nfs: prevent backlogging of renewd requests

If the renewd send queue gets backlogged (e.g., if the server goes down),
we will keep filling the queue with periodic RENEW/SEQUENCE requests.

This patch schedules a new renewd request if and only if the previous one
returns (either success or failure)

Signed-off-by: Alexandros Batsakis <batsakis@netapp.com>
[Trond.Myklebust@netapp.com: moved nfs4_schedule_state_renewal() into
separate nfs4_renew_release() and nfs41_sequence_release() callbacks
to ensure correct behaviour on call setup failure]
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>

2 files changed, 27 insertions, 21 deletions

diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c
index 84b53d38f50b..726bc195039d 100644
--- a/fs/nfs/nfs4proc.c
+++ b/fs/nfs/nfs4proc.c
@@ -3147,10 +3147,17 @@ static void nfs4_proc_commit_setup(struct nfs_write_data *data, struct rpc_messa
  * nfs4_proc_async_renew(): This is not one of the nfs_rpc_ops; it is a special
  * standalone procedure for queueing an asynchronous RENEW.
  */
+static void nfs4_renew_release(void *data)
+{
+        struct nfs_client *clp = data;
+
+        nfs4_schedule_state_renewal(clp);
+}
+
 static void nfs4_renew_done(struct rpc_task *task, void *data)
 {
-        struct nfs_client *clp = (struct nfs_client *)task->tk_msg.rpc_argp;
-        unsigned long timestamp = (unsigned long)data;
+        struct nfs_client *clp = data;
+        unsigned long timestamp = task->tk_start;
 
         if (task->tk_status < 0) {
                 /* Unless we're shutting down, schedule state recovery! */
@@ -3166,6 +3173,7 @@ static void nfs4_renew_done(struct rpc_task *task, void *data)
 
 static const struct rpc_call_ops nfs4_renew_ops = {
         .rpc_call_done = nfs4_renew_done,
+        .rpc_release = nfs4_renew_release,
 };
 
 int nfs4_proc_async_renew(struct nfs_client *clp, struct rpc_cred *cred)
@@ -3177,7 +3185,7 @@ int nfs4_proc_async_renew(struct nfs_client *clp, struct rpc_cred *cred)
         };
 
         return rpc_call_async(clp->cl_rpcclient, &msg, RPC_TASK_SOFT,
-                        &nfs4_renew_ops, (void *)jiffies);
+                        &nfs4_renew_ops, clp);
 }
 
 int nfs4_proc_renew(struct nfs_client *clp, struct rpc_cred *cred)
@@ -5023,7 +5031,14 @@ static int nfs4_proc_sequence(struct nfs_client *clp, struct rpc_cred *cred)
                                        &res, args.sa_cache_this, 1);
 }
 
-void nfs41_sequence_call_done(struct rpc_task *task, void *data)
+static void nfs41_sequence_release(void *data)
+{
+        struct nfs_client *clp = (struct nfs_client *)data;
+
+        nfs4_schedule_state_renewal(clp);
+}
+
+static void nfs41_sequence_call_done(struct rpc_task *task, void *data)
 {
         struct nfs_client *clp = (struct nfs_client *)data;
 
@@ -5064,6 +5079,7 @@ static void nfs41_sequence_prepare(struct rpc_task *task, void *data)
 static const struct rpc_call_ops nfs41_sequence_ops = {
         .rpc_call_done = nfs41_sequence_call_done,
         .rpc_call_prepare = nfs41_sequence_prepare,
+        .rpc_release = nfs41_sequence_release,
 };
 
 static int nfs41_proc_async_sequence(struct nfs_client *clp,
diff --git a/fs/nfs/nfs4renewd.c b/fs/nfs/nfs4renewd.c
index 0156c01c212c..d87f10327b72 100644
--- a/fs/nfs/nfs4renewd.c
+++ b/fs/nfs/nfs4renewd.c
@@ -36,11 +36,6 @@
  * as an rpc_task, not a real kernel thread, so it always runs in rpciod's
  * context.  There is one renewd per nfs_server.
  *
- * TODO: If the send queue gets backlogged (e.g., if the server goes down),
- * we will keep filling the queue with periodic RENEW requests.  We need a
- * mechanism for ensuring that if renewd successfully sends off a request,
- * then it only wakes up when the request is finished.  Maybe use the
- * child task framework of the RPC layer?
  */
 
 #include <linux/mm.h>
@@ -63,7 +58,7 @@ nfs4_renew_state(struct work_struct *work)
         struct nfs_client *clp =
                 container_of(work, struct nfs_client, cl_renewd.work);
         struct rpc_cred *cred;
-        long lease, timeout;
+        long lease;
         unsigned long last, now;
 
         ops = nfs4_state_renewal_ops[clp->cl_minorversion];
@@ -75,7 +70,6 @@ nfs4_renew_state(struct work_struct *work)
         lease = clp->cl_lease_time;
         last = clp->cl_last_renewal;
         now = jiffies;
-        timeout = (2 * lease) / 3 + (long)last - (long)now;
         /* Are we close to a lease timeout? */
         if (time_after(now, last + lease/3)) {
                 cred = ops->get_state_renewal_cred_locked(clp);
@@ -90,19 +84,15 @@ nfs4_renew_state(struct work_struct *work)
                         /* Queue an asynchronous RENEW. */
                         ops->sched_state_renewal(clp, cred);
                         put_rpccred(cred);
+                        goto out_exp;
                 }
-                timeout = (2 * lease) / 3;
-                spin_lock(&clp->cl_lock);
-        } else
+        } else {
                 dprintk("%s: failed to call renewd. Reason: lease not expired \n",
                                 __func__);
-        if (timeout < 5 * HZ)    /* safeguard */
-                timeout = 5 * HZ;
-        dprintk("%s: requeueing work. Lease period = %ld\n",
-                        __func__, (timeout + HZ - 1) / HZ);
-        cancel_delayed_work(&clp->cl_renewd);
-        schedule_delayed_work(&clp->cl_renewd, timeout);
-        spin_unlock(&clp->cl_lock);
+                spin_unlock(&clp->cl_lock);
+        }
+        nfs4_schedule_state_renewal(clp);
+out_exp:
         nfs_expire_unreferenced_delegations(clp);
 out:
         dprintk("%s: done\n", __func__);