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