libceph: optionally use bounce buffer on recv path in crc mode

Both msgr1 and msgr2 in crc mode are zero copy in the sense that
message data is read from the socket directly into the destination
buffer.  We assume that the destination buffer is stable (i.e. remains
unchanged while it is being read to) though.  Otherwise, CRC errors
ensue:

  libceph: read_partial_message 0000000048edf8ad data crc 1063286393 != exp. 228122706
  libceph: osd1 (1)192.168.122.1:6843 bad crc/signature

  libceph: bad data crc, calculated 57958023, expected 1805382778
  libceph: osd2 (2)192.168.122.1:6876 integrity error, bad crc

Introduce rxbounce option to enable use of a bounce buffer when
receiving message data.  In particular this is needed if a mapped
image is a Windows VM disk, passed to QEMU.  Windows has a system-wide
"dummy" page that may be mapped into the destination buffer (potentially
more than once into the same buffer) by the Windows Memory Manager in
an effort to generate a single large I/O [1][2].  QEMU makes a point of
preserving overlap relationships when cloning I/O vectors, so krbd gets
exposed to this behaviour.

[1] "What Is Really in That MDL?"
    https://docs.microsoft.com/en-us/previous-versions/windows/hardware/design/dn614012(v=vs.85)
[2] https://blogs.msmvps.com/kernelmustard/2005/05/04/dummy-pages/

URL: https://bugzilla.redhat.com/show_bug.cgi?id=1973317
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>

1 files changed, 48 insertions, 6 deletions

diff --git a/net/ceph/messenger_v1.c b/net/ceph/messenger_v1.c
index 2cb5ffdf071a..6b014eca3a13 100644
--- a/net/ceph/messenger_v1.c
+++ b/net/ceph/messenger_v1.c
@@ -992,8 +992,7 @@ static int read_partial_message_section(struct ceph_connection *con,
 
 static int read_partial_msg_data(struct ceph_connection *con)
 {
-	struct ceph_msg *msg = con->in_msg;
-	struct ceph_msg_data_cursor *cursor = &msg->cursor;
+	struct ceph_msg_data_cursor *cursor = &con->in_msg->cursor;
 	bool do_datacrc = !ceph_test_opt(from_msgr(con->msgr), NOCRC);
 	struct page *page;
 	size_t page_offset;
@@ -1001,9 +1000,6 @@ static int read_partial_msg_data(struct ceph_connection *con)
 	u32 crc = 0;
 	int ret;
 
-	if (!msg->num_data_items)
-		return -EIO;
-
 	if (do_datacrc)
 		crc = con->in_data_crc;
 	while (cursor->total_resid) {
@@ -1031,6 +1027,46 @@ static int read_partial_msg_data(struct ceph_connection *con)
 	return 1;	/* must return > 0 to indicate success */
 }
 
+static int read_partial_msg_data_bounce(struct ceph_connection *con)
+{
+	struct ceph_msg_data_cursor *cursor = &con->in_msg->cursor;
+	struct page *page;
+	size_t off, len;
+	u32 crc;
+	int ret;
+
+	if (unlikely(!con->bounce_page)) {
+		con->bounce_page = alloc_page(GFP_NOIO);
+		if (!con->bounce_page) {
+			pr_err("failed to allocate bounce page\n");
+			return -ENOMEM;
+		}
+	}
+
+	crc = con->in_data_crc;
+	while (cursor->total_resid) {
+		if (!cursor->resid) {
+			ceph_msg_data_advance(cursor, 0);
+			continue;
+		}
+
+		page = ceph_msg_data_next(cursor, &off, &len, NULL);
+		ret = ceph_tcp_recvpage(con->sock, con->bounce_page, 0, len);
+		if (ret <= 0) {
+			con->in_data_crc = crc;
+			return ret;
+		}
+
+		crc = crc32c(crc, page_address(con->bounce_page), ret);
+		memcpy_to_page(page, off, page_address(con->bounce_page), ret);
+
+		ceph_msg_data_advance(cursor, ret);
+	}
+	con->in_data_crc = crc;
+
+	return 1;	/* must return > 0 to indicate success */
+}
+
 /*
  * read (part of) a message.
  */
@@ -1141,7 +1177,13 @@ static int read_partial_message(struct ceph_connection *con)
 
 	/* (page) data */
 	if (data_len) {
-		ret = read_partial_msg_data(con);
+		if (!m->num_data_items)
+			return -EIO;
+
+		if (ceph_test_opt(from_msgr(con->msgr), RXBOUNCE))
+			ret = read_partial_msg_data_bounce(con);
+		else
+			ret = read_partial_msg_data(con);
 		if (ret <= 0)
 			return ret;
 	}