drivers/gpu/drm/i915/i915_gem_shrinker.c

   1 /*
   2  * Copyright © 2008-2015 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  */
  24
  25 #include <linux/oom.h>
  26 #include <linux/shmem_fs.h>
  27 #include <linux/slab.h>
  28 #include <linux/swap.h>
  29 #include <linux/pci.h>
  30 #include <linux/dma-buf.h>
  31 #include <linux/vmalloc.h>
  32 #include <drm/drmP.h>
  33 #include <drm/i915_drm.h>
  34
  35 #include "i915_drv.h"
  36 #include "i915_trace.h"
  37
  38 static bool shrinker_lock(struct drm_i915_private *dev_priv, bool *unlock)
  39 {
  40         switch (mutex_trylock_recursive(&dev_priv->drm.struct_mutex)) {
  41         case MUTEX_TRYLOCK_RECURSIVE:
  42                 *unlock = false;
  43                 return true;
  44
  45         case MUTEX_TRYLOCK_FAILED:
  46                 *unlock = false;
  47                 preempt_disable();
  48                 do {
  49                         cpu_relax();
  50                         if (mutex_trylock(&dev_priv->drm.struct_mutex)) {
  51                                 *unlock = true;
  52                                 break;
  53                         }
  54                 } while (!need_resched());
  55                 preempt_enable();
  56                 return *unlock;
  57
  58         case MUTEX_TRYLOCK_SUCCESS:
  59                 *unlock = true;
  60                 return true;
  61         }
  62
  63         BUG();
  64 }
  65
  66 static void shrinker_unlock(struct drm_i915_private *dev_priv, bool unlock)
  67 {
  68         if (!unlock)
  69                 return;
  70
  71         mutex_unlock(&dev_priv->drm.struct_mutex);
  72 }
  73
  74 static bool any_vma_pinned(struct drm_i915_gem_object *obj)
  75 {
  76         struct i915_vma *vma;
  77
  78         list_for_each_entry(vma, &obj->vma_list, obj_link) {
  79                 /* Only GGTT vma may be permanently pinned, and are always
  80                  * at the start of the list. We can stop hunting as soon
  81                  * as we see a ppGTT vma.
  82                  */
  83                 if (!i915_vma_is_ggtt(vma))
  84                         break;
  85
  86                 if (i915_vma_is_pinned(vma))
  87                         return true;
  88         }
  89
  90         return false;
  91 }
  92
  93 static bool swap_available(void)
  94 {
  95         return get_nr_swap_pages() > 0;
  96 }
  97
  98 static bool can_release_pages(struct drm_i915_gem_object *obj)
  99 {
 100         if (!obj->mm.pages)
 101                 return false;
 102
 103         /* Consider only shrinkable ojects. */
 104         if (!i915_gem_object_is_shrinkable(obj))
 105                 return false;
 106
 107         /* Only report true if by unbinding the object and putting its pages
 108          * we can actually make forward progress towards freeing physical
 109          * pages.
 110          *
 111          * If the pages are pinned for any other reason than being bound
 112          * to the GPU, simply unbinding from the GPU is not going to succeed
 113          * in releasing our pin count on the pages themselves.
 114          */
 115         if (atomic_read(&obj->mm.pages_pin_count) > obj->bind_count)
 116                 return false;
 117
 118         if (any_vma_pinned(obj))
 119                 return false;
 120
 121         /* We can only return physical pages to the system if we can either
 122          * discard the contents (because the user has marked them as being
 123          * purgeable) or if we can move their contents out to swap.
 124          */
 125         return swap_available() || obj->mm.madv == I915_MADV_DONTNEED;
 126 }
 127
 128 static bool unsafe_drop_pages(struct drm_i915_gem_object *obj)
 129 {
 130         if (i915_gem_object_unbind(obj) == 0)
 131                 __i915_gem_object_put_pages(obj, I915_MM_SHRINKER);
 132         return !READ_ONCE(obj->mm.pages);
 133 }
 134
 135 /**
 136  * i915_gem_shrink - Shrink buffer object caches
 137  * @dev_priv: i915 device
 138  * @target: amount of memory to make available, in pages
 139  * @nr_scanned: optional output for number of pages scanned (incremental)
 140  * @flags: control flags for selecting cache types
 141  *
 142  * This function is the main interface to the shrinker. It will try to release
 143  * up to @target pages of main memory backing storage from buffer objects.
 144  * Selection of the specific caches can be done with @flags. This is e.g. useful
 145  * when purgeable objects should be removed from caches preferentially.
 146  *
 147  * Note that it's not guaranteed that released amount is actually available as
 148  * free system memory - the pages might still be in-used to due to other reasons
 149  * (like cpu mmaps) or the mm core has reused them before we could grab them.
 150  * Therefore code that needs to explicitly shrink buffer objects caches (e.g. to
 151  * avoid deadlocks in memory reclaim) must fall back to i915_gem_shrink_all().
 152  *
 153  * Also note that any kind of pinning (both per-vma address space pins and
 154  * backing storage pins at the buffer object level) result in the shrinker code
 155  * having to skip the object.
 156  *
 157  * Returns:
 158  * The number of pages of backing storage actually released.
 159  */
 160 unsigned long
 161 i915_gem_shrink(struct drm_i915_private *dev_priv,
 162                 unsigned long target,
 163                 unsigned long *nr_scanned,
 164                 unsigned flags)
 165 {
 166         const struct {
 167                 struct list_head *list;
 168                 unsigned int bit;
 169         } phases[] = {
 170                 { &dev_priv->mm.unbound_list, I915_SHRINK_UNBOUND },
 171                 { &dev_priv->mm.bound_list, I915_SHRINK_BOUND },
 172                 { NULL, 0 },
 173         }, *phase;
 174         unsigned long count = 0;
 175         unsigned long scanned = 0;
 176         bool unlock;
 177
 178         if (!shrinker_lock(dev_priv, &unlock))
 179                 return 0;
 180
 181         trace_i915_gem_shrink(dev_priv, target, flags);
 182         i915_gem_retire_requests(dev_priv);
 183
 184         /*
 185          * Unbinding of objects will require HW access; Let us not wake the
 186          * device just to recover a little memory. If absolutely necessary,
 187          * we will force the wake during oom-notifier.
 188          */
 189         if ((flags & I915_SHRINK_BOUND) &&
 190             !intel_runtime_pm_get_if_in_use(dev_priv))
 191                 flags &= ~I915_SHRINK_BOUND;
 192
 193         /*
 194          * As we may completely rewrite the (un)bound list whilst unbinding
 195          * (due to retiring requests) we have to strictly process only
 196          * one element of the list at the time, and recheck the list
 197          * on every iteration.
 198          *
 199          * In particular, we must hold a reference whilst removing the
 200          * object as we may end up waiting for and/or retiring the objects.
 201          * This might release the final reference (held by the active list)
 202          * and result in the object being freed from under us. This is
 203          * similar to the precautions the eviction code must take whilst
 204          * removing objects.
 205          *
 206          * Also note that although these lists do not hold a reference to
 207          * the object we can safely grab one here: The final object
 208          * unreferencing and the bound_list are both protected by the
 209          * dev->struct_mutex and so we won't ever be able to observe an
 210          * object on the bound_list with a reference count equals 0.
 211          */
 212         for (phase = phases; phase->list; phase++) {
 213                 struct list_head still_in_list;
 214                 struct drm_i915_gem_object *obj;
 215
 216                 if ((flags & phase->bit) == 0)
 217                         continue;
 218
 219                 INIT_LIST_HEAD(&still_in_list);
 220                 while (count < target &&
 221                        (obj = list_first_entry_or_null(phase->list,
 222                                                        typeof(*obj),
 223                                                        global_link))) {
 224                         list_move_tail(&obj->global_link, &still_in_list);
 225                         if (!obj->mm.pages) {
 226                                 list_del_init(&obj->global_link);
 227                                 continue;
 228                         }
 229
 230                         if (flags & I915_SHRINK_PURGEABLE &&
 231                             obj->mm.madv != I915_MADV_DONTNEED)
 232                                 continue;
 233
 234                         if (flags & I915_SHRINK_VMAPS &&
 235                             !is_vmalloc_addr(obj->mm.mapping))
 236                                 continue;
 237
 238                         if (!(flags & I915_SHRINK_ACTIVE) &&
 239                             (i915_gem_object_is_active(obj) ||
 240                              i915_gem_object_is_framebuffer(obj)))
 241                                 continue;
 242
 243                         if (!can_release_pages(obj))
 244                                 continue;
 245
 246                         if (unsafe_drop_pages(obj)) {
 247                                 /* May arrive from get_pages on another bo */
 248                                 mutex_lock_nested(&obj->mm.lock,
 249                                                   I915_MM_SHRINKER);
 250                                 if (!obj->mm.pages) {
 251                                         __i915_gem_object_invalidate(obj);
 252                                         list_del_init(&obj->global_link);
 253                                         count += obj->base.size >> PAGE_SHIFT;
 254                                 }
 255                                 mutex_unlock(&obj->mm.lock);
 256                                 scanned += obj->base.size >> PAGE_SHIFT;
 257                         }
 258                 }
 259                 list_splice_tail(&still_in_list, phase->list);
 260         }
 261
 262         if (flags & I915_SHRINK_BOUND)
 263                 intel_runtime_pm_put(dev_priv);
 264
 265         i915_gem_retire_requests(dev_priv);
 266
 267         shrinker_unlock(dev_priv, unlock);
 268
 269         if (nr_scanned)
 270                 *nr_scanned += scanned;
 271         return count;
 272 }
 273
 274 /**
 275  * i915_gem_shrink_all - Shrink buffer object caches completely
 276  * @dev_priv: i915 device
 277  *
 278  * This is a simple wraper around i915_gem_shrink() to aggressively shrink all
 279  * caches completely. It also first waits for and retires all outstanding
 280  * requests to also be able to release backing storage for active objects.
 281  *
 282  * This should only be used in code to intentionally quiescent the gpu or as a
 283  * last-ditch effort when memory seems to have run out.
 284  *
 285  * Returns:
 286  * The number of pages of backing storage actually released.
 287  */
 288 unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv)
 289 {
 290         unsigned long freed;
 291
 292         intel_runtime_pm_get(dev_priv);
 293         freed = i915_gem_shrink(dev_priv, -1UL, NULL,
 294                                 I915_SHRINK_BOUND |
 295                                 I915_SHRINK_UNBOUND |
 296                                 I915_SHRINK_ACTIVE);
 297         intel_runtime_pm_put(dev_priv);
 298
 299         return freed;
 300 }
 301
 302 static unsigned long
 303 i915_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
 304 {
 305         struct drm_i915_private *dev_priv =
 306                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 307         struct drm_i915_gem_object *obj;
 308         unsigned long count;
 309         bool unlock;
 310
 311         if (!shrinker_lock(dev_priv, &unlock))
 312                 return 0;
 313
 314         i915_gem_retire_requests(dev_priv);
 315
 316         count = 0;
 317         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_link)
 318                 if (can_release_pages(obj))
 319                         count += obj->base.size >> PAGE_SHIFT;
 320
 321         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_link) {
 322                 if (!i915_gem_object_is_active(obj) && can_release_pages(obj))
 323                         count += obj->base.size >> PAGE_SHIFT;
 324         }
 325
 326         shrinker_unlock(dev_priv, unlock);
 327
 328         return count;
 329 }
 330
 331 static unsigned long
 332 i915_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 333 {
 334         struct drm_i915_private *dev_priv =
 335                 container_of(shrinker, struct drm_i915_private, mm.shrinker);
 336         unsigned long freed;
 337         bool unlock;
 338
 339         sc->nr_scanned = 0;
 340
 341         if (!shrinker_lock(dev_priv, &unlock))
 342                 return SHRINK_STOP;
 343
 344         freed = i915_gem_shrink(dev_priv,
 345                                 sc->nr_to_scan,
 346                                 &sc->nr_scanned,
 347                                 I915_SHRINK_BOUND |
 348                                 I915_SHRINK_UNBOUND |
 349                                 I915_SHRINK_PURGEABLE);
 350         if (freed < sc->nr_to_scan)
 351                 freed += i915_gem_shrink(dev_priv,
 352                                          sc->nr_to_scan - sc->nr_scanned,
 353                                          &sc->nr_scanned,
 354                                          I915_SHRINK_BOUND |
 355                                          I915_SHRINK_UNBOUND);
 356         if (freed < sc->nr_to_scan && current_is_kswapd()) {
 357                 intel_runtime_pm_get(dev_priv);
 358                 freed += i915_gem_shrink(dev_priv,
 359                                          sc->nr_to_scan - sc->nr_scanned,
 360                                          &sc->nr_scanned,
 361                                          I915_SHRINK_ACTIVE |
 362                                          I915_SHRINK_BOUND |
 363                                          I915_SHRINK_UNBOUND);
 364                 intel_runtime_pm_put(dev_priv);
 365         }
 366
 367         shrinker_unlock(dev_priv, unlock);
 368
 369         return sc->nr_scanned ? freed : SHRINK_STOP;
 370 }
 371
 372 static bool
 373 shrinker_lock_uninterruptible(struct drm_i915_private *dev_priv, bool *unlock,
 374                               int timeout_ms)
 375 {
 376         unsigned long timeout = jiffies + msecs_to_jiffies_timeout(timeout_ms);
 377
 378         do {
 379                 if (i915_gem_wait_for_idle(dev_priv, 0) == 0 &&
 380                     shrinker_lock(dev_priv, unlock))
 381                         break;
 382
 383                 schedule_timeout_killable(1);
 384                 if (fatal_signal_pending(current))
 385                         return false;
 386
 387                 if (time_after(jiffies, timeout)) {
 388                         pr_err("Unable to lock GPU to purge memory.\n");
 389                         return false;
 390                 }
 391         } while (1);
 392
 393         return true;
 394 }
 395
 396 static int
 397 i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
 398 {
 399         struct drm_i915_private *dev_priv =
 400                 container_of(nb, struct drm_i915_private, mm.oom_notifier);
 401         struct drm_i915_gem_object *obj;
 402         unsigned long unevictable, bound, unbound, freed_pages;
 403         bool unlock;
 404
 405         if (!shrinker_lock_uninterruptible(dev_priv, &unlock, 5000))
 406                 return NOTIFY_DONE;
 407
 408         freed_pages = i915_gem_shrink_all(dev_priv);
 409
 410         /* Because we may be allocating inside our own driver, we cannot
 411          * assert that there are no objects with pinned pages that are not
 412          * being pointed to by hardware.
 413          */
 414         unbound = bound = unevictable = 0;
 415         list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_link) {
 416                 if (!obj->mm.pages)
 417                         continue;
 418
 419                 if (!can_release_pages(obj))
 420                         unevictable += obj->base.size >> PAGE_SHIFT;
 421                 else
 422                         unbound += obj->base.size >> PAGE_SHIFT;
 423         }
 424         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_link) {
 425                 if (!obj->mm.pages)
 426                         continue;
 427
 428                 if (!can_release_pages(obj))
 429                         unevictable += obj->base.size >> PAGE_SHIFT;
 430                 else
 431                         bound += obj->base.size >> PAGE_SHIFT;
 432         }
 433
 434         shrinker_unlock(dev_priv, unlock);
 435
 436         if (freed_pages || unbound || bound)
 437                 pr_info("Purging GPU memory, %lu pages freed, "
 438                         "%lu pages still pinned.\n",
 439                         freed_pages, unevictable);
 440         if (unbound || bound)
 441                 pr_err("%lu and %lu pages still available in the "
 442                        "bound and unbound GPU page lists.\n",
 443                        bound, unbound);
 444
 445         *(unsigned long *)ptr += freed_pages;
 446         return NOTIFY_DONE;
 447 }
 448
 449 static int
 450 i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
 451 {
 452         struct drm_i915_private *dev_priv =
 453                 container_of(nb, struct drm_i915_private, mm.vmap_notifier);
 454         struct i915_vma *vma, *next;
 455         unsigned long freed_pages = 0;
 456         bool unlock;
 457         int ret;
 458
 459         if (!shrinker_lock_uninterruptible(dev_priv, &unlock, 5000))
 460                 return NOTIFY_DONE;
 461
 462         /* Force everything onto the inactive lists */
 463         ret = i915_gem_wait_for_idle(dev_priv, I915_WAIT_LOCKED);
 464         if (ret)
 465                 goto out;
 466
 467         intel_runtime_pm_get(dev_priv);
 468         freed_pages += i915_gem_shrink(dev_priv, -1UL, NULL,
 469                                        I915_SHRINK_BOUND |
 470                                        I915_SHRINK_UNBOUND |
 471                                        I915_SHRINK_ACTIVE |
 472                                        I915_SHRINK_VMAPS);
 473         intel_runtime_pm_put(dev_priv);
 474
 475         /* We also want to clear any cached iomaps as they wrap vmap */
 476         list_for_each_entry_safe(vma, next,
 477                                  &dev_priv->ggtt.base.inactive_list, vm_link) {
 478                 unsigned long count = vma->node.size >> PAGE_SHIFT;
 479                 if (vma->iomap && i915_vma_unbind(vma) == 0)
 480                         freed_pages += count;
 481         }
 482
 483 out:
 484         shrinker_unlock(dev_priv, unlock);
 485
 486         *(unsigned long *)ptr += freed_pages;
 487         return NOTIFY_DONE;
 488 }
 489
 490 /**
 491  * i915_gem_shrinker_init - Initialize i915 shrinker
 492  * @dev_priv: i915 device
 493  *
 494  * This function registers and sets up the i915 shrinker and OOM handler.
 495  */
 496 void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
 497 {
 498         dev_priv->mm.shrinker.scan_objects = i915_gem_shrinker_scan;
 499         dev_priv->mm.shrinker.count_objects = i915_gem_shrinker_count;
 500         dev_priv->mm.shrinker.seeks = DEFAULT_SEEKS;
 501         WARN_ON(register_shrinker(&dev_priv->mm.shrinker));
 502
 503         dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
 504         WARN_ON(register_oom_notifier(&dev_priv->mm.oom_notifier));
 505
 506         dev_priv->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
 507         WARN_ON(register_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
 508 }
 509
 510 /**
 511  * i915_gem_shrinker_cleanup - Clean up i915 shrinker
 512  * @dev_priv: i915 device
 513  *
 514  * This function unregisters the i915 shrinker and OOM handler.
 515  */
 516 void i915_gem_shrinker_cleanup(struct drm_i915_private *dev_priv)
 517 {
 518         WARN_ON(unregister_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
 519         WARN_ON(unregister_oom_notifier(&dev_priv->mm.oom_notifier));
 520         unregister_shrinker(&dev_priv->mm.shrinker);
 521 }