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