Linux Audio

Check our new training course

Linux debugging, profiling, tracing and performance analysis training

Apr 14-17, 2025
Register
Loading...
Note: File does not exist in v3.15.
  1// SPDX-License-Identifier: GPL-2.0-or-later
  2/*
  3 * Copyright (C) 2019 Oracle.  All Rights Reserved.
  4 * Author: Darrick J. Wong <darrick.wong@oracle.com>
  5 */
  6#include "xfs.h"
  7#include "xfs_fs.h"
  8#include "xfs_shared.h"
  9#include "xfs_format.h"
 10#include "xfs_log_format.h"
 11#include "xfs_trans_resv.h"
 12#include "xfs_mount.h"
 13#include "xfs_inode.h"
 14#include "xfs_btree.h"
 15#include "xfs_ialloc.h"
 16#include "xfs_ialloc_btree.h"
 17#include "xfs_iwalk.h"
 18#include "xfs_error.h"
 19#include "xfs_trace.h"
 20#include "xfs_icache.h"
 21#include "xfs_health.h"
 22#include "xfs_trans.h"
 23#include "xfs_pwork.h"
 24
 25/*
 26 * Walking Inodes in the Filesystem
 27 * ================================
 28 *
 29 * This iterator function walks a subset of filesystem inodes in increasing
 30 * order from @startino until there are no more inodes.  For each allocated
 31 * inode it finds, it calls a walk function with the relevant inode number and
 32 * a pointer to caller-provided data.  The walk function can return the usual
 33 * negative error code to stop the iteration; 0 to continue the iteration; or
 34 * -ECANCELED to stop the iteration.  This return value is returned to the
 35 * caller.
 36 *
 37 * Internally, we allow the walk function to do anything, which means that we
 38 * cannot maintain the inobt cursor or our lock on the AGI buffer.  We
 39 * therefore cache the inobt records in kernel memory and only call the walk
 40 * function when our memory buffer is full.  @nr_recs is the number of records
 41 * that we've cached, and @sz_recs is the size of our cache.
 42 *
 43 * It is the responsibility of the walk function to ensure it accesses
 44 * allocated inodes, as the inobt records may be stale by the time they are
 45 * acted upon.
 46 */
 47
 48struct xfs_iwalk_ag {
 49	/* parallel work control data; will be null if single threaded */
 50	struct xfs_pwork		pwork;
 51
 52	struct xfs_mount		*mp;
 53	struct xfs_trans		*tp;
 54
 55	/* Where do we start the traversal? */
 56	xfs_ino_t			startino;
 57
 58	/* Array of inobt records we cache. */
 59	struct xfs_inobt_rec_incore	*recs;
 60
 61	/* Number of entries allocated for the @recs array. */
 62	unsigned int			sz_recs;
 63
 64	/* Number of entries in the @recs array that are in use. */
 65	unsigned int			nr_recs;
 66
 67	/* Inode walk function and data pointer. */
 68	xfs_iwalk_fn			iwalk_fn;
 69	xfs_inobt_walk_fn		inobt_walk_fn;
 70	void				*data;
 71
 72	/*
 73	 * Make it look like the inodes up to startino are free so that
 74	 * bulkstat can start its inode iteration at the correct place without
 75	 * needing to special case everywhere.
 76	 */
 77	unsigned int			trim_start:1;
 78
 79	/* Skip empty inobt records? */
 80	unsigned int			skip_empty:1;
 81};
 82
 83/*
 84 * Loop over all clusters in a chunk for a given incore inode allocation btree
 85 * record.  Do a readahead if there are any allocated inodes in that cluster.
 86 */
 87STATIC void
 88xfs_iwalk_ichunk_ra(
 89	struct xfs_mount		*mp,
 90	xfs_agnumber_t			agno,
 91	struct xfs_inobt_rec_incore	*irec)
 92{
 93	struct xfs_ino_geometry		*igeo = M_IGEO(mp);
 94	xfs_agblock_t			agbno;
 95	struct blk_plug			plug;
 96	int				i;	/* inode chunk index */
 97
 98	agbno = XFS_AGINO_TO_AGBNO(mp, irec->ir_startino);
 99
100	blk_start_plug(&plug);
101	for (i = 0; i < XFS_INODES_PER_CHUNK; i += igeo->inodes_per_cluster) {
102		xfs_inofree_t	imask;
103
104		imask = xfs_inobt_maskn(i, igeo->inodes_per_cluster);
105		if (imask & ~irec->ir_free) {
106			xfs_btree_reada_bufs(mp, agno, agbno,
107					igeo->blocks_per_cluster,
108					&xfs_inode_buf_ops);
109		}
110		agbno += igeo->blocks_per_cluster;
111	}
112	blk_finish_plug(&plug);
113}
114
115/*
116 * Set the bits in @irec's free mask that correspond to the inodes before
117 * @agino so that we skip them.  This is how we restart an inode walk that was
118 * interrupted in the middle of an inode record.
119 */
120STATIC void
121xfs_iwalk_adjust_start(
122	xfs_agino_t			agino,	/* starting inode of chunk */
123	struct xfs_inobt_rec_incore	*irec)	/* btree record */
124{
125	int				idx;	/* index into inode chunk */
126	int				i;
127
128	idx = agino - irec->ir_startino;
129
130	/*
131	 * We got a right chunk with some left inodes allocated at it.  Grab
132	 * the chunk record.  Mark all the uninteresting inodes free because
133	 * they're before our start point.
134	 */
135	for (i = 0; i < idx; i++) {
136		if (XFS_INOBT_MASK(i) & ~irec->ir_free)
137			irec->ir_freecount++;
138	}
139
140	irec->ir_free |= xfs_inobt_maskn(0, idx);
141}
142
143/* Allocate memory for a walk. */
144STATIC int
145xfs_iwalk_alloc(
146	struct xfs_iwalk_ag	*iwag)
147{
148	size_t			size;
149
150	ASSERT(iwag->recs == NULL);
151	iwag->nr_recs = 0;
152
153	/* Allocate a prefetch buffer for inobt records. */
154	size = iwag->sz_recs * sizeof(struct xfs_inobt_rec_incore);
155	iwag->recs = kmem_alloc(size, KM_MAYFAIL);
156	if (iwag->recs == NULL)
157		return -ENOMEM;
158
159	return 0;
160}
161
162/* Free memory we allocated for a walk. */
163STATIC void
164xfs_iwalk_free(
165	struct xfs_iwalk_ag	*iwag)
166{
167	kmem_free(iwag->recs);
168	iwag->recs = NULL;
169}
170
171/* For each inuse inode in each cached inobt record, call our function. */
172STATIC int
173xfs_iwalk_ag_recs(
174	struct xfs_iwalk_ag		*iwag)
175{
176	struct xfs_mount		*mp = iwag->mp;
177	struct xfs_trans		*tp = iwag->tp;
178	xfs_ino_t			ino;
179	unsigned int			i, j;
180	xfs_agnumber_t			agno;
181	int				error;
182
183	agno = XFS_INO_TO_AGNO(mp, iwag->startino);
184	for (i = 0; i < iwag->nr_recs; i++) {
185		struct xfs_inobt_rec_incore	*irec = &iwag->recs[i];
186
187		trace_xfs_iwalk_ag_rec(mp, agno, irec);
188
189		if (xfs_pwork_want_abort(&iwag->pwork))
190			return 0;
191
192		if (iwag->inobt_walk_fn) {
193			error = iwag->inobt_walk_fn(mp, tp, agno, irec,
194					iwag->data);
195			if (error)
196				return error;
197		}
198
199		if (!iwag->iwalk_fn)
200			continue;
201
202		for (j = 0; j < XFS_INODES_PER_CHUNK; j++) {
203			if (xfs_pwork_want_abort(&iwag->pwork))
204				return 0;
205
206			/* Skip if this inode is free */
207			if (XFS_INOBT_MASK(j) & irec->ir_free)
208				continue;
209
210			/* Otherwise call our function. */
211			ino = XFS_AGINO_TO_INO(mp, agno, irec->ir_startino + j);
212			error = iwag->iwalk_fn(mp, tp, ino, iwag->data);
213			if (error)
214				return error;
215		}
216	}
217
218	return 0;
219}
220
221/* Delete cursor and let go of AGI. */
222static inline void
223xfs_iwalk_del_inobt(
224	struct xfs_trans	*tp,
225	struct xfs_btree_cur	**curpp,
226	struct xfs_buf		**agi_bpp,
227	int			error)
228{
229	if (*curpp) {
230		xfs_btree_del_cursor(*curpp, error);
231		*curpp = NULL;
232	}
233	if (*agi_bpp) {
234		xfs_trans_brelse(tp, *agi_bpp);
235		*agi_bpp = NULL;
236	}
237}
238
239/*
240 * Set ourselves up for walking inobt records starting from a given point in
241 * the filesystem.
242 *
243 * If caller passed in a nonzero start inode number, load the record from the
244 * inobt and make the record look like all the inodes before agino are free so
245 * that we skip them, and then move the cursor to the next inobt record.  This
246 * is how we support starting an iwalk in the middle of an inode chunk.
247 *
248 * If the caller passed in a start number of zero, move the cursor to the first
249 * inobt record.
250 *
251 * The caller is responsible for cleaning up the cursor and buffer pointer
252 * regardless of the error status.
253 */
254STATIC int
255xfs_iwalk_ag_start(
256	struct xfs_iwalk_ag	*iwag,
257	xfs_agnumber_t		agno,
258	xfs_agino_t		agino,
259	struct xfs_btree_cur	**curpp,
260	struct xfs_buf		**agi_bpp,
261	int			*has_more)
262{
263	struct xfs_mount	*mp = iwag->mp;
264	struct xfs_trans	*tp = iwag->tp;
265	struct xfs_inobt_rec_incore *irec;
266	int			error;
267
268	/* Set up a fresh cursor and empty the inobt cache. */
269	iwag->nr_recs = 0;
270	error = xfs_inobt_cur(mp, tp, agno, XFS_BTNUM_INO, curpp, agi_bpp);
271	if (error)
272		return error;
273
274	/* Starting at the beginning of the AG?  That's easy! */
275	if (agino == 0)
276		return xfs_inobt_lookup(*curpp, 0, XFS_LOOKUP_GE, has_more);
277
278	/*
279	 * Otherwise, we have to grab the inobt record where we left off, stuff
280	 * the record into our cache, and then see if there are more records.
281	 * We require a lookup cache of at least two elements so that the
282	 * caller doesn't have to deal with tearing down the cursor to walk the
283	 * records.
284	 */
285	error = xfs_inobt_lookup(*curpp, agino, XFS_LOOKUP_LE, has_more);
286	if (error)
287		return error;
288
289	/*
290	 * If the LE lookup at @agino yields no records, jump ahead to the
291	 * inobt cursor increment to see if there are more records to process.
292	 */
293	if (!*has_more)
294		goto out_advance;
295
296	/* Get the record, should always work */
297	irec = &iwag->recs[iwag->nr_recs];
298	error = xfs_inobt_get_rec(*curpp, irec, has_more);
299	if (error)
300		return error;
301	if (XFS_IS_CORRUPT(mp, *has_more != 1))
302		return -EFSCORRUPTED;
303
304	/*
305	 * If the LE lookup yielded an inobt record before the cursor position,
306	 * skip it and see if there's another one after it.
307	 */
308	if (irec->ir_startino + XFS_INODES_PER_CHUNK <= agino)
309		goto out_advance;
310
311	/*
312	 * If agino fell in the middle of the inode record, make it look like
313	 * the inodes up to agino are free so that we don't return them again.
314	 */
315	if (iwag->trim_start)
316		xfs_iwalk_adjust_start(agino, irec);
317
318	/*
319	 * The prefetch calculation is supposed to give us a large enough inobt
320	 * record cache that grab_ichunk can stage a partial first record and
321	 * the loop body can cache a record without having to check for cache
322	 * space until after it reads an inobt record.
323	 */
324	iwag->nr_recs++;
325	ASSERT(iwag->nr_recs < iwag->sz_recs);
326
327out_advance:
328	return xfs_btree_increment(*curpp, 0, has_more);
329}
330
331/*
332 * The inobt record cache is full, so preserve the inobt cursor state and
333 * run callbacks on the cached inobt records.  When we're done, restore the
334 * cursor state to wherever the cursor would have been had the cache not been
335 * full (and therefore we could've just incremented the cursor) if *@has_more
336 * is true.  On exit, *@has_more will indicate whether or not the caller should
337 * try for more inode records.
338 */
339STATIC int
340xfs_iwalk_run_callbacks(
341	struct xfs_iwalk_ag		*iwag,
342	xfs_agnumber_t			agno,
343	struct xfs_btree_cur		**curpp,
344	struct xfs_buf			**agi_bpp,
345	int				*has_more)
346{
347	struct xfs_mount		*mp = iwag->mp;
348	struct xfs_trans		*tp = iwag->tp;
349	struct xfs_inobt_rec_incore	*irec;
350	xfs_agino_t			restart;
351	int				error;
352
353	ASSERT(iwag->nr_recs > 0);
354
355	/* Delete cursor but remember the last record we cached... */
356	xfs_iwalk_del_inobt(tp, curpp, agi_bpp, 0);
357	irec = &iwag->recs[iwag->nr_recs - 1];
358	restart = irec->ir_startino + XFS_INODES_PER_CHUNK - 1;
359
360	error = xfs_iwalk_ag_recs(iwag);
361	if (error)
362		return error;
363
364	/* ...empty the cache... */
365	iwag->nr_recs = 0;
366
367	if (!has_more)
368		return 0;
369
370	/* ...and recreate the cursor just past where we left off. */
371	error = xfs_inobt_cur(mp, tp, agno, XFS_BTNUM_INO, curpp, agi_bpp);
372	if (error)
373		return error;
374
375	return xfs_inobt_lookup(*curpp, restart, XFS_LOOKUP_GE, has_more);
376}
377
378/* Walk all inodes in a single AG, from @iwag->startino to the end of the AG. */
379STATIC int
380xfs_iwalk_ag(
381	struct xfs_iwalk_ag		*iwag)
382{
383	struct xfs_mount		*mp = iwag->mp;
384	struct xfs_trans		*tp = iwag->tp;
385	struct xfs_buf			*agi_bp = NULL;
386	struct xfs_btree_cur		*cur = NULL;
387	xfs_agnumber_t			agno;
388	xfs_agino_t			agino;
389	int				has_more;
390	int				error = 0;
391
392	/* Set up our cursor at the right place in the inode btree. */
393	agno = XFS_INO_TO_AGNO(mp, iwag->startino);
394	agino = XFS_INO_TO_AGINO(mp, iwag->startino);
395	error = xfs_iwalk_ag_start(iwag, agno, agino, &cur, &agi_bp, &has_more);
396
397	while (!error && has_more) {
398		struct xfs_inobt_rec_incore	*irec;
399
400		cond_resched();
401		if (xfs_pwork_want_abort(&iwag->pwork))
402			goto out;
403
404		/* Fetch the inobt record. */
405		irec = &iwag->recs[iwag->nr_recs];
406		error = xfs_inobt_get_rec(cur, irec, &has_more);
407		if (error || !has_more)
408			break;
409
410		/* No allocated inodes in this chunk; skip it. */
411		if (iwag->skip_empty && irec->ir_freecount == irec->ir_count) {
412			error = xfs_btree_increment(cur, 0, &has_more);
413			if (error)
414				break;
415			continue;
416		}
417
418		/*
419		 * Start readahead for this inode chunk in anticipation of
420		 * walking the inodes.
421		 */
422		if (iwag->iwalk_fn)
423			xfs_iwalk_ichunk_ra(mp, agno, irec);
424
425		/*
426		 * If there's space in the buffer for more records, increment
427		 * the btree cursor and grab more.
428		 */
429		if (++iwag->nr_recs < iwag->sz_recs) {
430			error = xfs_btree_increment(cur, 0, &has_more);
431			if (error || !has_more)
432				break;
433			continue;
434		}
435
436		/*
437		 * Otherwise, we need to save cursor state and run the callback
438		 * function on the cached records.  The run_callbacks function
439		 * is supposed to return a cursor pointing to the record where
440		 * we would be if we had been able to increment like above.
441		 */
442		ASSERT(has_more);
443		error = xfs_iwalk_run_callbacks(iwag, agno, &cur, &agi_bp,
444				&has_more);
445	}
446
447	if (iwag->nr_recs == 0 || error)
448		goto out;
449
450	/* Walk the unprocessed records in the cache. */
451	error = xfs_iwalk_run_callbacks(iwag, agno, &cur, &agi_bp, &has_more);
452
453out:
454	xfs_iwalk_del_inobt(tp, &cur, &agi_bp, error);
455	return error;
456}
457
458/*
459 * We experimentally determined that the reduction in ioctl call overhead
460 * diminishes when userspace asks for more than 2048 inodes, so we'll cap
461 * prefetch at this point.
462 */
463#define IWALK_MAX_INODE_PREFETCH	(2048U)
464
465/*
466 * Given the number of inodes to prefetch, set the number of inobt records that
467 * we cache in memory, which controls the number of inodes we try to read
468 * ahead.  Set the maximum if @inodes == 0.
469 */
470static inline unsigned int
471xfs_iwalk_prefetch(
472	unsigned int		inodes)
473{
474	unsigned int		inobt_records;
475
476	/*
477	 * If the caller didn't tell us the number of inodes they wanted,
478	 * assume the maximum prefetch possible for best performance.
479	 * Otherwise, cap prefetch at that maximum so that we don't start an
480	 * absurd amount of prefetch.
481	 */
482	if (inodes == 0)
483		inodes = IWALK_MAX_INODE_PREFETCH;
484	inodes = min(inodes, IWALK_MAX_INODE_PREFETCH);
485
486	/* Round the inode count up to a full chunk. */
487	inodes = round_up(inodes, XFS_INODES_PER_CHUNK);
488
489	/*
490	 * In order to convert the number of inodes to prefetch into an
491	 * estimate of the number of inobt records to cache, we require a
492	 * conversion factor that reflects our expectations of the average
493	 * loading factor of an inode chunk.  Based on data gathered, most
494	 * (but not all) filesystems manage to keep the inode chunks totally
495	 * full, so we'll underestimate slightly so that our readahead will
496	 * still deliver the performance we want on aging filesystems:
497	 *
498	 * inobt = inodes / (INODES_PER_CHUNK * (4 / 5));
499	 *
500	 * The funny math is to avoid integer division.
501	 */
502	inobt_records = (inodes * 5) / (4 * XFS_INODES_PER_CHUNK);
503
504	/*
505	 * Allocate enough space to prefetch at least two inobt records so that
506	 * we can cache both the record where the iwalk started and the next
507	 * record.  This simplifies the AG inode walk loop setup code.
508	 */
509	return max(inobt_records, 2U);
510}
511
512/*
513 * Walk all inodes in the filesystem starting from @startino.  The @iwalk_fn
514 * will be called for each allocated inode, being passed the inode's number and
515 * @data.  @max_prefetch controls how many inobt records' worth of inodes we
516 * try to readahead.
517 */
518int
519xfs_iwalk(
520	struct xfs_mount	*mp,
521	struct xfs_trans	*tp,
522	xfs_ino_t		startino,
523	unsigned int		flags,
524	xfs_iwalk_fn		iwalk_fn,
525	unsigned int		inode_records,
526	void			*data)
527{
528	struct xfs_iwalk_ag	iwag = {
529		.mp		= mp,
530		.tp		= tp,
531		.iwalk_fn	= iwalk_fn,
532		.data		= data,
533		.startino	= startino,
534		.sz_recs	= xfs_iwalk_prefetch(inode_records),
535		.trim_start	= 1,
536		.skip_empty	= 1,
537		.pwork		= XFS_PWORK_SINGLE_THREADED,
538	};
539	xfs_agnumber_t		agno = XFS_INO_TO_AGNO(mp, startino);
540	int			error;
541
542	ASSERT(agno < mp->m_sb.sb_agcount);
543	ASSERT(!(flags & ~XFS_IWALK_FLAGS_ALL));
544
545	error = xfs_iwalk_alloc(&iwag);
546	if (error)
547		return error;
548
549	for (; agno < mp->m_sb.sb_agcount; agno++) {
550		error = xfs_iwalk_ag(&iwag);
551		if (error)
552			break;
553		iwag.startino = XFS_AGINO_TO_INO(mp, agno + 1, 0);
554		if (flags & XFS_INOBT_WALK_SAME_AG)
555			break;
556	}
557
558	xfs_iwalk_free(&iwag);
559	return error;
560}
561
562/* Run per-thread iwalk work. */
563static int
564xfs_iwalk_ag_work(
565	struct xfs_mount	*mp,
566	struct xfs_pwork	*pwork)
567{
568	struct xfs_iwalk_ag	*iwag;
569	int			error = 0;
570
571	iwag = container_of(pwork, struct xfs_iwalk_ag, pwork);
572	if (xfs_pwork_want_abort(pwork))
573		goto out;
574
575	error = xfs_iwalk_alloc(iwag);
576	if (error)
577		goto out;
578
579	error = xfs_iwalk_ag(iwag);
580	xfs_iwalk_free(iwag);
581out:
582	kmem_free(iwag);
583	return error;
584}
585
586/*
587 * Walk all the inodes in the filesystem using multiple threads to process each
588 * AG.
589 */
590int
591xfs_iwalk_threaded(
592	struct xfs_mount	*mp,
593	xfs_ino_t		startino,
594	unsigned int		flags,
595	xfs_iwalk_fn		iwalk_fn,
596	unsigned int		inode_records,
597	bool			polled,
598	void			*data)
599{
600	struct xfs_pwork_ctl	pctl;
601	xfs_agnumber_t		agno = XFS_INO_TO_AGNO(mp, startino);
602	unsigned int		nr_threads;
603	int			error;
604
605	ASSERT(agno < mp->m_sb.sb_agcount);
606	ASSERT(!(flags & ~XFS_IWALK_FLAGS_ALL));
607
608	nr_threads = xfs_pwork_guess_datadev_parallelism(mp);
609	error = xfs_pwork_init(mp, &pctl, xfs_iwalk_ag_work, "xfs_iwalk",
610			nr_threads);
611	if (error)
612		return error;
613
614	for (; agno < mp->m_sb.sb_agcount; agno++) {
615		struct xfs_iwalk_ag	*iwag;
616
617		if (xfs_pwork_ctl_want_abort(&pctl))
618			break;
619
620		iwag = kmem_zalloc(sizeof(struct xfs_iwalk_ag), 0);
621		iwag->mp = mp;
622		iwag->iwalk_fn = iwalk_fn;
623		iwag->data = data;
624		iwag->startino = startino;
625		iwag->sz_recs = xfs_iwalk_prefetch(inode_records);
626		xfs_pwork_queue(&pctl, &iwag->pwork);
627		startino = XFS_AGINO_TO_INO(mp, agno + 1, 0);
628		if (flags & XFS_INOBT_WALK_SAME_AG)
629			break;
630	}
631
632	if (polled)
633		xfs_pwork_poll(&pctl);
634	return xfs_pwork_destroy(&pctl);
635}
636
637/*
638 * Allow callers to cache up to a page's worth of inobt records.  This reflects
639 * the existing inumbers prefetching behavior.  Since the inobt walk does not
640 * itself do anything with the inobt records, we can set a fairly high limit
641 * here.
642 */
643#define MAX_INOBT_WALK_PREFETCH	\
644	(PAGE_SIZE / sizeof(struct xfs_inobt_rec_incore))
645
646/*
647 * Given the number of records that the user wanted, set the number of inobt
648 * records that we buffer in memory.  Set the maximum if @inobt_records == 0.
649 */
650static inline unsigned int
651xfs_inobt_walk_prefetch(
652	unsigned int		inobt_records)
653{
654	/*
655	 * If the caller didn't tell us the number of inobt records they
656	 * wanted, assume the maximum prefetch possible for best performance.
657	 */
658	if (inobt_records == 0)
659		inobt_records = MAX_INOBT_WALK_PREFETCH;
660
661	/*
662	 * Allocate enough space to prefetch at least two inobt records so that
663	 * we can cache both the record where the iwalk started and the next
664	 * record.  This simplifies the AG inode walk loop setup code.
665	 */
666	inobt_records = max(inobt_records, 2U);
667
668	/*
669	 * Cap prefetch at that maximum so that we don't use an absurd amount
670	 * of memory.
671	 */
672	return min_t(unsigned int, inobt_records, MAX_INOBT_WALK_PREFETCH);
673}
674
675/*
676 * Walk all inode btree records in the filesystem starting from @startino.  The
677 * @inobt_walk_fn will be called for each btree record, being passed the incore
678 * record and @data.  @max_prefetch controls how many inobt records we try to
679 * cache ahead of time.
680 */
681int
682xfs_inobt_walk(
683	struct xfs_mount	*mp,
684	struct xfs_trans	*tp,
685	xfs_ino_t		startino,
686	unsigned int		flags,
687	xfs_inobt_walk_fn	inobt_walk_fn,
688	unsigned int		inobt_records,
689	void			*data)
690{
691	struct xfs_iwalk_ag	iwag = {
692		.mp		= mp,
693		.tp		= tp,
694		.inobt_walk_fn	= inobt_walk_fn,
695		.data		= data,
696		.startino	= startino,
697		.sz_recs	= xfs_inobt_walk_prefetch(inobt_records),
698		.pwork		= XFS_PWORK_SINGLE_THREADED,
699	};
700	xfs_agnumber_t		agno = XFS_INO_TO_AGNO(mp, startino);
701	int			error;
702
703	ASSERT(agno < mp->m_sb.sb_agcount);
704	ASSERT(!(flags & ~XFS_INOBT_WALK_FLAGS_ALL));
705
706	error = xfs_iwalk_alloc(&iwag);
707	if (error)
708		return error;
709
710	for (; agno < mp->m_sb.sb_agcount; agno++) {
711		error = xfs_iwalk_ag(&iwag);
712		if (error)
713			break;
714		iwag.startino = XFS_AGINO_TO_INO(mp, agno + 1, 0);
715		if (flags & XFS_INOBT_WALK_SAME_AG)
716			break;
717	}
718
719	xfs_iwalk_free(&iwag);
720	return error;
721}