Stephen Hemminger <sthem...@microsoft.com> writes: > Do we need ACCESS_ONCE() here to avoid check/use issues? >
I think we don't: this is the only place in the function where we read the variable so we'll get normal read. We're not trying to syncronize with netvsc_init_buf() as that would require locking, if we read stale NULL value after it was already updated on a different CPU we're fine, we'll just return -EAGAIN. > -----Original Message----- > From: Vitaly Kuznetsov [mailto:vkuzn...@redhat.com] > Sent: Wednesday, October 19, 2016 2:53 PM > To: net...@vger.kernel.org > Cc: Stephen Hemminger <sthem...@microsoft.com>; de...@linuxdriverproject.org; > linux-kernel@vger.kernel.org; KY Srinivasan <k...@microsoft.com>; Haiyang > Zhang <haiya...@microsoft.com> > Subject: [PATCH net-next] hv_netvsc: fix a race between netvsc_send() and > netvsc_init_buf() > > Fix in commit 880988348270 ("hv_netvsc: set nvdev link after populating > chn_table") turns out to be incomplete. A crash in > netvsc_get_next_send_section() is observed on mtu change when the device is > under load. The race I identified is: if we get to netvsc_send() after we set > net_device_ctx->nvdev link in netvsc_device_add() but before we finish > netvsc_connect_vsp()->netvsc_init_buf() send_section_map is not allocated and > we crash. Unfortunately we can't set net_device_ctx->nvdev link after the > netvsc_init_buf() call as during the negotiation we need to receive packets > and on the receive path we check for it. It would probably be possible to > split nvdev into a pair of nvdev_in and nvdev_out links and check them > accordingly in get_outbound_net_device()/ > get_inbound_net_device() but this looks like an overkill. > > Check that send_section_map is allocated in netvsc_send(). > > Signed-off-by: Vitaly Kuznetsov <vkuzn...@redhat.com> > --- > drivers/net/hyperv/netvsc.c | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/drivers/net/hyperv/netvsc.c b/drivers/net/hyperv/netvsc.c index > 720b5fa..e2bfaac 100644 > --- a/drivers/net/hyperv/netvsc.c > +++ b/drivers/net/hyperv/netvsc.c > @@ -888,6 +888,13 @@ int netvsc_send(struct hv_device *device, > if (!net_device) > return -ENODEV; > > + /* We may race with netvsc_connect_vsp()/netvsc_init_buf() and get > + * here before the negotiation with the host is finished and > + * send_section_map may not be allocated yet. > + */ > + if (!net_device->send_section_map) > + return -EAGAIN; > + > out_channel = net_device->chn_table[q_idx]; > > packet->send_buf_index = NETVSC_INVALID_INDEX; > -- > 2.7.4 -- Vitaly
