On 5/18/22 9:14 AM, Lee Jones wrote: > On Wed, 18 May 2022, Jens Axboe wrote: > >> On 5/18/22 6:54 AM, Jens Axboe wrote: >>> On 5/18/22 6:52 AM, Jens Axboe wrote: >>>> On 5/18/22 6:50 AM, Lee Jones wrote: >>>>> On Tue, 17 May 2022, Jens Axboe wrote: >>>>> >>>>>> On 5/17/22 7:00 AM, Lee Jones wrote: >>>>>>> On Tue, 17 May 2022, Jens Axboe wrote: >>>>>>> >>>>>>>> On 5/17/22 6:36 AM, Lee Jones wrote: >>>>>>>>> On Tue, 17 May 2022, Jens Axboe wrote: >>>>>>>>> >>>>>>>>>> On 5/17/22 6:24 AM, Lee Jones wrote: >>>>>>>>>>> On Tue, 17 May 2022, Jens Axboe wrote: >>>>>>>>>>> >>>>>>>>>>>> On 5/17/22 5:41 AM, Lee Jones wrote: >>>>>>>>>>>>> Good afternoon Jens, Pavel, et al., >>>>>>>>>>>>> >>>>>>>>>>>>> Not sure if you are presently aware, but there appears to be a >>>>>>>>>>>>> use-after-free issue affecting the io_uring worker driver (fs/io-wq.c) >>>>>>>>>>>>> in Stable v5.10.y. >>>>>>>>>>>>> >>>>>>>>>>>>> The full sysbot report can be seen below [0]. >>>>>>>>>>>>> >>>>>>>>>>>>> The C-reproducer has been placed below that [1]. >>>>>>>>>>>>> >>>>>>>>>>>>> I had great success running this reproducer in an infinite loop. >>>>>>>>>>>>> >>>>>>>>>>>>> My colleague reverse-bisected the fixing commit to: >>>>>>>>>>>>> >>>>>>>>>>>>> commit fb3a1f6c745ccd896afadf6e2d6f073e871d38ba >>>>>>>>>>>>> Author: Jens Axboe <axboe@xxxxxxxxx> >>>>>>>>>>>>> Date: Fri Feb 26 09:47:20 2021 -0700 >>>>>>>>>>>>> >>>>>>>>>>>>> io-wq: have manager wait for all workers to exit >>>>>>>>>>>>> >>>>>>>>>>>>> Instead of having to wait separately on workers and manager, just have >>>>>>>>>>>>> the manager wait on the workers. We use an atomic_t for the reference >>>>>>>>>>>>> here, as we need to start at 0 and allow increment from that. Since the >>>>>>>>>>>>> number of workers is naturally capped by the allowed nr of processes, >>>>>>>>>>>>> and that uses an int, there is no risk of overflow. >>>>>>>>>>>>> >>>>>>>>>>>>> Signed-off-by: Jens Axboe <axboe@xxxxxxxxx> >>>>>>>>>>>>> >>>>>>>>>>>>> fs/io-wq.c | 30 ++++++++++++++++++++++-------- >>>>>>>>>>>>> 1 file changed, 22 insertions(+), 8 deletions(-) >>>>>>>>>>>> >>>>>>>>>>>> Does this fix it: >>>>>>>>>>>> >>>>>>>>>>>> commit 886d0137f104a440d9dfa1d16efc1db06c9a2c02 >>>>>>>>>>>> Author: Jens Axboe <axboe@xxxxxxxxx> >>>>>>>>>>>> Date: Fri Mar 5 12:59:30 2021 -0700 >>>>>>>>>>>> >>>>>>>>>>>> io-wq: fix race in freeing 'wq' and worker access >>>>>>>>>>>> >>>>>>>>>>>> Looks like it didn't make it into 5.10-stable, but we can certainly >>>>>>>>>>>> rectify that. >>>>>>>>>>> >>>>>>>>>>> Thanks for your quick response Jens. >>>>>>>>>>> >>>>>>>>>>> This patch doesn't apply cleanly to v5.10.y. >>>>>>>>>> >>>>>>>>>> This is probably why it never made it into 5.10-stable :-/ >>>>>>>>> >>>>>>>>> Right. It doesn't apply at all unfortunately. >>>>>>>>> >>>>>>>>>>> I'll have a go at back-porting it. Please bear with me. >>>>>>>>>> >>>>>>>>>> Let me know if you into issues with that and I can help out. >>>>>>>>> >>>>>>>>> I think the dependency list is too big. >>>>>>>>> >>>>>>>>> Too much has changed that was never back-ported. >>>>>>>>> >>>>>>>>> Actually the list of patches pertaining to fs/io-wq.c alone isn't so >>>>>>>>> bad, I did start to back-port them all but some of the big ones have >>>>>>>>> fs/io_uring.c changes incorporated and that list is huge (256 patches >>>>>>>>> from v5.10 to the fixing patch mentioned above). >>>>>>>> >>>>>>>> The problem is that 5.12 went to the new worker setup, and this patch >>>>>>>> landed after that even though it also applies to the pre-native workers. >>>>>>>> Hence the dependency chain isn't really as long as it seems, probably >>>>>>>> just a few patches backporting the change references and completions. >>>>>>>> >>>>>>>> I'll take a look this afternoon. >>>>>>> >>>>>>> Thanks Jens. I really appreciate it. >>>>>> >>>>>> Can you see if this helps? Untested... >>>>> >>>>> What base does this apply against please? >>>>> >>>>> I tried Mainline and v5.10.116 and both failed. >>>> >>>> It's against 5.10.116, so that's puzzling. Let me double check I sent >>>> the right one... >>> >>> Looks like I sent the one from the wrong directory, sorry about that. >>> This one should be better: >> >> Nope, both are the right one. Maybe your mailer is mangling the patch? >> I'll attach it gzip'ed here in case that helps. > > Okay, that applied, thanks. > > Unfortunately, I am still able to crash the kernel in the same way. Alright, maybe it's not enough. I can't get your reproducer to crash, unfortunately. I'll try on a different box. -- Jens Axboe
