nfsd: serialize state seqid morphing operations
authorJeff Layton <jlayton@poochiereds.net>
Thu, 17 Sep 2015 11:47:08 +0000 (07:47 -0400)
committerJ. Bruce Fields <bfields@redhat.com>
Mon, 12 Oct 2015 21:31:03 +0000 (17:31 -0400)
Andrew was seeing a race occur when an OPEN and OPEN_DOWNGRADE were
running in parallel. The server would receive the OPEN_DOWNGRADE first
and check its seqid, but then an OPEN would race in and bump it. The
OPEN_DOWNGRADE would then complete and bump the seqid again.  The result
was that the OPEN_DOWNGRADE would be applied after the OPEN, even though
it should have been rejected since the seqid changed.

The only recourse we have here I think is to serialize operations that
bump the seqid in a stateid, particularly when we're given a seqid in
the call. To address this, we add a new rw_semaphore to the
nfs4_ol_stateid struct. We do a down_write prior to checking the seqid
after looking up the stateid to ensure that nothing else is going to
bump it while we're operating on it.

In the case of OPEN, we do a down_read, as the call doesn't contain a
seqid. Those can run in parallel -- we just need to serialize them when
there is a concurrent OPEN_DOWNGRADE or CLOSE.

LOCK and LOCKU however always take the write lock as there is no
opportunity for parallelizing those.

Reported-and-Tested-by: Andrew W Elble <aweits@rit.edu>
Signed-off-by: Jeff Layton <jeff.layton@primarydata.com>
Cc: stable@vger.kernel.org
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
fs/nfsd/nfs4state.c
fs/nfsd/state.h

index 0f1d5691b795751553d02003c32c4ceb90a1ee5c..1b39edf10b67da594f6a9d8deea8ba29e49794d2 100644 (file)
@@ -3360,6 +3360,7 @@ static void init_open_stateid(struct nfs4_ol_stateid *stp, struct nfs4_file *fp,
        stp->st_access_bmap = 0;
        stp->st_deny_bmap = 0;
        stp->st_openstp = NULL;
+       init_rwsem(&stp->st_rwsem);
        spin_lock(&oo->oo_owner.so_client->cl_lock);
        list_add(&stp->st_perstateowner, &oo->oo_owner.so_stateids);
        spin_lock(&fp->fi_lock);
@@ -4187,15 +4188,20 @@ nfsd4_process_open2(struct svc_rqst *rqstp, struct svc_fh *current_fh, struct nf
         */
        if (stp) {
                /* Stateid was found, this is an OPEN upgrade */
+               down_read(&stp->st_rwsem);
                status = nfs4_upgrade_open(rqstp, fp, current_fh, stp, open);
-               if (status)
+               if (status) {
+                       up_read(&stp->st_rwsem);
                        goto out;
+               }
        } else {
                stp = open->op_stp;
                open->op_stp = NULL;
                init_open_stateid(stp, fp, open);
+               down_read(&stp->st_rwsem);
                status = nfs4_get_vfs_file(rqstp, fp, current_fh, stp, open);
                if (status) {
+                       up_read(&stp->st_rwsem);
                        release_open_stateid(stp);
                        goto out;
                }
@@ -4207,6 +4213,7 @@ nfsd4_process_open2(struct svc_rqst *rqstp, struct svc_fh *current_fh, struct nf
        }
        update_stateid(&stp->st_stid.sc_stateid);
        memcpy(&open->op_stateid, &stp->st_stid.sc_stateid, sizeof(stateid_t));
+       up_read(&stp->st_rwsem);
 
        if (nfsd4_has_session(&resp->cstate)) {
                if (open->op_deleg_want & NFS4_SHARE_WANT_NO_DELEG) {
@@ -4819,10 +4826,13 @@ static __be32 nfs4_seqid_op_checks(struct nfsd4_compound_state *cstate, stateid_
                 * revoked delegations are kept only for free_stateid.
                 */
                return nfserr_bad_stateid;
+       down_write(&stp->st_rwsem);
        status = check_stateid_generation(stateid, &stp->st_stid.sc_stateid, nfsd4_has_session(cstate));
-       if (status)
-               return status;
-       return nfs4_check_fh(current_fh, &stp->st_stid);
+       if (status == nfs_ok)
+               status = nfs4_check_fh(current_fh, &stp->st_stid);
+       if (status != nfs_ok)
+               up_write(&stp->st_rwsem);
+       return status;
 }
 
 /* 
@@ -4869,6 +4879,7 @@ static __be32 nfs4_preprocess_confirmed_seqid_op(struct nfsd4_compound_state *cs
                return status;
        oo = openowner(stp->st_stateowner);
        if (!(oo->oo_flags & NFS4_OO_CONFIRMED)) {
+               up_write(&stp->st_rwsem);
                nfs4_put_stid(&stp->st_stid);
                return nfserr_bad_stateid;
        }
@@ -4899,11 +4910,14 @@ nfsd4_open_confirm(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
                goto out;
        oo = openowner(stp->st_stateowner);
        status = nfserr_bad_stateid;
-       if (oo->oo_flags & NFS4_OO_CONFIRMED)
+       if (oo->oo_flags & NFS4_OO_CONFIRMED) {
+               up_write(&stp->st_rwsem);
                goto put_stateid;
+       }
        oo->oo_flags |= NFS4_OO_CONFIRMED;
        update_stateid(&stp->st_stid.sc_stateid);
        memcpy(&oc->oc_resp_stateid, &stp->st_stid.sc_stateid, sizeof(stateid_t));
+       up_write(&stp->st_rwsem);
        dprintk("NFSD: %s: success, seqid=%d stateid=" STATEID_FMT "\n",
                __func__, oc->oc_seqid, STATEID_VAL(&stp->st_stid.sc_stateid));
 
@@ -4982,6 +4996,7 @@ nfsd4_open_downgrade(struct svc_rqst *rqstp,
        memcpy(&od->od_stateid, &stp->st_stid.sc_stateid, sizeof(stateid_t));
        status = nfs_ok;
 put_stateid:
+       up_write(&stp->st_rwsem);
        nfs4_put_stid(&stp->st_stid);
 out:
        nfsd4_bump_seqid(cstate, status);
@@ -5035,6 +5050,7 @@ nfsd4_close(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
                goto out; 
        update_stateid(&stp->st_stid.sc_stateid);
        memcpy(&close->cl_stateid, &stp->st_stid.sc_stateid, sizeof(stateid_t));
+       up_write(&stp->st_rwsem);
 
        nfsd4_close_open_stateid(stp);
 
@@ -5260,6 +5276,7 @@ init_lock_stateid(struct nfs4_ol_stateid *stp, struct nfs4_lockowner *lo,
        stp->st_access_bmap = 0;
        stp->st_deny_bmap = open_stp->st_deny_bmap;
        stp->st_openstp = open_stp;
+       init_rwsem(&stp->st_rwsem);
        list_add(&stp->st_locks, &open_stp->st_locks);
        list_add(&stp->st_perstateowner, &lo->lo_owner.so_stateids);
        spin_lock(&fp->fi_lock);
@@ -5428,6 +5445,7 @@ nfsd4_lock(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
                                        &open_stp, nn);
                if (status)
                        goto out;
+               up_write(&open_stp->st_rwsem);
                open_sop = openowner(open_stp->st_stateowner);
                status = nfserr_bad_stateid;
                if (!same_clid(&open_sop->oo_owner.so_client->cl_clientid,
@@ -5435,6 +5453,8 @@ nfsd4_lock(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
                        goto out;
                status = lookup_or_create_lock_state(cstate, open_stp, lock,
                                                        &lock_stp, &new);
+               if (status == nfs_ok)
+                       down_write(&lock_stp->st_rwsem);
        } else {
                status = nfs4_preprocess_seqid_op(cstate,
                                       lock->lk_old_lock_seqid,
@@ -5540,6 +5560,8 @@ out:
                    seqid_mutating_err(ntohl(status)))
                        lock_sop->lo_owner.so_seqid++;
 
+               up_write(&lock_stp->st_rwsem);
+
                /*
                 * If this is a new, never-before-used stateid, and we are
                 * returning an error, then just go ahead and release it.
@@ -5709,6 +5731,7 @@ nfsd4_locku(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
 fput:
        fput(filp);
 put_stateid:
+       up_write(&stp->st_rwsem);
        nfs4_put_stid(&stp->st_stid);
 out:
        nfsd4_bump_seqid(cstate, status);
index 583ffc13cae27d165be052609f37dc7428b54dcc..31bde12feefeec8434ef0aa25d24d318c9a4f938 100644 (file)
@@ -534,15 +534,16 @@ struct nfs4_file {
  * Better suggestions welcome.
  */
 struct nfs4_ol_stateid {
-       struct nfs4_stid    st_stid; /* must be first field */
-       struct list_head              st_perfile;
-       struct list_head              st_perstateowner;
-       struct list_head              st_locks;
-       struct nfs4_stateowner      * st_stateowner;
-       struct nfs4_clnt_odstate    * st_clnt_odstate;
-       unsigned char                 st_access_bmap;
-       unsigned char                 st_deny_bmap;
-       struct nfs4_ol_stateid         * st_openstp;
+       struct nfs4_stid                st_stid;
+       struct list_head                st_perfile;
+       struct list_head                st_perstateowner;
+       struct list_head                st_locks;
+       struct nfs4_stateowner          *st_stateowner;
+       struct nfs4_clnt_odstate        *st_clnt_odstate;
+       unsigned char                   st_access_bmap;
+       unsigned char                   st_deny_bmap;
+       struct nfs4_ol_stateid          *st_openstp;
+       struct rw_semaphore             st_rwsem;
 };
 
 static inline struct nfs4_ol_stateid *openlockstateid(struct nfs4_stid *s)