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 #include "gt/intel_gt.h"
  18
  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 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 int try_to_writeback(struct drm_i915_gem_object *obj, unsigned int flags)
  60 {
  61         if (obj->ops->shrink) {
  62                 unsigned int shrink_flags = 0;
  63
  64                 if (!(flags & I915_SHRINK_ACTIVE))
  65                         shrink_flags |= I915_GEM_OBJECT_SHRINK_NO_GPU_WAIT;
  66
  67                 if (flags & I915_SHRINK_WRITEBACK)
  68                         shrink_flags |= I915_GEM_OBJECT_SHRINK_WRITEBACK;
  69
  70                 return obj->ops->shrink(obj, shrink_flags);
  71         }
  72
  73         return 0;
  74 }
  75
  76 /**
  77  * i915_gem_shrink - Shrink buffer object caches
  78  * @ww: i915 gem ww acquire ctx, or NULL
  79  * @i915: i915 device
  80  * @target: amount of memory to make available, in pages
  81  * @nr_scanned: optional output for number of pages scanned (incremental)
  82  * @shrink: control flags for selecting cache types
  83  *
  84  * This function is the main interface to the shrinker. It will try to release
  85  * up to @target pages of main memory backing storage from buffer objects.
  86  * Selection of the specific caches can be done with @flags. This is e.g. useful
  87  * when purgeable objects should be removed from caches preferentially.
  88  *
  89  * Note that it's not guaranteed that released amount is actually available as
  90  * free system memory - the pages might still be in-used to due to other reasons
  91  * (like cpu mmaps) or the mm core has reused them before we could grab them.
  92  * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
  93  * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
  94  *
  95  * Also note that any kind of pinning (both per-vma address space pins and
  96  * backing storage pins at the buffer object level) result in the shrinker code
  97  * having to skip the object.
  98  *
  99  * Returns:
 100  * The number of pages of backing storage actually released.
 101  */
 102 unsigned long
 103 i915_gem_shrink(struct i915_gem_ww_ctx *ww,
 104                 struct drm_i915_private *i915,
 105                 unsigned long target,
 106                 unsigned long *nr_scanned,
 107                 unsigned int shrink)
 108 {
 109         const struct {
 110                 struct list_head *list;
 111                 unsigned int bit;
 112         } phases[] = {
 113                 { &i915->mm.purge_list, ~0u },
 114                 {
 115                         &i915->mm.shrink_list,
 116                         I915_SHRINK_BOUND | I915_SHRINK_UNBOUND
 117                 },
 118                 { NULL, 0 },
 119         }, *phase;
 120         intel_wakeref_t wakeref = 0;
 121         unsigned long count = 0;
 122         unsigned long scanned = 0;
 123         int err = 0, i = 0;
 124         struct intel_gt *gt;
 125
 126         /* CHV + VTD workaround use stop_machine(); need to trylock vm->mutex */
 127         bool trylock_vm = !ww && intel_vm_no_concurrent_access_wa(i915);
 128
 129         trace_i915_gem_shrink(i915, target, shrink);
 130
 131         /*
 132          * Unbinding of objects will require HW access; Let us not wake the
 133          * device just to recover a little memory. If absolutely necessary,
 134          * we will force the wake during oom-notifier.
 135          */
 136         if (shrink & I915_SHRINK_BOUND) {
 137                 wakeref = intel_runtime_pm_get_if_in_use(&i915->runtime_pm);
 138                 if (!wakeref)
 139                         shrink &= ~I915_SHRINK_BOUND;
 140         }
 141
 142         /*
 143          * When shrinking the active list, we should also consider active
 144          * contexts. Active contexts are pinned until they are retired, and
 145          * so can not be simply unbound to retire and unpin their pages. To
 146          * shrink the contexts, we must wait until the gpu is idle and
 147          * completed its switch to the kernel context. In short, we do
 148          * not have a good mechanism for idling a specific context, but
 149          * what we can do is give them a kick so that we do not keep idle
 150          * contexts around longer than is necessary.
 151          */
 152         if (shrink & I915_SHRINK_ACTIVE) {
 153                 for_each_gt(gt, i915, i)
 154                         /* Retire requests to unpin all idle contexts */
 155                         intel_gt_retire_requests(gt);
 156         }
 157
 158         /*
 159          * As we may completely rewrite the (un)bound list whilst unbinding
 160          * (due to retiring requests) we have to strictly process only
 161          * one element of the list at the time, and recheck the list
 162          * on every iteration.
 163          *
 164          * In particular, we must hold a reference whilst removing the
 165          * object as we may end up waiting for and/or retiring the objects.
 166          * This might release the final reference (held by the active list)
 167          * and result in the object being freed from under us. This is
 168          * similar to the precautions the eviction code must take whilst
 169          * removing objects.
 170          *
 171          * Also note that although these lists do not hold a reference to
 172          * the object we can safely grab one here: The final object
 173          * unreferencing and the bound_list are both protected by the
 174          * dev->struct_mutex and so we won't ever be able to observe an
 175          * object on the bound_list with a reference count equals 0.
 176          */
 177         for (phase = phases; phase->list; phase++) {
 178                 struct list_head still_in_list;
 179                 struct drm_i915_gem_object *obj;
 180                 unsigned long flags;
 181
 182                 if ((shrink & phase->bit) == 0)
 183                         continue;
 184
 185                 INIT_LIST_HEAD(&still_in_list);
 186
 187                 /*
 188                  * We serialize our access to unreferenced objects through
 189                  * the use of the struct_mutex. While the objects are not
 190                  * yet freed (due to RCU then a workqueue) we still want
 191                  * to be able to shrink their pages, so they remain on
 192                  * the unbound/bound list until actually freed.
 193                  */
 194                 spin_lock_irqsave(&i915->mm.obj_lock, flags);
 195                 while (count < target &&
 196                        (obj = list_first_entry_or_null(phase->list,
 197                                                        typeof(*obj),
 198                                                        mm.link))) {
 199                         list_move_tail(&obj->mm.link, &still_in_list);
 200
 201                         if (shrink & I915_SHRINK_VMAPS &&
 202                             !is_vmalloc_addr(obj->mm.mapping))
 203                                 continue;
 204
 205                         if (!(shrink & I915_SHRINK_ACTIVE) &&
 206                             i915_gem_object_is_framebuffer(obj))
 207                                 continue;
 208
 209                         if (!can_release_pages(obj))
 210                                 continue;
 211
 212                         if (!kref_get_unless_zero(&obj->base.refcount))
 213                                 continue;
 214
 215                         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 216
 217                         /* May arrive from get_pages on another bo */
 218                         if (!ww) {
 219                                 if (!i915_gem_object_trylock(obj, NULL))
 220                                         goto skip;
 221                         } else {
 222                                 err = i915_gem_object_lock(obj, ww);
 223                                 if (err)
 224                                         goto skip;
 225                         }
 226
 227                         if (drop_pages(obj, shrink, trylock_vm) &&
 228                             !__i915_gem_object_put_pages(obj) &&
 229                             !try_to_writeback(obj, shrink))
 230                                 count += obj->base.size >> PAGE_SHIFT;
 231
 232                         if (!ww)
 233                                 i915_gem_object_unlock(obj);
 234
 235                         scanned += obj->base.size >> PAGE_SHIFT;
 236 skip:
 237                         i915_gem_object_put(obj);
 238
 239                         spin_lock_irqsave(&i915->mm.obj_lock, flags);
 240                         if (err)
 241                                 break;
 242                 }
 243                 list_splice_tail(&still_in_list, phase->list);
 244                 spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 245                 if (err)
 246                         break;
 247         }
 248
 249         if (shrink & I915_SHRINK_BOUND)
 250                 intel_runtime_pm_put(&i915->runtime_pm, wakeref);
 251
 252         if (err)
 253                 return err;
 254
 255         if (nr_scanned)
 256                 *nr_scanned += scanned;
 257         return count;
 258 }
 259
 260 /**
 261  * i915_gem_shrink_all - Shrink buffer object caches completely
 262  * @i915: i915 device
 263  *
 264  * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
 265  * caches completely. It also first waits for and retires all outstanding
 266  * requests to also be able to release backing storage for active objects.
 267  *
 268  * This should only be used in code to intentionally quiescent the gpu or as a
 269  * last-ditch effort when memory seems to have run out.
 270  *
 271  * Returns:
 272  * The number of pages of backing storage actually released.
 273  */
 274 unsigned long i915_gem_shrink_all(struct drm_i915_private *i915)
 275 {
 276         intel_wakeref_t wakeref;
 277         unsigned long freed = 0;
 278
 279         with_intel_runtime_pm(&i915->runtime_pm, wakeref) {
 280                 freed = i915_gem_shrink(NULL, i915, -1UL, NULL,
 281                                         I915_SHRINK_BOUND |
 282                                         I915_SHRINK_UNBOUND);
 283         }
 284
 285         return freed;
 286 }
 287
 288 static unsigned long
 289 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
 290 {
 291         struct drm_i915_private *i915 = shrinker->private_data;
 292         unsigned long num_objects;
 293         unsigned long count;
 294
 295         count = READ_ONCE(i915->mm.shrink_memory) >> PAGE_SHIFT;
 296         num_objects = READ_ONCE(i915->mm.shrink_count);
 297
 298         /*
 299          * Update our preferred vmscan batch size for the next pass.
 300          * Our rough guess for an effective batch size is roughly 2
 301          * available GEM objects worth of pages. That is we don't want
 302          * the shrinker to fire, until it is worth the cost of freeing an
 303          * entire GEM object.
 304          */
 305         if (num_objects) {
 306                 unsigned long avg = 2 * count / num_objects;
 307
 308                 i915->mm.shrinker->batch =
 309                         max((i915->mm.shrinker->batch + avg) >> 1,
 310                             128ul /* default SHRINK_BATCH */);
 311         }
 312
 313         return count;
 314 }
 315
 316 static unsigned long
 317 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 318 {
 319         struct drm_i915_private *i915 = shrinker->private_data;
 320         unsigned long freed;
 321
 322         sc->nr_scanned = 0;
 323
 324         freed = i915_gem_shrink(NULL, i915,
 325                                 sc->nr_to_scan,
 326                                 &sc->nr_scanned,
 327                                 I915_SHRINK_BOUND |
 328                                 I915_SHRINK_UNBOUND);
 329         if (sc->nr_scanned < sc->nr_to_scan && current_is_kswapd()) {
 330                 intel_wakeref_t wakeref;
 331
 332                 with_intel_runtime_pm(&i915->runtime_pm, wakeref) {
 333                         freed += i915_gem_shrink(NULL, i915,
 334                                                  sc->nr_to_scan - sc->nr_scanned,
 335                                                  &sc->nr_scanned,
 336                                                  I915_SHRINK_ACTIVE |
 337                                                  I915_SHRINK_BOUND |
 338                                                  I915_SHRINK_UNBOUND |
 339                                                  I915_SHRINK_WRITEBACK);
 340                 }
 341         }
 342
 343         return sc->nr_scanned ? freed : SHRINK_STOP;
 344 }
 345
 346 static int
 347 i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
 348 {
 349         struct drm_i915_private *i915 =
 350                 container_of(nb, struct drm_i915_private, mm.oom_notifier);
 351         struct drm_i915_gem_object *obj;
 352         unsigned long unevictable, available, freed_pages;
 353         intel_wakeref_t wakeref;
 354         unsigned long flags;
 355
 356         freed_pages = 0;
 357         with_intel_runtime_pm(&i915->runtime_pm, wakeref)
 358                 freed_pages += i915_gem_shrink(NULL, i915, -1UL, NULL,
 359                                                I915_SHRINK_BOUND |
 360                                                I915_SHRINK_UNBOUND |
 361                                                I915_SHRINK_WRITEBACK);
 362
 363         /* Because we may be allocating inside our own driver, we cannot
 364          * assert that there are no objects with pinned pages that are not
 365          * being pointed to by hardware.
 366          */
 367         available = unevictable = 0;
 368         spin_lock_irqsave(&i915->mm.obj_lock, flags);
 369         list_for_each_entry(obj, &i915->mm.shrink_list, mm.link) {
 370                 if (!can_release_pages(obj))
 371                         unevictable += obj->base.size >> PAGE_SHIFT;
 372                 else
 373                         available += obj->base.size >> PAGE_SHIFT;
 374         }
 375         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 376
 377         if (freed_pages || available)
 378                 pr_info("Purging GPU memory, %lu pages freed, "
 379                         "%lu pages still pinned, %lu pages left available.\n",
 380                         freed_pages, unevictable, available);
 381
 382         *(unsigned long *)ptr += freed_pages;
 383         return NOTIFY_DONE;
 384 }
 385
 386 static int
 387 i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
 388 {
 389         struct drm_i915_private *i915 =
 390                 container_of(nb, struct drm_i915_private, mm.vmap_notifier);
 391         struct i915_vma *vma, *next;
 392         unsigned long freed_pages = 0;
 393         intel_wakeref_t wakeref;
 394         struct intel_gt *gt;
 395         int i;
 396
 397         with_intel_runtime_pm(&i915->runtime_pm, wakeref)
 398                 freed_pages += i915_gem_shrink(NULL, i915, -1UL, NULL,
 399                                                I915_SHRINK_BOUND |
 400                                                I915_SHRINK_UNBOUND |
 401                                                I915_SHRINK_VMAPS);
 402
 403         /* We also want to clear any cached iomaps as they wrap vmap */
 404         for_each_gt(gt, i915, i) {
 405                 mutex_lock(&gt->ggtt->vm.mutex);
 406                 list_for_each_entry_safe(vma, next,
 407                                          &gt->ggtt->vm.bound_list, vm_link) {
 408                         unsigned long count = i915_vma_size(vma) >> PAGE_SHIFT;
 409                         struct drm_i915_gem_object *obj = vma->obj;
 410
 411                         if (!vma->iomap || i915_vma_is_active(vma))
 412                                 continue;
 413
 414                         if (!i915_gem_object_trylock(obj, NULL))
 415                                 continue;
 416
 417                         if (__i915_vma_unbind(vma) == 0)
 418                                 freed_pages += count;
 419
 420                         i915_gem_object_unlock(obj);
 421                 }
 422                 mutex_unlock(&gt->ggtt->vm.mutex);
 423         }
 424
 425         *(unsigned long *)ptr += freed_pages;
 426         return NOTIFY_DONE;
 427 }
 428
 429 void i915_gem_driver_register__shrinker(struct drm_i915_private *i915)
 430 {
 431         i915->mm.shrinker = shrinker_alloc(0, "drm-i915_gem");
 432         if (!i915->mm.shrinker) {
 433                 drm_WARN_ON(&i915->drm, 1);
 434         } else {
 435                 i915->mm.shrinker->scan_objects = i915_gem_shrinker_scan;
 436                 i915->mm.shrinker->count_objects = i915_gem_shrinker_count;
 437                 i915->mm.shrinker->batch = 4096;
 438                 i915->mm.shrinker->private_data = i915;
 439
 440                 shrinker_register(i915->mm.shrinker);
 441         }
 442
 443         i915->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
 444         drm_WARN_ON(&i915->drm, register_oom_notifier(&i915->mm.oom_notifier));
 445
 446         i915->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
 447         drm_WARN_ON(&i915->drm,
 448                     register_vmap_purge_notifier(&i915->mm.vmap_notifier));
 449 }
 450
 451 void i915_gem_driver_unregister__shrinker(struct drm_i915_private *i915)
 452 {
 453         drm_WARN_ON(&i915->drm,
 454                     unregister_vmap_purge_notifier(&i915->mm.vmap_notifier));
 455         drm_WARN_ON(&i915->drm,
 456                     unregister_oom_notifier(&i915->mm.oom_notifier));
 457         shrinker_free(i915->mm.shrinker);
 458 }
 459
 460 void i915_gem_shrinker_taints_mutex(struct drm_i915_private *i915,
 461                                     struct mutex *mutex)
 462 {
 463         if (!IS_ENABLED(CONFIG_LOCKDEP))
 464                 return;
 465
 466         fs_reclaim_acquire(GFP_KERNEL);
 467
 468         mutex_acquire(&mutex->dep_map, 0, 0, _RET_IP_);
 469         mutex_release(&mutex->dep_map, _RET_IP_);
 470
 471         fs_reclaim_release(GFP_KERNEL);
 472 }
 473
 474 /**
 475  * i915_gem_object_make_unshrinkable - Hide the object from the shrinker. By
 476  * default all object types that support shrinking(see IS_SHRINKABLE), will also
 477  * make the object visible to the shrinker after allocating the system memory
 478  * pages.
 479  * @obj: The GEM object.
 480  *
 481  * This is typically used for special kernel internal objects that can't be
 482  * easily processed by the shrinker, like if they are perma-pinned.
 483  */
 484 void i915_gem_object_make_unshrinkable(struct drm_i915_gem_object *obj)
 485 {
 486         struct drm_i915_private *i915 = obj_to_i915(obj);
 487         unsigned long flags;
 488
 489         /*
 490          * We can only be called while the pages are pinned or when
 491          * the pages are released. If pinned, we should only be called
 492          * from a single caller under controlled conditions; and on release
 493          * only one caller may release us. Neither the two may cross.
 494          */
 495         if (atomic_add_unless(&obj->mm.shrink_pin, 1, 0))
 496                 return;
 497
 498         spin_lock_irqsave(&i915->mm.obj_lock, flags);
 499         if (!atomic_fetch_inc(&obj->mm.shrink_pin) &&
 500             !list_empty(&obj->mm.link)) {
 501                 list_del_init(&obj->mm.link);
 502                 i915->mm.shrink_count--;
 503                 i915->mm.shrink_memory -= obj->base.size;
 504         }
 505         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 506 }
 507
 508 static void ___i915_gem_object_make_shrinkable(struct drm_i915_gem_object *obj,
 509                                                struct list_head *head)
 510 {
 511         struct drm_i915_private *i915 = obj_to_i915(obj);
 512         unsigned long flags;
 513
 514         if (!i915_gem_object_is_shrinkable(obj))
 515                 return;
 516
 517         if (atomic_add_unless(&obj->mm.shrink_pin, -1, 1))
 518                 return;
 519
 520         spin_lock_irqsave(&i915->mm.obj_lock, flags);
 521         GEM_BUG_ON(!kref_read(&obj->base.refcount));
 522         if (atomic_dec_and_test(&obj->mm.shrink_pin)) {
 523                 GEM_BUG_ON(!list_empty(&obj->mm.link));
 524
 525                 list_add_tail(&obj->mm.link, head);
 526                 i915->mm.shrink_count++;
 527                 i915->mm.shrink_memory += obj->base.size;
 528
 529         }
 530         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 531 }
 532
 533 /**
 534  * __i915_gem_object_make_shrinkable - Move the object to the tail of the
 535  * shrinkable list. Objects on this list might be swapped out. Used with
 536  * WILLNEED objects.
 537  * @obj: The GEM object.
 538  *
 539  * DO NOT USE. This is intended to be called on very special objects that don't
 540  * yet have mm.pages, but are guaranteed to have potentially reclaimable pages
 541  * underneath.
 542  */
 543 void __i915_gem_object_make_shrinkable(struct drm_i915_gem_object *obj)
 544 {
 545         ___i915_gem_object_make_shrinkable(obj,
 546                                            &obj_to_i915(obj)->mm.shrink_list);
 547 }
 548
 549 /**
 550  * __i915_gem_object_make_purgeable - Move the object to the tail of the
 551  * purgeable list. Objects on this list might be swapped out. Used with
 552  * DONTNEED objects.
 553  * @obj: The GEM object.
 554  *
 555  * DO NOT USE. This is intended to be called on very special objects that don't
 556  * yet have mm.pages, but are guaranteed to have potentially reclaimable pages
 557  * underneath.
 558  */
 559 void __i915_gem_object_make_purgeable(struct drm_i915_gem_object *obj)
 560 {
 561         ___i915_gem_object_make_shrinkable(obj,
 562                                            &obj_to_i915(obj)->mm.purge_list);
 563 }
 564
 565 /**
 566  * i915_gem_object_make_shrinkable - Move the object to the tail of the
 567  * shrinkable list. Objects on this list might be swapped out. Used with
 568  * WILLNEED objects.
 569  * @obj: The GEM object.
 570  *
 571  * MUST only be called on objects which have backing pages.
 572  *
 573  * MUST be balanced with previous call to i915_gem_object_make_unshrinkable().
 574  */
 575 void i915_gem_object_make_shrinkable(struct drm_i915_gem_object *obj)
 576 {
 577         GEM_BUG_ON(!i915_gem_object_has_pages(obj));
 578         __i915_gem_object_make_shrinkable(obj);
 579 }
 580
 581 /**
 582  * i915_gem_object_make_purgeable - Move the object to the tail of the purgeable
 583  * list. Used with DONTNEED objects. Unlike with shrinkable objects, the
 584  * shrinker will attempt to discard the backing pages, instead of trying to swap
 585  * them out.
 586  * @obj: The GEM object.
 587  *
 588  * MUST only be called on objects which have backing pages.
 589  *
 590  * MUST be balanced with previous call to i915_gem_object_make_unshrinkable().
 591  */
 592 void i915_gem_object_make_purgeable(struct drm_i915_gem_object *obj)
 593 {
 594         GEM_BUG_ON(!i915_gem_object_has_pages(obj));
 595         __i915_gem_object_make_purgeable(obj);
 596 }