xfs: xfs_swap_extents needs to handle dynamic fork offsets
[safe/jmp/linux-2.6] / fs / xfs / xfs_iget.c
index 1256746..155e798 100644 (file)
@@ -18,6 +18,7 @@
 #include "xfs.h"
 #include "xfs_fs.h"
 #include "xfs_types.h"
+#include "xfs_acl.h"
 #include "xfs_bit.h"
 #include "xfs_log.h"
 #include "xfs_inum.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_btree_trace.h"
+#include "xfs_trace.h"
+
 
 /*
- * 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, attach it to the provided
- * vnode.
- *
- * If it is not in core, read it in from the file system's device,
- * add it to the cache and attach the provided vnode.
- *
- * 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.
+ * Allocate and initialise an xfs_inode.
  */
-STATIC int
-xfs_iget_core(
-       struct inode    *inode,
-       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)
+STATIC struct xfs_inode *
+xfs_inode_alloc(
+       struct xfs_mount        *mp,
+       xfs_ino_t               ino)
 {
-       struct inode    *old_inode;
-       xfs_inode_t     *ip;
-       int             error;
-       unsigned long   first_index, mask;
-       xfs_perag_t     *pag;
-       xfs_agino_t     agino;
+       struct xfs_inode        *ip;
 
-       /* the radix tree exists only in inode capable AGs */
-       if (XFS_INO_TO_AGNO(mp, ino) >= mp->m_maxagi)
-               return EINVAL;
+       /*
+        * 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;
+       }
 
-       /* get the perag structure and ensure that it's inode capable */
-       pag = xfs_get_perag(mp, ino);
-       if (!pag->pagi_inodeok)
-               return EINVAL;
-       ASSERT(pag->pag_ici_init);
-       agino = XFS_INO_TO_AGINO(mp, ino);
+       ASSERT(atomic_read(&ip->i_iocount) == 0);
+       ASSERT(atomic_read(&ip->i_pincount) == 0);
+       ASSERT(!spin_is_locked(&ip->i_flags_lock));
+       ASSERT(completion_done(&ip->i_flush));
+
+       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_update_core = 0;
+       ip->i_delayed_blks = 0;
+       memset(&ip->i_d, 0, sizeof(xfs_icdinode_t));
+       ip->i_size = 0;
+       ip->i_new_size = 0;
+
+       /* prevent anyone from using this yet */
+       VFS_I(ip)->i_state = I_NEW;
 
-again:
-       read_lock(&pag->pag_ici_lock);
-       ip = radix_tree_lookup(&pag->pag_ici_root, agino);
+       return 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 != NULL) {
+       if (ip->i_afp)
+               xfs_idestroy_fork(ip, XFS_ATTR_FORK);
+
+       if (ip->i_itemp) {
                /*
-                * If INEW is set this inode is being set up
-                * we need to pause and try again.
+                * Only if we are shutting down the fs will we see an
+                * inode still in the AIL. If it is there, we should remove
+                * it to prevent a use-after-free from occurring.
                 */
-               if (xfs_iflags_test(ip, XFS_INEW)) {
-                       read_unlock(&pag->pag_ici_lock);
-                       delay(1);
-                       XFS_STATS_INC(xs_ig_frecycle);
-
-                       goto again;
+               xfs_log_item_t  *lip = &ip->i_itemp->ili_item;
+               struct xfs_ail  *ailp = lip->li_ailp;
+
+               ASSERT(((lip->li_flags & XFS_LI_IN_AIL) == 0) ||
+                                      XFS_FORCED_SHUTDOWN(ip->i_mount));
+               if (lip->li_flags & XFS_LI_IN_AIL) {
+                       spin_lock(&ailp->xa_lock);
+                       if (lip->li_flags & XFS_LI_IN_AIL)
+                               xfs_trans_ail_delete(ailp, lip);
+                       else
+                               spin_unlock(&ailp->xa_lock);
                }
+               xfs_inode_item_destroy(ip);
+               ip->i_itemp = NULL;
+       }
 
-               old_inode = ip->i_vnode;
-               if (old_inode == NULL) {
-                       /*
-                        * If IRECLAIM is set this inode is
-                        * on its way out of the system,
-                        * we need to pause and try again.
-                        */
-                       if (xfs_iflags_test(ip, XFS_IRECLAIM)) {
-                               read_unlock(&pag->pag_ici_lock);
-                               delay(1);
-                               XFS_STATS_INC(xs_ig_frecycle);
-
-                               goto again;
-                       }
-                       ASSERT(xfs_iflags_test(ip, XFS_IRECLAIMABLE));
+       /* asserts to verify all state is correct here */
+       ASSERT(atomic_read(&ip->i_iocount) == 0);
+       ASSERT(atomic_read(&ip->i_pincount) == 0);
+       ASSERT(!spin_is_locked(&ip->i_flags_lock));
+       ASSERT(completion_done(&ip->i_flush));
 
-                       /*
-                        * If lookup is racing with unlink, then we
-                        * should return an error immediately so we
-                        * don't remove it from the reclaim list and
-                        * potentially leak the inode.
-                        */
-                       if ((ip->i_d.di_mode == 0) &&
-                           !(flags & XFS_IGET_CREATE)) {
-                               read_unlock(&pag->pag_ici_lock);
-                               xfs_put_perag(mp, pag);
-                               return ENOENT;
-                       }
+       kmem_zone_free(xfs_inode_zone, ip);
+}
 
-                       xfs_itrace_exit_tag(ip, "xfs_iget.alloc");
+/*
+ * 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,
+       int                     flags,
+       int                     lock_flags) __releases(pag->pag_ici_lock)
+{
+       struct inode            *inode = VFS_I(ip);
+       struct xfs_mount        *mp = ip->i_mount;
+       int                     error;
 
-                       XFS_STATS_INC(xs_ig_found);
-                       xfs_iflags_clear(ip, XFS_IRECLAIMABLE);
-                       read_unlock(&pag->pag_ici_lock);
+       spin_lock(&ip->i_flags_lock);
 
-                       XFS_MOUNT_ILOCK(mp);
-                       list_del_init(&ip->i_reclaim);
-                       XFS_MOUNT_IUNLOCK(mp);
+       /*
+        * 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;
+       }
 
-                       goto finish_inode;
+       /*
+        * 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;
+       }
 
-               } else if (inode != old_inode) {
-                       /* The inode is being torn down, pause and
-                        * try again.
-                        */
-                       if (old_inode->i_state & (I_FREEING | I_CLEAR)) {
-                               read_unlock(&pag->pag_ici_lock);
-                               delay(1);
-                               XFS_STATS_INC(xs_ig_frecycle);
-
-                               goto again;
-                       }
-/* Chances are the other vnode (the one in the inode) is being torn
-* down right now, and we landed on top of it. Question is, what do
-* we do? Unhook the old inode and hook up the new one?
-*/
-                       cmn_err(CE_PANIC,
-               "xfs_iget_core: ambiguous vns: vp/0x%p, invp/0x%p",
-                                       old_inode, inode);
-               }
+       /*
+        * 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);
 
                /*
-                * Inode cache hit
+                * We need to set XFS_INEW atomically with clearing the
+                * reclaimable tag so that we do have an indicator of the
+                * inode still being initialized.
                 */
+               ip->i_flags |= XFS_INEW;
+               ip->i_flags &= ~XFS_IRECLAIMABLE;
+               __xfs_inode_clear_reclaim_tag(mp, pag, ip);
+
+               spin_unlock(&ip->i_flags_lock);
                read_unlock(&pag->pag_ici_lock);
-               XFS_STATS_INC(xs_ig_found);
 
-finish_inode:
-               if (ip->i_d.di_mode == 0 && !(flags & XFS_IGET_CREATE)) {
-                       xfs_put_perag(mp, pag);
-                       return ENOENT;
+               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.
+                        */
+                       read_lock(&pag->pag_ici_lock);
+                       spin_lock(&ip->i_flags_lock);
+
+                       ip->i_flags &= ~XFS_INEW;
+                       ip->i_flags |= XFS_IRECLAIMABLE;
+                       __xfs_inode_set_reclaim_tag(pag, ip);
+                       trace_xfs_iget_reclaim(ip);
+                       goto out_error;
+               }
+               inode->i_state = I_NEW;
+       } else {
+               /* If the VFS inode is being torn down, pause and try again. */
+               if (!igrab(inode)) {
+                       error = EAGAIN;
+                       goto out_error;
                }
 
-               if (lock_flags != 0)
-                       xfs_ilock(ip, lock_flags);
-
-               xfs_iflags_clear(ip, XFS_ISTALE);
-               xfs_itrace_exit_tag(ip, "xfs_iget.found");
-               goto return_ip;
+               /* We've got a live one. */
+               spin_unlock(&ip->i_flags_lock);
+               read_unlock(&pag->pag_ici_lock);
        }
 
-       /*
-        * Inode cache miss
-        */
+       if (lock_flags != 0)
+               xfs_ilock(ip, lock_flags);
+
+       xfs_iflags_clear(ip, XFS_ISTALE);
+       XFS_STATS_INC(xs_ig_found);
+
+       trace_xfs_iget_found(ip);
+       return 0;
+
+out_error:
+       spin_unlock(&ip->i_flags_lock);
        read_unlock(&pag->pag_ici_lock);
-       XFS_STATS_INC(xs_ig_missed);
+       return error;
+}
 
-       /*
-        * 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,
-                         (flags & XFS_IGET_BULKSTAT) ? XFS_IMAP_BULKSTAT : 0);
-       if (error) {
-               xfs_put_perag(mp, pag);
-               return error;
-       }
 
-       xfs_itrace_exit_tag(ip, "xfs_iget.alloc");
+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,
+       xfs_daddr_t             bno,
+       int                     flags,
+       int                     lock_flags)
+{
+       struct xfs_inode        *ip;
+       int                     error;
+       unsigned long           first_index, mask;
+       xfs_agino_t             agino = XFS_INO_TO_AGINO(mp, ino);
+
+       ip = xfs_inode_alloc(mp, ino);
+       if (!ip)
+               return ENOMEM;
+
+       error = xfs_iread(mp, tp, ip, bno, flags);
+       if (error)
+               goto out_destroy;
+
+       xfs_itrace_entry(ip);
 
        if ((ip->i_d.di_mode == 0) && !(flags & XFS_IGET_CREATE)) {
-               xfs_idestroy(ip);
-               xfs_put_perag(mp, pag);
-               return ENOENT;
+               error = ENOENT;
+               goto out_destroy;
        }
 
        /*
         * Preload the radix tree so we can insert safely under the
-        * write spinlock.
+        * write spinlock. Note that we cannot sleep inside the preload
+        * region.
         */
        if (radix_tree_preload(GFP_KERNEL)) {
-               xfs_idestroy(ip);
-               delay(1);
-               goto again;
+               error = EAGAIN;
+               goto out_destroy;
        }
 
-       if (lock_flags)
-               xfs_ilock(ip, lock_flags);
+       /*
+        * 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();
+       }
 
        mask = ~(((XFS_INODE_CLUSTER_SIZE(mp) >> mp->m_sb.sb_inodelog)) - 1);
        first_index = agino & mask;
        write_lock(&pag->pag_ici_lock);
-       /*
-        * insert the new inode
-        */
+
+       /* insert the new inode */
        error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
        if (unlikely(error)) {
-               BUG_ON(error != -EEXIST);
-               write_unlock(&pag->pag_ici_lock);
-               radix_tree_preload_end();
-               if (lock_flags)
-                       xfs_iunlock(ip, lock_flags);
-               xfs_idestroy(ip);
+               WARN_ON(error != -EEXIST);
                XFS_STATS_INC(xs_ig_dup);
-               goto again;
+               error = EAGAIN;
+               goto out_preload_end;
        }
 
-       /*
-        * These values _must_ be set before releasing the radix tree lock!
-        */
+       /* These values _must_ be set before releasing the radix tree lock! */
        ip->i_udquot = ip->i_gdquot = NULL;
        xfs_iflags_set(ip, XFS_INEW);
 
        write_unlock(&pag->pag_ici_lock);
        radix_tree_preload_end();
-       xfs_put_perag(mp, pag);
-
- return_ip:
-       ASSERT(ip->i_df.if_ext_max ==
-              XFS_IFORK_DSIZE(ip) / sizeof(xfs_bmbt_rec_t));
 
-       xfs_iflags_set(ip, XFS_IMODIFIED);
+       trace_xfs_iget_alloc(ip);
        *ipp = ip;
-
-       /*
-        * Set up the Linux with the Linux inode.
-        */
-       ip->i_vnode = inode;
-       inode->i_private = 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 (ip->i_d.di_mode != 0)
-               xfs_setup_inode(ip);
        return 0;
-}
 
+out_preload_end:
+       write_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;
+}
 
 /*
- * The 'normal' internal xfs_iget, if needed it will
- * 'allocate', or 'get', the vnode.
+ * 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.
+ * bno -- the block number starting the buffer containing the inode,
+ *       if known (as by bulkstat), else 0.
  */
 int
 xfs_iget(
@@ -293,84 +364,61 @@ xfs_iget(
        xfs_inode_t     **ipp,
        xfs_daddr_t     bno)
 {
-       struct inode    *inode;
        xfs_inode_t     *ip;
        int             error;
+       xfs_perag_t     *pag;
+       xfs_agino_t     agino;
 
-       XFS_STATS_INC(xs_ig_attempts);
+       /* the radix tree exists only in inode capable AGs */
+       if (XFS_INO_TO_AGNO(mp, ino) >= mp->m_maxagi)
+               return EINVAL;
 
-retry:
-       inode = iget_locked(mp->m_super, ino);
-       if (!inode)
-               /* If we got no inode we are out of memory */
-               return ENOMEM;
+       /* get the perag structure and ensure that it's inode capable */
+       pag = xfs_get_perag(mp, ino);
+       if (!pag->pagi_inodeok)
+               return EINVAL;
+       ASSERT(pag->pag_ici_init);
+       agino = XFS_INO_TO_AGINO(mp, ino);
 
-       if (inode->i_state & I_NEW) {
-               XFS_STATS_INC(vn_active);
-               XFS_STATS_INC(vn_alloc);
+again:
+       error = 0;
+       read_lock(&pag->pag_ici_lock);
+       ip = radix_tree_lookup(&pag->pag_ici_root, agino);
 
-               error = xfs_iget_core(inode, mp, tp, ino, flags,
-                               lock_flags, ipp, bno);
-               if (error) {
-                       make_bad_inode(inode);
-                       if (inode->i_state & I_NEW)
-                               unlock_new_inode(inode);
-                       iput(inode);
-               }
-               return error;
+       if (ip) {
+               error = xfs_iget_cache_hit(pag, ip, flags, lock_flags);
+               if (error)
+                       goto out_error_or_again;
+       } else {
+               read_unlock(&pag->pag_ici_lock);
+               XFS_STATS_INC(xs_ig_missed);
+
+               error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip, bno,
+                                                       flags, lock_flags);
+               if (error)
+                       goto out_error_or_again;
        }
+       xfs_put_perag(mp, pag);
+
+       *ipp = ip;
 
+       ASSERT(ip->i_df.if_ext_max ==
+              XFS_IFORK_DSIZE(ip) / sizeof(xfs_bmbt_rec_t));
        /*
-        * If the inode is not fully constructed due to
-        * filehandle mismatches wait for the inode to go
-        * away and try again.
-        *
-        * iget_locked will call __wait_on_freeing_inode
-        * to wait for the inode to go away.
+        * 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 (is_bad_inode(inode)) {
-               iput(inode);
-               delay(1);
-               goto retry;
-       }
+       if (xfs_iflags_test(ip, XFS_INEW) && ip->i_d.di_mode != 0)
+               xfs_setup_inode(ip);
+       return 0;
 
-       ip = XFS_I(inode);
-       if (!ip) {
-               iput(inode);
+out_error_or_again:
+       if (error == EAGAIN) {
                delay(1);
-               goto retry;
+               goto again;
        }
-
-       if (lock_flags != 0)
-               xfs_ilock(ip, lock_flags);
-       XFS_STATS_INC(xs_ig_found);
-       *ipp = ip;
-       return 0;
-}
-
-/*
- * Look for the inode corresponding to the given ino in the hash table.
- * If it is there and its i_transp pointer matches tp, return it.
- * Otherwise, return NULL.
- */
-xfs_inode_t *
-xfs_inode_incore(xfs_mount_t   *mp,
-                xfs_ino_t      ino,
-                xfs_trans_t    *tp)
-{
-       xfs_inode_t     *ip;
-       xfs_perag_t     *pag;
-
-       pag = xfs_get_perag(mp, ino);
-       read_lock(&pag->pag_ici_lock);
-       ip = radix_tree_lookup(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, ino));
-       read_unlock(&pag->pag_ici_lock);
        xfs_put_perag(mp, pag);
-
-       /* the returned inode must match the transaction */
-       if (ip && (ip->i_transp != tp))
-               return NULL;
-       return ip;
+       return error;
 }
 
 /*
@@ -413,78 +461,55 @@ xfs_iput_new(
        IRELE(ip);
 }
 
-
 /*
- * This routine embodies the part of the reclaim code that pulls
- * the inode from the inode hash table and the mount structure's
- * inode list.
- * This should only be called from xfs_reclaim().
+ * This is called free all the memory associated with an inode.
+ * It must free the inode itself and any buffers allocated for
+ * if_extents/if_data and if_broot.  It must also free the lock
+ * associated with the inode.
+ *
+ * Note: because we don't initialise everything on reallocation out
+ * of the zone, we must ensure we nullify everything correctly before
+ * freeing the structure.
  */
 void
-xfs_ireclaim(xfs_inode_t *ip)
+xfs_ireclaim(
+       struct xfs_inode        *ip)
 {
-       /*
-        * Remove from old hash list and mount list.
-        */
-       XFS_STATS_INC(xs_ig_reclaims);
-
-       xfs_iextract(ip);
-
-       /*
-        * Here we do a spurious inode lock in order to coordinate with
-        * xfs_sync().  This is because xfs_sync() references the inodes
-        * in the mount list without taking references on the corresponding
-        * vnodes.  We make that OK here by ensuring that we wait until
-        * the inode is unlocked in xfs_sync() before we go ahead and
-        * free it.  We get both the regular lock and the io lock because
-        * the xfs_sync() code may need to drop the regular one but will
-        * still hold the io lock.
-        */
-       xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
+       struct xfs_mount        *mp = ip->i_mount;
+       struct xfs_perag        *pag;
+       xfs_agino_t             agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
 
-       /*
-        * Release dquots (and their references) if any. An inode may escape
-        * xfs_inactive and get here via vn_alloc->vn_reclaim path.
-        */
-       XFS_QM_DQDETACH(ip->i_mount, ip);
+       XFS_STATS_INC(xs_ig_reclaims);
 
        /*
-        * Pull our behavior descriptor from the vnode chain.
+        * Remove the inode from the per-AG radix tree.
+        *
+        * Because radix_tree_delete won't complain even if the item was never
+        * added to the tree assert that it's been there before to catch
+        * problems with the inode life time early on.
         */
-       if (ip->i_vnode) {
-               ip->i_vnode->i_private = NULL;
-               ip->i_vnode = NULL;
-       }
+       pag = xfs_get_perag(mp, ip->i_ino);
+       write_lock(&pag->pag_ici_lock);
+       if (!radix_tree_delete(&pag->pag_ici_root, agino))
+               ASSERT(0);
+       write_unlock(&pag->pag_ici_lock);
+       xfs_put_perag(mp, pag);
 
        /*
-        * Free all memory associated with the inode.
+        * Here we do an (almost) spurious inode lock in order to coordinate
+        * with inode cache radix tree lookups.  This is because the lookup
+        * can reference the inodes in the cache without taking references.
+        *
+        * We make that OK here by ensuring that we wait until the inode is
+        * unlocked after the lookup before we go ahead and free it.  We get
+        * both the ilock and the iolock because the code may need to drop the
+        * ilock one but will still hold the iolock.
         */
+       xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
+       xfs_qm_dqdetach(ip);
        xfs_iunlock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
-       xfs_idestroy(ip);
-}
 
-/*
- * This routine removes an about-to-be-destroyed inode from
- * all of the lists in which it is located with the exception
- * of the behavior chain.
- */
-void
-xfs_iextract(
-       xfs_inode_t     *ip)
-{
-       xfs_mount_t     *mp = ip->i_mount;
-       xfs_perag_t     *pag = xfs_get_perag(mp, ip->i_ino);
-
-       write_lock(&pag->pag_ici_lock);
-       radix_tree_delete(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, ip->i_ino));
-       write_unlock(&pag->pag_ici_lock);
-       xfs_put_perag(mp, pag);
-
-       /* Deal with the deleted inodes list */
-       XFS_MOUNT_ILOCK(mp);
-       list_del_init(&ip->i_reclaim);
-       mp->m_ireclaims++;
-       XFS_MOUNT_IUNLOCK(mp);
+       xfs_inode_free(ip);
 }
 
 /*
@@ -581,7 +606,7 @@ xfs_ilock(
        else if (lock_flags & XFS_ILOCK_SHARED)
                mraccess_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
 
-       xfs_ilock_trace(ip, 1, lock_flags, (inst_t *)__return_address);
+       trace_xfs_ilock(ip, lock_flags, _RET_IP_);
 }
 
 /*
@@ -626,7 +651,7 @@ xfs_ilock_nowait(
                if (!mrtryaccess(&ip->i_lock))
                        goto out_undo_iolock;
        }
-       xfs_ilock_trace(ip, 2, lock_flags, (inst_t *)__return_address);
+       trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
        return 1;
 
  out_undo_iolock:
@@ -685,10 +710,10 @@ xfs_iunlock(
                 * it is in the AIL and anyone is waiting on it.  Don't do
                 * this if the caller has asked us not to.
                 */
-               xfs_trans_unlocked_item(ip->i_mount,
+               xfs_trans_unlocked_item(ip->i_itemp->ili_item.li_ailp,
                                        (xfs_log_item_t*)(ip->i_itemp));
        }
-       xfs_ilock_trace(ip, 3, lock_flags, (inst_t *)__return_address);
+       trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
 }
 
 /*
@@ -707,6 +732,8 @@ xfs_ilock_demote(
                mrdemote(&ip->i_lock);
        if (lock_flags & XFS_IOLOCK_EXCL)
                mrdemote(&ip->i_iolock);
+
+       trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
 }
 
 #ifdef DEBUG
@@ -737,4 +764,3 @@ xfs_isilocked(
        return 1;
 }
 #endif
-