xfs: remove xfs_iget.c
authorDave Chinner <dchinner@redhat.com>
Mon, 8 Oct 2012 10:56:11 +0000 (21:56 +1100)
committerBen Myers <bpm@sgi.com>
Wed, 17 Oct 2012 18:42:25 +0000 (13:42 -0500)
The inode cache functions remaining in xfs_iget.c can be moved to xfs_icache.c
along with the other inode cache functions. This removes all functionality from
xfs_iget.c, so the file can simply be removed.

This move results in various functions now only having the scope of a single
file (e.g. xfs_inode_free()), so clean up all the definitions and exported
prototypes in xfs_icache.[ch] and xfs_inode.h appropriately.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
12 files changed:
fs/xfs/Makefile
fs/xfs/xfs_export.c
fs/xfs/xfs_icache.c
fs/xfs/xfs_icache.h
fs/xfs/xfs_iget.c [deleted file]
fs/xfs/xfs_inode.c
fs/xfs/xfs_inode.h
fs/xfs/xfs_itable.c
fs/xfs/xfs_log_recover.c
fs/xfs/xfs_qm.c
fs/xfs/xfs_rtalloc.c
fs/xfs/xfs_vnodeops.c

index 442f256dbcac6a3a9b33fdcf65610e9233bf36f8..e65357bb3dc641ac9b880935d35c67e373102169 100644 (file)
@@ -40,7 +40,6 @@ xfs-y                         += xfs_aops.o \
                                   xfs_fs_subr.o \
                                   xfs_globals.o \
                                   xfs_icache.o \
-                                  xfs_iget.o \
                                   xfs_ioctl.o \
                                   xfs_iomap.o \
                                   xfs_iops.o \
index 8c6d1d70278cc6e71dbc093cbbfc5f2c5d44df19..a83611849cee292d6731d8a782778d6ebb2bf048 100644 (file)
@@ -29,6 +29,7 @@
 #include "xfs_inode.h"
 #include "xfs_inode_item.h"
 #include "xfs_trace.h"
+#include "xfs_icache.h"
 
 /*
  * Note that we only accept fileids which are long enough rather than allow
index eba216f11d5e91af23c5cdf110bd82b0cc7c6bb6..9c8703b5cd72819610a12c37aecf2c8e329fa25d 100644 (file)
 #include <linux/kthread.h>
 #include <linux/freezer.h>
 
+STATIC void __xfs_inode_clear_reclaim_tag(struct xfs_mount *mp,
+                               struct xfs_perag *pag, struct xfs_inode *ip);
+
+/*
+ * Allocate and initialise an xfs_inode.
+ */
+STATIC struct xfs_inode *
+xfs_inode_alloc(
+       struct xfs_mount        *mp,
+       xfs_ino_t               ino)
+{
+       struct xfs_inode        *ip;
+
+       /*
+        * if this didn't occur in transactions, we could use
+        * KM_MAYFAIL and return NULL here on ENOMEM. Set the
+        * code up to do this anyway.
+        */
+       ip = kmem_zone_alloc(xfs_inode_zone, KM_SLEEP);
+       if (!ip)
+               return NULL;
+       if (inode_init_always(mp->m_super, VFS_I(ip))) {
+               kmem_zone_free(xfs_inode_zone, ip);
+               return NULL;
+       }
+
+       ASSERT(atomic_read(&ip->i_pincount) == 0);
+       ASSERT(!spin_is_locked(&ip->i_flags_lock));
+       ASSERT(!xfs_isiflocked(ip));
+       ASSERT(ip->i_ino == 0);
+
+       mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino);
+
+       /* initialise the xfs inode */
+       ip->i_ino = ino;
+       ip->i_mount = mp;
+       memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
+       ip->i_afp = NULL;
+       memset(&ip->i_df, 0, sizeof(xfs_ifork_t));
+       ip->i_flags = 0;
+       ip->i_delayed_blks = 0;
+       memset(&ip->i_d, 0, sizeof(xfs_icdinode_t));
+
+       return ip;
+}
+
+STATIC void
+xfs_inode_free_callback(
+       struct rcu_head         *head)
+{
+       struct inode            *inode = container_of(head, struct inode, i_rcu);
+       struct xfs_inode        *ip = XFS_I(inode);
+
+       kmem_zone_free(xfs_inode_zone, ip);
+}
+
+STATIC void
+xfs_inode_free(
+       struct xfs_inode        *ip)
+{
+       switch (ip->i_d.di_mode & S_IFMT) {
+       case S_IFREG:
+       case S_IFDIR:
+       case S_IFLNK:
+               xfs_idestroy_fork(ip, XFS_DATA_FORK);
+               break;
+       }
+
+       if (ip->i_afp)
+               xfs_idestroy_fork(ip, XFS_ATTR_FORK);
+
+       if (ip->i_itemp) {
+               ASSERT(!(ip->i_itemp->ili_item.li_flags & XFS_LI_IN_AIL));
+               xfs_inode_item_destroy(ip);
+               ip->i_itemp = NULL;
+       }
+
+       /* asserts to verify all state is correct here */
+       ASSERT(atomic_read(&ip->i_pincount) == 0);
+       ASSERT(!spin_is_locked(&ip->i_flags_lock));
+       ASSERT(!xfs_isiflocked(ip));
+
+       /*
+        * Because we use RCU freeing we need to ensure the inode always
+        * appears to be reclaimed with an invalid inode number when in the
+        * free state. The ip->i_flags_lock provides the barrier against lookup
+        * races.
+        */
+       spin_lock(&ip->i_flags_lock);
+       ip->i_flags = XFS_IRECLAIM;
+       ip->i_ino = 0;
+       spin_unlock(&ip->i_flags_lock);
+
+       call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
+}
+
+/*
+ * Check the validity of the inode we just found it the cache
+ */
+static int
+xfs_iget_cache_hit(
+       struct xfs_perag        *pag,
+       struct xfs_inode        *ip,
+       xfs_ino_t               ino,
+       int                     flags,
+       int                     lock_flags) __releases(RCU)
+{
+       struct inode            *inode = VFS_I(ip);
+       struct xfs_mount        *mp = ip->i_mount;
+       int                     error;
+
+       /*
+        * check for re-use of an inode within an RCU grace period due to the
+        * radix tree nodes not being updated yet. We monitor for this by
+        * setting the inode number to zero before freeing the inode structure.
+        * If the inode has been reallocated and set up, then the inode number
+        * will not match, so check for that, too.
+        */
+       spin_lock(&ip->i_flags_lock);
+       if (ip->i_ino != ino) {
+               trace_xfs_iget_skip(ip);
+               XFS_STATS_INC(xs_ig_frecycle);
+               error = EAGAIN;
+               goto out_error;
+       }
+
+
+       /*
+        * If we are racing with another cache hit that is currently
+        * instantiating this inode or currently recycling it out of
+        * reclaimabe state, wait for the initialisation to complete
+        * before continuing.
+        *
+        * XXX(hch): eventually we should do something equivalent to
+        *           wait_on_inode to wait for these flags to be cleared
+        *           instead of polling for it.
+        */
+       if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) {
+               trace_xfs_iget_skip(ip);
+               XFS_STATS_INC(xs_ig_frecycle);
+               error = EAGAIN;
+               goto out_error;
+       }
+
+       /*
+        * If lookup is racing with unlink return an error immediately.
+        */
+       if (ip->i_d.di_mode == 0 && !(flags & XFS_IGET_CREATE)) {
+               error = ENOENT;
+               goto out_error;
+       }
+
+       /*
+        * If IRECLAIMABLE is set, we've torn down the VFS inode already.
+        * Need to carefully get it back into useable state.
+        */
+       if (ip->i_flags & XFS_IRECLAIMABLE) {
+               trace_xfs_iget_reclaim(ip);
+
+               /*
+                * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode
+                * from stomping over us while we recycle the inode.  We can't
+                * clear the radix tree reclaimable tag yet as it requires
+                * pag_ici_lock to be held exclusive.
+                */
+               ip->i_flags |= XFS_IRECLAIM;
+
+               spin_unlock(&ip->i_flags_lock);
+               rcu_read_unlock();
+
+               error = -inode_init_always(mp->m_super, inode);
+               if (error) {
+                       /*
+                        * Re-initializing the inode failed, and we are in deep
+                        * trouble.  Try to re-add it to the reclaim list.
+                        */
+                       rcu_read_lock();
+                       spin_lock(&ip->i_flags_lock);
+
+                       ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
+                       ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
+                       trace_xfs_iget_reclaim_fail(ip);
+                       goto out_error;
+               }
+
+               spin_lock(&pag->pag_ici_lock);
+               spin_lock(&ip->i_flags_lock);
+
+               /*
+                * Clear the per-lifetime state in the inode as we are now
+                * effectively a new inode and need to return to the initial
+                * state before reuse occurs.
+                */
+               ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
+               ip->i_flags |= XFS_INEW;
+               __xfs_inode_clear_reclaim_tag(mp, pag, ip);
+               inode->i_state = I_NEW;
+
+               ASSERT(!rwsem_is_locked(&ip->i_iolock.mr_lock));
+               mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino);
+
+               spin_unlock(&ip->i_flags_lock);
+               spin_unlock(&pag->pag_ici_lock);
+       } else {
+               /* If the VFS inode is being torn down, pause and try again. */
+               if (!igrab(inode)) {
+                       trace_xfs_iget_skip(ip);
+                       error = EAGAIN;
+                       goto out_error;
+               }
+
+               /* We've got a live one. */
+               spin_unlock(&ip->i_flags_lock);
+               rcu_read_unlock();
+               trace_xfs_iget_hit(ip);
+       }
+
+       if (lock_flags != 0)
+               xfs_ilock(ip, lock_flags);
+
+       xfs_iflags_clear(ip, XFS_ISTALE | XFS_IDONTCACHE);
+       XFS_STATS_INC(xs_ig_found);
+
+       return 0;
+
+out_error:
+       spin_unlock(&ip->i_flags_lock);
+       rcu_read_unlock();
+       return error;
+}
+
+
+static int
+xfs_iget_cache_miss(
+       struct xfs_mount        *mp,
+       struct xfs_perag        *pag,
+       xfs_trans_t             *tp,
+       xfs_ino_t               ino,
+       struct xfs_inode        **ipp,
+       int                     flags,
+       int                     lock_flags)
+{
+       struct xfs_inode        *ip;
+       int                     error;
+       xfs_agino_t             agino = XFS_INO_TO_AGINO(mp, ino);
+       int                     iflags;
+
+       ip = xfs_inode_alloc(mp, ino);
+       if (!ip)
+               return ENOMEM;
+
+       error = xfs_iread(mp, tp, ip, flags);
+       if (error)
+               goto out_destroy;
+
+       trace_xfs_iget_miss(ip);
+
+       if ((ip->i_d.di_mode == 0) && !(flags & XFS_IGET_CREATE)) {
+               error = ENOENT;
+               goto out_destroy;
+       }
+
+       /*
+        * Preload the radix tree so we can insert safely under the
+        * write spinlock. Note that we cannot sleep inside the preload
+        * region. Since we can be called from transaction context, don't
+        * recurse into the file system.
+        */
+       if (radix_tree_preload(GFP_NOFS)) {
+               error = EAGAIN;
+               goto out_destroy;
+       }
+
+       /*
+        * Because the inode hasn't been added to the radix-tree yet it can't
+        * be found by another thread, so we can do the non-sleeping lock here.
+        */
+       if (lock_flags) {
+               if (!xfs_ilock_nowait(ip, lock_flags))
+                       BUG();
+       }
+
+       /*
+        * These values must be set before inserting the inode into the radix
+        * tree as the moment it is inserted a concurrent lookup (allowed by the
+        * RCU locking mechanism) can find it and that lookup must see that this
+        * is an inode currently under construction (i.e. that XFS_INEW is set).
+        * The ip->i_flags_lock that protects the XFS_INEW flag forms the
+        * memory barrier that ensures this detection works correctly at lookup
+        * time.
+        */
+       iflags = XFS_INEW;
+       if (flags & XFS_IGET_DONTCACHE)
+               iflags |= XFS_IDONTCACHE;
+       ip->i_udquot = ip->i_gdquot = NULL;
+       xfs_iflags_set(ip, iflags);
+
+       /* insert the new inode */
+       spin_lock(&pag->pag_ici_lock);
+       error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
+       if (unlikely(error)) {
+               WARN_ON(error != -EEXIST);
+               XFS_STATS_INC(xs_ig_dup);
+               error = EAGAIN;
+               goto out_preload_end;
+       }
+       spin_unlock(&pag->pag_ici_lock);
+       radix_tree_preload_end();
+
+       *ipp = ip;
+       return 0;
+
+out_preload_end:
+       spin_unlock(&pag->pag_ici_lock);
+       radix_tree_preload_end();
+       if (lock_flags)
+               xfs_iunlock(ip, lock_flags);
+out_destroy:
+       __destroy_inode(VFS_I(ip));
+       xfs_inode_free(ip);
+       return error;
+}
+
+/*
+ * Look up an inode by number in the given file system.
+ * The inode is looked up in the cache held in each AG.
+ * If the inode is found in the cache, initialise the vfs inode
+ * if necessary.
+ *
+ * If it is not in core, read it in from the file system's device,
+ * add it to the cache and initialise the vfs inode.
+ *
+ * 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.
+ */
+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_inode_t     *ip;
+       int             error;
+       xfs_perag_t     *pag;
+       xfs_agino_t     agino;
+
+       /*
+        * xfs_reclaim_inode() uses the ILOCK to ensure an inode
+        * doesn't get freed while it's being referenced during a
+        * radix tree traversal here.  It assumes this function
+        * aqcuires only the ILOCK (and therefore it has no need to
+        * involve the IOLOCK in this synchronization).
+        */
+       ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
+
+       /* reject inode numbers outside existing AGs */
+       if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
+               return EINVAL;
+
+       /* get the perag structure and ensure that it's inode capable */
+       pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
+       agino = XFS_INO_TO_AGINO(mp, ino);
+
+again:
+       error = 0;
+       rcu_read_lock();
+       ip = radix_tree_lookup(&pag->pag_ici_root, agino);
+
+       if (ip) {
+               error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
+               if (error)
+                       goto out_error_or_again;
+       } else {
+               rcu_read_unlock();
+               XFS_STATS_INC(xs_ig_missed);
+
+               error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
+                                                       flags, lock_flags);
+               if (error)
+                       goto out_error_or_again;
+       }
+       xfs_perag_put(pag);
+
+       *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.
+        */
+       if (xfs_iflags_test(ip, XFS_INEW) && ip->i_d.di_mode != 0)
+               xfs_setup_inode(ip);
+       return 0;
+
+out_error_or_again:
+       if (error == EAGAIN) {
+               delay(1);
+               goto again;
+       }
+       xfs_perag_put(pag);
+       return error;
+}
+
 /*
  * The inode lookup is done in batches to keep the amount of lock traffic and
  * radix tree lookups to a minimum. The batch size is a trade off between
@@ -253,7 +668,7 @@ xfs_reclaim_worker(
        xfs_reclaim_work_queue(mp);
 }
 
-void
+static void
 __xfs_inode_set_reclaim_tag(
        struct xfs_perag        *pag,
        struct xfs_inode        *ip)
@@ -319,7 +734,7 @@ __xfs_inode_clear_reclaim(
        }
 }
 
-void
+STATIC void
 __xfs_inode_clear_reclaim_tag(
        xfs_mount_t     *mp,
        xfs_perag_t     *pag,
@@ -542,7 +957,7 @@ out:
  * then a shut down during filesystem unmount reclaim walk leak all the
  * unreclaimed inodes.
  */
-int
+STATIC int
 xfs_reclaim_inodes_ag(
        struct xfs_mount        *mp,
        int                     flags,
index 0ba9c89c316e1e2415e2e39f852e56ecc16d1c58..222e22f16b4ac9c97583aa69cff15c54393200f8 100644 (file)
@@ -24,6 +24,9 @@ struct xfs_perag;
 #define SYNC_WAIT              0x0001  /* wait for i/o to complete */
 #define SYNC_TRYLOCK           0x0002  /* only try to lock inodes */
 
+int xfs_iget(struct xfs_mount *mp, struct xfs_trans *tp, xfs_ino_t ino,
+            uint flags, uint lock_flags, xfs_inode_t **ipp);
+
 void xfs_reclaim_worker(struct work_struct *work);
 
 int xfs_reclaim_inodes(struct xfs_mount *mp, int mode);
@@ -31,9 +34,6 @@ int xfs_reclaim_inodes_count(struct xfs_mount *mp);
 void xfs_reclaim_inodes_nr(struct xfs_mount *mp, int nr_to_scan);
 
 void xfs_inode_set_reclaim_tag(struct xfs_inode *ip);
-void __xfs_inode_set_reclaim_tag(struct xfs_perag *pag, struct xfs_inode *ip);
-void __xfs_inode_clear_reclaim_tag(struct xfs_mount *mp, struct xfs_perag *pag,
-                               struct xfs_inode *ip);
 
 int xfs_sync_inode_grab(struct xfs_inode *ip);
 int xfs_inode_ag_iterator(struct xfs_mount *mp,
diff --git a/fs/xfs/xfs_iget.c b/fs/xfs/xfs_iget.c
deleted file mode 100644 (file)
index ea9a5fa..0000000
+++ /dev/null
@@ -1,455 +0,0 @@
-/*
- * Copyright (c) 2000-2005 Silicon Graphics, Inc.
- * All Rights Reserved.
- *
- * This program is free software; you can redistribute it and/or
- * modify it under the terms of the GNU General Public License as
- * published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it would be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU General Public License for more details.
- *
- * You should have received a copy of the GNU General Public License
- * along with this program; if not, write the Free Software Foundation,
- * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
- */
-#include "xfs.h"
-#include "xfs_fs.h"
-#include "xfs_types.h"
-#include "xfs_acl.h"
-#include "xfs_log.h"
-#include "xfs_inum.h"
-#include "xfs_trans.h"
-#include "xfs_sb.h"
-#include "xfs_ag.h"
-#include "xfs_mount.h"
-#include "xfs_bmap_btree.h"
-#include "xfs_alloc_btree.h"
-#include "xfs_ialloc_btree.h"
-#include "xfs_dinode.h"
-#include "xfs_inode.h"
-#include "xfs_btree.h"
-#include "xfs_ialloc.h"
-#include "xfs_quota.h"
-#include "xfs_utils.h"
-#include "xfs_trans_priv.h"
-#include "xfs_inode_item.h"
-#include "xfs_bmap.h"
-#include "xfs_trace.h"
-#include "xfs_icache.h"
-
-
-/*
- * Allocate and initialise an xfs_inode.
- */
-STATIC struct xfs_inode *
-xfs_inode_alloc(
-       struct xfs_mount        *mp,
-       xfs_ino_t               ino)
-{
-       struct xfs_inode        *ip;
-
-       /*
-        * if this didn't occur in transactions, we could use
-        * KM_MAYFAIL and return NULL here on ENOMEM. Set the
-        * code up to do this anyway.
-        */
-       ip = kmem_zone_alloc(xfs_inode_zone, KM_SLEEP);
-       if (!ip)
-               return NULL;
-       if (inode_init_always(mp->m_super, VFS_I(ip))) {
-               kmem_zone_free(xfs_inode_zone, ip);
-               return NULL;
-       }
-
-       ASSERT(atomic_read(&ip->i_pincount) == 0);
-       ASSERT(!spin_is_locked(&ip->i_flags_lock));
-       ASSERT(!xfs_isiflocked(ip));
-       ASSERT(ip->i_ino == 0);
-
-       mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino);
-
-       /* initialise the xfs inode */
-       ip->i_ino = ino;
-       ip->i_mount = mp;
-       memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
-       ip->i_afp = NULL;
-       memset(&ip->i_df, 0, sizeof(xfs_ifork_t));
-       ip->i_flags = 0;
-       ip->i_delayed_blks = 0;
-       memset(&ip->i_d, 0, sizeof(xfs_icdinode_t));
-
-       return ip;
-}
-
-STATIC void
-xfs_inode_free_callback(
-       struct rcu_head         *head)
-{
-       struct inode            *inode = container_of(head, struct inode, i_rcu);
-       struct xfs_inode        *ip = XFS_I(inode);
-
-       kmem_zone_free(xfs_inode_zone, ip);
-}
-
-void
-xfs_inode_free(
-       struct xfs_inode        *ip)
-{
-       switch (ip->i_d.di_mode & S_IFMT) {
-       case S_IFREG:
-       case S_IFDIR:
-       case S_IFLNK:
-               xfs_idestroy_fork(ip, XFS_DATA_FORK);
-               break;
-       }
-
-       if (ip->i_afp)
-               xfs_idestroy_fork(ip, XFS_ATTR_FORK);
-
-       if (ip->i_itemp) {
-               ASSERT(!(ip->i_itemp->ili_item.li_flags & XFS_LI_IN_AIL));
-               xfs_inode_item_destroy(ip);
-               ip->i_itemp = NULL;
-       }
-
-       /* asserts to verify all state is correct here */
-       ASSERT(atomic_read(&ip->i_pincount) == 0);
-       ASSERT(!spin_is_locked(&ip->i_flags_lock));
-       ASSERT(!xfs_isiflocked(ip));
-
-       /*
-        * Because we use RCU freeing we need to ensure the inode always
-        * appears to be reclaimed with an invalid inode number when in the
-        * free state. The ip->i_flags_lock provides the barrier against lookup
-        * races.
-        */
-       spin_lock(&ip->i_flags_lock);
-       ip->i_flags = XFS_IRECLAIM;
-       ip->i_ino = 0;
-       spin_unlock(&ip->i_flags_lock);
-
-       call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
-}
-
-/*
- * Check the validity of the inode we just found it the cache
- */
-static int
-xfs_iget_cache_hit(
-       struct xfs_perag        *pag,
-       struct xfs_inode        *ip,
-       xfs_ino_t               ino,
-       int                     flags,
-       int                     lock_flags) __releases(RCU)
-{
-       struct inode            *inode = VFS_I(ip);
-       struct xfs_mount        *mp = ip->i_mount;
-       int                     error;
-
-       /*
-        * check for re-use of an inode within an RCU grace period due to the
-        * radix tree nodes not being updated yet. We monitor for this by
-        * setting the inode number to zero before freeing the inode structure.
-        * If the inode has been reallocated and set up, then the inode number
-        * will not match, so check for that, too.
-        */
-       spin_lock(&ip->i_flags_lock);
-       if (ip->i_ino != ino) {
-               trace_xfs_iget_skip(ip);
-               XFS_STATS_INC(xs_ig_frecycle);
-               error = EAGAIN;
-               goto out_error;
-       }
-
-
-       /*
-        * If we are racing with another cache hit that is currently
-        * instantiating this inode or currently recycling it out of
-        * reclaimabe state, wait for the initialisation to complete
-        * before continuing.
-        *
-        * XXX(hch): eventually we should do something equivalent to
-        *           wait_on_inode to wait for these flags to be cleared
-        *           instead of polling for it.
-        */
-       if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) {
-               trace_xfs_iget_skip(ip);
-               XFS_STATS_INC(xs_ig_frecycle);
-               error = EAGAIN;
-               goto out_error;
-       }
-
-       /*
-        * If lookup is racing with unlink return an error immediately.
-        */
-       if (ip->i_d.di_mode == 0 && !(flags & XFS_IGET_CREATE)) {
-               error = ENOENT;
-               goto out_error;
-       }
-
-       /*
-        * If IRECLAIMABLE is set, we've torn down the VFS inode already.
-        * Need to carefully get it back into useable state.
-        */
-       if (ip->i_flags & XFS_IRECLAIMABLE) {
-               trace_xfs_iget_reclaim(ip);
-
-               /*
-                * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode
-                * from stomping over us while we recycle the inode.  We can't
-                * clear the radix tree reclaimable tag yet as it requires
-                * pag_ici_lock to be held exclusive.
-                */
-               ip->i_flags |= XFS_IRECLAIM;
-
-               spin_unlock(&ip->i_flags_lock);
-               rcu_read_unlock();
-
-               error = -inode_init_always(mp->m_super, inode);
-               if (error) {
-                       /*
-                        * Re-initializing the inode failed, and we are in deep
-                        * trouble.  Try to re-add it to the reclaim list.
-                        */
-                       rcu_read_lock();
-                       spin_lock(&ip->i_flags_lock);
-
-                       ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
-                       ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
-                       trace_xfs_iget_reclaim_fail(ip);
-                       goto out_error;
-               }
-
-               spin_lock(&pag->pag_ici_lock);
-               spin_lock(&ip->i_flags_lock);
-
-               /*
-                * Clear the per-lifetime state in the inode as we are now
-                * effectively a new inode and need to return to the initial
-                * state before reuse occurs.
-                */
-               ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
-               ip->i_flags |= XFS_INEW;
-               __xfs_inode_clear_reclaim_tag(mp, pag, ip);
-               inode->i_state = I_NEW;
-
-               ASSERT(!rwsem_is_locked(&ip->i_iolock.mr_lock));
-               mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino);
-
-               spin_unlock(&ip->i_flags_lock);
-               spin_unlock(&pag->pag_ici_lock);
-       } else {
-               /* If the VFS inode is being torn down, pause and try again. */
-               if (!igrab(inode)) {
-                       trace_xfs_iget_skip(ip);
-                       error = EAGAIN;
-                       goto out_error;
-               }
-
-               /* We've got a live one. */
-               spin_unlock(&ip->i_flags_lock);
-               rcu_read_unlock();
-               trace_xfs_iget_hit(ip);
-       }
-
-       if (lock_flags != 0)
-               xfs_ilock(ip, lock_flags);
-
-       xfs_iflags_clear(ip, XFS_ISTALE | XFS_IDONTCACHE);
-       XFS_STATS_INC(xs_ig_found);
-
-       return 0;
-
-out_error:
-       spin_unlock(&ip->i_flags_lock);
-       rcu_read_unlock();
-       return error;
-}
-
-
-static int
-xfs_iget_cache_miss(
-       struct xfs_mount        *mp,
-       struct xfs_perag        *pag,
-       xfs_trans_t             *tp,
-       xfs_ino_t               ino,
-       struct xfs_inode        **ipp,
-       int                     flags,
-       int                     lock_flags)
-{
-       struct xfs_inode        *ip;
-       int                     error;
-       xfs_agino_t             agino = XFS_INO_TO_AGINO(mp, ino);
-       int                     iflags;
-
-       ip = xfs_inode_alloc(mp, ino);
-       if (!ip)
-               return ENOMEM;
-
-       error = xfs_iread(mp, tp, ip, flags);
-       if (error)
-               goto out_destroy;
-
-       trace_xfs_iget_miss(ip);
-
-       if ((ip->i_d.di_mode == 0) && !(flags & XFS_IGET_CREATE)) {
-               error = ENOENT;
-               goto out_destroy;
-       }
-
-       /*
-        * Preload the radix tree so we can insert safely under the
-        * write spinlock. Note that we cannot sleep inside the preload
-        * region. Since we can be called from transaction context, don't
-        * recurse into the file system.
-        */
-       if (radix_tree_preload(GFP_NOFS)) {
-               error = EAGAIN;
-               goto out_destroy;
-       }
-
-       /*
-        * Because the inode hasn't been added to the radix-tree yet it can't
-        * be found by another thread, so we can do the non-sleeping lock here.
-        */
-       if (lock_flags) {
-               if (!xfs_ilock_nowait(ip, lock_flags))
-                       BUG();
-       }
-
-       /*
-        * These values must be set before inserting the inode into the radix
-        * tree as the moment it is inserted a concurrent lookup (allowed by the
-        * RCU locking mechanism) can find it and that lookup must see that this
-        * is an inode currently under construction (i.e. that XFS_INEW is set).
-        * The ip->i_flags_lock that protects the XFS_INEW flag forms the
-        * memory barrier that ensures this detection works correctly at lookup
-        * time.
-        */
-       iflags = XFS_INEW;
-       if (flags & XFS_IGET_DONTCACHE)
-               iflags |= XFS_IDONTCACHE;
-       ip->i_udquot = ip->i_gdquot = NULL;
-       xfs_iflags_set(ip, iflags);
-
-       /* insert the new inode */
-       spin_lock(&pag->pag_ici_lock);
-       error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
-       if (unlikely(error)) {
-               WARN_ON(error != -EEXIST);
-               XFS_STATS_INC(xs_ig_dup);
-               error = EAGAIN;
-               goto out_preload_end;
-       }
-       spin_unlock(&pag->pag_ici_lock);
-       radix_tree_preload_end();
-
-       *ipp = ip;
-       return 0;
-
-out_preload_end:
-       spin_unlock(&pag->pag_ici_lock);
-       radix_tree_preload_end();
-       if (lock_flags)
-               xfs_iunlock(ip, lock_flags);
-out_destroy:
-       __destroy_inode(VFS_I(ip));
-       xfs_inode_free(ip);
-       return error;
-}
-
-/*
- * Look up an inode by number in the given file system.
- * The inode is looked up in the cache held in each AG.
- * If the inode is found in the cache, initialise the vfs inode
- * if necessary.
- *
- * If it is not in core, read it in from the file system's device,
- * add it to the cache and initialise the vfs inode.
- *
- * 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.
- */
-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_inode_t     *ip;
-       int             error;
-       xfs_perag_t     *pag;
-       xfs_agino_t     agino;
-
-       /*
-        * xfs_reclaim_inode() uses the ILOCK to ensure an inode
-        * doesn't get freed while it's being referenced during a
-        * radix tree traversal here.  It assumes this function
-        * aqcuires only the ILOCK (and therefore it has no need to
-        * involve the IOLOCK in this synchronization).
-        */
-       ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
-
-       /* reject inode numbers outside existing AGs */
-       if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
-               return EINVAL;
-
-       /* get the perag structure and ensure that it's inode capable */
-       pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
-       agino = XFS_INO_TO_AGINO(mp, ino);
-
-again:
-       error = 0;
-       rcu_read_lock();
-       ip = radix_tree_lookup(&pag->pag_ici_root, agino);
-
-       if (ip) {
-               error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
-               if (error)
-                       goto out_error_or_again;
-       } else {
-               rcu_read_unlock();
-               XFS_STATS_INC(xs_ig_missed);
-
-               error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
-                                                       flags, lock_flags);
-               if (error)
-                       goto out_error_or_again;
-       }
-       xfs_perag_put(pag);
-
-       *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.
-        */
-       if (xfs_iflags_test(ip, XFS_INEW) && ip->i_d.di_mode != 0)
-               xfs_setup_inode(ip);
-       return 0;
-
-out_error_or_again:
-       if (error == EAGAIN) {
-               delay(1);
-               goto again;
-       }
-       xfs_perag_put(pag);
-       return error;
-}
-
index ba404e4b9f0cb334054ec4e219c0ae77e807242a..bba8f37525b3867576aea0acaf3adddbadbe7f2b 100644 (file)
@@ -45,6 +45,7 @@
 #include "xfs_filestream.h"
 #include "xfs_vnodeops.h"
 #include "xfs_trace.h"
+#include "xfs_icache.h"
 
 kmem_zone_t *xfs_ifork_zone;
 kmem_zone_t *xfs_inode_zone;
index 94b32f906e7903a32cc1e1474f3d4a6fed267f93..1fc2065e010b4ddf3bbd8889735a979588a4a857 100644 (file)
@@ -496,11 +496,10 @@ static inline int xfs_isiflocked(struct xfs_inode *ip)
        (((pip)->i_mount->m_flags & XFS_MOUNT_GRPID) || \
         ((pip)->i_d.di_mode & S_ISGID))
 
+
 /*
- * xfs_iget.c prototypes.
+ * xfs_inode.c prototypes.
  */
-int            xfs_iget(struct xfs_mount *, struct xfs_trans *, xfs_ino_t,
-                        uint, uint, xfs_inode_t **);
 void           xfs_ilock(xfs_inode_t *, uint);
 int            xfs_ilock_nowait(xfs_inode_t *, uint);
 void           xfs_iunlock(xfs_inode_t *, uint);
@@ -508,11 +507,6 @@ void               xfs_ilock_demote(xfs_inode_t *, uint);
 int            xfs_isilocked(xfs_inode_t *, uint);
 uint           xfs_ilock_map_shared(xfs_inode_t *);
 void           xfs_iunlock_map_shared(xfs_inode_t *, uint);
-void           xfs_inode_free(struct xfs_inode *ip);
-
-/*
- * xfs_inode.c prototypes.
- */
 int            xfs_ialloc(struct xfs_trans *, xfs_inode_t *, umode_t,
                           xfs_nlink_t, xfs_dev_t, prid_t, int,
                           struct xfs_buf **, xfs_inode_t **);
index 01d10a66e30243518d293f2c72a29b6135f7977b..3998fd2a7949475801898801e9009efbd5d4aeb9 100644 (file)
@@ -34,6 +34,7 @@
 #include "xfs_error.h"
 #include "xfs_btree.h"
 #include "xfs_trace.h"
+#include "xfs_icache.h"
 
 STATIC int
 xfs_internal_inum(
index 5da3ace352bffe6ca32d16ae76fc131291914a00..651c98859b045fd80bf40d463fc5ce9a908c8ad5 100644 (file)
@@ -42,6 +42,7 @@
 #include "xfs_quota.h"
 #include "xfs_utils.h"
 #include "xfs_trace.h"
+#include "xfs_icache.h"
 
 STATIC int
 xlog_find_zeroed(
index 2e86fa0cfc0d660374b2d8860f7c30e059b94e2b..48c750b0e830444dffcf9404ada2c919115673b1 100644 (file)
@@ -40,6 +40,7 @@
 #include "xfs_utils.h"
 #include "xfs_qm.h"
 #include "xfs_trace.h"
+#include "xfs_icache.h"
 
 /*
  * The global quota manager. There is only one of these for the entire
index ca28a4ba4b548f0c379291bfb0e716ac3a9ec54e..a69e0b4750a9ff8761f1096fd6473e3ab36509ed 100644 (file)
@@ -38,6 +38,7 @@
 #include "xfs_utils.h"
 #include "xfs_trace.h"
 #include "xfs_buf.h"
+#include "xfs_icache.h"
 
 
 /*
index 14928564f1068ca97a8c837785a74314da32f1ba..2ee1f49da0aa70eed076c03d4ebfef78f172acef 100644 (file)
@@ -47,6 +47,7 @@
 #include "xfs_filestream.h"
 #include "xfs_vnodeops.h"
 #include "xfs_trace.h"
+#include "xfs_icache.h"
 
 /*
  * The maximum pathlen is 1024 bytes. Since the minimum file system