2014-07-15 14:11:38

by Chuck Lever III

Subject: [PATCH] svcrdma: Add zero padding if the client doesn't send it

See RFC 5666 section 3.7: clients don't have to send zero XDR
padding.

BugLink: https://bugzilla.linux-nfs.org/show_bug.cgi?id=246
Signed-off-by: Chuck Lever <[email protected]>
---
Hi Bruce-

This is an alternative solution to changing the sanity check in the
XDR WRITE decoder. It adjusts the incoming xdr_buf to include a zero
pad just after the transport has received each RPC request.

Thoughts?

net/sunrpc/xprtrdma/svc_rdma_recvfrom.c | 30 ++++++++++++++++++++++++++++++
1 files changed, 30 insertions(+), 0 deletions(-)

diff --git a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
index 8f92a61..9a3465d 100644
--- a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
+++ b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
@@ -43,6 +43,7 @@
#include <linux/sunrpc/debug.h>
#include <linux/sunrpc/rpc_rdma.h>
#include <linux/spinlock.h>
+#include <linux/highmem.h>
#include <asm/unaligned.h>
#include <rdma/ib_verbs.h>
#include <rdma/rdma_cm.h>
@@ -435,6 +436,34 @@ static int rdma_read_chunks(struct svcxprt_rdma *xprt,
return ret;
}

+/*
+ * To avoid a separate RDMA READ just for a handful of zero bytes,
+ * RFC 5666 section 3.7 allows the client to omit the XDR zero pad
+ * in chunk lists.
+ */
+static void
+rdma_fix_xdr_pad(struct xdr_buf *buf)
+{
+ unsigned int page_len = buf->page_len;
+ unsigned int size = (XDR_QUADLEN(page_len) << 2) - page_len;
+ unsigned int offset, pg_no;
+ char *p;
+
+ if (size == 0)
+ return;
+
+ offset = page_len & ~PAGE_MASK;
+ pg_no = page_len >> PAGE_SHIFT;
+
+ p = kmap_atomic(buf->pages[pg_no]);
+ memset(p + offset, 0, size);
+ kunmap_atomic(p);
+
+ buf->page_len += size;
+ buf->buflen += size;
+ buf->len += size;
+}
+
static int rdma_read_complete(struct svc_rqst *rqstp,
struct svc_rdma_op_ctxt *head)
{
@@ -449,6 +478,7 @@ static int rdma_read_complete(struct svc_rqst *rqstp,
rqstp->rq_pages[page_no] = head->pages[page_no];
}
/* Point rq_arg.pages past header */
+ rdma_fix_xdr_pad(&head->arg);
rqstp->rq_arg.pages = &rqstp->rq_pages[head->hdr_count];
rqstp->rq_arg.page_len = head->arg.page_len;
rqstp->rq_arg.page_base = head->arg.page_base;

2014-07-15 14:28:32

by J. Bruce Fields

Subject: Re: [PATCH] svcrdma: Add zero padding if the client doesn't send it

On Tue, Jul 15, 2014 at 10:11:34AM -0400, Chuck Lever wrote:
> See RFC 5666 section 3.7: clients don't have to send zero XDR
> padding.
>
> BugLink: https://bugzilla.linux-nfs.org/show_bug.cgi?id=246
> Signed-off-by: Chuck Lever <[email protected]>
> ---
> Hi Bruce-
>
> This is an alternative solution to changing the sanity check in the
> XDR WRITE decoder. It adjusts the incoming xdr_buf to include a zero
> pad just after the transport has received each RPC request.
>
> Thoughts?

That looks pretty simple. So if this works then I think I'd be happier
doing this than worrying about other spots like the drc code where we
might have missed an assumption.

> net/sunrpc/xprtrdma/svc_rdma_recvfrom.c | 30 ++++++++++++++++++++++++++++++
> 1 files changed, 30 insertions(+), 0 deletions(-)
>
> diff --git a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
> index 8f92a61..9a3465d 100644
> --- a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
> +++ b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
> @@ -43,6 +43,7 @@
> #include <linux/sunrpc/debug.h>
> #include <linux/sunrpc/rpc_rdma.h>
> #include <linux/spinlock.h>
> +#include <linux/highmem.h>
> #include <asm/unaligned.h>
> #include <rdma/ib_verbs.h>
> #include <rdma/rdma_cm.h>
> @@ -435,6 +436,34 @@ static int rdma_read_chunks(struct svcxprt_rdma *xprt,
> return ret;
> }
>
> +/*
> + * To avoid a separate RDMA READ just for a handful of zero bytes,
> + * RFC 5666 section 3.7 allows the client to omit the XDR zero pad
> + * in chunk lists.
> + */
> +static void
> +rdma_fix_xdr_pad(struct xdr_buf *buf)
> +{
> + unsigned int page_len = buf->page_len;
> + unsigned int size = (XDR_QUADLEN(page_len) << 2) - page_len;
> + unsigned int offset, pg_no;
> + char *p;
> +
> + if (size == 0)
> + return;
> +
> + offset = page_len & ~PAGE_MASK;
> + pg_no = page_len >> PAGE_SHIFT;
> +
> + p = kmap_atomic(buf->pages[pg_no]);
> + memset(p + offset, 0, size);
> + kunmap_atomic(p);

If these are pages alloc'd by svc_recv() with alloc_page(GFP_KERNEL)
then the kmap/kunmap shouldn't be necessary.

--b.

> +
> + buf->page_len += size;
> + buf->buflen += size;
> + buf->len += size;
> +}
> +
> static int rdma_read_complete(struct svc_rqst *rqstp,
> struct svc_rdma_op_ctxt *head)
> {
> @@ -449,6 +478,7 @@ static int rdma_read_complete(struct svc_rqst *rqstp,
> rqstp->rq_pages[page_no] = head->pages[page_no];
> }
> /* Point rq_arg.pages past header */
> + rdma_fix_xdr_pad(&head->arg);
> rqstp->rq_arg.pages = &rqstp->rq_pages[head->hdr_count];
> rqstp->rq_arg.page_len = head->arg.page_len;
> rqstp->rq_arg.page_base = head->arg.page_base;
>