On 2018?09?28? 01:04, Michael S. Tsirkin wrote:> On Thu, Sep 27, 2018 at 08:22:04PM +0800, Jason Wang wrote: >> The access of vsock is not protected by vhost_vsock_lock. This may >> lead to use after free since vhost_vsock_dev_release() may free the >> pointer at the same time. >> >> Fix this by holding the lock during the access. >> >> Reported-by:syzbot+e3e074963495f92a89ed at syzkaller.appspotmail.com >> Fixes: 16320f363ae1 ("vhost-vsock: add pkt cancel capability") >> Fixes: 433fc58e6bf2 ("VSOCK: Introduce vhost_vsock.ko") >> Cc: Stefan Hajnoczi<stefanha at redhat.com> >> Signed-off-by: Jason Wang<jasowang at redhat.com> > Wow is that really the best we can do?For net/stable, probably yes.> A global lock on a data path > operation?It's already there, and the patch only increase the critical section.> Granted use after free is nasty but Stefan said he sees > a way to fix it using a per socket refcount. He's on vacation > until Oct 4 though ... >Stefan has acked the pacth, so I think it's ok? We can do optimization for -next on top. Thanks
On Fri, Sep 28, 2018 at 07:37:37AM +0800, Jason Wang wrote:> > > On 2018?09?28? 01:04, Michael S. Tsirkin wrote: > > On Thu, Sep 27, 2018 at 08:22:04PM +0800, Jason Wang wrote: > > > The access of vsock is not protected by vhost_vsock_lock. This may > > > lead to use after free since vhost_vsock_dev_release() may free the > > > pointer at the same time. > > > > > > Fix this by holding the lock during the access. > > > > > > Reported-by:syzbot+e3e074963495f92a89ed at syzkaller.appspotmail.com > > > Fixes: 16320f363ae1 ("vhost-vsock: add pkt cancel capability") > > > Fixes: 433fc58e6bf2 ("VSOCK: Introduce vhost_vsock.ko") > > > Cc: Stefan Hajnoczi<stefanha at redhat.com> > > > Signed-off-by: Jason Wang<jasowang at redhat.com> > > Wow is that really the best we can do? > > For net/stable, probably yes. > > > A global lock on a data path > > operation? > > It's already there,&vhost_vsock_lock? were is it takes on data path?> and the patch only increase the critical section. > > > Granted use after free is nasty but Stefan said he sees > > a way to fix it using a per socket refcount. He's on vacation > > until Oct 4 though ... > > > > Stefan has acked the pacth, so I think it's ok? We can do optimization for > -next on top. > > ThanksWell on high SMP serializing can drop performance as much as x100 so I'm not sure it's appropriate - seems to fix a bug but can introduce a regression. Let's see how does a proper fix look first? -- MST
On 2018?09?28? 07:50, Michael S. Tsirkin wrote:> On Fri, Sep 28, 2018 at 07:37:37AM +0800, Jason Wang wrote: >> >> On 2018?09?28? 01:04, Michael S. Tsirkin wrote: >>> On Thu, Sep 27, 2018 at 08:22:04PM +0800, Jason Wang wrote: >>>> The access of vsock is not protected by vhost_vsock_lock. This may >>>> lead to use after free since vhost_vsock_dev_release() may free the >>>> pointer at the same time. >>>> >>>> Fix this by holding the lock during the access. >>>> >>>> Reported-by:syzbot+e3e074963495f92a89ed at syzkaller.appspotmail.com >>>> Fixes: 16320f363ae1 ("vhost-vsock: add pkt cancel capability") >>>> Fixes: 433fc58e6bf2 ("VSOCK: Introduce vhost_vsock.ko") >>>> Cc: Stefan Hajnoczi<stefanha at redhat.com> >>>> Signed-off-by: Jason Wang<jasowang at redhat.com> >>> Wow is that really the best we can do? >> For net/stable, probably yes. >> >>> A global lock on a data path >>> operation? >> It's already there, > &vhost_vsock_lock? were is it takes on data path?Ok, but the current code use list which means a global lock is needed anyway here.> >> and the patch only increase the critical section. >> >>> Granted use after free is nasty but Stefan said he sees >>> a way to fix it using a per socket refcount. He's on vacation >>> until Oct 4 though ... >>> >> Stefan has acked the pacth, so I think it's ok? We can do optimization for >> -next on top. >> >> Thanks > > Well on high SMP serializing can drop performance as much as x100 so I'm > not sure it's appropriate - seems to fix a bug but can introduce a > regression. Let's see how does a proper fix look first? >It looks to me hlist + RCU is better. But I'm not sure it's suitable for -net/-stable. Thanks