1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2010, 2023 Red Hat, Inc.
7 #include "xfs_shared.h"
8 #include "xfs_format.h"
9 #include "xfs_log_format.h"
10 #include "xfs_trans_resv.h"
11 #include "xfs_trans.h"
12 #include "xfs_mount.h"
13 #include "xfs_btree.h"
14 #include "xfs_alloc_btree.h"
15 #include "xfs_alloc.h"
16 #include "xfs_discard.h"
17 #include "xfs_error.h"
18 #include "xfs_extent_busy.h"
19 #include "xfs_trace.h"
22 #include "xfs_health.h"
25 * Notes on an efficient, low latency fstrim algorithm
27 * We need to walk the filesystem free space and issue discards on the free
28 * space that meet the search criteria (size and location). We cannot issue
29 * discards on extents that might be in use, or are so recently in use they are
30 * still marked as busy. To serialise against extent state changes whilst we are
31 * gathering extents to trim, we must hold the AGF lock to lock out other
32 * allocations and extent free operations that might change extent state.
34 * However, we cannot just hold the AGF for the entire AG free space walk whilst
35 * we issue discards on each free space that is found. Storage devices can have
36 * extremely slow discard implementations (e.g. ceph RBD) and so walking a
37 * couple of million free extents and issuing synchronous discards on each
38 * extent can take a *long* time. Whilst we are doing this walk, nothing else
39 * can access the AGF, and we can stall transactions and hence the log whilst
40 * modifications wait for the AGF lock to be released. This can lead hung tasks
41 * kicking the hung task timer and rebooting the system. This is bad.
43 * Hence we need to take a leaf from the bulkstat playbook. It takes the AGI
44 * lock, gathers a range of inode cluster buffers that are allocated, drops the
45 * AGI lock and then reads all the inode cluster buffers and processes them. It
46 * loops doing this, using a cursor to keep track of where it is up to in the AG
47 * for each iteration to restart the INOBT lookup from.
49 * We can't do this exactly with free space - once we drop the AGF lock, the
50 * state of the free extent is out of our control and we cannot run a discard
51 * safely on it in this situation. Unless, of course, we've marked the free
52 * extent as busy and undergoing a discard operation whilst we held the AGF
55 * This is exactly how online discard works - free extents are marked busy when
56 * they are freed, and once the extent free has been committed to the journal,
57 * the busy extent record is marked as "undergoing discard" and the discard is
58 * then issued on the free extent. Once the discard completes, the busy extent
59 * record is removed and the extent is able to be allocated again.
61 * In the context of fstrim, if we find a free extent we need to discard, we
62 * don't have to discard it immediately. All we need to do it record that free
63 * extent as being busy and under discard, and all the allocation routines will
64 * now avoid trying to allocate it. Hence if we mark the extent as busy under
65 * the AGF lock, we can safely discard it without holding the AGF lock because
66 * nothing will attempt to allocate that free space until the discard completes.
68 * This also allows us to issue discards asynchronously like we do with online
69 * discard, and so for fast devices fstrim will run much faster as we can have
70 * multiple discard operations in flight at once, as well as pipeline the free
71 * extent search so that it overlaps in flight discard IO.
74 struct workqueue_struct *xfs_discard_wq;
77 xfs_discard_endio_work(
78 struct work_struct *work)
80 struct xfs_busy_extents *extents =
81 container_of(work, struct xfs_busy_extents, endio_work);
83 xfs_extent_busy_clear(extents->mount, &extents->extent_list, false);
84 kfree(extents->owner);
88 * Queue up the actual completion to a thread to avoid IRQ-safe locking for
95 struct xfs_busy_extents *extents = bio->bi_private;
97 INIT_WORK(&extents->endio_work, xfs_discard_endio_work);
98 queue_work(xfs_discard_wq, &extents->endio_work);
103 * Walk the discard list and issue discards on all the busy extents in the
104 * list. We plug and chain the bios so that we only need a single completion
105 * call to clear all the busy extents once the discards are complete.
109 struct xfs_mount *mp,
110 struct xfs_busy_extents *extents)
112 struct xfs_extent_busy *busyp;
113 struct bio *bio = NULL;
114 struct blk_plug plug;
117 blk_start_plug(&plug);
118 list_for_each_entry(busyp, &extents->extent_list, list) {
119 trace_xfs_discard_extent(mp, busyp->agno, busyp->bno,
122 error = __blkdev_issue_discard(mp->m_ddev_targp->bt_bdev,
123 XFS_AGB_TO_DADDR(mp, busyp->agno, busyp->bno),
124 XFS_FSB_TO_BB(mp, busyp->length),
126 if (error && error != -EOPNOTSUPP) {
128 "discard failed for extent [0x%llx,%u], error %d",
129 (unsigned long long)busyp->bno,
137 bio->bi_private = extents;
138 bio->bi_end_io = xfs_discard_endio;
141 xfs_discard_endio_work(&extents->endio_work);
143 blk_finish_plug(&plug);
150 xfs_trim_gather_extents(
151 struct xfs_perag *pag,
155 struct xfs_alloc_rec_incore *tcur,
156 struct xfs_busy_extents *extents,
157 uint64_t *blocks_trimmed)
159 struct xfs_mount *mp = pag->pag_mount;
160 struct xfs_trans *tp;
161 struct xfs_btree_cur *cur;
162 struct xfs_buf *agbp;
168 * Force out the log. This means any transactions that might have freed
169 * space before we take the AGF buffer lock are now on disk, and the
170 * volatile disk cache is flushed.
172 xfs_log_force(mp, XFS_LOG_SYNC);
174 error = xfs_trans_alloc_empty(mp, &tp);
178 error = xfs_alloc_read_agf(pag, tp, 0, &agbp);
180 goto out_trans_cancel;
182 cur = xfs_cntbt_init_cursor(mp, tp, agbp, pag);
185 * Look up the extent length requested in the AGF and start with it.
187 if (tcur->ar_startblock == NULLAGBLOCK)
188 error = xfs_alloc_lookup_ge(cur, 0, tcur->ar_blockcount, &i);
190 error = xfs_alloc_lookup_le(cur, tcur->ar_startblock,
191 tcur->ar_blockcount, &i);
195 /* nothing of that length left in the AG, we are done */
196 tcur->ar_blockcount = 0;
201 * Loop until we are done with all extents that are large
202 * enough to be worth discarding or we hit batch limits.
210 error = xfs_alloc_get_rec(cur, &fbno, &flen, &i);
213 if (XFS_IS_CORRUPT(mp, i != 1)) {
214 xfs_btree_mark_sick(cur);
215 error = -EFSCORRUPTED;
221 * Update the cursor to point at this extent so we
222 * restart the next batch from this extent.
224 tcur->ar_startblock = fbno;
225 tcur->ar_blockcount = flen;
230 * use daddr format for all range/len calculations as that is
231 * the format the range/len variables are supplied in by
234 dbno = XFS_AGB_TO_DADDR(mp, pag->pag_agno, fbno);
235 dlen = XFS_FSB_TO_BB(mp, flen);
238 * Too small? Give up.
241 trace_xfs_discard_toosmall(mp, pag->pag_agno, fbno, flen);
242 tcur->ar_blockcount = 0;
247 * If the extent is entirely outside of the range we are
248 * supposed to discard skip it. Do not bother to trim
249 * down partially overlapping ranges for now.
251 if (dbno + dlen < start || dbno > end) {
252 trace_xfs_discard_exclude(mp, pag->pag_agno, fbno, flen);
257 * If any blocks in the range are still busy, skip the
258 * discard and try again the next time.
260 if (xfs_extent_busy_search(mp, pag, fbno, flen)) {
261 trace_xfs_discard_busy(mp, pag->pag_agno, fbno, flen);
265 xfs_extent_busy_insert_discard(pag, fbno, flen,
266 &extents->extent_list);
267 *blocks_trimmed += flen;
269 error = xfs_btree_decrement(cur, 0, &i);
274 * If there's no more records in the tree, we are done. Set the
275 * cursor block count to 0 to indicate to the caller that there
276 * is no more extents to search.
279 tcur->ar_blockcount = 0;
283 * If there was an error, release all the gathered busy extents because
284 * we aren't going to issue a discard on them any more.
287 xfs_extent_busy_clear(mp, &extents->extent_list, false);
289 xfs_btree_del_cursor(cur, error);
291 xfs_trans_cancel(tp);
296 xfs_trim_should_stop(void)
298 return fatal_signal_pending(current) || freezing(current);
302 * Iterate the free list gathering extents and discarding them. We need a cursor
303 * for the repeated iteration of gather/discard loop, so use the longest extent
304 * we found in the last batch as the key to start the next.
308 struct xfs_perag *pag,
312 uint64_t *blocks_trimmed)
314 struct xfs_alloc_rec_incore tcur = {
315 .ar_blockcount = pag->pagf_longest,
316 .ar_startblock = NULLAGBLOCK,
321 struct xfs_busy_extents *extents;
323 extents = kzalloc(sizeof(*extents), GFP_KERNEL);
329 extents->mount = pag->pag_mount;
330 extents->owner = extents;
331 INIT_LIST_HEAD(&extents->extent_list);
333 error = xfs_trim_gather_extents(pag, start, end, minlen,
334 &tcur, extents, blocks_trimmed);
341 * We hand the extent list to the discard function here so the
342 * discarded extents can be removed from the busy extent list.
343 * This allows the discards to run asynchronously with gathering
344 * the next round of extents to discard.
346 * However, we must ensure that we do not reference the extent
347 * list after this function call, as it may have been freed by
348 * the time control returns to us.
350 error = xfs_discard_extents(pag->pag_mount, extents);
354 if (xfs_trim_should_stop())
357 } while (tcur.ar_blockcount != 0);
364 * trim a range of the filesystem.
366 * Note: the parameters passed from userspace are byte ranges into the
367 * filesystem which does not match to the format we use for filesystem block
368 * addressing. FSB addressing is sparse (AGNO|AGBNO), while the incoming format
369 * is a linear address range. Hence we need to use DADDR based conversions and
370 * comparisons for determining the correct offset and regions to trim.
374 struct xfs_mount *mp,
375 struct fstrim_range __user *urange)
377 struct xfs_perag *pag;
378 unsigned int granularity =
379 bdev_discard_granularity(mp->m_ddev_targp->bt_bdev);
380 struct fstrim_range range;
381 xfs_daddr_t start, end, minlen;
383 uint64_t blocks_trimmed = 0;
384 int error, last_error = 0;
386 if (!capable(CAP_SYS_ADMIN))
388 if (!bdev_max_discard_sectors(mp->m_ddev_targp->bt_bdev))
392 * We haven't recovered the log, so we cannot use our bnobt-guided
393 * storage zapping commands.
395 if (xfs_has_norecovery(mp))
398 if (copy_from_user(&range, urange, sizeof(range)))
401 range.minlen = max_t(u64, granularity, range.minlen);
402 minlen = BTOBB(range.minlen);
404 * Truncating down the len isn't actually quite correct, but using
405 * BBTOB would mean we trivially get overflows for values
406 * of ULLONG_MAX or slightly lower. And ULLONG_MAX is the default
407 * used by the fstrim application. In the end it really doesn't
408 * matter as trimming blocks is an advisory interface.
410 if (range.start >= XFS_FSB_TO_B(mp, mp->m_sb.sb_dblocks) ||
411 range.minlen > XFS_FSB_TO_B(mp, mp->m_ag_max_usable) ||
412 range.len < mp->m_sb.sb_blocksize)
415 start = BTOBB(range.start);
416 end = start + BTOBBT(range.len) - 1;
418 if (end > XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks) - 1)
419 end = XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks) - 1;
421 agno = xfs_daddr_to_agno(mp, start);
422 for_each_perag_range(mp, agno, xfs_daddr_to_agno(mp, end), pag) {
423 error = xfs_trim_extents(pag, start, end, minlen,
428 if (xfs_trim_should_stop()) {
437 range.len = XFS_FSB_TO_B(mp, blocks_trimmed);
438 if (copy_to_user(urange, &range, sizeof(range)))