xprtrdma: Properly handle exhaustion of the rb_mws list
authorChuck Lever <chuck.lever@oracle.com>
Tue, 29 Jul 2014 21:24:36 +0000 (17:24 -0400)
committerAnna Schumaker <Anna.Schumaker@Netapp.com>
Thu, 31 Jul 2014 20:22:55 +0000 (16:22 -0400)
If the rb_mws list is exhausted, clean up and return NULL so that
call_allocate() will delay and try again.

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Tested-by: Shirley Ma <shirley.ma@oracle.com>
Tested-by: Devesh Sharma <devesh.sharma@emulex.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
net/sunrpc/xprtrdma/verbs.c

index 0ad7d10f13a7fb1264e39b4e81d55647c24c46a9..017f0abb2a8675e98a3c3e1adee655ba304b5b6e 100644 (file)
@@ -1256,6 +1256,67 @@ rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
        kfree(buf->rb_pool);
 }
 
+/* "*mw" can be NULL when rpcrdma_buffer_get_mrs() fails, leaving
+ * some req segments uninitialized.
+ */
+static void
+rpcrdma_buffer_put_mr(struct rpcrdma_mw **mw, struct rpcrdma_buffer *buf)
+{
+       if (*mw) {
+               list_add_tail(&(*mw)->mw_list, &buf->rb_mws);
+               *mw = NULL;
+       }
+}
+
+/* Cycle mw's back in reverse order, and "spin" them.
+ * This delays and scrambles reuse as much as possible.
+ */
+static void
+rpcrdma_buffer_put_mrs(struct rpcrdma_req *req, struct rpcrdma_buffer *buf)
+{
+       struct rpcrdma_mr_seg *seg = req->rl_segments;
+       struct rpcrdma_mr_seg *seg1 = seg;
+       int i;
+
+       for (i = 1, seg++; i < RPCRDMA_MAX_SEGS; seg++, i++)
+               rpcrdma_buffer_put_mr(&seg->mr_chunk.rl_mw, buf);
+       rpcrdma_buffer_put_mr(&seg1->mr_chunk.rl_mw, buf);
+}
+
+static void
+rpcrdma_buffer_put_sendbuf(struct rpcrdma_req *req, struct rpcrdma_buffer *buf)
+{
+       buf->rb_send_bufs[--buf->rb_send_index] = req;
+       req->rl_niovs = 0;
+       if (req->rl_reply) {
+               buf->rb_recv_bufs[--buf->rb_recv_index] = req->rl_reply;
+               req->rl_reply->rr_func = NULL;
+               req->rl_reply = NULL;
+       }
+}
+
+static struct rpcrdma_req *
+rpcrdma_buffer_get_mrs(struct rpcrdma_req *req, struct rpcrdma_buffer *buf)
+{
+       struct rpcrdma_mw *r;
+       int i;
+
+       i = RPCRDMA_MAX_SEGS - 1;
+       while (!list_empty(&buf->rb_mws)) {
+               r = list_entry(buf->rb_mws.next,
+                              struct rpcrdma_mw, mw_list);
+               list_del(&r->mw_list);
+               req->rl_segments[i].mr_chunk.rl_mw = r;
+               if (unlikely(i-- == 0))
+                       return req;     /* Success */
+       }
+
+       /* Not enough entries on rb_mws for this req */
+       rpcrdma_buffer_put_sendbuf(req, buf);
+       rpcrdma_buffer_put_mrs(req, buf);
+       return NULL;
+}
+
 /*
  * Get a set of request/reply buffers.
  *
@@ -1268,10 +1329,9 @@ rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
 struct rpcrdma_req *
 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
 {
+       struct rpcrdma_ia *ia = rdmab_to_ia(buffers);
        struct rpcrdma_req *req;
        unsigned long flags;
-       int i;
-       struct rpcrdma_mw *r;
 
        spin_lock_irqsave(&buffers->rb_lock, flags);
        if (buffers->rb_send_index == buffers->rb_max_requests) {
@@ -1291,14 +1351,13 @@ rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
                buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
        }
        buffers->rb_send_bufs[buffers->rb_send_index++] = NULL;
-       if (!list_empty(&buffers->rb_mws)) {
-               i = RPCRDMA_MAX_SEGS - 1;
-               do {
-                       r = list_entry(buffers->rb_mws.next,
-                                       struct rpcrdma_mw, mw_list);
-                       list_del(&r->mw_list);
-                       req->rl_segments[i].mr_chunk.rl_mw = r;
-               } while (--i >= 0);
+       switch (ia->ri_memreg_strategy) {
+       case RPCRDMA_FRMR:
+       case RPCRDMA_MTHCAFMR:
+               req = rpcrdma_buffer_get_mrs(req, buffers);
+               break;
+       default:
+               break;
        }
        spin_unlock_irqrestore(&buffers->rb_lock, flags);
        return req;
@@ -1313,34 +1372,14 @@ rpcrdma_buffer_put(struct rpcrdma_req *req)
 {
        struct rpcrdma_buffer *buffers = req->rl_buffer;
        struct rpcrdma_ia *ia = rdmab_to_ia(buffers);
-       int i;
        unsigned long flags;
 
        spin_lock_irqsave(&buffers->rb_lock, flags);
-       buffers->rb_send_bufs[--buffers->rb_send_index] = req;
-       req->rl_niovs = 0;
-       if (req->rl_reply) {
-               buffers->rb_recv_bufs[--buffers->rb_recv_index] = req->rl_reply;
-               req->rl_reply->rr_func = NULL;
-               req->rl_reply = NULL;
-       }
+       rpcrdma_buffer_put_sendbuf(req, buffers);
        switch (ia->ri_memreg_strategy) {
        case RPCRDMA_FRMR:
        case RPCRDMA_MTHCAFMR:
-               /*
-                * Cycle mw's back in reverse order, and "spin" them.
-                * This delays and scrambles reuse as much as possible.
-                */
-               i = 1;
-               do {
-                       struct rpcrdma_mw **mw;
-                       mw = &req->rl_segments[i].mr_chunk.rl_mw;
-                       list_add_tail(&(*mw)->mw_list, &buffers->rb_mws);
-                       *mw = NULL;
-               } while (++i < RPCRDMA_MAX_SEGS);
-               list_add_tail(&req->rl_segments[0].mr_chunk.rl_mw->mw_list,
-                                       &buffers->rb_mws);
-               req->rl_segments[0].mr_chunk.rl_mw = NULL;
+               rpcrdma_buffer_put_mrs(req, buffers);
                break;
        default:
                break;