xfs: reduce AGF hold times during fstrim operations A recent log space overflow and recovery failure was root caused to a long running truncate blocking on the AGF and ending up pinning the tail of the log. The filesystem then hung, the machine was rebooted, and log recoery then refused to run because there wasn't enough space in the log for EFI transaction reservation. The reason the long running truncate got blocked on the AGF for so long was that an fstrim was being run. THe underlying block device was large and very slow (10TB ceph rbd volume) and so discarding all the free space in the AG took a really long time. The current fstrim implementation holds the AGF across the entire operations - both the freee space scan and the issuing of all the discards. The discards are synchronous and single depth, so if there are millions of free spaces, we hold the AGF lock across millions of discard operations. It doesn't really need to be said that this is a Bad Thing. This series reworks the fstrim discard path to use the same mechanisms as online discard. This allows discards to be issued asynchronously without holding the AGF locked, enabling higher discard queue depths (much faster on fast devices) and only requiring the AGF lock to be held whilst we are scanning free space. To do this, we make use of busy extents - we lock the AGF, mark all the extents we want to discard as "busy under discard" so that nothing will be allowed to allocate them, and then drop the AGF lock. We then issue discards on the gathered busy extents and on discard completion remove them from the busy list. This results in AGF lock holds times for fstrim dropping to a few milliseconds each batch of free extents we scan, and so the hours long hold times that can currently occur on large, slow, badly fragmented device no longer occur. Signed-off-by: Dave Chinner <dchinner@redhat.com> Signed-off-by: Chandan Babu R <chandanbabu@kernel.org> * tag 'xfs-fstrim-busy-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/dgc/linux-xfs: xfs: abort fstrim if kernel is suspending xfs: reduce AGF hold times during fstrim operations xfs: move log discard work to xfs_discard.c
Unified · Split
| @@ -1,6 +1,6 @@ | ||
| 1 | 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | 2 | /* |
| 3 | − * Copyright (C) 2010 Red Hat, Inc. | |
| 3 | + * Copyright (C) 2010, 2023 Red Hat, Inc. | |
| 4 | 4 | * All Rights Reserved. |
| 5 | 5 | */ |
| 6 | 6 | #include "xfs.h" |
| @@ -19,21 +19,147 @@ · 12 unmodified lines | ||
| 19 | 19 | #include "xfs_log.h" |
| 20 | 20 | #include "xfs_ag.h" |
| 21 | 21 | |
| 22 | −STATIC int | |
| 23 | −xfs_trim_extents( | |
| 22 | +/* | |
| 23 | + * Notes on an efficient, low latency fstrim algorithm | |
| 24 | + * | |
| 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. | |
| 31 | + * | |
| 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. | |
| 40 | + * | |
| 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. | |
| 46 | + * | |
| 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 | |
| 51 | + * locked. | |
| 52 | + * | |
| 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. | |
| 58 | + * | |
| 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. | |
| 65 | + * | |
| 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. | |
| 70 | + */ | |
| 71 | + | |
| 72 | +struct workqueue_struct *xfs_discard_wq; | |
| 73 | + | |
| 74 | +static void | |
| 75 | +xfs_discard_endio_work( | |
| 76 | + struct work_struct *work) | |
| 77 | +{ | |
| 78 | + struct xfs_busy_extents *extents = | |
| 79 | + container_of(work, struct xfs_busy_extents, endio_work); | |
| 80 | + | |
| 81 | + xfs_extent_busy_clear(extents->mount, &extents->extent_list, false); | |
| 82 | + kmem_free(extents->owner); | |
| 83 | +} | |
| 84 | + | |
| 85 | +/* | |
| 86 | + * Queue up the actual completion to a thread to avoid IRQ-safe locking for | |
| 87 | + * pagb_lock. | |
| 88 | + */ | |
| 89 | +static void | |
| 90 | +xfs_discard_endio( | |
| 91 | + struct bio *bio) | |
| 92 | +{ | |
| 93 | + struct xfs_busy_extents *extents = bio->bi_private; | |
| 94 | + | |
| 95 | + INIT_WORK(&extents->endio_work, xfs_discard_endio_work); | |
| 96 | + queue_work(xfs_discard_wq, &extents->endio_work); | |
| 97 | + bio_put(bio); | |
| 98 | +} | |
| 99 | + | |
| 100 | +/* | |
| 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. | |
| 104 | + */ | |
| 105 | +int | |
| 106 | +xfs_discard_extents( | |
| 107 | + struct xfs_mount *mp, | |
| 108 | + struct xfs_busy_extents *extents) | |
| 109 | +{ | |
| 110 | + struct xfs_extent_busy *busyp; | |
| 111 | + struct bio *bio = NULL; | |
| 112 | + struct blk_plug plug; | |
| 113 | + int error = 0; | |
| 114 | + | |
| 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, | |
| 118 | + busyp->length); | |
| 119 | + | |
| 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), | |
| 123 | + GFP_NOFS, &bio); | |
| 124 | + if (error && error != -EOPNOTSUPP) { | |
| 125 | + xfs_info(mp, | |
| 126 | + "discard failed for extent [0x%llx,%u], error %d", | |
| 127 | + (unsigned long long)busyp->bno, | |
| 128 | + busyp->length, | |
| 129 | + error); | |
| 130 | + break; | |
| 131 | + } | |
| 132 | + } | |
| 133 | + | |
| 134 | + if (bio) { | |
| 135 | + bio->bi_private = extents; | |
| 136 | + bio->bi_end_io = xfs_discard_endio; | |
| 137 | + submit_bio(bio); | |
| 138 | + } else { | |
| 139 | + xfs_discard_endio_work(&extents->endio_work); | |
| 140 | + } | |
| 141 | + blk_finish_plug(&plug); | |
| 142 | + | |
| 143 | + return error; | |
| 144 | +} | |
| 145 | + | |
| 146 | + | |
| 147 | +static int | |
| 148 | +xfs_trim_gather_extents( | |
| 24 | 149 | struct xfs_perag *pag, |
| 25 | 150 | xfs_daddr_t start, |
| 26 | 151 | xfs_daddr_t end, |
| 27 | 152 | xfs_daddr_t minlen, |
| 153 | + struct xfs_alloc_rec_incore *tcur, | |
| 154 | + struct xfs_busy_extents *extents, | |
| 28 | 155 | uint64_t *blocks_trimmed) |
| 29 | 156 | { |
| 30 | 157 | struct xfs_mount *mp = pag->pag_mount; |
| 31 | − struct block_device *bdev = mp->m_ddev_targp->bt_bdev; | |
| 32 | 158 | struct xfs_btree_cur *cur; |
| 33 | 159 | struct xfs_buf *agbp; |
| 34 | − struct xfs_agf *agf; | |
| 35 | 160 | int error; |
| 36 | 161 | int i; |
| 162 | + int batch = 100; | |
| 37 | 163 | |
| 38 | 164 | /* |
| 39 | 165 | * Force out the log. This means any transactions that might have freed |
| @@ -45,20 +171,28 @@ xfs_trim_extents( · 5 unmodified lines | ||
| 45 | 171 | error = xfs_alloc_read_agf(pag, NULL, 0, &agbp); |
| 46 | 172 | if (error) |
| 47 | 173 | return error; |
| 48 | − agf = agbp->b_addr; | |
| 49 | 174 | |
| 50 | 175 | cur = xfs_allocbt_init_cursor(mp, NULL, agbp, pag, XFS_BTNUM_CNT); |
| 51 | 176 | |
| 52 | 177 | /* |
| 53 | − * Look up the longest btree in the AGF and start with it. | |
| 178 | + * Look up the extent length requested in the AGF and start with it. | |
| 54 | 179 | */ |
| 55 | − error = xfs_alloc_lookup_ge(cur, 0, be32_to_cpu(agf->agf_longest), &i); | |
| 180 | + if (tcur->ar_startblock == NULLAGBLOCK) | |
| 181 | + error = xfs_alloc_lookup_ge(cur, 0, tcur->ar_blockcount, &i); | |
| 182 | + else | |
| 183 | + error = xfs_alloc_lookup_le(cur, tcur->ar_startblock, | |
| 184 | + tcur->ar_blockcount, &i); | |
| 56 | 185 | if (error) |
| 57 | 186 | goto out_del_cursor; |
| 187 | + if (i == 0) { | |
| 188 | + /* nothing of that length left in the AG, we are done */ | |
| 189 | + tcur->ar_blockcount = 0; | |
| 190 | + goto out_del_cursor; | |
| 191 | + } | |
| 58 | 192 | |
| 59 | 193 | /* |
| 60 | 194 | * Loop until we are done with all extents that are large |
| 61 | − * enough to be worth discarding. | |
| 195 | + * enough to be worth discarding or we hit batch limits. | |
| 62 | 196 | */ |
| 63 | 197 | while (i) { |
| 64 | 198 | xfs_agblock_t fbno; |
| @@ -73,7 +207,16 @@ xfs_trim_extents( · 8 unmodified lines | ||
| 73 | 207 | error = -EFSCORRUPTED; |
| 74 | 208 | break; |
| 75 | 209 | } |
| 76 | − ASSERT(flen <= be32_to_cpu(agf->agf_longest)); | |
| 210 | + | |
| 211 | + if (--batch <= 0) { | |
| 212 | + /* | |
| 213 | + * Update the cursor to point at this extent so we | |
| 214 | + * restart the next batch from this extent. | |
| 215 | + */ | |
| 216 | + tcur->ar_startblock = fbno; | |
| 217 | + tcur->ar_blockcount = flen; | |
| 218 | + break; | |
| 219 | + } | |
| 77 | 220 | |
| 78 | 221 | /* |
| 79 | 222 | * use daddr format for all range/len calculations as that is |
| @@ -88,6 +231,7 @@ xfs_trim_extents( · 8 unmodified lines | ||
| 88 | 231 | */ |
| 89 | 232 | if (dlen < minlen) { |
| 90 | 233 | trace_xfs_discard_toosmall(mp, pag->pag_agno, fbno, flen); |
| 234 | + tcur->ar_blockcount = 0; | |
| 91 | 235 | break; |
| 92 | 236 | } |
| 93 | 237 | |
| @@ -110,29 +254,103 @@ xfs_trim_extents( · 16 unmodified lines | ||
| 110 | 254 | goto next_extent; |
| 111 | 255 | } |
| 112 | 256 | |
| 113 | − trace_xfs_discard_extent(mp, pag->pag_agno, fbno, flen); | |
| 114 | − error = blkdev_issue_discard(bdev, dbno, dlen, GFP_NOFS); | |
| 115 | − if (error) | |
| 116 | − break; | |
| 257 | + xfs_extent_busy_insert_discard(pag, fbno, flen, | |
| 258 | + &extents->extent_list); | |
| 117 | 259 | *blocks_trimmed += flen; |
| 118 | − | |
| 119 | 260 | next_extent: |
| 120 | 261 | error = xfs_btree_decrement(cur, 0, &i); |
| 121 | 262 | if (error) |
| 122 | 263 | break; |
| 123 | 264 | |
| 124 | − if (fatal_signal_pending(current)) { | |
| 125 | − error = -ERESTARTSYS; | |
| 126 | − break; | |
| 127 | − } | |
| 265 | + /* | |
| 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. | |
| 269 | + */ | |
| 270 | + if (i == 0) | |
| 271 | + tcur->ar_blockcount = 0; | |
| 128 | 272 | } |
| 129 | 273 | |
| 274 | + /* | |
| 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. | |
| 277 | + */ | |
| 278 | + if (error) | |
| 279 | + xfs_extent_busy_clear(mp, &extents->extent_list, false); | |
| 130 | 280 | out_del_cursor: |
| 131 | 281 | xfs_btree_del_cursor(cur, error); |
| 132 | 282 | xfs_buf_relse(agbp); |
| 133 | 283 | return error; |
| 134 | 284 | } |
| 135 | 285 | |
| 286 | +static bool | |
| 287 | +xfs_trim_should_stop(void) | |
| 288 | +{ | |
| 289 | + return fatal_signal_pending(current) || freezing(current); | |
| 290 | +} | |
| 291 | + | |
| 292 | +/* | |
| 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. | |
| 296 | + */ | |
| 297 | +static int | |
| 298 | +xfs_trim_extents( | |
| 299 | + struct xfs_perag *pag, | |
| 300 | + xfs_daddr_t start, | |
| 301 | + xfs_daddr_t end, | |
| 302 | + xfs_daddr_t minlen, | |
| 303 | + uint64_t *blocks_trimmed) | |
| 304 | +{ | |
| 305 | + struct xfs_alloc_rec_incore tcur = { | |
| 306 | + .ar_blockcount = pag->pagf_longest, | |
| 307 | + .ar_startblock = NULLAGBLOCK, | |
| 308 | + }; | |
| 309 | + int error = 0; | |
| 310 | + | |
| 311 | + do { | |
| 312 | + struct xfs_busy_extents *extents; | |
| 313 | + | |
| 314 | + extents = kzalloc(sizeof(*extents), GFP_KERNEL); | |
| 315 | + if (!extents) { | |
| 316 | + error = -ENOMEM; | |
| 317 | + break; | |
| 318 | + } | |
| 319 | + | |
| 320 | + extents->mount = pag->pag_mount; | |
| 321 | + extents->owner = extents; | |
| 322 | + INIT_LIST_HEAD(&extents->extent_list); | |
| 323 | + | |
| 324 | + error = xfs_trim_gather_extents(pag, start, end, minlen, | |
| 325 | + &tcur, extents, blocks_trimmed); | |
| 326 | + if (error) { | |
| 327 | + kfree(extents); | |
| 328 | + break; | |
| 329 | + } | |
| 330 | + | |
| 331 | + /* | |
| 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. | |
| 336 | + * | |
| 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. | |
| 340 | + */ | |
| 341 | + error = xfs_discard_extents(pag->pag_mount, extents); | |
| 342 | + if (error) | |
| 343 | + break; | |
| 344 | + | |
| 345 | + if (xfs_trim_should_stop()) | |
| 346 | + break; | |
| 347 | + | |
| 348 | + } while (tcur.ar_blockcount != 0); | |
| 349 | + | |
| 350 | + return error; | |
| 351 | + | |
| 352 | +} | |
| 353 | + | |
| 136 | 354 | /* |
| 137 | 355 | * trim a range of the filesystem. |
| 138 | 356 | * |
| @@ -195,12 +413,12 @@ xfs_ioc_trim( · 56 unmodified lines | ||
| 195 | 413 | for_each_perag_range(mp, agno, xfs_daddr_to_agno(mp, end), pag) { |
| 196 | 414 | error = xfs_trim_extents(pag, start, end, minlen, |
| 197 | 415 | &blocks_trimmed); |
| 198 | − if (error) { | |
| 416 | + if (error) | |
| 199 | 417 | last_error = error; |
| 200 | − if (error == -ERESTARTSYS) { | |
| 201 | − xfs_perag_rele(pag); | |
| 202 | − break; | |
| 203 | − } | |
| 418 | + | |
| 419 | + if (xfs_trim_should_stop()) { | |
| 420 | + xfs_perag_rele(pag); | |
| 421 | + break; | |
| 204 | 422 | } |
| 205 | 423 | } |
| 206 | 424 | |
| @@ -3,8 +3,10 @@ · 2 unmodified lines | ||
| 3 | 3 | #define XFS_DISCARD_H 1 |
| 4 | 4 | |
| 5 | 5 | struct fstrim_range; |
| 6 | −struct list_head; | |
| 6 | +struct xfs_mount; | |
| 7 | +struct xfs_busy_extents; | |
| 7 | 8 | |
| 8 | −extern int xfs_ioc_trim(struct xfs_mount *, struct fstrim_range __user *); | |
| 9 | +int xfs_discard_extents(struct xfs_mount *mp, struct xfs_busy_extents *busy); | |
| 10 | +int xfs_ioc_trim(struct xfs_mount *mp, struct fstrim_range __user *fstrim); | |
| 9 | 11 | |
| 10 | 12 | #endif /* XFS_DISCARD_H */ |
| @@ -19,13 +19,13 @@ · 18 unmodified lines | ||
| 19 | 19 | #include "xfs_log.h" |
| 20 | 20 | #include "xfs_ag.h" |
| 21 | 21 | |
| 22 | −void | |
| 23 | −xfs_extent_busy_insert( | |
| 24 | − struct xfs_trans *tp, | |
| 22 | +static void | |
| 23 | +xfs_extent_busy_insert_list( | |
| 25 | 24 | struct xfs_perag *pag, |
| 26 | 25 | xfs_agblock_t bno, |
| 27 | 26 | xfs_extlen_t len, |
| 28 | − unsigned int flags) | |
| 27 | + unsigned int flags, | |
| 28 | + struct list_head *busy_list) | |
| 29 | 29 | { |
| 30 | 30 | struct xfs_extent_busy *new; |
| 31 | 31 | struct xfs_extent_busy *busyp; |
| @@ -40,7 +40,7 @@ xfs_extent_busy_insert( · 8 unmodified lines | ||
| 40 | 40 | new->flags = flags; |
| 41 | 41 | |
| 42 | 42 | /* trace before insert to be able to see failed inserts */ |
| 43 | − trace_xfs_extent_busy(tp->t_mountp, pag->pag_agno, bno, len); | |
| 43 | + trace_xfs_extent_busy(pag->pag_mount, pag->pag_agno, bno, len); | |
| 44 | 44 | |
| 45 | 45 | spin_lock(&pag->pagb_lock); |
| 46 | 46 | rbp = &pag->pagb_tree.rb_node; |
| @@ -62,10 +62,32 @@ xfs_extent_busy_insert( · 15 unmodified lines | ||
| 62 | 62 | rb_link_node(&new->rb_node, parent, rbp); |
| 63 | 63 | rb_insert_color(&new->rb_node, &pag->pagb_tree); |
| 64 | 64 | |
| 65 | − list_add(&new->list, &tp->t_busy); | |
| 65 | + list_add(&new->list, busy_list); | |
| 66 | 66 | spin_unlock(&pag->pagb_lock); |
| 67 | 67 | } |
| 68 | 68 | |
| 69 | +void | |
| 70 | +xfs_extent_busy_insert( | |
| 71 | + struct xfs_trans *tp, | |
| 72 | + struct xfs_perag *pag, | |
| 73 | + xfs_agblock_t bno, | |
| 74 | + xfs_extlen_t len, | |
| 75 | + unsigned int flags) | |
| 76 | +{ | |
| 77 | + xfs_extent_busy_insert_list(pag, bno, len, flags, &tp->t_busy); | |
| 78 | +} | |
| 79 | + | |
| 80 | +void | |
| 81 | +xfs_extent_busy_insert_discard( | |
| 82 | + struct xfs_perag *pag, | |
| 83 | + xfs_agblock_t bno, | |
| 84 | + xfs_extlen_t len, | |
| 85 | + struct list_head *busy_list) | |
| 86 | +{ | |
| 87 | + xfs_extent_busy_insert_list(pag, bno, len, XFS_EXTENT_BUSY_DISCARDED, | |
| 88 | + busy_list); | |
| 89 | +} | |
| 90 | + | |
| 69 | 91 | /* |
| 70 | 92 | * Search for a busy extent within the range of the extent we are about to |
| 71 | 93 | * allocate. You need to be holding the busy extent tree lock when calling |
| @@ -16,9 +16,6 @@ struct xfs_alloc_arg; · 15 unmodified lines | ||
| 16 | 16 | /* |
| 17 | 17 | * Busy block/extent entry. Indexed by a rbtree in perag to mark blocks that |
| 18 | 18 | * have been freed but whose transactions aren't committed to disk yet. |
| 19 | − * | |
| 20 | − * Note that we use the transaction ID to record the transaction, not the | |
| 21 | − * transaction structure itself. See xfs_extent_busy_insert() for details. | |
| 22 | 19 | */ |
| 23 | 20 | struct xfs_extent_busy { |
| 24 | 21 | struct rb_node rb_node; /* ag by-bno indexed search tree */ |
| @@ -31,10 +28,31 @@ struct xfs_extent_busy { · 6 unmodified lines | ||
| 31 | 28 | #define XFS_EXTENT_BUSY_SKIP_DISCARD 0x02 /* do not discard */ |
| 32 | 29 | }; |
| 33 | 30 | |
| 31 | +/* | |
| 32 | + * List used to track groups of related busy extents all the way through | |
| 33 | + * to discard completion. | |
| 34 | + */ | |
| 35 | +struct xfs_busy_extents { | |
| 36 | + struct xfs_mount *mount; | |
| 37 | + struct list_head extent_list; | |
| 38 | + struct work_struct endio_work; | |
| 39 | + | |
| 40 | + /* | |
| 41 | + * Owner is the object containing the struct xfs_busy_extents to free | |
| 42 | + * once the busy extents have been processed. If only the | |
| 43 | + * xfs_busy_extents object needs freeing, then point this at itself. | |
| 44 | + */ | |
| 45 | + void *owner; | |
| 46 | +}; | |
| 47 | + | |
| 34 | 48 | void |
| 35 | 49 | xfs_extent_busy_insert(struct xfs_trans *tp, struct xfs_perag *pag, |
| 36 | 50 | xfs_agblock_t bno, xfs_extlen_t len, unsigned int flags); |
| 37 | 51 | |
| 52 | +void | |
| 53 | +xfs_extent_busy_insert_discard(struct xfs_perag *pag, xfs_agblock_t bno, | |
| 54 | + xfs_extlen_t len, struct list_head *busy_list); | |
| 55 | + | |
| 38 | 56 | void |
| 39 | 57 | xfs_extent_busy_clear(struct xfs_mount *mp, struct list_head *list, |
| 40 | 58 | bool do_discard); |
| @@ -16,8 +16,7 @@ · 15 unmodified lines | ||
| 16 | 16 | #include "xfs_log.h" |
| 17 | 17 | #include "xfs_log_priv.h" |
| 18 | 18 | #include "xfs_trace.h" |
| 19 | − | |
| 20 | −struct workqueue_struct *xfs_discard_wq; | |
| 19 | +#include "xfs_discard.h" | |
| 21 | 20 | |
| 22 | 21 | /* |
| 23 | 22 | * Allocate a new ticket. Failing to get a new ticket makes it really hard to |
| @@ -103,7 +102,7 @@ xlog_cil_ctx_alloc(void) · 79 unmodified lines | ||
| 103 | 102 | |
| 104 | 103 | ctx = kmem_zalloc(sizeof(*ctx), KM_NOFS); |
| 105 | 104 | INIT_LIST_HEAD(&ctx->committing); |
| 106 | − INIT_LIST_HEAD(&ctx->busy_extents); | |
| 105 | + INIT_LIST_HEAD(&ctx->busy_extents.extent_list); | |
| 107 | 106 | INIT_LIST_HEAD(&ctx->log_items); |
| 108 | 107 | INIT_LIST_HEAD(&ctx->lv_chain); |
| 109 | 108 | INIT_WORK(&ctx->push_work, xlog_cil_push_work); |
| @@ -132,7 +131,7 @@ xlog_cil_push_pcp_aggregate( · 22 unmodified lines | ||
| 132 | 131 | |
| 133 | 132 | if (!list_empty(&cilpcp->busy_extents)) { |
| 134 | 133 | list_splice_init(&cilpcp->busy_extents, |
| 135 | − &ctx->busy_extents); | |
| 134 | + &ctx->busy_extents.extent_list); | |
| 136 | 135 | } |
| 137 | 136 | if (!list_empty(&cilpcp->log_items)) |
| 138 | 137 | list_splice_init(&cilpcp->log_items, &ctx->log_items); |
| @@ -708,76 +707,6 @@ xlog_cil_free_logvec( · 569 unmodified lines | ||
| 708 | 707 | } |
| 709 | 708 | } |
| 710 | 709 | |
| 711 | −static void | |
| 712 | −xlog_discard_endio_work( | |
| 713 | − struct work_struct *work) | |
| 714 | −{ | |
| 715 | − struct xfs_cil_ctx *ctx = | |
| 716 | − container_of(work, struct xfs_cil_ctx, discard_endio_work); | |
| 717 | − struct xfs_mount *mp = ctx->cil->xc_log->l_mp; | |
| 718 | − | |
| 719 | − xfs_extent_busy_clear(mp, &ctx->busy_extents, false); | |
| 720 | − kmem_free(ctx); | |
| 721 | −} | |
| 722 | − | |
| 723 | −/* | |
| 724 | − * Queue up the actual completion to a thread to avoid IRQ-safe locking for | |
| 725 | − * pagb_lock. Note that we need a unbounded workqueue, otherwise we might | |
| 726 | − * get the execution delayed up to 30 seconds for weird reasons. | |
| 727 | − */ | |
| 728 | −static void | |
| 729 | −xlog_discard_endio( | |
| 730 | − struct bio *bio) | |
| 731 | −{ | |
| 732 | − struct xfs_cil_ctx *ctx = bio->bi_private; | |
| 733 | − | |
| 734 | − INIT_WORK(&ctx->discard_endio_work, xlog_discard_endio_work); | |
| 735 | − queue_work(xfs_discard_wq, &ctx->discard_endio_work); | |
| 736 | − bio_put(bio); | |
| 737 | −} | |
| 738 | − | |
| 739 | −static void | |
| 740 | −xlog_discard_busy_extents( | |
| 741 | − struct xfs_mount *mp, | |
| 742 | − struct xfs_cil_ctx *ctx) | |
| 743 | −{ | |
| 744 | − struct list_head *list = &ctx->busy_extents; | |
| 745 | − struct xfs_extent_busy *busyp; | |
| 746 | − struct bio *bio = NULL; | |
| 747 | − struct blk_plug plug; | |
| 748 | − int error = 0; | |
| 749 | − | |
| 750 | − ASSERT(xfs_has_discard(mp)); | |
| 751 | − | |
| 752 | − blk_start_plug(&plug); | |
| 753 | − list_for_each_entry(busyp, list, list) { | |
| 754 | − trace_xfs_discard_extent(mp, busyp->agno, busyp->bno, | |
| 755 | − busyp->length); | |
| 756 | − | |
| 757 | − error = __blkdev_issue_discard(mp->m_ddev_targp->bt_bdev, | |
| 758 | − XFS_AGB_TO_DADDR(mp, busyp->agno, busyp->bno), | |
| 759 | − XFS_FSB_TO_BB(mp, busyp->length), | |
| 760 | − GFP_NOFS, &bio); | |
| 761 | − if (error && error != -EOPNOTSUPP) { | |
| 762 | − xfs_info(mp, | |
| 763 | − "discard failed for extent [0x%llx,%u], error %d", | |
| 764 | − (unsigned long long)busyp->bno, | |
| 765 | − busyp->length, | |
| 766 | − error); | |
| 767 | − break; | |
| 768 | − } | |
| 769 | − } | |
| 770 | − | |
| 771 | − if (bio) { | |
| 772 | − bio->bi_private = ctx; | |
| 773 | − bio->bi_end_io = xlog_discard_endio; | |
| 774 | − submit_bio(bio); | |
| 775 | − } else { | |
| 776 | − xlog_discard_endio_work(&ctx->discard_endio_work); | |
| 777 | − } | |
| 778 | − blk_finish_plug(&plug); | |
| 779 | −} | |
| 780 | − | |
| 781 | 710 | /* |
| 782 | 711 | * Mark all items committed and clear busy extents. We free the log vector |
| 783 | 712 | * chains in a separate pass so that we unpin the log items as quickly as |
| @@ -807,8 +736,8 @@ xlog_cil_committed( · 23 unmodified lines | ||
| 807 | 736 | xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, &ctx->lv_chain, |
| 808 | 737 | ctx->start_lsn, abort); |
| 809 | 738 | |
| 810 | − xfs_extent_busy_sort(&ctx->busy_extents); | |
| 811 | − xfs_extent_busy_clear(mp, &ctx->busy_extents, | |
| 739 | + xfs_extent_busy_sort(&ctx->busy_extents.extent_list); | |
| 740 | + xfs_extent_busy_clear(mp, &ctx->busy_extents.extent_list, | |
| 812 | 741 | xfs_has_discard(mp) && !abort); |
| 813 | 742 | |
| 814 | 743 | spin_lock(&ctx->cil->xc_push_lock); |
| @@ -817,10 +746,14 @@ xlog_cil_committed( · 2 unmodified lines | ||
| 817 | 746 | |
| 818 | 747 | xlog_cil_free_logvec(&ctx->lv_chain); |
| 819 | 748 | |
| 820 | − if (!list_empty(&ctx->busy_extents)) | |
| 821 | − xlog_discard_busy_extents(mp, ctx); | |
| 822 | − else | |
| 823 | − kmem_free(ctx); | |
| 749 | + if (!list_empty(&ctx->busy_extents.extent_list)) { | |
| 750 | + ctx->busy_extents.mount = mp; | |
| 751 | + ctx->busy_extents.owner = ctx; | |
| 752 | + xfs_discard_extents(mp, &ctx->busy_extents); | |
| 753 | + return; | |
| 754 | + } | |
| 755 | + | |
| 756 | + kmem_free(ctx); | |
| 824 | 757 | } |
| 825 | 758 | |
| 826 | 759 | void |
| @@ -6,6 +6,8 @@ · 5 unmodified lines | ||
| 6 | 6 | #ifndef __XFS_LOG_PRIV_H__ |
| 7 | 7 | #define __XFS_LOG_PRIV_H__ |
| 8 | 8 | |
| 9 | +#include "xfs_extent_busy.h" /* for struct xfs_busy_extents */ | |
| 10 | + | |
| 9 | 11 | struct xfs_buf; |
| 10 | 12 | struct xlog; |
| 11 | 13 | struct xlog_ticket; |
| @@ -223,12 +225,11 @@ struct xfs_cil_ctx { · 211 unmodified lines | ||
| 223 | 225 | struct xlog_in_core *commit_iclog; |
| 224 | 226 | struct xlog_ticket *ticket; /* chkpt ticket */ |
| 225 | 227 | atomic_t space_used; /* aggregate size of regions */ |
| 226 | − struct list_head busy_extents; /* busy extents in chkpt */ | |
| 228 | + struct xfs_busy_extents busy_extents; | |
| 227 | 229 | struct list_head log_items; /* log items in chkpt */ |
| 228 | 230 | struct list_head lv_chain; /* logvecs being pushed */ |
| 229 | 231 | struct list_head iclog_entry; |
| 230 | 232 | struct list_head committing; /* ctx committing list */ |
| 231 | − struct work_struct discard_endio_work; | |
| 232 | 233 | struct work_struct push_work; |
| 233 | 234 | atomic_t order_id; |
| 234 | 235 | |