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