On Tue, Nov 20, 2018 at 07:03:17PM +0000, Will Deacon wrote: > On Tue, Nov 20, 2018 at 06:28:54PM +0000, Will Deacon wrote: > > On Sat, Nov 10, 2018 at 11:17:03AM +0000, Jan Glauber wrote: > > > On Fri, Nov 09, 2018 at 03:58:56PM +0000, Will Deacon wrote: > > > > On Fri, Nov 09, 2018 at 02:37:51PM +0000, Jan Glauber wrote: > > > > > I'm seeing the following oops reproducible with upstream kernel on arm64 > > > > > (ThunderX2): > > > > > > > > [...] > > > > > > > > > It happens after 1-3 hours of running 'stress-ng --dev 128'. This testcase > > > > > does a scandir of /dev and then calls random stuff like ioctl, lseek, > > > > > open/close etc. on the entries. I assume no files are deleted under /dev > > > > > during the testcase. > > > > > > > > > > The NULL pointer is the inode pointer of next. The next dentry->d_flags is > > > > > DCACHE_RCUACCESS when this happens. > > > > > > > > > > Any hints on how to further debug this? > > > > > > > > Can you reproduce the issue with vanilla -rc1 and do you have a "known good" > > > > kernel? > > > > > > I can try out -rc1, but IIRC this wasn't bisectible as the bug was present at > > > least back to 4.14. I need to double check that as there were other issues > > > that are resolved now so I may confuse things here. I've defintely seen > > > the same bug with 4.18. > > > > > > Unfortunately I lost access to the machine as our data center seems to be > > > moving currently so it might take some days until I can try -rc1. > > > > Ok, I've just managed to reproduce this in a KVM guest running v4.20-rc3 on > > both the host and the guest, so if anybody has any ideas of things to try then > > I'm happy to give them a shot. In the meantime, I'll try again with a bunch of > > debug checks enabled. Hi Will, good that you can reproduce the issue. I've verified that the issue is indeed reproducible with 4.14. > > Weee, I eventually hit a use-after-free from KASAN. See below. I ran KASAN (and all the other debug stuff) but didn't trigger anything in the host. --Jan
