]> git.kernelconcepts.de Git - karo-tx-linux.git/commitdiff
nfs4: treat lock owners as opaque values
authorJeff Layton <jlayton@poochiereds.net>
Thu, 1 May 2014 10:28:45 +0000 (06:28 -0400)
committerTrond Myklebust <trond.myklebust@primarydata.com>
Sat, 12 Jul 2014 22:36:31 +0000 (18:36 -0400)
Do the following set of ops with a file on a NFSv4 mount:

    exec 3>>/file/on/nfsv4
    flock -x 3
    exec 3>&-

You'll see the LOCK request go across the wire, but no LOCKU when the
file is closed.

What happens is that the fd is passed across a fork, and the final close
is done in a different process than the opener. That makes
__nfs4_find_lock_state miss finding the correct lock state because it
uses the fl_pid as a search key. A new one is created, and the locking
code treats it as a delegation stateid (because NFS_LOCK_INITIALIZED
isn't set).

The root cause of this breakage seems to be commit 77041ed9b49a9e
(NFSv4: Ensure the lockowners are labelled using the fl_owner and/or
fl_pid).

That changed it so that flock lockowners are allocated based on the
fl_pid. I think this is incorrect. flock locks should be "owned" by the
struct file, and that is already accounted for in the fl_owner field of
the lock request when it comes through nfs_flock.

This patch basically reverts the above commit and with it, a LOCKU is
sent in the above reproducer.

Signed-off-by: Jeff Layton <jlayton@poochiereds.net>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
fs/nfs/nfs4_fs.h
fs/nfs/nfs4state.c

index b8ea4a26998c875fac678b1cf34ebbd5db38f9d2..19f567c3967096874f875ee78def1b952426e688 100644 (file)
@@ -129,17 +129,6 @@ enum {
  * LOCK: one nfs4_state (LOCK) to hold the lock stateid nfs4_state(OPEN)
  */
 
-struct nfs4_lock_owner {
-       unsigned int lo_type;
-#define NFS4_ANY_LOCK_TYPE     (0U)
-#define NFS4_FLOCK_LOCK_TYPE   (1U << 0)
-#define NFS4_POSIX_LOCK_TYPE   (1U << 1)
-       union {
-               fl_owner_t posix_owner;
-               pid_t flock_owner;
-       } lo_u;
-};
-
 struct nfs4_lock_state {
        struct list_head        ls_locks;       /* Other lock stateids */
        struct nfs4_state *     ls_state;       /* Pointer to open state */
@@ -149,7 +138,7 @@ struct nfs4_lock_state {
        struct nfs_seqid_counter        ls_seqid;
        nfs4_stateid            ls_stateid;
        atomic_t                ls_count;
-       struct nfs4_lock_owner  ls_owner;
+       fl_owner_t              ls_owner;
 };
 
 /* bits for nfs4_state->flags */
index 848f6853c59e49c8b3b570ddc99626bfab935212..544040835482bf2d00a855d9a124abc6e10db245 100644 (file)
@@ -787,21 +787,12 @@ void nfs4_close_sync(struct nfs4_state *state, fmode_t fmode)
  * that is compatible with current->files
  */
 static struct nfs4_lock_state *
-__nfs4_find_lock_state(struct nfs4_state *state, fl_owner_t fl_owner, pid_t fl_pid, unsigned int type)
+__nfs4_find_lock_state(struct nfs4_state *state, fl_owner_t fl_owner)
 {
        struct nfs4_lock_state *pos;
        list_for_each_entry(pos, &state->lock_states, ls_locks) {
-               if (type != NFS4_ANY_LOCK_TYPE && pos->ls_owner.lo_type != type)
+               if (pos->ls_owner != fl_owner)
                        continue;
-               switch (pos->ls_owner.lo_type) {
-               case NFS4_POSIX_LOCK_TYPE:
-                       if (pos->ls_owner.lo_u.posix_owner != fl_owner)
-                               continue;
-                       break;
-               case NFS4_FLOCK_LOCK_TYPE:
-                       if (pos->ls_owner.lo_u.flock_owner != fl_pid)
-                               continue;
-               }
                atomic_inc(&pos->ls_count);
                return pos;
        }
@@ -813,7 +804,7 @@ __nfs4_find_lock_state(struct nfs4_state *state, fl_owner_t fl_owner, pid_t fl_p
  * exists, return an uninitialized one.
  *
  */
-static struct nfs4_lock_state *nfs4_alloc_lock_state(struct nfs4_state *state, fl_owner_t fl_owner, pid_t fl_pid, unsigned int type)
+static struct nfs4_lock_state *nfs4_alloc_lock_state(struct nfs4_state *state, fl_owner_t fl_owner)
 {
        struct nfs4_lock_state *lsp;
        struct nfs_server *server = state->owner->so_server;
@@ -824,17 +815,7 @@ static struct nfs4_lock_state *nfs4_alloc_lock_state(struct nfs4_state *state, f
        nfs4_init_seqid_counter(&lsp->ls_seqid);
        atomic_set(&lsp->ls_count, 1);
        lsp->ls_state = state;
-       lsp->ls_owner.lo_type = type;
-       switch (lsp->ls_owner.lo_type) {
-       case NFS4_FLOCK_LOCK_TYPE:
-               lsp->ls_owner.lo_u.flock_owner = fl_pid;
-               break;
-       case NFS4_POSIX_LOCK_TYPE:
-               lsp->ls_owner.lo_u.posix_owner = fl_owner;
-               break;
-       default:
-               goto out_free;
-       }
+       lsp->ls_owner = fl_owner;
        lsp->ls_seqid.owner_id = ida_simple_get(&server->lockowner_id, 0, 0, GFP_NOFS);
        if (lsp->ls_seqid.owner_id < 0)
                goto out_free;
@@ -857,13 +838,13 @@ void nfs4_free_lock_state(struct nfs_server *server, struct nfs4_lock_state *lsp
  * exists, return an uninitialized one.
  *
  */
-static struct nfs4_lock_state *nfs4_get_lock_state(struct nfs4_state *state, fl_owner_t owner, pid_t pid, unsigned int type)
+static struct nfs4_lock_state *nfs4_get_lock_state(struct nfs4_state *state, fl_owner_t owner)
 {
        struct nfs4_lock_state *lsp, *new = NULL;
        
        for(;;) {
                spin_lock(&state->state_lock);
-               lsp = __nfs4_find_lock_state(state, owner, pid, type);
+               lsp = __nfs4_find_lock_state(state, owner);
                if (lsp != NULL)
                        break;
                if (new != NULL) {
@@ -874,7 +855,7 @@ static struct nfs4_lock_state *nfs4_get_lock_state(struct nfs4_state *state, fl_
                        break;
                }
                spin_unlock(&state->state_lock);
-               new = nfs4_alloc_lock_state(state, owner, pid, type);
+               new = nfs4_alloc_lock_state(state, owner);
                if (new == NULL)
                        return NULL;
        }
@@ -935,13 +916,7 @@ int nfs4_set_lock_state(struct nfs4_state *state, struct file_lock *fl)
 
        if (fl->fl_ops != NULL)
                return 0;
-       if (fl->fl_flags & FL_POSIX)
-               lsp = nfs4_get_lock_state(state, fl->fl_owner, 0, NFS4_POSIX_LOCK_TYPE);
-       else if (fl->fl_flags & FL_FLOCK)
-               lsp = nfs4_get_lock_state(state, NULL, fl->fl_pid,
-                               NFS4_FLOCK_LOCK_TYPE);
-       else
-               return -EINVAL;
+       lsp = nfs4_get_lock_state(state, fl->fl_owner);
        if (lsp == NULL)
                return -ENOMEM;
        fl->fl_u.nfs4_fl.owner = lsp;
@@ -955,7 +930,6 @@ static int nfs4_copy_lock_stateid(nfs4_stateid *dst,
 {
        struct nfs4_lock_state *lsp;
        fl_owner_t fl_owner;
-       pid_t fl_pid;
        int ret = -ENOENT;
 
 
@@ -966,9 +940,8 @@ static int nfs4_copy_lock_stateid(nfs4_stateid *dst,
                goto out;
 
        fl_owner = lockowner->l_owner;
-       fl_pid = lockowner->l_pid;
        spin_lock(&state->state_lock);
-       lsp = __nfs4_find_lock_state(state, fl_owner, fl_pid, NFS4_ANY_LOCK_TYPE);
+       lsp = __nfs4_find_lock_state(state, fl_owner);
        if (lsp && test_bit(NFS_LOCK_LOST, &lsp->ls_flags))
                ret = -EIO;
        else if (lsp != NULL && test_bit(NFS_LOCK_INITIALIZED, &lsp->ls_flags) != 0) {