drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c

   1 /*
   2  * Copyright 2014 Advanced Micro Devices, Inc.
   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  *    Christian König <[email protected]>
  29  */
  30
  31 #include <drm/drmP.h>
  32 #include "amdgpu.h"
  33 #include "amdgpu_trace.h"
  34
  35 struct amdgpu_sync_entry {
  36         struct hlist_node       node;
  37         struct fence            *fence;
  38 };
  39
  40 /**
  41  * amdgpu_sync_create - zero init sync object
  42  *
  43  * @sync: sync object to initialize
  44  *
  45  * Just clear the sync object for now.
  46  */
  47 void amdgpu_sync_create(struct amdgpu_sync *sync)
  48 {
  49         unsigned i;
  50
  51         for (i = 0; i < AMDGPU_NUM_SYNCS; ++i)
  52                 sync->semaphores[i] = NULL;
  53
  54         for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
  55                 sync->sync_to[i] = NULL;
  56
  57         hash_init(sync->fences);
  58         sync->last_vm_update = NULL;
  59 }
  60
  61 /**
  62  * amdgpu_sync_fence - remember to sync to this fence
  63  *
  64  * @sync: sync object to add fence to
  65  * @fence: fence to sync to
  66  *
  67  */
  68 int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
  69                       struct fence *f)
  70 {
  71         struct amdgpu_sync_entry *e;
  72         struct amdgpu_fence *fence;
  73         struct amdgpu_fence *other;
  74
  75         if (!f)
  76                 return 0;
  77
  78         fence = to_amdgpu_fence(f);
  79         if (!fence || fence->ring->adev != adev) {
  80                 hash_for_each_possible(sync->fences, e, node, f->context) {
  81                         struct fence *new;
  82                         if (unlikely(e->fence->context != f->context))
  83                                 continue;
  84                         new = fence_get(fence_later(e->fence, f));
  85                         if (new) {
  86                                 fence_put(e->fence);
  87                                 e->fence = new;
  88                         }
  89                         return 0;
  90                 }
  91
  92                 e = kmalloc(sizeof(struct amdgpu_sync_entry), GFP_KERNEL);
  93                 if (!e)
  94                         return -ENOMEM;
  95
  96                 hash_add(sync->fences, &e->node, f->context);
  97                 e->fence = fence_get(f);
  98                 return 0;
  99         }
 100
 101         other = sync->sync_to[fence->ring->idx];
 102         sync->sync_to[fence->ring->idx] = amdgpu_fence_ref(
 103                 amdgpu_fence_later(fence, other));
 104         amdgpu_fence_unref(&other);
 105
 106         if (fence->owner == AMDGPU_FENCE_OWNER_VM) {
 107                 other = sync->last_vm_update;
 108                 sync->last_vm_update = amdgpu_fence_ref(
 109                         amdgpu_fence_later(fence, other));
 110                 amdgpu_fence_unref(&other);
 111         }
 112
 113         return 0;
 114 }
 115
 116 /**
 117  * amdgpu_sync_resv - use the semaphores to sync to a reservation object
 118  *
 119  * @sync: sync object to add fences from reservation object to
 120  * @resv: reservation object with embedded fence
 121  * @shared: true if we should only sync to the exclusive fence
 122  *
 123  * Sync to the fence using the semaphore objects
 124  */
 125 int amdgpu_sync_resv(struct amdgpu_device *adev,
 126                      struct amdgpu_sync *sync,
 127                      struct reservation_object *resv,
 128                      void *owner)
 129 {
 130         struct reservation_object_list *flist;
 131         struct fence *f;
 132         struct amdgpu_fence *fence;
 133         unsigned i;
 134         int r = 0;
 135
 136         if (resv == NULL)
 137                 return -EINVAL;
 138
 139         /* always sync to the exclusive fence */
 140         f = reservation_object_get_excl(resv);
 141         r = amdgpu_sync_fence(adev, sync, f);
 142
 143         flist = reservation_object_get_list(resv);
 144         if (!flist || r)
 145                 return r;
 146
 147         for (i = 0; i < flist->shared_count; ++i) {
 148                 f = rcu_dereference_protected(flist->shared[i],
 149                                               reservation_object_held(resv));
 150                 fence = f ? to_amdgpu_fence(f) : NULL;
 151                 if (fence && fence->ring->adev == adev) {
 152                         /* VM updates are only interesting
 153                          * for other VM updates and moves.
 154                          */
 155                         if ((owner != AMDGPU_FENCE_OWNER_MOVE) &&
 156                             (fence->owner != AMDGPU_FENCE_OWNER_MOVE) &&
 157                             ((owner == AMDGPU_FENCE_OWNER_VM) !=
 158                              (fence->owner == AMDGPU_FENCE_OWNER_VM)))
 159                                 continue;
 160
 161                         /* Ignore fence from the same owner as
 162                          * long as it isn't undefined.
 163                          */
 164                         if (owner != AMDGPU_FENCE_OWNER_UNDEFINED &&
 165                             fence->owner == owner)
 166                                 continue;
 167                 }
 168
 169                 r = amdgpu_sync_fence(adev, sync, f);
 170                 if (r)
 171                         break;
 172         }
 173         return r;
 174 }
 175
 176 int amdgpu_sync_wait(struct amdgpu_sync *sync)
 177 {
 178         struct amdgpu_sync_entry *e;
 179         struct hlist_node *tmp;
 180         int i, r;
 181
 182         hash_for_each_safe(sync->fences, i, tmp, e, node) {
 183                 r = fence_wait(e->fence, false);
 184                 if (r)
 185                         return r;
 186
 187                 hash_del(&e->node);
 188                 fence_put(e->fence);
 189                 kfree(e);
 190         }
 191         return 0;
 192 }
 193
 194 /**
 195  * amdgpu_sync_rings - sync ring to all registered fences
 196  *
 197  * @sync: sync object to use
 198  * @ring: ring that needs sync
 199  *
 200  * Ensure that all registered fences are signaled before letting
 201  * the ring continue. The caller must hold the ring lock.
 202  */
 203 int amdgpu_sync_rings(struct amdgpu_sync *sync,
 204                       struct amdgpu_ring *ring)
 205 {
 206         struct amdgpu_device *adev = ring->adev;
 207         unsigned count = 0;
 208         int i, r;
 209
 210         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
 211                 struct amdgpu_fence *fence = sync->sync_to[i];
 212                 struct amdgpu_semaphore *semaphore;
 213                 struct amdgpu_ring *other = adev->rings[i];
 214
 215                 /* check if we really need to sync */
 216                 if (!amdgpu_fence_need_sync(fence, ring))
 217                         continue;
 218
 219                 /* prevent GPU deadlocks */
 220                 if (!other->ready) {
 221                         dev_err(adev->dev, "Syncing to a disabled ring!");
 222                         return -EINVAL;
 223                 }
 224
 225                 if (amdgpu_enable_scheduler || (count >= AMDGPU_NUM_SYNCS)) {
 226                         /* not enough room, wait manually */
 227                         r = fence_wait(&fence->base, false);
 228                         if (r)
 229                                 return r;
 230                         continue;
 231                 }
 232                 r = amdgpu_semaphore_create(adev, &semaphore);
 233                 if (r)
 234                         return r;
 235
 236                 sync->semaphores[count++] = semaphore;
 237
 238                 /* allocate enough space for sync command */
 239                 r = amdgpu_ring_alloc(other, 16);
 240                 if (r)
 241                         return r;
 242
 243                 /* emit the signal semaphore */
 244                 if (!amdgpu_semaphore_emit_signal(other, semaphore)) {
 245                         /* signaling wasn't successful wait manually */
 246                         amdgpu_ring_undo(other);
 247                         r = fence_wait(&fence->base, false);
 248                         if (r)
 249                                 return r;
 250                         continue;
 251                 }
 252
 253                 /* we assume caller has already allocated space on waiters ring */
 254                 if (!amdgpu_semaphore_emit_wait(ring, semaphore)) {
 255                         /* waiting wasn't successful wait manually */
 256                         amdgpu_ring_undo(other);
 257                         r = fence_wait(&fence->base, false);
 258                         if (r)
 259                                 return r;
 260                         continue;
 261                 }
 262
 263                 amdgpu_ring_commit(other);
 264                 amdgpu_fence_note_sync(fence, ring);
 265         }
 266
 267         return 0;
 268 }
 269
 270 /**
 271  * amdgpu_sync_free - free the sync object
 272  *
 273  * @adev: amdgpu_device pointer
 274  * @sync: sync object to use
 275  * @fence: fence to use for the free
 276  *
 277  * Free the sync object by freeing all semaphores in it.
 278  */
 279 void amdgpu_sync_free(struct amdgpu_device *adev,
 280                       struct amdgpu_sync *sync,
 281                       struct fence *fence)
 282 {
 283         struct amdgpu_sync_entry *e;
 284         struct hlist_node *tmp;
 285         unsigned i;
 286
 287         hash_for_each_safe(sync->fences, i, tmp, e, node) {
 288                 hash_del(&e->node);
 289                 fence_put(e->fence);
 290                 kfree(e);
 291         }
 292
 293         for (i = 0; i < AMDGPU_NUM_SYNCS; ++i)
 294                 amdgpu_semaphore_free(adev, &sync->semaphores[i], fence);
 295
 296         for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
 297                 amdgpu_fence_unref(&sync->sync_to[i]);
 298
 299         amdgpu_fence_unref(&sync->last_vm_update);
 300 }