lkml.org 
[lkml]   [2019]   [Nov]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: KCSAN: data-race in taskstats_exit / taskstats_exit
On Fri, Nov 8, 2019 at 1:54 AM Balbir Singh <bsingharora@gmail.com> wrote:
>
> On Wed, 2019-11-06 at 11:23 +0100, Marco Elver wrote:
> > On Wed, 6 Nov 2019 at 01:10, Balbir Singh <bsingharora@gmail.com> wrote:
> > >
> > > On Fri, 2019-10-04 at 21:26 -0700, syzbot wrote:
> > > > Hello,
> > > >
> > > > syzbot found the following crash on:
> > > >
> > > > HEAD commit: b4bd9343 x86, kcsan: Enable KCSAN for x86
> > > > git tree: https://github.com/google/ktsan.git kcsan
> > > > console output: https://syzkaller.appspot.com/x/log.txt?x=125329db600000
> > > > kernel config:
> > > > https://syzkaller.appspot.com/x/.config?x=c0906aa620713d80
> > > > dashboard link:
> > > > https://syzkaller.appspot.com/bug?extid=c5d03165a1bd1dead0c1
> > > > compiler: gcc (GCC) 9.0.0 20181231 (experimental)
> > > >
> > > > Unfortunately, I don't have any reproducer for this crash yet.
> > > >
> > > > IMPORTANT: if you fix the bug, please add the following tag to the
> > > > commit:
> > > > Reported-by: syzbot+c5d03165a1bd1dead0c1@syzkaller.appspotmail.com
> > > >
> > > > ==================================================================
> > > > BUG: KCSAN: data-race in taskstats_exit / taskstats_exit
> > > >
> > > > write to 0xffff8881157bbe10 of 8 bytes by task 7951 on cpu 0:
> > > > taskstats_tgid_alloc kernel/taskstats.c:567 [inline]
> > > > taskstats_exit+0x6b7/0x717 kernel/taskstats.c:596
> > > > do_exit+0x2c2/0x18e0 kernel/exit.c:864
> > > > do_group_exit+0xb4/0x1c0 kernel/exit.c:983
> > > > get_signal+0x2a2/0x1320 kernel/signal.c:2734
> > > > do_signal+0x3b/0xc00 arch/x86/kernel/signal.c:815
> > > > exit_to_usermode_loop+0x250/0x2c0 arch/x86/entry/common.c:159
> > > > prepare_exit_to_usermode arch/x86/entry/common.c:194 [inline]
> > > > syscall_return_slowpath arch/x86/entry/common.c:274 [inline]
> > > > do_syscall_64+0x2d7/0x2f0 arch/x86/entry/common.c:299
> > > > entry_SYSCALL_64_after_hwframe+0x44/0xa9
> > > >
> > > > read to 0xffff8881157bbe10 of 8 bytes by task 7949 on cpu 1:
> > > > taskstats_tgid_alloc kernel/taskstats.c:559 [inline]
> > > > taskstats_exit+0xb2/0x717 kernel/taskstats.c:596
> > > > do_exit+0x2c2/0x18e0 kernel/exit.c:864
> > > > do_group_exit+0xb4/0x1c0 kernel/exit.c:983
> > > > __do_sys_exit_group kernel/exit.c:994 [inline]
> > > > __se_sys_exit_group kernel/exit.c:992 [inline]
> > > > __x64_sys_exit_group+0x2e/0x30 kernel/exit.c:992
> > > > do_syscall_64+0xcf/0x2f0 arch/x86/entry/common.c:296
> > > > entry_SYSCALL_64_after_hwframe+0x44/0xa9
> > > >
> > >
> > > Sorry I've been away and just catching up with email
> > >
> > > I don't think this is a bug, if I interpret the report correctly it shows
> > > a
> > > race
> > >
> > > static struct taskstats *taskstats_tgid_alloc(struct task_struct *tsk)
> > > {
> > > struct signal_struct *sig = tsk->signal;
> > > struct taskstats *stats;
> > >
> > > #1 if (sig->stats || thread_group_empty(tsk)) <- the check of sig-
> > > >stats
> > > goto ret;
> > >
> > > /* No problem if kmem_cache_zalloc() fails */
> > > stats = kmem_cache_zalloc(taskstats_cache, GFP_KERNEL);
> > >
> > > spin_lock_irq(&tsk->sighand->siglock);
> > > if (!sig->stats) {
> > > #2 sig->stats = stats; <- here in setting sig->stats
> > > stats = NULL;
> > > }
> > > spin_unlock_irq(&tsk->sighand->siglock);
> > >
> > > if (stats)
> > > kmem_cache_free(taskstats_cache, stats);
> > > ret:
> > > return sig->stats;
> > > }
> > >
> > > The worst case scenario is that we might see sig->stats as being NULL when
> > > two
> > > threads belonging to the same tgid. We do free up stats if we got that
> > > wrong
> > >
> > > Am I misinterpreting the report?
> > >
> > > Balbir Singh.
> >
> > The plain concurrent reads/writes are a data race, which may manifest
> > in various undefined behaviour due to compiler optimizations [1, 2].
> > Note that, "data race" does not necessarily imply "race condition";
> > some data races are race conditions (usually the more interesting
> > bugs) -- but not *all* data races are race conditions (sometimes
> > referred to as "benign races"). KCSAN reports data races according to
> > the LKMM.
> > [1] https://lwn.net/Articles/793253/
> > [2] https://lwn.net/Articles/799218/
> >
> > If there is no race condition here that warrants heavier
> > synchronization (locking etc.), we still have the data race which
> > needs fixing by using marked atomic operations (READ_ONCE, WRITE_ONCE,
> > atomic_t, etc.). We also need to consider memory ordering requirements
> > (do we need smp_*mb(), smp_load_acquire/smp_store_release, ..)?
> >
> > In the case here, the pattern is double-checked locking, which is
> > incorrect without atomic operations and the correct memory ordering.
> > There is a lengthy discussion here:
> >
> https://lore.kernel.org/lkml/20191021113327.22365-1-christian.brauner@ubuntu.com/
> >
> >
>
> I am still not convinced unless someone can prove that unsigned long reads are
> non-atomic,

None of the relevant standards guarantee this. C standard very
explicitly states the opposite - any data race renders behavior of the
program as undefined. LKMM does not give any defined behavior to data
races either. QED ;)

> acquire/release and barriers semantics don't matter because the
> code deals with the race inside of a lock if the read was spurious, The
> assumption is based on the face that sig->stats can be NULL or the kzalloc'ed
> value in all cases.
>
> Balbir Singh.
>
> --
> You received this message because you are subscribed to the Google Groups "syzkaller-bugs" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to syzkaller-bugs+unsubscribe@googlegroups.com.
> To view this discussion on the web visit https://groups.google.com/d/msgid/syzkaller-bugs/acd6b0d98a7ebcb4ead9b263ec5c568c5a747166.camel%40gmail.com.

\
 
 \ /
  Last update: 2019-11-08 09:56    [W:0.115 / U:11.116 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site