xref: /openbmc/linux/fs/xfs/xfs_buf.c (revision 3286f88f)
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, !(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(&bp->b_lock);
106  	if (!(bp->b_state & XFS_BSTATE_IN_FLIGHT)) {
107  		bp->b_state |= XFS_BSTATE_IN_FLIGHT;
108  		percpu_counter_inc(&bp->b_target->bt_io_count);
109  	}
110  	spin_unlock(&bp->b_lock);
111  }
112  
113  /*
114   * Clear the in-flight state on a buffer about to be released to the LRU or
115   * freed and unaccount from the buftarg.
116   */
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(&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(&bp->b_lock);
134  	__xfs_buf_ioacct_dec(bp);
135  	spin_unlock(&bp->b_lock);
136  }
137  
138  /*
139   * When we mark a buffer stale, we remove the buffer from the LRU and clear the
140   * b_lru_ref count so that the buffer is freed immediately when the buffer
141   * reference count falls to zero. If the buffer is already on the LRU, we need
142   * to remove the reference that LRU holds on the buffer.
143   *
144   * This prevents build-up of stale buffers on the LRU.
145   */
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(&bp->b_lock);
168  	__xfs_buf_ioacct_dec(bp);
169  
170  	atomic_set(&bp->b_lru_ref, 0);
171  	if (!(bp->b_state & XFS_BSTATE_DISPOSE) &&
172  	    (list_lru_del(&bp->b_target->bt_lru, &bp->b_lru)))
173  		atomic_dec(&bp->b_hold);
174  
175  	ASSERT(atomic_read(&bp->b_hold) >= 1);
176  	spin_unlock(&bp->b_lock);
177  }
178  
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(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(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(xfs_buf_cache, GFP_NOFS | __GFP_NOFAIL);
226  
227  	/*
228  	 * We don't want certain flags to appear in b_flags unless they are
229  	 * specifically set by later operations on the buffer.
230  	 */
231  	flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD);
232  
233  	atomic_set(&bp->b_hold, 1);
234  	atomic_set(&bp->b_lru_ref, 1);
235  	init_completion(&bp->b_iowait);
236  	INIT_LIST_HEAD(&bp->b_lru);
237  	INIT_LIST_HEAD(&bp->b_list);
238  	INIT_LIST_HEAD(&bp->b_li_list);
239  	sema_init(&bp->b_sema, 0); /* held, no waiters */
240  	spin_lock_init(&bp->b_lock);
241  	bp->b_target = target;
242  	bp->b_mount = target->bt_mount;
243  	bp->b_flags = flags;
244  
245  	/*
246  	 * Set length and io_length to the same value initially.
247  	 * I/O routines should use io_length, which will be the same in
248  	 * most cases but may be reset (e.g. XFS recovery).
249  	 */
250  	error = xfs_buf_get_maps(bp, nmaps);
251  	if (error)  {
252  		kmem_cache_free(xfs_buf_cache, bp);
253  		return error;
254  	}
255  
256  	bp->b_rhash_key = map[0].bm_bn;
257  	bp->b_length = 0;
258  	for (i = 0; i < nmaps; i++) {
259  		bp->b_maps[i].bm_bn = map[i].bm_bn;
260  		bp->b_maps[i].bm_len = map[i].bm_len;
261  		bp->b_length += map[i].bm_len;
262  	}
263  
264  	atomic_set(&bp->b_pin_count, 0);
265  	init_waitqueue_head(&bp->b_waiters);
266  
267  	XFS_STATS_INC(bp->b_mount, xb_create);
268  	trace_xfs_buf_init(bp, _RET_IP_);
269  
270  	*bpp = bp;
271  	return 0;
272  }
273  
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(bp->b_addr, bp->b_page_count);
284  
285  	for (i = 0; i < bp->b_page_count; i++) {
286  		if (bp->b_pages[i])
287  			__free_page(bp->b_pages[i]);
288  	}
289  	mm_account_reclaimed_pages(bp->b_page_count);
290  
291  	if (bp->b_pages != bp->b_page_array)
292  		kmem_free(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(xfs_buf_cache, 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(bp->b_addr);
319  
320  	call_rcu(&bp->b_rcu, 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(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(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(sizeof(struct page *) * bp->b_page_count,
373  					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_mask, bp->b_page_count,
392  						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_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(bp->b_pages, bp->b_page_count,
441  						-1);
442  			if (bp->b_addr)
443  				break;
444  			vm_unmap_aliases();
445  		} while (retried++ <= 1);
446  		memalloc_nofs_restore(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  		ASSERT(bp->b_flags & XBF_STALE);
485  		return 1;
486  	}
487  	return 0;
488  }
489  
490  static const struct rhashtable_params xfs_buf_hash_params = {
491  	.min_size		= 32,	/* empty AGs have minimal footprint */
492  	.nelem_hint		= 16,
493  	.key_len		= sizeof(xfs_daddr_t),
494  	.key_offset		= offsetof(struct xfs_buf, b_rhash_key),
495  	.head_offset		= offsetof(struct xfs_buf, b_rhash_head),
496  	.automatic_shrinking	= true,
497  	.obj_cmpfn		= _xfs_buf_obj_cmp,
498  };
499  
500  int
501  xfs_buf_hash_init(
502  	struct xfs_perag	*pag)
503  {
504  	spin_lock_init(&pag->pag_buf_lock);
505  	return rhashtable_init(&pag->pag_buf_hash, &xfs_buf_hash_params);
506  }
507  
508  void
509  xfs_buf_hash_destroy(
510  	struct xfs_perag	*pag)
511  {
512  	rhashtable_destroy(&pag->pag_buf_hash);
513  }
514  
515  static int
516  xfs_buf_map_verify(
517  	struct xfs_buftarg	*btp,
518  	struct xfs_buf_map	*map)
519  {
520  	xfs_daddr_t		eofs;
521  
522  	/* Check for IOs smaller than the sector size / not sector aligned */
523  	ASSERT(!(BBTOB(map->bm_len) < btp->bt_meta_sectorsize));
524  	ASSERT(!(BBTOB(map->bm_bn) & (xfs_off_t)btp->bt_meta_sectormask));
525  
526  	/*
527  	 * Corrupted block numbers can get through to here, unfortunately, so we
528  	 * have to check that the buffer falls within the filesystem bounds.
529  	 */
530  	eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks);
531  	if (map->bm_bn < 0 || map->bm_bn >= eofs) {
532  		xfs_alert(btp->bt_mount,
533  			  "%s: daddr 0x%llx out of range, EOFS 0x%llx",
534  			  __func__, map->bm_bn, eofs);
535  		WARN_ON(1);
536  		return -EFSCORRUPTED;
537  	}
538  	return 0;
539  }
540  
541  static int
542  xfs_buf_find_lock(
543  	struct xfs_buf          *bp,
544  	xfs_buf_flags_t		flags)
545  {
546  	if (flags & XBF_TRYLOCK) {
547  		if (!xfs_buf_trylock(bp)) {
548  			XFS_STATS_INC(bp->b_mount, xb_busy_locked);
549  			return -EAGAIN;
550  		}
551  	} else {
552  		xfs_buf_lock(bp);
553  		XFS_STATS_INC(bp->b_mount, xb_get_locked_waited);
554  	}
555  
556  	/*
557  	 * if the buffer is stale, clear all the external state associated with
558  	 * it. We need to keep flags such as how we allocated the buffer memory
559  	 * intact here.
560  	 */
561  	if (bp->b_flags & XBF_STALE) {
562  		ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
563  		bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
564  		bp->b_ops = NULL;
565  	}
566  	return 0;
567  }
568  
569  static inline int
570  xfs_buf_lookup(
571  	struct xfs_perag	*pag,
572  	struct xfs_buf_map	*map,
573  	xfs_buf_flags_t		flags,
574  	struct xfs_buf		**bpp)
575  {
576  	struct xfs_buf          *bp;
577  	int			error;
578  
579  	rcu_read_lock();
580  	bp = rhashtable_lookup(&pag->pag_buf_hash, map, xfs_buf_hash_params);
581  	if (!bp || !atomic_inc_not_zero(&bp->b_hold)) {
582  		rcu_read_unlock();
583  		return -ENOENT;
584  	}
585  	rcu_read_unlock();
586  
587  	error = xfs_buf_find_lock(bp, flags);
588  	if (error) {
589  		xfs_buf_rele(bp);
590  		return error;
591  	}
592  
593  	trace_xfs_buf_find(bp, flags, _RET_IP_);
594  	*bpp = bp;
595  	return 0;
596  }
597  
598  /*
599   * Insert the new_bp into the hash table. This consumes the perag reference
600   * taken for the lookup regardless of the result of the insert.
601   */
602  static int
603  xfs_buf_find_insert(
604  	struct xfs_buftarg	*btp,
605  	struct xfs_perag	*pag,
606  	struct xfs_buf_map	*cmap,
607  	struct xfs_buf_map	*map,
608  	int			nmaps,
609  	xfs_buf_flags_t		flags,
610  	struct xfs_buf		**bpp)
611  {
612  	struct xfs_buf		*new_bp;
613  	struct xfs_buf		*bp;
614  	int			error;
615  
616  	error = _xfs_buf_alloc(btp, map, nmaps, flags, &new_bp);
617  	if (error)
618  		goto out_drop_pag;
619  
620  	/*
621  	 * For buffers that fit entirely within a single page, first attempt to
622  	 * allocate the memory from the heap to minimise memory usage. If we
623  	 * can't get heap memory for these small buffers, we fall back to using
624  	 * the page allocator.
625  	 */
626  	if (BBTOB(new_bp->b_length) >= PAGE_SIZE ||
627  	    xfs_buf_alloc_kmem(new_bp, flags) < 0) {
628  		error = xfs_buf_alloc_pages(new_bp, flags);
629  		if (error)
630  			goto out_free_buf;
631  	}
632  
633  	spin_lock(&pag->pag_buf_lock);
634  	bp = rhashtable_lookup_get_insert_fast(&pag->pag_buf_hash,
635  			&new_bp->b_rhash_head, xfs_buf_hash_params);
636  	if (IS_ERR(bp)) {
637  		error = PTR_ERR(bp);
638  		spin_unlock(&pag->pag_buf_lock);
639  		goto out_free_buf;
640  	}
641  	if (bp) {
642  		/* found an existing buffer */
643  		atomic_inc(&bp->b_hold);
644  		spin_unlock(&pag->pag_buf_lock);
645  		error = xfs_buf_find_lock(bp, flags);
646  		if (error)
647  			xfs_buf_rele(bp);
648  		else
649  			*bpp = bp;
650  		goto out_free_buf;
651  	}
652  
653  	/* The new buffer keeps the perag reference until it is freed. */
654  	new_bp->b_pag = pag;
655  	spin_unlock(&pag->pag_buf_lock);
656  	*bpp = new_bp;
657  	return 0;
658  
659  out_free_buf:
660  	xfs_buf_free(new_bp);
661  out_drop_pag:
662  	xfs_perag_put(pag);
663  	return error;
664  }
665  
666  /*
667   * Assembles a buffer covering the specified range. The code is optimised for
668   * cache hits, as metadata intensive workloads will see 3 orders of magnitude
669   * more hits than misses.
670   */
671  int
672  xfs_buf_get_map(
673  	struct xfs_buftarg	*btp,
674  	struct xfs_buf_map	*map,
675  	int			nmaps,
676  	xfs_buf_flags_t		flags,
677  	struct xfs_buf		**bpp)
678  {
679  	struct xfs_perag	*pag;
680  	struct xfs_buf		*bp = NULL;
681  	struct xfs_buf_map	cmap = { .bm_bn = map[0].bm_bn };
682  	int			error;
683  	int			i;
684  
685  	for (i = 0; i < nmaps; i++)
686  		cmap.bm_len += map[i].bm_len;
687  
688  	error = xfs_buf_map_verify(btp, &cmap);
689  	if (error)
690  		return error;
691  
692  	pag = xfs_perag_get(btp->bt_mount,
693  			    xfs_daddr_to_agno(btp->bt_mount, cmap.bm_bn));
694  
695  	error = xfs_buf_lookup(pag, &cmap, flags, &bp);
696  	if (error && error != -ENOENT)
697  		goto out_put_perag;
698  
699  	/* cache hits always outnumber misses by at least 10:1 */
700  	if (unlikely(!bp)) {
701  		XFS_STATS_INC(btp->bt_mount, xb_miss_locked);
702  
703  		if (flags & XBF_INCORE)
704  			goto out_put_perag;
705  
706  		/* xfs_buf_find_insert() consumes the perag reference. */
707  		error = xfs_buf_find_insert(btp, pag, &cmap, map, nmaps,
708  				flags, &bp);
709  		if (error)
710  			return error;
711  	} else {
712  		XFS_STATS_INC(btp->bt_mount, xb_get_locked);
713  		xfs_perag_put(pag);
714  	}
715  
716  	/* We do not hold a perag reference anymore. */
717  	if (!bp->b_addr) {
718  		error = _xfs_buf_map_pages(bp, flags);
719  		if (unlikely(error)) {
720  			xfs_warn_ratelimited(btp->bt_mount,
721  				"%s: failed to map %u pages", __func__,
722  				bp->b_page_count);
723  			xfs_buf_relse(bp);
724  			return error;
725  		}
726  	}
727  
728  	/*
729  	 * Clear b_error if this is a lookup from a caller that doesn't expect
730  	 * valid data to be found in the buffer.
731  	 */
732  	if (!(flags & XBF_READ))
733  		xfs_buf_ioerror(bp, 0);
734  
735  	XFS_STATS_INC(btp->bt_mount, xb_get);
736  	trace_xfs_buf_get(bp, flags, _RET_IP_);
737  	*bpp = bp;
738  	return 0;
739  
740  out_put_perag:
741  	xfs_perag_put(pag);
742  	return error;
743  }
744  
745  int
746  _xfs_buf_read(
747  	struct xfs_buf		*bp,
748  	xfs_buf_flags_t		flags)
749  {
750  	ASSERT(!(flags & XBF_WRITE));
751  	ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL);
752  
753  	bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD | XBF_DONE);
754  	bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
755  
756  	return xfs_buf_submit(bp);
757  }
758  
759  /*
760   * Reverify a buffer found in cache without an attached ->b_ops.
761   *
762   * If the caller passed an ops structure and the buffer doesn't have ops
763   * assigned, set the ops and use it to verify the contents. If verification
764   * fails, clear XBF_DONE. We assume the buffer has no recorded errors and is
765   * already in XBF_DONE state on entry.
766   *
767   * Under normal operations, every in-core buffer is verified on read I/O
768   * completion. There are two scenarios that can lead to in-core buffers without
769   * an assigned ->b_ops. The first is during log recovery of buffers on a V4
770   * filesystem, though these buffers are purged at the end of recovery. The
771   * other is online repair, which intentionally reads with a NULL buffer ops to
772   * run several verifiers across an in-core buffer in order to establish buffer
773   * type.  If repair can't establish that, the buffer will be left in memory
774   * with NULL buffer ops.
775   */
776  int
777  xfs_buf_reverify(
778  	struct xfs_buf		*bp,
779  	const struct xfs_buf_ops *ops)
780  {
781  	ASSERT(bp->b_flags & XBF_DONE);
782  	ASSERT(bp->b_error == 0);
783  
784  	if (!ops || bp->b_ops)
785  		return 0;
786  
787  	bp->b_ops = ops;
788  	bp->b_ops->verify_read(bp);
789  	if (bp->b_error)
790  		bp->b_flags &= ~XBF_DONE;
791  	return bp->b_error;
792  }
793  
794  int
795  xfs_buf_read_map(
796  	struct xfs_buftarg	*target,
797  	struct xfs_buf_map	*map,
798  	int			nmaps,
799  	xfs_buf_flags_t		flags,
800  	struct xfs_buf		**bpp,
801  	const struct xfs_buf_ops *ops,
802  	xfs_failaddr_t		fa)
803  {
804  	struct xfs_buf		*bp;
805  	int			error;
806  
807  	flags |= XBF_READ;
808  	*bpp = NULL;
809  
810  	error = xfs_buf_get_map(target, map, nmaps, flags, &bp);
811  	if (error)
812  		return error;
813  
814  	trace_xfs_buf_read(bp, flags, _RET_IP_);
815  
816  	if (!(bp->b_flags & XBF_DONE)) {
817  		/* Initiate the buffer read and wait. */
818  		XFS_STATS_INC(target->bt_mount, xb_get_read);
819  		bp->b_ops = ops;
820  		error = _xfs_buf_read(bp, flags);
821  
822  		/* Readahead iodone already dropped the buffer, so exit. */
823  		if (flags & XBF_ASYNC)
824  			return 0;
825  	} else {
826  		/* Buffer already read; all we need to do is check it. */
827  		error = xfs_buf_reverify(bp, ops);
828  
829  		/* Readahead already finished; drop the buffer and exit. */
830  		if (flags & XBF_ASYNC) {
831  			xfs_buf_relse(bp);
832  			return 0;
833  		}
834  
835  		/* We do not want read in the flags */
836  		bp->b_flags &= ~XBF_READ;
837  		ASSERT(bp->b_ops != NULL || ops == NULL);
838  	}
839  
840  	/*
841  	 * If we've had a read error, then the contents of the buffer are
842  	 * invalid and should not be used. To ensure that a followup read tries
843  	 * to pull the buffer from disk again, we clear the XBF_DONE flag and
844  	 * mark the buffer stale. This ensures that anyone who has a current
845  	 * reference to the buffer will interpret it's contents correctly and
846  	 * future cache lookups will also treat it as an empty, uninitialised
847  	 * buffer.
848  	 */
849  	if (error) {
850  		/*
851  		 * Check against log shutdown for error reporting because
852  		 * metadata writeback may require a read first and we need to
853  		 * report errors in metadata writeback until the log is shut
854  		 * down. High level transaction read functions already check
855  		 * against mount shutdown, anyway, so we only need to be
856  		 * concerned about low level IO interactions here.
857  		 */
858  		if (!xlog_is_shutdown(target->bt_mount->m_log))
859  			xfs_buf_ioerror_alert(bp, fa);
860  
861  		bp->b_flags &= ~XBF_DONE;
862  		xfs_buf_stale(bp);
863  		xfs_buf_relse(bp);
864  
865  		/* bad CRC means corrupted metadata */
866  		if (error == -EFSBADCRC)
867  			error = -EFSCORRUPTED;
868  		return error;
869  	}
870  
871  	*bpp = bp;
872  	return 0;
873  }
874  
875  /*
876   *	If we are not low on memory then do the readahead in a deadlock
877   *	safe manner.
878   */
879  void
880  xfs_buf_readahead_map(
881  	struct xfs_buftarg	*target,
882  	struct xfs_buf_map	*map,
883  	int			nmaps,
884  	const struct xfs_buf_ops *ops)
885  {
886  	struct xfs_buf		*bp;
887  
888  	xfs_buf_read_map(target, map, nmaps,
889  		     XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD, &bp, ops,
890  		     __this_address);
891  }
892  
893  /*
894   * Read an uncached buffer from disk. Allocates and returns a locked
895   * buffer containing the disk contents or nothing. Uncached buffers always have
896   * a cache index of XFS_BUF_DADDR_NULL so we can easily determine if the buffer
897   * is cached or uncached during fault diagnosis.
898   */
899  int
900  xfs_buf_read_uncached(
901  	struct xfs_buftarg	*target,
902  	xfs_daddr_t		daddr,
903  	size_t			numblks,
904  	xfs_buf_flags_t		flags,
905  	struct xfs_buf		**bpp,
906  	const struct xfs_buf_ops *ops)
907  {
908  	struct xfs_buf		*bp;
909  	int			error;
910  
911  	*bpp = NULL;
912  
913  	error = xfs_buf_get_uncached(target, numblks, flags, &bp);
914  	if (error)
915  		return error;
916  
917  	/* set up the buffer for a read IO */
918  	ASSERT(bp->b_map_count == 1);
919  	bp->b_rhash_key = XFS_BUF_DADDR_NULL;
920  	bp->b_maps[0].bm_bn = daddr;
921  	bp->b_flags |= XBF_READ;
922  	bp->b_ops = ops;
923  
924  	xfs_buf_submit(bp);
925  	if (bp->b_error) {
926  		error = bp->b_error;
927  		xfs_buf_relse(bp);
928  		return error;
929  	}
930  
931  	*bpp = bp;
932  	return 0;
933  }
934  
935  int
936  xfs_buf_get_uncached(
937  	struct xfs_buftarg	*target,
938  	size_t			numblks,
939  	xfs_buf_flags_t		flags,
940  	struct xfs_buf		**bpp)
941  {
942  	int			error;
943  	struct xfs_buf		*bp;
944  	DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
945  
946  	*bpp = NULL;
947  
948  	/* flags might contain irrelevant bits, pass only what we care about */
949  	error = _xfs_buf_alloc(target, &map, 1, flags & XBF_NO_IOACCT, &bp);
950  	if (error)
951  		return error;
952  
953  	error = xfs_buf_alloc_pages(bp, flags);
954  	if (error)
955  		goto fail_free_buf;
956  
957  	error = _xfs_buf_map_pages(bp, 0);
958  	if (unlikely(error)) {
959  		xfs_warn(target->bt_mount,
960  			"%s: failed to map pages", __func__);
961  		goto fail_free_buf;
962  	}
963  
964  	trace_xfs_buf_get_uncached(bp, _RET_IP_);
965  	*bpp = bp;
966  	return 0;
967  
968  fail_free_buf:
969  	xfs_buf_free(bp);
970  	return error;
971  }
972  
973  /*
974   *	Increment reference count on buffer, to hold the buffer concurrently
975   *	with another thread which may release (free) the buffer asynchronously.
976   *	Must hold the buffer already to call this function.
977   */
978  void
979  xfs_buf_hold(
980  	struct xfs_buf		*bp)
981  {
982  	trace_xfs_buf_hold(bp, _RET_IP_);
983  	atomic_inc(&bp->b_hold);
984  }
985  
986  /*
987   * Release a hold on the specified buffer. If the hold count is 1, the buffer is
988   * placed on LRU or freed (depending on b_lru_ref).
989   */
990  void
991  xfs_buf_rele(
992  	struct xfs_buf		*bp)
993  {
994  	struct xfs_perag	*pag = bp->b_pag;
995  	bool			release;
996  	bool			freebuf = false;
997  
998  	trace_xfs_buf_rele(bp, _RET_IP_);
999  
1000  	if (!pag) {
1001  		ASSERT(list_empty(&bp->b_lru));
1002  		if (atomic_dec_and_test(&bp->b_hold)) {
1003  			xfs_buf_ioacct_dec(bp);
1004  			xfs_buf_free(bp);
1005  		}
1006  		return;
1007  	}
1008  
1009  	ASSERT(atomic_read(&bp->b_hold) > 0);
1010  
1011  	/*
1012  	 * We grab the b_lock here first to serialise racing xfs_buf_rele()
1013  	 * calls. The pag_buf_lock being taken on the last reference only
1014  	 * serialises against racing lookups in xfs_buf_find(). IOWs, the second
1015  	 * to last reference we drop here is not serialised against the last
1016  	 * reference until we take bp->b_lock. Hence if we don't grab b_lock
1017  	 * first, the last "release" reference can win the race to the lock and
1018  	 * free the buffer before the second-to-last reference is processed,
1019  	 * leading to a use-after-free scenario.
1020  	 */
1021  	spin_lock(&bp->b_lock);
1022  	release = atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock);
1023  	if (!release) {
1024  		/*
1025  		 * Drop the in-flight state if the buffer is already on the LRU
1026  		 * and it holds the only reference. This is racy because we
1027  		 * haven't acquired the pag lock, but the use of _XBF_IN_FLIGHT
1028  		 * ensures the decrement occurs only once per-buf.
1029  		 */
1030  		if ((atomic_read(&bp->b_hold) == 1) && !list_empty(&bp->b_lru))
1031  			__xfs_buf_ioacct_dec(bp);
1032  		goto out_unlock;
1033  	}
1034  
1035  	/* the last reference has been dropped ... */
1036  	__xfs_buf_ioacct_dec(bp);
1037  	if (!(bp->b_flags & XBF_STALE) && atomic_read(&bp->b_lru_ref)) {
1038  		/*
1039  		 * If the buffer is added to the LRU take a new reference to the
1040  		 * buffer for the LRU and clear the (now stale) dispose list
1041  		 * state flag
1042  		 */
1043  		if (list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) {
1044  			bp->b_state &= ~XFS_BSTATE_DISPOSE;
1045  			atomic_inc(&bp->b_hold);
1046  		}
1047  		spin_unlock(&pag->pag_buf_lock);
1048  	} else {
1049  		/*
1050  		 * most of the time buffers will already be removed from the
1051  		 * LRU, so optimise that case by checking for the
1052  		 * XFS_BSTATE_DISPOSE flag indicating the last list the buffer
1053  		 * was on was the disposal list
1054  		 */
1055  		if (!(bp->b_state & XFS_BSTATE_DISPOSE)) {
1056  			list_lru_del(&bp->b_target->bt_lru, &bp->b_lru);
1057  		} else {
1058  			ASSERT(list_empty(&bp->b_lru));
1059  		}
1060  
1061  		ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1062  		rhashtable_remove_fast(&pag->pag_buf_hash, &bp->b_rhash_head,
1063  				       xfs_buf_hash_params);
1064  		spin_unlock(&pag->pag_buf_lock);
1065  		xfs_perag_put(pag);
1066  		freebuf = true;
1067  	}
1068  
1069  out_unlock:
1070  	spin_unlock(&bp->b_lock);
1071  
1072  	if (freebuf)
1073  		xfs_buf_free(bp);
1074  }
1075  
1076  
1077  /*
1078   *	Lock a buffer object, if it is not already locked.
1079   *
1080   *	If we come across a stale, pinned, locked buffer, we know that we are
1081   *	being asked to lock a buffer that has been reallocated. Because it is
1082   *	pinned, we know that the log has not been pushed to disk and hence it
1083   *	will still be locked.  Rather than continuing to have trylock attempts
1084   *	fail until someone else pushes the log, push it ourselves before
1085   *	returning.  This means that the xfsaild will not get stuck trying
1086   *	to push on stale inode buffers.
1087   */
1088  int
1089  xfs_buf_trylock(
1090  	struct xfs_buf		*bp)
1091  {
1092  	int			locked;
1093  
1094  	locked = down_trylock(&bp->b_sema) == 0;
1095  	if (locked)
1096  		trace_xfs_buf_trylock(bp, _RET_IP_);
1097  	else
1098  		trace_xfs_buf_trylock_fail(bp, _RET_IP_);
1099  	return locked;
1100  }
1101  
1102  /*
1103   *	Lock a buffer object.
1104   *
1105   *	If we come across a stale, pinned, locked buffer, we know that we
1106   *	are being asked to lock a buffer that has been reallocated. Because
1107   *	it is pinned, we know that the log has not been pushed to disk and
1108   *	hence it will still be locked. Rather than sleeping until someone
1109   *	else pushes the log, push it ourselves before trying to get the lock.
1110   */
1111  void
1112  xfs_buf_lock(
1113  	struct xfs_buf		*bp)
1114  {
1115  	trace_xfs_buf_lock(bp, _RET_IP_);
1116  
1117  	if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
1118  		xfs_log_force(bp->b_mount, 0);
1119  	down(&bp->b_sema);
1120  
1121  	trace_xfs_buf_lock_done(bp, _RET_IP_);
1122  }
1123  
1124  void
1125  xfs_buf_unlock(
1126  	struct xfs_buf		*bp)
1127  {
1128  	ASSERT(xfs_buf_islocked(bp));
1129  
1130  	up(&bp->b_sema);
1131  	trace_xfs_buf_unlock(bp, _RET_IP_);
1132  }
1133  
1134  STATIC void
1135  xfs_buf_wait_unpin(
1136  	struct xfs_buf		*bp)
1137  {
1138  	DECLARE_WAITQUEUE	(wait, current);
1139  
1140  	if (atomic_read(&bp->b_pin_count) == 0)
1141  		return;
1142  
1143  	add_wait_queue(&bp->b_waiters, &wait);
1144  	for (;;) {
1145  		set_current_state(TASK_UNINTERRUPTIBLE);
1146  		if (atomic_read(&bp->b_pin_count) == 0)
1147  			break;
1148  		io_schedule();
1149  	}
1150  	remove_wait_queue(&bp->b_waiters, &wait);
1151  	set_current_state(TASK_RUNNING);
1152  }
1153  
1154  static void
1155  xfs_buf_ioerror_alert_ratelimited(
1156  	struct xfs_buf		*bp)
1157  {
1158  	static unsigned long	lasttime;
1159  	static struct xfs_buftarg *lasttarg;
1160  
1161  	if (bp->b_target != lasttarg ||
1162  	    time_after(jiffies, (lasttime + 5*HZ))) {
1163  		lasttime = jiffies;
1164  		xfs_buf_ioerror_alert(bp, __this_address);
1165  	}
1166  	lasttarg = bp->b_target;
1167  }
1168  
1169  /*
1170   * Account for this latest trip around the retry handler, and decide if
1171   * we've failed enough times to constitute a permanent failure.
1172   */
1173  static bool
1174  xfs_buf_ioerror_permanent(
1175  	struct xfs_buf		*bp,
1176  	struct xfs_error_cfg	*cfg)
1177  {
1178  	struct xfs_mount	*mp = bp->b_mount;
1179  
1180  	if (cfg->max_retries != XFS_ERR_RETRY_FOREVER &&
1181  	    ++bp->b_retries > cfg->max_retries)
1182  		return true;
1183  	if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER &&
1184  	    time_after(jiffies, cfg->retry_timeout + bp->b_first_retry_time))
1185  		return true;
1186  
1187  	/* At unmount we may treat errors differently */
1188  	if (xfs_is_unmounting(mp) && mp->m_fail_unmount)
1189  		return true;
1190  
1191  	return false;
1192  }
1193  
1194  /*
1195   * On a sync write or shutdown we just want to stale the buffer and let the
1196   * caller handle the error in bp->b_error appropriately.
1197   *
1198   * If the write was asynchronous then no one will be looking for the error.  If
1199   * this is the first failure of this type, clear the error state and write the
1200   * buffer out again. This means we always retry an async write failure at least
1201   * once, but we also need to set the buffer up to behave correctly now for
1202   * repeated failures.
1203   *
1204   * If we get repeated async write failures, then we take action according to the
1205   * error configuration we have been set up to use.
1206   *
1207   * Returns true if this function took care of error handling and the caller must
1208   * not touch the buffer again.  Return false if the caller should proceed with
1209   * normal I/O completion handling.
1210   */
1211  static bool
1212  xfs_buf_ioend_handle_error(
1213  	struct xfs_buf		*bp)
1214  {
1215  	struct xfs_mount	*mp = bp->b_mount;
1216  	struct xfs_error_cfg	*cfg;
1217  
1218  	/*
1219  	 * If we've already shutdown the journal because of I/O errors, there's
1220  	 * no point in giving this a retry.
1221  	 */
1222  	if (xlog_is_shutdown(mp->m_log))
1223  		goto out_stale;
1224  
1225  	xfs_buf_ioerror_alert_ratelimited(bp);
1226  
1227  	/*
1228  	 * We're not going to bother about retrying this during recovery.
1229  	 * One strike!
1230  	 */
1231  	if (bp->b_flags & _XBF_LOGRECOVERY) {
1232  		xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1233  		return false;
1234  	}
1235  
1236  	/*
1237  	 * Synchronous writes will have callers process the error.
1238  	 */
1239  	if (!(bp->b_flags & XBF_ASYNC))
1240  		goto out_stale;
1241  
1242  	trace_xfs_buf_iodone_async(bp, _RET_IP_);
1243  
1244  	cfg = xfs_error_get_cfg(mp, XFS_ERR_METADATA, bp->b_error);
1245  	if (bp->b_last_error != bp->b_error ||
1246  	    !(bp->b_flags & (XBF_STALE | XBF_WRITE_FAIL))) {
1247  		bp->b_last_error = bp->b_error;
1248  		if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER &&
1249  		    !bp->b_first_retry_time)
1250  			bp->b_first_retry_time = jiffies;
1251  		goto resubmit;
1252  	}
1253  
1254  	/*
1255  	 * Permanent error - we need to trigger a shutdown if we haven't already
1256  	 * to indicate that inconsistency will result from this action.
1257  	 */
1258  	if (xfs_buf_ioerror_permanent(bp, cfg)) {
1259  		xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1260  		goto out_stale;
1261  	}
1262  
1263  	/* Still considered a transient error. Caller will schedule retries. */
1264  	if (bp->b_flags & _XBF_INODES)
1265  		xfs_buf_inode_io_fail(bp);
1266  	else if (bp->b_flags & _XBF_DQUOTS)
1267  		xfs_buf_dquot_io_fail(bp);
1268  	else
1269  		ASSERT(list_empty(&bp->b_li_list));
1270  	xfs_buf_ioerror(bp, 0);
1271  	xfs_buf_relse(bp);
1272  	return true;
1273  
1274  resubmit:
1275  	xfs_buf_ioerror(bp, 0);
1276  	bp->b_flags |= (XBF_DONE | XBF_WRITE_FAIL);
1277  	xfs_buf_submit(bp);
1278  	return true;
1279  out_stale:
1280  	xfs_buf_stale(bp);
1281  	bp->b_flags |= XBF_DONE;
1282  	bp->b_flags &= ~XBF_WRITE;
1283  	trace_xfs_buf_error_relse(bp, _RET_IP_);
1284  	return false;
1285  }
1286  
1287  static void
1288  xfs_buf_ioend(
1289  	struct xfs_buf	*bp)
1290  {
1291  	trace_xfs_buf_iodone(bp, _RET_IP_);
1292  
1293  	/*
1294  	 * Pull in IO completion errors now. We are guaranteed to be running
1295  	 * single threaded, so we don't need the lock to read b_io_error.
1296  	 */
1297  	if (!bp->b_error && bp->b_io_error)
1298  		xfs_buf_ioerror(bp, bp->b_io_error);
1299  
1300  	if (bp->b_flags & XBF_READ) {
1301  		if (!bp->b_error && bp->b_ops)
1302  			bp->b_ops->verify_read(bp);
1303  		if (!bp->b_error)
1304  			bp->b_flags |= XBF_DONE;
1305  	} else {
1306  		if (!bp->b_error) {
1307  			bp->b_flags &= ~XBF_WRITE_FAIL;
1308  			bp->b_flags |= XBF_DONE;
1309  		}
1310  
1311  		if (unlikely(bp->b_error) && xfs_buf_ioend_handle_error(bp))
1312  			return;
1313  
1314  		/* clear the retry state */
1315  		bp->b_last_error = 0;
1316  		bp->b_retries = 0;
1317  		bp->b_first_retry_time = 0;
1318  
1319  		/*
1320  		 * Note that for things like remote attribute buffers, there may
1321  		 * not be a buffer log item here, so processing the buffer log
1322  		 * item must remain optional.
1323  		 */
1324  		if (bp->b_log_item)
1325  			xfs_buf_item_done(bp);
1326  
1327  		if (bp->b_flags & _XBF_INODES)
1328  			xfs_buf_inode_iodone(bp);
1329  		else if (bp->b_flags & _XBF_DQUOTS)
1330  			xfs_buf_dquot_iodone(bp);
1331  
1332  	}
1333  
1334  	bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD |
1335  			 _XBF_LOGRECOVERY);
1336  
1337  	if (bp->b_flags & XBF_ASYNC)
1338  		xfs_buf_relse(bp);
1339  	else
1340  		complete(&bp->b_iowait);
1341  }
1342  
1343  static void
1344  xfs_buf_ioend_work(
1345  	struct work_struct	*work)
1346  {
1347  	struct xfs_buf		*bp =
1348  		container_of(work, struct xfs_buf, b_ioend_work);
1349  
1350  	xfs_buf_ioend(bp);
1351  }
1352  
1353  static void
1354  xfs_buf_ioend_async(
1355  	struct xfs_buf	*bp)
1356  {
1357  	INIT_WORK(&bp->b_ioend_work, xfs_buf_ioend_work);
1358  	queue_work(bp->b_mount->m_buf_workqueue, &bp->b_ioend_work);
1359  }
1360  
1361  void
1362  __xfs_buf_ioerror(
1363  	struct xfs_buf		*bp,
1364  	int			error,
1365  	xfs_failaddr_t		failaddr)
1366  {
1367  	ASSERT(error <= 0 && error >= -1000);
1368  	bp->b_error = error;
1369  	trace_xfs_buf_ioerror(bp, error, failaddr);
1370  }
1371  
1372  void
1373  xfs_buf_ioerror_alert(
1374  	struct xfs_buf		*bp,
1375  	xfs_failaddr_t		func)
1376  {
1377  	xfs_buf_alert_ratelimited(bp, "XFS: metadata IO error",
1378  		"metadata I/O error in \"%pS\" at daddr 0x%llx len %d error %d",
1379  				  func, (uint64_t)xfs_buf_daddr(bp),
1380  				  bp->b_length, -bp->b_error);
1381  }
1382  
1383  /*
1384   * To simulate an I/O failure, the buffer must be locked and held with at least
1385   * three references. The LRU reference is dropped by the stale call. The buf
1386   * item reference is dropped via ioend processing. The third reference is owned
1387   * by the caller and is dropped on I/O completion if the buffer is XBF_ASYNC.
1388   */
1389  void
1390  xfs_buf_ioend_fail(
1391  	struct xfs_buf	*bp)
1392  {
1393  	bp->b_flags &= ~XBF_DONE;
1394  	xfs_buf_stale(bp);
1395  	xfs_buf_ioerror(bp, -EIO);
1396  	xfs_buf_ioend(bp);
1397  }
1398  
1399  int
1400  xfs_bwrite(
1401  	struct xfs_buf		*bp)
1402  {
1403  	int			error;
1404  
1405  	ASSERT(xfs_buf_islocked(bp));
1406  
1407  	bp->b_flags |= XBF_WRITE;
1408  	bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q |
1409  			 XBF_DONE);
1410  
1411  	error = xfs_buf_submit(bp);
1412  	if (error)
1413  		xfs_force_shutdown(bp->b_mount, SHUTDOWN_META_IO_ERROR);
1414  	return error;
1415  }
1416  
1417  static void
1418  xfs_buf_bio_end_io(
1419  	struct bio		*bio)
1420  {
1421  	struct xfs_buf		*bp = (struct xfs_buf *)bio->bi_private;
1422  
1423  	if (!bio->bi_status &&
1424  	    (bp->b_flags & XBF_WRITE) && (bp->b_flags & XBF_ASYNC) &&
1425  	    XFS_TEST_ERROR(false, bp->b_mount, XFS_ERRTAG_BUF_IOERROR))
1426  		bio->bi_status = BLK_STS_IOERR;
1427  
1428  	/*
1429  	 * don't overwrite existing errors - otherwise we can lose errors on
1430  	 * buffers that require multiple bios to complete.
1431  	 */
1432  	if (bio->bi_status) {
1433  		int error = blk_status_to_errno(bio->bi_status);
1434  
1435  		cmpxchg(&bp->b_io_error, 0, error);
1436  	}
1437  
1438  	if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1439  		invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1440  
1441  	if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1442  		xfs_buf_ioend_async(bp);
1443  	bio_put(bio);
1444  }
1445  
1446  static void
1447  xfs_buf_ioapply_map(
1448  	struct xfs_buf	*bp,
1449  	int		map,
1450  	int		*buf_offset,
1451  	int		*count,
1452  	blk_opf_t	op)
1453  {
1454  	int		page_index;
1455  	unsigned int	total_nr_pages = bp->b_page_count;
1456  	int		nr_pages;
1457  	struct bio	*bio;
1458  	sector_t	sector =  bp->b_maps[map].bm_bn;
1459  	int		size;
1460  	int		offset;
1461  
1462  	/* skip the pages in the buffer before the start offset */
1463  	page_index = 0;
1464  	offset = *buf_offset;
1465  	while (offset >= PAGE_SIZE) {
1466  		page_index++;
1467  		offset -= PAGE_SIZE;
1468  	}
1469  
1470  	/*
1471  	 * Limit the IO size to the length of the current vector, and update the
1472  	 * remaining IO count for the next time around.
1473  	 */
1474  	size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1475  	*count -= size;
1476  	*buf_offset += size;
1477  
1478  next_chunk:
1479  	atomic_inc(&bp->b_io_remaining);
1480  	nr_pages = bio_max_segs(total_nr_pages);
1481  
1482  	bio = bio_alloc(bp->b_target->bt_bdev, nr_pages, op, GFP_NOIO);
1483  	bio->bi_iter.bi_sector = sector;
1484  	bio->bi_end_io = xfs_buf_bio_end_io;
1485  	bio->bi_private = bp;
1486  
1487  	for (; size && nr_pages; nr_pages--, page_index++) {
1488  		int	rbytes, nbytes = PAGE_SIZE - offset;
1489  
1490  		if (nbytes > size)
1491  			nbytes = size;
1492  
1493  		rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1494  				      offset);
1495  		if (rbytes < nbytes)
1496  			break;
1497  
1498  		offset = 0;
1499  		sector += BTOBB(nbytes);
1500  		size -= nbytes;
1501  		total_nr_pages--;
1502  	}
1503  
1504  	if (likely(bio->bi_iter.bi_size)) {
1505  		if (xfs_buf_is_vmapped(bp)) {
1506  			flush_kernel_vmap_range(bp->b_addr,
1507  						xfs_buf_vmap_len(bp));
1508  		}
1509  		submit_bio(bio);
1510  		if (size)
1511  			goto next_chunk;
1512  	} else {
1513  		/*
1514  		 * This is guaranteed not to be the last io reference count
1515  		 * because the caller (xfs_buf_submit) holds a count itself.
1516  		 */
1517  		atomic_dec(&bp->b_io_remaining);
1518  		xfs_buf_ioerror(bp, -EIO);
1519  		bio_put(bio);
1520  	}
1521  
1522  }
1523  
1524  STATIC void
1525  _xfs_buf_ioapply(
1526  	struct xfs_buf	*bp)
1527  {
1528  	struct blk_plug	plug;
1529  	blk_opf_t	op;
1530  	int		offset;
1531  	int		size;
1532  	int		i;
1533  
1534  	/*
1535  	 * Make sure we capture only current IO errors rather than stale errors
1536  	 * left over from previous use of the buffer (e.g. failed readahead).
1537  	 */
1538  	bp->b_error = 0;
1539  
1540  	if (bp->b_flags & XBF_WRITE) {
1541  		op = REQ_OP_WRITE;
1542  
1543  		/*
1544  		 * Run the write verifier callback function if it exists. If
1545  		 * this function fails it will mark the buffer with an error and
1546  		 * the IO should not be dispatched.
1547  		 */
1548  		if (bp->b_ops) {
1549  			bp->b_ops->verify_write(bp);
1550  			if (bp->b_error) {
1551  				xfs_force_shutdown(bp->b_mount,
1552  						   SHUTDOWN_CORRUPT_INCORE);
1553  				return;
1554  			}
1555  		} else if (bp->b_rhash_key != XFS_BUF_DADDR_NULL) {
1556  			struct xfs_mount *mp = bp->b_mount;
1557  
1558  			/*
1559  			 * non-crc filesystems don't attach verifiers during
1560  			 * log recovery, so don't warn for such filesystems.
1561  			 */
1562  			if (xfs_has_crc(mp)) {
1563  				xfs_warn(mp,
1564  					"%s: no buf ops on daddr 0x%llx len %d",
1565  					__func__, xfs_buf_daddr(bp),
1566  					bp->b_length);
1567  				xfs_hex_dump(bp->b_addr,
1568  						XFS_CORRUPTION_DUMP_LEN);
1569  				dump_stack();
1570  			}
1571  		}
1572  	} else {
1573  		op = REQ_OP_READ;
1574  		if (bp->b_flags & XBF_READ_AHEAD)
1575  			op |= REQ_RAHEAD;
1576  	}
1577  
1578  	/* we only use the buffer cache for meta-data */
1579  	op |= REQ_META;
1580  
1581  	/*
1582  	 * Walk all the vectors issuing IO on them. Set up the initial offset
1583  	 * into the buffer and the desired IO size before we start -
1584  	 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1585  	 * subsequent call.
1586  	 */
1587  	offset = bp->b_offset;
1588  	size = BBTOB(bp->b_length);
1589  	blk_start_plug(&plug);
1590  	for (i = 0; i < bp->b_map_count; i++) {
1591  		xfs_buf_ioapply_map(bp, i, &offset, &size, op);
1592  		if (bp->b_error)
1593  			break;
1594  		if (size <= 0)
1595  			break;	/* all done */
1596  	}
1597  	blk_finish_plug(&plug);
1598  }
1599  
1600  /*
1601   * Wait for I/O completion of a sync buffer and return the I/O error code.
1602   */
1603  static int
1604  xfs_buf_iowait(
1605  	struct xfs_buf	*bp)
1606  {
1607  	ASSERT(!(bp->b_flags & XBF_ASYNC));
1608  
1609  	trace_xfs_buf_iowait(bp, _RET_IP_);
1610  	wait_for_completion(&bp->b_iowait);
1611  	trace_xfs_buf_iowait_done(bp, _RET_IP_);
1612  
1613  	return bp->b_error;
1614  }
1615  
1616  /*
1617   * Buffer I/O submission path, read or write. Asynchronous submission transfers
1618   * the buffer lock ownership and the current reference to the IO. It is not
1619   * safe to reference the buffer after a call to this function unless the caller
1620   * holds an additional reference itself.
1621   */
1622  static int
1623  __xfs_buf_submit(
1624  	struct xfs_buf	*bp,
1625  	bool		wait)
1626  {
1627  	int		error = 0;
1628  
1629  	trace_xfs_buf_submit(bp, _RET_IP_);
1630  
1631  	ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1632  
1633  	/*
1634  	 * On log shutdown we stale and complete the buffer immediately. We can
1635  	 * be called to read the superblock before the log has been set up, so
1636  	 * be careful checking the log state.
1637  	 *
1638  	 * Checking the mount shutdown state here can result in the log tail
1639  	 * moving inappropriately on disk as the log may not yet be shut down.
1640  	 * i.e. failing this buffer on mount shutdown can remove it from the AIL
1641  	 * and move the tail of the log forwards without having written this
1642  	 * buffer to disk. This corrupts the log tail state in memory, and
1643  	 * because the log may not be shut down yet, it can then be propagated
1644  	 * to disk before the log is shutdown. Hence we check log shutdown
1645  	 * state here rather than mount state to avoid corrupting the log tail
1646  	 * on shutdown.
1647  	 */
1648  	if (bp->b_mount->m_log &&
1649  	    xlog_is_shutdown(bp->b_mount->m_log)) {
1650  		xfs_buf_ioend_fail(bp);
1651  		return -EIO;
1652  	}
1653  
1654  	/*
1655  	 * Grab a reference so the buffer does not go away underneath us. For
1656  	 * async buffers, I/O completion drops the callers reference, which
1657  	 * could occur before submission returns.
1658  	 */
1659  	xfs_buf_hold(bp);
1660  
1661  	if (bp->b_flags & XBF_WRITE)
1662  		xfs_buf_wait_unpin(bp);
1663  
1664  	/* clear the internal error state to avoid spurious errors */
1665  	bp->b_io_error = 0;
1666  
1667  	/*
1668  	 * Set the count to 1 initially, this will stop an I/O completion
1669  	 * callout which happens before we have started all the I/O from calling
1670  	 * xfs_buf_ioend too early.
1671  	 */
1672  	atomic_set(&bp->b_io_remaining, 1);
1673  	if (bp->b_flags & XBF_ASYNC)
1674  		xfs_buf_ioacct_inc(bp);
1675  	_xfs_buf_ioapply(bp);
1676  
1677  	/*
1678  	 * If _xfs_buf_ioapply failed, we can get back here with only the IO
1679  	 * reference we took above. If we drop it to zero, run completion so
1680  	 * that we don't return to the caller with completion still pending.
1681  	 */
1682  	if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1683  		if (bp->b_error || !(bp->b_flags & XBF_ASYNC))
1684  			xfs_buf_ioend(bp);
1685  		else
1686  			xfs_buf_ioend_async(bp);
1687  	}
1688  
1689  	if (wait)
1690  		error = xfs_buf_iowait(bp);
1691  
1692  	/*
1693  	 * Release the hold that keeps the buffer referenced for the entire
1694  	 * I/O. Note that if the buffer is async, it is not safe to reference
1695  	 * after this release.
1696  	 */
1697  	xfs_buf_rele(bp);
1698  	return error;
1699  }
1700  
1701  void *
1702  xfs_buf_offset(
1703  	struct xfs_buf		*bp,
1704  	size_t			offset)
1705  {
1706  	struct page		*page;
1707  
1708  	if (bp->b_addr)
1709  		return bp->b_addr + offset;
1710  
1711  	page = bp->b_pages[offset >> PAGE_SHIFT];
1712  	return page_address(page) + (offset & (PAGE_SIZE-1));
1713  }
1714  
1715  void
1716  xfs_buf_zero(
1717  	struct xfs_buf		*bp,
1718  	size_t			boff,
1719  	size_t			bsize)
1720  {
1721  	size_t			bend;
1722  
1723  	bend = boff + bsize;
1724  	while (boff < bend) {
1725  		struct page	*page;
1726  		int		page_index, page_offset, csize;
1727  
1728  		page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1729  		page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1730  		page = bp->b_pages[page_index];
1731  		csize = min_t(size_t, PAGE_SIZE - page_offset,
1732  				      BBTOB(bp->b_length) - boff);
1733  
1734  		ASSERT((csize + page_offset) <= PAGE_SIZE);
1735  
1736  		memset(page_address(page) + page_offset, 0, csize);
1737  
1738  		boff += csize;
1739  	}
1740  }
1741  
1742  /*
1743   * Log a message about and stale a buffer that a caller has decided is corrupt.
1744   *
1745   * This function should be called for the kinds of metadata corruption that
1746   * cannot be detect from a verifier, such as incorrect inter-block relationship
1747   * data.  Do /not/ call this function from a verifier function.
1748   *
1749   * The buffer must be XBF_DONE prior to the call.  Afterwards, the buffer will
1750   * be marked stale, but b_error will not be set.  The caller is responsible for
1751   * releasing the buffer or fixing it.
1752   */
1753  void
1754  __xfs_buf_mark_corrupt(
1755  	struct xfs_buf		*bp,
1756  	xfs_failaddr_t		fa)
1757  {
1758  	ASSERT(bp->b_flags & XBF_DONE);
1759  
1760  	xfs_buf_corruption_error(bp, fa);
1761  	xfs_buf_stale(bp);
1762  }
1763  
1764  /*
1765   *	Handling of buffer targets (buftargs).
1766   */
1767  
1768  /*
1769   * Wait for any bufs with callbacks that have been submitted but have not yet
1770   * returned. These buffers will have an elevated hold count, so wait on those
1771   * while freeing all the buffers only held by the LRU.
1772   */
1773  static enum lru_status
1774  xfs_buftarg_drain_rele(
1775  	struct list_head	*item,
1776  	struct list_lru_one	*lru,
1777  	spinlock_t		*lru_lock,
1778  	void			*arg)
1779  
1780  {
1781  	struct xfs_buf		*bp = container_of(item, struct xfs_buf, b_lru);
1782  	struct list_head	*dispose = arg;
1783  
1784  	if (atomic_read(&bp->b_hold) > 1) {
1785  		/* need to wait, so skip it this pass */
1786  		trace_xfs_buf_drain_buftarg(bp, _RET_IP_);
1787  		return LRU_SKIP;
1788  	}
1789  	if (!spin_trylock(&bp->b_lock))
1790  		return LRU_SKIP;
1791  
1792  	/*
1793  	 * clear the LRU reference count so the buffer doesn't get
1794  	 * ignored in xfs_buf_rele().
1795  	 */
1796  	atomic_set(&bp->b_lru_ref, 0);
1797  	bp->b_state |= XFS_BSTATE_DISPOSE;
1798  	list_lru_isolate_move(lru, item, dispose);
1799  	spin_unlock(&bp->b_lock);
1800  	return LRU_REMOVED;
1801  }
1802  
1803  /*
1804   * Wait for outstanding I/O on the buftarg to complete.
1805   */
1806  void
1807  xfs_buftarg_wait(
1808  	struct xfs_buftarg	*btp)
1809  {
1810  	/*
1811  	 * First wait on the buftarg I/O count for all in-flight buffers to be
1812  	 * released. This is critical as new buffers do not make the LRU until
1813  	 * they are released.
1814  	 *
1815  	 * Next, flush the buffer workqueue to ensure all completion processing
1816  	 * has finished. Just waiting on buffer locks is not sufficient for
1817  	 * async IO as the reference count held over IO is not released until
1818  	 * after the buffer lock is dropped. Hence we need to ensure here that
1819  	 * all reference counts have been dropped before we start walking the
1820  	 * LRU list.
1821  	 */
1822  	while (percpu_counter_sum(&btp->bt_io_count))
1823  		delay(100);
1824  	flush_workqueue(btp->bt_mount->m_buf_workqueue);
1825  }
1826  
1827  void
1828  xfs_buftarg_drain(
1829  	struct xfs_buftarg	*btp)
1830  {
1831  	LIST_HEAD(dispose);
1832  	int			loop = 0;
1833  	bool			write_fail = false;
1834  
1835  	xfs_buftarg_wait(btp);
1836  
1837  	/* loop until there is nothing left on the lru list. */
1838  	while (list_lru_count(&btp->bt_lru)) {
1839  		list_lru_walk(&btp->bt_lru, xfs_buftarg_drain_rele,
1840  			      &dispose, LONG_MAX);
1841  
1842  		while (!list_empty(&dispose)) {
1843  			struct xfs_buf *bp;
1844  			bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1845  			list_del_init(&bp->b_lru);
1846  			if (bp->b_flags & XBF_WRITE_FAIL) {
1847  				write_fail = true;
1848  				xfs_buf_alert_ratelimited(bp,
1849  					"XFS: Corruption Alert",
1850  "Corruption Alert: Buffer at daddr 0x%llx had permanent write failures!",
1851  					(long long)xfs_buf_daddr(bp));
1852  			}
1853  			xfs_buf_rele(bp);
1854  		}
1855  		if (loop++ != 0)
1856  			delay(100);
1857  	}
1858  
1859  	/*
1860  	 * If one or more failed buffers were freed, that means dirty metadata
1861  	 * was thrown away. This should only ever happen after I/O completion
1862  	 * handling has elevated I/O error(s) to permanent failures and shuts
1863  	 * down the journal.
1864  	 */
1865  	if (write_fail) {
1866  		ASSERT(xlog_is_shutdown(btp->bt_mount->m_log));
1867  		xfs_alert(btp->bt_mount,
1868  	      "Please run xfs_repair to determine the extent of the problem.");
1869  	}
1870  }
1871  
1872  static enum lru_status
1873  xfs_buftarg_isolate(
1874  	struct list_head	*item,
1875  	struct list_lru_one	*lru,
1876  	spinlock_t		*lru_lock,
1877  	void			*arg)
1878  {
1879  	struct xfs_buf		*bp = container_of(item, struct xfs_buf, b_lru);
1880  	struct list_head	*dispose = arg;
1881  
1882  	/*
1883  	 * we are inverting the lru lock/bp->b_lock here, so use a trylock.
1884  	 * If we fail to get the lock, just skip it.
1885  	 */
1886  	if (!spin_trylock(&bp->b_lock))
1887  		return LRU_SKIP;
1888  	/*
1889  	 * Decrement the b_lru_ref count unless the value is already
1890  	 * zero. If the value is already zero, we need to reclaim the
1891  	 * buffer, otherwise it gets another trip through the LRU.
1892  	 */
1893  	if (atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1894  		spin_unlock(&bp->b_lock);
1895  		return LRU_ROTATE;
1896  	}
1897  
1898  	bp->b_state |= XFS_BSTATE_DISPOSE;
1899  	list_lru_isolate_move(lru, item, dispose);
1900  	spin_unlock(&bp->b_lock);
1901  	return LRU_REMOVED;
1902  }
1903  
1904  static unsigned long
1905  xfs_buftarg_shrink_scan(
1906  	struct shrinker		*shrink,
1907  	struct shrink_control	*sc)
1908  {
1909  	struct xfs_buftarg	*btp = container_of(shrink,
1910  					struct xfs_buftarg, bt_shrinker);
1911  	LIST_HEAD(dispose);
1912  	unsigned long		freed;
1913  
1914  	freed = list_lru_shrink_walk(&btp->bt_lru, sc,
1915  				     xfs_buftarg_isolate, &dispose);
1916  
1917  	while (!list_empty(&dispose)) {
1918  		struct xfs_buf *bp;
1919  		bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1920  		list_del_init(&bp->b_lru);
1921  		xfs_buf_rele(bp);
1922  	}
1923  
1924  	return freed;
1925  }
1926  
1927  static unsigned long
1928  xfs_buftarg_shrink_count(
1929  	struct shrinker		*shrink,
1930  	struct shrink_control	*sc)
1931  {
1932  	struct xfs_buftarg	*btp = container_of(shrink,
1933  					struct xfs_buftarg, bt_shrinker);
1934  	return list_lru_shrink_count(&btp->bt_lru, sc);
1935  }
1936  
1937  void
1938  xfs_free_buftarg(
1939  	struct xfs_buftarg	*btp)
1940  {
1941  	unregister_shrinker(&btp->bt_shrinker);
1942  	ASSERT(percpu_counter_sum(&btp->bt_io_count) == 0);
1943  	percpu_counter_destroy(&btp->bt_io_count);
1944  	list_lru_destroy(&btp->bt_lru);
1945  
1946  	blkdev_issue_flush(btp->bt_bdev);
1947  	invalidate_bdev(btp->bt_bdev);
1948  	fs_put_dax(btp->bt_daxdev, btp->bt_mount);
1949  
1950  	kmem_free(btp);
1951  }
1952  
1953  int
1954  xfs_setsize_buftarg(
1955  	xfs_buftarg_t		*btp,
1956  	unsigned int		sectorsize)
1957  {
1958  	/* Set up metadata sector size info */
1959  	btp->bt_meta_sectorsize = sectorsize;
1960  	btp->bt_meta_sectormask = sectorsize - 1;
1961  
1962  	if (set_blocksize(btp->bt_bdev, sectorsize)) {
1963  		xfs_warn(btp->bt_mount,
1964  			"Cannot set_blocksize to %u on device %pg",
1965  			sectorsize, btp->bt_bdev);
1966  		return -EINVAL;
1967  	}
1968  
1969  	/* Set up device logical sector size mask */
1970  	btp->bt_logical_sectorsize = bdev_logical_block_size(btp->bt_bdev);
1971  	btp->bt_logical_sectormask = bdev_logical_block_size(btp->bt_bdev) - 1;
1972  
1973  	return 0;
1974  }
1975  
1976  /*
1977   * When allocating the initial buffer target we have not yet
1978   * read in the superblock, so don't know what sized sectors
1979   * are being used at this early stage.  Play safe.
1980   */
1981  STATIC int
1982  xfs_setsize_buftarg_early(
1983  	xfs_buftarg_t		*btp,
1984  	struct block_device	*bdev)
1985  {
1986  	return xfs_setsize_buftarg(btp, bdev_logical_block_size(bdev));
1987  }
1988  
1989  struct xfs_buftarg *
1990  xfs_alloc_buftarg(
1991  	struct xfs_mount	*mp,
1992  	struct block_device	*bdev)
1993  {
1994  	xfs_buftarg_t		*btp;
1995  	const struct dax_holder_operations *ops = NULL;
1996  
1997  #if defined(CONFIG_FS_DAX) && defined(CONFIG_MEMORY_FAILURE)
1998  	ops = &xfs_dax_holder_operations;
1999  #endif
2000  	btp = kmem_zalloc(sizeof(*btp), KM_NOFS);
2001  
2002  	btp->bt_mount = mp;
2003  	btp->bt_dev =  bdev->bd_dev;
2004  	btp->bt_bdev = bdev;
2005  	btp->bt_daxdev = fs_dax_get_by_bdev(bdev, &btp->bt_dax_part_off,
2006  					    mp, ops);
2007  
2008  	/*
2009  	 * Buffer IO error rate limiting. Limit it to no more than 10 messages
2010  	 * per 30 seconds so as to not spam logs too much on repeated errors.
2011  	 */
2012  	ratelimit_state_init(&btp->bt_ioerror_rl, 30 * HZ,
2013  			     DEFAULT_RATELIMIT_BURST);
2014  
2015  	if (xfs_setsize_buftarg_early(btp, bdev))
2016  		goto error_free;
2017  
2018  	if (list_lru_init(&btp->bt_lru))
2019  		goto error_free;
2020  
2021  	if (percpu_counter_init(&btp->bt_io_count, 0, GFP_KERNEL))
2022  		goto error_lru;
2023  
2024  	btp->bt_shrinker.count_objects = xfs_buftarg_shrink_count;
2025  	btp->bt_shrinker.scan_objects = xfs_buftarg_shrink_scan;
2026  	btp->bt_shrinker.seeks = DEFAULT_SEEKS;
2027  	btp->bt_shrinker.flags = SHRINKER_NUMA_AWARE;
2028  	if (register_shrinker(&btp->bt_shrinker, "xfs-buf:%s",
2029  			      mp->m_super->s_id))
2030  		goto error_pcpu;
2031  	return btp;
2032  
2033  error_pcpu:
2034  	percpu_counter_destroy(&btp->bt_io_count);
2035  error_lru:
2036  	list_lru_destroy(&btp->bt_lru);
2037  error_free:
2038  	kmem_free(btp);
2039  	return NULL;
2040  }
2041  
2042  /*
2043   * Cancel a delayed write list.
2044   *
2045   * Remove each buffer from the list, clear the delwri queue flag and drop the
2046   * associated buffer reference.
2047   */
2048  void
2049  xfs_buf_delwri_cancel(
2050  	struct list_head	*list)
2051  {
2052  	struct xfs_buf		*bp;
2053  
2054  	while (!list_empty(list)) {
2055  		bp = list_first_entry(list, struct xfs_buf, b_list);
2056  
2057  		xfs_buf_lock(bp);
2058  		bp->b_flags &= ~_XBF_DELWRI_Q;
2059  		list_del_init(&bp->b_list);
2060  		xfs_buf_relse(bp);
2061  	}
2062  }
2063  
2064  /*
2065   * Add a buffer to the delayed write list.
2066   *
2067   * This queues a buffer for writeout if it hasn't already been.  Note that
2068   * neither this routine nor the buffer list submission functions perform
2069   * any internal synchronization.  It is expected that the lists are thread-local
2070   * to the callers.
2071   *
2072   * Returns true if we queued up the buffer, or false if it already had
2073   * been on the buffer list.
2074   */
2075  bool
2076  xfs_buf_delwri_queue(
2077  	struct xfs_buf		*bp,
2078  	struct list_head	*list)
2079  {
2080  	ASSERT(xfs_buf_islocked(bp));
2081  	ASSERT(!(bp->b_flags & XBF_READ));
2082  
2083  	/*
2084  	 * If the buffer is already marked delwri it already is queued up
2085  	 * by someone else for imediate writeout.  Just ignore it in that
2086  	 * case.
2087  	 */
2088  	if (bp->b_flags & _XBF_DELWRI_Q) {
2089  		trace_xfs_buf_delwri_queued(bp, _RET_IP_);
2090  		return false;
2091  	}
2092  
2093  	trace_xfs_buf_delwri_queue(bp, _RET_IP_);
2094  
2095  	/*
2096  	 * If a buffer gets written out synchronously or marked stale while it
2097  	 * is on a delwri list we lazily remove it. To do this, the other party
2098  	 * clears the  _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
2099  	 * It remains referenced and on the list.  In a rare corner case it
2100  	 * might get readded to a delwri list after the synchronous writeout, in
2101  	 * which case we need just need to re-add the flag here.
2102  	 */
2103  	bp->b_flags |= _XBF_DELWRI_Q;
2104  	if (list_empty(&bp->b_list)) {
2105  		atomic_inc(&bp->b_hold);
2106  		list_add_tail(&bp->b_list, list);
2107  	}
2108  
2109  	return true;
2110  }
2111  
2112  /*
2113   * Compare function is more complex than it needs to be because
2114   * the return value is only 32 bits and we are doing comparisons
2115   * on 64 bit values
2116   */
2117  static int
2118  xfs_buf_cmp(
2119  	void			*priv,
2120  	const struct list_head	*a,
2121  	const struct list_head	*b)
2122  {
2123  	struct xfs_buf	*ap = container_of(a, struct xfs_buf, b_list);
2124  	struct xfs_buf	*bp = container_of(b, struct xfs_buf, b_list);
2125  	xfs_daddr_t		diff;
2126  
2127  	diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn;
2128  	if (diff < 0)
2129  		return -1;
2130  	if (diff > 0)
2131  		return 1;
2132  	return 0;
2133  }
2134  
2135  /*
2136   * Submit buffers for write. If wait_list is specified, the buffers are
2137   * submitted using sync I/O and placed on the wait list such that the caller can
2138   * iowait each buffer. Otherwise async I/O is used and the buffers are released
2139   * at I/O completion time. In either case, buffers remain locked until I/O
2140   * completes and the buffer is released from the queue.
2141   */
2142  static int
2143  xfs_buf_delwri_submit_buffers(
2144  	struct list_head	*buffer_list,
2145  	struct list_head	*wait_list)
2146  {
2147  	struct xfs_buf		*bp, *n;
2148  	int			pinned = 0;
2149  	struct blk_plug		plug;
2150  
2151  	list_sort(NULL, buffer_list, xfs_buf_cmp);
2152  
2153  	blk_start_plug(&plug);
2154  	list_for_each_entry_safe(bp, n, buffer_list, b_list) {
2155  		if (!wait_list) {
2156  			if (!xfs_buf_trylock(bp))
2157  				continue;
2158  			if (xfs_buf_ispinned(bp)) {
2159  				xfs_buf_unlock(bp);
2160  				pinned++;
2161  				continue;
2162  			}
2163  		} else {
2164  			xfs_buf_lock(bp);
2165  		}
2166  
2167  		/*
2168  		 * Someone else might have written the buffer synchronously or
2169  		 * marked it stale in the meantime.  In that case only the
2170  		 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
2171  		 * reference and remove it from the list here.
2172  		 */
2173  		if (!(bp->b_flags & _XBF_DELWRI_Q)) {
2174  			list_del_init(&bp->b_list);
2175  			xfs_buf_relse(bp);
2176  			continue;
2177  		}
2178  
2179  		trace_xfs_buf_delwri_split(bp, _RET_IP_);
2180  
2181  		/*
2182  		 * If we have a wait list, each buffer (and associated delwri
2183  		 * queue reference) transfers to it and is submitted
2184  		 * synchronously. Otherwise, drop the buffer from the delwri
2185  		 * queue and submit async.
2186  		 */
2187  		bp->b_flags &= ~_XBF_DELWRI_Q;
2188  		bp->b_flags |= XBF_WRITE;
2189  		if (wait_list) {
2190  			bp->b_flags &= ~XBF_ASYNC;
2191  			list_move_tail(&bp->b_list, wait_list);
2192  		} else {
2193  			bp->b_flags |= XBF_ASYNC;
2194  			list_del_init(&bp->b_list);
2195  		}
2196  		__xfs_buf_submit(bp, false);
2197  	}
2198  	blk_finish_plug(&plug);
2199  
2200  	return pinned;
2201  }
2202  
2203  /*
2204   * Write out a buffer list asynchronously.
2205   *
2206   * This will take the @buffer_list, write all non-locked and non-pinned buffers
2207   * out and not wait for I/O completion on any of the buffers.  This interface
2208   * is only safely useable for callers that can track I/O completion by higher
2209   * level means, e.g. AIL pushing as the @buffer_list is consumed in this
2210   * function.
2211   *
2212   * Note: this function will skip buffers it would block on, and in doing so
2213   * leaves them on @buffer_list so they can be retried on a later pass. As such,
2214   * it is up to the caller to ensure that the buffer list is fully submitted or
2215   * cancelled appropriately when they are finished with the list. Failure to
2216   * cancel or resubmit the list until it is empty will result in leaked buffers
2217   * at unmount time.
2218   */
2219  int
2220  xfs_buf_delwri_submit_nowait(
2221  	struct list_head	*buffer_list)
2222  {
2223  	return xfs_buf_delwri_submit_buffers(buffer_list, NULL);
2224  }
2225  
2226  /*
2227   * Write out a buffer list synchronously.
2228   *
2229   * This will take the @buffer_list, write all buffers out and wait for I/O
2230   * completion on all of the buffers. @buffer_list is consumed by the function,
2231   * so callers must have some other way of tracking buffers if they require such
2232   * functionality.
2233   */
2234  int
2235  xfs_buf_delwri_submit(
2236  	struct list_head	*buffer_list)
2237  {
2238  	LIST_HEAD		(wait_list);
2239  	int			error = 0, error2;
2240  	struct xfs_buf		*bp;
2241  
2242  	xfs_buf_delwri_submit_buffers(buffer_list, &wait_list);
2243  
2244  	/* Wait for IO to complete. */
2245  	while (!list_empty(&wait_list)) {
2246  		bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
2247  
2248  		list_del_init(&bp->b_list);
2249  
2250  		/*
2251  		 * Wait on the locked buffer, check for errors and unlock and
2252  		 * release the delwri queue reference.
2253  		 */
2254  		error2 = xfs_buf_iowait(bp);
2255  		xfs_buf_relse(bp);
2256  		if (!error)
2257  			error = error2;
2258  	}
2259  
2260  	return error;
2261  }
2262  
2263  /*
2264   * Push a single buffer on a delwri queue.
2265   *
2266   * The purpose of this function is to submit a single buffer of a delwri queue
2267   * and return with the buffer still on the original queue. The waiting delwri
2268   * buffer submission infrastructure guarantees transfer of the delwri queue
2269   * buffer reference to a temporary wait list. We reuse this infrastructure to
2270   * transfer the buffer back to the original queue.
2271   *
2272   * Note the buffer transitions from the queued state, to the submitted and wait
2273   * listed state and back to the queued state during this call. The buffer
2274   * locking and queue management logic between _delwri_pushbuf() and
2275   * _delwri_queue() guarantee that the buffer cannot be queued to another list
2276   * before returning.
2277   */
2278  int
2279  xfs_buf_delwri_pushbuf(
2280  	struct xfs_buf		*bp,
2281  	struct list_head	*buffer_list)
2282  {
2283  	LIST_HEAD		(submit_list);
2284  	int			error;
2285  
2286  	ASSERT(bp->b_flags & _XBF_DELWRI_Q);
2287  
2288  	trace_xfs_buf_delwri_pushbuf(bp, _RET_IP_);
2289  
2290  	/*
2291  	 * Isolate the buffer to a new local list so we can submit it for I/O
2292  	 * independently from the rest of the original list.
2293  	 */
2294  	xfs_buf_lock(bp);
2295  	list_move(&bp->b_list, &submit_list);
2296  	xfs_buf_unlock(bp);
2297  
2298  	/*
2299  	 * Delwri submission clears the DELWRI_Q buffer flag and returns with
2300  	 * the buffer on the wait list with the original reference. Rather than
2301  	 * bounce the buffer from a local wait list back to the original list
2302  	 * after I/O completion, reuse the original list as the wait list.
2303  	 */
2304  	xfs_buf_delwri_submit_buffers(&submit_list, buffer_list);
2305  
2306  	/*
2307  	 * The buffer is now locked, under I/O and wait listed on the original
2308  	 * delwri queue. Wait for I/O completion, restore the DELWRI_Q flag and
2309  	 * return with the buffer unlocked and on the original queue.
2310  	 */
2311  	error = xfs_buf_iowait(bp);
2312  	bp->b_flags |= _XBF_DELWRI_Q;
2313  	xfs_buf_unlock(bp);
2314  
2315  	return error;
2316  }
2317  
2318  void xfs_buf_set_ref(struct xfs_buf *bp, int lru_ref)
2319  {
2320  	/*
2321  	 * Set the lru reference count to 0 based on the error injection tag.
2322  	 * This allows userspace to disrupt buffer caching for debug/testing
2323  	 * purposes.
2324  	 */
2325  	if (XFS_TEST_ERROR(false, bp->b_mount, XFS_ERRTAG_BUF_LRU_REF))
2326  		lru_ref = 0;
2327  
2328  	atomic_set(&bp->b_lru_ref, lru_ref);
2329  }
2330  
2331  /*
2332   * Verify an on-disk magic value against the magic value specified in the
2333   * verifier structure. The verifier magic is in disk byte order so the caller is
2334   * expected to pass the value directly from disk.
2335   */
2336  bool
2337  xfs_verify_magic(
2338  	struct xfs_buf		*bp,
2339  	__be32			dmagic)
2340  {
2341  	struct xfs_mount	*mp = bp->b_mount;
2342  	int			idx;
2343  
2344  	idx = xfs_has_crc(mp);
2345  	if (WARN_ON(!bp->b_ops || !bp->b_ops->magic[idx]))
2346  		return false;
2347  	return dmagic == bp->b_ops->magic[idx];
2348  }
2349  /*
2350   * Verify an on-disk magic value against the magic value specified in the
2351   * verifier structure. The verifier magic is in disk byte order so the caller is
2352   * expected to pass the value directly from disk.
2353   */
2354  bool
2355  xfs_verify_magic16(
2356  	struct xfs_buf		*bp,
2357  	__be16			dmagic)
2358  {
2359  	struct xfs_mount	*mp = bp->b_mount;
2360  	int			idx;
2361  
2362  	idx = xfs_has_crc(mp);
2363  	if (WARN_ON(!bp->b_ops || !bp->b_ops->magic16[idx]))
2364  		return false;
2365  	return dmagic == bp->b_ops->magic16[idx];
2366  }
2367