xprtrdma: rpcrdma_regbuf alignment

Allocate the struct rpcrdma_regbuf separately from the I/O buffer
to better guarantee the alignment of the I/O buffer and eliminate
the wasted space between the rpcrdma_regbuf metadata and the buffer
itself.

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>

net/sunrpc/xprtrdma/backchannel.c

@@ -45,10 +45,10 @@ static int rpcrdma_bc_setup_reqs(struct rpcrdma_xprt *r_xprt,
 
 		size = r_xprt->rx_data.inline_rsize;
 		rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, GFP_KERNEL);
-		if (IS_ERR(rb))
+		if (!rb)
 			goto out_fail;
 		req->rl_sendbuf = rb;
-		xdr_buf_init(&rqst->rq_snd_buf, rb->rg_base,
+		xdr_buf_init(&rqst->rq_snd_buf, rdmab_data(rb),
 			     min_t(size_t, size, PAGE_SIZE));
 	}
 	return 0;
@@ -123,7 +123,7 @@ static int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst)
 
 	rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
 	xdr_init_encode(&req->rl_stream, &req->rl_hdrbuf,
-			req->rl_rdmabuf->rg_base, rqst);
+			rdmab_data(req->rl_rdmabuf), rqst);
 
 	p = xdr_reserve_space(&req->rl_stream, 28);
 	if (unlikely(!p))

net/sunrpc/xprtrdma/rpc_rdma.c

@@ -747,8 +747,8 @@ rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
 	int ret;
 
 	rpcrdma_set_xdrlen(&req->rl_hdrbuf, 0);
-	xdr_init_encode(xdr, &req->rl_hdrbuf,
+	xdr_init_encode(xdr, &req->rl_hdrbuf, rdmab_data(req->rl_rdmabuf),
-			req->rl_rdmabuf->rg_base, rqst);
+			rqst);
 
 	/* Fixed header fields */
 	ret = -EMSGSIZE;

net/sunrpc/xprtrdma/transport.c

@@ -595,7 +595,7 @@ rpcrdma_get_sendbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 		return true;
 
 	rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
-	if (IS_ERR(rb))
+	if (!rb)
 		return false;
 
 	rpcrdma_free_regbuf(req->rl_sendbuf);
@@ -625,7 +625,7 @@ rpcrdma_get_recvbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 		return true;
 
 	rb = rpcrdma_alloc_regbuf(size, DMA_NONE, flags);
-	if (IS_ERR(rb))
+	if (!rb)
 		return false;
 
 	rpcrdma_free_regbuf(req->rl_recvbuf);
@@ -660,8 +660,8 @@ xprt_rdma_allocate(struct rpc_task *task)
 	if (!rpcrdma_get_recvbuf(r_xprt, req, rqst->rq_rcvsize, flags))
 		goto out_fail;
 
-	rqst->rq_buffer = req->rl_sendbuf->rg_base;
+	rqst->rq_buffer = rdmab_data(req->rl_sendbuf);
-	rqst->rq_rbuffer = req->rl_recvbuf->rg_base;
+	rqst->rq_rbuffer = rdmab_data(req->rl_recvbuf);
 	trace_xprtrdma_op_allocate(task, req);
 	return 0;
 

net/sunrpc/xprtrdma/verbs.c

@@ -1013,12 +1013,12 @@ struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, gfp_t flags)
 		return NULL;
 
 	rb = rpcrdma_alloc_regbuf(RPCRDMA_HDRBUF_SIZE, DMA_TO_DEVICE, flags);
-	if (IS_ERR(rb)) {
+	if (!rb) {
 		kfree(req);
 		return NULL;
 	}
 	req->rl_rdmabuf = rb;
-	xdr_buf_init(&req->rl_hdrbuf, rb->rg_base, rdmab_length(rb));
+	xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
 	req->rl_buffer = buffer;
 	INIT_LIST_HEAD(&req->rl_registered);
 
@@ -1040,9 +1040,9 @@ static bool rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt, bool temp)
 
 	rep->rr_rdmabuf = rpcrdma_alloc_regbuf(cdata->inline_rsize,
 					       DMA_FROM_DEVICE, GFP_KERNEL);
-	if (IS_ERR(rep->rr_rdmabuf))
+	if (!rep->rr_rdmabuf)
 		goto out_free;
-	xdr_buf_init(&rep->rr_hdrbuf, rep->rr_rdmabuf->rg_base,
+	xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
 		     rdmab_length(rep->rr_rdmabuf));
 
 	rep->rr_cqe.done = rpcrdma_wc_receive;
@@ -1354,8 +1354,7 @@ rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
  * @direction: direction of data movement
  * @flags: GFP flags
  *
- * Returns an ERR_PTR, or a pointer to a regbuf, a buffer that
+ * Returns a pointer to a rpcrdma_regbuf object, or NULL.
- * can be persistently DMA-mapped for I/O.
  *
  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
  * receiving the payload of RDMA RECV operations. During Long Calls
@@ -1367,14 +1366,18 @@ rpcrdma_alloc_regbuf(size_t size, enum dma_data_direction direction,
 {
 	struct rpcrdma_regbuf *rb;
 
-	rb = kmalloc(sizeof(*rb) + size, flags);
+	rb = kmalloc(sizeof(*rb), flags);
-	if (rb == NULL)
+	if (!rb)
-		return ERR_PTR(-ENOMEM);
+		return NULL;
+	rb->rg_data = kmalloc(size, flags);
+	if (!rb->rg_data) {
+		kfree(rb);
+		return NULL;
+	}
 
 	rb->rg_device = NULL;
 	rb->rg_direction = direction;
 	rb->rg_iov.length = size;
-
 	return rb;
 }
 
@@ -1392,7 +1395,7 @@ __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
 		return false;
 
 	rb->rg_iov.addr = ib_dma_map_single(device,
-					    (void *)rb->rg_base,
+					    rdmab_data(rb),
 					    rdmab_length(rb),
 					    rb->rg_direction);
 	if (ib_dma_mapping_error(device, rdmab_addr(rb))) {
@@ -1427,6 +1430,8 @@ void
 rpcrdma_free_regbuf(struct rpcrdma_regbuf *rb)
 {
 	rpcrdma_dma_unmap_regbuf(rb);
+	if (rb)
+		kfree(rb->rg_data);
 	kfree(rb);
 }
 

net/sunrpc/xprtrdma/xprt_rdma.h

@@ -121,33 +121,34 @@ struct rpcrdma_regbuf {
 	struct ib_sge		rg_iov;
 	struct ib_device	*rg_device;
 	enum dma_data_direction	rg_direction;
-	__be32			rg_base[0] __attribute__ ((aligned(256)));
+	void			*rg_data;
 };
 
-static inline u64
+static inline u64 rdmab_addr(struct rpcrdma_regbuf *rb)
-rdmab_addr(struct rpcrdma_regbuf *rb)
 {
 	return rb->rg_iov.addr;
 }
 
-static inline u32
+static inline u32 rdmab_length(struct rpcrdma_regbuf *rb)
-rdmab_length(struct rpcrdma_regbuf *rb)
 {
 	return rb->rg_iov.length;
 }
 
-static inline u32
+static inline u32 rdmab_lkey(struct rpcrdma_regbuf *rb)
-rdmab_lkey(struct rpcrdma_regbuf *rb)
 {
 	return rb->rg_iov.lkey;
 }
 
-static inline struct ib_device *
+static inline struct ib_device *rdmab_device(struct rpcrdma_regbuf *rb)
-rdmab_device(struct rpcrdma_regbuf *rb)
 {
 	return rb->rg_device;
 }
 
+static inline void *rdmab_data(const struct rpcrdma_regbuf *rb)
+{
+	return rb->rg_data;
+}
+
 #define RPCRDMA_DEF_GFP		(GFP_NOIO | __GFP_NOWARN)
 
 /* To ensure a transport can always make forward progress,