Merge tag 'omap-for-v4.6/dt-ti81xx-signed' of git://git.kernel.org/pub/scm/linux...
[linux-2.6-block.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 static atomic_t amdgpu_fence_slab_ref = ATOMIC_INIT(0);
59
60 /*
61  * Cast helper
62  */
63 static const struct fence_ops amdgpu_fence_ops;
64 static inline struct amdgpu_fence *to_amdgpu_fence(struct fence *f)
65 {
66         struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base);
67
68         if (__f->base.ops == &amdgpu_fence_ops)
69                 return __f;
70
71         return NULL;
72 }
73
74 /**
75  * amdgpu_fence_write - write a fence value
76  *
77  * @ring: ring the fence is associated with
78  * @seq: sequence number to write
79  *
80  * Writes a fence value to memory (all asics).
81  */
82 static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq)
83 {
84         struct amdgpu_fence_driver *drv = &ring->fence_drv;
85
86         if (drv->cpu_addr)
87                 *drv->cpu_addr = cpu_to_le32(seq);
88 }
89
90 /**
91  * amdgpu_fence_read - read a fence value
92  *
93  * @ring: ring the fence is associated with
94  *
95  * Reads a fence value from memory (all asics).
96  * Returns the value of the fence read from memory.
97  */
98 static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
99 {
100         struct amdgpu_fence_driver *drv = &ring->fence_drv;
101         u32 seq = 0;
102
103         if (drv->cpu_addr)
104                 seq = le32_to_cpu(*drv->cpu_addr);
105         else
106                 seq = atomic_read(&drv->last_seq);
107
108         return seq;
109 }
110
111 /**
112  * amdgpu_fence_emit - emit a fence on the requested ring
113  *
114  * @ring: ring the fence is associated with
115  * @f: resulting fence object
116  *
117  * Emits a fence command on the requested ring (all asics).
118  * Returns 0 on success, -ENOMEM on failure.
119  */
120 int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **f)
121 {
122         struct amdgpu_device *adev = ring->adev;
123         struct amdgpu_fence *fence;
124         struct fence **ptr;
125         uint32_t seq;
126
127         fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
128         if (fence == NULL)
129                 return -ENOMEM;
130
131         seq = ++ring->fence_drv.sync_seq;
132         fence->ring = ring;
133         fence_init(&fence->base, &amdgpu_fence_ops,
134                    &ring->fence_drv.lock,
135                    adev->fence_context + ring->idx,
136                    seq);
137         amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
138                                seq, AMDGPU_FENCE_FLAG_INT);
139
140         ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
141         /* This function can't be called concurrently anyway, otherwise
142          * emitting the fence would mess up the hardware ring buffer.
143          */
144         BUG_ON(rcu_dereference_protected(*ptr, 1));
145
146         rcu_assign_pointer(*ptr, fence_get(&fence->base));
147
148         *f = &fence->base;
149
150         return 0;
151 }
152
153 /**
154  * amdgpu_fence_schedule_fallback - schedule fallback check
155  *
156  * @ring: pointer to struct amdgpu_ring
157  *
158  * Start a timer as fallback to our interrupts.
159  */
160 static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring)
161 {
162         mod_timer(&ring->fence_drv.fallback_timer,
163                   jiffies + AMDGPU_FENCE_JIFFIES_TIMEOUT);
164 }
165
166 /**
167  * amdgpu_fence_process - check for fence activity
168  *
169  * @ring: pointer to struct amdgpu_ring
170  *
171  * Checks the current fence value and calculates the last
172  * signalled fence value. Wakes the fence queue if the
173  * sequence number has increased.
174  */
175 void amdgpu_fence_process(struct amdgpu_ring *ring)
176 {
177         struct amdgpu_fence_driver *drv = &ring->fence_drv;
178         uint32_t seq, last_seq;
179         int r;
180
181         do {
182                 last_seq = atomic_read(&ring->fence_drv.last_seq);
183                 seq = amdgpu_fence_read(ring);
184
185         } while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
186
187         if (seq != ring->fence_drv.sync_seq)
188                 amdgpu_fence_schedule_fallback(ring);
189
190         while (last_seq != seq) {
191                 struct fence *fence, **ptr;
192
193                 ptr = &drv->fences[++last_seq & drv->num_fences_mask];
194
195                 /* There is always exactly one thread signaling this fence slot */
196                 fence = rcu_dereference_protected(*ptr, 1);
197                 rcu_assign_pointer(*ptr, NULL);
198
199                 BUG_ON(!fence);
200
201                 r = fence_signal(fence);
202                 if (!r)
203                         FENCE_TRACE(fence, "signaled from irq context\n");
204                 else
205                         BUG();
206
207                 fence_put(fence);
208         }
209 }
210
211 /**
212  * amdgpu_fence_fallback - fallback for hardware interrupts
213  *
214  * @work: delayed work item
215  *
216  * Checks for fence activity.
217  */
218 static void amdgpu_fence_fallback(unsigned long arg)
219 {
220         struct amdgpu_ring *ring = (void *)arg;
221
222         amdgpu_fence_process(ring);
223 }
224
225 /**
226  * amdgpu_fence_wait_empty - wait for all fences to signal
227  *
228  * @adev: amdgpu device pointer
229  * @ring: ring index the fence is associated with
230  *
231  * Wait for all fences on the requested ring to signal (all asics).
232  * Returns 0 if the fences have passed, error for all other cases.
233  */
234 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
235 {
236         uint64_t seq = ACCESS_ONCE(ring->fence_drv.sync_seq);
237         struct fence *fence, **ptr;
238         int r;
239
240         if (!seq)
241                 return 0;
242
243         ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
244         rcu_read_lock();
245         fence = rcu_dereference(*ptr);
246         if (!fence || !fence_get_rcu(fence)) {
247                 rcu_read_unlock();
248                 return 0;
249         }
250         rcu_read_unlock();
251
252         r = fence_wait(fence, false);
253         fence_put(fence);
254         return r;
255 }
256
257 /**
258  * amdgpu_fence_count_emitted - get the count of emitted fences
259  *
260  * @ring: ring the fence is associated with
261  *
262  * Get the number of fences emitted on the requested ring (all asics).
263  * Returns the number of emitted fences on the ring.  Used by the
264  * dynpm code to ring track activity.
265  */
266 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
267 {
268         uint64_t emitted;
269
270         /* We are not protected by ring lock when reading the last sequence
271          * but it's ok to report slightly wrong fence count here.
272          */
273         amdgpu_fence_process(ring);
274         emitted = 0x100000000ull;
275         emitted -= atomic_read(&ring->fence_drv.last_seq);
276         emitted += ACCESS_ONCE(ring->fence_drv.sync_seq);
277         return lower_32_bits(emitted);
278 }
279
280 /**
281  * amdgpu_fence_driver_start_ring - make the fence driver
282  * ready for use on the requested ring.
283  *
284  * @ring: ring to start the fence driver on
285  * @irq_src: interrupt source to use for this ring
286  * @irq_type: interrupt type to use for this ring
287  *
288  * Make the fence driver ready for processing (all asics).
289  * Not all asics have all rings, so each asic will only
290  * start the fence driver on the rings it has.
291  * Returns 0 for success, errors for failure.
292  */
293 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
294                                    struct amdgpu_irq_src *irq_src,
295                                    unsigned irq_type)
296 {
297         struct amdgpu_device *adev = ring->adev;
298         uint64_t index;
299
300         if (ring != &adev->uvd.ring) {
301                 ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs];
302                 ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4);
303         } else {
304                 /* put fence directly behind firmware */
305                 index = ALIGN(adev->uvd.fw->size, 8);
306                 ring->fence_drv.cpu_addr = adev->uvd.cpu_addr + index;
307                 ring->fence_drv.gpu_addr = adev->uvd.gpu_addr + index;
308         }
309         amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
310         amdgpu_irq_get(adev, irq_src, irq_type);
311
312         ring->fence_drv.irq_src = irq_src;
313         ring->fence_drv.irq_type = irq_type;
314         ring->fence_drv.initialized = true;
315
316         dev_info(adev->dev, "fence driver on ring %d use gpu addr 0x%016llx, "
317                  "cpu addr 0x%p\n", ring->idx,
318                  ring->fence_drv.gpu_addr, ring->fence_drv.cpu_addr);
319         return 0;
320 }
321
322 /**
323  * amdgpu_fence_driver_init_ring - init the fence driver
324  * for the requested ring.
325  *
326  * @ring: ring to init the fence driver on
327  * @num_hw_submission: number of entries on the hardware queue
328  *
329  * Init the fence driver for the requested ring (all asics).
330  * Helper function for amdgpu_fence_driver_init().
331  */
332 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
333                                   unsigned num_hw_submission)
334 {
335         long timeout;
336         int r;
337
338         /* Check that num_hw_submission is a power of two */
339         if ((num_hw_submission & (num_hw_submission - 1)) != 0)
340                 return -EINVAL;
341
342         ring->fence_drv.cpu_addr = NULL;
343         ring->fence_drv.gpu_addr = 0;
344         ring->fence_drv.sync_seq = 0;
345         atomic_set(&ring->fence_drv.last_seq, 0);
346         ring->fence_drv.initialized = false;
347
348         setup_timer(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback,
349                     (unsigned long)ring);
350
351         ring->fence_drv.num_fences_mask = num_hw_submission - 1;
352         spin_lock_init(&ring->fence_drv.lock);
353         ring->fence_drv.fences = kcalloc(num_hw_submission, sizeof(void *),
354                                          GFP_KERNEL);
355         if (!ring->fence_drv.fences)
356                 return -ENOMEM;
357
358         timeout = msecs_to_jiffies(amdgpu_lockup_timeout);
359         if (timeout == 0) {
360                 /*
361                  * FIXME:
362                  * Delayed workqueue cannot use it directly,
363                  * so the scheduler will not use delayed workqueue if
364                  * MAX_SCHEDULE_TIMEOUT is set.
365                  * Currently keep it simple and silly.
366                  */
367                 timeout = MAX_SCHEDULE_TIMEOUT;
368         }
369         r = amd_sched_init(&ring->sched, &amdgpu_sched_ops,
370                            num_hw_submission,
371                            timeout, ring->name);
372         if (r) {
373                 DRM_ERROR("Failed to create scheduler on ring %s.\n",
374                           ring->name);
375                 return r;
376         }
377
378         return 0;
379 }
380
381 /**
382  * amdgpu_fence_driver_init - init the fence driver
383  * for all possible rings.
384  *
385  * @adev: amdgpu device pointer
386  *
387  * Init the fence driver for all possible rings (all asics).
388  * Not all asics have all rings, so each asic will only
389  * start the fence driver on the rings it has using
390  * amdgpu_fence_driver_start_ring().
391  * Returns 0 for success.
392  */
393 int amdgpu_fence_driver_init(struct amdgpu_device *adev)
394 {
395         if (atomic_inc_return(&amdgpu_fence_slab_ref) == 1) {
396                 amdgpu_fence_slab = kmem_cache_create(
397                         "amdgpu_fence", sizeof(struct amdgpu_fence), 0,
398                         SLAB_HWCACHE_ALIGN, NULL);
399                 if (!amdgpu_fence_slab)
400                         return -ENOMEM;
401         }
402         if (amdgpu_debugfs_fence_init(adev))
403                 dev_err(adev->dev, "fence debugfs file creation failed\n");
404
405         return 0;
406 }
407
408 /**
409  * amdgpu_fence_driver_fini - tear down the fence driver
410  * for all possible rings.
411  *
412  * @adev: amdgpu device pointer
413  *
414  * Tear down the fence driver for all possible rings (all asics).
415  */
416 void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
417 {
418         unsigned i, j;
419         int r;
420
421         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
422                 struct amdgpu_ring *ring = adev->rings[i];
423
424                 if (!ring || !ring->fence_drv.initialized)
425                         continue;
426                 r = amdgpu_fence_wait_empty(ring);
427                 if (r) {
428                         /* no need to trigger GPU reset as we are unloading */
429                         amdgpu_fence_driver_force_completion(adev);
430                 }
431                 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
432                                ring->fence_drv.irq_type);
433                 amd_sched_fini(&ring->sched);
434                 del_timer_sync(&ring->fence_drv.fallback_timer);
435                 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
436                         fence_put(ring->fence_drv.fences[i]);
437                 kfree(ring->fence_drv.fences);
438                 ring->fence_drv.initialized = false;
439         }
440
441         if (atomic_dec_and_test(&amdgpu_fence_slab_ref))
442                 kmem_cache_destroy(amdgpu_fence_slab);
443 }
444
445 /**
446  * amdgpu_fence_driver_suspend - suspend the fence driver
447  * for all possible rings.
448  *
449  * @adev: amdgpu device pointer
450  *
451  * Suspend the fence driver for all possible rings (all asics).
452  */
453 void amdgpu_fence_driver_suspend(struct amdgpu_device *adev)
454 {
455         int i, r;
456
457         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
458                 struct amdgpu_ring *ring = adev->rings[i];
459                 if (!ring || !ring->fence_drv.initialized)
460                         continue;
461
462                 /* wait for gpu to finish processing current batch */
463                 r = amdgpu_fence_wait_empty(ring);
464                 if (r) {
465                         /* delay GPU reset to resume */
466                         amdgpu_fence_driver_force_completion(adev);
467                 }
468
469                 /* disable the interrupt */
470                 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
471                                ring->fence_drv.irq_type);
472         }
473 }
474
475 /**
476  * amdgpu_fence_driver_resume - resume the fence driver
477  * for all possible rings.
478  *
479  * @adev: amdgpu device pointer
480  *
481  * Resume the fence driver for all possible rings (all asics).
482  * Not all asics have all rings, so each asic will only
483  * start the fence driver on the rings it has using
484  * amdgpu_fence_driver_start_ring().
485  * Returns 0 for success.
486  */
487 void amdgpu_fence_driver_resume(struct amdgpu_device *adev)
488 {
489         int i;
490
491         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
492                 struct amdgpu_ring *ring = adev->rings[i];
493                 if (!ring || !ring->fence_drv.initialized)
494                         continue;
495
496                 /* enable the interrupt */
497                 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
498                                ring->fence_drv.irq_type);
499         }
500 }
501
502 /**
503  * amdgpu_fence_driver_force_completion - force all fence waiter to complete
504  *
505  * @adev: amdgpu device pointer
506  *
507  * In case of GPU reset failure make sure no process keep waiting on fence
508  * that will never complete.
509  */
510 void amdgpu_fence_driver_force_completion(struct amdgpu_device *adev)
511 {
512         int i;
513
514         for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
515                 struct amdgpu_ring *ring = adev->rings[i];
516                 if (!ring || !ring->fence_drv.initialized)
517                         continue;
518
519                 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
520         }
521 }
522
523 /*
524  * Common fence implementation
525  */
526
527 static const char *amdgpu_fence_get_driver_name(struct fence *fence)
528 {
529         return "amdgpu";
530 }
531
532 static const char *amdgpu_fence_get_timeline_name(struct fence *f)
533 {
534         struct amdgpu_fence *fence = to_amdgpu_fence(f);
535         return (const char *)fence->ring->name;
536 }
537
538 /**
539  * amdgpu_fence_enable_signaling - enable signalling on fence
540  * @fence: fence
541  *
542  * This function is called with fence_queue lock held, and adds a callback
543  * to fence_queue that checks if this fence is signaled, and if so it
544  * signals the fence and removes itself.
545  */
546 static bool amdgpu_fence_enable_signaling(struct fence *f)
547 {
548         struct amdgpu_fence *fence = to_amdgpu_fence(f);
549         struct amdgpu_ring *ring = fence->ring;
550
551         if (!timer_pending(&ring->fence_drv.fallback_timer))
552                 amdgpu_fence_schedule_fallback(ring);
553
554         FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
555
556         return true;
557 }
558
559 /**
560  * amdgpu_fence_free - free up the fence memory
561  *
562  * @rcu: RCU callback head
563  *
564  * Free up the fence memory after the RCU grace period.
565  */
566 static void amdgpu_fence_free(struct rcu_head *rcu)
567 {
568         struct fence *f = container_of(rcu, struct fence, rcu);
569         struct amdgpu_fence *fence = to_amdgpu_fence(f);
570         kmem_cache_free(amdgpu_fence_slab, fence);
571 }
572
573 /**
574  * amdgpu_fence_release - callback that fence can be freed
575  *
576  * @fence: fence
577  *
578  * This function is called when the reference count becomes zero.
579  * It just RCU schedules freeing up the fence.
580  */
581 static void amdgpu_fence_release(struct fence *f)
582 {
583         call_rcu(&f->rcu, amdgpu_fence_free);
584 }
585
586 static const struct fence_ops amdgpu_fence_ops = {
587         .get_driver_name = amdgpu_fence_get_driver_name,
588         .get_timeline_name = amdgpu_fence_get_timeline_name,
589         .enable_signaling = amdgpu_fence_enable_signaling,
590         .wait = fence_default_wait,
591         .release = amdgpu_fence_release,
592 };
593
594 /*
595  * Fence debugfs
596  */
597 #if defined(CONFIG_DEBUG_FS)
598 static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data)
599 {
600         struct drm_info_node *node = (struct drm_info_node *)m->private;
601         struct drm_device *dev = node->minor->dev;
602         struct amdgpu_device *adev = dev->dev_private;
603         int i;
604
605         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
606                 struct amdgpu_ring *ring = adev->rings[i];
607                 if (!ring || !ring->fence_drv.initialized)
608                         continue;
609
610                 amdgpu_fence_process(ring);
611
612                 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
613                 seq_printf(m, "Last signaled fence 0x%08x\n",
614                            atomic_read(&ring->fence_drv.last_seq));
615                 seq_printf(m, "Last emitted        0x%08x\n",
616                            ring->fence_drv.sync_seq);
617         }
618         return 0;
619 }
620
621 /**
622  * amdgpu_debugfs_gpu_reset - manually trigger a gpu reset
623  *
624  * Manually trigger a gpu reset at the next fence wait.
625  */
626 static int amdgpu_debugfs_gpu_reset(struct seq_file *m, void *data)
627 {
628         struct drm_info_node *node = (struct drm_info_node *) m->private;
629         struct drm_device *dev = node->minor->dev;
630         struct amdgpu_device *adev = dev->dev_private;
631
632         seq_printf(m, "gpu reset\n");
633         amdgpu_gpu_reset(adev);
634
635         return 0;
636 }
637
638 static struct drm_info_list amdgpu_debugfs_fence_list[] = {
639         {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
640         {"amdgpu_gpu_reset", &amdgpu_debugfs_gpu_reset, 0, NULL}
641 };
642 #endif
643
644 int amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
645 {
646 #if defined(CONFIG_DEBUG_FS)
647         return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 2);
648 #else
649         return 0;
650 #endif
651 }
652