1 /* 2 * Copyright © 2008-2010 Intel Corporation 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 21 * IN THE SOFTWARE. 22 * 23 * Authors: 24 * Eric Anholt <eric@anholt.net> 25 * Chris Wilson <chris@chris-wilson.co.uuk> 26 * 27 */ 28 29 #include <drm/drmP.h> 30 #include <drm/i915_drm.h> 31 32 #include "i915_drv.h" 33 #include "intel_drv.h" 34 #include "i915_trace.h" 35 36 static bool ggtt_is_idle(struct drm_i915_private *dev_priv) 37 { 38 struct i915_ggtt *ggtt = &dev_priv->ggtt; 39 struct intel_engine_cs *engine; 40 enum intel_engine_id id; 41 42 for_each_engine(engine, dev_priv, id) { 43 struct intel_timeline *tl; 44 45 tl = &ggtt->base.timeline.engine[engine->id]; 46 if (i915_gem_active_isset(&tl->last_request)) 47 return false; 48 } 49 50 return true; 51 } 52 53 static bool 54 mark_free(struct drm_mm_scan *scan, 55 struct i915_vma *vma, 56 unsigned int flags, 57 struct list_head *unwind) 58 { 59 if (i915_vma_is_pinned(vma)) 60 return false; 61 62 if (WARN_ON(!list_empty(&vma->exec_list))) 63 return false; 64 65 if (flags & PIN_NONFAULT && !list_empty(&vma->obj->userfault_link)) 66 return false; 67 68 list_add(&vma->exec_list, unwind); 69 return drm_mm_scan_add_block(scan, &vma->node); 70 } 71 72 /** 73 * i915_gem_evict_something - Evict vmas to make room for binding a new one 74 * @vm: address space to evict from 75 * @min_size: size of the desired free space 76 * @alignment: alignment constraint of the desired free space 77 * @cache_level: cache_level for the desired space 78 * @start: start (inclusive) of the range from which to evict objects 79 * @end: end (exclusive) of the range from which to evict objects 80 * @flags: additional flags to control the eviction algorithm 81 * 82 * This function will try to evict vmas until a free space satisfying the 83 * requirements is found. Callers must check first whether any such hole exists 84 * already before calling this function. 85 * 86 * This function is used by the object/vma binding code. 87 * 88 * Since this function is only used to free up virtual address space it only 89 * ignores pinned vmas, and not object where the backing storage itself is 90 * pinned. Hence obj->pages_pin_count does not protect against eviction. 91 * 92 * To clarify: This is for freeing up virtual address space, not for freeing 93 * memory in e.g. the shrinker. 94 */ 95 int 96 i915_gem_evict_something(struct i915_address_space *vm, 97 u64 min_size, u64 alignment, 98 unsigned cache_level, 99 u64 start, u64 end, 100 unsigned flags) 101 { 102 struct drm_i915_private *dev_priv = vm->i915; 103 struct drm_mm_scan scan; 104 struct list_head eviction_list; 105 struct list_head *phases[] = { 106 &vm->inactive_list, 107 &vm->active_list, 108 NULL, 109 }, **phase; 110 struct i915_vma *vma, *next; 111 struct drm_mm_node *node; 112 enum drm_mm_insert_mode mode; 113 int ret; 114 115 lockdep_assert_held(&vm->i915->drm.struct_mutex); 116 trace_i915_gem_evict(vm, min_size, alignment, flags); 117 118 /* 119 * The goal is to evict objects and amalgamate space in LRU order. 120 * The oldest idle objects reside on the inactive list, which is in 121 * retirement order. The next objects to retire are those in flight, 122 * on the active list, again in retirement order. 123 * 124 * The retirement sequence is thus: 125 * 1. Inactive objects (already retired) 126 * 2. Active objects (will stall on unbinding) 127 * 128 * On each list, the oldest objects lie at the HEAD with the freshest 129 * object on the TAIL. 130 */ 131 kprintf("i915_gem_evict_something: %016llx-%016llx\n", start, end); 132 mode = DRM_MM_INSERT_BEST; 133 if (flags & PIN_HIGH) 134 mode = DRM_MM_INSERT_HIGH; 135 if (flags & PIN_MAPPABLE) 136 mode = DRM_MM_INSERT_LOW; 137 drm_mm_scan_init_with_range(&scan, &vm->mm, 138 min_size, alignment, cache_level, 139 start, end, mode); 140 141 /* Retire before we search the active list. Although we have 142 * reasonable accuracy in our retirement lists, we may have 143 * a stray pin (preventing eviction) that can only be resolved by 144 * retiring. 145 */ 146 if (!(flags & PIN_NONBLOCK)) 147 i915_gem_retire_requests(dev_priv); 148 else 149 phases[1] = NULL; 150 151 search_again: 152 INIT_LIST_HEAD(&eviction_list); 153 phase = phases; 154 do { 155 list_for_each_entry(vma, *phase, vm_link) 156 if (mark_free(&scan, vma, flags, &eviction_list)) 157 goto found; 158 } while (*++phase); 159 160 /* Nothing found, clean up and bail out! */ 161 list_for_each_entry_safe(vma, next, &eviction_list, exec_list) { 162 ret = drm_mm_scan_remove_block(&scan, &vma->node); 163 BUG_ON(ret); 164 165 INIT_LIST_HEAD(&vma->exec_list); 166 } 167 168 /* Can we unpin some objects such as idle hw contents, 169 * or pending flips? But since only the GGTT has global entries 170 * such as scanouts, rinbuffers and contexts, we can skip the 171 * purge when inspecting per-process local address spaces. 172 */ 173 if (!i915_is_ggtt(vm) || flags & PIN_NONBLOCK) { 174 kprintf("i915_gem_evict_something: ENOSPC %d,%d\n", 175 !i915_is_ggtt(vm), ((flags & PIN_NONBLOCK) != 0)); 176 return -ENOSPC; 177 } 178 kprintf("i915_gem_evict_something: Nothing found %d,%d\n", 179 ggtt_is_idle(dev_priv), 180 intel_has_pending_fb_unpin(dev_priv)); 181 182 if (ggtt_is_idle(dev_priv)) { 183 /* If we still have pending pageflip completions, drop 184 * back to userspace to give our workqueues time to 185 * acquire our locks and unpin the old scanouts. 186 */ 187 return intel_has_pending_fb_unpin(dev_priv) ? -EAGAIN : -ENOSPC; 188 } 189 190 /* Not everything in the GGTT is tracked via vma (otherwise we 191 * could evict as required with minimal stalling) so we are forced 192 * to idle the GPU and explicitly retire outstanding requests in 193 * the hopes that we can then remove contexts and the like only 194 * bound by their active reference. 195 */ 196 ret = i915_gem_switch_to_kernel_context(dev_priv); 197 if (ret) 198 return ret; 199 kprintf("i915_gem_evict_something: (A) Switch to kernel context and retire everything %d\n", ret); 200 201 ret = i915_gem_wait_for_idle(dev_priv, 202 I915_WAIT_INTERRUPTIBLE | 203 I915_WAIT_LOCKED); 204 kprintf("i915_gem_evict_something: (B) Switch to kernel context and retire everything %d\n", ret); 205 if (ret) 206 return ret; 207 208 goto search_again; 209 210 found: 211 /* drm_mm doesn't allow any other other operations while 212 * scanning, therefore store to-be-evicted objects on a 213 * temporary list and take a reference for all before 214 * calling unbind (which may remove the active reference 215 * of any of our objects, thus corrupting the list). 216 */ 217 list_for_each_entry_safe(vma, next, &eviction_list, exec_list) { 218 if (drm_mm_scan_remove_block(&scan, &vma->node)) 219 __i915_vma_pin(vma); 220 else 221 list_del_init(&vma->exec_list); 222 } 223 224 /* Unbinding will emit any required flushes */ 225 ret = 0; 226 while (!list_empty(&eviction_list)) { 227 vma = list_first_entry(&eviction_list, 228 struct i915_vma, 229 exec_list); 230 231 list_del_init(&vma->exec_list); 232 kprintf("i915_gem_evict_something: EVICT VMA size=%lld (%016llx,%lld) ", 233 vma->size, vma->node.start, vma->node.size); 234 __i915_vma_unpin(vma); 235 if (ret == 0) 236 ret = i915_vma_unbind(vma); 237 kprintf("ret=%d\n", ret); 238 } 239 240 while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) { 241 vma = container_of(node, struct i915_vma, node); 242 ret = i915_vma_unbind(vma); 243 } 244 245 return ret; 246 } 247 248 /** 249 * i915_gem_evict_for_vma - Evict vmas to make room for binding a new one 250 * @vm: address space to evict from 251 * @target: range (and color) to evict for 252 * @flags: additional flags to control the eviction algorithm 253 * 254 * This function will try to evict vmas that overlap the target node. 255 * 256 * To clarify: This is for freeing up virtual address space, not for freeing 257 * memory in e.g. the shrinker. 258 */ 259 int i915_gem_evict_for_node(struct i915_address_space *vm, 260 struct drm_mm_node *target, 261 unsigned int flags) 262 { 263 LINUX_LIST_HEAD(eviction_list); 264 struct drm_mm_node *node; 265 u64 start = target->start; 266 u64 end = start + target->size; 267 struct i915_vma *vma, *next; 268 bool check_color; 269 int ret = 0; 270 271 lockdep_assert_held(&vm->i915->drm.struct_mutex); 272 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE)); 273 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE)); 274 275 trace_i915_gem_evict_node(vm, target, flags); 276 277 /* Retire before we search the active list. Although we have 278 * reasonable accuracy in our retirement lists, we may have 279 * a stray pin (preventing eviction) that can only be resolved by 280 * retiring. 281 */ 282 if (!(flags & PIN_NONBLOCK)) 283 i915_gem_retire_requests(vm->i915); 284 285 check_color = vm->mm.color_adjust; 286 if (check_color) { 287 /* Expand search to cover neighbouring guard pages (or lack!) */ 288 if (start) 289 start -= I915_GTT_PAGE_SIZE; 290 291 /* Always look at the page afterwards to avoid the end-of-GTT */ 292 end += I915_GTT_PAGE_SIZE; 293 } 294 GEM_BUG_ON(start >= end); 295 296 drm_mm_for_each_node_in_range(node, &vm->mm, start, end) { 297 /* If we find any non-objects (!vma), we cannot evict them */ 298 if (node->color == I915_COLOR_UNEVICTABLE) { 299 ret = -ENOSPC; 300 break; 301 } 302 303 GEM_BUG_ON(!node->allocated); 304 vma = container_of(node, typeof(*vma), node); 305 306 /* If we are using coloring to insert guard pages between 307 * different cache domains within the address space, we have 308 * to check whether the objects on either side of our range 309 * abutt and conflict. If they are in conflict, then we evict 310 * those as well to make room for our guard pages. 311 */ 312 if (check_color) { 313 if (node->start + node->size == target->start) { 314 if (node->color == target->color) 315 continue; 316 } 317 if (node->start == target->start + target->size) { 318 if (node->color == target->color) 319 continue; 320 } 321 } 322 323 if (flags & PIN_NONBLOCK && 324 (i915_vma_is_pinned(vma) || i915_vma_is_active(vma))) { 325 ret = -ENOSPC; 326 break; 327 } 328 329 /* Overlap of objects in the same batch? */ 330 if (i915_vma_is_pinned(vma) || !list_empty(&vma->exec_list)) { 331 ret = -ENOSPC; 332 if (vma->exec_entry && 333 vma->exec_entry->flags & EXEC_OBJECT_PINNED) 334 ret = -EINVAL; 335 break; 336 } 337 338 /* Never show fear in the face of dragons! 339 * 340 * We cannot directly remove this node from within this 341 * iterator and as with i915_gem_evict_something() we employ 342 * the vma pin_count in order to prevent the action of 343 * unbinding one vma from freeing (by dropping its active 344 * reference) another in our eviction list. 345 */ 346 __i915_vma_pin(vma); 347 list_add(&vma->exec_list, &eviction_list); 348 } 349 350 list_for_each_entry_safe(vma, next, &eviction_list, exec_list) { 351 list_del_init(&vma->exec_list); 352 __i915_vma_unpin(vma); 353 if (ret == 0) 354 ret = i915_vma_unbind(vma); 355 } 356 357 return ret; 358 } 359 360 /** 361 * i915_gem_evict_vm - Evict all idle vmas from a vm 362 * @vm: Address space to cleanse 363 * @do_idle: Boolean directing whether to idle first. 364 * 365 * This function evicts all idles vmas from a vm. If all unpinned vmas should be 366 * evicted the @do_idle needs to be set to true. 367 * 368 * This is used by the execbuf code as a last-ditch effort to defragment the 369 * address space. 370 * 371 * To clarify: This is for freeing up virtual address space, not for freeing 372 * memory in e.g. the shrinker. 373 */ 374 int i915_gem_evict_vm(struct i915_address_space *vm, bool do_idle) 375 { 376 struct i915_vma *vma, *next; 377 int ret; 378 379 lockdep_assert_held(&vm->i915->drm.struct_mutex); 380 trace_i915_gem_evict_vm(vm); 381 382 if (do_idle) { 383 struct drm_i915_private *dev_priv = vm->i915; 384 385 if (i915_is_ggtt(vm)) { 386 ret = i915_gem_switch_to_kernel_context(dev_priv); 387 if (ret) 388 return ret; 389 } 390 391 ret = i915_gem_wait_for_idle(dev_priv, 392 I915_WAIT_INTERRUPTIBLE | 393 I915_WAIT_LOCKED); 394 if (ret) 395 return ret; 396 397 WARN_ON(!list_empty(&vm->active_list)); 398 } 399 400 list_for_each_entry_safe(vma, next, &vm->inactive_list, vm_link) 401 if (!i915_vma_is_pinned(vma)) 402 WARN_ON(i915_vma_unbind(vma)); 403 404 return 0; 405 } 406 407 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 408 #include "selftests/i915_gem_evict.c" 409 #endif 410