1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright (C) 2019 Oracle. All Rights Reserved.
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"
26 * Walking Inodes in the Filesystem
27 * ================================
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
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.
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
49 /* parallel work control data; will be null if single threaded */
50 struct xfs_pwork pwork;
55 /* Where do we start the traversal? */
58 /* What was the last inode number we saw when iterating the inobt? */
61 /* Array of inobt records we cache. */
62 struct xfs_inobt_rec_incore *recs;
64 /* Number of entries allocated for the @recs array. */
67 /* Number of entries in the @recs array that are in use. */
70 /* Inode walk function and data pointer. */
71 xfs_iwalk_fn iwalk_fn;
72 xfs_inobt_walk_fn inobt_walk_fn;
76 * Make it look like the inodes up to startino are free so that
77 * bulkstat can start its inode iteration at the correct place without
78 * needing to special case everywhere.
80 unsigned int trim_start:1;
82 /* Skip empty inobt records? */
83 unsigned int skip_empty:1;
87 * Loop over all clusters in a chunk for a given incore inode allocation btree
88 * record. Do a readahead if there are any allocated inodes in that cluster.
94 struct xfs_inobt_rec_incore *irec)
96 struct xfs_ino_geometry *igeo = M_IGEO(mp);
99 int i; /* inode chunk index */
101 agbno = XFS_AGINO_TO_AGBNO(mp, irec->ir_startino);
103 blk_start_plug(&plug);
104 for (i = 0; i < XFS_INODES_PER_CHUNK; i += igeo->inodes_per_cluster) {
107 imask = xfs_inobt_maskn(i, igeo->inodes_per_cluster);
108 if (imask & ~irec->ir_free) {
109 xfs_btree_reada_bufs(mp, agno, agbno,
110 igeo->blocks_per_cluster,
113 agbno += igeo->blocks_per_cluster;
115 blk_finish_plug(&plug);
119 * Set the bits in @irec's free mask that correspond to the inodes before
120 * @agino so that we skip them. This is how we restart an inode walk that was
121 * interrupted in the middle of an inode record.
124 xfs_iwalk_adjust_start(
125 xfs_agino_t agino, /* starting inode of chunk */
126 struct xfs_inobt_rec_incore *irec) /* btree record */
128 int idx; /* index into inode chunk */
131 idx = agino - irec->ir_startino;
134 * We got a right chunk with some left inodes allocated at it. Grab
135 * the chunk record. Mark all the uninteresting inodes free because
136 * they're before our start point.
138 for (i = 0; i < idx; i++) {
139 if (XFS_INOBT_MASK(i) & ~irec->ir_free)
140 irec->ir_freecount++;
143 irec->ir_free |= xfs_inobt_maskn(0, idx);
146 /* Allocate memory for a walk. */
149 struct xfs_iwalk_ag *iwag)
153 ASSERT(iwag->recs == NULL);
156 /* Allocate a prefetch buffer for inobt records. */
157 size = iwag->sz_recs * sizeof(struct xfs_inobt_rec_incore);
158 iwag->recs = kmem_alloc(size, KM_MAYFAIL);
159 if (iwag->recs == NULL)
165 /* Free memory we allocated for a walk. */
168 struct xfs_iwalk_ag *iwag)
170 kmem_free(iwag->recs);
174 /* For each inuse inode in each cached inobt record, call our function. */
177 struct xfs_iwalk_ag *iwag)
179 struct xfs_mount *mp = iwag->mp;
180 struct xfs_trans *tp = iwag->tp;
186 agno = XFS_INO_TO_AGNO(mp, iwag->startino);
187 for (i = 0; i < iwag->nr_recs; i++) {
188 struct xfs_inobt_rec_incore *irec = &iwag->recs[i];
190 trace_xfs_iwalk_ag_rec(mp, agno, irec);
192 if (xfs_pwork_want_abort(&iwag->pwork))
195 if (iwag->inobt_walk_fn) {
196 error = iwag->inobt_walk_fn(mp, tp, agno, irec,
205 for (j = 0; j < XFS_INODES_PER_CHUNK; j++) {
206 if (xfs_pwork_want_abort(&iwag->pwork))
209 /* Skip if this inode is free */
210 if (XFS_INOBT_MASK(j) & irec->ir_free)
213 /* Otherwise call our function. */
214 ino = XFS_AGINO_TO_INO(mp, agno, irec->ir_startino + j);
215 error = iwag->iwalk_fn(mp, tp, ino, iwag->data);
224 /* Delete cursor and let go of AGI. */
227 struct xfs_trans *tp,
228 struct xfs_btree_cur **curpp,
229 struct xfs_buf **agi_bpp,
233 xfs_btree_del_cursor(*curpp, error);
237 xfs_trans_brelse(tp, *agi_bpp);
243 * Set ourselves up for walking inobt records starting from a given point in
246 * If caller passed in a nonzero start inode number, load the record from the
247 * inobt and make the record look like all the inodes before agino are free so
248 * that we skip them, and then move the cursor to the next inobt record. This
249 * is how we support starting an iwalk in the middle of an inode chunk.
251 * If the caller passed in a start number of zero, move the cursor to the first
254 * The caller is responsible for cleaning up the cursor and buffer pointer
255 * regardless of the error status.
259 struct xfs_iwalk_ag *iwag,
262 struct xfs_btree_cur **curpp,
263 struct xfs_buf **agi_bpp,
266 struct xfs_mount *mp = iwag->mp;
267 struct xfs_trans *tp = iwag->tp;
268 struct xfs_inobt_rec_incore *irec;
271 /* Set up a fresh cursor and empty the inobt cache. */
273 error = xfs_inobt_cur(mp, tp, agno, XFS_BTNUM_INO, curpp, agi_bpp);
277 /* Starting at the beginning of the AG? That's easy! */
279 return xfs_inobt_lookup(*curpp, 0, XFS_LOOKUP_GE, has_more);
282 * Otherwise, we have to grab the inobt record where we left off, stuff
283 * the record into our cache, and then see if there are more records.
284 * We require a lookup cache of at least two elements so that the
285 * caller doesn't have to deal with tearing down the cursor to walk the
288 error = xfs_inobt_lookup(*curpp, agino, XFS_LOOKUP_LE, has_more);
293 * If the LE lookup at @agino yields no records, jump ahead to the
294 * inobt cursor increment to see if there are more records to process.
299 /* Get the record, should always work */
300 irec = &iwag->recs[iwag->nr_recs];
301 error = xfs_inobt_get_rec(*curpp, irec, has_more);
304 if (XFS_IS_CORRUPT(mp, *has_more != 1))
305 return -EFSCORRUPTED;
307 iwag->lastino = XFS_AGINO_TO_INO(mp, agno,
308 irec->ir_startino + XFS_INODES_PER_CHUNK - 1);
311 * If the LE lookup yielded an inobt record before the cursor position,
312 * skip it and see if there's another one after it.
314 if (irec->ir_startino + XFS_INODES_PER_CHUNK <= agino)
318 * If agino fell in the middle of the inode record, make it look like
319 * the inodes up to agino are free so that we don't return them again.
321 if (iwag->trim_start)
322 xfs_iwalk_adjust_start(agino, irec);
325 * The prefetch calculation is supposed to give us a large enough inobt
326 * record cache that grab_ichunk can stage a partial first record and
327 * the loop body can cache a record without having to check for cache
328 * space until after it reads an inobt record.
331 ASSERT(iwag->nr_recs < iwag->sz_recs);
334 return xfs_btree_increment(*curpp, 0, has_more);
338 * The inobt record cache is full, so preserve the inobt cursor state and
339 * run callbacks on the cached inobt records. When we're done, restore the
340 * cursor state to wherever the cursor would have been had the cache not been
341 * full (and therefore we could've just incremented the cursor) if *@has_more
342 * is true. On exit, *@has_more will indicate whether or not the caller should
343 * try for more inode records.
346 xfs_iwalk_run_callbacks(
347 struct xfs_iwalk_ag *iwag,
349 struct xfs_btree_cur **curpp,
350 struct xfs_buf **agi_bpp,
353 struct xfs_mount *mp = iwag->mp;
354 struct xfs_trans *tp = iwag->tp;
355 struct xfs_inobt_rec_incore *irec;
356 xfs_agino_t next_agino;
359 next_agino = XFS_INO_TO_AGINO(mp, iwag->lastino) + 1;
361 ASSERT(iwag->nr_recs > 0);
363 /* Delete cursor but remember the last record we cached... */
364 xfs_iwalk_del_inobt(tp, curpp, agi_bpp, 0);
365 irec = &iwag->recs[iwag->nr_recs - 1];
366 ASSERT(next_agino >= irec->ir_startino + XFS_INODES_PER_CHUNK);
368 error = xfs_iwalk_ag_recs(iwag);
372 /* ...empty the cache... */
378 /* ...and recreate the cursor just past where we left off. */
379 error = xfs_inobt_cur(mp, tp, agno, XFS_BTNUM_INO, curpp, agi_bpp);
383 return xfs_inobt_lookup(*curpp, next_agino, XFS_LOOKUP_GE, has_more);
386 /* Walk all inodes in a single AG, from @iwag->startino to the end of the AG. */
389 struct xfs_iwalk_ag *iwag)
391 struct xfs_mount *mp = iwag->mp;
392 struct xfs_trans *tp = iwag->tp;
393 struct xfs_buf *agi_bp = NULL;
394 struct xfs_btree_cur *cur = NULL;
400 /* Set up our cursor at the right place in the inode btree. */
401 agno = XFS_INO_TO_AGNO(mp, iwag->startino);
402 agino = XFS_INO_TO_AGINO(mp, iwag->startino);
403 error = xfs_iwalk_ag_start(iwag, agno, agino, &cur, &agi_bp, &has_more);
405 while (!error && has_more) {
406 struct xfs_inobt_rec_incore *irec;
410 if (xfs_pwork_want_abort(&iwag->pwork))
413 /* Fetch the inobt record. */
414 irec = &iwag->recs[iwag->nr_recs];
415 error = xfs_inobt_get_rec(cur, irec, &has_more);
416 if (error || !has_more)
419 /* Make sure that we always move forward. */
420 rec_fsino = XFS_AGINO_TO_INO(mp, agno, irec->ir_startino);
421 if (iwag->lastino != NULLFSINO &&
422 XFS_IS_CORRUPT(mp, iwag->lastino >= rec_fsino)) {
423 error = -EFSCORRUPTED;
426 iwag->lastino = rec_fsino + XFS_INODES_PER_CHUNK - 1;
428 /* No allocated inodes in this chunk; skip it. */
429 if (iwag->skip_empty && irec->ir_freecount == irec->ir_count) {
430 error = xfs_btree_increment(cur, 0, &has_more);
437 * Start readahead for this inode chunk in anticipation of
438 * walking the inodes.
441 xfs_iwalk_ichunk_ra(mp, agno, irec);
444 * If there's space in the buffer for more records, increment
445 * the btree cursor and grab more.
447 if (++iwag->nr_recs < iwag->sz_recs) {
448 error = xfs_btree_increment(cur, 0, &has_more);
449 if (error || !has_more)
455 * Otherwise, we need to save cursor state and run the callback
456 * function on the cached records. The run_callbacks function
457 * is supposed to return a cursor pointing to the record where
458 * we would be if we had been able to increment like above.
461 error = xfs_iwalk_run_callbacks(iwag, agno, &cur, &agi_bp,
465 if (iwag->nr_recs == 0 || error)
468 /* Walk the unprocessed records in the cache. */
469 error = xfs_iwalk_run_callbacks(iwag, agno, &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 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, startino);
561 ASSERT(agno < mp->m_sb.sb_agcount);
562 ASSERT(!(flags & ~XFS_IWALK_FLAGS_ALL));
564 error = xfs_iwalk_alloc(&iwag);
568 for (; agno < mp->m_sb.sb_agcount; agno++) {
569 error = xfs_iwalk_ag(&iwag);
572 iwag.startino = XFS_AGINO_TO_INO(mp, agno + 1, 0);
573 if (flags & XFS_INOBT_WALK_SAME_AG)
577 xfs_iwalk_free(&iwag);
581 /* Run per-thread iwalk work. */
584 struct xfs_mount *mp,
585 struct xfs_pwork *pwork)
587 struct xfs_iwalk_ag *iwag;
590 iwag = container_of(pwork, struct xfs_iwalk_ag, pwork);
591 if (xfs_pwork_want_abort(pwork))
594 error = xfs_iwalk_alloc(iwag);
598 error = xfs_iwalk_ag(iwag);
599 xfs_iwalk_free(iwag);
606 * Walk all the inodes in the filesystem using multiple threads to process each
611 struct xfs_mount *mp,
614 xfs_iwalk_fn iwalk_fn,
615 unsigned int inode_records,
619 struct xfs_pwork_ctl pctl;
620 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, startino);
621 unsigned int nr_threads;
624 ASSERT(agno < mp->m_sb.sb_agcount);
625 ASSERT(!(flags & ~XFS_IWALK_FLAGS_ALL));
627 nr_threads = xfs_pwork_guess_datadev_parallelism(mp);
628 error = xfs_pwork_init(mp, &pctl, xfs_iwalk_ag_work, "xfs_iwalk",
633 for (; agno < mp->m_sb.sb_agcount; agno++) {
634 struct xfs_iwalk_ag *iwag;
636 if (xfs_pwork_ctl_want_abort(&pctl))
639 iwag = kmem_zalloc(sizeof(struct xfs_iwalk_ag), 0);
641 iwag->iwalk_fn = iwalk_fn;
643 iwag->startino = startino;
644 iwag->sz_recs = xfs_iwalk_prefetch(inode_records);
645 iwag->lastino = NULLFSINO;
646 xfs_pwork_queue(&pctl, &iwag->pwork);
647 startino = XFS_AGINO_TO_INO(mp, agno + 1, 0);
648 if (flags & XFS_INOBT_WALK_SAME_AG)
653 xfs_pwork_poll(&pctl);
654 return xfs_pwork_destroy(&pctl);
658 * Allow callers to cache up to a page's worth of inobt records. This reflects
659 * the existing inumbers prefetching behavior. Since the inobt walk does not
660 * itself do anything with the inobt records, we can set a fairly high limit
663 #define MAX_INOBT_WALK_PREFETCH \
664 (PAGE_SIZE / sizeof(struct xfs_inobt_rec_incore))
667 * Given the number of records that the user wanted, set the number of inobt
668 * records that we buffer in memory. Set the maximum if @inobt_records == 0.
670 static inline unsigned int
671 xfs_inobt_walk_prefetch(
672 unsigned int inobt_records)
675 * If the caller didn't tell us the number of inobt records they
676 * wanted, assume the maximum prefetch possible for best performance.
678 if (inobt_records == 0)
679 inobt_records = MAX_INOBT_WALK_PREFETCH;
682 * Allocate enough space to prefetch at least two inobt records so that
683 * we can cache both the record where the iwalk started and the next
684 * record. This simplifies the AG inode walk loop setup code.
686 inobt_records = max(inobt_records, 2U);
689 * Cap prefetch at that maximum so that we don't use an absurd amount
692 return min_t(unsigned int, inobt_records, MAX_INOBT_WALK_PREFETCH);
696 * Walk all inode btree records in the filesystem starting from @startino. The
697 * @inobt_walk_fn will be called for each btree record, being passed the incore
698 * record and @data. @max_prefetch controls how many inobt records we try to
699 * cache ahead of time.
703 struct xfs_mount *mp,
704 struct xfs_trans *tp,
707 xfs_inobt_walk_fn inobt_walk_fn,
708 unsigned int inobt_records,
711 struct xfs_iwalk_ag iwag = {
714 .inobt_walk_fn = inobt_walk_fn,
716 .startino = startino,
717 .sz_recs = xfs_inobt_walk_prefetch(inobt_records),
718 .pwork = XFS_PWORK_SINGLE_THREADED,
719 .lastino = NULLFSINO,
721 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, startino);
724 ASSERT(agno < mp->m_sb.sb_agcount);
725 ASSERT(!(flags & ~XFS_INOBT_WALK_FLAGS_ALL));
727 error = xfs_iwalk_alloc(&iwag);
731 for (; agno < mp->m_sb.sb_agcount; agno++) {
732 error = xfs_iwalk_ag(&iwag);
735 iwag.startino = XFS_AGINO_TO_INO(mp, agno + 1, 0);
736 if (flags & XFS_INOBT_WALK_SAME_AG)
740 xfs_iwalk_free(&iwag);