]> rtime.felk.cvut.cz Git - linux-imx.git/blob - drivers/gpu/drm/radeon/radeon_ring.c
drm/radeon: implement ring saving on reset v4
[linux-imx.git] / drivers / gpu / drm / radeon / radeon_ring.c
1 /*
2  * Copyright 2008 Advanced Micro Devices, Inc.
3  * Copyright 2008 Red Hat Inc.
4  * Copyright 2009 Jerome Glisse.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the "Software"),
8  * to deal in the Software without restriction, including without limitation
9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10  * and/or sell copies of the Software, and to permit persons to whom the
11  * Software is furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22  * OTHER DEALINGS IN THE SOFTWARE.
23  *
24  * Authors: Dave Airlie
25  *          Alex Deucher
26  *          Jerome Glisse
27  *          Christian König
28  */
29 #include <linux/seq_file.h>
30 #include <linux/slab.h>
31 #include "drmP.h"
32 #include "radeon_drm.h"
33 #include "radeon_reg.h"
34 #include "radeon.h"
35 #include "atom.h"
36
37 /*
38  * IB.
39  */
40 int radeon_debugfs_sa_init(struct radeon_device *rdev);
41
42 int radeon_ib_get(struct radeon_device *rdev, int ring,
43                   struct radeon_ib *ib, unsigned size)
44 {
45         int i, r;
46
47         r = radeon_sa_bo_new(rdev, &rdev->ring_tmp_bo, &ib->sa_bo, size, 256, true);
48         if (r) {
49                 dev_err(rdev->dev, "failed to get a new IB (%d)\n", r);
50                 return r;
51         }
52
53         r = radeon_semaphore_create(rdev, &ib->semaphore);
54         if (r) {
55                 return r;
56         }
57
58         ib->ring = ring;
59         ib->fence = NULL;
60         ib->ptr = radeon_sa_bo_cpu_addr(ib->sa_bo);
61         ib->gpu_addr = radeon_sa_bo_gpu_addr(ib->sa_bo);
62         ib->vm_id = 0;
63         ib->is_const_ib = false;
64         for (i = 0; i < RADEON_NUM_RINGS; ++i)
65                 ib->sync_to[i] = NULL;
66
67         return 0;
68 }
69
70 void radeon_ib_free(struct radeon_device *rdev, struct radeon_ib *ib)
71 {
72         radeon_semaphore_free(rdev, &ib->semaphore, ib->fence);
73         radeon_sa_bo_free(rdev, &ib->sa_bo, ib->fence);
74         radeon_fence_unref(&ib->fence);
75 }
76
77 int radeon_ib_schedule(struct radeon_device *rdev, struct radeon_ib *ib)
78 {
79         struct radeon_ring *ring = &rdev->ring[ib->ring];
80         bool need_sync = false;
81         int i, r = 0;
82
83         if (!ib->length_dw || !ring->ready) {
84                 /* TODO: Nothings in the ib we should report. */
85                 dev_err(rdev->dev, "couldn't schedule ib\n");
86                 return -EINVAL;
87         }
88
89         /* 64 dwords should be enough for fence too */
90         r = radeon_ring_lock(rdev, ring, 64 + RADEON_NUM_RINGS * 8);
91         if (r) {
92                 dev_err(rdev->dev, "scheduling IB failed (%d).\n", r);
93                 return r;
94         }
95         for (i = 0; i < RADEON_NUM_RINGS; ++i) {
96                 struct radeon_fence *fence = ib->sync_to[i];
97                 if (radeon_fence_need_sync(fence, ib->ring)) {
98                         need_sync = true;
99                         radeon_semaphore_sync_rings(rdev, ib->semaphore,
100                                                     fence->ring, ib->ring);
101                         radeon_fence_note_sync(fence, ib->ring);
102                 }
103         }
104         /* immediately free semaphore when we don't need to sync */
105         if (!need_sync) {
106                 radeon_semaphore_free(rdev, &ib->semaphore, NULL);
107         }
108         radeon_ring_ib_execute(rdev, ib->ring, ib);
109         r = radeon_fence_emit(rdev, &ib->fence, ib->ring);
110         if (r) {
111                 dev_err(rdev->dev, "failed to emit fence for new IB (%d)\n", r);
112                 radeon_ring_unlock_undo(rdev, ring);
113                 return r;
114         }
115         radeon_ring_unlock_commit(rdev, ring);
116         return 0;
117 }
118
119 int radeon_ib_pool_init(struct radeon_device *rdev)
120 {
121         int r;
122
123         if (rdev->ib_pool_ready) {
124                 return 0;
125         }
126         r = radeon_sa_bo_manager_init(rdev, &rdev->ring_tmp_bo,
127                                       RADEON_IB_POOL_SIZE*64*1024,
128                                       RADEON_GEM_DOMAIN_GTT);
129         if (r) {
130                 return r;
131         }
132
133         r = radeon_sa_bo_manager_start(rdev, &rdev->ring_tmp_bo);
134         if (r) {
135                 return r;
136         }
137
138         rdev->ib_pool_ready = true;
139         if (radeon_debugfs_sa_init(rdev)) {
140                 dev_err(rdev->dev, "failed to register debugfs file for SA\n");
141         }
142         return 0;
143 }
144
145 void radeon_ib_pool_fini(struct radeon_device *rdev)
146 {
147         if (rdev->ib_pool_ready) {
148                 radeon_sa_bo_manager_suspend(rdev, &rdev->ring_tmp_bo);
149                 radeon_sa_bo_manager_fini(rdev, &rdev->ring_tmp_bo);
150                 rdev->ib_pool_ready = false;
151         }
152 }
153
154 int radeon_ib_ring_tests(struct radeon_device *rdev)
155 {
156         unsigned i;
157         int r;
158
159         for (i = 0; i < RADEON_NUM_RINGS; ++i) {
160                 struct radeon_ring *ring = &rdev->ring[i];
161
162                 if (!ring->ready)
163                         continue;
164
165                 r = radeon_ib_test(rdev, i, ring);
166                 if (r) {
167                         ring->ready = false;
168
169                         if (i == RADEON_RING_TYPE_GFX_INDEX) {
170                                 /* oh, oh, that's really bad */
171                                 DRM_ERROR("radeon: failed testing IB on GFX ring (%d).\n", r);
172                                 rdev->accel_working = false;
173                                 return r;
174
175                         } else {
176                                 /* still not good, but we can live with it */
177                                 DRM_ERROR("radeon: failed testing IB on ring %d (%d).\n", i, r);
178                         }
179                 }
180         }
181         return 0;
182 }
183
184 /*
185  * Ring.
186  */
187 int radeon_debugfs_ring_init(struct radeon_device *rdev, struct radeon_ring *ring);
188
189 void radeon_ring_write(struct radeon_ring *ring, uint32_t v)
190 {
191 #if DRM_DEBUG_CODE
192         if (ring->count_dw <= 0) {
193                 DRM_ERROR("radeon: writting more dword to ring than expected !\n");
194         }
195 #endif
196         ring->ring[ring->wptr++] = v;
197         ring->wptr &= ring->ptr_mask;
198         ring->count_dw--;
199         ring->ring_free_dw--;
200 }
201
202 int radeon_ring_index(struct radeon_device *rdev, struct radeon_ring *ring)
203 {
204         /* r1xx-r5xx only has CP ring */
205         if (rdev->family < CHIP_R600)
206                 return RADEON_RING_TYPE_GFX_INDEX;
207
208         if (rdev->family >= CHIP_CAYMAN) {
209                 if (ring == &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX])
210                         return CAYMAN_RING_TYPE_CP1_INDEX;
211                 else if (ring == &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX])
212                         return CAYMAN_RING_TYPE_CP2_INDEX;
213         }
214         return RADEON_RING_TYPE_GFX_INDEX;
215 }
216
217 void radeon_ring_free_size(struct radeon_device *rdev, struct radeon_ring *ring)
218 {
219         u32 rptr;
220
221         if (rdev->wb.enabled)
222                 rptr = le32_to_cpu(rdev->wb.wb[ring->rptr_offs/4]);
223         else
224                 rptr = RREG32(ring->rptr_reg);
225         ring->rptr = (rptr & ring->ptr_reg_mask) >> ring->ptr_reg_shift;
226         /* This works because ring_size is a power of 2 */
227         ring->ring_free_dw = (ring->rptr + (ring->ring_size / 4));
228         ring->ring_free_dw -= ring->wptr;
229         ring->ring_free_dw &= ring->ptr_mask;
230         if (!ring->ring_free_dw) {
231                 ring->ring_free_dw = ring->ring_size / 4;
232         }
233 }
234
235
236 int radeon_ring_alloc(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ndw)
237 {
238         int r;
239
240         /* Align requested size with padding so unlock_commit can
241          * pad safely */
242         ndw = (ndw + ring->align_mask) & ~ring->align_mask;
243         while (ndw > (ring->ring_free_dw - 1)) {
244                 radeon_ring_free_size(rdev, ring);
245                 if (ndw < ring->ring_free_dw) {
246                         break;
247                 }
248                 r = radeon_fence_wait_next_locked(rdev, radeon_ring_index(rdev, ring));
249                 if (r)
250                         return r;
251         }
252         ring->count_dw = ndw;
253         ring->wptr_old = ring->wptr;
254         return 0;
255 }
256
257 int radeon_ring_lock(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ndw)
258 {
259         int r;
260
261         mutex_lock(&rdev->ring_lock);
262         r = radeon_ring_alloc(rdev, ring, ndw);
263         if (r) {
264                 mutex_unlock(&rdev->ring_lock);
265                 return r;
266         }
267         return 0;
268 }
269
270 void radeon_ring_commit(struct radeon_device *rdev, struct radeon_ring *ring)
271 {
272         /* We pad to match fetch size */
273         while (ring->wptr & ring->align_mask) {
274                 radeon_ring_write(ring, ring->nop);
275         }
276         DRM_MEMORYBARRIER();
277         WREG32(ring->wptr_reg, (ring->wptr << ring->ptr_reg_shift) & ring->ptr_reg_mask);
278         (void)RREG32(ring->wptr_reg);
279 }
280
281 void radeon_ring_unlock_commit(struct radeon_device *rdev, struct radeon_ring *ring)
282 {
283         radeon_ring_commit(rdev, ring);
284         mutex_unlock(&rdev->ring_lock);
285 }
286
287 void radeon_ring_undo(struct radeon_ring *ring)
288 {
289         ring->wptr = ring->wptr_old;
290 }
291
292 void radeon_ring_unlock_undo(struct radeon_device *rdev, struct radeon_ring *ring)
293 {
294         radeon_ring_undo(ring);
295         mutex_unlock(&rdev->ring_lock);
296 }
297
298 void radeon_ring_force_activity(struct radeon_device *rdev, struct radeon_ring *ring)
299 {
300         int r;
301
302         radeon_ring_free_size(rdev, ring);
303         if (ring->rptr == ring->wptr) {
304                 r = radeon_ring_alloc(rdev, ring, 1);
305                 if (!r) {
306                         radeon_ring_write(ring, ring->nop);
307                         radeon_ring_commit(rdev, ring);
308                 }
309         }
310 }
311
312 void radeon_ring_lockup_update(struct radeon_ring *ring)
313 {
314         ring->last_rptr = ring->rptr;
315         ring->last_activity = jiffies;
316 }
317
318 /**
319  * radeon_ring_test_lockup() - check if ring is lockedup by recording information
320  * @rdev:       radeon device structure
321  * @ring:       radeon_ring structure holding ring information
322  *
323  * We don't need to initialize the lockup tracking information as we will either
324  * have CP rptr to a different value of jiffies wrap around which will force
325  * initialization of the lockup tracking informations.
326  *
327  * A possible false positivie is if we get call after while and last_cp_rptr ==
328  * the current CP rptr, even if it's unlikely it might happen. To avoid this
329  * if the elapsed time since last call is bigger than 2 second than we return
330  * false and update the tracking information. Due to this the caller must call
331  * radeon_ring_test_lockup several time in less than 2sec for lockup to be reported
332  * the fencing code should be cautious about that.
333  *
334  * Caller should write to the ring to force CP to do something so we don't get
335  * false positive when CP is just gived nothing to do.
336  *
337  **/
338 bool radeon_ring_test_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
339 {
340         unsigned long cjiffies, elapsed;
341         uint32_t rptr;
342
343         cjiffies = jiffies;
344         if (!time_after(cjiffies, ring->last_activity)) {
345                 /* likely a wrap around */
346                 radeon_ring_lockup_update(ring);
347                 return false;
348         }
349         rptr = RREG32(ring->rptr_reg);
350         ring->rptr = (rptr & ring->ptr_reg_mask) >> ring->ptr_reg_shift;
351         if (ring->rptr != ring->last_rptr) {
352                 /* CP is still working no lockup */
353                 radeon_ring_lockup_update(ring);
354                 return false;
355         }
356         elapsed = jiffies_to_msecs(cjiffies - ring->last_activity);
357         if (radeon_lockup_timeout && elapsed >= radeon_lockup_timeout) {
358                 dev_err(rdev->dev, "GPU lockup CP stall for more than %lumsec\n", elapsed);
359                 return true;
360         }
361         /* give a chance to the GPU ... */
362         return false;
363 }
364
365 /**
366  * radeon_ring_backup - Back up the content of a ring
367  *
368  * @rdev: radeon_device pointer
369  * @ring: the ring we want to back up
370  *
371  * Saves all unprocessed commits from a ring, returns the number of dwords saved.
372  */
373 unsigned radeon_ring_backup(struct radeon_device *rdev, struct radeon_ring *ring,
374                             uint32_t **data)
375 {
376         unsigned size, ptr, i;
377         int ridx = radeon_ring_index(rdev, ring);
378
379         /* just in case lock the ring */
380         mutex_lock(&rdev->ring_lock);
381         *data = NULL;
382
383         if (ring->ring_obj == NULL || !ring->rptr_save_reg) {
384                 mutex_unlock(&rdev->ring_lock);
385                 return 0;
386         }
387
388         /* it doesn't make sense to save anything if all fences are signaled */
389         if (!radeon_fence_count_emitted(rdev, ridx)) {
390                 mutex_unlock(&rdev->ring_lock);
391                 return 0;
392         }
393
394         /* calculate the number of dw on the ring */
395         ptr = RREG32(ring->rptr_save_reg);
396         size = ring->wptr + (ring->ring_size / 4);
397         size -= ptr;
398         size &= ring->ptr_mask;
399         if (size == 0) {
400                 mutex_unlock(&rdev->ring_lock);
401                 return 0;
402         }
403
404         /* and then save the content of the ring */
405         *data = kmalloc(size * 4, GFP_KERNEL);
406         for (i = 0; i < size; ++i) {
407                 (*data)[i] = ring->ring[ptr++];
408                 ptr &= ring->ptr_mask;
409         }
410
411         mutex_unlock(&rdev->ring_lock);
412         return size;
413 }
414
415 /**
416  * radeon_ring_restore - append saved commands to the ring again
417  *
418  * @rdev: radeon_device pointer
419  * @ring: ring to append commands to
420  * @size: number of dwords we want to write
421  * @data: saved commands
422  *
423  * Allocates space on the ring and restore the previously saved commands.
424  */
425 int radeon_ring_restore(struct radeon_device *rdev, struct radeon_ring *ring,
426                         unsigned size, uint32_t *data)
427 {
428         int i, r;
429
430         if (!size || !data)
431                 return 0;
432
433         /* restore the saved ring content */
434         r = radeon_ring_lock(rdev, ring, size);
435         if (r)
436                 return r;
437
438         for (i = 0; i < size; ++i) {
439                 radeon_ring_write(ring, data[i]);
440         }
441
442         radeon_ring_unlock_commit(rdev, ring);
443         kfree(data);
444         return 0;
445 }
446
447 int radeon_ring_init(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ring_size,
448                      unsigned rptr_offs, unsigned rptr_reg, unsigned wptr_reg,
449                      u32 ptr_reg_shift, u32 ptr_reg_mask, u32 nop)
450 {
451         int r;
452
453         ring->ring_size = ring_size;
454         ring->rptr_offs = rptr_offs;
455         ring->rptr_reg = rptr_reg;
456         ring->wptr_reg = wptr_reg;
457         ring->ptr_reg_shift = ptr_reg_shift;
458         ring->ptr_reg_mask = ptr_reg_mask;
459         ring->nop = nop;
460         /* Allocate ring buffer */
461         if (ring->ring_obj == NULL) {
462                 r = radeon_bo_create(rdev, ring->ring_size, PAGE_SIZE, true,
463                                      RADEON_GEM_DOMAIN_GTT,
464                                      NULL, &ring->ring_obj);
465                 if (r) {
466                         dev_err(rdev->dev, "(%d) ring create failed\n", r);
467                         return r;
468                 }
469                 r = radeon_bo_reserve(ring->ring_obj, false);
470                 if (unlikely(r != 0))
471                         return r;
472                 r = radeon_bo_pin(ring->ring_obj, RADEON_GEM_DOMAIN_GTT,
473                                         &ring->gpu_addr);
474                 if (r) {
475                         radeon_bo_unreserve(ring->ring_obj);
476                         dev_err(rdev->dev, "(%d) ring pin failed\n", r);
477                         return r;
478                 }
479                 r = radeon_bo_kmap(ring->ring_obj,
480                                        (void **)&ring->ring);
481                 radeon_bo_unreserve(ring->ring_obj);
482                 if (r) {
483                         dev_err(rdev->dev, "(%d) ring map failed\n", r);
484                         return r;
485                 }
486         }
487         ring->ptr_mask = (ring->ring_size / 4) - 1;
488         ring->ring_free_dw = ring->ring_size / 4;
489         if (radeon_debugfs_ring_init(rdev, ring)) {
490                 DRM_ERROR("Failed to register debugfs file for rings !\n");
491         }
492         return 0;
493 }
494
495 void radeon_ring_fini(struct radeon_device *rdev, struct radeon_ring *ring)
496 {
497         int r;
498         struct radeon_bo *ring_obj;
499
500         mutex_lock(&rdev->ring_lock);
501         ring_obj = ring->ring_obj;
502         ring->ready = false;
503         ring->ring = NULL;
504         ring->ring_obj = NULL;
505         mutex_unlock(&rdev->ring_lock);
506
507         if (ring_obj) {
508                 r = radeon_bo_reserve(ring_obj, false);
509                 if (likely(r == 0)) {
510                         radeon_bo_kunmap(ring_obj);
511                         radeon_bo_unpin(ring_obj);
512                         radeon_bo_unreserve(ring_obj);
513                 }
514                 radeon_bo_unref(&ring_obj);
515         }
516 }
517
518 /*
519  * Debugfs info
520  */
521 #if defined(CONFIG_DEBUG_FS)
522
523 static int radeon_debugfs_ring_info(struct seq_file *m, void *data)
524 {
525         struct drm_info_node *node = (struct drm_info_node *) m->private;
526         struct drm_device *dev = node->minor->dev;
527         struct radeon_device *rdev = dev->dev_private;
528         int ridx = *(int*)node->info_ent->data;
529         struct radeon_ring *ring = &rdev->ring[ridx];
530         unsigned count, i, j;
531
532         radeon_ring_free_size(rdev, ring);
533         count = (ring->ring_size / 4) - ring->ring_free_dw;
534         seq_printf(m, "wptr(0x%04x): 0x%08x\n", ring->wptr_reg, RREG32(ring->wptr_reg));
535         seq_printf(m, "rptr(0x%04x): 0x%08x\n", ring->rptr_reg, RREG32(ring->rptr_reg));
536         if (ring->rptr_save_reg) {
537                 seq_printf(m, "rptr next(0x%04x): 0x%08x\n", ring->rptr_save_reg,
538                            RREG32(ring->rptr_save_reg));
539         }
540         seq_printf(m, "driver's copy of the wptr: 0x%08x\n", ring->wptr);
541         seq_printf(m, "driver's copy of the rptr: 0x%08x\n", ring->rptr);
542         seq_printf(m, "%u free dwords in ring\n", ring->ring_free_dw);
543         seq_printf(m, "%u dwords in ring\n", count);
544         i = ring->rptr;
545         for (j = 0; j <= count; j++) {
546                 seq_printf(m, "r[%04d]=0x%08x\n", i, ring->ring[i]);
547                 i = (i + 1) & ring->ptr_mask;
548         }
549         return 0;
550 }
551
552 static int radeon_ring_type_gfx_index = RADEON_RING_TYPE_GFX_INDEX;
553 static int cayman_ring_type_cp1_index = CAYMAN_RING_TYPE_CP1_INDEX;
554 static int cayman_ring_type_cp2_index = CAYMAN_RING_TYPE_CP2_INDEX;
555
556 static struct drm_info_list radeon_debugfs_ring_info_list[] = {
557         {"radeon_ring_gfx", radeon_debugfs_ring_info, 0, &radeon_ring_type_gfx_index},
558         {"radeon_ring_cp1", radeon_debugfs_ring_info, 0, &cayman_ring_type_cp1_index},
559         {"radeon_ring_cp2", radeon_debugfs_ring_info, 0, &cayman_ring_type_cp2_index},
560 };
561
562 static int radeon_debugfs_sa_info(struct seq_file *m, void *data)
563 {
564         struct drm_info_node *node = (struct drm_info_node *) m->private;
565         struct drm_device *dev = node->minor->dev;
566         struct radeon_device *rdev = dev->dev_private;
567
568         radeon_sa_bo_dump_debug_info(&rdev->ring_tmp_bo, m);
569
570         return 0;
571
572 }
573
574 static struct drm_info_list radeon_debugfs_sa_list[] = {
575         {"radeon_sa_info", &radeon_debugfs_sa_info, 0, NULL},
576 };
577
578 #endif
579
580 int radeon_debugfs_ring_init(struct radeon_device *rdev, struct radeon_ring *ring)
581 {
582 #if defined(CONFIG_DEBUG_FS)
583         unsigned i;
584         for (i = 0; i < ARRAY_SIZE(radeon_debugfs_ring_info_list); ++i) {
585                 struct drm_info_list *info = &radeon_debugfs_ring_info_list[i];
586                 int ridx = *(int*)radeon_debugfs_ring_info_list[i].data;
587                 unsigned r;
588
589                 if (&rdev->ring[ridx] != ring)
590                         continue;
591
592                 r = radeon_debugfs_add_files(rdev, info, 1);
593                 if (r)
594                         return r;
595         }
596 #endif
597         return 0;
598 }
599
600 int radeon_debugfs_sa_init(struct radeon_device *rdev)
601 {
602 #if defined(CONFIG_DEBUG_FS)
603         return radeon_debugfs_add_files(rdev, radeon_debugfs_sa_list, 1);
604 #else
605         return 0;
606 #endif
607 }