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 /* default destructor */
  48 static void ttm_bo_default_destroy(struct ttm_buffer_object *bo)
  49 {
  50         kfree(bo);
  51 }
  52
  53 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
  54                                         struct ttm_placement *placement)
  55 {
  56         struct drm_printer p = drm_debug_printer(TTM_PFX);
  57         struct ttm_resource_manager *man;
  58         int i, mem_type;
  59
  60         drm_printf(&p, "No space for %p (%lu pages, %zuK, %zuM)\n",
  61                    bo, bo->resource->num_pages, bo->base.size >> 10,
  62                    bo->base.size >> 20);
  63         for (i = 0; i < placement->num_placement; i++) {
  64                 mem_type = placement->placement[i].mem_type;
  65                 drm_printf(&p, "  placement[%d]=0x%08X (%d)\n",
  66                            i, placement->placement[i].flags, mem_type);
  67                 man = ttm_manager_type(bo->bdev, mem_type);
  68                 ttm_resource_manager_debug(man, &p);
  69         }
  70 }
  71
  72 /**
  73  * ttm_bo_move_to_lru_tail
  74  *
  75  * @bo: The buffer object.
  76  *
  77  * Move this BO to the tail of all lru lists used to lookup and reserve an
  78  * object. This function must be called with struct ttm_global::lru_lock
  79  * held, and is used to make a BO less likely to be considered for eviction.
  80  */
  81 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo)
  82 {
  83         dma_resv_assert_held(bo->base.resv);
  84
  85         if (bo->resource)
  86                 ttm_resource_move_to_lru_tail(bo->resource);
  87 }
  88 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
  89
  90 /**
  91  * ttm_bo_set_bulk_move - update BOs bulk move object
  92  *
  93  * @bo: The buffer object.
  94  *
  95  * Update the BOs bulk move object, making sure that resources are added/removed
  96  * as well. A bulk move allows to move many resource on the LRU at once,
  97  * resulting in much less overhead of maintaining the LRU.
  98  * The only requirement is that the resources stay together on the LRU and are
  99  * never separated. This is enforces by setting the bulk_move structure on a BO.
 100  * ttm_lru_bulk_move_tail() should be used to move all resources to the tail of
 101  * their LRU list.
 102  */
 103 void ttm_bo_set_bulk_move(struct ttm_buffer_object *bo,
 104                           struct ttm_lru_bulk_move *bulk)
 105 {
 106         dma_resv_assert_held(bo->base.resv);
 107
 108         if (bo->bulk_move == bulk)
 109                 return;
 110
 111         spin_lock(&bo->bdev->lru_lock);
 112         if (bo->bulk_move && bo->resource)
 113                 ttm_lru_bulk_move_del(bo->bulk_move, bo->resource);
 114         bo->bulk_move = bulk;
 115         if (bo->bulk_move && bo->resource)
 116                 ttm_lru_bulk_move_add(bo->bulk_move, bo->resource);
 117         spin_unlock(&bo->bdev->lru_lock);
 118 }
 119 EXPORT_SYMBOL(ttm_bo_set_bulk_move);
 120
 121 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
 122                                   struct ttm_resource *mem, bool evict,
 123                                   struct ttm_operation_ctx *ctx,
 124                                   struct ttm_place *hop)
 125 {
 126         struct ttm_resource_manager *old_man, *new_man;
 127         struct ttm_device *bdev = bo->bdev;
 128         int ret;
 129
 130         old_man = ttm_manager_type(bdev, bo->resource->mem_type);
 131         new_man = ttm_manager_type(bdev, mem->mem_type);
 132
 133         ttm_bo_unmap_virtual(bo);
 134
 135         /*
 136          * Create and bind a ttm if required.
 137          */
 138
 139         if (new_man->use_tt) {
 140                 /* Zero init the new TTM structure if the old location should
 141                  * have used one as well.
 142                  */
 143                 ret = ttm_tt_create(bo, old_man->use_tt);
 144                 if (ret)
 145                         goto out_err;
 146
 147                 if (mem->mem_type != TTM_PL_SYSTEM) {
 148                         ret = ttm_tt_populate(bo->bdev, bo->ttm, ctx);
 149                         if (ret)
 150                                 goto out_err;
 151                 }
 152         }
 153
 154         ret = dma_resv_reserve_fences(bo->base.resv, 1);
 155         if (ret)
 156                 goto out_err;
 157
 158         ret = bdev->funcs->move(bo, evict, ctx, mem, hop);
 159         if (ret) {
 160                 if (ret == -EMULTIHOP)
 161                         return ret;
 162                 goto out_err;
 163         }
 164
 165         ctx->bytes_moved += bo->base.size;
 166         return 0;
 167
 168 out_err:
 169         new_man = ttm_manager_type(bdev, bo->resource->mem_type);
 170         if (!new_man->use_tt)
 171                 ttm_bo_tt_destroy(bo);
 172
 173         return ret;
 174 }
 175
 176 /*
 177  * Call bo::reserved.
 178  * Will release GPU memory type usage on destruction.
 179  * This is the place to put in driver specific hooks to release
 180  * driver private resources.
 181  * Will release the bo::reserved lock.
 182  */
 183
 184 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
 185 {
 186         if (bo->bdev->funcs->delete_mem_notify)
 187                 bo->bdev->funcs->delete_mem_notify(bo);
 188
 189         ttm_bo_tt_destroy(bo);
 190         ttm_resource_free(bo, &bo->resource);
 191 }
 192
 193 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
 194 {
 195         int r;
 196
 197         if (bo->base.resv == &bo->base._resv)
 198                 return 0;
 199
 200         BUG_ON(!dma_resv_trylock(&bo->base._resv));
 201
 202         r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
 203         dma_resv_unlock(&bo->base._resv);
 204         if (r)
 205                 return r;
 206
 207         if (bo->type != ttm_bo_type_sg) {
 208                 /* This works because the BO is about to be destroyed and nobody
 209                  * reference it any more. The only tricky case is the trylock on
 210                  * the resv object while holding the lru_lock.
 211                  */
 212                 spin_lock(&bo->bdev->lru_lock);
 213                 bo->base.resv = &bo->base._resv;
 214                 spin_unlock(&bo->bdev->lru_lock);
 215         }
 216
 217         return r;
 218 }
 219
 220 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
 221 {
 222         struct dma_resv *resv = &bo->base._resv;
 223         struct dma_resv_iter cursor;
 224         struct dma_fence *fence;
 225
 226         dma_resv_iter_begin(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP);
 227         dma_resv_for_each_fence_unlocked(&cursor, fence) {
 228                 if (!fence->ops->signaled)
 229                         dma_fence_enable_sw_signaling(fence);
 230         }
 231         dma_resv_iter_end(&cursor);
 232 }
 233
 234 /**
 235  * ttm_bo_cleanup_refs
 236  * If bo idle, remove from lru lists, and unref.
 237  * If not idle, block if possible.
 238  *
 239  * Must be called with lru_lock and reservation held, this function
 240  * will drop the lru lock and optionally the reservation lock before returning.
 241  *
 242  * @bo:                    The buffer object to clean-up
 243  * @interruptible:         Any sleeps should occur interruptibly.
 244  * @no_wait_gpu:           Never wait for gpu. Return -EBUSY instead.
 245  * @unlock_resv:           Unlock the reservation lock as well.
 246  */
 247
 248 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
 249                                bool interruptible, bool no_wait_gpu,
 250                                bool unlock_resv)
 251 {
 252         struct dma_resv *resv = &bo->base._resv;
 253         int ret;
 254
 255         if (dma_resv_test_signaled(resv, DMA_RESV_USAGE_BOOKKEEP))
 256                 ret = 0;
 257         else
 258                 ret = -EBUSY;
 259
 260         if (ret && !no_wait_gpu) {
 261                 long lret;
 262
 263                 if (unlock_resv)
 264                         dma_resv_unlock(bo->base.resv);
 265                 spin_unlock(&bo->bdev->lru_lock);
 266
 267                 lret = dma_resv_wait_timeout(resv, DMA_RESV_USAGE_BOOKKEEP,
 268                                              interruptible,
 269                                              30 * HZ);
 270
 271                 if (lret < 0)
 272                         return lret;
 273                 else if (lret == 0)
 274                         return -EBUSY;
 275
 276                 spin_lock(&bo->bdev->lru_lock);
 277                 if (unlock_resv && !dma_resv_trylock(bo->base.resv)) {
 278                         /*
 279                          * We raced, and lost, someone else holds the reservation now,
 280                          * and is probably busy in ttm_bo_cleanup_memtype_use.
 281                          *
 282                          * Even if it's not the case, because we finished waiting any
 283                          * delayed destruction would succeed, so just return success
 284                          * here.
 285                          */
 286                         spin_unlock(&bo->bdev->lru_lock);
 287                         return 0;
 288                 }
 289                 ret = 0;
 290         }
 291
 292         if (ret || unlikely(list_empty(&bo->ddestroy))) {
 293                 if (unlock_resv)
 294                         dma_resv_unlock(bo->base.resv);
 295                 spin_unlock(&bo->bdev->lru_lock);
 296                 return ret;
 297         }
 298
 299         list_del_init(&bo->ddestroy);
 300         spin_unlock(&bo->bdev->lru_lock);
 301         ttm_bo_cleanup_memtype_use(bo);
 302
 303         if (unlock_resv)
 304                 dma_resv_unlock(bo->base.resv);
 305
 306         ttm_bo_put(bo);
 307
 308         return 0;
 309 }
 310
 311 /*
 312  * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
 313  * encountered buffers.
 314  */
 315 bool ttm_bo_delayed_delete(struct ttm_device *bdev, bool remove_all)
 316 {
 317         struct list_head removed;
 318         bool empty;
 319
 320         INIT_LIST_HEAD(&removed);
 321
 322         spin_lock(&bdev->lru_lock);
 323         while (!list_empty(&bdev->ddestroy)) {
 324                 struct ttm_buffer_object *bo;
 325
 326                 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
 327                                       ddestroy);
 328                 list_move_tail(&bo->ddestroy, &removed);
 329                 if (!ttm_bo_get_unless_zero(bo))
 330                         continue;
 331
 332                 if (remove_all || bo->base.resv != &bo->base._resv) {
 333                         spin_unlock(&bdev->lru_lock);
 334                         dma_resv_lock(bo->base.resv, NULL);
 335
 336                         spin_lock(&bdev->lru_lock);
 337                         ttm_bo_cleanup_refs(bo, false, !remove_all, true);
 338
 339                 } else if (dma_resv_trylock(bo->base.resv)) {
 340                         ttm_bo_cleanup_refs(bo, false, !remove_all, true);
 341                 } else {
 342                         spin_unlock(&bdev->lru_lock);
 343                 }
 344
 345                 ttm_bo_put(bo);
 346                 spin_lock(&bdev->lru_lock);
 347         }
 348         list_splice_tail(&removed, &bdev->ddestroy);
 349         empty = list_empty(&bdev->ddestroy);
 350         spin_unlock(&bdev->lru_lock);
 351
 352         return empty;
 353 }
 354
 355 static void ttm_bo_release(struct kref *kref)
 356 {
 357         struct ttm_buffer_object *bo =
 358             container_of(kref, struct ttm_buffer_object, kref);
 359         struct ttm_device *bdev = bo->bdev;
 360         int ret;
 361
 362         WARN_ON_ONCE(bo->pin_count);
 363         WARN_ON_ONCE(bo->bulk_move);
 364
 365         if (!bo->deleted) {
 366                 ret = ttm_bo_individualize_resv(bo);
 367                 if (ret) {
 368                         /* Last resort, if we fail to allocate memory for the
 369                          * fences block for the BO to become idle
 370                          */
 371                         dma_resv_wait_timeout(bo->base.resv,
 372                                               DMA_RESV_USAGE_BOOKKEEP, false,
 373                                               30 * HZ);
 374                 }
 375
 376                 if (bo->bdev->funcs->release_notify)
 377                         bo->bdev->funcs->release_notify(bo);
 378
 379                 drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
 380                 ttm_mem_io_free(bdev, bo->resource);
 381         }
 382
 383         if (!dma_resv_test_signaled(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP) ||
 384             !dma_resv_trylock(bo->base.resv)) {
 385                 /* The BO is not idle, resurrect it for delayed destroy */
 386                 ttm_bo_flush_all_fences(bo);
 387                 bo->deleted = true;
 388
 389                 spin_lock(&bo->bdev->lru_lock);
 390
 391                 /*
 392                  * Make pinned bos immediately available to
 393                  * shrinkers, now that they are queued for
 394                  * destruction.
 395                  *
 396                  * FIXME: QXL is triggering this. Can be removed when the
 397                  * driver is fixed.
 398                  */
 399                 if (bo->pin_count) {
 400                         bo->pin_count = 0;
 401                         ttm_resource_move_to_lru_tail(bo->resource);
 402                 }
 403
 404                 kref_init(&bo->kref);
 405                 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
 406                 spin_unlock(&bo->bdev->lru_lock);
 407
 408                 schedule_delayed_work(&bdev->wq,
 409                                       ((HZ / 100) < 1) ? 1 : HZ / 100);
 410                 return;
 411         }
 412
 413         spin_lock(&bo->bdev->lru_lock);
 414         list_del(&bo->ddestroy);
 415         spin_unlock(&bo->bdev->lru_lock);
 416
 417         ttm_bo_cleanup_memtype_use(bo);
 418         dma_resv_unlock(bo->base.resv);
 419
 420         atomic_dec(&ttm_glob.bo_count);
 421         bo->destroy(bo);
 422 }
 423
 424 void ttm_bo_put(struct ttm_buffer_object *bo)
 425 {
 426         kref_put(&bo->kref, ttm_bo_release);
 427 }
 428 EXPORT_SYMBOL(ttm_bo_put);
 429
 430 int ttm_bo_lock_delayed_workqueue(struct ttm_device *bdev)
 431 {
 432         return cancel_delayed_work_sync(&bdev->wq);
 433 }
 434 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
 435
 436 void ttm_bo_unlock_delayed_workqueue(struct ttm_device *bdev, int resched)
 437 {
 438         if (resched)
 439                 schedule_delayed_work(&bdev->wq,
 440                                       ((HZ / 100) < 1) ? 1 : HZ / 100);
 441 }
 442 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
 443
 444 static int ttm_bo_bounce_temp_buffer(struct ttm_buffer_object *bo,
 445                                      struct ttm_resource **mem,
 446                                      struct ttm_operation_ctx *ctx,
 447                                      struct ttm_place *hop)
 448 {
 449         struct ttm_placement hop_placement;
 450         struct ttm_resource *hop_mem;
 451         int ret;
 452
 453         hop_placement.num_placement = hop_placement.num_busy_placement = 1;
 454         hop_placement.placement = hop_placement.busy_placement = hop;
 455
 456         /* find space in the bounce domain */
 457         ret = ttm_bo_mem_space(bo, &hop_placement, &hop_mem, ctx);
 458         if (ret)
 459                 return ret;
 460         /* move to the bounce domain */
 461         ret = ttm_bo_handle_move_mem(bo, hop_mem, false, ctx, NULL);
 462         if (ret) {
 463                 ttm_resource_free(bo, &hop_mem);
 464                 return ret;
 465         }
 466         return 0;
 467 }
 468
 469 static int ttm_bo_evict(struct ttm_buffer_object *bo,
 470                         struct ttm_operation_ctx *ctx)
 471 {
 472         struct ttm_device *bdev = bo->bdev;
 473         struct ttm_resource *evict_mem;
 474         struct ttm_placement placement;
 475         struct ttm_place hop;
 476         int ret = 0;
 477
 478         memset(&hop, 0, sizeof(hop));
 479
 480         dma_resv_assert_held(bo->base.resv);
 481
 482         placement.num_placement = 0;
 483         placement.num_busy_placement = 0;
 484         bdev->funcs->evict_flags(bo, &placement);
 485
 486         if (!placement.num_placement && !placement.num_busy_placement) {
 487                 ret = ttm_bo_wait(bo, true, false);
 488                 if (ret)
 489                         return ret;
 490
 491                 /*
 492                  * Since we've already synced, this frees backing store
 493                  * immediately.
 494                  */
 495                 return ttm_bo_pipeline_gutting(bo);
 496         }
 497
 498         ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
 499         if (ret) {
 500                 if (ret != -ERESTARTSYS) {
 501                         pr_err("Failed to find memory space for buffer 0x%p eviction\n",
 502                                bo);
 503                         ttm_bo_mem_space_debug(bo, &placement);
 504                 }
 505                 goto out;
 506         }
 507
 508 bounce:
 509         ret = ttm_bo_handle_move_mem(bo, evict_mem, true, ctx, &hop);
 510         if (ret == -EMULTIHOP) {
 511                 ret = ttm_bo_bounce_temp_buffer(bo, &evict_mem, ctx, &hop);
 512                 if (ret) {
 513                         pr_err("Buffer eviction failed\n");
 514                         ttm_resource_free(bo, &evict_mem);
 515                         goto out;
 516                 }
 517                 /* try and move to final place now. */
 518                 goto bounce;
 519         }
 520 out:
 521         return ret;
 522 }
 523
 524 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
 525                               const struct ttm_place *place)
 526 {
 527         dma_resv_assert_held(bo->base.resv);
 528         if (bo->resource->mem_type == TTM_PL_SYSTEM)
 529                 return true;
 530
 531         /* Don't evict this BO if it's outside of the
 532          * requested placement range
 533          */
 534         if (place->fpfn >= (bo->resource->start + bo->resource->num_pages) ||
 535             (place->lpfn && place->lpfn <= bo->resource->start))
 536                 return false;
 537
 538         return true;
 539 }
 540 EXPORT_SYMBOL(ttm_bo_eviction_valuable);
 541
 542 /*
 543  * Check the target bo is allowable to be evicted or swapout, including cases:
 544  *
 545  * a. if share same reservation object with ctx->resv, have assumption
 546  * reservation objects should already be locked, so not lock again and
 547  * return true directly when either the opreation allow_reserved_eviction
 548  * or the target bo already is in delayed free list;
 549  *
 550  * b. Otherwise, trylock it.
 551  */
 552 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
 553                                            struct ttm_operation_ctx *ctx,
 554                                            const struct ttm_place *place,
 555                                            bool *locked, bool *busy)
 556 {
 557         bool ret = false;
 558
 559         if (bo->base.resv == ctx->resv) {
 560                 dma_resv_assert_held(bo->base.resv);
 561                 if (ctx->allow_res_evict)
 562                         ret = true;
 563                 *locked = false;
 564                 if (busy)
 565                         *busy = false;
 566         } else {
 567                 ret = dma_resv_trylock(bo->base.resv);
 568                 *locked = ret;
 569                 if (busy)
 570                         *busy = !ret;
 571         }
 572
 573         if (ret && place && (bo->resource->mem_type != place->mem_type ||
 574                 !bo->bdev->funcs->eviction_valuable(bo, place))) {
 575                 ret = false;
 576                 if (*locked) {
 577                         dma_resv_unlock(bo->base.resv);
 578                         *locked = false;
 579                 }
 580         }
 581
 582         return ret;
 583 }
 584
 585 /**
 586  * ttm_mem_evict_wait_busy - wait for a busy BO to become available
 587  *
 588  * @busy_bo: BO which couldn't be locked with trylock
 589  * @ctx: operation context
 590  * @ticket: acquire ticket
 591  *
 592  * Try to lock a busy buffer object to avoid failing eviction.
 593  */
 594 static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
 595                                    struct ttm_operation_ctx *ctx,
 596                                    struct ww_acquire_ctx *ticket)
 597 {
 598         int r;
 599
 600         if (!busy_bo || !ticket)
 601                 return -EBUSY;
 602
 603         if (ctx->interruptible)
 604                 r = dma_resv_lock_interruptible(busy_bo->base.resv,
 605                                                           ticket);
 606         else
 607                 r = dma_resv_lock(busy_bo->base.resv, ticket);
 608
 609         /*
 610          * TODO: It would be better to keep the BO locked until allocation is at
 611          * least tried one more time, but that would mean a much larger rework
 612          * of TTM.
 613          */
 614         if (!r)
 615                 dma_resv_unlock(busy_bo->base.resv);
 616
 617         return r == -EDEADLK ? -EBUSY : r;
 618 }
 619
 620 int ttm_mem_evict_first(struct ttm_device *bdev,
 621                         struct ttm_resource_manager *man,
 622                         const struct ttm_place *place,
 623                         struct ttm_operation_ctx *ctx,
 624                         struct ww_acquire_ctx *ticket)
 625 {
 626         struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
 627         struct ttm_resource_cursor cursor;
 628         struct ttm_resource *res;
 629         bool locked = false;
 630         int ret;
 631
 632         spin_lock(&bdev->lru_lock);
 633         ttm_resource_manager_for_each_res(man, &cursor, res) {
 634                 bool busy;
 635
 636                 if (!ttm_bo_evict_swapout_allowable(res->bo, ctx, place,
 637                                                     &locked, &busy)) {
 638                         if (busy && !busy_bo && ticket !=
 639                             dma_resv_locking_ctx(res->bo->base.resv))
 640                                 busy_bo = res->bo;
 641                         continue;
 642                 }
 643
 644                 if (ttm_bo_get_unless_zero(res->bo)) {
 645                         bo = res->bo;
 646                         break;
 647                 }
 648                 if (locked)
 649                         dma_resv_unlock(res->bo->base.resv);
 650         }
 651
 652         if (!bo) {
 653                 if (busy_bo && !ttm_bo_get_unless_zero(busy_bo))
 654                         busy_bo = NULL;
 655                 spin_unlock(&bdev->lru_lock);
 656                 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
 657                 if (busy_bo)
 658                         ttm_bo_put(busy_bo);
 659                 return ret;
 660         }
 661
 662         if (bo->deleted) {
 663                 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
 664                                           ctx->no_wait_gpu, locked);
 665                 ttm_bo_put(bo);
 666                 return ret;
 667         }
 668
 669         spin_unlock(&bdev->lru_lock);
 670
 671         ret = ttm_bo_evict(bo, ctx);
 672         if (locked)
 673                 ttm_bo_unreserve(bo);
 674         else
 675                 ttm_bo_move_to_lru_tail_unlocked(bo);
 676
 677         ttm_bo_put(bo);
 678         return ret;
 679 }
 680
 681 /**
 682  * ttm_bo_pin - Pin the buffer object.
 683  * @bo: The buffer object to pin
 684  *
 685  * Make sure the buffer is not evicted any more during memory pressure.
 686  * @bo must be unpinned again by calling ttm_bo_unpin().
 687  */
 688 void ttm_bo_pin(struct ttm_buffer_object *bo)
 689 {
 690         dma_resv_assert_held(bo->base.resv);
 691         WARN_ON_ONCE(!kref_read(&bo->kref));
 692         if (!(bo->pin_count++) && bo->bulk_move && bo->resource)
 693                 ttm_lru_bulk_move_del(bo->bulk_move, bo->resource);
 694 }
 695 EXPORT_SYMBOL(ttm_bo_pin);
 696
 697 /**
 698  * ttm_bo_unpin - Unpin the buffer object.
 699  * @bo: The buffer object to unpin
 700  *
 701  * Allows the buffer object to be evicted again during memory pressure.
 702  */
 703 void ttm_bo_unpin(struct ttm_buffer_object *bo)
 704 {
 705         dma_resv_assert_held(bo->base.resv);
 706         WARN_ON_ONCE(!kref_read(&bo->kref));
 707         if (WARN_ON_ONCE(!bo->pin_count))
 708                 return;
 709
 710         if (!(--bo->pin_count) && bo->bulk_move && bo->resource)
 711                 ttm_lru_bulk_move_add(bo->bulk_move, bo->resource);
 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 ? destroy : ttm_bo_default_destroy;
 940
 941         kref_init(&bo->kref);
 942         INIT_LIST_HEAD(&bo->ddestroy);
 943         bo->bdev = bdev;
 944         bo->type = type;
 945         bo->page_alignment = page_alignment;
 946         bo->pin_count = 0;
 947         bo->sg = sg;
 948         bo->bulk_move = NULL;
 949         if (resv) {
 950                 bo->base.resv = resv;
 951                 dma_resv_assert_held(bo->base.resv);
 952         } else {
 953                 bo->base.resv = &bo->base._resv;
 954         }
 955         atomic_inc(&ttm_glob.bo_count);
 956
 957         ret = ttm_resource_alloc(bo, &sys_mem, &bo->resource);
 958         if (unlikely(ret)) {
 959                 ttm_bo_put(bo);
 960                 return ret;
 961         }
 962
 963         /*
 964          * For ttm_bo_type_device buffers, allocate
 965          * address space from the device.
 966          */
 967         if (bo->type == ttm_bo_type_device ||
 968             bo->type == ttm_bo_type_sg)
 969                 ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
 970                                          bo->resource->num_pages);
 971
 972         /* passed reservation objects should already be locked,
 973          * since otherwise lockdep will be angered in radeon.
 974          */
 975         if (!resv) {
 976                 locked = dma_resv_trylock(bo->base.resv);
 977                 WARN_ON(!locked);
 978         }
 979
 980         if (likely(!ret))
 981                 ret = ttm_bo_validate(bo, placement, ctx);
 982
 983         if (unlikely(ret)) {
 984                 if (!resv)
 985                         ttm_bo_unreserve(bo);
 986
 987                 ttm_bo_put(bo);
 988                 return ret;
 989         }
 990
 991         return ret;
 992 }
 993 EXPORT_SYMBOL(ttm_bo_init_reserved);
 994
 995 int ttm_bo_init(struct ttm_device *bdev,
 996                 struct ttm_buffer_object *bo,
 997                 size_t size,
 998                 enum ttm_bo_type type,
 999                 struct ttm_placement *placement,
1000                 uint32_t page_alignment,
1001                 bool interruptible,
1002                 struct sg_table *sg,
1003                 struct dma_resv *resv,
1004                 void (*destroy) (struct ttm_buffer_object *))
1005 {
1006         struct ttm_operation_ctx ctx = { interruptible, false };
1007         int ret;
1008
1009         ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1010                                    page_alignment, &ctx, sg, resv, destroy);
1011         if (ret)
1012                 return ret;
1013
1014         if (!resv)
1015                 ttm_bo_unreserve(bo);
1016
1017         return 0;
1018 }
1019 EXPORT_SYMBOL(ttm_bo_init);
1020
1021 /*
1022  * buffer object vm functions.
1023  */
1024
1025 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1026 {
1027         struct ttm_device *bdev = bo->bdev;
1028
1029         drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1030         ttm_mem_io_free(bdev, bo->resource);
1031 }
1032 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1033
1034 int ttm_bo_wait(struct ttm_buffer_object *bo,
1035                 bool interruptible, bool no_wait)
1036 {
1037         long timeout = 15 * HZ;
1038
1039         if (no_wait) {
1040                 if (dma_resv_test_signaled(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP))
1041                         return 0;
1042                 else
1043                         return -EBUSY;
1044         }
1045
1046         timeout = dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP,
1047                                         interruptible, timeout);
1048         if (timeout < 0)
1049                 return timeout;
1050
1051         if (timeout == 0)
1052                 return -EBUSY;
1053
1054         return 0;
1055 }
1056 EXPORT_SYMBOL(ttm_bo_wait);
1057
1058 int ttm_bo_swapout(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx,
1059                    gfp_t gfp_flags)
1060 {
1061         struct ttm_place place;
1062         bool locked;
1063         int ret;
1064
1065         /*
1066          * While the bo may already reside in SYSTEM placement, set
1067          * SYSTEM as new placement to cover also the move further below.
1068          * The driver may use the fact that we're moving from SYSTEM
1069          * as an indication that we're about to swap out.
1070          */
1071         memset(&place, 0, sizeof(place));
1072         place.mem_type = bo->resource->mem_type;
1073         if (!ttm_bo_evict_swapout_allowable(bo, ctx, &place, &locked, NULL))
1074                 return -EBUSY;
1075
1076         if (!bo->ttm || !ttm_tt_is_populated(bo->ttm) ||
1077             bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL ||
1078             bo->ttm->page_flags & TTM_TT_FLAG_SWAPPED ||
1079             !ttm_bo_get_unless_zero(bo)) {
1080                 if (locked)
1081                         dma_resv_unlock(bo->base.resv);
1082                 return -EBUSY;
1083         }
1084
1085         if (bo->deleted) {
1086                 ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1087                 ttm_bo_put(bo);
1088                 return ret == -EBUSY ? -ENOSPC : ret;
1089         }
1090
1091         /* TODO: Cleanup the locking */
1092         spin_unlock(&bo->bdev->lru_lock);
1093
1094         /*
1095          * Move to system cached
1096          */
1097         if (bo->resource->mem_type != TTM_PL_SYSTEM) {
1098                 struct ttm_operation_ctx ctx = { false, false };
1099                 struct ttm_resource *evict_mem;
1100                 struct ttm_place hop;
1101
1102                 memset(&hop, 0, sizeof(hop));
1103                 place.mem_type = TTM_PL_SYSTEM;
1104                 ret = ttm_resource_alloc(bo, &place, &evict_mem);
1105                 if (unlikely(ret))
1106                         goto out;
1107
1108                 ret = ttm_bo_handle_move_mem(bo, evict_mem, true, &ctx, &hop);
1109                 if (unlikely(ret != 0)) {
1110                         WARN(ret == -EMULTIHOP, "Unexpected multihop in swaput - likely driver bug.\n");
1111                         goto out;
1112                 }
1113         }
1114
1115         /*
1116          * Make sure BO is idle.
1117          */
1118         ret = ttm_bo_wait(bo, false, false);
1119         if (unlikely(ret != 0))
1120                 goto out;
1121
1122         ttm_bo_unmap_virtual(bo);
1123
1124         /*
1125          * Swap out. Buffer will be swapped in again as soon as
1126          * anyone tries to access a ttm page.
1127          */
1128         if (bo->bdev->funcs->swap_notify)
1129                 bo->bdev->funcs->swap_notify(bo);
1130
1131         if (ttm_tt_is_populated(bo->ttm))
1132                 ret = ttm_tt_swapout(bo->bdev, bo->ttm, gfp_flags);
1133 out:
1134
1135         /*
1136          * Unreserve without putting on LRU to avoid swapping out an
1137          * already swapped buffer.
1138          */
1139         if (locked)
1140                 dma_resv_unlock(bo->base.resv);
1141         ttm_bo_put(bo);
1142         return ret == -EBUSY ? -ENOSPC : ret;
1143 }
1144
1145 void ttm_bo_tt_destroy(struct ttm_buffer_object *bo)
1146 {
1147         if (bo->ttm == NULL)
1148                 return;
1149
1150         ttm_tt_unpopulate(bo->bdev, bo->ttm);
1151         ttm_tt_destroy(bo->bdev, bo->ttm);
1152         bo->ttm = NULL;
1153 }