fs/xfs/xfs_vfsops.c

   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_bit.h"
  22 #include "xfs_log.h"
  23 #include "xfs_inum.h"
  24 #include "xfs_trans.h"
  25 #include "xfs_sb.h"
  26 #include "xfs_ag.h"
  27 #include "xfs_dir2.h"
  28 #include "xfs_dmapi.h"
  29 #include "xfs_mount.h"
  30 #include "xfs_da_btree.h"
  31 #include "xfs_bmap_btree.h"
  32 #include "xfs_ialloc_btree.h"
  33 #include "xfs_alloc_btree.h"
  34 #include "xfs_dir2_sf.h"
  35 #include "xfs_attr_sf.h"
  36 #include "xfs_dinode.h"
  37 #include "xfs_inode.h"
  38 #include "xfs_inode_item.h"
  39 #include "xfs_btree.h"
  40 #include "xfs_alloc.h"
  41 #include "xfs_ialloc.h"
  42 #include "xfs_quota.h"
  43 #include "xfs_error.h"
  44 #include "xfs_bmap.h"
  45 #include "xfs_rw.h"
  46 #include "xfs_buf_item.h"
  47 #include "xfs_log_priv.h"
  48 #include "xfs_dir2_trace.h"
  49 #include "xfs_extfree_item.h"
  50 #include "xfs_acl.h"
  51 #include "xfs_attr.h"
  52 #include "xfs_clnt.h"
  53 #include "xfs_mru_cache.h"
  54 #include "xfs_filestream.h"
  55 #include "xfs_fsops.h"
  56 #include "xfs_vnodeops.h"
  57 #include "xfs_vfsops.h"
  58 #include "xfs_utils.h"
  59
  60
  61 int __init
  62 xfs_init(void)
  63 {
  64 #ifdef XFS_DABUF_DEBUG
  65         extern spinlock_t        xfs_dabuf_global_lock;
  66         spin_lock_init(&xfs_dabuf_global_lock);
  67 #endif
  68
  69         /*
  70          * Initialize all of the zone allocators we use.
  71          */
  72         xfs_log_ticket_zone = kmem_zone_init(sizeof(xlog_ticket_t),
  73                                                 "xfs_log_ticket");
  74         xfs_bmap_free_item_zone = kmem_zone_init(sizeof(xfs_bmap_free_item_t),
  75                                                 "xfs_bmap_free_item");
  76         xfs_btree_cur_zone = kmem_zone_init(sizeof(xfs_btree_cur_t),
  77                                                 "xfs_btree_cur");
  78         xfs_da_state_zone = kmem_zone_init(sizeof(xfs_da_state_t),
  79                                                 "xfs_da_state");
  80         xfs_dabuf_zone = kmem_zone_init(sizeof(xfs_dabuf_t), "xfs_dabuf");
  81         xfs_ifork_zone = kmem_zone_init(sizeof(xfs_ifork_t), "xfs_ifork");
  82         xfs_trans_zone = kmem_zone_init(sizeof(xfs_trans_t), "xfs_trans");
  83         xfs_acl_zone_init(xfs_acl_zone, "xfs_acl");
  84         xfs_mru_cache_init();
  85         xfs_filestream_init();
  86
  87         /*
  88          * The size of the zone allocated buf log item is the maximum
  89          * size possible under XFS.  This wastes a little bit of memory,
  90          * but it is much faster.
  91          */
  92         xfs_buf_item_zone =
  93                 kmem_zone_init((sizeof(xfs_buf_log_item_t) +
  94                                 (((XFS_MAX_BLOCKSIZE / XFS_BLI_CHUNK) /
  95                                   NBWORD) * sizeof(int))),
  96                                "xfs_buf_item");
  97         xfs_efd_zone =
  98                 kmem_zone_init((sizeof(xfs_efd_log_item_t) +
  99                                ((XFS_EFD_MAX_FAST_EXTENTS - 1) *
 100                                  sizeof(xfs_extent_t))),
 101                                       "xfs_efd_item");
 102         xfs_efi_zone =
 103                 kmem_zone_init((sizeof(xfs_efi_log_item_t) +
 104                                ((XFS_EFI_MAX_FAST_EXTENTS - 1) *
 105                                  sizeof(xfs_extent_t))),
 106                                       "xfs_efi_item");
 107
 108         /*
 109          * These zones warrant special memory allocator hints
 110          */
 111         xfs_inode_zone =
 112                 kmem_zone_init_flags(sizeof(xfs_inode_t), "xfs_inode",
 113                                         KM_ZONE_HWALIGN | KM_ZONE_RECLAIM |
 114                                         KM_ZONE_SPREAD, NULL);
 115         xfs_ili_zone =
 116                 kmem_zone_init_flags(sizeof(xfs_inode_log_item_t), "xfs_ili",
 117                                         KM_ZONE_SPREAD, NULL);
 118
 119         /*
 120          * Allocate global trace buffers.
 121          */
 122 #ifdef XFS_ALLOC_TRACE
 123         xfs_alloc_trace_buf = ktrace_alloc(XFS_ALLOC_TRACE_SIZE, KM_SLEEP);
 124 #endif
 125 #ifdef XFS_BMAP_TRACE
 126         xfs_bmap_trace_buf = ktrace_alloc(XFS_BMAP_TRACE_SIZE, KM_SLEEP);
 127 #endif
 128 #ifdef XFS_BMBT_TRACE
 129         xfs_bmbt_trace_buf = ktrace_alloc(XFS_BMBT_TRACE_SIZE, KM_SLEEP);
 130 #endif
 131 #ifdef XFS_ATTR_TRACE
 132         xfs_attr_trace_buf = ktrace_alloc(XFS_ATTR_TRACE_SIZE, KM_SLEEP);
 133 #endif
 134 #ifdef XFS_DIR2_TRACE
 135         xfs_dir2_trace_buf = ktrace_alloc(XFS_DIR2_GTRACE_SIZE, KM_SLEEP);
 136 #endif
 137
 138         xfs_dir_startup();
 139
 140 #if (defined(DEBUG) || defined(INDUCE_IO_ERROR))
 141         xfs_error_test_init();
 142 #endif /* DEBUG || INDUCE_IO_ERROR */
 143
 144         xfs_init_procfs();
 145         xfs_sysctl_register();
 146         return 0;
 147 }
 148
 149 void __exit
 150 xfs_cleanup(void)
 151 {
 152         extern kmem_zone_t      *xfs_inode_zone;
 153         extern kmem_zone_t      *xfs_efd_zone;
 154         extern kmem_zone_t      *xfs_efi_zone;
 155
 156         xfs_cleanup_procfs();
 157         xfs_sysctl_unregister();
 158         xfs_filestream_uninit();
 159         xfs_mru_cache_uninit();
 160         xfs_acl_zone_destroy(xfs_acl_zone);
 161
 162 #ifdef XFS_DIR2_TRACE
 163         ktrace_free(xfs_dir2_trace_buf);
 164 #endif
 165 #ifdef XFS_ATTR_TRACE
 166         ktrace_free(xfs_attr_trace_buf);
 167 #endif
 168 #ifdef XFS_BMBT_TRACE
 169         ktrace_free(xfs_bmbt_trace_buf);
 170 #endif
 171 #ifdef XFS_BMAP_TRACE
 172         ktrace_free(xfs_bmap_trace_buf);
 173 #endif
 174 #ifdef XFS_ALLOC_TRACE
 175         ktrace_free(xfs_alloc_trace_buf);
 176 #endif
 177
 178         kmem_zone_destroy(xfs_bmap_free_item_zone);
 179         kmem_zone_destroy(xfs_btree_cur_zone);
 180         kmem_zone_destroy(xfs_inode_zone);
 181         kmem_zone_destroy(xfs_trans_zone);
 182         kmem_zone_destroy(xfs_da_state_zone);
 183         kmem_zone_destroy(xfs_dabuf_zone);
 184         kmem_zone_destroy(xfs_buf_item_zone);
 185         kmem_zone_destroy(xfs_efd_zone);
 186         kmem_zone_destroy(xfs_efi_zone);
 187         kmem_zone_destroy(xfs_ifork_zone);
 188         kmem_zone_destroy(xfs_ili_zone);
 189         kmem_zone_destroy(xfs_log_ticket_zone);
 190 }
 191
 192 STATIC void
 193 xfs_quiesce_fs(
 194         xfs_mount_t             *mp)
 195 {
 196         int                     count = 0, pincount;
 197
 198         xfs_flush_buftarg(mp->m_ddev_targp, 0);
 199         xfs_finish_reclaim_all(mp, 0);
 200
 201         /* This loop must run at least twice.
 202          * The first instance of the loop will flush
 203          * most meta data but that will generate more
 204          * meta data (typically directory updates).
 205          * Which then must be flushed and logged before
 206          * we can write the unmount record.
 207          */
 208         do {
 209                 xfs_syncsub(mp, SYNC_INODE_QUIESCE, NULL);
 210                 pincount = xfs_flush_buftarg(mp->m_ddev_targp, 1);
 211                 if (!pincount) {
 212                         delay(50);
 213                         count++;
 214                 }
 215         } while (count < 2);
 216 }
 217
 218 /*
 219  * Second stage of a quiesce. The data is already synced, now we have to take
 220  * care of the metadata. New transactions are already blocked, so we need to
 221  * wait for any remaining transactions to drain out before proceding.
 222  */
 223 void
 224 xfs_attr_quiesce(
 225         xfs_mount_t     *mp)
 226 {
 227         int     error = 0;
 228
 229         /* wait for all modifications to complete */
 230         while (atomic_read(&mp->m_active_trans) > 0)
 231                 delay(100);
 232
 233         /* flush inodes and push all remaining buffers out to disk */
 234         xfs_quiesce_fs(mp);
 235
 236         ASSERT_ALWAYS(atomic_read(&mp->m_active_trans) == 0);
 237
 238         /* Push the superblock and write an unmount record */
 239         error = xfs_log_sbcount(mp, 1);
 240         if (error)
 241                 xfs_fs_cmn_err(CE_WARN, mp,
 242                                 "xfs_attr_quiesce: failed to log sb changes. "
 243                                 "Frozen image may not be consistent.");
 244         xfs_log_unmount_write(mp);
 245         xfs_unmountfs_writesb(mp);
 246 }
 247
 248 /*
 249  * xfs_unmount_flush implements a set of flush operation on special
 250  * inodes, which are needed as a separate set of operations so that
 251  * they can be called as part of relocation process.
 252  */
 253 int
 254 xfs_unmount_flush(
 255         xfs_mount_t     *mp,            /* Mount structure we are getting
 256                                            rid of. */
 257         int             relocation)     /* Called from vfs relocation. */
 258 {
 259         xfs_inode_t     *rip = mp->m_rootip;
 260         xfs_inode_t     *rbmip;
 261         xfs_inode_t     *rsumip = NULL;
 262         bhv_vnode_t     *rvp = XFS_ITOV(rip);
 263         int             error;
 264
 265         xfs_ilock(rip, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
 266         xfs_iflock(rip);
 267
 268         /*
 269          * Flush out the real time inodes.
 270          */
 271         if ((rbmip = mp->m_rbmip) != NULL) {
 272                 xfs_ilock(rbmip, XFS_ILOCK_EXCL);
 273                 xfs_iflock(rbmip);
 274                 error = xfs_iflush(rbmip, XFS_IFLUSH_SYNC);
 275                 xfs_iunlock(rbmip, XFS_ILOCK_EXCL);
 276
 277                 if (error == EFSCORRUPTED)
 278                         goto fscorrupt_out;
 279
 280                 ASSERT(vn_count(XFS_ITOV(rbmip)) == 1);
 281
 282                 rsumip = mp->m_rsumip;
 283                 xfs_ilock(rsumip, XFS_ILOCK_EXCL);
 284                 xfs_iflock(rsumip);
 285                 error = xfs_iflush(rsumip, XFS_IFLUSH_SYNC);
 286                 xfs_iunlock(rsumip, XFS_ILOCK_EXCL);
 287
 288                 if (error == EFSCORRUPTED)
 289                         goto fscorrupt_out;
 290
 291                 ASSERT(vn_count(XFS_ITOV(rsumip)) == 1);
 292         }
 293
 294         /*
 295          * Synchronously flush root inode to disk
 296          */
 297         error = xfs_iflush(rip, XFS_IFLUSH_SYNC);
 298         if (error == EFSCORRUPTED)
 299                 goto fscorrupt_out2;
 300
 301         if (vn_count(rvp) != 1 && !relocation) {
 302                 xfs_iunlock(rip, XFS_ILOCK_EXCL);
 303                 return XFS_ERROR(EBUSY);
 304         }
 305
 306         /*
 307          * Release dquot that rootinode, rbmino and rsumino might be holding,
 308          * flush and purge the quota inodes.
 309          */
 310         error = XFS_QM_UNMOUNT(mp);
 311         if (error == EFSCORRUPTED)
 312                 goto fscorrupt_out2;
 313
 314         if (rbmip) {
 315                 IRELE(rbmip);
 316                 IRELE(rsumip);
 317         }
 318
 319         xfs_iunlock(rip, XFS_ILOCK_EXCL);
 320         return 0;
 321
 322 fscorrupt_out:
 323         xfs_ifunlock(rip);
 324
 325 fscorrupt_out2:
 326         xfs_iunlock(rip, XFS_ILOCK_EXCL);
 327
 328         return XFS_ERROR(EFSCORRUPTED);
 329 }
 330
 331 /*
 332  * xfs_sync flushes any pending I/O to file system vfsp.
 333  *
 334  * This routine is called by vfs_sync() to make sure that things make it
 335  * out to disk eventually, on sync() system calls to flush out everything,
 336  * and when the file system is unmounted.  For the vfs_sync() case, all
 337  * we really need to do is sync out the log to make all of our meta-data
 338  * updates permanent (except for timestamps).  For calls from pflushd(),
 339  * dirty pages are kept moving by calling pdflush() on the inodes
 340  * containing them.  We also flush the inodes that we can lock without
 341  * sleeping and the superblock if we can lock it without sleeping from
 342  * vfs_sync() so that items at the tail of the log are always moving out.
 343  *
 344  * Flags:
 345  *      SYNC_BDFLUSH - We're being called from vfs_sync() so we don't want
 346  *                     to sleep if we can help it.  All we really need
 347  *                     to do is ensure that the log is synced at least
 348  *                     periodically.  We also push the inodes and
 349  *                     superblock if we can lock them without sleeping
 350  *                      and they are not pinned.
 351  *      SYNC_ATTR    - We need to flush the inodes.  If SYNC_BDFLUSH is not
 352  *                     set, then we really want to lock each inode and flush
 353  *                     it.
 354  *      SYNC_WAIT    - All the flushes that take place in this call should
 355  *                     be synchronous.
 356  *      SYNC_DELWRI  - This tells us to push dirty pages associated with
 357  *                     inodes.  SYNC_WAIT and SYNC_BDFLUSH are used to
 358  *                     determine if they should be flushed sync, async, or
 359  *                     delwri.
 360  *      SYNC_CLOSE   - This flag is passed when the system is being
 361  *                     unmounted.  We should sync and invalidate everything.
 362  *      SYNC_FSDATA  - This indicates that the caller would like to make
 363  *                     sure the superblock is safe on disk.  We can ensure
 364  *                     this by simply making sure the log gets flushed
 365  *                     if SYNC_BDFLUSH is set, and by actually writing it
 366  *                     out otherwise.
 367  *      SYNC_IOWAIT  - The caller wants us to wait for all data I/O to complete
 368  *                     before we return (including direct I/O). Forms the drain
 369  *                     side of the write barrier needed to safely quiesce the
 370  *                     filesystem.
 371  *
 372  */
 373 int
 374 xfs_sync(
 375         xfs_mount_t     *mp,
 376         int             flags)
 377 {
 378         int             error;
 379
 380         /*
 381          * Get the Quota Manager to flush the dquots.
 382          *
 383          * If XFS quota support is not enabled or this filesystem
 384          * instance does not use quotas XFS_QM_DQSYNC will always
 385          * return zero.
 386          */
 387         error = XFS_QM_DQSYNC(mp, flags);
 388         if (error) {
 389                 /*
 390                  * If we got an IO error, we will be shutting down.
 391                  * So, there's nothing more for us to do here.
 392                  */
 393                 ASSERT(error != EIO || XFS_FORCED_SHUTDOWN(mp));
 394                 if (XFS_FORCED_SHUTDOWN(mp))
 395                         return XFS_ERROR(error);
 396         }
 397
 398         if (flags & SYNC_IOWAIT)
 399                 xfs_filestream_flush(mp);
 400
 401         return xfs_syncsub(mp, flags, NULL);
 402 }
 403
 404 /*
 405  * xfs sync routine for internal use
 406  *
 407  * This routine supports all of the flags defined for the generic vfs_sync
 408  * interface as explained above under xfs_sync.
 409  *
 410  */
 411 int
 412 xfs_sync_inodes(
 413         xfs_mount_t     *mp,
 414         int             flags,
 415         int             *bypassed)
 416 {
 417         xfs_inode_t     *ip = NULL;
 418         bhv_vnode_t     *vp = NULL;
 419         int             error;
 420         int             last_error;
 421         uint64_t        fflag;
 422         uint            lock_flags;
 423         uint            base_lock_flags;
 424         boolean_t       mount_locked;
 425         boolean_t       vnode_refed;
 426         int             preempt;
 427         xfs_iptr_t      *ipointer;
 428 #ifdef DEBUG
 429         boolean_t       ipointer_in = B_FALSE;
 430
 431 #define IPOINTER_SET    ipointer_in = B_TRUE
 432 #define IPOINTER_CLR    ipointer_in = B_FALSE
 433 #else
 434 #define IPOINTER_SET
 435 #define IPOINTER_CLR
 436 #endif
 437
 438
 439 /* Insert a marker record into the inode list after inode ip. The list
 440  * must be locked when this is called. After the call the list will no
 441  * longer be locked.
 442  */
 443 #define IPOINTER_INSERT(ip, mp) { \
 444                 ASSERT(ipointer_in == B_FALSE); \
 445                 ipointer->ip_mnext = ip->i_mnext; \
 446                 ipointer->ip_mprev = ip; \
 447                 ip->i_mnext = (xfs_inode_t *)ipointer; \
 448                 ipointer->ip_mnext->i_mprev = (xfs_inode_t *)ipointer; \
 449                 preempt = 0; \
 450                 XFS_MOUNT_IUNLOCK(mp); \
 451                 mount_locked = B_FALSE; \
 452                 IPOINTER_SET; \
 453         }
 454
 455 /* Remove the marker from the inode list. If the marker was the only item
 456  * in the list then there are no remaining inodes and we should zero out
 457  * the whole list. If we are the current head of the list then move the head
 458  * past us.
 459  */
 460 #define IPOINTER_REMOVE(ip, mp) { \
 461                 ASSERT(ipointer_in == B_TRUE); \
 462                 if (ipointer->ip_mnext != (xfs_inode_t *)ipointer) { \
 463                         ip = ipointer->ip_mnext; \
 464                         ip->i_mprev = ipointer->ip_mprev; \
 465                         ipointer->ip_mprev->i_mnext = ip; \
 466                         if (mp->m_inodes == (xfs_inode_t *)ipointer) { \
 467                                 mp->m_inodes = ip; \
 468                         } \
 469                 } else { \
 470                         ASSERT(mp->m_inodes == (xfs_inode_t *)ipointer); \
 471                         mp->m_inodes = NULL; \
 472                         ip = NULL; \
 473                 } \
 474                 IPOINTER_CLR; \
 475         }
 476
 477 #define XFS_PREEMPT_MASK        0x7f
 478
 479         ASSERT(!(flags & SYNC_BDFLUSH));
 480
 481         if (bypassed)
 482                 *bypassed = 0;
 483         if (mp->m_flags & XFS_MOUNT_RDONLY)
 484                 return 0;
 485         error = 0;
 486         last_error = 0;
 487         preempt = 0;
 488
 489         /* Allocate a reference marker */
 490         ipointer = (xfs_iptr_t *)kmem_zalloc(sizeof(xfs_iptr_t), KM_SLEEP);
 491
 492         fflag = XFS_B_ASYNC;            /* default is don't wait */
 493         if (flags & SYNC_DELWRI)
 494                 fflag = XFS_B_DELWRI;
 495         if (flags & SYNC_WAIT)
 496                 fflag = 0;              /* synchronous overrides all */
 497
 498         base_lock_flags = XFS_ILOCK_SHARED;
 499         if (flags & (SYNC_DELWRI | SYNC_CLOSE)) {
 500                 /*
 501                  * We need the I/O lock if we're going to call any of
 502                  * the flush/inval routines.
 503                  */
 504                 base_lock_flags |= XFS_IOLOCK_SHARED;
 505         }
 506
 507         XFS_MOUNT_ILOCK(mp);
 508
 509         ip = mp->m_inodes;
 510
 511         mount_locked = B_TRUE;
 512         vnode_refed  = B_FALSE;
 513
 514         IPOINTER_CLR;
 515
 516         do {
 517                 ASSERT(ipointer_in == B_FALSE);
 518                 ASSERT(vnode_refed == B_FALSE);
 519
 520                 lock_flags = base_lock_flags;
 521
 522                 /*
 523                  * There were no inodes in the list, just break out
 524                  * of the loop.
 525                  */
 526                 if (ip == NULL) {
 527                         break;
 528                 }
 529
 530                 /*
 531                  * We found another sync thread marker - skip it
 532                  */
 533                 if (ip->i_mount == NULL) {
 534                         ip = ip->i_mnext;
 535                         continue;
 536                 }
 537
 538                 vp = XFS_ITOV_NULL(ip);
 539
 540                 /*
 541                  * If the vnode is gone then this is being torn down,
 542                  * call reclaim if it is flushed, else let regular flush
 543                  * code deal with it later in the loop.
 544                  */
 545
 546                 if (vp == NULL) {
 547                         /* Skip ones already in reclaim */
 548                         if (ip->i_flags & XFS_IRECLAIM) {
 549                                 ip = ip->i_mnext;
 550                                 continue;
 551                         }
 552                         if (xfs_ilock_nowait(ip, XFS_ILOCK_EXCL) == 0) {
 553                                 ip = ip->i_mnext;
 554                         } else if ((xfs_ipincount(ip) == 0) &&
 555                                     xfs_iflock_nowait(ip)) {
 556                                 IPOINTER_INSERT(ip, mp);
 557
 558                                 xfs_finish_reclaim(ip, 1,
 559                                                 XFS_IFLUSH_DELWRI_ELSE_ASYNC);
 560
 561                                 XFS_MOUNT_ILOCK(mp);
 562                                 mount_locked = B_TRUE;
 563                                 IPOINTER_REMOVE(ip, mp);
 564                         } else {
 565                                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 566                                 ip = ip->i_mnext;
 567                         }
 568                         continue;
 569                 }
 570
 571                 if (VN_BAD(vp)) {
 572                         ip = ip->i_mnext;
 573                         continue;
 574                 }
 575
 576                 if (XFS_FORCED_SHUTDOWN(mp) && !(flags & SYNC_CLOSE)) {
 577                         XFS_MOUNT_IUNLOCK(mp);
 578                         kmem_free(ipointer);
 579                         return 0;
 580                 }
 581
 582                 /*
 583                  * Try to lock without sleeping.  We're out of order with
 584                  * the inode list lock here, so if we fail we need to drop
 585                  * the mount lock and try again.  If we're called from
 586                  * bdflush() here, then don't bother.
 587                  *
 588                  * The inode lock here actually coordinates with the
 589                  * almost spurious inode lock in xfs_ireclaim() to prevent
 590                  * the vnode we handle here without a reference from
 591                  * being freed while we reference it.  If we lock the inode
 592                  * while it's on the mount list here, then the spurious inode
 593                  * lock in xfs_ireclaim() after the inode is pulled from
 594                  * the mount list will sleep until we release it here.
 595                  * This keeps the vnode from being freed while we reference
 596                  * it.
 597                  */
 598                 if (xfs_ilock_nowait(ip, lock_flags) == 0) {
 599                         if (vp == NULL) {
 600                                 ip = ip->i_mnext;
 601                                 continue;
 602                         }
 603
 604                         vp = vn_grab(vp);
 605                         if (vp == NULL) {
 606                                 ip = ip->i_mnext;
 607                                 continue;
 608                         }
 609
 610                         IPOINTER_INSERT(ip, mp);
 611                         xfs_ilock(ip, lock_flags);
 612
 613                         ASSERT(vp == XFS_ITOV(ip));
 614                         ASSERT(ip->i_mount == mp);
 615
 616                         vnode_refed = B_TRUE;
 617                 }
 618
 619                 /* From here on in the loop we may have a marker record
 620                  * in the inode list.
 621                  */
 622
 623                 /*
 624                  * If we have to flush data or wait for I/O completion
 625                  * we need to drop the ilock that we currently hold.
 626                  * If we need to drop the lock, insert a marker if we
 627                  * have not already done so.
 628                  */
 629                 if ((flags & (SYNC_CLOSE|SYNC_IOWAIT)) ||
 630                     ((flags & SYNC_DELWRI) && VN_DIRTY(vp))) {
 631                         if (mount_locked) {
 632                                 IPOINTER_INSERT(ip, mp);
 633                         }
 634                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 635
 636                         if (flags & SYNC_CLOSE) {
 637                                 /* Shutdown case. Flush and invalidate. */
 638                                 if (XFS_FORCED_SHUTDOWN(mp))
 639                                         xfs_tosspages(ip, 0, -1,
 640                                                              FI_REMAPF);
 641                                 else
 642                                         error = xfs_flushinval_pages(ip,
 643                                                         0, -1, FI_REMAPF);
 644                         } else if ((flags & SYNC_DELWRI) && VN_DIRTY(vp)) {
 645                                 error = xfs_flush_pages(ip, 0,
 646                                                         -1, fflag, FI_NONE);
 647                         }
 648
 649                         /*
 650                          * When freezing, we need to wait ensure all I/O (including direct
 651                          * I/O) is complete to ensure no further data modification can take
 652                          * place after this point
 653                          */
 654                         if (flags & SYNC_IOWAIT)
 655                                 vn_iowait(ip);
 656
 657                         xfs_ilock(ip, XFS_ILOCK_SHARED);
 658                 }
 659
 660                 if ((flags & SYNC_ATTR) &&
 661                     (ip->i_update_core ||
 662                      (ip->i_itemp && ip->i_itemp->ili_format.ilf_fields))) {
 663                         if (mount_locked)
 664                                 IPOINTER_INSERT(ip, mp);
 665
 666                         if (flags & SYNC_WAIT) {
 667                                 xfs_iflock(ip);
 668                                 error = xfs_iflush(ip, XFS_IFLUSH_SYNC);
 669
 670                         /*
 671                          * If we can't acquire the flush lock, then the inode
 672                          * is already being flushed so don't bother waiting.
 673                          *
 674                          * If we can lock it then do a delwri flush so we can
 675                          * combine multiple inode flushes in each disk write.
 676                          */
 677                         } else if (xfs_iflock_nowait(ip)) {
 678                                 error = xfs_iflush(ip, XFS_IFLUSH_DELWRI);
 679                         } else if (bypassed) {
 680                                 (*bypassed)++;
 681                         }
 682                 }
 683
 684                 if (lock_flags != 0) {
 685                         xfs_iunlock(ip, lock_flags);
 686                 }
 687
 688                 if (vnode_refed) {
 689                         /*
 690                          * If we had to take a reference on the vnode
 691                          * above, then wait until after we've unlocked
 692                          * the inode to release the reference.  This is
 693                          * because we can be already holding the inode
 694                          * lock when IRELE() calls xfs_inactive().
 695                          *
 696                          * Make sure to drop the mount lock before calling
 697                          * IRELE() so that we don't trip over ourselves if
 698                          * we have to go for the mount lock again in the
 699                          * inactive code.
 700                          */
 701                         if (mount_locked) {
 702                                 IPOINTER_INSERT(ip, mp);
 703                         }
 704
 705                         IRELE(ip);
 706
 707                         vnode_refed = B_FALSE;
 708                 }
 709
 710                 if (error) {
 711                         last_error = error;
 712                 }
 713
 714                 /*
 715                  * bail out if the filesystem is corrupted.
 716                  */
 717                 if (error == EFSCORRUPTED)  {
 718                         if (!mount_locked) {
 719                                 XFS_MOUNT_ILOCK(mp);
 720                                 IPOINTER_REMOVE(ip, mp);
 721                         }
 722                         XFS_MOUNT_IUNLOCK(mp);
 723                         ASSERT(ipointer_in == B_FALSE);
 724                         kmem_free(ipointer);
 725                         return XFS_ERROR(error);
 726                 }
 727
 728                 /* Let other threads have a chance at the mount lock
 729                  * if we have looped many times without dropping the
 730                  * lock.
 731                  */
 732                 if ((++preempt & XFS_PREEMPT_MASK) == 0) {
 733                         if (mount_locked) {
 734                                 IPOINTER_INSERT(ip, mp);
 735                         }
 736                 }
 737
 738                 if (mount_locked == B_FALSE) {
 739                         XFS_MOUNT_ILOCK(mp);
 740                         mount_locked = B_TRUE;
 741                         IPOINTER_REMOVE(ip, mp);
 742                         continue;
 743                 }
 744
 745                 ASSERT(ipointer_in == B_FALSE);
 746                 ip = ip->i_mnext;
 747
 748         } while (ip != mp->m_inodes);
 749
 750         XFS_MOUNT_IUNLOCK(mp);
 751
 752         ASSERT(ipointer_in == B_FALSE);
 753
 754         kmem_free(ipointer);
 755         return XFS_ERROR(last_error);
 756 }
 757
 758 /*
 759  * xfs sync routine for internal use
 760  *
 761  * This routine supports all of the flags defined for the generic vfs_sync
 762  * interface as explained above under xfs_sync.
 763  *
 764  */
 765 int
 766 xfs_syncsub(
 767         xfs_mount_t     *mp,
 768         int             flags,
 769         int             *bypassed)
 770 {
 771         int             error = 0;
 772         int             last_error = 0;
 773         uint            log_flags = XFS_LOG_FORCE;
 774         xfs_buf_t       *bp;
 775         xfs_buf_log_item_t      *bip;
 776
 777         /*
 778          * Sync out the log.  This ensures that the log is periodically
 779          * flushed even if there is not enough activity to fill it up.
 780          */
 781         if (flags & SYNC_WAIT)
 782                 log_flags |= XFS_LOG_SYNC;
 783
 784         xfs_log_force(mp, (xfs_lsn_t)0, log_flags);
 785
 786         if (flags & (SYNC_ATTR|SYNC_DELWRI)) {
 787                 if (flags & SYNC_BDFLUSH)
 788                         xfs_finish_reclaim_all(mp, 1);
 789                 else
 790                         error = xfs_sync_inodes(mp, flags, bypassed);
 791         }
 792
 793         /*
 794          * Flushing out dirty data above probably generated more
 795          * log activity, so if this isn't vfs_sync() then flush
 796          * the log again.
 797          */
 798         if (flags & SYNC_DELWRI) {
 799                 xfs_log_force(mp, (xfs_lsn_t)0, log_flags);
 800         }
 801
 802         if (flags & SYNC_FSDATA) {
 803                 /*
 804                  * If this is vfs_sync() then only sync the superblock
 805                  * if we can lock it without sleeping and it is not pinned.
 806                  */
 807                 if (flags & SYNC_BDFLUSH) {
 808                         bp = xfs_getsb(mp, XFS_BUF_TRYLOCK);
 809                         if (bp != NULL) {
 810                                 bip = XFS_BUF_FSPRIVATE(bp,xfs_buf_log_item_t*);
 811                                 if ((bip != NULL) &&
 812                                     xfs_buf_item_dirty(bip)) {
 813                                         if (!(XFS_BUF_ISPINNED(bp))) {
 814                                                 XFS_BUF_ASYNC(bp);
 815                                                 error = xfs_bwrite(mp, bp);
 816                                         } else {
 817                                                 xfs_buf_relse(bp);
 818                                         }
 819                                 } else {
 820                                         xfs_buf_relse(bp);
 821                                 }
 822                         }
 823                 } else {
 824                         bp = xfs_getsb(mp, 0);
 825                         /*
 826                          * If the buffer is pinned then push on the log so
 827                          * we won't get stuck waiting in the write for
 828                          * someone, maybe ourselves, to flush the log.
 829                          * Even though we just pushed the log above, we
 830                          * did not have the superblock buffer locked at
 831                          * that point so it can become pinned in between
 832                          * there and here.
 833                          */
 834                         if (XFS_BUF_ISPINNED(bp))
 835                                 xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE);
 836                         if (flags & SYNC_WAIT)
 837                                 XFS_BUF_UNASYNC(bp);
 838                         else
 839                                 XFS_BUF_ASYNC(bp);
 840                         error = xfs_bwrite(mp, bp);
 841                 }
 842                 if (error) {
 843                         last_error = error;
 844                 }
 845         }
 846
 847         /*
 848          * Now check to see if the log needs a "dummy" transaction.
 849          */
 850         if (!(flags & SYNC_REMOUNT) && xfs_log_need_covered(mp)) {
 851                 xfs_trans_t *tp;
 852                 xfs_inode_t *ip;
 853
 854                 /*
 855                  * Put a dummy transaction in the log to tell
 856                  * recovery that all others are OK.
 857                  */
 858                 tp = xfs_trans_alloc(mp, XFS_TRANS_DUMMY1);
 859                 if ((error = xfs_trans_reserve(tp, 0,
 860                                 XFS_ICHANGE_LOG_RES(mp),
 861                                 0, 0, 0)))  {
 862                         xfs_trans_cancel(tp, 0);
 863                         return error;
 864                 }
 865
 866                 ip = mp->m_rootip;
 867                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 868
 869                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
 870                 xfs_trans_ihold(tp, ip);
 871                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 872                 error = xfs_trans_commit(tp, 0);
 873                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 874                 xfs_log_force(mp, (xfs_lsn_t)0, log_flags);
 875         }
 876
 877         /*
 878          * When shutting down, we need to insure that the AIL is pushed
 879          * to disk or the filesystem can appear corrupt from the PROM.
 880          */
 881         if ((flags & (SYNC_CLOSE|SYNC_WAIT)) == (SYNC_CLOSE|SYNC_WAIT)) {
 882                 XFS_bflush(mp->m_ddev_targp);
 883                 if (mp->m_rtdev_targp) {
 884                         XFS_bflush(mp->m_rtdev_targp);
 885                 }
 886         }
 887
 888         return XFS_ERROR(last_error);
 889 }