drivers/gpu/drm/i915/gem/i915_gem_shrinker.c

   1 /*
   2  * SPDX-License-Identifier: MIT
   3  *
   4  * Copyright © 2008-2015 Intel Corporation
   5  */
   6
   7 #include <linux/oom.h>
   8 #include <linux/sched/mm.h>
   9 #include <linux/shmem_fs.h>
  10 #include <linux/slab.h>
  11 #include <linux/swap.h>
  12 #include <linux/pci.h>
  13 #include <linux/dma-buf.h>
  14 #include <linux/vmalloc.h>
  15
  16 #include "gt/intel_gt_requests.h"
  17
  18 #include "dma_resv_utils.h"
  19 #include "i915_trace.h"
  20
  21 static bool swap_available(void)
  22 {
  23         return get_nr_swap_pages() > 0;
  24 }
  25
  26 static bool can_release_pages(struct drm_i915_gem_object *obj)
  27 {
  28         /* Consider only shrinkable ojects. */
  29         if (!i915_gem_object_is_shrinkable(obj))
  30                 return false;
  31
  32         /*
  33          * We can only return physical pages to the system if we can either
  34          * discard the contents (because the user has marked them as being
  35          * purgeable) or if we can move their contents out to swap.
  36          */
  37         return swap_available() || obj->mm.madv == I915_MADV_DONTNEED;
  38 }
  39
  40 static bool unsafe_drop_pages(struct drm_i915_gem_object *obj,
  41                               unsigned long shrink)
  42 {
  43         unsigned long flags;
  44
  45         flags = 0;
  46         if (shrink & I915_SHRINK_ACTIVE)
  47                 flags = I915_GEM_OBJECT_UNBIND_ACTIVE;
  48         if (!(shrink & I915_SHRINK_BOUND))
  49                 flags = I915_GEM_OBJECT_UNBIND_TEST;
  50
  51         if (i915_gem_object_unbind(obj, flags) == 0)
  52                 return true;
  53
  54         return false;
  55 }
  56
  57 static void try_to_writeback(struct drm_i915_gem_object *obj,
  58                              unsigned int flags)
  59 {
  60         switch (obj->mm.madv) {
  61         case I915_MADV_DONTNEED:
  62                 i915_gem_object_truncate(obj);
  63         case __I915_MADV_PURGED:
  64                 return;
  65         }
  66
  67         if (flags & I915_SHRINK_WRITEBACK)
  68                 i915_gem_object_writeback(obj);
  69 }
  70
  71 /**
  72  * i915_gem_shrink - Shrink buffer object caches
  73  * @ww: i915 gem ww acquire ctx, or NULL
  74  * @i915: i915 device
  75  * @target: amount of memory to make available, in pages
  76  * @nr_scanned: optional output for number of pages scanned (incremental)
  77  * @shrink: control flags for selecting cache types
  78  *
  79  * This function is the main interface to the shrinker. It will try to release
  80  * up to @target pages of main memory backing storage from buffer objects.
  81  * Selection of the specific caches can be done with @flags. This is e.g. useful
  82  * when purgeable objects should be removed from caches preferentially.
  83  *
  84  * Note that it's not guaranteed that released amount is actually available as
  85  * free system memory - the pages might still be in-used to due to other reasons
  86  * (like cpu mmaps) or the mm core has reused them before we could grab them.
  87  * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
  88  * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
  89  *
  90  * Also note that any kind of pinning (both per-vma address space pins and
  91  * backing storage pins at the buffer object level) result in the shrinker code
  92  * having to skip the object.
  93  *
  94  * Returns:
  95  * The number of pages of backing storage actually released.
  96  */
  97 unsigned long
  98 i915_gem_shrink(struct i915_gem_ww_ctx *ww,
  99                 struct drm_i915_private *i915,
 100                 unsigned long target,
 101                 unsigned long *nr_scanned,
 102                 unsigned int shrink)
 103 {
 104         const struct {
 105                 struct list_head *list;
 106                 unsigned int bit;
 107         } phases[] = {
 108                 { &i915->mm.purge_list, ~0u },
 109                 {
 110                         &i915->mm.shrink_list,
 111                         I915_SHRINK_BOUND | I915_SHRINK_UNBOUND
 112                 },
 113                 { NULL, 0 },
 114         }, *phase;
 115         intel_wakeref_t wakeref = 0;
 116         unsigned long count = 0;
 117         unsigned long scanned = 0;
 118         int err;
 119
 120         trace_i915_gem_shrink(i915, target, shrink);
 121
 122         /*
 123          * Unbinding of objects will require HW access; Let us not wake the
 124          * device just to recover a little memory. If absolutely necessary,
 125          * we will force the wake during oom-notifier.
 126          */
 127         if (shrink & I915_SHRINK_BOUND) {
 128                 wakeref = intel_runtime_pm_get_if_in_use(&i915->runtime_pm);
 129                 if (!wakeref)
 130                         shrink &= ~I915_SHRINK_BOUND;
 131         }
 132
 133         /*
 134          * When shrinking the active list, we should also consider active
 135          * contexts. Active contexts are pinned until they are retired, and
 136          * so can not be simply unbound to retire and unpin their pages. To
 137          * shrink the contexts, we must wait until the gpu is idle and
 138          * completed its switch to the kernel context. In short, we do
 139          * not have a good mechanism for idling a specific context, but
 140          * what we can do is give them a kick so that we do not keep idle
 141          * contexts around longer than is necessary.
 142          */
 143         if (shrink & I915_SHRINK_ACTIVE)
 144                 /* Retire requests to unpin all idle contexts */
 145                 intel_gt_retire_requests(&i915->gt);
 146
 147         /*
 148          * As we may completely rewrite the (un)bound list whilst unbinding
 149          * (due to retiring requests) we have to strictly process only
 150          * one element of the list at the time, and recheck the list
 151          * on every iteration.
 152          *
 153          * In particular, we must hold a reference whilst removing the
 154          * object as we may end up waiting for and/or retiring the objects.
 155          * This might release the final reference (held by the active list)
 156          * and result in the object being freed from under us. This is
 157          * similar to the precautions the eviction code must take whilst
 158          * removing objects.
 159          *
 160          * Also note that although these lists do not hold a reference to
 161          * the object we can safely grab one here: The final object
 162          * unreferencing and the bound_list are both protected by the
 163          * dev->struct_mutex and so we won't ever be able to observe an
 164          * object on the bound_list with a reference count equals 0.
 165          */
 166         for (phase = phases; phase->list; phase++) {
 167                 struct list_head still_in_list;
 168                 struct drm_i915_gem_object *obj;
 169                 unsigned long flags;
 170
 171                 if ((shrink & phase->bit) == 0)
 172                         continue;
 173
 174                 INIT_LIST_HEAD(&still_in_list);
 175
 176                 /*
 177                  * We serialize our access to unreferenced objects through
 178                  * the use of the struct_mutex. While the objects are not
 179                  * yet freed (due to RCU then a workqueue) we still want
 180                  * to be able to shrink their pages, so they remain on
 181                  * the unbound/bound list until actually freed.
 182                  */
 183                 spin_lock_irqsave(&i915->mm.obj_lock, flags);
 184                 while (count < target &&
 185                        (obj = list_first_entry_or_null(phase->list,
 186                                                        typeof(*obj),
 187                                                        mm.link))) {
 188                         list_move_tail(&obj->mm.link, &still_in_list);
 189
 190                         if (shrink & I915_SHRINK_VMAPS &&
 191                             !is_vmalloc_addr(obj->mm.mapping))
 192                                 continue;
 193
 194                         if (!(shrink & I915_SHRINK_ACTIVE) &&
 195                             i915_gem_object_is_framebuffer(obj))
 196                                 continue;
 197
 198                         if (!can_release_pages(obj))
 199                                 continue;
 200
 201                         if (!kref_get_unless_zero(&obj->base.refcount))
 202                                 continue;
 203
 204                         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 205
 206                         err = 0;
 207                         if (unsafe_drop_pages(obj, shrink)) {
 208                                 /* May arrive from get_pages on another bo */
 209                                 if (!ww) {
 210                                         if (!i915_gem_object_trylock(obj))
 211                                                 goto skip;
 212                                 } else {
 213                                         err = i915_gem_object_lock(obj, ww);
 214                                         if (err)
 215                                                 goto skip;
 216                                 }
 217
 218                                 if (!__i915_gem_object_put_pages(obj)) {
 219                                         try_to_writeback(obj, shrink);
 220                                         count += obj->base.size >> PAGE_SHIFT;
 221                                 }
 222                                 if (!ww)
 223                                         i915_gem_object_unlock(obj);
 224                         }
 225
 226                         dma_resv_prune(obj->base.resv);
 227
 228                         scanned += obj->base.size >> PAGE_SHIFT;
 229 skip:
 230                         i915_gem_object_put(obj);
 231
 232                         spin_lock_irqsave(&i915->mm.obj_lock, flags);
 233                         if (err)
 234                                 break;
 235                 }
 236                 list_splice_tail(&still_in_list, phase->list);
 237                 spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 238                 if (err)
 239                         return err;
 240         }
 241
 242         if (shrink & I915_SHRINK_BOUND)
 243                 intel_runtime_pm_put(&i915->runtime_pm, wakeref);
 244
 245         if (nr_scanned)
 246                 *nr_scanned += scanned;
 247         return count;
 248 }
 249
 250 /**
 251  * i915_gem_shrink_all - Shrink buffer object caches completely
 252  * @i915: i915 device
 253  *
 254  * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
 255  * caches completely. It also first waits for and retires all outstanding
 256  * requests to also be able to release backing storage for active objects.
 257  *
 258  * This should only be used in code to intentionally quiescent the gpu or as a
 259  * last-ditch effort when memory seems to have run out.
 260  *
 261  * Returns:
 262  * The number of pages of backing storage actually released.
 263  */
 264 unsigned long i915_gem_shrink_all(struct drm_i915_private *i915)
 265 {
 266         intel_wakeref_t wakeref;
 267         unsigned long freed = 0;
 268
 269         with_intel_runtime_pm(&i915->runtime_pm, wakeref) {
 270                 freed = i915_gem_shrink(NULL, i915, -1UL, NULL,
 271                                         I915_SHRINK_BOUND |
 272                                         I915_SHRINK_UNBOUND);
 273         }
 274
 275         return freed;
 276 }
 277
 278 static unsigned long
 279 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
 280 {
 281         struct drm_i915_private *i915 =
 282                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 283         unsigned long num_objects;
 284         unsigned long count;
 285
 286         count = READ_ONCE(i915->mm.shrink_memory) >> PAGE_SHIFT;
 287         num_objects = READ_ONCE(i915->mm.shrink_count);
 288
 289         /*
 290          * Update our preferred vmscan batch size for the next pass.
 291          * Our rough guess for an effective batch size is roughly 2
 292          * available GEM objects worth of pages. That is we don't want
 293          * the shrinker to fire, until it is worth the cost of freeing an
 294          * entire GEM object.
 295          */
 296         if (num_objects) {
 297                 unsigned long avg = 2 * count / num_objects;
 298
 299                 i915->mm.shrinker.batch =
 300                         max((i915->mm.shrinker.batch + avg) >> 1,
 301                             128ul /* default SHRINK_BATCH */);
 302         }
 303
 304         return count;
 305 }
 306
 307 static unsigned long
 308 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 309 {
 310         struct drm_i915_private *i915 =
 311                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 312         unsigned long freed;
 313
 314         sc->nr_scanned = 0;
 315
 316         freed = i915_gem_shrink(NULL, i915,
 317                                 sc->nr_to_scan,
 318                                 &sc->nr_scanned,
 319                                 I915_SHRINK_BOUND |
 320                                 I915_SHRINK_UNBOUND);
 321         if (sc->nr_scanned < sc->nr_to_scan && current_is_kswapd()) {
 322                 intel_wakeref_t wakeref;
 323
 324                 with_intel_runtime_pm(&i915->runtime_pm, wakeref) {
 325                         freed += i915_gem_shrink(NULL, i915,
 326                                                  sc->nr_to_scan - sc->nr_scanned,
 327                                                  &sc->nr_scanned,
 328                                                  I915_SHRINK_ACTIVE |
 329                                                  I915_SHRINK_BOUND |
 330                                                  I915_SHRINK_UNBOUND |
 331                                                  I915_SHRINK_WRITEBACK);
 332                 }
 333         }
 334
 335         return sc->nr_scanned ? freed : SHRINK_STOP;
 336 }
 337
 338 static int
 339 i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
 340 {
 341         struct drm_i915_private *i915 =
 342                 container_of(nb, struct drm_i915_private, mm.oom_notifier);
 343         struct drm_i915_gem_object *obj;
 344         unsigned long unevictable, available, freed_pages;
 345         intel_wakeref_t wakeref;
 346         unsigned long flags;
 347
 348         freed_pages = 0;
 349         with_intel_runtime_pm(&i915->runtime_pm, wakeref)
 350                 freed_pages += i915_gem_shrink(NULL, i915, -1UL, NULL,
 351                                                I915_SHRINK_BOUND |
 352                                                I915_SHRINK_UNBOUND |
 353                                                I915_SHRINK_WRITEBACK);
 354
 355         /* Because we may be allocating inside our own driver, we cannot
 356          * assert that there are no objects with pinned pages that are not
 357          * being pointed to by hardware.
 358          */
 359         available = unevictable = 0;
 360         spin_lock_irqsave(&i915->mm.obj_lock, flags);
 361         list_for_each_entry(obj, &i915->mm.shrink_list, mm.link) {
 362                 if (!can_release_pages(obj))
 363                         unevictable += obj->base.size >> PAGE_SHIFT;
 364                 else
 365                         available += obj->base.size >> PAGE_SHIFT;
 366         }
 367         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 368
 369         if (freed_pages || available)
 370                 pr_info("Purging GPU memory, %lu pages freed, "
 371                         "%lu pages still pinned, %lu pages left available.\n",
 372                         freed_pages, unevictable, available);
 373
 374         *(unsigned long *)ptr += freed_pages;
 375         return NOTIFY_DONE;
 376 }
 377
 378 static int
 379 i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
 380 {
 381         struct drm_i915_private *i915 =
 382                 container_of(nb, struct drm_i915_private, mm.vmap_notifier);
 383         struct i915_vma *vma, *next;
 384         unsigned long freed_pages = 0;
 385         intel_wakeref_t wakeref;
 386
 387         with_intel_runtime_pm(&i915->runtime_pm, wakeref)
 388                 freed_pages += i915_gem_shrink(NULL, i915, -1UL, NULL,
 389                                                I915_SHRINK_BOUND |
 390                                                I915_SHRINK_UNBOUND |
 391                                                I915_SHRINK_VMAPS);
 392
 393         /* We also want to clear any cached iomaps as they wrap vmap */
 394         mutex_lock(&i915->ggtt.vm.mutex);
 395         list_for_each_entry_safe(vma, next,
 396                                  &i915->ggtt.vm.bound_list, vm_link) {
 397                 unsigned long count = vma->node.size >> PAGE_SHIFT;
 398
 399                 if (!vma->iomap || i915_vma_is_active(vma))
 400                         continue;
 401
 402                 if (__i915_vma_unbind(vma) == 0)
 403                         freed_pages += count;
 404         }
 405         mutex_unlock(&i915->ggtt.vm.mutex);
 406
 407         *(unsigned long *)ptr += freed_pages;
 408         return NOTIFY_DONE;
 409 }
 410
 411 void i915_gem_driver_register__shrinker(struct drm_i915_private *i915)
 412 {
 413         i915->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
 414         i915->mm.shrinker.count_objects = i915_gem_shrinker_count;
 415         i915->mm.shrinker.seeks = DEFAULT_SEEKS;
 416         i915->mm.shrinker.batch = 4096;
 417         drm_WARN_ON(&i915->drm, register_shrinker(&i915->mm.shrinker));
 418
 419         i915->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
 420         drm_WARN_ON(&i915->drm, register_oom_notifier(&i915->mm.oom_notifier));
 421
 422         i915->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
 423         drm_WARN_ON(&i915->drm,
 424                     register_vmap_purge_notifier(&i915->mm.vmap_notifier));
 425 }
 426
 427 void i915_gem_driver_unregister__shrinker(struct drm_i915_private *i915)
 428 {
 429         drm_WARN_ON(&i915->drm,
 430                     unregister_vmap_purge_notifier(&i915->mm.vmap_notifier));
 431         drm_WARN_ON(&i915->drm,
 432                     unregister_oom_notifier(&i915->mm.oom_notifier));
 433         unregister_shrinker(&i915->mm.shrinker);
 434 }
 435
 436 void i915_gem_shrinker_taints_mutex(struct drm_i915_private *i915,
 437                                     struct mutex *mutex)
 438 {
 439         if (!IS_ENABLED(CONFIG_LOCKDEP))
 440                 return;
 441
 442         fs_reclaim_acquire(GFP_KERNEL);
 443
 444         mutex_acquire(&mutex->dep_map, 0, 0, _RET_IP_);
 445         mutex_release(&mutex->dep_map, _RET_IP_);
 446
 447         fs_reclaim_release(GFP_KERNEL);
 448 }
 449
 450 #define obj_to_i915(obj__) to_i915((obj__)->base.dev)
 451
 452 void i915_gem_object_make_unshrinkable(struct drm_i915_gem_object *obj)
 453 {
 454         struct drm_i915_private *i915 = obj_to_i915(obj);
 455         unsigned long flags;
 456
 457         /*
 458          * We can only be called while the pages are pinned or when
 459          * the pages are released. If pinned, we should only be called
 460          * from a single caller under controlled conditions; and on release
 461          * only one caller may release us. Neither the two may cross.
 462          */
 463         if (atomic_add_unless(&obj->mm.shrink_pin, 1, 0))
 464                 return;
 465
 466         spin_lock_irqsave(&i915->mm.obj_lock, flags);
 467         if (!atomic_fetch_inc(&obj->mm.shrink_pin) &&
 468             !list_empty(&obj->mm.link)) {
 469                 list_del_init(&obj->mm.link);
 470                 i915->mm.shrink_count--;
 471                 i915->mm.shrink_memory -= obj->base.size;
 472         }
 473         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 474 }
 475
 476 static void __i915_gem_object_make_shrinkable(struct drm_i915_gem_object *obj,
 477                                               struct list_head *head)
 478 {
 479         struct drm_i915_private *i915 = obj_to_i915(obj);
 480         unsigned long flags;
 481
 482         GEM_BUG_ON(!i915_gem_object_has_pages(obj));
 483         if (!i915_gem_object_is_shrinkable(obj))
 484                 return;
 485
 486         if (atomic_add_unless(&obj->mm.shrink_pin, -1, 1))
 487                 return;
 488
 489         spin_lock_irqsave(&i915->mm.obj_lock, flags);
 490         GEM_BUG_ON(!kref_read(&obj->base.refcount));
 491         if (atomic_dec_and_test(&obj->mm.shrink_pin)) {
 492                 GEM_BUG_ON(!list_empty(&obj->mm.link));
 493
 494                 list_add_tail(&obj->mm.link, head);
 495                 i915->mm.shrink_count++;
 496                 i915->mm.shrink_memory += obj->base.size;
 497
 498         }
 499         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 500 }
 501
 502 void i915_gem_object_make_shrinkable(struct drm_i915_gem_object *obj)
 503 {
 504         __i915_gem_object_make_shrinkable(obj,
 505                                           &obj_to_i915(obj)->mm.shrink_list);
 506 }
 507
 508 void i915_gem_object_make_purgeable(struct drm_i915_gem_object *obj)
 509 {
 510         __i915_gem_object_make_shrinkable(obj,
 511                                           &obj_to_i915(obj)->mm.purge_list);
 512 }