/*
* This routine is called to allocate a transaction structure.
* The type parameter indicates the type of the transaction. These
* are enumerated in xfs_trans.h.
*
* Dynamically allocate the transaction structure from the transaction
* zone, initialize it, and return it to the caller.
*/
xfs_trans_t *
xfs_trans_alloc(
xfs_mount_t *mp,
uint type)
{
vfs_wait_for_freeze(XFS_MTOVFS(mp), SB_FREEZE_TRANS);
return _xfs_trans_alloc(mp, type);
}
static int
xfs_vn_allocate(xfs_mount_t *mp, xfs_inode_t *ip, struct xfs_vnode **vpp)
{
struct vnode *vp;
struct xfs_vnode *vdata;
int error;
/* Use zone allocator here? */
vdata = kmem_zalloc(sizeof(*vdata), KM_SLEEP);
error = getnewvnode("xfs", XVFSTOMNT(XFS_MTOVFS(mp)),
&xfs_vnops, &vp);
if (error) {
kmem_free(vdata, sizeof(*vdata));
return (error);
}
vp->v_vnlock->lk_flags |= LK_CANRECURSE;
vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curthread);
error = insmntque(vp, XVFSTOMNT(XFS_MTOVFS(mp)));
if (error != 0) {
kmem_free(vdata, sizeof(*vdata));
return (error);
}
vp->v_data = (void *)vdata;
vdata->v_number= 0;
vdata->v_inode = ip;
vdata->v_vfsp = XFS_MTOVFS(mp);
vdata->v_vnode = vp;
vn_bhv_head_init(VN_BHV_HEAD(vdata), "vnode");
#ifdef CONFIG_XFS_VNODE_TRACING
vp->v_trace = ktrace_alloc(VNODE_TRACE_SIZE, KM_SLEEP);
#endif /* CONFIG_XFS_VNODE_TRACING */
vn_trace_exit(vp, "vn_initialize", (inst_t *)__return_address);
if (error == 0)
*vpp = vdata;
return (error);
}
/*
* This routine is called to allocate a transaction structure.
* The type parameter indicates the type of the transaction. These
* are enumerated in xfs_trans.h.
*
* Dynamically allocate the transaction structure from the transaction
* zone, initialize it, and return it to the caller.
*/
xfs_trans_t *
xfs_trans_alloc(
xfs_mount_t *mp,
uint type)
{
fs_check_frozen(XFS_MTOVFS(mp), SB_FREEZE_TRANS);
atomic_inc(&mp->m_active_trans);
return (_xfs_trans_alloc(mp, type));
}
/*
* Clear the quotaflags in memory and in the superblock.
*/
int
xfs_mount_reset_sbqflags(xfs_mount_t *mp)
{
int error;
xfs_trans_t *tp;
unsigned long s;
mp->m_qflags = 0;
/*
* It is OK to look at sb_qflags here in mount path,
* without SB_LOCK.
*/
if (mp->m_sb.sb_qflags == 0)
return 0;
s = XFS_SB_LOCK(mp);
mp->m_sb.sb_qflags = 0;
XFS_SB_UNLOCK(mp, s);
/*
* if the fs is readonly, let the incore superblock run
* with quotas off but don't flush the update out to disk
*/
if (XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY)
return 0;
#ifdef QUOTADEBUG
xfs_fs_cmn_err(CE_NOTE, mp, "Writing superblock quota changes");
#endif
tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE);
if ((error = xfs_trans_reserve(tp, 0, mp->m_sb.sb_sectsize + 128, 0, 0,
XFS_DEFAULT_LOG_COUNT))) {
xfs_trans_cancel(tp, 0);
xfs_fs_cmn_err(CE_ALERT, mp,
"xfs_mount_reset_sbqflags: Superblock update failed!");
return error;
}
xfs_mod_sb(tp, XFS_SB_QFLAGS);
error = xfs_trans_commit(tp, 0, NULL);
return error;
}
/*
* xfs_trans_apply_sb_deltas() is called from the commit code
* to bring the superblock buffer into the current transaction
* and modify it as requested by earlier calls to xfs_trans_mod_sb().
*
* For now we just look at each field allowed to change and change
* it if necessary.
*/
STATIC void
xfs_trans_apply_sb_deltas(
xfs_trans_t *tp)
{
xfs_sb_t *sbp;
xfs_buf_t *bp;
int whole = 0;
bp = xfs_trans_getsb(tp, tp->t_mountp, 0);
sbp = XFS_BUF_TO_SBP(bp);
/*
* Check that superblock mods match the mods made to AGF counters.
*/
ASSERT((tp->t_fdblocks_delta + tp->t_res_fdblocks_delta) ==
(tp->t_ag_freeblks_delta + tp->t_ag_flist_delta +
tp->t_ag_btree_delta));
if (tp->t_icount_delta != 0) {
INT_MOD(sbp->sb_icount, ARCH_CONVERT, tp->t_icount_delta);
}
if (tp->t_ifree_delta != 0) {
INT_MOD(sbp->sb_ifree, ARCH_CONVERT, tp->t_ifree_delta);
}
if (tp->t_fdblocks_delta != 0) {
INT_MOD(sbp->sb_fdblocks, ARCH_CONVERT, tp->t_fdblocks_delta);
}
if (tp->t_res_fdblocks_delta != 0) {
INT_MOD(sbp->sb_fdblocks, ARCH_CONVERT, tp->t_res_fdblocks_delta);
}
if (tp->t_frextents_delta != 0) {
INT_MOD(sbp->sb_frextents, ARCH_CONVERT, tp->t_frextents_delta);
}
if (tp->t_res_frextents_delta != 0) {
INT_MOD(sbp->sb_frextents, ARCH_CONVERT, tp->t_res_frextents_delta);
}
if (tp->t_dblocks_delta != 0) {
INT_MOD(sbp->sb_dblocks, ARCH_CONVERT, tp->t_dblocks_delta);
whole = 1;
}
if (tp->t_agcount_delta != 0) {
INT_MOD(sbp->sb_agcount, ARCH_CONVERT, tp->t_agcount_delta);
whole = 1;
}
if (tp->t_imaxpct_delta != 0) {
INT_MOD(sbp->sb_imax_pct, ARCH_CONVERT, tp->t_imaxpct_delta);
whole = 1;
}
if (tp->t_rextsize_delta != 0) {
INT_MOD(sbp->sb_rextsize, ARCH_CONVERT, tp->t_rextsize_delta);
whole = 1;
}
if (tp->t_rbmblocks_delta != 0) {
INT_MOD(sbp->sb_rbmblocks, ARCH_CONVERT, tp->t_rbmblocks_delta);
whole = 1;
}
if (tp->t_rblocks_delta != 0) {
INT_MOD(sbp->sb_rblocks, ARCH_CONVERT, tp->t_rblocks_delta);
whole = 1;
}
if (tp->t_rextents_delta != 0) {
INT_MOD(sbp->sb_rextents, ARCH_CONVERT, tp->t_rextents_delta);
whole = 1;
}
if (tp->t_rextslog_delta != 0) {
INT_MOD(sbp->sb_rextslog, ARCH_CONVERT, tp->t_rextslog_delta);
whole = 1;
}
if (whole)
/*
* Log the whole thing, the fields are noncontiguous.
*/
xfs_trans_log_buf(tp, bp, 0, sizeof(xfs_sb_t) - 1);
else
/*
* Since all the modifiable fields are contiguous, we
* can get away with this.
*/
xfs_trans_log_buf(tp, bp, offsetof(xfs_sb_t, sb_icount),
offsetof(xfs_sb_t, sb_frextents) +
sizeof(sbp->sb_frextents) - 1);
#ifdef XXXKAN
XFS_MTOVFS(tp->t_mountp)->vfs_super->s_dirt = 1;
#endif
}
vfs_t *
xfs_mtovfs(xfs_mount_t *mp)
{
return XFS_MTOVFS(mp);
}
/*
* Look up an inode by number in the given file system.
* The inode is looked up in the hash table for the file system
* represented by the mount point parameter mp. Each bucket of
* the hash table is guarded by an individual semaphore.
*
* If the inode is found in the hash table, its corresponding vnode
* is obtained with a call to vn_get(). This call takes care of
* coordination with the reclamation of the inode and vnode. Note
* that the vmap structure is filled in while holding the hash lock.
* This gives us the state of the inode/vnode when we found it and
* is used for coordination in vn_get().
*
* If it is not in core, read it in from the file system's device and
* add the inode into the hash table.
*
* The inode is locked according to the value of the lock_flags parameter.
* This flag parameter indicates how and if the inode's IO lock and inode lock
* should be taken.
*
* mp -- the mount point structure for the current file system. It points
* to the inode hash table.
* tp -- a pointer to the current transaction if there is one. This is
* simply passed through to the xfs_iread() call.
* ino -- the number of the inode desired. This is the unique identifier
* within the file system for the inode being requested.
* lock_flags -- flags indicating how to lock the inode. See the comment
* for xfs_ilock() for a list of valid values.
* bno -- the block number starting the buffer containing the inode,
* if known (as by bulkstat), else 0.
*/
int
xfs_iget(
xfs_mount_t *mp,
xfs_trans_t *tp,
xfs_ino_t ino,
uint flags,
uint lock_flags,
xfs_inode_t **ipp,
xfs_daddr_t bno)
{
xfs_ihash_t *ih;
xfs_inode_t *ip;
xfs_inode_t *iq;
xfs_vnode_t *vp;
ulong version;
int error;
/* REFERENCED */
int newnode;
xfs_chash_t *ch;
xfs_chashlist_t *chl, *chlnew;
vmap_t vmap;
SPLDECL(s);
XFS_STATS_INC(xs_ig_attempts);
ih = XFS_IHASH(mp, ino);
again:
read_lock(&ih->ih_lock);
for (ip = ih->ih_next; ip != NULL; ip = ip->i_next) {
if (ip->i_ino == ino) {
vp = XFS_ITOV(ip);
VMAP(vp, vmap);
/*
* Inode cache hit: if ip is not at the front of
* its hash chain, move it there now.
* Do this with the lock held for update, but
* do statistics after releasing the lock.
*/
if (ip->i_prevp != &ih->ih_next
&& rwlock_trypromote(&ih->ih_lock)) {
if ((iq = ip->i_next)) {
iq->i_prevp = ip->i_prevp;
}
*ip->i_prevp = iq;
iq = ih->ih_next;
iq->i_prevp = &ip->i_next;
ip->i_next = iq;
ip->i_prevp = &ih->ih_next;
ih->ih_next = ip;
write_unlock(&ih->ih_lock);
} else {
read_unlock(&ih->ih_lock);
}
XFS_STATS_INC(xs_ig_found);
/*
* Get a reference to the vnode/inode.
* vn_get() takes care of coordination with
* the file system inode release and reclaim
* functions. If it returns NULL, the inode
* has been reclaimed so just start the search
* over again. We probably won't find it,
* but we could be racing with another cpu
* looking for the same inode so we have to at
* least look.
*/
if (!(vp = vn_get(vp, &vmap))) {
XFS_STATS_INC(xs_ig_frecycle);
goto again;
}
if (lock_flags != 0) {
ip->i_flags &= ~XFS_IRECLAIM;
xfs_ilock(ip, lock_flags);
}
newnode = (ip->i_d.di_mode == 0);
if (newnode) {
xfs_iocore_inode_reinit(ip);
}
ip->i_flags &= ~XFS_ISTALE;
vn_trace_exit(vp, "xfs_iget.found",
(inst_t *)__return_address);
goto return_ip;
}
}
/*
* Inode cache miss: save the hash chain version stamp and unlock
* the chain, so we don't deadlock in vn_alloc.
*/
XFS_STATS_INC(xs_ig_missed);
version = ih->ih_version;
read_unlock(&ih->ih_lock);
/*
* Read the disk inode attributes into a new inode structure and get
* a new vnode for it. This should also initialize i_ino and i_mount.
*/
error = xfs_iread(mp, tp, ino, &ip, bno);
if (error) {
return error;
}
error = xfs_vn_allocate(mp, ip, &vp);
if (error) {
return error;
}
vn_trace_exit(vp, "xfs_iget.alloc", (inst_t *)__return_address);
xfs_inode_lock_init(ip, vp);
xfs_iocore_inode_init(ip);
if (lock_flags != 0) {
xfs_ilock(ip, lock_flags);
}
/*
* Put ip on its hash chain, unless someone else hashed a duplicate
* after we released the hash lock.
*/
write_lock(&ih->ih_lock);
if (ih->ih_version != version) {
for (iq = ih->ih_next; iq != NULL; iq = iq->i_next) {
if (iq->i_ino == ino) {
write_unlock(&ih->ih_lock);
xfs_idestroy(ip);
XFS_STATS_INC(xs_ig_dup);
goto again;
}
}
}
/*
* These values _must_ be set before releasing ihlock!
*/
ip->i_hash = ih;
if ((iq = ih->ih_next)) {
iq->i_prevp = &ip->i_next;
}
ip->i_next = iq;
ip->i_prevp = &ih->ih_next;
ih->ih_next = ip;
ip->i_udquot = ip->i_gdquot = NULL;
ih->ih_version++;
write_unlock(&ih->ih_lock);
/*
* put ip on its cluster's hash chain
*/
ASSERT(ip->i_chash == NULL && ip->i_cprev == NULL &&
ip->i_cnext == NULL);
chlnew = NULL;
ch = XFS_CHASH(mp, ip->i_blkno);
chlredo:
s = mutex_spinlock(&ch->ch_lock);
for (chl = ch->ch_list; chl != NULL; chl = chl->chl_next) {
if (chl->chl_blkno == ip->i_blkno) {
/* insert this inode into the doubly-linked list
* where chl points */
if ((iq = chl->chl_ip)) {
ip->i_cprev = iq->i_cprev;
iq->i_cprev->i_cnext = ip;
iq->i_cprev = ip;
ip->i_cnext = iq;
} else {
ip->i_cnext = ip;
ip->i_cprev = ip;
}
chl->chl_ip = ip;
ip->i_chash = chl;
break;
}
}
/* no hash list found for this block; add a new hash list */
if (chl == NULL) {
if (chlnew == NULL) {
mutex_spinunlock(&ch->ch_lock, s);
ASSERT(xfs_chashlist_zone != NULL);
chlnew = (xfs_chashlist_t *)
kmem_zone_alloc(xfs_chashlist_zone,
KM_SLEEP);
ASSERT(chlnew != NULL);
goto chlredo;
} else {
ip->i_cnext = ip;
ip->i_cprev = ip;
ip->i_chash = chlnew;
chlnew->chl_ip = ip;
chlnew->chl_blkno = ip->i_blkno;
chlnew->chl_next = ch->ch_list;
ch->ch_list = chlnew;
chlnew = NULL;
}
} else {
if (chlnew != NULL) {
kmem_zone_free(xfs_chashlist_zone, chlnew);
}
}
mutex_spinunlock(&ch->ch_lock, s);
/*
* Link ip to its mount and thread it on the mount's inode list.
*/
XFS_MOUNT_ILOCK(mp);
if ((iq = mp->m_inodes)) {
ASSERT(iq->i_mprev->i_mnext == iq);
ip->i_mprev = iq->i_mprev;
iq->i_mprev->i_mnext = ip;
iq->i_mprev = ip;
ip->i_mnext = iq;
} else {
ip->i_mnext = ip;
ip->i_mprev = ip;
}
mp->m_inodes = ip;
XFS_MOUNT_IUNLOCK(mp);
newnode = 1;
return_ip:
ASSERT(ip->i_df.if_ext_max ==
XFS_IFORK_DSIZE(ip) / sizeof(xfs_bmbt_rec_t));
ASSERT(((ip->i_d.di_flags & XFS_DIFLAG_REALTIME) != 0) ==
((ip->i_iocore.io_flags & XFS_IOCORE_RT) != 0));
*ipp = ip;
/*
* If we have a real type for an on-disk inode, we can set ops(&unlock)
* now. If it's a new inode being created, xfs_ialloc will handle it.
*/
XVFS_INIT_VNODE(XFS_MTOVFS(mp), vp, XFS_ITOBHV(ip), 1);
return 0;
}
/*
* xfs sync routine for internal use
*
* This routine supports all of the flags defined for the generic VFS_SYNC
* interface as explained above under xfs_sync. In the interests of not
* changing interfaces within the 6.5 family, additional internallly-
* required functions are specified within a separate xflags parameter,
* only available by calling this routine.
*
*/
STATIC int
xfs_sync_inodes(
xfs_mount_t *mp,
int flags,
int xflags,
int *bypassed)
{
xfs_inode_t *ip = NULL;
xfs_inode_t *ip_next;
xfs_buf_t *bp;
vnode_t *vp = NULL;
vmap_t vmap;
int error;
int last_error;
uint64_t fflag;
uint lock_flags;
uint base_lock_flags;
boolean_t mount_locked;
boolean_t vnode_refed;
int preempt;
xfs_dinode_t *dip;
xfs_iptr_t *ipointer;
#ifdef DEBUG
boolean_t ipointer_in = B_FALSE;
#define IPOINTER_SET ipointer_in = B_TRUE
#define IPOINTER_CLR ipointer_in = B_FALSE
#else
#define IPOINTER_SET
#define IPOINTER_CLR
#endif
/* Insert a marker record into the inode list after inode ip. The list
* must be locked when this is called. After the call the list will no
* longer be locked.
*/
#define IPOINTER_INSERT(ip, mp) { \
ASSERT(ipointer_in == B_FALSE); \
ipointer->ip_mnext = ip->i_mnext; \
ipointer->ip_mprev = ip; \
ip->i_mnext = (xfs_inode_t *)ipointer; \
ipointer->ip_mnext->i_mprev = (xfs_inode_t *)ipointer; \
preempt = 0; \
XFS_MOUNT_IUNLOCK(mp); \
mount_locked = B_FALSE; \
IPOINTER_SET; \
}
/* Remove the marker from the inode list. If the marker was the only item
* in the list then there are no remaining inodes and we should zero out
* the whole list. If we are the current head of the list then move the head
* past us.
*/
#define IPOINTER_REMOVE(ip, mp) { \
ASSERT(ipointer_in == B_TRUE); \
if (ipointer->ip_mnext != (xfs_inode_t *)ipointer) { \
ip = ipointer->ip_mnext; \
ip->i_mprev = ipointer->ip_mprev; \
ipointer->ip_mprev->i_mnext = ip; \
if (mp->m_inodes == (xfs_inode_t *)ipointer) { \
mp->m_inodes = ip; \
} \
} else { \
ASSERT(mp->m_inodes == (xfs_inode_t *)ipointer); \
mp->m_inodes = NULL; \
ip = NULL; \
} \
IPOINTER_CLR; \
}
#define XFS_PREEMPT_MASK 0x7f
if (bypassed)
*bypassed = 0;
if (XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY)
return 0;
error = 0;
last_error = 0;
preempt = 0;
/* Allocate a reference marker */
ipointer = (xfs_iptr_t *)kmem_zalloc(sizeof(xfs_iptr_t), KM_SLEEP);
fflag = XFS_B_ASYNC; /* default is don't wait */
if (flags & SYNC_BDFLUSH)
fflag = XFS_B_DELWRI;
if (flags & SYNC_WAIT)
fflag = 0; /* synchronous overrides all */
base_lock_flags = XFS_ILOCK_SHARED;
if (flags & (SYNC_DELWRI | SYNC_CLOSE)) {
/*
* We need the I/O lock if we're going to call any of
* the flush/inval routines.
*/
base_lock_flags |= XFS_IOLOCK_SHARED;
}
XFS_MOUNT_ILOCK(mp);
ip = mp->m_inodes;
mount_locked = B_TRUE;
vnode_refed = B_FALSE;
IPOINTER_CLR;
do {
ASSERT(ipointer_in == B_FALSE);
ASSERT(vnode_refed == B_FALSE);
lock_flags = base_lock_flags;
/*
* There were no inodes in the list, just break out
* of the loop.
*/
if (ip == NULL) {
break;
}
/*
* We found another sync thread marker - skip it
*/
if (ip->i_mount == NULL) {
ip = ip->i_mnext;
continue;
}
vp = XFS_ITOV_NULL(ip);
/*
* If the vnode is gone then this is being torn down,
* call reclaim if it is flushed, else let regular flush
* code deal with it later in the loop.
*/
if (vp == NULL) {
/* Skip ones already in reclaim */
if (ip->i_flags & XFS_IRECLAIM) {
ip = ip->i_mnext;
continue;
}
if (xfs_ilock_nowait(ip, XFS_ILOCK_EXCL) == 0) {
ip = ip->i_mnext;
} else if ((xfs_ipincount(ip) == 0) &&
xfs_iflock_nowait(ip)) {
IPOINTER_INSERT(ip, mp);
xfs_finish_reclaim(ip, 1,
XFS_IFLUSH_DELWRI_ELSE_ASYNC);
XFS_MOUNT_ILOCK(mp);
mount_locked = B_TRUE;
IPOINTER_REMOVE(ip, mp);
} else {
xfs_iunlock(ip, XFS_ILOCK_EXCL);
ip = ip->i_mnext;
}
continue;
}
if (XFS_FORCED_SHUTDOWN(mp) && !(flags & SYNC_CLOSE)) {
XFS_MOUNT_IUNLOCK(mp);
kmem_free(ipointer, sizeof(xfs_iptr_t));
return 0;
}
/*
* If this is just vfs_sync() or pflushd() calling
* then we can skip inodes for which it looks like
* there is nothing to do. Since we don't have the
* inode locked this is racey, but these are periodic
* calls so it doesn't matter. For the others we want
* to know for sure, so we at least try to lock them.
*/
if (flags & SYNC_BDFLUSH) {
if (((ip->i_itemp == NULL) ||
!(ip->i_itemp->ili_format.ilf_fields &
XFS_ILOG_ALL)) &&
(ip->i_update_core == 0)) {
ip = ip->i_mnext;
continue;
}
}
/*
* Try to lock without sleeping. We're out of order with
* the inode list lock here, so if we fail we need to drop
* the mount lock and try again. If we're called from
* bdflush() here, then don't bother.
*
* The inode lock here actually coordinates with the
* almost spurious inode lock in xfs_ireclaim() to prevent
* the vnode we handle here without a reference from
* being freed while we reference it. If we lock the inode
* while it's on the mount list here, then the spurious inode
* lock in xfs_ireclaim() after the inode is pulled from
* the mount list will sleep until we release it here.
* This keeps the vnode from being freed while we reference
* it. It is also cheaper and simpler than actually doing
* a vn_get() for every inode we touch here.
*/
if (xfs_ilock_nowait(ip, lock_flags) == 0) {
if ((flags & SYNC_BDFLUSH) || (vp == NULL)) {
ip = ip->i_mnext;
continue;
}
/*
* We need to unlock the inode list lock in order
* to lock the inode. Insert a marker record into
* the inode list to remember our position, dropping
* the lock is now done inside the IPOINTER_INSERT
* macro.
*
* We also use the inode list lock to protect us
* in taking a snapshot of the vnode version number
* for use in calling vn_get().
*/
VMAP(vp, vmap);
IPOINTER_INSERT(ip, mp);
vp = vn_get(vp, &vmap);
if (vp == NULL) {
/*
* The vnode was reclaimed once we let go
* of the inode list lock. Skip to the
* next list entry. Remove the marker.
*/
XFS_MOUNT_ILOCK(mp);
mount_locked = B_TRUE;
vnode_refed = B_FALSE;
IPOINTER_REMOVE(ip, mp);
continue;
}
xfs_ilock(ip, lock_flags);
ASSERT(vp == XFS_ITOV(ip));
ASSERT(ip->i_mount == mp);
vnode_refed = B_TRUE;
}
/* From here on in the loop we may have a marker record
* in the inode list.
*/
if ((flags & SYNC_CLOSE) && (vp != NULL)) {
/*
* This is the shutdown case. We just need to
* flush and invalidate all the pages associated
* with the inode. Drop the inode lock since
* we can't hold it across calls to the buffer
* cache.
*
* We don't set the VREMAPPING bit in the vnode
* here, because we don't hold the vnode lock
* exclusively. It doesn't really matter, though,
* because we only come here when we're shutting
* down anyway.
*/
xfs_iunlock(ip, XFS_ILOCK_SHARED);
if (XFS_FORCED_SHUTDOWN(mp)) {
VOP_TOSS_PAGES(vp, 0, -1, FI_REMAPF);
} else {
VOP_FLUSHINVAL_PAGES(vp, 0, -1, FI_REMAPF);
}
xfs_ilock(ip, XFS_ILOCK_SHARED);
} else if ((flags & SYNC_DELWRI) && (vp != NULL)) {
if (VN_DIRTY(vp)) {
/* We need to have dropped the lock here,
* so insert a marker if we have not already
* done so.
*/
if (mount_locked) {
IPOINTER_INSERT(ip, mp);
}
/*
* Drop the inode lock since we can't hold it
* across calls to the buffer cache.
*/
xfs_iunlock(ip, XFS_ILOCK_SHARED);
VOP_FLUSH_PAGES(vp, (xfs_off_t)0, -1,
fflag, FI_NONE, error);
xfs_ilock(ip, XFS_ILOCK_SHARED);
}
}
if (flags & SYNC_BDFLUSH) {
if ((flags & SYNC_ATTR) &&
((ip->i_update_core) ||
((ip->i_itemp != NULL) &&
(ip->i_itemp->ili_format.ilf_fields != 0)))) {
/* Insert marker and drop lock if not already
* done.
*/
if (mount_locked) {
IPOINTER_INSERT(ip, mp);
}
/*
* We don't want the periodic flushing of the
* inodes by vfs_sync() to interfere with
* I/O to the file, especially read I/O
* where it is only the access time stamp
* that is being flushed out. To prevent
* long periods where we have both inode
* locks held shared here while reading the
* inode's buffer in from disk, we drop the
* inode lock while reading in the inode
* buffer. We have to release the buffer
* and reacquire the inode lock so that they
* are acquired in the proper order (inode
* locks first). The buffer will go at the
* end of the lru chain, though, so we can
* expect it to still be there when we go
* for it again in xfs_iflush().
*/
if ((xfs_ipincount(ip) == 0) &&
xfs_iflock_nowait(ip)) {
xfs_ifunlock(ip);
xfs_iunlock(ip, XFS_ILOCK_SHARED);
error = xfs_itobp(mp, NULL, ip,
&dip, &bp, 0);
if (!error) {
xfs_buf_relse(bp);
} else {
/* Bailing out, remove the
* marker and free it.
*/
XFS_MOUNT_ILOCK(mp);
IPOINTER_REMOVE(ip, mp);
XFS_MOUNT_IUNLOCK(mp);
ASSERT(!(lock_flags &
XFS_IOLOCK_SHARED));
kmem_free(ipointer,
sizeof(xfs_iptr_t));
return (0);
}
/*
* Since we dropped the inode lock,
* the inode may have been reclaimed.
* Therefore, we reacquire the mount
* lock and check to see if we were the
* inode reclaimed. If this happened
* then the ipointer marker will no
* longer point back at us. In this
* case, move ip along to the inode
* after the marker, remove the marker
* and continue.
*/
XFS_MOUNT_ILOCK(mp);
mount_locked = B_TRUE;
if (ip != ipointer->ip_mprev) {
IPOINTER_REMOVE(ip, mp);
ASSERT(!vnode_refed);
ASSERT(!(lock_flags &
XFS_IOLOCK_SHARED));
continue;
}
ASSERT(ip->i_mount == mp);
if (xfs_ilock_nowait(ip,
XFS_ILOCK_SHARED) == 0) {
ASSERT(ip->i_mount == mp);
/*
* We failed to reacquire
* the inode lock without
* sleeping, so just skip
* the inode for now. We
* clear the ILOCK bit from
* the lock_flags so that we
* won't try to drop a lock
* we don't hold below.
*/
lock_flags &= ~XFS_ILOCK_SHARED;
IPOINTER_REMOVE(ip_next, mp);
} else if ((xfs_ipincount(ip) == 0) &&
xfs_iflock_nowait(ip)) {
ASSERT(ip->i_mount == mp);
/*
* Since this is vfs_sync()
* calling we only flush the
* inode out if we can lock
* it without sleeping and
* it is not pinned. Drop
* the mount lock here so
* that we don't hold it for
* too long. We already have
* a marker in the list here.
*/
XFS_MOUNT_IUNLOCK(mp);
mount_locked = B_FALSE;
error = xfs_iflush(ip,
XFS_IFLUSH_DELWRI);
} else {
ASSERT(ip->i_mount == mp);
IPOINTER_REMOVE(ip_next, mp);
}
}
}
} else {
if ((flags & SYNC_ATTR) &&
((ip->i_update_core) ||
((ip->i_itemp != NULL) &&
(ip->i_itemp->ili_format.ilf_fields != 0)))) {
if (mount_locked) {
IPOINTER_INSERT(ip, mp);
}
if (flags & SYNC_WAIT) {
xfs_iflock(ip);
error = xfs_iflush(ip,
XFS_IFLUSH_SYNC);
} else {
/*
* If we can't acquire the flush
* lock, then the inode is already
* being flushed so don't bother
* waiting. If we can lock it then
* do a delwri flush so we can
* combine multiple inode flushes
* in each disk write.
*/
if (xfs_iflock_nowait(ip)) {
error = xfs_iflush(ip,
XFS_IFLUSH_DELWRI);
}
else if (bypassed)
(*bypassed)++;
}
}
}
if (lock_flags != 0) {
xfs_iunlock(ip, lock_flags);
}
if (vnode_refed) {
/*
* If we had to take a reference on the vnode
* above, then wait until after we've unlocked
* the inode to release the reference. This is
* because we can be already holding the inode
* lock when VN_RELE() calls xfs_inactive().
*
* Make sure to drop the mount lock before calling
* VN_RELE() so that we don't trip over ourselves if
* we have to go for the mount lock again in the
* inactive code.
*/
if (mount_locked) {
IPOINTER_INSERT(ip, mp);
}
VN_RELE(vp);
vnode_refed = B_FALSE;
}
if (error) {
last_error = error;
}
/*
* bail out if the filesystem is corrupted.
*/
if (error == EFSCORRUPTED) {
if (!mount_locked) {
XFS_MOUNT_ILOCK(mp);
IPOINTER_REMOVE(ip, mp);
}
XFS_MOUNT_IUNLOCK(mp);
ASSERT(ipointer_in == B_FALSE);
kmem_free(ipointer, sizeof(xfs_iptr_t));
return XFS_ERROR(error);
}
/* Let other threads have a chance at the mount lock
* if we have looped many times without dropping the
* lock.
*/
if ((++preempt & XFS_PREEMPT_MASK) == 0) {
if (mount_locked) {
IPOINTER_INSERT(ip, mp);
}
}
if (mount_locked == B_FALSE) {
XFS_MOUNT_ILOCK(mp);
mount_locked = B_TRUE;
IPOINTER_REMOVE(ip, mp);
continue;
}
ASSERT(ipointer_in == B_FALSE);
ip = ip->i_mnext;
} while (ip != mp->m_inodes);
XFS_MOUNT_IUNLOCK(mp);
ASSERT(ipointer_in == B_FALSE);
kmem_free(ipointer, sizeof(xfs_iptr_t));
return XFS_ERROR(last_error);
}