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 <christian.koenig@amd.com> 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 dma_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, 64 struct dma_fence *f) 65 { 66 struct amd_sched_fence *s_fence = to_amd_sched_fence(f); 67 68 if (s_fence) { 69 struct amdgpu_ring *ring; 70 71 ring = container_of(s_fence->sched, struct amdgpu_ring, sched); 72 return ring->adev == adev; 73 } 74 75 return false; 76 } 77 78 /** 79 * amdgpu_sync_get_owner - extract the owner of a fence 80 * 81 * @fence: fence get the owner from 82 * 83 * Extract who originally created the fence. 84 */ 85 static void *amdgpu_sync_get_owner(struct dma_fence *f) 86 { 87 struct amd_sched_fence *s_fence = to_amd_sched_fence(f); 88 89 if (s_fence) 90 return s_fence->owner; 91 92 return AMDGPU_FENCE_OWNER_UNDEFINED; 93 } 94 95 /** 96 * amdgpu_sync_keep_later - Keep the later fence 97 * 98 * @keep: existing fence to test 99 * @fence: new fence 100 * 101 * Either keep the existing fence or the new one, depending which one is later. 102 */ 103 static void amdgpu_sync_keep_later(struct dma_fence **keep, 104 struct dma_fence *fence) 105 { 106 if (*keep && dma_fence_is_later(*keep, fence)) 107 return; 108 109 dma_fence_put(*keep); 110 *keep = dma_fence_get(fence); 111 } 112 113 /** 114 * amdgpu_sync_add_later - add the fence to the hash 115 * 116 * @sync: sync object to add the fence to 117 * @f: fence to add 118 * 119 * Tries to add the fence to an existing hash entry. Returns true when an entry 120 * was found, false otherwise. 121 */ 122 static bool amdgpu_sync_add_later(struct amdgpu_sync *sync, struct dma_fence *f) 123 { 124 struct amdgpu_sync_entry *e; 125 126 hash_for_each_possible(sync->fences, e, node, f->context) { 127 if (unlikely(e->fence->context != f->context)) 128 continue; 129 130 amdgpu_sync_keep_later(&e->fence, f); 131 return true; 132 } 133 return false; 134 } 135 136 /** 137 * amdgpu_sync_fence - remember to sync to this fence 138 * 139 * @sync: sync object to add fence to 140 * @fence: fence to sync to 141 * 142 */ 143 int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync, 144 struct dma_fence *f) 145 { 146 struct amdgpu_sync_entry *e; 147 148 if (!f) 149 return 0; 150 151 if (amdgpu_sync_same_dev(adev, f) && 152 amdgpu_sync_get_owner(f) == AMDGPU_FENCE_OWNER_VM) 153 amdgpu_sync_keep_later(&sync->last_vm_update, f); 154 155 if (amdgpu_sync_add_later(sync, f)) 156 return 0; 157 158 e = kmem_cache_alloc(amdgpu_sync_slab, GFP_KERNEL); 159 if (!e) 160 return -ENOMEM; 161 162 hash_add(sync->fences, &e->node, f->context); 163 e->fence = dma_fence_get(f); 164 return 0; 165 } 166 167 /** 168 * amdgpu_sync_resv - sync to a reservation object 169 * 170 * @sync: sync object to add fences from reservation object to 171 * @resv: reservation object with embedded fence 172 * @shared: true if we should only sync to the exclusive fence 173 * 174 * Sync to the fence 175 */ 176 int amdgpu_sync_resv(struct amdgpu_device *adev, 177 struct amdgpu_sync *sync, 178 struct reservation_object *resv, 179 void *owner) 180 { 181 struct reservation_object_list *flist; 182 struct dma_fence *f; 183 void *fence_owner; 184 unsigned i; 185 int r = 0; 186 187 if (resv == NULL) 188 return -EINVAL; 189 190 /* always sync to the exclusive fence */ 191 f = reservation_object_get_excl(resv); 192 r = amdgpu_sync_fence(adev, sync, f); 193 194 flist = reservation_object_get_list(resv); 195 if (!flist || r) 196 return r; 197 198 for (i = 0; i < flist->shared_count; ++i) { 199 f = rcu_dereference_protected(flist->shared[i], 200 reservation_object_held(resv)); 201 if (amdgpu_sync_same_dev(adev, f)) { 202 /* VM updates are only interesting 203 * for other VM updates and moves. 204 */ 205 fence_owner = amdgpu_sync_get_owner(f); 206 if ((owner != AMDGPU_FENCE_OWNER_UNDEFINED) && 207 (fence_owner != AMDGPU_FENCE_OWNER_UNDEFINED) && 208 ((owner == AMDGPU_FENCE_OWNER_VM) != 209 (fence_owner == AMDGPU_FENCE_OWNER_VM))) 210 continue; 211 212 /* Ignore fence from the same owner as 213 * long as it isn't undefined. 214 */ 215 if (owner != AMDGPU_FENCE_OWNER_UNDEFINED && 216 fence_owner == owner) 217 continue; 218 } 219 220 r = amdgpu_sync_fence(adev, sync, f); 221 if (r) 222 break; 223 } 224 return r; 225 } 226 227 /** 228 * amdgpu_sync_peek_fence - get the next fence not signaled yet 229 * 230 * @sync: the sync object 231 * @ring: optional ring to use for test 232 * 233 * Returns the next fence not signaled yet without removing it from the sync 234 * object. 235 */ 236 struct dma_fence *amdgpu_sync_peek_fence(struct amdgpu_sync *sync, 237 struct amdgpu_ring *ring) 238 { 239 struct amdgpu_sync_entry *e; 240 struct hlist_node *tmp; 241 int i; 242 243 hash_for_each_safe(sync->fences, i, tmp, e, node) { 244 struct dma_fence *f = e->fence; 245 struct amd_sched_fence *s_fence = to_amd_sched_fence(f); 246 247 if (dma_fence_is_signaled(f)) { 248 hash_del(&e->node); 249 dma_fence_put(f); 250 kmem_cache_free(amdgpu_sync_slab, e); 251 continue; 252 } 253 if (ring && s_fence) { 254 /* For fences from the same ring it is sufficient 255 * when they are scheduled. 256 */ 257 if (s_fence->sched == &ring->sched) { 258 if (dma_fence_is_signaled(&s_fence->scheduled)) 259 continue; 260 261 return &s_fence->scheduled; 262 } 263 } 264 265 return f; 266 } 267 268 return NULL; 269 } 270 271 /** 272 * amdgpu_sync_get_fence - get the next fence from the sync object 273 * 274 * @sync: sync object to use 275 * 276 * Get and removes the next fence from the sync object not signaled yet. 277 */ 278 struct dma_fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync) 279 { 280 struct amdgpu_sync_entry *e; 281 struct hlist_node *tmp; 282 struct dma_fence *f; 283 int i; 284 285 hash_for_each_safe(sync->fences, i, tmp, e, node) { 286 287 f = e->fence; 288 289 hash_del(&e->node); 290 kmem_cache_free(amdgpu_sync_slab, e); 291 292 if (!dma_fence_is_signaled(f)) 293 return f; 294 295 dma_fence_put(f); 296 } 297 return NULL; 298 } 299 300 int amdgpu_sync_wait(struct amdgpu_sync *sync, bool intr) 301 { 302 struct amdgpu_sync_entry *e; 303 struct hlist_node *tmp; 304 int i, r; 305 306 hash_for_each_safe(sync->fences, i, tmp, e, node) { 307 r = dma_fence_wait(e->fence, intr); 308 if (r) 309 return r; 310 311 hash_del(&e->node); 312 dma_fence_put(e->fence); 313 kmem_cache_free(amdgpu_sync_slab, e); 314 } 315 316 return 0; 317 } 318 319 /** 320 * amdgpu_sync_free - free the sync object 321 * 322 * @sync: sync object to use 323 * 324 * Free the sync object. 325 */ 326 void amdgpu_sync_free(struct amdgpu_sync *sync) 327 { 328 struct amdgpu_sync_entry *e; 329 struct hlist_node *tmp; 330 unsigned i; 331 332 hash_for_each_safe(sync->fences, i, tmp, e, node) { 333 hash_del(&e->node); 334 dma_fence_put(e->fence); 335 kmem_cache_free(amdgpu_sync_slab, e); 336 } 337 338 dma_fence_put(sync->last_vm_update); 339 } 340 341 /** 342 * amdgpu_sync_init - init sync object subsystem 343 * 344 * Allocate the slab allocator. 345 */ 346 int amdgpu_sync_init(void) 347 { 348 amdgpu_sync_slab = kmem_cache_create( 349 "amdgpu_sync", sizeof(struct amdgpu_sync_entry), 0, 350 SLAB_HWCACHE_ALIGN, NULL); 351 if (!amdgpu_sync_slab) 352 return -ENOMEM; 353 354 return 0; 355 } 356 357 /** 358 * amdgpu_sync_fini - fini sync object subsystem 359 * 360 * Free the slab allocator. 361 */ 362 void amdgpu_sync_fini(void) 363 { 364 kmem_cache_destroy(amdgpu_sync_slab); 365 } 366