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"
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 /* Array of inobt records we cache. */
59 struct xfs_inobt_rec_incore *recs;
61 /* Number of entries allocated for the @recs array. */
64 /* Number of entries in the @recs array that are in use. */
67 /* Inode walk function and data pointer. */
68 xfs_iwalk_fn iwalk_fn;
69 xfs_inobt_walk_fn inobt_walk_fn;
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.
77 unsigned int trim_start:1;
79 /* Skip empty inobt records? */
80 unsigned int skip_empty:1;
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.
91 struct xfs_inobt_rec_incore *irec)
93 struct xfs_ino_geometry *igeo = M_IGEO(mp);
96 int i; /* inode chunk index */
98 agbno = XFS_AGINO_TO_AGBNO(mp, irec->ir_startino);
100 blk_start_plug(&plug);
101 for (i = 0; i < XFS_INODES_PER_CHUNK; i += igeo->inodes_per_cluster) {
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,
110 agbno += igeo->blocks_per_cluster;
112 blk_finish_plug(&plug);
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.
121 xfs_iwalk_adjust_start(
122 xfs_agino_t agino, /* starting inode of chunk */
123 struct xfs_inobt_rec_incore *irec) /* btree record */
125 int idx; /* index into inode chunk */
128 idx = agino - irec->ir_startino;
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.
135 for (i = 0; i < idx; i++) {
136 if (XFS_INOBT_MASK(i) & ~irec->ir_free)
137 irec->ir_freecount++;
140 irec->ir_free |= xfs_inobt_maskn(0, idx);
143 /* Allocate memory for a walk. */
146 struct xfs_iwalk_ag *iwag)
150 ASSERT(iwag->recs == NULL);
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)
162 /* Free memory we allocated for a walk. */
165 struct xfs_iwalk_ag *iwag)
167 kmem_free(iwag->recs);
171 /* For each inuse inode in each cached inobt record, call our function. */
174 struct xfs_iwalk_ag *iwag)
176 struct xfs_mount *mp = iwag->mp;
177 struct xfs_trans *tp = iwag->tp;
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];
187 trace_xfs_iwalk_ag_rec(mp, agno, irec);
189 if (xfs_pwork_want_abort(&iwag->pwork))
192 if (iwag->inobt_walk_fn) {
193 error = iwag->inobt_walk_fn(mp, tp, agno, irec,
202 for (j = 0; j < XFS_INODES_PER_CHUNK; j++) {
203 if (xfs_pwork_want_abort(&iwag->pwork))
206 /* Skip if this inode is free */
207 if (XFS_INOBT_MASK(j) & irec->ir_free)
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);
221 /* Delete cursor and let go of AGI. */
224 struct xfs_trans *tp,
225 struct xfs_btree_cur **curpp,
226 struct xfs_buf **agi_bpp,
230 xfs_btree_del_cursor(*curpp, error);
234 xfs_trans_brelse(tp, *agi_bpp);
240 * Set ourselves up for walking inobt records starting from a given point in
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.
248 * If the caller passed in a start number of zero, move the cursor to the first
251 * The caller is responsible for cleaning up the cursor and buffer pointer
252 * regardless of the error status.
256 struct xfs_iwalk_ag *iwag,
259 struct xfs_btree_cur **curpp,
260 struct xfs_buf **agi_bpp,
263 struct xfs_mount *mp = iwag->mp;
264 struct xfs_trans *tp = iwag->tp;
265 struct xfs_inobt_rec_incore *irec;
268 /* Set up a fresh cursor and empty the inobt cache. */
270 error = xfs_inobt_cur(mp, tp, agno, XFS_BTNUM_INO, curpp, agi_bpp);
274 /* Starting at the beginning of the AG? That's easy! */
276 return xfs_inobt_lookup(*curpp, 0, XFS_LOOKUP_GE, has_more);
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
285 error = xfs_inobt_lookup(*curpp, agino, XFS_LOOKUP_LE, has_more);
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.
296 /* Get the record, should always work */
297 irec = &iwag->recs[iwag->nr_recs];
298 error = xfs_inobt_get_rec(*curpp, irec, has_more);
301 XFS_WANT_CORRUPTED_RETURN(mp, *has_more == 1);
304 * If the LE lookup yielded an inobt record before the cursor position,
305 * skip it and see if there's another one after it.
307 if (irec->ir_startino + XFS_INODES_PER_CHUNK <= agino)
311 * If agino fell in the middle of the inode record, make it look like
312 * the inodes up to agino are free so that we don't return them again.
314 if (iwag->trim_start)
315 xfs_iwalk_adjust_start(agino, irec);
318 * The prefetch calculation is supposed to give us a large enough inobt
319 * record cache that grab_ichunk can stage a partial first record and
320 * the loop body can cache a record without having to check for cache
321 * space until after it reads an inobt record.
324 ASSERT(iwag->nr_recs < iwag->sz_recs);
327 return xfs_btree_increment(*curpp, 0, has_more);
331 * The inobt record cache is full, so preserve the inobt cursor state and
332 * run callbacks on the cached inobt records. When we're done, restore the
333 * cursor state to wherever the cursor would have been had the cache not been
334 * full (and therefore we could've just incremented the cursor) if *@has_more
335 * is true. On exit, *@has_more will indicate whether or not the caller should
336 * try for more inode records.
339 xfs_iwalk_run_callbacks(
340 struct xfs_iwalk_ag *iwag,
342 struct xfs_btree_cur **curpp,
343 struct xfs_buf **agi_bpp,
346 struct xfs_mount *mp = iwag->mp;
347 struct xfs_trans *tp = iwag->tp;
348 struct xfs_inobt_rec_incore *irec;
352 ASSERT(iwag->nr_recs > 0);
354 /* Delete cursor but remember the last record we cached... */
355 xfs_iwalk_del_inobt(tp, curpp, agi_bpp, 0);
356 irec = &iwag->recs[iwag->nr_recs - 1];
357 restart = irec->ir_startino + XFS_INODES_PER_CHUNK - 1;
359 error = xfs_iwalk_ag_recs(iwag);
363 /* ...empty the cache... */
369 /* ...and recreate the cursor just past where we left off. */
370 error = xfs_inobt_cur(mp, tp, agno, XFS_BTNUM_INO, curpp, agi_bpp);
374 return xfs_inobt_lookup(*curpp, restart, XFS_LOOKUP_GE, has_more);
377 /* Walk all inodes in a single AG, from @iwag->startino to the end of the AG. */
380 struct xfs_iwalk_ag *iwag)
382 struct xfs_mount *mp = iwag->mp;
383 struct xfs_trans *tp = iwag->tp;
384 struct xfs_buf *agi_bp = NULL;
385 struct xfs_btree_cur *cur = NULL;
391 /* Set up our cursor at the right place in the inode btree. */
392 agno = XFS_INO_TO_AGNO(mp, iwag->startino);
393 agino = XFS_INO_TO_AGINO(mp, iwag->startino);
394 error = xfs_iwalk_ag_start(iwag, agno, agino, &cur, &agi_bp, &has_more);
396 while (!error && has_more) {
397 struct xfs_inobt_rec_incore *irec;
400 if (xfs_pwork_want_abort(&iwag->pwork))
403 /* Fetch the inobt record. */
404 irec = &iwag->recs[iwag->nr_recs];
405 error = xfs_inobt_get_rec(cur, irec, &has_more);
406 if (error || !has_more)
409 /* No allocated inodes in this chunk; skip it. */
410 if (iwag->skip_empty && irec->ir_freecount == irec->ir_count) {
411 error = xfs_btree_increment(cur, 0, &has_more);
418 * Start readahead for this inode chunk in anticipation of
419 * walking the inodes.
422 xfs_iwalk_ichunk_ra(mp, agno, irec);
425 * If there's space in the buffer for more records, increment
426 * the btree cursor and grab more.
428 if (++iwag->nr_recs < iwag->sz_recs) {
429 error = xfs_btree_increment(cur, 0, &has_more);
430 if (error || !has_more)
436 * Otherwise, we need to save cursor state and run the callback
437 * function on the cached records. The run_callbacks function
438 * is supposed to return a cursor pointing to the record where
439 * we would be if we had been able to increment like above.
442 error = xfs_iwalk_run_callbacks(iwag, agno, &cur, &agi_bp,
446 if (iwag->nr_recs == 0 || error)
449 /* Walk the unprocessed records in the cache. */
450 error = xfs_iwalk_run_callbacks(iwag, agno, &cur, &agi_bp, &has_more);
453 xfs_iwalk_del_inobt(tp, &cur, &agi_bp, error);
458 * We experimentally determined that the reduction in ioctl call overhead
459 * diminishes when userspace asks for more than 2048 inodes, so we'll cap
460 * prefetch at this point.
462 #define IWALK_MAX_INODE_PREFETCH (2048U)
465 * Given the number of inodes to prefetch, set the number of inobt records that
466 * we cache in memory, which controls the number of inodes we try to read
467 * ahead. Set the maximum if @inodes == 0.
469 static inline unsigned int
473 unsigned int inobt_records;
476 * If the caller didn't tell us the number of inodes they wanted,
477 * assume the maximum prefetch possible for best performance.
478 * Otherwise, cap prefetch at that maximum so that we don't start an
479 * absurd amount of prefetch.
482 inodes = IWALK_MAX_INODE_PREFETCH;
483 inodes = min(inodes, IWALK_MAX_INODE_PREFETCH);
485 /* Round the inode count up to a full chunk. */
486 inodes = round_up(inodes, XFS_INODES_PER_CHUNK);
489 * In order to convert the number of inodes to prefetch into an
490 * estimate of the number of inobt records to cache, we require a
491 * conversion factor that reflects our expectations of the average
492 * loading factor of an inode chunk. Based on data gathered, most
493 * (but not all) filesystems manage to keep the inode chunks totally
494 * full, so we'll underestimate slightly so that our readahead will
495 * still deliver the performance we want on aging filesystems:
497 * inobt = inodes / (INODES_PER_CHUNK * (4 / 5));
499 * The funny math is to avoid integer division.
501 inobt_records = (inodes * 5) / (4 * XFS_INODES_PER_CHUNK);
504 * Allocate enough space to prefetch at least two inobt records so that
505 * we can cache both the record where the iwalk started and the next
506 * record. This simplifies the AG inode walk loop setup code.
508 return max(inobt_records, 2U);
512 * Walk all inodes in the filesystem starting from @startino. The @iwalk_fn
513 * will be called for each allocated inode, being passed the inode's number and
514 * @data. @max_prefetch controls how many inobt records' worth of inodes we
519 struct xfs_mount *mp,
520 struct xfs_trans *tp,
523 xfs_iwalk_fn iwalk_fn,
524 unsigned int inode_records,
527 struct xfs_iwalk_ag iwag = {
530 .iwalk_fn = iwalk_fn,
532 .startino = startino,
533 .sz_recs = xfs_iwalk_prefetch(inode_records),
536 .pwork = XFS_PWORK_SINGLE_THREADED,
538 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, startino);
541 ASSERT(agno < mp->m_sb.sb_agcount);
542 ASSERT(!(flags & ~XFS_IWALK_FLAGS_ALL));
544 error = xfs_iwalk_alloc(&iwag);
548 for (; agno < mp->m_sb.sb_agcount; agno++) {
549 error = xfs_iwalk_ag(&iwag);
552 iwag.startino = XFS_AGINO_TO_INO(mp, agno + 1, 0);
553 if (flags & XFS_INOBT_WALK_SAME_AG)
557 xfs_iwalk_free(&iwag);
561 /* Run per-thread iwalk work. */
564 struct xfs_mount *mp,
565 struct xfs_pwork *pwork)
567 struct xfs_iwalk_ag *iwag;
570 iwag = container_of(pwork, struct xfs_iwalk_ag, pwork);
571 if (xfs_pwork_want_abort(pwork))
574 error = xfs_iwalk_alloc(iwag);
578 error = xfs_iwalk_ag(iwag);
579 xfs_iwalk_free(iwag);
586 * Walk all the inodes in the filesystem using multiple threads to process each
591 struct xfs_mount *mp,
594 xfs_iwalk_fn iwalk_fn,
595 unsigned int inode_records,
599 struct xfs_pwork_ctl pctl;
600 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, startino);
601 unsigned int nr_threads;
604 ASSERT(agno < mp->m_sb.sb_agcount);
605 ASSERT(!(flags & ~XFS_IWALK_FLAGS_ALL));
607 nr_threads = xfs_pwork_guess_datadev_parallelism(mp);
608 error = xfs_pwork_init(mp, &pctl, xfs_iwalk_ag_work, "xfs_iwalk",
613 for (; agno < mp->m_sb.sb_agcount; agno++) {
614 struct xfs_iwalk_ag *iwag;
616 if (xfs_pwork_ctl_want_abort(&pctl))
619 iwag = kmem_zalloc(sizeof(struct xfs_iwalk_ag), 0);
621 iwag->iwalk_fn = iwalk_fn;
623 iwag->startino = startino;
624 iwag->sz_recs = xfs_iwalk_prefetch(inode_records);
625 xfs_pwork_queue(&pctl, &iwag->pwork);
626 startino = XFS_AGINO_TO_INO(mp, agno + 1, 0);
627 if (flags & XFS_INOBT_WALK_SAME_AG)
632 xfs_pwork_poll(&pctl);
633 return xfs_pwork_destroy(&pctl);
637 * Allow callers to cache up to a page's worth of inobt records. This reflects
638 * the existing inumbers prefetching behavior. Since the inobt walk does not
639 * itself do anything with the inobt records, we can set a fairly high limit
642 #define MAX_INOBT_WALK_PREFETCH \
643 (PAGE_SIZE / sizeof(struct xfs_inobt_rec_incore))
646 * Given the number of records that the user wanted, set the number of inobt
647 * records that we buffer in memory. Set the maximum if @inobt_records == 0.
649 static inline unsigned int
650 xfs_inobt_walk_prefetch(
651 unsigned int inobt_records)
654 * If the caller didn't tell us the number of inobt records they
655 * wanted, assume the maximum prefetch possible for best performance.
657 if (inobt_records == 0)
658 inobt_records = MAX_INOBT_WALK_PREFETCH;
661 * Allocate enough space to prefetch at least two inobt records so that
662 * we can cache both the record where the iwalk started and the next
663 * record. This simplifies the AG inode walk loop setup code.
665 inobt_records = max(inobt_records, 2U);
668 * Cap prefetch at that maximum so that we don't use an absurd amount
671 return min_t(unsigned int, inobt_records, MAX_INOBT_WALK_PREFETCH);
675 * Walk all inode btree records in the filesystem starting from @startino. The
676 * @inobt_walk_fn will be called for each btree record, being passed the incore
677 * record and @data. @max_prefetch controls how many inobt records we try to
678 * cache ahead of time.
682 struct xfs_mount *mp,
683 struct xfs_trans *tp,
686 xfs_inobt_walk_fn inobt_walk_fn,
687 unsigned int inobt_records,
690 struct xfs_iwalk_ag iwag = {
693 .inobt_walk_fn = inobt_walk_fn,
695 .startino = startino,
696 .sz_recs = xfs_inobt_walk_prefetch(inobt_records),
697 .pwork = XFS_PWORK_SINGLE_THREADED,
699 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, startino);
702 ASSERT(agno < mp->m_sb.sb_agcount);
703 ASSERT(!(flags & ~XFS_INOBT_WALK_FLAGS_ALL));
705 error = xfs_iwalk_alloc(&iwag);
709 for (; agno < mp->m_sb.sb_agcount; agno++) {
710 error = xfs_iwalk_ag(&iwag);
713 iwag.startino = XFS_AGINO_TO_INO(mp, agno + 1, 0);
714 if (flags & XFS_INOBT_WALK_SAME_AG)
718 xfs_iwalk_free(&iwag);