xref: /linux/fs/xfs/xfs_iomap.c (revision b48543c451c30387b53ee6e202dda8d5303f6268)
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 #include "xfs_health.h"
31 #include "xfs_rtbitmap.h"
32 
33 #define XFS_ALLOC_ALIGN(mp, off) \
34 	(((off) >> mp->m_allocsize_log) << mp->m_allocsize_log)
35 
36 static int
37 xfs_alert_fsblock_zero(
38 	xfs_inode_t	*ip,
39 	xfs_bmbt_irec_t	*imap)
40 {
41 	xfs_alert_tag(ip->i_mount, XFS_PTAG_FSBLOCK_ZERO,
42 			"Access to block zero in inode %llu "
43 			"start_block: %llx start_off: %llx "
44 			"blkcnt: %llx extent-state: %x",
45 		(unsigned long long)ip->i_ino,
46 		(unsigned long long)imap->br_startblock,
47 		(unsigned long long)imap->br_startoff,
48 		(unsigned long long)imap->br_blockcount,
49 		imap->br_state);
50 	xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
51 	return -EFSCORRUPTED;
52 }
53 
54 u64
55 xfs_iomap_inode_sequence(
56 	struct xfs_inode	*ip,
57 	u16			iomap_flags)
58 {
59 	u64			cookie = 0;
60 
61 	if (iomap_flags & IOMAP_F_XATTR)
62 		return READ_ONCE(ip->i_af.if_seq);
63 	if ((iomap_flags & IOMAP_F_SHARED) && ip->i_cowfp)
64 		cookie = (u64)READ_ONCE(ip->i_cowfp->if_seq) << 32;
65 	return cookie | READ_ONCE(ip->i_df.if_seq);
66 }
67 
68 /*
69  * Check that the iomap passed to us is still valid for the given offset and
70  * length.
71  */
72 static bool
73 xfs_iomap_valid(
74 	struct inode		*inode,
75 	const struct iomap	*iomap)
76 {
77 	struct xfs_inode	*ip = XFS_I(inode);
78 
79 	if (iomap->validity_cookie !=
80 			xfs_iomap_inode_sequence(ip, iomap->flags)) {
81 		trace_xfs_iomap_invalid(ip, iomap);
82 		return false;
83 	}
84 
85 	XFS_ERRORTAG_DELAY(ip->i_mount, XFS_ERRTAG_WRITE_DELAY_MS);
86 	return true;
87 }
88 
89 static const struct iomap_folio_ops xfs_iomap_folio_ops = {
90 	.iomap_valid		= xfs_iomap_valid,
91 };
92 
93 int
94 xfs_bmbt_to_iomap(
95 	struct xfs_inode	*ip,
96 	struct iomap		*iomap,
97 	struct xfs_bmbt_irec	*imap,
98 	unsigned int		mapping_flags,
99 	u16			iomap_flags,
100 	u64			sequence_cookie)
101 {
102 	struct xfs_mount	*mp = ip->i_mount;
103 	struct xfs_buftarg	*target = xfs_inode_buftarg(ip);
104 
105 	if (unlikely(!xfs_valid_startblock(ip, imap->br_startblock))) {
106 		xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
107 		return xfs_alert_fsblock_zero(ip, imap);
108 	}
109 
110 	if (imap->br_startblock == HOLESTARTBLOCK) {
111 		iomap->addr = IOMAP_NULL_ADDR;
112 		iomap->type = IOMAP_HOLE;
113 	} else if (imap->br_startblock == DELAYSTARTBLOCK ||
114 		   isnullstartblock(imap->br_startblock)) {
115 		iomap->addr = IOMAP_NULL_ADDR;
116 		iomap->type = IOMAP_DELALLOC;
117 	} else {
118 		iomap->addr = BBTOB(xfs_fsb_to_db(ip, imap->br_startblock));
119 		if (mapping_flags & IOMAP_DAX)
120 			iomap->addr += target->bt_dax_part_off;
121 
122 		if (imap->br_state == XFS_EXT_UNWRITTEN)
123 			iomap->type = IOMAP_UNWRITTEN;
124 		else
125 			iomap->type = IOMAP_MAPPED;
126 
127 	}
128 	iomap->offset = XFS_FSB_TO_B(mp, imap->br_startoff);
129 	iomap->length = XFS_FSB_TO_B(mp, imap->br_blockcount);
130 	if (mapping_flags & IOMAP_DAX)
131 		iomap->dax_dev = target->bt_daxdev;
132 	else
133 		iomap->bdev = target->bt_bdev;
134 	iomap->flags = iomap_flags;
135 
136 	if (xfs_ipincount(ip) &&
137 	    (ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
138 		iomap->flags |= IOMAP_F_DIRTY;
139 
140 	iomap->validity_cookie = sequence_cookie;
141 	iomap->folio_ops = &xfs_iomap_folio_ops;
142 	return 0;
143 }
144 
145 static void
146 xfs_hole_to_iomap(
147 	struct xfs_inode	*ip,
148 	struct iomap		*iomap,
149 	xfs_fileoff_t		offset_fsb,
150 	xfs_fileoff_t		end_fsb)
151 {
152 	struct xfs_buftarg	*target = xfs_inode_buftarg(ip);
153 
154 	iomap->addr = IOMAP_NULL_ADDR;
155 	iomap->type = IOMAP_HOLE;
156 	iomap->offset = XFS_FSB_TO_B(ip->i_mount, offset_fsb);
157 	iomap->length = XFS_FSB_TO_B(ip->i_mount, end_fsb - offset_fsb);
158 	iomap->bdev = target->bt_bdev;
159 	iomap->dax_dev = target->bt_daxdev;
160 }
161 
162 static inline xfs_fileoff_t
163 xfs_iomap_end_fsb(
164 	struct xfs_mount	*mp,
165 	loff_t			offset,
166 	loff_t			count)
167 {
168 	ASSERT(offset <= mp->m_super->s_maxbytes);
169 	return min(XFS_B_TO_FSB(mp, offset + count),
170 		   XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes));
171 }
172 
173 static xfs_extlen_t
174 xfs_eof_alignment(
175 	struct xfs_inode	*ip)
176 {
177 	struct xfs_mount	*mp = ip->i_mount;
178 	xfs_extlen_t		align = 0;
179 
180 	if (!XFS_IS_REALTIME_INODE(ip)) {
181 		/*
182 		 * Round up the allocation request to a stripe unit
183 		 * (m_dalign) boundary if the file size is >= stripe unit
184 		 * size, and we are allocating past the allocation eof.
185 		 *
186 		 * If mounted with the "-o swalloc" option the alignment is
187 		 * increased from the strip unit size to the stripe width.
188 		 */
189 		if (mp->m_swidth && xfs_has_swalloc(mp))
190 			align = mp->m_swidth;
191 		else if (mp->m_dalign)
192 			align = mp->m_dalign;
193 
194 		if (align && XFS_ISIZE(ip) < XFS_FSB_TO_B(mp, align))
195 			align = 0;
196 	}
197 
198 	return align;
199 }
200 
201 /*
202  * Check if last_fsb is outside the last extent, and if so grow it to the next
203  * stripe unit boundary.
204  */
205 xfs_fileoff_t
206 xfs_iomap_eof_align_last_fsb(
207 	struct xfs_inode	*ip,
208 	xfs_fileoff_t		end_fsb)
209 {
210 	struct xfs_ifork	*ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);
211 	xfs_extlen_t		extsz = xfs_get_extsz_hint(ip);
212 	xfs_extlen_t		align = xfs_eof_alignment(ip);
213 	struct xfs_bmbt_irec	irec;
214 	struct xfs_iext_cursor	icur;
215 
216 	ASSERT(!xfs_need_iread_extents(ifp));
217 
218 	/*
219 	 * Always round up the allocation request to the extent hint boundary.
220 	 */
221 	if (extsz) {
222 		if (align)
223 			align = roundup_64(align, extsz);
224 		else
225 			align = extsz;
226 	}
227 
228 	if (align) {
229 		xfs_fileoff_t	aligned_end_fsb = roundup_64(end_fsb, align);
230 
231 		xfs_iext_last(ifp, &icur);
232 		if (!xfs_iext_get_extent(ifp, &icur, &irec) ||
233 		    aligned_end_fsb >= irec.br_startoff + irec.br_blockcount)
234 			return aligned_end_fsb;
235 	}
236 
237 	return end_fsb;
238 }
239 
240 int
241 xfs_iomap_write_direct(
242 	struct xfs_inode	*ip,
243 	xfs_fileoff_t		offset_fsb,
244 	xfs_fileoff_t		count_fsb,
245 	unsigned int		flags,
246 	struct xfs_bmbt_irec	*imap,
247 	u64			*seq)
248 {
249 	struct xfs_mount	*mp = ip->i_mount;
250 	struct xfs_trans	*tp;
251 	xfs_filblks_t		resaligned;
252 	int			nimaps;
253 	unsigned int		dblocks, rblocks;
254 	bool			force = false;
255 	int			error;
256 	int			bmapi_flags = XFS_BMAPI_PREALLOC;
257 	int			nr_exts = XFS_IEXT_ADD_NOSPLIT_CNT;
258 
259 	ASSERT(count_fsb > 0);
260 
261 	resaligned = xfs_aligned_fsb_count(offset_fsb, count_fsb,
262 					   xfs_get_extsz_hint(ip));
263 	if (unlikely(XFS_IS_REALTIME_INODE(ip))) {
264 		dblocks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
265 		rblocks = resaligned;
266 	} else {
267 		dblocks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
268 		rblocks = 0;
269 	}
270 
271 	error = xfs_qm_dqattach(ip);
272 	if (error)
273 		return error;
274 
275 	/*
276 	 * For DAX, we do not allocate unwritten extents, but instead we zero
277 	 * the block before we commit the transaction.  Ideally we'd like to do
278 	 * this outside the transaction context, but if we commit and then crash
279 	 * we may not have zeroed the blocks and this will be exposed on
280 	 * recovery of the allocation. Hence we must zero before commit.
281 	 *
282 	 * Further, if we are mapping unwritten extents here, we need to zero
283 	 * and convert them to written so that we don't need an unwritten extent
284 	 * callback for DAX. This also means that we need to be able to dip into
285 	 * the reserve block pool for bmbt block allocation if there is no space
286 	 * left but we need to do unwritten extent conversion.
287 	 */
288 	if (flags & IOMAP_DAX) {
289 		bmapi_flags = XFS_BMAPI_CONVERT | XFS_BMAPI_ZERO;
290 		if (imap->br_state == XFS_EXT_UNWRITTEN) {
291 			force = true;
292 			nr_exts = XFS_IEXT_WRITE_UNWRITTEN_CNT;
293 			dblocks = XFS_DIOSTRAT_SPACE_RES(mp, 0) << 1;
294 		}
295 	}
296 
297 	error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write, dblocks,
298 			rblocks, force, &tp);
299 	if (error)
300 		return error;
301 
302 	error = xfs_iext_count_extend(tp, ip, XFS_DATA_FORK, nr_exts);
303 	if (error)
304 		goto out_trans_cancel;
305 
306 	/*
307 	 * From this point onwards we overwrite the imap pointer that the
308 	 * caller gave to us.
309 	 */
310 	nimaps = 1;
311 	error = xfs_bmapi_write(tp, ip, offset_fsb, count_fsb, bmapi_flags, 0,
312 				imap, &nimaps);
313 	if (error)
314 		goto out_trans_cancel;
315 
316 	/*
317 	 * Complete the transaction
318 	 */
319 	error = xfs_trans_commit(tp);
320 	if (error)
321 		goto out_unlock;
322 
323 	if (unlikely(!xfs_valid_startblock(ip, imap->br_startblock))) {
324 		xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
325 		error = xfs_alert_fsblock_zero(ip, imap);
326 	}
327 
328 out_unlock:
329 	*seq = xfs_iomap_inode_sequence(ip, 0);
330 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
331 	return error;
332 
333 out_trans_cancel:
334 	xfs_trans_cancel(tp);
335 	goto out_unlock;
336 }
337 
338 STATIC bool
339 xfs_quota_need_throttle(
340 	struct xfs_inode	*ip,
341 	xfs_dqtype_t		type,
342 	xfs_fsblock_t		alloc_blocks)
343 {
344 	struct xfs_dquot	*dq = xfs_inode_dquot(ip, type);
345 
346 	if (!dq || !xfs_this_quota_on(ip->i_mount, type))
347 		return false;
348 
349 	/* no hi watermark, no throttle */
350 	if (!dq->q_prealloc_hi_wmark)
351 		return false;
352 
353 	/* under the lo watermark, no throttle */
354 	if (dq->q_blk.reserved + alloc_blocks < dq->q_prealloc_lo_wmark)
355 		return false;
356 
357 	return true;
358 }
359 
360 STATIC void
361 xfs_quota_calc_throttle(
362 	struct xfs_inode	*ip,
363 	xfs_dqtype_t		type,
364 	xfs_fsblock_t		*qblocks,
365 	int			*qshift,
366 	int64_t			*qfreesp)
367 {
368 	struct xfs_dquot	*dq = xfs_inode_dquot(ip, type);
369 	int64_t			freesp;
370 	int			shift = 0;
371 
372 	/* no dq, or over hi wmark, squash the prealloc completely */
373 	if (!dq || dq->q_blk.reserved >= dq->q_prealloc_hi_wmark) {
374 		*qblocks = 0;
375 		*qfreesp = 0;
376 		return;
377 	}
378 
379 	freesp = dq->q_prealloc_hi_wmark - dq->q_blk.reserved;
380 	if (freesp < dq->q_low_space[XFS_QLOWSP_5_PCNT]) {
381 		shift = 2;
382 		if (freesp < dq->q_low_space[XFS_QLOWSP_3_PCNT])
383 			shift += 2;
384 		if (freesp < dq->q_low_space[XFS_QLOWSP_1_PCNT])
385 			shift += 2;
386 	}
387 
388 	if (freesp < *qfreesp)
389 		*qfreesp = freesp;
390 
391 	/* only overwrite the throttle values if we are more aggressive */
392 	if ((freesp >> shift) < (*qblocks >> *qshift)) {
393 		*qblocks = freesp;
394 		*qshift = shift;
395 	}
396 }
397 
398 static int64_t
399 xfs_iomap_freesp(
400 	struct percpu_counter	*counter,
401 	uint64_t		low_space[XFS_LOWSP_MAX],
402 	int			*shift)
403 {
404 	int64_t			freesp;
405 
406 	freesp = percpu_counter_read_positive(counter);
407 	if (freesp < low_space[XFS_LOWSP_5_PCNT]) {
408 		*shift = 2;
409 		if (freesp < low_space[XFS_LOWSP_4_PCNT])
410 			(*shift)++;
411 		if (freesp < low_space[XFS_LOWSP_3_PCNT])
412 			(*shift)++;
413 		if (freesp < low_space[XFS_LOWSP_2_PCNT])
414 			(*shift)++;
415 		if (freesp < low_space[XFS_LOWSP_1_PCNT])
416 			(*shift)++;
417 	}
418 	return freesp;
419 }
420 
421 /*
422  * If we don't have a user specified preallocation size, dynamically increase
423  * the preallocation size as the size of the file grows.  Cap the maximum size
424  * at a single extent or less if the filesystem is near full. The closer the
425  * filesystem is to being full, the smaller the maximum preallocation.
426  */
427 STATIC xfs_fsblock_t
428 xfs_iomap_prealloc_size(
429 	struct xfs_inode	*ip,
430 	int			whichfork,
431 	loff_t			offset,
432 	loff_t			count,
433 	struct xfs_iext_cursor	*icur)
434 {
435 	struct xfs_iext_cursor	ncur = *icur;
436 	struct xfs_bmbt_irec	prev, got;
437 	struct xfs_mount	*mp = ip->i_mount;
438 	struct xfs_ifork	*ifp = xfs_ifork_ptr(ip, whichfork);
439 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
440 	int64_t			freesp;
441 	xfs_fsblock_t		qblocks;
442 	xfs_fsblock_t		alloc_blocks = 0;
443 	xfs_extlen_t		plen;
444 	int			shift = 0;
445 	int			qshift = 0;
446 
447 	/*
448 	 * As an exception we don't do any preallocation at all if the file is
449 	 * smaller than the minimum preallocation and we are using the default
450 	 * dynamic preallocation scheme, as it is likely this is the only write
451 	 * to the file that is going to be done.
452 	 */
453 	if (XFS_ISIZE(ip) < XFS_FSB_TO_B(mp, mp->m_allocsize_blocks))
454 		return 0;
455 
456 	/*
457 	 * Use the minimum preallocation size for small files or if we are
458 	 * writing right after a hole.
459 	 */
460 	if (XFS_ISIZE(ip) < XFS_FSB_TO_B(mp, mp->m_dalign) ||
461 	    !xfs_iext_prev_extent(ifp, &ncur, &prev) ||
462 	    prev.br_startoff + prev.br_blockcount < offset_fsb)
463 		return mp->m_allocsize_blocks;
464 
465 	/*
466 	 * Take the size of the preceding data extents as the basis for the
467 	 * preallocation size. Note that we don't care if the previous extents
468 	 * are written or not.
469 	 */
470 	plen = prev.br_blockcount;
471 	while (xfs_iext_prev_extent(ifp, &ncur, &got)) {
472 		if (plen > XFS_MAX_BMBT_EXTLEN / 2 ||
473 		    isnullstartblock(got.br_startblock) ||
474 		    got.br_startoff + got.br_blockcount != prev.br_startoff ||
475 		    got.br_startblock + got.br_blockcount != prev.br_startblock)
476 			break;
477 		plen += got.br_blockcount;
478 		prev = got;
479 	}
480 
481 	/*
482 	 * If the size of the extents is greater than half the maximum extent
483 	 * length, then use the current offset as the basis.  This ensures that
484 	 * for large files the preallocation size always extends to
485 	 * XFS_BMBT_MAX_EXTLEN rather than falling short due to things like stripe
486 	 * unit/width alignment of real extents.
487 	 */
488 	alloc_blocks = plen * 2;
489 	if (alloc_blocks > XFS_MAX_BMBT_EXTLEN)
490 		alloc_blocks = XFS_B_TO_FSB(mp, offset);
491 	qblocks = alloc_blocks;
492 
493 	/*
494 	 * XFS_BMBT_MAX_EXTLEN is not a power of two value but we round the prealloc
495 	 * down to the nearest power of two value after throttling. To prevent
496 	 * the round down from unconditionally reducing the maximum supported
497 	 * prealloc size, we round up first, apply appropriate throttling, round
498 	 * down and cap the value to XFS_BMBT_MAX_EXTLEN.
499 	 */
500 	alloc_blocks = XFS_FILEOFF_MIN(roundup_pow_of_two(XFS_MAX_BMBT_EXTLEN),
501 				       alloc_blocks);
502 
503 	if (unlikely(XFS_IS_REALTIME_INODE(ip)))
504 		freesp = xfs_rtx_to_rtb(mp,
505 			xfs_iomap_freesp(&mp->m_frextents,
506 					mp->m_low_rtexts, &shift));
507 	else
508 		freesp = xfs_iomap_freesp(&mp->m_fdblocks, mp->m_low_space,
509 				&shift);
510 
511 	/*
512 	 * Check each quota to cap the prealloc size, provide a shift value to
513 	 * throttle with and adjust amount of available space.
514 	 */
515 	if (xfs_quota_need_throttle(ip, XFS_DQTYPE_USER, alloc_blocks))
516 		xfs_quota_calc_throttle(ip, XFS_DQTYPE_USER, &qblocks, &qshift,
517 					&freesp);
518 	if (xfs_quota_need_throttle(ip, XFS_DQTYPE_GROUP, alloc_blocks))
519 		xfs_quota_calc_throttle(ip, XFS_DQTYPE_GROUP, &qblocks, &qshift,
520 					&freesp);
521 	if (xfs_quota_need_throttle(ip, XFS_DQTYPE_PROJ, alloc_blocks))
522 		xfs_quota_calc_throttle(ip, XFS_DQTYPE_PROJ, &qblocks, &qshift,
523 					&freesp);
524 
525 	/*
526 	 * The final prealloc size is set to the minimum of free space available
527 	 * in each of the quotas and the overall filesystem.
528 	 *
529 	 * The shift throttle value is set to the maximum value as determined by
530 	 * the global low free space values and per-quota low free space values.
531 	 */
532 	alloc_blocks = min(alloc_blocks, qblocks);
533 	shift = max(shift, qshift);
534 
535 	if (shift)
536 		alloc_blocks >>= shift;
537 	/*
538 	 * rounddown_pow_of_two() returns an undefined result if we pass in
539 	 * alloc_blocks = 0.
540 	 */
541 	if (alloc_blocks)
542 		alloc_blocks = rounddown_pow_of_two(alloc_blocks);
543 	if (alloc_blocks > XFS_MAX_BMBT_EXTLEN)
544 		alloc_blocks = XFS_MAX_BMBT_EXTLEN;
545 
546 	/*
547 	 * If we are still trying to allocate more space than is
548 	 * available, squash the prealloc hard. This can happen if we
549 	 * have a large file on a small filesystem and the above
550 	 * lowspace thresholds are smaller than XFS_BMBT_MAX_EXTLEN.
551 	 */
552 	while (alloc_blocks && alloc_blocks >= freesp)
553 		alloc_blocks >>= 4;
554 	if (alloc_blocks < mp->m_allocsize_blocks)
555 		alloc_blocks = mp->m_allocsize_blocks;
556 	trace_xfs_iomap_prealloc_size(ip, alloc_blocks, shift,
557 				      mp->m_allocsize_blocks);
558 	return alloc_blocks;
559 }
560 
561 int
562 xfs_iomap_write_unwritten(
563 	xfs_inode_t	*ip,
564 	xfs_off_t	offset,
565 	xfs_off_t	count,
566 	bool		update_isize)
567 {
568 	xfs_mount_t	*mp = ip->i_mount;
569 	xfs_fileoff_t	offset_fsb;
570 	xfs_filblks_t	count_fsb;
571 	xfs_filblks_t	numblks_fsb;
572 	int		nimaps;
573 	xfs_trans_t	*tp;
574 	xfs_bmbt_irec_t imap;
575 	struct inode	*inode = VFS_I(ip);
576 	xfs_fsize_t	i_size;
577 	uint		resblks;
578 	int		error;
579 
580 	trace_xfs_unwritten_convert(ip, offset, count);
581 
582 	offset_fsb = XFS_B_TO_FSBT(mp, offset);
583 	count_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
584 	count_fsb = (xfs_filblks_t)(count_fsb - offset_fsb);
585 
586 	/*
587 	 * Reserve enough blocks in this transaction for two complete extent
588 	 * btree splits.  We may be converting the middle part of an unwritten
589 	 * extent and in this case we will insert two new extents in the btree
590 	 * each of which could cause a full split.
591 	 *
592 	 * This reservation amount will be used in the first call to
593 	 * xfs_bmbt_split() to select an AG with enough space to satisfy the
594 	 * rest of the operation.
595 	 */
596 	resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0) << 1;
597 
598 	/* Attach dquots so that bmbt splits are accounted correctly. */
599 	error = xfs_qm_dqattach(ip);
600 	if (error)
601 		return error;
602 
603 	do {
604 		/*
605 		 * Set up a transaction to convert the range of extents
606 		 * from unwritten to real. Do allocations in a loop until
607 		 * we have covered the range passed in.
608 		 *
609 		 * Note that we can't risk to recursing back into the filesystem
610 		 * here as we might be asked to write out the same inode that we
611 		 * complete here and might deadlock on the iolock.
612 		 */
613 		error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write, resblks,
614 				0, true, &tp);
615 		if (error)
616 			return error;
617 
618 		error = xfs_iext_count_extend(tp, ip, XFS_DATA_FORK,
619 				XFS_IEXT_WRITE_UNWRITTEN_CNT);
620 		if (error)
621 			goto error_on_bmapi_transaction;
622 
623 		/*
624 		 * Modify the unwritten extent state of the buffer.
625 		 */
626 		nimaps = 1;
627 		error = xfs_bmapi_write(tp, ip, offset_fsb, count_fsb,
628 					XFS_BMAPI_CONVERT, resblks, &imap,
629 					&nimaps);
630 		if (error)
631 			goto error_on_bmapi_transaction;
632 
633 		/*
634 		 * Log the updated inode size as we go.  We have to be careful
635 		 * to only log it up to the actual write offset if it is
636 		 * halfway into a block.
637 		 */
638 		i_size = XFS_FSB_TO_B(mp, offset_fsb + count_fsb);
639 		if (i_size > offset + count)
640 			i_size = offset + count;
641 		if (update_isize && i_size > i_size_read(inode))
642 			i_size_write(inode, i_size);
643 		i_size = xfs_new_eof(ip, i_size);
644 		if (i_size) {
645 			ip->i_disk_size = i_size;
646 			xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
647 		}
648 
649 		error = xfs_trans_commit(tp);
650 		xfs_iunlock(ip, XFS_ILOCK_EXCL);
651 		if (error)
652 			return error;
653 
654 		if (unlikely(!xfs_valid_startblock(ip, imap.br_startblock))) {
655 			xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
656 			return xfs_alert_fsblock_zero(ip, &imap);
657 		}
658 
659 		if ((numblks_fsb = imap.br_blockcount) == 0) {
660 			/*
661 			 * The numblks_fsb value should always get
662 			 * smaller, otherwise the loop is stuck.
663 			 */
664 			ASSERT(imap.br_blockcount);
665 			break;
666 		}
667 		offset_fsb += numblks_fsb;
668 		count_fsb -= numblks_fsb;
669 	} while (count_fsb > 0);
670 
671 	return 0;
672 
673 error_on_bmapi_transaction:
674 	xfs_trans_cancel(tp);
675 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
676 	return error;
677 }
678 
679 static inline bool
680 imap_needs_alloc(
681 	struct inode		*inode,
682 	unsigned		flags,
683 	struct xfs_bmbt_irec	*imap,
684 	int			nimaps)
685 {
686 	/* don't allocate blocks when just zeroing */
687 	if (flags & IOMAP_ZERO)
688 		return false;
689 	if (!nimaps ||
690 	    imap->br_startblock == HOLESTARTBLOCK ||
691 	    imap->br_startblock == DELAYSTARTBLOCK)
692 		return true;
693 	/* we convert unwritten extents before copying the data for DAX */
694 	if ((flags & IOMAP_DAX) && imap->br_state == XFS_EXT_UNWRITTEN)
695 		return true;
696 	return false;
697 }
698 
699 static inline bool
700 imap_needs_cow(
701 	struct xfs_inode	*ip,
702 	unsigned int		flags,
703 	struct xfs_bmbt_irec	*imap,
704 	int			nimaps)
705 {
706 	if (!xfs_is_cow_inode(ip))
707 		return false;
708 
709 	/* when zeroing we don't have to COW holes or unwritten extents */
710 	if (flags & IOMAP_ZERO) {
711 		if (!nimaps ||
712 		    imap->br_startblock == HOLESTARTBLOCK ||
713 		    imap->br_state == XFS_EXT_UNWRITTEN)
714 			return false;
715 	}
716 
717 	return true;
718 }
719 
720 /*
721  * Extents not yet cached requires exclusive access, don't block for
722  * IOMAP_NOWAIT.
723  *
724  * This is basically an opencoded xfs_ilock_data_map_shared() call, but with
725  * support for IOMAP_NOWAIT.
726  */
727 static int
728 xfs_ilock_for_iomap(
729 	struct xfs_inode	*ip,
730 	unsigned		flags,
731 	unsigned		*lockmode)
732 {
733 	if (flags & IOMAP_NOWAIT) {
734 		if (xfs_need_iread_extents(&ip->i_df))
735 			return -EAGAIN;
736 		if (!xfs_ilock_nowait(ip, *lockmode))
737 			return -EAGAIN;
738 	} else {
739 		if (xfs_need_iread_extents(&ip->i_df))
740 			*lockmode = XFS_ILOCK_EXCL;
741 		xfs_ilock(ip, *lockmode);
742 	}
743 
744 	return 0;
745 }
746 
747 /*
748  * Check that the imap we are going to return to the caller spans the entire
749  * range that the caller requested for the IO.
750  */
751 static bool
752 imap_spans_range(
753 	struct xfs_bmbt_irec	*imap,
754 	xfs_fileoff_t		offset_fsb,
755 	xfs_fileoff_t		end_fsb)
756 {
757 	if (imap->br_startoff > offset_fsb)
758 		return false;
759 	if (imap->br_startoff + imap->br_blockcount < end_fsb)
760 		return false;
761 	return true;
762 }
763 
764 static int
765 xfs_direct_write_iomap_begin(
766 	struct inode		*inode,
767 	loff_t			offset,
768 	loff_t			length,
769 	unsigned		flags,
770 	struct iomap		*iomap,
771 	struct iomap		*srcmap)
772 {
773 	struct xfs_inode	*ip = XFS_I(inode);
774 	struct xfs_mount	*mp = ip->i_mount;
775 	struct xfs_bmbt_irec	imap, cmap;
776 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
777 	xfs_fileoff_t		end_fsb = xfs_iomap_end_fsb(mp, offset, length);
778 	int			nimaps = 1, error = 0;
779 	bool			shared = false;
780 	u16			iomap_flags = 0;
781 	unsigned int		lockmode;
782 	u64			seq;
783 
784 	ASSERT(flags & (IOMAP_WRITE | IOMAP_ZERO));
785 
786 	if (xfs_is_shutdown(mp))
787 		return -EIO;
788 
789 	/*
790 	 * Writes that span EOF might trigger an IO size update on completion,
791 	 * so consider them to be dirty for the purposes of O_DSYNC even if
792 	 * there is no other metadata changes pending or have been made here.
793 	 */
794 	if (offset + length > i_size_read(inode))
795 		iomap_flags |= IOMAP_F_DIRTY;
796 
797 	/*
798 	 * COW writes may allocate delalloc space or convert unwritten COW
799 	 * extents, so we need to make sure to take the lock exclusively here.
800 	 */
801 	if (xfs_is_cow_inode(ip))
802 		lockmode = XFS_ILOCK_EXCL;
803 	else
804 		lockmode = XFS_ILOCK_SHARED;
805 
806 relock:
807 	error = xfs_ilock_for_iomap(ip, flags, &lockmode);
808 	if (error)
809 		return error;
810 
811 	/*
812 	 * The reflink iflag could have changed since the earlier unlocked
813 	 * check, check if it again and relock if needed.
814 	 */
815 	if (xfs_is_cow_inode(ip) && lockmode == XFS_ILOCK_SHARED) {
816 		xfs_iunlock(ip, lockmode);
817 		lockmode = XFS_ILOCK_EXCL;
818 		goto relock;
819 	}
820 
821 	error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, &imap,
822 			       &nimaps, 0);
823 	if (error)
824 		goto out_unlock;
825 
826 	if (imap_needs_cow(ip, flags, &imap, nimaps)) {
827 		error = -EAGAIN;
828 		if (flags & IOMAP_NOWAIT)
829 			goto out_unlock;
830 
831 		/* may drop and re-acquire the ilock */
832 		error = xfs_reflink_allocate_cow(ip, &imap, &cmap, &shared,
833 				&lockmode,
834 				(flags & IOMAP_DIRECT) || IS_DAX(inode));
835 		if (error)
836 			goto out_unlock;
837 		if (shared)
838 			goto out_found_cow;
839 		end_fsb = imap.br_startoff + imap.br_blockcount;
840 		length = XFS_FSB_TO_B(mp, end_fsb) - offset;
841 	}
842 
843 	if (imap_needs_alloc(inode, flags, &imap, nimaps))
844 		goto allocate_blocks;
845 
846 	/*
847 	 * NOWAIT and OVERWRITE I/O needs to span the entire requested I/O with
848 	 * a single map so that we avoid partial IO failures due to the rest of
849 	 * the I/O range not covered by this map triggering an EAGAIN condition
850 	 * when it is subsequently mapped and aborting the I/O.
851 	 */
852 	if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY)) {
853 		error = -EAGAIN;
854 		if (!imap_spans_range(&imap, offset_fsb, end_fsb))
855 			goto out_unlock;
856 	}
857 
858 	/*
859 	 * For overwrite only I/O, we cannot convert unwritten extents without
860 	 * requiring sub-block zeroing.  This can only be done under an
861 	 * exclusive IOLOCK, hence return -EAGAIN if this is not a written
862 	 * extent to tell the caller to try again.
863 	 */
864 	if (flags & IOMAP_OVERWRITE_ONLY) {
865 		error = -EAGAIN;
866 		if (imap.br_state != XFS_EXT_NORM &&
867 	            ((offset | length) & mp->m_blockmask))
868 			goto out_unlock;
869 	}
870 
871 	seq = xfs_iomap_inode_sequence(ip, iomap_flags);
872 	xfs_iunlock(ip, lockmode);
873 	trace_xfs_iomap_found(ip, offset, length, XFS_DATA_FORK, &imap);
874 	return xfs_bmbt_to_iomap(ip, iomap, &imap, flags, iomap_flags, seq);
875 
876 allocate_blocks:
877 	error = -EAGAIN;
878 	if (flags & (IOMAP_NOWAIT | IOMAP_OVERWRITE_ONLY))
879 		goto out_unlock;
880 
881 	/*
882 	 * We cap the maximum length we map to a sane size  to keep the chunks
883 	 * of work done where somewhat symmetric with the work writeback does.
884 	 * This is a completely arbitrary number pulled out of thin air as a
885 	 * best guess for initial testing.
886 	 *
887 	 * Note that the values needs to be less than 32-bits wide until the
888 	 * lower level functions are updated.
889 	 */
890 	length = min_t(loff_t, length, 1024 * PAGE_SIZE);
891 	end_fsb = xfs_iomap_end_fsb(mp, offset, length);
892 
893 	if (offset + length > XFS_ISIZE(ip))
894 		end_fsb = xfs_iomap_eof_align_last_fsb(ip, end_fsb);
895 	else if (nimaps && imap.br_startblock == HOLESTARTBLOCK)
896 		end_fsb = min(end_fsb, imap.br_startoff + imap.br_blockcount);
897 	xfs_iunlock(ip, lockmode);
898 
899 	error = xfs_iomap_write_direct(ip, offset_fsb, end_fsb - offset_fsb,
900 			flags, &imap, &seq);
901 	if (error)
902 		return error;
903 
904 	trace_xfs_iomap_alloc(ip, offset, length, XFS_DATA_FORK, &imap);
905 	return xfs_bmbt_to_iomap(ip, iomap, &imap, flags,
906 				 iomap_flags | IOMAP_F_NEW, seq);
907 
908 out_found_cow:
909 	length = XFS_FSB_TO_B(mp, cmap.br_startoff + cmap.br_blockcount);
910 	trace_xfs_iomap_found(ip, offset, length - offset, XFS_COW_FORK, &cmap);
911 	if (imap.br_startblock != HOLESTARTBLOCK) {
912 		seq = xfs_iomap_inode_sequence(ip, 0);
913 		error = xfs_bmbt_to_iomap(ip, srcmap, &imap, flags, 0, seq);
914 		if (error)
915 			goto out_unlock;
916 	}
917 	seq = xfs_iomap_inode_sequence(ip, IOMAP_F_SHARED);
918 	xfs_iunlock(ip, lockmode);
919 	return xfs_bmbt_to_iomap(ip, iomap, &cmap, flags, IOMAP_F_SHARED, seq);
920 
921 out_unlock:
922 	if (lockmode)
923 		xfs_iunlock(ip, lockmode);
924 	return error;
925 }
926 
927 const struct iomap_ops xfs_direct_write_iomap_ops = {
928 	.iomap_begin		= xfs_direct_write_iomap_begin,
929 };
930 
931 static int
932 xfs_dax_write_iomap_end(
933 	struct inode		*inode,
934 	loff_t			pos,
935 	loff_t			length,
936 	ssize_t			written,
937 	unsigned		flags,
938 	struct iomap		*iomap)
939 {
940 	struct xfs_inode	*ip = XFS_I(inode);
941 
942 	if (!xfs_is_cow_inode(ip))
943 		return 0;
944 
945 	if (!written) {
946 		xfs_reflink_cancel_cow_range(ip, pos, length, true);
947 		return 0;
948 	}
949 
950 	return xfs_reflink_end_cow(ip, pos, written);
951 }
952 
953 const struct iomap_ops xfs_dax_write_iomap_ops = {
954 	.iomap_begin	= xfs_direct_write_iomap_begin,
955 	.iomap_end	= xfs_dax_write_iomap_end,
956 };
957 
958 static int
959 xfs_buffered_write_iomap_begin(
960 	struct inode		*inode,
961 	loff_t			offset,
962 	loff_t			count,
963 	unsigned		flags,
964 	struct iomap		*iomap,
965 	struct iomap		*srcmap)
966 {
967 	struct xfs_inode	*ip = XFS_I(inode);
968 	struct xfs_mount	*mp = ip->i_mount;
969 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
970 	xfs_fileoff_t		end_fsb = xfs_iomap_end_fsb(mp, offset, count);
971 	struct xfs_bmbt_irec	imap, cmap;
972 	struct xfs_iext_cursor	icur, ccur;
973 	xfs_fsblock_t		prealloc_blocks = 0;
974 	bool			eof = false, cow_eof = false, shared = false;
975 	int			allocfork = XFS_DATA_FORK;
976 	int			error = 0;
977 	unsigned int		lockmode = XFS_ILOCK_EXCL;
978 	u64			seq;
979 
980 	if (xfs_is_shutdown(mp))
981 		return -EIO;
982 
983 	/* we can't use delayed allocations when using extent size hints */
984 	if (xfs_get_extsz_hint(ip))
985 		return xfs_direct_write_iomap_begin(inode, offset, count,
986 				flags, iomap, srcmap);
987 
988 	error = xfs_qm_dqattach(ip);
989 	if (error)
990 		return error;
991 
992 	error = xfs_ilock_for_iomap(ip, flags, &lockmode);
993 	if (error)
994 		return error;
995 
996 	if (XFS_IS_CORRUPT(mp, !xfs_ifork_has_extents(&ip->i_df)) ||
997 	    XFS_TEST_ERROR(false, mp, XFS_ERRTAG_BMAPIFORMAT)) {
998 		xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
999 		error = -EFSCORRUPTED;
1000 		goto out_unlock;
1001 	}
1002 
1003 	XFS_STATS_INC(mp, xs_blk_mapw);
1004 
1005 	error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK);
1006 	if (error)
1007 		goto out_unlock;
1008 
1009 	/*
1010 	 * Search the data fork first to look up our source mapping.  We
1011 	 * always need the data fork map, as we have to return it to the
1012 	 * iomap code so that the higher level write code can read data in to
1013 	 * perform read-modify-write cycles for unaligned writes.
1014 	 */
1015 	eof = !xfs_iext_lookup_extent(ip, &ip->i_df, offset_fsb, &icur, &imap);
1016 	if (eof)
1017 		imap.br_startoff = end_fsb; /* fake hole until the end */
1018 
1019 	/* We never need to allocate blocks for zeroing or unsharing a hole. */
1020 	if ((flags & (IOMAP_UNSHARE | IOMAP_ZERO)) &&
1021 	    imap.br_startoff > offset_fsb) {
1022 		xfs_hole_to_iomap(ip, iomap, offset_fsb, imap.br_startoff);
1023 		goto out_unlock;
1024 	}
1025 
1026 	/*
1027 	 * For zeroing, trim a delalloc extent that extends beyond the EOF
1028 	 * block.  If it starts beyond the EOF block, convert it to an
1029 	 * unwritten extent.
1030 	 */
1031 	if ((flags & IOMAP_ZERO) && imap.br_startoff <= offset_fsb &&
1032 	    isnullstartblock(imap.br_startblock)) {
1033 		xfs_fileoff_t eof_fsb = XFS_B_TO_FSB(mp, XFS_ISIZE(ip));
1034 
1035 		if (offset_fsb >= eof_fsb)
1036 			goto convert_delay;
1037 		if (end_fsb > eof_fsb) {
1038 			end_fsb = eof_fsb;
1039 			xfs_trim_extent(&imap, offset_fsb,
1040 					end_fsb - offset_fsb);
1041 		}
1042 	}
1043 
1044 	/*
1045 	 * Search the COW fork extent list even if we did not find a data fork
1046 	 * extent.  This serves two purposes: first this implements the
1047 	 * speculative preallocation using cowextsize, so that we also unshare
1048 	 * block adjacent to shared blocks instead of just the shared blocks
1049 	 * themselves.  Second the lookup in the extent list is generally faster
1050 	 * than going out to the shared extent tree.
1051 	 */
1052 	if (xfs_is_cow_inode(ip)) {
1053 		if (!ip->i_cowfp) {
1054 			ASSERT(!xfs_is_reflink_inode(ip));
1055 			xfs_ifork_init_cow(ip);
1056 		}
1057 		cow_eof = !xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb,
1058 				&ccur, &cmap);
1059 		if (!cow_eof && cmap.br_startoff <= offset_fsb) {
1060 			trace_xfs_reflink_cow_found(ip, &cmap);
1061 			goto found_cow;
1062 		}
1063 	}
1064 
1065 	if (imap.br_startoff <= offset_fsb) {
1066 		/*
1067 		 * For reflink files we may need a delalloc reservation when
1068 		 * overwriting shared extents.   This includes zeroing of
1069 		 * existing extents that contain data.
1070 		 */
1071 		if (!xfs_is_cow_inode(ip) ||
1072 		    ((flags & IOMAP_ZERO) && imap.br_state != XFS_EXT_NORM)) {
1073 			trace_xfs_iomap_found(ip, offset, count, XFS_DATA_FORK,
1074 					&imap);
1075 			goto found_imap;
1076 		}
1077 
1078 		xfs_trim_extent(&imap, offset_fsb, end_fsb - offset_fsb);
1079 
1080 		/* Trim the mapping to the nearest shared extent boundary. */
1081 		error = xfs_bmap_trim_cow(ip, &imap, &shared);
1082 		if (error)
1083 			goto out_unlock;
1084 
1085 		/* Not shared?  Just report the (potentially capped) extent. */
1086 		if (!shared) {
1087 			trace_xfs_iomap_found(ip, offset, count, XFS_DATA_FORK,
1088 					&imap);
1089 			goto found_imap;
1090 		}
1091 
1092 		/*
1093 		 * Fork all the shared blocks from our write offset until the
1094 		 * end of the extent.
1095 		 */
1096 		allocfork = XFS_COW_FORK;
1097 		end_fsb = imap.br_startoff + imap.br_blockcount;
1098 	} else {
1099 		/*
1100 		 * We cap the maximum length we map here to MAX_WRITEBACK_PAGES
1101 		 * pages to keep the chunks of work done where somewhat
1102 		 * symmetric with the work writeback does.  This is a completely
1103 		 * arbitrary number pulled out of thin air.
1104 		 *
1105 		 * Note that the values needs to be less than 32-bits wide until
1106 		 * the lower level functions are updated.
1107 		 */
1108 		count = min_t(loff_t, count, 1024 * PAGE_SIZE);
1109 		end_fsb = xfs_iomap_end_fsb(mp, offset, count);
1110 
1111 		if (xfs_is_always_cow_inode(ip))
1112 			allocfork = XFS_COW_FORK;
1113 	}
1114 
1115 	if (eof && offset + count > XFS_ISIZE(ip)) {
1116 		/*
1117 		 * Determine the initial size of the preallocation.
1118 		 * We clean up any extra preallocation when the file is closed.
1119 		 */
1120 		if (xfs_has_allocsize(mp))
1121 			prealloc_blocks = mp->m_allocsize_blocks;
1122 		else if (allocfork == XFS_DATA_FORK)
1123 			prealloc_blocks = xfs_iomap_prealloc_size(ip, allocfork,
1124 						offset, count, &icur);
1125 		else
1126 			prealloc_blocks = xfs_iomap_prealloc_size(ip, allocfork,
1127 						offset, count, &ccur);
1128 		if (prealloc_blocks) {
1129 			xfs_extlen_t	align;
1130 			xfs_off_t	end_offset;
1131 			xfs_fileoff_t	p_end_fsb;
1132 
1133 			end_offset = XFS_ALLOC_ALIGN(mp, offset + count - 1);
1134 			p_end_fsb = XFS_B_TO_FSBT(mp, end_offset) +
1135 					prealloc_blocks;
1136 
1137 			align = xfs_eof_alignment(ip);
1138 			if (align)
1139 				p_end_fsb = roundup_64(p_end_fsb, align);
1140 
1141 			p_end_fsb = min(p_end_fsb,
1142 				XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes));
1143 			ASSERT(p_end_fsb > offset_fsb);
1144 			prealloc_blocks = p_end_fsb - end_fsb;
1145 		}
1146 	}
1147 
1148 	if (allocfork == XFS_COW_FORK) {
1149 		error = xfs_bmapi_reserve_delalloc(ip, allocfork, offset_fsb,
1150 				end_fsb - offset_fsb, prealloc_blocks, &cmap,
1151 				&ccur, cow_eof);
1152 		if (error)
1153 			goto out_unlock;
1154 
1155 		trace_xfs_iomap_alloc(ip, offset, count, allocfork, &cmap);
1156 		goto found_cow;
1157 	}
1158 
1159 	error = xfs_bmapi_reserve_delalloc(ip, allocfork, offset_fsb,
1160 			end_fsb - offset_fsb, prealloc_blocks, &imap, &icur,
1161 			eof);
1162 	if (error)
1163 		goto out_unlock;
1164 
1165 	/*
1166 	 * Flag newly allocated delalloc blocks with IOMAP_F_NEW so we punch
1167 	 * them out if the write happens to fail.
1168 	 */
1169 	seq = xfs_iomap_inode_sequence(ip, IOMAP_F_NEW);
1170 	xfs_iunlock(ip, lockmode);
1171 	trace_xfs_iomap_alloc(ip, offset, count, allocfork, &imap);
1172 	return xfs_bmbt_to_iomap(ip, iomap, &imap, flags, IOMAP_F_NEW, seq);
1173 
1174 found_imap:
1175 	seq = xfs_iomap_inode_sequence(ip, 0);
1176 	xfs_iunlock(ip, lockmode);
1177 	return xfs_bmbt_to_iomap(ip, iomap, &imap, flags, 0, seq);
1178 
1179 convert_delay:
1180 	xfs_iunlock(ip, lockmode);
1181 	truncate_pagecache(inode, offset);
1182 	error = xfs_bmapi_convert_delalloc(ip, XFS_DATA_FORK, offset,
1183 					   iomap, NULL);
1184 	if (error)
1185 		return error;
1186 
1187 	trace_xfs_iomap_alloc(ip, offset, count, XFS_DATA_FORK, &imap);
1188 	return 0;
1189 
1190 found_cow:
1191 	seq = xfs_iomap_inode_sequence(ip, 0);
1192 	if (imap.br_startoff <= offset_fsb) {
1193 		error = xfs_bmbt_to_iomap(ip, srcmap, &imap, flags, 0, seq);
1194 		if (error)
1195 			goto out_unlock;
1196 		seq = xfs_iomap_inode_sequence(ip, IOMAP_F_SHARED);
1197 		xfs_iunlock(ip, lockmode);
1198 		return xfs_bmbt_to_iomap(ip, iomap, &cmap, flags,
1199 					 IOMAP_F_SHARED, seq);
1200 	}
1201 
1202 	xfs_trim_extent(&cmap, offset_fsb, imap.br_startoff - offset_fsb);
1203 	xfs_iunlock(ip, lockmode);
1204 	return xfs_bmbt_to_iomap(ip, iomap, &cmap, flags, 0, seq);
1205 
1206 out_unlock:
1207 	xfs_iunlock(ip, lockmode);
1208 	return error;
1209 }
1210 
1211 static int
1212 xfs_buffered_write_delalloc_punch(
1213 	struct inode		*inode,
1214 	loff_t			offset,
1215 	loff_t			length)
1216 {
1217 	xfs_bmap_punch_delalloc_range(XFS_I(inode), offset, offset + length);
1218 	return 0;
1219 }
1220 
1221 static int
1222 xfs_buffered_write_iomap_end(
1223 	struct inode		*inode,
1224 	loff_t			offset,
1225 	loff_t			length,
1226 	ssize_t			written,
1227 	unsigned		flags,
1228 	struct iomap		*iomap)
1229 {
1230 
1231 	struct xfs_mount	*mp = XFS_M(inode->i_sb);
1232 	int			error;
1233 
1234 	error = iomap_file_buffered_write_punch_delalloc(inode, iomap, offset,
1235 			length, written, &xfs_buffered_write_delalloc_punch);
1236 	if (error && !xfs_is_shutdown(mp)) {
1237 		xfs_alert(mp, "%s: unable to clean up ino 0x%llx",
1238 			__func__, XFS_I(inode)->i_ino);
1239 		return error;
1240 	}
1241 	return 0;
1242 }
1243 
1244 const struct iomap_ops xfs_buffered_write_iomap_ops = {
1245 	.iomap_begin		= xfs_buffered_write_iomap_begin,
1246 	.iomap_end		= xfs_buffered_write_iomap_end,
1247 };
1248 
1249 /*
1250  * iomap_page_mkwrite() will never fail in a way that requires delalloc extents
1251  * that it allocated to be revoked. Hence we do not need an .iomap_end method
1252  * for this operation.
1253  */
1254 const struct iomap_ops xfs_page_mkwrite_iomap_ops = {
1255 	.iomap_begin		= xfs_buffered_write_iomap_begin,
1256 };
1257 
1258 static int
1259 xfs_read_iomap_begin(
1260 	struct inode		*inode,
1261 	loff_t			offset,
1262 	loff_t			length,
1263 	unsigned		flags,
1264 	struct iomap		*iomap,
1265 	struct iomap		*srcmap)
1266 {
1267 	struct xfs_inode	*ip = XFS_I(inode);
1268 	struct xfs_mount	*mp = ip->i_mount;
1269 	struct xfs_bmbt_irec	imap;
1270 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
1271 	xfs_fileoff_t		end_fsb = xfs_iomap_end_fsb(mp, offset, length);
1272 	int			nimaps = 1, error = 0;
1273 	bool			shared = false;
1274 	unsigned int		lockmode = XFS_ILOCK_SHARED;
1275 	u64			seq;
1276 
1277 	ASSERT(!(flags & (IOMAP_WRITE | IOMAP_ZERO)));
1278 
1279 	if (xfs_is_shutdown(mp))
1280 		return -EIO;
1281 
1282 	error = xfs_ilock_for_iomap(ip, flags, &lockmode);
1283 	if (error)
1284 		return error;
1285 	error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, &imap,
1286 			       &nimaps, 0);
1287 	if (!error && ((flags & IOMAP_REPORT) || IS_DAX(inode)))
1288 		error = xfs_reflink_trim_around_shared(ip, &imap, &shared);
1289 	seq = xfs_iomap_inode_sequence(ip, shared ? IOMAP_F_SHARED : 0);
1290 	xfs_iunlock(ip, lockmode);
1291 
1292 	if (error)
1293 		return error;
1294 	trace_xfs_iomap_found(ip, offset, length, XFS_DATA_FORK, &imap);
1295 	return xfs_bmbt_to_iomap(ip, iomap, &imap, flags,
1296 				 shared ? IOMAP_F_SHARED : 0, seq);
1297 }
1298 
1299 const struct iomap_ops xfs_read_iomap_ops = {
1300 	.iomap_begin		= xfs_read_iomap_begin,
1301 };
1302 
1303 static int
1304 xfs_seek_iomap_begin(
1305 	struct inode		*inode,
1306 	loff_t			offset,
1307 	loff_t			length,
1308 	unsigned		flags,
1309 	struct iomap		*iomap,
1310 	struct iomap		*srcmap)
1311 {
1312 	struct xfs_inode	*ip = XFS_I(inode);
1313 	struct xfs_mount	*mp = ip->i_mount;
1314 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
1315 	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + length);
1316 	xfs_fileoff_t		cow_fsb = NULLFILEOFF, data_fsb = NULLFILEOFF;
1317 	struct xfs_iext_cursor	icur;
1318 	struct xfs_bmbt_irec	imap, cmap;
1319 	int			error = 0;
1320 	unsigned		lockmode;
1321 	u64			seq;
1322 
1323 	if (xfs_is_shutdown(mp))
1324 		return -EIO;
1325 
1326 	lockmode = xfs_ilock_data_map_shared(ip);
1327 	error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK);
1328 	if (error)
1329 		goto out_unlock;
1330 
1331 	if (xfs_iext_lookup_extent(ip, &ip->i_df, offset_fsb, &icur, &imap)) {
1332 		/*
1333 		 * If we found a data extent we are done.
1334 		 */
1335 		if (imap.br_startoff <= offset_fsb)
1336 			goto done;
1337 		data_fsb = imap.br_startoff;
1338 	} else {
1339 		/*
1340 		 * Fake a hole until the end of the file.
1341 		 */
1342 		data_fsb = xfs_iomap_end_fsb(mp, offset, length);
1343 	}
1344 
1345 	/*
1346 	 * If a COW fork extent covers the hole, report it - capped to the next
1347 	 * data fork extent:
1348 	 */
1349 	if (xfs_inode_has_cow_data(ip) &&
1350 	    xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &cmap))
1351 		cow_fsb = cmap.br_startoff;
1352 	if (cow_fsb != NULLFILEOFF && cow_fsb <= offset_fsb) {
1353 		if (data_fsb < cow_fsb + cmap.br_blockcount)
1354 			end_fsb = min(end_fsb, data_fsb);
1355 		xfs_trim_extent(&cmap, offset_fsb, end_fsb - offset_fsb);
1356 		seq = xfs_iomap_inode_sequence(ip, IOMAP_F_SHARED);
1357 		error = xfs_bmbt_to_iomap(ip, iomap, &cmap, flags,
1358 				IOMAP_F_SHARED, seq);
1359 		/*
1360 		 * This is a COW extent, so we must probe the page cache
1361 		 * because there could be dirty page cache being backed
1362 		 * by this extent.
1363 		 */
1364 		iomap->type = IOMAP_UNWRITTEN;
1365 		goto out_unlock;
1366 	}
1367 
1368 	/*
1369 	 * Else report a hole, capped to the next found data or COW extent.
1370 	 */
1371 	if (cow_fsb != NULLFILEOFF && cow_fsb < data_fsb)
1372 		imap.br_blockcount = cow_fsb - offset_fsb;
1373 	else
1374 		imap.br_blockcount = data_fsb - offset_fsb;
1375 	imap.br_startoff = offset_fsb;
1376 	imap.br_startblock = HOLESTARTBLOCK;
1377 	imap.br_state = XFS_EXT_NORM;
1378 done:
1379 	seq = xfs_iomap_inode_sequence(ip, 0);
1380 	xfs_trim_extent(&imap, offset_fsb, end_fsb - offset_fsb);
1381 	error = xfs_bmbt_to_iomap(ip, iomap, &imap, flags, 0, seq);
1382 out_unlock:
1383 	xfs_iunlock(ip, lockmode);
1384 	return error;
1385 }
1386 
1387 const struct iomap_ops xfs_seek_iomap_ops = {
1388 	.iomap_begin		= xfs_seek_iomap_begin,
1389 };
1390 
1391 static int
1392 xfs_xattr_iomap_begin(
1393 	struct inode		*inode,
1394 	loff_t			offset,
1395 	loff_t			length,
1396 	unsigned		flags,
1397 	struct iomap		*iomap,
1398 	struct iomap		*srcmap)
1399 {
1400 	struct xfs_inode	*ip = XFS_I(inode);
1401 	struct xfs_mount	*mp = ip->i_mount;
1402 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
1403 	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + length);
1404 	struct xfs_bmbt_irec	imap;
1405 	int			nimaps = 1, error = 0;
1406 	unsigned		lockmode;
1407 	int			seq;
1408 
1409 	if (xfs_is_shutdown(mp))
1410 		return -EIO;
1411 
1412 	lockmode = xfs_ilock_attr_map_shared(ip);
1413 
1414 	/* if there are no attribute fork or extents, return ENOENT */
1415 	if (!xfs_inode_has_attr_fork(ip) || !ip->i_af.if_nextents) {
1416 		error = -ENOENT;
1417 		goto out_unlock;
1418 	}
1419 
1420 	ASSERT(ip->i_af.if_format != XFS_DINODE_FMT_LOCAL);
1421 	error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb, &imap,
1422 			       &nimaps, XFS_BMAPI_ATTRFORK);
1423 out_unlock:
1424 
1425 	seq = xfs_iomap_inode_sequence(ip, IOMAP_F_XATTR);
1426 	xfs_iunlock(ip, lockmode);
1427 
1428 	if (error)
1429 		return error;
1430 	ASSERT(nimaps);
1431 	return xfs_bmbt_to_iomap(ip, iomap, &imap, flags, IOMAP_F_XATTR, seq);
1432 }
1433 
1434 const struct iomap_ops xfs_xattr_iomap_ops = {
1435 	.iomap_begin		= xfs_xattr_iomap_begin,
1436 };
1437 
1438 int
1439 xfs_zero_range(
1440 	struct xfs_inode	*ip,
1441 	loff_t			pos,
1442 	loff_t			len,
1443 	bool			*did_zero)
1444 {
1445 	struct inode		*inode = VFS_I(ip);
1446 
1447 	if (IS_DAX(inode))
1448 		return dax_zero_range(inode, pos, len, did_zero,
1449 				      &xfs_dax_write_iomap_ops);
1450 	return iomap_zero_range(inode, pos, len, did_zero,
1451 				&xfs_buffered_write_iomap_ops);
1452 }
1453 
1454 int
1455 xfs_truncate_page(
1456 	struct xfs_inode	*ip,
1457 	loff_t			pos,
1458 	bool			*did_zero)
1459 {
1460 	struct inode		*inode = VFS_I(ip);
1461 
1462 	if (IS_DAX(inode))
1463 		return dax_truncate_page(inode, pos, did_zero,
1464 					&xfs_dax_write_iomap_ops);
1465 	return iomap_truncate_page(inode, pos, did_zero,
1466 				   &xfs_buffered_write_iomap_ops);
1467 }
1468