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;
89 * Loop over all clusters in a chunk for a given incore inode allocation btree
90 * record. Do a readahead if there are any allocated inodes in that cluster.
95 struct xfs_perag *pag,
96 struct xfs_inobt_rec_incore *irec)
98 struct xfs_ino_geometry *igeo = M_IGEO(mp);
100 struct blk_plug plug;
101 int i; /* inode chunk index */
103 agbno = XFS_AGINO_TO_AGBNO(mp, irec->ir_startino);
105 blk_start_plug(&plug);
106 for (i = 0; i < XFS_INODES_PER_CHUNK; i += igeo->inodes_per_cluster) {
109 imask = xfs_inobt_maskn(i, igeo->inodes_per_cluster);
110 if (imask & ~irec->ir_free) {
111 xfs_btree_reada_bufs(mp, pag->pag_agno, agbno,
112 igeo->blocks_per_cluster,
115 agbno += igeo->blocks_per_cluster;
117 blk_finish_plug(&plug);
121 * Set the bits in @irec's free mask that correspond to the inodes before
122 * @agino so that we skip them. This is how we restart an inode walk that was
123 * interrupted in the middle of an inode record.
126 xfs_iwalk_adjust_start(
127 xfs_agino_t agino, /* starting inode of chunk */
128 struct xfs_inobt_rec_incore *irec) /* btree record */
130 int idx; /* index into inode chunk */
133 idx = agino - irec->ir_startino;
136 * We got a right chunk with some left inodes allocated at it. Grab
137 * the chunk record. Mark all the uninteresting inodes free because
138 * they're before our start point.
140 for (i = 0; i < idx; i++) {
141 if (XFS_INOBT_MASK(i) & ~irec->ir_free)
142 irec->ir_freecount++;
145 irec->ir_free |= xfs_inobt_maskn(0, idx);
148 /* Allocate memory for a walk. */
151 struct xfs_iwalk_ag *iwag)
155 ASSERT(iwag->recs == NULL);
158 /* Allocate a prefetch buffer for inobt records. */
159 size = iwag->sz_recs * sizeof(struct xfs_inobt_rec_incore);
160 iwag->recs = kmem_alloc(size, KM_MAYFAIL);
161 if (iwag->recs == NULL)
167 /* Free memory we allocated for a walk. */
170 struct xfs_iwalk_ag *iwag)
172 kmem_free(iwag->recs);
176 /* For each inuse inode in each cached inobt record, call our function. */
179 struct xfs_iwalk_ag *iwag)
181 struct xfs_mount *mp = iwag->mp;
182 struct xfs_trans *tp = iwag->tp;
183 struct xfs_perag *pag = iwag->pag;
188 for (i = 0; i < iwag->nr_recs; i++) {
189 struct xfs_inobt_rec_incore *irec = &iwag->recs[i];
191 trace_xfs_iwalk_ag_rec(mp, pag->pag_agno, irec);
193 if (xfs_pwork_want_abort(&iwag->pwork))
196 if (iwag->inobt_walk_fn) {
197 error = iwag->inobt_walk_fn(mp, tp, pag->pag_agno, irec,
206 for (j = 0; j < XFS_INODES_PER_CHUNK; j++) {
207 if (xfs_pwork_want_abort(&iwag->pwork))
210 /* Skip if this inode is free */
211 if (XFS_INOBT_MASK(j) & irec->ir_free)
214 /* Otherwise call our function. */
215 ino = XFS_AGINO_TO_INO(mp, pag->pag_agno,
216 irec->ir_startino + j);
217 error = iwag->iwalk_fn(mp, tp, ino, iwag->data);
226 /* Delete cursor and let go of AGI. */
229 struct xfs_trans *tp,
230 struct xfs_btree_cur **curpp,
231 struct xfs_buf **agi_bpp,
235 xfs_btree_del_cursor(*curpp, error);
239 xfs_trans_brelse(tp, *agi_bpp);
245 * Set ourselves up for walking inobt records starting from a given point in
248 * If caller passed in a nonzero start inode number, load the record from the
249 * inobt and make the record look like all the inodes before agino are free so
250 * that we skip them, and then move the cursor to the next inobt record. This
251 * is how we support starting an iwalk in the middle of an inode chunk.
253 * If the caller passed in a start number of zero, move the cursor to the first
256 * The caller is responsible for cleaning up the cursor and buffer pointer
257 * regardless of the error status.
261 struct xfs_iwalk_ag *iwag,
263 struct xfs_btree_cur **curpp,
264 struct xfs_buf **agi_bpp,
267 struct xfs_mount *mp = iwag->mp;
268 struct xfs_trans *tp = iwag->tp;
269 struct xfs_perag *pag = iwag->pag;
270 struct xfs_inobt_rec_incore *irec;
273 /* Set up a fresh cursor and empty the inobt cache. */
275 error = xfs_inobt_cur(mp, tp, pag, XFS_BTNUM_INO, curpp, agi_bpp);
279 /* Starting at the beginning of the AG? That's easy! */
281 return xfs_inobt_lookup(*curpp, 0, XFS_LOOKUP_GE, has_more);
284 * Otherwise, we have to grab the inobt record where we left off, stuff
285 * the record into our cache, and then see if there are more records.
286 * We require a lookup cache of at least two elements so that the
287 * caller doesn't have to deal with tearing down the cursor to walk the
290 error = xfs_inobt_lookup(*curpp, agino, XFS_LOOKUP_LE, has_more);
295 * If the LE lookup at @agino yields no records, jump ahead to the
296 * inobt cursor increment to see if there are more records to process.
301 /* Get the record, should always work */
302 irec = &iwag->recs[iwag->nr_recs];
303 error = xfs_inobt_get_rec(*curpp, irec, has_more);
306 if (XFS_IS_CORRUPT(mp, *has_more != 1))
307 return -EFSCORRUPTED;
309 iwag->lastino = XFS_AGINO_TO_INO(mp, pag->pag_agno,
310 irec->ir_startino + XFS_INODES_PER_CHUNK - 1);
313 * If the LE lookup yielded an inobt record before the cursor position,
314 * skip it and see if there's another one after it.
316 if (irec->ir_startino + XFS_INODES_PER_CHUNK <= agino)
320 * If agino fell in the middle of the inode record, make it look like
321 * the inodes up to agino are free so that we don't return them again.
323 if (iwag->trim_start)
324 xfs_iwalk_adjust_start(agino, irec);
327 * The prefetch calculation is supposed to give us a large enough inobt
328 * record cache that grab_ichunk can stage a partial first record and
329 * the loop body can cache a record without having to check for cache
330 * space until after it reads an inobt record.
333 ASSERT(iwag->nr_recs < iwag->sz_recs);
336 return xfs_btree_increment(*curpp, 0, has_more);
340 * The inobt record cache is full, so preserve the inobt cursor state and
341 * run callbacks on the cached inobt records. When we're done, restore the
342 * cursor state to wherever the cursor would have been had the cache not been
343 * full (and therefore we could've just incremented the cursor) if *@has_more
344 * is true. On exit, *@has_more will indicate whether or not the caller should
345 * try for more inode records.
348 xfs_iwalk_run_callbacks(
349 struct xfs_iwalk_ag *iwag,
350 struct xfs_btree_cur **curpp,
351 struct xfs_buf **agi_bpp,
354 struct xfs_mount *mp = iwag->mp;
355 struct xfs_trans *tp = iwag->tp;
356 struct xfs_inobt_rec_incore *irec;
357 xfs_agino_t next_agino;
360 next_agino = XFS_INO_TO_AGINO(mp, iwag->lastino) + 1;
362 ASSERT(iwag->nr_recs > 0);
364 /* Delete cursor but remember the last record we cached... */
365 xfs_iwalk_del_inobt(tp, curpp, agi_bpp, 0);
366 irec = &iwag->recs[iwag->nr_recs - 1];
367 ASSERT(next_agino >= irec->ir_startino + XFS_INODES_PER_CHUNK);
369 error = xfs_iwalk_ag_recs(iwag);
373 /* ...empty the cache... */
379 /* ...and recreate the cursor just past where we left off. */
380 error = xfs_inobt_cur(mp, tp, iwag->pag, XFS_BTNUM_INO, curpp, agi_bpp);
384 return xfs_inobt_lookup(*curpp, next_agino, XFS_LOOKUP_GE, has_more);
387 /* Walk all inodes in a single AG, from @iwag->startino to the end of the AG. */
390 struct xfs_iwalk_ag *iwag)
392 struct xfs_mount *mp = iwag->mp;
393 struct xfs_trans *tp = iwag->tp;
394 struct xfs_perag *pag = iwag->pag;
395 struct xfs_buf *agi_bp = NULL;
396 struct xfs_btree_cur *cur = NULL;
401 /* Set up our cursor at the right place in the inode btree. */
402 ASSERT(pag->pag_agno == XFS_INO_TO_AGNO(mp, iwag->startino));
403 agino = XFS_INO_TO_AGINO(mp, iwag->startino);
404 error = xfs_iwalk_ag_start(iwag, agino, &cur, &agi_bp, &has_more);
406 while (!error && has_more) {
407 struct xfs_inobt_rec_incore *irec;
411 if (xfs_pwork_want_abort(&iwag->pwork))
414 /* Fetch the inobt record. */
415 irec = &iwag->recs[iwag->nr_recs];
416 error = xfs_inobt_get_rec(cur, irec, &has_more);
417 if (error || !has_more)
420 /* Make sure that we always move forward. */
421 rec_fsino = XFS_AGINO_TO_INO(mp, pag->pag_agno, irec->ir_startino);
422 if (iwag->lastino != NULLFSINO &&
423 XFS_IS_CORRUPT(mp, iwag->lastino >= rec_fsino)) {
424 error = -EFSCORRUPTED;
427 iwag->lastino = rec_fsino + XFS_INODES_PER_CHUNK - 1;
429 /* No allocated inodes in this chunk; skip it. */
430 if (iwag->skip_empty && irec->ir_freecount == irec->ir_count) {
431 error = xfs_btree_increment(cur, 0, &has_more);
438 * Start readahead for this inode chunk in anticipation of
439 * walking the inodes.
442 xfs_iwalk_ichunk_ra(mp, pag, irec);
445 * If there's space in the buffer for more records, increment
446 * the btree cursor and grab more.
448 if (++iwag->nr_recs < iwag->sz_recs) {
449 error = xfs_btree_increment(cur, 0, &has_more);
450 if (error || !has_more)
456 * Otherwise, we need to save cursor state and run the callback
457 * function on the cached records. The run_callbacks function
458 * is supposed to return a cursor pointing to the record where
459 * we would be if we had been able to increment like above.
462 error = xfs_iwalk_run_callbacks(iwag, &cur, &agi_bp, &has_more);
465 if (iwag->nr_recs == 0 || error)
468 /* Walk the unprocessed records in the cache. */
469 error = xfs_iwalk_run_callbacks(iwag, &cur, &agi_bp, &has_more);
472 xfs_iwalk_del_inobt(tp, &cur, &agi_bp, error);
477 * We experimentally determined that the reduction in ioctl call overhead
478 * diminishes when userspace asks for more than 2048 inodes, so we'll cap
479 * prefetch at this point.
481 #define IWALK_MAX_INODE_PREFETCH (2048U)
484 * Given the number of inodes to prefetch, set the number of inobt records that
485 * we cache in memory, which controls the number of inodes we try to read
486 * ahead. Set the maximum if @inodes == 0.
488 static inline unsigned int
492 unsigned int inobt_records;
495 * If the caller didn't tell us the number of inodes they wanted,
496 * assume the maximum prefetch possible for best performance.
497 * Otherwise, cap prefetch at that maximum so that we don't start an
498 * absurd amount of prefetch.
501 inodes = IWALK_MAX_INODE_PREFETCH;
502 inodes = min(inodes, IWALK_MAX_INODE_PREFETCH);
504 /* Round the inode count up to a full chunk. */
505 inodes = round_up(inodes, XFS_INODES_PER_CHUNK);
508 * In order to convert the number of inodes to prefetch into an
509 * estimate of the number of inobt records to cache, we require a
510 * conversion factor that reflects our expectations of the average
511 * loading factor of an inode chunk. Based on data gathered, most
512 * (but not all) filesystems manage to keep the inode chunks totally
513 * full, so we'll underestimate slightly so that our readahead will
514 * still deliver the performance we want on aging filesystems:
516 * inobt = inodes / (INODES_PER_CHUNK * (4 / 5));
518 * The funny math is to avoid integer division.
520 inobt_records = (inodes * 5) / (4 * XFS_INODES_PER_CHUNK);
523 * Allocate enough space to prefetch at least two inobt records so that
524 * we can cache both the record where the iwalk started and the next
525 * record. This simplifies the AG inode walk loop setup code.
527 return max(inobt_records, 2U);
531 * Walk all inodes in the filesystem starting from @startino. The @iwalk_fn
532 * will be called for each allocated inode, being passed the inode's number and
533 * @data. @max_prefetch controls how many inobt records' worth of inodes we
538 struct xfs_mount *mp,
539 struct xfs_trans *tp,
542 xfs_iwalk_fn iwalk_fn,
543 unsigned int inode_records,
546 struct xfs_iwalk_ag iwag = {
549 .iwalk_fn = iwalk_fn,
551 .startino = startino,
552 .sz_recs = xfs_iwalk_prefetch(inode_records),
555 .pwork = XFS_PWORK_SINGLE_THREADED,
556 .lastino = NULLFSINO,
558 struct xfs_perag *pag;
559 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, startino);
562 ASSERT(agno < mp->m_sb.sb_agcount);
563 ASSERT(!(flags & ~XFS_IWALK_FLAGS_ALL));
565 error = xfs_iwalk_alloc(&iwag);
569 for_each_perag_from(mp, agno, pag) {
571 error = xfs_iwalk_ag(&iwag);
574 iwag.startino = XFS_AGINO_TO_INO(mp, agno + 1, 0);
575 if (flags & XFS_INOBT_WALK_SAME_AG)
582 xfs_iwalk_free(&iwag);
586 /* Run per-thread iwalk work. */
589 struct xfs_mount *mp,
590 struct xfs_pwork *pwork)
592 struct xfs_iwalk_ag *iwag;
595 iwag = container_of(pwork, struct xfs_iwalk_ag, pwork);
596 if (xfs_pwork_want_abort(pwork))
599 error = xfs_iwalk_alloc(iwag);
603 error = xfs_iwalk_ag(iwag);
604 xfs_iwalk_free(iwag);
606 xfs_perag_put(iwag->pag);
612 * Walk all the inodes in the filesystem using multiple threads to process each
617 struct xfs_mount *mp,
620 xfs_iwalk_fn iwalk_fn,
621 unsigned int inode_records,
625 struct xfs_pwork_ctl pctl;
626 struct xfs_perag *pag;
627 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, startino);
630 ASSERT(agno < mp->m_sb.sb_agcount);
631 ASSERT(!(flags & ~XFS_IWALK_FLAGS_ALL));
633 error = xfs_pwork_init(mp, &pctl, xfs_iwalk_ag_work, "xfs_iwalk");
637 for_each_perag_from(mp, agno, pag) {
638 struct xfs_iwalk_ag *iwag;
640 if (xfs_pwork_ctl_want_abort(&pctl))
643 iwag = kmem_zalloc(sizeof(struct xfs_iwalk_ag), 0);
647 * perag is being handed off to async work, so take another
648 * reference for the async work to release.
650 atomic_inc(&pag->pag_ref);
652 iwag->iwalk_fn = iwalk_fn;
654 iwag->startino = startino;
655 iwag->sz_recs = xfs_iwalk_prefetch(inode_records);
656 iwag->lastino = NULLFSINO;
657 xfs_pwork_queue(&pctl, &iwag->pwork);
658 startino = XFS_AGINO_TO_INO(mp, pag->pag_agno + 1, 0);
659 if (flags & XFS_INOBT_WALK_SAME_AG)
665 xfs_pwork_poll(&pctl);
666 return xfs_pwork_destroy(&pctl);
670 * Allow callers to cache up to a page's worth of inobt records. This reflects
671 * the existing inumbers prefetching behavior. Since the inobt walk does not
672 * itself do anything with the inobt records, we can set a fairly high limit
675 #define MAX_INOBT_WALK_PREFETCH \
676 (PAGE_SIZE / sizeof(struct xfs_inobt_rec_incore))
679 * Given the number of records that the user wanted, set the number of inobt
680 * records that we buffer in memory. Set the maximum if @inobt_records == 0.
682 static inline unsigned int
683 xfs_inobt_walk_prefetch(
684 unsigned int inobt_records)
687 * If the caller didn't tell us the number of inobt records they
688 * wanted, assume the maximum prefetch possible for best performance.
690 if (inobt_records == 0)
691 inobt_records = MAX_INOBT_WALK_PREFETCH;
694 * Allocate enough space to prefetch at least two inobt records so that
695 * we can cache both the record where the iwalk started and the next
696 * record. This simplifies the AG inode walk loop setup code.
698 inobt_records = max(inobt_records, 2U);
701 * Cap prefetch at that maximum so that we don't use an absurd amount
704 return min_t(unsigned int, inobt_records, MAX_INOBT_WALK_PREFETCH);
708 * Walk all inode btree records in the filesystem starting from @startino. The
709 * @inobt_walk_fn will be called for each btree record, being passed the incore
710 * record and @data. @max_prefetch controls how many inobt records we try to
711 * cache ahead of time.
715 struct xfs_mount *mp,
716 struct xfs_trans *tp,
719 xfs_inobt_walk_fn inobt_walk_fn,
720 unsigned int inobt_records,
723 struct xfs_iwalk_ag iwag = {
726 .inobt_walk_fn = inobt_walk_fn,
728 .startino = startino,
729 .sz_recs = xfs_inobt_walk_prefetch(inobt_records),
730 .pwork = XFS_PWORK_SINGLE_THREADED,
731 .lastino = NULLFSINO,
733 struct xfs_perag *pag;
734 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, startino);
737 ASSERT(agno < mp->m_sb.sb_agcount);
738 ASSERT(!(flags & ~XFS_INOBT_WALK_FLAGS_ALL));
740 error = xfs_iwalk_alloc(&iwag);
744 for_each_perag_from(mp, agno, pag) {
746 error = xfs_iwalk_ag(&iwag);
749 iwag.startino = XFS_AGINO_TO_INO(mp, pag->pag_agno + 1, 0);
750 if (flags & XFS_INOBT_WALK_SAME_AG)
757 xfs_iwalk_free(&iwag);