1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2000-2002,2005 Silicon Graphics, Inc.
4 * Copyright (c) 2010 David Chinner.
5 * Copyright (c) 2011 Christoph Hellwig.
10 #include "xfs_format.h"
11 #include "xfs_log_format.h"
12 #include "xfs_shared.h"
13 #include "xfs_trans_resv.h"
14 #include "xfs_mount.h"
15 #include "xfs_alloc.h"
16 #include "xfs_extent_busy.h"
17 #include "xfs_trace.h"
18 #include "xfs_trans.h"
21 #include "xfs_rtgroup.h"
23 struct xfs_extent_busy_tree {
25 struct rb_root eb_tree;
27 wait_queue_head_t eb_wait;
31 xfs_extent_busy_insert_list(
36 struct list_head *busy_list)
38 struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
39 struct xfs_extent_busy *new;
40 struct xfs_extent_busy *busyp;
42 struct rb_node *parent = NULL;
44 new = kzalloc(sizeof(struct xfs_extent_busy),
45 GFP_KERNEL | __GFP_NOFAIL);
46 new->group = xfs_group_hold(xg);
49 INIT_LIST_HEAD(&new->list);
52 /* trace before insert to be able to see failed inserts */
53 trace_xfs_extent_busy(xg, bno, len);
55 spin_lock(&eb->eb_lock);
56 rbp = &eb->eb_tree.rb_node;
59 busyp = rb_entry(parent, struct xfs_extent_busy, rb_node);
61 if (new->bno < busyp->bno) {
62 rbp = &(*rbp)->rb_left;
63 ASSERT(new->bno + new->length <= busyp->bno);
64 } else if (new->bno > busyp->bno) {
65 rbp = &(*rbp)->rb_right;
66 ASSERT(bno >= busyp->bno + busyp->length);
72 rb_link_node(&new->rb_node, parent, rbp);
73 rb_insert_color(&new->rb_node, &eb->eb_tree);
75 /* always process discard lists in fifo order */
76 list_add_tail(&new->list, busy_list);
77 spin_unlock(&eb->eb_lock);
81 xfs_extent_busy_insert(
88 xfs_extent_busy_insert_list(xg, bno, len, flags, &tp->t_busy);
92 xfs_extent_busy_insert_discard(
96 struct list_head *busy_list)
98 xfs_extent_busy_insert_list(xg, bno, len, XFS_EXTENT_BUSY_DISCARDED,
103 * Search for a busy extent within the range of the extent we are about to
104 * allocate. You need to be holding the busy extent tree lock when calling
105 * xfs_extent_busy_search(). This function returns 0 for no overlapping busy
106 * extent, -1 for an overlapping but not exact busy extent, and 1 for an exact
107 * match. This is done so that a non-zero return indicates an overlap that
108 * will require a synchronous transaction, but it can still be
109 * used to distinguish between a partial or exact match.
112 xfs_extent_busy_search(
113 struct xfs_group *xg,
117 struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
119 struct xfs_extent_busy *busyp;
122 /* find closest start bno overlap */
123 spin_lock(&eb->eb_lock);
124 rbp = eb->eb_tree.rb_node;
126 busyp = rb_entry(rbp, struct xfs_extent_busy, rb_node);
127 if (bno < busyp->bno) {
128 /* may overlap, but exact start block is lower */
129 if (bno + len > busyp->bno)
132 } else if (bno > busyp->bno) {
133 /* may overlap, but exact start block is higher */
134 if (bno < busyp->bno + busyp->length)
138 /* bno matches busyp, length determines exact match */
139 match = (busyp->length == len) ? 1 : -1;
143 spin_unlock(&eb->eb_lock);
148 * The found free extent [fbno, fend] overlaps part or all of the given busy
149 * extent. If the overlap covers the beginning, the end, or all of the busy
150 * extent, the overlapping portion can be made unbusy and used for the
151 * allocation. We can't split a busy extent because we can't modify a
152 * transaction/CIL context busy list, but we can update an entry's block
155 * Returns true if the extent can safely be reused, or false if the search
156 * needs to be restarted.
159 xfs_extent_busy_update_extent(
160 struct xfs_group *xg,
161 struct xfs_extent_busy *busyp,
165 __releases(&eb->eb_lock)
166 __acquires(&eb->eb_lock)
168 struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
169 xfs_agblock_t fend = fbno + flen;
170 xfs_agblock_t bbno = busyp->bno;
171 xfs_agblock_t bend = bbno + busyp->length;
174 * This extent is currently being discarded. Give the thread
175 * performing the discard a chance to mark the extent unbusy
178 if (busyp->flags & XFS_EXTENT_BUSY_DISCARDED) {
179 spin_unlock(&eb->eb_lock);
181 spin_lock(&eb->eb_lock);
186 * If there is a busy extent overlapping a user allocation, we have
187 * no choice but to force the log and retry the search.
189 * Fortunately this does not happen during normal operation, but
190 * only if the filesystem is very low on space and has to dip into
191 * the AGFL for normal allocations.
196 if (bbno < fbno && bend > fend) {
200 * +BBBBBBBBBBBBBBBBB+
206 * We would have to split the busy extent to be able to track
207 * it correct, which we cannot do because we would have to
208 * modify the list of busy extents attached to the transaction
209 * or CIL context, which is immutable.
211 * Force out the log to clear the busy extent and retry the
215 } else if (bbno >= fbno && bend <= fend) {
219 * +BBBBBBBBBBBBBBBBB+
220 * +-----------------+
225 * +BBBBBBBBBBBBBBBBB+
226 * +--------------------------+
231 * +BBBBBBBBBBBBBBBBB+
232 * +--------------------------+
237 * +BBBBBBBBBBBBBBBBB+
238 * +-----------------------------------+
244 * The busy extent is fully covered by the extent we are
245 * allocating, and can simply be removed from the rbtree.
246 * However we cannot remove it from the immutable list
247 * tracking busy extents in the transaction or CIL context,
248 * so set the length to zero to mark it invalid.
250 * We also need to restart the busy extent search from the
251 * tree root, because erasing the node can rearrange the
254 rb_erase(&busyp->rb_node, &eb->eb_tree);
257 } else if (fend < bend) {
261 * +BBBBBBBBBBBBBBBBB+
267 * +BBBBBBBBBBBBBBBBB+
268 * +------------------+
273 busyp->length = bend - fend;
274 } else if (bbno < fbno) {
278 * +BBBBBBBBBBBBBBBBB+
284 * +BBBBBBBBBBBBBBBBB+
285 * +----------------------+
288 busyp->length = fbno - busyp->bno;
293 trace_xfs_extent_busy_reuse(xg, fbno, flen);
297 spin_unlock(&eb->eb_lock);
298 xfs_log_force(xg->xg_mount, XFS_LOG_SYNC);
299 trace_xfs_extent_busy_force(xg, fbno, flen);
300 spin_lock(&eb->eb_lock);
305 * For a given extent [fbno, flen], make sure we can reuse it safely.
308 xfs_extent_busy_reuse(
309 struct xfs_group *xg,
314 struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
318 spin_lock(&eb->eb_lock);
320 rbp = eb->eb_tree.rb_node;
322 struct xfs_extent_busy *busyp =
323 rb_entry(rbp, struct xfs_extent_busy, rb_node);
324 xfs_agblock_t bbno = busyp->bno;
325 xfs_agblock_t bend = bbno + busyp->length;
327 if (fbno + flen <= bbno) {
330 } else if (fbno >= bend) {
335 if (!xfs_extent_busy_update_extent(xg, busyp, fbno, flen,
339 spin_unlock(&eb->eb_lock);
343 * For a given extent [fbno, flen], search the busy extent list to find a
344 * subset of the extent that is not busy. If *rlen is smaller than
345 * args->minlen no suitable extent could be found, and the higher level
346 * code needs to force out the log and retry the allocation.
348 * Return the current busy generation for the group if the extent is busy. This
349 * value can be used to wait for at least one of the currently busy extents
350 * to be cleared. Note that the busy list is not guaranteed to be empty after
351 * the gen is woken. The state of a specific extent must always be confirmed
352 * with another call to xfs_extent_busy_trim() before it can be used.
355 xfs_extent_busy_trim(
356 struct xfs_group *xg,
363 struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
371 spin_lock(&eb->eb_lock);
374 rbp = eb->eb_tree.rb_node;
375 while (rbp && flen >= minlen) {
376 struct xfs_extent_busy *busyp =
377 rb_entry(rbp, struct xfs_extent_busy, rb_node);
378 xfs_agblock_t fend = fbno + flen;
379 xfs_agblock_t bbno = busyp->bno;
380 xfs_agblock_t bend = bbno + busyp->length;
385 } else if (fbno >= bend) {
396 * +BBBBBBBBBBBBBBBBB+
402 * +BBBBBBBBBBBBBBBBB+
408 * +BBBBBBBBBBBBBBBBB+
414 * +BBBBBBBBBBBBBBBBB+
415 * +-----------------+
418 * No unbusy region in extent, return failure.
426 * +BBBBBBBBBBBBBBBBB+
427 * +----------------------+
432 * +BBBBBBBBBBBBBBBBB+
433 * +--------------------------+
436 * Needs to be trimmed to:
441 } else if (bend >= fend) {
447 * +BBBBBBBBBBBBBBBBB+
448 * +------------------+
453 * +BBBBBBBBBBBBBBBBB+
454 * +--------------------------+
457 * Needs to be trimmed to:
468 * +BBBBBBBBBBBBBBBBB+
469 * +-----------------------------------+
473 * +-------+ OR +-------+
474 * fbno fend fbno fend
476 * Backward allocation leads to significant
477 * fragmentation of directories, which degrades
478 * directory performance, therefore we always want to
479 * choose the option that produces forward allocation
481 * Preferring the lower bno extent will make the next
482 * request use "fend" as the start of the next
483 * allocation; if the segment is no longer busy at
484 * that point, we'll get a contiguous allocation, but
485 * even if it is still busy, we will get a forward
487 * We try to avoid choosing the segment at "bend",
488 * because that can lead to the next allocation
489 * taking the segment at "fbno", which would be a
490 * backward allocation. We only use the segment at
491 * "fbno" if it is much larger than the current
492 * requested size, because in that case there's a
493 * good chance subsequent allocations will be
496 if (bbno - fbno >= maxlen) {
497 /* left candidate fits perfect */
499 } else if (fend - bend >= maxlen * 4) {
500 /* right candidate has enough free space */
502 } else if (bbno - fbno >= minlen) {
503 /* left candidate fits minimum requirement */
514 if (fbno != *bno || flen != *len) {
515 trace_xfs_extent_busy_trim(xg, *bno, *len, fbno, flen);
518 *busy_gen = eb->eb_gen;
521 spin_unlock(&eb->eb_lock);
525 * Return a zero extent length as failure indications. All callers
526 * re-check if the trimmed extent satisfies the minlen requirement.
533 xfs_extent_busy_clear_one(
534 struct xfs_extent_busy *busyp,
537 struct xfs_extent_busy_tree *eb = busyp->group->xg_busy_extents;
541 !(busyp->flags & XFS_EXTENT_BUSY_SKIP_DISCARD)) {
542 busyp->flags = XFS_EXTENT_BUSY_DISCARDED;
545 trace_xfs_extent_busy_clear(busyp->group, busyp->bno,
547 rb_erase(&busyp->rb_node, &eb->eb_tree);
550 list_del_init(&busyp->list);
551 xfs_group_put(busyp->group);
557 * Remove all extents on the passed in list from the busy extents tree.
558 * If do_discard is set skip extents that need to be discarded, and mark
559 * these as undergoing a discard operation instead.
562 xfs_extent_busy_clear(
563 struct list_head *list,
566 struct xfs_extent_busy *busyp, *next;
568 busyp = list_first_entry_or_null(list, typeof(*busyp), list);
573 struct xfs_group *xg = xfs_group_hold(busyp->group);
574 struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
577 spin_lock(&eb->eb_lock);
579 next = list_next_entry(busyp, list);
580 if (xfs_extent_busy_clear_one(busyp, do_discard))
583 } while (!list_entry_is_head(busyp, list, list) &&
588 wake_up_all(&eb->eb_wait);
590 spin_unlock(&eb->eb_lock);
592 } while (!list_entry_is_head(busyp, list, list));
596 * Flush out all busy extents for this group.
598 * If the current transaction is holding busy extents, the caller may not want
599 * to wait for committed busy extents to resolve. If we are being told just to
600 * try a flush or progress has been made since we last skipped a busy extent,
601 * return immediately to allow the caller to try again.
603 * If we are freeing extents, we might actually be holding the only free extents
604 * in the transaction busy list and the log force won't resolve that situation.
605 * In this case, we must return -EAGAIN to avoid a deadlock by informing the
606 * caller it needs to commit the busy extents it holds before retrying the
607 * extent free operation.
610 xfs_extent_busy_flush(
611 struct xfs_trans *tp,
612 struct xfs_group *xg,
614 uint32_t alloc_flags)
616 struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
620 error = xfs_log_force(tp->t_mountp, XFS_LOG_SYNC);
624 /* Avoid deadlocks on uncommitted busy extents. */
625 if (!list_empty(&tp->t_busy)) {
626 if (alloc_flags & XFS_ALLOC_FLAG_TRYFLUSH)
629 if (busy_gen != READ_ONCE(eb->eb_gen))
632 if (alloc_flags & XFS_ALLOC_FLAG_FREEING)
636 /* Wait for committed busy extents to resolve. */
638 prepare_to_wait(&eb->eb_wait, &wait, TASK_KILLABLE);
639 if (busy_gen != READ_ONCE(eb->eb_gen))
644 finish_wait(&eb->eb_wait, &wait);
649 xfs_extent_busy_wait_group(
650 struct xfs_group *xg)
653 struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
656 prepare_to_wait(&eb->eb_wait, &wait, TASK_KILLABLE);
657 if (RB_EMPTY_ROOT(&eb->eb_tree))
661 finish_wait(&eb->eb_wait, &wait);
665 xfs_extent_busy_wait_all(
666 struct xfs_mount *mp)
668 struct xfs_perag *pag = NULL;
669 struct xfs_rtgroup *rtg = NULL;
671 while ((pag = xfs_perag_next(mp, pag)))
672 xfs_extent_busy_wait_group(pag_group(pag));
674 if (xfs_has_rtgroups(mp))
675 while ((rtg = xfs_rtgroup_next(mp, rtg)))
676 xfs_extent_busy_wait_group(rtg_group(rtg));
680 * Callback for list_sort to sort busy extents by the group they reside in.
683 xfs_extent_busy_ag_cmp(
685 const struct list_head *l1,
686 const struct list_head *l2)
688 struct xfs_extent_busy *b1 =
689 container_of(l1, struct xfs_extent_busy, list);
690 struct xfs_extent_busy *b2 =
691 container_of(l2, struct xfs_extent_busy, list);
694 diff = b1->group->xg_gno - b2->group->xg_gno;
696 diff = b1->bno - b2->bno;
700 /* Are there any busy extents in this group? */
702 xfs_extent_busy_list_empty(
703 struct xfs_group *xg,
706 struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
709 spin_lock(&eb->eb_lock);
710 res = RB_EMPTY_ROOT(&eb->eb_tree);
711 *busy_gen = READ_ONCE(eb->eb_gen);
712 spin_unlock(&eb->eb_lock);
716 struct xfs_extent_busy_tree *
717 xfs_extent_busy_alloc(void)
719 struct xfs_extent_busy_tree *eb;
721 eb = kzalloc(sizeof(*eb), GFP_KERNEL);
724 spin_lock_init(&eb->eb_lock);
725 init_waitqueue_head(&eb->eb_wait);
726 eb->eb_tree = RB_ROOT;