715 lines · c
1// SPDX-License-Identifier: GPL-2.02/*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#include "xfs_rtbitmap.h"24 25/*26 * Notes on an efficient, low latency fstrim algorithm27 *28 * We need to walk the filesystem free space and issue discards on the free29 * space that meet the search criteria (size and location). We cannot issue30 * discards on extents that might be in use, or are so recently in use they are31 * still marked as busy. To serialise against extent state changes whilst we are32 * gathering extents to trim, we must hold the AGF lock to lock out other33 * allocations and extent free operations that might change extent state.34 *35 * However, we cannot just hold the AGF for the entire AG free space walk whilst36 * we issue discards on each free space that is found. Storage devices can have37 * extremely slow discard implementations (e.g. ceph RBD) and so walking a38 * couple of million free extents and issuing synchronous discards on each39 * extent can take a *long* time. Whilst we are doing this walk, nothing else40 * can access the AGF, and we can stall transactions and hence the log whilst41 * modifications wait for the AGF lock to be released. This can lead hung tasks42 * kicking the hung task timer and rebooting the system. This is bad.43 *44 * Hence we need to take a leaf from the bulkstat playbook. It takes the AGI45 * lock, gathers a range of inode cluster buffers that are allocated, drops the46 * AGI lock and then reads all the inode cluster buffers and processes them. It47 * loops doing this, using a cursor to keep track of where it is up to in the AG48 * for each iteration to restart the INOBT lookup from.49 *50 * We can't do this exactly with free space - once we drop the AGF lock, the51 * state of the free extent is out of our control and we cannot run a discard52 * safely on it in this situation. Unless, of course, we've marked the free53 * extent as busy and undergoing a discard operation whilst we held the AGF54 * locked.55 *56 * This is exactly how online discard works - free extents are marked busy when57 * they are freed, and once the extent free has been committed to the journal,58 * the busy extent record is marked as "undergoing discard" and the discard is59 * then issued on the free extent. Once the discard completes, the busy extent60 * record is removed and the extent is able to be allocated again.61 *62 * In the context of fstrim, if we find a free extent we need to discard, we63 * don't have to discard it immediately. All we need to do it record that free64 * extent as being busy and under discard, and all the allocation routines will65 * now avoid trying to allocate it. Hence if we mark the extent as busy under66 * the AGF lock, we can safely discard it without holding the AGF lock because67 * nothing will attempt to allocate that free space until the discard completes.68 *69 * This also allows us to issue discards asynchronously like we do with online70 * discard, and so for fast devices fstrim will run much faster as we can have71 * multiple discard operations in flight at once, as well as pipeline the free72 * extent search so that it overlaps in flight discard IO.73 */74 75struct workqueue_struct *xfs_discard_wq;76 77static void78xfs_discard_endio_work(79 struct work_struct *work)80{81 struct xfs_busy_extents *extents =82 container_of(work, struct xfs_busy_extents, endio_work);83 84 xfs_extent_busy_clear(extents->mount, &extents->extent_list, false);85 kfree(extents->owner);86}87 88/*89 * Queue up the actual completion to a thread to avoid IRQ-safe locking for90 * pagb_lock.91 */92static void93xfs_discard_endio(94 struct bio *bio)95{96 struct xfs_busy_extents *extents = bio->bi_private;97 98 INIT_WORK(&extents->endio_work, xfs_discard_endio_work);99 queue_work(xfs_discard_wq, &extents->endio_work);100 bio_put(bio);101}102 103/*104 * Walk the discard list and issue discards on all the busy extents in the105 * list. We plug and chain the bios so that we only need a single completion106 * call to clear all the busy extents once the discards are complete.107 */108int109xfs_discard_extents(110 struct xfs_mount *mp,111 struct xfs_busy_extents *extents)112{113 struct xfs_extent_busy *busyp;114 struct bio *bio = NULL;115 struct blk_plug plug;116 int error = 0;117 118 blk_start_plug(&plug);119 list_for_each_entry(busyp, &extents->extent_list, list) {120 trace_xfs_discard_extent(mp, busyp->agno, busyp->bno,121 busyp->length);122 123 error = __blkdev_issue_discard(mp->m_ddev_targp->bt_bdev,124 XFS_AGB_TO_DADDR(mp, busyp->agno, busyp->bno),125 XFS_FSB_TO_BB(mp, busyp->length),126 GFP_KERNEL, &bio);127 if (error && error != -EOPNOTSUPP) {128 xfs_info(mp,129 "discard failed for extent [0x%llx,%u], error %d",130 (unsigned long long)busyp->bno,131 busyp->length,132 error);133 break;134 }135 }136 137 if (bio) {138 bio->bi_private = extents;139 bio->bi_end_io = xfs_discard_endio;140 submit_bio(bio);141 } else {142 xfs_discard_endio_work(&extents->endio_work);143 }144 blk_finish_plug(&plug);145 146 return error;147}148 149struct xfs_trim_cur {150 xfs_agblock_t start;151 xfs_extlen_t count;152 xfs_agblock_t end;153 xfs_extlen_t minlen;154 bool by_bno;155};156 157static int158xfs_trim_gather_extents(159 struct xfs_perag *pag,160 struct xfs_trim_cur *tcur,161 struct xfs_busy_extents *extents)162{163 struct xfs_mount *mp = pag->pag_mount;164 struct xfs_trans *tp;165 struct xfs_btree_cur *cur;166 struct xfs_buf *agbp;167 int error;168 int i;169 int batch = 100;170 171 /*172 * Force out the log. This means any transactions that might have freed173 * space before we take the AGF buffer lock are now on disk, and the174 * volatile disk cache is flushed.175 */176 xfs_log_force(mp, XFS_LOG_SYNC);177 178 error = xfs_trans_alloc_empty(mp, &tp);179 if (error)180 return error;181 182 error = xfs_alloc_read_agf(pag, tp, 0, &agbp);183 if (error)184 goto out_trans_cancel;185 186 if (tcur->by_bno) {187 /* sub-AG discard request always starts at tcur->start */188 cur = xfs_bnobt_init_cursor(mp, tp, agbp, pag);189 error = xfs_alloc_lookup_le(cur, tcur->start, 0, &i);190 if (!error && !i)191 error = xfs_alloc_lookup_ge(cur, tcur->start, 0, &i);192 } else if (tcur->start == 0) {193 /* first time through a by-len starts with max length */194 cur = xfs_cntbt_init_cursor(mp, tp, agbp, pag);195 error = xfs_alloc_lookup_ge(cur, 0, tcur->count, &i);196 } else {197 /* nth time through a by-len starts where we left off */198 cur = xfs_cntbt_init_cursor(mp, tp, agbp, pag);199 error = xfs_alloc_lookup_le(cur, tcur->start, tcur->count, &i);200 }201 if (error)202 goto out_del_cursor;203 if (i == 0) {204 /* nothing of that length left in the AG, we are done */205 tcur->count = 0;206 goto out_del_cursor;207 }208 209 /*210 * Loop until we are done with all extents that are large211 * enough to be worth discarding or we hit batch limits.212 */213 while (i) {214 xfs_agblock_t fbno;215 xfs_extlen_t flen;216 217 error = xfs_alloc_get_rec(cur, &fbno, &flen, &i);218 if (error)219 break;220 if (XFS_IS_CORRUPT(mp, i != 1)) {221 xfs_btree_mark_sick(cur);222 error = -EFSCORRUPTED;223 break;224 }225 226 if (--batch <= 0) {227 /*228 * Update the cursor to point at this extent so we229 * restart the next batch from this extent.230 */231 tcur->start = fbno;232 tcur->count = flen;233 break;234 }235 236 /*237 * If the extent is entirely outside of the range we are238 * supposed to skip it. Do not bother to trim down partially239 * overlapping ranges for now.240 */241 if (fbno + flen < tcur->start) {242 trace_xfs_discard_exclude(mp, pag->pag_agno, fbno, flen);243 goto next_extent;244 }245 if (fbno > tcur->end) {246 trace_xfs_discard_exclude(mp, pag->pag_agno, fbno, flen);247 if (tcur->by_bno) {248 tcur->count = 0;249 break;250 }251 goto next_extent;252 }253 254 /* Trim the extent returned to the range we want. */255 if (fbno < tcur->start) {256 flen -= tcur->start - fbno;257 fbno = tcur->start;258 }259 if (fbno + flen > tcur->end + 1)260 flen = tcur->end - fbno + 1;261 262 /* Too small? Give up. */263 if (flen < tcur->minlen) {264 trace_xfs_discard_toosmall(mp, pag->pag_agno, fbno, flen);265 if (tcur->by_bno)266 goto next_extent;267 tcur->count = 0;268 break;269 }270 271 /*272 * If any blocks in the range are still busy, skip the273 * discard and try again the next time.274 */275 if (xfs_extent_busy_search(mp, pag, fbno, flen)) {276 trace_xfs_discard_busy(mp, pag->pag_agno, fbno, flen);277 goto next_extent;278 }279 280 xfs_extent_busy_insert_discard(pag, fbno, flen,281 &extents->extent_list);282next_extent:283 if (tcur->by_bno)284 error = xfs_btree_increment(cur, 0, &i);285 else286 error = xfs_btree_decrement(cur, 0, &i);287 if (error)288 break;289 290 /*291 * If there's no more records in the tree, we are done. Set the292 * cursor block count to 0 to indicate to the caller that there293 * is no more extents to search.294 */295 if (i == 0)296 tcur->count = 0;297 }298 299 /*300 * If there was an error, release all the gathered busy extents because301 * we aren't going to issue a discard on them any more.302 */303 if (error)304 xfs_extent_busy_clear(mp, &extents->extent_list, false);305out_del_cursor:306 xfs_btree_del_cursor(cur, error);307out_trans_cancel:308 xfs_trans_cancel(tp);309 return error;310}311 312static bool313xfs_trim_should_stop(void)314{315 return fatal_signal_pending(current) || freezing(current);316}317 318/*319 * Iterate the free list gathering extents and discarding them. We need a cursor320 * for the repeated iteration of gather/discard loop, so use the longest extent321 * we found in the last batch as the key to start the next.322 */323static int324xfs_trim_perag_extents(325 struct xfs_perag *pag,326 xfs_agblock_t start,327 xfs_agblock_t end,328 xfs_extlen_t minlen)329{330 struct xfs_trim_cur tcur = {331 .start = start,332 .count = pag->pagf_longest,333 .end = end,334 .minlen = minlen,335 };336 int error = 0;337 338 if (start != 0 || end != pag->block_count)339 tcur.by_bno = true;340 341 do {342 struct xfs_busy_extents *extents;343 344 extents = kzalloc(sizeof(*extents), GFP_KERNEL);345 if (!extents) {346 error = -ENOMEM;347 break;348 }349 350 extents->mount = pag->pag_mount;351 extents->owner = extents;352 INIT_LIST_HEAD(&extents->extent_list);353 354 error = xfs_trim_gather_extents(pag, &tcur, extents);355 if (error) {356 kfree(extents);357 break;358 }359 360 /*361 * We hand the extent list to the discard function here so the362 * discarded extents can be removed from the busy extent list.363 * This allows the discards to run asynchronously with gathering364 * the next round of extents to discard.365 *366 * However, we must ensure that we do not reference the extent367 * list after this function call, as it may have been freed by368 * the time control returns to us.369 */370 error = xfs_discard_extents(pag->pag_mount, extents);371 if (error)372 break;373 374 if (xfs_trim_should_stop())375 break;376 377 } while (tcur.count != 0);378 379 return error;380 381}382 383static int384xfs_trim_datadev_extents(385 struct xfs_mount *mp,386 xfs_daddr_t start,387 xfs_daddr_t end,388 xfs_extlen_t minlen)389{390 xfs_agnumber_t start_agno, end_agno;391 xfs_agblock_t start_agbno, end_agbno;392 xfs_daddr_t ddev_end;393 struct xfs_perag *pag;394 int last_error = 0, error;395 396 ddev_end = min_t(xfs_daddr_t, end,397 XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks) - 1);398 399 start_agno = xfs_daddr_to_agno(mp, start);400 start_agbno = xfs_daddr_to_agbno(mp, start);401 end_agno = xfs_daddr_to_agno(mp, ddev_end);402 end_agbno = xfs_daddr_to_agbno(mp, ddev_end);403 404 for_each_perag_range(mp, start_agno, end_agno, pag) {405 xfs_agblock_t agend = pag->block_count;406 407 if (start_agno == end_agno)408 agend = end_agbno;409 error = xfs_trim_perag_extents(pag, start_agbno, agend, minlen);410 if (error)411 last_error = error;412 413 if (xfs_trim_should_stop()) {414 xfs_perag_rele(pag);415 break;416 }417 start_agbno = 0;418 }419 420 return last_error;421}422 423#ifdef CONFIG_XFS_RT424struct xfs_trim_rtdev {425 /* list of rt extents to free */426 struct list_head extent_list;427 428 /* minimum length that caller allows us to trim */429 xfs_rtblock_t minlen_fsb;430 431 /* restart point for the rtbitmap walk */432 xfs_rtxnum_t restart_rtx;433 434 /* stopping point for the current rtbitmap walk */435 xfs_rtxnum_t stop_rtx;436};437 438struct xfs_rtx_busy {439 struct list_head list;440 xfs_rtblock_t bno;441 xfs_rtblock_t length;442};443 444static void445xfs_discard_free_rtdev_extents(446 struct xfs_trim_rtdev *tr)447{448 struct xfs_rtx_busy *busyp, *n;449 450 list_for_each_entry_safe(busyp, n, &tr->extent_list, list) {451 list_del_init(&busyp->list);452 kfree(busyp);453 }454}455 456/*457 * Walk the discard list and issue discards on all the busy extents in the458 * list. We plug and chain the bios so that we only need a single completion459 * call to clear all the busy extents once the discards are complete.460 */461static int462xfs_discard_rtdev_extents(463 struct xfs_mount *mp,464 struct xfs_trim_rtdev *tr)465{466 struct block_device *bdev = mp->m_rtdev_targp->bt_bdev;467 struct xfs_rtx_busy *busyp;468 struct bio *bio = NULL;469 struct blk_plug plug;470 xfs_rtblock_t start = NULLRTBLOCK, length = 0;471 int error = 0;472 473 blk_start_plug(&plug);474 list_for_each_entry(busyp, &tr->extent_list, list) {475 if (start == NULLRTBLOCK)476 start = busyp->bno;477 length += busyp->length;478 479 trace_xfs_discard_rtextent(mp, busyp->bno, busyp->length);480 481 error = __blkdev_issue_discard(bdev,482 XFS_FSB_TO_BB(mp, busyp->bno),483 XFS_FSB_TO_BB(mp, busyp->length),484 GFP_NOFS, &bio);485 if (error)486 break;487 }488 xfs_discard_free_rtdev_extents(tr);489 490 if (bio) {491 error = submit_bio_wait(bio);492 if (error == -EOPNOTSUPP)493 error = 0;494 if (error)495 xfs_info(mp,496 "discard failed for rtextent [0x%llx,%llu], error %d",497 (unsigned long long)start,498 (unsigned long long)length,499 error);500 bio_put(bio);501 }502 blk_finish_plug(&plug);503 504 return error;505}506 507static int508xfs_trim_gather_rtextent(509 struct xfs_mount *mp,510 struct xfs_trans *tp,511 const struct xfs_rtalloc_rec *rec,512 void *priv)513{514 struct xfs_trim_rtdev *tr = priv;515 struct xfs_rtx_busy *busyp;516 xfs_rtblock_t rbno, rlen;517 518 if (rec->ar_startext > tr->stop_rtx) {519 /*520 * If we've scanned a large number of rtbitmap blocks, update521 * the cursor to point at this extent so we restart the next522 * batch from this extent.523 */524 tr->restart_rtx = rec->ar_startext;525 return -ECANCELED;526 }527 528 rbno = xfs_rtx_to_rtb(mp, rec->ar_startext);529 rlen = xfs_rtx_to_rtb(mp, rec->ar_extcount);530 531 /* Ignore too small. */532 if (rlen < tr->minlen_fsb) {533 trace_xfs_discard_rttoosmall(mp, rbno, rlen);534 return 0;535 }536 537 busyp = kzalloc(sizeof(struct xfs_rtx_busy), GFP_KERNEL);538 if (!busyp)539 return -ENOMEM;540 541 busyp->bno = rbno;542 busyp->length = rlen;543 INIT_LIST_HEAD(&busyp->list);544 list_add_tail(&busyp->list, &tr->extent_list);545 546 tr->restart_rtx = rec->ar_startext + rec->ar_extcount;547 return 0;548}549 550static int551xfs_trim_rtdev_extents(552 struct xfs_mount *mp,553 xfs_daddr_t start,554 xfs_daddr_t end,555 xfs_daddr_t minlen)556{557 struct xfs_trim_rtdev tr = {558 .minlen_fsb = XFS_BB_TO_FSB(mp, minlen),559 };560 xfs_rtxnum_t low, high;561 struct xfs_trans *tp;562 xfs_daddr_t rtdev_daddr;563 int error;564 565 INIT_LIST_HEAD(&tr.extent_list);566 567 /* Shift the start and end downwards to match the rt device. */568 rtdev_daddr = XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks);569 if (start > rtdev_daddr)570 start -= rtdev_daddr;571 else572 start = 0;573 574 if (end <= rtdev_daddr)575 return 0;576 end -= rtdev_daddr;577 578 error = xfs_trans_alloc_empty(mp, &tp);579 if (error)580 return error;581 582 end = min_t(xfs_daddr_t, end,583 XFS_FSB_TO_BB(mp, mp->m_sb.sb_rblocks) - 1);584 585 /* Convert the rt blocks to rt extents */586 low = xfs_rtb_to_rtxup(mp, XFS_BB_TO_FSB(mp, start));587 high = xfs_rtb_to_rtx(mp, XFS_BB_TO_FSBT(mp, end));588 589 /*590 * Walk the free ranges between low and high. The query_range function591 * trims the extents returned.592 */593 do {594 tr.stop_rtx = low + (mp->m_sb.sb_blocksize * NBBY);595 xfs_rtbitmap_lock_shared(mp, XFS_RBMLOCK_BITMAP);596 error = xfs_rtalloc_query_range(mp, tp, low, high,597 xfs_trim_gather_rtextent, &tr);598 599 if (error == -ECANCELED)600 error = 0;601 if (error) {602 xfs_rtbitmap_unlock_shared(mp, XFS_RBMLOCK_BITMAP);603 xfs_discard_free_rtdev_extents(&tr);604 break;605 }606 607 if (list_empty(&tr.extent_list)) {608 xfs_rtbitmap_unlock_shared(mp, XFS_RBMLOCK_BITMAP);609 break;610 }611 612 error = xfs_discard_rtdev_extents(mp, &tr);613 xfs_rtbitmap_unlock_shared(mp, XFS_RBMLOCK_BITMAP);614 if (error)615 break;616 617 low = tr.restart_rtx;618 } while (!xfs_trim_should_stop() && low <= high);619 620 xfs_trans_cancel(tp);621 return error;622}623#else624# define xfs_trim_rtdev_extents(...) (-EOPNOTSUPP)625#endif /* CONFIG_XFS_RT */626 627/*628 * trim a range of the filesystem.629 *630 * Note: the parameters passed from userspace are byte ranges into the631 * filesystem which does not match to the format we use for filesystem block632 * addressing. FSB addressing is sparse (AGNO|AGBNO), while the incoming format633 * is a linear address range. Hence we need to use DADDR based conversions and634 * comparisons for determining the correct offset and regions to trim.635 *636 * The realtime device is mapped into the FITRIM "address space" immediately637 * after the data device.638 */639int640xfs_ioc_trim(641 struct xfs_mount *mp,642 struct fstrim_range __user *urange)643{644 unsigned int granularity =645 bdev_discard_granularity(mp->m_ddev_targp->bt_bdev);646 struct block_device *rt_bdev = NULL;647 struct fstrim_range range;648 xfs_daddr_t start, end;649 xfs_extlen_t minlen;650 xfs_rfsblock_t max_blocks;651 int error, last_error = 0;652 653 if (!capable(CAP_SYS_ADMIN))654 return -EPERM;655 if (mp->m_rtdev_targp &&656 bdev_max_discard_sectors(mp->m_rtdev_targp->bt_bdev))657 rt_bdev = mp->m_rtdev_targp->bt_bdev;658 if (!bdev_max_discard_sectors(mp->m_ddev_targp->bt_bdev) && !rt_bdev)659 return -EOPNOTSUPP;660 661 if (rt_bdev)662 granularity = max(granularity,663 bdev_discard_granularity(rt_bdev));664 665 /*666 * We haven't recovered the log, so we cannot use our bnobt-guided667 * storage zapping commands.668 */669 if (xfs_has_norecovery(mp))670 return -EROFS;671 672 if (copy_from_user(&range, urange, sizeof(range)))673 return -EFAULT;674 675 range.minlen = max_t(u64, granularity, range.minlen);676 minlen = XFS_B_TO_FSB(mp, range.minlen);677 678 /*679 * Truncating down the len isn't actually quite correct, but using680 * BBTOB would mean we trivially get overflows for values681 * of ULLONG_MAX or slightly lower. And ULLONG_MAX is the default682 * used by the fstrim application. In the end it really doesn't683 * matter as trimming blocks is an advisory interface.684 */685 max_blocks = mp->m_sb.sb_dblocks + mp->m_sb.sb_rblocks;686 if (range.start >= XFS_FSB_TO_B(mp, max_blocks) ||687 range.minlen > XFS_FSB_TO_B(mp, mp->m_ag_max_usable) ||688 range.len < mp->m_sb.sb_blocksize)689 return -EINVAL;690 691 start = BTOBB(range.start);692 end = start + BTOBBT(range.len) - 1;693 694 if (bdev_max_discard_sectors(mp->m_ddev_targp->bt_bdev)) {695 error = xfs_trim_datadev_extents(mp, start, end, minlen);696 if (error)697 last_error = error;698 }699 700 if (rt_bdev && !xfs_trim_should_stop()) {701 error = xfs_trim_rtdev_extents(mp, start, end, minlen);702 if (error)703 last_error = error;704 }705 706 if (last_error)707 return last_error;708 709 range.len = min_t(unsigned long long, range.len,710 XFS_FSB_TO_B(mp, max_blocks) - range.start);711 if (copy_to_user(urange, &range, sizeof(range)))712 return -EFAULT;713 return 0;714}715