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 #include <drm/i915_drm.h>
  16
  17 #include "i915_trace.h"
  18
  19 static bool swap_available(void)
  20 {
  21         return get_nr_swap_pages() > 0;
  22 }
  23
  24 static bool can_release_pages(struct drm_i915_gem_object *obj)
  25 {
  26         /* Consider only shrinkable ojects. */
  27         if (!i915_gem_object_is_shrinkable(obj))
  28                 return false;
  29
  30         /*
  31          * Only report true if by unbinding the object and putting its pages
  32          * we can actually make forward progress towards freeing physical
  33          * pages.
  34          *
  35          * If the pages are pinned for any other reason than being bound
  36          * to the GPU, simply unbinding from the GPU is not going to succeed
  37          * in releasing our pin count on the pages themselves.
  38          */
  39         if (atomic_read(&obj->mm.pages_pin_count) > atomic_read(&obj->bind_count))
  40                 return false;
  41
  42         /*
  43          * We can only return physical pages to the system if we can either
  44          * discard the contents (because the user has marked them as being
  45          * purgeable) or if we can move their contents out to swap.
  46          */
  47         return swap_available() || obj->mm.madv == I915_MADV_DONTNEED;
  48 }
  49
  50 static bool unsafe_drop_pages(struct drm_i915_gem_object *obj,
  51                               unsigned long shrink)
  52 {
  53         unsigned long flags;
  54
  55         flags = 0;
  56         if (shrink & I915_SHRINK_ACTIVE)
  57                 flags = I915_GEM_OBJECT_UNBIND_ACTIVE;
  58
  59         if (i915_gem_object_unbind(obj, flags) == 0)
  60                 __i915_gem_object_put_pages(obj);
  61
  62         return !i915_gem_object_has_pages(obj);
  63 }
  64
  65 static void try_to_writeback(struct drm_i915_gem_object *obj,
  66                              unsigned int flags)
  67 {
  68         switch (obj->mm.madv) {
  69         case I915_MADV_DONTNEED:
  70                 i915_gem_object_truncate(obj);
  71         case __I915_MADV_PURGED:
  72                 return;
  73         }
  74
  75         if (flags & I915_SHRINK_WRITEBACK)
  76                 i915_gem_object_writeback(obj);
  77 }
  78
  79 /**
  80  * i915_gem_shrink - Shrink buffer object caches
  81  * @i915: i915 device
  82  * @target: amount of memory to make available, in pages
  83  * @nr_scanned: optional output for number of pages scanned (incremental)
  84  * @shrink: control flags for selecting cache types
  85  *
  86  * This function is the main interface to the shrinker. It will try to release
  87  * up to @target pages of main memory backing storage from buffer objects.
  88  * Selection of the specific caches can be done with @flags. This is e.g. useful
  89  * when purgeable objects should be removed from caches preferentially.
  90  *
  91  * Note that it's not guaranteed that released amount is actually available as
  92  * free system memory - the pages might still be in-used to due to other reasons
  93  * (like cpu mmaps) or the mm core has reused them before we could grab them.
  94  * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
  95  * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
  96  *
  97  * Also note that any kind of pinning (both per-vma address space pins and
  98  * backing storage pins at the buffer object level) result in the shrinker code
  99  * having to skip the object.
 100  *
 101  * Returns:
 102  * The number of pages of backing storage actually released.
 103  */
 104 unsigned long
 105 i915_gem_shrink(struct drm_i915_private *i915,
 106                 unsigned long target,
 107                 unsigned long *nr_scanned,
 108                 unsigned int shrink)
 109 {
 110         const struct {
 111                 struct list_head *list;
 112                 unsigned int bit;
 113         } phases[] = {
 114                 { &i915->mm.purge_list, ~0u },
 115                 {
 116                         &i915->mm.shrink_list,
 117                         I915_SHRINK_BOUND | I915_SHRINK_UNBOUND
 118                 },
 119                 { NULL, 0 },
 120         }, *phase;
 121         intel_wakeref_t wakeref = 0;
 122         unsigned long count = 0;
 123         unsigned long scanned = 0;
 124
 125         /*
 126          * When shrinking the active list, we should also consider active
 127          * contexts. Active contexts are pinned until they are retired, and
 128          * so can not be simply unbound to retire and unpin their pages. To
 129          * shrink the contexts, we must wait until the gpu is idle and
 130          * completed its switch to the kernel context. In short, we do
 131          * not have a good mechanism for idling a specific context.
 132          */
 133
 134         trace_i915_gem_shrink(i915, target, shrink);
 135
 136         /*
 137          * Unbinding of objects will require HW access; Let us not wake the
 138          * device just to recover a little memory. If absolutely necessary,
 139          * we will force the wake during oom-notifier.
 140          */
 141         if (shrink & I915_SHRINK_BOUND) {
 142                 wakeref = intel_runtime_pm_get_if_in_use(&i915->runtime_pm);
 143                 if (!wakeref)
 144                         shrink &= ~I915_SHRINK_BOUND;
 145         }
 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 (!(shrink & I915_SHRINK_BOUND) &&
 199                             atomic_read(&obj->bind_count))
 200                                 continue;
 201
 202                         if (!can_release_pages(obj))
 203                                 continue;
 204
 205                         if (!kref_get_unless_zero(&obj->base.refcount))
 206                                 continue;
 207
 208                         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 209
 210                         if (unsafe_drop_pages(obj, shrink)) {
 211                                 /* May arrive from get_pages on another bo */
 212                                 mutex_lock(&obj->mm.lock);
 213                                 if (!i915_gem_object_has_pages(obj)) {
 214                                         try_to_writeback(obj, shrink);
 215                                         count += obj->base.size >> PAGE_SHIFT;
 216                                 }
 217                                 mutex_unlock(&obj->mm.lock);
 218                         }
 219
 220                         scanned += obj->base.size >> PAGE_SHIFT;
 221                         i915_gem_object_put(obj);
 222
 223                         spin_lock_irqsave(&i915->mm.obj_lock, flags);
 224                 }
 225                 list_splice_tail(&still_in_list, phase->list);
 226                 spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 227         }
 228
 229         if (shrink & I915_SHRINK_BOUND)
 230                 intel_runtime_pm_put(&i915->runtime_pm, wakeref);
 231
 232         if (nr_scanned)
 233                 *nr_scanned += scanned;
 234         return count;
 235 }
 236
 237 /**
 238  * i915_gem_shrink_all - Shrink buffer object caches completely
 239  * @i915: i915 device
 240  *
 241  * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
 242  * caches completely. It also first waits for and retires all outstanding
 243  * requests to also be able to release backing storage for active objects.
 244  *
 245  * This should only be used in code to intentionally quiescent the gpu or as a
 246  * last-ditch effort when memory seems to have run out.
 247  *
 248  * Returns:
 249  * The number of pages of backing storage actually released.
 250  */
 251 unsigned long i915_gem_shrink_all(struct drm_i915_private *i915)
 252 {
 253         intel_wakeref_t wakeref;
 254         unsigned long freed = 0;
 255
 256         with_intel_runtime_pm(&i915->runtime_pm, wakeref) {
 257                 freed = i915_gem_shrink(i915, -1UL, NULL,
 258                                         I915_SHRINK_BOUND |
 259                                         I915_SHRINK_UNBOUND);
 260         }
 261
 262         return freed;
 263 }
 264
 265 static unsigned long
 266 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
 267 {
 268         struct drm_i915_private *i915 =
 269                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 270         unsigned long num_objects;
 271         unsigned long count;
 272
 273         count = READ_ONCE(i915->mm.shrink_memory) >> PAGE_SHIFT;
 274         num_objects = READ_ONCE(i915->mm.shrink_count);
 275
 276         /*
 277          * Update our preferred vmscan batch size for the next pass.
 278          * Our rough guess for an effective batch size is roughly 2
 279          * available GEM objects worth of pages. That is we don't want
 280          * the shrinker to fire, until it is worth the cost of freeing an
 281          * entire GEM object.
 282          */
 283         if (num_objects) {
 284                 unsigned long avg = 2 * count / num_objects;
 285
 286                 i915->mm.shrinker.batch =
 287                         max((i915->mm.shrinker.batch + avg) >> 1,
 288                             128ul /* default SHRINK_BATCH */);
 289         }
 290
 291         return count;
 292 }
 293
 294 static unsigned long
 295 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 296 {
 297         struct drm_i915_private *i915 =
 298                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 299         unsigned long freed;
 300
 301         sc->nr_scanned = 0;
 302
 303         freed = i915_gem_shrink(i915,
 304                                 sc->nr_to_scan,
 305                                 &sc->nr_scanned,
 306                                 I915_SHRINK_BOUND |
 307                                 I915_SHRINK_UNBOUND);
 308         if (sc->nr_scanned < sc->nr_to_scan && current_is_kswapd()) {
 309                 intel_wakeref_t wakeref;
 310
 311                 with_intel_runtime_pm(&i915->runtime_pm, wakeref) {
 312                         freed += i915_gem_shrink(i915,
 313                                                  sc->nr_to_scan - sc->nr_scanned,
 314                                                  &sc->nr_scanned,
 315                                                  I915_SHRINK_ACTIVE |
 316                                                  I915_SHRINK_BOUND |
 317                                                  I915_SHRINK_UNBOUND |
 318                                                  I915_SHRINK_WRITEBACK);
 319                 }
 320         }
 321
 322         return sc->nr_scanned ? freed : SHRINK_STOP;
 323 }
 324
 325 static int
 326 i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
 327 {
 328         struct drm_i915_private *i915 =
 329                 container_of(nb, struct drm_i915_private, mm.oom_notifier);
 330         struct drm_i915_gem_object *obj;
 331         unsigned long unevictable, available, freed_pages;
 332         intel_wakeref_t wakeref;
 333         unsigned long flags;
 334
 335         freed_pages = 0;
 336         with_intel_runtime_pm(&i915->runtime_pm, wakeref)
 337                 freed_pages += i915_gem_shrink(i915, -1UL, NULL,
 338                                                I915_SHRINK_BOUND |
 339                                                I915_SHRINK_UNBOUND |
 340                                                I915_SHRINK_WRITEBACK);
 341
 342         /* Because we may be allocating inside our own driver, we cannot
 343          * assert that there are no objects with pinned pages that are not
 344          * being pointed to by hardware.
 345          */
 346         available = unevictable = 0;
 347         spin_lock_irqsave(&i915->mm.obj_lock, flags);
 348         list_for_each_entry(obj, &i915->mm.shrink_list, mm.link) {
 349                 if (!can_release_pages(obj))
 350                         unevictable += obj->base.size >> PAGE_SHIFT;
 351                 else
 352                         available += obj->base.size >> PAGE_SHIFT;
 353         }
 354         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 355
 356         if (freed_pages || available)
 357                 pr_info("Purging GPU memory, %lu pages freed, "
 358                         "%lu pages still pinned, %lu pages left available.\n",
 359                         freed_pages, unevictable, available);
 360
 361         *(unsigned long *)ptr += freed_pages;
 362         return NOTIFY_DONE;
 363 }
 364
 365 static int
 366 i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
 367 {
 368         struct drm_i915_private *i915 =
 369                 container_of(nb, struct drm_i915_private, mm.vmap_notifier);
 370         struct i915_vma *vma, *next;
 371         unsigned long freed_pages = 0;
 372         intel_wakeref_t wakeref;
 373
 374         with_intel_runtime_pm(&i915->runtime_pm, wakeref)
 375                 freed_pages += i915_gem_shrink(i915, -1UL, NULL,
 376                                                I915_SHRINK_BOUND |
 377                                                I915_SHRINK_UNBOUND |
 378                                                I915_SHRINK_VMAPS);
 379
 380         /* We also want to clear any cached iomaps as they wrap vmap */
 381         mutex_lock(&i915->ggtt.vm.mutex);
 382         list_for_each_entry_safe(vma, next,
 383                                  &i915->ggtt.vm.bound_list, vm_link) {
 384                 unsigned long count = vma->node.size >> PAGE_SHIFT;
 385
 386                 if (!vma->iomap || i915_vma_is_active(vma))
 387                         continue;
 388
 389                 if (__i915_vma_unbind(vma) == 0)
 390                         freed_pages += count;
 391         }
 392         mutex_unlock(&i915->ggtt.vm.mutex);
 393
 394         *(unsigned long *)ptr += freed_pages;
 395         return NOTIFY_DONE;
 396 }
 397
 398 void i915_gem_driver_register__shrinker(struct drm_i915_private *i915)
 399 {
 400         i915->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
 401         i915->mm.shrinker.count_objects = i915_gem_shrinker_count;
 402         i915->mm.shrinker.seeks = DEFAULT_SEEKS;
 403         i915->mm.shrinker.batch = 4096;
 404         WARN_ON(register_shrinker(&i915->mm.shrinker));
 405
 406         i915->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
 407         WARN_ON(register_oom_notifier(&i915->mm.oom_notifier));
 408
 409         i915->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
 410         WARN_ON(register_vmap_purge_notifier(&i915->mm.vmap_notifier));
 411 }
 412
 413 void i915_gem_driver_unregister__shrinker(struct drm_i915_private *i915)
 414 {
 415         WARN_ON(unregister_vmap_purge_notifier(&i915->mm.vmap_notifier));
 416         WARN_ON(unregister_oom_notifier(&i915->mm.oom_notifier));
 417         unregister_shrinker(&i915->mm.shrinker);
 418 }
 419
 420 void i915_gem_shrinker_taints_mutex(struct drm_i915_private *i915,
 421                                     struct mutex *mutex)
 422 {
 423         bool unlock = false;
 424
 425         if (!IS_ENABLED(CONFIG_LOCKDEP))
 426                 return;
 427
 428         if (!lockdep_is_held_type(&i915->drm.struct_mutex, -1)) {
 429                 mutex_acquire(&i915->drm.struct_mutex.dep_map,
 430                               I915_MM_NORMAL, 0, _RET_IP_);
 431                 unlock = true;
 432         }
 433
 434         fs_reclaim_acquire(GFP_KERNEL);
 435
 436         mutex_acquire(&mutex->dep_map, 0, 0, _RET_IP_);
 437         mutex_release(&mutex->dep_map, _RET_IP_);
 438
 439         fs_reclaim_release(GFP_KERNEL);
 440
 441         if (unlock)
 442                 mutex_release(&i915->drm.struct_mutex.dep_map, _RET_IP_);
 443 }
 444
 445 #define obj_to_i915(obj__) to_i915((obj__)->base.dev)
 446
 447 void i915_gem_object_make_unshrinkable(struct drm_i915_gem_object *obj)
 448 {
 449         struct drm_i915_private *i915 = obj_to_i915(obj);
 450         unsigned long flags;
 451
 452         /*
 453          * We can only be called while the pages are pinned or when
 454          * the pages are released. If pinned, we should only be called
 455          * from a single caller under controlled conditions; and on release
 456          * only one caller may release us. Neither the two may cross.
 457          */
 458         if (atomic_add_unless(&obj->mm.shrink_pin, 1, 0))
 459                 return;
 460
 461         spin_lock_irqsave(&i915->mm.obj_lock, flags);
 462         if (!atomic_fetch_inc(&obj->mm.shrink_pin) &&
 463             !list_empty(&obj->mm.link)) {
 464                 list_del_init(&obj->mm.link);
 465                 i915->mm.shrink_count--;
 466                 i915->mm.shrink_memory -= obj->base.size;
 467         }
 468         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 469 }
 470
 471 static void __i915_gem_object_make_shrinkable(struct drm_i915_gem_object *obj,
 472                                               struct list_head *head)
 473 {
 474         struct drm_i915_private *i915 = obj_to_i915(obj);
 475         unsigned long flags;
 476
 477         GEM_BUG_ON(!i915_gem_object_has_pages(obj));
 478         if (!i915_gem_object_is_shrinkable(obj))
 479                 return;
 480
 481         if (atomic_add_unless(&obj->mm.shrink_pin, -1, 1))
 482                 return;
 483
 484         spin_lock_irqsave(&i915->mm.obj_lock, flags);
 485         GEM_BUG_ON(!kref_read(&obj->base.refcount));
 486         if (atomic_dec_and_test(&obj->mm.shrink_pin)) {
 487                 GEM_BUG_ON(!list_empty(&obj->mm.link));
 488
 489                 list_add_tail(&obj->mm.link, head);
 490                 i915->mm.shrink_count++;
 491                 i915->mm.shrink_memory += obj->base.size;
 492
 493         }
 494         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 495 }
 496
 497 void i915_gem_object_make_shrinkable(struct drm_i915_gem_object *obj)
 498 {
 499         __i915_gem_object_make_shrinkable(obj,
 500                                           &obj_to_i915(obj)->mm.shrink_list);
 501 }
 502
 503 void i915_gem_object_make_purgeable(struct drm_i915_gem_object *obj)
 504 {
 505         __i915_gem_object_make_shrinkable(obj,
 506                                           &obj_to_i915(obj)->mm.purge_list);
 507 }