Re: [PATCH 0/3] kasan: memorize and print call_rcu stack

From: Dmitry Vyukov
Date: Wed May 06 2020 - 08:17:13 EST


On Wed, May 6, 2020 at 2:01 PM Walter Wu <walter-zh.wu@xxxxxxxxxxxx> wrote:
>
> On Wed, 2020-05-06 at 11:37 +0200, 'Dmitry Vyukov' via kasan-dev wrote:
> > On Wed, May 6, 2020 at 8:23 AM Walter Wu <walter-zh.wu@xxxxxxxxxxxx> wrote:
> > > > > This patchset improves KASAN reports by making them to have
> > > > > call_rcu() call stack information. It is helpful for programmers
> > > > > to solve use-after-free or double-free memory issue.
> > > > >
> > > > > The KASAN report was as follows(cleaned up slightly):
> > > > >
> > > > > BUG: KASAN: use-after-free in kasan_rcu_reclaim+0x58/0x60
> > > > >
> > > > > Freed by task 0:
> > > > > save_stack+0x24/0x50
> > > > > __kasan_slab_free+0x110/0x178
> > > > > kasan_slab_free+0x10/0x18
> > > > > kfree+0x98/0x270
> > > > > kasan_rcu_reclaim+0x1c/0x60
> > > > > rcu_core+0x8b4/0x10f8
> > > > > rcu_core_si+0xc/0x18
> > > > > efi_header_end+0x238/0xa6c
> > > > >
> > > > > First call_rcu() call stack:
> > > > > save_stack+0x24/0x50
> > > > > kasan_record_callrcu+0xc8/0xd8
> > > > > call_rcu+0x190/0x580
> > > > > kasan_rcu_uaf+0x1d8/0x278
> > > > >
> > > > > Last call_rcu() call stack:
> > > > > (stack is not available)
> > > > >
> > > > >
> > > > > Add new CONFIG option to record first and last call_rcu() call stack
> > > > > and KASAN report prints two call_rcu() call stack.
> > > > >
> > > > > This option doesn't increase the cost of memory consumption. It is
> > > > > only suitable for generic KASAN.
> > > >
> > > > I donât understand why this needs to be a Kconfig option at all. If call_rcu() stacks are useful in general, then just always gather those information. How do developers judge if they need to select this option or not?
> > >
> > > Because we don't want to increase slub meta-data size, so enabling this
> > > option can print call_rcu() stacks, but the in-use slub object doesn't
> > > print free stack. So if have out-of-bound issue, then it will not print
> > > free stack. It is a trade-off, see [1].
> > >
> > > [1] https://bugzilla.kernel.org/show_bug.cgi?id=198437
> >
> > Hi Walter,
> >
> > Great you are tackling this!
> >
> > I have the same general sentiment as Qian. I would enable this
> > unconditionally because:
> >
> > 1. We still can't get both rcu stack and free stack. I would assume
> > most kernel testing systems need to enable this (we definitely enable
> > on syzbot). This means we do not have free stack for allocation
> > objects in any reports coming from testing systems. Which greatly
> > diminishes the value of the other mode.
> >
> > 2. Kernel is undertested. Introducing any additional configuration
> > options is a problem in such context. Chances are that some of the
> > modes are not working or will break in future.
> >
> > 3. That free stack actually causes lots of confusion and I never found
> > it useful:
> > https://bugzilla.kernel.org/show_bug.cgi?id=198425
> > If it's a very delayed UAF, either one may get another report for the
> > same bug with not so delayed UAF, or if it's way too delayed, then the
> > previous free stack is wrong as well.
> >
> > 4. Most users don't care that much about debugging tools to learn
> > every bit of every debugging tool and spend time fine-tuning it for
> > their context. Most KASAN users won't even be aware of this choice,
> > and they will just use whatever is the default.
> >
> > 5. Each configuration option increases implementation complexity.
> >
> > What would have value is if we figure out how to make both of them
> > work at the same time without increasing memory consumption. But I
> > don't see any way to do this.
> >
> > I propose to make this the only mode. I am sure lots of users will
> > find this additional stack useful, whereas the free stack is even
> > frequently confusing.
> >
>
> Ok.
> If we want to have a default enabling it, but it should only work in
> generic KASAN, because we need to get object status(allocation or
> freeing) from shadow memory, tag-based KASAN can't do it. So we should
> have a default enabling it in generic KASAN?

Yes, let's do generic KASAN always memorizes rcu stack; tags KASAN
never memorizes rcu stacks. No new configurations.