Merge tag 'mac80211-for-davem-2016-07-06' of git://git.kernel.org/pub/scm/linux/kerne...
[cascardo/linux.git] / drivers / gpu / drm / amd / amdgpu / amdgpu_fence.c
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 #include <linux/seq_file.h>
32 #include <linux/atomic.h>
33 #include <linux/wait.h>
34 #include <linux/kref.h>
35 #include <linux/slab.h>
36 #include <linux/firmware.h>
37 #include <drm/drmP.h>
38 #include "amdgpu.h"
39 #include "amdgpu_trace.h"
40
41 /*
42  * Fences
43  * Fences mark an event in the GPUs pipeline and are used
44  * for GPU/CPU synchronization.  When the fence is written,
45  * it is expected that all buffers associated with that fence
46  * are no longer in use by the associated ring on the GPU and
47  * that the the relevant GPU caches have been flushed.
48  */
49
50 struct amdgpu_fence {
51         struct fence base;
52
53         /* RB, DMA, etc. */
54         struct amdgpu_ring              *ring;
55 };
56
57 static struct kmem_cache *amdgpu_fence_slab;
58
59 int amdgpu_fence_slab_init(void)
60 {
61         amdgpu_fence_slab = kmem_cache_create(
62                 "amdgpu_fence", sizeof(struct amdgpu_fence), 0,
63                 SLAB_HWCACHE_ALIGN, NULL);
64         if (!amdgpu_fence_slab)
65                 return -ENOMEM;
66         return 0;
67 }
68
69 void amdgpu_fence_slab_fini(void)
70 {
71         kmem_cache_destroy(amdgpu_fence_slab);
72 }
73 /*
74  * Cast helper
75  */
76 static const struct fence_ops amdgpu_fence_ops;
77 static inline struct amdgpu_fence *to_amdgpu_fence(struct fence *f)
78 {
79         struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base);
80
81         if (__f->base.ops == &amdgpu_fence_ops)
82                 return __f;
83
84         return NULL;
85 }
86
87 /**
88  * amdgpu_fence_write - write a fence value
89  *
90  * @ring: ring the fence is associated with
91  * @seq: sequence number to write
92  *
93  * Writes a fence value to memory (all asics).
94  */
95 static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq)
96 {
97         struct amdgpu_fence_driver *drv = &ring->fence_drv;
98
99         if (drv->cpu_addr)
100                 *drv->cpu_addr = cpu_to_le32(seq);
101 }
102
103 /**
104  * amdgpu_fence_read - read a fence value
105  *
106  * @ring: ring the fence is associated with
107  *
108  * Reads a fence value from memory (all asics).
109  * Returns the value of the fence read from memory.
110  */
111 static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
112 {
113         struct amdgpu_fence_driver *drv = &ring->fence_drv;
114         u32 seq = 0;
115
116         if (drv->cpu_addr)
117                 seq = le32_to_cpu(*drv->cpu_addr);
118         else
119                 seq = atomic_read(&drv->last_seq);
120
121         return seq;
122 }
123
124 /**
125  * amdgpu_fence_emit - emit a fence on the requested ring
126  *
127  * @ring: ring the fence is associated with
128  * @f: resulting fence object
129  *
130  * Emits a fence command on the requested ring (all asics).
131  * Returns 0 on success, -ENOMEM on failure.
132  */
133 int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **f)
134 {
135         struct amdgpu_device *adev = ring->adev;
136         struct amdgpu_fence *fence;
137         struct fence *old, **ptr;
138         uint32_t seq;
139
140         fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
141         if (fence == NULL)
142                 return -ENOMEM;
143
144         seq = ++ring->fence_drv.sync_seq;
145         fence->ring = ring;
146         fence_init(&fence->base, &amdgpu_fence_ops,
147                    &ring->fence_drv.lock,
148                    adev->fence_context + ring->idx,
149                    seq);
150         amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
151                                seq, AMDGPU_FENCE_FLAG_INT);
152
153         ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
154         /* This function can't be called concurrently anyway, otherwise
155          * emitting the fence would mess up the hardware ring buffer.
156          */
157         old = rcu_dereference_protected(*ptr, 1);
158         if (old && !fence_is_signaled(old)) {
159                 DRM_INFO("rcu slot is busy\n");
160                 fence_wait(old, false);
161         }
162
163         rcu_assign_pointer(*ptr, fence_get(&fence->base));
164
165         *f = &fence->base;
166
167         return 0;
168 }
169
170 /**
171  * amdgpu_fence_schedule_fallback - schedule fallback check
172  *
173  * @ring: pointer to struct amdgpu_ring
174  *
175  * Start a timer as fallback to our interrupts.
176  */
177 static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring)
178 {
179         mod_timer(&ring->fence_drv.fallback_timer,
180                   jiffies + AMDGPU_FENCE_JIFFIES_TIMEOUT);
181 }
182
183 /**
184  * amdgpu_fence_process - check for fence activity
185  *
186  * @ring: pointer to struct amdgpu_ring
187  *
188  * Checks the current fence value and calculates the last
189  * signalled fence value. Wakes the fence queue if the
190  * sequence number has increased.
191  */
192 void amdgpu_fence_process(struct amdgpu_ring *ring)
193 {
194         struct amdgpu_fence_driver *drv = &ring->fence_drv;
195         uint32_t seq, last_seq;
196         int r;
197
198         do {
199                 last_seq = atomic_read(&ring->fence_drv.last_seq);
200                 seq = amdgpu_fence_read(ring);
201
202         } while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
203
204         if (seq != ring->fence_drv.sync_seq)
205                 amdgpu_fence_schedule_fallback(ring);
206
207         while (last_seq != seq) {
208                 struct fence *fence, **ptr;
209
210                 ptr = &drv->fences[++last_seq & drv->num_fences_mask];
211
212                 /* There is always exactly one thread signaling this fence slot */
213                 fence = rcu_dereference_protected(*ptr, 1);
214                 RCU_INIT_POINTER(*ptr, NULL);
215
216                 BUG_ON(!fence);
217
218                 r = fence_signal(fence);
219                 if (!r)
220                         FENCE_TRACE(fence, "signaled from irq context\n");
221                 else
222                         BUG();
223
224                 fence_put(fence);
225         }
226 }
227
228 /**
229  * amdgpu_fence_fallback - fallback for hardware interrupts
230  *
231  * @work: delayed work item
232  *
233  * Checks for fence activity.
234  */
235 static void amdgpu_fence_fallback(unsigned long arg)
236 {
237         struct amdgpu_ring *ring = (void *)arg;
238
239         amdgpu_fence_process(ring);
240 }
241
242 /**
243  * amdgpu_fence_wait_empty - wait for all fences to signal
244  *
245  * @adev: amdgpu device pointer
246  * @ring: ring index the fence is associated with
247  *
248  * Wait for all fences on the requested ring to signal (all asics).
249  * Returns 0 if the fences have passed, error for all other cases.
250  */
251 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
252 {
253         uint64_t seq = ACCESS_ONCE(ring->fence_drv.sync_seq);
254         struct fence *fence, **ptr;
255         int r;
256
257         if (!seq)
258                 return 0;
259
260         ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
261         rcu_read_lock();
262         fence = rcu_dereference(*ptr);
263         if (!fence || !fence_get_rcu(fence)) {
264                 rcu_read_unlock();
265                 return 0;
266         }
267         rcu_read_unlock();
268
269         r = fence_wait(fence, false);
270         fence_put(fence);
271         return r;
272 }
273
274 /**
275  * amdgpu_fence_count_emitted - get the count of emitted fences
276  *
277  * @ring: ring the fence is associated with
278  *
279  * Get the number of fences emitted on the requested ring (all asics).
280  * Returns the number of emitted fences on the ring.  Used by the
281  * dynpm code to ring track activity.
282  */
283 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
284 {
285         uint64_t emitted;
286
287         /* We are not protected by ring lock when reading the last sequence
288          * but it's ok to report slightly wrong fence count here.
289          */
290         amdgpu_fence_process(ring);
291         emitted = 0x100000000ull;
292         emitted -= atomic_read(&ring->fence_drv.last_seq);
293         emitted += ACCESS_ONCE(ring->fence_drv.sync_seq);
294         return lower_32_bits(emitted);
295 }
296
297 /**
298  * amdgpu_fence_driver_start_ring - make the fence driver
299  * ready for use on the requested ring.
300  *
301  * @ring: ring to start the fence driver on
302  * @irq_src: interrupt source to use for this ring
303  * @irq_type: interrupt type to use for this ring
304  *
305  * Make the fence driver ready for processing (all asics).
306  * Not all asics have all rings, so each asic will only
307  * start the fence driver on the rings it has.
308  * Returns 0 for success, errors for failure.
309  */
310 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
311                                    struct amdgpu_irq_src *irq_src,
312                                    unsigned irq_type)
313 {
314         struct amdgpu_device *adev = ring->adev;
315         uint64_t index;
316
317         if (ring != &adev->uvd.ring) {
318                 ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs];
319                 ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4);
320         } else {
321                 /* put fence directly behind firmware */
322                 index = ALIGN(adev->uvd.fw->size, 8);
323                 ring->fence_drv.cpu_addr = adev->uvd.cpu_addr + index;
324                 ring->fence_drv.gpu_addr = adev->uvd.gpu_addr + index;
325         }
326         amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
327         amdgpu_irq_get(adev, irq_src, irq_type);
328
329         ring->fence_drv.irq_src = irq_src;
330         ring->fence_drv.irq_type = irq_type;
331         ring->fence_drv.initialized = true;
332
333         dev_info(adev->dev, "fence driver on ring %d use gpu addr 0x%016llx, "
334                  "cpu addr 0x%p\n", ring->idx,
335                  ring->fence_drv.gpu_addr, ring->fence_drv.cpu_addr);
336         return 0;
337 }
338
339 /**
340  * amdgpu_fence_driver_init_ring - init the fence driver
341  * for the requested ring.
342  *
343  * @ring: ring to init the fence driver on
344  * @num_hw_submission: number of entries on the hardware queue
345  *
346  * Init the fence driver for the requested ring (all asics).
347  * Helper function for amdgpu_fence_driver_init().
348  */
349 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
350                                   unsigned num_hw_submission)
351 {
352         long timeout;
353         int r;
354
355         /* Check that num_hw_submission is a power of two */
356         if ((num_hw_submission & (num_hw_submission - 1)) != 0)
357                 return -EINVAL;
358
359         ring->fence_drv.cpu_addr = NULL;
360         ring->fence_drv.gpu_addr = 0;
361         ring->fence_drv.sync_seq = 0;
362         atomic_set(&ring->fence_drv.last_seq, 0);
363         ring->fence_drv.initialized = false;
364
365         setup_timer(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback,
366                     (unsigned long)ring);
367
368         ring->fence_drv.num_fences_mask = num_hw_submission * 2 - 1;
369         spin_lock_init(&ring->fence_drv.lock);
370         ring->fence_drv.fences = kcalloc(num_hw_submission * 2, sizeof(void *),
371                                          GFP_KERNEL);
372         if (!ring->fence_drv.fences)
373                 return -ENOMEM;
374
375         timeout = msecs_to_jiffies(amdgpu_lockup_timeout);
376         if (timeout == 0) {
377                 /*
378                  * FIXME:
379                  * Delayed workqueue cannot use it directly,
380                  * so the scheduler will not use delayed workqueue if
381                  * MAX_SCHEDULE_TIMEOUT is set.
382                  * Currently keep it simple and silly.
383                  */
384                 timeout = MAX_SCHEDULE_TIMEOUT;
385         }
386         r = amd_sched_init(&ring->sched, &amdgpu_sched_ops,
387                            num_hw_submission,
388                            timeout, ring->name);
389         if (r) {
390                 DRM_ERROR("Failed to create scheduler on ring %s.\n",
391                           ring->name);
392                 return r;
393         }
394
395         return 0;
396 }
397
398 /**
399  * amdgpu_fence_driver_init - init the fence driver
400  * for all possible rings.
401  *
402  * @adev: amdgpu device pointer
403  *
404  * Init the fence driver for all possible rings (all asics).
405  * Not all asics have all rings, so each asic will only
406  * start the fence driver on the rings it has using
407  * amdgpu_fence_driver_start_ring().
408  * Returns 0 for success.
409  */
410 int amdgpu_fence_driver_init(struct amdgpu_device *adev)
411 {
412         if (amdgpu_debugfs_fence_init(adev))
413                 dev_err(adev->dev, "fence debugfs file creation failed\n");
414
415         return 0;
416 }
417
418 /**
419  * amdgpu_fence_driver_fini - tear down the fence driver
420  * for all possible rings.
421  *
422  * @adev: amdgpu device pointer
423  *
424  * Tear down the fence driver for all possible rings (all asics).
425  */
426 void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
427 {
428         unsigned i, j;
429         int r;
430
431         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
432                 struct amdgpu_ring *ring = adev->rings[i];
433
434                 if (!ring || !ring->fence_drv.initialized)
435                         continue;
436                 r = amdgpu_fence_wait_empty(ring);
437                 if (r) {
438                         /* no need to trigger GPU reset as we are unloading */
439                         amdgpu_fence_driver_force_completion(adev);
440                 }
441                 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
442                                ring->fence_drv.irq_type);
443                 amd_sched_fini(&ring->sched);
444                 del_timer_sync(&ring->fence_drv.fallback_timer);
445                 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
446                         fence_put(ring->fence_drv.fences[j]);
447                 kfree(ring->fence_drv.fences);
448                 ring->fence_drv.initialized = false;
449         }
450 }
451
452 /**
453  * amdgpu_fence_driver_suspend - suspend the fence driver
454  * for all possible rings.
455  *
456  * @adev: amdgpu device pointer
457  *
458  * Suspend the fence driver for all possible rings (all asics).
459  */
460 void amdgpu_fence_driver_suspend(struct amdgpu_device *adev)
461 {
462         int i, r;
463
464         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
465                 struct amdgpu_ring *ring = adev->rings[i];
466                 if (!ring || !ring->fence_drv.initialized)
467                         continue;
468
469                 /* wait for gpu to finish processing current batch */
470                 r = amdgpu_fence_wait_empty(ring);
471                 if (r) {
472                         /* delay GPU reset to resume */
473                         amdgpu_fence_driver_force_completion(adev);
474                 }
475
476                 /* disable the interrupt */
477                 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
478                                ring->fence_drv.irq_type);
479         }
480 }
481
482 /**
483  * amdgpu_fence_driver_resume - resume the fence driver
484  * for all possible rings.
485  *
486  * @adev: amdgpu device pointer
487  *
488  * Resume the fence driver for all possible rings (all asics).
489  * Not all asics have all rings, so each asic will only
490  * start the fence driver on the rings it has using
491  * amdgpu_fence_driver_start_ring().
492  * Returns 0 for success.
493  */
494 void amdgpu_fence_driver_resume(struct amdgpu_device *adev)
495 {
496         int i;
497
498         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
499                 struct amdgpu_ring *ring = adev->rings[i];
500                 if (!ring || !ring->fence_drv.initialized)
501                         continue;
502
503                 /* enable the interrupt */
504                 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
505                                ring->fence_drv.irq_type);
506         }
507 }
508
509 /**
510  * amdgpu_fence_driver_force_completion - force all fence waiter to complete
511  *
512  * @adev: amdgpu device pointer
513  *
514  * In case of GPU reset failure make sure no process keep waiting on fence
515  * that will never complete.
516  */
517 void amdgpu_fence_driver_force_completion(struct amdgpu_device *adev)
518 {
519         int i;
520
521         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
522                 struct amdgpu_ring *ring = adev->rings[i];
523                 if (!ring || !ring->fence_drv.initialized)
524                         continue;
525
526                 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
527         }
528 }
529
530 /*
531  * Common fence implementation
532  */
533
534 static const char *amdgpu_fence_get_driver_name(struct fence *fence)
535 {
536         return "amdgpu";
537 }
538
539 static const char *amdgpu_fence_get_timeline_name(struct fence *f)
540 {
541         struct amdgpu_fence *fence = to_amdgpu_fence(f);
542         return (const char *)fence->ring->name;
543 }
544
545 /**
546  * amdgpu_fence_enable_signaling - enable signalling on fence
547  * @fence: fence
548  *
549  * This function is called with fence_queue lock held, and adds a callback
550  * to fence_queue that checks if this fence is signaled, and if so it
551  * signals the fence and removes itself.
552  */
553 static bool amdgpu_fence_enable_signaling(struct fence *f)
554 {
555         struct amdgpu_fence *fence = to_amdgpu_fence(f);
556         struct amdgpu_ring *ring = fence->ring;
557
558         if (!timer_pending(&ring->fence_drv.fallback_timer))
559                 amdgpu_fence_schedule_fallback(ring);
560
561         FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
562
563         return true;
564 }
565
566 /**
567  * amdgpu_fence_free - free up the fence memory
568  *
569  * @rcu: RCU callback head
570  *
571  * Free up the fence memory after the RCU grace period.
572  */
573 static void amdgpu_fence_free(struct rcu_head *rcu)
574 {
575         struct fence *f = container_of(rcu, struct fence, rcu);
576         struct amdgpu_fence *fence = to_amdgpu_fence(f);
577         kmem_cache_free(amdgpu_fence_slab, fence);
578 }
579
580 /**
581  * amdgpu_fence_release - callback that fence can be freed
582  *
583  * @fence: fence
584  *
585  * This function is called when the reference count becomes zero.
586  * It just RCU schedules freeing up the fence.
587  */
588 static void amdgpu_fence_release(struct fence *f)
589 {
590         call_rcu(&f->rcu, amdgpu_fence_free);
591 }
592
593 static const struct fence_ops amdgpu_fence_ops = {
594         .get_driver_name = amdgpu_fence_get_driver_name,
595         .get_timeline_name = amdgpu_fence_get_timeline_name,
596         .enable_signaling = amdgpu_fence_enable_signaling,
597         .wait = fence_default_wait,
598         .release = amdgpu_fence_release,
599 };
600
601 /*
602  * Fence debugfs
603  */
604 #if defined(CONFIG_DEBUG_FS)
605 static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data)
606 {
607         struct drm_info_node *node = (struct drm_info_node *)m->private;
608         struct drm_device *dev = node->minor->dev;
609         struct amdgpu_device *adev = dev->dev_private;
610         int i;
611
612         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
613                 struct amdgpu_ring *ring = adev->rings[i];
614                 if (!ring || !ring->fence_drv.initialized)
615                         continue;
616
617                 amdgpu_fence_process(ring);
618
619                 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
620                 seq_printf(m, "Last signaled fence 0x%08x\n",
621                            atomic_read(&ring->fence_drv.last_seq));
622                 seq_printf(m, "Last emitted        0x%08x\n",
623                            ring->fence_drv.sync_seq);
624         }
625         return 0;
626 }
627
628 /**
629  * amdgpu_debugfs_gpu_reset - manually trigger a gpu reset
630  *
631  * Manually trigger a gpu reset at the next fence wait.
632  */
633 static int amdgpu_debugfs_gpu_reset(struct seq_file *m, void *data)
634 {
635         struct drm_info_node *node = (struct drm_info_node *) m->private;
636         struct drm_device *dev = node->minor->dev;
637         struct amdgpu_device *adev = dev->dev_private;
638
639         seq_printf(m, "gpu reset\n");
640         amdgpu_gpu_reset(adev);
641
642         return 0;
643 }
644
645 static const struct drm_info_list amdgpu_debugfs_fence_list[] = {
646         {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
647         {"amdgpu_gpu_reset", &amdgpu_debugfs_gpu_reset, 0, NULL}
648 };
649 #endif
650
651 int amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
652 {
653 #if defined(CONFIG_DEBUG_FS)
654         return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 2);
655 #else
656         return 0;
657 #endif
658 }
659