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 (engine->last_retired_context != i915->kernel_context) 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 if (err) 74 return err; 75 76 return 0; 77 } 78 79 static bool 80 mark_free(struct drm_mm_scan *scan, 81 struct i915_vma *vma, 82 unsigned int flags, 83 struct list_head *unwind) 84 { 85 if (i915_vma_is_pinned(vma)) 86 return false; 87 88 if (flags & PIN_NONFAULT && i915_vma_has_userfault(vma)) 89 return false; 90 91 list_add(&vma->evict_link, unwind); 92 return drm_mm_scan_add_block(scan, &vma->node); 93 } 94 95 /** 96 * i915_gem_evict_something - Evict vmas to make room for binding a new one 97 * @vm: address space to evict from 98 * @min_size: size of the desired free space 99 * @alignment: alignment constraint of the desired free space 100 * @cache_level: cache_level for the desired space 101 * @start: start (inclusive) of the range from which to evict objects 102 * @end: end (exclusive) of the range from which to evict objects 103 * @flags: additional flags to control the eviction algorithm 104 * 105 * This function will try to evict vmas until a free space satisfying the 106 * requirements is found. Callers must check first whether any such hole exists 107 * already before calling this function. 108 * 109 * This function is used by the object/vma binding code. 110 * 111 * Since this function is only used to free up virtual address space it only 112 * ignores pinned vmas, and not object where the backing storage itself is 113 * pinned. Hence obj->pages_pin_count does not protect against eviction. 114 * 115 * To clarify: This is for freeing up virtual address space, not for freeing 116 * memory in e.g. the shrinker. 117 */ 118 int 119 i915_gem_evict_something(struct i915_address_space *vm, 120 u64 min_size, u64 alignment, 121 unsigned cache_level, 122 u64 start, u64 end, 123 unsigned flags) 124 { 125 struct drm_i915_private *dev_priv = vm->i915; 126 struct drm_mm_scan scan; 127 struct list_head eviction_list; 128 struct list_head *phases[] = { 129 &vm->inactive_list, 130 &vm->active_list, 131 NULL, 132 }, **phase; 133 struct i915_vma *vma, *next; 134 struct drm_mm_node *node; 135 enum drm_mm_insert_mode mode; 136 int ret; 137 138 lockdep_assert_held(&vm->i915->drm.struct_mutex); 139 trace_i915_gem_evict(vm, min_size, alignment, flags); 140 141 /* 142 * The goal is to evict objects and amalgamate space in LRU order. 143 * The oldest idle objects reside on the inactive list, which is in 144 * retirement order. The next objects to retire are those in flight, 145 * on the active list, again in retirement order. 146 * 147 * The retirement sequence is thus: 148 * 1. Inactive objects (already retired) 149 * 2. Active objects (will stall on unbinding) 150 * 151 * On each list, the oldest objects lie at the HEAD with the freshest 152 * object on the TAIL. 153 */ 154 mode = DRM_MM_INSERT_BEST; 155 if (flags & PIN_HIGH) 156 mode = DRM_MM_INSERT_HIGH; 157 if (flags & PIN_MAPPABLE) 158 mode = DRM_MM_INSERT_LOW; 159 drm_mm_scan_init_with_range(&scan, &vm->mm, 160 min_size, alignment, cache_level, 161 start, end, mode); 162 163 /* 164 * Retire before we search the active list. Although we have 165 * reasonable accuracy in our retirement lists, we may have 166 * a stray pin (preventing eviction) that can only be resolved by 167 * retiring. 168 */ 169 if (!(flags & PIN_NONBLOCK)) 170 i915_gem_retire_requests(dev_priv); 171 else 172 phases[1] = NULL; 173 174 search_again: 175 INIT_LIST_HEAD(&eviction_list); 176 phase = phases; 177 do { 178 list_for_each_entry(vma, *phase, vm_link) 179 if (mark_free(&scan, vma, flags, &eviction_list)) 180 goto found; 181 } while (*++phase); 182 183 /* Nothing found, clean up and bail out! */ 184 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 185 ret = drm_mm_scan_remove_block(&scan, &vma->node); 186 BUG_ON(ret); 187 } 188 189 /* 190 * Can we unpin some objects such as idle hw contents, 191 * or pending flips? But since only the GGTT has global entries 192 * such as scanouts, rinbuffers and contexts, we can skip the 193 * purge when inspecting per-process local address spaces. 194 */ 195 if (!i915_is_ggtt(vm) || flags & PIN_NONBLOCK) 196 return -ENOSPC; 197 198 /* 199 * Not everything in the GGTT is tracked via VMA using 200 * i915_vma_move_to_active(), otherwise we could evict as required 201 * with minimal stalling. Instead we are forced to idle the GPU and 202 * explicitly retire outstanding requests which will then remove 203 * the pinning for active objects such as contexts and ring, 204 * enabling us to evict them on the next iteration. 205 * 206 * To ensure that all user contexts are evictable, we perform 207 * a switch to the perma-pinned kernel context. This all also gives 208 * us a termination condition, when the last retired context is 209 * the kernel's there is no more we can evict. 210 */ 211 if (!ggtt_is_idle(dev_priv)) { 212 if (I915_SELFTEST_ONLY(igt_evict_ctl.fail_if_busy)) 213 return -EBUSY; 214 215 ret = ggtt_flush(dev_priv); 216 if (ret) 217 return ret; 218 219 goto search_again; 220 } 221 222 /* 223 * If we still have pending pageflip completions, drop 224 * back to userspace to give our workqueues time to 225 * acquire our locks and unpin the old scanouts. 226 */ 227 return intel_has_pending_fb_unpin(dev_priv) ? -EAGAIN : -ENOSPC; 228 229 found: 230 /* drm_mm doesn't allow any other other operations while 231 * scanning, therefore store to-be-evicted objects on a 232 * temporary list and take a reference for all before 233 * calling unbind (which may remove the active reference 234 * of any of our objects, thus corrupting the list). 235 */ 236 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 237 if (drm_mm_scan_remove_block(&scan, &vma->node)) 238 __i915_vma_pin(vma); 239 else 240 list_del(&vma->evict_link); 241 } 242 243 /* Unbinding will emit any required flushes */ 244 ret = 0; 245 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 246 __i915_vma_unpin(vma); 247 if (ret == 0) 248 ret = i915_vma_unbind(vma); 249 } 250 251 while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) { 252 vma = container_of(node, struct i915_vma, node); 253 ret = i915_vma_unbind(vma); 254 } 255 256 return ret; 257 } 258 259 /** 260 * i915_gem_evict_for_vma - Evict vmas to make room for binding a new one 261 * @vm: address space to evict from 262 * @target: range (and color) to evict for 263 * @flags: additional flags to control the eviction algorithm 264 * 265 * This function will try to evict vmas that overlap the target node. 266 * 267 * To clarify: This is for freeing up virtual address space, not for freeing 268 * memory in e.g. the shrinker. 269 */ 270 int i915_gem_evict_for_node(struct i915_address_space *vm, 271 struct drm_mm_node *target, 272 unsigned int flags) 273 { 274 LIST_HEAD(eviction_list); 275 struct drm_mm_node *node; 276 u64 start = target->start; 277 u64 end = start + target->size; 278 struct i915_vma *vma, *next; 279 bool check_color; 280 int ret = 0; 281 282 lockdep_assert_held(&vm->i915->drm.struct_mutex); 283 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE)); 284 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE)); 285 286 trace_i915_gem_evict_node(vm, target, flags); 287 288 /* Retire before we search the active list. Although we have 289 * reasonable accuracy in our retirement lists, we may have 290 * a stray pin (preventing eviction) that can only be resolved by 291 * retiring. 292 */ 293 if (!(flags & PIN_NONBLOCK)) 294 i915_gem_retire_requests(vm->i915); 295 296 check_color = vm->mm.color_adjust; 297 if (check_color) { 298 /* Expand search to cover neighbouring guard pages (or lack!) */ 299 if (start) 300 start -= I915_GTT_PAGE_SIZE; 301 302 /* Always look at the page afterwards to avoid the end-of-GTT */ 303 end += I915_GTT_PAGE_SIZE; 304 } 305 GEM_BUG_ON(start >= end); 306 307 drm_mm_for_each_node_in_range(node, &vm->mm, start, end) { 308 /* If we find any non-objects (!vma), we cannot evict them */ 309 if (node->color == I915_COLOR_UNEVICTABLE) { 310 ret = -ENOSPC; 311 break; 312 } 313 314 GEM_BUG_ON(!node->allocated); 315 vma = container_of(node, typeof(*vma), node); 316 317 /* If we are using coloring to insert guard pages between 318 * different cache domains within the address space, we have 319 * to check whether the objects on either side of our range 320 * abutt and conflict. If they are in conflict, then we evict 321 * those as well to make room for our guard pages. 322 */ 323 if (check_color) { 324 if (node->start + node->size == target->start) { 325 if (node->color == target->color) 326 continue; 327 } 328 if (node->start == target->start + target->size) { 329 if (node->color == target->color) 330 continue; 331 } 332 } 333 334 if (flags & PIN_NONBLOCK && 335 (i915_vma_is_pinned(vma) || i915_vma_is_active(vma))) { 336 ret = -ENOSPC; 337 break; 338 } 339 340 if (flags & PIN_NONFAULT && i915_vma_has_userfault(vma)) { 341 ret = -ENOSPC; 342 break; 343 } 344 345 /* Overlap of objects in the same batch? */ 346 if (i915_vma_is_pinned(vma)) { 347 ret = -ENOSPC; 348 if (vma->exec_flags && 349 *vma->exec_flags & EXEC_OBJECT_PINNED) 350 ret = -EINVAL; 351 break; 352 } 353 354 /* Never show fear in the face of dragons! 355 * 356 * We cannot directly remove this node from within this 357 * iterator and as with i915_gem_evict_something() we employ 358 * the vma pin_count in order to prevent the action of 359 * unbinding one vma from freeing (by dropping its active 360 * reference) another in our eviction list. 361 */ 362 __i915_vma_pin(vma); 363 list_add(&vma->evict_link, &eviction_list); 364 } 365 366 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 367 __i915_vma_unpin(vma); 368 if (ret == 0) 369 ret = i915_vma_unbind(vma); 370 } 371 372 return ret; 373 } 374 375 /** 376 * i915_gem_evict_vm - Evict all idle vmas from a vm 377 * @vm: Address space to cleanse 378 * 379 * This function evicts all vmas from a vm. 380 * 381 * This is used by the execbuf code as a last-ditch effort to defragment the 382 * address space. 383 * 384 * To clarify: This is for freeing up virtual address space, not for freeing 385 * memory in e.g. the shrinker. 386 */ 387 int i915_gem_evict_vm(struct i915_address_space *vm) 388 { 389 struct list_head *phases[] = { 390 &vm->inactive_list, 391 &vm->active_list, 392 NULL 393 }, **phase; 394 struct list_head eviction_list; 395 struct i915_vma *vma, *next; 396 int ret; 397 398 lockdep_assert_held(&vm->i915->drm.struct_mutex); 399 trace_i915_gem_evict_vm(vm); 400 401 /* Switch back to the default context in order to unpin 402 * the existing context objects. However, such objects only 403 * pin themselves inside the global GTT and performing the 404 * switch otherwise is ineffective. 405 */ 406 if (i915_is_ggtt(vm)) { 407 ret = ggtt_flush(vm->i915); 408 if (ret) 409 return ret; 410 } 411 412 INIT_LIST_HEAD(&eviction_list); 413 phase = phases; 414 do { 415 list_for_each_entry(vma, *phase, vm_link) { 416 if (i915_vma_is_pinned(vma)) 417 continue; 418 419 __i915_vma_pin(vma); 420 list_add(&vma->evict_link, &eviction_list); 421 } 422 } while (*++phase); 423 424 ret = 0; 425 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 426 __i915_vma_unpin(vma); 427 if (ret == 0) 428 ret = i915_vma_unbind(vma); 429 } 430 return ret; 431 } 432 433 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 434 #include "selftests/i915_gem_evict.c" 435 #endif 436