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->mem.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->swap);
  77         list_del_init(&bo->lru);
  78
  79         if (bdev->funcs->del_from_lru_notify)
  80                 bdev->funcs->del_from_lru_notify(bo);
  81 }
  82
  83 static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos,
  84                                      struct ttm_buffer_object *bo)
  85 {
  86         if (!pos->first)
  87                 pos->first = bo;
  88         pos->last = bo;
  89 }
  90
  91 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
  92                              struct ttm_resource *mem,
  93                              struct ttm_lru_bulk_move *bulk)
  94 {
  95         struct ttm_device *bdev = bo->bdev;
  96         struct ttm_resource_manager *man;
  97
  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         man = ttm_manager_type(bdev, mem->mem_type);
 106         list_move_tail(&bo->lru, &man->lru[bo->priority]);
 107         if (man->use_tt && bo->ttm &&
 108             !(bo->ttm->page_flags & (TTM_PAGE_FLAG_SG |
 109                                      TTM_PAGE_FLAG_SWAPPED))) {
 110                 struct list_head *swap;
 111
 112                 swap = &ttm_glob.swap_lru[bo->priority];
 113                 list_move_tail(&bo->swap, swap);
 114         } else {
 115                 list_del_init(&bo->swap);
 116         }
 117
 118         if (bdev->funcs->del_from_lru_notify)
 119                 bdev->funcs->del_from_lru_notify(bo);
 120
 121         if (bulk && !bo->pin_count) {
 122                 switch (bo->mem.mem_type) {
 123                 case TTM_PL_TT:
 124                         ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo);
 125                         break;
 126
 127                 case TTM_PL_VRAM:
 128                         ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo);
 129                         break;
 130                 }
 131                 if (bo->ttm && !(bo->ttm->page_flags &
 132                                  (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED)))
 133                         ttm_bo_bulk_move_set_pos(&bulk->swap[bo->priority], bo);
 134         }
 135 }
 136 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
 137
 138 void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk)
 139 {
 140         unsigned i;
 141
 142         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
 143                 struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i];
 144                 struct ttm_resource_manager *man;
 145
 146                 if (!pos->first)
 147                         continue;
 148
 149                 dma_resv_assert_held(pos->first->base.resv);
 150                 dma_resv_assert_held(pos->last->base.resv);
 151
 152                 man = ttm_manager_type(pos->first->bdev, TTM_PL_TT);
 153                 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
 154                                     &pos->last->lru);
 155         }
 156
 157         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
 158                 struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i];
 159                 struct ttm_resource_manager *man;
 160
 161                 if (!pos->first)
 162                         continue;
 163
 164                 dma_resv_assert_held(pos->first->base.resv);
 165                 dma_resv_assert_held(pos->last->base.resv);
 166
 167                 man = ttm_manager_type(pos->first->bdev, TTM_PL_VRAM);
 168                 list_bulk_move_tail(&man->lru[i], &pos->first->lru,
 169                                     &pos->last->lru);
 170         }
 171
 172         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
 173                 struct ttm_lru_bulk_move_pos *pos = &bulk->swap[i];
 174                 struct list_head *lru;
 175
 176                 if (!pos->first)
 177                         continue;
 178
 179                 dma_resv_assert_held(pos->first->base.resv);
 180                 dma_resv_assert_held(pos->last->base.resv);
 181
 182                 lru = &ttm_glob.swap_lru[i];
 183                 list_bulk_move_tail(lru, &pos->first->swap, &pos->last->swap);
 184         }
 185 }
 186 EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail);
 187
 188 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
 189                                   struct ttm_resource *mem, bool evict,
 190                                   struct ttm_operation_ctx *ctx,
 191                                   struct ttm_place *hop)
 192 {
 193         struct ttm_device *bdev = bo->bdev;
 194         struct ttm_resource_manager *old_man = ttm_manager_type(bdev, bo->mem.mem_type);
 195         struct ttm_resource_manager *new_man = ttm_manager_type(bdev, mem->mem_type);
 196         int ret;
 197
 198         ttm_bo_unmap_virtual(bo);
 199
 200         /*
 201          * Create and bind a ttm if required.
 202          */
 203
 204         if (new_man->use_tt) {
 205                 /* Zero init the new TTM structure if the old location should
 206                  * have used one as well.
 207                  */
 208                 ret = ttm_tt_create(bo, old_man->use_tt);
 209                 if (ret)
 210                         goto out_err;
 211
 212                 if (mem->mem_type != TTM_PL_SYSTEM) {
 213                         ret = ttm_tt_populate(bo->bdev, bo->ttm, ctx);
 214                         if (ret)
 215                                 goto out_err;
 216                 }
 217         }
 218
 219         ret = bdev->funcs->move(bo, evict, ctx, mem, hop);
 220         if (ret) {
 221                 if (ret == -EMULTIHOP)
 222                         return ret;
 223                 goto out_err;
 224         }
 225
 226         ctx->bytes_moved += bo->base.size;
 227         return 0;
 228
 229 out_err:
 230         new_man = ttm_manager_type(bdev, bo->mem.mem_type);
 231         if (!new_man->use_tt)
 232                 ttm_bo_tt_destroy(bo);
 233
 234         return ret;
 235 }
 236
 237 /*
 238  * Call bo::reserved.
 239  * Will release GPU memory type usage on destruction.
 240  * This is the place to put in driver specific hooks to release
 241  * driver private resources.
 242  * Will release the bo::reserved lock.
 243  */
 244
 245 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
 246 {
 247         if (bo->bdev->funcs->delete_mem_notify)
 248                 bo->bdev->funcs->delete_mem_notify(bo);
 249
 250         ttm_bo_tt_destroy(bo);
 251         ttm_resource_free(bo, &bo->mem);
 252 }
 253
 254 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
 255 {
 256         int r;
 257
 258         if (bo->base.resv == &bo->base._resv)
 259                 return 0;
 260
 261         BUG_ON(!dma_resv_trylock(&bo->base._resv));
 262
 263         r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
 264         dma_resv_unlock(&bo->base._resv);
 265         if (r)
 266                 return r;
 267
 268         if (bo->type != ttm_bo_type_sg) {
 269                 /* This works because the BO is about to be destroyed and nobody
 270                  * reference it any more. The only tricky case is the trylock on
 271                  * the resv object while holding the lru_lock.
 272                  */
 273                 spin_lock(&ttm_glob.lru_lock);
 274                 bo->base.resv = &bo->base._resv;
 275                 spin_unlock(&ttm_glob.lru_lock);
 276         }
 277
 278         return r;
 279 }
 280
 281 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
 282 {
 283         struct dma_resv *resv = &bo->base._resv;
 284         struct dma_resv_list *fobj;
 285         struct dma_fence *fence;
 286         int i;
 287
 288         rcu_read_lock();
 289         fobj = rcu_dereference(resv->fence);
 290         fence = rcu_dereference(resv->fence_excl);
 291         if (fence && !fence->ops->signaled)
 292                 dma_fence_enable_sw_signaling(fence);
 293
 294         for (i = 0; fobj && i < fobj->shared_count; ++i) {
 295                 fence = rcu_dereference(fobj->shared[i]);
 296
 297                 if (!fence->ops->signaled)
 298                         dma_fence_enable_sw_signaling(fence);
 299         }
 300         rcu_read_unlock();
 301 }
 302
 303 /**
 304  * function ttm_bo_cleanup_refs
 305  * If bo idle, remove from lru lists, and unref.
 306  * If not idle, block if possible.
 307  *
 308  * Must be called with lru_lock and reservation held, this function
 309  * will drop the lru lock and optionally the reservation lock before returning.
 310  *
 311  * @bo:                    The buffer object to clean-up
 312  * @interruptible:         Any sleeps should occur interruptibly.
 313  * @no_wait_gpu:           Never wait for gpu. Return -EBUSY instead.
 314  * @unlock_resv:           Unlock the reservation lock as well.
 315  */
 316
 317 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
 318                                bool interruptible, bool no_wait_gpu,
 319                                bool unlock_resv)
 320 {
 321         struct dma_resv *resv = &bo->base._resv;
 322         int ret;
 323
 324         if (dma_resv_test_signaled_rcu(resv, true))
 325                 ret = 0;
 326         else
 327                 ret = -EBUSY;
 328
 329         if (ret && !no_wait_gpu) {
 330                 long lret;
 331
 332                 if (unlock_resv)
 333                         dma_resv_unlock(bo->base.resv);
 334                 spin_unlock(&ttm_glob.lru_lock);
 335
 336                 lret = dma_resv_wait_timeout_rcu(resv, true, interruptible,
 337                                                  30 * HZ);
 338
 339                 if (lret < 0)
 340                         return lret;
 341                 else if (lret == 0)
 342                         return -EBUSY;
 343
 344                 spin_lock(&ttm_glob.lru_lock);
 345                 if (unlock_resv && !dma_resv_trylock(bo->base.resv)) {
 346                         /*
 347                          * We raced, and lost, someone else holds the reservation now,
 348                          * and is probably busy in ttm_bo_cleanup_memtype_use.
 349                          *
 350                          * Even if it's not the case, because we finished waiting any
 351                          * delayed destruction would succeed, so just return success
 352                          * here.
 353                          */
 354                         spin_unlock(&ttm_glob.lru_lock);
 355                         return 0;
 356                 }
 357                 ret = 0;
 358         }
 359
 360         if (ret || unlikely(list_empty(&bo->ddestroy))) {
 361                 if (unlock_resv)
 362                         dma_resv_unlock(bo->base.resv);
 363                 spin_unlock(&ttm_glob.lru_lock);
 364                 return ret;
 365         }
 366
 367         ttm_bo_del_from_lru(bo);
 368         list_del_init(&bo->ddestroy);
 369         spin_unlock(&ttm_glob.lru_lock);
 370         ttm_bo_cleanup_memtype_use(bo);
 371
 372         if (unlock_resv)
 373                 dma_resv_unlock(bo->base.resv);
 374
 375         ttm_bo_put(bo);
 376
 377         return 0;
 378 }
 379
 380 /*
 381  * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
 382  * encountered buffers.
 383  */
 384 bool ttm_bo_delayed_delete(struct ttm_device *bdev, bool remove_all)
 385 {
 386         struct ttm_global *glob = &ttm_glob;
 387         struct list_head removed;
 388         bool empty;
 389
 390         INIT_LIST_HEAD(&removed);
 391
 392         spin_lock(&glob->lru_lock);
 393         while (!list_empty(&bdev->ddestroy)) {
 394                 struct ttm_buffer_object *bo;
 395
 396                 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
 397                                       ddestroy);
 398                 list_move_tail(&bo->ddestroy, &removed);
 399                 if (!ttm_bo_get_unless_zero(bo))
 400                         continue;
 401
 402                 if (remove_all || bo->base.resv != &bo->base._resv) {
 403                         spin_unlock(&glob->lru_lock);
 404                         dma_resv_lock(bo->base.resv, NULL);
 405
 406                         spin_lock(&glob->lru_lock);
 407                         ttm_bo_cleanup_refs(bo, false, !remove_all, true);
 408
 409                 } else if (dma_resv_trylock(bo->base.resv)) {
 410                         ttm_bo_cleanup_refs(bo, false, !remove_all, true);
 411                 } else {
 412                         spin_unlock(&glob->lru_lock);
 413                 }
 414
 415                 ttm_bo_put(bo);
 416                 spin_lock(&glob->lru_lock);
 417         }
 418         list_splice_tail(&removed, &bdev->ddestroy);
 419         empty = list_empty(&bdev->ddestroy);
 420         spin_unlock(&glob->lru_lock);
 421
 422         return empty;
 423 }
 424
 425 static void ttm_bo_release(struct kref *kref)
 426 {
 427         struct ttm_buffer_object *bo =
 428             container_of(kref, struct ttm_buffer_object, kref);
 429         struct ttm_device *bdev = bo->bdev;
 430         int ret;
 431
 432         if (!bo->deleted) {
 433                 ret = ttm_bo_individualize_resv(bo);
 434                 if (ret) {
 435                         /* Last resort, if we fail to allocate memory for the
 436                          * fences block for the BO to become idle
 437                          */
 438                         dma_resv_wait_timeout_rcu(bo->base.resv, true, false,
 439                                                   30 * HZ);
 440                 }
 441
 442                 if (bo->bdev->funcs->release_notify)
 443                         bo->bdev->funcs->release_notify(bo);
 444
 445                 drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
 446                 ttm_mem_io_free(bdev, &bo->mem);
 447         }
 448
 449         if (!dma_resv_test_signaled_rcu(bo->base.resv, true) ||
 450             !dma_resv_trylock(bo->base.resv)) {
 451                 /* The BO is not idle, resurrect it for delayed destroy */
 452                 ttm_bo_flush_all_fences(bo);
 453                 bo->deleted = true;
 454
 455                 spin_lock(&ttm_glob.lru_lock);
 456
 457                 /*
 458                  * Make pinned bos immediately available to
 459                  * shrinkers, now that they are queued for
 460                  * destruction.
 461                  */
 462                 if (WARN_ON(bo->pin_count)) {
 463                         bo->pin_count = 0;
 464                         ttm_bo_move_to_lru_tail(bo, &bo->mem, NULL);
 465                 }
 466
 467                 kref_init(&bo->kref);
 468                 list_add_tail(&bo->ddestroy, &bdev->ddestroy);
 469                 spin_unlock(&ttm_glob.lru_lock);
 470
 471                 schedule_delayed_work(&bdev->wq,
 472                                       ((HZ / 100) < 1) ? 1 : HZ / 100);
 473                 return;
 474         }
 475
 476         spin_lock(&ttm_glob.lru_lock);
 477         ttm_bo_del_from_lru(bo);
 478         list_del(&bo->ddestroy);
 479         spin_unlock(&ttm_glob.lru_lock);
 480
 481         ttm_bo_cleanup_memtype_use(bo);
 482         dma_resv_unlock(bo->base.resv);
 483
 484         atomic_dec(&ttm_glob.bo_count);
 485         dma_fence_put(bo->moving);
 486         if (!ttm_bo_uses_embedded_gem_object(bo))
 487                 dma_resv_fini(&bo->base._resv);
 488         bo->destroy(bo);
 489 }
 490
 491 void ttm_bo_put(struct ttm_buffer_object *bo)
 492 {
 493         kref_put(&bo->kref, ttm_bo_release);
 494 }
 495 EXPORT_SYMBOL(ttm_bo_put);
 496
 497 int ttm_bo_lock_delayed_workqueue(struct ttm_device *bdev)
 498 {
 499         return cancel_delayed_work_sync(&bdev->wq);
 500 }
 501 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
 502
 503 void ttm_bo_unlock_delayed_workqueue(struct ttm_device *bdev, int resched)
 504 {
 505         if (resched)
 506                 schedule_delayed_work(&bdev->wq,
 507                                       ((HZ / 100) < 1) ? 1 : HZ / 100);
 508 }
 509 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
 510
 511 static int ttm_bo_evict(struct ttm_buffer_object *bo,
 512                         struct ttm_operation_ctx *ctx)
 513 {
 514         struct ttm_device *bdev = bo->bdev;
 515         struct ttm_resource evict_mem;
 516         struct ttm_placement placement;
 517         struct ttm_place hop;
 518         int ret = 0;
 519
 520         memset(&hop, 0, sizeof(hop));
 521
 522         dma_resv_assert_held(bo->base.resv);
 523
 524         placement.num_placement = 0;
 525         placement.num_busy_placement = 0;
 526         bdev->funcs->evict_flags(bo, &placement);
 527
 528         if (!placement.num_placement && !placement.num_busy_placement) {
 529                 ttm_bo_wait(bo, false, false);
 530
 531                 ttm_bo_cleanup_memtype_use(bo);
 532                 return ttm_tt_create(bo, false);
 533         }
 534
 535         evict_mem = bo->mem;
 536         evict_mem.mm_node = NULL;
 537         evict_mem.bus.offset = 0;
 538         evict_mem.bus.addr = NULL;
 539
 540         ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
 541         if (ret) {
 542                 if (ret != -ERESTARTSYS) {
 543                         pr_err("Failed to find memory space for buffer 0x%p eviction\n",
 544                                bo);
 545                         ttm_bo_mem_space_debug(bo, &placement);
 546                 }
 547                 goto out;
 548         }
 549
 550         ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx, &hop);
 551         if (unlikely(ret)) {
 552                 WARN(ret == -EMULTIHOP, "Unexpected multihop in eviction - likely driver bug\n");
 553                 if (ret != -ERESTARTSYS)
 554                         pr_err("Buffer eviction failed\n");
 555                 ttm_resource_free(bo, &evict_mem);
 556         }
 557 out:
 558         return ret;
 559 }
 560
 561 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
 562                               const struct ttm_place *place)
 563 {
 564         /* Don't evict this BO if it's outside of the
 565          * requested placement range
 566          */
 567         if (place->fpfn >= (bo->mem.start + bo->mem.num_pages) ||
 568             (place->lpfn && place->lpfn <= bo->mem.start))
 569                 return false;
 570
 571         return true;
 572 }
 573 EXPORT_SYMBOL(ttm_bo_eviction_valuable);
 574
 575 /*
 576  * Check the target bo is allowable to be evicted or swapout, including cases:
 577  *
 578  * a. if share same reservation object with ctx->resv, have assumption
 579  * reservation objects should already be locked, so not lock again and
 580  * return true directly when either the opreation allow_reserved_eviction
 581  * or the target bo already is in delayed free list;
 582  *
 583  * b. Otherwise, trylock it.
 584  */
 585 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
 586                         struct ttm_operation_ctx *ctx, bool *locked, bool *busy)
 587 {
 588         bool ret = false;
 589
 590         if (bo->base.resv == ctx->resv) {
 591                 dma_resv_assert_held(bo->base.resv);
 592                 if (ctx->allow_res_evict)
 593                         ret = true;
 594                 *locked = false;
 595                 if (busy)
 596                         *busy = false;
 597         } else {
 598                 ret = dma_resv_trylock(bo->base.resv);
 599                 *locked = ret;
 600                 if (busy)
 601                         *busy = !ret;
 602         }
 603
 604         return ret;
 605 }
 606
 607 /**
 608  * ttm_mem_evict_wait_busy - wait for a busy BO to become available
 609  *
 610  * @busy_bo: BO which couldn't be locked with trylock
 611  * @ctx: operation context
 612  * @ticket: acquire ticket
 613  *
 614  * Try to lock a busy buffer object to avoid failing eviction.
 615  */
 616 static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
 617                                    struct ttm_operation_ctx *ctx,
 618                                    struct ww_acquire_ctx *ticket)
 619 {
 620         int r;
 621
 622         if (!busy_bo || !ticket)
 623                 return -EBUSY;
 624
 625         if (ctx->interruptible)
 626                 r = dma_resv_lock_interruptible(busy_bo->base.resv,
 627                                                           ticket);
 628         else
 629                 r = dma_resv_lock(busy_bo->base.resv, ticket);
 630
 631         /*
 632          * TODO: It would be better to keep the BO locked until allocation is at
 633          * least tried one more time, but that would mean a much larger rework
 634          * of TTM.
 635          */
 636         if (!r)
 637                 dma_resv_unlock(busy_bo->base.resv);
 638
 639         return r == -EDEADLK ? -EBUSY : r;
 640 }
 641
 642 int ttm_mem_evict_first(struct ttm_device *bdev,
 643                         struct ttm_resource_manager *man,
 644                         const struct ttm_place *place,
 645                         struct ttm_operation_ctx *ctx,
 646                         struct ww_acquire_ctx *ticket)
 647 {
 648         struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
 649         bool locked = false;
 650         unsigned i;
 651         int ret;
 652
 653         spin_lock(&ttm_glob.lru_lock);
 654         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
 655                 list_for_each_entry(bo, &man->lru[i], lru) {
 656                         bool busy;
 657
 658                         if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
 659                                                             &busy)) {
 660                                 if (busy && !busy_bo && ticket !=
 661                                     dma_resv_locking_ctx(bo->base.resv))
 662                                         busy_bo = bo;
 663                                 continue;
 664                         }
 665
 666                         if (place && !bdev->funcs->eviction_valuable(bo,
 667                                                                       place)) {
 668                                 if (locked)
 669                                         dma_resv_unlock(bo->base.resv);
 670                                 continue;
 671                         }
 672                         if (!ttm_bo_get_unless_zero(bo)) {
 673                                 if (locked)
 674                                         dma_resv_unlock(bo->base.resv);
 675                                 continue;
 676                         }
 677                         break;
 678                 }
 679
 680                 /* If the inner loop terminated early, we have our candidate */
 681                 if (&bo->lru != &man->lru[i])
 682                         break;
 683
 684                 bo = NULL;
 685         }
 686
 687         if (!bo) {
 688                 if (busy_bo && !ttm_bo_get_unless_zero(busy_bo))
 689                         busy_bo = NULL;
 690                 spin_unlock(&ttm_glob.lru_lock);
 691                 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
 692                 if (busy_bo)
 693                         ttm_bo_put(busy_bo);
 694                 return ret;
 695         }
 696
 697         if (bo->deleted) {
 698                 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
 699                                           ctx->no_wait_gpu, locked);
 700                 ttm_bo_put(bo);
 701                 return ret;
 702         }
 703
 704         spin_unlock(&ttm_glob.lru_lock);
 705
 706         ret = ttm_bo_evict(bo, ctx);
 707         if (locked)
 708                 ttm_bo_unreserve(bo);
 709
 710         ttm_bo_put(bo);
 711         return ret;
 712 }
 713
 714 /*
 715  * Add the last move fence to the BO and reserve a new shared slot.
 716  */
 717 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
 718                                  struct ttm_resource_manager *man,
 719                                  struct ttm_resource *mem,
 720                                  bool no_wait_gpu)
 721 {
 722         struct dma_fence *fence;
 723         int ret;
 724
 725         spin_lock(&man->move_lock);
 726         fence = dma_fence_get(man->move);
 727         spin_unlock(&man->move_lock);
 728
 729         if (!fence)
 730                 return 0;
 731
 732         if (no_wait_gpu) {
 733                 dma_fence_put(fence);
 734                 return -EBUSY;
 735         }
 736
 737         dma_resv_add_shared_fence(bo->base.resv, fence);
 738
 739         ret = dma_resv_reserve_shared(bo->base.resv, 1);
 740         if (unlikely(ret)) {
 741                 dma_fence_put(fence);
 742                 return ret;
 743         }
 744
 745         dma_fence_put(bo->moving);
 746         bo->moving = fence;
 747         return 0;
 748 }
 749
 750 /*
 751  * Repeatedly evict memory from the LRU for @mem_type until we create enough
 752  * space, or we've evicted everything and there isn't enough space.
 753  */
 754 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
 755                                   const struct ttm_place *place,
 756                                   struct ttm_resource *mem,
 757                                   struct ttm_operation_ctx *ctx)
 758 {
 759         struct ttm_device *bdev = bo->bdev;
 760         struct ttm_resource_manager *man = ttm_manager_type(bdev, mem->mem_type);
 761         struct ww_acquire_ctx *ticket;
 762         int ret;
 763
 764         ticket = dma_resv_locking_ctx(bo->base.resv);
 765         do {
 766                 ret = ttm_resource_alloc(bo, place, mem);
 767                 if (likely(!ret))
 768                         break;
 769                 if (unlikely(ret != -ENOSPC))
 770                         return ret;
 771                 ret = ttm_mem_evict_first(bdev, man, place, ctx,
 772                                           ticket);
 773                 if (unlikely(ret != 0))
 774                         return ret;
 775         } while (1);
 776
 777         return ttm_bo_add_move_fence(bo, man, mem, ctx->no_wait_gpu);
 778 }
 779
 780 /**
 781  * ttm_bo_mem_placement - check if placement is compatible
 782  * @bo: BO to find memory for
 783  * @place: where to search
 784  * @mem: the memory object to fill in
 785  *
 786  * Check if placement is compatible and fill in mem structure.
 787  * Returns -EBUSY if placement won't work or negative error code.
 788  * 0 when placement can be used.
 789  */
 790 static int ttm_bo_mem_placement(struct ttm_buffer_object *bo,
 791                                 const struct ttm_place *place,
 792                                 struct ttm_resource *mem)
 793 {
 794         struct ttm_device *bdev = bo->bdev;
 795         struct ttm_resource_manager *man;
 796
 797         man = ttm_manager_type(bdev, place->mem_type);
 798         if (!man || !ttm_resource_manager_used(man))
 799                 return -EBUSY;
 800
 801         mem->mem_type = place->mem_type;
 802         mem->placement = place->flags;
 803
 804         spin_lock(&ttm_glob.lru_lock);
 805         ttm_bo_move_to_lru_tail(bo, mem, NULL);
 806         spin_unlock(&ttm_glob.lru_lock);
 807
 808         return 0;
 809 }
 810
 811 /*
 812  * Creates space for memory region @mem according to its type.
 813  *
 814  * This function first searches for free space in compatible memory types in
 815  * the priority order defined by the driver.  If free space isn't found, then
 816  * ttm_bo_mem_force_space is attempted in priority order to evict and find
 817  * space.
 818  */
 819 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
 820                         struct ttm_placement *placement,
 821                         struct ttm_resource *mem,
 822                         struct ttm_operation_ctx *ctx)
 823 {
 824         struct ttm_device *bdev = bo->bdev;
 825         bool type_found = false;
 826         int i, ret;
 827
 828         ret = dma_resv_reserve_shared(bo->base.resv, 1);
 829         if (unlikely(ret))
 830                 return ret;
 831
 832         for (i = 0; i < placement->num_placement; ++i) {
 833                 const struct ttm_place *place = &placement->placement[i];
 834                 struct ttm_resource_manager *man;
 835
 836                 ret = ttm_bo_mem_placement(bo, place, mem);
 837                 if (ret)
 838                         continue;
 839
 840                 type_found = true;
 841                 ret = ttm_resource_alloc(bo, place, mem);
 842                 if (ret == -ENOSPC)
 843                         continue;
 844                 if (unlikely(ret))
 845                         goto error;
 846
 847                 man = ttm_manager_type(bdev, mem->mem_type);
 848                 ret = ttm_bo_add_move_fence(bo, man, mem, ctx->no_wait_gpu);
 849                 if (unlikely(ret)) {
 850                         ttm_resource_free(bo, mem);
 851                         if (ret == -EBUSY)
 852                                 continue;
 853
 854                         goto error;
 855                 }
 856                 return 0;
 857         }
 858
 859         for (i = 0; i < placement->num_busy_placement; ++i) {
 860                 const struct ttm_place *place = &placement->busy_placement[i];
 861
 862                 ret = ttm_bo_mem_placement(bo, place, mem);
 863                 if (ret)
 864                         continue;
 865
 866                 type_found = true;
 867                 ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
 868                 if (likely(!ret))
 869                         return 0;
 870
 871                 if (ret && ret != -EBUSY)
 872                         goto error;
 873         }
 874
 875         ret = -ENOMEM;
 876         if (!type_found) {
 877                 pr_err(TTM_PFX "No compatible memory type found\n");
 878                 ret = -EINVAL;
 879         }
 880
 881 error:
 882         if (bo->mem.mem_type == TTM_PL_SYSTEM && !bo->pin_count)
 883                 ttm_bo_move_to_lru_tail_unlocked(bo);
 884
 885         return ret;
 886 }
 887 EXPORT_SYMBOL(ttm_bo_mem_space);
 888
 889 static int ttm_bo_bounce_temp_buffer(struct ttm_buffer_object *bo,
 890                                      struct ttm_resource *mem,
 891                                      struct ttm_operation_ctx *ctx,
 892                                      struct ttm_place *hop)
 893 {
 894         struct ttm_placement hop_placement;
 895         int ret;
 896         struct ttm_resource hop_mem = *mem;
 897
 898         hop_mem.mm_node = NULL;
 899         hop_mem.mem_type = TTM_PL_SYSTEM;
 900         hop_mem.placement = 0;
 901
 902         hop_placement.num_placement = hop_placement.num_busy_placement = 1;
 903         hop_placement.placement = hop_placement.busy_placement = hop;
 904
 905         /* find space in the bounce domain */
 906         ret = ttm_bo_mem_space(bo, &hop_placement, &hop_mem, ctx);
 907         if (ret)
 908                 return ret;
 909         /* move to the bounce domain */
 910         ret = ttm_bo_handle_move_mem(bo, &hop_mem, false, ctx, NULL);
 911         if (ret)
 912                 return ret;
 913         return 0;
 914 }
 915
 916 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
 917                               struct ttm_placement *placement,
 918                               struct ttm_operation_ctx *ctx)
 919 {
 920         int ret = 0;
 921         struct ttm_place hop;
 922         struct ttm_resource mem;
 923
 924         dma_resv_assert_held(bo->base.resv);
 925
 926         memset(&hop, 0, sizeof(hop));
 927
 928         mem.num_pages = PAGE_ALIGN(bo->base.size) >> PAGE_SHIFT;
 929         mem.page_alignment = bo->mem.page_alignment;
 930         mem.bus.offset = 0;
 931         mem.bus.addr = NULL;
 932         mem.mm_node = NULL;
 933
 934         /*
 935          * Determine where to move the buffer.
 936          *
 937          * If driver determines move is going to need
 938          * an extra step then it will return -EMULTIHOP
 939          * and the buffer will be moved to the temporary
 940          * stop and the driver will be called to make
 941          * the second hop.
 942          */
 943 bounce:
 944         ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
 945         if (ret)
 946                 return ret;
 947         ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx, &hop);
 948         if (ret == -EMULTIHOP) {
 949                 ret = ttm_bo_bounce_temp_buffer(bo, &mem, ctx, &hop);
 950                 if (ret)
 951                         return ret;
 952                 /* try and move to final place now. */
 953                 goto bounce;
 954         }
 955         if (ret)
 956                 ttm_resource_free(bo, &mem);
 957         return ret;
 958 }
 959
 960 static bool ttm_bo_places_compat(const struct ttm_place *places,
 961                                  unsigned num_placement,
 962                                  struct ttm_resource *mem,
 963                                  uint32_t *new_flags)
 964 {
 965         unsigned i;
 966
 967         for (i = 0; i < num_placement; i++) {
 968                 const struct ttm_place *heap = &places[i];
 969
 970                 if ((mem->start < heap->fpfn ||
 971                      (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
 972                         continue;
 973
 974                 *new_flags = heap->flags;
 975                 if ((mem->mem_type == heap->mem_type) &&
 976                     (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
 977                      (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
 978                         return true;
 979         }
 980         return false;
 981 }
 982
 983 bool ttm_bo_mem_compat(struct ttm_placement *placement,
 984                        struct ttm_resource *mem,
 985                        uint32_t *new_flags)
 986 {
 987         if (ttm_bo_places_compat(placement->placement, placement->num_placement,
 988                                  mem, new_flags))
 989                 return true;
 990
 991         if ((placement->busy_placement != placement->placement ||
 992              placement->num_busy_placement > placement->num_placement) &&
 993             ttm_bo_places_compat(placement->busy_placement,
 994                                  placement->num_busy_placement,
 995                                  mem, new_flags))
 996                 return true;
 997
 998         return false;
 999 }
1000 EXPORT_SYMBOL(ttm_bo_mem_compat);
1001
1002 int ttm_bo_validate(struct ttm_buffer_object *bo,
1003                     struct ttm_placement *placement,
1004                     struct ttm_operation_ctx *ctx)
1005 {
1006         int ret;
1007         uint32_t new_flags;
1008
1009         dma_resv_assert_held(bo->base.resv);
1010
1011         /*
1012          * Remove the backing store if no placement is given.
1013          */
1014         if (!placement->num_placement && !placement->num_busy_placement) {
1015                 ret = ttm_bo_pipeline_gutting(bo);
1016                 if (ret)
1017                         return ret;
1018
1019                 return ttm_tt_create(bo, false);
1020         }
1021
1022         /*
1023          * Check whether we need to move buffer.
1024          */
1025         if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1026                 ret = ttm_bo_move_buffer(bo, placement, ctx);
1027                 if (ret)
1028                         return ret;
1029         }
1030         /*
1031          * We might need to add a TTM.
1032          */
1033         if (bo->mem.mem_type == TTM_PL_SYSTEM) {
1034                 ret = ttm_tt_create(bo, true);
1035                 if (ret)
1036                         return ret;
1037         }
1038         return 0;
1039 }
1040 EXPORT_SYMBOL(ttm_bo_validate);
1041
1042 int ttm_bo_init_reserved(struct ttm_device *bdev,
1043                          struct ttm_buffer_object *bo,
1044                          size_t size,
1045                          enum ttm_bo_type type,
1046                          struct ttm_placement *placement,
1047                          uint32_t page_alignment,
1048                          struct ttm_operation_ctx *ctx,
1049                          struct sg_table *sg,
1050                          struct dma_resv *resv,
1051                          void (*destroy) (struct ttm_buffer_object *))
1052 {
1053         bool locked;
1054         int ret = 0;
1055
1056         bo->destroy = destroy ? destroy : ttm_bo_default_destroy;
1057
1058         kref_init(&bo->kref);
1059         INIT_LIST_HEAD(&bo->lru);
1060         INIT_LIST_HEAD(&bo->ddestroy);
1061         INIT_LIST_HEAD(&bo->swap);
1062         bo->bdev = bdev;
1063         bo->type = type;
1064         bo->mem.mem_type = TTM_PL_SYSTEM;
1065         bo->mem.num_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
1066         bo->mem.mm_node = NULL;
1067         bo->mem.page_alignment = page_alignment;
1068         bo->mem.bus.offset = 0;
1069         bo->mem.bus.addr = NULL;
1070         bo->moving = NULL;
1071         bo->mem.placement = 0;
1072         bo->pin_count = 0;
1073         bo->sg = sg;
1074         if (resv) {
1075                 bo->base.resv = resv;
1076                 dma_resv_assert_held(bo->base.resv);
1077         } else {
1078                 bo->base.resv = &bo->base._resv;
1079         }
1080         if (!ttm_bo_uses_embedded_gem_object(bo)) {
1081                 /*
1082                  * bo.base is not initialized, so we have to setup the
1083                  * struct elements we want use regardless.
1084                  */
1085                 bo->base.size = size;
1086                 dma_resv_init(&bo->base._resv);
1087                 drm_vma_node_reset(&bo->base.vma_node);
1088         }
1089         atomic_inc(&ttm_glob.bo_count);
1090
1091         /*
1092          * For ttm_bo_type_device buffers, allocate
1093          * address space from the device.
1094          */
1095         if (bo->type == ttm_bo_type_device ||
1096             bo->type == ttm_bo_type_sg)
1097                 ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
1098                                          bo->mem.num_pages);
1099
1100         /* passed reservation objects should already be locked,
1101          * since otherwise lockdep will be angered in radeon.
1102          */
1103         if (!resv) {
1104                 locked = dma_resv_trylock(bo->base.resv);
1105                 WARN_ON(!locked);
1106         }
1107
1108         if (likely(!ret))
1109                 ret = ttm_bo_validate(bo, placement, ctx);
1110
1111         if (unlikely(ret)) {
1112                 if (!resv)
1113                         ttm_bo_unreserve(bo);
1114
1115                 ttm_bo_put(bo);
1116                 return ret;
1117         }
1118
1119         ttm_bo_move_to_lru_tail_unlocked(bo);
1120
1121         return ret;
1122 }
1123 EXPORT_SYMBOL(ttm_bo_init_reserved);
1124
1125 int ttm_bo_init(struct ttm_device *bdev,
1126                 struct ttm_buffer_object *bo,
1127                 size_t size,
1128                 enum ttm_bo_type type,
1129                 struct ttm_placement *placement,
1130                 uint32_t page_alignment,
1131                 bool interruptible,
1132                 struct sg_table *sg,
1133                 struct dma_resv *resv,
1134                 void (*destroy) (struct ttm_buffer_object *))
1135 {
1136         struct ttm_operation_ctx ctx = { interruptible, false };
1137         int ret;
1138
1139         ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1140                                    page_alignment, &ctx, sg, resv, destroy);
1141         if (ret)
1142                 return ret;
1143
1144         if (!resv)
1145                 ttm_bo_unreserve(bo);
1146
1147         return 0;
1148 }
1149 EXPORT_SYMBOL(ttm_bo_init);
1150
1151 /*
1152  * buffer object vm functions.
1153  */
1154
1155 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1156 {
1157         struct ttm_device *bdev = bo->bdev;
1158
1159         drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1160         ttm_mem_io_free(bdev, &bo->mem);
1161 }
1162 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1163
1164 int ttm_bo_wait(struct ttm_buffer_object *bo,
1165                 bool interruptible, bool no_wait)
1166 {
1167         long timeout = 15 * HZ;
1168
1169         if (no_wait) {
1170                 if (dma_resv_test_signaled_rcu(bo->base.resv, true))
1171                         return 0;
1172                 else
1173                         return -EBUSY;
1174         }
1175
1176         timeout = dma_resv_wait_timeout_rcu(bo->base.resv, true,
1177                                                       interruptible, timeout);
1178         if (timeout < 0)
1179                 return timeout;
1180
1181         if (timeout == 0)
1182                 return -EBUSY;
1183
1184         dma_resv_add_excl_fence(bo->base.resv, NULL);
1185         return 0;
1186 }
1187 EXPORT_SYMBOL(ttm_bo_wait);
1188
1189 /*
1190  * A buffer object shrink method that tries to swap out the first
1191  * buffer object on the bo_global::swap_lru list.
1192  */
1193 int ttm_bo_swapout(struct ttm_operation_ctx *ctx, gfp_t gfp_flags)
1194 {
1195         struct ttm_global *glob = &ttm_glob;
1196         struct ttm_buffer_object *bo;
1197         int ret = -EBUSY;
1198         bool locked;
1199         unsigned i;
1200
1201         spin_lock(&glob->lru_lock);
1202         for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1203                 list_for_each_entry(bo, &glob->swap_lru[i], swap) {
1204                         if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
1205                                                             NULL))
1206                                 continue;
1207
1208                         if (!ttm_bo_get_unless_zero(bo)) {
1209                                 if (locked)
1210                                         dma_resv_unlock(bo->base.resv);
1211                                 continue;
1212                         }
1213
1214                         ret = 0;
1215                         break;
1216                 }
1217                 if (!ret)
1218                         break;
1219         }
1220
1221         if (ret) {
1222                 spin_unlock(&glob->lru_lock);
1223                 return ret;
1224         }
1225
1226         if (bo->deleted) {
1227                 ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1228                 ttm_bo_put(bo);
1229                 return ret;
1230         }
1231
1232         ttm_bo_del_from_lru(bo);
1233         spin_unlock(&glob->lru_lock);
1234
1235         /**
1236          * Move to system cached
1237          */
1238
1239         if (bo->mem.mem_type != TTM_PL_SYSTEM) {
1240                 struct ttm_operation_ctx ctx = { false, false };
1241                 struct ttm_resource evict_mem;
1242                 struct ttm_place hop;
1243
1244                 memset(&hop, 0, sizeof(hop));
1245
1246                 evict_mem = bo->mem;
1247                 evict_mem.mm_node = NULL;
1248                 evict_mem.placement = 0;
1249                 evict_mem.mem_type = TTM_PL_SYSTEM;
1250
1251                 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx, &hop);
1252                 if (unlikely(ret != 0)) {
1253                         WARN(ret == -EMULTIHOP, "Unexpected multihop in swaput - likely driver bug.\n");
1254                         goto out;
1255                 }
1256         }
1257
1258         /**
1259          * Make sure BO is idle.
1260          */
1261
1262         ret = ttm_bo_wait(bo, false, false);
1263         if (unlikely(ret != 0))
1264                 goto out;
1265
1266         ttm_bo_unmap_virtual(bo);
1267
1268         /**
1269          * Swap out. Buffer will be swapped in again as soon as
1270          * anyone tries to access a ttm page.
1271          */
1272
1273         if (bo->bdev->funcs->swap_notify)
1274                 bo->bdev->funcs->swap_notify(bo);
1275
1276         ret = ttm_tt_swapout(bo->bdev, bo->ttm, gfp_flags);
1277 out:
1278
1279         /**
1280          *
1281          * Unreserve without putting on LRU to avoid swapping out an
1282          * already swapped buffer.
1283          */
1284         if (locked)
1285                 dma_resv_unlock(bo->base.resv);
1286         ttm_bo_put(bo);
1287         return ret;
1288 }
1289 EXPORT_SYMBOL(ttm_bo_swapout);
1290
1291 void ttm_bo_tt_destroy(struct ttm_buffer_object *bo)
1292 {
1293         if (bo->ttm == NULL)
1294                 return;
1295
1296         ttm_tt_destroy(bo->bdev, bo->ttm);
1297         bo->ttm = NULL;
1298 }