Re: [PATCH] hv_netvsc: Make sure out channel is fully opened on send
flat view
From: Stephen Hemminger <stephen@networkplumber.org>
Date: 2018-03-13 19:35:21
Also in:
lkml
On Tue, 13 Mar 2018 20:06:50 +0100 Mohammed Gamal [off-list ref] wrote:
quoted hunk ↗ jump to hunk
Dring high network traffic changes to network interface parameters such as number of channels or MTU can cause a kernel panic with a NULL pointer dereference. This is due to netvsc_device_remove() being called and deallocating the channel ring buffers, which can then be accessed by netvsc_send_pkt() before they're allocated on calling netvsc_device_add() The patch fixes this problem by checking the channel state and returning ENODEV if not yet opened. We also move the call to hv_ringbuf_avail_percent() which may access the uninitialized ring buffer. Signed-off-by: Mohammed Gamal <redacted> --- drivers/net/hyperv/netvsc.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)diff --git a/drivers/net/hyperv/netvsc.c b/drivers/net/hyperv/netvsc.c index 0265d70..44a8358 100644 --- a/drivers/net/hyperv/netvsc.c +++ b/drivers/net/hyperv/netvsc.c@@ -757,7 +757,7 @@ static inline int netvsc_send_pkt( struct netdev_queue *txq = netdev_get_tx_queue(ndev, packet->q_idx); u64 req_id; int ret; - u32 ring_avail = hv_ringbuf_avail_percent(&out_channel->outbound); + u32 ring_avail; nvmsg.hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT; if (skb)@@ -773,7 +773,7 @@ static inline int netvsc_send_pkt( req_id = (ulong)skb; - if (out_channel->rescind) + if (out_channel->rescind || out_channel->state != CHANNEL_OPENED_STATE) return -ENODEV; if (packet->page_buf_cnt) {@@ -791,6 +791,7 @@ static inline int netvsc_send_pkt( VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED); } + ring_avail = hv_ringbuf_avail_percent(&out_channel->outbound); if (ret == 0) { atomic_inc_return(&nvchan->queue_sends);
Thanks for your patch. Yes there are races with the current update logic. The root cause goes higher up in the flow; the send queues should be stopped before netvsc_device_remove is called. Solving it where you tried to is racy and not going to work reliably. Network patches should go to netdev@vger.kernel.org You can't move the ring_avail check until after the vmbus_sendpacket because that will break the flow control logic. Instead, you should just move the avail_read check until just after the existing rescind check. Also, you shouldn't need to check for OPENED_STATE, just rescind is enough.