Linux Audio

Check our new training course

Loading...
v6.13.7
   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.h>
  35#include <drm/ttm/ttm_placement.h>
  36#include <drm/ttm/ttm_tt.h>
  37
  38#include <linux/jiffies.h>
  39#include <linux/slab.h>
  40#include <linux/sched.h>
  41#include <linux/mm.h>
  42#include <linux/file.h>
  43#include <linux/module.h>
  44#include <linux/atomic.h>
  45#include <linux/dma-resv.h>
  46
  47#include "ttm_module.h"
  48
 
 
 
 
 
 
  49static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
  50					struct ttm_placement *placement)
  51{
  52	struct drm_printer p = drm_dbg_printer(NULL, DRM_UT_CORE, TTM_PFX);
  53	struct ttm_resource_manager *man;
  54	int i, mem_type;
  55
 
 
 
  56	for (i = 0; i < placement->num_placement; i++) {
  57		mem_type = placement->placement[i].mem_type;
  58		drm_printf(&p, "  placement[%d]=0x%08X (%d)\n",
  59			   i, placement->placement[i].flags, mem_type);
  60		man = ttm_manager_type(bo->bdev, mem_type);
  61		ttm_resource_manager_debug(man, &p);
  62	}
  63}
  64
  65/**
  66 * ttm_bo_move_to_lru_tail
  67 *
  68 * @bo: The buffer object.
  69 *
  70 * Move this BO to the tail of all lru lists used to lookup and reserve an
  71 * object. This function must be called with struct ttm_global::lru_lock
  72 * held, and is used to make a BO less likely to be considered for eviction.
  73 */
  74void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo)
  75{
  76	dma_resv_assert_held(bo->base.resv);
  77
  78	if (bo->resource)
  79		ttm_resource_move_to_lru_tail(bo->resource);
 
 
  80}
  81EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
  82
  83/**
  84 * ttm_bo_set_bulk_move - update BOs bulk move object
  85 *
  86 * @bo: The buffer object.
  87 * @bulk: bulk move structure
  88 *
  89 * Update the BOs bulk move object, making sure that resources are added/removed
  90 * as well. A bulk move allows to move many resource on the LRU at once,
  91 * resulting in much less overhead of maintaining the LRU.
  92 * The only requirement is that the resources stay together on the LRU and are
  93 * never separated. This is enforces by setting the bulk_move structure on a BO.
  94 * ttm_lru_bulk_move_tail() should be used to move all resources to the tail of
  95 * their LRU list.
  96 */
  97void ttm_bo_set_bulk_move(struct ttm_buffer_object *bo,
  98			  struct ttm_lru_bulk_move *bulk)
  99{
 100	dma_resv_assert_held(bo->base.resv);
 
 
 
 
 
 
 
 
 
 
 101
 102	if (bo->bulk_move == bulk)
 
 
 
 
 103		return;
 
 104
 105	spin_lock(&bo->bdev->lru_lock);
 106	if (bo->resource)
 107		ttm_resource_del_bulk_move(bo->resource, bo);
 108	bo->bulk_move = bulk;
 109	if (bo->resource)
 110		ttm_resource_add_bulk_move(bo->resource, bo);
 111	spin_unlock(&bo->bdev->lru_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 112}
 113EXPORT_SYMBOL(ttm_bo_set_bulk_move);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 114
 115static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
 116				  struct ttm_resource *mem, bool evict,
 117				  struct ttm_operation_ctx *ctx,
 118				  struct ttm_place *hop)
 119{
 
 120	struct ttm_device *bdev = bo->bdev;
 121	bool old_use_tt, new_use_tt;
 122	int ret;
 123
 124	old_use_tt = !bo->resource || ttm_manager_type(bdev, bo->resource->mem_type)->use_tt;
 125	new_use_tt = ttm_manager_type(bdev, mem->mem_type)->use_tt;
 126
 127	ttm_bo_unmap_virtual(bo);
 128
 129	/*
 130	 * Create and bind a ttm if required.
 131	 */
 132
 133	if (new_use_tt) {
 134		/* Zero init the new TTM structure if the old location should
 135		 * have used one as well.
 136		 */
 137		ret = ttm_tt_create(bo, old_use_tt);
 138		if (ret)
 139			goto out_err;
 140
 141		if (mem->mem_type != TTM_PL_SYSTEM) {
 142			ret = ttm_bo_populate(bo, ctx);
 143			if (ret)
 144				goto out_err;
 145		}
 146	}
 147
 148	ret = dma_resv_reserve_fences(bo->base.resv, 1);
 149	if (ret)
 150		goto out_err;
 151
 152	ret = bdev->funcs->move(bo, evict, ctx, mem, hop);
 153	if (ret) {
 154		if (ret == -EMULTIHOP)
 155			return ret;
 156		goto out_err;
 157	}
 158
 159	ctx->bytes_moved += bo->base.size;
 160	return 0;
 161
 162out_err:
 163	if (!old_use_tt)
 
 164		ttm_bo_tt_destroy(bo);
 165
 166	return ret;
 167}
 168
 169/*
 170 * Call bo::reserved.
 171 * Will release GPU memory type usage on destruction.
 172 * This is the place to put in driver specific hooks to release
 173 * driver private resources.
 174 * Will release the bo::reserved lock.
 175 */
 176
 177static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
 178{
 179	if (bo->bdev->funcs->delete_mem_notify)
 180		bo->bdev->funcs->delete_mem_notify(bo);
 181
 182	ttm_bo_tt_destroy(bo);
 183	ttm_resource_free(bo, &bo->resource);
 184}
 185
 186static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
 187{
 188	int r;
 189
 190	if (bo->base.resv == &bo->base._resv)
 191		return 0;
 192
 193	BUG_ON(!dma_resv_trylock(&bo->base._resv));
 194
 195	r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
 196	dma_resv_unlock(&bo->base._resv);
 197	if (r)
 198		return r;
 199
 200	if (bo->type != ttm_bo_type_sg) {
 201		/* This works because the BO is about to be destroyed and nobody
 202		 * reference it any more. The only tricky case is the trylock on
 203		 * the resv object while holding the lru_lock.
 204		 */
 205		spin_lock(&bo->bdev->lru_lock);
 206		bo->base.resv = &bo->base._resv;
 207		spin_unlock(&bo->bdev->lru_lock);
 208	}
 209
 210	return r;
 211}
 212
 213static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
 214{
 215	struct dma_resv *resv = &bo->base._resv;
 216	struct dma_resv_iter cursor;
 217	struct dma_fence *fence;
 
 
 
 
 
 
 
 
 
 
 218
 219	dma_resv_iter_begin(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP);
 220	dma_resv_for_each_fence_unlocked(&cursor, fence) {
 221		if (!fence->ops->signaled)
 222			dma_fence_enable_sw_signaling(fence);
 223	}
 224	dma_resv_iter_end(&cursor);
 225}
 226
 227/*
 228 * Block for the dma_resv object to become idle, lock the buffer and clean up
 229 * the resource and tt object.
 230 */
 231static void ttm_bo_delayed_delete(struct work_struct *work)
 
 
 
 
 
 
 
 
 
 
 
 
 232{
 233	struct ttm_buffer_object *bo;
 
 
 
 
 
 
 234
 235	bo = container_of(work, typeof(*bo), delayed_delete);
 
 236
 237	dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP, false,
 238			      MAX_SCHEDULE_TIMEOUT);
 239	dma_resv_lock(bo->base.resv, NULL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 240	ttm_bo_cleanup_memtype_use(bo);
 241	dma_resv_unlock(bo->base.resv);
 
 
 
 242	ttm_bo_put(bo);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 243}
 244
 245static void ttm_bo_release(struct kref *kref)
 246{
 247	struct ttm_buffer_object *bo =
 248	    container_of(kref, struct ttm_buffer_object, kref);
 249	struct ttm_device *bdev = bo->bdev;
 250	int ret;
 251
 252	WARN_ON_ONCE(bo->pin_count);
 253	WARN_ON_ONCE(bo->bulk_move);
 254
 255	if (!bo->deleted) {
 256		ret = ttm_bo_individualize_resv(bo);
 257		if (ret) {
 258			/* Last resort, if we fail to allocate memory for the
 259			 * fences block for the BO to become idle
 260			 */
 261			dma_resv_wait_timeout(bo->base.resv,
 262					      DMA_RESV_USAGE_BOOKKEEP, false,
 263					      30 * HZ);
 264		}
 265
 266		if (bo->bdev->funcs->release_notify)
 267			bo->bdev->funcs->release_notify(bo);
 268
 269		drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
 270		ttm_mem_io_free(bdev, bo->resource);
 
 271
 272		if (!dma_resv_test_signaled(bo->base.resv,
 273					    DMA_RESV_USAGE_BOOKKEEP) ||
 274		    (want_init_on_free() && (bo->ttm != NULL)) ||
 275		    bo->type == ttm_bo_type_sg ||
 276		    !dma_resv_trylock(bo->base.resv)) {
 277			/* The BO is not idle, resurrect it for delayed destroy */
 278			ttm_bo_flush_all_fences(bo);
 279			bo->deleted = true;
 280
 281			spin_lock(&bo->bdev->lru_lock);
 282
 283			/*
 284			 * Make pinned bos immediately available to
 285			 * shrinkers, now that they are queued for
 286			 * destruction.
 287			 *
 288			 * FIXME: QXL is triggering this. Can be removed when the
 289			 * driver is fixed.
 290			 */
 291			if (bo->pin_count) {
 292				bo->pin_count = 0;
 293				ttm_resource_move_to_lru_tail(bo->resource);
 294			}
 295
 296			kref_init(&bo->kref);
 297			spin_unlock(&bo->bdev->lru_lock);
 
 298
 299			INIT_WORK(&bo->delayed_delete, ttm_bo_delayed_delete);
 
 
 
 300
 301			/* Schedule the worker on the closest NUMA node. This
 302			 * improves performance since system memory might be
 303			 * cleared on free and that is best done on a CPU core
 304			 * close to it.
 305			 */
 306			queue_work_node(bdev->pool.nid, bdev->wq, &bo->delayed_delete);
 307			return;
 308		}
 309
 310		ttm_bo_cleanup_memtype_use(bo);
 311		dma_resv_unlock(bo->base.resv);
 312	}
 313
 314	atomic_dec(&ttm_glob.bo_count);
 
 315	bo->destroy(bo);
 316}
 317
 318/**
 319 * ttm_bo_put
 320 *
 321 * @bo: The buffer object.
 322 *
 323 * Unreference a buffer object.
 324 */
 325void ttm_bo_put(struct ttm_buffer_object *bo)
 326{
 327	kref_put(&bo->kref, ttm_bo_release);
 328}
 329EXPORT_SYMBOL(ttm_bo_put);
 330
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 331static int ttm_bo_bounce_temp_buffer(struct ttm_buffer_object *bo,
 
 332				     struct ttm_operation_ctx *ctx,
 333				     struct ttm_place *hop)
 334{
 335	struct ttm_placement hop_placement;
 336	struct ttm_resource *hop_mem;
 337	int ret;
 338
 339	hop_placement.num_placement = 1;
 340	hop_placement.placement = hop;
 341
 342	/* find space in the bounce domain */
 343	ret = ttm_bo_mem_space(bo, &hop_placement, &hop_mem, ctx);
 344	if (ret)
 345		return ret;
 346	/* move to the bounce domain */
 347	ret = ttm_bo_handle_move_mem(bo, hop_mem, false, ctx, NULL);
 348	if (ret) {
 349		ttm_resource_free(bo, &hop_mem);
 350		return ret;
 351	}
 352	return 0;
 353}
 354
 355static int ttm_bo_evict(struct ttm_buffer_object *bo,
 356			struct ttm_operation_ctx *ctx)
 357{
 358	struct ttm_device *bdev = bo->bdev;
 359	struct ttm_resource *evict_mem;
 360	struct ttm_placement placement;
 361	struct ttm_place hop;
 362	int ret = 0;
 363
 364	memset(&hop, 0, sizeof(hop));
 365
 366	dma_resv_assert_held(bo->base.resv);
 367
 368	placement.num_placement = 0;
 
 369	bdev->funcs->evict_flags(bo, &placement);
 370
 371	if (!placement.num_placement) {
 372		ret = ttm_bo_wait_ctx(bo, ctx);
 373		if (ret)
 374			return ret;
 375
 376		/*
 377		 * Since we've already synced, this frees backing store
 378		 * immediately.
 379		 */
 380		return ttm_bo_pipeline_gutting(bo);
 381	}
 382
 383	ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
 384	if (ret) {
 385		if (ret != -ERESTARTSYS) {
 386			pr_err("Failed to find memory space for buffer 0x%p eviction\n",
 387			       bo);
 388			ttm_bo_mem_space_debug(bo, &placement);
 389		}
 390		goto out;
 391	}
 392
 393	do {
 394		ret = ttm_bo_handle_move_mem(bo, evict_mem, true, ctx, &hop);
 395		if (ret != -EMULTIHOP)
 396			break;
 397
 398		ret = ttm_bo_bounce_temp_buffer(bo, ctx, &hop);
 399	} while (!ret);
 400
 401	if (ret) {
 402		ttm_resource_free(bo, &evict_mem);
 403		if (ret != -ERESTARTSYS && ret != -EINTR)
 404			pr_err("Buffer eviction failed\n");
 
 
 
 
 
 405	}
 406out:
 407	return ret;
 408}
 409
 410/**
 411 * ttm_bo_eviction_valuable
 412 *
 413 * @bo: The buffer object to evict
 414 * @place: the placement we need to make room for
 415 *
 416 * Check if it is valuable to evict the BO to make room for the given placement.
 417 */
 418bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
 419			      const struct ttm_place *place)
 420{
 421	struct ttm_resource *res = bo->resource;
 422	struct ttm_device *bdev = bo->bdev;
 423
 424	dma_resv_assert_held(bo->base.resv);
 425	if (bo->resource->mem_type == TTM_PL_SYSTEM)
 426		return true;
 427
 428	/* Don't evict this BO if it's outside of the
 429	 * requested placement range
 430	 */
 431	return ttm_resource_intersects(bdev, res, place, bo->base.size);
 
 
 
 
 432}
 433EXPORT_SYMBOL(ttm_bo_eviction_valuable);
 434
 435/**
 436 * ttm_bo_evict_first() - Evict the first bo on the manager's LRU list.
 437 * @bdev: The ttm device.
 438 * @man: The manager whose bo to evict.
 439 * @ctx: The TTM operation ctx governing the eviction.
 440 *
 441 * Return: 0 if successful or the resource disappeared. Negative error code on error.
 442 */
 443int ttm_bo_evict_first(struct ttm_device *bdev, struct ttm_resource_manager *man,
 444		       struct ttm_operation_ctx *ctx)
 445{
 446	struct ttm_resource_cursor cursor;
 447	struct ttm_buffer_object *bo;
 448	struct ttm_resource *res;
 449	unsigned int mem_type;
 450	int ret = 0;
 451
 452	spin_lock(&bdev->lru_lock);
 453	res = ttm_resource_manager_first(man, &cursor);
 454	ttm_resource_cursor_fini(&cursor);
 455	if (!res) {
 456		ret = -ENOENT;
 457		goto out_no_ref;
 458	}
 459	bo = res->bo;
 460	if (!ttm_bo_get_unless_zero(bo))
 461		goto out_no_ref;
 462	mem_type = res->mem_type;
 463	spin_unlock(&bdev->lru_lock);
 464	ret = ttm_bo_reserve(bo, ctx->interruptible, ctx->no_wait_gpu, NULL);
 465	if (ret)
 466		goto out_no_lock;
 467	if (!bo->resource || bo->resource->mem_type != mem_type)
 468		goto out_bo_moved;
 469
 470	if (bo->deleted) {
 471		ret = ttm_bo_wait_ctx(bo, ctx);
 472		if (!ret)
 473			ttm_bo_cleanup_memtype_use(bo);
 474	} else {
 475		ret = ttm_bo_evict(bo, ctx);
 
 
 
 
 
 
 
 
 
 
 
 476	}
 477out_bo_moved:
 478	dma_resv_unlock(bo->base.resv);
 479out_no_lock:
 480	ttm_bo_put(bo);
 481	return ret;
 482
 483out_no_ref:
 484	spin_unlock(&bdev->lru_lock);
 485	return ret;
 486}
 487
 488/**
 489 * struct ttm_bo_evict_walk - Parameters for the evict walk.
 
 
 
 
 
 
 490 */
 491struct ttm_bo_evict_walk {
 492	/** @walk: The walk base parameters. */
 493	struct ttm_lru_walk walk;
 494	/** @place: The place passed to the resource allocation. */
 495	const struct ttm_place *place;
 496	/** @evictor: The buffer object we're trying to make room for. */
 497	struct ttm_buffer_object *evictor;
 498	/** @res: The allocated resource if any. */
 499	struct ttm_resource **res;
 500	/** @evicted: Number of successful evictions. */
 501	unsigned long evicted;
 502};
 503
 504static s64 ttm_bo_evict_cb(struct ttm_lru_walk *walk, struct ttm_buffer_object *bo)
 505{
 506	struct ttm_bo_evict_walk *evict_walk =
 507		container_of(walk, typeof(*evict_walk), walk);
 508	s64 lret;
 509
 510	if (bo->pin_count || !bo->bdev->funcs->eviction_valuable(bo, evict_walk->place))
 511		return 0;
 512
 513	if (bo->deleted) {
 514		lret = ttm_bo_wait_ctx(bo, walk->ctx);
 515		if (!lret)
 516			ttm_bo_cleanup_memtype_use(bo);
 517	} else {
 518		lret = ttm_bo_evict(bo, walk->ctx);
 519	}
 520
 521	if (lret)
 522		goto out;
 523
 524	evict_walk->evicted++;
 525	if (evict_walk->res)
 526		lret = ttm_resource_alloc(evict_walk->evictor, evict_walk->place,
 527					  evict_walk->res);
 528	if (lret == 0)
 529		return 1;
 530out:
 531	/* Errors that should terminate the walk. */
 532	if (lret == -ENOSPC)
 533		return -EBUSY;
 534
 535	return lret;
 536}
 
 
 
 537
 538static const struct ttm_lru_walk_ops ttm_evict_walk_ops = {
 539	.process_bo = ttm_bo_evict_cb,
 540};
 541
 542static int ttm_bo_evict_alloc(struct ttm_device *bdev,
 543			      struct ttm_resource_manager *man,
 544			      const struct ttm_place *place,
 545			      struct ttm_buffer_object *evictor,
 546			      struct ttm_operation_ctx *ctx,
 547			      struct ww_acquire_ctx *ticket,
 548			      struct ttm_resource **res)
 549{
 550	struct ttm_bo_evict_walk evict_walk = {
 551		.walk = {
 552			.ops = &ttm_evict_walk_ops,
 553			.ctx = ctx,
 554			.ticket = ticket,
 555		},
 556		.place = place,
 557		.evictor = evictor,
 558		.res = res,
 559	};
 560	s64 lret;
 561
 562	evict_walk.walk.trylock_only = true;
 563	lret = ttm_lru_walk_for_evict(&evict_walk.walk, bdev, man, 1);
 564	if (lret || !ticket)
 565		goto out;
 566
 567	/* If ticket-locking, repeat while making progress. */
 568	evict_walk.walk.trylock_only = false;
 569	do {
 570		/* The walk may clear the evict_walk.walk.ticket field */
 571		evict_walk.walk.ticket = ticket;
 572		evict_walk.evicted = 0;
 573		lret = ttm_lru_walk_for_evict(&evict_walk.walk, bdev, man, 1);
 574	} while (!lret && evict_walk.evicted);
 575out:
 576	if (lret < 0)
 577		return lret;
 578	if (lret == 0)
 579		return -EBUSY;
 580	return 0;
 581}
 582
 583/**
 584 * ttm_bo_pin - Pin the buffer object.
 585 * @bo: The buffer object to pin
 586 *
 587 * Make sure the buffer is not evicted any more during memory pressure.
 588 * @bo must be unpinned again by calling ttm_bo_unpin().
 589 */
 590void ttm_bo_pin(struct ttm_buffer_object *bo)
 591{
 592	dma_resv_assert_held(bo->base.resv);
 593	WARN_ON_ONCE(!kref_read(&bo->kref));
 594	spin_lock(&bo->bdev->lru_lock);
 595	if (bo->resource)
 596		ttm_resource_del_bulk_move(bo->resource, bo);
 597	if (!bo->pin_count++ && bo->resource)
 598		ttm_resource_move_to_lru_tail(bo->resource);
 599	spin_unlock(&bo->bdev->lru_lock);
 600}
 601EXPORT_SYMBOL(ttm_bo_pin);
 602
 603/**
 604 * ttm_bo_unpin - Unpin the buffer object.
 605 * @bo: The buffer object to unpin
 606 *
 607 * Allows the buffer object to be evicted again during memory pressure.
 608 */
 609void ttm_bo_unpin(struct ttm_buffer_object *bo)
 610{
 611	dma_resv_assert_held(bo->base.resv);
 612	WARN_ON_ONCE(!kref_read(&bo->kref));
 613	if (WARN_ON_ONCE(!bo->pin_count))
 614		return;
 615
 616	spin_lock(&bo->bdev->lru_lock);
 617	if (!--bo->pin_count && bo->resource) {
 618		ttm_resource_add_bulk_move(bo->resource, bo);
 619		ttm_resource_move_to_lru_tail(bo->resource);
 
 
 
 
 
 
 
 
 
 620	}
 621	spin_unlock(&bo->bdev->lru_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 622}
 623EXPORT_SYMBOL(ttm_bo_unpin);
 624
 625/*
 626 * Add the last move fence to the BO as kernel dependency and reserve a new
 627 * fence slot.
 
 628 */
 629static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
 630				 struct ttm_resource_manager *man,
 
 631				 bool no_wait_gpu)
 632{
 633	struct dma_fence *fence;
 634	int ret;
 635
 636	spin_lock(&man->move_lock);
 637	fence = dma_fence_get(man->move);
 638	spin_unlock(&man->move_lock);
 639
 640	if (!fence)
 641		return 0;
 642
 643	if (no_wait_gpu) {
 644		ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
 645		dma_fence_put(fence);
 646		return ret;
 647	}
 648
 649	dma_resv_add_fence(bo->base.resv, fence, DMA_RESV_USAGE_KERNEL);
 650
 651	ret = dma_resv_reserve_fences(bo->base.resv, 1);
 652	dma_fence_put(fence);
 653	return ret;
 
 
 
 
 
 
 654}
 655
 656/**
 657 * ttm_bo_alloc_resource - Allocate backing store for a BO
 658 *
 659 * @bo: Pointer to a struct ttm_buffer_object of which we want a resource for
 660 * @placement: Proposed new placement for the buffer object
 661 * @ctx: if and how to sleep, lock buffers and alloc memory
 662 * @force_space: If we should evict buffers to force space
 663 * @res: The resulting struct ttm_resource.
 664 *
 665 * Allocates a resource for the buffer object pointed to by @bo, using the
 666 * placement flags in @placement, potentially evicting other buffer objects when
 667 * @force_space is true.
 668 * This function may sleep while waiting for resources to become available.
 669 * Returns:
 670 * -EBUSY: No space available (only if no_wait == true).
 671 * -ENOSPC: Could not allocate space for the buffer object, either due to
 672 * fragmentation or concurrent allocators.
 673 * -ERESTARTSYS: An interruptible sleep was interrupted by a signal.
 674 */
 675static int ttm_bo_alloc_resource(struct ttm_buffer_object *bo,
 676				 struct ttm_placement *placement,
 677				 struct ttm_operation_ctx *ctx,
 678				 bool force_space,
 679				 struct ttm_resource **res)
 680{
 681	struct ttm_device *bdev = bo->bdev;
 
 682	struct ww_acquire_ctx *ticket;
 683	int i, ret;
 684
 
 685	ticket = dma_resv_locking_ctx(bo->base.resv);
 686	ret = dma_resv_reserve_fences(bo->base.resv, 1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 687	if (unlikely(ret))
 688		return ret;
 689
 690	for (i = 0; i < placement->num_placement; ++i) {
 691		const struct ttm_place *place = &placement->placement[i];
 692		struct ttm_resource_manager *man;
 693		bool may_evict;
 694
 695		man = ttm_manager_type(bdev, place->mem_type);
 696		if (!man || !ttm_resource_manager_used(man))
 697			continue;
 698
 699		if (place->flags & (force_space ? TTM_PL_FLAG_DESIRED :
 700				    TTM_PL_FLAG_FALLBACK))
 
 701			continue;
 
 
 702
 703		may_evict = (force_space && place->mem_type != TTM_PL_SYSTEM);
 704		ret = ttm_resource_alloc(bo, place, res);
 705		if (ret) {
 706			if (ret != -ENOSPC)
 707				return ret;
 708			if (!may_evict)
 709				continue;
 710
 711			ret = ttm_bo_evict_alloc(bdev, man, place, bo, ctx,
 712						 ticket, res);
 713			if (ret == -EBUSY)
 714				continue;
 715			if (ret)
 716				return ret;
 717		}
 718
 719		ret = ttm_bo_add_move_fence(bo, man, ctx->no_wait_gpu);
 720		if (unlikely(ret)) {
 721			ttm_resource_free(bo, res);
 722			if (ret == -EBUSY)
 723				continue;
 724
 725			return ret;
 726		}
 727		return 0;
 728	}
 729
 730	return -ENOSPC;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 731}
 
 732
 733/*
 734 * ttm_bo_mem_space - Wrapper around ttm_bo_alloc_resource
 735 *
 736 * @bo: Pointer to a struct ttm_buffer_object of which we want a resource for
 737 * @placement: Proposed new placement for the buffer object
 738 * @res: The resulting struct ttm_resource.
 739 * @ctx: if and how to sleep, lock buffers and alloc memory
 740 *
 741 * Tries both idle allocation and forcefully eviction of buffers. See
 742 * ttm_bo_alloc_resource for details.
 743 */
 744int ttm_bo_mem_space(struct ttm_buffer_object *bo,
 745		     struct ttm_placement *placement,
 746		     struct ttm_resource **res,
 747		     struct ttm_operation_ctx *ctx)
 748{
 749	bool force_space = false;
 
 750	int ret;
 751
 752	do {
 753		ret = ttm_bo_alloc_resource(bo, placement, ctx,
 754					    force_space, res);
 755		force_space = !force_space;
 756	} while (ret == -ENOSPC && force_space);
 757
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 758	return ret;
 759}
 760EXPORT_SYMBOL(ttm_bo_mem_space);
 761
 762/**
 763 * ttm_bo_validate
 764 *
 765 * @bo: The buffer object.
 766 * @placement: Proposed placement for the buffer object.
 767 * @ctx: validation parameters.
 768 *
 769 * Changes placement and caching policy of the buffer object
 770 * according proposed placement.
 771 * Returns
 772 * -EINVAL on invalid proposed placement.
 773 * -ENOMEM on out-of-memory condition.
 774 * -EBUSY if no_wait is true and buffer busy.
 775 * -ERESTARTSYS if interrupted by a signal.
 776 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 777int ttm_bo_validate(struct ttm_buffer_object *bo,
 778		    struct ttm_placement *placement,
 779		    struct ttm_operation_ctx *ctx)
 780{
 781	struct ttm_resource *res;
 782	struct ttm_place hop;
 783	bool force_space;
 784	int ret;
 
 785
 786	dma_resv_assert_held(bo->base.resv);
 787
 788	/*
 789	 * Remove the backing store if no placement is given.
 790	 */
 791	if (!placement->num_placement)
 792		return ttm_bo_pipeline_gutting(bo);
 793
 794	force_space = false;
 795	do {
 796		/* Check whether we need to move buffer. */
 797		if (bo->resource &&
 798		    ttm_resource_compatible(bo->resource, placement,
 799					    force_space))
 800			return 0;
 801
 802		/* Moving of pinned BOs is forbidden */
 803		if (bo->pin_count)
 804			return -EINVAL;
 805
 806		/*
 807		 * Determine where to move the buffer.
 808		 *
 809		 * If driver determines move is going to need
 810		 * an extra step then it will return -EMULTIHOP
 811		 * and the buffer will be moved to the temporary
 812		 * stop and the driver will be called to make
 813		 * the second hop.
 814		 */
 815		ret = ttm_bo_alloc_resource(bo, placement, ctx, force_space,
 816					    &res);
 817		force_space = !force_space;
 818		if (ret == -ENOSPC)
 819			continue;
 820		if (ret)
 821			return ret;
 822
 823bounce:
 824		ret = ttm_bo_handle_move_mem(bo, res, false, ctx, &hop);
 825		if (ret == -EMULTIHOP) {
 826			ret = ttm_bo_bounce_temp_buffer(bo, ctx, &hop);
 827			/* try and move to final place now. */
 828			if (!ret)
 829				goto bounce;
 830		}
 831		if (ret) {
 832			ttm_resource_free(bo, &res);
 833			return ret;
 834		}
 835
 836	} while (ret && force_space);
 837
 838	/* For backward compatibility with userspace */
 839	if (ret == -ENOSPC)
 840		return -ENOMEM;
 841
 842	/*
 843	 * We might need to add a TTM.
 844	 */
 845	if (!bo->resource || bo->resource->mem_type == TTM_PL_SYSTEM) {
 846		ret = ttm_tt_create(bo, true);
 847		if (ret)
 848			return ret;
 849	}
 850	return 0;
 851}
 852EXPORT_SYMBOL(ttm_bo_validate);
 853
 854/**
 855 * ttm_bo_init_reserved
 856 *
 857 * @bdev: Pointer to a ttm_device struct.
 858 * @bo: Pointer to a ttm_buffer_object to be initialized.
 859 * @type: Requested type of buffer object.
 860 * @placement: Initial placement for buffer object.
 861 * @alignment: Data alignment in pages.
 862 * @ctx: TTM operation context for memory allocation.
 863 * @sg: Scatter-gather table.
 864 * @resv: Pointer to a dma_resv, or NULL to let ttm allocate one.
 865 * @destroy: Destroy function. Use NULL for kfree().
 866 *
 867 * This function initializes a pre-allocated struct ttm_buffer_object.
 868 * As this object may be part of a larger structure, this function,
 869 * together with the @destroy function, enables driver-specific objects
 870 * derived from a ttm_buffer_object.
 871 *
 872 * On successful return, the caller owns an object kref to @bo. The kref and
 873 * list_kref are usually set to 1, but note that in some situations, other
 874 * tasks may already be holding references to @bo as well.
 875 * Furthermore, if resv == NULL, the buffer's reservation lock will be held,
 876 * and it is the caller's responsibility to call ttm_bo_unreserve.
 877 *
 878 * If a failure occurs, the function will call the @destroy function. Thus,
 879 * after a failure, dereferencing @bo is illegal and will likely cause memory
 880 * corruption.
 881 *
 882 * Returns
 883 * -ENOMEM: Out of memory.
 884 * -EINVAL: Invalid placement flags.
 885 * -ERESTARTSYS: Interrupted by signal while sleeping waiting for resources.
 886 */
 887int ttm_bo_init_reserved(struct ttm_device *bdev, struct ttm_buffer_object *bo,
 888			 enum ttm_bo_type type, struct ttm_placement *placement,
 889			 uint32_t alignment, struct ttm_operation_ctx *ctx,
 890			 struct sg_table *sg, struct dma_resv *resv,
 891			 void (*destroy) (struct ttm_buffer_object *))
 892{
 
 
 893	int ret;
 894
 
 
 895	kref_init(&bo->kref);
 
 
 896	bo->bdev = bdev;
 897	bo->type = type;
 898	bo->page_alignment = alignment;
 899	bo->destroy = destroy;
 900	bo->pin_count = 0;
 901	bo->sg = sg;
 902	bo->bulk_move = NULL;
 903	if (resv)
 904		bo->base.resv = resv;
 905	else
 
 906		bo->base.resv = &bo->base._resv;
 
 907	atomic_inc(&ttm_glob.bo_count);
 908
 
 
 
 
 
 
 909	/*
 910	 * For ttm_bo_type_device buffers, allocate
 911	 * address space from the device.
 912	 */
 913	if (bo->type == ttm_bo_type_device || bo->type == ttm_bo_type_sg) {
 
 914		ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
 915					 PFN_UP(bo->base.size));
 916		if (ret)
 917			goto err_put;
 918	}
 919
 920	/* passed reservation objects should already be locked,
 921	 * since otherwise lockdep will be angered in radeon.
 922	 */
 923	if (!resv)
 924		WARN_ON(!dma_resv_trylock(bo->base.resv));
 925	else
 926		dma_resv_assert_held(resv);
 927
 928	ret = ttm_bo_validate(bo, placement, ctx);
 929	if (unlikely(ret))
 930		goto err_unlock;
 931
 932	return 0;
 
 
 933
 934err_unlock:
 935	if (!resv)
 936		dma_resv_unlock(bo->base.resv);
 
 
 937
 938err_put:
 939	ttm_bo_put(bo);
 940	return ret;
 941}
 942EXPORT_SYMBOL(ttm_bo_init_reserved);
 943
 944/**
 945 * ttm_bo_init_validate
 946 *
 947 * @bdev: Pointer to a ttm_device struct.
 948 * @bo: Pointer to a ttm_buffer_object to be initialized.
 949 * @type: Requested type of buffer object.
 950 * @placement: Initial placement for buffer object.
 951 * @alignment: Data alignment in pages.
 952 * @interruptible: If needing to sleep to wait for GPU resources,
 953 * sleep interruptible.
 954 * pinned in physical memory. If this behaviour is not desired, this member
 955 * holds a pointer to a persistent shmem object. Typically, this would
 956 * point to the shmem object backing a GEM object if TTM is used to back a
 957 * GEM user interface.
 958 * @sg: Scatter-gather table.
 959 * @resv: Pointer to a dma_resv, or NULL to let ttm allocate one.
 960 * @destroy: Destroy function. Use NULL for kfree().
 961 *
 962 * This function initializes a pre-allocated struct ttm_buffer_object.
 963 * As this object may be part of a larger structure, this function,
 964 * together with the @destroy function,
 965 * enables driver-specific objects derived from a ttm_buffer_object.
 966 *
 967 * On successful return, the caller owns an object kref to @bo. The kref and
 968 * list_kref are usually set to 1, but note that in some situations, other
 969 * tasks may already be holding references to @bo as well.
 970 *
 971 * If a failure occurs, the function will call the @destroy function, Thus,
 972 * after a failure, dereferencing @bo is illegal and will likely cause memory
 973 * corruption.
 974 *
 975 * Returns
 976 * -ENOMEM: Out of memory.
 977 * -EINVAL: Invalid placement flags.
 978 * -ERESTARTSYS: Interrupted by signal while sleeping waiting for resources.
 979 */
 980int ttm_bo_init_validate(struct ttm_device *bdev, struct ttm_buffer_object *bo,
 981			 enum ttm_bo_type type, struct ttm_placement *placement,
 982			 uint32_t alignment, bool interruptible,
 983			 struct sg_table *sg, struct dma_resv *resv,
 984			 void (*destroy) (struct ttm_buffer_object *))
 985{
 986	struct ttm_operation_ctx ctx = { interruptible, false };
 987	int ret;
 988
 989	ret = ttm_bo_init_reserved(bdev, bo, type, placement, alignment, &ctx,
 990				   sg, resv, destroy);
 991	if (ret)
 992		return ret;
 993
 994	if (!resv)
 995		ttm_bo_unreserve(bo);
 996
 997	return 0;
 998}
 999EXPORT_SYMBOL(ttm_bo_init_validate);
1000
1001/*
1002 * buffer object vm functions.
1003 */
1004
1005/**
1006 * ttm_bo_unmap_virtual
1007 *
1008 * @bo: tear down the virtual mappings for this BO
1009 */
1010void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1011{
1012	struct ttm_device *bdev = bo->bdev;
1013
1014	drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1015	ttm_mem_io_free(bdev, bo->resource);
1016}
1017EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1018
1019/**
1020 * ttm_bo_wait_ctx - wait for buffer idle.
1021 *
1022 * @bo:  The buffer object.
1023 * @ctx: defines how to wait
1024 *
1025 * Waits for the buffer to be idle. Used timeout depends on the context.
1026 * Returns -EBUSY if wait timed outt, -ERESTARTSYS if interrupted by a signal or
1027 * zero on success.
1028 */
1029int ttm_bo_wait_ctx(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx)
1030{
1031	long ret;
1032
1033	if (ctx->no_wait_gpu) {
1034		if (dma_resv_test_signaled(bo->base.resv,
1035					   DMA_RESV_USAGE_BOOKKEEP))
1036			return 0;
1037		else
1038			return -EBUSY;
1039	}
1040
1041	ret = dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP,
1042				    ctx->interruptible, 15 * HZ);
1043	if (unlikely(ret < 0))
1044		return ret;
1045	if (unlikely(ret == 0))
 
1046		return -EBUSY;
 
 
1047	return 0;
1048}
1049EXPORT_SYMBOL(ttm_bo_wait_ctx);
1050
1051/**
1052 * struct ttm_bo_swapout_walk - Parameters for the swapout walk
1053 */
1054struct ttm_bo_swapout_walk {
1055	/** @walk: The walk base parameters. */
1056	struct ttm_lru_walk walk;
1057	/** @gfp_flags: The gfp flags to use for ttm_tt_swapout() */
1058	gfp_t gfp_flags;
1059};
1060
1061static s64
1062ttm_bo_swapout_cb(struct ttm_lru_walk *walk, struct ttm_buffer_object *bo)
1063{
1064	struct ttm_place place = {.mem_type = bo->resource->mem_type};
1065	struct ttm_bo_swapout_walk *swapout_walk =
1066		container_of(walk, typeof(*swapout_walk), walk);
1067	struct ttm_operation_ctx *ctx = walk->ctx;
1068	s64 ret;
1069
1070	/*
1071	 * While the bo may already reside in SYSTEM placement, set
1072	 * SYSTEM as new placement to cover also the move further below.
1073	 * The driver may use the fact that we're moving from SYSTEM
1074	 * as an indication that we're about to swap out.
1075	 */
1076	if (bo->pin_count || !bo->bdev->funcs->eviction_valuable(bo, &place)) {
1077		ret = -EBUSY;
1078		goto out;
1079	}
1080
1081	if (!bo->ttm || !ttm_tt_is_populated(bo->ttm) ||
1082	    bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL ||
1083	    bo->ttm->page_flags & TTM_TT_FLAG_SWAPPED) {
1084		ret = -EBUSY;
1085		goto out;
 
 
1086	}
1087
1088	if (bo->deleted) {
1089		pgoff_t num_pages = bo->ttm->num_pages;
1090
1091		ret = ttm_bo_wait_ctx(bo, ctx);
1092		if (ret)
1093			goto out;
1094
1095		ttm_bo_cleanup_memtype_use(bo);
1096		ret = num_pages;
1097		goto out;
1098	}
1099
 
 
 
 
1100	/*
1101	 * Move to system cached
1102	 */
1103	if (bo->resource->mem_type != TTM_PL_SYSTEM) {
 
1104		struct ttm_resource *evict_mem;
1105		struct ttm_place hop;
1106
1107		memset(&hop, 0, sizeof(hop));
1108		place.mem_type = TTM_PL_SYSTEM;
1109		ret = ttm_resource_alloc(bo, &place, &evict_mem);
1110		if (ret)
1111			goto out;
1112
1113		ret = ttm_bo_handle_move_mem(bo, evict_mem, true, ctx, &hop);
1114		if (ret) {
1115			WARN(ret == -EMULTIHOP,
1116			     "Unexpected multihop in swapout - likely driver bug.\n");
1117			ttm_resource_free(bo, &evict_mem);
1118			goto out;
1119		}
1120	}
1121
1122	/*
1123	 * Make sure BO is idle.
1124	 */
1125	ret = ttm_bo_wait_ctx(bo, ctx);
1126	if (ret)
1127		goto out;
1128
1129	ttm_bo_unmap_virtual(bo);
 
 
 
 
 
1130	if (bo->bdev->funcs->swap_notify)
1131		bo->bdev->funcs->swap_notify(bo);
1132
1133	if (ttm_tt_is_populated(bo->ttm)) {
1134		spin_lock(&bo->bdev->lru_lock);
1135		ttm_resource_del_bulk_move(bo->resource, bo);
1136		spin_unlock(&bo->bdev->lru_lock);
1137
1138		ret = ttm_tt_swapout(bo->bdev, bo->ttm, swapout_walk->gfp_flags);
1139
1140		spin_lock(&bo->bdev->lru_lock);
1141		if (ret)
1142			ttm_resource_add_bulk_move(bo->resource, bo);
1143		ttm_resource_move_to_lru_tail(bo->resource);
1144		spin_unlock(&bo->bdev->lru_lock);
1145	}
1146
1147out:
1148	/* Consider -ENOMEM and -ENOSPC non-fatal. */
1149	if (ret == -ENOMEM || ret == -ENOSPC)
1150		ret = -EBUSY;
1151
1152	return ret;
1153}
1154
1155const struct ttm_lru_walk_ops ttm_swap_ops = {
1156	.process_bo = ttm_bo_swapout_cb,
1157};
1158
1159/**
1160 * ttm_bo_swapout() - Swap out buffer objects on the LRU list to shmem.
1161 * @bdev: The ttm device.
1162 * @ctx: The ttm_operation_ctx governing the swapout operation.
1163 * @man: The resource manager whose resources / buffer objects are
1164 * goint to be swapped out.
1165 * @gfp_flags: The gfp flags used for shmem page allocations.
1166 * @target: The desired number of bytes to swap out.
1167 *
1168 * Return: The number of bytes actually swapped out, or negative error code
1169 * on error.
1170 */
1171s64 ttm_bo_swapout(struct ttm_device *bdev, struct ttm_operation_ctx *ctx,
1172		   struct ttm_resource_manager *man, gfp_t gfp_flags,
1173		   s64 target)
1174{
1175	struct ttm_bo_swapout_walk swapout_walk = {
1176		.walk = {
1177			.ops = &ttm_swap_ops,
1178			.ctx = ctx,
1179			.trylock_only = true,
1180		},
1181		.gfp_flags = gfp_flags,
1182	};
1183
1184	return ttm_lru_walk_for_evict(&swapout_walk.walk, bdev, man, target);
 
 
 
 
 
 
 
1185}
1186
1187void ttm_bo_tt_destroy(struct ttm_buffer_object *bo)
1188{
1189	if (bo->ttm == NULL)
1190		return;
1191
1192	ttm_tt_unpopulate(bo->bdev, bo->ttm);
1193	ttm_tt_destroy(bo->bdev, bo->ttm);
1194	bo->ttm = NULL;
1195}
1196
1197/**
1198 * ttm_bo_populate() - Ensure that a buffer object has backing pages
1199 * @bo: The buffer object
1200 * @ctx: The ttm_operation_ctx governing the operation.
1201 *
1202 * For buffer objects in a memory type whose manager uses
1203 * struct ttm_tt for backing pages, ensure those backing pages
1204 * are present and with valid content. The bo's resource is also
1205 * placed on the correct LRU list if it was previously swapped
1206 * out.
1207 *
1208 * Return: 0 if successful, negative error code on failure.
1209 * Note: May return -EINTR or -ERESTARTSYS if @ctx::interruptible
1210 * is set to true.
1211 */
1212int ttm_bo_populate(struct ttm_buffer_object *bo,
1213		    struct ttm_operation_ctx *ctx)
1214{
1215	struct ttm_tt *tt = bo->ttm;
1216	bool swapped;
1217	int ret;
1218
1219	dma_resv_assert_held(bo->base.resv);
1220
1221	if (!tt)
1222		return 0;
1223
1224	swapped = ttm_tt_is_swapped(tt);
1225	ret = ttm_tt_populate(bo->bdev, tt, ctx);
1226	if (ret)
1227		return ret;
1228
1229	if (swapped && !ttm_tt_is_swapped(tt) && !bo->pin_count &&
1230	    bo->resource) {
1231		spin_lock(&bo->bdev->lru_lock);
1232		ttm_resource_add_bulk_move(bo->resource, bo);
1233		ttm_resource_move_to_lru_tail(bo->resource);
1234		spin_unlock(&bo->bdev->lru_lock);
1235	}
1236
1237	return 0;
1238}
1239EXPORT_SYMBOL(ttm_bo_populate);
v5.14.15
   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 */
  48static void ttm_bo_default_destroy(struct ttm_buffer_object *bo)
  49{
  50	kfree(bo);
  51}
  52
  53static 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
  72static 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
  82static 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
  90void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo,
  91			     struct ttm_resource *mem,
  92			     struct ttm_lru_bulk_move *bulk)
  93{
  94	struct ttm_device *bdev = bo->bdev;
  95	struct ttm_resource_manager *man;
  96
  97	if (!bo->deleted)
  98		dma_resv_assert_held(bo->base.resv);
  99
 100	if (bo->pin_count) {
 101		ttm_bo_del_from_lru(bo);
 102		return;
 103	}
 104
 105	if (!mem)
 106		return;
 107
 108	man = ttm_manager_type(bdev, mem->mem_type);
 109	list_move_tail(&bo->lru, &man->lru[bo->priority]);
 110
 111	if (bdev->funcs->del_from_lru_notify)
 112		bdev->funcs->del_from_lru_notify(bo);
 113
 114	if (bulk && !bo->pin_count) {
 115		switch (bo->resource->mem_type) {
 116		case TTM_PL_TT:
 117			ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo);
 118			break;
 119
 120		case TTM_PL_VRAM:
 121			ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo);
 122			break;
 123		}
 124	}
 125}
 126EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
 127
 128void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk)
 129{
 130	unsigned i;
 131
 132	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
 133		struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i];
 134		struct ttm_resource_manager *man;
 135
 136		if (!pos->first)
 137			continue;
 138
 139		dma_resv_assert_held(pos->first->base.resv);
 140		dma_resv_assert_held(pos->last->base.resv);
 141
 142		man = ttm_manager_type(pos->first->bdev, TTM_PL_TT);
 143		list_bulk_move_tail(&man->lru[i], &pos->first->lru,
 144				    &pos->last->lru);
 145	}
 146
 147	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
 148		struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i];
 149		struct ttm_resource_manager *man;
 150
 151		if (!pos->first)
 152			continue;
 153
 154		dma_resv_assert_held(pos->first->base.resv);
 155		dma_resv_assert_held(pos->last->base.resv);
 156
 157		man = ttm_manager_type(pos->first->bdev, TTM_PL_VRAM);
 158		list_bulk_move_tail(&man->lru[i], &pos->first->lru,
 159				    &pos->last->lru);
 160	}
 161}
 162EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail);
 163
 164static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
 165				  struct ttm_resource *mem, bool evict,
 166				  struct ttm_operation_ctx *ctx,
 167				  struct ttm_place *hop)
 168{
 169	struct ttm_resource_manager *old_man, *new_man;
 170	struct ttm_device *bdev = bo->bdev;
 
 171	int ret;
 172
 173	old_man = ttm_manager_type(bdev, bo->resource->mem_type);
 174	new_man = ttm_manager_type(bdev, mem->mem_type);
 175
 176	ttm_bo_unmap_virtual(bo);
 177
 178	/*
 179	 * Create and bind a ttm if required.
 180	 */
 181
 182	if (new_man->use_tt) {
 183		/* Zero init the new TTM structure if the old location should
 184		 * have used one as well.
 185		 */
 186		ret = ttm_tt_create(bo, old_man->use_tt);
 187		if (ret)
 188			goto out_err;
 189
 190		if (mem->mem_type != TTM_PL_SYSTEM) {
 191			ret = ttm_tt_populate(bo->bdev, bo->ttm, ctx);
 192			if (ret)
 193				goto out_err;
 194		}
 195	}
 196
 
 
 
 
 197	ret = bdev->funcs->move(bo, evict, ctx, mem, hop);
 198	if (ret) {
 199		if (ret == -EMULTIHOP)
 200			return ret;
 201		goto out_err;
 202	}
 203
 204	ctx->bytes_moved += bo->base.size;
 205	return 0;
 206
 207out_err:
 208	new_man = ttm_manager_type(bdev, bo->resource->mem_type);
 209	if (!new_man->use_tt)
 210		ttm_bo_tt_destroy(bo);
 211
 212	return ret;
 213}
 214
 215/*
 216 * Call bo::reserved.
 217 * Will release GPU memory type usage on destruction.
 218 * This is the place to put in driver specific hooks to release
 219 * driver private resources.
 220 * Will release the bo::reserved lock.
 221 */
 222
 223static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
 224{
 225	if (bo->bdev->funcs->delete_mem_notify)
 226		bo->bdev->funcs->delete_mem_notify(bo);
 227
 228	ttm_bo_tt_destroy(bo);
 229	ttm_resource_free(bo, &bo->resource);
 230}
 231
 232static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
 233{
 234	int r;
 235
 236	if (bo->base.resv == &bo->base._resv)
 237		return 0;
 238
 239	BUG_ON(!dma_resv_trylock(&bo->base._resv));
 240
 241	r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv);
 242	dma_resv_unlock(&bo->base._resv);
 243	if (r)
 244		return r;
 245
 246	if (bo->type != ttm_bo_type_sg) {
 247		/* This works because the BO is about to be destroyed and nobody
 248		 * reference it any more. The only tricky case is the trylock on
 249		 * the resv object while holding the lru_lock.
 250		 */
 251		spin_lock(&bo->bdev->lru_lock);
 252		bo->base.resv = &bo->base._resv;
 253		spin_unlock(&bo->bdev->lru_lock);
 254	}
 255
 256	return r;
 257}
 258
 259static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
 260{
 261	struct dma_resv *resv = &bo->base._resv;
 262	struct dma_resv_list *fobj;
 263	struct dma_fence *fence;
 264	int i;
 265
 266	rcu_read_lock();
 267	fobj = dma_resv_shared_list(resv);
 268	fence = dma_resv_excl_fence(resv);
 269	if (fence && !fence->ops->signaled)
 270		dma_fence_enable_sw_signaling(fence);
 271
 272	for (i = 0; fobj && i < fobj->shared_count; ++i) {
 273		fence = rcu_dereference(fobj->shared[i]);
 274
 
 
 275		if (!fence->ops->signaled)
 276			dma_fence_enable_sw_signaling(fence);
 277	}
 278	rcu_read_unlock();
 279}
 280
 281/**
 282 * ttm_bo_cleanup_refs
 283 * If bo idle, remove from lru lists, and unref.
 284 * If not idle, block if possible.
 285 *
 286 * Must be called with lru_lock and reservation held, this function
 287 * will drop the lru lock and optionally the reservation lock before returning.
 288 *
 289 * @bo:                    The buffer object to clean-up
 290 * @interruptible:         Any sleeps should occur interruptibly.
 291 * @no_wait_gpu:           Never wait for gpu. Return -EBUSY instead.
 292 * @unlock_resv:           Unlock the reservation lock as well.
 293 */
 294
 295static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
 296			       bool interruptible, bool no_wait_gpu,
 297			       bool unlock_resv)
 298{
 299	struct dma_resv *resv = &bo->base._resv;
 300	int ret;
 301
 302	if (dma_resv_test_signaled(resv, true))
 303		ret = 0;
 304	else
 305		ret = -EBUSY;
 306
 307	if (ret && !no_wait_gpu) {
 308		long lret;
 309
 310		if (unlock_resv)
 311			dma_resv_unlock(bo->base.resv);
 312		spin_unlock(&bo->bdev->lru_lock);
 313
 314		lret = dma_resv_wait_timeout(resv, true, interruptible,
 315					     30 * HZ);
 316
 317		if (lret < 0)
 318			return lret;
 319		else if (lret == 0)
 320			return -EBUSY;
 321
 322		spin_lock(&bo->bdev->lru_lock);
 323		if (unlock_resv && !dma_resv_trylock(bo->base.resv)) {
 324			/*
 325			 * We raced, and lost, someone else holds the reservation now,
 326			 * and is probably busy in ttm_bo_cleanup_memtype_use.
 327			 *
 328			 * Even if it's not the case, because we finished waiting any
 329			 * delayed destruction would succeed, so just return success
 330			 * here.
 331			 */
 332			spin_unlock(&bo->bdev->lru_lock);
 333			return 0;
 334		}
 335		ret = 0;
 336	}
 337
 338	if (ret || unlikely(list_empty(&bo->ddestroy))) {
 339		if (unlock_resv)
 340			dma_resv_unlock(bo->base.resv);
 341		spin_unlock(&bo->bdev->lru_lock);
 342		return ret;
 343	}
 344
 345	ttm_bo_del_from_lru(bo);
 346	list_del_init(&bo->ddestroy);
 347	spin_unlock(&bo->bdev->lru_lock);
 348	ttm_bo_cleanup_memtype_use(bo);
 349
 350	if (unlock_resv)
 351		dma_resv_unlock(bo->base.resv);
 352
 353	ttm_bo_put(bo);
 354
 355	return 0;
 356}
 357
 358/*
 359 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
 360 * encountered buffers.
 361 */
 362bool ttm_bo_delayed_delete(struct ttm_device *bdev, bool remove_all)
 363{
 364	struct list_head removed;
 365	bool empty;
 366
 367	INIT_LIST_HEAD(&removed);
 368
 369	spin_lock(&bdev->lru_lock);
 370	while (!list_empty(&bdev->ddestroy)) {
 371		struct ttm_buffer_object *bo;
 372
 373		bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
 374				      ddestroy);
 375		list_move_tail(&bo->ddestroy, &removed);
 376		if (!ttm_bo_get_unless_zero(bo))
 377			continue;
 378
 379		if (remove_all || bo->base.resv != &bo->base._resv) {
 380			spin_unlock(&bdev->lru_lock);
 381			dma_resv_lock(bo->base.resv, NULL);
 382
 383			spin_lock(&bdev->lru_lock);
 384			ttm_bo_cleanup_refs(bo, false, !remove_all, true);
 385
 386		} else if (dma_resv_trylock(bo->base.resv)) {
 387			ttm_bo_cleanup_refs(bo, false, !remove_all, true);
 388		} else {
 389			spin_unlock(&bdev->lru_lock);
 390		}
 391
 392		ttm_bo_put(bo);
 393		spin_lock(&bdev->lru_lock);
 394	}
 395	list_splice_tail(&removed, &bdev->ddestroy);
 396	empty = list_empty(&bdev->ddestroy);
 397	spin_unlock(&bdev->lru_lock);
 398
 399	return empty;
 400}
 401
 402static void ttm_bo_release(struct kref *kref)
 403{
 404	struct ttm_buffer_object *bo =
 405	    container_of(kref, struct ttm_buffer_object, kref);
 406	struct ttm_device *bdev = bo->bdev;
 407	int ret;
 408
 409	WARN_ON_ONCE(bo->pin_count);
 
 410
 411	if (!bo->deleted) {
 412		ret = ttm_bo_individualize_resv(bo);
 413		if (ret) {
 414			/* Last resort, if we fail to allocate memory for the
 415			 * fences block for the BO to become idle
 416			 */
 417			dma_resv_wait_timeout(bo->base.resv, true, false,
 
 418					      30 * HZ);
 419		}
 420
 421		if (bo->bdev->funcs->release_notify)
 422			bo->bdev->funcs->release_notify(bo);
 423
 424		drm_vma_offset_remove(bdev->vma_manager, &bo->base.vma_node);
 425		ttm_mem_io_free(bdev, bo->resource);
 426	}
 427
 428	if (!dma_resv_test_signaled(bo->base.resv, true) ||
 429	    !dma_resv_trylock(bo->base.resv)) {
 430		/* The BO is not idle, resurrect it for delayed destroy */
 431		ttm_bo_flush_all_fences(bo);
 432		bo->deleted = true;
 
 
 
 433
 434		spin_lock(&bo->bdev->lru_lock);
 435
 436		/*
 437		 * Make pinned bos immediately available to
 438		 * shrinkers, now that they are queued for
 439		 * destruction.
 440		 *
 441		 * FIXME: QXL is triggering this. Can be removed when the
 442		 * driver is fixed.
 443		 */
 444		if (bo->pin_count) {
 445			bo->pin_count = 0;
 446			ttm_bo_move_to_lru_tail(bo, bo->resource, NULL);
 447		}
 448
 449		kref_init(&bo->kref);
 450		list_add_tail(&bo->ddestroy, &bdev->ddestroy);
 451		spin_unlock(&bo->bdev->lru_lock);
 452
 453		schedule_delayed_work(&bdev->wq,
 454				      ((HZ / 100) < 1) ? 1 : HZ / 100);
 455		return;
 456	}
 457
 458	spin_lock(&bo->bdev->lru_lock);
 459	ttm_bo_del_from_lru(bo);
 460	list_del(&bo->ddestroy);
 461	spin_unlock(&bo->bdev->lru_lock);
 
 
 
 
 462
 463	ttm_bo_cleanup_memtype_use(bo);
 464	dma_resv_unlock(bo->base.resv);
 
 465
 466	atomic_dec(&ttm_glob.bo_count);
 467	dma_fence_put(bo->moving);
 468	bo->destroy(bo);
 469}
 470
 
 
 
 
 
 
 
 471void ttm_bo_put(struct ttm_buffer_object *bo)
 472{
 473	kref_put(&bo->kref, ttm_bo_release);
 474}
 475EXPORT_SYMBOL(ttm_bo_put);
 476
 477int ttm_bo_lock_delayed_workqueue(struct ttm_device *bdev)
 478{
 479	return cancel_delayed_work_sync(&bdev->wq);
 480}
 481EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
 482
 483void ttm_bo_unlock_delayed_workqueue(struct ttm_device *bdev, int resched)
 484{
 485	if (resched)
 486		schedule_delayed_work(&bdev->wq,
 487				      ((HZ / 100) < 1) ? 1 : HZ / 100);
 488}
 489EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
 490
 491static int ttm_bo_bounce_temp_buffer(struct ttm_buffer_object *bo,
 492				     struct ttm_resource **mem,
 493				     struct ttm_operation_ctx *ctx,
 494				     struct ttm_place *hop)
 495{
 496	struct ttm_placement hop_placement;
 497	struct ttm_resource *hop_mem;
 498	int ret;
 499
 500	hop_placement.num_placement = hop_placement.num_busy_placement = 1;
 501	hop_placement.placement = hop_placement.busy_placement = hop;
 502
 503	/* find space in the bounce domain */
 504	ret = ttm_bo_mem_space(bo, &hop_placement, &hop_mem, ctx);
 505	if (ret)
 506		return ret;
 507	/* move to the bounce domain */
 508	ret = ttm_bo_handle_move_mem(bo, hop_mem, false, ctx, NULL);
 509	if (ret) {
 510		ttm_resource_free(bo, &hop_mem);
 511		return ret;
 512	}
 513	return 0;
 514}
 515
 516static int ttm_bo_evict(struct ttm_buffer_object *bo,
 517			struct ttm_operation_ctx *ctx)
 518{
 519	struct ttm_device *bdev = bo->bdev;
 520	struct ttm_resource *evict_mem;
 521	struct ttm_placement placement;
 522	struct ttm_place hop;
 523	int ret = 0;
 524
 525	memset(&hop, 0, sizeof(hop));
 526
 527	dma_resv_assert_held(bo->base.resv);
 528
 529	placement.num_placement = 0;
 530	placement.num_busy_placement = 0;
 531	bdev->funcs->evict_flags(bo, &placement);
 532
 533	if (!placement.num_placement && !placement.num_busy_placement) {
 534		ret = ttm_bo_wait(bo, true, false);
 535		if (ret)
 536			return ret;
 537
 538		/*
 539		 * Since we've already synced, this frees backing store
 540		 * immediately.
 541		 */
 542		return ttm_bo_pipeline_gutting(bo);
 543	}
 544
 545	ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
 546	if (ret) {
 547		if (ret != -ERESTARTSYS) {
 548			pr_err("Failed to find memory space for buffer 0x%p eviction\n",
 549			       bo);
 550			ttm_bo_mem_space_debug(bo, &placement);
 551		}
 552		goto out;
 553	}
 554
 555bounce:
 556	ret = ttm_bo_handle_move_mem(bo, evict_mem, true, ctx, &hop);
 557	if (ret == -EMULTIHOP) {
 558		ret = ttm_bo_bounce_temp_buffer(bo, &evict_mem, ctx, &hop);
 559		if (ret) {
 
 
 
 
 
 
 560			pr_err("Buffer eviction failed\n");
 561			ttm_resource_free(bo, &evict_mem);
 562			goto out;
 563		}
 564		/* try and move to final place now. */
 565		goto bounce;
 566	}
 567out:
 568	return ret;
 569}
 570
 
 
 
 
 
 
 
 
 571bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
 572			      const struct ttm_place *place)
 573{
 
 
 
 574	dma_resv_assert_held(bo->base.resv);
 575	if (bo->resource->mem_type == TTM_PL_SYSTEM)
 576		return true;
 577
 578	/* Don't evict this BO if it's outside of the
 579	 * requested placement range
 580	 */
 581	if (place->fpfn >= (bo->resource->start + bo->resource->num_pages) ||
 582	    (place->lpfn && place->lpfn <= bo->resource->start))
 583		return false;
 584
 585	return true;
 586}
 587EXPORT_SYMBOL(ttm_bo_eviction_valuable);
 588
 589/*
 590 * Check the target bo is allowable to be evicted or swapout, including cases:
 
 
 
 591 *
 592 * a. if share same reservation object with ctx->resv, have assumption
 593 * reservation objects should already be locked, so not lock again and
 594 * return true directly when either the opreation allow_reserved_eviction
 595 * or the target bo already is in delayed free list;
 596 *
 597 * b. Otherwise, trylock it.
 598 */
 599static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
 600					   struct ttm_operation_ctx *ctx,
 601					   const struct ttm_place *place,
 602					   bool *locked, bool *busy)
 603{
 604	bool ret = false;
 605
 606	if (bo->base.resv == ctx->resv) {
 607		dma_resv_assert_held(bo->base.resv);
 608		if (ctx->allow_res_evict)
 609			ret = true;
 610		*locked = false;
 611		if (busy)
 612			*busy = false;
 
 
 
 
 
 
 
 
 
 
 
 
 613	} else {
 614		ret = dma_resv_trylock(bo->base.resv);
 615		*locked = ret;
 616		if (busy)
 617			*busy = !ret;
 618	}
 619
 620	if (ret && place && !bo->bdev->funcs->eviction_valuable(bo, place)) {
 621		ret = false;
 622		if (*locked) {
 623			dma_resv_unlock(bo->base.resv);
 624			*locked = false;
 625		}
 626	}
 
 
 
 
 
 627
 
 
 628	return ret;
 629}
 630
 631/**
 632 * ttm_mem_evict_wait_busy - wait for a busy BO to become available
 633 *
 634 * @busy_bo: BO which couldn't be locked with trylock
 635 * @ctx: operation context
 636 * @ticket: acquire ticket
 637 *
 638 * Try to lock a busy buffer object to avoid failing eviction.
 639 */
 640static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo,
 641				   struct ttm_operation_ctx *ctx,
 642				   struct ww_acquire_ctx *ticket)
 643{
 644	int r;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 645
 646	if (!busy_bo || !ticket)
 
 
 
 
 
 
 
 
 647		return -EBUSY;
 648
 649	if (ctx->interruptible)
 650		r = dma_resv_lock_interruptible(busy_bo->base.resv,
 651							  ticket);
 652	else
 653		r = dma_resv_lock(busy_bo->base.resv, ticket);
 654
 655	/*
 656	 * TODO: It would be better to keep the BO locked until allocation is at
 657	 * least tried one more time, but that would mean a much larger rework
 658	 * of TTM.
 659	 */
 660	if (!r)
 661		dma_resv_unlock(busy_bo->base.resv);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 662
 663	return r == -EDEADLK ? -EBUSY : r;
 
 
 
 
 
 
 
 
 
 
 
 
 
 664}
 665
 666int ttm_mem_evict_first(struct ttm_device *bdev,
 667			struct ttm_resource_manager *man,
 668			const struct ttm_place *place,
 669			struct ttm_operation_ctx *ctx,
 670			struct ww_acquire_ctx *ticket)
 
 
 
 671{
 672	struct ttm_buffer_object *bo = NULL, *busy_bo = NULL;
 673	bool locked = false;
 674	unsigned i;
 675	int ret;
 
 
 
 
 
 
 676
 677	spin_lock(&bdev->lru_lock);
 678	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
 679		list_for_each_entry(bo, &man->lru[i], lru) {
 680			bool busy;
 681
 682			if (!ttm_bo_evict_swapout_allowable(bo, ctx, place,
 683							    &locked, &busy)) {
 684				if (busy && !busy_bo && ticket !=
 685				    dma_resv_locking_ctx(bo->base.resv))
 686					busy_bo = bo;
 687				continue;
 688			}
 689
 690			if (!ttm_bo_get_unless_zero(bo)) {
 691				if (locked)
 692					dma_resv_unlock(bo->base.resv);
 693				continue;
 694			}
 695			break;
 696		}
 697
 698		/* If the inner loop terminated early, we have our candidate */
 699		if (&bo->lru != &man->lru[i])
 700			break;
 701
 702		bo = NULL;
 703	}
 704
 705	if (!bo) {
 706		if (busy_bo && !ttm_bo_get_unless_zero(busy_bo))
 707			busy_bo = NULL;
 708		spin_unlock(&bdev->lru_lock);
 709		ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket);
 710		if (busy_bo)
 711			ttm_bo_put(busy_bo);
 712		return ret;
 713	}
 714
 715	if (bo->deleted) {
 716		ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
 717					  ctx->no_wait_gpu, locked);
 718		ttm_bo_put(bo);
 719		return ret;
 720	}
 721
 722	spin_unlock(&bdev->lru_lock);
 723
 724	ret = ttm_bo_evict(bo, ctx);
 725	if (locked)
 726		ttm_bo_unreserve(bo);
 727
 728	ttm_bo_put(bo);
 729	return ret;
 730}
 
 731
 732/*
 733 * Add the last move fence to the BO and reserve a new shared slot. We only use
 734 * a shared slot to avoid unecessary sync and rely on the subsequent bo move to
 735 * either stall or use an exclusive fence respectively set bo->moving.
 736 */
 737static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
 738				 struct ttm_resource_manager *man,
 739				 struct ttm_resource *mem,
 740				 bool no_wait_gpu)
 741{
 742	struct dma_fence *fence;
 743	int ret;
 744
 745	spin_lock(&man->move_lock);
 746	fence = dma_fence_get(man->move);
 747	spin_unlock(&man->move_lock);
 748
 749	if (!fence)
 750		return 0;
 751
 752	if (no_wait_gpu) {
 753		ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
 754		dma_fence_put(fence);
 755		return ret;
 756	}
 757
 758	dma_resv_add_shared_fence(bo->base.resv, fence);
 759
 760	ret = dma_resv_reserve_shared(bo->base.resv, 1);
 761	if (unlikely(ret)) {
 762		dma_fence_put(fence);
 763		return ret;
 764	}
 765
 766	dma_fence_put(bo->moving);
 767	bo->moving = fence;
 768	return 0;
 769}
 770
 771/*
 772 * Repeatedly evict memory from the LRU for @mem_type until we create enough
 773 * space, or we've evicted everything and there isn't enough space.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 774 */
 775static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
 776				  const struct ttm_place *place,
 777				  struct ttm_resource **mem,
 778				  struct ttm_operation_ctx *ctx)
 
 779{
 780	struct ttm_device *bdev = bo->bdev;
 781	struct ttm_resource_manager *man;
 782	struct ww_acquire_ctx *ticket;
 783	int ret;
 784
 785	man = ttm_manager_type(bdev, place->mem_type);
 786	ticket = dma_resv_locking_ctx(bo->base.resv);
 787	do {
 788		ret = ttm_resource_alloc(bo, place, mem);
 789		if (likely(!ret))
 790			break;
 791		if (unlikely(ret != -ENOSPC))
 792			return ret;
 793		ret = ttm_mem_evict_first(bdev, man, place, ctx,
 794					  ticket);
 795		if (unlikely(ret != 0))
 796			return ret;
 797	} while (1);
 798
 799	return ttm_bo_add_move_fence(bo, man, *mem, ctx->no_wait_gpu);
 800}
 801
 802/*
 803 * Creates space for memory region @mem according to its type.
 804 *
 805 * This function first searches for free space in compatible memory types in
 806 * the priority order defined by the driver.  If free space isn't found, then
 807 * ttm_bo_mem_force_space is attempted in priority order to evict and find
 808 * space.
 809 */
 810int ttm_bo_mem_space(struct ttm_buffer_object *bo,
 811			struct ttm_placement *placement,
 812			struct ttm_resource **mem,
 813			struct ttm_operation_ctx *ctx)
 814{
 815	struct ttm_device *bdev = bo->bdev;
 816	bool type_found = false;
 817	int i, ret;
 818
 819	ret = dma_resv_reserve_shared(bo->base.resv, 1);
 820	if (unlikely(ret))
 821		return ret;
 822
 823	for (i = 0; i < placement->num_placement; ++i) {
 824		const struct ttm_place *place = &placement->placement[i];
 825		struct ttm_resource_manager *man;
 
 826
 827		man = ttm_manager_type(bdev, place->mem_type);
 828		if (!man || !ttm_resource_manager_used(man))
 829			continue;
 830
 831		type_found = true;
 832		ret = ttm_resource_alloc(bo, place, mem);
 833		if (ret == -ENOSPC)
 834			continue;
 835		if (unlikely(ret))
 836			goto error;
 837
 838		ret = ttm_bo_add_move_fence(bo, man, *mem, ctx->no_wait_gpu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 839		if (unlikely(ret)) {
 840			ttm_resource_free(bo, mem);
 841			if (ret == -EBUSY)
 842				continue;
 843
 844			goto error;
 845		}
 846		return 0;
 847	}
 848
 849	for (i = 0; i < placement->num_busy_placement; ++i) {
 850		const struct ttm_place *place = &placement->busy_placement[i];
 851		struct ttm_resource_manager *man;
 852
 853		man = ttm_manager_type(bdev, place->mem_type);
 854		if (!man || !ttm_resource_manager_used(man))
 855			continue;
 856
 857		type_found = true;
 858		ret = ttm_bo_mem_force_space(bo, place, mem, ctx);
 859		if (likely(!ret))
 860			return 0;
 861
 862		if (ret && ret != -EBUSY)
 863			goto error;
 864	}
 865
 866	ret = -ENOMEM;
 867	if (!type_found) {
 868		pr_err(TTM_PFX "No compatible memory type found\n");
 869		ret = -EINVAL;
 870	}
 871
 872error:
 873	if (bo->resource->mem_type == TTM_PL_SYSTEM && !bo->pin_count)
 874		ttm_bo_move_to_lru_tail_unlocked(bo);
 875
 876	return ret;
 877}
 878EXPORT_SYMBOL(ttm_bo_mem_space);
 879
 880static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
 881			      struct ttm_placement *placement,
 882			      struct ttm_operation_ctx *ctx)
 
 
 
 
 
 
 
 
 
 
 
 
 883{
 884	struct ttm_resource *mem;
 885	struct ttm_place hop;
 886	int ret;
 887
 888	dma_resv_assert_held(bo->base.resv);
 
 
 
 
 889
 890	/*
 891	 * Determine where to move the buffer.
 892	 *
 893	 * If driver determines move is going to need
 894	 * an extra step then it will return -EMULTIHOP
 895	 * and the buffer will be moved to the temporary
 896	 * stop and the driver will be called to make
 897	 * the second hop.
 898	 */
 899	ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
 900	if (ret)
 901		return ret;
 902bounce:
 903	ret = ttm_bo_handle_move_mem(bo, mem, false, ctx, &hop);
 904	if (ret == -EMULTIHOP) {
 905		ret = ttm_bo_bounce_temp_buffer(bo, &mem, ctx, &hop);
 906		if (ret)
 907			goto out;
 908		/* try and move to final place now. */
 909		goto bounce;
 910	}
 911out:
 912	if (ret)
 913		ttm_resource_free(bo, &mem);
 914	return ret;
 915}
 
 916
 917static bool ttm_bo_places_compat(const struct ttm_place *places,
 918				 unsigned num_placement,
 919				 struct ttm_resource *mem,
 920				 uint32_t *new_flags)
 921{
 922	unsigned i;
 923
 924	for (i = 0; i < num_placement; i++) {
 925		const struct ttm_place *heap = &places[i];
 926
 927		if ((mem->start < heap->fpfn ||
 928		     (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
 929			continue;
 930
 931		*new_flags = heap->flags;
 932		if ((mem->mem_type == heap->mem_type) &&
 933		    (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
 934		     (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
 935			return true;
 936	}
 937	return false;
 938}
 939
 940bool ttm_bo_mem_compat(struct ttm_placement *placement,
 941		       struct ttm_resource *mem,
 942		       uint32_t *new_flags)
 943{
 944	if (ttm_bo_places_compat(placement->placement, placement->num_placement,
 945				 mem, new_flags))
 946		return true;
 947
 948	if ((placement->busy_placement != placement->placement ||
 949	     placement->num_busy_placement > placement->num_placement) &&
 950	    ttm_bo_places_compat(placement->busy_placement,
 951				 placement->num_busy_placement,
 952				 mem, new_flags))
 953		return true;
 954
 955	return false;
 956}
 957EXPORT_SYMBOL(ttm_bo_mem_compat);
 958
 959int ttm_bo_validate(struct ttm_buffer_object *bo,
 960		    struct ttm_placement *placement,
 961		    struct ttm_operation_ctx *ctx)
 962{
 
 
 
 963	int ret;
 964	uint32_t new_flags;
 965
 966	dma_resv_assert_held(bo->base.resv);
 967
 968	/*
 969	 * Remove the backing store if no placement is given.
 970	 */
 971	if (!placement->num_placement && !placement->num_busy_placement)
 972		return ttm_bo_pipeline_gutting(bo);
 973
 974	/*
 975	 * Check whether we need to move buffer.
 976	 */
 977	if (!ttm_bo_mem_compat(placement, bo->resource, &new_flags)) {
 978		ret = ttm_bo_move_buffer(bo, placement, ctx);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 979		if (ret)
 980			return ret;
 981	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 982	/*
 983	 * We might need to add a TTM.
 984	 */
 985	if (bo->resource->mem_type == TTM_PL_SYSTEM) {
 986		ret = ttm_tt_create(bo, true);
 987		if (ret)
 988			return ret;
 989	}
 990	return 0;
 991}
 992EXPORT_SYMBOL(ttm_bo_validate);
 993
 994int ttm_bo_init_reserved(struct ttm_device *bdev,
 995			 struct ttm_buffer_object *bo,
 996			 size_t size,
 997			 enum ttm_bo_type type,
 998			 struct ttm_placement *placement,
 999			 uint32_t page_alignment,
1000			 struct ttm_operation_ctx *ctx,
1001			 struct sg_table *sg,
1002			 struct dma_resv *resv,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1003			 void (*destroy) (struct ttm_buffer_object *))
1004{
1005	static const struct ttm_place sys_mem = { .mem_type = TTM_PL_SYSTEM };
1006	bool locked;
1007	int ret;
1008
1009	bo->destroy = destroy ? destroy : ttm_bo_default_destroy;
1010
1011	kref_init(&bo->kref);
1012	INIT_LIST_HEAD(&bo->lru);
1013	INIT_LIST_HEAD(&bo->ddestroy);
1014	bo->bdev = bdev;
1015	bo->type = type;
1016	bo->page_alignment = page_alignment;
1017	bo->moving = NULL;
1018	bo->pin_count = 0;
1019	bo->sg = sg;
1020	if (resv) {
 
1021		bo->base.resv = resv;
1022		dma_resv_assert_held(bo->base.resv);
1023	} else {
1024		bo->base.resv = &bo->base._resv;
1025	}
1026	atomic_inc(&ttm_glob.bo_count);
1027
1028	ret = ttm_resource_alloc(bo, &sys_mem, &bo->resource);
1029	if (unlikely(ret)) {
1030		ttm_bo_put(bo);
1031		return ret;
1032	}
1033
1034	/*
1035	 * For ttm_bo_type_device buffers, allocate
1036	 * address space from the device.
1037	 */
1038	if (bo->type == ttm_bo_type_device ||
1039	    bo->type == ttm_bo_type_sg)
1040		ret = drm_vma_offset_add(bdev->vma_manager, &bo->base.vma_node,
1041					 bo->resource->num_pages);
 
 
 
1042
1043	/* passed reservation objects should already be locked,
1044	 * since otherwise lockdep will be angered in radeon.
1045	 */
1046	if (!resv) {
1047		locked = dma_resv_trylock(bo->base.resv);
1048		WARN_ON(!locked);
1049	}
1050
1051	if (likely(!ret))
1052		ret = ttm_bo_validate(bo, placement, ctx);
 
1053
1054	if (unlikely(ret)) {
1055		if (!resv)
1056			ttm_bo_unreserve(bo);
1057
1058		ttm_bo_put(bo);
1059		return ret;
1060	}
1061
1062	ttm_bo_move_to_lru_tail_unlocked(bo);
1063
 
 
1064	return ret;
1065}
1066EXPORT_SYMBOL(ttm_bo_init_reserved);
1067
1068int ttm_bo_init(struct ttm_device *bdev,
1069		struct ttm_buffer_object *bo,
1070		size_t size,
1071		enum ttm_bo_type type,
1072		struct ttm_placement *placement,
1073		uint32_t page_alignment,
1074		bool interruptible,
1075		struct sg_table *sg,
1076		struct dma_resv *resv,
1077		void (*destroy) (struct ttm_buffer_object *))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1078{
1079	struct ttm_operation_ctx ctx = { interruptible, false };
1080	int ret;
1081
1082	ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1083				   page_alignment, &ctx, sg, resv, destroy);
1084	if (ret)
1085		return ret;
1086
1087	if (!resv)
1088		ttm_bo_unreserve(bo);
1089
1090	return 0;
1091}
1092EXPORT_SYMBOL(ttm_bo_init);
1093
1094/*
1095 * buffer object vm functions.
1096 */
1097
 
 
 
 
 
1098void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1099{
1100	struct ttm_device *bdev = bo->bdev;
1101
1102	drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping);
1103	ttm_mem_io_free(bdev, bo->resource);
1104}
1105EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1106
1107int ttm_bo_wait(struct ttm_buffer_object *bo,
1108		bool interruptible, bool no_wait)
 
 
 
 
 
 
 
 
 
1109{
1110	long timeout = 15 * HZ;
1111
1112	if (no_wait) {
1113		if (dma_resv_test_signaled(bo->base.resv, true))
 
1114			return 0;
1115		else
1116			return -EBUSY;
1117	}
1118
1119	timeout = dma_resv_wait_timeout(bo->base.resv, true, interruptible,
1120					timeout);
1121	if (timeout < 0)
1122		return timeout;
1123
1124	if (timeout == 0)
1125		return -EBUSY;
1126
1127	dma_resv_add_excl_fence(bo->base.resv, NULL);
1128	return 0;
1129}
1130EXPORT_SYMBOL(ttm_bo_wait);
1131
1132int ttm_bo_swapout(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx,
1133		   gfp_t gfp_flags)
1134{
1135	struct ttm_place place;
1136	bool locked;
1137	int ret;
 
 
 
 
 
 
 
 
 
 
 
 
1138
1139	/*
1140	 * While the bo may already reside in SYSTEM placement, set
1141	 * SYSTEM as new placement to cover also the move further below.
1142	 * The driver may use the fact that we're moving from SYSTEM
1143	 * as an indication that we're about to swap out.
1144	 */
1145	memset(&place, 0, sizeof(place));
1146	place.mem_type = TTM_PL_SYSTEM;
1147	if (!ttm_bo_evict_swapout_allowable(bo, ctx, &place, &locked, NULL))
1148		return -EBUSY;
1149
1150	if (!bo->ttm || !ttm_tt_is_populated(bo->ttm) ||
1151	    bo->ttm->page_flags & TTM_PAGE_FLAG_SG ||
1152	    bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED ||
1153	    !ttm_bo_get_unless_zero(bo)) {
1154		if (locked)
1155			dma_resv_unlock(bo->base.resv);
1156		return -EBUSY;
1157	}
1158
1159	if (bo->deleted) {
1160		ret = ttm_bo_cleanup_refs(bo, false, false, locked);
1161		ttm_bo_put(bo);
1162		return ret == -EBUSY ? -ENOSPC : ret;
 
 
 
 
 
 
1163	}
1164
1165	ttm_bo_del_from_lru(bo);
1166	/* TODO: Cleanup the locking */
1167	spin_unlock(&bo->bdev->lru_lock);
1168
1169	/*
1170	 * Move to system cached
1171	 */
1172	if (bo->resource->mem_type != TTM_PL_SYSTEM) {
1173		struct ttm_operation_ctx ctx = { false, false };
1174		struct ttm_resource *evict_mem;
1175		struct ttm_place hop;
1176
1177		memset(&hop, 0, sizeof(hop));
 
1178		ret = ttm_resource_alloc(bo, &place, &evict_mem);
1179		if (unlikely(ret))
1180			goto out;
1181
1182		ret = ttm_bo_handle_move_mem(bo, evict_mem, true, &ctx, &hop);
1183		if (unlikely(ret != 0)) {
1184			WARN(ret == -EMULTIHOP, "Unexpected multihop in swaput - likely driver bug.\n");
 
 
1185			goto out;
1186		}
1187	}
1188
1189	/*
1190	 * Make sure BO is idle.
1191	 */
1192	ret = ttm_bo_wait(bo, false, false);
1193	if (unlikely(ret != 0))
1194		goto out;
1195
1196	ttm_bo_unmap_virtual(bo);
1197
1198	/*
1199	 * Swap out. Buffer will be swapped in again as soon as
1200	 * anyone tries to access a ttm page.
1201	 */
1202	if (bo->bdev->funcs->swap_notify)
1203		bo->bdev->funcs->swap_notify(bo);
1204
1205	if (ttm_tt_is_populated(bo->ttm))
1206		ret = ttm_tt_swapout(bo->bdev, bo->ttm, gfp_flags);
 
 
 
 
 
 
 
 
 
 
 
 
1207out:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1208
1209	/*
1210	 * Unreserve without putting on LRU to avoid swapping out an
1211	 * already swapped buffer.
1212	 */
1213	if (locked)
1214		dma_resv_unlock(bo->base.resv);
1215	ttm_bo_put(bo);
1216	return ret == -EBUSY ? -ENOSPC : ret;
1217}
1218
1219void ttm_bo_tt_destroy(struct ttm_buffer_object *bo)
1220{
1221	if (bo->ttm == NULL)
1222		return;
1223
 
1224	ttm_tt_destroy(bo->bdev, bo->ttm);
1225	bo->ttm = NULL;
1226}