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