@@ -28,10 +28,10 @@ static int rpcrdma_bc_setup_reqs(struct rpcrdma_xprt *r_xprt,
unsigned int i;
for (i = 0; i < (count << 1); i++) {
- struct rpcrdma_regbuf *rb;
size_t size;
- req = rpcrdma_req_create(r_xprt, GFP_KERNEL);
+ size = min_t(size_t, r_xprt->rx_data.inline_rsize, PAGE_SIZE);
+ req = rpcrdma_req_create(r_xprt, size, GFP_KERNEL);
if (!req)
return -ENOMEM;
rqst = &req->rl_slot;
@@ -42,20 +42,10 @@ static int rpcrdma_bc_setup_reqs(struct rpcrdma_xprt *r_xprt,
spin_lock(&xprt->bc_pa_lock);
list_add(&rqst->rq_bc_pa_list, &xprt->bc_pa_list);
spin_unlock(&xprt->bc_pa_lock);
-
- size = r_xprt->rx_data.inline_rsize;
- rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, GFP_KERNEL);
- if (!rb)
- goto out_fail;
- req->rl_sendbuf = rb;
- xdr_buf_init(&rqst->rq_snd_buf, rdmab_data(rb),
- min_t(size_t, size, PAGE_SIZE));
+ xdr_buf_init(&rqst->rq_snd_buf, rdmab_data(req->rl_sendbuf),
+ size);
}
return 0;
-
-out_fail:
- rpcrdma_req_destroy(req);
- return -ENOMEM;
}
/**
@@ -591,7 +591,7 @@ void xprt_rdma_close(struct rpc_xprt *xprt)
{
struct rpcrdma_regbuf *rb;
- if (req->rl_sendbuf && rdmab_length(req->rl_sendbuf) >= size)
+ if (likely(rdmab_length(req->rl_sendbuf) >= size))
return true;
rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
@@ -621,7 +621,7 @@ void xprt_rdma_close(struct rpc_xprt *xprt)
{
struct rpcrdma_regbuf *rb;
- if (req->rl_recvbuf && rdmab_length(req->rl_recvbuf) >= size)
+ if (likely(rdmab_length(req->rl_recvbuf) >= size))
return true;
rb = rpcrdma_alloc_regbuf(size, DMA_NONE, flags);
@@ -998,11 +998,13 @@ struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf)
/**
* rpcrdma_req_create - Allocate an rpcrdma_req object
* @r_xprt: controlling r_xprt
+ * @size: initial size, in bytes, of send and receive buffers
* @flags: GFP flags passed to memory allocators
*
* Returns an allocated and fully initialized rpcrdma_req or NULL.
*/
-struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, gfp_t flags)
+struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
+ gfp_t flags)
{
struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
struct rpcrdma_regbuf *rb;
@@ -1010,22 +1012,37 @@ struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, gfp_t flags)
req = kzalloc(sizeof(*req), flags);
if (req == NULL)
- return NULL;
+ goto out1;
rb = rpcrdma_alloc_regbuf(RPCRDMA_HDRBUF_SIZE, DMA_TO_DEVICE, flags);
- if (!rb) {
- kfree(req);
- return NULL;
- }
+ if (!rb)
+ goto out2;
req->rl_rdmabuf = rb;
xdr_buf_init(&req->rl_hdrbuf, rdmab_data(rb), rdmab_length(rb));
+
+ req->rl_sendbuf = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
+ if (!req->rl_sendbuf)
+ goto out3;
+
+ req->rl_recvbuf = rpcrdma_alloc_regbuf(size, DMA_NONE, flags);
+ if (!req->rl_recvbuf)
+ goto out4;
+
req->rl_buffer = buffer;
INIT_LIST_HEAD(&req->rl_registered);
-
spin_lock(&buffer->rb_lock);
list_add(&req->rl_all, &buffer->rb_allreqs);
spin_unlock(&buffer->rb_lock);
return req;
+
+out4:
+ kfree(req->rl_sendbuf);
+out3:
+ kfree(req->rl_rdmabuf);
+out2:
+ kfree(req);
+out1:
+ return NULL;
}
static bool rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt, bool temp)
@@ -1090,7 +1107,8 @@ static bool rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt, bool temp)
for (i = 0; i < buf->rb_max_requests; i++) {
struct rpcrdma_req *req;
- req = rpcrdma_req_create(r_xprt, GFP_KERNEL);
+ req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE,
+ GFP_KERNEL);
if (!req)
goto out;
list_add(&req->rl_list, &buf->rb_send_bufs);
@@ -529,7 +529,7 @@ int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
/*
* Buffer calls - xprtrdma/verbs.c
*/
-struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt,
+struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
gfp_t flags);
void rpcrdma_req_destroy(struct rpcrdma_req *req);
int rpcrdma_buffer_create(struct rpcrdma_xprt *);
Allocating an rpcrdma_req's regbufs at xprt create time enables a pair of micro-optimizations: First, if these regbufs are always there, we can eliminate two conditional branches from the hot xprt_rdma_allocate path. Second, by allocating a 1KB buffer, it places a lower bound on the size of these buffers, without adding yet another conditional branch. The lower bound reduces the number of hardway re- allocations. In fact, for some workloads it completely eliminates hardway allocations. Signed-off-by: Chuck Lever <chuck.lever@oracle.com> --- net/sunrpc/xprtrdma/backchannel.c | 18 ++++-------------- net/sunrpc/xprtrdma/transport.c | 4 ++-- net/sunrpc/xprtrdma/verbs.c | 34 ++++++++++++++++++++++++++-------- net/sunrpc/xprtrdma/xprt_rdma.h | 2 +- 4 files changed, 33 insertions(+), 25 deletions(-)