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 I915_SELFTEST_DECLARE(static struct igt_evict_ctl { 37 bool fail_if_busy:1; 38 } igt_evict_ctl;) 39 40 static bool ggtt_is_idle(struct drm_i915_private *i915) 41 { 42 struct intel_engine_cs *engine; 43 enum intel_engine_id id; 44 45 if (i915->gt.active_requests) 46 return false; 47 48 for_each_engine(engine, i915, id) { 49 if (!intel_engine_has_kernel_context(engine)) 50 return false; 51 } 52 53 return true; 54 } 55 56 static int ggtt_flush(struct drm_i915_private *i915) 57 { 58 int err; 59 60 /* Not everything in the GGTT is tracked via vma (otherwise we 61 * could evict as required with minimal stalling) so we are forced 62 * to idle the GPU and explicitly retire outstanding requests in 63 * the hopes that we can then remove contexts and the like only 64 * bound by their active reference. 65 */ 66 err = i915_gem_switch_to_kernel_context(i915); 67 if (err) 68 return err; 69 70 err = i915_gem_wait_for_idle(i915, 71 I915_WAIT_INTERRUPTIBLE | 72 I915_WAIT_LOCKED, 73 MAX_SCHEDULE_TIMEOUT); 74 if (err) 75 return err; 76 77 GEM_BUG_ON(!ggtt_is_idle(i915)); 78 return 0; 79 } 80 81 static bool 82 mark_free(struct drm_mm_scan *scan, 83 struct i915_vma *vma, 84 unsigned int flags, 85 struct list_head *unwind) 86 { 87 if (i915_vma_is_pinned(vma)) 88 return false; 89 90 if (flags & PIN_NONFAULT && i915_vma_has_userfault(vma)) 91 return false; 92 93 list_add(&vma->evict_link, unwind); 94 return drm_mm_scan_add_block(scan, &vma->node); 95 } 96 97 /** 98 * i915_gem_evict_something - Evict vmas to make room for binding a new one 99 * @vm: address space to evict from 100 * @min_size: size of the desired free space 101 * @alignment: alignment constraint of the desired free space 102 * @cache_level: cache_level for the desired space 103 * @start: start (inclusive) of the range from which to evict objects 104 * @end: end (exclusive) of the range from which to evict objects 105 * @flags: additional flags to control the eviction algorithm 106 * 107 * This function will try to evict vmas until a free space satisfying the 108 * requirements is found. Callers must check first whether any such hole exists 109 * already before calling this function. 110 * 111 * This function is used by the object/vma binding code. 112 * 113 * Since this function is only used to free up virtual address space it only 114 * ignores pinned vmas, and not object where the backing storage itself is 115 * pinned. Hence obj->pages_pin_count does not protect against eviction. 116 * 117 * To clarify: This is for freeing up virtual address space, not for freeing 118 * memory in e.g. the shrinker. 119 */ 120 int 121 i915_gem_evict_something(struct i915_address_space *vm, 122 u64 min_size, u64 alignment, 123 unsigned cache_level, 124 u64 start, u64 end, 125 unsigned flags) 126 { 127 struct drm_i915_private *dev_priv = vm->i915; 128 struct drm_mm_scan scan; 129 struct list_head eviction_list; 130 struct list_head *phases[] = { 131 &vm->inactive_list, 132 &vm->active_list, 133 NULL, 134 }, **phase; 135 struct i915_vma *vma, *next; 136 struct drm_mm_node *node; 137 enum drm_mm_insert_mode mode; 138 int ret; 139 140 lockdep_assert_held(&vm->i915->drm.struct_mutex); 141 trace_i915_gem_evict(vm, min_size, alignment, flags); 142 143 /* 144 * The goal is to evict objects and amalgamate space in LRU order. 145 * The oldest idle objects reside on the inactive list, which is in 146 * retirement order. The next objects to retire are those in flight, 147 * on the active list, again in retirement order. 148 * 149 * The retirement sequence is thus: 150 * 1. Inactive objects (already retired) 151 * 2. Active objects (will stall on unbinding) 152 * 153 * On each list, the oldest objects lie at the HEAD with the freshest 154 * object on the TAIL. 155 */ 156 mode = DRM_MM_INSERT_BEST; 157 if (flags & PIN_HIGH) 158 mode = DRM_MM_INSERT_HIGH; 159 if (flags & PIN_MAPPABLE) 160 mode = DRM_MM_INSERT_LOW; 161 drm_mm_scan_init_with_range(&scan, &vm->mm, 162 min_size, alignment, cache_level, 163 start, end, mode); 164 165 /* 166 * Retire before we search the active list. Although we have 167 * reasonable accuracy in our retirement lists, we may have 168 * a stray pin (preventing eviction) that can only be resolved by 169 * retiring. 170 */ 171 if (!(flags & PIN_NONBLOCK)) 172 i915_retire_requests(dev_priv); 173 else 174 phases[1] = NULL; 175 176 search_again: 177 INIT_LIST_HEAD(&eviction_list); 178 phase = phases; 179 do { 180 list_for_each_entry(vma, *phase, vm_link) 181 if (mark_free(&scan, vma, flags, &eviction_list)) 182 goto found; 183 } while (*++phase); 184 185 /* Nothing found, clean up and bail out! */ 186 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 187 ret = drm_mm_scan_remove_block(&scan, &vma->node); 188 BUG_ON(ret); 189 } 190 191 /* 192 * Can we unpin some objects such as idle hw contents, 193 * or pending flips? But since only the GGTT has global entries 194 * such as scanouts, rinbuffers and contexts, we can skip the 195 * purge when inspecting per-process local address spaces. 196 */ 197 if (!i915_is_ggtt(vm) || flags & PIN_NONBLOCK) 198 return -ENOSPC; 199 200 /* 201 * Not everything in the GGTT is tracked via VMA using 202 * i915_vma_move_to_active(), otherwise we could evict as required 203 * with minimal stalling. Instead we are forced to idle the GPU and 204 * explicitly retire outstanding requests which will then remove 205 * the pinning for active objects such as contexts and ring, 206 * enabling us to evict them on the next iteration. 207 * 208 * To ensure that all user contexts are evictable, we perform 209 * a switch to the perma-pinned kernel context. This all also gives 210 * us a termination condition, when the last retired context is 211 * the kernel's there is no more we can evict. 212 */ 213 if (!ggtt_is_idle(dev_priv)) { 214 if (I915_SELFTEST_ONLY(igt_evict_ctl.fail_if_busy)) 215 return -EBUSY; 216 217 ret = ggtt_flush(dev_priv); 218 if (ret) 219 return ret; 220 221 cond_resched(); 222 goto search_again; 223 } 224 225 /* 226 * If we still have pending pageflip completions, drop 227 * back to userspace to give our workqueues time to 228 * acquire our locks and unpin the old scanouts. 229 */ 230 return intel_has_pending_fb_unpin(dev_priv) ? -EAGAIN : -ENOSPC; 231 232 found: 233 /* drm_mm doesn't allow any other other operations while 234 * scanning, therefore store to-be-evicted objects on a 235 * temporary list and take a reference for all before 236 * calling unbind (which may remove the active reference 237 * of any of our objects, thus corrupting the list). 238 */ 239 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 240 if (drm_mm_scan_remove_block(&scan, &vma->node)) 241 __i915_vma_pin(vma); 242 else 243 list_del(&vma->evict_link); 244 } 245 246 /* Unbinding will emit any required flushes */ 247 ret = 0; 248 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 249 __i915_vma_unpin(vma); 250 if (ret == 0) 251 ret = i915_vma_unbind(vma); 252 } 253 254 while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) { 255 vma = container_of(node, struct i915_vma, node); 256 ret = i915_vma_unbind(vma); 257 } 258 259 return ret; 260 } 261 262 /** 263 * i915_gem_evict_for_vma - Evict vmas to make room for binding a new one 264 * @vm: address space to evict from 265 * @target: range (and color) to evict for 266 * @flags: additional flags to control the eviction algorithm 267 * 268 * This function will try to evict vmas that overlap the target node. 269 * 270 * To clarify: This is for freeing up virtual address space, not for freeing 271 * memory in e.g. the shrinker. 272 */ 273 int i915_gem_evict_for_node(struct i915_address_space *vm, 274 struct drm_mm_node *target, 275 unsigned int flags) 276 { 277 LIST_HEAD(eviction_list); 278 struct drm_mm_node *node; 279 u64 start = target->start; 280 u64 end = start + target->size; 281 struct i915_vma *vma, *next; 282 bool check_color; 283 int ret = 0; 284 285 lockdep_assert_held(&vm->i915->drm.struct_mutex); 286 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE)); 287 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE)); 288 289 trace_i915_gem_evict_node(vm, target, flags); 290 291 /* Retire before we search the active list. Although we have 292 * reasonable accuracy in our retirement lists, we may have 293 * a stray pin (preventing eviction) that can only be resolved by 294 * retiring. 295 */ 296 if (!(flags & PIN_NONBLOCK)) 297 i915_retire_requests(vm->i915); 298 299 check_color = vm->mm.color_adjust; 300 if (check_color) { 301 /* Expand search to cover neighbouring guard pages (or lack!) */ 302 if (start) 303 start -= I915_GTT_PAGE_SIZE; 304 305 /* Always look at the page afterwards to avoid the end-of-GTT */ 306 end += I915_GTT_PAGE_SIZE; 307 } 308 GEM_BUG_ON(start >= end); 309 310 drm_mm_for_each_node_in_range(node, &vm->mm, start, end) { 311 /* If we find any non-objects (!vma), we cannot evict them */ 312 if (node->color == I915_COLOR_UNEVICTABLE) { 313 ret = -ENOSPC; 314 break; 315 } 316 317 GEM_BUG_ON(!node->allocated); 318 vma = container_of(node, typeof(*vma), node); 319 320 /* If we are using coloring to insert guard pages between 321 * different cache domains within the address space, we have 322 * to check whether the objects on either side of our range 323 * abutt and conflict. If they are in conflict, then we evict 324 * those as well to make room for our guard pages. 325 */ 326 if (check_color) { 327 if (node->start + node->size == target->start) { 328 if (node->color == target->color) 329 continue; 330 } 331 if (node->start == target->start + target->size) { 332 if (node->color == target->color) 333 continue; 334 } 335 } 336 337 if (flags & PIN_NONBLOCK && 338 (i915_vma_is_pinned(vma) || i915_vma_is_active(vma))) { 339 ret = -ENOSPC; 340 break; 341 } 342 343 if (flags & PIN_NONFAULT && i915_vma_has_userfault(vma)) { 344 ret = -ENOSPC; 345 break; 346 } 347 348 /* Overlap of objects in the same batch? */ 349 if (i915_vma_is_pinned(vma)) { 350 ret = -ENOSPC; 351 if (vma->exec_flags && 352 *vma->exec_flags & EXEC_OBJECT_PINNED) 353 ret = -EINVAL; 354 break; 355 } 356 357 /* Never show fear in the face of dragons! 358 * 359 * We cannot directly remove this node from within this 360 * iterator and as with i915_gem_evict_something() we employ 361 * the vma pin_count in order to prevent the action of 362 * unbinding one vma from freeing (by dropping its active 363 * reference) another in our eviction list. 364 */ 365 __i915_vma_pin(vma); 366 list_add(&vma->evict_link, &eviction_list); 367 } 368 369 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 370 __i915_vma_unpin(vma); 371 if (ret == 0) 372 ret = i915_vma_unbind(vma); 373 } 374 375 return ret; 376 } 377 378 /** 379 * i915_gem_evict_vm - Evict all idle vmas from a vm 380 * @vm: Address space to cleanse 381 * 382 * This function evicts all vmas from a vm. 383 * 384 * This is used by the execbuf code as a last-ditch effort to defragment the 385 * address space. 386 * 387 * To clarify: This is for freeing up virtual address space, not for freeing 388 * memory in e.g. the shrinker. 389 */ 390 int i915_gem_evict_vm(struct i915_address_space *vm) 391 { 392 struct list_head *phases[] = { 393 &vm->inactive_list, 394 &vm->active_list, 395 NULL 396 }, **phase; 397 struct list_head eviction_list; 398 struct i915_vma *vma, *next; 399 int ret; 400 401 lockdep_assert_held(&vm->i915->drm.struct_mutex); 402 trace_i915_gem_evict_vm(vm); 403 404 /* Switch back to the default context in order to unpin 405 * the existing context objects. However, such objects only 406 * pin themselves inside the global GTT and performing the 407 * switch otherwise is ineffective. 408 */ 409 if (i915_is_ggtt(vm)) { 410 ret = ggtt_flush(vm->i915); 411 if (ret) 412 return ret; 413 } 414 415 INIT_LIST_HEAD(&eviction_list); 416 phase = phases; 417 do { 418 list_for_each_entry(vma, *phase, vm_link) { 419 if (i915_vma_is_pinned(vma)) 420 continue; 421 422 __i915_vma_pin(vma); 423 list_add(&vma->evict_link, &eviction_list); 424 } 425 } while (*++phase); 426 427 ret = 0; 428 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 429 __i915_vma_unpin(vma); 430 if (ret == 0) 431 ret = i915_vma_unbind(vma); 432 } 433 return ret; 434 } 435 436 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 437 #include "selftests/i915_gem_evict.c" 438 #endif 439