On Fri, Mar 12, 2021 at 10:47:53PM +0800, lyl2019@xxxxxxxxxxxxxxxx wrote:
>
>
>
> > -----原始邮件-----
> > 发件人: "Tom Parkin" <tparkin@xxxxxxxxxxx>
> > 发送时间: 2021-03-12 18:12:58 (星期五)
> > 收件人: lyl2019@xxxxxxxxxxxxxxxx
> > 抄送: paulus@xxxxxxxxx, davem@xxxxxxxxxxxxx, linux-ppp@xxxxxxxxxxxxxxx, netdev@xxxxxxxxxxxxxxx, linux-kernel@xxxxxxxxxxxxxxx
> > 主题: Re: [BUG] net/ppp: A use after free in ppp_unregister_channe
> >
> > Thanks for the report!
> >
> > On Thu, Mar 11, 2021 at 20:34:44 +0800, lyl2019@xxxxxxxxxxxxxxxx wrote:
> > > File: drivers/net/ppp/ppp_generic.c
> > >
> > > In ppp_unregister_channel, pch could be freed in ppp_unbridge_channels()
> > > but after that pch is still in use. Inside the function ppp_unbridge_channels,
> > > if "pchbb == pch" is true and then pch will be freed.
> >
> > Do you have a way to reproduce a use-after-free scenario?
> >
> > From static analysis I'm not sure how pch would be freed in
> > ppp_unbridge_channels when called via. ppp_unregister_channel.
> >
> > In theory (at least!) the caller of ppp_register_net_channel holds
> > a reference on struct channel which ppp_unregister_channel drops.
> >
> > Each channel in a bridged pair holds a reference on the other.
> >
> > Hence on return from ppp_unbridge_channels, the channel should not have
> > been freed (in this code path) because the ppp_register_net_channel
> > reference has not yet been dropped.
> >
> > Maybe there is an issue with the reference counting or a race of some
> > sort?
> >
> > > I checked the commit history and found that this problem is introduced from
> > > 4cf476ced45d7 ("ppp: add PPPIOCBRIDGECHAN and PPPIOCUNBRIDGECHAN ioctls").
> > >
> > > I have no idea about how to generate a suitable patch, sorry.
>
> This issue was reported by a path-sensitive static analyzer developed by our Lab,
> thus i have not a crash or bug log.
>
> As the return type of ppp_unbridge_channels() is a int, can we return a value to
> inform caller that the channel is freed?
I don't think this is going to improve anything, as
ppp_unregister_channel() couldn't take any corrective action anyway.
