1/*
   2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   3 * All Rights Reserved.
   4 *
   5 * This program is free software; you can redistribute it and/or
   6 * modify it under the terms of the GNU General Public License as
   7 * published by the Free Software Foundation.
   8 *
   9 * This program is distributed in the hope that it would be useful,
  10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12 * GNU General Public License for more details.
  13 *
  14 * You should have received a copy of the GNU General Public License
  15 * along with this program; if not, write the Free Software Foundation,
  16 * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17 */
  18#include "xfs.h"
  19#include <linux/stddef.h>
  20#include <linux/errno.h>
  21#include <linux/gfp.h>
  22#include <linux/pagemap.h>
  23#include <linux/init.h>
  24#include <linux/vmalloc.h>
  25#include <linux/bio.h>
  26#include <linux/sysctl.h>
  27#include <linux/proc_fs.h>
  28#include <linux/workqueue.h>
  29#include <linux/percpu.h>
  30#include <linux/blkdev.h>
  31#include <linux/hash.h>
  32#include <linux/kthread.h>
  33#include <linux/migrate.h>
  34#include <linux/backing-dev.h>
  35#include <linux/freezer.h>
  36
  37#include "xfs_format.h"
  38#include "xfs_log_format.h"
  39#include "xfs_trans_resv.h"
  40#include "xfs_sb.h"
  41#include "xfs_mount.h"
  42#include "xfs_trace.h"
  43#include "xfs_log.h"
  44
  45static kmem_zone_t *xfs_buf_zone;
  46
  47#ifdef XFS_BUF_LOCK_TRACKING
  48# define XB_SET_OWNER(bp)	((bp)->b_last_holder = current->pid)
  49# define XB_CLEAR_OWNER(bp)	((bp)->b_last_holder = -1)
  50# define XB_GET_OWNER(bp)	((bp)->b_last_holder)
  51#else
  52# define XB_SET_OWNER(bp)	do { } while (0)
  53# define XB_CLEAR_OWNER(bp)	do { } while (0)
  54# define XB_GET_OWNER(bp)	do { } while (0)
  55#endif
  56
  57#define xb_to_gfp(flags) \
  58	((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : GFP_NOFS) | __GFP_NOWARN)
  59
  60
  61static inline int
  62xfs_buf_is_vmapped(
  63	struct xfs_buf	*bp)
  64{
  65	/*
  66	 * Return true if the buffer is vmapped.
  67	 *
  68	 * b_addr is null if the buffer is not mapped, but the code is clever
  69	 * enough to know it doesn't have to map a single page, so the check has
  70	 * to be both for b_addr and bp->b_page_count > 1.
  71	 */
  72	return bp->b_addr && bp->b_page_count > 1;
  73}
  74
  75static inline int
  76xfs_buf_vmap_len(
  77	struct xfs_buf	*bp)
  78{
  79	return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
  80}
  81
  82/*
  83 * Bump the I/O in flight count on the buftarg if we haven't yet done so for
  84 * this buffer. The count is incremented once per buffer (per hold cycle)
  85 * because the corresponding decrement is deferred to buffer release. Buffers
  86 * can undergo I/O multiple times in a hold-release cycle and per buffer I/O
  87 * tracking adds unnecessary overhead. This is used for sychronization purposes
  88 * with unmount (see xfs_wait_buftarg()), so all we really need is a count of
  89 * in-flight buffers.
  90 *
  91 * Buffers that are never released (e.g., superblock, iclog buffers) must set
  92 * the XBF_NO_IOACCT flag before I/O submission. Otherwise, the buftarg count
  93 * never reaches zero and unmount hangs indefinitely.
  94 */
  95static inline void
  96xfs_buf_ioacct_inc(
  97	struct xfs_buf	*bp)
  98{
  99	if (bp->b_flags & (XBF_NO_IOACCT|_XBF_IN_FLIGHT))
 100		return;
 101
 102	ASSERT(bp->b_flags & XBF_ASYNC);
 103	bp->b_flags |= _XBF_IN_FLIGHT;
 104	percpu_counter_inc(&bp->b_target->bt_io_count);
 105}
 106
 107/*
 108 * Clear the in-flight state on a buffer about to be released to the LRU or
 109 * freed and unaccount from the buftarg.
 110 */
 111static inline void
 112xfs_buf_ioacct_dec(
 113	struct xfs_buf	*bp)
 114{
 115	if (!(bp->b_flags & _XBF_IN_FLIGHT))
 116		return;
 117
 118	bp->b_flags &= ~_XBF_IN_FLIGHT;
 119	percpu_counter_dec(&bp->b_target->bt_io_count);
 120}
 121
 122/*
 123 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
 124 * b_lru_ref count so that the buffer is freed immediately when the buffer
 125 * reference count falls to zero. If the buffer is already on the LRU, we need
 126 * to remove the reference that LRU holds on the buffer.
 127 *
 128 * This prevents build-up of stale buffers on the LRU.
 129 */
 130void
 131xfs_buf_stale(
 132	struct xfs_buf	*bp)
 133{
 134	ASSERT(xfs_buf_islocked(bp));
 135
 136	bp->b_flags |= XBF_STALE;
 137
 138	/*
 139	 * Clear the delwri status so that a delwri queue walker will not
 140	 * flush this buffer to disk now that it is stale. The delwri queue has
 141	 * a reference to the buffer, so this is safe to do.
 142	 */
 143	bp->b_flags &= ~_XBF_DELWRI_Q;
 144
 145	/*
 146	 * Once the buffer is marked stale and unlocked, a subsequent lookup
 147	 * could reset b_flags. There is no guarantee that the buffer is
 148	 * unaccounted (released to LRU) before that occurs. Drop in-flight
 149	 * status now to preserve accounting consistency.
 150	 */
 151	xfs_buf_ioacct_dec(bp);
 152
 153	spin_lock(&bp->b_lock);
 154	atomic_set(&bp->b_lru_ref, 0);
 155	if (!(bp->b_state & XFS_BSTATE_DISPOSE) &&
 156	    (list_lru_del(&bp->b_target->bt_lru, &bp->b_lru)))
 157		atomic_dec(&bp->b_hold);
 158
 159	ASSERT(atomic_read(&bp->b_hold) >= 1);
 160	spin_unlock(&bp->b_lock);
 161}
 162
 163static int
 164xfs_buf_get_maps(
 165	struct xfs_buf		*bp,
 166	int			map_count)
 167{
 168	ASSERT(bp->b_maps == NULL);
 169	bp->b_map_count = map_count;
 170
 171	if (map_count == 1) {
 172		bp->b_maps = &bp->__b_map;
 173		return 0;
 174	}
 175
 176	bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map),
 177				KM_NOFS);
 178	if (!bp->b_maps)
 179		return -ENOMEM;
 180	return 0;
 181}
 182
 183/*
 184 *	Frees b_pages if it was allocated.
 185 */
 186static void
 187xfs_buf_free_maps(
 188	struct xfs_buf	*bp)
 189{
 190	if (bp->b_maps != &bp->__b_map) {
 191		kmem_free(bp->b_maps);
 192		bp->b_maps = NULL;
 193	}
 194}
 195
 196struct xfs_buf *
 197_xfs_buf_alloc(
 198	struct xfs_buftarg	*target,
 199	struct xfs_buf_map	*map,
 200	int			nmaps,
 201	xfs_buf_flags_t		flags)
 202{
 203	struct xfs_buf		*bp;
 204	int			error;
 205	int			i;
 206
 207	bp = kmem_zone_zalloc(xfs_buf_zone, KM_NOFS);
 208	if (unlikely(!bp))
 209		return NULL;
 210
 211	/*
 212	 * We don't want certain flags to appear in b_flags unless they are
 213	 * specifically set by later operations on the buffer.
 214	 */
 215	flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD);
 216
 217	atomic_set(&bp->b_hold, 1);
 218	atomic_set(&bp->b_lru_ref, 1);
 219	init_completion(&bp->b_iowait);
 220	INIT_LIST_HEAD(&bp->b_lru);
 221	INIT_LIST_HEAD(&bp->b_list);
 
 222	sema_init(&bp->b_sema, 0); /* held, no waiters */
 223	spin_lock_init(&bp->b_lock);
 224	XB_SET_OWNER(bp);
 225	bp->b_target = target;
 226	bp->b_flags = flags;
 227
 228	/*
 229	 * Set length and io_length to the same value initially.
 230	 * I/O routines should use io_length, which will be the same in
 231	 * most cases but may be reset (e.g. XFS recovery).
 232	 */
 233	error = xfs_buf_get_maps(bp, nmaps);
 234	if (error)  {
 235		kmem_zone_free(xfs_buf_zone, bp);
 236		return NULL;
 237	}
 238
 239	bp->b_bn = map[0].bm_bn;
 240	bp->b_length = 0;
 241	for (i = 0; i < nmaps; i++) {
 242		bp->b_maps[i].bm_bn = map[i].bm_bn;
 243		bp->b_maps[i].bm_len = map[i].bm_len;
 244		bp->b_length += map[i].bm_len;
 245	}
 246	bp->b_io_length = bp->b_length;
 247
 248	atomic_set(&bp->b_pin_count, 0);
 249	init_waitqueue_head(&bp->b_waiters);
 250
 251	XFS_STATS_INC(target->bt_mount, xb_create);
 252	trace_xfs_buf_init(bp, _RET_IP_);
 253
 254	return bp;
 255}
 256
 257/*
 258 *	Allocate a page array capable of holding a specified number
 259 *	of pages, and point the page buf at it.
 260 */
 261STATIC int
 262_xfs_buf_get_pages(
 263	xfs_buf_t		*bp,
 264	int			page_count)
 265{
 266	/* Make sure that we have a page list */
 267	if (bp->b_pages == NULL) {
 268		bp->b_page_count = page_count;
 269		if (page_count <= XB_PAGES) {
 270			bp->b_pages = bp->b_page_array;
 271		} else {
 272			bp->b_pages = kmem_alloc(sizeof(struct page *) *
 273						 page_count, KM_NOFS);
 274			if (bp->b_pages == NULL)
 275				return -ENOMEM;
 276		}
 277		memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
 278	}
 279	return 0;
 280}
 281
 282/*
 283 *	Frees b_pages if it was allocated.
 284 */
 285STATIC void
 286_xfs_buf_free_pages(
 287	xfs_buf_t	*bp)
 288{
 289	if (bp->b_pages != bp->b_page_array) {
 290		kmem_free(bp->b_pages);
 291		bp->b_pages = NULL;
 292	}
 293}
 294
 295/*
 296 *	Releases the specified buffer.
 297 *
 298 * 	The modification state of any associated pages is left unchanged.
 299 * 	The buffer must not be on any hash - use xfs_buf_rele instead for
 300 * 	hashed and refcounted buffers
 301 */
 302void
 303xfs_buf_free(
 304	xfs_buf_t		*bp)
 305{
 306	trace_xfs_buf_free(bp, _RET_IP_);
 307
 308	ASSERT(list_empty(&bp->b_lru));
 309
 310	if (bp->b_flags & _XBF_PAGES) {
 311		uint		i;
 312
 313		if (xfs_buf_is_vmapped(bp))
 314			vm_unmap_ram(bp->b_addr - bp->b_offset,
 315					bp->b_page_count);
 316
 317		for (i = 0; i < bp->b_page_count; i++) {
 318			struct page	*page = bp->b_pages[i];
 319
 320			__free_page(page);
 321		}
 322	} else if (bp->b_flags & _XBF_KMEM)
 323		kmem_free(bp->b_addr);
 324	_xfs_buf_free_pages(bp);
 325	xfs_buf_free_maps(bp);
 326	kmem_zone_free(xfs_buf_zone, bp);
 327}
 328
 329/*
 330 * Allocates all the pages for buffer in question and builds it's page list.
 331 */
 332STATIC int
 333xfs_buf_allocate_memory(
 334	xfs_buf_t		*bp,
 335	uint			flags)
 336{
 337	size_t			size;
 338	size_t			nbytes, offset;
 339	gfp_t			gfp_mask = xb_to_gfp(flags);
 340	unsigned short		page_count, i;
 341	xfs_off_t		start, end;
 342	int			error;
 343
 344	/*
 345	 * for buffers that are contained within a single page, just allocate
 346	 * the memory from the heap - there's no need for the complexity of
 347	 * page arrays to keep allocation down to order 0.
 348	 */
 349	size = BBTOB(bp->b_length);
 350	if (size < PAGE_SIZE) {
 351		bp->b_addr = kmem_alloc(size, KM_NOFS);
 352		if (!bp->b_addr) {
 353			/* low memory - use alloc_page loop instead */
 354			goto use_alloc_page;
 355		}
 356
 357		if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) !=
 358		    ((unsigned long)bp->b_addr & PAGE_MASK)) {
 359			/* b_addr spans two pages - use alloc_page instead */
 360			kmem_free(bp->b_addr);
 361			bp->b_addr = NULL;
 362			goto use_alloc_page;
 363		}
 364		bp->b_offset = offset_in_page(bp->b_addr);
 365		bp->b_pages = bp->b_page_array;
 366		bp->b_pages[0] = virt_to_page(bp->b_addr);
 367		bp->b_page_count = 1;
 368		bp->b_flags |= _XBF_KMEM;
 369		return 0;
 370	}
 371
 372use_alloc_page:
 373	start = BBTOB(bp->b_maps[0].bm_bn) >> PAGE_SHIFT;
 374	end = (BBTOB(bp->b_maps[0].bm_bn + bp->b_length) + PAGE_SIZE - 1)
 375								>> PAGE_SHIFT;
 376	page_count = end - start;
 377	error = _xfs_buf_get_pages(bp, page_count);
 378	if (unlikely(error))
 379		return error;
 380
 381	offset = bp->b_offset;
 382	bp->b_flags |= _XBF_PAGES;
 383
 384	for (i = 0; i < bp->b_page_count; i++) {
 385		struct page	*page;
 386		uint		retries = 0;
 387retry:
 388		page = alloc_page(gfp_mask);
 389		if (unlikely(page == NULL)) {
 390			if (flags & XBF_READ_AHEAD) {
 391				bp->b_page_count = i;
 392				error = -ENOMEM;
 393				goto out_free_pages;
 394			}
 395
 396			/*
 397			 * This could deadlock.
 398			 *
 399			 * But until all the XFS lowlevel code is revamped to
 400			 * handle buffer allocation failures we can't do much.
 401			 */
 402			if (!(++retries % 100))
 403				xfs_err(NULL,
 404		"%s(%u) possible memory allocation deadlock in %s (mode:0x%x)",
 405					current->comm, current->pid,
 406					__func__, gfp_mask);
 407
 408			XFS_STATS_INC(bp->b_target->bt_mount, xb_page_retries);
 409			congestion_wait(BLK_RW_ASYNC, HZ/50);
 410			goto retry;
 411		}
 412
 413		XFS_STATS_INC(bp->b_target->bt_mount, xb_page_found);
 414
 415		nbytes = min_t(size_t, size, PAGE_SIZE - offset);
 416		size -= nbytes;
 417		bp->b_pages[i] = page;
 418		offset = 0;
 419	}
 420	return 0;
 421
 422out_free_pages:
 423	for (i = 0; i < bp->b_page_count; i++)
 424		__free_page(bp->b_pages[i]);
 425	bp->b_flags &= ~_XBF_PAGES;
 426	return error;
 427}
 428
 429/*
 430 *	Map buffer into kernel address-space if necessary.
 431 */
 432STATIC int
 433_xfs_buf_map_pages(
 434	xfs_buf_t		*bp,
 435	uint			flags)
 436{
 437	ASSERT(bp->b_flags & _XBF_PAGES);
 438	if (bp->b_page_count == 1) {
 439		/* A single page buffer is always mappable */
 440		bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
 441	} else if (flags & XBF_UNMAPPED) {
 442		bp->b_addr = NULL;
 443	} else {
 444		int retried = 0;
 445		unsigned noio_flag;
 446
 447		/*
 448		 * vm_map_ram() will allocate auxillary structures (e.g.
 449		 * pagetables) with GFP_KERNEL, yet we are likely to be under
 450		 * GFP_NOFS context here. Hence we need to tell memory reclaim
 451		 * that we are in such a context via PF_MEMALLOC_NOIO to prevent
 452		 * memory reclaim re-entering the filesystem here and
 453		 * potentially deadlocking.
 454		 */
 455		noio_flag = memalloc_noio_save();
 456		do {
 457			bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
 458						-1, PAGE_KERNEL);
 459			if (bp->b_addr)
 460				break;
 461			vm_unmap_aliases();
 462		} while (retried++ <= 1);
 463		memalloc_noio_restore(noio_flag);
 464
 465		if (!bp->b_addr)
 466			return -ENOMEM;
 467		bp->b_addr += bp->b_offset;
 468	}
 469
 470	return 0;
 471}
 472
 473/*
 474 *	Finding and Reading Buffers
 475 */
 476static int
 477_xfs_buf_obj_cmp(
 478	struct rhashtable_compare_arg	*arg,
 479	const void			*obj)
 480{
 481	const struct xfs_buf_map	*map = arg->key;
 482	const struct xfs_buf		*bp = obj;
 483
 484	/*
 485	 * The key hashing in the lookup path depends on the key being the
 486	 * first element of the compare_arg, make sure to assert this.
 487	 */
 488	BUILD_BUG_ON(offsetof(struct xfs_buf_map, bm_bn) != 0);
 489
 490	if (bp->b_bn != map->bm_bn)
 491		return 1;
 492
 493	if (unlikely(bp->b_length != map->bm_len)) {
 494		/*
 495		 * found a block number match. If the range doesn't
 496		 * match, the only way this is allowed is if the buffer
 497		 * in the cache is stale and the transaction that made
 498		 * it stale has not yet committed. i.e. we are
 499		 * reallocating a busy extent. Skip this buffer and
 500		 * continue searching for an exact match.
 501		 */
 502		ASSERT(bp->b_flags & XBF_STALE);
 503		return 1;
 504	}
 505	return 0;
 506}
 507
 508static const struct rhashtable_params xfs_buf_hash_params = {
 509	.min_size		= 32,	/* empty AGs have minimal footprint */
 510	.nelem_hint		= 16,
 511	.key_len		= sizeof(xfs_daddr_t),
 512	.key_offset		= offsetof(struct xfs_buf, b_bn),
 513	.head_offset		= offsetof(struct xfs_buf, b_rhash_head),
 514	.automatic_shrinking	= true,
 515	.obj_cmpfn		= _xfs_buf_obj_cmp,
 516};
 517
 518int
 519xfs_buf_hash_init(
 520	struct xfs_perag	*pag)
 521{
 522	spin_lock_init(&pag->pag_buf_lock);
 523	return rhashtable_init(&pag->pag_buf_hash, &xfs_buf_hash_params);
 524}
 525
 526void
 527xfs_buf_hash_destroy(
 528	struct xfs_perag	*pag)
 529{
 530	rhashtable_destroy(&pag->pag_buf_hash);
 531}
 532
 533/*
 534 *	Look up, and creates if absent, a lockable buffer for
 535 *	a given range of an inode.  The buffer is returned
 536 *	locked.	No I/O is implied by this call.
 537 */
 538xfs_buf_t *
 539_xfs_buf_find(
 540	struct xfs_buftarg	*btp,
 541	struct xfs_buf_map	*map,
 542	int			nmaps,
 543	xfs_buf_flags_t		flags,
 544	xfs_buf_t		*new_bp)
 545{
 546	struct xfs_perag	*pag;
 
 
 547	xfs_buf_t		*bp;
 548	struct xfs_buf_map	cmap = { .bm_bn = map[0].bm_bn };
 549	xfs_daddr_t		eofs;
 
 550	int			i;
 551
 552	for (i = 0; i < nmaps; i++)
 553		cmap.bm_len += map[i].bm_len;
 554
 555	/* Check for IOs smaller than the sector size / not sector aligned */
 556	ASSERT(!(BBTOB(cmap.bm_len) < btp->bt_meta_sectorsize));
 557	ASSERT(!(BBTOB(cmap.bm_bn) & (xfs_off_t)btp->bt_meta_sectormask));
 558
 559	/*
 560	 * Corrupted block numbers can get through to here, unfortunately, so we
 561	 * have to check that the buffer falls within the filesystem bounds.
 562	 */
 563	eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks);
 564	if (cmap.bm_bn < 0 || cmap.bm_bn >= eofs) {
 565		/*
 566		 * XXX (dgc): we should really be returning -EFSCORRUPTED here,
 567		 * but none of the higher level infrastructure supports
 568		 * returning a specific error on buffer lookup failures.
 569		 */
 570		xfs_alert(btp->bt_mount,
 571			  "%s: Block out of range: block 0x%llx, EOFS 0x%llx ",
 572			  __func__, cmap.bm_bn, eofs);
 573		WARN_ON(1);
 574		return NULL;
 575	}
 576
 
 577	pag = xfs_perag_get(btp->bt_mount,
 578			    xfs_daddr_to_agno(btp->bt_mount, cmap.bm_bn));
 579
 
 580	spin_lock(&pag->pag_buf_lock);
 581	bp = rhashtable_lookup_fast(&pag->pag_buf_hash, &cmap,
 582				    xfs_buf_hash_params);
 583	if (bp) {
 584		atomic_inc(&bp->b_hold);
 585		goto found;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 586	}
 587
 588	/* No match found */
 589	if (new_bp) {
 
 
 590		/* the buffer keeps the perag reference until it is freed */
 591		new_bp->b_pag = pag;
 592		rhashtable_insert_fast(&pag->pag_buf_hash,
 593				       &new_bp->b_rhash_head,
 594				       xfs_buf_hash_params);
 595		spin_unlock(&pag->pag_buf_lock);
 596	} else {
 597		XFS_STATS_INC(btp->bt_mount, xb_miss_locked);
 598		spin_unlock(&pag->pag_buf_lock);
 599		xfs_perag_put(pag);
 600	}
 601	return new_bp;
 602
 603found:
 604	spin_unlock(&pag->pag_buf_lock);
 605	xfs_perag_put(pag);
 606
 607	if (!xfs_buf_trylock(bp)) {
 608		if (flags & XBF_TRYLOCK) {
 609			xfs_buf_rele(bp);
 610			XFS_STATS_INC(btp->bt_mount, xb_busy_locked);
 611			return NULL;
 612		}
 613		xfs_buf_lock(bp);
 614		XFS_STATS_INC(btp->bt_mount, xb_get_locked_waited);
 615	}
 616
 617	/*
 618	 * if the buffer is stale, clear all the external state associated with
 619	 * it. We need to keep flags such as how we allocated the buffer memory
 620	 * intact here.
 621	 */
 622	if (bp->b_flags & XBF_STALE) {
 623		ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
 624		ASSERT(bp->b_iodone == NULL);
 625		bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
 626		bp->b_ops = NULL;
 627	}
 628
 629	trace_xfs_buf_find(bp, flags, _RET_IP_);
 630	XFS_STATS_INC(btp->bt_mount, xb_get_locked);
 631	return bp;
 632}
 633
 634/*
 635 * Assembles a buffer covering the specified range. The code is optimised for
 636 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
 637 * more hits than misses.
 638 */
 639struct xfs_buf *
 640xfs_buf_get_map(
 641	struct xfs_buftarg	*target,
 642	struct xfs_buf_map	*map,
 643	int			nmaps,
 644	xfs_buf_flags_t		flags)
 645{
 646	struct xfs_buf		*bp;
 647	struct xfs_buf		*new_bp;
 648	int			error = 0;
 649
 650	bp = _xfs_buf_find(target, map, nmaps, flags, NULL);
 651	if (likely(bp))
 652		goto found;
 653
 654	new_bp = _xfs_buf_alloc(target, map, nmaps, flags);
 655	if (unlikely(!new_bp))
 656		return NULL;
 657
 658	error = xfs_buf_allocate_memory(new_bp, flags);
 659	if (error) {
 660		xfs_buf_free(new_bp);
 661		return NULL;
 662	}
 663
 664	bp = _xfs_buf_find(target, map, nmaps, flags, new_bp);
 665	if (!bp) {
 666		xfs_buf_free(new_bp);
 667		return NULL;
 668	}
 669
 670	if (bp != new_bp)
 671		xfs_buf_free(new_bp);
 672
 673found:
 674	if (!bp->b_addr) {
 675		error = _xfs_buf_map_pages(bp, flags);
 676		if (unlikely(error)) {
 677			xfs_warn(target->bt_mount,
 678				"%s: failed to map pagesn", __func__);
 679			xfs_buf_relse(bp);
 680			return NULL;
 681		}
 682	}
 683
 684	/*
 685	 * Clear b_error if this is a lookup from a caller that doesn't expect
 686	 * valid data to be found in the buffer.
 687	 */
 688	if (!(flags & XBF_READ))
 689		xfs_buf_ioerror(bp, 0);
 690
 691	XFS_STATS_INC(target->bt_mount, xb_get);
 692	trace_xfs_buf_get(bp, flags, _RET_IP_);
 693	return bp;
 694}
 695
 696STATIC int
 697_xfs_buf_read(
 698	xfs_buf_t		*bp,
 699	xfs_buf_flags_t		flags)
 700{
 701	ASSERT(!(flags & XBF_WRITE));
 702	ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL);
 703
 704	bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD);
 705	bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
 706
 707	if (flags & XBF_ASYNC) {
 708		xfs_buf_submit(bp);
 709		return 0;
 710	}
 711	return xfs_buf_submit_wait(bp);
 712}
 713
 714xfs_buf_t *
 715xfs_buf_read_map(
 716	struct xfs_buftarg	*target,
 717	struct xfs_buf_map	*map,
 718	int			nmaps,
 719	xfs_buf_flags_t		flags,
 720	const struct xfs_buf_ops *ops)
 721{
 722	struct xfs_buf		*bp;
 723
 724	flags |= XBF_READ;
 725
 726	bp = xfs_buf_get_map(target, map, nmaps, flags);
 727	if (bp) {
 728		trace_xfs_buf_read(bp, flags, _RET_IP_);
 729
 730		if (!(bp->b_flags & XBF_DONE)) {
 731			XFS_STATS_INC(target->bt_mount, xb_get_read);
 732			bp->b_ops = ops;
 733			_xfs_buf_read(bp, flags);
 734		} else if (flags & XBF_ASYNC) {
 735			/*
 736			 * Read ahead call which is already satisfied,
 737			 * drop the buffer
 738			 */
 739			xfs_buf_relse(bp);
 740			return NULL;
 741		} else {
 742			/* We do not want read in the flags */
 743			bp->b_flags &= ~XBF_READ;
 744		}
 745	}
 746
 747	return bp;
 748}
 749
 750/*
 751 *	If we are not low on memory then do the readahead in a deadlock
 752 *	safe manner.
 753 */
 754void
 755xfs_buf_readahead_map(
 756	struct xfs_buftarg	*target,
 757	struct xfs_buf_map	*map,
 758	int			nmaps,
 759	const struct xfs_buf_ops *ops)
 760{
 761	if (bdi_read_congested(target->bt_bdi))
 762		return;
 763
 764	xfs_buf_read_map(target, map, nmaps,
 765		     XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD, ops);
 766}
 767
 768/*
 769 * Read an uncached buffer from disk. Allocates and returns a locked
 770 * buffer containing the disk contents or nothing.
 771 */
 772int
 773xfs_buf_read_uncached(
 774	struct xfs_buftarg	*target,
 775	xfs_daddr_t		daddr,
 776	size_t			numblks,
 777	int			flags,
 778	struct xfs_buf		**bpp,
 779	const struct xfs_buf_ops *ops)
 780{
 781	struct xfs_buf		*bp;
 782
 783	*bpp = NULL;
 784
 785	bp = xfs_buf_get_uncached(target, numblks, flags);
 786	if (!bp)
 787		return -ENOMEM;
 788
 789	/* set up the buffer for a read IO */
 790	ASSERT(bp->b_map_count == 1);
 791	bp->b_bn = XFS_BUF_DADDR_NULL;  /* always null for uncached buffers */
 792	bp->b_maps[0].bm_bn = daddr;
 793	bp->b_flags |= XBF_READ;
 794	bp->b_ops = ops;
 795
 796	xfs_buf_submit_wait(bp);
 797	if (bp->b_error) {
 798		int	error = bp->b_error;
 799		xfs_buf_relse(bp);
 800		return error;
 801	}
 802
 803	*bpp = bp;
 804	return 0;
 805}
 806
 807/*
 808 * Return a buffer allocated as an empty buffer and associated to external
 809 * memory via xfs_buf_associate_memory() back to it's empty state.
 810 */
 811void
 812xfs_buf_set_empty(
 813	struct xfs_buf		*bp,
 814	size_t			numblks)
 815{
 816	if (bp->b_pages)
 817		_xfs_buf_free_pages(bp);
 818
 819	bp->b_pages = NULL;
 820	bp->b_page_count = 0;
 821	bp->b_addr = NULL;
 822	bp->b_length = numblks;
 823	bp->b_io_length = numblks;
 824
 825	ASSERT(bp->b_map_count == 1);
 826	bp->b_bn = XFS_BUF_DADDR_NULL;
 827	bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL;
 828	bp->b_maps[0].bm_len = bp->b_length;
 829}
 830
 831static inline struct page *
 832mem_to_page(
 833	void			*addr)
 834{
 835	if ((!is_vmalloc_addr(addr))) {
 836		return virt_to_page(addr);
 837	} else {
 838		return vmalloc_to_page(addr);
 839	}
 840}
 841
 842int
 843xfs_buf_associate_memory(
 844	xfs_buf_t		*bp,
 845	void			*mem,
 846	size_t			len)
 847{
 848	int			rval;
 849	int			i = 0;
 850	unsigned long		pageaddr;
 851	unsigned long		offset;
 852	size_t			buflen;
 853	int			page_count;
 854
 855	pageaddr = (unsigned long)mem & PAGE_MASK;
 856	offset = (unsigned long)mem - pageaddr;
 857	buflen = PAGE_ALIGN(len + offset);
 858	page_count = buflen >> PAGE_SHIFT;
 859
 860	/* Free any previous set of page pointers */
 861	if (bp->b_pages)
 862		_xfs_buf_free_pages(bp);
 863
 864	bp->b_pages = NULL;
 865	bp->b_addr = mem;
 866
 867	rval = _xfs_buf_get_pages(bp, page_count);
 868	if (rval)
 869		return rval;
 870
 871	bp->b_offset = offset;
 872
 873	for (i = 0; i < bp->b_page_count; i++) {
 874		bp->b_pages[i] = mem_to_page((void *)pageaddr);
 875		pageaddr += PAGE_SIZE;
 876	}
 877
 878	bp->b_io_length = BTOBB(len);
 879	bp->b_length = BTOBB(buflen);
 880
 881	return 0;
 882}
 883
 884xfs_buf_t *
 885xfs_buf_get_uncached(
 886	struct xfs_buftarg	*target,
 887	size_t			numblks,
 888	int			flags)
 889{
 890	unsigned long		page_count;
 891	int			error, i;
 892	struct xfs_buf		*bp;
 893	DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
 894
 895	/* flags might contain irrelevant bits, pass only what we care about */
 896	bp = _xfs_buf_alloc(target, &map, 1, flags & XBF_NO_IOACCT);
 897	if (unlikely(bp == NULL))
 898		goto fail;
 899
 900	page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT;
 901	error = _xfs_buf_get_pages(bp, page_count);
 902	if (error)
 903		goto fail_free_buf;
 904
 905	for (i = 0; i < page_count; i++) {
 906		bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
 907		if (!bp->b_pages[i])
 908			goto fail_free_mem;
 909	}
 910	bp->b_flags |= _XBF_PAGES;
 911
 912	error = _xfs_buf_map_pages(bp, 0);
 913	if (unlikely(error)) {
 914		xfs_warn(target->bt_mount,
 915			"%s: failed to map pages", __func__);
 916		goto fail_free_mem;
 917	}
 918
 919	trace_xfs_buf_get_uncached(bp, _RET_IP_);
 920	return bp;
 921
 922 fail_free_mem:
 923	while (--i >= 0)
 924		__free_page(bp->b_pages[i]);
 925	_xfs_buf_free_pages(bp);
 926 fail_free_buf:
 927	xfs_buf_free_maps(bp);
 928	kmem_zone_free(xfs_buf_zone, bp);
 929 fail:
 930	return NULL;
 931}
 932
 933/*
 934 *	Increment reference count on buffer, to hold the buffer concurrently
 935 *	with another thread which may release (free) the buffer asynchronously.
 936 *	Must hold the buffer already to call this function.
 937 */
 938void
 939xfs_buf_hold(
 940	xfs_buf_t		*bp)
 941{
 942	trace_xfs_buf_hold(bp, _RET_IP_);
 943	atomic_inc(&bp->b_hold);
 944}
 945
 946/*
 947 * Release a hold on the specified buffer. If the hold count is 1, the buffer is
 948 * placed on LRU or freed (depending on b_lru_ref).
 949 */
 950void
 951xfs_buf_rele(
 952	xfs_buf_t		*bp)
 953{
 954	struct xfs_perag	*pag = bp->b_pag;
 955	bool			release;
 956	bool			freebuf = false;
 957
 958	trace_xfs_buf_rele(bp, _RET_IP_);
 959
 960	if (!pag) {
 961		ASSERT(list_empty(&bp->b_lru));
 962		if (atomic_dec_and_test(&bp->b_hold)) {
 963			xfs_buf_ioacct_dec(bp);
 964			xfs_buf_free(bp);
 965		}
 966		return;
 967	}
 968
 969	ASSERT(atomic_read(&bp->b_hold) > 0);
 970
 971	release = atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock);
 972	spin_lock(&bp->b_lock);
 973	if (!release) {
 974		/*
 975		 * Drop the in-flight state if the buffer is already on the LRU
 976		 * and it holds the only reference. This is racy because we
 977		 * haven't acquired the pag lock, but the use of _XBF_IN_FLIGHT
 978		 * ensures the decrement occurs only once per-buf.
 979		 */
 980		if ((atomic_read(&bp->b_hold) == 1) && !list_empty(&bp->b_lru))
 981			xfs_buf_ioacct_dec(bp);
 982		goto out_unlock;
 983	}
 984
 985	/* the last reference has been dropped ... */
 986	xfs_buf_ioacct_dec(bp);
 987	if (!(bp->b_flags & XBF_STALE) && atomic_read(&bp->b_lru_ref)) {
 988		/*
 989		 * If the buffer is added to the LRU take a new reference to the
 990		 * buffer for the LRU and clear the (now stale) dispose list
 991		 * state flag
 992		 */
 993		if (list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) {
 994			bp->b_state &= ~XFS_BSTATE_DISPOSE;
 995			atomic_inc(&bp->b_hold);
 996		}
 997		spin_unlock(&pag->pag_buf_lock);
 998	} else {
 999		/*
1000		 * most of the time buffers will already be removed from the
1001		 * LRU, so optimise that case by checking for the
1002		 * XFS_BSTATE_DISPOSE flag indicating the last list the buffer
1003		 * was on was the disposal list
1004		 */
1005		if (!(bp->b_state & XFS_BSTATE_DISPOSE)) {
1006			list_lru_del(&bp->b_target->bt_lru, &bp->b_lru);
1007		} else {
1008			ASSERT(list_empty(&bp->b_lru));
1009		}
 
 
 
 
 
 
 
 
 
 
1010
1011		ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1012		rhashtable_remove_fast(&pag->pag_buf_hash, &bp->b_rhash_head,
1013				       xfs_buf_hash_params);
1014		spin_unlock(&pag->pag_buf_lock);
1015		xfs_perag_put(pag);
1016		freebuf = true;
1017	}
1018
1019out_unlock:
1020	spin_unlock(&bp->b_lock);
1021
1022	if (freebuf)
1023		xfs_buf_free(bp);
1024}
1025
1026
1027/*
1028 *	Lock a buffer object, if it is not already locked.
1029 *
1030 *	If we come across a stale, pinned, locked buffer, we know that we are
1031 *	being asked to lock a buffer that has been reallocated. Because it is
1032 *	pinned, we know that the log has not been pushed to disk and hence it
1033 *	will still be locked.  Rather than continuing to have trylock attempts
1034 *	fail until someone else pushes the log, push it ourselves before
1035 *	returning.  This means that the xfsaild will not get stuck trying
1036 *	to push on stale inode buffers.
1037 */
1038int
1039xfs_buf_trylock(
1040	struct xfs_buf		*bp)
1041{
1042	int			locked;
1043
1044	locked = down_trylock(&bp->b_sema) == 0;
1045	if (locked) {
1046		XB_SET_OWNER(bp);
1047		trace_xfs_buf_trylock(bp, _RET_IP_);
1048	} else {
1049		trace_xfs_buf_trylock_fail(bp, _RET_IP_);
1050	}
1051	return locked;
1052}
1053
1054/*
1055 *	Lock a buffer object.
1056 *
1057 *	If we come across a stale, pinned, locked buffer, we know that we
1058 *	are being asked to lock a buffer that has been reallocated. Because
1059 *	it is pinned, we know that the log has not been pushed to disk and
1060 *	hence it will still be locked. Rather than sleeping until someone
1061 *	else pushes the log, push it ourselves before trying to get the lock.
1062 */
1063void
1064xfs_buf_lock(
1065	struct xfs_buf		*bp)
1066{
1067	trace_xfs_buf_lock(bp, _RET_IP_);
1068
1069	if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
1070		xfs_log_force(bp->b_target->bt_mount, 0);
1071	down(&bp->b_sema);
1072	XB_SET_OWNER(bp);
1073
1074	trace_xfs_buf_lock_done(bp, _RET_IP_);
1075}
1076
1077void
1078xfs_buf_unlock(
1079	struct xfs_buf		*bp)
1080{
1081	XB_CLEAR_OWNER(bp);
1082	up(&bp->b_sema);
1083
1084	trace_xfs_buf_unlock(bp, _RET_IP_);
1085}
1086
1087STATIC void
1088xfs_buf_wait_unpin(
1089	xfs_buf_t		*bp)
1090{
1091	DECLARE_WAITQUEUE	(wait, current);
1092
1093	if (atomic_read(&bp->b_pin_count) == 0)
1094		return;
1095
1096	add_wait_queue(&bp->b_waiters, &wait);
1097	for (;;) {
1098		set_current_state(TASK_UNINTERRUPTIBLE);
1099		if (atomic_read(&bp->b_pin_count) == 0)
1100			break;
1101		io_schedule();
1102	}
1103	remove_wait_queue(&bp->b_waiters, &wait);
1104	set_current_state(TASK_RUNNING);
1105}
1106
1107/*
1108 *	Buffer Utility Routines
1109 */
1110
1111void
1112xfs_buf_ioend(
1113	struct xfs_buf	*bp)
1114{
1115	bool		read = bp->b_flags & XBF_READ;
1116
1117	trace_xfs_buf_iodone(bp, _RET_IP_);
1118
1119	bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1120
1121	/*
1122	 * Pull in IO completion errors now. We are guaranteed to be running
1123	 * single threaded, so we don't need the lock to read b_io_error.
1124	 */
1125	if (!bp->b_error && bp->b_io_error)
1126		xfs_buf_ioerror(bp, bp->b_io_error);
1127
1128	/* Only validate buffers that were read without errors */
1129	if (read && !bp->b_error && bp->b_ops) {
1130		ASSERT(!bp->b_iodone);
1131		bp->b_ops->verify_read(bp);
1132	}
1133
1134	if (!bp->b_error)
1135		bp->b_flags |= XBF_DONE;
1136
1137	if (bp->b_iodone)
1138		(*(bp->b_iodone))(bp);
1139	else if (bp->b_flags & XBF_ASYNC)
1140		xfs_buf_relse(bp);
1141	else
1142		complete(&bp->b_iowait);
1143}
1144
1145static void
1146xfs_buf_ioend_work(
1147	struct work_struct	*work)
1148{
1149	struct xfs_buf		*bp =
1150		container_of(work, xfs_buf_t, b_ioend_work);
1151
1152	xfs_buf_ioend(bp);
1153}
1154
1155static void
1156xfs_buf_ioend_async(
1157	struct xfs_buf	*bp)
1158{
1159	INIT_WORK(&bp->b_ioend_work, xfs_buf_ioend_work);
1160	queue_work(bp->b_ioend_wq, &bp->b_ioend_work);
1161}
1162
1163void
1164xfs_buf_ioerror(
1165	xfs_buf_t		*bp,
1166	int			error)
1167{
1168	ASSERT(error <= 0 && error >= -1000);
1169	bp->b_error = error;
1170	trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1171}
1172
1173void
1174xfs_buf_ioerror_alert(
1175	struct xfs_buf		*bp,
1176	const char		*func)
1177{
1178	xfs_alert(bp->b_target->bt_mount,
1179"metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d",
1180		(__uint64_t)XFS_BUF_ADDR(bp), func, -bp->b_error, bp->b_length);
1181}
1182
1183int
1184xfs_bwrite(
1185	struct xfs_buf		*bp)
1186{
1187	int			error;
1188
1189	ASSERT(xfs_buf_islocked(bp));
1190
1191	bp->b_flags |= XBF_WRITE;
1192	bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q |
1193			 XBF_WRITE_FAIL | XBF_DONE);
1194
1195	error = xfs_buf_submit_wait(bp);
1196	if (error) {
1197		xfs_force_shutdown(bp->b_target->bt_mount,
1198				   SHUTDOWN_META_IO_ERROR);
1199	}
1200	return error;
1201}
1202
1203static void
1204xfs_buf_bio_end_io(
1205	struct bio		*bio)
1206{
1207	struct xfs_buf		*bp = (struct xfs_buf *)bio->bi_private;
1208
1209	/*
1210	 * don't overwrite existing errors - otherwise we can lose errors on
1211	 * buffers that require multiple bios to complete.
1212	 */
1213	if (bio->bi_error)
1214		cmpxchg(&bp->b_io_error, 0, bio->bi_error);
 
 
 
 
1215
1216	if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1217		invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1218
1219	if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1220		xfs_buf_ioend_async(bp);
1221	bio_put(bio);
1222}
1223
1224static void
1225xfs_buf_ioapply_map(
1226	struct xfs_buf	*bp,
1227	int		map,
1228	int		*buf_offset,
1229	int		*count,
1230	int		op,
1231	int		op_flags)
1232{
1233	int		page_index;
1234	int		total_nr_pages = bp->b_page_count;
1235	int		nr_pages;
1236	struct bio	*bio;
1237	sector_t	sector =  bp->b_maps[map].bm_bn;
1238	int		size;
1239	int		offset;
1240
1241	total_nr_pages = bp->b_page_count;
1242
1243	/* skip the pages in the buffer before the start offset */
1244	page_index = 0;
1245	offset = *buf_offset;
1246	while (offset >= PAGE_SIZE) {
1247		page_index++;
1248		offset -= PAGE_SIZE;
1249	}
1250
1251	/*
1252	 * Limit the IO size to the length of the current vector, and update the
1253	 * remaining IO count for the next time around.
1254	 */
1255	size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1256	*count -= size;
1257	*buf_offset += size;
1258
1259next_chunk:
1260	atomic_inc(&bp->b_io_remaining);
1261	nr_pages = min(total_nr_pages, BIO_MAX_PAGES);
 
 
1262
1263	bio = bio_alloc(GFP_NOIO, nr_pages);
1264	bio->bi_bdev = bp->b_target->bt_bdev;
1265	bio->bi_iter.bi_sector = sector;
1266	bio->bi_end_io = xfs_buf_bio_end_io;
1267	bio->bi_private = bp;
1268	bio_set_op_attrs(bio, op, op_flags);
1269
1270	for (; size && nr_pages; nr_pages--, page_index++) {
1271		int	rbytes, nbytes = PAGE_SIZE - offset;
1272
1273		if (nbytes > size)
1274			nbytes = size;
1275
1276		rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1277				      offset);
1278		if (rbytes < nbytes)
1279			break;
1280
1281		offset = 0;
1282		sector += BTOBB(nbytes);
1283		size -= nbytes;
1284		total_nr_pages--;
1285	}
1286
1287	if (likely(bio->bi_iter.bi_size)) {
1288		if (xfs_buf_is_vmapped(bp)) {
1289			flush_kernel_vmap_range(bp->b_addr,
1290						xfs_buf_vmap_len(bp));
1291		}
1292		submit_bio(bio);
1293		if (size)
1294			goto next_chunk;
1295	} else {
1296		/*
1297		 * This is guaranteed not to be the last io reference count
1298		 * because the caller (xfs_buf_submit) holds a count itself.
1299		 */
1300		atomic_dec(&bp->b_io_remaining);
1301		xfs_buf_ioerror(bp, -EIO);
1302		bio_put(bio);
1303	}
1304
1305}
1306
1307STATIC void
1308_xfs_buf_ioapply(
1309	struct xfs_buf	*bp)
1310{
1311	struct blk_plug	plug;
1312	int		op;
1313	int		op_flags = 0;
1314	int		offset;
1315	int		size;
1316	int		i;
1317
1318	/*
1319	 * Make sure we capture only current IO errors rather than stale errors
1320	 * left over from previous use of the buffer (e.g. failed readahead).
1321	 */
1322	bp->b_error = 0;
1323
1324	/*
1325	 * Initialize the I/O completion workqueue if we haven't yet or the
1326	 * submitter has not opted to specify a custom one.
1327	 */
1328	if (!bp->b_ioend_wq)
1329		bp->b_ioend_wq = bp->b_target->bt_mount->m_buf_workqueue;
1330
1331	if (bp->b_flags & XBF_WRITE) {
1332		op = REQ_OP_WRITE;
1333		if (bp->b_flags & XBF_SYNCIO)
1334			op_flags = REQ_SYNC;
 
 
1335		if (bp->b_flags & XBF_FUA)
1336			op_flags |= REQ_FUA;
1337		if (bp->b_flags & XBF_FLUSH)
1338			op_flags |= REQ_PREFLUSH;
1339
1340		/*
1341		 * Run the write verifier callback function if it exists. If
1342		 * this function fails it will mark the buffer with an error and
1343		 * the IO should not be dispatched.
1344		 */
1345		if (bp->b_ops) {
1346			bp->b_ops->verify_write(bp);
1347			if (bp->b_error) {
1348				xfs_force_shutdown(bp->b_target->bt_mount,
1349						   SHUTDOWN_CORRUPT_INCORE);
1350				return;
1351			}
1352		} else if (bp->b_bn != XFS_BUF_DADDR_NULL) {
1353			struct xfs_mount *mp = bp->b_target->bt_mount;
1354
1355			/*
1356			 * non-crc filesystems don't attach verifiers during
1357			 * log recovery, so don't warn for such filesystems.
1358			 */
1359			if (xfs_sb_version_hascrc(&mp->m_sb)) {
1360				xfs_warn(mp,
1361					"%s: no ops on block 0x%llx/0x%x",
1362					__func__, bp->b_bn, bp->b_length);
1363				xfs_hex_dump(bp->b_addr, 64);
1364				dump_stack();
1365			}
1366		}
1367	} else if (bp->b_flags & XBF_READ_AHEAD) {
1368		op = REQ_OP_READ;
1369		op_flags = REQ_RAHEAD;
1370	} else {
1371		op = REQ_OP_READ;
1372	}
1373
1374	/* we only use the buffer cache for meta-data */
1375	op_flags |= REQ_META;
1376
1377	/*
1378	 * Walk all the vectors issuing IO on them. Set up the initial offset
1379	 * into the buffer and the desired IO size before we start -
1380	 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1381	 * subsequent call.
1382	 */
1383	offset = bp->b_offset;
1384	size = BBTOB(bp->b_io_length);
1385	blk_start_plug(&plug);
1386	for (i = 0; i < bp->b_map_count; i++) {
1387		xfs_buf_ioapply_map(bp, i, &offset, &size, op, op_flags);
1388		if (bp->b_error)
1389			break;
1390		if (size <= 0)
1391			break;	/* all done */
1392	}
1393	blk_finish_plug(&plug);
1394}
1395
1396/*
1397 * Asynchronous IO submission path. This transfers the buffer lock ownership and
1398 * the current reference to the IO. It is not safe to reference the buffer after
1399 * a call to this function unless the caller holds an additional reference
1400 * itself.
1401 */
1402void
1403xfs_buf_submit(
1404	struct xfs_buf	*bp)
1405{
1406	trace_xfs_buf_submit(bp, _RET_IP_);
1407
1408	ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1409	ASSERT(bp->b_flags & XBF_ASYNC);
1410
1411	/* on shutdown we stale and complete the buffer immediately */
1412	if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1413		xfs_buf_ioerror(bp, -EIO);
1414		bp->b_flags &= ~XBF_DONE;
1415		xfs_buf_stale(bp);
1416		xfs_buf_ioend(bp);
1417		return;
1418	}
1419
1420	if (bp->b_flags & XBF_WRITE)
1421		xfs_buf_wait_unpin(bp);
1422
1423	/* clear the internal error state to avoid spurious errors */
1424	bp->b_io_error = 0;
1425
1426	/*
1427	 * The caller's reference is released during I/O completion.
1428	 * This occurs some time after the last b_io_remaining reference is
1429	 * released, so after we drop our Io reference we have to have some
1430	 * other reference to ensure the buffer doesn't go away from underneath
1431	 * us. Take a direct reference to ensure we have safe access to the
1432	 * buffer until we are finished with it.
1433	 */
1434	xfs_buf_hold(bp);
1435
1436	/*
1437	 * Set the count to 1 initially, this will stop an I/O completion
1438	 * callout which happens before we have started all the I/O from calling
1439	 * xfs_buf_ioend too early.
1440	 */
1441	atomic_set(&bp->b_io_remaining, 1);
1442	xfs_buf_ioacct_inc(bp);
1443	_xfs_buf_ioapply(bp);
1444
1445	/*
1446	 * If _xfs_buf_ioapply failed, we can get back here with only the IO
1447	 * reference we took above. If we drop it to zero, run completion so
1448	 * that we don't return to the caller with completion still pending.
1449	 */
1450	if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1451		if (bp->b_error)
1452			xfs_buf_ioend(bp);
1453		else
1454			xfs_buf_ioend_async(bp);
1455	}
1456
1457	xfs_buf_rele(bp);
1458	/* Note: it is not safe to reference bp now we've dropped our ref */
1459}
1460
1461/*
1462 * Synchronous buffer IO submission path, read or write.
1463 */
1464int
1465xfs_buf_submit_wait(
1466	struct xfs_buf	*bp)
1467{
1468	int		error;
1469
1470	trace_xfs_buf_submit_wait(bp, _RET_IP_);
1471
1472	ASSERT(!(bp->b_flags & (_XBF_DELWRI_Q | XBF_ASYNC)));
1473
1474	if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1475		xfs_buf_ioerror(bp, -EIO);
1476		xfs_buf_stale(bp);
1477		bp->b_flags &= ~XBF_DONE;
1478		return -EIO;
1479	}
1480
1481	if (bp->b_flags & XBF_WRITE)
1482		xfs_buf_wait_unpin(bp);
1483
1484	/* clear the internal error state to avoid spurious errors */
1485	bp->b_io_error = 0;
1486
1487	/*
1488	 * For synchronous IO, the IO does not inherit the submitters reference
1489	 * count, nor the buffer lock. Hence we cannot release the reference we
1490	 * are about to take until we've waited for all IO completion to occur,
1491	 * including any xfs_buf_ioend_async() work that may be pending.
1492	 */
1493	xfs_buf_hold(bp);
1494
1495	/*
1496	 * Set the count to 1 initially, this will stop an I/O completion
1497	 * callout which happens before we have started all the I/O from calling
1498	 * xfs_buf_ioend too early.
1499	 */
1500	atomic_set(&bp->b_io_remaining, 1);
1501	_xfs_buf_ioapply(bp);
1502
1503	/*
1504	 * make sure we run completion synchronously if it raced with us and is
1505	 * already complete.
1506	 */
1507	if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1508		xfs_buf_ioend(bp);
1509
1510	/* wait for completion before gathering the error from the buffer */
1511	trace_xfs_buf_iowait(bp, _RET_IP_);
1512	wait_for_completion(&bp->b_iowait);
1513	trace_xfs_buf_iowait_done(bp, _RET_IP_);
1514	error = bp->b_error;
1515
1516	/*
1517	 * all done now, we can release the hold that keeps the buffer
1518	 * referenced for the entire IO.
1519	 */
1520	xfs_buf_rele(bp);
1521	return error;
1522}
1523
1524void *
1525xfs_buf_offset(
1526	struct xfs_buf		*bp,
1527	size_t			offset)
1528{
1529	struct page		*page;
1530
1531	if (bp->b_addr)
1532		return bp->b_addr + offset;
1533
1534	offset += bp->b_offset;
1535	page = bp->b_pages[offset >> PAGE_SHIFT];
1536	return page_address(page) + (offset & (PAGE_SIZE-1));
1537}
1538
1539/*
1540 *	Move data into or out of a buffer.
1541 */
1542void
1543xfs_buf_iomove(
1544	xfs_buf_t		*bp,	/* buffer to process		*/
1545	size_t			boff,	/* starting buffer offset	*/
1546	size_t			bsize,	/* length to copy		*/
1547	void			*data,	/* data address			*/
1548	xfs_buf_rw_t		mode)	/* read/write/zero flag		*/
1549{
1550	size_t			bend;
1551
1552	bend = boff + bsize;
1553	while (boff < bend) {
1554		struct page	*page;
1555		int		page_index, page_offset, csize;
1556
1557		page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1558		page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1559		page = bp->b_pages[page_index];
1560		csize = min_t(size_t, PAGE_SIZE - page_offset,
1561				      BBTOB(bp->b_io_length) - boff);
1562
1563		ASSERT((csize + page_offset) <= PAGE_SIZE);
1564
1565		switch (mode) {
1566		case XBRW_ZERO:
1567			memset(page_address(page) + page_offset, 0, csize);
1568			break;
1569		case XBRW_READ:
1570			memcpy(data, page_address(page) + page_offset, csize);
1571			break;
1572		case XBRW_WRITE:
1573			memcpy(page_address(page) + page_offset, data, csize);
1574		}
1575
1576		boff += csize;
1577		data += csize;
1578	}
1579}
1580
1581/*
1582 *	Handling of buffer targets (buftargs).
1583 */
1584
1585/*
1586 * Wait for any bufs with callbacks that have been submitted but have not yet
1587 * returned. These buffers will have an elevated hold count, so wait on those
1588 * while freeing all the buffers only held by the LRU.
1589 */
1590static enum lru_status
1591xfs_buftarg_wait_rele(
1592	struct list_head	*item,
1593	struct list_lru_one	*lru,
1594	spinlock_t		*lru_lock,
1595	void			*arg)
1596
1597{
1598	struct xfs_buf		*bp = container_of(item, struct xfs_buf, b_lru);
1599	struct list_head	*dispose = arg;
1600
1601	if (atomic_read(&bp->b_hold) > 1) {
1602		/* need to wait, so skip it this pass */
1603		trace_xfs_buf_wait_buftarg(bp, _RET_IP_);
1604		return LRU_SKIP;
1605	}
1606	if (!spin_trylock(&bp->b_lock))
1607		return LRU_SKIP;
1608
1609	/*
1610	 * clear the LRU reference count so the buffer doesn't get
1611	 * ignored in xfs_buf_rele().
1612	 */
1613	atomic_set(&bp->b_lru_ref, 0);
1614	bp->b_state |= XFS_BSTATE_DISPOSE;
1615	list_lru_isolate_move(lru, item, dispose);
1616	spin_unlock(&bp->b_lock);
1617	return LRU_REMOVED;
1618}
1619
1620void
1621xfs_wait_buftarg(
1622	struct xfs_buftarg	*btp)
1623{
1624	LIST_HEAD(dispose);
1625	int loop = 0;
1626
1627	/*
1628	 * First wait on the buftarg I/O count for all in-flight buffers to be
1629	 * released. This is critical as new buffers do not make the LRU until
1630	 * they are released.
1631	 *
1632	 * Next, flush the buffer workqueue to ensure all completion processing
1633	 * has finished. Just waiting on buffer locks is not sufficient for
1634	 * async IO as the reference count held over IO is not released until
1635	 * after the buffer lock is dropped. Hence we need to ensure here that
1636	 * all reference counts have been dropped before we start walking the
1637	 * LRU list.
1638	 */
1639	while (percpu_counter_sum(&btp->bt_io_count))
1640		delay(100);
1641	flush_workqueue(btp->bt_mount->m_buf_workqueue);
1642
1643	/* loop until there is nothing left on the lru list. */
1644	while (list_lru_count(&btp->bt_lru)) {
1645		list_lru_walk(&btp->bt_lru, xfs_buftarg_wait_rele,
1646			      &dispose, LONG_MAX);
1647
1648		while (!list_empty(&dispose)) {
1649			struct xfs_buf *bp;
1650			bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1651			list_del_init(&bp->b_lru);
1652			if (bp->b_flags & XBF_WRITE_FAIL) {
1653				xfs_alert(btp->bt_mount,
1654"Corruption Alert: Buffer at block 0x%llx had permanent write failures!",
1655					(long long)bp->b_bn);
1656				xfs_alert(btp->bt_mount,
1657"Please run xfs_repair to determine the extent of the problem.");
1658			}
1659			xfs_buf_rele(bp);
1660		}
1661		if (loop++ != 0)
1662			delay(100);
1663	}
1664}
1665
1666static enum lru_status
1667xfs_buftarg_isolate(
1668	struct list_head	*item,
1669	struct list_lru_one	*lru,
1670	spinlock_t		*lru_lock,
1671	void			*arg)
1672{
1673	struct xfs_buf		*bp = container_of(item, struct xfs_buf, b_lru);
1674	struct list_head	*dispose = arg;
1675
1676	/*
1677	 * we are inverting the lru lock/bp->b_lock here, so use a trylock.
1678	 * If we fail to get the lock, just skip it.
1679	 */
1680	if (!spin_trylock(&bp->b_lock))
1681		return LRU_SKIP;
1682	/*
1683	 * Decrement the b_lru_ref count unless the value is already
1684	 * zero. If the value is already zero, we need to reclaim the
1685	 * buffer, otherwise it gets another trip through the LRU.
1686	 */
1687	if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1688		spin_unlock(&bp->b_lock);
1689		return LRU_ROTATE;
1690	}
1691
1692	bp->b_state |= XFS_BSTATE_DISPOSE;
1693	list_lru_isolate_move(lru, item, dispose);
1694	spin_unlock(&bp->b_lock);
1695	return LRU_REMOVED;
1696}
1697
1698static unsigned long
1699xfs_buftarg_shrink_scan(
1700	struct shrinker		*shrink,
1701	struct shrink_control	*sc)
1702{
1703	struct xfs_buftarg	*btp = container_of(shrink,
1704					struct xfs_buftarg, bt_shrinker);
1705	LIST_HEAD(dispose);
1706	unsigned long		freed;
1707
1708	freed = list_lru_shrink_walk(&btp->bt_lru, sc,
1709				     xfs_buftarg_isolate, &dispose);
1710
1711	while (!list_empty(&dispose)) {
1712		struct xfs_buf *bp;
1713		bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1714		list_del_init(&bp->b_lru);
1715		xfs_buf_rele(bp);
1716	}
1717
1718	return freed;
1719}
1720
1721static unsigned long
1722xfs_buftarg_shrink_count(
1723	struct shrinker		*shrink,
1724	struct shrink_control	*sc)
1725{
1726	struct xfs_buftarg	*btp = container_of(shrink,
1727					struct xfs_buftarg, bt_shrinker);
1728	return list_lru_shrink_count(&btp->bt_lru, sc);
1729}
1730
1731void
1732xfs_free_buftarg(
1733	struct xfs_mount	*mp,
1734	struct xfs_buftarg	*btp)
1735{
1736	unregister_shrinker(&btp->bt_shrinker);
1737	ASSERT(percpu_counter_sum(&btp->bt_io_count) == 0);
1738	percpu_counter_destroy(&btp->bt_io_count);
1739	list_lru_destroy(&btp->bt_lru);
1740
1741	xfs_blkdev_issue_flush(btp);
 
1742
1743	kmem_free(btp);
1744}
1745
1746int
1747xfs_setsize_buftarg(
1748	xfs_buftarg_t		*btp,
1749	unsigned int		sectorsize)
1750{
1751	/* Set up metadata sector size info */
1752	btp->bt_meta_sectorsize = sectorsize;
1753	btp->bt_meta_sectormask = sectorsize - 1;
1754
1755	if (set_blocksize(btp->bt_bdev, sectorsize)) {
1756		xfs_warn(btp->bt_mount,
1757			"Cannot set_blocksize to %u on device %pg",
1758			sectorsize, btp->bt_bdev);
1759		return -EINVAL;
1760	}
1761
1762	/* Set up device logical sector size mask */
1763	btp->bt_logical_sectorsize = bdev_logical_block_size(btp->bt_bdev);
1764	btp->bt_logical_sectormask = bdev_logical_block_size(btp->bt_bdev) - 1;
1765
1766	return 0;
1767}
1768
1769/*
1770 * When allocating the initial buffer target we have not yet
1771 * read in the superblock, so don't know what sized sectors
1772 * are being used at this early stage.  Play safe.
1773 */
1774STATIC int
1775xfs_setsize_buftarg_early(
1776	xfs_buftarg_t		*btp,
1777	struct block_device	*bdev)
1778{
1779	return xfs_setsize_buftarg(btp, bdev_logical_block_size(bdev));
1780}
1781
1782xfs_buftarg_t *
1783xfs_alloc_buftarg(
1784	struct xfs_mount	*mp,
1785	struct block_device	*bdev)
1786{
1787	xfs_buftarg_t		*btp;
1788
1789	btp = kmem_zalloc(sizeof(*btp), KM_SLEEP | KM_NOFS);
1790
1791	btp->bt_mount = mp;
1792	btp->bt_dev =  bdev->bd_dev;
1793	btp->bt_bdev = bdev;
1794	btp->bt_bdi = blk_get_backing_dev_info(bdev);
1795
1796	if (xfs_setsize_buftarg_early(btp, bdev))
1797		goto error;
1798
1799	if (list_lru_init(&btp->bt_lru))
1800		goto error;
1801
1802	if (percpu_counter_init(&btp->bt_io_count, 0, GFP_KERNEL))
1803		goto error;
1804
1805	btp->bt_shrinker.count_objects = xfs_buftarg_shrink_count;
1806	btp->bt_shrinker.scan_objects = xfs_buftarg_shrink_scan;
1807	btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1808	btp->bt_shrinker.flags = SHRINKER_NUMA_AWARE;
1809	register_shrinker(&btp->bt_shrinker);
1810	return btp;
1811
1812error:
1813	kmem_free(btp);
1814	return NULL;
1815}
1816
1817/*
1818 * Add a buffer to the delayed write list.
1819 *
1820 * This queues a buffer for writeout if it hasn't already been.  Note that
1821 * neither this routine nor the buffer list submission functions perform
1822 * any internal synchronization.  It is expected that the lists are thread-local
1823 * to the callers.
1824 *
1825 * Returns true if we queued up the buffer, or false if it already had
1826 * been on the buffer list.
1827 */
1828bool
1829xfs_buf_delwri_queue(
1830	struct xfs_buf		*bp,
1831	struct list_head	*list)
1832{
1833	ASSERT(xfs_buf_islocked(bp));
1834	ASSERT(!(bp->b_flags & XBF_READ));
1835
1836	/*
1837	 * If the buffer is already marked delwri it already is queued up
1838	 * by someone else for imediate writeout.  Just ignore it in that
1839	 * case.
1840	 */
1841	if (bp->b_flags & _XBF_DELWRI_Q) {
1842		trace_xfs_buf_delwri_queued(bp, _RET_IP_);
1843		return false;
1844	}
1845
1846	trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1847
1848	/*
1849	 * If a buffer gets written out synchronously or marked stale while it
1850	 * is on a delwri list we lazily remove it. To do this, the other party
1851	 * clears the  _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
1852	 * It remains referenced and on the list.  In a rare corner case it
1853	 * might get readded to a delwri list after the synchronous writeout, in
1854	 * which case we need just need to re-add the flag here.
1855	 */
1856	bp->b_flags |= _XBF_DELWRI_Q;
1857	if (list_empty(&bp->b_list)) {
1858		atomic_inc(&bp->b_hold);
1859		list_add_tail(&bp->b_list, list);
1860	}
1861
1862	return true;
1863}
1864
1865/*
1866 * Compare function is more complex than it needs to be because
1867 * the return value is only 32 bits and we are doing comparisons
1868 * on 64 bit values
1869 */
1870static int
1871xfs_buf_cmp(
1872	void		*priv,
1873	struct list_head *a,
1874	struct list_head *b)
1875{
1876	struct xfs_buf	*ap = container_of(a, struct xfs_buf, b_list);
1877	struct xfs_buf	*bp = container_of(b, struct xfs_buf, b_list);
1878	xfs_daddr_t		diff;
1879
1880	diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn;
1881	if (diff < 0)
1882		return -1;
1883	if (diff > 0)
1884		return 1;
1885	return 0;
1886}
1887
1888/*
1889 * submit buffers for write.
1890 *
1891 * When we have a large buffer list, we do not want to hold all the buffers
1892 * locked while we block on the request queue waiting for IO dispatch. To avoid
1893 * this problem, we lock and submit buffers in groups of 50, thereby minimising
1894 * the lock hold times for lists which may contain thousands of objects.
1895 *
1896 * To do this, we sort the buffer list before we walk the list to lock and
1897 * submit buffers, and we plug and unplug around each group of buffers we
1898 * submit.
1899 */
1900static int
1901xfs_buf_delwri_submit_buffers(
1902	struct list_head	*buffer_list,
1903	struct list_head	*wait_list)
 
1904{
 
1905	struct xfs_buf		*bp, *n;
1906	LIST_HEAD		(submit_list);
1907	int			pinned = 0;
1908	struct blk_plug		plug;
1909
1910	list_sort(NULL, buffer_list, xfs_buf_cmp);
1911
1912	blk_start_plug(&plug);
1913	list_for_each_entry_safe(bp, n, buffer_list, b_list) {
1914		if (!wait_list) {
1915			if (xfs_buf_ispinned(bp)) {
1916				pinned++;
1917				continue;
1918			}
1919			if (!xfs_buf_trylock(bp))
1920				continue;
1921		} else {
1922			xfs_buf_lock(bp);
1923		}
1924
1925		/*
1926		 * Someone else might have written the buffer synchronously or
1927		 * marked it stale in the meantime.  In that case only the
1928		 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
1929		 * reference and remove it from the list here.
1930		 */
1931		if (!(bp->b_flags & _XBF_DELWRI_Q)) {
1932			list_del_init(&bp->b_list);
1933			xfs_buf_relse(bp);
1934			continue;
1935		}
1936
 
1937		trace_xfs_buf_delwri_split(bp, _RET_IP_);
 
 
 
 
 
 
 
 
1938
1939		/*
1940		 * We do all IO submission async. This means if we need
1941		 * to wait for IO completion we need to take an extra
1942		 * reference so the buffer is still valid on the other
1943		 * side. We need to move the buffer onto the io_list
1944		 * at this point so the caller can still access it.
1945		 */
1946		bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_WRITE_FAIL);
1947		bp->b_flags |= XBF_WRITE | XBF_ASYNC;
1948		if (wait_list) {
1949			xfs_buf_hold(bp);
1950			list_move_tail(&bp->b_list, wait_list);
1951		} else
1952			list_del_init(&bp->b_list);
1953
1954		xfs_buf_submit(bp);
1955	}
1956	blk_finish_plug(&plug);
1957
1958	return pinned;
1959}
1960
1961/*
1962 * Write out a buffer list asynchronously.
1963 *
1964 * This will take the @buffer_list, write all non-locked and non-pinned buffers
1965 * out and not wait for I/O completion on any of the buffers.  This interface
1966 * is only safely useable for callers that can track I/O completion by higher
1967 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
1968 * function.
1969 */
1970int
1971xfs_buf_delwri_submit_nowait(
1972	struct list_head	*buffer_list)
1973{
1974	return xfs_buf_delwri_submit_buffers(buffer_list, NULL);
 
1975}
1976
1977/*
1978 * Write out a buffer list synchronously.
1979 *
1980 * This will take the @buffer_list, write all buffers out and wait for I/O
1981 * completion on all of the buffers. @buffer_list is consumed by the function,
1982 * so callers must have some other way of tracking buffers if they require such
1983 * functionality.
1984 */
1985int
1986xfs_buf_delwri_submit(
1987	struct list_head	*buffer_list)
1988{
1989	LIST_HEAD		(wait_list);
1990	int			error = 0, error2;
1991	struct xfs_buf		*bp;
1992
1993	xfs_buf_delwri_submit_buffers(buffer_list, &wait_list);
1994
1995	/* Wait for IO to complete. */
1996	while (!list_empty(&wait_list)) {
1997		bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
1998
1999		list_del_init(&bp->b_list);
2000
2001		/* locking the buffer will wait for async IO completion. */
2002		xfs_buf_lock(bp);
2003		error2 = bp->b_error;
2004		xfs_buf_relse(bp);
2005		if (!error)
2006			error = error2;
2007	}
2008
2009	return error;
2010}
2011
2012int __init
2013xfs_buf_init(void)
2014{
2015	xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
2016						KM_ZONE_HWALIGN, NULL);
2017	if (!xfs_buf_zone)
2018		goto out;
2019
2020	return 0;
2021
2022 out:
2023	return -ENOMEM;
2024}
2025
2026void
2027xfs_buf_terminate(void)
2028{
2029	kmem_zone_destroy(xfs_buf_zone);
2030}