nfsd: replace rp_mutex to avoid deadlock in move_to_close_lru()
authorNeilBrown <neilb@suse.de>
Mon, 8 Apr 2024 02:09:17 +0000 (12:09 +1000)
committerChuck Lever <chuck.lever@oracle.com>
Mon, 6 May 2024 13:07:16 +0000 (09:07 -0400)
move_to_close_lru() waits for sc_count to become zero while holding
rp_mutex.  This can deadlock if another thread holds a reference and is
waiting for rp_mutex.

By the time we get to move_to_close_lru() the openowner is unhashed and
cannot be found any more.  So code waiting for the mutex can safely
retry the lookup if move_to_close_lru() has started.

So change rp_mutex to an atomic_t with three states:

 RP_UNLOCK   - state is still hashed, not locked for reply
 RP_LOCKED   - state is still hashed, is locked for reply
 RP_UNHASHED - state is not hashed, no code can get a lock.

Use wait_var_event() to wait for either a lock, or for the owner to be
unhashed.  In the latter case, retry the lookup.

Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
fs/nfsd/nfs4state.c
fs/nfsd/state.h

index 43bdff49aa03b7ea1af040076a6d29650a1c7f98..e91e741719a0234b45c23fb94018fae9c939f11a 100644 (file)
@@ -4650,21 +4650,32 @@ nfsd4_init_leases_net(struct nfsd_net *nn)
        atomic_set(&nn->nfsd_courtesy_clients, 0);
 }
 
+enum rp_lock {
+       RP_UNLOCKED,
+       RP_LOCKED,
+       RP_UNHASHED,
+};
+
 static void init_nfs4_replay(struct nfs4_replay *rp)
 {
        rp->rp_status = nfserr_serverfault;
        rp->rp_buflen = 0;
        rp->rp_buf = rp->rp_ibuf;
-       mutex_init(&rp->rp_mutex);
+       atomic_set(&rp->rp_locked, RP_UNLOCKED);
 }
 
-static void nfsd4_cstate_assign_replay(struct nfsd4_compound_state *cstate,
-               struct nfs4_stateowner *so)
+static int nfsd4_cstate_assign_replay(struct nfsd4_compound_state *cstate,
+                                     struct nfs4_stateowner *so)
 {
        if (!nfsd4_has_session(cstate)) {
-               mutex_lock(&so->so_replay.rp_mutex);
+               wait_var_event(&so->so_replay.rp_locked,
+                              atomic_cmpxchg(&so->so_replay.rp_locked,
+                                             RP_UNLOCKED, RP_LOCKED) != RP_LOCKED);
+               if (atomic_read(&so->so_replay.rp_locked) == RP_UNHASHED)
+                       return -EAGAIN;
                cstate->replay_owner = nfs4_get_stateowner(so);
        }
+       return 0;
 }
 
 void nfsd4_cstate_clear_replay(struct nfsd4_compound_state *cstate)
@@ -4673,7 +4684,8 @@ void nfsd4_cstate_clear_replay(struct nfsd4_compound_state *cstate)
 
        if (so != NULL) {
                cstate->replay_owner = NULL;
-               mutex_unlock(&so->so_replay.rp_mutex);
+               atomic_set(&so->so_replay.rp_locked, RP_UNLOCKED);
+               wake_up_var(&so->so_replay.rp_locked);
                nfs4_put_stateowner(so);
        }
 }
@@ -4969,7 +4981,11 @@ move_to_close_lru(struct nfs4_ol_stateid *s, struct net *net)
         * Wait for the refcount to drop to 2. Since it has been unhashed,
         * there should be no danger of the refcount going back up again at
         * this point.
+        * Some threads with a reference might be waiting for rp_locked,
+        * so tell them to stop waiting.
         */
+       atomic_set(&oo->oo_owner.so_replay.rp_locked, RP_UNHASHED);
+       wake_up_var(&oo->oo_owner.so_replay.rp_locked);
        wait_event(close_wq, refcount_read(&s->st_stid.sc_count) == 2);
 
        release_all_access(s);
@@ -5342,11 +5358,15 @@ nfsd4_process_open1(struct nfsd4_compound_state *cstate,
        clp = cstate->clp;
 
        strhashval = ownerstr_hashval(&open->op_owner);
+retry:
        oo = find_or_alloc_open_stateowner(strhashval, open, cstate);
        open->op_openowner = oo;
        if (!oo)
                return nfserr_jukebox;
-       nfsd4_cstate_assign_replay(cstate, &oo->oo_owner);
+       if (nfsd4_cstate_assign_replay(cstate, &oo->oo_owner) == -EAGAIN) {
+               nfs4_put_stateowner(&oo->oo_owner);
+               goto retry;
+       }
        status = nfsd4_check_seqid(cstate, &oo->oo_owner, open->op_seqid);
        if (status)
                return status;
@@ -7186,12 +7206,16 @@ nfs4_preprocess_seqid_op(struct nfsd4_compound_state *cstate, u32 seqid,
        trace_nfsd_preprocess(seqid, stateid);
 
        *stpp = NULL;
+retry:
        status = nfsd4_lookup_stateid(cstate, stateid,
                                      typemask, statusmask, &s, nn);
        if (status)
                return status;
        stp = openlockstateid(s);
-       nfsd4_cstate_assign_replay(cstate, stp->st_stateowner);
+       if (nfsd4_cstate_assign_replay(cstate, stp->st_stateowner) == -EAGAIN) {
+               nfs4_put_stateowner(stp->st_stateowner);
+               goto retry;
+       }
 
        status = nfs4_seqid_op_checks(cstate, stateid, seqid, stp);
        if (!status)
index 2ed0fcf879fd17be57d1d963ad74e864d73315f5..a6261754deede4eb97f228c0c79aa6b906b60abc 100644 (file)
@@ -486,7 +486,7 @@ struct nfs4_replay {
        unsigned int            rp_buflen;
        char                    *rp_buf;
        struct knfsd_fh         rp_openfh;
-       struct mutex            rp_mutex;
+       atomic_t                rp_locked;
        char                    rp_ibuf[NFSD4_REPLAY_ISIZE];
 };