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 /** 65 * i915_gem_evict_something - Evict vmas to make room for binding a new one 66 * @vm: address space to evict from 67 * @min_size: size of the desired free space 68 * @alignment: alignment constraint of the desired free space 69 * @color: color for the desired space 70 * @start: start (inclusive) of the range from which to evict objects 71 * @end: end (exclusive) of the range from which to evict objects 72 * @flags: additional flags to control the eviction algorithm 73 * 74 * This function will try to evict vmas until a free space satisfying the 75 * requirements is found. Callers must check first whether any such hole exists 76 * already before calling this function. 77 * 78 * This function is used by the object/vma binding code. 79 * 80 * Since this function is only used to free up virtual address space it only 81 * ignores pinned vmas, and not object where the backing storage itself is 82 * pinned. Hence obj->pages_pin_count does not protect against eviction. 83 * 84 * To clarify: This is for freeing up virtual address space, not for freeing 85 * memory in e.g. the shrinker. 86 */ 87 int 88 i915_gem_evict_something(struct i915_address_space *vm, 89 u64 min_size, u64 alignment, 90 unsigned long color, 91 u64 start, u64 end, 92 unsigned flags) 93 { 94 struct drm_mm_scan scan; 95 struct list_head eviction_list; 96 struct i915_vma *vma, *next; 97 struct drm_mm_node *node; 98 enum drm_mm_insert_mode mode; 99 struct i915_vma *active; 100 int ret; 101 102 lockdep_assert_held(&vm->mutex); 103 trace_i915_gem_evict(vm, min_size, alignment, flags); 104 105 /* 106 * The goal is to evict objects and amalgamate space in rough LRU order. 107 * Since both active and inactive objects reside on the same list, 108 * in a mix of creation and last scanned order, as we process the list 109 * we sort it into inactive/active, which keeps the active portion 110 * in a rough MRU order. 111 * 112 * The retirement sequence is thus: 113 * 1. Inactive objects (already retired, random order) 114 * 2. Active objects (will stall on unbinding, oldest scanned first) 115 */ 116 mode = DRM_MM_INSERT_BEST; 117 if (flags & PIN_HIGH) 118 mode = DRM_MM_INSERT_HIGH; 119 if (flags & PIN_MAPPABLE) 120 mode = DRM_MM_INSERT_LOW; 121 drm_mm_scan_init_with_range(&scan, &vm->mm, 122 min_size, alignment, color, 123 start, end, mode); 124 125 intel_gt_retire_requests(vm->gt); 126 127 search_again: 128 active = NULL; 129 INIT_LIST_HEAD(&eviction_list); 130 list_for_each_entry_safe(vma, next, &vm->bound_list, vm_link) { 131 if (vma == active) { /* now seen this vma twice */ 132 if (flags & PIN_NONBLOCK) 133 break; 134 135 active = ERR_PTR(-EAGAIN); 136 } 137 138 /* 139 * We keep this list in a rough least-recently scanned order 140 * of active elements (inactive elements are cheap to reap). 141 * New entries are added to the end, and we move anything we 142 * scan to the end. The assumption is that the working set 143 * of applications is either steady state (and thanks to the 144 * userspace bo cache it almost always is) or volatile and 145 * frequently replaced after a frame, which are self-evicting! 146 * Given that assumption, the MRU order of the scan list is 147 * fairly static, and keeping it in least-recently scan order 148 * is suitable. 149 * 150 * To notice when we complete one full cycle, we record the 151 * first active element seen, before moving it to the tail. 152 */ 153 if (active != ERR_PTR(-EAGAIN) && i915_vma_is_active(vma)) { 154 if (!active) 155 active = vma; 156 157 list_move_tail(&vma->vm_link, &vm->bound_list); 158 continue; 159 } 160 161 if (mark_free(&scan, vma, flags, &eviction_list)) 162 goto found; 163 } 164 165 /* Nothing found, clean up and bail out! */ 166 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 167 ret = drm_mm_scan_remove_block(&scan, &vma->node); 168 BUG_ON(ret); 169 } 170 171 /* 172 * Can we unpin some objects such as idle hw contents, 173 * or pending flips? But since only the GGTT has global entries 174 * such as scanouts, rinbuffers and contexts, we can skip the 175 * purge when inspecting per-process local address spaces. 176 */ 177 if (!i915_is_ggtt(vm) || flags & PIN_NONBLOCK) 178 return -ENOSPC; 179 180 /* 181 * Not everything in the GGTT is tracked via VMA using 182 * i915_vma_move_to_active(), otherwise we could evict as required 183 * with minimal stalling. Instead we are forced to idle the GPU and 184 * explicitly retire outstanding requests which will then remove 185 * the pinning for active objects such as contexts and ring, 186 * enabling us to evict them on the next iteration. 187 * 188 * To ensure that all user contexts are evictable, we perform 189 * a switch to the perma-pinned kernel context. This all also gives 190 * us a termination condition, when the last retired context is 191 * the kernel's there is no more we can evict. 192 */ 193 if (I915_SELFTEST_ONLY(igt_evict_ctl.fail_if_busy)) 194 return -EBUSY; 195 196 ret = ggtt_flush(vm->gt); 197 if (ret) 198 return ret; 199 200 cond_resched(); 201 202 flags |= PIN_NONBLOCK; 203 goto search_again; 204 205 found: 206 /* drm_mm doesn't allow any other other operations while 207 * scanning, therefore store to-be-evicted objects on a 208 * temporary list and take a reference for all before 209 * calling unbind (which may remove the active reference 210 * of any of our objects, thus corrupting the list). 211 */ 212 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 213 if (drm_mm_scan_remove_block(&scan, &vma->node)) 214 __i915_vma_pin(vma); 215 else 216 list_del(&vma->evict_link); 217 } 218 219 /* Unbinding will emit any required flushes */ 220 ret = 0; 221 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 222 __i915_vma_unpin(vma); 223 if (ret == 0) 224 ret = __i915_vma_unbind(vma); 225 } 226 227 while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) { 228 vma = container_of(node, struct i915_vma, node); 229 230 /* If we find any non-objects (!vma), we cannot evict them */ 231 if (vma->node.color != I915_COLOR_UNEVICTABLE) 232 ret = __i915_vma_unbind(vma); 233 else 234 ret = -ENOSPC; /* XXX search failed, try again? */ 235 } 236 237 return ret; 238 } 239 240 /** 241 * i915_gem_evict_for_node - Evict vmas to make room for binding a new one 242 * @vm: address space to evict from 243 * @target: range (and color) to evict for 244 * @flags: additional flags to control the eviction algorithm 245 * 246 * This function will try to evict vmas that overlap the target node. 247 * 248 * To clarify: This is for freeing up virtual address space, not for freeing 249 * memory in e.g. the shrinker. 250 */ 251 int i915_gem_evict_for_node(struct i915_address_space *vm, 252 struct drm_mm_node *target, 253 unsigned int flags) 254 { 255 LIST_HEAD(eviction_list); 256 struct drm_mm_node *node; 257 u64 start = target->start; 258 u64 end = start + target->size; 259 struct i915_vma *vma, *next; 260 int ret = 0; 261 262 lockdep_assert_held(&vm->mutex); 263 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE)); 264 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE)); 265 266 trace_i915_gem_evict_node(vm, target, flags); 267 268 /* 269 * Retire before we search the active list. Although we have 270 * reasonable accuracy in our retirement lists, we may have 271 * a stray pin (preventing eviction) that can only be resolved by 272 * retiring. 273 */ 274 intel_gt_retire_requests(vm->gt); 275 276 if (i915_vm_has_cache_coloring(vm)) { 277 /* Expand search to cover neighbouring guard pages (or lack!) */ 278 if (start) 279 start -= I915_GTT_PAGE_SIZE; 280 281 /* Always look at the page afterwards to avoid the end-of-GTT */ 282 end += I915_GTT_PAGE_SIZE; 283 } 284 GEM_BUG_ON(start >= end); 285 286 drm_mm_for_each_node_in_range(node, &vm->mm, start, end) { 287 /* If we find any non-objects (!vma), we cannot evict them */ 288 if (node->color == I915_COLOR_UNEVICTABLE) { 289 ret = -ENOSPC; 290 break; 291 } 292 293 GEM_BUG_ON(!drm_mm_node_allocated(node)); 294 vma = container_of(node, typeof(*vma), node); 295 296 /* 297 * If we are using coloring to insert guard pages between 298 * different cache domains within the address space, we have 299 * to check whether the objects on either side of our range 300 * abutt and conflict. If they are in conflict, then we evict 301 * those as well to make room for our guard pages. 302 */ 303 if (i915_vm_has_cache_coloring(vm)) { 304 if (node->start + node->size == target->start) { 305 if (node->color == target->color) 306 continue; 307 } 308 if (node->start == target->start + target->size) { 309 if (node->color == target->color) 310 continue; 311 } 312 } 313 314 if (i915_vma_is_pinned(vma)) { 315 ret = -ENOSPC; 316 break; 317 } 318 319 if (flags & PIN_NONBLOCK && i915_vma_is_active(vma)) { 320 ret = -ENOSPC; 321 break; 322 } 323 324 /* 325 * Never show fear in the face of dragons! 326 * 327 * We cannot directly remove this node from within this 328 * iterator and as with i915_gem_evict_something() we employ 329 * the vma pin_count in order to prevent the action of 330 * unbinding one vma from freeing (by dropping its active 331 * reference) another in our eviction list. 332 */ 333 __i915_vma_pin(vma); 334 list_add(&vma->evict_link, &eviction_list); 335 } 336 337 list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { 338 __i915_vma_unpin(vma); 339 if (ret == 0) 340 ret = __i915_vma_unbind(vma); 341 } 342 343 return ret; 344 } 345 346 /** 347 * i915_gem_evict_vm - Evict all idle vmas from a vm 348 * @vm: Address space to cleanse 349 * 350 * This function evicts all vmas from a vm. 351 * 352 * This is used by the execbuf code as a last-ditch effort to defragment the 353 * address space. 354 * 355 * To clarify: This is for freeing up virtual address space, not for freeing 356 * memory in e.g. the shrinker. 357 */ 358 int i915_gem_evict_vm(struct i915_address_space *vm) 359 { 360 int ret = 0; 361 362 lockdep_assert_held(&vm->mutex); 363 trace_i915_gem_evict_vm(vm); 364 365 /* Switch back to the default context in order to unpin 366 * the existing context objects. However, such objects only 367 * pin themselves inside the global GTT and performing the 368 * switch otherwise is ineffective. 369 */ 370 if (i915_is_ggtt(vm)) { 371 ret = ggtt_flush(vm->gt); 372 if (ret) 373 return ret; 374 } 375 376 do { 377 struct i915_vma *vma, *vn; 378 LIST_HEAD(eviction_list); 379 380 list_for_each_entry(vma, &vm->bound_list, vm_link) { 381 if (i915_vma_is_pinned(vma)) 382 continue; 383 384 __i915_vma_pin(vma); 385 list_add(&vma->evict_link, &eviction_list); 386 } 387 if (list_empty(&eviction_list)) 388 break; 389 390 ret = 0; 391 list_for_each_entry_safe(vma, vn, &eviction_list, evict_link) { 392 __i915_vma_unpin(vma); 393 if (ret == 0) 394 ret = __i915_vma_unbind(vma); 395 if (ret != -EINTR) /* "Get me out of here!" */ 396 ret = 0; 397 } 398 } while (ret == 0); 399 400 return ret; 401 } 402 403 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 404 #include "selftests/i915_gem_evict.c" 405 #endif 406