]> git.kernelconcepts.de Git - karo-tx-linux.git/commitdiff
9P/RDMA: count posted buffers without a pending request
authorSimon Derr <simon.derr@bull.net>
Fri, 21 Jun 2013 13:32:42 +0000 (15:32 +0200)
committerEric Van Hensbergen <ericvh@gmail.com>
Mon, 8 Jul 2013 03:04:36 +0000 (22:04 -0500)
In rdma_request():

If an error occurs between posting the recv and the send,
there will be a reply context posted without a pending
request.
Since there is no way to "un-post" it, we remember it and
skip post_recv() for the next request.

Signed-off-by: Simon Derr <simon.derr@bull.net>
Signed-off-by: Eric Van Hensbergen <ericvh@gmail.com>
net/9p/client.c
net/9p/trans_rdma.c

index db5bf2480a336d3bdf7a2b357ca5e3cc6f765fe4..d18a0b22f62cd7383651259f5a49cbd589f00c3f 100644 (file)
@@ -656,8 +656,10 @@ static int p9_client_flush(struct p9_client *c, struct p9_req_t *oldreq)
                return PTR_ERR(req);
 
 
-       /* if we haven't received a response for oldreq,
-          remove it from the list. */
+       /*
+        * if we haven't received a response for oldreq,
+        * remove it from the list.
+        */
        spin_lock(&c->lock);
        if (oldreq->status == REQ_STATUS_FLSH)
                list_del(&oldreq->req_list);
index 926e72d00e5719830564a49d2105434d5f006aa5..8f68df5d29731cf23b6de1bbee5284c904006c93 100644 (file)
@@ -74,6 +74,8 @@
  * @sq_sem: Semaphore for the SQ
  * @rq_depth: The depth of the Receive Queue.
  * @rq_sem: Semaphore for the RQ
+ * @excess_rc : Amount of posted Receive Contexts without a pending request.
+ *             See rdma_request()
  * @addr: The remote peer's address
  * @req_lock: Protects the active request list
  * @cm_done: Completion event for connection management tracking
@@ -99,6 +101,7 @@ struct p9_trans_rdma {
        struct semaphore sq_sem;
        int rq_depth;
        struct semaphore rq_sem;
+       atomic_t excess_rc;
        struct sockaddr_in addr;
        spinlock_t req_lock;
 
@@ -426,6 +429,26 @@ static int rdma_request(struct p9_client *client, struct p9_req_t *req)
        struct p9_rdma_context *c = NULL;
        struct p9_rdma_context *rpl_context = NULL;
 
+       /* When an error occurs between posting the recv and the send,
+        * there will be a receive context posted without a pending request.
+        * Since there is no way to "un-post" it, we remember it and skip
+        * post_recv() for the next request.
+        * So here,
+        * see if we are this `next request' and need to absorb an excess rc.
+        * If yes, then drop and free our own, and do not recv_post().
+        **/
+       if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
+               if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
+                       /* Got one ! */
+                       kfree(req->rc);
+                       req->rc = NULL;
+                       goto dont_need_post_recv;
+               } else {
+                       /* We raced and lost. */
+                       atomic_inc(&rdma->excess_rc);
+               }
+       }
+
        /* Allocate an fcall for the reply */
        rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
        if (!rpl_context) {
@@ -451,10 +474,10 @@ static int rdma_request(struct p9_client *client, struct p9_req_t *req)
                p9_debug(P9_DEBUG_FCALL, "POST RECV failed\n");
                goto recv_error;
        }
-
        /* remove posted receive buffer from request structure */
        req->rc = NULL;
 
+dont_need_post_recv:
        /* Post the request */
        c = kmalloc(sizeof *c, GFP_NOFS);
        if (!c) {
@@ -499,6 +522,11 @@ static int rdma_request(struct p9_client *client, struct p9_req_t *req)
  send_error:
        kfree(c);
        p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
+
+       /* Ach.
+        *  We did recv_post(), but not send. We have one recv_post in excess.
+        */
+       atomic_inc(&rdma->excess_rc);
        return err;
 
  /* Handle errors that happened during or while preparing post_recv(): */
@@ -549,6 +577,7 @@ static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
        init_completion(&rdma->cm_done);
        sema_init(&rdma->sq_sem, rdma->sq_depth);
        sema_init(&rdma->rq_sem, rdma->rq_depth);
+       atomic_set(&rdma->excess_rc, 0);
 
        return rdma;
 }