[PATCH] NLM: fix a client-side race on blocking locks.
authorTrond Myklebust <Trond.Myklebust@netapp.com>
Wed, 22 Jun 2005 17:16:31 +0000 (17:16 +0000)
committerTrond Myklebust <Trond.Myklebust@netapp.com>
Wed, 22 Jun 2005 20:07:42 +0000 (16:07 -0400)
 If the lock blocks, the server may send us a GRANTED message that
 races with the reply to our LOCK request. Make sure that we catch
 the GRANTED by queueing up our request on the nlm_blocked list
 before we send off the first LOCK rpc call.

Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
fs/lockd/clntlock.c
fs/lockd/clntproc.c
include/linux/lockd/lockd.h

index 44adb84183b6b4e3c56300e8142c8cd49c472582..006bb9e14579e42c37cd5954dac3a34b1eacf25c 100644 (file)
@@ -42,23 +42,51 @@ struct nlm_wait {
 static LIST_HEAD(nlm_blocked);
 
 /*
- * Block on a lock
+ * Queue up a lock for blocking so that the GRANTED request can see it
  */
-int
-nlmclnt_block(struct nlm_host *host, struct file_lock *fl, u32 *statp)
+int nlmclnt_prepare_block(struct nlm_rqst *req, struct nlm_host *host, struct file_lock *fl)
+{
+       struct nlm_wait *block;
+
+       BUG_ON(req->a_block != NULL);
+       block = kmalloc(sizeof(*block), GFP_KERNEL);
+       if (block == NULL)
+               return -ENOMEM;
+       block->b_host = host;
+       block->b_lock = fl;
+       init_waitqueue_head(&block->b_wait);
+       block->b_status = NLM_LCK_BLOCKED;
+
+       list_add(&block->b_list, &nlm_blocked);
+       req->a_block = block;
+
+       return 0;
+}
+
+void nlmclnt_finish_block(struct nlm_rqst *req)
 {
-       struct nlm_wait block, **head;
-       int             err;
-       u32             pstate;
+       struct nlm_wait *block = req->a_block;
 
-       block.b_host   = host;
-       block.b_lock   = fl;
-       init_waitqueue_head(&block.b_wait);
-       block.b_status = NLM_LCK_BLOCKED;
-       list_add(&block.b_list, &nlm_blocked);
+       if (block == NULL)
+               return;
+       req->a_block = NULL;
+       list_del(&block->b_list);
+       kfree(block);
+}
 
-       /* Remember pseudo nsm state */
-       pstate = host->h_state;
+/*
+ * Block on a lock
+ */
+long nlmclnt_block(struct nlm_rqst *req, long timeout)
+{
+       struct nlm_wait *block = req->a_block;
+       long ret;
+
+       /* A borken server might ask us to block even if we didn't
+        * request it. Just say no!
+        */
+       if (!req->a_args.block)
+               return -EAGAIN;
 
        /* Go to sleep waiting for GRANT callback. Some servers seem
         * to lose callbacks, however, so we're going to poll from
@@ -68,23 +96,16 @@ nlmclnt_block(struct nlm_host *host, struct file_lock *fl, u32 *statp)
         * a 1 minute timeout would do. See the comment before
         * nlmclnt_lock for an explanation.
         */
-       sleep_on_timeout(&block.b_wait, 30*HZ);
+       ret = wait_event_interruptible_timeout(block->b_wait,
+                       block->b_status != NLM_LCK_BLOCKED,
+                       timeout);
 
-       list_del(&block.b_list);
-
-       if (!signalled()) {
-               *statp = block.b_status;
-               return 0;
+       if (block->b_status != NLM_LCK_BLOCKED) {
+               req->a_res.status = block->b_status;
+               block->b_status = NLM_LCK_BLOCKED;
        }
 
-       /* Okay, we were interrupted. Cancel the pending request
-        * unless the server has rebooted.
-        */
-       if (pstate == host->h_state && (err = nlmclnt_cancel(host, fl)) < 0)
-               printk(KERN_NOTICE
-                       "lockd: CANCEL call failed (errno %d)\n", -err);
-
-       return -ERESTARTSYS;
+       return ret;
 }
 
 /*
@@ -94,27 +115,23 @@ u32
 nlmclnt_grant(struct nlm_lock *lock)
 {
        struct nlm_wait *block;
+       u32 res = nlm_lck_denied;
 
        /*
         * Look up blocked request based on arguments. 
         * Warning: must not use cookie to match it!
         */
        list_for_each_entry(block, &nlm_blocked, b_list) {
-               if (nlm_compare_locks(block->b_lock, &lock->fl))
-                       break;
+               if (nlm_compare_locks(block->b_lock, &lock->fl)) {
+                       /* Alright, we found a lock. Set the return status
+                        * and wake up the caller
+                        */
+                       block->b_status = NLM_LCK_GRANTED;
+                       wake_up(&block->b_wait);
+                       res = nlm_granted;
+               }
        }
-
-       /* Ooops, no blocked request found. */
-       if (block == NULL)
-               return nlm_lck_denied;
-
-       /* Alright, we found the lock. Set the return status and
-        * wake up the caller.
-        */
-       block->b_status = NLM_LCK_GRANTED;
-       wake_up(&block->b_wait);
-
-       return nlm_granted;
+       return res;
 }
 
 /*
index a4407619b1f16f98b68e6ff9487a7724f1cb08b4..fd77ed1d710de4f198c275960f1bb4898f19ef3e 100644 (file)
@@ -21,6 +21,7 @@
 
 #define NLMDBG_FACILITY                NLMDBG_CLIENT
 #define NLMCLNT_GRACE_WAIT     (5*HZ)
+#define NLMCLNT_POLL_TIMEOUT   (30*HZ)
 
 static int     nlmclnt_test(struct nlm_rqst *, struct file_lock *);
 static int     nlmclnt_lock(struct nlm_rqst *, struct file_lock *);
@@ -553,7 +554,8 @@ nlmclnt_lock(struct nlm_rqst *req, struct file_lock *fl)
 {
        struct nlm_host *host = req->a_host;
        struct nlm_res  *resp = &req->a_res;
-       int             status;
+       long timeout;
+       int status;
 
        if (!host->h_monitored && nsm_monitor(host) < 0) {
                printk(KERN_NOTICE "lockd: failed to monitor %s\n",
@@ -562,15 +564,32 @@ nlmclnt_lock(struct nlm_rqst *req, struct file_lock *fl)
                goto out;
        }
 
-       do {
-               if ((status = nlmclnt_call(req, NLMPROC_LOCK)) >= 0) {
-                       if (resp->status != NLM_LCK_BLOCKED)
-                               break;
-                       status = nlmclnt_block(host, fl, &resp->status);
-               }
+       if (req->a_args.block) {
+               status = nlmclnt_prepare_block(req, host, fl);
                if (status < 0)
                        goto out;
-       } while (resp->status == NLM_LCK_BLOCKED && req->a_args.block);
+       }
+       for(;;) {
+               status = nlmclnt_call(req, NLMPROC_LOCK);
+               if (status < 0)
+                       goto out_unblock;
+               if (resp->status != NLM_LCK_BLOCKED)
+                       break;
+               /* Wait on an NLM blocking lock */
+               timeout = nlmclnt_block(req, NLMCLNT_POLL_TIMEOUT);
+               /* Did a reclaimer thread notify us of a server reboot? */
+               if (resp->status ==  NLM_LCK_DENIED_GRACE_PERIOD)
+                       continue;
+               if (resp->status != NLM_LCK_BLOCKED)
+                       break;
+               if (timeout >= 0)
+                       continue;
+               /* We were interrupted. Send a CANCEL request to the server
+                * and exit
+                */
+               status = (int)timeout;
+               goto out_unblock;
+       }
 
        if (resp->status == NLM_LCK_GRANTED) {
                fl->fl_u.nfs_fl.state = host->h_state;
@@ -579,6 +598,11 @@ nlmclnt_lock(struct nlm_rqst *req, struct file_lock *fl)
                do_vfs_lock(fl);
        }
        status = nlm_stat_to_errno(resp->status);
+out_unblock:
+       nlmclnt_finish_block(req);
+       /* Cancel the blocked request if it is still pending */
+       if (resp->status == NLM_LCK_BLOCKED)
+               nlmclnt_cancel(host, fl);
 out:
        nlmclnt_release_lockargs(req);
        return status;
index 0d9d22578212c237701479609498e84a423cd063..16d4e5a08e1dd75e2daaee997392bc88aeaa47ca 100644 (file)
@@ -72,6 +72,8 @@ struct nlm_lockowner {
        uint32_t pid;
 };
 
+struct nlm_wait;
+
 /*
  * Memory chunk for NLM client RPC request.
  */
@@ -81,6 +83,7 @@ struct nlm_rqst {
        struct nlm_host *       a_host;         /* host handle */
        struct nlm_args         a_args;         /* arguments */
        struct nlm_res          a_res;          /* result */
+       struct nlm_wait *       a_block;
        char                    a_owner[NLMCLNT_OHSIZE];
 };
 
@@ -142,7 +145,9 @@ extern unsigned long                nlmsvc_timeout;
  * Lockd client functions
  */
 struct nlm_rqst * nlmclnt_alloc_call(void);
-int              nlmclnt_block(struct nlm_host *, struct file_lock *, u32 *);
+int              nlmclnt_prepare_block(struct nlm_rqst *req, struct nlm_host *host, struct file_lock *fl);
+void             nlmclnt_finish_block(struct nlm_rqst *req);
+long             nlmclnt_block(struct nlm_rqst *req, long timeout);
 int              nlmclnt_cancel(struct nlm_host *, struct file_lock *);
 u32              nlmclnt_grant(struct nlm_lock *);
 void             nlmclnt_recovery(struct nlm_host *, u32);