Allow to allocate pages in the receive buffers lazily
Patch from Olaf Kirch <[email protected]>: Replies to the GETACL remote
procedure call can become quite big, yet in the common case replies will
be very small. This patch checks of argument pages have already been
allocated, and allocates pages up to the maximum length of the xdr_buf
when this is not the case.
Andreas Gruenbacher <[email protected]>, SUSE Labs
Index: linux-2.6.6-rc2/include/linux/sunrpc/xdr.h
===================================================================
--- linux-2.6.6-rc2.orig/include/linux/sunrpc/xdr.h
+++ linux-2.6.6-rc2/include/linux/sunrpc/xdr.h
@@ -169,7 +169,7 @@ typedef struct {
typedef size_t (*skb_read_actor_t)(skb_reader_t *desc, void *to, size_t len);
-extern void xdr_partial_copy_from_skb(struct xdr_buf *, unsigned int,
+extern int xdr_partial_copy_from_skb(struct xdr_buf *, unsigned int,
skb_reader_t *, skb_read_actor_t);
struct socket;
Index: linux-2.6.6-rc2/net/sunrpc/xdr.c
===================================================================
--- linux-2.6.6-rc2.orig/net/sunrpc/xdr.c
+++ linux-2.6.6-rc2/net/sunrpc/xdr.c
@@ -303,7 +303,7 @@ void xdr_kunmap(struct xdr_buf *xdr, siz
}
}
-void
+int
xdr_partial_copy_from_skb(struct xdr_buf *xdr, unsigned int base,
skb_reader_t *desc,
skb_read_actor_t copy_actor)
@@ -317,7 +317,7 @@ xdr_partial_copy_from_skb(struct xdr_buf
len -= base;
ret = copy_actor(desc, (char *)xdr->head[0].iov_base + base, len);
if (ret != len || !desc->count)
- return;
+ return 0;
base = 0;
} else
base -= len;
@@ -337,6 +337,13 @@ xdr_partial_copy_from_skb(struct xdr_buf
do {
char *kaddr;
+ /* ACL likes to be lazy in allocating pages - ACLs
+ * are small by default but can get huge. */
+ if (unlikely(*ppage == NULL)) {
+ if (!(*ppage = alloc_page(GFP_ATOMIC)))
+ return -ENOMEM;
+ }
+
len = PAGE_CACHE_SIZE;
kaddr = kmap_atomic(*ppage, KM_SKB_SUNRPC_DATA);
if (base) {
@@ -353,13 +360,15 @@ xdr_partial_copy_from_skb(struct xdr_buf
flush_dcache_page(*ppage);
kunmap_atomic(kaddr, KM_SKB_SUNRPC_DATA);
if (ret != len || !desc->count)
- return;
+ return 0;
ppage++;
} while ((pglen -= len) != 0);
copy_tail:
len = xdr->tail[0].iov_len;
if (base < len)
copy_actor(desc, (char *)xdr->tail[0].iov_base + base, len - base);
+
+ return 0;
}
Index: linux-2.6.6-rc2/net/sunrpc/xprt.c
===================================================================
--- linux-2.6.6-rc2.orig/net/sunrpc/xprt.c
+++ linux-2.6.6-rc2/net/sunrpc/xprt.c
@@ -693,7 +693,8 @@ csum_partial_copy_to_xdr(struct xdr_buf
goto no_checksum;
desc.csum = csum_partial(skb->data, desc.offset, skb->csum);
- xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_and_csum_bits);
+ if (xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_and_csum_bits) < 0)
+ return -1;
if (desc.offset != skb->len) {
unsigned int csum2;
csum2 = skb_checksum(skb, desc.offset, skb->len - desc.offset, 0);
@@ -705,7 +706,8 @@ csum_partial_copy_to_xdr(struct xdr_buf
return -1;
return 0;
no_checksum:
- xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_bits);
+ if (xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_bits) < 0)
+ return -1;
if (desc.count)
return -1;
return 0;
@@ -872,6 +874,7 @@ tcp_read_request(struct rpc_xprt *xprt,
struct rpc_rqst *req;
struct xdr_buf *rcvbuf;
size_t len;
+ int r;
/* Find and lock the request corresponding to this xid */
spin_lock(&xprt->sock_lock);
@@ -892,16 +895,30 @@ tcp_read_request(struct rpc_xprt *xprt,
len = xprt->tcp_reclen - xprt->tcp_offset;
memcpy(&my_desc, desc, sizeof(my_desc));
my_desc.count = len;
- xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
- &my_desc, tcp_copy_data);
+ r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
+ &my_desc, tcp_copy_data);
desc->count -= len;
desc->offset += len;
} else
- xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
- desc, tcp_copy_data);
+ r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
+ desc, tcp_copy_data);
xprt->tcp_copied += len;
xprt->tcp_offset += len;
+ if (r < 0) {
+ /* Error when copying to the receive buffer,
+ * usually because we weren't able to allocate
+ * additional buffer pages. All we can do now
+ * is turn off XPRT_COPY_DATA, so the request
+ * will not receive any additional updates,
+ * and time out.
+ * Any remaining data from this record will
+ * be discarded.
+ */
+ xprt->tcp_flags &= ~XPRT_COPY_DATA;
+ goto out;
+ }
+
if (xprt->tcp_copied == req->rq_private_buf.buflen)
xprt->tcp_flags &= ~XPRT_COPY_DATA;
else if (xprt->tcp_offset == xprt->tcp_reclen) {
@@ -914,6 +931,7 @@ tcp_read_request(struct rpc_xprt *xprt,
req->rq_task->tk_pid);
xprt_complete_rqst(xprt, req, xprt->tcp_copied);
}
+out:
spin_unlock(&xprt->sock_lock);
tcp_check_recm(xprt);
}
On Mon, 2004-04-26 at 06:28, Andreas Gruenbacher wrote:
> Allow to allocate pages in the receive buffers lazily
>
> Patch from Olaf Kirch <[email protected]>: Replies to the GETACL remote
> procedure call can become quite big, yet in the common case replies will
> be very small. This patch checks of argument pages have already been
> allocated, and allocates pages up to the maximum length of the xdr_buf
> when this is not the case.
AFAICS, there is nothing to stop xdr_partial_copy_from_skb() from
writing beyond the end of xdr->pages[]. How do you propose to prevent
this?
Cheers,
Trond
On Mon, 2004-04-26 at 22:38, Trond Myklebust wrote:
> On Mon, 2004-04-26 at 06:28, Andreas Gruenbacher wrote:
> > Allow to allocate pages in the receive buffers lazily
> >
> > Patch from Olaf Kirch <[email protected]>: Replies to the GETACL remote
> > procedure call can become quite big, yet in the common case replies will
> > be very small. This patch checks of argument pages have already been
> > allocated, and allocates pages up to the maximum length of the xdr_buf
> > when this is not the case.
>
> AFAICS, there is nothing to stop xdr_partial_copy_from_skb() from
> writing beyond the end of xdr->pages[]. How do you propose to prevent
> this?
xdr->page_len determines the maximum length. The pages array must
contain enough entries to hold that many pages; this is no different
from pre-allocating the pages.
Cheers,
--
Andreas Gruenbacher <[email protected]>
SUSE Labs, SUSE LINUX AG