summaryrefslogtreecommitdiffstats
path: root/net
diff options
context:
space:
mode:
authorChuck Lever <chuck.lever@oracle.com>2016-09-15 16:56:10 +0200
committerAnna Schumaker <Anna.Schumaker@Netapp.com>2016-09-19 19:08:37 +0200
commit99ef4db329f1ee2413dad49346e72a6c902474d1 (patch)
treee1ba252d10de1bac6fe4ad83b4a6f5339a3d235d /net
parentxprtrdma: Use smaller buffers for RPC-over-RDMA headers (diff)
downloadlinux-99ef4db329f1ee2413dad49346e72a6c902474d1.tar.xz
linux-99ef4db329f1ee2413dad49346e72a6c902474d1.zip
xprtrdma: Replace DMA_BIDIRECTIONAL
The use of DMA_BIDIRECTIONAL is discouraged by DMA-API.txt. Fortunately, xprtrdma now knows which direction I/O is going as soon as it allocates each regbuf. The RPC Call and Reply buffers are no longer the same regbuf. They can each be labeled correctly now. The RPC Reply buffer is never part of either a Send or Receive WR, but it can be part of Reply chunk, which is mapped and registered via ->ro_map . So it is not DMA mapped when it is allocated (DMA_NONE), to avoid a double- mapping. Since Receive buffers are no longer DMA_BIDIRECTIONAL and their contents are never modified by the host CPU, DMA-API-HOWTO.txt suggests that a DMA sync before posting each buffer should be unnecessary. (See my_card_interrupt_handler). Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Diffstat (limited to 'net')
-rw-r--r--net/sunrpc/xprtrdma/backchannel.c5
-rw-r--r--net/sunrpc/xprtrdma/transport.c7
-rw-r--r--net/sunrpc/xprtrdma/verbs.c55
-rw-r--r--net/sunrpc/xprtrdma/xprt_rdma.h4
4 files changed, 36 insertions, 35 deletions
diff --git a/net/sunrpc/xprtrdma/backchannel.c b/net/sunrpc/xprtrdma/backchannel.c
index 60fc9915292f..ceae87206347 100644
--- a/net/sunrpc/xprtrdma/backchannel.c
+++ b/net/sunrpc/xprtrdma/backchannel.c
@@ -45,13 +45,14 @@ static int rpcrdma_bc_setup_rqst(struct rpcrdma_xprt *r_xprt,
return PTR_ERR(req);
req->rl_backchannel = true;
- rb = rpcrdma_alloc_regbuf(ia, RPCRDMA_HDRBUF_SIZE, GFP_KERNEL);
+ rb = rpcrdma_alloc_regbuf(ia, RPCRDMA_HDRBUF_SIZE,
+ DMA_TO_DEVICE, GFP_KERNEL);
if (IS_ERR(rb))
goto out_fail;
req->rl_rdmabuf = rb;
size = r_xprt->rx_data.inline_rsize;
- rb = rpcrdma_alloc_regbuf(ia, size, GFP_KERNEL);
+ rb = rpcrdma_alloc_regbuf(ia, size, DMA_TO_DEVICE, GFP_KERNEL);
if (IS_ERR(rb))
goto out_fail;
req->rl_sendbuf = rb;
diff --git a/net/sunrpc/xprtrdma/transport.c b/net/sunrpc/xprtrdma/transport.c
index 94dbfd3e89a7..34246916434b 100644
--- a/net/sunrpc/xprtrdma/transport.c
+++ b/net/sunrpc/xprtrdma/transport.c
@@ -490,7 +490,7 @@ rpcrdma_get_rdmabuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
if (req->rl_rdmabuf)
return true;
- rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, size, flags);
+ rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, size, DMA_TO_DEVICE, flags);
if (IS_ERR(rb))
return false;
@@ -517,7 +517,8 @@ rpcrdma_get_sendbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
return true;
min_size = max_t(size_t, size, r_xprt->rx_data.inline_wsize);
- rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, min_size, flags);
+ rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, min_size,
+ DMA_TO_DEVICE, flags);
if (IS_ERR(rb))
return false;
@@ -547,7 +548,7 @@ rpcrdma_get_recvbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
if (req->rl_recvbuf && rdmab_length(req->rl_recvbuf) >= size)
return true;
- rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, size, flags);
+ rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, size, DMA_NONE, flags);
if (IS_ERR(rb))
return false;
diff --git a/net/sunrpc/xprtrdma/verbs.c b/net/sunrpc/xprtrdma/verbs.c
index fc6b4ea8b7ec..9edea34aeb36 100644
--- a/net/sunrpc/xprtrdma/verbs.c
+++ b/net/sunrpc/xprtrdma/verbs.c
@@ -866,7 +866,7 @@ rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
goto out;
rep->rr_rdmabuf = rpcrdma_alloc_regbuf(ia, cdata->inline_rsize,
- GFP_KERNEL);
+ DMA_FROM_DEVICE, GFP_KERNEL);
if (IS_ERR(rep->rr_rdmabuf)) {
rc = PTR_ERR(rep->rr_rdmabuf);
goto out_free;
@@ -1172,27 +1172,24 @@ rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
spin_unlock(&buffers->rb_lock);
}
-/*
- * Wrappers for internal-use kmalloc memory registration, used by buffer code.
- */
-
/**
- * rpcrdma_alloc_regbuf - kmalloc and register memory for SEND/RECV buffers
+ * rpcrdma_alloc_regbuf - allocate and DMA-map memory for SEND/RECV buffers
* @ia: controlling rpcrdma_ia
* @size: size of buffer to be allocated, in bytes
+ * @direction: direction of data movement
* @flags: GFP flags
*
- * Returns pointer to private header of an area of internally
- * registered memory, or an ERR_PTR. The registered buffer follows
- * the end of the private header.
+ * Returns an ERR_PTR, or a pointer to a regbuf, which is a
+ * contiguous memory region that is DMA mapped persistently, and
+ * is registered for local I/O.
*
* xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
- * receiving the payload of RDMA RECV operations. regbufs are not
- * used for RDMA READ/WRITE operations, thus are registered only for
- * LOCAL access.
+ * receiving the payload of RDMA RECV operations. During Long Calls
+ * or Replies they may be registered externally via ro_map.
*/
struct rpcrdma_regbuf *
-rpcrdma_alloc_regbuf(struct rpcrdma_ia *ia, size_t size, gfp_t flags)
+rpcrdma_alloc_regbuf(struct rpcrdma_ia *ia, size_t size,
+ enum dma_data_direction direction, gfp_t flags)
{
struct rpcrdma_regbuf *rb;
struct ib_sge *iov;
@@ -1201,15 +1198,20 @@ rpcrdma_alloc_regbuf(struct rpcrdma_ia *ia, size_t size, gfp_t flags)
if (rb == NULL)
goto out;
+ rb->rg_direction = direction;
iov = &rb->rg_iov;
- iov->addr = ib_dma_map_single(ia->ri_device,
- (void *)rb->rg_base, size,
- DMA_BIDIRECTIONAL);
- if (ib_dma_mapping_error(ia->ri_device, iov->addr))
- goto out_free;
-
iov->length = size;
iov->lkey = ia->ri_pd->local_dma_lkey;
+
+ if (direction != DMA_NONE) {
+ iov->addr = ib_dma_map_single(ia->ri_device,
+ (void *)rb->rg_base,
+ rdmab_length(rb),
+ rb->rg_direction);
+ if (ib_dma_mapping_error(ia->ri_device, iov->addr))
+ goto out_free;
+ }
+
return rb;
out_free:
@@ -1226,14 +1228,14 @@ out:
void
rpcrdma_free_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
{
- struct ib_sge *iov;
-
if (!rb)
return;
- iov = &rb->rg_iov;
- ib_dma_unmap_single(ia->ri_device,
- iov->addr, iov->length, DMA_BIDIRECTIONAL);
+ if (rb->rg_direction != DMA_NONE) {
+ ib_dma_unmap_single(ia->ri_device, rdmab_addr(rb),
+ rdmab_length(rb), rb->rg_direction);
+ }
+
kfree(rb);
}
@@ -1305,11 +1307,6 @@ rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
recv_wr.num_sge = 1;
- ib_dma_sync_single_for_cpu(ia->ri_device,
- rdmab_addr(rep->rr_rdmabuf),
- rdmab_length(rep->rr_rdmabuf),
- DMA_BIDIRECTIONAL);
-
rc = ib_post_recv(ia->ri_id->qp, &recv_wr, &recv_wr_fail);
if (rc)
goto out_postrecv;
diff --git a/net/sunrpc/xprtrdma/xprt_rdma.h b/net/sunrpc/xprtrdma/xprt_rdma.h
index cc426b165c09..9569b212be54 100644
--- a/net/sunrpc/xprtrdma/xprt_rdma.h
+++ b/net/sunrpc/xprtrdma/xprt_rdma.h
@@ -113,6 +113,7 @@ struct rpcrdma_ep {
struct rpcrdma_regbuf {
struct ib_sge rg_iov;
+ enum dma_data_direction rg_direction;
__be32 rg_base[0] __attribute__ ((aligned(256)));
};
@@ -477,7 +478,8 @@ void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
void rpcrdma_defer_mr_recovery(struct rpcrdma_mw *);
struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(struct rpcrdma_ia *,
- size_t, gfp_t);
+ size_t, enum dma_data_direction,
+ gfp_t);
void rpcrdma_free_regbuf(struct rpcrdma_ia *,
struct rpcrdma_regbuf *);