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 static struct kmem_cache *amdgpu_sync_slab;
  41
  42 /**
  43  * amdgpu_sync_create - zero init sync object
  44  *
  45  * @sync: sync object to initialize
  46  *
  47  * Just clear the sync object for now.
  48  */
  49 void amdgpu_sync_create(struct amdgpu_sync *sync)
  50 {
  51         hash_init(sync->fences);
  52         sync->last_vm_update = NULL;
  53 }
  54
  55 /**
  56  * amdgpu_sync_same_dev - test if fence belong to us
  57  *
  58  * @adev: amdgpu device to use for the test
  59  * @f: fence to test
  60  *
  61  * Test if the fence was issued by us.
  62  */
  63 static bool amdgpu_sync_same_dev(struct amdgpu_device *adev, struct fence *f)
  64 {
  65         struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
  66
  67         if (s_fence) {
  68                 struct amdgpu_ring *ring;
  69
  70                 ring = container_of(s_fence->sched, struct amdgpu_ring, sched);
  71                 return ring->adev == adev;
  72         }
  73
  74         return false;
  75 }
  76
  77 /**
  78  * amdgpu_sync_get_owner - extract the owner of a fence
  79  *
  80  * @fence: fence get the owner from
  81  *
  82  * Extract who originally created the fence.
  83  */
  84 static void *amdgpu_sync_get_owner(struct fence *f)
  85 {
  86         struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
  87
  88         if (s_fence)
  89                 return s_fence->owner;
  90
  91         return AMDGPU_FENCE_OWNER_UNDEFINED;
  92 }
  93
  94 /**
  95  * amdgpu_sync_keep_later - Keep the later fence
  96  *
  97  * @keep: existing fence to test
  98  * @fence: new fence
  99  *
 100  * Either keep the existing fence or the new one, depending which one is later.
 101  */
 102 static void amdgpu_sync_keep_later(struct fence **keep, struct fence *fence)
 103 {
 104         if (*keep && fence_is_later(*keep, fence))
 105                 return;
 106
 107         fence_put(*keep);
 108         *keep = fence_get(fence);
 109 }
 110
 111 /**
 112  * amdgpu_sync_add_later - add the fence to the hash
 113  *
 114  * @sync: sync object to add the fence to
 115  * @f: fence to add
 116  *
 117  * Tries to add the fence to an existing hash entry. Returns true when an entry
 118  * was found, false otherwise.
 119  */
 120 static bool amdgpu_sync_add_later(struct amdgpu_sync *sync, struct fence *f)
 121 {
 122         struct amdgpu_sync_entry *e;
 123
 124         hash_for_each_possible(sync->fences, e, node, f->context) {
 125                 if (unlikely(e->fence->context != f->context))
 126                         continue;
 127
 128                 amdgpu_sync_keep_later(&e->fence, f);
 129                 return true;
 130         }
 131         return false;
 132 }
 133
 134 /**
 135  * amdgpu_sync_fence - remember to sync to this fence
 136  *
 137  * @sync: sync object to add fence to
 138  * @fence: fence to sync to
 139  *
 140  */
 141 int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
 142                       struct fence *f)
 143 {
 144         struct amdgpu_sync_entry *e;
 145
 146         if (!f)
 147                 return 0;
 148
 149         if (amdgpu_sync_same_dev(adev, f) &&
 150             amdgpu_sync_get_owner(f) == AMDGPU_FENCE_OWNER_VM)
 151                 amdgpu_sync_keep_later(&sync->last_vm_update, f);
 152
 153         if (amdgpu_sync_add_later(sync, f))
 154                 return 0;
 155
 156         e = kmem_cache_alloc(amdgpu_sync_slab, GFP_KERNEL);
 157         if (!e)
 158                 return -ENOMEM;
 159
 160         hash_add(sync->fences, &e->node, f->context);
 161         e->fence = fence_get(f);
 162         return 0;
 163 }
 164
 165 /**
 166  * amdgpu_sync_resv - sync to a reservation object
 167  *
 168  * @sync: sync object to add fences from reservation object to
 169  * @resv: reservation object with embedded fence
 170  * @shared: true if we should only sync to the exclusive fence
 171  *
 172  * Sync to the fence
 173  */
 174 int amdgpu_sync_resv(struct amdgpu_device *adev,
 175                      struct amdgpu_sync *sync,
 176                      struct reservation_object *resv,
 177                      void *owner)
 178 {
 179         struct reservation_object_list *flist;
 180         struct fence *f;
 181         void *fence_owner;
 182         unsigned i;
 183         int r = 0;
 184
 185         if (resv == NULL)
 186                 return -EINVAL;
 187
 188         /* always sync to the exclusive fence */
 189         f = reservation_object_get_excl(resv);
 190         r = amdgpu_sync_fence(adev, sync, f);
 191
 192         flist = reservation_object_get_list(resv);
 193         if (!flist || r)
 194                 return r;
 195
 196         for (i = 0; i < flist->shared_count; ++i) {
 197                 f = rcu_dereference_protected(flist->shared[i],
 198                                               reservation_object_held(resv));
 199                 if (amdgpu_sync_same_dev(adev, f)) {
 200                         /* VM updates are only interesting
 201                          * for other VM updates and moves.
 202                          */
 203                         fence_owner = amdgpu_sync_get_owner(f);
 204                         if ((owner != AMDGPU_FENCE_OWNER_UNDEFINED) &&
 205                             (fence_owner != AMDGPU_FENCE_OWNER_UNDEFINED) &&
 206                             ((owner == AMDGPU_FENCE_OWNER_VM) !=
 207                              (fence_owner == AMDGPU_FENCE_OWNER_VM)))
 208                                 continue;
 209
 210                         /* Ignore fence from the same owner as
 211                          * long as it isn't undefined.
 212                          */
 213                         if (owner != AMDGPU_FENCE_OWNER_UNDEFINED &&
 214                             fence_owner == owner)
 215                                 continue;
 216                 }
 217
 218                 r = amdgpu_sync_fence(adev, sync, f);
 219                 if (r)
 220                         break;
 221         }
 222         return r;
 223 }
 224
 225 /**
 226  * amdgpu_sync_is_idle - test if all fences are signaled
 227  *
 228  * @sync: the sync object
 229  *
 230  * Returns true if all fences in the sync object are signaled.
 231  */
 232 bool amdgpu_sync_is_idle(struct amdgpu_sync *sync)
 233 {
 234         struct amdgpu_sync_entry *e;
 235         struct hlist_node *tmp;
 236         int i;
 237
 238         hash_for_each_safe(sync->fences, i, tmp, e, node) {
 239                 struct fence *f = e->fence;
 240
 241                 if (fence_is_signaled(f)) {
 242                         hash_del(&e->node);
 243                         fence_put(f);
 244                         kmem_cache_free(amdgpu_sync_slab, e);
 245                         continue;
 246                 }
 247
 248                 return false;
 249         }
 250
 251         return true;
 252 }
 253
 254 /**
 255  * amdgpu_sync_cycle_fences - move fences from one sync object into another
 256  *
 257  * @dst: the destination sync object
 258  * @src: the source sync object
 259  * @fence: fence to add to source
 260  *
 261  * Remove all fences from source and put them into destination and add
 262  * fence as new one into source.
 263  */
 264 int amdgpu_sync_cycle_fences(struct amdgpu_sync *dst, struct amdgpu_sync *src,
 265                              struct fence *fence)
 266 {
 267         struct amdgpu_sync_entry *e, *newone;
 268         struct hlist_node *tmp;
 269         int i;
 270
 271         /* Allocate the new entry before moving the old ones */
 272         newone = kmem_cache_alloc(amdgpu_sync_slab, GFP_KERNEL);
 273         if (!newone)
 274                 return -ENOMEM;
 275
 276         hash_for_each_safe(src->fences, i, tmp, e, node) {
 277                 struct fence *f = e->fence;
 278
 279                 hash_del(&e->node);
 280                 if (fence_is_signaled(f)) {
 281                         fence_put(f);
 282                         kmem_cache_free(amdgpu_sync_slab, e);
 283                         continue;
 284                 }
 285
 286                 if (amdgpu_sync_add_later(dst, f)) {
 287                         kmem_cache_free(amdgpu_sync_slab, e);
 288                         continue;
 289                 }
 290
 291                 hash_add(dst->fences, &e->node, f->context);
 292         }
 293
 294         hash_add(src->fences, &newone->node, fence->context);
 295         newone->fence = fence_get(fence);
 296
 297         return 0;
 298 }
 299
 300 struct fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync)
 301 {
 302         struct amdgpu_sync_entry *e;
 303         struct hlist_node *tmp;
 304         struct fence *f;
 305         int i;
 306
 307         hash_for_each_safe(sync->fences, i, tmp, e, node) {
 308
 309                 f = e->fence;
 310
 311                 hash_del(&e->node);
 312                 kmem_cache_free(amdgpu_sync_slab, e);
 313
 314                 if (!fence_is_signaled(f))
 315                         return f;
 316
 317                 fence_put(f);
 318         }
 319         return NULL;
 320 }
 321
 322 int amdgpu_sync_wait(struct amdgpu_sync *sync)
 323 {
 324         struct amdgpu_sync_entry *e;
 325         struct hlist_node *tmp;
 326         int i, r;
 327
 328         hash_for_each_safe(sync->fences, i, tmp, e, node) {
 329                 r = fence_wait(e->fence, false);
 330                 if (r)
 331                         return r;
 332
 333                 hash_del(&e->node);
 334                 fence_put(e->fence);
 335                 kmem_cache_free(amdgpu_sync_slab, e);
 336         }
 337
 338         return 0;
 339 }
 340
 341 /**
 342  * amdgpu_sync_free - free the sync object
 343  *
 344  * @sync: sync object to use
 345  *
 346  * Free the sync object.
 347  */
 348 void amdgpu_sync_free(struct amdgpu_sync *sync)
 349 {
 350         struct amdgpu_sync_entry *e;
 351         struct hlist_node *tmp;
 352         unsigned i;
 353
 354         hash_for_each_safe(sync->fences, i, tmp, e, node) {
 355                 hash_del(&e->node);
 356                 fence_put(e->fence);
 357                 kmem_cache_free(amdgpu_sync_slab, e);
 358         }
 359
 360         fence_put(sync->last_vm_update);
 361 }
 362
 363 /**
 364  * amdgpu_sync_init - init sync object subsystem
 365  *
 366  * Allocate the slab allocator.
 367  */
 368 int amdgpu_sync_init(void)
 369 {
 370         amdgpu_sync_slab = kmem_cache_create(
 371                 "amdgpu_sync", sizeof(struct amdgpu_sync_entry), 0,
 372                 SLAB_HWCACHE_ALIGN, NULL);
 373         if (!amdgpu_sync_slab)
 374                 return -ENOMEM;
 375
 376         return 0;
 377 }
 378
 379 /**
 380  * amdgpu_sync_fini - fini sync object subsystem
 381  *
 382  * Free the slab allocator.
 383  */
 384 void amdgpu_sync_fini(void)
 385 {
 386         kmem_cache_destroy(amdgpu_sync_slab);
 387 }