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_mount.h"
12 #include "xfs_btree.h"
13 #include "xfs_alloc_btree.h"
14 #include "xfs_alloc.h"
15 #include "xfs_discard.h"
16 #include "xfs_error.h"
17 #include "xfs_extent_busy.h"
18 #include "xfs_trace.h"
23 * Notes on an efficient, low latency fstrim algorithm
25 * We need to walk the filesystem free space and issue discards on the free
26 * space that meet the search criteria (size and location). We cannot issue
27 * discards on extents that might be in use, or are so recently in use they are
28 * still marked as busy. To serialise against extent state changes whilst we are
29 * gathering extents to trim, we must hold the AGF lock to lock out other
30 * allocations and extent free operations that might change extent state.
32 * However, we cannot just hold the AGF for the entire AG free space walk whilst
33 * we issue discards on each free space that is found. Storage devices can have
34 * extremely slow discard implementations (e.g. ceph RBD) and so walking a
35 * couple of million free extents and issuing synchronous discards on each
36 * extent can take a *long* time. Whilst we are doing this walk, nothing else
37 * can access the AGF, and we can stall transactions and hence the log whilst
38 * modifications wait for the AGF lock to be released. This can lead hung tasks
39 * kicking the hung task timer and rebooting the system. This is bad.
41 * Hence we need to take a leaf from the bulkstat playbook. It takes the AGI
42 * lock, gathers a range of inode cluster buffers that are allocated, drops the
43 * AGI lock and then reads all the inode cluster buffers and processes them. It
44 * loops doing this, using a cursor to keep track of where it is up to in the AG
45 * for each iteration to restart the INOBT lookup from.
47 * We can't do this exactly with free space - once we drop the AGF lock, the
48 * state of the free extent is out of our control and we cannot run a discard
49 * safely on it in this situation. Unless, of course, we've marked the free
50 * extent as busy and undergoing a discard operation whilst we held the AGF
53 * This is exactly how online discard works - free extents are marked busy when
54 * they are freed, and once the extent free has been committed to the journal,
55 * the busy extent record is marked as "undergoing discard" and the discard is
56 * then issued on the free extent. Once the discard completes, the busy extent
57 * record is removed and the extent is able to be allocated again.
59 * In the context of fstrim, if we find a free extent we need to discard, we
60 * don't have to discard it immediately. All we need to do it record that free
61 * extent as being busy and under discard, and all the allocation routines will
62 * now avoid trying to allocate it. Hence if we mark the extent as busy under
63 * the AGF lock, we can safely discard it without holding the AGF lock because
64 * nothing will attempt to allocate that free space until the discard completes.
66 * This also allows us to issue discards asynchronously like we do with online
67 * discard, and so for fast devices fstrim will run much faster as we can have
68 * multiple discard operations in flight at once, as well as pipeline the free
69 * extent search so that it overlaps in flight discard IO.
72 struct workqueue_struct *xfs_discard_wq;
75 xfs_discard_endio_work(
76 struct work_struct *work)
78 struct xfs_busy_extents *extents =
79 container_of(work, struct xfs_busy_extents, endio_work);
81 xfs_extent_busy_clear(extents->mount, &extents->extent_list, false);
82 kmem_free(extents->owner);
86 * Queue up the actual completion to a thread to avoid IRQ-safe locking for
93 struct xfs_busy_extents *extents = bio->bi_private;
95 INIT_WORK(&extents->endio_work, xfs_discard_endio_work);
96 queue_work(xfs_discard_wq, &extents->endio_work);
101 * Walk the discard list and issue discards on all the busy extents in the
102 * list. We plug and chain the bios so that we only need a single completion
103 * call to clear all the busy extents once the discards are complete.
107 struct xfs_mount *mp,
108 struct xfs_busy_extents *extents)
110 struct xfs_extent_busy *busyp;
111 struct bio *bio = NULL;
112 struct blk_plug plug;
115 blk_start_plug(&plug);
116 list_for_each_entry(busyp, &extents->extent_list, list) {
117 trace_xfs_discard_extent(mp, busyp->agno, busyp->bno,
120 error = __blkdev_issue_discard(mp->m_ddev_targp->bt_bdev,
121 XFS_AGB_TO_DADDR(mp, busyp->agno, busyp->bno),
122 XFS_FSB_TO_BB(mp, busyp->length),
124 if (error && error != -EOPNOTSUPP) {
126 "discard failed for extent [0x%llx,%u], error %d",
127 (unsigned long long)busyp->bno,
135 bio->bi_private = extents;
136 bio->bi_end_io = xfs_discard_endio;
139 xfs_discard_endio_work(&extents->endio_work);
141 blk_finish_plug(&plug);
148 xfs_trim_gather_extents(
149 struct xfs_perag *pag,
153 struct xfs_alloc_rec_incore *tcur,
154 struct xfs_busy_extents *extents,
155 uint64_t *blocks_trimmed)
157 struct xfs_mount *mp = pag->pag_mount;
158 struct xfs_btree_cur *cur;
159 struct xfs_buf *agbp;
165 * Force out the log. This means any transactions that might have freed
166 * space before we take the AGF buffer lock are now on disk, and the
167 * volatile disk cache is flushed.
169 xfs_log_force(mp, XFS_LOG_SYNC);
171 error = xfs_alloc_read_agf(pag, NULL, 0, &agbp);
175 cur = xfs_allocbt_init_cursor(mp, NULL, agbp, pag, XFS_BTNUM_CNT);
178 * Look up the extent length requested in the AGF and start with it.
180 if (tcur->ar_startblock == NULLAGBLOCK)
181 error = xfs_alloc_lookup_ge(cur, 0, tcur->ar_blockcount, &i);
183 error = xfs_alloc_lookup_le(cur, tcur->ar_startblock,
184 tcur->ar_blockcount, &i);
188 /* nothing of that length left in the AG, we are done */
189 tcur->ar_blockcount = 0;
194 * Loop until we are done with all extents that are large
195 * enough to be worth discarding or we hit batch limits.
203 error = xfs_alloc_get_rec(cur, &fbno, &flen, &i);
206 if (XFS_IS_CORRUPT(mp, i != 1)) {
207 error = -EFSCORRUPTED;
213 * Update the cursor to point at this extent so we
214 * restart the next batch from this extent.
216 tcur->ar_startblock = fbno;
217 tcur->ar_blockcount = flen;
222 * use daddr format for all range/len calculations as that is
223 * the format the range/len variables are supplied in by
226 dbno = XFS_AGB_TO_DADDR(mp, pag->pag_agno, fbno);
227 dlen = XFS_FSB_TO_BB(mp, flen);
230 * Too small? Give up.
233 trace_xfs_discard_toosmall(mp, pag->pag_agno, fbno, flen);
234 tcur->ar_blockcount = 0;
239 * If the extent is entirely outside of the range we are
240 * supposed to discard skip it. Do not bother to trim
241 * down partially overlapping ranges for now.
243 if (dbno + dlen < start || dbno > end) {
244 trace_xfs_discard_exclude(mp, pag->pag_agno, fbno, flen);
249 * If any blocks in the range are still busy, skip the
250 * discard and try again the next time.
252 if (xfs_extent_busy_search(mp, pag, fbno, flen)) {
253 trace_xfs_discard_busy(mp, pag->pag_agno, fbno, flen);
257 xfs_extent_busy_insert_discard(pag, fbno, flen,
258 &extents->extent_list);
259 *blocks_trimmed += flen;
261 error = xfs_btree_decrement(cur, 0, &i);
266 * If there's no more records in the tree, we are done. Set the
267 * cursor block count to 0 to indicate to the caller that there
268 * is no more extents to search.
271 tcur->ar_blockcount = 0;
275 * If there was an error, release all the gathered busy extents because
276 * we aren't going to issue a discard on them any more.
279 xfs_extent_busy_clear(mp, &extents->extent_list, false);
281 xfs_btree_del_cursor(cur, error);
287 xfs_trim_should_stop(void)
289 return fatal_signal_pending(current) || freezing(current);
293 * Iterate the free list gathering extents and discarding them. We need a cursor
294 * for the repeated iteration of gather/discard loop, so use the longest extent
295 * we found in the last batch as the key to start the next.
299 struct xfs_perag *pag,
303 uint64_t *blocks_trimmed)
305 struct xfs_alloc_rec_incore tcur = {
306 .ar_blockcount = pag->pagf_longest,
307 .ar_startblock = NULLAGBLOCK,
312 struct xfs_busy_extents *extents;
314 extents = kzalloc(sizeof(*extents), GFP_KERNEL);
320 extents->mount = pag->pag_mount;
321 extents->owner = extents;
322 INIT_LIST_HEAD(&extents->extent_list);
324 error = xfs_trim_gather_extents(pag, start, end, minlen,
325 &tcur, extents, blocks_trimmed);
332 * We hand the extent list to the discard function here so the
333 * discarded extents can be removed from the busy extent list.
334 * This allows the discards to run asynchronously with gathering
335 * the next round of extents to discard.
337 * However, we must ensure that we do not reference the extent
338 * list after this function call, as it may have been freed by
339 * the time control returns to us.
341 error = xfs_discard_extents(pag->pag_mount, extents);
345 if (xfs_trim_should_stop())
348 } while (tcur.ar_blockcount != 0);
355 * trim a range of the filesystem.
357 * Note: the parameters passed from userspace are byte ranges into the
358 * filesystem which does not match to the format we use for filesystem block
359 * addressing. FSB addressing is sparse (AGNO|AGBNO), while the incoming format
360 * is a linear address range. Hence we need to use DADDR based conversions and
361 * comparisons for determining the correct offset and regions to trim.
365 struct xfs_mount *mp,
366 struct fstrim_range __user *urange)
368 struct xfs_perag *pag;
369 unsigned int granularity =
370 bdev_discard_granularity(mp->m_ddev_targp->bt_bdev);
371 struct fstrim_range range;
372 xfs_daddr_t start, end, minlen;
374 uint64_t blocks_trimmed = 0;
375 int error, last_error = 0;
377 if (!capable(CAP_SYS_ADMIN))
379 if (!bdev_max_discard_sectors(mp->m_ddev_targp->bt_bdev))
383 * We haven't recovered the log, so we cannot use our bnobt-guided
384 * storage zapping commands.
386 if (xfs_has_norecovery(mp))
389 if (copy_from_user(&range, urange, sizeof(range)))
392 range.minlen = max_t(u64, granularity, range.minlen);
393 minlen = BTOBB(range.minlen);
395 * Truncating down the len isn't actually quite correct, but using
396 * BBTOB would mean we trivially get overflows for values
397 * of ULLONG_MAX or slightly lower. And ULLONG_MAX is the default
398 * used by the fstrim application. In the end it really doesn't
399 * matter as trimming blocks is an advisory interface.
401 if (range.start >= XFS_FSB_TO_B(mp, mp->m_sb.sb_dblocks) ||
402 range.minlen > XFS_FSB_TO_B(mp, mp->m_ag_max_usable) ||
403 range.len < mp->m_sb.sb_blocksize)
406 start = BTOBB(range.start);
407 end = start + BTOBBT(range.len) - 1;
409 if (end > XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks) - 1)
410 end = XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks) - 1;
412 agno = xfs_daddr_to_agno(mp, start);
413 for_each_perag_range(mp, agno, xfs_daddr_to_agno(mp, end), pag) {
414 error = xfs_trim_extents(pag, start, end, minlen,
419 if (xfs_trim_should_stop()) {
428 range.len = XFS_FSB_TO_B(mp, blocks_trimmed);
429 if (copy_to_user(urange, &range, sizeof(range)))