drivers/gpu/drm/ttm/ttm_bo.c

   1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
   2 /**************************************************************************
   3  *
   4  * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  20  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  21  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  22  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  23  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  24  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  25  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28 /*
  29  * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
  30  */
  31
  32 #define pr_fmt(fmt) "[TTM] " fmt
  33
  34 #include <drm/ttm/ttm_bo_driver.h>
  35 #include <drm/ttm/ttm_placement.h>
  36 #include <linux/jiffies.h>
  37 #include <linux/slab.h>
  38 #include <linux/sched.h>
  39 #include <linux/mm.h>
  40 #include <linux/file.h>
  41 #include <linux/module.h>
  42 #include <linux/atomic.h>
  43 #include <linux/dma-resv.h>
  44
  45 #include "ttm_module.h"
  46
  47 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
  48                                         struct ttm_placement *placement)
  49 {
  50         struct drm_printer p = drm_debug_printer(TTM_PFX);
  51         struct ttm_resource_manager *man;
  52         int i, mem_type;
  53
  54         drm_printf(&p, "No space for %p (%lu pages, %zuK, %zuM)\n",
  55                    bo, bo->resource->num_pages, bo->base.size >> 10,
  56                    bo->base.size >> 20);
  57         for (i = 0; i < placement->num_placement; i++) {
  58                 mem_type = placement->placement[i].mem_type;
  59                 drm_printf(&p, "  placement[%d]=0x%08X (%d)\n",
  60                            i, placement->placement[i].flags, mem_type);
  61                 man = ttm_manager_type(bo->bdev, mem_type);
  62                 ttm_resource_manager_debug(man, &p);
  63         }
  64 }
  65
  66 /**
  67  * ttm_bo_move_to_lru_tail
  68  *
  69  * @bo: The buffer object.
  70  *
  71  * Move this BO to the tail of all lru lists used to lookup and reserve an
  72  * object. This function must be called with struct ttm_global::lru_lock
  73  * held, and is used to make a BO less likely to be considered for eviction.
  74  */
  75 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo)
  76 {
  77         dma_resv_assert_held(bo->base.resv);
  78
  79         if (bo->resource)
  80                 ttm_resource_move_to_lru_tail(bo->resource);
  81 }
  82 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
  83
  84 /**
  85  * ttm_bo_set_bulk_move - update BOs bulk move object
  86  *
  87  * @bo: The buffer object.
  88  *
  89  * Update the BOs bulk move object, making sure that resources are added/removed
  90  * as well. A bulk move allows to move many resource on the LRU at once,
  91  * resulting in much less overhead of maintaining the LRU.
  92  * The only requirement is that the resources stay together on the LRU and are
  93  * never separated. This is enforces by setting the bulk_move structure on a BO.
  94  * ttm_lru_bulk_move_tail() should be used to move all resources to the tail of
  95  * their LRU list.
  96  */
  97 void ttm_bo_set_bulk_move(struct ttm_buffer_object *bo,
  98                           struct ttm_lru_bulk_move *bulk)
  99 {
 100         dma_resv_assert_held(bo->base.resv);
 101
 102         if (bo->bulk_move == bulk)
 103                 return;
 104
 105         spin_lock(&bo->bdev->lru_lock);
 106         if (bo->resource)
 107                 ttm_resource_del_bulk_move(bo->resource, bo);
 108         bo->bulk_move = bulk;
 109         if (bo->resource)
 110                 ttm_resource_add_bulk_move(bo->resource, bo);
 111         spin_unlock(&bo->bdev->lru_lock);
 112 }
 113 EXPORT_SYMBOL(ttm_bo_set_bulk_move);
 114
 115 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
 116                                   struct ttm_resource *mem, bool evict,
 117                                   struct ttm_operation_ctx *ctx,
 118                                   struct ttm_place *hop)
 119 {
 120         struct ttm_resource_manager *old_man, *new_man;
 121         struct ttm_device *bdev = bo->bdev;
 122         int ret;
 123
 124         old_man = ttm_manager_type(bdev, bo->resource->mem_type);
 125         new_man = ttm_manager_type(bdev, mem->mem_type);
 126
 127         ttm_bo_unmap_virtual(bo);
 128
 129         /*
 130          * Create and bind a ttm if required.
 131          */
 132
 133         if (new_man->use_tt) {
 134                 /* Zero init the new TTM structure if the old location should
 135                  * have used one as well.
 136                  */
 137                 ret = ttm_tt_create(bo, old_man->use_tt);
 138                 if (ret)
 139                         goto out_err;
 140
 141                 if (mem->mem_type != TTM_PL_SYSTEM) {
 142                         ret = ttm_tt_populate(bo->bdev, bo->ttm, ctx);
 143                         if (ret)
 144                                 goto out_err;
 145                 }
 146         }
 147
 148         ret = dma_resv_reserve_fences(bo->base.resv, 1);
 149         if (ret)
 150                 goto out_err;
 151
 152         ret = bdev->funcs->move(bo, evict, ctx, mem, hop);
 153         if (ret) {
 154                 if (ret == -EMULTIHOP)
 155                         return ret;
 156                 goto out_err;
 157         }
 158
 159         ctx->bytes_moved += bo->base.size;
 160         return 0;
 161
 162 out_err:
 163         new_man = ttm_manager_type(bdev, bo->resource->mem_type);
 164         if (!new_man->use_tt)
 165                 ttm_bo_tt_destroy(bo);
 166
 167         return ret;
 168 }
 169
 170 /*
 171  * Call bo::reserved.
 172  * Will release GPU memory type usage on destruction.
 173  * This is the place to put in driver specific hooks to release
 174  * driver private resources.
 175  * Will release the bo::reserved lock.
 176  */
 177
 178 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
 179 {
 180         if (bo->bdev->funcs->delete_mem_notify)
 181                 bo->bdev->funcs->delete_mem_notify(bo);
 182
 183         ttm_bo_tt_destroy(bo);
 184         ttm_resource_free(bo, &bo->resource);
 185 }
 186
 187 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
 188 {
 189         int r;
 190
 191         if (bo->base.resv == &bo->base._resv)
 192                 return 0;
 193
 194         BUG_ON(!dma_resv_trylock(&bo->base._resv));
 195
 196         r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
 197         dma_resv_unlock(&bo->base._resv);
 198         if (r)
 199                 return r;
 200
 201         if (bo->type != ttm_bo_type_sg) {
 202                 /* This works because the BO is about to be destroyed and nobody
 203                  * reference it any more. The only tricky case is the trylock on
 204                  * the resv object while holding the lru_lock.
 205                  */
 206                 spin_lock(&bo->bdev->lru_lock);
 207                 bo->base.resv = &bo->base._resv;
 208                 spin_unlock(&bo->bdev->lru_lock);
 209         }
 210
 211         return r;
 212 }
 213
 214 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
 215 {
 216         struct dma_resv *resv = &bo->base._resv;
 217         struct dma_resv_iter cursor;
 218         struct dma_fence *fence;
 219
 220         dma_resv_iter_begin(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP);
 221         dma_resv_for_each_fence_unlocked(&cursor, fence) {
 222                 if (!fence->ops->signaled)
 223                         dma_fence_enable_sw_signaling(fence);
 224         }
 225         dma_resv_iter_end(&cursor);
 226 }
 227
 228 /**
 229  * ttm_bo_cleanup_refs
 230  * If bo idle, remove from lru lists, and unref.
 231  * If not idle, block if possible.
 232  *
 233  * Must be called with lru_lock and reservation held, this function
 234  * will drop the lru lock and optionally the reservation lock before returning.
 235  *
 236  * @bo:                    The buffer object to clean-up
 237  * @interruptible:         Any sleeps should occur interruptibly.
 238  * @no_wait_gpu:           Never wait for gpu. Return -EBUSY instead.
 239  * @unlock_resv:           Unlock the reservation lock as well.
 240  */
 241
 242 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
 243                                bool interruptible, bool no_wait_gpu,
 244                                bool unlock_resv)
 245 {
 246         struct dma_resv *resv = &bo->base._resv;
 247         int ret;
 248
 249         if (dma_resv_test_signaled(resv, DMA_RESV_USAGE_BOOKKEEP))
 250                 ret = 0;
 251         else
 252                 ret = -EBUSY;
 253
 254         if (ret && !no_wait_gpu) {
 255                 long lret;
 256
 257                 if (unlock_resv)
 258                         dma_resv_unlock(bo->base.resv);
 259                 spin_unlock(&bo->bdev->lru_lock);
 260
 261                 lret = dma_resv_wait_timeout(resv, DMA_RESV_USAGE_BOOKKEEP,
 262                                              interruptible,
 263                                              30 * HZ);
 264
 265                 if (lret < 0)
 266                         return lret;
 267                 else if (lret == 0)
 268                         return -EBUSY;
 269
 270                 spin_lock(&bo->bdev->lru_lock);
 271                 if (unlock_resv && !dma_resv_trylock(bo->base.resv)) {
 272                         /*
 273                          * We raced, and lost, someone else holds the reservation now,
 274                          * and is probably busy in ttm_bo_cleanup_memtype_use.
 275                          *
 276                          * Even if it's not the case, because we finished waiting any
 277                          * delayed destruction would succeed, so just return success
 278                          * here.
 279                          */
 280                         spin_unlock(&bo->bdev->lru_lock);
 281                         return 0;
 282                 }
 283                 ret = 0;
 284         }
 285
 286         if (ret || unlikely(list_empty(&bo->ddestroy))) {
 287                 if (unlock_resv)
 288                         dma_resv_unlock(bo->base.resv);
 289                 spin_unlock(&bo->bdev->lru_lock);
 290                 return ret;
 291         }
 292
 293         list_del_init(&bo->ddestroy);
 294         spin_unlock(&bo->bdev->lru_lock);
 295         ttm_bo_cleanup_memtype_use(bo);
 296
 297         if (unlock_resv)
 298                 dma_resv_unlock(bo->base.resv);
 299
 300         ttm_bo_put(bo);
 301
 302         return 0;
 303 }
 304
 305 /*
 306  * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
 307  * encountered buffers.
 308  */
 309 bool ttm_bo_delayed_delete(struct ttm_device *bdev, bool remove_all)
 310 {
 311         struct list_head removed;
 312         bool empty;
 313
 314         INIT_LIST_HEAD(&removed);
 315
 316         spin_lock(&bdev->lru_lock);
 317         while (!list_empty(&bdev->ddestroy)) {
 318                 struct ttm_buffer_object *bo;
 319
 320                 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
 321                                       ddestroy);
 322                 list_move_tail(&bo->ddestroy, &removed);
 323                 if (!ttm_bo_get_unless_zero(bo))
 324                         continue;
 325
 326                 if (remove_all || bo->base.resv != &bo->base._resv) {
 327                         spin_unlock(&bdev->lru_lock);
 328                         dma_resv_lock(bo->base.resv, NULL);
 329
 330                         spin_lock(&bdev->lru_lock);
 331                         ttm_bo_cleanup_refs(bo, false, !remove_all, true);
 332
 333                 } else if (dma_resv_trylock(bo->base.resv)) {
 334                         ttm_bo_cleanup_refs(bo, false, !remove_all, true);
 335                 } else {
 336                         spin_unlock(&bdev->lru_lock);
 337                 }
 338
 339                 ttm_bo_put(bo);
 340                 spin_lock(&bdev->lru_lock);
 341         }
 342         list_splice_tail(&removed, &bdev->ddestroy);
 343         empty = list_empty(&bdev->ddestroy);
 344         spin_unlock(&bdev->lru_lock);
 345
 346         return empty;
 347 }
 348
 349 static void ttm_bo_release(struct kref *kref)
 350 {
 351         struct ttm_buffer_object *bo =
 352             container_of(kref, struct ttm_buffer_object, kref);
 353         struct ttm_device *bdev = bo->bdev;
 354         int ret;
 355
 356         WARN_ON_ONCE(bo->pin_count);
 357         WARN_ON_ONCE(bo->bulk_move);
 358
 359         if (!bo->deleted) {
 360                 ret = ttm_bo_individualize_resv(bo);
 361                 if (ret) {
 362                         /* Last resort, if we fail to allocate memory for the
 363                          * fences block for the BO to become idle
 364                          */
 365                         dma_resv_wait_timeout(bo->base.resv,
 366                                               DMA_RESV_USAGE_BOOKKEEP, false,
 367                                               30 * HZ);
 368                 }
 369
 370                 if (bo->bdev->funcs->release_notify)
 371                         bo->bdev->funcs->release_notify(bo);
 372
 373                 drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
 374                 ttm_mem_io_free(bdev, bo->resource);
 375         }
 376
 377         if (!dma_resv_test_signaled(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP) ||
 378             !dma_resv_trylock(bo->base.resv)) {
 379                 /* The BO is not idle, resurrect it for delayed destroy */
 380                 ttm_bo_flush_all_fences(bo);
 381                 bo->deleted = true;
 382
 383                 spin_lock(&bo->bdev->lru_lock);
 384
 385                 /*
 386                  * Make pinned bos immediately available to
 387                  * shrinkers, now that they are queued for
 388                  * destruction.
 389                  *
 390                  * FIXME: QXL is triggering this. Can be removed when the
 391                  * driver is fixed.
 392                  */
 393                 if (bo->pin_count) {
 394                         bo->pin_count = 0;
 395                         ttm_resource_move_to_lru_tail(bo->resource);
 396                 }
 397
 398                 kref_init(&bo->kref);
 399                 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
 400                 spin_unlock(&bo->bdev->lru_lock);
 401
 402                 schedule_delayed_work(&bdev->wq,
 403                                       ((HZ / 100) < 1) ? 1 : HZ / 100);
 404                 return;
 405         }
 406
 407         spin_lock(&bo->bdev->lru_lock);
 408         list_del(&bo->ddestroy);
 409         spin_unlock(&bo->bdev->lru_lock);
 410
 411         ttm_bo_cleanup_memtype_use(bo);
 412         dma_resv_unlock(bo->base.resv);
 413
 414         atomic_dec(&ttm_glob.bo_count);
 415         bo->destroy(bo);
 416 }
 417
 418 void ttm_bo_put(struct ttm_buffer_object *bo)
 419 {
 420         kref_put(&bo->kref, ttm_bo_release);
 421 }
 422 EXPORT_SYMBOL(ttm_bo_put);
 423
 424 int ttm_bo_lock_delayed_workqueue(struct ttm_device *bdev)
 425 {
 426         return cancel_delayed_work_sync(&bdev->wq);
 427 }
 428 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
 429
 430 void ttm_bo_unlock_delayed_workqueue(struct ttm_device *bdev, int resched)
 431 {
 432         if (resched)
 433                 schedule_delayed_work(&bdev->wq,
 434                                       ((HZ / 100) < 1) ? 1 : HZ / 100);
 435 }
 436 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
 437
 438 static int ttm_bo_bounce_temp_buffer(struct ttm_buffer_object *bo,
 439                                      struct ttm_resource **mem,
 440                                      struct ttm_operation_ctx *ctx,
 441                                      struct ttm_place *hop)
 442 {
 443         struct ttm_placement hop_placement;
 444         struct ttm_resource *hop_mem;
 445         int ret;
 446
 447         hop_placement.num_placement = hop_placement.num_busy_placement = 1;
 448         hop_placement.placement = hop_placement.busy_placement = hop;
 449
 450         /* find space in the bounce domain */
 451         ret = ttm_bo_mem_space(bo, &hop_placement, &hop_mem, ctx);
 452         if (ret)
 453                 return ret;
 454         /* move to the bounce domain */
 455         ret = ttm_bo_handle_move_mem(bo, hop_mem, false, ctx, NULL);
 456         if (ret) {
 457                 ttm_resource_free(bo, &hop_mem);
 458                 return ret;
 459         }
 460         return 0;
 461 }
 462
 463 static int ttm_bo_evict(struct ttm_buffer_object *bo,
 464                         struct ttm_operation_ctx *ctx)
 465 {
 466         struct ttm_device *bdev = bo->bdev;
 467         struct ttm_resource *evict_mem;
 468         struct ttm_placement placement;
 469         struct ttm_place hop;
 470         int ret = 0;
 471
 472         memset(&hop, 0, sizeof(hop));
 473
 474         dma_resv_assert_held(bo->base.resv);
 475
 476         placement.num_placement = 0;
 477         placement.num_busy_placement = 0;
 478         bdev->funcs->evict_flags(bo, &placement);
 479
 480         if (!placement.num_placement && !placement.num_busy_placement) {
 481                 ret = ttm_bo_wait(bo, true, false);
 482                 if (ret)
 483                         return ret;
 484
 485                 /*
 486                  * Since we've already synced, this frees backing store
 487                  * immediately.
 488                  */
 489                 return ttm_bo_pipeline_gutting(bo);
 490         }
 491
 492         ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
 493         if (ret) {
 494                 if (ret != -ERESTARTSYS) {
 495                         pr_err("Failed to find memory space for buffer 0x%p eviction\n",
 496                                bo);
 497                         ttm_bo_mem_space_debug(bo, &placement);
 498                 }
 499                 goto out;
 500         }
 501
 502 bounce:
 503         ret = ttm_bo_handle_move_mem(bo, evict_mem, true, ctx, &hop);
 504         if (ret == -EMULTIHOP) {
 505                 ret = ttm_bo_bounce_temp_buffer(bo, &evict_mem, ctx, &hop);
 506                 if (ret) {
 507                         pr_err("Buffer eviction failed\n");
 508                         ttm_resource_free(bo, &evict_mem);
 509                         goto out;
 510                 }
 511                 /* try and move to final place now. */
 512                 goto bounce;
 513         }
 514 out:
 515         return ret;
 516 }
 517
 518 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
 519                               const struct ttm_place *place)
 520 {
 521         dma_resv_assert_held(bo->base.resv);
 522         if (bo->resource->mem_type == TTM_PL_SYSTEM)
 523                 return true;
 524
 525         /* Don't evict this BO if it's outside of the
 526          * requested placement range
 527          */
 528         if (place->fpfn >= (bo->resource->start + bo->resource->num_pages) ||
 529             (place->lpfn && place->lpfn <= bo->resource->start))
 530                 return false;
 531
 532         return true;
 533 }
 534 EXPORT_SYMBOL(ttm_bo_eviction_valuable);
 535
 536 /*
 537  * Check the target bo is allowable to be evicted or swapout, including cases:
 538  *
 539  * a. if share same reservation object with ctx->resv, have assumption
 540  * reservation objects should already be locked, so not lock again and
 541  * return true directly when either the opreation allow_reserved_eviction
 542  * or the target bo already is in delayed free list;
 543  *
 544  * b. Otherwise, trylock it.
 545  */
 546 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
 547                                            struct ttm_operation_ctx *ctx,
 548                                            const struct ttm_place *place,
 549                                            bool *locked, bool *busy)
 550 {
 551         bool ret = false;
 552
 553         if (bo->base.resv == ctx->resv) {
 554                 dma_resv_assert_held(bo->base.resv);
 555                 if (ctx->allow_res_evict)
 556                         ret = true;
 557                 *locked = false;
 558                 if (busy)
 559                         *busy = false;
 560         } else {
 561                 ret = dma_resv_trylock(bo->base.resv);
 562                 *locked = ret;
 563                 if (busy)
 564                         *busy = !ret;
 565         }
 566
 567         if (ret && place && (bo->resource->mem_type != place->mem_type ||
 568                 !bo->bdev->funcs->eviction_valuable(bo, place))) {
 569                 ret = false;
 570                 if (*locked) {
 571                         dma_resv_unlock(bo->base.resv);
 572                         *locked = false;
 573                 }
 574         }
 575
 576         return ret;
 577 }
 578
 579 /**
 580  * ttm_mem_evict_wait_busy - wait for a busy BO to become available
 581  *
 582  * @busy_bo: BO which couldn't be locked with trylock
 583  * @ctx: operation context
 584  * @ticket: acquire ticket
 585  *
 586  * Try to lock a busy buffer object to avoid failing eviction.
 587  */
 588 static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
 589                                    struct ttm_operation_ctx *ctx,
 590                                    struct ww_acquire_ctx *ticket)
 591 {
 592         int r;
 593
 594         if (!busy_bo || !ticket)
 595                 return -EBUSY;
 596
 597         if (ctx->interruptible)
 598                 r = dma_resv_lock_interruptible(busy_bo->base.resv,
 599                                                           ticket);
 600         else
 601                 r = dma_resv_lock(busy_bo->base.resv, ticket);
 602
 603         /*
 604          * TODO: It would be better to keep the BO locked until allocation is at
 605          * least tried one more time, but that would mean a much larger rework
 606          * of TTM.
 607          */
 608         if (!r)
 609                 dma_resv_unlock(busy_bo->base.resv);
 610
 611         return r == -EDEADLK ? -EBUSY : r;
 612 }
 613
 614 int ttm_mem_evict_first(struct ttm_device *bdev,
 615                         struct ttm_resource_manager *man,
 616                         const struct ttm_place *place,
 617                         struct ttm_operation_ctx *ctx,
 618                         struct ww_acquire_ctx *ticket)
 619 {
 620         struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
 621         struct ttm_resource_cursor cursor;
 622         struct ttm_resource *res;
 623         bool locked = false;
 624         int ret;
 625
 626         spin_lock(&bdev->lru_lock);
 627         ttm_resource_manager_for_each_res(man, &cursor, res) {
 628                 bool busy;
 629
 630                 if (!ttm_bo_evict_swapout_allowable(res->bo, ctx, place,
 631                                                     &locked, &busy)) {
 632                         if (busy && !busy_bo && ticket !=
 633                             dma_resv_locking_ctx(res->bo->base.resv))
 634                                 busy_bo = res->bo;
 635                         continue;
 636                 }
 637
 638                 if (ttm_bo_get_unless_zero(res->bo)) {
 639                         bo = res->bo;
 640                         break;
 641                 }
 642                 if (locked)
 643                         dma_resv_unlock(res->bo->base.resv);
 644         }
 645
 646         if (!bo) {
 647                 if (busy_bo && !ttm_bo_get_unless_zero(busy_bo))
 648                         busy_bo = NULL;
 649                 spin_unlock(&bdev->lru_lock);
 650                 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
 651                 if (busy_bo)
 652                         ttm_bo_put(busy_bo);
 653                 return ret;
 654         }
 655
 656         if (bo->deleted) {
 657                 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
 658                                           ctx->no_wait_gpu, locked);
 659                 ttm_bo_put(bo);
 660                 return ret;
 661         }
 662
 663         spin_unlock(&bdev->lru_lock);
 664
 665         ret = ttm_bo_evict(bo, ctx);
 666         if (locked)
 667                 ttm_bo_unreserve(bo);
 668         else
 669                 ttm_bo_move_to_lru_tail_unlocked(bo);
 670
 671         ttm_bo_put(bo);
 672         return ret;
 673 }
 674
 675 /**
 676  * ttm_bo_pin - Pin the buffer object.
 677  * @bo: The buffer object to pin
 678  *
 679  * Make sure the buffer is not evicted any more during memory pressure.
 680  * @bo must be unpinned again by calling ttm_bo_unpin().
 681  */
 682 void ttm_bo_pin(struct ttm_buffer_object *bo)
 683 {
 684         dma_resv_assert_held(bo->base.resv);
 685         WARN_ON_ONCE(!kref_read(&bo->kref));
 686         spin_lock(&bo->bdev->lru_lock);
 687         if (bo->resource)
 688                 ttm_resource_del_bulk_move(bo->resource, bo);
 689         ++bo->pin_count;
 690         spin_unlock(&bo->bdev->lru_lock);
 691 }
 692 EXPORT_SYMBOL(ttm_bo_pin);
 693
 694 /**
 695  * ttm_bo_unpin - Unpin the buffer object.
 696  * @bo: The buffer object to unpin
 697  *
 698  * Allows the buffer object to be evicted again during memory pressure.
 699  */
 700 void ttm_bo_unpin(struct ttm_buffer_object *bo)
 701 {
 702         dma_resv_assert_held(bo->base.resv);
 703         WARN_ON_ONCE(!kref_read(&bo->kref));
 704         if (WARN_ON_ONCE(!bo->pin_count))
 705                 return;
 706
 707         spin_lock(&bo->bdev->lru_lock);
 708         --bo->pin_count;
 709         if (bo->resource)
 710                 ttm_resource_add_bulk_move(bo->resource, bo);
 711         spin_unlock(&bo->bdev->lru_lock);
 712 }
 713 EXPORT_SYMBOL(ttm_bo_unpin);
 714
 715 /*
 716  * Add the last move fence to the BO as kernel dependency and reserve a new
 717  * fence slot.
 718  */
 719 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
 720                                  struct ttm_resource_manager *man,
 721                                  struct ttm_resource *mem,
 722                                  bool no_wait_gpu)
 723 {
 724         struct dma_fence *fence;
 725         int ret;
 726
 727         spin_lock(&man->move_lock);
 728         fence = dma_fence_get(man->move);
 729         spin_unlock(&man->move_lock);
 730
 731         if (!fence)
 732                 return 0;
 733
 734         if (no_wait_gpu) {
 735                 ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
 736                 dma_fence_put(fence);
 737                 return ret;
 738         }
 739
 740         dma_resv_add_fence(bo->base.resv, fence, DMA_RESV_USAGE_KERNEL);
 741
 742         ret = dma_resv_reserve_fences(bo->base.resv, 1);
 743         dma_fence_put(fence);
 744         return ret;
 745 }
 746
 747 /*
 748  * Repeatedly evict memory from the LRU for @mem_type until we create enough
 749  * space, or we've evicted everything and there isn't enough space.
 750  */
 751 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
 752                                   const struct ttm_place *place,
 753                                   struct ttm_resource **mem,
 754                                   struct ttm_operation_ctx *ctx)
 755 {
 756         struct ttm_device *bdev = bo->bdev;
 757         struct ttm_resource_manager *man;
 758         struct ww_acquire_ctx *ticket;
 759         int ret;
 760
 761         man = ttm_manager_type(bdev, place->mem_type);
 762         ticket = dma_resv_locking_ctx(bo->base.resv);
 763         do {
 764                 ret = ttm_resource_alloc(bo, place, mem);
 765                 if (likely(!ret))
 766                         break;
 767                 if (unlikely(ret != -ENOSPC))
 768                         return ret;
 769                 ret = ttm_mem_evict_first(bdev, man, place, ctx,
 770                                           ticket);
 771                 if (unlikely(ret != 0))
 772                         return ret;
 773         } while (1);
 774
 775         return ttm_bo_add_move_fence(bo, man, *mem, ctx->no_wait_gpu);
 776 }
 777
 778 /*
 779  * Creates space for memory region @mem according to its type.
 780  *
 781  * This function first searches for free space in compatible memory types in
 782  * the priority order defined by the driver.  If free space isn't found, then
 783  * ttm_bo_mem_force_space is attempted in priority order to evict and find
 784  * space.
 785  */
 786 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
 787                         struct ttm_placement *placement,
 788                         struct ttm_resource **mem,
 789                         struct ttm_operation_ctx *ctx)
 790 {
 791         struct ttm_device *bdev = bo->bdev;
 792         bool type_found = false;
 793         int i, ret;
 794
 795         ret = dma_resv_reserve_fences(bo->base.resv, 1);
 796         if (unlikely(ret))
 797                 return ret;
 798
 799         for (i = 0; i < placement->num_placement; ++i) {
 800                 const struct ttm_place *place = &placement->placement[i];
 801                 struct ttm_resource_manager *man;
 802
 803                 man = ttm_manager_type(bdev, place->mem_type);
 804                 if (!man || !ttm_resource_manager_used(man))
 805                         continue;
 806
 807                 type_found = true;
 808                 ret = ttm_resource_alloc(bo, place, mem);
 809                 if (ret == -ENOSPC)
 810                         continue;
 811                 if (unlikely(ret))
 812                         goto error;
 813
 814                 ret = ttm_bo_add_move_fence(bo, man, *mem, ctx->no_wait_gpu);
 815                 if (unlikely(ret)) {
 816                         ttm_resource_free(bo, mem);
 817                         if (ret == -EBUSY)
 818                                 continue;
 819
 820                         goto error;
 821                 }
 822                 return 0;
 823         }
 824
 825         for (i = 0; i < placement->num_busy_placement; ++i) {
 826                 const struct ttm_place *place = &placement->busy_placement[i];
 827                 struct ttm_resource_manager *man;
 828
 829                 man = ttm_manager_type(bdev, place->mem_type);
 830                 if (!man || !ttm_resource_manager_used(man))
 831                         continue;
 832
 833                 type_found = true;
 834                 ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
 835                 if (likely(!ret))
 836                         return 0;
 837
 838                 if (ret && ret != -EBUSY)
 839                         goto error;
 840         }
 841
 842         ret = -ENOMEM;
 843         if (!type_found) {
 844                 pr_err(TTM_PFX "No compatible memory type found\n");
 845                 ret = -EINVAL;
 846         }
 847
 848 error:
 849         return ret;
 850 }
 851 EXPORT_SYMBOL(ttm_bo_mem_space);
 852
 853 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
 854                               struct ttm_placement *placement,
 855                               struct ttm_operation_ctx *ctx)
 856 {
 857         struct ttm_resource *mem;
 858         struct ttm_place hop;
 859         int ret;
 860
 861         dma_resv_assert_held(bo->base.resv);
 862
 863         /*
 864          * Determine where to move the buffer.
 865          *
 866          * If driver determines move is going to need
 867          * an extra step then it will return -EMULTIHOP
 868          * and the buffer will be moved to the temporary
 869          * stop and the driver will be called to make
 870          * the second hop.
 871          */
 872         ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
 873         if (ret)
 874                 return ret;
 875 bounce:
 876         ret = ttm_bo_handle_move_mem(bo, mem, false, ctx, &hop);
 877         if (ret == -EMULTIHOP) {
 878                 ret = ttm_bo_bounce_temp_buffer(bo, &mem, ctx, &hop);
 879                 if (ret)
 880                         goto out;
 881                 /* try and move to final place now. */
 882                 goto bounce;
 883         }
 884 out:
 885         if (ret)
 886                 ttm_resource_free(bo, &mem);
 887         return ret;
 888 }
 889
 890 int ttm_bo_validate(struct ttm_buffer_object *bo,
 891                     struct ttm_placement *placement,
 892                     struct ttm_operation_ctx *ctx)
 893 {
 894         int ret;
 895
 896         dma_resv_assert_held(bo->base.resv);
 897
 898         /*
 899          * Remove the backing store if no placement is given.
 900          */
 901         if (!placement->num_placement && !placement->num_busy_placement)
 902                 return ttm_bo_pipeline_gutting(bo);
 903
 904         /*
 905          * Check whether we need to move buffer.
 906          */
 907         if (!ttm_resource_compat(bo->resource, placement)) {
 908                 ret = ttm_bo_move_buffer(bo, placement, ctx);
 909                 if (ret)
 910                         return ret;
 911         }
 912         /*
 913          * We might need to add a TTM.
 914          */
 915         if (bo->resource->mem_type == TTM_PL_SYSTEM) {
 916                 ret = ttm_tt_create(bo, true);
 917                 if (ret)
 918                         return ret;
 919         }
 920         return 0;
 921 }
 922 EXPORT_SYMBOL(ttm_bo_validate);
 923
 924 int ttm_bo_init_reserved(struct ttm_device *bdev,
 925                          struct ttm_buffer_object *bo,
 926                          size_t size,
 927                          enum ttm_bo_type type,
 928                          struct ttm_placement *placement,
 929                          uint32_t page_alignment,
 930                          struct ttm_operation_ctx *ctx,
 931                          struct sg_table *sg,
 932                          struct dma_resv *resv,
 933                          void (*destroy) (struct ttm_buffer_object *))
 934 {
 935         static const struct ttm_place sys_mem = { .mem_type = TTM_PL_SYSTEM };
 936         bool locked;
 937         int ret;
 938
 939         bo->destroy = destroy;
 940         kref_init(&bo->kref);
 941         INIT_LIST_HEAD(&bo->ddestroy);
 942         bo->bdev = bdev;
 943         bo->type = type;
 944         bo->page_alignment = page_alignment;
 945         bo->pin_count = 0;
 946         bo->sg = sg;
 947         bo->bulk_move = NULL;
 948         if (resv) {
 949                 bo->base.resv = resv;
 950                 dma_resv_assert_held(bo->base.resv);
 951         } else {
 952                 bo->base.resv = &bo->base._resv;
 953         }
 954         atomic_inc(&ttm_glob.bo_count);
 955
 956         ret = ttm_resource_alloc(bo, &sys_mem, &bo->resource);
 957         if (unlikely(ret)) {
 958                 ttm_bo_put(bo);
 959                 return ret;
 960         }
 961
 962         /*
 963          * For ttm_bo_type_device buffers, allocate
 964          * address space from the device.
 965          */
 966         if (bo->type == ttm_bo_type_device ||
 967             bo->type == ttm_bo_type_sg)
 968                 ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
 969                                          bo->resource->num_pages);
 970
 971         /* passed reservation objects should already be locked,
 972          * since otherwise lockdep will be angered in radeon.
 973          */
 974         if (!resv) {
 975                 locked = dma_resv_trylock(bo->base.resv);
 976                 WARN_ON(!locked);
 977         }
 978
 979         if (likely(!ret))
 980                 ret = ttm_bo_validate(bo, placement, ctx);
 981
 982         if (unlikely(ret)) {
 983                 if (!resv)
 984                         ttm_bo_unreserve(bo);
 985
 986                 ttm_bo_put(bo);
 987                 return ret;
 988         }
 989
 990         return ret;
 991 }
 992 EXPORT_SYMBOL(ttm_bo_init_reserved);
 993
 994 int ttm_bo_init(struct ttm_device *bdev,
 995                 struct ttm_buffer_object *bo,
 996                 size_t size,
 997                 enum ttm_bo_type type,
 998                 struct ttm_placement *placement,
 999                 uint32_t page_alignment,
1000                 bool interruptible,
1001                 struct sg_table *sg,
1002                 struct dma_resv *resv,
1003                 void (*destroy) (struct ttm_buffer_object *))
1004 {
1005         struct ttm_operation_ctx ctx = { interruptible, false };
1006         int ret;
1007
1008         ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1009                                    page_alignment, &ctx, sg, resv, destroy);
1010         if (ret)
1011                 return ret;
1012
1013         if (!resv)
1014                 ttm_bo_unreserve(bo);
1015
1016         return 0;
1017 }
1018 EXPORT_SYMBOL(ttm_bo_init);
1019
1020 /*
1021  * buffer object vm functions.
1022  */
1023
1024 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1025 {
1026         struct ttm_device *bdev = bo->bdev;
1027
1028         drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1029         ttm_mem_io_free(bdev, bo->resource);
1030 }
1031 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1032
1033 int ttm_bo_wait(struct ttm_buffer_object *bo,
1034                 bool interruptible, bool no_wait)
1035 {
1036         long timeout = 15 * HZ;
1037
1038         if (no_wait) {
1039                 if (dma_resv_test_signaled(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP))
1040                         return 0;
1041                 else
1042                         return -EBUSY;
1043         }
1044
1045         timeout = dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP,
1046                                         interruptible, timeout);
1047         if (timeout < 0)
1048                 return timeout;
1049
1050         if (timeout == 0)
1051                 return -EBUSY;
1052
1053         return 0;
1054 }
1055 EXPORT_SYMBOL(ttm_bo_wait);
1056
1057 int ttm_bo_swapout(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx,
1058                    gfp_t gfp_flags)
1059 {
1060         struct ttm_place place;
1061         bool locked;
1062         int ret;
1063
1064         /*
1065          * While the bo may already reside in SYSTEM placement, set
1066          * SYSTEM as new placement to cover also the move further below.
1067          * The driver may use the fact that we're moving from SYSTEM
1068          * as an indication that we're about to swap out.
1069          */
1070         memset(&place, 0, sizeof(place));
1071         place.mem_type = bo->resource->mem_type;
1072         if (!ttm_bo_evict_swapout_allowable(bo, ctx, &place, &locked, NULL))
1073                 return -EBUSY;
1074
1075         if (!bo->ttm || !ttm_tt_is_populated(bo->ttm) ||
1076             bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL ||
1077             bo->ttm->page_flags & TTM_TT_FLAG_SWAPPED ||
1078             !ttm_bo_get_unless_zero(bo)) {
1079                 if (locked)
1080                         dma_resv_unlock(bo->base.resv);
1081                 return -EBUSY;
1082         }
1083
1084         if (bo->deleted) {
1085                 ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1086                 ttm_bo_put(bo);
1087                 return ret == -EBUSY ? -ENOSPC : ret;
1088         }
1089
1090         /* TODO: Cleanup the locking */
1091         spin_unlock(&bo->bdev->lru_lock);
1092
1093         /*
1094          * Move to system cached
1095          */
1096         if (bo->resource->mem_type != TTM_PL_SYSTEM) {
1097                 struct ttm_operation_ctx ctx = { false, false };
1098                 struct ttm_resource *evict_mem;
1099                 struct ttm_place hop;
1100
1101                 memset(&hop, 0, sizeof(hop));
1102                 place.mem_type = TTM_PL_SYSTEM;
1103                 ret = ttm_resource_alloc(bo, &place, &evict_mem);
1104                 if (unlikely(ret))
1105                         goto out;
1106
1107                 ret = ttm_bo_handle_move_mem(bo, evict_mem, true, &ctx, &hop);
1108                 if (unlikely(ret != 0)) {
1109                         WARN(ret == -EMULTIHOP, "Unexpected multihop in swaput - likely driver bug.\n");
1110                         goto out;
1111                 }
1112         }
1113
1114         /*
1115          * Make sure BO is idle.
1116          */
1117         ret = ttm_bo_wait(bo, false, false);
1118         if (unlikely(ret != 0))
1119                 goto out;
1120
1121         ttm_bo_unmap_virtual(bo);
1122
1123         /*
1124          * Swap out. Buffer will be swapped in again as soon as
1125          * anyone tries to access a ttm page.
1126          */
1127         if (bo->bdev->funcs->swap_notify)
1128                 bo->bdev->funcs->swap_notify(bo);
1129
1130         if (ttm_tt_is_populated(bo->ttm))
1131                 ret = ttm_tt_swapout(bo->bdev, bo->ttm, gfp_flags);
1132 out:
1133
1134         /*
1135          * Unreserve without putting on LRU to avoid swapping out an
1136          * already swapped buffer.
1137          */
1138         if (locked)
1139                 dma_resv_unlock(bo->base.resv);
1140         ttm_bo_put(bo);
1141         return ret == -EBUSY ? -ENOSPC : ret;
1142 }
1143
1144 void ttm_bo_tt_destroy(struct ttm_buffer_object *bo)
1145 {
1146         if (bo->ttm == NULL)
1147                 return;
1148
1149         ttm_tt_unpopulate(bo->bdev, bo->ttm);
1150         ttm_tt_destroy(bo->bdev, bo->ttm);
1151         bo->ttm = NULL;
1152 }