xref: /linux/fs/xfs/xfs_discard.c (revision e6a901a00822659181c93c86d8bbc2a17779fddc)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2010, 2023 Red Hat, Inc.
4  * All Rights Reserved.
5  */
6 #include "xfs.h"
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"
20 #include "xfs_log.h"
21 #include "xfs_ag.h"
22 #include "xfs_health.h"
23 
24 /*
25  * Notes on an efficient, low latency fstrim algorithm
26  *
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.
33  *
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.
42  *
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.
48  *
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
53  * locked.
54  *
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.
60  *
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.
67  *
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.
72  */
73 
74 struct workqueue_struct *xfs_discard_wq;
75 
76 static void
77 xfs_discard_endio_work(
78 	struct work_struct	*work)
79 {
80 	struct xfs_busy_extents	*extents =
81 		container_of(work, struct xfs_busy_extents, endio_work);
82 
83 	xfs_extent_busy_clear(extents->mount, &extents->extent_list, false);
84 	kfree(extents->owner);
85 }
86 
87 /*
88  * Queue up the actual completion to a thread to avoid IRQ-safe locking for
89  * pagb_lock.
90  */
91 static void
92 xfs_discard_endio(
93 	struct bio		*bio)
94 {
95 	struct xfs_busy_extents	*extents = bio->bi_private;
96 
97 	INIT_WORK(&extents->endio_work, xfs_discard_endio_work);
98 	queue_work(xfs_discard_wq, &extents->endio_work);
99 	bio_put(bio);
100 }
101 
102 /*
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.
106  */
107 int
108 xfs_discard_extents(
109 	struct xfs_mount	*mp,
110 	struct xfs_busy_extents	*extents)
111 {
112 	struct xfs_extent_busy	*busyp;
113 	struct bio		*bio = NULL;
114 	struct blk_plug		plug;
115 	int			error = 0;
116 
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,
120 					 busyp->length);
121 
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),
125 				GFP_KERNEL, &bio);
126 		if (error && error != -EOPNOTSUPP) {
127 			xfs_info(mp,
128 	 "discard failed for extent [0x%llx,%u], error %d",
129 				 (unsigned long long)busyp->bno,
130 				 busyp->length,
131 				 error);
132 			break;
133 		}
134 	}
135 
136 	if (bio) {
137 		bio->bi_private = extents;
138 		bio->bi_end_io = xfs_discard_endio;
139 		submit_bio(bio);
140 	} else {
141 		xfs_discard_endio_work(&extents->endio_work);
142 	}
143 	blk_finish_plug(&plug);
144 
145 	return error;
146 }
147 
148 
149 static int
150 xfs_trim_gather_extents(
151 	struct xfs_perag	*pag,
152 	xfs_daddr_t		start,
153 	xfs_daddr_t		end,
154 	xfs_daddr_t		minlen,
155 	struct xfs_alloc_rec_incore *tcur,
156 	struct xfs_busy_extents	*extents,
157 	uint64_t		*blocks_trimmed)
158 {
159 	struct xfs_mount	*mp = pag->pag_mount;
160 	struct xfs_trans	*tp;
161 	struct xfs_btree_cur	*cur;
162 	struct xfs_buf		*agbp;
163 	int			error;
164 	int			i;
165 	int			batch = 100;
166 
167 	/*
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.
171 	 */
172 	xfs_log_force(mp, XFS_LOG_SYNC);
173 
174 	error = xfs_trans_alloc_empty(mp, &tp);
175 	if (error)
176 		return error;
177 
178 	error = xfs_alloc_read_agf(pag, tp, 0, &agbp);
179 	if (error)
180 		goto out_trans_cancel;
181 
182 	cur = xfs_cntbt_init_cursor(mp, tp, agbp, pag);
183 
184 	/*
185 	 * Look up the extent length requested in the AGF and start with it.
186 	 */
187 	if (tcur->ar_startblock == NULLAGBLOCK)
188 		error = xfs_alloc_lookup_ge(cur, 0, tcur->ar_blockcount, &i);
189 	else
190 		error = xfs_alloc_lookup_le(cur, tcur->ar_startblock,
191 				tcur->ar_blockcount, &i);
192 	if (error)
193 		goto out_del_cursor;
194 	if (i == 0) {
195 		/* nothing of that length left in the AG, we are done */
196 		tcur->ar_blockcount = 0;
197 		goto out_del_cursor;
198 	}
199 
200 	/*
201 	 * Loop until we are done with all extents that are large
202 	 * enough to be worth discarding or we hit batch limits.
203 	 */
204 	while (i) {
205 		xfs_agblock_t	fbno;
206 		xfs_extlen_t	flen;
207 		xfs_daddr_t	dbno;
208 		xfs_extlen_t	dlen;
209 
210 		error = xfs_alloc_get_rec(cur, &fbno, &flen, &i);
211 		if (error)
212 			break;
213 		if (XFS_IS_CORRUPT(mp, i != 1)) {
214 			xfs_btree_mark_sick(cur);
215 			error = -EFSCORRUPTED;
216 			break;
217 		}
218 
219 		if (--batch <= 0) {
220 			/*
221 			 * Update the cursor to point at this extent so we
222 			 * restart the next batch from this extent.
223 			 */
224 			tcur->ar_startblock = fbno;
225 			tcur->ar_blockcount = flen;
226 			break;
227 		}
228 
229 		/*
230 		 * use daddr format for all range/len calculations as that is
231 		 * the format the range/len variables are supplied in by
232 		 * userspace.
233 		 */
234 		dbno = XFS_AGB_TO_DADDR(mp, pag->pag_agno, fbno);
235 		dlen = XFS_FSB_TO_BB(mp, flen);
236 
237 		/*
238 		 * Too small?  Give up.
239 		 */
240 		if (dlen < minlen) {
241 			trace_xfs_discard_toosmall(mp, pag->pag_agno, fbno, flen);
242 			tcur->ar_blockcount = 0;
243 			break;
244 		}
245 
246 		/*
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.
250 		 */
251 		if (dbno + dlen < start || dbno > end) {
252 			trace_xfs_discard_exclude(mp, pag->pag_agno, fbno, flen);
253 			goto next_extent;
254 		}
255 
256 		/*
257 		 * If any blocks in the range are still busy, skip the
258 		 * discard and try again the next time.
259 		 */
260 		if (xfs_extent_busy_search(mp, pag, fbno, flen)) {
261 			trace_xfs_discard_busy(mp, pag->pag_agno, fbno, flen);
262 			goto next_extent;
263 		}
264 
265 		xfs_extent_busy_insert_discard(pag, fbno, flen,
266 				&extents->extent_list);
267 		*blocks_trimmed += flen;
268 next_extent:
269 		error = xfs_btree_decrement(cur, 0, &i);
270 		if (error)
271 			break;
272 
273 		/*
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.
277 		 */
278 		if (i == 0)
279 			tcur->ar_blockcount = 0;
280 	}
281 
282 	/*
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.
285 	 */
286 	if (error)
287 		xfs_extent_busy_clear(mp, &extents->extent_list, false);
288 out_del_cursor:
289 	xfs_btree_del_cursor(cur, error);
290 out_trans_cancel:
291 	xfs_trans_cancel(tp);
292 	return error;
293 }
294 
295 static bool
296 xfs_trim_should_stop(void)
297 {
298 	return fatal_signal_pending(current) || freezing(current);
299 }
300 
301 /*
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.
305  */
306 static int
307 xfs_trim_extents(
308 	struct xfs_perag	*pag,
309 	xfs_daddr_t		start,
310 	xfs_daddr_t		end,
311 	xfs_daddr_t		minlen,
312 	uint64_t		*blocks_trimmed)
313 {
314 	struct xfs_alloc_rec_incore tcur = {
315 		.ar_blockcount = pag->pagf_longest,
316 		.ar_startblock = NULLAGBLOCK,
317 	};
318 	int			error = 0;
319 
320 	do {
321 		struct xfs_busy_extents	*extents;
322 
323 		extents = kzalloc(sizeof(*extents), GFP_KERNEL);
324 		if (!extents) {
325 			error = -ENOMEM;
326 			break;
327 		}
328 
329 		extents->mount = pag->pag_mount;
330 		extents->owner = extents;
331 		INIT_LIST_HEAD(&extents->extent_list);
332 
333 		error = xfs_trim_gather_extents(pag, start, end, minlen,
334 				&tcur, extents, blocks_trimmed);
335 		if (error) {
336 			kfree(extents);
337 			break;
338 		}
339 
340 		/*
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.
345 		 *
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.
349 		 */
350 		error = xfs_discard_extents(pag->pag_mount, extents);
351 		if (error)
352 			break;
353 
354 		if (xfs_trim_should_stop())
355 			break;
356 
357 	} while (tcur.ar_blockcount != 0);
358 
359 	return error;
360 
361 }
362 
363 /*
364  * trim a range of the filesystem.
365  *
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.
371  */
372 int
373 xfs_ioc_trim(
374 	struct xfs_mount		*mp,
375 	struct fstrim_range __user	*urange)
376 {
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;
382 	xfs_agnumber_t		agno;
383 	uint64_t		blocks_trimmed = 0;
384 	int			error, last_error = 0;
385 
386 	if (!capable(CAP_SYS_ADMIN))
387 		return -EPERM;
388 	if (!bdev_max_discard_sectors(mp->m_ddev_targp->bt_bdev))
389 		return -EOPNOTSUPP;
390 
391 	/*
392 	 * We haven't recovered the log, so we cannot use our bnobt-guided
393 	 * storage zapping commands.
394 	 */
395 	if (xfs_has_norecovery(mp))
396 		return -EROFS;
397 
398 	if (copy_from_user(&range, urange, sizeof(range)))
399 		return -EFAULT;
400 
401 	range.minlen = max_t(u64, granularity, range.minlen);
402 	minlen = BTOBB(range.minlen);
403 	/*
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.
409 	 */
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)
413 		return -EINVAL;
414 
415 	start = BTOBB(range.start);
416 	end = start + BTOBBT(range.len) - 1;
417 
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;
420 
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,
424 					  &blocks_trimmed);
425 		if (error)
426 			last_error = error;
427 
428 		if (xfs_trim_should_stop()) {
429 			xfs_perag_rele(pag);
430 			break;
431 		}
432 	}
433 
434 	if (last_error)
435 		return last_error;
436 
437 	range.len = XFS_FSB_TO_B(mp, blocks_trimmed);
438 	if (copy_to_user(urange, &range, sizeof(range)))
439 		return -EFAULT;
440 	return 0;
441 }
442