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