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 static void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
  73 {
  74         struct ttm_device *bdev = bo->bdev;
  75
  76         list_del_init(&bo->lru);
  77
  78         if (bdev->funcs->del_from_lru_notify)
  79                 bdev->funcs->del_from_lru_notify(bo);
  80 }
  81
  82 static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos,
  83                                      struct ttm_buffer_object *bo)
  84 {
  85         if (!pos->first)
  86                 pos->first = bo;
  87         pos->last = bo;
  88 }
  89
  90 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
  91                              struct ttm_resource *mem,
  92                              struct ttm_lru_bulk_move *bulk)
  93 {
  94         struct ttm_device *bdev = bo->bdev;
  95         struct ttm_resource_manager *man;
  96
  97         if (!bo->deleted)
  98                 dma_resv_assert_held(bo->base.resv);
  99
 100         if (bo->pin_count) {
 101                 ttm_bo_del_from_lru(bo);
 102                 return;
 103         }
 104
 105         if (!mem)
 106                 return;
 107
 108         man = ttm_manager_type(bdev, mem->mem_type);
 109         list_move_tail(&bo->lru, &man->lru[bo->priority]);
 110
 111         if (bdev->funcs->del_from_lru_notify)
 112                 bdev->funcs->del_from_lru_notify(bo);
 113
 114         if (bulk && !bo->pin_count) {
 115                 switch (bo->resource->mem_type) {
 116                 case TTM_PL_TT:
 117                         ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo);
 118                         break;
 119
 120                 case TTM_PL_VRAM:
 121                         ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo);
 122                         break;
 123                 }
 124         }
 125 }
 126 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
 127
 128 void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk)
 129 {
 130         unsigned i;
 131
 132         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
 133                 struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i];
 134                 struct ttm_resource_manager *man;
 135
 136                 if (!pos->first)
 137                         continue;
 138
 139                 dma_resv_assert_held(pos->first->base.resv);
 140                 dma_resv_assert_held(pos->last->base.resv);
 141
 142                 man = ttm_manager_type(pos->first->bdev, TTM_PL_TT);
 143                 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
 144                                     &pos->last->lru);
 145         }
 146
 147         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
 148                 struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i];
 149                 struct ttm_resource_manager *man;
 150
 151                 if (!pos->first)
 152                         continue;
 153
 154                 dma_resv_assert_held(pos->first->base.resv);
 155                 dma_resv_assert_held(pos->last->base.resv);
 156
 157                 man = ttm_manager_type(pos->first->bdev, TTM_PL_VRAM);
 158                 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
 159                                     &pos->last->lru);
 160         }
 161 }
 162 EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail);
 163
 164 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
 165                                   struct ttm_resource *mem, bool evict,
 166                                   struct ttm_operation_ctx *ctx,
 167                                   struct ttm_place *hop)
 168 {
 169         struct ttm_resource_manager *old_man, *new_man;
 170         struct ttm_device *bdev = bo->bdev;
 171         int ret;
 172
 173         old_man = ttm_manager_type(bdev, bo->resource->mem_type);
 174         new_man = ttm_manager_type(bdev, mem->mem_type);
 175
 176         ttm_bo_unmap_virtual(bo);
 177
 178         /*
 179          * Create and bind a ttm if required.
 180          */
 181
 182         if (new_man->use_tt) {
 183                 /* Zero init the new TTM structure if the old location should
 184                  * have used one as well.
 185                  */
 186                 ret = ttm_tt_create(bo, old_man->use_tt);
 187                 if (ret)
 188                         goto out_err;
 189
 190                 if (mem->mem_type != TTM_PL_SYSTEM) {
 191                         ret = ttm_tt_populate(bo->bdev, bo->ttm, ctx);
 192                         if (ret)
 193                                 goto out_err;
 194                 }
 195         }
 196
 197         ret = bdev->funcs->move(bo, evict, ctx, mem, hop);
 198         if (ret) {
 199                 if (ret == -EMULTIHOP)
 200                         return ret;
 201                 goto out_err;
 202         }
 203
 204         ctx->bytes_moved += bo->base.size;
 205         return 0;
 206
 207 out_err:
 208         new_man = ttm_manager_type(bdev, bo->resource->mem_type);
 209         if (!new_man->use_tt)
 210                 ttm_bo_tt_destroy(bo);
 211
 212         return ret;
 213 }
 214
 215 /*
 216  * Call bo::reserved.
 217  * Will release GPU memory type usage on destruction.
 218  * This is the place to put in driver specific hooks to release
 219  * driver private resources.
 220  * Will release the bo::reserved lock.
 221  */
 222
 223 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
 224 {
 225         if (bo->bdev->funcs->delete_mem_notify)
 226                 bo->bdev->funcs->delete_mem_notify(bo);
 227
 228         ttm_bo_tt_destroy(bo);
 229         ttm_resource_free(bo, &bo->resource);
 230 }
 231
 232 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
 233 {
 234         int r;
 235
 236         if (bo->base.resv == &bo->base._resv)
 237                 return 0;
 238
 239         BUG_ON(!dma_resv_trylock(&bo->base._resv));
 240
 241         r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
 242         dma_resv_unlock(&bo->base._resv);
 243         if (r)
 244                 return r;
 245
 246         if (bo->type != ttm_bo_type_sg) {
 247                 /* This works because the BO is about to be destroyed and nobody
 248                  * reference it any more. The only tricky case is the trylock on
 249                  * the resv object while holding the lru_lock.
 250                  */
 251                 spin_lock(&bo->bdev->lru_lock);
 252                 bo->base.resv = &bo->base._resv;
 253                 spin_unlock(&bo->bdev->lru_lock);
 254         }
 255
 256         return r;
 257 }
 258
 259 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
 260 {
 261         struct dma_resv *resv = &bo->base._resv;
 262         struct dma_resv_list *fobj;
 263         struct dma_fence *fence;
 264         int i;
 265
 266         rcu_read_lock();
 267         fobj = dma_resv_shared_list(resv);
 268         fence = dma_resv_excl_fence(resv);
 269         if (fence && !fence->ops->signaled)
 270                 dma_fence_enable_sw_signaling(fence);
 271
 272         for (i = 0; fobj && i < fobj->shared_count; ++i) {
 273                 fence = rcu_dereference(fobj->shared[i]);
 274
 275                 if (!fence->ops->signaled)
 276                         dma_fence_enable_sw_signaling(fence);
 277         }
 278         rcu_read_unlock();
 279 }
 280
 281 /**
 282  * ttm_bo_cleanup_refs
 283  * If bo idle, remove from lru lists, and unref.
 284  * If not idle, block if possible.
 285  *
 286  * Must be called with lru_lock and reservation held, this function
 287  * will drop the lru lock and optionally the reservation lock before returning.
 288  *
 289  * @bo:                    The buffer object to clean-up
 290  * @interruptible:         Any sleeps should occur interruptibly.
 291  * @no_wait_gpu:           Never wait for gpu. Return -EBUSY instead.
 292  * @unlock_resv:           Unlock the reservation lock as well.
 293  */
 294
 295 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
 296                                bool interruptible, bool no_wait_gpu,
 297                                bool unlock_resv)
 298 {
 299         struct dma_resv *resv = &bo->base._resv;
 300         int ret;
 301
 302         if (dma_resv_test_signaled(resv, true))
 303                 ret = 0;
 304         else
 305                 ret = -EBUSY;
 306
 307         if (ret && !no_wait_gpu) {
 308                 long lret;
 309
 310                 if (unlock_resv)
 311                         dma_resv_unlock(bo->base.resv);
 312                 spin_unlock(&bo->bdev->lru_lock);
 313
 314                 lret = dma_resv_wait_timeout(resv, true, interruptible,
 315                                              30 * HZ);
 316
 317                 if (lret < 0)
 318                         return lret;
 319                 else if (lret == 0)
 320                         return -EBUSY;
 321
 322                 spin_lock(&bo->bdev->lru_lock);
 323                 if (unlock_resv && !dma_resv_trylock(bo->base.resv)) {
 324                         /*
 325                          * We raced, and lost, someone else holds the reservation now,
 326                          * and is probably busy in ttm_bo_cleanup_memtype_use.
 327                          *
 328                          * Even if it's not the case, because we finished waiting any
 329                          * delayed destruction would succeed, so just return success
 330                          * here.
 331                          */
 332                         spin_unlock(&bo->bdev->lru_lock);
 333                         return 0;
 334                 }
 335                 ret = 0;
 336         }
 337
 338         if (ret || unlikely(list_empty(&bo->ddestroy))) {
 339                 if (unlock_resv)
 340                         dma_resv_unlock(bo->base.resv);
 341                 spin_unlock(&bo->bdev->lru_lock);
 342                 return ret;
 343         }
 344
 345         ttm_bo_del_from_lru(bo);
 346         list_del_init(&bo->ddestroy);
 347         spin_unlock(&bo->bdev->lru_lock);
 348         ttm_bo_cleanup_memtype_use(bo);
 349
 350         if (unlock_resv)
 351                 dma_resv_unlock(bo->base.resv);
 352
 353         ttm_bo_put(bo);
 354
 355         return 0;
 356 }
 357
 358 /*
 359  * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
 360  * encountered buffers.
 361  */
 362 bool ttm_bo_delayed_delete(struct ttm_device *bdev, bool remove_all)
 363 {
 364         struct list_head removed;
 365         bool empty;
 366
 367         INIT_LIST_HEAD(&removed);
 368
 369         spin_lock(&bdev->lru_lock);
 370         while (!list_empty(&bdev->ddestroy)) {
 371                 struct ttm_buffer_object *bo;
 372
 373                 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
 374                                       ddestroy);
 375                 list_move_tail(&bo->ddestroy, &removed);
 376                 if (!ttm_bo_get_unless_zero(bo))
 377                         continue;
 378
 379                 if (remove_all || bo->base.resv != &bo->base._resv) {
 380                         spin_unlock(&bdev->lru_lock);
 381                         dma_resv_lock(bo->base.resv, NULL);
 382
 383                         spin_lock(&bdev->lru_lock);
 384                         ttm_bo_cleanup_refs(bo, false, !remove_all, true);
 385
 386                 } else if (dma_resv_trylock(bo->base.resv)) {
 387                         ttm_bo_cleanup_refs(bo, false, !remove_all, true);
 388                 } else {
 389                         spin_unlock(&bdev->lru_lock);
 390                 }
 391
 392                 ttm_bo_put(bo);
 393                 spin_lock(&bdev->lru_lock);
 394         }
 395         list_splice_tail(&removed, &bdev->ddestroy);
 396         empty = list_empty(&bdev->ddestroy);
 397         spin_unlock(&bdev->lru_lock);
 398
 399         return empty;
 400 }
 401
 402 static void ttm_bo_release(struct kref *kref)
 403 {
 404         struct ttm_buffer_object *bo =
 405             container_of(kref, struct ttm_buffer_object, kref);
 406         struct ttm_device *bdev = bo->bdev;
 407         int ret;
 408
 409         WARN_ON_ONCE(bo->pin_count);
 410
 411         if (!bo->deleted) {
 412                 ret = ttm_bo_individualize_resv(bo);
 413                 if (ret) {
 414                         /* Last resort, if we fail to allocate memory for the
 415                          * fences block for the BO to become idle
 416                          */
 417                         dma_resv_wait_timeout(bo->base.resv, true, false,
 418                                               30 * HZ);
 419                 }
 420
 421                 if (bo->bdev->funcs->release_notify)
 422                         bo->bdev->funcs->release_notify(bo);
 423
 424                 drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
 425                 ttm_mem_io_free(bdev, bo->resource);
 426         }
 427
 428         if (!dma_resv_test_signaled(bo->base.resv, true) ||
 429             !dma_resv_trylock(bo->base.resv)) {
 430                 /* The BO is not idle, resurrect it for delayed destroy */
 431                 ttm_bo_flush_all_fences(bo);
 432                 bo->deleted = true;
 433
 434                 spin_lock(&bo->bdev->lru_lock);
 435
 436                 /*
 437                  * Make pinned bos immediately available to
 438                  * shrinkers, now that they are queued for
 439                  * destruction.
 440                  *
 441                  * FIXME: QXL is triggering this. Can be removed when the
 442                  * driver is fixed.
 443                  */
 444                 if (bo->pin_count) {
 445                         bo->pin_count = 0;
 446                         ttm_bo_move_to_lru_tail(bo, bo->resource, NULL);
 447                 }
 448
 449                 kref_init(&bo->kref);
 450                 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
 451                 spin_unlock(&bo->bdev->lru_lock);
 452
 453                 schedule_delayed_work(&bdev->wq,
 454                                       ((HZ / 100) < 1) ? 1 : HZ / 100);
 455                 return;
 456         }
 457
 458         spin_lock(&bo->bdev->lru_lock);
 459         ttm_bo_del_from_lru(bo);
 460         list_del(&bo->ddestroy);
 461         spin_unlock(&bo->bdev->lru_lock);
 462
 463         ttm_bo_cleanup_memtype_use(bo);
 464         dma_resv_unlock(bo->base.resv);
 465
 466         atomic_dec(&ttm_glob.bo_count);
 467         dma_fence_put(bo->moving);
 468         bo->destroy(bo);
 469 }
 470
 471 void ttm_bo_put(struct ttm_buffer_object *bo)
 472 {
 473         kref_put(&bo->kref, ttm_bo_release);
 474 }
 475 EXPORT_SYMBOL(ttm_bo_put);
 476
 477 int ttm_bo_lock_delayed_workqueue(struct ttm_device *bdev)
 478 {
 479         return cancel_delayed_work_sync(&bdev->wq);
 480 }
 481 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
 482
 483 void ttm_bo_unlock_delayed_workqueue(struct ttm_device *bdev, int resched)
 484 {
 485         if (resched)
 486                 schedule_delayed_work(&bdev->wq,
 487                                       ((HZ / 100) < 1) ? 1 : HZ / 100);
 488 }
 489 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
 490
 491 static int ttm_bo_bounce_temp_buffer(struct ttm_buffer_object *bo,
 492                                      struct ttm_resource **mem,
 493                                      struct ttm_operation_ctx *ctx,
 494                                      struct ttm_place *hop)
 495 {
 496         struct ttm_placement hop_placement;
 497         struct ttm_resource *hop_mem;
 498         int ret;
 499
 500         hop_placement.num_placement = hop_placement.num_busy_placement = 1;
 501         hop_placement.placement = hop_placement.busy_placement = hop;
 502
 503         /* find space in the bounce domain */
 504         ret = ttm_bo_mem_space(bo, &hop_placement, &hop_mem, ctx);
 505         if (ret)
 506                 return ret;
 507         /* move to the bounce domain */
 508         ret = ttm_bo_handle_move_mem(bo, hop_mem, false, ctx, NULL);
 509         if (ret) {
 510                 ttm_resource_free(bo, &hop_mem);
 511                 return ret;
 512         }
 513         return 0;
 514 }
 515
 516 static int ttm_bo_evict(struct ttm_buffer_object *bo,
 517                         struct ttm_operation_ctx *ctx)
 518 {
 519         struct ttm_device *bdev = bo->bdev;
 520         struct ttm_resource *evict_mem;
 521         struct ttm_placement placement;
 522         struct ttm_place hop;
 523         int ret = 0;
 524
 525         memset(&hop, 0, sizeof(hop));
 526
 527         dma_resv_assert_held(bo->base.resv);
 528
 529         placement.num_placement = 0;
 530         placement.num_busy_placement = 0;
 531         bdev->funcs->evict_flags(bo, &placement);
 532
 533         if (!placement.num_placement && !placement.num_busy_placement) {
 534                 ret = ttm_bo_wait(bo, true, false);
 535                 if (ret)
 536                         return ret;
 537
 538                 /*
 539                  * Since we've already synced, this frees backing store
 540                  * immediately.
 541                  */
 542                 return ttm_bo_pipeline_gutting(bo);
 543         }
 544
 545         ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
 546         if (ret) {
 547                 if (ret != -ERESTARTSYS) {
 548                         pr_err("Failed to find memory space for buffer 0x%p eviction\n",
 549                                bo);
 550                         ttm_bo_mem_space_debug(bo, &placement);
 551                 }
 552                 goto out;
 553         }
 554
 555 bounce:
 556         ret = ttm_bo_handle_move_mem(bo, evict_mem, true, ctx, &hop);
 557         if (ret == -EMULTIHOP) {
 558                 ret = ttm_bo_bounce_temp_buffer(bo, &evict_mem, ctx, &hop);
 559                 if (ret) {
 560                         pr_err("Buffer eviction failed\n");
 561                         ttm_resource_free(bo, &evict_mem);
 562                         goto out;
 563                 }
 564                 /* try and move to final place now. */
 565                 goto bounce;
 566         }
 567 out:
 568         return ret;
 569 }
 570
 571 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
 572                               const struct ttm_place *place)
 573 {
 574         dma_resv_assert_held(bo->base.resv);
 575         if (bo->resource->mem_type == TTM_PL_SYSTEM)
 576                 return true;
 577
 578         /* Don't evict this BO if it's outside of the
 579          * requested placement range
 580          */
 581         if (place->fpfn >= (bo->resource->start + bo->resource->num_pages) ||
 582             (place->lpfn && place->lpfn <= bo->resource->start))
 583                 return false;
 584
 585         return true;
 586 }
 587 EXPORT_SYMBOL(ttm_bo_eviction_valuable);
 588
 589 /*
 590  * Check the target bo is allowable to be evicted or swapout, including cases:
 591  *
 592  * a. if share same reservation object with ctx->resv, have assumption
 593  * reservation objects should already be locked, so not lock again and
 594  * return true directly when either the opreation allow_reserved_eviction
 595  * or the target bo already is in delayed free list;
 596  *
 597  * b. Otherwise, trylock it.
 598  */
 599 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
 600                                            struct ttm_operation_ctx *ctx,
 601                                            const struct ttm_place *place,
 602                                            bool *locked, bool *busy)
 603 {
 604         bool ret = false;
 605
 606         if (bo->base.resv == ctx->resv) {
 607                 dma_resv_assert_held(bo->base.resv);
 608                 if (ctx->allow_res_evict)
 609                         ret = true;
 610                 *locked = false;
 611                 if (busy)
 612                         *busy = false;
 613         } else {
 614                 ret = dma_resv_trylock(bo->base.resv);
 615                 *locked = ret;
 616                 if (busy)
 617                         *busy = !ret;
 618         }
 619
 620         if (ret && place && !bo->bdev->funcs->eviction_valuable(bo, place)) {
 621                 ret = false;
 622                 if (*locked) {
 623                         dma_resv_unlock(bo->base.resv);
 624                         *locked = false;
 625                 }
 626         }
 627
 628         return ret;
 629 }
 630
 631 /**
 632  * ttm_mem_evict_wait_busy - wait for a busy BO to become available
 633  *
 634  * @busy_bo: BO which couldn't be locked with trylock
 635  * @ctx: operation context
 636  * @ticket: acquire ticket
 637  *
 638  * Try to lock a busy buffer object to avoid failing eviction.
 639  */
 640 static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
 641                                    struct ttm_operation_ctx *ctx,
 642                                    struct ww_acquire_ctx *ticket)
 643 {
 644         int r;
 645
 646         if (!busy_bo || !ticket)
 647                 return -EBUSY;
 648
 649         if (ctx->interruptible)
 650                 r = dma_resv_lock_interruptible(busy_bo->base.resv,
 651                                                           ticket);
 652         else
 653                 r = dma_resv_lock(busy_bo->base.resv, ticket);
 654
 655         /*
 656          * TODO: It would be better to keep the BO locked until allocation is at
 657          * least tried one more time, but that would mean a much larger rework
 658          * of TTM.
 659          */
 660         if (!r)
 661                 dma_resv_unlock(busy_bo->base.resv);
 662
 663         return r == -EDEADLK ? -EBUSY : r;
 664 }
 665
 666 int ttm_mem_evict_first(struct ttm_device *bdev,
 667                         struct ttm_resource_manager *man,
 668                         const struct ttm_place *place,
 669                         struct ttm_operation_ctx *ctx,
 670                         struct ww_acquire_ctx *ticket)
 671 {
 672         struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
 673         bool locked = false;
 674         unsigned i;
 675         int ret;
 676
 677         spin_lock(&bdev->lru_lock);
 678         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
 679                 list_for_each_entry(bo, &man->lru[i], lru) {
 680                         bool busy;
 681
 682                         if (!ttm_bo_evict_swapout_allowable(bo, ctx, place,
 683                                                             &locked, &busy)) {
 684                                 if (busy && !busy_bo && ticket !=
 685                                     dma_resv_locking_ctx(bo->base.resv))
 686                                         busy_bo = bo;
 687                                 continue;
 688                         }
 689
 690                         if (!ttm_bo_get_unless_zero(bo)) {
 691                                 if (locked)
 692                                         dma_resv_unlock(bo->base.resv);
 693                                 continue;
 694                         }
 695                         break;
 696                 }
 697
 698                 /* If the inner loop terminated early, we have our candidate */
 699                 if (&bo->lru != &man->lru[i])
 700                         break;
 701
 702                 bo = NULL;
 703         }
 704
 705         if (!bo) {
 706                 if (busy_bo && !ttm_bo_get_unless_zero(busy_bo))
 707                         busy_bo = NULL;
 708                 spin_unlock(&bdev->lru_lock);
 709                 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
 710                 if (busy_bo)
 711                         ttm_bo_put(busy_bo);
 712                 return ret;
 713         }
 714
 715         if (bo->deleted) {
 716                 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
 717                                           ctx->no_wait_gpu, locked);
 718                 ttm_bo_put(bo);
 719                 return ret;
 720         }
 721
 722         spin_unlock(&bdev->lru_lock);
 723
 724         ret = ttm_bo_evict(bo, ctx);
 725         if (locked)
 726                 ttm_bo_unreserve(bo);
 727
 728         ttm_bo_put(bo);
 729         return ret;
 730 }
 731
 732 /*
 733  * Add the last move fence to the BO and reserve a new shared slot. We only use
 734  * a shared slot to avoid unecessary sync and rely on the subsequent bo move to
 735  * either stall or use an exclusive fence respectively set bo->moving.
 736  */
 737 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
 738                                  struct ttm_resource_manager *man,
 739                                  struct ttm_resource *mem,
 740                                  bool no_wait_gpu)
 741 {
 742         struct dma_fence *fence;
 743         int ret;
 744
 745         spin_lock(&man->move_lock);
 746         fence = dma_fence_get(man->move);
 747         spin_unlock(&man->move_lock);
 748
 749         if (!fence)
 750                 return 0;
 751
 752         if (no_wait_gpu) {
 753                 ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
 754                 dma_fence_put(fence);
 755                 return ret;
 756         }
 757
 758         dma_resv_add_shared_fence(bo->base.resv, fence);
 759
 760         ret = dma_resv_reserve_shared(bo->base.resv, 1);
 761         if (unlikely(ret)) {
 762                 dma_fence_put(fence);
 763                 return ret;
 764         }
 765
 766         dma_fence_put(bo->moving);
 767         bo->moving = fence;
 768         return 0;
 769 }
 770
 771 /*
 772  * Repeatedly evict memory from the LRU for @mem_type until we create enough
 773  * space, or we've evicted everything and there isn't enough space.
 774  */
 775 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
 776                                   const struct ttm_place *place,
 777                                   struct ttm_resource **mem,
 778                                   struct ttm_operation_ctx *ctx)
 779 {
 780         struct ttm_device *bdev = bo->bdev;
 781         struct ttm_resource_manager *man;
 782         struct ww_acquire_ctx *ticket;
 783         int ret;
 784
 785         man = ttm_manager_type(bdev, place->mem_type);
 786         ticket = dma_resv_locking_ctx(bo->base.resv);
 787         do {
 788                 ret = ttm_resource_alloc(bo, place, mem);
 789                 if (likely(!ret))
 790                         break;
 791                 if (unlikely(ret != -ENOSPC))
 792                         return ret;
 793                 ret = ttm_mem_evict_first(bdev, man, place, ctx,
 794                                           ticket);
 795                 if (unlikely(ret != 0))
 796                         return ret;
 797         } while (1);
 798
 799         return ttm_bo_add_move_fence(bo, man, *mem, ctx->no_wait_gpu);
 800 }
 801
 802 /*
 803  * Creates space for memory region @mem according to its type.
 804  *
 805  * This function first searches for free space in compatible memory types in
 806  * the priority order defined by the driver.  If free space isn't found, then
 807  * ttm_bo_mem_force_space is attempted in priority order to evict and find
 808  * space.
 809  */
 810 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
 811                         struct ttm_placement *placement,
 812                         struct ttm_resource **mem,
 813                         struct ttm_operation_ctx *ctx)
 814 {
 815         struct ttm_device *bdev = bo->bdev;
 816         bool type_found = false;
 817         int i, ret;
 818
 819         ret = dma_resv_reserve_shared(bo->base.resv, 1);
 820         if (unlikely(ret))
 821                 return ret;
 822
 823         for (i = 0; i < placement->num_placement; ++i) {
 824                 const struct ttm_place *place = &placement->placement[i];
 825                 struct ttm_resource_manager *man;
 826
 827                 man = ttm_manager_type(bdev, place->mem_type);
 828                 if (!man || !ttm_resource_manager_used(man))
 829                         continue;
 830
 831                 type_found = true;
 832                 ret = ttm_resource_alloc(bo, place, mem);
 833                 if (ret == -ENOSPC)
 834                         continue;
 835                 if (unlikely(ret))
 836                         goto error;
 837
 838                 ret = ttm_bo_add_move_fence(bo, man, *mem, ctx->no_wait_gpu);
 839                 if (unlikely(ret)) {
 840                         ttm_resource_free(bo, mem);
 841                         if (ret == -EBUSY)
 842                                 continue;
 843
 844                         goto error;
 845                 }
 846                 return 0;
 847         }
 848
 849         for (i = 0; i < placement->num_busy_placement; ++i) {
 850                 const struct ttm_place *place = &placement->busy_placement[i];
 851                 struct ttm_resource_manager *man;
 852
 853                 man = ttm_manager_type(bdev, place->mem_type);
 854                 if (!man || !ttm_resource_manager_used(man))
 855                         continue;
 856
 857                 type_found = true;
 858                 ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
 859                 if (likely(!ret))
 860                         return 0;
 861
 862                 if (ret && ret != -EBUSY)
 863                         goto error;
 864         }
 865
 866         ret = -ENOMEM;
 867         if (!type_found) {
 868                 pr_err(TTM_PFX "No compatible memory type found\n");
 869                 ret = -EINVAL;
 870         }
 871
 872 error:
 873         if (bo->resource->mem_type == TTM_PL_SYSTEM && !bo->pin_count)
 874                 ttm_bo_move_to_lru_tail_unlocked(bo);
 875
 876         return ret;
 877 }
 878 EXPORT_SYMBOL(ttm_bo_mem_space);
 879
 880 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
 881                               struct ttm_placement *placement,
 882                               struct ttm_operation_ctx *ctx)
 883 {
 884         struct ttm_resource *mem;
 885         struct ttm_place hop;
 886         int ret;
 887
 888         dma_resv_assert_held(bo->base.resv);
 889
 890         /*
 891          * Determine where to move the buffer.
 892          *
 893          * If driver determines move is going to need
 894          * an extra step then it will return -EMULTIHOP
 895          * and the buffer will be moved to the temporary
 896          * stop and the driver will be called to make
 897          * the second hop.
 898          */
 899         ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
 900         if (ret)
 901                 return ret;
 902 bounce:
 903         ret = ttm_bo_handle_move_mem(bo, mem, false, ctx, &hop);
 904         if (ret == -EMULTIHOP) {
 905                 ret = ttm_bo_bounce_temp_buffer(bo, &mem, ctx, &hop);
 906                 if (ret)
 907                         goto out;
 908                 /* try and move to final place now. */
 909                 goto bounce;
 910         }
 911 out:
 912         if (ret)
 913                 ttm_resource_free(bo, &mem);
 914         return ret;
 915 }
 916
 917 static bool ttm_bo_places_compat(const struct ttm_place *places,
 918                                  unsigned num_placement,
 919                                  struct ttm_resource *mem,
 920                                  uint32_t *new_flags)
 921 {
 922         unsigned i;
 923
 924         if (mem->placement & TTM_PL_FLAG_TEMPORARY)
 925                 return false;
 926
 927         for (i = 0; i < num_placement; i++) {
 928                 const struct ttm_place *heap = &places[i];
 929
 930                 if ((mem->start < heap->fpfn ||
 931                      (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
 932                         continue;
 933
 934                 *new_flags = heap->flags;
 935                 if ((mem->mem_type == heap->mem_type) &&
 936                     (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
 937                      (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
 938                         return true;
 939         }
 940         return false;
 941 }
 942
 943 bool ttm_bo_mem_compat(struct ttm_placement *placement,
 944                        struct ttm_resource *mem,
 945                        uint32_t *new_flags)
 946 {
 947         if (ttm_bo_places_compat(placement->placement, placement->num_placement,
 948                                  mem, new_flags))
 949                 return true;
 950
 951         if ((placement->busy_placement != placement->placement ||
 952              placement->num_busy_placement > placement->num_placement) &&
 953             ttm_bo_places_compat(placement->busy_placement,
 954                                  placement->num_busy_placement,
 955                                  mem, new_flags))
 956                 return true;
 957
 958         return false;
 959 }
 960 EXPORT_SYMBOL(ttm_bo_mem_compat);
 961
 962 int ttm_bo_validate(struct ttm_buffer_object *bo,
 963                     struct ttm_placement *placement,
 964                     struct ttm_operation_ctx *ctx)
 965 {
 966         int ret;
 967         uint32_t new_flags;
 968
 969         dma_resv_assert_held(bo->base.resv);
 970
 971         /*
 972          * Remove the backing store if no placement is given.
 973          */
 974         if (!placement->num_placement && !placement->num_busy_placement)
 975                 return ttm_bo_pipeline_gutting(bo);
 976
 977         /*
 978          * Check whether we need to move buffer.
 979          */
 980         if (!ttm_bo_mem_compat(placement, bo->resource, &new_flags)) {
 981                 ret = ttm_bo_move_buffer(bo, placement, ctx);
 982                 if (ret)
 983                         return ret;
 984         }
 985         /*
 986          * We might need to add a TTM.
 987          */
 988         if (bo->resource->mem_type == TTM_PL_SYSTEM) {
 989                 ret = ttm_tt_create(bo, true);
 990                 if (ret)
 991                         return ret;
 992         }
 993         return 0;
 994 }
 995 EXPORT_SYMBOL(ttm_bo_validate);
 996
 997 int ttm_bo_init_reserved(struct ttm_device *bdev,
 998                          struct ttm_buffer_object *bo,
 999                          size_t size,
1000                          enum ttm_bo_type type,
1001                          struct ttm_placement *placement,
1002                          uint32_t page_alignment,
1003                          struct ttm_operation_ctx *ctx,
1004                          struct sg_table *sg,
1005                          struct dma_resv *resv,
1006                          void (*destroy) (struct ttm_buffer_object *))
1007 {
1008         static const struct ttm_place sys_mem = { .mem_type = TTM_PL_SYSTEM };
1009         bool locked;
1010         int ret;
1011
1012         bo->destroy = destroy ? destroy : ttm_bo_default_destroy;
1013
1014         kref_init(&bo->kref);
1015         INIT_LIST_HEAD(&bo->lru);
1016         INIT_LIST_HEAD(&bo->ddestroy);
1017         bo->bdev = bdev;
1018         bo->type = type;
1019         bo->page_alignment = page_alignment;
1020         bo->moving = NULL;
1021         bo->pin_count = 0;
1022         bo->sg = sg;
1023         if (resv) {
1024                 bo->base.resv = resv;
1025                 dma_resv_assert_held(bo->base.resv);
1026         } else {
1027                 bo->base.resv = &bo->base._resv;
1028         }
1029         atomic_inc(&ttm_glob.bo_count);
1030
1031         ret = ttm_resource_alloc(bo, &sys_mem, &bo->resource);
1032         if (unlikely(ret)) {
1033                 ttm_bo_put(bo);
1034                 return ret;
1035         }
1036
1037         /*
1038          * For ttm_bo_type_device buffers, allocate
1039          * address space from the device.
1040          */
1041         if (bo->type == ttm_bo_type_device ||
1042             bo->type == ttm_bo_type_sg)
1043                 ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
1044                                          bo->resource->num_pages);
1045
1046         /* passed reservation objects should already be locked,
1047          * since otherwise lockdep will be angered in radeon.
1048          */
1049         if (!resv) {
1050                 locked = dma_resv_trylock(bo->base.resv);
1051                 WARN_ON(!locked);
1052         }
1053
1054         if (likely(!ret))
1055                 ret = ttm_bo_validate(bo, placement, ctx);
1056
1057         if (unlikely(ret)) {
1058                 if (!resv)
1059                         ttm_bo_unreserve(bo);
1060
1061                 ttm_bo_put(bo);
1062                 return ret;
1063         }
1064
1065         ttm_bo_move_to_lru_tail_unlocked(bo);
1066
1067         return ret;
1068 }
1069 EXPORT_SYMBOL(ttm_bo_init_reserved);
1070
1071 int ttm_bo_init(struct ttm_device *bdev,
1072                 struct ttm_buffer_object *bo,
1073                 size_t size,
1074                 enum ttm_bo_type type,
1075                 struct ttm_placement *placement,
1076                 uint32_t page_alignment,
1077                 bool interruptible,
1078                 struct sg_table *sg,
1079                 struct dma_resv *resv,
1080                 void (*destroy) (struct ttm_buffer_object *))
1081 {
1082         struct ttm_operation_ctx ctx = { interruptible, false };
1083         int ret;
1084
1085         ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1086                                    page_alignment, &ctx, sg, resv, destroy);
1087         if (ret)
1088                 return ret;
1089
1090         if (!resv)
1091                 ttm_bo_unreserve(bo);
1092
1093         return 0;
1094 }
1095 EXPORT_SYMBOL(ttm_bo_init);
1096
1097 /*
1098  * buffer object vm functions.
1099  */
1100
1101 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1102 {
1103         struct ttm_device *bdev = bo->bdev;
1104
1105         drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1106         ttm_mem_io_free(bdev, bo->resource);
1107 }
1108 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1109
1110 int ttm_bo_wait(struct ttm_buffer_object *bo,
1111                 bool interruptible, bool no_wait)
1112 {
1113         long timeout = 15 * HZ;
1114
1115         if (no_wait) {
1116                 if (dma_resv_test_signaled(bo->base.resv, true))
1117                         return 0;
1118                 else
1119                         return -EBUSY;
1120         }
1121
1122         timeout = dma_resv_wait_timeout(bo->base.resv, true, interruptible,
1123                                         timeout);
1124         if (timeout < 0)
1125                 return timeout;
1126
1127         if (timeout == 0)
1128                 return -EBUSY;
1129
1130         dma_resv_add_excl_fence(bo->base.resv, NULL);
1131         return 0;
1132 }
1133 EXPORT_SYMBOL(ttm_bo_wait);
1134
1135 int ttm_bo_swapout(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx,
1136                    gfp_t gfp_flags)
1137 {
1138         struct ttm_place place;
1139         bool locked;
1140         int ret;
1141
1142         /*
1143          * While the bo may already reside in SYSTEM placement, set
1144          * SYSTEM as new placement to cover also the move further below.
1145          * The driver may use the fact that we're moving from SYSTEM
1146          * as an indication that we're about to swap out.
1147          */
1148         memset(&place, 0, sizeof(place));
1149         place.mem_type = TTM_PL_SYSTEM;
1150         if (!ttm_bo_evict_swapout_allowable(bo, ctx, &place, &locked, NULL))
1151                 return -EBUSY;
1152
1153         if (!bo->ttm || !ttm_tt_is_populated(bo->ttm) ||
1154             bo->ttm->page_flags & TTM_PAGE_FLAG_SG ||
1155             bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED ||
1156             !ttm_bo_get_unless_zero(bo)) {
1157                 if (locked)
1158                         dma_resv_unlock(bo->base.resv);
1159                 return -EBUSY;
1160         }
1161
1162         if (bo->deleted) {
1163                 ttm_bo_cleanup_refs(bo, false, false, locked);
1164                 ttm_bo_put(bo);
1165                 return 0;
1166         }
1167
1168         ttm_bo_del_from_lru(bo);
1169         /* TODO: Cleanup the locking */
1170         spin_unlock(&bo->bdev->lru_lock);
1171
1172         /*
1173          * Move to system cached
1174          */
1175         if (bo->resource->mem_type != TTM_PL_SYSTEM) {
1176                 struct ttm_operation_ctx ctx = { false, false };
1177                 struct ttm_resource *evict_mem;
1178                 struct ttm_place hop;
1179
1180                 memset(&hop, 0, sizeof(hop));
1181                 ret = ttm_resource_alloc(bo, &place, &evict_mem);
1182                 if (unlikely(ret))
1183                         goto out;
1184
1185                 ret = ttm_bo_handle_move_mem(bo, evict_mem, true, &ctx, &hop);
1186                 if (unlikely(ret != 0)) {
1187                         WARN(ret == -EMULTIHOP, "Unexpected multihop in swaput - likely driver bug.\n");
1188                         goto out;
1189                 }
1190         }
1191
1192         /*
1193          * Make sure BO is idle.
1194          */
1195         ret = ttm_bo_wait(bo, false, false);
1196         if (unlikely(ret != 0))
1197                 goto out;
1198
1199         ttm_bo_unmap_virtual(bo);
1200
1201         /*
1202          * Swap out. Buffer will be swapped in again as soon as
1203          * anyone tries to access a ttm page.
1204          */
1205         if (bo->bdev->funcs->swap_notify)
1206                 bo->bdev->funcs->swap_notify(bo);
1207
1208         if (ttm_tt_is_populated(bo->ttm))
1209                 ret = ttm_tt_swapout(bo->bdev, bo->ttm, gfp_flags);
1210 out:
1211
1212         /*
1213          * Unreserve without putting on LRU to avoid swapping out an
1214          * already swapped buffer.
1215          */
1216         if (locked)
1217                 dma_resv_unlock(bo->base.resv);
1218         ttm_bo_put(bo);
1219         return ret;
1220 }
1221
1222 void ttm_bo_tt_destroy(struct ttm_buffer_object *bo)
1223 {
1224         if (bo->ttm == NULL)
1225                 return;
1226
1227         ttm_tt_destroy(bo->bdev, bo->ttm);
1228         bo->ttm = NULL;
1229 }