Skip to contentSkip to conversation

libceph: Partially revert changes to support MSG_SPLICE_PAGES

PATCH Superseded David Howells 2 participants 2 messages Jun 26, 2023 – Jun 27, 2023

Subsystems (3)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 libceph: Partially revert changes to support MSG_SPLICE_PAGES

Fix the mishandling of MSG_DONTWAIT and also reinstates the per-page checking of the source pages (which might have come from a DIO write by userspace) by partially reverting the changes to support MSG_SPLICE_PAGES and doing things a little differently. In messenger_v1:

(1) The ceph_tcp_sendpage() is resurrected and the callers reverted to use that.

(2) The callers now pass MSG_MORE unconditionally. Previously, they were passing in MSG_MORE|MSG_SENDPAGE_NOTLAST and then degrading that to just MSG_MORE on the last call to ->sendpage().

(3) Make ceph_tcp_sendpage() a wrapper around sendmsg() rather than sendpage(), setting MSG_SPLICE_PAGES if sendpage_ok() returns true on the page.

In messenger_v2:

(4) Bring back do_try_sendpage() and make the callers use that.

(5) Make do_try_sendpage() use sendmsg() for both cases and set MSG_SPLICE_PAGES if sendpage_ok() is set.

Fixes 40a8c17aa770 ("ceph: Use sendmsg(MSG_SPLICE_PAGES) rather than sendpage")
Fixes fa094ccae1e7 ("ceph: Use sendmsg(MSG_SPLICE_PAGES) rather than sendpage()")
cc Ilya Dryomov <[email protected]>
cc Xiubo Li <[email protected]>
cc Jeff Layton <[email protected]>
cc "David S. Miller" <[email protected]>
cc Eric Dumazet <[email protected]>
cc Jakub Kicinski <[email protected]>
cc Paolo Abeni <[email protected]>
cc Jens Axboe <[email protected]>
cc Matthew Wilcox <[email protected]>

---
Notes:
ver #2)
- Removed mention of MSG_SENDPAGE_NOTLAST in comments.
- Changed some refs to sendpage to MSG_SPLICE_PAGES in comments.
- Init msg_iter in ceph_tcp_sendpage().
- Move setting of MSG_SPLICE_PAGES in do_try_sendpage() next to comment and adjust how it is cleared.

 net/ceph/messenger_v1.c |   58 ++++++++++++++++++++-----------
 net/ceph/messenger_v2.c |   88 ++++++++++++++++++++++++++++++++++++++----------
 2 files changed, 107 insertions(+), 39 deletions(-)
net/ceph/messenger_v1.c +37 -21
index 814579f27f04..51a6f28aa798 100644
@@ -74,6 +74,39 @@ static int ceph_tcp_sendmsg(struct socket *sock, struct kvec *iov,
 	return r;
 }
 
+/*
+ * @more: MSG_MORE or 0.
+ */
+static int ceph_tcp_sendpage(struct socket *sock, struct page *page,
+			     int offset, size_t size, int more)
+{
+	struct msghdr msg = {
+		.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL | more,
+	};
+	struct bio_vec bvec;
+	int ret;
+
+	/*
+	 * MSG_SPLICE_PAGES cannot properly handle pages with page_count == 0,
+	 * we need to fall back to sendmsg if that's the case.
+	 *
+	 * Same goes for slab pages: skb_can_coalesce() allows
+	 * coalescing neighboring slab objects into a single frag which
+	 * triggers one of hardened usercopy checks.
+	 */
+	if (sendpage_ok(page))
+		msg.msg_flags |= MSG_SPLICE_PAGES;
+
+	bvec_set_page(&bvec, page, size, offset);
+	iov_iter_bvec(&msg.msg_iter, ITER_DEST, &bvec, 1, size);
+
+	ret = sock_sendmsg(sock, &msg);
+	if (ret == -EAGAIN)
+		ret = 0;
+
+	return ret;
+}
+
 static void con_out_kvec_reset(struct ceph_connection *con)
 {
 	BUG_ON(con->v1.out_skip);
@@ -450,10 +483,6 @@ static int write_partial_message_data(struct ceph_connection *con)
 	 */
 	crc = do_datacrc ? le32_to_cpu(msg->footer.data_crc) : 0;
 	while (cursor->total_resid) {
-		struct bio_vec bvec;
-		struct msghdr msghdr = {
-			.msg_flags = MSG_SPLICE_PAGES,
-		};
 		struct page *page;
 		size_t page_offset;
 		size_t length;
@@ -465,13 +494,8 @@ static int write_partial_message_data(struct ceph_connection *con)
 		}
 
 		page = ceph_msg_data_next(cursor, &page_offset, &length);
-		if (length != cursor->total_resid)
-			msghdr.msg_flags |= MSG_MORE;
-
-		bvec_set_page(&bvec, page, length, page_offset);
-		iov_iter_bvec(&msghdr.msg_iter, ITER_SOURCE, &bvec, 1, length);
-
-		ret = sock_sendmsg(con->sock, &msghdr);
+		ret = ceph_tcp_sendpage(con->sock, page, page_offset, length,
+					MSG_MORE);
 		if (ret <= 0) {
 			if (do_datacrc)
 				msg->footer.data_crc = cpu_to_le32(crc);
@@ -501,22 +525,14 @@ static int write_partial_message_data(struct ceph_connection *con)
  */
 static int write_partial_skip(struct ceph_connection *con)
 {
-	struct bio_vec bvec;
-	struct msghdr msghdr = {
-		.msg_flags = MSG_SPLICE_PAGES | MSG_MORE,
-	};
 	int ret;
 
 	dout("%s %p %d left\n", __func__, con, con->v1.out_skip);
 	while (con->v1.out_skip > 0) {
 		size_t size = min(con->v1.out_skip, (int)PAGE_SIZE);
 
-		if (size == con->v1.out_skip)
-			msghdr.msg_flags &= ~MSG_MORE;
-		bvec_set_page(&bvec, ZERO_PAGE(0), size, 0);
-		iov_iter_bvec(&msghdr.msg_iter, ITER_SOURCE, &bvec, 1, size);
-
-		ret = sock_sendmsg(con->sock, &msghdr);
+		ret = ceph_tcp_sendpage(con->sock, ceph_zero_page, 0, size,
+					MSG_MORE);
 		if (ret <= 0)
 			goto out;
 		con->v1.out_skip -= ret;
net/ceph/messenger_v2.c +70 -18

108 lines not on this page · raw message

maintainer
#1 lore raw
Quoted 78 lines · David Howells

Fix the mishandling of MSG_DONTWAIT and also reinstates the per-page checking of the source pages (which might have come from a DIO write by userspace) by partially reverting the changes to support MSG_SPLICE_PAGES and doing things a little differently. In messenger_v1:

(1) The ceph_tcp_sendpage() is resurrected and the callers reverted to use that.

(2) The callers now pass MSG_MORE unconditionally. Previously, they were passing in MSG_MORE|MSG_SENDPAGE_NOTLAST and then degrading that to just MSG_MORE on the last call to ->sendpage().

(3) Make ceph_tcp_sendpage() a wrapper around sendmsg() rather than sendpage(), setting MSG_SPLICE_PAGES if sendpage_ok() returns true on the page.

In messenger_v2:

(4) Bring back do_try_sendpage() and make the callers use that.

(5) Make do_try_sendpage() use sendmsg() for both cases and set MSG_SPLICE_PAGES if sendpage_ok() is set.

Fixes 40a8c17aa770 ("ceph: Use sendmsg(MSG_SPLICE_PAGES) rather than sendpage")
Fixes fa094ccae1e7 ("ceph: Use sendmsg(MSG_SPLICE_PAGES) rather than sendpage()")
cc Ilya Dryomov <[email protected]>
cc Xiubo Li <[email protected]>
cc Jeff Layton <[email protected]>
cc "David S. Miller" <[email protected]>
cc Eric Dumazet <[email protected]>
cc Jakub Kicinski <[email protected]>
cc Paolo Abeni <[email protected]>
cc Jens Axboe <[email protected]>
cc Matthew Wilcox <[email protected]>

---
Notes:
ver #2)
- Removed mention of MSG_SENDPAGE_NOTLAST in comments.
- Changed some refs to sendpage to MSG_SPLICE_PAGES in comments.
- Init msg_iter in ceph_tcp_sendpage().
- Move setting of MSG_SPLICE_PAGES in do_try_sendpage() next to comment and adjust how it is cleared.

 net/ceph/messenger_v1.c |   58 ++++++++++++++++++++-----------
 net/ceph/messenger_v2.c |   88 ++++++++++++++++++++++++++++++++++++++----------
 2 files changed, 107 insertions(+), 39 deletions(-)
net/ceph/messenger_v1.c +15 -0
index 814579f27f04..51a6f28aa798 100644
@@ -74,6 +74,39 @@ static int ceph_tcp_sendmsg(struct socket *sock, struct kvec *iov,
        return r;
 }
 
+/*
+ * @more: MSG_MORE or 0.
+ */
+static int ceph_tcp_sendpage(struct socket *sock, struct page *page,
+                            int offset, size_t size, int more)
+{
+       struct msghdr msg = {
+               .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL | more,
+       };
+       struct bio_vec bvec;
+       int ret;
+
+       /*
+        * MSG_SPLICE_PAGES cannot properly handle pages with page_count == 0,
+        * we need to fall back to sendmsg if that's the case.

net/ceph/messenger_v1.c static int ceph_tcp_sendmsg(struct socket *sock, struct kvec *iov, · line 101

Quoted 10 lines
net/ceph/messenger_v1.c +10 -0
+        *
+        * Same goes for slab pages: skb_can_coalesce() allows
+        * coalescing neighboring slab objects into a single frag which
+        * triggers one of hardened usercopy checks.
+        */
+       if (sendpage_ok(page))
+               msg.msg_flags |= MSG_SPLICE_PAGES;
+
+       bvec_set_page(&bvec, page, size, offset);
+       iov_iter_bvec(&msg.msg_iter, ITER_DEST, &bvec, 1, size);

Hi David,

Shouldn't this be ITER_SOURCE?

Thanks,

Ilya