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