1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   4 * All Rights Reserved.
   5 */
   6#include "xfs.h"
   7#include <linux/backing-dev.h>
   8#include <linux/dax.h>
   9
  10#include "xfs_shared.h"
  11#include "xfs_format.h"
  12#include "xfs_log_format.h"
  13#include "xfs_trans_resv.h"
  14#include "xfs_mount.h"
  15#include "xfs_trace.h"
  16#include "xfs_log.h"
  17#include "xfs_log_recover.h"
  18#include "xfs_log_priv.h"
  19#include "xfs_trans.h"
  20#include "xfs_buf_item.h"
  21#include "xfs_errortag.h"
  22#include "xfs_error.h"
  23#include "xfs_ag.h"
  24#include "xfs_buf_mem.h"
  25
  26struct kmem_cache *xfs_buf_cache;
  27
  28/*
  29 * Locking orders
  30 *
  31 * xfs_buf_ioacct_inc:
  32 * xfs_buf_ioacct_dec:
  33 *	b_sema (caller holds)
  34 *	  b_lock
  35 *
  36 * xfs_buf_stale:
  37 *	b_sema (caller holds)
  38 *	  b_lock
  39 *	    lru_lock
  40 *
  41 * xfs_buf_rele:
  42 *	b_lock
  43 *	  pag_buf_lock
  44 *	    lru_lock
  45 *
  46 * xfs_buftarg_drain_rele
  47 *	lru_lock
  48 *	  b_lock (trylock due to inversion)
  49 *
  50 * xfs_buftarg_isolate
  51 *	lru_lock
  52 *	  b_lock (trylock due to inversion)
  53 */
  54
  55static int __xfs_buf_submit(struct xfs_buf *bp, bool wait);
  56
  57static inline int
  58xfs_buf_submit(
  59	struct xfs_buf		*bp)
  60{
  61	return __xfs_buf_submit(bp, !(bp->b_flags & XBF_ASYNC));
  62}
  63
  64static inline bool xfs_buf_is_uncached(struct xfs_buf *bp)
  65{
  66	return bp->b_rhash_key == XFS_BUF_DADDR_NULL;
  67}
  68
  69static inline int
  70xfs_buf_is_vmapped(
  71	struct xfs_buf	*bp)
  72{
  73	/*
  74	 * Return true if the buffer is vmapped.
  75	 *
  76	 * b_addr is null if the buffer is not mapped, but the code is clever
  77	 * enough to know it doesn't have to map a single page, so the check has
  78	 * to be both for b_addr and bp->b_page_count > 1.
  79	 */
  80	return bp->b_addr && bp->b_page_count > 1;
  81}
  82
  83static inline int
  84xfs_buf_vmap_len(
  85	struct xfs_buf	*bp)
  86{
  87	return (bp->b_page_count * PAGE_SIZE);
  88}
  89
  90/*
  91 * Bump the I/O in flight count on the buftarg if we haven't yet done so for
  92 * this buffer. The count is incremented once per buffer (per hold cycle)
  93 * because the corresponding decrement is deferred to buffer release. Buffers
  94 * can undergo I/O multiple times in a hold-release cycle and per buffer I/O
  95 * tracking adds unnecessary overhead. This is used for sychronization purposes
  96 * with unmount (see xfs_buftarg_drain()), so all we really need is a count of
  97 * in-flight buffers.
  98 *
  99 * Buffers that are never released (e.g., superblock, iclog buffers) must set
 100 * the XBF_NO_IOACCT flag before I/O submission. Otherwise, the buftarg count
 101 * never reaches zero and unmount hangs indefinitely.
 102 */
 103static inline void
 104xfs_buf_ioacct_inc(
 105	struct xfs_buf	*bp)
 106{
 107	if (bp->b_flags & XBF_NO_IOACCT)
 108		return;
 109
 110	ASSERT(bp->b_flags & XBF_ASYNC);
 111	spin_lock(&bp->b_lock);
 112	if (!(bp->b_state & XFS_BSTATE_IN_FLIGHT)) {
 113		bp->b_state |= XFS_BSTATE_IN_FLIGHT;
 114		percpu_counter_inc(&bp->b_target->bt_io_count);
 115	}
 116	spin_unlock(&bp->b_lock);
 117}
 118
 119/*
 120 * Clear the in-flight state on a buffer about to be released to the LRU or
 121 * freed and unaccount from the buftarg.
 122 */
 123static inline void
 124__xfs_buf_ioacct_dec(
 125	struct xfs_buf	*bp)
 126{
 127	lockdep_assert_held(&bp->b_lock);
 128
 129	if (bp->b_state & XFS_BSTATE_IN_FLIGHT) {
 130		bp->b_state &= ~XFS_BSTATE_IN_FLIGHT;
 131		percpu_counter_dec(&bp->b_target->bt_io_count);
 132	}
 133}
 134
 135static inline void
 136xfs_buf_ioacct_dec(
 137	struct xfs_buf	*bp)
 138{
 139	spin_lock(&bp->b_lock);
 140	__xfs_buf_ioacct_dec(bp);
 141	spin_unlock(&bp->b_lock);
 142}
 143
 144/*
 145 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
 146 * b_lru_ref count so that the buffer is freed immediately when the buffer
 147 * reference count falls to zero. If the buffer is already on the LRU, we need
 148 * to remove the reference that LRU holds on the buffer.
 149 *
 150 * This prevents build-up of stale buffers on the LRU.
 151 */
 152void
 153xfs_buf_stale(
 154	struct xfs_buf	*bp)
 155{
 156	ASSERT(xfs_buf_islocked(bp));
 157
 158	bp->b_flags |= XBF_STALE;
 159
 160	/*
 161	 * Clear the delwri status so that a delwri queue walker will not
 162	 * flush this buffer to disk now that it is stale. The delwri queue has
 163	 * a reference to the buffer, so this is safe to do.
 164	 */
 165	bp->b_flags &= ~_XBF_DELWRI_Q;
 166
 167	/*
 168	 * Once the buffer is marked stale and unlocked, a subsequent lookup
 169	 * could reset b_flags. There is no guarantee that the buffer is
 170	 * unaccounted (released to LRU) before that occurs. Drop in-flight
 171	 * status now to preserve accounting consistency.
 172	 */
 173	spin_lock(&bp->b_lock);
 174	__xfs_buf_ioacct_dec(bp);
 175
 176	atomic_set(&bp->b_lru_ref, 0);
 177	if (!(bp->b_state & XFS_BSTATE_DISPOSE) &&
 178	    (list_lru_del_obj(&bp->b_target->bt_lru, &bp->b_lru)))
 179		atomic_dec(&bp->b_hold);
 180
 181	ASSERT(atomic_read(&bp->b_hold) >= 1);
 182	spin_unlock(&bp->b_lock);
 183}
 184
 185static int
 186xfs_buf_get_maps(
 187	struct xfs_buf		*bp,
 188	int			map_count)
 189{
 190	ASSERT(bp->b_maps == NULL);
 191	bp->b_map_count = map_count;
 192
 193	if (map_count == 1) {
 194		bp->b_maps = &bp->__b_map;
 195		return 0;
 196	}
 197
 198	bp->b_maps = kzalloc(map_count * sizeof(struct xfs_buf_map),
 199			GFP_KERNEL | __GFP_NOLOCKDEP | __GFP_NOFAIL);
 200	if (!bp->b_maps)
 201		return -ENOMEM;
 202	return 0;
 203}
 204
 205/*
 206 *	Frees b_pages if it was allocated.
 207 */
 208static void
 209xfs_buf_free_maps(
 210	struct xfs_buf	*bp)
 211{
 212	if (bp->b_maps != &bp->__b_map) {
 213		kfree(bp->b_maps);
 214		bp->b_maps = NULL;
 215	}
 216}
 217
 218static int
 219_xfs_buf_alloc(
 220	struct xfs_buftarg	*target,
 221	struct xfs_buf_map	*map,
 222	int			nmaps,
 223	xfs_buf_flags_t		flags,
 224	struct xfs_buf		**bpp)
 225{
 226	struct xfs_buf		*bp;
 227	int			error;
 228	int			i;
 229
 230	*bpp = NULL;
 231	bp = kmem_cache_zalloc(xfs_buf_cache,
 232			GFP_KERNEL | __GFP_NOLOCKDEP | __GFP_NOFAIL);
 233
 234	/*
 235	 * We don't want certain flags to appear in b_flags unless they are
 236	 * specifically set by later operations on the buffer.
 237	 */
 238	flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD);
 239
 240	atomic_set(&bp->b_hold, 1);
 241	atomic_set(&bp->b_lru_ref, 1);
 242	init_completion(&bp->b_iowait);
 243	INIT_LIST_HEAD(&bp->b_lru);
 244	INIT_LIST_HEAD(&bp->b_list);
 245	INIT_LIST_HEAD(&bp->b_li_list);
 246	sema_init(&bp->b_sema, 0); /* held, no waiters */
 247	spin_lock_init(&bp->b_lock);
 248	bp->b_target = target;
 249	bp->b_mount = target->bt_mount;
 250	bp->b_flags = flags;
 251
 252	/*
 253	 * Set length and io_length to the same value initially.
 254	 * I/O routines should use io_length, which will be the same in
 255	 * most cases but may be reset (e.g. XFS recovery).
 256	 */
 257	error = xfs_buf_get_maps(bp, nmaps);
 258	if (error)  {
 259		kmem_cache_free(xfs_buf_cache, bp);
 260		return error;
 261	}
 262
 263	bp->b_rhash_key = map[0].bm_bn;
 264	bp->b_length = 0;
 265	for (i = 0; i < nmaps; i++) {
 266		bp->b_maps[i].bm_bn = map[i].bm_bn;
 267		bp->b_maps[i].bm_len = map[i].bm_len;
 268		bp->b_length += map[i].bm_len;
 269	}
 270
 271	atomic_set(&bp->b_pin_count, 0);
 272	init_waitqueue_head(&bp->b_waiters);
 273
 274	XFS_STATS_INC(bp->b_mount, xb_create);
 275	trace_xfs_buf_init(bp, _RET_IP_);
 276
 277	*bpp = bp;
 278	return 0;
 279}
 280
 281static void
 282xfs_buf_free_pages(
 283	struct xfs_buf	*bp)
 284{
 285	uint		i;
 286
 287	ASSERT(bp->b_flags & _XBF_PAGES);
 288
 289	if (xfs_buf_is_vmapped(bp))
 290		vm_unmap_ram(bp->b_addr, bp->b_page_count);
 291
 292	for (i = 0; i < bp->b_page_count; i++) {
 293		if (bp->b_pages[i])
 294			__free_page(bp->b_pages[i]);
 295	}
 296	mm_account_reclaimed_pages(bp->b_page_count);
 297
 298	if (bp->b_pages != bp->b_page_array)
 299		kfree(bp->b_pages);
 300	bp->b_pages = NULL;
 301	bp->b_flags &= ~_XBF_PAGES;
 302}
 303
 304static void
 305xfs_buf_free_callback(
 306	struct callback_head	*cb)
 307{
 308	struct xfs_buf		*bp = container_of(cb, struct xfs_buf, b_rcu);
 309
 310	xfs_buf_free_maps(bp);
 311	kmem_cache_free(xfs_buf_cache, bp);
 312}
 313
 314static void
 315xfs_buf_free(
 316	struct xfs_buf		*bp)
 317{
 318	trace_xfs_buf_free(bp, _RET_IP_);
 319
 320	ASSERT(list_empty(&bp->b_lru));
 321
 322	if (xfs_buftarg_is_mem(bp->b_target))
 323		xmbuf_unmap_page(bp);
 324	else if (bp->b_flags & _XBF_PAGES)
 325		xfs_buf_free_pages(bp);
 326	else if (bp->b_flags & _XBF_KMEM)
 327		kfree(bp->b_addr);
 328
 329	call_rcu(&bp->b_rcu, xfs_buf_free_callback);
 330}
 331
 332static int
 333xfs_buf_alloc_kmem(
 334	struct xfs_buf	*bp,
 335	xfs_buf_flags_t	flags)
 336{
 337	gfp_t		gfp_mask = GFP_KERNEL | __GFP_NOLOCKDEP | __GFP_NOFAIL;
 338	size_t		size = BBTOB(bp->b_length);
 339
 340	/* Assure zeroed buffer for non-read cases. */
 341	if (!(flags & XBF_READ))
 342		gfp_mask |= __GFP_ZERO;
 343
 344	bp->b_addr = kmalloc(size, gfp_mask);
 345	if (!bp->b_addr)
 346		return -ENOMEM;
 347
 348	if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) !=
 349	    ((unsigned long)bp->b_addr & PAGE_MASK)) {
 350		/* b_addr spans two pages - use alloc_page instead */
 351		kfree(bp->b_addr);
 352		bp->b_addr = NULL;
 353		return -ENOMEM;
 354	}
 355	bp->b_offset = offset_in_page(bp->b_addr);
 356	bp->b_pages = bp->b_page_array;
 357	bp->b_pages[0] = kmem_to_page(bp->b_addr);
 358	bp->b_page_count = 1;
 359	bp->b_flags |= _XBF_KMEM;
 360	return 0;
 361}
 362
 363static int
 364xfs_buf_alloc_pages(
 365	struct xfs_buf	*bp,
 366	xfs_buf_flags_t	flags)
 367{
 368	gfp_t		gfp_mask = GFP_KERNEL | __GFP_NOLOCKDEP | __GFP_NOWARN;
 369	long		filled = 0;
 370
 371	if (flags & XBF_READ_AHEAD)
 372		gfp_mask |= __GFP_NORETRY;
 373
 374	/* Make sure that we have a page list */
 375	bp->b_page_count = DIV_ROUND_UP(BBTOB(bp->b_length), PAGE_SIZE);
 376	if (bp->b_page_count <= XB_PAGES) {
 377		bp->b_pages = bp->b_page_array;
 378	} else {
 379		bp->b_pages = kzalloc(sizeof(struct page *) * bp->b_page_count,
 380					gfp_mask);
 381		if (!bp->b_pages)
 382			return -ENOMEM;
 383	}
 384	bp->b_flags |= _XBF_PAGES;
 385
 386	/* Assure zeroed buffer for non-read cases. */
 387	if (!(flags & XBF_READ))
 388		gfp_mask |= __GFP_ZERO;
 389
 390	/*
 391	 * Bulk filling of pages can take multiple calls. Not filling the entire
 392	 * array is not an allocation failure, so don't back off if we get at
 393	 * least one extra page.
 394	 */
 395	for (;;) {
 396		long	last = filled;
 397
 398		filled = alloc_pages_bulk_array(gfp_mask, bp->b_page_count,
 399						bp->b_pages);
 400		if (filled == bp->b_page_count) {
 401			XFS_STATS_INC(bp->b_mount, xb_page_found);
 402			break;
 403		}
 404
 405		if (filled != last)
 406			continue;
 407
 408		if (flags & XBF_READ_AHEAD) {
 409			xfs_buf_free_pages(bp);
 410			return -ENOMEM;
 411		}
 412
 413		XFS_STATS_INC(bp->b_mount, xb_page_retries);
 414		memalloc_retry_wait(gfp_mask);
 415	}
 416	return 0;
 417}
 418
 419/*
 420 *	Map buffer into kernel address-space if necessary.
 421 */
 422STATIC int
 423_xfs_buf_map_pages(
 424	struct xfs_buf		*bp,
 425	xfs_buf_flags_t		flags)
 426{
 427	ASSERT(bp->b_flags & _XBF_PAGES);
 428	if (bp->b_page_count == 1) {
 429		/* A single page buffer is always mappable */
 430		bp->b_addr = page_address(bp->b_pages[0]);
 431	} else if (flags & XBF_UNMAPPED) {
 432		bp->b_addr = NULL;
 433	} else {
 434		int retried = 0;
 435		unsigned nofs_flag;
 436
 437		/*
 438		 * vm_map_ram() will allocate auxiliary structures (e.g.
 439		 * pagetables) with GFP_KERNEL, yet we often under a scoped nofs
 440		 * context here. Mixing GFP_KERNEL with GFP_NOFS allocations
 441		 * from the same call site that can be run from both above and
 442		 * below memory reclaim causes lockdep false positives. Hence we
 443		 * always need to force this allocation to nofs context because
 444		 * we can't pass __GFP_NOLOCKDEP down to auxillary structures to
 445		 * prevent false positive lockdep reports.
 446		 *
 447		 * XXX(dgc): I think dquot reclaim is the only place we can get
 448		 * to this function from memory reclaim context now. If we fix
 449		 * that like we've fixed inode reclaim to avoid writeback from
 450		 * reclaim, this nofs wrapping can go away.
 451		 */
 452		nofs_flag = memalloc_nofs_save();
 453		do {
 454			bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
 455						-1);
 456			if (bp->b_addr)
 457				break;
 458			vm_unmap_aliases();
 459		} while (retried++ <= 1);
 460		memalloc_nofs_restore(nofs_flag);
 461
 462		if (!bp->b_addr)
 463			return -ENOMEM;
 464	}
 465
 466	return 0;
 467}
 468
 469/*
 470 *	Finding and Reading Buffers
 471 */
 472static int
 473_xfs_buf_obj_cmp(
 474	struct rhashtable_compare_arg	*arg,
 475	const void			*obj)
 476{
 477	const struct xfs_buf_map	*map = arg->key;
 478	const struct xfs_buf		*bp = obj;
 479
 480	/*
 481	 * The key hashing in the lookup path depends on the key being the
 482	 * first element of the compare_arg, make sure to assert this.
 483	 */
 484	BUILD_BUG_ON(offsetof(struct xfs_buf_map, bm_bn) != 0);
 485
 486	if (bp->b_rhash_key != map->bm_bn)
 487		return 1;
 488
 489	if (unlikely(bp->b_length != map->bm_len)) {
 490		/*
 491		 * found a block number match. If the range doesn't
 492		 * match, the only way this is allowed is if the buffer
 493		 * in the cache is stale and the transaction that made
 494		 * it stale has not yet committed. i.e. we are
 495		 * reallocating a busy extent. Skip this buffer and
 496		 * continue searching for an exact match.
 
 
 
 497		 */
 498		if (!(map->bm_flags & XBM_LIVESCAN))
 499			ASSERT(bp->b_flags & XBF_STALE);
 500		return 1;
 501	}
 502	return 0;
 503}
 504
 505static const struct rhashtable_params xfs_buf_hash_params = {
 506	.min_size		= 32,	/* empty AGs have minimal footprint */
 507	.nelem_hint		= 16,
 508	.key_len		= sizeof(xfs_daddr_t),
 509	.key_offset		= offsetof(struct xfs_buf, b_rhash_key),
 510	.head_offset		= offsetof(struct xfs_buf, b_rhash_head),
 511	.automatic_shrinking	= true,
 512	.obj_cmpfn		= _xfs_buf_obj_cmp,
 513};
 514
 515int
 516xfs_buf_cache_init(
 517	struct xfs_buf_cache	*bch)
 518{
 519	spin_lock_init(&bch->bc_lock);
 520	return rhashtable_init(&bch->bc_hash, &xfs_buf_hash_params);
 521}
 522
 523void
 524xfs_buf_cache_destroy(
 525	struct xfs_buf_cache	*bch)
 526{
 527	rhashtable_destroy(&bch->bc_hash);
 528}
 529
 530static int
 531xfs_buf_map_verify(
 532	struct xfs_buftarg	*btp,
 533	struct xfs_buf_map	*map)
 534{
 535	xfs_daddr_t		eofs;
 536
 537	/* Check for IOs smaller than the sector size / not sector aligned */
 538	ASSERT(!(BBTOB(map->bm_len) < btp->bt_meta_sectorsize));
 539	ASSERT(!(BBTOB(map->bm_bn) & (xfs_off_t)btp->bt_meta_sectormask));
 540
 541	/*
 542	 * Corrupted block numbers can get through to here, unfortunately, so we
 543	 * have to check that the buffer falls within the filesystem bounds.
 544	 */
 545	eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks);
 546	if (map->bm_bn < 0 || map->bm_bn >= eofs) {
 547		xfs_alert(btp->bt_mount,
 548			  "%s: daddr 0x%llx out of range, EOFS 0x%llx",
 549			  __func__, map->bm_bn, eofs);
 550		WARN_ON(1);
 551		return -EFSCORRUPTED;
 552	}
 553	return 0;
 554}
 555
 556static int
 557xfs_buf_find_lock(
 558	struct xfs_buf          *bp,
 559	xfs_buf_flags_t		flags)
 560{
 561	if (flags & XBF_TRYLOCK) {
 562		if (!xfs_buf_trylock(bp)) {
 563			XFS_STATS_INC(bp->b_mount, xb_busy_locked);
 564			return -EAGAIN;
 565		}
 566	} else {
 567		xfs_buf_lock(bp);
 568		XFS_STATS_INC(bp->b_mount, xb_get_locked_waited);
 569	}
 570
 571	/*
 572	 * if the buffer is stale, clear all the external state associated with
 573	 * it. We need to keep flags such as how we allocated the buffer memory
 574	 * intact here.
 575	 */
 576	if (bp->b_flags & XBF_STALE) {
 577		if (flags & XBF_LIVESCAN) {
 578			xfs_buf_unlock(bp);
 579			return -ENOENT;
 580		}
 581		ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
 582		bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
 583		bp->b_ops = NULL;
 584	}
 585	return 0;
 586}
 587
 588static inline int
 589xfs_buf_lookup(
 590	struct xfs_buf_cache	*bch,
 591	struct xfs_buf_map	*map,
 592	xfs_buf_flags_t		flags,
 593	struct xfs_buf		**bpp)
 594{
 595	struct xfs_buf          *bp;
 596	int			error;
 597
 598	rcu_read_lock();
 599	bp = rhashtable_lookup(&bch->bc_hash, map, xfs_buf_hash_params);
 600	if (!bp || !atomic_inc_not_zero(&bp->b_hold)) {
 601		rcu_read_unlock();
 602		return -ENOENT;
 603	}
 604	rcu_read_unlock();
 605
 606	error = xfs_buf_find_lock(bp, flags);
 607	if (error) {
 608		xfs_buf_rele(bp);
 609		return error;
 610	}
 611
 612	trace_xfs_buf_find(bp, flags, _RET_IP_);
 613	*bpp = bp;
 614	return 0;
 615}
 616
 617/*
 618 * Insert the new_bp into the hash table. This consumes the perag reference
 619 * taken for the lookup regardless of the result of the insert.
 620 */
 621static int
 622xfs_buf_find_insert(
 623	struct xfs_buftarg	*btp,
 624	struct xfs_buf_cache	*bch,
 625	struct xfs_perag	*pag,
 626	struct xfs_buf_map	*cmap,
 627	struct xfs_buf_map	*map,
 628	int			nmaps,
 629	xfs_buf_flags_t		flags,
 630	struct xfs_buf		**bpp)
 631{
 632	struct xfs_buf		*new_bp;
 633	struct xfs_buf		*bp;
 634	int			error;
 635
 636	error = _xfs_buf_alloc(btp, map, nmaps, flags, &new_bp);
 637	if (error)
 638		goto out_drop_pag;
 639
 640	if (xfs_buftarg_is_mem(new_bp->b_target)) {
 641		error = xmbuf_map_page(new_bp);
 642	} else if (BBTOB(new_bp->b_length) >= PAGE_SIZE ||
 643		   xfs_buf_alloc_kmem(new_bp, flags) < 0) {
 644		/*
 645		 * For buffers that fit entirely within a single page, first
 646		 * attempt to allocate the memory from the heap to minimise
 647		 * memory usage. If we can't get heap memory for these small
 648		 * buffers, we fall back to using the page allocator.
 649		 */
 650		error = xfs_buf_alloc_pages(new_bp, flags);
 651	}
 652	if (error)
 653		goto out_free_buf;
 654
 655	spin_lock(&bch->bc_lock);
 656	bp = rhashtable_lookup_get_insert_fast(&bch->bc_hash,
 657			&new_bp->b_rhash_head, xfs_buf_hash_params);
 658	if (IS_ERR(bp)) {
 659		error = PTR_ERR(bp);
 660		spin_unlock(&bch->bc_lock);
 661		goto out_free_buf;
 662	}
 663	if (bp) {
 664		/* found an existing buffer */
 665		atomic_inc(&bp->b_hold);
 666		spin_unlock(&bch->bc_lock);
 667		error = xfs_buf_find_lock(bp, flags);
 668		if (error)
 669			xfs_buf_rele(bp);
 670		else
 671			*bpp = bp;
 672		goto out_free_buf;
 673	}
 674
 675	/* The new buffer keeps the perag reference until it is freed. */
 676	new_bp->b_pag = pag;
 677	spin_unlock(&bch->bc_lock);
 678	*bpp = new_bp;
 679	return 0;
 680
 681out_free_buf:
 682	xfs_buf_free(new_bp);
 683out_drop_pag:
 684	if (pag)
 685		xfs_perag_put(pag);
 686	return error;
 687}
 688
 689static inline struct xfs_perag *
 690xfs_buftarg_get_pag(
 691	struct xfs_buftarg		*btp,
 692	const struct xfs_buf_map	*map)
 693{
 694	struct xfs_mount		*mp = btp->bt_mount;
 695
 696	if (xfs_buftarg_is_mem(btp))
 697		return NULL;
 698	return xfs_perag_get(mp, xfs_daddr_to_agno(mp, map->bm_bn));
 699}
 700
 701static inline struct xfs_buf_cache *
 702xfs_buftarg_buf_cache(
 703	struct xfs_buftarg		*btp,
 704	struct xfs_perag		*pag)
 705{
 706	if (pag)
 707		return &pag->pag_bcache;
 708	return btp->bt_cache;
 709}
 710
 711/*
 712 * Assembles a buffer covering the specified range. The code is optimised for
 713 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
 714 * more hits than misses.
 715 */
 716int
 717xfs_buf_get_map(
 718	struct xfs_buftarg	*btp,
 719	struct xfs_buf_map	*map,
 720	int			nmaps,
 721	xfs_buf_flags_t		flags,
 722	struct xfs_buf		**bpp)
 723{
 724	struct xfs_buf_cache	*bch;
 725	struct xfs_perag	*pag;
 726	struct xfs_buf		*bp = NULL;
 727	struct xfs_buf_map	cmap = { .bm_bn = map[0].bm_bn };
 728	int			error;
 729	int			i;
 730
 731	if (flags & XBF_LIVESCAN)
 732		cmap.bm_flags |= XBM_LIVESCAN;
 733	for (i = 0; i < nmaps; i++)
 734		cmap.bm_len += map[i].bm_len;
 735
 736	error = xfs_buf_map_verify(btp, &cmap);
 737	if (error)
 738		return error;
 739
 740	pag = xfs_buftarg_get_pag(btp, &cmap);
 741	bch = xfs_buftarg_buf_cache(btp, pag);
 742
 743	error = xfs_buf_lookup(bch, &cmap, flags, &bp);
 744	if (error && error != -ENOENT)
 745		goto out_put_perag;
 746
 747	/* cache hits always outnumber misses by at least 10:1 */
 748	if (unlikely(!bp)) {
 749		XFS_STATS_INC(btp->bt_mount, xb_miss_locked);
 750
 751		if (flags & XBF_INCORE)
 752			goto out_put_perag;
 753
 754		/* xfs_buf_find_insert() consumes the perag reference. */
 755		error = xfs_buf_find_insert(btp, bch, pag, &cmap, map, nmaps,
 756				flags, &bp);
 757		if (error)
 758			return error;
 759	} else {
 760		XFS_STATS_INC(btp->bt_mount, xb_get_locked);
 761		if (pag)
 762			xfs_perag_put(pag);
 763	}
 764
 765	/* We do not hold a perag reference anymore. */
 766	if (!bp->b_addr) {
 767		error = _xfs_buf_map_pages(bp, flags);
 768		if (unlikely(error)) {
 769			xfs_warn_ratelimited(btp->bt_mount,
 770				"%s: failed to map %u pages", __func__,
 771				bp->b_page_count);
 772			xfs_buf_relse(bp);
 773			return error;
 774		}
 775	}
 776
 777	/*
 778	 * Clear b_error if this is a lookup from a caller that doesn't expect
 779	 * valid data to be found in the buffer.
 780	 */
 781	if (!(flags & XBF_READ))
 782		xfs_buf_ioerror(bp, 0);
 783
 784	XFS_STATS_INC(btp->bt_mount, xb_get);
 785	trace_xfs_buf_get(bp, flags, _RET_IP_);
 786	*bpp = bp;
 787	return 0;
 788
 789out_put_perag:
 790	if (pag)
 791		xfs_perag_put(pag);
 792	return error;
 793}
 794
 795int
 796_xfs_buf_read(
 797	struct xfs_buf		*bp,
 798	xfs_buf_flags_t		flags)
 799{
 800	ASSERT(!(flags & XBF_WRITE));
 801	ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL);
 802
 803	bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD | XBF_DONE);
 804	bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
 805
 806	return xfs_buf_submit(bp);
 807}
 808
 809/*
 810 * Reverify a buffer found in cache without an attached ->b_ops.
 811 *
 812 * If the caller passed an ops structure and the buffer doesn't have ops
 813 * assigned, set the ops and use it to verify the contents. If verification
 814 * fails, clear XBF_DONE. We assume the buffer has no recorded errors and is
 815 * already in XBF_DONE state on entry.
 816 *
 817 * Under normal operations, every in-core buffer is verified on read I/O
 818 * completion. There are two scenarios that can lead to in-core buffers without
 819 * an assigned ->b_ops. The first is during log recovery of buffers on a V4
 820 * filesystem, though these buffers are purged at the end of recovery. The
 821 * other is online repair, which intentionally reads with a NULL buffer ops to
 822 * run several verifiers across an in-core buffer in order to establish buffer
 823 * type.  If repair can't establish that, the buffer will be left in memory
 824 * with NULL buffer ops.
 825 */
 826int
 827xfs_buf_reverify(
 828	struct xfs_buf		*bp,
 829	const struct xfs_buf_ops *ops)
 830{
 831	ASSERT(bp->b_flags & XBF_DONE);
 832	ASSERT(bp->b_error == 0);
 833
 834	if (!ops || bp->b_ops)
 835		return 0;
 836
 837	bp->b_ops = ops;
 838	bp->b_ops->verify_read(bp);
 839	if (bp->b_error)
 840		bp->b_flags &= ~XBF_DONE;
 841	return bp->b_error;
 842}
 843
 844int
 845xfs_buf_read_map(
 846	struct xfs_buftarg	*target,
 847	struct xfs_buf_map	*map,
 848	int			nmaps,
 849	xfs_buf_flags_t		flags,
 850	struct xfs_buf		**bpp,
 851	const struct xfs_buf_ops *ops,
 852	xfs_failaddr_t		fa)
 853{
 854	struct xfs_buf		*bp;
 855	int			error;
 856
 857	flags |= XBF_READ;
 858	*bpp = NULL;
 859
 860	error = xfs_buf_get_map(target, map, nmaps, flags, &bp);
 861	if (error)
 862		return error;
 863
 864	trace_xfs_buf_read(bp, flags, _RET_IP_);
 865
 866	if (!(bp->b_flags & XBF_DONE)) {
 867		/* Initiate the buffer read and wait. */
 868		XFS_STATS_INC(target->bt_mount, xb_get_read);
 869		bp->b_ops = ops;
 870		error = _xfs_buf_read(bp, flags);
 871
 872		/* Readahead iodone already dropped the buffer, so exit. */
 873		if (flags & XBF_ASYNC)
 874			return 0;
 875	} else {
 876		/* Buffer already read; all we need to do is check it. */
 877		error = xfs_buf_reverify(bp, ops);
 878
 879		/* Readahead already finished; drop the buffer and exit. */
 880		if (flags & XBF_ASYNC) {
 881			xfs_buf_relse(bp);
 882			return 0;
 883		}
 884
 885		/* We do not want read in the flags */
 886		bp->b_flags &= ~XBF_READ;
 887		ASSERT(bp->b_ops != NULL || ops == NULL);
 888	}
 889
 890	/*
 891	 * If we've had a read error, then the contents of the buffer are
 892	 * invalid and should not be used. To ensure that a followup read tries
 893	 * to pull the buffer from disk again, we clear the XBF_DONE flag and
 894	 * mark the buffer stale. This ensures that anyone who has a current
 895	 * reference to the buffer will interpret it's contents correctly and
 896	 * future cache lookups will also treat it as an empty, uninitialised
 897	 * buffer.
 898	 */
 899	if (error) {
 900		/*
 901		 * Check against log shutdown for error reporting because
 902		 * metadata writeback may require a read first and we need to
 903		 * report errors in metadata writeback until the log is shut
 904		 * down. High level transaction read functions already check
 905		 * against mount shutdown, anyway, so we only need to be
 906		 * concerned about low level IO interactions here.
 907		 */
 908		if (!xlog_is_shutdown(target->bt_mount->m_log))
 909			xfs_buf_ioerror_alert(bp, fa);
 910
 911		bp->b_flags &= ~XBF_DONE;
 912		xfs_buf_stale(bp);
 913		xfs_buf_relse(bp);
 914
 915		/* bad CRC means corrupted metadata */
 916		if (error == -EFSBADCRC)
 917			error = -EFSCORRUPTED;
 918		return error;
 919	}
 920
 921	*bpp = bp;
 922	return 0;
 923}
 924
 925/*
 926 *	If we are not low on memory then do the readahead in a deadlock
 927 *	safe manner.
 928 */
 929void
 930xfs_buf_readahead_map(
 931	struct xfs_buftarg	*target,
 932	struct xfs_buf_map	*map,
 933	int			nmaps,
 934	const struct xfs_buf_ops *ops)
 935{
 936	struct xfs_buf		*bp;
 937
 938	/*
 939	 * Currently we don't have a good means or justification for performing
 940	 * xmbuf_map_page asynchronously, so we don't do readahead.
 941	 */
 942	if (xfs_buftarg_is_mem(target))
 943		return;
 944
 945	xfs_buf_read_map(target, map, nmaps,
 946		     XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD, &bp, ops,
 947		     __this_address);
 948}
 949
 950/*
 951 * Read an uncached buffer from disk. Allocates and returns a locked
 952 * buffer containing the disk contents or nothing. Uncached buffers always have
 953 * a cache index of XFS_BUF_DADDR_NULL so we can easily determine if the buffer
 954 * is cached or uncached during fault diagnosis.
 955 */
 956int
 957xfs_buf_read_uncached(
 958	struct xfs_buftarg	*target,
 959	xfs_daddr_t		daddr,
 960	size_t			numblks,
 961	xfs_buf_flags_t		flags,
 962	struct xfs_buf		**bpp,
 963	const struct xfs_buf_ops *ops)
 964{
 965	struct xfs_buf		*bp;
 966	int			error;
 967
 968	*bpp = NULL;
 969
 970	error = xfs_buf_get_uncached(target, numblks, flags, &bp);
 971	if (error)
 972		return error;
 973
 974	/* set up the buffer for a read IO */
 975	ASSERT(bp->b_map_count == 1);
 976	bp->b_rhash_key = XFS_BUF_DADDR_NULL;
 977	bp->b_maps[0].bm_bn = daddr;
 978	bp->b_flags |= XBF_READ;
 979	bp->b_ops = ops;
 980
 981	xfs_buf_submit(bp);
 982	if (bp->b_error) {
 983		error = bp->b_error;
 984		xfs_buf_relse(bp);
 985		return error;
 986	}
 987
 988	*bpp = bp;
 989	return 0;
 990}
 991
 992int
 993xfs_buf_get_uncached(
 994	struct xfs_buftarg	*target,
 995	size_t			numblks,
 996	xfs_buf_flags_t		flags,
 997	struct xfs_buf		**bpp)
 998{
 999	int			error;
1000	struct xfs_buf		*bp;
1001	DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
1002
1003	*bpp = NULL;
1004
1005	/* flags might contain irrelevant bits, pass only what we care about */
1006	error = _xfs_buf_alloc(target, &map, 1, flags & XBF_NO_IOACCT, &bp);
1007	if (error)
1008		return error;
1009
1010	if (xfs_buftarg_is_mem(bp->b_target))
1011		error = xmbuf_map_page(bp);
1012	else
1013		error = xfs_buf_alloc_pages(bp, flags);
1014	if (error)
1015		goto fail_free_buf;
1016
1017	error = _xfs_buf_map_pages(bp, 0);
1018	if (unlikely(error)) {
1019		xfs_warn(target->bt_mount,
1020			"%s: failed to map pages", __func__);
1021		goto fail_free_buf;
1022	}
1023
1024	trace_xfs_buf_get_uncached(bp, _RET_IP_);
1025	*bpp = bp;
1026	return 0;
1027
1028fail_free_buf:
1029	xfs_buf_free(bp);
1030	return error;
1031}
1032
1033/*
1034 *	Increment reference count on buffer, to hold the buffer concurrently
1035 *	with another thread which may release (free) the buffer asynchronously.
1036 *	Must hold the buffer already to call this function.
1037 */
1038void
1039xfs_buf_hold(
1040	struct xfs_buf		*bp)
1041{
1042	trace_xfs_buf_hold(bp, _RET_IP_);
1043	atomic_inc(&bp->b_hold);
1044}
1045
1046static void
1047xfs_buf_rele_uncached(
1048	struct xfs_buf		*bp)
1049{
1050	ASSERT(list_empty(&bp->b_lru));
1051	if (atomic_dec_and_test(&bp->b_hold)) {
1052		xfs_buf_ioacct_dec(bp);
1053		xfs_buf_free(bp);
1054	}
1055}
1056
1057static void
1058xfs_buf_rele_cached(
1059	struct xfs_buf		*bp)
1060{
1061	struct xfs_buftarg	*btp = bp->b_target;
1062	struct xfs_perag	*pag = bp->b_pag;
1063	struct xfs_buf_cache	*bch = xfs_buftarg_buf_cache(btp, pag);
1064	bool			release;
1065	bool			freebuf = false;
1066
1067	trace_xfs_buf_rele(bp, _RET_IP_);
1068
1069	ASSERT(atomic_read(&bp->b_hold) > 0);
1070
1071	/*
1072	 * We grab the b_lock here first to serialise racing xfs_buf_rele()
1073	 * calls. The pag_buf_lock being taken on the last reference only
1074	 * serialises against racing lookups in xfs_buf_find(). IOWs, the second
1075	 * to last reference we drop here is not serialised against the last
1076	 * reference until we take bp->b_lock. Hence if we don't grab b_lock
1077	 * first, the last "release" reference can win the race to the lock and
1078	 * free the buffer before the second-to-last reference is processed,
1079	 * leading to a use-after-free scenario.
1080	 */
1081	spin_lock(&bp->b_lock);
1082	release = atomic_dec_and_lock(&bp->b_hold, &bch->bc_lock);
1083	if (!release) {
1084		/*
1085		 * Drop the in-flight state if the buffer is already on the LRU
1086		 * and it holds the only reference. This is racy because we
1087		 * haven't acquired the pag lock, but the use of _XBF_IN_FLIGHT
1088		 * ensures the decrement occurs only once per-buf.
1089		 */
1090		if ((atomic_read(&bp->b_hold) == 1) && !list_empty(&bp->b_lru))
1091			__xfs_buf_ioacct_dec(bp);
1092		goto out_unlock;
1093	}
1094
1095	/* the last reference has been dropped ... */
1096	__xfs_buf_ioacct_dec(bp);
1097	if (!(bp->b_flags & XBF_STALE) && atomic_read(&bp->b_lru_ref)) {
1098		/*
1099		 * If the buffer is added to the LRU take a new reference to the
1100		 * buffer for the LRU and clear the (now stale) dispose list
1101		 * state flag
1102		 */
1103		if (list_lru_add_obj(&btp->bt_lru, &bp->b_lru)) {
1104			bp->b_state &= ~XFS_BSTATE_DISPOSE;
1105			atomic_inc(&bp->b_hold);
1106		}
1107		spin_unlock(&bch->bc_lock);
1108	} else {
1109		/*
1110		 * most of the time buffers will already be removed from the
1111		 * LRU, so optimise that case by checking for the
1112		 * XFS_BSTATE_DISPOSE flag indicating the last list the buffer
1113		 * was on was the disposal list
1114		 */
1115		if (!(bp->b_state & XFS_BSTATE_DISPOSE)) {
1116			list_lru_del_obj(&btp->bt_lru, &bp->b_lru);
1117		} else {
1118			ASSERT(list_empty(&bp->b_lru));
1119		}
1120
1121		ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1122		rhashtable_remove_fast(&bch->bc_hash, &bp->b_rhash_head,
1123				xfs_buf_hash_params);
1124		spin_unlock(&bch->bc_lock);
1125		if (pag)
1126			xfs_perag_put(pag);
1127		freebuf = true;
1128	}
1129
1130out_unlock:
1131	spin_unlock(&bp->b_lock);
1132
1133	if (freebuf)
1134		xfs_buf_free(bp);
1135}
1136
1137/*
1138 * Release a hold on the specified buffer.
1139 */
1140void
1141xfs_buf_rele(
1142	struct xfs_buf		*bp)
1143{
1144	trace_xfs_buf_rele(bp, _RET_IP_);
1145	if (xfs_buf_is_uncached(bp))
1146		xfs_buf_rele_uncached(bp);
1147	else
1148		xfs_buf_rele_cached(bp);
1149}
1150
1151/*
1152 *	Lock a buffer object, if it is not already locked.
1153 *
1154 *	If we come across a stale, pinned, locked buffer, we know that we are
1155 *	being asked to lock a buffer that has been reallocated. Because it is
1156 *	pinned, we know that the log has not been pushed to disk and hence it
1157 *	will still be locked.  Rather than continuing to have trylock attempts
1158 *	fail until someone else pushes the log, push it ourselves before
1159 *	returning.  This means that the xfsaild will not get stuck trying
1160 *	to push on stale inode buffers.
1161 */
1162int
1163xfs_buf_trylock(
1164	struct xfs_buf		*bp)
1165{
1166	int			locked;
1167
1168	locked = down_trylock(&bp->b_sema) == 0;
1169	if (locked)
1170		trace_xfs_buf_trylock(bp, _RET_IP_);
1171	else
1172		trace_xfs_buf_trylock_fail(bp, _RET_IP_);
1173	return locked;
1174}
1175
1176/*
1177 *	Lock a buffer object.
1178 *
1179 *	If we come across a stale, pinned, locked buffer, we know that we
1180 *	are being asked to lock a buffer that has been reallocated. Because
1181 *	it is pinned, we know that the log has not been pushed to disk and
1182 *	hence it will still be locked. Rather than sleeping until someone
1183 *	else pushes the log, push it ourselves before trying to get the lock.
1184 */
1185void
1186xfs_buf_lock(
1187	struct xfs_buf		*bp)
1188{
1189	trace_xfs_buf_lock(bp, _RET_IP_);
1190
1191	if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
1192		xfs_log_force(bp->b_mount, 0);
1193	down(&bp->b_sema);
1194
1195	trace_xfs_buf_lock_done(bp, _RET_IP_);
1196}
1197
1198void
1199xfs_buf_unlock(
1200	struct xfs_buf		*bp)
1201{
1202	ASSERT(xfs_buf_islocked(bp));
1203
1204	up(&bp->b_sema);
1205	trace_xfs_buf_unlock(bp, _RET_IP_);
1206}
1207
1208STATIC void
1209xfs_buf_wait_unpin(
1210	struct xfs_buf		*bp)
1211{
1212	DECLARE_WAITQUEUE	(wait, current);
1213
1214	if (atomic_read(&bp->b_pin_count) == 0)
1215		return;
1216
1217	add_wait_queue(&bp->b_waiters, &wait);
1218	for (;;) {
1219		set_current_state(TASK_UNINTERRUPTIBLE);
1220		if (atomic_read(&bp->b_pin_count) == 0)
1221			break;
1222		io_schedule();
1223	}
1224	remove_wait_queue(&bp->b_waiters, &wait);
1225	set_current_state(TASK_RUNNING);
1226}
1227
1228static void
1229xfs_buf_ioerror_alert_ratelimited(
1230	struct xfs_buf		*bp)
1231{
1232	static unsigned long	lasttime;
1233	static struct xfs_buftarg *lasttarg;
1234
1235	if (bp->b_target != lasttarg ||
1236	    time_after(jiffies, (lasttime + 5*HZ))) {
1237		lasttime = jiffies;
1238		xfs_buf_ioerror_alert(bp, __this_address);
1239	}
1240	lasttarg = bp->b_target;
1241}
1242
1243/*
1244 * Account for this latest trip around the retry handler, and decide if
1245 * we've failed enough times to constitute a permanent failure.
1246 */
1247static bool
1248xfs_buf_ioerror_permanent(
1249	struct xfs_buf		*bp,
1250	struct xfs_error_cfg	*cfg)
1251{
1252	struct xfs_mount	*mp = bp->b_mount;
1253
1254	if (cfg->max_retries != XFS_ERR_RETRY_FOREVER &&
1255	    ++bp->b_retries > cfg->max_retries)
1256		return true;
1257	if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER &&
1258	    time_after(jiffies, cfg->retry_timeout + bp->b_first_retry_time))
1259		return true;
1260
1261	/* At unmount we may treat errors differently */
1262	if (xfs_is_unmounting(mp) && mp->m_fail_unmount)
1263		return true;
1264
1265	return false;
1266}
1267
1268/*
1269 * On a sync write or shutdown we just want to stale the buffer and let the
1270 * caller handle the error in bp->b_error appropriately.
1271 *
1272 * If the write was asynchronous then no one will be looking for the error.  If
1273 * this is the first failure of this type, clear the error state and write the
1274 * buffer out again. This means we always retry an async write failure at least
1275 * once, but we also need to set the buffer up to behave correctly now for
1276 * repeated failures.
1277 *
1278 * If we get repeated async write failures, then we take action according to the
1279 * error configuration we have been set up to use.
1280 *
1281 * Returns true if this function took care of error handling and the caller must
1282 * not touch the buffer again.  Return false if the caller should proceed with
1283 * normal I/O completion handling.
1284 */
1285static bool
1286xfs_buf_ioend_handle_error(
1287	struct xfs_buf		*bp)
1288{
1289	struct xfs_mount	*mp = bp->b_mount;
1290	struct xfs_error_cfg	*cfg;
1291
1292	/*
1293	 * If we've already shutdown the journal because of I/O errors, there's
1294	 * no point in giving this a retry.
1295	 */
1296	if (xlog_is_shutdown(mp->m_log))
1297		goto out_stale;
1298
1299	xfs_buf_ioerror_alert_ratelimited(bp);
1300
1301	/*
1302	 * We're not going to bother about retrying this during recovery.
1303	 * One strike!
1304	 */
1305	if (bp->b_flags & _XBF_LOGRECOVERY) {
1306		xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1307		return false;
1308	}
1309
1310	/*
1311	 * Synchronous writes will have callers process the error.
1312	 */
1313	if (!(bp->b_flags & XBF_ASYNC))
1314		goto out_stale;
1315
1316	trace_xfs_buf_iodone_async(bp, _RET_IP_);
1317
1318	cfg = xfs_error_get_cfg(mp, XFS_ERR_METADATA, bp->b_error);
1319	if (bp->b_last_error != bp->b_error ||
1320	    !(bp->b_flags & (XBF_STALE | XBF_WRITE_FAIL))) {
1321		bp->b_last_error = bp->b_error;
1322		if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER &&
1323		    !bp->b_first_retry_time)
1324			bp->b_first_retry_time = jiffies;
1325		goto resubmit;
1326	}
1327
1328	/*
1329	 * Permanent error - we need to trigger a shutdown if we haven't already
1330	 * to indicate that inconsistency will result from this action.
1331	 */
1332	if (xfs_buf_ioerror_permanent(bp, cfg)) {
1333		xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1334		goto out_stale;
1335	}
1336
1337	/* Still considered a transient error. Caller will schedule retries. */
1338	if (bp->b_flags & _XBF_INODES)
1339		xfs_buf_inode_io_fail(bp);
1340	else if (bp->b_flags & _XBF_DQUOTS)
1341		xfs_buf_dquot_io_fail(bp);
1342	else
1343		ASSERT(list_empty(&bp->b_li_list));
1344	xfs_buf_ioerror(bp, 0);
1345	xfs_buf_relse(bp);
1346	return true;
1347
1348resubmit:
1349	xfs_buf_ioerror(bp, 0);
1350	bp->b_flags |= (XBF_DONE | XBF_WRITE_FAIL);
1351	xfs_buf_submit(bp);
1352	return true;
1353out_stale:
1354	xfs_buf_stale(bp);
1355	bp->b_flags |= XBF_DONE;
1356	bp->b_flags &= ~XBF_WRITE;
1357	trace_xfs_buf_error_relse(bp, _RET_IP_);
1358	return false;
1359}
1360
1361static void
1362xfs_buf_ioend(
1363	struct xfs_buf	*bp)
1364{
1365	trace_xfs_buf_iodone(bp, _RET_IP_);
1366
1367	/*
1368	 * Pull in IO completion errors now. We are guaranteed to be running
1369	 * single threaded, so we don't need the lock to read b_io_error.
1370	 */
1371	if (!bp->b_error && bp->b_io_error)
1372		xfs_buf_ioerror(bp, bp->b_io_error);
1373
1374	if (bp->b_flags & XBF_READ) {
1375		if (!bp->b_error && bp->b_ops)
1376			bp->b_ops->verify_read(bp);
1377		if (!bp->b_error)
1378			bp->b_flags |= XBF_DONE;
1379	} else {
1380		if (!bp->b_error) {
1381			bp->b_flags &= ~XBF_WRITE_FAIL;
1382			bp->b_flags |= XBF_DONE;
1383		}
1384
1385		if (unlikely(bp->b_error) && xfs_buf_ioend_handle_error(bp))
1386			return;
1387
1388		/* clear the retry state */
1389		bp->b_last_error = 0;
1390		bp->b_retries = 0;
1391		bp->b_first_retry_time = 0;
1392
1393		/*
1394		 * Note that for things like remote attribute buffers, there may
1395		 * not be a buffer log item here, so processing the buffer log
1396		 * item must remain optional.
1397		 */
1398		if (bp->b_log_item)
1399			xfs_buf_item_done(bp);
1400
1401		if (bp->b_flags & _XBF_INODES)
1402			xfs_buf_inode_iodone(bp);
1403		else if (bp->b_flags & _XBF_DQUOTS)
1404			xfs_buf_dquot_iodone(bp);
1405
1406	}
1407
1408	bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD |
1409			 _XBF_LOGRECOVERY);
1410
1411	if (bp->b_flags & XBF_ASYNC)
1412		xfs_buf_relse(bp);
1413	else
1414		complete(&bp->b_iowait);
1415}
1416
1417static void
1418xfs_buf_ioend_work(
1419	struct work_struct	*work)
1420{
1421	struct xfs_buf		*bp =
1422		container_of(work, struct xfs_buf, b_ioend_work);
1423
1424	xfs_buf_ioend(bp);
1425}
1426
1427static void
1428xfs_buf_ioend_async(
1429	struct xfs_buf	*bp)
1430{
1431	INIT_WORK(&bp->b_ioend_work, xfs_buf_ioend_work);
1432	queue_work(bp->b_mount->m_buf_workqueue, &bp->b_ioend_work);
1433}
1434
1435void
1436__xfs_buf_ioerror(
1437	struct xfs_buf		*bp,
1438	int			error,
1439	xfs_failaddr_t		failaddr)
1440{
1441	ASSERT(error <= 0 && error >= -1000);
1442	bp->b_error = error;
1443	trace_xfs_buf_ioerror(bp, error, failaddr);
1444}
1445
1446void
1447xfs_buf_ioerror_alert(
1448	struct xfs_buf		*bp,
1449	xfs_failaddr_t		func)
1450{
1451	xfs_buf_alert_ratelimited(bp, "XFS: metadata IO error",
1452		"metadata I/O error in \"%pS\" at daddr 0x%llx len %d error %d",
1453				  func, (uint64_t)xfs_buf_daddr(bp),
1454				  bp->b_length, -bp->b_error);
1455}
1456
1457/*
1458 * To simulate an I/O failure, the buffer must be locked and held with at least
1459 * three references. The LRU reference is dropped by the stale call. The buf
1460 * item reference is dropped via ioend processing. The third reference is owned
1461 * by the caller and is dropped on I/O completion if the buffer is XBF_ASYNC.
1462 */
1463void
1464xfs_buf_ioend_fail(
1465	struct xfs_buf	*bp)
1466{
1467	bp->b_flags &= ~XBF_DONE;
1468	xfs_buf_stale(bp);
1469	xfs_buf_ioerror(bp, -EIO);
1470	xfs_buf_ioend(bp);
1471}
1472
1473int
1474xfs_bwrite(
1475	struct xfs_buf		*bp)
1476{
1477	int			error;
1478
1479	ASSERT(xfs_buf_islocked(bp));
1480
1481	bp->b_flags |= XBF_WRITE;
1482	bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q |
1483			 XBF_DONE);
1484
1485	error = xfs_buf_submit(bp);
1486	if (error)
1487		xfs_force_shutdown(bp->b_mount, SHUTDOWN_META_IO_ERROR);
1488	return error;
1489}
1490
1491static void
1492xfs_buf_bio_end_io(
1493	struct bio		*bio)
1494{
1495	struct xfs_buf		*bp = (struct xfs_buf *)bio->bi_private;
1496
1497	if (!bio->bi_status &&
1498	    (bp->b_flags & XBF_WRITE) && (bp->b_flags & XBF_ASYNC) &&
1499	    XFS_TEST_ERROR(false, bp->b_mount, XFS_ERRTAG_BUF_IOERROR))
1500		bio->bi_status = BLK_STS_IOERR;
1501
1502	/*
1503	 * don't overwrite existing errors - otherwise we can lose errors on
1504	 * buffers that require multiple bios to complete.
1505	 */
1506	if (bio->bi_status) {
1507		int error = blk_status_to_errno(bio->bi_status);
1508
1509		cmpxchg(&bp->b_io_error, 0, error);
1510	}
1511
1512	if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1513		invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1514
1515	if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1516		xfs_buf_ioend_async(bp);
1517	bio_put(bio);
1518}
1519
1520static void
1521xfs_buf_ioapply_map(
1522	struct xfs_buf	*bp,
1523	int		map,
1524	int		*buf_offset,
1525	int		*count,
1526	blk_opf_t	op)
1527{
1528	int		page_index;
1529	unsigned int	total_nr_pages = bp->b_page_count;
1530	int		nr_pages;
1531	struct bio	*bio;
1532	sector_t	sector =  bp->b_maps[map].bm_bn;
1533	int		size;
1534	int		offset;
1535
1536	/* skip the pages in the buffer before the start offset */
1537	page_index = 0;
1538	offset = *buf_offset;
1539	while (offset >= PAGE_SIZE) {
1540		page_index++;
1541		offset -= PAGE_SIZE;
1542	}
1543
1544	/*
1545	 * Limit the IO size to the length of the current vector, and update the
1546	 * remaining IO count for the next time around.
1547	 */
1548	size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1549	*count -= size;
1550	*buf_offset += size;
1551
1552next_chunk:
1553	atomic_inc(&bp->b_io_remaining);
1554	nr_pages = bio_max_segs(total_nr_pages);
1555
1556	bio = bio_alloc(bp->b_target->bt_bdev, nr_pages, op, GFP_NOIO);
1557	bio->bi_iter.bi_sector = sector;
1558	bio->bi_end_io = xfs_buf_bio_end_io;
1559	bio->bi_private = bp;
1560
1561	for (; size && nr_pages; nr_pages--, page_index++) {
1562		int	rbytes, nbytes = PAGE_SIZE - offset;
1563
1564		if (nbytes > size)
1565			nbytes = size;
1566
1567		rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1568				      offset);
1569		if (rbytes < nbytes)
1570			break;
1571
1572		offset = 0;
1573		sector += BTOBB(nbytes);
1574		size -= nbytes;
1575		total_nr_pages--;
1576	}
1577
1578	if (likely(bio->bi_iter.bi_size)) {
1579		if (xfs_buf_is_vmapped(bp)) {
1580			flush_kernel_vmap_range(bp->b_addr,
1581						xfs_buf_vmap_len(bp));
1582		}
1583		submit_bio(bio);
1584		if (size)
1585			goto next_chunk;
1586	} else {
1587		/*
1588		 * This is guaranteed not to be the last io reference count
1589		 * because the caller (xfs_buf_submit) holds a count itself.
1590		 */
1591		atomic_dec(&bp->b_io_remaining);
1592		xfs_buf_ioerror(bp, -EIO);
1593		bio_put(bio);
1594	}
1595
1596}
1597
1598STATIC void
1599_xfs_buf_ioapply(
1600	struct xfs_buf	*bp)
1601{
1602	struct blk_plug	plug;
1603	blk_opf_t	op;
1604	int		offset;
1605	int		size;
1606	int		i;
1607
1608	/*
1609	 * Make sure we capture only current IO errors rather than stale errors
1610	 * left over from previous use of the buffer (e.g. failed readahead).
1611	 */
1612	bp->b_error = 0;
1613
1614	if (bp->b_flags & XBF_WRITE) {
1615		op = REQ_OP_WRITE;
1616
1617		/*
1618		 * Run the write verifier callback function if it exists. If
1619		 * this function fails it will mark the buffer with an error and
1620		 * the IO should not be dispatched.
1621		 */
1622		if (bp->b_ops) {
1623			bp->b_ops->verify_write(bp);
1624			if (bp->b_error) {
1625				xfs_force_shutdown(bp->b_mount,
1626						   SHUTDOWN_CORRUPT_INCORE);
1627				return;
1628			}
1629		} else if (bp->b_rhash_key != XFS_BUF_DADDR_NULL) {
1630			struct xfs_mount *mp = bp->b_mount;
1631
1632			/*
1633			 * non-crc filesystems don't attach verifiers during
1634			 * log recovery, so don't warn for such filesystems.
1635			 */
1636			if (xfs_has_crc(mp)) {
1637				xfs_warn(mp,
1638					"%s: no buf ops on daddr 0x%llx len %d",
1639					__func__, xfs_buf_daddr(bp),
1640					bp->b_length);
1641				xfs_hex_dump(bp->b_addr,
1642						XFS_CORRUPTION_DUMP_LEN);
1643				dump_stack();
1644			}
1645		}
1646	} else {
1647		op = REQ_OP_READ;
1648		if (bp->b_flags & XBF_READ_AHEAD)
1649			op |= REQ_RAHEAD;
1650	}
1651
1652	/* we only use the buffer cache for meta-data */
1653	op |= REQ_META;
1654
1655	/* in-memory targets are directly mapped, no IO required. */
1656	if (xfs_buftarg_is_mem(bp->b_target)) {
1657		xfs_buf_ioend(bp);
1658		return;
1659	}
1660
1661	/*
1662	 * Walk all the vectors issuing IO on them. Set up the initial offset
1663	 * into the buffer and the desired IO size before we start -
1664	 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1665	 * subsequent call.
1666	 */
1667	offset = bp->b_offset;
1668	size = BBTOB(bp->b_length);
1669	blk_start_plug(&plug);
1670	for (i = 0; i < bp->b_map_count; i++) {
1671		xfs_buf_ioapply_map(bp, i, &offset, &size, op);
1672		if (bp->b_error)
1673			break;
1674		if (size <= 0)
1675			break;	/* all done */
1676	}
1677	blk_finish_plug(&plug);
1678}
1679
1680/*
1681 * Wait for I/O completion of a sync buffer and return the I/O error code.
1682 */
1683static int
1684xfs_buf_iowait(
1685	struct xfs_buf	*bp)
1686{
1687	ASSERT(!(bp->b_flags & XBF_ASYNC));
1688
1689	trace_xfs_buf_iowait(bp, _RET_IP_);
1690	wait_for_completion(&bp->b_iowait);
1691	trace_xfs_buf_iowait_done(bp, _RET_IP_);
1692
1693	return bp->b_error;
1694}
1695
1696/*
1697 * Buffer I/O submission path, read or write. Asynchronous submission transfers
1698 * the buffer lock ownership and the current reference to the IO. It is not
1699 * safe to reference the buffer after a call to this function unless the caller
1700 * holds an additional reference itself.
1701 */
1702static int
1703__xfs_buf_submit(
1704	struct xfs_buf	*bp,
1705	bool		wait)
1706{
1707	int		error = 0;
1708
1709	trace_xfs_buf_submit(bp, _RET_IP_);
1710
1711	ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1712
1713	/*
1714	 * On log shutdown we stale and complete the buffer immediately. We can
1715	 * be called to read the superblock before the log has been set up, so
1716	 * be careful checking the log state.
1717	 *
1718	 * Checking the mount shutdown state here can result in the log tail
1719	 * moving inappropriately on disk as the log may not yet be shut down.
1720	 * i.e. failing this buffer on mount shutdown can remove it from the AIL
1721	 * and move the tail of the log forwards without having written this
1722	 * buffer to disk. This corrupts the log tail state in memory, and
1723	 * because the log may not be shut down yet, it can then be propagated
1724	 * to disk before the log is shutdown. Hence we check log shutdown
1725	 * state here rather than mount state to avoid corrupting the log tail
1726	 * on shutdown.
1727	 */
1728	if (bp->b_mount->m_log &&
1729	    xlog_is_shutdown(bp->b_mount->m_log)) {
1730		xfs_buf_ioend_fail(bp);
1731		return -EIO;
1732	}
1733
1734	/*
1735	 * Grab a reference so the buffer does not go away underneath us. For
1736	 * async buffers, I/O completion drops the callers reference, which
1737	 * could occur before submission returns.
1738	 */
1739	xfs_buf_hold(bp);
1740
1741	if (bp->b_flags & XBF_WRITE)
1742		xfs_buf_wait_unpin(bp);
1743
1744	/* clear the internal error state to avoid spurious errors */
1745	bp->b_io_error = 0;
1746
1747	/*
1748	 * Set the count to 1 initially, this will stop an I/O completion
1749	 * callout which happens before we have started all the I/O from calling
1750	 * xfs_buf_ioend too early.
1751	 */
1752	atomic_set(&bp->b_io_remaining, 1);
1753	if (bp->b_flags & XBF_ASYNC)
1754		xfs_buf_ioacct_inc(bp);
1755	_xfs_buf_ioapply(bp);
1756
1757	/*
1758	 * If _xfs_buf_ioapply failed, we can get back here with only the IO
1759	 * reference we took above. If we drop it to zero, run completion so
1760	 * that we don't return to the caller with completion still pending.
1761	 */
1762	if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1763		if (bp->b_error || !(bp->b_flags & XBF_ASYNC))
1764			xfs_buf_ioend(bp);
1765		else
1766			xfs_buf_ioend_async(bp);
1767	}
1768
1769	if (wait)
1770		error = xfs_buf_iowait(bp);
1771
1772	/*
1773	 * Release the hold that keeps the buffer referenced for the entire
1774	 * I/O. Note that if the buffer is async, it is not safe to reference
1775	 * after this release.
1776	 */
1777	xfs_buf_rele(bp);
1778	return error;
1779}
1780
1781void *
1782xfs_buf_offset(
1783	struct xfs_buf		*bp,
1784	size_t			offset)
1785{
1786	struct page		*page;
1787
1788	if (bp->b_addr)
1789		return bp->b_addr + offset;
1790
1791	page = bp->b_pages[offset >> PAGE_SHIFT];
1792	return page_address(page) + (offset & (PAGE_SIZE-1));
1793}
1794
1795void
1796xfs_buf_zero(
1797	struct xfs_buf		*bp,
1798	size_t			boff,
1799	size_t			bsize)
1800{
1801	size_t			bend;
1802
1803	bend = boff + bsize;
1804	while (boff < bend) {
1805		struct page	*page;
1806		int		page_index, page_offset, csize;
1807
1808		page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1809		page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1810		page = bp->b_pages[page_index];
1811		csize = min_t(size_t, PAGE_SIZE - page_offset,
1812				      BBTOB(bp->b_length) - boff);
1813
1814		ASSERT((csize + page_offset) <= PAGE_SIZE);
1815
1816		memset(page_address(page) + page_offset, 0, csize);
1817
1818		boff += csize;
1819	}
1820}
1821
1822/*
1823 * Log a message about and stale a buffer that a caller has decided is corrupt.
1824 *
1825 * This function should be called for the kinds of metadata corruption that
1826 * cannot be detect from a verifier, such as incorrect inter-block relationship
1827 * data.  Do /not/ call this function from a verifier function.
1828 *
1829 * The buffer must be XBF_DONE prior to the call.  Afterwards, the buffer will
1830 * be marked stale, but b_error will not be set.  The caller is responsible for
1831 * releasing the buffer or fixing it.
1832 */
1833void
1834__xfs_buf_mark_corrupt(
1835	struct xfs_buf		*bp,
1836	xfs_failaddr_t		fa)
1837{
1838	ASSERT(bp->b_flags & XBF_DONE);
1839
1840	xfs_buf_corruption_error(bp, fa);
1841	xfs_buf_stale(bp);
1842}
1843
1844/*
1845 *	Handling of buffer targets (buftargs).
1846 */
1847
1848/*
1849 * Wait for any bufs with callbacks that have been submitted but have not yet
1850 * returned. These buffers will have an elevated hold count, so wait on those
1851 * while freeing all the buffers only held by the LRU.
1852 */
1853static enum lru_status
1854xfs_buftarg_drain_rele(
1855	struct list_head	*item,
1856	struct list_lru_one	*lru,
1857	spinlock_t		*lru_lock,
1858	void			*arg)
1859
1860{
1861	struct xfs_buf		*bp = container_of(item, struct xfs_buf, b_lru);
1862	struct list_head	*dispose = arg;
1863
1864	if (atomic_read(&bp->b_hold) > 1) {
1865		/* need to wait, so skip it this pass */
1866		trace_xfs_buf_drain_buftarg(bp, _RET_IP_);
1867		return LRU_SKIP;
1868	}
1869	if (!spin_trylock(&bp->b_lock))
1870		return LRU_SKIP;
1871
1872	/*
1873	 * clear the LRU reference count so the buffer doesn't get
1874	 * ignored in xfs_buf_rele().
1875	 */
1876	atomic_set(&bp->b_lru_ref, 0);
1877	bp->b_state |= XFS_BSTATE_DISPOSE;
1878	list_lru_isolate_move(lru, item, dispose);
1879	spin_unlock(&bp->b_lock);
1880	return LRU_REMOVED;
1881}
1882
1883/*
1884 * Wait for outstanding I/O on the buftarg to complete.
1885 */
1886void
1887xfs_buftarg_wait(
1888	struct xfs_buftarg	*btp)
1889{
1890	/*
1891	 * First wait on the buftarg I/O count for all in-flight buffers to be
1892	 * released. This is critical as new buffers do not make the LRU until
1893	 * they are released.
1894	 *
1895	 * Next, flush the buffer workqueue to ensure all completion processing
1896	 * has finished. Just waiting on buffer locks is not sufficient for
1897	 * async IO as the reference count held over IO is not released until
1898	 * after the buffer lock is dropped. Hence we need to ensure here that
1899	 * all reference counts have been dropped before we start walking the
1900	 * LRU list.
1901	 */
1902	while (percpu_counter_sum(&btp->bt_io_count))
1903		delay(100);
1904	flush_workqueue(btp->bt_mount->m_buf_workqueue);
1905}
1906
1907void
1908xfs_buftarg_drain(
1909	struct xfs_buftarg	*btp)
1910{
1911	LIST_HEAD(dispose);
1912	int			loop = 0;
1913	bool			write_fail = false;
1914
1915	xfs_buftarg_wait(btp);
1916
1917	/* loop until there is nothing left on the lru list. */
1918	while (list_lru_count(&btp->bt_lru)) {
1919		list_lru_walk(&btp->bt_lru, xfs_buftarg_drain_rele,
1920			      &dispose, LONG_MAX);
1921
1922		while (!list_empty(&dispose)) {
1923			struct xfs_buf *bp;
1924			bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1925			list_del_init(&bp->b_lru);
1926			if (bp->b_flags & XBF_WRITE_FAIL) {
1927				write_fail = true;
1928				xfs_buf_alert_ratelimited(bp,
1929					"XFS: Corruption Alert",
1930"Corruption Alert: Buffer at daddr 0x%llx had permanent write failures!",
1931					(long long)xfs_buf_daddr(bp));
1932			}
1933			xfs_buf_rele(bp);
1934		}
1935		if (loop++ != 0)
1936			delay(100);
1937	}
1938
1939	/*
1940	 * If one or more failed buffers were freed, that means dirty metadata
1941	 * was thrown away. This should only ever happen after I/O completion
1942	 * handling has elevated I/O error(s) to permanent failures and shuts
1943	 * down the journal.
1944	 */
1945	if (write_fail) {
1946		ASSERT(xlog_is_shutdown(btp->bt_mount->m_log));
1947		xfs_alert(btp->bt_mount,
1948	      "Please run xfs_repair to determine the extent of the problem.");
1949	}
1950}
1951
1952static enum lru_status
1953xfs_buftarg_isolate(
1954	struct list_head	*item,
1955	struct list_lru_one	*lru,
1956	spinlock_t		*lru_lock,
1957	void			*arg)
1958{
1959	struct xfs_buf		*bp = container_of(item, struct xfs_buf, b_lru);
1960	struct list_head	*dispose = arg;
1961
1962	/*
1963	 * we are inverting the lru lock/bp->b_lock here, so use a trylock.
1964	 * If we fail to get the lock, just skip it.
1965	 */
1966	if (!spin_trylock(&bp->b_lock))
1967		return LRU_SKIP;
1968	/*
1969	 * Decrement the b_lru_ref count unless the value is already
1970	 * zero. If the value is already zero, we need to reclaim the
1971	 * buffer, otherwise it gets another trip through the LRU.
1972	 */
1973	if (atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1974		spin_unlock(&bp->b_lock);
1975		return LRU_ROTATE;
1976	}
1977
1978	bp->b_state |= XFS_BSTATE_DISPOSE;
1979	list_lru_isolate_move(lru, item, dispose);
1980	spin_unlock(&bp->b_lock);
1981	return LRU_REMOVED;
1982}
1983
1984static unsigned long
1985xfs_buftarg_shrink_scan(
1986	struct shrinker		*shrink,
1987	struct shrink_control	*sc)
1988{
1989	struct xfs_buftarg	*btp = shrink->private_data;
1990	LIST_HEAD(dispose);
1991	unsigned long		freed;
1992
1993	freed = list_lru_shrink_walk(&btp->bt_lru, sc,
1994				     xfs_buftarg_isolate, &dispose);
1995
1996	while (!list_empty(&dispose)) {
1997		struct xfs_buf *bp;
1998		bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1999		list_del_init(&bp->b_lru);
2000		xfs_buf_rele(bp);
2001	}
2002
2003	return freed;
2004}
2005
2006static unsigned long
2007xfs_buftarg_shrink_count(
2008	struct shrinker		*shrink,
2009	struct shrink_control	*sc)
2010{
2011	struct xfs_buftarg	*btp = shrink->private_data;
2012	return list_lru_shrink_count(&btp->bt_lru, sc);
2013}
2014
2015void
2016xfs_destroy_buftarg(
2017	struct xfs_buftarg	*btp)
2018{
2019	shrinker_free(btp->bt_shrinker);
2020	ASSERT(percpu_counter_sum(&btp->bt_io_count) == 0);
2021	percpu_counter_destroy(&btp->bt_io_count);
2022	list_lru_destroy(&btp->bt_lru);
2023}
2024
2025void
2026xfs_free_buftarg(
2027	struct xfs_buftarg	*btp)
2028{
2029	xfs_destroy_buftarg(btp);
2030	fs_put_dax(btp->bt_daxdev, btp->bt_mount);
2031	/* the main block device is closed by kill_block_super */
2032	if (btp->bt_bdev != btp->bt_mount->m_super->s_bdev)
2033		bdev_fput(btp->bt_bdev_file);
2034	kfree(btp);
2035}
2036
2037int
2038xfs_setsize_buftarg(
2039	struct xfs_buftarg	*btp,
2040	unsigned int		sectorsize)
2041{
2042	/* Set up metadata sector size info */
2043	btp->bt_meta_sectorsize = sectorsize;
2044	btp->bt_meta_sectormask = sectorsize - 1;
2045
2046	if (set_blocksize(btp->bt_bdev, sectorsize)) {
2047		xfs_warn(btp->bt_mount,
2048			"Cannot set_blocksize to %u on device %pg",
2049			sectorsize, btp->bt_bdev);
2050		return -EINVAL;
2051	}
2052
2053	return 0;
2054}
2055
2056int
2057xfs_init_buftarg(
2058	struct xfs_buftarg		*btp,
2059	size_t				logical_sectorsize,
2060	const char			*descr)
2061{
2062	/* Set up device logical sector size mask */
2063	btp->bt_logical_sectorsize = logical_sectorsize;
2064	btp->bt_logical_sectormask = logical_sectorsize - 1;
2065
2066	/*
2067	 * Buffer IO error rate limiting. Limit it to no more than 10 messages
2068	 * per 30 seconds so as to not spam logs too much on repeated errors.
2069	 */
2070	ratelimit_state_init(&btp->bt_ioerror_rl, 30 * HZ,
2071			     DEFAULT_RATELIMIT_BURST);
2072
2073	if (list_lru_init(&btp->bt_lru))
2074		return -ENOMEM;
2075	if (percpu_counter_init(&btp->bt_io_count, 0, GFP_KERNEL))
2076		goto out_destroy_lru;
2077
2078	btp->bt_shrinker =
2079		shrinker_alloc(SHRINKER_NUMA_AWARE, "xfs-buf:%s", descr);
2080	if (!btp->bt_shrinker)
2081		goto out_destroy_io_count;
2082	btp->bt_shrinker->count_objects = xfs_buftarg_shrink_count;
2083	btp->bt_shrinker->scan_objects = xfs_buftarg_shrink_scan;
2084	btp->bt_shrinker->private_data = btp;
2085	shrinker_register(btp->bt_shrinker);
2086	return 0;
2087
2088out_destroy_io_count:
2089	percpu_counter_destroy(&btp->bt_io_count);
2090out_destroy_lru:
2091	list_lru_destroy(&btp->bt_lru);
2092	return -ENOMEM;
2093}
2094
2095struct xfs_buftarg *
2096xfs_alloc_buftarg(
2097	struct xfs_mount	*mp,
2098	struct file		*bdev_file)
2099{
2100	struct xfs_buftarg	*btp;
2101	const struct dax_holder_operations *ops = NULL;
2102
2103#if defined(CONFIG_FS_DAX) && defined(CONFIG_MEMORY_FAILURE)
2104	ops = &xfs_dax_holder_operations;
2105#endif
2106	btp = kzalloc(sizeof(*btp), GFP_KERNEL | __GFP_NOFAIL);
2107
2108	btp->bt_mount = mp;
2109	btp->bt_bdev_file = bdev_file;
2110	btp->bt_bdev = file_bdev(bdev_file);
2111	btp->bt_dev = btp->bt_bdev->bd_dev;
2112	btp->bt_daxdev = fs_dax_get_by_bdev(btp->bt_bdev, &btp->bt_dax_part_off,
2113					    mp, ops);
 
 
 
 
 
 
 
2114
2115	/*
2116	 * When allocating the buftargs we have not yet read the super block and
2117	 * thus don't know the file system sector size yet.
2118	 */
2119	if (xfs_setsize_buftarg(btp, bdev_logical_block_size(btp->bt_bdev)))
2120		goto error_free;
2121	if (xfs_init_buftarg(btp, bdev_logical_block_size(btp->bt_bdev),
2122			mp->m_super->s_id))
2123		goto error_free;
2124
2125	return btp;
2126
2127error_free:
2128	kfree(btp);
2129	return NULL;
2130}
2131
2132static inline void
2133xfs_buf_list_del(
2134	struct xfs_buf		*bp)
2135{
2136	list_del_init(&bp->b_list);
2137	wake_up_var(&bp->b_list);
2138}
2139
2140/*
2141 * Cancel a delayed write list.
2142 *
2143 * Remove each buffer from the list, clear the delwri queue flag and drop the
2144 * associated buffer reference.
2145 */
2146void
2147xfs_buf_delwri_cancel(
2148	struct list_head	*list)
2149{
2150	struct xfs_buf		*bp;
2151
2152	while (!list_empty(list)) {
2153		bp = list_first_entry(list, struct xfs_buf, b_list);
2154
2155		xfs_buf_lock(bp);
2156		bp->b_flags &= ~_XBF_DELWRI_Q;
2157		xfs_buf_list_del(bp);
2158		xfs_buf_relse(bp);
2159	}
2160}
2161
2162/*
2163 * Add a buffer to the delayed write list.
2164 *
2165 * This queues a buffer for writeout if it hasn't already been.  Note that
2166 * neither this routine nor the buffer list submission functions perform
2167 * any internal synchronization.  It is expected that the lists are thread-local
2168 * to the callers.
2169 *
2170 * Returns true if we queued up the buffer, or false if it already had
2171 * been on the buffer list.
2172 */
2173bool
2174xfs_buf_delwri_queue(
2175	struct xfs_buf		*bp,
2176	struct list_head	*list)
2177{
2178	ASSERT(xfs_buf_islocked(bp));
2179	ASSERT(!(bp->b_flags & XBF_READ));
2180
2181	/*
2182	 * If the buffer is already marked delwri it already is queued up
2183	 * by someone else for imediate writeout.  Just ignore it in that
2184	 * case.
2185	 */
2186	if (bp->b_flags & _XBF_DELWRI_Q) {
2187		trace_xfs_buf_delwri_queued(bp, _RET_IP_);
2188		return false;
2189	}
2190
2191	trace_xfs_buf_delwri_queue(bp, _RET_IP_);
2192
2193	/*
2194	 * If a buffer gets written out synchronously or marked stale while it
2195	 * is on a delwri list we lazily remove it. To do this, the other party
2196	 * clears the  _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
2197	 * It remains referenced and on the list.  In a rare corner case it
2198	 * might get readded to a delwri list after the synchronous writeout, in
2199	 * which case we need just need to re-add the flag here.
2200	 */
2201	bp->b_flags |= _XBF_DELWRI_Q;
2202	if (list_empty(&bp->b_list)) {
2203		atomic_inc(&bp->b_hold);
2204		list_add_tail(&bp->b_list, list);
2205	}
2206
2207	return true;
2208}
2209
2210/*
2211 * Queue a buffer to this delwri list as part of a data integrity operation.
2212 * If the buffer is on any other delwri list, we'll wait for that to clear
2213 * so that the caller can submit the buffer for IO and wait for the result.
2214 * Callers must ensure the buffer is not already on the list.
2215 */
2216void
2217xfs_buf_delwri_queue_here(
2218	struct xfs_buf		*bp,
2219	struct list_head	*buffer_list)
2220{
2221	/*
2222	 * We need this buffer to end up on the /caller's/ delwri list, not any
2223	 * old list.  This can happen if the buffer is marked stale (which
2224	 * clears DELWRI_Q) after the AIL queues the buffer to its list but
2225	 * before the AIL has a chance to submit the list.
2226	 */
2227	while (!list_empty(&bp->b_list)) {
2228		xfs_buf_unlock(bp);
2229		wait_var_event(&bp->b_list, list_empty(&bp->b_list));
2230		xfs_buf_lock(bp);
2231	}
2232
2233	ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
2234
2235	xfs_buf_delwri_queue(bp, buffer_list);
2236}
2237
2238/*
2239 * Compare function is more complex than it needs to be because
2240 * the return value is only 32 bits and we are doing comparisons
2241 * on 64 bit values
2242 */
2243static int
2244xfs_buf_cmp(
2245	void			*priv,
2246	const struct list_head	*a,
2247	const struct list_head	*b)
2248{
2249	struct xfs_buf	*ap = container_of(a, struct xfs_buf, b_list);
2250	struct xfs_buf	*bp = container_of(b, struct xfs_buf, b_list);
2251	xfs_daddr_t		diff;
2252
2253	diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn;
2254	if (diff < 0)
2255		return -1;
2256	if (diff > 0)
2257		return 1;
2258	return 0;
2259}
2260
2261/*
2262 * Submit buffers for write. If wait_list is specified, the buffers are
2263 * submitted using sync I/O and placed on the wait list such that the caller can
2264 * iowait each buffer. Otherwise async I/O is used and the buffers are released
2265 * at I/O completion time. In either case, buffers remain locked until I/O
2266 * completes and the buffer is released from the queue.
2267 */
2268static int
2269xfs_buf_delwri_submit_buffers(
2270	struct list_head	*buffer_list,
2271	struct list_head	*wait_list)
2272{
2273	struct xfs_buf		*bp, *n;
2274	int			pinned = 0;
2275	struct blk_plug		plug;
2276
2277	list_sort(NULL, buffer_list, xfs_buf_cmp);
2278
2279	blk_start_plug(&plug);
2280	list_for_each_entry_safe(bp, n, buffer_list, b_list) {
2281		if (!wait_list) {
2282			if (!xfs_buf_trylock(bp))
2283				continue;
2284			if (xfs_buf_ispinned(bp)) {
2285				xfs_buf_unlock(bp);
2286				pinned++;
2287				continue;
2288			}
2289		} else {
2290			xfs_buf_lock(bp);
2291		}
2292
2293		/*
2294		 * Someone else might have written the buffer synchronously or
2295		 * marked it stale in the meantime.  In that case only the
2296		 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
2297		 * reference and remove it from the list here.
2298		 */
2299		if (!(bp->b_flags & _XBF_DELWRI_Q)) {
2300			xfs_buf_list_del(bp);
2301			xfs_buf_relse(bp);
2302			continue;
2303		}
2304
2305		trace_xfs_buf_delwri_split(bp, _RET_IP_);
2306
2307		/*
2308		 * If we have a wait list, each buffer (and associated delwri
2309		 * queue reference) transfers to it and is submitted
2310		 * synchronously. Otherwise, drop the buffer from the delwri
2311		 * queue and submit async.
2312		 */
2313		bp->b_flags &= ~_XBF_DELWRI_Q;
2314		bp->b_flags |= XBF_WRITE;
2315		if (wait_list) {
2316			bp->b_flags &= ~XBF_ASYNC;
2317			list_move_tail(&bp->b_list, wait_list);
2318		} else {
2319			bp->b_flags |= XBF_ASYNC;
2320			xfs_buf_list_del(bp);
2321		}
2322		__xfs_buf_submit(bp, false);
2323	}
2324	blk_finish_plug(&plug);
2325
2326	return pinned;
2327}
2328
2329/*
2330 * Write out a buffer list asynchronously.
2331 *
2332 * This will take the @buffer_list, write all non-locked and non-pinned buffers
2333 * out and not wait for I/O completion on any of the buffers.  This interface
2334 * is only safely useable for callers that can track I/O completion by higher
2335 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
2336 * function.
2337 *
2338 * Note: this function will skip buffers it would block on, and in doing so
2339 * leaves them on @buffer_list so they can be retried on a later pass. As such,
2340 * it is up to the caller to ensure that the buffer list is fully submitted or
2341 * cancelled appropriately when they are finished with the list. Failure to
2342 * cancel or resubmit the list until it is empty will result in leaked buffers
2343 * at unmount time.
2344 */
2345int
2346xfs_buf_delwri_submit_nowait(
2347	struct list_head	*buffer_list)
2348{
2349	return xfs_buf_delwri_submit_buffers(buffer_list, NULL);
2350}
2351
2352/*
2353 * Write out a buffer list synchronously.
2354 *
2355 * This will take the @buffer_list, write all buffers out and wait for I/O
2356 * completion on all of the buffers. @buffer_list is consumed by the function,
2357 * so callers must have some other way of tracking buffers if they require such
2358 * functionality.
2359 */
2360int
2361xfs_buf_delwri_submit(
2362	struct list_head	*buffer_list)
2363{
2364	LIST_HEAD		(wait_list);
2365	int			error = 0, error2;
2366	struct xfs_buf		*bp;
2367
2368	xfs_buf_delwri_submit_buffers(buffer_list, &wait_list);
2369
2370	/* Wait for IO to complete. */
2371	while (!list_empty(&wait_list)) {
2372		bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
2373
2374		xfs_buf_list_del(bp);
2375
2376		/*
2377		 * Wait on the locked buffer, check for errors and unlock and
2378		 * release the delwri queue reference.
2379		 */
2380		error2 = xfs_buf_iowait(bp);
2381		xfs_buf_relse(bp);
2382		if (!error)
2383			error = error2;
2384	}
2385
2386	return error;
2387}
2388
2389/*
2390 * Push a single buffer on a delwri queue.
2391 *
2392 * The purpose of this function is to submit a single buffer of a delwri queue
2393 * and return with the buffer still on the original queue. The waiting delwri
2394 * buffer submission infrastructure guarantees transfer of the delwri queue
2395 * buffer reference to a temporary wait list. We reuse this infrastructure to
2396 * transfer the buffer back to the original queue.
2397 *
2398 * Note the buffer transitions from the queued state, to the submitted and wait
2399 * listed state and back to the queued state during this call. The buffer
2400 * locking and queue management logic between _delwri_pushbuf() and
2401 * _delwri_queue() guarantee that the buffer cannot be queued to another list
2402 * before returning.
2403 */
2404int
2405xfs_buf_delwri_pushbuf(
2406	struct xfs_buf		*bp,
2407	struct list_head	*buffer_list)
2408{
2409	LIST_HEAD		(submit_list);
2410	int			error;
2411
2412	ASSERT(bp->b_flags & _XBF_DELWRI_Q);
2413
2414	trace_xfs_buf_delwri_pushbuf(bp, _RET_IP_);
2415
2416	/*
2417	 * Isolate the buffer to a new local list so we can submit it for I/O
2418	 * independently from the rest of the original list.
2419	 */
2420	xfs_buf_lock(bp);
2421	list_move(&bp->b_list, &submit_list);
2422	xfs_buf_unlock(bp);
2423
2424	/*
2425	 * Delwri submission clears the DELWRI_Q buffer flag and returns with
2426	 * the buffer on the wait list with the original reference. Rather than
2427	 * bounce the buffer from a local wait list back to the original list
2428	 * after I/O completion, reuse the original list as the wait list.
2429	 */
2430	xfs_buf_delwri_submit_buffers(&submit_list, buffer_list);
2431
2432	/*
2433	 * The buffer is now locked, under I/O and wait listed on the original
2434	 * delwri queue. Wait for I/O completion, restore the DELWRI_Q flag and
2435	 * return with the buffer unlocked and on the original queue.
2436	 */
2437	error = xfs_buf_iowait(bp);
2438	bp->b_flags |= _XBF_DELWRI_Q;
2439	xfs_buf_unlock(bp);
2440
2441	return error;
2442}
2443
2444void xfs_buf_set_ref(struct xfs_buf *bp, int lru_ref)
2445{
2446	/*
2447	 * Set the lru reference count to 0 based on the error injection tag.
2448	 * This allows userspace to disrupt buffer caching for debug/testing
2449	 * purposes.
2450	 */
2451	if (XFS_TEST_ERROR(false, bp->b_mount, XFS_ERRTAG_BUF_LRU_REF))
2452		lru_ref = 0;
2453
2454	atomic_set(&bp->b_lru_ref, lru_ref);
2455}
2456
2457/*
2458 * Verify an on-disk magic value against the magic value specified in the
2459 * verifier structure. The verifier magic is in disk byte order so the caller is
2460 * expected to pass the value directly from disk.
2461 */
2462bool
2463xfs_verify_magic(
2464	struct xfs_buf		*bp,
2465	__be32			dmagic)
2466{
2467	struct xfs_mount	*mp = bp->b_mount;
2468	int			idx;
2469
2470	idx = xfs_has_crc(mp);
2471	if (WARN_ON(!bp->b_ops || !bp->b_ops->magic[idx]))
2472		return false;
2473	return dmagic == bp->b_ops->magic[idx];
2474}
2475/*
2476 * Verify an on-disk magic value against the magic value specified in the
2477 * verifier structure. The verifier magic is in disk byte order so the caller is
2478 * expected to pass the value directly from disk.
2479 */
2480bool
2481xfs_verify_magic16(
2482	struct xfs_buf		*bp,
2483	__be16			dmagic)
2484{
2485	struct xfs_mount	*mp = bp->b_mount;
2486	int			idx;
2487
2488	idx = xfs_has_crc(mp);
2489	if (WARN_ON(!bp->b_ops || !bp->b_ops->magic16[idx]))
2490		return false;
2491	return dmagic == bp->b_ops->magic16[idx];
2492}