Re: general protection fault in watchdog

[Dmitry Vyukov]

On Fri, Dec 14, 2018 at 5:53 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
>
> On Fri 14-12-18 15:31:44, Dmitry Vyukov wrote:
> > On Fri, Dec 14, 2018 at 2:54 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > >
> > > On Fri 14-12-18 14:42:33, Dmitry Vyukov wrote:
> > > > On Fri, Dec 14, 2018 at 2:28 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > > > >
> > > > > On Fri 14-12-18 14:11:05, Dmitry Vyukov wrote:
> > > > > > On Fri, Dec 14, 2018 at 1:51 PM syzbot
> > > > > > <syzbot+7713f3aa67be76b1552c@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > Hello,
> > > > > > >
> > > > > > > syzbot found the following crash on:
> > > > > > >
> > > > > > > HEAD commit:    f5d582777bcb Merge branch 'for-linus' of git://git.kernel...
> > > > > > > git tree:       upstream
> > > > > > > console output: https://syzkaller.appspot.com/x/log.txt?x=16aca143400000
> > > > > > > kernel config:  https://syzkaller.appspot.com/x/.config?x=c8970c89a0efbb23
> > > > > > > dashboard link: https://syzkaller.appspot.com/bug?extid=7713f3aa67be76b1552c
> > > > > > > compiler:       gcc (GCC) 8.0.1 20180413 (experimental)
> > > > > > > syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=1131381b400000
> > > > > > > C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=13bae593400000
> > > > > > >
> > > > > > > IMPORTANT: if you fix the bug, please add the following tag to the commit:
> > > > > > > Reported-by: syzbot+7713f3aa67be76b1552c@xxxxxxxxxxxxxxxxxxxxxxxxx
> > > > > >
> > > > > > +linux-mm for memcg question
> > > > > >
> > > > > > What the repro does is effectively just
> > > > > > setsockopt(EBT_SO_SET_ENTRIES). This eats all machine memory and
> > > > > > causes OOMs. Somehow it also caused the GPF in watchdog when it
> > > > > > iterates over task list, perhaps some scheduler code leaves a dangling
> > > > > > pointer on OOM failures.
> > > > > >
> > > > > > But what bothers me is a different thing. syzkaller test processes are
> > > > > > sandboxed with a restrictive memcg which should prevent them from
> > > > > > eating all memory. do_replace_finish calls vmalloc, which uses
> > > > > > GFP_KERNEL, which does not include GFP_ACCOUNT (GFP_KERNEL_ACCOUNT
> > > > > > does). And page alloc seems to change memory against memcg iff
> > > > > > GFP_ACCOUNT is provided.
> > > > > > Am I missing something or vmalloc is indeed not accounted (DoS)? I see
> > > > > > some explicit uses of GFP_KERNEL_ACCOUNT, e.g. the one below, but they
> > > > > > seem to be very sparse.
> > > > > >
> > > > > > static void *seq_buf_alloc(unsigned long size)
> > > > > > {
> > > > > >      return kvmalloc(size, GFP_KERNEL_ACCOUNT);
> > > > > > }
> > > > > >
> > > > > > Now looking at the code I also don't see how kmalloc(GFP_KERNEL) is
> > > > > > accounted... Which makes me think I am still missing something.
> > > > >
> > > > > You are not missing anything. We do not account all allocations and you
> > > > > have to explicitly opt-in by __GFP_ACCOUNT. This is a deliberate
> > > > > decision. If the allocation is directly controlable by an untrusted user
> > > > > and the memory is associated with a process life time then this looks
> > > > > like a good usecase for __GFP_ACCOUNT. If an allocation outlives a
> > > > > process then there the flag should be considered with a great care
> > > > > because oom killer is not able to resolve the memcg pressure and so the
> > > > > limit enforcement is not effective.
> > > >
> > > > Interesting.
> > > > I understand that namespaces, memcg's and processes (maybe even
> > > > threads) can have arbitrary overlapping. But I naively thought that in
> > > > canonical hierarchical cases it should all somehow work.
> > > > Question 1: is there some other, stricter sandboxing mechanism?
> > >
> > > I do not think so
> > >
> > > > We try
> > > > to sandbox syzkaller processes with everything available , because
> > > > these OOMs usually leads either to dead machines or hang/stall false
> > > > positives, which are nasty.
> > >
> > > Which is a useful test on its own. If you are able to trigger the global
> > > OOM from a restricted environment then you have a good candidate to
> > > consider a new __GFP_ACCOUNT user.
> > >
> > > > Question 2: this is a simple DoS vector, right? If I put a container
> > > > into a 1MB memcg, it can still eat arbitrary amount of non-pagable
> > > > kernel memory?
> > >
> > > As I've said. If there is a direct vector to allocated an unbounded
> > > amount of memory from the userspace (trusted users aside) then yes this
> > > sounds like a DoS to me.
> >
> > +netfilter maintainers for this easy DoS vector
> > short story: vmalloc in do_replace_finish allows unbounded memory
> > allocation not accounted to memcg
>
> Haven't we discussed that recently?

I just gave an executive summary to the new people I added to CC.

> > Looks pretty unbounded:
> >
> > [  763.451796] syz-executor681: vmalloc: allocation failure, allocated
> > 1027440640 of 1879052288 bytes, mode:0x6000c0(GFP_KERNEL),
> > nodemask=(null)
> >
> > But how does this play with what you said about memory outliving
> > process? Netfilter tables can definitely outlive the process, they are
> > attached to net ns.
>
> Not very well, but it is not hopeless either. The charged memory
> wouldn't go away with oom victims so it will stay there until somebody
> destroys those objects. On the other hand any process within that memcg
> would be killed so a DoS should be somehow contained.
>
> > Also, am I reading this correctly that potentially thousands of kernel
> > memory allocations need to be converted to ACCOUNT? I mean for small
> > ones we maybe care less, but they _should_ be accounted. They can also
> > build up, or just simply allow small repeated allocations.
> >
> > GFP_KERNEL Referenced in 11070 files:
> > https://elixir.bootlin.com/linux/latest/ident/GFP_KERNEL
> >
> > GFP_KERNEL_ACCOUNT Referenced in 19 files:
> > https://elixir.bootlin.com/linux/latest/ident/GFP_KERNEL_ACCOUNT
>
> Yes, we do not care about all kernel objects. Most of them are contained
> by some high-level objects already. The purpose of the kmem accounting
> is to limit those objects that can grow really large and are reclaimable
> in some way.

I see, thanks for the explanations.