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