xref: /openbmc/linux/fs/xfs/xfs_iomap.c (revision 0e17c50f)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
4  * Copyright (c) 2016-2018 Christoph Hellwig.
5  * All Rights Reserved.
6  */
7 #include "xfs.h"
8 #include "xfs_fs.h"
9 #include "xfs_shared.h"
10 #include "xfs_format.h"
11 #include "xfs_log_format.h"
12 #include "xfs_trans_resv.h"
13 #include "xfs_mount.h"
14 #include "xfs_inode.h"
15 #include "xfs_btree.h"
16 #include "xfs_bmap_btree.h"
17 #include "xfs_bmap.h"
18 #include "xfs_bmap_util.h"
19 #include "xfs_errortag.h"
20 #include "xfs_error.h"
21 #include "xfs_trans.h"
22 #include "xfs_trans_space.h"
23 #include "xfs_inode_item.h"
24 #include "xfs_iomap.h"
25 #include "xfs_trace.h"
26 #include "xfs_quota.h"
27 #include "xfs_dquot_item.h"
28 #include "xfs_dquot.h"
29 #include "xfs_reflink.h"
30 
31 
32 #define XFS_ALLOC_ALIGN(mp, off) \
33 	(((off) >> mp->m_allocsize_log) << mp->m_allocsize_log)
34 
35 static int
36 xfs_alert_fsblock_zero(
37 	xfs_inode_t	*ip,
38 	xfs_bmbt_irec_t	*imap)
39 {
40 	xfs_alert_tag(ip->i_mount, XFS_PTAG_FSBLOCK_ZERO,
41 			"Access to block zero in inode %llu "
42 			"start_block: %llx start_off: %llx "
43 			"blkcnt: %llx extent-state: %x",
44 		(unsigned long long)ip->i_ino,
45 		(unsigned long long)imap->br_startblock,
46 		(unsigned long long)imap->br_startoff,
47 		(unsigned long long)imap->br_blockcount,
48 		imap->br_state);
49 	return -EFSCORRUPTED;
50 }
51 
52 int
53 xfs_bmbt_to_iomap(
54 	struct xfs_inode	*ip,
55 	struct iomap		*iomap,
56 	struct xfs_bmbt_irec	*imap,
57 	u16			flags)
58 {
59 	struct xfs_mount	*mp = ip->i_mount;
60 	struct xfs_buftarg	*target = xfs_inode_buftarg(ip);
61 
62 	if (unlikely(!xfs_valid_startblock(ip, imap->br_startblock)))
63 		return xfs_alert_fsblock_zero(ip, imap);
64 
65 	if (imap->br_startblock == HOLESTARTBLOCK) {
66 		iomap->addr = IOMAP_NULL_ADDR;
67 		iomap->type = IOMAP_HOLE;
68 	} else if (imap->br_startblock == DELAYSTARTBLOCK ||
69 		   isnullstartblock(imap->br_startblock)) {
70 		iomap->addr = IOMAP_NULL_ADDR;
71 		iomap->type = IOMAP_DELALLOC;
72 	} else {
73 		iomap->addr = BBTOB(xfs_fsb_to_db(ip, imap->br_startblock));
74 		if (imap->br_state == XFS_EXT_UNWRITTEN)
75 			iomap->type = IOMAP_UNWRITTEN;
76 		else
77 			iomap->type = IOMAP_MAPPED;
78 	}
79 	iomap->offset = XFS_FSB_TO_B(mp, imap->br_startoff);
80 	iomap->length = XFS_FSB_TO_B(mp, imap->br_blockcount);
81 	iomap->bdev = target->bt_bdev;
82 	iomap->dax_dev = target->bt_daxdev;
83 	iomap->flags = flags;
84 
85 	if (xfs_ipincount(ip) &&
86 	    (ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
87 		iomap->flags |= IOMAP_F_DIRTY;
88 	return 0;
89 }
90 
91 static void
92 xfs_hole_to_iomap(
93 	struct xfs_inode	*ip,
94 	struct iomap		*iomap,
95 	xfs_fileoff_t		offset_fsb,
96 	xfs_fileoff_t		end_fsb)
97 {
98 	struct xfs_buftarg	*target = xfs_inode_buftarg(ip);
99 
100 	iomap->addr = IOMAP_NULL_ADDR;
101 	iomap->type = IOMAP_HOLE;
102 	iomap->offset = XFS_FSB_TO_B(ip->i_mount, offset_fsb);
103 	iomap->length = XFS_FSB_TO_B(ip->i_mount, end_fsb - offset_fsb);
104 	iomap->bdev = target->bt_bdev;
105 	iomap->dax_dev = target->bt_daxdev;
106 }
107 
108 static inline xfs_fileoff_t
109 xfs_iomap_end_fsb(
110 	struct xfs_mount	*mp,
111 	loff_t			offset,
112 	loff_t			count)
113 {
114 	ASSERT(offset <= mp->m_super->s_maxbytes);
115 	return min(XFS_B_TO_FSB(mp, offset + count),
116 		   XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes));
117 }
118 
119 static xfs_extlen_t
120 xfs_eof_alignment(
121 	struct xfs_inode	*ip)
122 {
123 	struct xfs_mount	*mp = ip->i_mount;
124 	xfs_extlen_t		align = 0;
125 
126 	if (!XFS_IS_REALTIME_INODE(ip)) {
127 		/*
128 		 * Round up the allocation request to a stripe unit
129 		 * (m_dalign) boundary if the file size is >= stripe unit
130 		 * size, and we are allocating past the allocation eof.
131 		 *
132 		 * If mounted with the "-o swalloc" option the alignment is
133 		 * increased from the strip unit size to the stripe width.
134 		 */
135 		if (mp->m_swidth && (mp->m_flags & XFS_MOUNT_SWALLOC))
136 			align = mp->m_swidth;
137 		else if (mp->m_dalign)
138 			align = mp->m_dalign;
139 
140 		if (align && XFS_ISIZE(ip) < XFS_FSB_TO_B(mp, align))
141 			align = 0;
142 	}
143 
144 	return align;
145 }
146 
147 /*
148  * Check if last_fsb is outside the last extent, and if so grow it to the next
149  * stripe unit boundary.
150  */
151 xfs_fileoff_t
152 xfs_iomap_eof_align_last_fsb(
153 	struct xfs_inode	*ip,
154 	xfs_fileoff_t		end_fsb)
155 {
156 	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
157 	xfs_extlen_t		extsz = xfs_get_extsz_hint(ip);
158 	xfs_extlen_t		align = xfs_eof_alignment(ip);
159 	struct xfs_bmbt_irec	irec;
160 	struct xfs_iext_cursor	icur;
161 
162 	ASSERT(ifp->if_flags & XFS_IFEXTENTS);
163 
164 	/*
165 	 * Always round up the allocation request to the extent hint boundary.
166 	 */
167 	if (extsz) {
168 		if (align)
169 			align = roundup_64(align, extsz);
170 		else
171 			align = extsz;
172 	}
173 
174 	if (align) {
175 		xfs_fileoff_t	aligned_end_fsb = roundup_64(end_fsb, align);
176 
177 		xfs_iext_last(ifp, &icur);
178 		if (!xfs_iext_get_extent(ifp, &icur, &irec) ||
179 		    aligned_end_fsb >= irec.br_startoff + irec.br_blockcount)
180 			return aligned_end_fsb;
181 	}
182 
183 	return end_fsb;
184 }
185 
186 int
187 xfs_iomap_write_direct(
188 	struct xfs_inode	*ip,
189 	xfs_fileoff_t		offset_fsb,
190 	xfs_fileoff_t		count_fsb,
191 	struct xfs_bmbt_irec	*imap)
192 {
193 	struct xfs_mount	*mp = ip->i_mount;
194 	struct xfs_trans	*tp;
195 	xfs_filblks_t		resaligned;
196 	int			nimaps;
197 	int			quota_flag;
198 	uint			qblocks, resblks;
199 	unsigned int		resrtextents = 0;
200 	int			error;
201 	int			bmapi_flags = XFS_BMAPI_PREALLOC;
202 	uint			tflags = 0;
203 
204 	ASSERT(count_fsb > 0);
205 
206 	resaligned = xfs_aligned_fsb_count(offset_fsb, count_fsb,
207 					   xfs_get_extsz_hint(ip));
208 	if (unlikely(XFS_IS_REALTIME_INODE(ip))) {
209 		resrtextents = qblocks = resaligned;
210 		resrtextents /= mp->m_sb.sb_rextsize;
211 		resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
212 		quota_flag = XFS_QMOPT_RES_RTBLKS;
213 	} else {
214 		resblks = qblocks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
215 		quota_flag = XFS_QMOPT_RES_REGBLKS;
216 	}
217 
218 	error = xfs_qm_dqattach(ip);
219 	if (error)
220 		return error;
221 
222 	/*
223 	 * For DAX, we do not allocate unwritten extents, but instead we zero
224 	 * the block before we commit the transaction.  Ideally we'd like to do
225 	 * this outside the transaction context, but if we commit and then crash
226 	 * we may not have zeroed the blocks and this will be exposed on
227 	 * recovery of the allocation. Hence we must zero before commit.
228 	 *
229 	 * Further, if we are mapping unwritten extents here, we need to zero
230 	 * and convert them to written so that we don't need an unwritten extent
231 	 * callback for DAX. This also means that we need to be able to dip into
232 	 * the reserve block pool for bmbt block allocation if there is no space
233 	 * left but we need to do unwritten extent conversion.
234 	 */
235 	if (IS_DAX(VFS_I(ip))) {
236 		bmapi_flags = XFS_BMAPI_CONVERT | XFS_BMAPI_ZERO;
237 		if (imap->br_state == XFS_EXT_UNWRITTEN) {
238 			tflags |= XFS_TRANS_RESERVE;
239 			resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0) << 1;
240 		}
241 	}
242 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, resrtextents,
243 			tflags, &tp);
244 	if (error)
245 		return error;
246 
247 	xfs_ilock(ip, XFS_ILOCK_EXCL);
248 
249 	error = xfs_trans_reserve_quota_nblks(tp, ip, qblocks, 0, quota_flag);
250 	if (error)
251 		goto out_trans_cancel;
252 
253 	xfs_trans_ijoin(tp, ip, 0);
254 
255 	/*
256 	 * From this point onwards we overwrite the imap pointer that the
257 	 * caller gave to us.
258 	 */
259 	nimaps = 1;
260 	error = xfs_bmapi_write(tp, ip, offset_fsb, count_fsb, bmapi_flags, 0,
261 				imap, &nimaps);
262 	if (error)
263 		goto out_res_cancel;
264 
265 	/*
266 	 * Complete the transaction
267 	 */
268 	error = xfs_trans_commit(tp);
269 	if (error)
270 		goto out_unlock;
271 
272 	/*
273 	 * Copy any maps to caller's array and return any error.
274 	 */
275 	if (nimaps == 0) {
276 		error = -ENOSPC;
277 		goto out_unlock;
278 	}
279 
280 	if (unlikely(!xfs_valid_startblock(ip, imap->br_startblock)))
281 		error = xfs_alert_fsblock_zero(ip, imap);
282 
283 out_unlock:
284 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
285 	return error;
286 
287 out_res_cancel:
288 	xfs_trans_unreserve_quota_nblks(tp, ip, (long)qblocks, 0, quota_flag);
289 out_trans_cancel:
290 	xfs_trans_cancel(tp);
291 	goto out_unlock;
292 }
293 
294 STATIC bool
295 xfs_quota_need_throttle(
296 	struct xfs_inode *ip,
297 	int type,
298 	xfs_fsblock_t alloc_blocks)
299 {
300 	struct xfs_dquot *dq = xfs_inode_dquot(ip, type);
301 
302 	if (!dq || !xfs_this_quota_on(ip->i_mount, type))
303 		return false;
304 
305 	/* no hi watermark, no throttle */
306 	if (!dq->q_prealloc_hi_wmark)
307 		return false;
308 
309 	/* under the lo watermark, no throttle */
310 	if (dq->q_res_bcount + alloc_blocks < dq->q_prealloc_lo_wmark)
311 		return false;
312 
313 	return true;
314 }
315 
316 STATIC void
317 xfs_quota_calc_throttle(
318 	struct xfs_inode *ip,
319 	int type,
320 	xfs_fsblock_t *qblocks,
321 	int *qshift,
322 	int64_t	*qfreesp)
323 {
324 	int64_t freesp;
325 	int shift = 0;
326 	struct xfs_dquot *dq = xfs_inode_dquot(ip, type);
327 
328 	/* no dq, or over hi wmark, squash the prealloc completely */
329 	if (!dq || dq->q_res_bcount >= dq->q_prealloc_hi_wmark) {
330 		*qblocks = 0;
331 		*qfreesp = 0;
332 		return;
333 	}
334 
335 	freesp = dq->q_prealloc_hi_wmark - dq->q_res_bcount;
336 	if (freesp < dq->q_low_space[XFS_QLOWSP_5_PCNT]) {
337 		shift = 2;
338 		if (freesp < dq->q_low_space[XFS_QLOWSP_3_PCNT])
339 			shift += 2;
340 		if (freesp < dq->q_low_space[XFS_QLOWSP_1_PCNT])
341 			shift += 2;
342 	}
343 
344 	if (freesp < *qfreesp)
345 		*qfreesp = freesp;
346 
347 	/* only overwrite the throttle values if we are more aggressive */
348 	if ((freesp >> shift) < (*qblocks >> *qshift)) {
349 		*qblocks = freesp;
350 		*qshift = shift;
351 	}
352 }
353 
354 /*
355  * If we are doing a write at the end of the file and there are no allocations
356  * past this one, then extend the allocation out to the file system's write
357  * iosize.
358  *
359  * If we don't have a user specified preallocation size, dynamically increase
360  * the preallocation size as the size of the file grows.  Cap the maximum size
361  * at a single extent or less if the filesystem is near full. The closer the
362  * filesystem is to full, the smaller the maximum prealocation.
363  *
364  * As an exception we don't do any preallocation at all if the file is smaller
365  * than the minimum preallocation and we are using the default dynamic
366  * preallocation scheme, as it is likely this is the only write to the file that
367  * is going to be done.
368  *
369  * We clean up any extra space left over when the file is closed in
370  * xfs_inactive().
371  */
372 STATIC xfs_fsblock_t
373 xfs_iomap_prealloc_size(
374 	struct xfs_inode	*ip,
375 	int			whichfork,
376 	loff_t			offset,
377 	loff_t			count,
378 	struct xfs_iext_cursor	*icur)
379 {
380 	struct xfs_mount	*mp = ip->i_mount;
381 	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, whichfork);
382 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
383 	struct xfs_bmbt_irec	prev;
384 	int			shift = 0;
385 	int64_t			freesp;
386 	xfs_fsblock_t		qblocks;
387 	int			qshift = 0;
388 	xfs_fsblock_t		alloc_blocks = 0;
389 
390 	if (offset + count <= XFS_ISIZE(ip))
391 		return 0;
392 
393 	if (!(mp->m_flags & XFS_MOUNT_ALLOCSIZE) &&
394 	    (XFS_ISIZE(ip) < XFS_FSB_TO_B(mp, mp->m_allocsize_blocks)))
395 		return 0;
396 
397 	/*
398 	 * If an explicit allocsize is set, the file is small, or we
399 	 * are writing behind a hole, then use the minimum prealloc:
400 	 */
401 	if ((mp->m_flags & XFS_MOUNT_ALLOCSIZE) ||
402 	    XFS_ISIZE(ip) < XFS_FSB_TO_B(mp, mp->m_dalign) ||
403 	    !xfs_iext_peek_prev_extent(ifp, icur, &prev) ||
404 	    prev.br_startoff + prev.br_blockcount < offset_fsb)
405 		return mp->m_allocsize_blocks;
406 
407 	/*
408 	 * Determine the initial size of the preallocation. We are beyond the
409 	 * current EOF here, but we need to take into account whether this is
410 	 * a sparse write or an extending write when determining the
411 	 * preallocation size.  Hence we need to look up the extent that ends
412 	 * at the current write offset and use the result to determine the
413 	 * preallocation size.
414 	 *
415 	 * If the extent is a hole, then preallocation is essentially disabled.
416 	 * Otherwise we take the size of the preceding data extent as the basis
417 	 * for the preallocation size. If the size of the extent is greater than
418 	 * half the maximum extent length, then use the current offset as the
419 	 * basis. This ensures that for large files the preallocation size
420 	 * always extends to MAXEXTLEN rather than falling short due to things
421 	 * like stripe unit/width alignment of real extents.
422 	 */
423 	if (prev.br_blockcount <= (MAXEXTLEN >> 1))
424 		alloc_blocks = prev.br_blockcount << 1;
425 	else
426 		alloc_blocks = XFS_B_TO_FSB(mp, offset);
427 	if (!alloc_blocks)
428 		goto check_writeio;
429 	qblocks = alloc_blocks;
430 
431 	/*
432 	 * MAXEXTLEN is not a power of two value but we round the prealloc down
433 	 * to the nearest power of two value after throttling. To prevent the
434 	 * round down from unconditionally reducing the maximum supported prealloc
435 	 * size, we round up first, apply appropriate throttling, round down and
436 	 * cap the value to MAXEXTLEN.
437 	 */
438 	alloc_blocks = XFS_FILEOFF_MIN(roundup_pow_of_two(MAXEXTLEN),
439 				       alloc_blocks);
440 
441 	freesp = percpu_counter_read_positive(&mp->m_fdblocks);
442 	if (freesp < mp->m_low_space[XFS_LOWSP_5_PCNT]) {
443 		shift = 2;
444 		if (freesp < mp->m_low_space[XFS_LOWSP_4_PCNT])
445 			shift++;
446 		if (freesp < mp->m_low_space[XFS_LOWSP_3_PCNT])
447 			shift++;
448 		if (freesp < mp->m_low_space[XFS_LOWSP_2_PCNT])
449 			shift++;
450 		if (freesp < mp->m_low_space[XFS_LOWSP_1_PCNT])
451 			shift++;
452 	}
453 
454 	/*
455 	 * Check each quota to cap the prealloc size, provide a shift value to
456 	 * throttle with and adjust amount of available space.
457 	 */
458 	if (xfs_quota_need_throttle(ip, XFS_DQ_USER, alloc_blocks))
459 		xfs_quota_calc_throttle(ip, XFS_DQ_USER, &qblocks, &qshift,
460 					&freesp);
461 	if (xfs_quota_need_throttle(ip, XFS_DQ_GROUP, alloc_blocks))
462 		xfs_quota_calc_throttle(ip, XFS_DQ_GROUP, &qblocks, &qshift,
463 					&freesp);
464 	if (xfs_quota_need_throttle(ip, XFS_DQ_PROJ, alloc_blocks))
465 		xfs_quota_calc_throttle(ip, XFS_DQ_PROJ, &qblocks, &qshift,
466 					&freesp);
467 
468 	/*
469 	 * The final prealloc size is set to the minimum of free space available
470 	 * in each of the quotas and the overall filesystem.
471 	 *
472 	 * The shift throttle value is set to the maximum value as determined by
473 	 * the global low free space values and per-quota low free space values.
474 	 */
475 	alloc_blocks = min(alloc_blocks, qblocks);
476 	shift = max(shift, qshift);
477 
478 	if (shift)
479 		alloc_blocks >>= shift;
480 	/*
481 	 * rounddown_pow_of_two() returns an undefined result if we pass in
482 	 * alloc_blocks = 0.
483 	 */
484 	if (alloc_blocks)
485 		alloc_blocks = rounddown_pow_of_two(alloc_blocks);
486 	if (alloc_blocks > MAXEXTLEN)
487 		alloc_blocks = MAXEXTLEN;
488 
489 	/*
490 	 * If we are still trying to allocate more space than is
491 	 * available, squash the prealloc hard. This can happen if we
492 	 * have a large file on a small filesystem and the above
493 	 * lowspace thresholds are smaller than MAXEXTLEN.
494 	 */
495 	while (alloc_blocks && alloc_blocks >= freesp)
496 		alloc_blocks >>= 4;
497 check_writeio:
498 	if (alloc_blocks < mp->m_allocsize_blocks)
499 		alloc_blocks = mp->m_allocsize_blocks;
500 	trace_xfs_iomap_prealloc_size(ip, alloc_blocks, shift,
501 				      mp->m_allocsize_blocks);
502 	return alloc_blocks;
503 }
504 
505 int
506 xfs_iomap_write_unwritten(
507 	xfs_inode_t	*ip,
508 	xfs_off_t	offset,
509 	xfs_off_t	count,
510 	bool		update_isize)
511 {
512 	xfs_mount_t	*mp = ip->i_mount;
513 	xfs_fileoff_t	offset_fsb;
514 	xfs_filblks_t	count_fsb;
515 	xfs_filblks_t	numblks_fsb;
516 	int		nimaps;
517 	xfs_trans_t	*tp;
518 	xfs_bmbt_irec_t imap;
519 	struct inode	*inode = VFS_I(ip);
520 	xfs_fsize_t	i_size;
521 	uint		resblks;
522 	int		error;
523 
524 	trace_xfs_unwritten_convert(ip, offset, count);
525 
526 	offset_fsb = XFS_B_TO_FSBT(mp, offset);
527 	count_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
528 	count_fsb = (xfs_filblks_t)(count_fsb - offset_fsb);
529 
530 	/*
531 	 * Reserve enough blocks in this transaction for two complete extent
532 	 * btree splits.  We may be converting the middle part of an unwritten
533 	 * extent and in this case we will insert two new extents in the btree
534 	 * each of which could cause a full split.
535 	 *
536 	 * This reservation amount will be used in the first call to
537 	 * xfs_bmbt_split() to select an AG with enough space to satisfy the
538 	 * rest of the operation.
539 	 */
540 	resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0) << 1;
541 
542 	/* Attach dquots so that bmbt splits are accounted correctly. */
543 	error = xfs_qm_dqattach(ip);
544 	if (error)
545 		return error;
546 
547 	do {
548 		/*
549 		 * Set up a transaction to convert the range of extents
550 		 * from unwritten to real. Do allocations in a loop until
551 		 * we have covered the range passed in.
552 		 *
553 		 * Note that we can't risk to recursing back into the filesystem
554 		 * here as we might be asked to write out the same inode that we
555 		 * complete here and might deadlock on the iolock.
556 		 */
557 		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0,
558 				XFS_TRANS_RESERVE, &tp);
559 		if (error)
560 			return error;
561 
562 		xfs_ilock(ip, XFS_ILOCK_EXCL);
563 		xfs_trans_ijoin(tp, ip, 0);
564 
565 		error = xfs_trans_reserve_quota_nblks(tp, ip, resblks, 0,
566 				XFS_QMOPT_RES_REGBLKS);
567 		if (error)
568 			goto error_on_bmapi_transaction;
569 
570 		/*
571 		 * Modify the unwritten extent state of the buffer.
572 		 */
573 		nimaps = 1;
574 		error = xfs_bmapi_write(tp, ip, offset_fsb, count_fsb,
575 					XFS_BMAPI_CONVERT, resblks, &imap,
576 					&nimaps);
577 		if (error)
578 			goto error_on_bmapi_transaction;
579 
580 		/*
581 		 * Log the updated inode size as we go.  We have to be careful
582 		 * to only log it up to the actual write offset if it is
583 		 * halfway into a block.
584 		 */
585 		i_size = XFS_FSB_TO_B(mp, offset_fsb + count_fsb);
586 		if (i_size > offset + count)
587 			i_size = offset + count;
588 		if (update_isize && i_size > i_size_read(inode))
589 			i_size_write(inode, i_size);
590 		i_size = xfs_new_eof(ip, i_size);
591 		if (i_size) {
592 			ip->i_d.di_size = i_size;
593 			xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
594 		}
595 
596 		error = xfs_trans_commit(tp);
597 		xfs_iunlock(ip, XFS_ILOCK_EXCL);
598 		if (error)
599 			return error;
600 
601 		if (unlikely(!xfs_valid_startblock(ip, imap.br_startblock)))
602 			return xfs_alert_fsblock_zero(ip, &imap);
603 
604 		if ((numblks_fsb = imap.br_blockcount) == 0) {
605 			/*
606 			 * The numblks_fsb value should always get
607 			 * smaller, otherwise the loop is stuck.
608 			 */
609 			ASSERT(imap.br_blockcount);
610 			break;
611 		}
612 		offset_fsb += numblks_fsb;
613 		count_fsb -= numblks_fsb;
614 	} while (count_fsb > 0);
615 
616 	return 0;
617 
618 error_on_bmapi_transaction:
619 	xfs_trans_cancel(tp);
620 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
621 	return error;
622 }
623 
624 static inline bool
625 imap_needs_alloc(
626 	struct inode		*inode,
627 	unsigned		flags,
628 	struct xfs_bmbt_irec	*imap,
629 	int			nimaps)
630 {
631 	/* don't allocate blocks when just zeroing */
632 	if (flags & IOMAP_ZERO)
633 		return false;
634 	if (!nimaps ||
635 	    imap->br_startblock == HOLESTARTBLOCK ||
636 	    imap->br_startblock == DELAYSTARTBLOCK)
637 		return true;
638 	/* we convert unwritten extents before copying the data for DAX */
639 	if (IS_DAX(inode) && imap->br_state == XFS_EXT_UNWRITTEN)
640 		return true;
641 	return false;
642 }
643 
644 static inline bool
645 imap_needs_cow(
646 	struct xfs_inode	*ip,
647 	unsigned int		flags,
648 	struct xfs_bmbt_irec	*imap,
649 	int			nimaps)
650 {
651 	if (!xfs_is_cow_inode(ip))
652 		return false;
653 
654 	/* when zeroing we don't have to COW holes or unwritten extents */
655 	if (flags & IOMAP_ZERO) {
656 		if (!nimaps ||
657 		    imap->br_startblock == HOLESTARTBLOCK ||
658 		    imap->br_state == XFS_EXT_UNWRITTEN)
659 			return false;
660 	}
661 
662 	return true;
663 }
664 
665 static int
666 xfs_ilock_for_iomap(
667 	struct xfs_inode	*ip,
668 	unsigned		flags,
669 	unsigned		*lockmode)
670 {
671 	unsigned		mode = XFS_ILOCK_SHARED;
672 	bool			is_write = flags & (IOMAP_WRITE | IOMAP_ZERO);
673 
674 	/*
675 	 * COW writes may allocate delalloc space or convert unwritten COW
676 	 * extents, so we need to make sure to take the lock exclusively here.
677 	 */
678 	if (xfs_is_cow_inode(ip) && is_write)
679 		mode = XFS_ILOCK_EXCL;
680 
681 	/*
682 	 * Extents not yet cached requires exclusive access, don't block.  This
683 	 * is an opencoded xfs_ilock_data_map_shared() call but with
684 	 * non-blocking behaviour.
685 	 */
686 	if (!(ip->i_df.if_flags & XFS_IFEXTENTS)) {
687 		if (flags & IOMAP_NOWAIT)
688 			return -EAGAIN;
689 		mode = XFS_ILOCK_EXCL;
690 	}
691 
692 relock:
693 	if (flags & IOMAP_NOWAIT) {
694 		if (!xfs_ilock_nowait(ip, mode))
695 			return -EAGAIN;
696 	} else {
697 		xfs_ilock(ip, mode);
698 	}
699 
700 	/*
701 	 * The reflink iflag could have changed since the earlier unlocked
702 	 * check, so if we got ILOCK_SHARED for a write and but we're now a
703 	 * reflink inode we have to switch to ILOCK_EXCL and relock.
704 	 */
705 	if (mode == XFS_ILOCK_SHARED && is_write && xfs_is_cow_inode(ip)) {
706 		xfs_iunlock(ip, mode);
707 		mode = XFS_ILOCK_EXCL;
708 		goto relock;
709 	}
710 
711 	*lockmode = mode;
712 	return 0;
713 }
714 
715 static int
716 xfs_direct_write_iomap_begin(
717 	struct inode		*inode,
718 	loff_t			offset,
719 	loff_t			length,
720 	unsigned		flags,
721 	struct iomap		*iomap,
722 	struct iomap		*srcmap)
723 {
724 	struct xfs_inode	*ip = XFS_I(inode);
725 	struct xfs_mount	*mp = ip->i_mount;
726 	struct xfs_bmbt_irec	imap, cmap;
727 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
728 	xfs_fileoff_t		end_fsb = xfs_iomap_end_fsb(mp, offset, length);
729 	int			nimaps = 1, error = 0;
730 	bool			shared = false;
731 	u16			iomap_flags = 0;
732 	unsigned		lockmode;
733 
734 	ASSERT(flags & (IOMAP_WRITE | IOMAP_ZERO));
735 
736 	if (XFS_FORCED_SHUTDOWN(mp))
737 		return -EIO;
738 
739 	/*
740 	 * Writes that span EOF might trigger an IO size update on completion,
741 	 * so consider them to be dirty for the purposes of O_DSYNC even if
742 	 * there is no other metadata changes pending or have been made here.
743 	 */
744 	if (offset + length > i_size_read(inode))
745 		iomap_flags |= IOMAP_F_DIRTY;
746 
747 	error = xfs_ilock_for_iomap(ip, flags, &lockmode);
748 	if (error)
749 		return error;
750 
751 	error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, &imap,
752 			       &nimaps, 0);
753 	if (error)
754 		goto out_unlock;
755 
756 	if (imap_needs_cow(ip, flags, &imap, nimaps)) {
757 		error = -EAGAIN;
758 		if (flags & IOMAP_NOWAIT)
759 			goto out_unlock;
760 
761 		/* may drop and re-acquire the ilock */
762 		error = xfs_reflink_allocate_cow(ip, &imap, &cmap, &shared,
763 				&lockmode, flags & IOMAP_DIRECT);
764 		if (error)
765 			goto out_unlock;
766 		if (shared)
767 			goto out_found_cow;
768 		end_fsb = imap.br_startoff + imap.br_blockcount;
769 		length = XFS_FSB_TO_B(mp, end_fsb) - offset;
770 	}
771 
772 	if (imap_needs_alloc(inode, flags, &imap, nimaps))
773 		goto allocate_blocks;
774 
775 	xfs_iunlock(ip, lockmode);
776 	trace_xfs_iomap_found(ip, offset, length, XFS_DATA_FORK, &imap);
777 	return xfs_bmbt_to_iomap(ip, iomap, &imap, iomap_flags);
778 
779 allocate_blocks:
780 	error = -EAGAIN;
781 	if (flags & IOMAP_NOWAIT)
782 		goto out_unlock;
783 
784 	/*
785 	 * We cap the maximum length we map to a sane size  to keep the chunks
786 	 * of work done where somewhat symmetric with the work writeback does.
787 	 * This is a completely arbitrary number pulled out of thin air as a
788 	 * best guess for initial testing.
789 	 *
790 	 * Note that the values needs to be less than 32-bits wide until the
791 	 * lower level functions are updated.
792 	 */
793 	length = min_t(loff_t, length, 1024 * PAGE_SIZE);
794 	end_fsb = xfs_iomap_end_fsb(mp, offset, length);
795 
796 	if (offset + length > XFS_ISIZE(ip))
797 		end_fsb = xfs_iomap_eof_align_last_fsb(ip, end_fsb);
798 	else if (nimaps && imap.br_startblock == HOLESTARTBLOCK)
799 		end_fsb = min(end_fsb, imap.br_startoff + imap.br_blockcount);
800 	xfs_iunlock(ip, lockmode);
801 
802 	error = xfs_iomap_write_direct(ip, offset_fsb, end_fsb - offset_fsb,
803 			&imap);
804 	if (error)
805 		return error;
806 
807 	trace_xfs_iomap_alloc(ip, offset, length, XFS_DATA_FORK, &imap);
808 	return xfs_bmbt_to_iomap(ip, iomap, &imap, iomap_flags | IOMAP_F_NEW);
809 
810 out_found_cow:
811 	xfs_iunlock(ip, lockmode);
812 	length = XFS_FSB_TO_B(mp, cmap.br_startoff + cmap.br_blockcount);
813 	trace_xfs_iomap_found(ip, offset, length - offset, XFS_COW_FORK, &cmap);
814 	if (imap.br_startblock != HOLESTARTBLOCK) {
815 		error = xfs_bmbt_to_iomap(ip, srcmap, &imap, 0);
816 		if (error)
817 			return error;
818 	}
819 	return xfs_bmbt_to_iomap(ip, iomap, &cmap, IOMAP_F_SHARED);
820 
821 out_unlock:
822 	xfs_iunlock(ip, lockmode);
823 	return error;
824 }
825 
826 const struct iomap_ops xfs_direct_write_iomap_ops = {
827 	.iomap_begin		= xfs_direct_write_iomap_begin,
828 };
829 
830 static int
831 xfs_buffered_write_iomap_begin(
832 	struct inode		*inode,
833 	loff_t			offset,
834 	loff_t			count,
835 	unsigned		flags,
836 	struct iomap		*iomap,
837 	struct iomap		*srcmap)
838 {
839 	struct xfs_inode	*ip = XFS_I(inode);
840 	struct xfs_mount	*mp = ip->i_mount;
841 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
842 	xfs_fileoff_t		end_fsb = xfs_iomap_end_fsb(mp, offset, count);
843 	struct xfs_bmbt_irec	imap, cmap;
844 	struct xfs_iext_cursor	icur, ccur;
845 	xfs_fsblock_t		prealloc_blocks = 0;
846 	bool			eof = false, cow_eof = false, shared = false;
847 	int			allocfork = XFS_DATA_FORK;
848 	int			error = 0;
849 
850 	/* we can't use delayed allocations when using extent size hints */
851 	if (xfs_get_extsz_hint(ip))
852 		return xfs_direct_write_iomap_begin(inode, offset, count,
853 				flags, iomap, srcmap);
854 
855 	ASSERT(!XFS_IS_REALTIME_INODE(ip));
856 
857 	xfs_ilock(ip, XFS_ILOCK_EXCL);
858 
859 	if (XFS_IS_CORRUPT(mp, !xfs_ifork_has_extents(ip, XFS_DATA_FORK)) ||
860 	    XFS_TEST_ERROR(false, mp, XFS_ERRTAG_BMAPIFORMAT)) {
861 		error = -EFSCORRUPTED;
862 		goto out_unlock;
863 	}
864 
865 	XFS_STATS_INC(mp, xs_blk_mapw);
866 
867 	if (!(ip->i_df.if_flags & XFS_IFEXTENTS)) {
868 		error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK);
869 		if (error)
870 			goto out_unlock;
871 	}
872 
873 	/*
874 	 * Search the data fork fork first to look up our source mapping.  We
875 	 * always need the data fork map, as we have to return it to the
876 	 * iomap code so that the higher level write code can read data in to
877 	 * perform read-modify-write cycles for unaligned writes.
878 	 */
879 	eof = !xfs_iext_lookup_extent(ip, &ip->i_df, offset_fsb, &icur, &imap);
880 	if (eof)
881 		imap.br_startoff = end_fsb; /* fake hole until the end */
882 
883 	/* We never need to allocate blocks for zeroing a hole. */
884 	if ((flags & IOMAP_ZERO) && imap.br_startoff > offset_fsb) {
885 		xfs_hole_to_iomap(ip, iomap, offset_fsb, imap.br_startoff);
886 		goto out_unlock;
887 	}
888 
889 	/*
890 	 * Search the COW fork extent list even if we did not find a data fork
891 	 * extent.  This serves two purposes: first this implements the
892 	 * speculative preallocation using cowextsize, so that we also unshare
893 	 * block adjacent to shared blocks instead of just the shared blocks
894 	 * themselves.  Second the lookup in the extent list is generally faster
895 	 * than going out to the shared extent tree.
896 	 */
897 	if (xfs_is_cow_inode(ip)) {
898 		if (!ip->i_cowfp) {
899 			ASSERT(!xfs_is_reflink_inode(ip));
900 			xfs_ifork_init_cow(ip);
901 		}
902 		cow_eof = !xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb,
903 				&ccur, &cmap);
904 		if (!cow_eof && cmap.br_startoff <= offset_fsb) {
905 			trace_xfs_reflink_cow_found(ip, &cmap);
906 			goto found_cow;
907 		}
908 	}
909 
910 	if (imap.br_startoff <= offset_fsb) {
911 		/*
912 		 * For reflink files we may need a delalloc reservation when
913 		 * overwriting shared extents.   This includes zeroing of
914 		 * existing extents that contain data.
915 		 */
916 		if (!xfs_is_cow_inode(ip) ||
917 		    ((flags & IOMAP_ZERO) && imap.br_state != XFS_EXT_NORM)) {
918 			trace_xfs_iomap_found(ip, offset, count, XFS_DATA_FORK,
919 					&imap);
920 			goto found_imap;
921 		}
922 
923 		xfs_trim_extent(&imap, offset_fsb, end_fsb - offset_fsb);
924 
925 		/* Trim the mapping to the nearest shared extent boundary. */
926 		error = xfs_bmap_trim_cow(ip, &imap, &shared);
927 		if (error)
928 			goto out_unlock;
929 
930 		/* Not shared?  Just report the (potentially capped) extent. */
931 		if (!shared) {
932 			trace_xfs_iomap_found(ip, offset, count, XFS_DATA_FORK,
933 					&imap);
934 			goto found_imap;
935 		}
936 
937 		/*
938 		 * Fork all the shared blocks from our write offset until the
939 		 * end of the extent.
940 		 */
941 		allocfork = XFS_COW_FORK;
942 		end_fsb = imap.br_startoff + imap.br_blockcount;
943 	} else {
944 		/*
945 		 * We cap the maximum length we map here to MAX_WRITEBACK_PAGES
946 		 * pages to keep the chunks of work done where somewhat
947 		 * symmetric with the work writeback does.  This is a completely
948 		 * arbitrary number pulled out of thin air.
949 		 *
950 		 * Note that the values needs to be less than 32-bits wide until
951 		 * the lower level functions are updated.
952 		 */
953 		count = min_t(loff_t, count, 1024 * PAGE_SIZE);
954 		end_fsb = xfs_iomap_end_fsb(mp, offset, count);
955 
956 		if (xfs_is_always_cow_inode(ip))
957 			allocfork = XFS_COW_FORK;
958 	}
959 
960 	error = xfs_qm_dqattach_locked(ip, false);
961 	if (error)
962 		goto out_unlock;
963 
964 	if (eof) {
965 		prealloc_blocks = xfs_iomap_prealloc_size(ip, allocfork, offset,
966 				count, &icur);
967 		if (prealloc_blocks) {
968 			xfs_extlen_t	align;
969 			xfs_off_t	end_offset;
970 			xfs_fileoff_t	p_end_fsb;
971 
972 			end_offset = XFS_ALLOC_ALIGN(mp, offset + count - 1);
973 			p_end_fsb = XFS_B_TO_FSBT(mp, end_offset) +
974 					prealloc_blocks;
975 
976 			align = xfs_eof_alignment(ip);
977 			if (align)
978 				p_end_fsb = roundup_64(p_end_fsb, align);
979 
980 			p_end_fsb = min(p_end_fsb,
981 				XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes));
982 			ASSERT(p_end_fsb > offset_fsb);
983 			prealloc_blocks = p_end_fsb - end_fsb;
984 		}
985 	}
986 
987 retry:
988 	error = xfs_bmapi_reserve_delalloc(ip, allocfork, offset_fsb,
989 			end_fsb - offset_fsb, prealloc_blocks,
990 			allocfork == XFS_DATA_FORK ? &imap : &cmap,
991 			allocfork == XFS_DATA_FORK ? &icur : &ccur,
992 			allocfork == XFS_DATA_FORK ? eof : cow_eof);
993 	switch (error) {
994 	case 0:
995 		break;
996 	case -ENOSPC:
997 	case -EDQUOT:
998 		/* retry without any preallocation */
999 		trace_xfs_delalloc_enospc(ip, offset, count);
1000 		if (prealloc_blocks) {
1001 			prealloc_blocks = 0;
1002 			goto retry;
1003 		}
1004 		/*FALLTHRU*/
1005 	default:
1006 		goto out_unlock;
1007 	}
1008 
1009 	if (allocfork == XFS_COW_FORK) {
1010 		trace_xfs_iomap_alloc(ip, offset, count, allocfork, &cmap);
1011 		goto found_cow;
1012 	}
1013 
1014 	/*
1015 	 * Flag newly allocated delalloc blocks with IOMAP_F_NEW so we punch
1016 	 * them out if the write happens to fail.
1017 	 */
1018 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1019 	trace_xfs_iomap_alloc(ip, offset, count, allocfork, &imap);
1020 	return xfs_bmbt_to_iomap(ip, iomap, &imap, IOMAP_F_NEW);
1021 
1022 found_imap:
1023 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1024 	return xfs_bmbt_to_iomap(ip, iomap, &imap, 0);
1025 
1026 found_cow:
1027 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1028 	if (imap.br_startoff <= offset_fsb) {
1029 		error = xfs_bmbt_to_iomap(ip, srcmap, &imap, 0);
1030 		if (error)
1031 			return error;
1032 	} else {
1033 		xfs_trim_extent(&cmap, offset_fsb,
1034 				imap.br_startoff - offset_fsb);
1035 	}
1036 	return xfs_bmbt_to_iomap(ip, iomap, &cmap, IOMAP_F_SHARED);
1037 
1038 out_unlock:
1039 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1040 	return error;
1041 }
1042 
1043 static int
1044 xfs_buffered_write_iomap_end(
1045 	struct inode		*inode,
1046 	loff_t			offset,
1047 	loff_t			length,
1048 	ssize_t			written,
1049 	unsigned		flags,
1050 	struct iomap		*iomap)
1051 {
1052 	struct xfs_inode	*ip = XFS_I(inode);
1053 	struct xfs_mount	*mp = ip->i_mount;
1054 	xfs_fileoff_t		start_fsb;
1055 	xfs_fileoff_t		end_fsb;
1056 	int			error = 0;
1057 
1058 	if (iomap->type != IOMAP_DELALLOC)
1059 		return 0;
1060 
1061 	/*
1062 	 * Behave as if the write failed if drop writes is enabled. Set the NEW
1063 	 * flag to force delalloc cleanup.
1064 	 */
1065 	if (XFS_TEST_ERROR(false, mp, XFS_ERRTAG_DROP_WRITES)) {
1066 		iomap->flags |= IOMAP_F_NEW;
1067 		written = 0;
1068 	}
1069 
1070 	/*
1071 	 * start_fsb refers to the first unused block after a short write. If
1072 	 * nothing was written, round offset down to point at the first block in
1073 	 * the range.
1074 	 */
1075 	if (unlikely(!written))
1076 		start_fsb = XFS_B_TO_FSBT(mp, offset);
1077 	else
1078 		start_fsb = XFS_B_TO_FSB(mp, offset + written);
1079 	end_fsb = XFS_B_TO_FSB(mp, offset + length);
1080 
1081 	/*
1082 	 * Trim delalloc blocks if they were allocated by this write and we
1083 	 * didn't manage to write the whole range.
1084 	 *
1085 	 * We don't need to care about racing delalloc as we hold i_mutex
1086 	 * across the reserve/allocate/unreserve calls. If there are delalloc
1087 	 * blocks in the range, they are ours.
1088 	 */
1089 	if ((iomap->flags & IOMAP_F_NEW) && start_fsb < end_fsb) {
1090 		truncate_pagecache_range(VFS_I(ip), XFS_FSB_TO_B(mp, start_fsb),
1091 					 XFS_FSB_TO_B(mp, end_fsb) - 1);
1092 
1093 		error = xfs_bmap_punch_delalloc_range(ip, start_fsb,
1094 					       end_fsb - start_fsb);
1095 		if (error && !XFS_FORCED_SHUTDOWN(mp)) {
1096 			xfs_alert(mp, "%s: unable to clean up ino %lld",
1097 				__func__, ip->i_ino);
1098 			return error;
1099 		}
1100 	}
1101 
1102 	return 0;
1103 }
1104 
1105 const struct iomap_ops xfs_buffered_write_iomap_ops = {
1106 	.iomap_begin		= xfs_buffered_write_iomap_begin,
1107 	.iomap_end		= xfs_buffered_write_iomap_end,
1108 };
1109 
1110 static int
1111 xfs_read_iomap_begin(
1112 	struct inode		*inode,
1113 	loff_t			offset,
1114 	loff_t			length,
1115 	unsigned		flags,
1116 	struct iomap		*iomap,
1117 	struct iomap		*srcmap)
1118 {
1119 	struct xfs_inode	*ip = XFS_I(inode);
1120 	struct xfs_mount	*mp = ip->i_mount;
1121 	struct xfs_bmbt_irec	imap;
1122 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
1123 	xfs_fileoff_t		end_fsb = xfs_iomap_end_fsb(mp, offset, length);
1124 	int			nimaps = 1, error = 0;
1125 	bool			shared = false;
1126 	unsigned		lockmode;
1127 
1128 	ASSERT(!(flags & (IOMAP_WRITE | IOMAP_ZERO)));
1129 
1130 	if (XFS_FORCED_SHUTDOWN(mp))
1131 		return -EIO;
1132 
1133 	error = xfs_ilock_for_iomap(ip, flags, &lockmode);
1134 	if (error)
1135 		return error;
1136 	error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, &imap,
1137 			       &nimaps, 0);
1138 	if (!error && (flags & IOMAP_REPORT))
1139 		error = xfs_reflink_trim_around_shared(ip, &imap, &shared);
1140 	xfs_iunlock(ip, lockmode);
1141 
1142 	if (error)
1143 		return error;
1144 	trace_xfs_iomap_found(ip, offset, length, XFS_DATA_FORK, &imap);
1145 	return xfs_bmbt_to_iomap(ip, iomap, &imap, shared ? IOMAP_F_SHARED : 0);
1146 }
1147 
1148 const struct iomap_ops xfs_read_iomap_ops = {
1149 	.iomap_begin		= xfs_read_iomap_begin,
1150 };
1151 
1152 static int
1153 xfs_seek_iomap_begin(
1154 	struct inode		*inode,
1155 	loff_t			offset,
1156 	loff_t			length,
1157 	unsigned		flags,
1158 	struct iomap		*iomap,
1159 	struct iomap		*srcmap)
1160 {
1161 	struct xfs_inode	*ip = XFS_I(inode);
1162 	struct xfs_mount	*mp = ip->i_mount;
1163 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
1164 	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + length);
1165 	xfs_fileoff_t		cow_fsb = NULLFILEOFF, data_fsb = NULLFILEOFF;
1166 	struct xfs_iext_cursor	icur;
1167 	struct xfs_bmbt_irec	imap, cmap;
1168 	int			error = 0;
1169 	unsigned		lockmode;
1170 
1171 	if (XFS_FORCED_SHUTDOWN(mp))
1172 		return -EIO;
1173 
1174 	lockmode = xfs_ilock_data_map_shared(ip);
1175 	if (!(ip->i_df.if_flags & XFS_IFEXTENTS)) {
1176 		error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK);
1177 		if (error)
1178 			goto out_unlock;
1179 	}
1180 
1181 	if (xfs_iext_lookup_extent(ip, &ip->i_df, offset_fsb, &icur, &imap)) {
1182 		/*
1183 		 * If we found a data extent we are done.
1184 		 */
1185 		if (imap.br_startoff <= offset_fsb)
1186 			goto done;
1187 		data_fsb = imap.br_startoff;
1188 	} else {
1189 		/*
1190 		 * Fake a hole until the end of the file.
1191 		 */
1192 		data_fsb = xfs_iomap_end_fsb(mp, offset, length);
1193 	}
1194 
1195 	/*
1196 	 * If a COW fork extent covers the hole, report it - capped to the next
1197 	 * data fork extent:
1198 	 */
1199 	if (xfs_inode_has_cow_data(ip) &&
1200 	    xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &cmap))
1201 		cow_fsb = cmap.br_startoff;
1202 	if (cow_fsb != NULLFILEOFF && cow_fsb <= offset_fsb) {
1203 		if (data_fsb < cow_fsb + cmap.br_blockcount)
1204 			end_fsb = min(end_fsb, data_fsb);
1205 		xfs_trim_extent(&cmap, offset_fsb, end_fsb);
1206 		error = xfs_bmbt_to_iomap(ip, iomap, &cmap, IOMAP_F_SHARED);
1207 		/*
1208 		 * This is a COW extent, so we must probe the page cache
1209 		 * because there could be dirty page cache being backed
1210 		 * by this extent.
1211 		 */
1212 		iomap->type = IOMAP_UNWRITTEN;
1213 		goto out_unlock;
1214 	}
1215 
1216 	/*
1217 	 * Else report a hole, capped to the next found data or COW extent.
1218 	 */
1219 	if (cow_fsb != NULLFILEOFF && cow_fsb < data_fsb)
1220 		imap.br_blockcount = cow_fsb - offset_fsb;
1221 	else
1222 		imap.br_blockcount = data_fsb - offset_fsb;
1223 	imap.br_startoff = offset_fsb;
1224 	imap.br_startblock = HOLESTARTBLOCK;
1225 	imap.br_state = XFS_EXT_NORM;
1226 done:
1227 	xfs_trim_extent(&imap, offset_fsb, end_fsb);
1228 	error = xfs_bmbt_to_iomap(ip, iomap, &imap, 0);
1229 out_unlock:
1230 	xfs_iunlock(ip, lockmode);
1231 	return error;
1232 }
1233 
1234 const struct iomap_ops xfs_seek_iomap_ops = {
1235 	.iomap_begin		= xfs_seek_iomap_begin,
1236 };
1237 
1238 static int
1239 xfs_xattr_iomap_begin(
1240 	struct inode		*inode,
1241 	loff_t			offset,
1242 	loff_t			length,
1243 	unsigned		flags,
1244 	struct iomap		*iomap,
1245 	struct iomap		*srcmap)
1246 {
1247 	struct xfs_inode	*ip = XFS_I(inode);
1248 	struct xfs_mount	*mp = ip->i_mount;
1249 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
1250 	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + length);
1251 	struct xfs_bmbt_irec	imap;
1252 	int			nimaps = 1, error = 0;
1253 	unsigned		lockmode;
1254 
1255 	if (XFS_FORCED_SHUTDOWN(mp))
1256 		return -EIO;
1257 
1258 	lockmode = xfs_ilock_attr_map_shared(ip);
1259 
1260 	/* if there are no attribute fork or extents, return ENOENT */
1261 	if (!XFS_IFORK_Q(ip) || !ip->i_d.di_anextents) {
1262 		error = -ENOENT;
1263 		goto out_unlock;
1264 	}
1265 
1266 	ASSERT(ip->i_d.di_aformat != XFS_DINODE_FMT_LOCAL);
1267 	error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, &imap,
1268 			       &nimaps, XFS_BMAPI_ATTRFORK);
1269 out_unlock:
1270 	xfs_iunlock(ip, lockmode);
1271 
1272 	if (error)
1273 		return error;
1274 	ASSERT(nimaps);
1275 	return xfs_bmbt_to_iomap(ip, iomap, &imap, 0);
1276 }
1277 
1278 const struct iomap_ops xfs_xattr_iomap_ops = {
1279 	.iomap_begin		= xfs_xattr_iomap_begin,
1280 };
1281