063 	struct radeon_wait_cb cb;
1064 
1065 	cb.task = current;
1066 
1067 	if (dma_fence_add_callback(f, &cb.base, radeon_fence_wait_cb))
1068 		return t;
1069 
1070 	while (t > 0) {
1071 		if (intr)
1072 			set_current_state(TASK_INTERRUPTIBLE);
1073 		else
1074 			set_current_state(TASK_UNINTERRUPTIBLE);
1075 
1076 		/*
1077 		 * radeon_test_signaled must be called after
1078 		 * set_current_state to prevent a race with wake_up_process
1079 		 */
1080 		if (radeon_test_signaled(fence))
1081 			break;
1082 
1083 		if (rdev->needs_reset) {
1084 			t = -EDEADLK;
1085 			break;
1086 		}
1087 
1088 		t = schedule_timeout(t);
1089 
1090 		if (t > 0 && intr && signal_pending(current))
1091 			t = -ERESTARTSYS;
1092 	}
1093 
1094 	__set_current_state(TASK_RUNNING);
1095 	dma_fence_remove_callback(f, &cb.base);
1096 
1097 	return t;
1098 }
1099 
1100 const struct dma_fence_ops radeon_fence_ops = {
1101 	.get_driver_name = radeon_fence_get_driver_name,
1102 	.get_timeline_name = radeon_fence_get_timeline_name,
1103 	.enable_signaling = radeon_fence_enable_signaling,
1104 	.signaled = radeon_fence_is_signaled,
1105 	.wait = radeon_fence_default_wait,
1106 	.release = NULL,
1107 };
1108