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/i915_drm.h> 30 31 #include "i915_drv.h" 32 #include "intel_drv.h" 33 #include "i915_trace.h" 34 35 I915_SELFTEST_DECLARE(static struct igt_evict_ctl { 36 bool fail_if_busy:1; 37 } igt_evict_ctl;) 38 39 static int ggtt_flush(struct drm_i915_private *i915) 40 { 41 /* 42 * Not everything in the GGTT is tracked via vma (otherwise we 43 * could evict as required with minimal stalling) so we are forced 44 * to idle the GPU and explicitly retire outstanding requests in 45 * the hopes that we can then remove contexts and the like only 46 * bound by their active reference. 47 */ 48 return i915_gem_wait_for_idle(i915, 49 I915_WAIT_INTERRUPTIBLE | 50 I915_WAIT_LOCKED, 51 MAX_SCHEDULE_TIMEOUT); 52 } 53 54 static bool 55 mark_free(struct drm_mm_scan *scan, 56 struct i915_vma *vma, 57 unsigned int flags, 58 struct list_head *unwind) 59 { 60 if (i915_vma_is_pinned(vma)) 61 return false; 62 63 if (flags & PIN_NONFAULT && i915_vma_has_userfault(vma)) 64 return false; 65 66 list_add(&vma->evict_link, unwind); 67 return drm_mm_scan_add_block(scan, &vma->node); 68 } 69 70 /** 71 * i915_gem_evict_something - Evict vmas to make room for binding a new one 72 * @vm: address space to evict from 73 * @min_size: size of the desired free space 74 * @alignment: alignment constraint of the desired free space 75 * @cache_level: cache_level for the desired space 76 * @start: start (inclusive) of the range from which to evict objects 77 * @end: end (exclusive) of the range from which to evict objects 78 * @flags: additional flags to control the eviction algorithm 79 * 80 * This function will try to evict vmas until a free space satisfying the 81 * requirements is found. Callers must check first whether any such hole exists 82 * already before calling this function. 83 * 84 * This function is used by the object/vma binding code. 85 * 86 * Since this function is only used to free up virtual address space it only 87 * ignores pinned vmas, and not object where the backing storage itself is 88 * pinned. Hence obj->pages_pin_count does not protect against eviction. 89 * 90 * To clarify: This is for freeing up virtual address space, not for freeing 91 * memory in e.g. the shrinker. 92 */ 93 int 94 i915_gem_evict_something(struct i915_address_space *vm, 95 u64 min_size, u64 alignment, 96 unsigned cache_level, 97 u64 start, u64 end, 98 unsigned flags) 99 { 100 struct drm_i915_private *dev_priv = vm->i915; 101 struct drm_mm_scan scan; 102 struct list_head eviction_list; 103 struct i915_vma *vma, *next; 104 struct drm_mm_node *node; 105 enum drm_mm_insert_mode mode; 106 struct i915_vma *active; 107 int ret; 108 109 lockdep_assert_held(&vm->i915->drm.struct_mutex); 110 trace_i915_gem_evict(vm, min_size, alignment, flags); 111 112 /* 113 * The goal is to evict objects and amalgamate space in rough LRU order. 114 * Since both active and inactive objects reside on the same list, 115 * in a mix of creation and last scanned order, as we process the list 116 * we sort it into inactive/active, which keeps the active portion 117 * in a rough MRU order. 118 * 119 * The retirement sequence is thus: 120 * 1. Inactive objects (already retired, random order) 121 * 2. Active objects (will stall on unbinding, oldest scanned first) 122 */ 123 mode = DRM_MM_INSERT_BEST; 124 if (flags & PIN_HIGH) 125 mode = DRM_MM_INSERT_HIGH; 126 if (flags & PIN_MAPPABLE) 127 mode = DRM_MM_INSERT_LOW; 128 drm_mm_scan_init_with_range(&scan, &vm->mm, 129 min_size, alignment, cache_level, 130 start, end, mode); 131 132 /* 133 * Retire before we search the active list. Although we have 134 * reasonable accuracy in our retirement lists, we may have 135 * a stray pin (preventing eviction) that can only be resolved by 136 * retiring. 137 */ 138 if (!(flags & PIN_NONBLOCK)) 139 i915_retire_requests(dev_priv); 140 141 search_again: 142 active = NULL; 143 INIT_LIST_HEAD(&eviction_list); 144 list_for_each_entry_safe(vma, next, &vm->bound_list, vm_link) { 145 /* 146 * We keep this list in a rough least-recently scanned order 147 * of active elements (inactive elements are cheap to reap). 148 * New entries are added to the end, and we move anything we 149 * scan to the end. The assumption is that the working set 150 * of applications is either steady state (and thanks to the 151 * userspace bo cache it almost always is) or volatile and 152 * frequently replaced after a frame, which are self-evicting! 153 * Given that assumption, the MRU order of the scan list is 154 * fairly static, and keeping it in least-recently scan order 155 * is suitable. 156 * 157 * To notice when we complete one full cycle, we record the 158 * first active element seen, before moving it to the tail. 159 */ 160 if (i915_vma_is_active(vma)) { 161 if (vma == active) { 162 if (flags & PIN_NONBLOCK) 163 break; 164 165 active = ERR_PTR(-EAGAIN); 166 } 167 168 if (active != ERR_PTR(-EAGAIN)) { 169 if (!active) 170 active = vma; 171 172 list_move_tail(&vma->vm_link, &vm->bound_list); 173 continue; 174 } 175 } 176 177 if (mark_free(&scan, vma, flags, &eviction_list)) 178 goto found; 179 } 180 181 /* Nothing found, clean up and bail out! */ 182 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 183 ret = drm_mm_scan_remove_block(&scan, &vma->node); 184 BUG_ON(ret); 185 } 186 187 /* 188 * Can we unpin some objects such as idle hw contents, 189 * or pending flips? But since only the GGTT has global entries 190 * such as scanouts, rinbuffers and contexts, we can skip the 191 * purge when inspecting per-process local address spaces. 192 */ 193 if (!i915_is_ggtt(vm) || flags & PIN_NONBLOCK) 194 return -ENOSPC; 195 196 /* 197 * Not everything in the GGTT is tracked via VMA using 198 * i915_vma_move_to_active(), otherwise we could evict as required 199 * with minimal stalling. Instead we are forced to idle the GPU and 200 * explicitly retire outstanding requests which will then remove 201 * the pinning for active objects such as contexts and ring, 202 * enabling us to evict them on the next iteration. 203 * 204 * To ensure that all user contexts are evictable, we perform 205 * a switch to the perma-pinned kernel context. This all also gives 206 * us a termination condition, when the last retired context is 207 * the kernel's there is no more we can evict. 208 */ 209 if (I915_SELFTEST_ONLY(igt_evict_ctl.fail_if_busy)) 210 return -EBUSY; 211 212 ret = ggtt_flush(dev_priv); 213 if (ret) 214 return ret; 215 216 cond_resched(); 217 218 flags |= PIN_NONBLOCK; 219 goto search_again; 220 221 found: 222 /* drm_mm doesn't allow any other other operations while 223 * scanning, therefore store to-be-evicted objects on a 224 * temporary list and take a reference for all before 225 * calling unbind (which may remove the active reference 226 * of any of our objects, thus corrupting the list). 227 */ 228 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 229 if (drm_mm_scan_remove_block(&scan, &vma->node)) 230 __i915_vma_pin(vma); 231 else 232 list_del(&vma->evict_link); 233 } 234 235 /* Unbinding will emit any required flushes */ 236 ret = 0; 237 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 238 __i915_vma_unpin(vma); 239 if (ret == 0) 240 ret = i915_vma_unbind(vma); 241 } 242 243 while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) { 244 vma = container_of(node, struct i915_vma, node); 245 ret = i915_vma_unbind(vma); 246 } 247 248 return ret; 249 } 250 251 /** 252 * i915_gem_evict_for_vma - Evict vmas to make room for binding a new one 253 * @vm: address space to evict from 254 * @target: range (and color) to evict for 255 * @flags: additional flags to control the eviction algorithm 256 * 257 * This function will try to evict vmas that overlap the target node. 258 * 259 * To clarify: This is for freeing up virtual address space, not for freeing 260 * memory in e.g. the shrinker. 261 */ 262 int i915_gem_evict_for_node(struct i915_address_space *vm, 263 struct drm_mm_node *target, 264 unsigned int flags) 265 { 266 LIST_HEAD(eviction_list); 267 struct drm_mm_node *node; 268 u64 start = target->start; 269 u64 end = start + target->size; 270 struct i915_vma *vma, *next; 271 bool check_color; 272 int ret = 0; 273 274 lockdep_assert_held(&vm->i915->drm.struct_mutex); 275 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE)); 276 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE)); 277 278 trace_i915_gem_evict_node(vm, target, flags); 279 280 /* Retire before we search the active list. Although we have 281 * reasonable accuracy in our retirement lists, we may have 282 * a stray pin (preventing eviction) that can only be resolved by 283 * retiring. 284 */ 285 if (!(flags & PIN_NONBLOCK)) 286 i915_retire_requests(vm->i915); 287 288 check_color = vm->mm.color_adjust; 289 if (check_color) { 290 /* Expand search to cover neighbouring guard pages (or lack!) */ 291 if (start) 292 start -= I915_GTT_PAGE_SIZE; 293 294 /* Always look at the page afterwards to avoid the end-of-GTT */ 295 end += I915_GTT_PAGE_SIZE; 296 } 297 GEM_BUG_ON(start >= end); 298 299 drm_mm_for_each_node_in_range(node, &vm->mm, start, end) { 300 /* If we find any non-objects (!vma), we cannot evict them */ 301 if (node->color == I915_COLOR_UNEVICTABLE) { 302 ret = -ENOSPC; 303 break; 304 } 305 306 GEM_BUG_ON(!node->allocated); 307 vma = container_of(node, typeof(*vma), node); 308 309 /* If we are using coloring to insert guard pages between 310 * different cache domains within the address space, we have 311 * to check whether the objects on either side of our range 312 * abutt and conflict. If they are in conflict, then we evict 313 * those as well to make room for our guard pages. 314 */ 315 if (check_color) { 316 if (node->start + node->size == target->start) { 317 if (node->color == target->color) 318 continue; 319 } 320 if (node->start == target->start + target->size) { 321 if (node->color == target->color) 322 continue; 323 } 324 } 325 326 if (flags & PIN_NONBLOCK && 327 (i915_vma_is_pinned(vma) || i915_vma_is_active(vma))) { 328 ret = -ENOSPC; 329 break; 330 } 331 332 if (flags & PIN_NONFAULT && i915_vma_has_userfault(vma)) { 333 ret = -ENOSPC; 334 break; 335 } 336 337 /* Overlap of objects in the same batch? */ 338 if (i915_vma_is_pinned(vma)) { 339 ret = -ENOSPC; 340 if (vma->exec_flags && 341 *vma->exec_flags & EXEC_OBJECT_PINNED) 342 ret = -EINVAL; 343 break; 344 } 345 346 /* Never show fear in the face of dragons! 347 * 348 * We cannot directly remove this node from within this 349 * iterator and as with i915_gem_evict_something() we employ 350 * the vma pin_count in order to prevent the action of 351 * unbinding one vma from freeing (by dropping its active 352 * reference) another in our eviction list. 353 */ 354 __i915_vma_pin(vma); 355 list_add(&vma->evict_link, &eviction_list); 356 } 357 358 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 359 __i915_vma_unpin(vma); 360 if (ret == 0) 361 ret = i915_vma_unbind(vma); 362 } 363 364 return ret; 365 } 366 367 /** 368 * i915_gem_evict_vm - Evict all idle vmas from a vm 369 * @vm: Address space to cleanse 370 * 371 * This function evicts all vmas from a vm. 372 * 373 * This is used by the execbuf code as a last-ditch effort to defragment the 374 * address space. 375 * 376 * To clarify: This is for freeing up virtual address space, not for freeing 377 * memory in e.g. the shrinker. 378 */ 379 int i915_gem_evict_vm(struct i915_address_space *vm) 380 { 381 struct list_head eviction_list; 382 struct i915_vma *vma, *next; 383 int ret; 384 385 lockdep_assert_held(&vm->i915->drm.struct_mutex); 386 trace_i915_gem_evict_vm(vm); 387 388 /* Switch back to the default context in order to unpin 389 * the existing context objects. However, such objects only 390 * pin themselves inside the global GTT and performing the 391 * switch otherwise is ineffective. 392 */ 393 if (i915_is_ggtt(vm)) { 394 ret = ggtt_flush(vm->i915); 395 if (ret) 396 return ret; 397 } 398 399 INIT_LIST_HEAD(&eviction_list); 400 mutex_lock(&vm->mutex); 401 list_for_each_entry(vma, &vm->bound_list, vm_link) { 402 if (i915_vma_is_pinned(vma)) 403 continue; 404 405 __i915_vma_pin(vma); 406 list_add(&vma->evict_link, &eviction_list); 407 } 408 mutex_unlock(&vm->mutex); 409 410 ret = 0; 411 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 412 __i915_vma_unpin(vma); 413 if (ret == 0) 414 ret = i915_vma_unbind(vma); 415 } 416 return ret; 417 } 418 419 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 420 #include "selftests/i915_gem_evict.c" 421 #endif 422