| 1 | /* |
| 2 | * Copyright (c) 2000-2005 Silicon Graphics, Inc. |
| 3 | * All Rights Reserved. |
| 4 | * |
| 5 | * This program is free software; you can redistribute it and/or |
| 6 | * modify it under the terms of the GNU General Public License as |
| 7 | * published by the Free Software Foundation. |
| 8 | * |
| 9 | * This program is distributed in the hope that it would be useful, |
| 10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 12 | * GNU General Public License for more details. |
| 13 | * |
| 14 | * You should have received a copy of the GNU General Public License |
| 15 | * along with this program; if not, write the Free Software Foundation, |
| 16 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA |
| 17 | */ |
| 18 | #include "xfs.h" |
| 19 | #include "xfs_fs.h" |
| 20 | #include "xfs_types.h" |
| 21 | #include "xfs_log.h" |
| 22 | #include "xfs_inum.h" |
| 23 | #include "xfs_trans.h" |
| 24 | #include "xfs_trans_priv.h" |
| 25 | #include "xfs_sb.h" |
| 26 | #include "xfs_ag.h" |
| 27 | #include "xfs_mount.h" |
| 28 | #include "xfs_bmap_btree.h" |
| 29 | #include "xfs_inode.h" |
| 30 | #include "xfs_dinode.h" |
| 31 | #include "xfs_error.h" |
| 32 | #include "xfs_filestream.h" |
| 33 | #include "xfs_vnodeops.h" |
| 34 | #include "xfs_inode_item.h" |
| 35 | #include "xfs_quota.h" |
| 36 | #include "xfs_trace.h" |
| 37 | #include "xfs_fsops.h" |
| 38 | |
| 39 | #include <linux/kthread.h> |
| 40 | #include <linux/freezer.h> |
| 41 | |
| 42 | struct workqueue_struct *xfs_syncd_wq; /* sync workqueue */ |
| 43 | |
| 44 | /* |
| 45 | * The inode lookup is done in batches to keep the amount of lock traffic and |
| 46 | * radix tree lookups to a minimum. The batch size is a trade off between |
| 47 | * lookup reduction and stack usage. This is in the reclaim path, so we can't |
| 48 | * be too greedy. |
| 49 | */ |
| 50 | #define XFS_LOOKUP_BATCH 32 |
| 51 | |
| 52 | STATIC int |
| 53 | xfs_inode_ag_walk_grab( |
| 54 | struct xfs_inode *ip) |
| 55 | { |
| 56 | struct inode *inode = VFS_I(ip); |
| 57 | |
| 58 | ASSERT(rcu_read_lock_held()); |
| 59 | |
| 60 | /* |
| 61 | * check for stale RCU freed inode |
| 62 | * |
| 63 | * If the inode has been reallocated, it doesn't matter if it's not in |
| 64 | * the AG we are walking - we are walking for writeback, so if it |
| 65 | * passes all the "valid inode" checks and is dirty, then we'll write |
| 66 | * it back anyway. If it has been reallocated and still being |
| 67 | * initialised, the XFS_INEW check below will catch it. |
| 68 | */ |
| 69 | spin_lock(&ip->i_flags_lock); |
| 70 | if (!ip->i_ino) |
| 71 | goto out_unlock_noent; |
| 72 | |
| 73 | /* avoid new or reclaimable inodes. Leave for reclaim code to flush */ |
| 74 | if (__xfs_iflags_test(ip, XFS_INEW | XFS_IRECLAIMABLE | XFS_IRECLAIM)) |
| 75 | goto out_unlock_noent; |
| 76 | spin_unlock(&ip->i_flags_lock); |
| 77 | |
| 78 | /* nothing to sync during shutdown */ |
| 79 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) |
| 80 | return EFSCORRUPTED; |
| 81 | |
| 82 | /* If we can't grab the inode, it must on it's way to reclaim. */ |
| 83 | if (!igrab(inode)) |
| 84 | return ENOENT; |
| 85 | |
| 86 | if (is_bad_inode(inode)) { |
| 87 | IRELE(ip); |
| 88 | return ENOENT; |
| 89 | } |
| 90 | |
| 91 | /* inode is valid */ |
| 92 | return 0; |
| 93 | |
| 94 | out_unlock_noent: |
| 95 | spin_unlock(&ip->i_flags_lock); |
| 96 | return ENOENT; |
| 97 | } |
| 98 | |
| 99 | STATIC int |
| 100 | xfs_inode_ag_walk( |
| 101 | struct xfs_mount *mp, |
| 102 | struct xfs_perag *pag, |
| 103 | int (*execute)(struct xfs_inode *ip, |
| 104 | struct xfs_perag *pag, int flags), |
| 105 | int flags) |
| 106 | { |
| 107 | uint32_t first_index; |
| 108 | int last_error = 0; |
| 109 | int skipped; |
| 110 | int done; |
| 111 | int nr_found; |
| 112 | |
| 113 | restart: |
| 114 | done = 0; |
| 115 | skipped = 0; |
| 116 | first_index = 0; |
| 117 | nr_found = 0; |
| 118 | do { |
| 119 | struct xfs_inode *batch[XFS_LOOKUP_BATCH]; |
| 120 | int error = 0; |
| 121 | int i; |
| 122 | |
| 123 | rcu_read_lock(); |
| 124 | nr_found = radix_tree_gang_lookup(&pag->pag_ici_root, |
| 125 | (void **)batch, first_index, |
| 126 | XFS_LOOKUP_BATCH); |
| 127 | if (!nr_found) { |
| 128 | rcu_read_unlock(); |
| 129 | break; |
| 130 | } |
| 131 | |
| 132 | /* |
| 133 | * Grab the inodes before we drop the lock. if we found |
| 134 | * nothing, nr == 0 and the loop will be skipped. |
| 135 | */ |
| 136 | for (i = 0; i < nr_found; i++) { |
| 137 | struct xfs_inode *ip = batch[i]; |
| 138 | |
| 139 | if (done || xfs_inode_ag_walk_grab(ip)) |
| 140 | batch[i] = NULL; |
| 141 | |
| 142 | /* |
| 143 | * Update the index for the next lookup. Catch |
| 144 | * overflows into the next AG range which can occur if |
| 145 | * we have inodes in the last block of the AG and we |
| 146 | * are currently pointing to the last inode. |
| 147 | * |
| 148 | * Because we may see inodes that are from the wrong AG |
| 149 | * due to RCU freeing and reallocation, only update the |
| 150 | * index if it lies in this AG. It was a race that lead |
| 151 | * us to see this inode, so another lookup from the |
| 152 | * same index will not find it again. |
| 153 | */ |
| 154 | if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno) |
| 155 | continue; |
| 156 | first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1); |
| 157 | if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino)) |
| 158 | done = 1; |
| 159 | } |
| 160 | |
| 161 | /* unlock now we've grabbed the inodes. */ |
| 162 | rcu_read_unlock(); |
| 163 | |
| 164 | for (i = 0; i < nr_found; i++) { |
| 165 | if (!batch[i]) |
| 166 | continue; |
| 167 | error = execute(batch[i], pag, flags); |
| 168 | IRELE(batch[i]); |
| 169 | if (error == EAGAIN) { |
| 170 | skipped++; |
| 171 | continue; |
| 172 | } |
| 173 | if (error && last_error != EFSCORRUPTED) |
| 174 | last_error = error; |
| 175 | } |
| 176 | |
| 177 | /* bail out if the filesystem is corrupted. */ |
| 178 | if (error == EFSCORRUPTED) |
| 179 | break; |
| 180 | |
| 181 | cond_resched(); |
| 182 | |
| 183 | } while (nr_found && !done); |
| 184 | |
| 185 | if (skipped) { |
| 186 | delay(1); |
| 187 | goto restart; |
| 188 | } |
| 189 | return last_error; |
| 190 | } |
| 191 | |
| 192 | int |
| 193 | xfs_inode_ag_iterator( |
| 194 | struct xfs_mount *mp, |
| 195 | int (*execute)(struct xfs_inode *ip, |
| 196 | struct xfs_perag *pag, int flags), |
| 197 | int flags) |
| 198 | { |
| 199 | struct xfs_perag *pag; |
| 200 | int error = 0; |
| 201 | int last_error = 0; |
| 202 | xfs_agnumber_t ag; |
| 203 | |
| 204 | ag = 0; |
| 205 | while ((pag = xfs_perag_get(mp, ag))) { |
| 206 | ag = pag->pag_agno + 1; |
| 207 | error = xfs_inode_ag_walk(mp, pag, execute, flags); |
| 208 | xfs_perag_put(pag); |
| 209 | if (error) { |
| 210 | last_error = error; |
| 211 | if (error == EFSCORRUPTED) |
| 212 | break; |
| 213 | } |
| 214 | } |
| 215 | return XFS_ERROR(last_error); |
| 216 | } |
| 217 | |
| 218 | STATIC int |
| 219 | xfs_sync_inode_data( |
| 220 | struct xfs_inode *ip, |
| 221 | struct xfs_perag *pag, |
| 222 | int flags) |
| 223 | { |
| 224 | struct inode *inode = VFS_I(ip); |
| 225 | struct address_space *mapping = inode->i_mapping; |
| 226 | int error = 0; |
| 227 | |
| 228 | if (!mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) |
| 229 | return 0; |
| 230 | |
| 231 | if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED)) { |
| 232 | if (flags & SYNC_TRYLOCK) |
| 233 | return 0; |
| 234 | xfs_ilock(ip, XFS_IOLOCK_SHARED); |
| 235 | } |
| 236 | |
| 237 | error = xfs_flush_pages(ip, 0, -1, (flags & SYNC_WAIT) ? |
| 238 | 0 : XBF_ASYNC, FI_NONE); |
| 239 | xfs_iunlock(ip, XFS_IOLOCK_SHARED); |
| 240 | return error; |
| 241 | } |
| 242 | |
| 243 | /* |
| 244 | * Write out pagecache data for the whole filesystem. |
| 245 | */ |
| 246 | STATIC int |
| 247 | xfs_sync_data( |
| 248 | struct xfs_mount *mp, |
| 249 | int flags) |
| 250 | { |
| 251 | int error; |
| 252 | |
| 253 | ASSERT((flags & ~(SYNC_TRYLOCK|SYNC_WAIT)) == 0); |
| 254 | |
| 255 | error = xfs_inode_ag_iterator(mp, xfs_sync_inode_data, flags); |
| 256 | if (error) |
| 257 | return XFS_ERROR(error); |
| 258 | |
| 259 | xfs_log_force(mp, (flags & SYNC_WAIT) ? XFS_LOG_SYNC : 0); |
| 260 | return 0; |
| 261 | } |
| 262 | |
| 263 | STATIC int |
| 264 | xfs_sync_fsdata( |
| 265 | struct xfs_mount *mp) |
| 266 | { |
| 267 | struct xfs_buf *bp; |
| 268 | int error; |
| 269 | |
| 270 | /* |
| 271 | * If the buffer is pinned then push on the log so we won't get stuck |
| 272 | * waiting in the write for someone, maybe ourselves, to flush the log. |
| 273 | * |
| 274 | * Even though we just pushed the log above, we did not have the |
| 275 | * superblock buffer locked at that point so it can become pinned in |
| 276 | * between there and here. |
| 277 | */ |
| 278 | bp = xfs_getsb(mp, 0); |
| 279 | if (xfs_buf_ispinned(bp)) |
| 280 | xfs_log_force(mp, 0); |
| 281 | error = xfs_bwrite(bp); |
| 282 | xfs_buf_relse(bp); |
| 283 | return error; |
| 284 | } |
| 285 | |
| 286 | /* |
| 287 | * When remounting a filesystem read-only or freezing the filesystem, we have |
| 288 | * two phases to execute. This first phase is syncing the data before we |
| 289 | * quiesce the filesystem, and the second is flushing all the inodes out after |
| 290 | * we've waited for all the transactions created by the first phase to |
| 291 | * complete. The second phase ensures that the inodes are written to their |
| 292 | * location on disk rather than just existing in transactions in the log. This |
| 293 | * means after a quiesce there is no log replay required to write the inodes to |
| 294 | * disk (this is the main difference between a sync and a quiesce). |
| 295 | */ |
| 296 | /* |
| 297 | * First stage of freeze - no writers will make progress now we are here, |
| 298 | * so we flush delwri and delalloc buffers here, then wait for all I/O to |
| 299 | * complete. Data is frozen at that point. Metadata is not frozen, |
| 300 | * transactions can still occur here so don't bother emptying the AIL |
| 301 | * because it'll just get dirty again. |
| 302 | */ |
| 303 | int |
| 304 | xfs_quiesce_data( |
| 305 | struct xfs_mount *mp) |
| 306 | { |
| 307 | int error, error2 = 0; |
| 308 | |
| 309 | /* force out the log */ |
| 310 | xfs_log_force(mp, XFS_LOG_SYNC); |
| 311 | |
| 312 | /* write superblock and hoover up shutdown errors */ |
| 313 | error = xfs_sync_fsdata(mp); |
| 314 | |
| 315 | /* mark the log as covered if needed */ |
| 316 | if (xfs_log_need_covered(mp)) |
| 317 | error2 = xfs_fs_log_dummy(mp); |
| 318 | |
| 319 | return error ? error : error2; |
| 320 | } |
| 321 | |
| 322 | /* |
| 323 | * Second stage of a quiesce. The data is already synced, now we have to take |
| 324 | * care of the metadata. New transactions are already blocked, so we need to |
| 325 | * wait for any remaining transactions to drain out before proceeding. |
| 326 | */ |
| 327 | void |
| 328 | xfs_quiesce_attr( |
| 329 | struct xfs_mount *mp) |
| 330 | { |
| 331 | int error = 0; |
| 332 | |
| 333 | /* wait for all modifications to complete */ |
| 334 | while (atomic_read(&mp->m_active_trans) > 0) |
| 335 | delay(100); |
| 336 | |
| 337 | /* reclaim inodes to do any IO before the freeze completes */ |
| 338 | xfs_reclaim_inodes(mp, 0); |
| 339 | xfs_reclaim_inodes(mp, SYNC_WAIT); |
| 340 | |
| 341 | /* flush all pending changes from the AIL */ |
| 342 | xfs_ail_push_all_sync(mp->m_ail); |
| 343 | |
| 344 | /* |
| 345 | * Just warn here till VFS can correctly support |
| 346 | * read-only remount without racing. |
| 347 | */ |
| 348 | WARN_ON(atomic_read(&mp->m_active_trans) != 0); |
| 349 | |
| 350 | /* Push the superblock and write an unmount record */ |
| 351 | error = xfs_log_sbcount(mp); |
| 352 | if (error) |
| 353 | xfs_warn(mp, "xfs_attr_quiesce: failed to log sb changes. " |
| 354 | "Frozen image may not be consistent."); |
| 355 | xfs_log_unmount_write(mp); |
| 356 | |
| 357 | /* |
| 358 | * At this point we might have modified the superblock again and thus |
| 359 | * added an item to the AIL, thus flush it again. |
| 360 | */ |
| 361 | xfs_ail_push_all_sync(mp->m_ail); |
| 362 | } |
| 363 | |
| 364 | static void |
| 365 | xfs_syncd_queue_sync( |
| 366 | struct xfs_mount *mp) |
| 367 | { |
| 368 | queue_delayed_work(xfs_syncd_wq, &mp->m_sync_work, |
| 369 | msecs_to_jiffies(xfs_syncd_centisecs * 10)); |
| 370 | } |
| 371 | |
| 372 | /* |
| 373 | * Every sync period we need to unpin all items, reclaim inodes and sync |
| 374 | * disk quotas. We might need to cover the log to indicate that the |
| 375 | * filesystem is idle and not frozen. |
| 376 | */ |
| 377 | STATIC void |
| 378 | xfs_sync_worker( |
| 379 | struct work_struct *work) |
| 380 | { |
| 381 | struct xfs_mount *mp = container_of(to_delayed_work(work), |
| 382 | struct xfs_mount, m_sync_work); |
| 383 | int error; |
| 384 | |
| 385 | /* |
| 386 | * We shouldn't write/force the log if we are in the mount/unmount |
| 387 | * process or on a read only filesystem. The workqueue still needs to be |
| 388 | * active in both cases, however, because it is used for inode reclaim |
| 389 | * during these times. Use the MS_ACTIVE flag to avoid doing anything |
| 390 | * during mount. Doing work during unmount is avoided by calling |
| 391 | * cancel_delayed_work_sync on this work queue before tearing down |
| 392 | * the ail and the log in xfs_log_unmount. |
| 393 | */ |
| 394 | if (!(mp->m_super->s_flags & MS_ACTIVE) && |
| 395 | !(mp->m_flags & XFS_MOUNT_RDONLY)) { |
| 396 | /* dgc: errors ignored here */ |
| 397 | if (mp->m_super->s_frozen == SB_UNFROZEN && |
| 398 | xfs_log_need_covered(mp)) |
| 399 | error = xfs_fs_log_dummy(mp); |
| 400 | else |
| 401 | xfs_log_force(mp, 0); |
| 402 | |
| 403 | /* start pushing all the metadata that is currently |
| 404 | * dirty */ |
| 405 | xfs_ail_push_all(mp->m_ail); |
| 406 | } |
| 407 | |
| 408 | /* queue us up again */ |
| 409 | xfs_syncd_queue_sync(mp); |
| 410 | } |
| 411 | |
| 412 | /* |
| 413 | * Queue a new inode reclaim pass if there are reclaimable inodes and there |
| 414 | * isn't a reclaim pass already in progress. By default it runs every 5s based |
| 415 | * on the xfs syncd work default of 30s. Perhaps this should have it's own |
| 416 | * tunable, but that can be done if this method proves to be ineffective or too |
| 417 | * aggressive. |
| 418 | */ |
| 419 | static void |
| 420 | xfs_syncd_queue_reclaim( |
| 421 | struct xfs_mount *mp) |
| 422 | { |
| 423 | |
| 424 | rcu_read_lock(); |
| 425 | if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) { |
| 426 | queue_delayed_work(xfs_syncd_wq, &mp->m_reclaim_work, |
| 427 | msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10)); |
| 428 | } |
| 429 | rcu_read_unlock(); |
| 430 | } |
| 431 | |
| 432 | /* |
| 433 | * This is a fast pass over the inode cache to try to get reclaim moving on as |
| 434 | * many inodes as possible in a short period of time. It kicks itself every few |
| 435 | * seconds, as well as being kicked by the inode cache shrinker when memory |
| 436 | * goes low. It scans as quickly as possible avoiding locked inodes or those |
| 437 | * already being flushed, and once done schedules a future pass. |
| 438 | */ |
| 439 | STATIC void |
| 440 | xfs_reclaim_worker( |
| 441 | struct work_struct *work) |
| 442 | { |
| 443 | struct xfs_mount *mp = container_of(to_delayed_work(work), |
| 444 | struct xfs_mount, m_reclaim_work); |
| 445 | |
| 446 | xfs_reclaim_inodes(mp, SYNC_TRYLOCK); |
| 447 | xfs_syncd_queue_reclaim(mp); |
| 448 | } |
| 449 | |
| 450 | /* |
| 451 | * Flush delayed allocate data, attempting to free up reserved space |
| 452 | * from existing allocations. At this point a new allocation attempt |
| 453 | * has failed with ENOSPC and we are in the process of scratching our |
| 454 | * heads, looking about for more room. |
| 455 | * |
| 456 | * Queue a new data flush if there isn't one already in progress and |
| 457 | * wait for completion of the flush. This means that we only ever have one |
| 458 | * inode flush in progress no matter how many ENOSPC events are occurring and |
| 459 | * so will prevent the system from bogging down due to every concurrent |
| 460 | * ENOSPC event scanning all the active inodes in the system for writeback. |
| 461 | */ |
| 462 | void |
| 463 | xfs_flush_inodes( |
| 464 | struct xfs_inode *ip) |
| 465 | { |
| 466 | struct xfs_mount *mp = ip->i_mount; |
| 467 | |
| 468 | queue_work(xfs_syncd_wq, &mp->m_flush_work); |
| 469 | flush_work_sync(&mp->m_flush_work); |
| 470 | } |
| 471 | |
| 472 | STATIC void |
| 473 | xfs_flush_worker( |
| 474 | struct work_struct *work) |
| 475 | { |
| 476 | struct xfs_mount *mp = container_of(work, |
| 477 | struct xfs_mount, m_flush_work); |
| 478 | |
| 479 | xfs_sync_data(mp, SYNC_TRYLOCK); |
| 480 | xfs_sync_data(mp, SYNC_TRYLOCK | SYNC_WAIT); |
| 481 | } |
| 482 | |
| 483 | int |
| 484 | xfs_syncd_init( |
| 485 | struct xfs_mount *mp) |
| 486 | { |
| 487 | INIT_WORK(&mp->m_flush_work, xfs_flush_worker); |
| 488 | INIT_DELAYED_WORK(&mp->m_sync_work, xfs_sync_worker); |
| 489 | INIT_DELAYED_WORK(&mp->m_reclaim_work, xfs_reclaim_worker); |
| 490 | |
| 491 | xfs_syncd_queue_sync(mp); |
| 492 | |
| 493 | return 0; |
| 494 | } |
| 495 | |
| 496 | void |
| 497 | xfs_syncd_stop( |
| 498 | struct xfs_mount *mp) |
| 499 | { |
| 500 | cancel_delayed_work_sync(&mp->m_sync_work); |
| 501 | cancel_delayed_work_sync(&mp->m_reclaim_work); |
| 502 | cancel_work_sync(&mp->m_flush_work); |
| 503 | } |
| 504 | |
| 505 | void |
| 506 | __xfs_inode_set_reclaim_tag( |
| 507 | struct xfs_perag *pag, |
| 508 | struct xfs_inode *ip) |
| 509 | { |
| 510 | radix_tree_tag_set(&pag->pag_ici_root, |
| 511 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), |
| 512 | XFS_ICI_RECLAIM_TAG); |
| 513 | |
| 514 | if (!pag->pag_ici_reclaimable) { |
| 515 | /* propagate the reclaim tag up into the perag radix tree */ |
| 516 | spin_lock(&ip->i_mount->m_perag_lock); |
| 517 | radix_tree_tag_set(&ip->i_mount->m_perag_tree, |
| 518 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), |
| 519 | XFS_ICI_RECLAIM_TAG); |
| 520 | spin_unlock(&ip->i_mount->m_perag_lock); |
| 521 | |
| 522 | /* schedule periodic background inode reclaim */ |
| 523 | xfs_syncd_queue_reclaim(ip->i_mount); |
| 524 | |
| 525 | trace_xfs_perag_set_reclaim(ip->i_mount, pag->pag_agno, |
| 526 | -1, _RET_IP_); |
| 527 | } |
| 528 | pag->pag_ici_reclaimable++; |
| 529 | } |
| 530 | |
| 531 | /* |
| 532 | * We set the inode flag atomically with the radix tree tag. |
| 533 | * Once we get tag lookups on the radix tree, this inode flag |
| 534 | * can go away. |
| 535 | */ |
| 536 | void |
| 537 | xfs_inode_set_reclaim_tag( |
| 538 | xfs_inode_t *ip) |
| 539 | { |
| 540 | struct xfs_mount *mp = ip->i_mount; |
| 541 | struct xfs_perag *pag; |
| 542 | |
| 543 | pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino)); |
| 544 | spin_lock(&pag->pag_ici_lock); |
| 545 | spin_lock(&ip->i_flags_lock); |
| 546 | __xfs_inode_set_reclaim_tag(pag, ip); |
| 547 | __xfs_iflags_set(ip, XFS_IRECLAIMABLE); |
| 548 | spin_unlock(&ip->i_flags_lock); |
| 549 | spin_unlock(&pag->pag_ici_lock); |
| 550 | xfs_perag_put(pag); |
| 551 | } |
| 552 | |
| 553 | STATIC void |
| 554 | __xfs_inode_clear_reclaim( |
| 555 | xfs_perag_t *pag, |
| 556 | xfs_inode_t *ip) |
| 557 | { |
| 558 | pag->pag_ici_reclaimable--; |
| 559 | if (!pag->pag_ici_reclaimable) { |
| 560 | /* clear the reclaim tag from the perag radix tree */ |
| 561 | spin_lock(&ip->i_mount->m_perag_lock); |
| 562 | radix_tree_tag_clear(&ip->i_mount->m_perag_tree, |
| 563 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), |
| 564 | XFS_ICI_RECLAIM_TAG); |
| 565 | spin_unlock(&ip->i_mount->m_perag_lock); |
| 566 | trace_xfs_perag_clear_reclaim(ip->i_mount, pag->pag_agno, |
| 567 | -1, _RET_IP_); |
| 568 | } |
| 569 | } |
| 570 | |
| 571 | void |
| 572 | __xfs_inode_clear_reclaim_tag( |
| 573 | xfs_mount_t *mp, |
| 574 | xfs_perag_t *pag, |
| 575 | xfs_inode_t *ip) |
| 576 | { |
| 577 | radix_tree_tag_clear(&pag->pag_ici_root, |
| 578 | XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG); |
| 579 | __xfs_inode_clear_reclaim(pag, ip); |
| 580 | } |
| 581 | |
| 582 | /* |
| 583 | * Grab the inode for reclaim exclusively. |
| 584 | * Return 0 if we grabbed it, non-zero otherwise. |
| 585 | */ |
| 586 | STATIC int |
| 587 | xfs_reclaim_inode_grab( |
| 588 | struct xfs_inode *ip, |
| 589 | int flags) |
| 590 | { |
| 591 | ASSERT(rcu_read_lock_held()); |
| 592 | |
| 593 | /* quick check for stale RCU freed inode */ |
| 594 | if (!ip->i_ino) |
| 595 | return 1; |
| 596 | |
| 597 | /* |
| 598 | * If we are asked for non-blocking operation, do unlocked checks to |
| 599 | * see if the inode already is being flushed or in reclaim to avoid |
| 600 | * lock traffic. |
| 601 | */ |
| 602 | if ((flags & SYNC_TRYLOCK) && |
| 603 | __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM)) |
| 604 | return 1; |
| 605 | |
| 606 | /* |
| 607 | * The radix tree lock here protects a thread in xfs_iget from racing |
| 608 | * with us starting reclaim on the inode. Once we have the |
| 609 | * XFS_IRECLAIM flag set it will not touch us. |
| 610 | * |
| 611 | * Due to RCU lookup, we may find inodes that have been freed and only |
| 612 | * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that |
| 613 | * aren't candidates for reclaim at all, so we must check the |
| 614 | * XFS_IRECLAIMABLE is set first before proceeding to reclaim. |
| 615 | */ |
| 616 | spin_lock(&ip->i_flags_lock); |
| 617 | if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) || |
| 618 | __xfs_iflags_test(ip, XFS_IRECLAIM)) { |
| 619 | /* not a reclaim candidate. */ |
| 620 | spin_unlock(&ip->i_flags_lock); |
| 621 | return 1; |
| 622 | } |
| 623 | __xfs_iflags_set(ip, XFS_IRECLAIM); |
| 624 | spin_unlock(&ip->i_flags_lock); |
| 625 | return 0; |
| 626 | } |
| 627 | |
| 628 | /* |
| 629 | * Inodes in different states need to be treated differently. The following |
| 630 | * table lists the inode states and the reclaim actions necessary: |
| 631 | * |
| 632 | * inode state iflush ret required action |
| 633 | * --------------- ---------- --------------- |
| 634 | * bad - reclaim |
| 635 | * shutdown EIO unpin and reclaim |
| 636 | * clean, unpinned 0 reclaim |
| 637 | * stale, unpinned 0 reclaim |
| 638 | * clean, pinned(*) 0 requeue |
| 639 | * stale, pinned EAGAIN requeue |
| 640 | * dirty, async - requeue |
| 641 | * dirty, sync 0 reclaim |
| 642 | * |
| 643 | * (*) dgc: I don't think the clean, pinned state is possible but it gets |
| 644 | * handled anyway given the order of checks implemented. |
| 645 | * |
| 646 | * Also, because we get the flush lock first, we know that any inode that has |
| 647 | * been flushed delwri has had the flush completed by the time we check that |
| 648 | * the inode is clean. |
| 649 | * |
| 650 | * Note that because the inode is flushed delayed write by AIL pushing, the |
| 651 | * flush lock may already be held here and waiting on it can result in very |
| 652 | * long latencies. Hence for sync reclaims, where we wait on the flush lock, |
| 653 | * the caller should push the AIL first before trying to reclaim inodes to |
| 654 | * minimise the amount of time spent waiting. For background relaim, we only |
| 655 | * bother to reclaim clean inodes anyway. |
| 656 | * |
| 657 | * Hence the order of actions after gaining the locks should be: |
| 658 | * bad => reclaim |
| 659 | * shutdown => unpin and reclaim |
| 660 | * pinned, async => requeue |
| 661 | * pinned, sync => unpin |
| 662 | * stale => reclaim |
| 663 | * clean => reclaim |
| 664 | * dirty, async => requeue |
| 665 | * dirty, sync => flush, wait and reclaim |
| 666 | */ |
| 667 | STATIC int |
| 668 | xfs_reclaim_inode( |
| 669 | struct xfs_inode *ip, |
| 670 | struct xfs_perag *pag, |
| 671 | int sync_mode) |
| 672 | { |
| 673 | struct xfs_buf *bp = NULL; |
| 674 | int error; |
| 675 | |
| 676 | restart: |
| 677 | error = 0; |
| 678 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
| 679 | if (!xfs_iflock_nowait(ip)) { |
| 680 | if (!(sync_mode & SYNC_WAIT)) |
| 681 | goto out; |
| 682 | xfs_iflock(ip); |
| 683 | } |
| 684 | |
| 685 | if (is_bad_inode(VFS_I(ip))) |
| 686 | goto reclaim; |
| 687 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) { |
| 688 | xfs_iunpin_wait(ip); |
| 689 | xfs_iflush_abort(ip, false); |
| 690 | goto reclaim; |
| 691 | } |
| 692 | if (xfs_ipincount(ip)) { |
| 693 | if (!(sync_mode & SYNC_WAIT)) |
| 694 | goto out_ifunlock; |
| 695 | xfs_iunpin_wait(ip); |
| 696 | } |
| 697 | if (xfs_iflags_test(ip, XFS_ISTALE)) |
| 698 | goto reclaim; |
| 699 | if (xfs_inode_clean(ip)) |
| 700 | goto reclaim; |
| 701 | |
| 702 | /* |
| 703 | * Never flush out dirty data during non-blocking reclaim, as it would |
| 704 | * just contend with AIL pushing trying to do the same job. |
| 705 | */ |
| 706 | if (!(sync_mode & SYNC_WAIT)) |
| 707 | goto out_ifunlock; |
| 708 | |
| 709 | /* |
| 710 | * Now we have an inode that needs flushing. |
| 711 | * |
| 712 | * Note that xfs_iflush will never block on the inode buffer lock, as |
| 713 | * xfs_ifree_cluster() can lock the inode buffer before it locks the |
| 714 | * ip->i_lock, and we are doing the exact opposite here. As a result, |
| 715 | * doing a blocking xfs_itobp() to get the cluster buffer would result |
| 716 | * in an ABBA deadlock with xfs_ifree_cluster(). |
| 717 | * |
| 718 | * As xfs_ifree_cluser() must gather all inodes that are active in the |
| 719 | * cache to mark them stale, if we hit this case we don't actually want |
| 720 | * to do IO here - we want the inode marked stale so we can simply |
| 721 | * reclaim it. Hence if we get an EAGAIN error here, just unlock the |
| 722 | * inode, back off and try again. Hopefully the next pass through will |
| 723 | * see the stale flag set on the inode. |
| 724 | */ |
| 725 | error = xfs_iflush(ip, &bp); |
| 726 | if (error == EAGAIN) { |
| 727 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
| 728 | /* backoff longer than in xfs_ifree_cluster */ |
| 729 | delay(2); |
| 730 | goto restart; |
| 731 | } |
| 732 | |
| 733 | if (!error) { |
| 734 | error = xfs_bwrite(bp); |
| 735 | xfs_buf_relse(bp); |
| 736 | } |
| 737 | |
| 738 | xfs_iflock(ip); |
| 739 | reclaim: |
| 740 | xfs_ifunlock(ip); |
| 741 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
| 742 | |
| 743 | XFS_STATS_INC(xs_ig_reclaims); |
| 744 | /* |
| 745 | * Remove the inode from the per-AG radix tree. |
| 746 | * |
| 747 | * Because radix_tree_delete won't complain even if the item was never |
| 748 | * added to the tree assert that it's been there before to catch |
| 749 | * problems with the inode life time early on. |
| 750 | */ |
| 751 | spin_lock(&pag->pag_ici_lock); |
| 752 | if (!radix_tree_delete(&pag->pag_ici_root, |
| 753 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino))) |
| 754 | ASSERT(0); |
| 755 | __xfs_inode_clear_reclaim(pag, ip); |
| 756 | spin_unlock(&pag->pag_ici_lock); |
| 757 | |
| 758 | /* |
| 759 | * Here we do an (almost) spurious inode lock in order to coordinate |
| 760 | * with inode cache radix tree lookups. This is because the lookup |
| 761 | * can reference the inodes in the cache without taking references. |
| 762 | * |
| 763 | * We make that OK here by ensuring that we wait until the inode is |
| 764 | * unlocked after the lookup before we go ahead and free it. |
| 765 | */ |
| 766 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
| 767 | xfs_qm_dqdetach(ip); |
| 768 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
| 769 | |
| 770 | xfs_inode_free(ip); |
| 771 | return error; |
| 772 | |
| 773 | out_ifunlock: |
| 774 | xfs_ifunlock(ip); |
| 775 | out: |
| 776 | xfs_iflags_clear(ip, XFS_IRECLAIM); |
| 777 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
| 778 | /* |
| 779 | * We could return EAGAIN here to make reclaim rescan the inode tree in |
| 780 | * a short while. However, this just burns CPU time scanning the tree |
| 781 | * waiting for IO to complete and xfssyncd never goes back to the idle |
| 782 | * state. Instead, return 0 to let the next scheduled background reclaim |
| 783 | * attempt to reclaim the inode again. |
| 784 | */ |
| 785 | return 0; |
| 786 | } |
| 787 | |
| 788 | /* |
| 789 | * Walk the AGs and reclaim the inodes in them. Even if the filesystem is |
| 790 | * corrupted, we still want to try to reclaim all the inodes. If we don't, |
| 791 | * then a shut down during filesystem unmount reclaim walk leak all the |
| 792 | * unreclaimed inodes. |
| 793 | */ |
| 794 | int |
| 795 | xfs_reclaim_inodes_ag( |
| 796 | struct xfs_mount *mp, |
| 797 | int flags, |
| 798 | int *nr_to_scan) |
| 799 | { |
| 800 | struct xfs_perag *pag; |
| 801 | int error = 0; |
| 802 | int last_error = 0; |
| 803 | xfs_agnumber_t ag; |
| 804 | int trylock = flags & SYNC_TRYLOCK; |
| 805 | int skipped; |
| 806 | |
| 807 | restart: |
| 808 | ag = 0; |
| 809 | skipped = 0; |
| 810 | while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) { |
| 811 | unsigned long first_index = 0; |
| 812 | int done = 0; |
| 813 | int nr_found = 0; |
| 814 | |
| 815 | ag = pag->pag_agno + 1; |
| 816 | |
| 817 | if (trylock) { |
| 818 | if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) { |
| 819 | skipped++; |
| 820 | xfs_perag_put(pag); |
| 821 | continue; |
| 822 | } |
| 823 | first_index = pag->pag_ici_reclaim_cursor; |
| 824 | } else |
| 825 | mutex_lock(&pag->pag_ici_reclaim_lock); |
| 826 | |
| 827 | do { |
| 828 | struct xfs_inode *batch[XFS_LOOKUP_BATCH]; |
| 829 | int i; |
| 830 | |
| 831 | rcu_read_lock(); |
| 832 | nr_found = radix_tree_gang_lookup_tag( |
| 833 | &pag->pag_ici_root, |
| 834 | (void **)batch, first_index, |
| 835 | XFS_LOOKUP_BATCH, |
| 836 | XFS_ICI_RECLAIM_TAG); |
| 837 | if (!nr_found) { |
| 838 | done = 1; |
| 839 | rcu_read_unlock(); |
| 840 | break; |
| 841 | } |
| 842 | |
| 843 | /* |
| 844 | * Grab the inodes before we drop the lock. if we found |
| 845 | * nothing, nr == 0 and the loop will be skipped. |
| 846 | */ |
| 847 | for (i = 0; i < nr_found; i++) { |
| 848 | struct xfs_inode *ip = batch[i]; |
| 849 | |
| 850 | if (done || xfs_reclaim_inode_grab(ip, flags)) |
| 851 | batch[i] = NULL; |
| 852 | |
| 853 | /* |
| 854 | * Update the index for the next lookup. Catch |
| 855 | * overflows into the next AG range which can |
| 856 | * occur if we have inodes in the last block of |
| 857 | * the AG and we are currently pointing to the |
| 858 | * last inode. |
| 859 | * |
| 860 | * Because we may see inodes that are from the |
| 861 | * wrong AG due to RCU freeing and |
| 862 | * reallocation, only update the index if it |
| 863 | * lies in this AG. It was a race that lead us |
| 864 | * to see this inode, so another lookup from |
| 865 | * the same index will not find it again. |
| 866 | */ |
| 867 | if (XFS_INO_TO_AGNO(mp, ip->i_ino) != |
| 868 | pag->pag_agno) |
| 869 | continue; |
| 870 | first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1); |
| 871 | if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino)) |
| 872 | done = 1; |
| 873 | } |
| 874 | |
| 875 | /* unlock now we've grabbed the inodes. */ |
| 876 | rcu_read_unlock(); |
| 877 | |
| 878 | for (i = 0; i < nr_found; i++) { |
| 879 | if (!batch[i]) |
| 880 | continue; |
| 881 | error = xfs_reclaim_inode(batch[i], pag, flags); |
| 882 | if (error && last_error != EFSCORRUPTED) |
| 883 | last_error = error; |
| 884 | } |
| 885 | |
| 886 | *nr_to_scan -= XFS_LOOKUP_BATCH; |
| 887 | |
| 888 | cond_resched(); |
| 889 | |
| 890 | } while (nr_found && !done && *nr_to_scan > 0); |
| 891 | |
| 892 | if (trylock && !done) |
| 893 | pag->pag_ici_reclaim_cursor = first_index; |
| 894 | else |
| 895 | pag->pag_ici_reclaim_cursor = 0; |
| 896 | mutex_unlock(&pag->pag_ici_reclaim_lock); |
| 897 | xfs_perag_put(pag); |
| 898 | } |
| 899 | |
| 900 | /* |
| 901 | * if we skipped any AG, and we still have scan count remaining, do |
| 902 | * another pass this time using blocking reclaim semantics (i.e |
| 903 | * waiting on the reclaim locks and ignoring the reclaim cursors). This |
| 904 | * ensure that when we get more reclaimers than AGs we block rather |
| 905 | * than spin trying to execute reclaim. |
| 906 | */ |
| 907 | if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) { |
| 908 | trylock = 0; |
| 909 | goto restart; |
| 910 | } |
| 911 | return XFS_ERROR(last_error); |
| 912 | } |
| 913 | |
| 914 | int |
| 915 | xfs_reclaim_inodes( |
| 916 | xfs_mount_t *mp, |
| 917 | int mode) |
| 918 | { |
| 919 | int nr_to_scan = INT_MAX; |
| 920 | |
| 921 | return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan); |
| 922 | } |
| 923 | |
| 924 | /* |
| 925 | * Scan a certain number of inodes for reclaim. |
| 926 | * |
| 927 | * When called we make sure that there is a background (fast) inode reclaim in |
| 928 | * progress, while we will throttle the speed of reclaim via doing synchronous |
| 929 | * reclaim of inodes. That means if we come across dirty inodes, we wait for |
| 930 | * them to be cleaned, which we hope will not be very long due to the |
| 931 | * background walker having already kicked the IO off on those dirty inodes. |
| 932 | */ |
| 933 | void |
| 934 | xfs_reclaim_inodes_nr( |
| 935 | struct xfs_mount *mp, |
| 936 | int nr_to_scan) |
| 937 | { |
| 938 | /* kick background reclaimer and push the AIL */ |
| 939 | xfs_syncd_queue_reclaim(mp); |
| 940 | xfs_ail_push_all(mp->m_ail); |
| 941 | |
| 942 | xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan); |
| 943 | } |
| 944 | |
| 945 | /* |
| 946 | * Return the number of reclaimable inodes in the filesystem for |
| 947 | * the shrinker to determine how much to reclaim. |
| 948 | */ |
| 949 | int |
| 950 | xfs_reclaim_inodes_count( |
| 951 | struct xfs_mount *mp) |
| 952 | { |
| 953 | struct xfs_perag *pag; |
| 954 | xfs_agnumber_t ag = 0; |
| 955 | int reclaimable = 0; |
| 956 | |
| 957 | while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) { |
| 958 | ag = pag->pag_agno + 1; |
| 959 | reclaimable += pag->pag_ici_reclaimable; |
| 960 | xfs_perag_put(pag); |
| 961 | } |
| 962 | return reclaimable; |
| 963 | } |
| 964 | |