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 "gem/i915_gem_context.h" 30 #include "gt/intel_gt_requests.h" 31 32 #include "i915_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 intel_gt *gt) 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 intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT); 49 } 50 51 static bool 52 mark_free(struct drm_mm_scan *scan, 53 struct i915_vma *vma, 54 unsigned int flags, 55 struct list_head *unwind) 56 { 57 if (i915_vma_is_pinned(vma)) 58 return false; 59 60 list_add(&vma->evict_link, unwind); 61 return drm_mm_scan_add_block(scan, &vma->node); 62 } 63 64 static bool defer_evict(struct i915_vma *vma) 65 { 66 if (i915_vma_is_active(vma)) 67 return true; 68 69 if (i915_vma_is_scanout(vma)) 70 return true; 71 72 return false; 73 } 74 75 /** 76 * i915_gem_evict_something - Evict vmas to make room for binding a new one 77 * @vm: address space to evict from 78 * @min_size: size of the desired free space 79 * @alignment: alignment constraint of the desired free space 80 * @color: color for the desired space 81 * @start: start (inclusive) of the range from which to evict objects 82 * @end: end (exclusive) of the range from which to evict objects 83 * @flags: additional flags to control the eviction algorithm 84 * 85 * This function will try to evict vmas until a free space satisfying the 86 * requirements is found. Callers must check first whether any such hole exists 87 * already before calling this function. 88 * 89 * This function is used by the object/vma binding code. 90 * 91 * Since this function is only used to free up virtual address space it only 92 * ignores pinned vmas, and not object where the backing storage itself is 93 * pinned. Hence obj->pages_pin_count does not protect against eviction. 94 * 95 * To clarify: This is for freeing up virtual address space, not for freeing 96 * memory in e.g. the shrinker. 97 */ 98 int 99 i915_gem_evict_something(struct i915_address_space *vm, 100 u64 min_size, u64 alignment, 101 unsigned long color, 102 u64 start, u64 end, 103 unsigned flags) 104 { 105 struct drm_mm_scan scan; 106 struct list_head eviction_list; 107 struct i915_vma *vma, *next; 108 struct drm_mm_node *node; 109 enum drm_mm_insert_mode mode; 110 struct i915_vma *active; 111 int ret; 112 113 lockdep_assert_held(&vm->mutex); 114 trace_i915_gem_evict(vm, min_size, alignment, flags); 115 116 /* 117 * The goal is to evict objects and amalgamate space in rough LRU order. 118 * Since both active and inactive objects reside on the same list, 119 * in a mix of creation and last scanned order, as we process the list 120 * we sort it into inactive/active, which keeps the active portion 121 * in a rough MRU order. 122 * 123 * The retirement sequence is thus: 124 * 1. Inactive objects (already retired, random order) 125 * 2. Active objects (will stall on unbinding, oldest scanned first) 126 */ 127 mode = DRM_MM_INSERT_BEST; 128 if (flags & PIN_HIGH) 129 mode = DRM_MM_INSERT_HIGH; 130 if (flags & PIN_MAPPABLE) 131 mode = DRM_MM_INSERT_LOW; 132 drm_mm_scan_init_with_range(&scan, &vm->mm, 133 min_size, alignment, color, 134 start, end, mode); 135 136 intel_gt_retire_requests(vm->gt); 137 138 search_again: 139 active = NULL; 140 INIT_LIST_HEAD(&eviction_list); 141 list_for_each_entry_safe(vma, next, &vm->bound_list, vm_link) { 142 if (vma == active) { /* now seen this vma twice */ 143 if (flags & PIN_NONBLOCK) 144 break; 145 146 active = ERR_PTR(-EAGAIN); 147 } 148 149 /* 150 * We keep this list in a rough least-recently scanned order 151 * of active elements (inactive elements are cheap to reap). 152 * New entries are added to the end, and we move anything we 153 * scan to the end. The assumption is that the working set 154 * of applications is either steady state (and thanks to the 155 * userspace bo cache it almost always is) or volatile and 156 * frequently replaced after a frame, which are self-evicting! 157 * Given that assumption, the MRU order of the scan list is 158 * fairly static, and keeping it in least-recently scan order 159 * is suitable. 160 * 161 * To notice when we complete one full cycle, we record the 162 * first active element seen, before moving it to the tail. 163 */ 164 if (active != ERR_PTR(-EAGAIN) && defer_evict(vma)) { 165 if (!active) 166 active = vma; 167 168 list_move_tail(&vma->vm_link, &vm->bound_list); 169 continue; 170 } 171 172 if (mark_free(&scan, vma, flags, &eviction_list)) 173 goto found; 174 } 175 176 /* Nothing found, clean up and bail out! */ 177 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 178 ret = drm_mm_scan_remove_block(&scan, &vma->node); 179 BUG_ON(ret); 180 } 181 182 /* 183 * Can we unpin some objects such as idle hw contents, 184 * or pending flips? But since only the GGTT has global entries 185 * such as scanouts, rinbuffers and contexts, we can skip the 186 * purge when inspecting per-process local address spaces. 187 */ 188 if (!i915_is_ggtt(vm) || flags & PIN_NONBLOCK) 189 return -ENOSPC; 190 191 /* 192 * Not everything in the GGTT is tracked via VMA using 193 * i915_vma_move_to_active(), otherwise we could evict as required 194 * with minimal stalling. Instead we are forced to idle the GPU and 195 * explicitly retire outstanding requests which will then remove 196 * the pinning for active objects such as contexts and ring, 197 * enabling us to evict them on the next iteration. 198 * 199 * To ensure that all user contexts are evictable, we perform 200 * a switch to the perma-pinned kernel context. This all also gives 201 * us a termination condition, when the last retired context is 202 * the kernel's there is no more we can evict. 203 */ 204 if (I915_SELFTEST_ONLY(igt_evict_ctl.fail_if_busy)) 205 return -EBUSY; 206 207 ret = ggtt_flush(vm->gt); 208 if (ret) 209 return ret; 210 211 cond_resched(); 212 213 flags |= PIN_NONBLOCK; 214 goto search_again; 215 216 found: 217 /* drm_mm doesn't allow any other other operations while 218 * scanning, therefore store to-be-evicted objects on a 219 * temporary list and take a reference for all before 220 * calling unbind (which may remove the active reference 221 * of any of our objects, thus corrupting the list). 222 */ 223 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 224 if (drm_mm_scan_remove_block(&scan, &vma->node)) 225 __i915_vma_pin(vma); 226 else 227 list_del(&vma->evict_link); 228 } 229 230 /* Unbinding will emit any required flushes */ 231 ret = 0; 232 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 233 __i915_vma_unpin(vma); 234 if (ret == 0) 235 ret = __i915_vma_unbind(vma); 236 } 237 238 while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) { 239 vma = container_of(node, struct i915_vma, node); 240 241 /* If we find any non-objects (!vma), we cannot evict them */ 242 if (vma->node.color != I915_COLOR_UNEVICTABLE) 243 ret = __i915_vma_unbind(vma); 244 else 245 ret = -ENOSPC; /* XXX search failed, try again? */ 246 } 247 248 return ret; 249 } 250 251 /** 252 * i915_gem_evict_for_node - 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 int ret = 0; 272 273 lockdep_assert_held(&vm->mutex); 274 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE)); 275 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE)); 276 277 trace_i915_gem_evict_node(vm, target, flags); 278 279 /* 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 intel_gt_retire_requests(vm->gt); 286 287 if (i915_vm_has_cache_coloring(vm)) { 288 /* Expand search to cover neighbouring guard pages (or lack!) */ 289 if (start) 290 start -= I915_GTT_PAGE_SIZE; 291 292 /* Always look at the page afterwards to avoid the end-of-GTT */ 293 end += I915_GTT_PAGE_SIZE; 294 } 295 GEM_BUG_ON(start >= end); 296 297 drm_mm_for_each_node_in_range(node, &vm->mm, start, end) { 298 /* If we find any non-objects (!vma), we cannot evict them */ 299 if (node->color == I915_COLOR_UNEVICTABLE) { 300 ret = -ENOSPC; 301 break; 302 } 303 304 GEM_BUG_ON(!drm_mm_node_allocated(node)); 305 vma = container_of(node, typeof(*vma), node); 306 307 /* 308 * If we are using coloring to insert guard pages between 309 * different cache domains within the address space, we have 310 * to check whether the objects on either side of our range 311 * abutt and conflict. If they are in conflict, then we evict 312 * those as well to make room for our guard pages. 313 */ 314 if (i915_vm_has_cache_coloring(vm)) { 315 if (node->start + node->size == target->start) { 316 if (node->color == target->color) 317 continue; 318 } 319 if (node->start == target->start + target->size) { 320 if (node->color == target->color) 321 continue; 322 } 323 } 324 325 if (i915_vma_is_pinned(vma)) { 326 ret = -ENOSPC; 327 break; 328 } 329 330 if (flags & PIN_NONBLOCK && i915_vma_is_active(vma)) { 331 ret = -ENOSPC; 332 break; 333 } 334 335 /* 336 * Never show fear in the face of dragons! 337 * 338 * We cannot directly remove this node from within this 339 * iterator and as with i915_gem_evict_something() we employ 340 * the vma pin_count in order to prevent the action of 341 * unbinding one vma from freeing (by dropping its active 342 * reference) another in our eviction list. 343 */ 344 __i915_vma_pin(vma); 345 list_add(&vma->evict_link, &eviction_list); 346 } 347 348 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 349 __i915_vma_unpin(vma); 350 if (ret == 0) 351 ret = __i915_vma_unbind(vma); 352 } 353 354 return ret; 355 } 356 357 /** 358 * i915_gem_evict_vm - Evict all idle vmas from a vm 359 * @vm: Address space to cleanse 360 * 361 * This function evicts all vmas from a vm. 362 * 363 * This is used by the execbuf code as a last-ditch effort to defragment the 364 * address space. 365 * 366 * To clarify: This is for freeing up virtual address space, not for freeing 367 * memory in e.g. the shrinker. 368 */ 369 int i915_gem_evict_vm(struct i915_address_space *vm) 370 { 371 int ret = 0; 372 373 lockdep_assert_held(&vm->mutex); 374 trace_i915_gem_evict_vm(vm); 375 376 /* Switch back to the default context in order to unpin 377 * the existing context objects. However, such objects only 378 * pin themselves inside the global GTT and performing the 379 * switch otherwise is ineffective. 380 */ 381 if (i915_is_ggtt(vm)) { 382 ret = ggtt_flush(vm->gt); 383 if (ret) 384 return ret; 385 } 386 387 do { 388 struct i915_vma *vma, *vn; 389 LIST_HEAD(eviction_list); 390 391 list_for_each_entry(vma, &vm->bound_list, vm_link) { 392 if (i915_vma_is_pinned(vma)) 393 continue; 394 395 __i915_vma_pin(vma); 396 list_add(&vma->evict_link, &eviction_list); 397 } 398 if (list_empty(&eviction_list)) 399 break; 400 401 ret = 0; 402 list_for_each_entry_safe(vma, vn, &eviction_list, evict_link) { 403 __i915_vma_unpin(vma); 404 if (ret == 0) 405 ret = __i915_vma_unbind(vma); 406 if (ret != -EINTR) /* "Get me out of here!" */ 407 ret = 0; 408 } 409 } while (ret == 0); 410 411 return ret; 412 } 413 414 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 415 #include "selftests/i915_gem_evict.c" 416 #endif 417