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 74struct workqueue_struct *xfs_discard_wq; 75 76static void 77xfs_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 */ 91static void 92xfs_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 */ 107int 108xfs_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 148struct xfs_trim_cur { 149 xfs_agblock_t start; 150 xfs_extlen_t count; 151 xfs_agblock_t end; 152 xfs_extlen_t minlen; 153 bool by_bno; 154}; 155 156static int 157xfs_trim_gather_extents( 158 struct xfs_perag *pag, 159 struct xfs_trim_cur *tcur, 160 struct xfs_busy_extents *extents, 161 uint64_t *blocks_trimmed) 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 freed 173 * space before we take the AGF buffer lock are now on disk, and the 174 * 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 large 211 * 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 we 229 * 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 are 238 * supposed to skip it. Do not bother to trim down partially 239 * 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 the 273 * 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); 282 *blocks_trimmed += flen; 283next_extent: 284 if (tcur->by_bno) 285 error = xfs_btree_increment(cur, 0, &i); 286 else 287 error = xfs_btree_decrement(cur, 0, &i); 288 if (error) 289 break; 290 291 /* 292 * If there's no more records in the tree, we are done. Set the 293 * cursor block count to 0 to indicate to the caller that there 294 * is no more extents to search. 295 */ 296 if (i == 0) 297 tcur->count = 0; 298 } 299 300 /* 301 * If there was an error, release all the gathered busy extents because 302 * we aren't going to issue a discard on them any more. 303 */ 304 if (error) 305 xfs_extent_busy_clear(mp, &extents->extent_list, false); 306out_del_cursor: 307 xfs_btree_del_cursor(cur, error); 308out_trans_cancel: 309 xfs_trans_cancel(tp); 310 return error; 311} 312 313static bool 314xfs_trim_should_stop(void) 315{ 316 return fatal_signal_pending(current) || freezing(current); 317} 318 319/* 320 * Iterate the free list gathering extents and discarding them. We need a cursor 321 * for the repeated iteration of gather/discard loop, so use the longest extent 322 * we found in the last batch as the key to start the next. 323 */ 324static int 325xfs_trim_extents( 326 struct xfs_perag *pag, 327 xfs_agblock_t start, 328 xfs_agblock_t end, 329 xfs_extlen_t minlen, 330 uint64_t *blocks_trimmed) 331{ 332 struct xfs_trim_cur tcur = { 333 .start = start, 334 .count = pag->pagf_longest, 335 .end = end, 336 .minlen = minlen, 337 }; 338 int error = 0; 339 340 if (start != 0 || end != pag->block_count) 341 tcur.by_bno = true; 342 343 do { 344 struct xfs_busy_extents *extents; 345 346 extents = kzalloc(sizeof(*extents), GFP_KERNEL); 347 if (!extents) { 348 error = -ENOMEM; 349 break; 350 } 351 352 extents->mount = pag->pag_mount; 353 extents->owner = extents; 354 INIT_LIST_HEAD(&extents->extent_list); 355 356 error = xfs_trim_gather_extents(pag, &tcur, extents, 357 blocks_trimmed); 358 if (error) { 359 kfree(extents); 360 break; 361 } 362 363 /* 364 * We hand the extent list to the discard function here so the 365 * discarded extents can be removed from the busy extent list. 366 * This allows the discards to run asynchronously with gathering 367 * the next round of extents to discard. 368 * 369 * However, we must ensure that we do not reference the extent 370 * list after this function call, as it may have been freed by 371 * the time control returns to us. 372 */ 373 error = xfs_discard_extents(pag->pag_mount, extents); 374 if (error) 375 break; 376 377 if (xfs_trim_should_stop()) 378 break; 379 380 } while (tcur.count != 0); 381 382 return error; 383 384} 385 386/* 387 * trim a range of the filesystem. 388 * 389 * Note: the parameters passed from userspace are byte ranges into the 390 * filesystem which does not match to the format we use for filesystem block 391 * addressing. FSB addressing is sparse (AGNO|AGBNO), while the incoming format 392 * is a linear address range. Hence we need to use DADDR based conversions and 393 * comparisons for determining the correct offset and regions to trim. 394 */ 395int 396xfs_ioc_trim( 397 struct xfs_mount *mp, 398 struct fstrim_range __user *urange) 399{ 400 struct xfs_perag *pag; 401 unsigned int granularity = 402 bdev_discard_granularity(mp->m_ddev_targp->bt_bdev); 403 struct fstrim_range range; 404 xfs_daddr_t start, end; 405 xfs_extlen_t minlen; 406 xfs_agnumber_t start_agno, end_agno; 407 xfs_agblock_t start_agbno, end_agbno; 408 uint64_t blocks_trimmed = 0; 409 int error, last_error = 0; 410 411 if (!capable(CAP_SYS_ADMIN)) 412 return -EPERM; 413 if (!bdev_max_discard_sectors(mp->m_ddev_targp->bt_bdev)) 414 return -EOPNOTSUPP; 415 416 /* 417 * We haven't recovered the log, so we cannot use our bnobt-guided 418 * storage zapping commands. 419 */ 420 if (xfs_has_norecovery(mp)) 421 return -EROFS; 422 423 if (copy_from_user(&range, urange, sizeof(range))) 424 return -EFAULT; 425 426 range.minlen = max_t(u64, granularity, range.minlen); 427 minlen = XFS_B_TO_FSB(mp, range.minlen); 428 429 /* 430 * Truncating down the len isn't actually quite correct, but using 431 * BBTOB would mean we trivially get overflows for values 432 * of ULLONG_MAX or slightly lower. And ULLONG_MAX is the default 433 * used by the fstrim application. In the end it really doesn't 434 * matter as trimming blocks is an advisory interface. 435 */ 436 if (range.start >= XFS_FSB_TO_B(mp, mp->m_sb.sb_dblocks) || 437 range.minlen > XFS_FSB_TO_B(mp, mp->m_ag_max_usable) || 438 range.len < mp->m_sb.sb_blocksize) 439 return -EINVAL; 440 441 start = BTOBB(range.start); 442 end = min_t(xfs_daddr_t, start + BTOBBT(range.len), 443 XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks)) - 1; 444 445 start_agno = xfs_daddr_to_agno(mp, start); 446 start_agbno = xfs_daddr_to_agbno(mp, start); 447 end_agno = xfs_daddr_to_agno(mp, end); 448 end_agbno = xfs_daddr_to_agbno(mp, end); 449 450 for_each_perag_range(mp, start_agno, end_agno, pag) { 451 xfs_agblock_t agend = pag->block_count; 452 453 if (start_agno == end_agno) 454 agend = end_agbno; 455 error = xfs_trim_extents(pag, start_agbno, agend, minlen, 456 &blocks_trimmed); 457 if (error) 458 last_error = error; 459 460 if (xfs_trim_should_stop()) { 461 xfs_perag_rele(pag); 462 break; 463 } 464 start_agbno = 0; 465 } 466 467 if (last_error) 468 return last_error; 469 470 range.len = XFS_FSB_TO_B(mp, blocks_trimmed); 471 if (copy_to_user(urange, &range, sizeof(range))) 472 return -EFAULT; 473 return 0; 474} 475