xprtrdma: Reset FRMRs after a flushed LOCAL_INV Work Request
authorChuck Lever <chuck.lever@oracle.com>
Tue, 29 Jul 2014 21:24:54 +0000 (17:24 -0400)
committerAnna Schumaker <Anna.Schumaker@Netapp.com>
Thu, 31 Jul 2014 20:22:55 +0000 (16:22 -0400)
When a LOCAL_INV Work Request is flushed, it leaves an FRMR in the
VALID state. This FRMR can be returned by rpcrdma_buffer_get(), and
must be knocked down in rpcrdma_register_frmr_external() before it
can be re-used.

Instead, capture these in rpcrdma_buffer_get(), and reset them.

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 3a6376a..ca55acf 100644 (file)
@@ -1358,8 +1358,91 @@ rpcrdma_buffer_put_sendbuf(struct rpcrdma_req *req, struct rpcrdma_buffer *buf)
        }
 }
 
+/* rpcrdma_unmap_one() was already done by rpcrdma_deregister_frmr_external().
+ * Redo only the ib_post_send().
+ */
+static void
+rpcrdma_retry_local_inv(struct rpcrdma_mw *r, struct rpcrdma_ia *ia)
+{
+       struct rpcrdma_xprt *r_xprt =
+                               container_of(ia, struct rpcrdma_xprt, rx_ia);
+       struct ib_send_wr invalidate_wr, *bad_wr;
+       int rc;
+
+       dprintk("RPC:       %s: FRMR %p is stale\n", __func__, r);
+
+       /* When this FRMR is re-inserted into rb_mws, it is no longer stale */
+       r->r.frmr.fr_state = FRMR_IS_VALID;
+
+       memset(&invalidate_wr, 0, sizeof(invalidate_wr));
+       invalidate_wr.wr_id = (unsigned long)(void *)r;
+       invalidate_wr.opcode = IB_WR_LOCAL_INV;
+       invalidate_wr.send_flags = IB_SEND_SIGNALED;
+       invalidate_wr.ex.invalidate_rkey = r->r.frmr.fr_mr->rkey;
+       DECR_CQCOUNT(&r_xprt->rx_ep);
+
+       dprintk("RPC:       %s: frmr %p invalidating rkey %08x\n",
+               __func__, r, r->r.frmr.fr_mr->rkey);
+
+       read_lock(&ia->ri_qplock);
+       rc = ib_post_send(ia->ri_id->qp, &invalidate_wr, &bad_wr);
+       read_unlock(&ia->ri_qplock);
+       if (rc) {
+               /* Force rpcrdma_buffer_get() to retry */
+               r->r.frmr.fr_state = FRMR_IS_STALE;
+               dprintk("RPC:       %s: ib_post_send failed, %i\n",
+                       __func__, rc);
+       }
+}
+
+static void
+rpcrdma_retry_flushed_linv(struct list_head *stale,
+                          struct rpcrdma_buffer *buf)
+{
+       struct rpcrdma_ia *ia = rdmab_to_ia(buf);
+       struct list_head *pos;
+       struct rpcrdma_mw *r;
+       unsigned long flags;
+
+       list_for_each(pos, stale) {
+               r = list_entry(pos, struct rpcrdma_mw, mw_list);
+               rpcrdma_retry_local_inv(r, ia);
+       }
+
+       spin_lock_irqsave(&buf->rb_lock, flags);
+       list_splice_tail(stale, &buf->rb_mws);
+       spin_unlock_irqrestore(&buf->rb_lock, flags);
+}
+
 static struct rpcrdma_req *
-rpcrdma_buffer_get_mrs(struct rpcrdma_req *req, struct rpcrdma_buffer *buf)
+rpcrdma_buffer_get_frmrs(struct rpcrdma_req *req, struct rpcrdma_buffer *buf,
+                        struct list_head *stale)
+{
+       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);
+               if (r->r.frmr.fr_state == FRMR_IS_STALE) {
+                       list_add(&r->mw_list, stale);
+                       continue;
+               }
+               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;
+}
+
+static struct rpcrdma_req *
+rpcrdma_buffer_get_fmrs(struct rpcrdma_req *req, struct rpcrdma_buffer *buf)
 {
        struct rpcrdma_mw *r;
        int i;
@@ -1393,6 +1476,7 @@ struct rpcrdma_req *
 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
 {
        struct rpcrdma_ia *ia = rdmab_to_ia(buffers);
+       struct list_head stale;
        struct rpcrdma_req *req;
        unsigned long flags;
 
@@ -1414,15 +1498,21 @@ 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;
+
+       INIT_LIST_HEAD(&stale);
        switch (ia->ri_memreg_strategy) {
        case RPCRDMA_FRMR:
+               req = rpcrdma_buffer_get_frmrs(req, buffers, &stale);
+               break;
        case RPCRDMA_MTHCAFMR:
-               req = rpcrdma_buffer_get_mrs(req, buffers);
+               req = rpcrdma_buffer_get_fmrs(req, buffers);
                break;
        default:
                break;
        }
        spin_unlock_irqrestore(&buffers->rb_lock, flags);
+       if (!list_empty(&stale))
+               rpcrdma_retry_flushed_linv(&stale, buffers);
        return req;
 }