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, bool trylock_vm)
  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         if (trylock_vm)
  51                 flags |= I915_GEM_OBJECT_UNBIND_VM_TRYLOCK;
  52
  53         if (i915_gem_object_unbind(obj, flags) == 0)
  54                 return true;
  55
  56         return false;
  57 }
  58
  59 static void try_to_writeback(struct drm_i915_gem_object *obj,
  60                              unsigned int flags)
  61 {
  62         switch (obj->mm.madv) {
  63         case I915_MADV_DONTNEED:
  64                 i915_gem_object_truncate(obj);
  65                 return;
  66         case __I915_MADV_PURGED:
  67                 return;
  68         }
  69
  70         if (flags & I915_SHRINK_WRITEBACK)
  71                 i915_gem_object_writeback(obj);
  72 }
  73
  74 /**
  75  * i915_gem_shrink - Shrink buffer object caches
  76  * @ww: i915 gem ww acquire ctx, or NULL
  77  * @i915: i915 device
  78  * @target: amount of memory to make available, in pages
  79  * @nr_scanned: optional output for number of pages scanned (incremental)
  80  * @shrink: control flags for selecting cache types
  81  *
  82  * This function is the main interface to the shrinker. It will try to release
  83  * up to @target pages of main memory backing storage from buffer objects.
  84  * Selection of the specific caches can be done with @flags. This is e.g. useful
  85  * when purgeable objects should be removed from caches preferentially.
  86  *
  87  * Note that it's not guaranteed that released amount is actually available as
  88  * free system memory - the pages might still be in-used to due to other reasons
  89  * (like cpu mmaps) or the mm core has reused them before we could grab them.
  90  * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
  91  * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
  92  *
  93  * Also note that any kind of pinning (both per-vma address space pins and
  94  * backing storage pins at the buffer object level) result in the shrinker code
  95  * having to skip the object.
  96  *
  97  * Returns:
  98  * The number of pages of backing storage actually released.
  99  */
 100 unsigned long
 101 i915_gem_shrink(struct i915_gem_ww_ctx *ww,
 102                 struct drm_i915_private *i915,
 103                 unsigned long target,
 104                 unsigned long *nr_scanned,
 105                 unsigned int shrink)
 106 {
 107         const struct {
 108                 struct list_head *list;
 109                 unsigned int bit;
 110         } phases[] = {
 111                 { &i915->mm.purge_list, ~0u },
 112                 {
 113                         &i915->mm.shrink_list,
 114                         I915_SHRINK_BOUND | I915_SHRINK_UNBOUND
 115                 },
 116                 { NULL, 0 },
 117         }, *phase;
 118         intel_wakeref_t wakeref = 0;
 119         unsigned long count = 0;
 120         unsigned long scanned = 0;
 121         int err;
 122
 123         /* CHV + VTD workaround use stop_machine(); need to trylock vm->mutex */
 124         bool trylock_vm = !ww && intel_vm_no_concurrent_access_wa(i915);
 125
 126         trace_i915_gem_shrink(i915, target, shrink);
 127
 128         /*
 129          * Unbinding of objects will require HW access; Let us not wake the
 130          * device just to recover a little memory. If absolutely necessary,
 131          * we will force the wake during oom-notifier.
 132          */
 133         if (shrink & I915_SHRINK_BOUND) {
 134                 wakeref = intel_runtime_pm_get_if_in_use(&i915->runtime_pm);
 135                 if (!wakeref)
 136                         shrink &= ~I915_SHRINK_BOUND;
 137         }
 138
 139         /*
 140          * When shrinking the active list, we should also consider active
 141          * contexts. Active contexts are pinned until they are retired, and
 142          * so can not be simply unbound to retire and unpin their pages. To
 143          * shrink the contexts, we must wait until the gpu is idle and
 144          * completed its switch to the kernel context. In short, we do
 145          * not have a good mechanism for idling a specific context, but
 146          * what we can do is give them a kick so that we do not keep idle
 147          * contexts around longer than is necessary.
 148          */
 149         if (shrink & I915_SHRINK_ACTIVE)
 150                 /* Retire requests to unpin all idle contexts */
 151                 intel_gt_retire_requests(&i915->gt);
 152
 153         /*
 154          * As we may completely rewrite the (un)bound list whilst unbinding
 155          * (due to retiring requests) we have to strictly process only
 156          * one element of the list at the time, and recheck the list
 157          * on every iteration.
 158          *
 159          * In particular, we must hold a reference whilst removing the
 160          * object as we may end up waiting for and/or retiring the objects.
 161          * This might release the final reference (held by the active list)
 162          * and result in the object being freed from under us. This is
 163          * similar to the precautions the eviction code must take whilst
 164          * removing objects.
 165          *
 166          * Also note that although these lists do not hold a reference to
 167          * the object we can safely grab one here: The final object
 168          * unreferencing and the bound_list are both protected by the
 169          * dev->struct_mutex and so we won't ever be able to observe an
 170          * object on the bound_list with a reference count equals 0.
 171          */
 172         for (phase = phases; phase->list; phase++) {
 173                 struct list_head still_in_list;
 174                 struct drm_i915_gem_object *obj;
 175                 unsigned long flags;
 176
 177                 if ((shrink & phase->bit) == 0)
 178                         continue;
 179
 180                 INIT_LIST_HEAD(&still_in_list);
 181
 182                 /*
 183                  * We serialize our access to unreferenced objects through
 184                  * the use of the struct_mutex. While the objects are not
 185                  * yet freed (due to RCU then a workqueue) we still want
 186                  * to be able to shrink their pages, so they remain on
 187                  * the unbound/bound list until actually freed.
 188                  */
 189                 spin_lock_irqsave(&i915->mm.obj_lock, flags);
 190                 while (count < target &&
 191                        (obj = list_first_entry_or_null(phase->list,
 192                                                        typeof(*obj),
 193                                                        mm.link))) {
 194                         list_move_tail(&obj->mm.link, &still_in_list);
 195
 196                         if (shrink & I915_SHRINK_VMAPS &&
 197                             !is_vmalloc_addr(obj->mm.mapping))
 198                                 continue;
 199
 200                         if (!(shrink & I915_SHRINK_ACTIVE) &&
 201                             i915_gem_object_is_framebuffer(obj))
 202                                 continue;
 203
 204                         if (!can_release_pages(obj))
 205                                 continue;
 206
 207                         if (!kref_get_unless_zero(&obj->base.refcount))
 208                                 continue;
 209
 210                         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 211
 212                         err = 0;
 213                         if (unsafe_drop_pages(obj, shrink, trylock_vm)) {
 214                                 /* May arrive from get_pages on another bo */
 215                                 if (!ww) {
 216                                         if (!i915_gem_object_trylock(obj))
 217                                                 goto skip;
 218                                 } else {
 219                                         err = i915_gem_object_lock(obj, ww);
 220                                         if (err)
 221                                                 goto skip;
 222                                 }
 223
 224                                 if (!__i915_gem_object_put_pages(obj)) {
 225                                         try_to_writeback(obj, shrink);
 226                                         count += obj->base.size >> PAGE_SHIFT;
 227                                 }
 228                                 if (!ww)
 229                                         i915_gem_object_unlock(obj);
 230                         }
 231
 232                         dma_resv_prune(obj->base.resv);
 233
 234                         scanned += obj->base.size >> PAGE_SHIFT;
 235 skip:
 236                         i915_gem_object_put(obj);
 237
 238                         spin_lock_irqsave(&i915->mm.obj_lock, flags);
 239                         if (err)
 240                                 break;
 241                 }
 242                 list_splice_tail(&still_in_list, phase->list);
 243                 spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 244                 if (err)
 245                         return err;
 246         }
 247
 248         if (shrink & I915_SHRINK_BOUND)
 249                 intel_runtime_pm_put(&i915->runtime_pm, wakeref);
 250
 251         if (nr_scanned)
 252                 *nr_scanned += scanned;
 253         return count;
 254 }
 255
 256 /**
 257  * i915_gem_shrink_all - Shrink buffer object caches completely
 258  * @i915: i915 device
 259  *
 260  * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
 261  * caches completely. It also first waits for and retires all outstanding
 262  * requests to also be able to release backing storage for active objects.
 263  *
 264  * This should only be used in code to intentionally quiescent the gpu or as a
 265  * last-ditch effort when memory seems to have run out.
 266  *
 267  * Returns:
 268  * The number of pages of backing storage actually released.
 269  */
 270 unsigned long i915_gem_shrink_all(struct drm_i915_private *i915)
 271 {
 272         intel_wakeref_t wakeref;
 273         unsigned long freed = 0;
 274
 275         with_intel_runtime_pm(&i915->runtime_pm, wakeref) {
 276                 freed = i915_gem_shrink(NULL, i915, -1UL, NULL,
 277                                         I915_SHRINK_BOUND |
 278                                         I915_SHRINK_UNBOUND);
 279         }
 280
 281         return freed;
 282 }
 283
 284 static unsigned long
 285 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
 286 {
 287         struct drm_i915_private *i915 =
 288                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 289         unsigned long num_objects;
 290         unsigned long count;
 291
 292         count = READ_ONCE(i915->mm.shrink_memory) >> PAGE_SHIFT;
 293         num_objects = READ_ONCE(i915->mm.shrink_count);
 294
 295         /*
 296          * Update our preferred vmscan batch size for the next pass.
 297          * Our rough guess for an effective batch size is roughly 2
 298          * available GEM objects worth of pages. That is we don't want
 299          * the shrinker to fire, until it is worth the cost of freeing an
 300          * entire GEM object.
 301          */
 302         if (num_objects) {
 303                 unsigned long avg = 2 * count / num_objects;
 304
 305                 i915->mm.shrinker.batch =
 306                         max((i915->mm.shrinker.batch + avg) >> 1,
 307                             128ul /* default SHRINK_BATCH */);
 308         }
 309
 310         return count;
 311 }
 312
 313 static unsigned long
 314 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 315 {
 316         struct drm_i915_private *i915 =
 317                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 318         unsigned long freed;
 319
 320         sc->nr_scanned = 0;
 321
 322         freed = i915_gem_shrink(NULL, i915,
 323                                 sc->nr_to_scan,
 324                                 &sc->nr_scanned,
 325                                 I915_SHRINK_BOUND |
 326                                 I915_SHRINK_UNBOUND);
 327         if (sc->nr_scanned < sc->nr_to_scan && current_is_kswapd()) {
 328                 intel_wakeref_t wakeref;
 329
 330                 with_intel_runtime_pm(&i915->runtime_pm, wakeref) {
 331                         freed += i915_gem_shrink(NULL, i915,
 332                                                  sc->nr_to_scan - sc->nr_scanned,
 333                                                  &sc->nr_scanned,
 334                                                  I915_SHRINK_ACTIVE |
 335                                                  I915_SHRINK_BOUND |
 336                                                  I915_SHRINK_UNBOUND |
 337                                                  I915_SHRINK_WRITEBACK);
 338                 }
 339         }
 340
 341         return sc->nr_scanned ? freed : SHRINK_STOP;
 342 }
 343
 344 static int
 345 i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
 346 {
 347         struct drm_i915_private *i915 =
 348                 container_of(nb, struct drm_i915_private, mm.oom_notifier);
 349         struct drm_i915_gem_object *obj;
 350         unsigned long unevictable, available, freed_pages;
 351         intel_wakeref_t wakeref;
 352         unsigned long flags;
 353
 354         freed_pages = 0;
 355         with_intel_runtime_pm(&i915->runtime_pm, wakeref)
 356                 freed_pages += i915_gem_shrink(NULL, i915, -1UL, NULL,
 357                                                I915_SHRINK_BOUND |
 358                                                I915_SHRINK_UNBOUND |
 359                                                I915_SHRINK_WRITEBACK);
 360
 361         /* Because we may be allocating inside our own driver, we cannot
 362          * assert that there are no objects with pinned pages that are not
 363          * being pointed to by hardware.
 364          */
 365         available = unevictable = 0;
 366         spin_lock_irqsave(&i915->mm.obj_lock, flags);
 367         list_for_each_entry(obj, &i915->mm.shrink_list, mm.link) {
 368                 if (!can_release_pages(obj))
 369                         unevictable += obj->base.size >> PAGE_SHIFT;
 370                 else
 371                         available += obj->base.size >> PAGE_SHIFT;
 372         }
 373         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 374
 375         if (freed_pages || available)
 376                 pr_info("Purging GPU memory, %lu pages freed, "
 377                         "%lu pages still pinned, %lu pages left available.\n",
 378                         freed_pages, unevictable, available);
 379
 380         *(unsigned long *)ptr += freed_pages;
 381         return NOTIFY_DONE;
 382 }
 383
 384 static int
 385 i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
 386 {
 387         struct drm_i915_private *i915 =
 388                 container_of(nb, struct drm_i915_private, mm.vmap_notifier);
 389         struct i915_vma *vma, *next;
 390         unsigned long freed_pages = 0;
 391         intel_wakeref_t wakeref;
 392
 393         with_intel_runtime_pm(&i915->runtime_pm, wakeref)
 394                 freed_pages += i915_gem_shrink(NULL, i915, -1UL, NULL,
 395                                                I915_SHRINK_BOUND |
 396                                                I915_SHRINK_UNBOUND |
 397                                                I915_SHRINK_VMAPS);
 398
 399         /* We also want to clear any cached iomaps as they wrap vmap */
 400         mutex_lock(&i915->ggtt.vm.mutex);
 401         list_for_each_entry_safe(vma, next,
 402                                  &i915->ggtt.vm.bound_list, vm_link) {
 403                 unsigned long count = vma->node.size >> PAGE_SHIFT;
 404
 405                 if (!vma->iomap || i915_vma_is_active(vma))
 406                         continue;
 407
 408                 if (__i915_vma_unbind(vma) == 0)
 409                         freed_pages += count;
 410         }
 411         mutex_unlock(&i915->ggtt.vm.mutex);
 412
 413         *(unsigned long *)ptr += freed_pages;
 414         return NOTIFY_DONE;
 415 }
 416
 417 void i915_gem_driver_register__shrinker(struct drm_i915_private *i915)
 418 {
 419         i915->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
 420         i915->mm.shrinker.count_objects = i915_gem_shrinker_count;
 421         i915->mm.shrinker.seeks = DEFAULT_SEEKS;
 422         i915->mm.shrinker.batch = 4096;
 423         drm_WARN_ON(&i915->drm, register_shrinker(&i915->mm.shrinker));
 424
 425         i915->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
 426         drm_WARN_ON(&i915->drm, register_oom_notifier(&i915->mm.oom_notifier));
 427
 428         i915->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
 429         drm_WARN_ON(&i915->drm,
 430                     register_vmap_purge_notifier(&i915->mm.vmap_notifier));
 431 }
 432
 433 void i915_gem_driver_unregister__shrinker(struct drm_i915_private *i915)
 434 {
 435         drm_WARN_ON(&i915->drm,
 436                     unregister_vmap_purge_notifier(&i915->mm.vmap_notifier));
 437         drm_WARN_ON(&i915->drm,
 438                     unregister_oom_notifier(&i915->mm.oom_notifier));
 439         unregister_shrinker(&i915->mm.shrinker);
 440 }
 441
 442 void i915_gem_shrinker_taints_mutex(struct drm_i915_private *i915,
 443                                     struct mutex *mutex)
 444 {
 445         if (!IS_ENABLED(CONFIG_LOCKDEP))
 446                 return;
 447
 448         fs_reclaim_acquire(GFP_KERNEL);
 449
 450         mutex_acquire(&mutex->dep_map, 0, 0, _RET_IP_);
 451         mutex_release(&mutex->dep_map, _RET_IP_);
 452
 453         fs_reclaim_release(GFP_KERNEL);
 454 }
 455
 456 #define obj_to_i915(obj__) to_i915((obj__)->base.dev)
 457
 458 void i915_gem_object_make_unshrinkable(struct drm_i915_gem_object *obj)
 459 {
 460         struct drm_i915_private *i915 = obj_to_i915(obj);
 461         unsigned long flags;
 462
 463         /*
 464          * We can only be called while the pages are pinned or when
 465          * the pages are released. If pinned, we should only be called
 466          * from a single caller under controlled conditions; and on release
 467          * only one caller may release us. Neither the two may cross.
 468          */
 469         if (atomic_add_unless(&obj->mm.shrink_pin, 1, 0))
 470                 return;
 471
 472         spin_lock_irqsave(&i915->mm.obj_lock, flags);
 473         if (!atomic_fetch_inc(&obj->mm.shrink_pin) &&
 474             !list_empty(&obj->mm.link)) {
 475                 list_del_init(&obj->mm.link);
 476                 i915->mm.shrink_count--;
 477                 i915->mm.shrink_memory -= obj->base.size;
 478         }
 479         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 480 }
 481
 482 static void __i915_gem_object_make_shrinkable(struct drm_i915_gem_object *obj,
 483                                               struct list_head *head)
 484 {
 485         struct drm_i915_private *i915 = obj_to_i915(obj);
 486         unsigned long flags;
 487
 488         GEM_BUG_ON(!i915_gem_object_has_pages(obj));
 489         if (!i915_gem_object_is_shrinkable(obj))
 490                 return;
 491
 492         if (atomic_add_unless(&obj->mm.shrink_pin, -1, 1))
 493                 return;
 494
 495         spin_lock_irqsave(&i915->mm.obj_lock, flags);
 496         GEM_BUG_ON(!kref_read(&obj->base.refcount));
 497         if (atomic_dec_and_test(&obj->mm.shrink_pin)) {
 498                 GEM_BUG_ON(!list_empty(&obj->mm.link));
 499
 500                 list_add_tail(&obj->mm.link, head);
 501                 i915->mm.shrink_count++;
 502                 i915->mm.shrink_memory += obj->base.size;
 503
 504         }
 505         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 506 }
 507
 508 void i915_gem_object_make_shrinkable(struct drm_i915_gem_object *obj)
 509 {
 510         __i915_gem_object_make_shrinkable(obj,
 511                                           &obj_to_i915(obj)->mm.shrink_list);
 512 }
 513
 514 void i915_gem_object_make_purgeable(struct drm_i915_gem_object *obj)
 515 {
 516         __i915_gem_object_make_shrinkable(obj,
 517                                           &obj_to_i915(obj)->mm.purge_list);
 518 }