1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright (C) 2019 Oracle. All Rights Reserved.
4 * Author: Darrick J. Wong <darrick.wong@oracle.com>
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"
27 * Walking Inodes in the Filesystem
28 * ================================
30 * This iterator function walks a subset of filesystem inodes in increasing
31 * order from @startino until there are no more inodes. For each allocated
32 * inode it finds, it calls a walk function with the relevant inode number and
33 * a pointer to caller-provided data. The walk function can return the usual
34 * negative error code to stop the iteration; 0 to continue the iteration; or
35 * -ECANCELED to stop the iteration. This return value is returned to the
38 * Internally, we allow the walk function to do anything, which means that we
39 * cannot maintain the inobt cursor or our lock on the AGI buffer. We
40 * therefore cache the inobt records in kernel memory and only call the walk
41 * function when our memory buffer is full. @nr_recs is the number of records
42 * that we've cached, and @sz_recs is the size of our cache.
44 * It is the responsibility of the walk function to ensure it accesses
45 * allocated inodes, as the inobt records may be stale by the time they are
50 /* parallel work control data; will be null if single threaded */
51 struct xfs_pwork pwork;
55 struct xfs_perag *pag;
57 /* Where do we start the traversal? */
60 /* What was the last inode number we saw when iterating the inobt? */
63 /* Array of inobt records we cache. */
64 struct xfs_inobt_rec_incore *recs;
66 /* Number of entries allocated for the @recs array. */
69 /* Number of entries in the @recs array that are in use. */
72 /* Inode walk function and data pointer. */
73 xfs_iwalk_fn iwalk_fn;
74 xfs_inobt_walk_fn inobt_walk_fn;
78 * Make it look like the inodes up to startino are free so that
79 * bulkstat can start its inode iteration at the correct place without
80 * needing to special case everywhere.
82 unsigned int trim_start:1;
84 /* Skip empty inobt records? */
85 unsigned int skip_empty:1;
87 /* Drop the (hopefully empty) transaction when calling iwalk_fn. */
88 unsigned int drop_trans:1;
92 * Loop over all clusters in a chunk for a given incore inode allocation btree
93 * record. Do a readahead if there are any allocated inodes in that cluster.
98 struct xfs_perag *pag,
99 struct xfs_inobt_rec_incore *irec)
101 struct xfs_ino_geometry *igeo = M_IGEO(mp);
103 struct blk_plug plug;
104 int i; /* inode chunk index */
106 agbno = XFS_AGINO_TO_AGBNO(mp, irec->ir_startino);
108 blk_start_plug(&plug);
109 for (i = 0; i < XFS_INODES_PER_CHUNK; i += igeo->inodes_per_cluster) {
112 imask = xfs_inobt_maskn(i, igeo->inodes_per_cluster);
113 if (imask & ~irec->ir_free) {
114 xfs_btree_reada_bufs(mp, pag->pag_agno, agbno,
115 igeo->blocks_per_cluster,
118 agbno += igeo->blocks_per_cluster;
120 blk_finish_plug(&plug);
124 * Set the bits in @irec's free mask that correspond to the inodes before
125 * @agino so that we skip them. This is how we restart an inode walk that was
126 * interrupted in the middle of an inode record.
129 xfs_iwalk_adjust_start(
130 xfs_agino_t agino, /* starting inode of chunk */
131 struct xfs_inobt_rec_incore *irec) /* btree record */
133 int idx; /* index into inode chunk */
136 idx = agino - irec->ir_startino;
139 * We got a right chunk with some left inodes allocated at it. Grab
140 * the chunk record. Mark all the uninteresting inodes free because
141 * they're before our start point.
143 for (i = 0; i < idx; i++) {
144 if (XFS_INOBT_MASK(i) & ~irec->ir_free)
145 irec->ir_freecount++;
148 irec->ir_free |= xfs_inobt_maskn(0, idx);
151 /* Allocate memory for a walk. */
154 struct xfs_iwalk_ag *iwag)
158 ASSERT(iwag->recs == NULL);
161 /* Allocate a prefetch buffer for inobt records. */
162 size = iwag->sz_recs * sizeof(struct xfs_inobt_rec_incore);
163 iwag->recs = kmem_alloc(size, KM_MAYFAIL);
164 if (iwag->recs == NULL)
170 /* Free memory we allocated for a walk. */
173 struct xfs_iwalk_ag *iwag)
175 kmem_free(iwag->recs);
179 /* For each inuse inode in each cached inobt record, call our function. */
182 struct xfs_iwalk_ag *iwag)
184 struct xfs_mount *mp = iwag->mp;
185 struct xfs_trans *tp = iwag->tp;
186 struct xfs_perag *pag = iwag->pag;
191 for (i = 0; i < iwag->nr_recs; i++) {
192 struct xfs_inobt_rec_incore *irec = &iwag->recs[i];
194 trace_xfs_iwalk_ag_rec(mp, pag->pag_agno, irec);
196 if (xfs_pwork_want_abort(&iwag->pwork))
199 if (iwag->inobt_walk_fn) {
200 error = iwag->inobt_walk_fn(mp, tp, pag->pag_agno, irec,
209 for (j = 0; j < XFS_INODES_PER_CHUNK; j++) {
210 if (xfs_pwork_want_abort(&iwag->pwork))
213 /* Skip if this inode is free */
214 if (XFS_INOBT_MASK(j) & irec->ir_free)
217 /* Otherwise call our function. */
218 ino = XFS_AGINO_TO_INO(mp, pag->pag_agno,
219 irec->ir_startino + j);
220 error = iwag->iwalk_fn(mp, tp, ino, iwag->data);
229 /* Delete cursor and let go of AGI. */
232 struct xfs_trans *tp,
233 struct xfs_btree_cur **curpp,
234 struct xfs_buf **agi_bpp,
238 xfs_btree_del_cursor(*curpp, error);
242 xfs_trans_brelse(tp, *agi_bpp);
248 * Set ourselves up for walking inobt records starting from a given point in
251 * If caller passed in a nonzero start inode number, load the record from the
252 * inobt and make the record look like all the inodes before agino are free so
253 * that we skip them, and then move the cursor to the next inobt record. This
254 * is how we support starting an iwalk in the middle of an inode chunk.
256 * If the caller passed in a start number of zero, move the cursor to the first
259 * The caller is responsible for cleaning up the cursor and buffer pointer
260 * regardless of the error status.
264 struct xfs_iwalk_ag *iwag,
266 struct xfs_btree_cur **curpp,
267 struct xfs_buf **agi_bpp,
270 struct xfs_mount *mp = iwag->mp;
271 struct xfs_trans *tp = iwag->tp;
272 struct xfs_perag *pag = iwag->pag;
273 struct xfs_inobt_rec_incore *irec;
276 /* Set up a fresh cursor and empty the inobt cache. */
278 error = xfs_inobt_cur(mp, tp, pag, XFS_BTNUM_INO, curpp, agi_bpp);
282 /* Starting at the beginning of the AG? That's easy! */
284 return xfs_inobt_lookup(*curpp, 0, XFS_LOOKUP_GE, has_more);
287 * Otherwise, we have to grab the inobt record where we left off, stuff
288 * the record into our cache, and then see if there are more records.
289 * We require a lookup cache of at least two elements so that the
290 * caller doesn't have to deal with tearing down the cursor to walk the
293 error = xfs_inobt_lookup(*curpp, agino, XFS_LOOKUP_LE, has_more);
298 * If the LE lookup at @agino yields no records, jump ahead to the
299 * inobt cursor increment to see if there are more records to process.
304 /* Get the record, should always work */
305 irec = &iwag->recs[iwag->nr_recs];
306 error = xfs_inobt_get_rec(*curpp, irec, has_more);
309 if (XFS_IS_CORRUPT(mp, *has_more != 1))
310 return -EFSCORRUPTED;
312 iwag->lastino = XFS_AGINO_TO_INO(mp, pag->pag_agno,
313 irec->ir_startino + XFS_INODES_PER_CHUNK - 1);
316 * If the LE lookup yielded an inobt record before the cursor position,
317 * skip it and see if there's another one after it.
319 if (irec->ir_startino + XFS_INODES_PER_CHUNK <= agino)
323 * If agino fell in the middle of the inode record, make it look like
324 * the inodes up to agino are free so that we don't return them again.
326 if (iwag->trim_start)
327 xfs_iwalk_adjust_start(agino, irec);
330 * The prefetch calculation is supposed to give us a large enough inobt
331 * record cache that grab_ichunk can stage a partial first record and
332 * the loop body can cache a record without having to check for cache
333 * space until after it reads an inobt record.
336 ASSERT(iwag->nr_recs < iwag->sz_recs);
339 return xfs_btree_increment(*curpp, 0, has_more);
343 * The inobt record cache is full, so preserve the inobt cursor state and
344 * run callbacks on the cached inobt records. When we're done, restore the
345 * cursor state to wherever the cursor would have been had the cache not been
346 * full (and therefore we could've just incremented the cursor) if *@has_more
347 * is true. On exit, *@has_more will indicate whether or not the caller should
348 * try for more inode records.
351 xfs_iwalk_run_callbacks(
352 struct xfs_iwalk_ag *iwag,
353 struct xfs_btree_cur **curpp,
354 struct xfs_buf **agi_bpp,
357 struct xfs_mount *mp = iwag->mp;
358 struct xfs_inobt_rec_incore *irec;
359 xfs_agino_t next_agino;
362 next_agino = XFS_INO_TO_AGINO(mp, iwag->lastino) + 1;
364 ASSERT(iwag->nr_recs > 0);
366 /* Delete cursor but remember the last record we cached... */
367 xfs_iwalk_del_inobt(iwag->tp, curpp, agi_bpp, 0);
368 irec = &iwag->recs[iwag->nr_recs - 1];
369 ASSERT(next_agino >= irec->ir_startino + XFS_INODES_PER_CHUNK);
371 if (iwag->drop_trans) {
372 xfs_trans_cancel(iwag->tp);
376 error = xfs_iwalk_ag_recs(iwag);
380 /* ...empty the cache... */
386 if (iwag->drop_trans) {
387 error = xfs_trans_alloc_empty(mp, &iwag->tp);
392 /* ...and recreate the cursor just past where we left off. */
393 error = xfs_inobt_cur(mp, iwag->tp, iwag->pag, XFS_BTNUM_INO, curpp,
398 return xfs_inobt_lookup(*curpp, next_agino, XFS_LOOKUP_GE, has_more);
401 /* Walk all inodes in a single AG, from @iwag->startino to the end of the AG. */
404 struct xfs_iwalk_ag *iwag)
406 struct xfs_mount *mp = iwag->mp;
407 struct xfs_perag *pag = iwag->pag;
408 struct xfs_buf *agi_bp = NULL;
409 struct xfs_btree_cur *cur = NULL;
414 /* Set up our cursor at the right place in the inode btree. */
415 ASSERT(pag->pag_agno == XFS_INO_TO_AGNO(mp, iwag->startino));
416 agino = XFS_INO_TO_AGINO(mp, iwag->startino);
417 error = xfs_iwalk_ag_start(iwag, agino, &cur, &agi_bp, &has_more);
419 while (!error && has_more) {
420 struct xfs_inobt_rec_incore *irec;
424 if (xfs_pwork_want_abort(&iwag->pwork))
427 /* Fetch the inobt record. */
428 irec = &iwag->recs[iwag->nr_recs];
429 error = xfs_inobt_get_rec(cur, irec, &has_more);
430 if (error || !has_more)
433 /* Make sure that we always move forward. */
434 rec_fsino = XFS_AGINO_TO_INO(mp, pag->pag_agno, irec->ir_startino);
435 if (iwag->lastino != NULLFSINO &&
436 XFS_IS_CORRUPT(mp, iwag->lastino >= rec_fsino)) {
437 error = -EFSCORRUPTED;
440 iwag->lastino = rec_fsino + XFS_INODES_PER_CHUNK - 1;
442 /* No allocated inodes in this chunk; skip it. */
443 if (iwag->skip_empty && irec->ir_freecount == irec->ir_count) {
444 error = xfs_btree_increment(cur, 0, &has_more);
451 * Start readahead for this inode chunk in anticipation of
452 * walking the inodes.
455 xfs_iwalk_ichunk_ra(mp, pag, irec);
458 * If there's space in the buffer for more records, increment
459 * the btree cursor and grab more.
461 if (++iwag->nr_recs < iwag->sz_recs) {
462 error = xfs_btree_increment(cur, 0, &has_more);
463 if (error || !has_more)
469 * Otherwise, we need to save cursor state and run the callback
470 * function on the cached records. The run_callbacks function
471 * is supposed to return a cursor pointing to the record where
472 * we would be if we had been able to increment like above.
475 error = xfs_iwalk_run_callbacks(iwag, &cur, &agi_bp, &has_more);
478 if (iwag->nr_recs == 0 || error)
481 /* Walk the unprocessed records in the cache. */
482 error = xfs_iwalk_run_callbacks(iwag, &cur, &agi_bp, &has_more);
485 xfs_iwalk_del_inobt(iwag->tp, &cur, &agi_bp, error);
490 * We experimentally determined that the reduction in ioctl call overhead
491 * diminishes when userspace asks for more than 2048 inodes, so we'll cap
492 * prefetch at this point.
494 #define IWALK_MAX_INODE_PREFETCH (2048U)
497 * Given the number of inodes to prefetch, set the number of inobt records that
498 * we cache in memory, which controls the number of inodes we try to read
499 * ahead. Set the maximum if @inodes == 0.
501 static inline unsigned int
505 unsigned int inobt_records;
508 * If the caller didn't tell us the number of inodes they wanted,
509 * assume the maximum prefetch possible for best performance.
510 * Otherwise, cap prefetch at that maximum so that we don't start an
511 * absurd amount of prefetch.
514 inodes = IWALK_MAX_INODE_PREFETCH;
515 inodes = min(inodes, IWALK_MAX_INODE_PREFETCH);
517 /* Round the inode count up to a full chunk. */
518 inodes = round_up(inodes, XFS_INODES_PER_CHUNK);
521 * In order to convert the number of inodes to prefetch into an
522 * estimate of the number of inobt records to cache, we require a
523 * conversion factor that reflects our expectations of the average
524 * loading factor of an inode chunk. Based on data gathered, most
525 * (but not all) filesystems manage to keep the inode chunks totally
526 * full, so we'll underestimate slightly so that our readahead will
527 * still deliver the performance we want on aging filesystems:
529 * inobt = inodes / (INODES_PER_CHUNK * (4 / 5));
531 * The funny math is to avoid integer division.
533 inobt_records = (inodes * 5) / (4 * XFS_INODES_PER_CHUNK);
536 * Allocate enough space to prefetch at least two inobt records so that
537 * we can cache both the record where the iwalk started and the next
538 * record. This simplifies the AG inode walk loop setup code.
540 return max(inobt_records, 2U);
544 * Walk all inodes in the filesystem starting from @startino. The @iwalk_fn
545 * will be called for each allocated inode, being passed the inode's number and
546 * @data. @max_prefetch controls how many inobt records' worth of inodes we
551 struct xfs_mount *mp,
552 struct xfs_trans *tp,
555 xfs_iwalk_fn iwalk_fn,
556 unsigned int inode_records,
559 struct xfs_iwalk_ag iwag = {
562 .iwalk_fn = iwalk_fn,
564 .startino = startino,
565 .sz_recs = xfs_iwalk_prefetch(inode_records),
568 .pwork = XFS_PWORK_SINGLE_THREADED,
569 .lastino = NULLFSINO,
571 struct xfs_perag *pag;
572 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, startino);
575 ASSERT(agno < mp->m_sb.sb_agcount);
576 ASSERT(!(flags & ~XFS_IWALK_FLAGS_ALL));
578 error = xfs_iwalk_alloc(&iwag);
582 for_each_perag_from(mp, agno, pag) {
584 error = xfs_iwalk_ag(&iwag);
587 iwag.startino = XFS_AGINO_TO_INO(mp, agno + 1, 0);
588 if (flags & XFS_INOBT_WALK_SAME_AG)
595 xfs_iwalk_free(&iwag);
599 /* Run per-thread iwalk work. */
602 struct xfs_mount *mp,
603 struct xfs_pwork *pwork)
605 struct xfs_iwalk_ag *iwag;
608 iwag = container_of(pwork, struct xfs_iwalk_ag, pwork);
609 if (xfs_pwork_want_abort(pwork))
612 error = xfs_iwalk_alloc(iwag);
616 * Grab an empty transaction so that we can use its recursive buffer
617 * locking abilities to detect cycles in the inobt without deadlocking.
619 error = xfs_trans_alloc_empty(mp, &iwag->tp);
622 iwag->drop_trans = 1;
624 error = xfs_iwalk_ag(iwag);
626 xfs_trans_cancel(iwag->tp);
627 xfs_iwalk_free(iwag);
629 xfs_perag_put(iwag->pag);
635 * Walk all the inodes in the filesystem using multiple threads to process each
640 struct xfs_mount *mp,
643 xfs_iwalk_fn iwalk_fn,
644 unsigned int inode_records,
648 struct xfs_pwork_ctl pctl;
649 struct xfs_perag *pag;
650 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, startino);
653 ASSERT(agno < mp->m_sb.sb_agcount);
654 ASSERT(!(flags & ~XFS_IWALK_FLAGS_ALL));
656 error = xfs_pwork_init(mp, &pctl, xfs_iwalk_ag_work, "xfs_iwalk");
660 for_each_perag_from(mp, agno, pag) {
661 struct xfs_iwalk_ag *iwag;
663 if (xfs_pwork_ctl_want_abort(&pctl))
666 iwag = kmem_zalloc(sizeof(struct xfs_iwalk_ag), 0);
670 * perag is being handed off to async work, so take another
671 * reference for the async work to release.
673 atomic_inc(&pag->pag_ref);
675 iwag->iwalk_fn = iwalk_fn;
677 iwag->startino = startino;
678 iwag->sz_recs = xfs_iwalk_prefetch(inode_records);
679 iwag->lastino = NULLFSINO;
680 xfs_pwork_queue(&pctl, &iwag->pwork);
681 startino = XFS_AGINO_TO_INO(mp, pag->pag_agno + 1, 0);
682 if (flags & XFS_INOBT_WALK_SAME_AG)
688 xfs_pwork_poll(&pctl);
689 return xfs_pwork_destroy(&pctl);
693 * Allow callers to cache up to a page's worth of inobt records. This reflects
694 * the existing inumbers prefetching behavior. Since the inobt walk does not
695 * itself do anything with the inobt records, we can set a fairly high limit
698 #define MAX_INOBT_WALK_PREFETCH \
699 (PAGE_SIZE / sizeof(struct xfs_inobt_rec_incore))
702 * Given the number of records that the user wanted, set the number of inobt
703 * records that we buffer in memory. Set the maximum if @inobt_records == 0.
705 static inline unsigned int
706 xfs_inobt_walk_prefetch(
707 unsigned int inobt_records)
710 * If the caller didn't tell us the number of inobt records they
711 * wanted, assume the maximum prefetch possible for best performance.
713 if (inobt_records == 0)
714 inobt_records = MAX_INOBT_WALK_PREFETCH;
717 * Allocate enough space to prefetch at least two inobt records so that
718 * we can cache both the record where the iwalk started and the next
719 * record. This simplifies the AG inode walk loop setup code.
721 inobt_records = max(inobt_records, 2U);
724 * Cap prefetch at that maximum so that we don't use an absurd amount
727 return min_t(unsigned int, inobt_records, MAX_INOBT_WALK_PREFETCH);
731 * Walk all inode btree records in the filesystem starting from @startino. The
732 * @inobt_walk_fn will be called for each btree record, being passed the incore
733 * record and @data. @max_prefetch controls how many inobt records we try to
734 * cache ahead of time.
738 struct xfs_mount *mp,
739 struct xfs_trans *tp,
742 xfs_inobt_walk_fn inobt_walk_fn,
743 unsigned int inobt_records,
746 struct xfs_iwalk_ag iwag = {
749 .inobt_walk_fn = inobt_walk_fn,
751 .startino = startino,
752 .sz_recs = xfs_inobt_walk_prefetch(inobt_records),
753 .pwork = XFS_PWORK_SINGLE_THREADED,
754 .lastino = NULLFSINO,
756 struct xfs_perag *pag;
757 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, startino);
760 ASSERT(agno < mp->m_sb.sb_agcount);
761 ASSERT(!(flags & ~XFS_INOBT_WALK_FLAGS_ALL));
763 error = xfs_iwalk_alloc(&iwag);
767 for_each_perag_from(mp, agno, pag) {
769 error = xfs_iwalk_ag(&iwag);
772 iwag.startino = XFS_AGINO_TO_INO(mp, pag->pag_agno + 1, 0);
773 if (flags & XFS_INOBT_WALK_SAME_AG)
780 xfs_iwalk_free(&iwag);