pNFS: Fix missing layoutreturn calls
authorTrond Myklebust <trond.myklebust@primarydata.com>
Wed, 27 Jan 2016 04:12:11 +0000 (23:12 -0500)
committerTrond Myklebust <trond.myklebust@primarydata.com>
Wed, 27 Jan 2016 04:12:11 +0000 (23:12 -0500)
The layoutreturn code currently relies on pnfs_put_lseg() to initiate the
RPC call when conditions are right. A problem arises when we want to
free the layout segment from inside an inode->i_lock section (e.g. in
pnfs_clear_request_commit()), since we cannot sleep.

The workaround is to move the actual call to pnfs_send_layoutreturn()
to pnfs_put_layout_hdr(), which doesn't have this restriction.

Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
fs/nfs/pnfs.c

index a3592cc34a20b76341c9c9b9af3378bdddebd477..9c20685d496186298ad7c92acff03bb82ec24c55 100644 (file)
@@ -52,9 +52,7 @@ static DEFINE_SPINLOCK(pnfs_spinlock);
  */
 static LIST_HEAD(pnfs_modules_tbl);
 
-static int
-pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
-                      enum pnfs_iomode iomode, bool sync);
+static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
 
 /* Return the registered pnfs layout driver module matching given id */
 static struct pnfs_layoutdriver_type *
@@ -243,6 +241,8 @@ pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 {
        struct inode *inode = lo->plh_inode;
 
+       pnfs_layoutreturn_before_put_layout_hdr(lo);
+
        if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
                if (!list_empty(&lo->plh_segs))
                        WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
@@ -345,58 +345,6 @@ pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
        rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 }
 
-/* Return true if layoutreturn is needed */
-static bool
-pnfs_layout_need_return(struct pnfs_layout_hdr *lo,
-                       struct pnfs_layout_segment *lseg)
-{
-       struct pnfs_layout_segment *s;
-
-       if (!test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
-               return false;
-
-       list_for_each_entry(s, &lo->plh_segs, pls_list)
-               if (s != lseg && test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
-                       return false;
-
-       return true;
-}
-
-static bool
-pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo)
-{
-       if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
-               return false;
-       lo->plh_return_iomode = 0;
-       pnfs_get_layout_hdr(lo);
-       clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE, &lo->plh_flags);
-       return true;
-}
-
-static void pnfs_layoutreturn_before_put_lseg(struct pnfs_layout_segment *lseg,
-               struct pnfs_layout_hdr *lo, struct inode *inode)
-{
-       lo = lseg->pls_layout;
-       inode = lo->plh_inode;
-
-       spin_lock(&inode->i_lock);
-       if (pnfs_layout_need_return(lo, lseg)) {
-               nfs4_stateid stateid;
-               enum pnfs_iomode iomode;
-               bool send;
-
-               nfs4_stateid_copy(&stateid, &lo->plh_stateid);
-               iomode = lo->plh_return_iomode;
-               send = pnfs_prepare_layoutreturn(lo);
-               spin_unlock(&inode->i_lock);
-               if (send) {
-                       /* Send an async layoutreturn so we dont deadlock */
-                       pnfs_send_layoutreturn(lo, &stateid, iomode, false);
-               }
-       } else
-               spin_unlock(&inode->i_lock);
-}
-
 void
 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 {
@@ -410,15 +358,8 @@ pnfs_put_lseg(struct pnfs_layout_segment *lseg)
                atomic_read(&lseg->pls_refcount),
                test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 
-       /* Handle the case where refcount != 1 */
-       if (atomic_add_unless(&lseg->pls_refcount, -1, 1))
-               return;
-
        lo = lseg->pls_layout;
        inode = lo->plh_inode;
-       /* Do we need a layoutreturn? */
-       if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
-               pnfs_layoutreturn_before_put_lseg(lseg, lo, inode);
 
        if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
                if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
@@ -937,6 +878,17 @@ void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
        rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 }
 
+static bool
+pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo)
+{
+       if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
+               return false;
+       lo->plh_return_iomode = 0;
+       pnfs_get_layout_hdr(lo);
+       clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE, &lo->plh_flags);
+       return true;
+}
+
 static int
 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
                       enum pnfs_iomode iomode, bool sync)
@@ -971,6 +923,48 @@ out:
        return status;
 }
 
+/* Return true if layoutreturn is needed */
+static bool
+pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
+{
+       struct pnfs_layout_segment *s;
+
+       if (!test_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE, &lo->plh_flags))
+               return false;
+
+       /* Defer layoutreturn until all lsegs are done */
+       list_for_each_entry(s, &lo->plh_segs, pls_list) {
+               if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
+                       return false;
+       }
+
+       return true;
+}
+
+static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
+{
+       struct inode *inode= lo->plh_inode;
+
+       if (!test_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE, &lo->plh_flags))
+               return;
+       spin_lock(&inode->i_lock);
+       if (pnfs_layout_need_return(lo)) {
+               nfs4_stateid stateid;
+               enum pnfs_iomode iomode;
+               bool send;
+
+               nfs4_stateid_copy(&stateid, &lo->plh_stateid);
+               iomode = lo->plh_return_iomode;
+               send = pnfs_prepare_layoutreturn(lo);
+               spin_unlock(&inode->i_lock);
+               if (send) {
+                       /* Send an async layoutreturn so we dont deadlock */
+                       pnfs_send_layoutreturn(lo, &stateid, iomode, false);
+               }
+       } else
+               spin_unlock(&inode->i_lock);
+}
+
 /*
  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
  * when the layout segment list is empty.