Re: KASAN: use-after-free Read in vgem_gem_dumb_create

[Christian KÃnig]

Am 03.02.20 um 10:06 schrieb Dan Carpenter:
> On Sun, Feb 02, 2020 at 02:19:18PM +0100, Daniel Vetter wrote:
> > On Fri, Jan 31, 2020 at 11:28 PM syzbot
> > <syzbot+0dc4444774d419e916c8@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
> > > Hello,
> > >
> > > syzbot found the following crash on:
> > >
> > > HEAD commit:    39bed42d Merge tag 'for-linus-hmm' of git://git.kernel.org..
> > > git tree:       upstream
> > > console output: https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fsyzkaller.appspot.com%2Fx%2Flog.txt%3Fx%3D179465bee00000&amp;data=02%7C01%7Cchristian.koenig%40amd.com%7C529f2273b8374f38560108d7a88862eb%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637163176051177627&amp;sdata=3goGqBs4%2BjkjCeV2bX5VTB%2F1PRLEP5bzq5Ec%2BN7fKHs%3D&amp;reserved=0
> > > kernel config:  https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fsyzkaller.appspot.com%2Fx%2F.config%3Fx%3D2646535f8818ae25&amp;data=02%7C01%7Cchristian.koenig%40amd.com%7C529f2273b8374f38560108d7a88862eb%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637163176051177627&amp;sdata=SnlKln%2FAG%2BVRVjSrOSJjUE%2BhSDf35wTqzWLCAyGQVss%3D&amp;reserved=0
> > > dashboard link: https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fsyzkaller.appspot.com%2Fbug%3Fextid%3D0dc4444774d419e916c8&amp;data=02%7C01%7Cchristian.koenig%40amd.com%7C529f2273b8374f38560108d7a88862eb%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637163176051177627&amp;sdata=33EJNAWjTm6Edi1J0oPBfs8epb%2BQ2cpAKlzl1sT40CQ%3D&amp;reserved=0
> > > compiler:       gcc (GCC) 9.0.0 20181231 (experimental)
> > > syz repro:      https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fsyzkaller.appspot.com%2Fx%2Frepro.syz%3Fx%3D16251279e00000&amp;data=02%7C01%7Cchristian.koenig%40amd.com%7C529f2273b8374f38560108d7a88862eb%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637163176051177627&amp;sdata=zmUyyp7znqQfLzzNZ80bNgCILAjeMeCVVr7xf7CHaWk%3D&amp;reserved=0
> > >
> > > The bug was bisected to:
> > >
> > > commit 7611750784664db46d0db95631e322aeb263dde7
> > > Author: Alex Deucher <alexander.deucher@xxxxxxx>
> > > Date:   Wed Jun 21 16:31:41 2017 +0000
> > >
> > >      drm/amdgpu: use kernel is_power_of_2 rather than local version
> > >
> > > bisection log:  https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fsyzkaller.appspot.com%2Fx%2Fbisect.txt%3Fx%3D11628df1e00000&amp;data=02%7C01%7Cchristian.koenig%40amd.com%7C529f2273b8374f38560108d7a88862eb%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637163176051177627&amp;sdata=5QpTG4iU%2FOt22L3jxRbNxtVPZZ2EvBAcFGZdqVnVCbU%3D&amp;reserved=0
> > > final crash:    https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fsyzkaller.appspot.com%2Fx%2Freport.txt%3Fx%3D13628df1e00000&amp;data=02%7C01%7Cchristian.koenig%40amd.com%7C529f2273b8374f38560108d7a88862eb%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637163176051177627&amp;sdata=hN6UZnFR2nIMPMspjIF7S82oXstaRl%2BLAzmz5yujPac%3D&amp;reserved=0
> > > console output: https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fsyzkaller.appspot.com%2Fx%2Flog.txt%3Fx%3D15628df1e00000&amp;data=02%7C01%7Cchristian.koenig%40amd.com%7C529f2273b8374f38560108d7a88862eb%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637163176051177627&amp;sdata=LHXMANOURDv3EsqTSvHSBZnPEzGQoJU1RbeqYExCaGk%3D&amp;reserved=0
> > >
> > > IMPORTANT: if you fix the bug, please add the following tag to the commit:
> > > Reported-by: syzbot+0dc4444774d419e916c8@xxxxxxxxxxxxxxxxxxxxxxxxx
> > > Fixes: 761175078466 ("drm/amdgpu: use kernel is_power_of_2 rather than local version")
> > Aside: This bisect line is complete nonsense ... I'm kinda at the
> > point where I'm assuming that syzbot bisect results are garbage, which
> > is maybe not what we want. I guess much stricter filtering for noise
> > is needed, dunno.
> > -Danile
> With race conditions the git bisect is often nonsense.

Which makes sense, but we can still try to sanitize the result. I'm not 
familiar with the test case, but I think it doesn't even compile the 
amdgpu driver.

So skipping all patches of stuff you don't even compile would make not 
only the result of bisecting quite a bit more reliable, but also speed 
the process up quite a bit.

But no good idea to how teach that to a compile bot or the git bisect 
command.

Regards,
Christian.

> regards,
> dan carpenter