We allocate extra metadata SKBs in case of a zerocopy send. This metadata
memory is accounted for in the OPTMEM limit. If there is any error while
sending zerocopy packets or if zerocopy is skipped, these metadata SKBs are
queued in the socket error queue. This error queue is freed when userspace
reads it.
Usually, if there are continuous failures, we merge the metadata into a single
SKB and free another one. As a result, it never exceeds the OPTMEM limit.
However, if there is any out-of-order processing or intermittent zerocopy
failures, this error chain can grow significantly, exhausting the OPTMEM limit.
As a result, all new sendmsg requests fail to allocate any new SKB, leading to
an ENOBUF error. Depending on the amount of data queued before the flush
(i.e., large live migration iterations), even large OPTMEM limits are prone to
failure.
To work around this, if we encounter an ENOBUF error with a zerocopy sendmsg,
we flush the error queue and retry once more.
Signed-off-by: Manish Mishra <manish.mis...@nutanix.com>
---
include/io/channel-socket.h | 5 +++
io/channel-socket.c | 74 ++++++++++++++++++++++++++++++-------
2 files changed, 65 insertions(+), 14 deletions(-)
V2:
1. Removed the dirty_sync_missed_zero_copy migration stat.
2. Made the call to qio_channel_socket_flush_internal() from
qio_channel_socket_writev() non-blocking.
V3:
1. Add the dirty_sync_missed_zero_copy migration stat again.
V4:
1. Minor nit to rename s/zero_copy_flush_pending/zerocopy_flushed_once.
diff --git a/include/io/channel-socket.h b/include/io/channel-socket.h
index ab15577d38..2c48b972e8 100644
--- a/include/io/channel-socket.h
+++ b/include/io/channel-socket.h
@@ -49,6 +49,11 @@ struct QIOChannelSocket {
socklen_t remoteAddrLen;
ssize_t zero_copy_queued;
ssize_t zero_copy_sent;
+ /**
+ * This flag indicates whether any new data was successfully sent with
+ * zerocopy since the last qio_channel_socket_flush() call.
+ */
+ bool new_zero_copy_sent_success;
};
diff --git a/io/channel-socket.c b/io/channel-socket.c
index 608bcf066e..d5882c16fe 100644
--- a/io/channel-socket.c
+++ b/io/channel-socket.c
@@ -37,6 +37,12 @@
#define SOCKET_MAX_FDS 16
+#ifdef QEMU_MSG_ZEROCOPY
+static int qio_channel_socket_flush_internal(QIOChannel *ioc,
+ bool block,
+ Error **errp);
+#endif
+
SocketAddress *
qio_channel_socket_get_local_address(QIOChannelSocket *ioc,
Error **errp)
@@ -65,6 +71,7 @@ qio_channel_socket_new(void)
sioc->fd = -1;
sioc->zero_copy_queued = 0;
sioc->zero_copy_sent = 0;
+ sioc->new_zero_copy_sent_success = FALSE;
ioc = QIO_CHANNEL(sioc);
qio_channel_set_feature(ioc, QIO_CHANNEL_FEATURE_SHUTDOWN);
@@ -566,6 +573,7 @@ static ssize_t qio_channel_socket_writev(QIOChannel *ioc,
size_t fdsize = sizeof(int) * nfds;
struct cmsghdr *cmsg;
int sflags = 0;
+ bool zerocopy_flushed_once = FALSE;
memset(control, 0, CMSG_SPACE(sizeof(int) * SOCKET_MAX_FDS));
@@ -612,9 +620,25 @@ static ssize_t qio_channel_socket_writev(QIOChannel *ioc,
goto retry;
case ENOBUFS:
if (flags & QIO_CHANNEL_WRITE_FLAG_ZERO_COPY) {
- error_setg_errno(errp, errno,
- "Process can't lock enough memory for using
MSG_ZEROCOPY");
- return -1;
+ /**
+ * Socket error queueing may exhaust the OPTMEM limit. Try
+ * flushing the error queue once.
+ */
+ if (!zerocopy_flushed_once) {
+ ret = qio_channel_socket_flush_internal(ioc, false, errp);
