Re: [tip:locking/urgent] locking/rwsem-spinlock: Fix EINTR branch in __down_write_common()
From: Ingo Molnar
Date: Thu Jul 06 2017 - 03:29:09 EST
* Niklas Cassel <niklas.cassel@xxxxxxxx> wrote:
> On 07/05/2017 04:27 PM, tip-bot for Kirill Tkhai wrote:
> > Commit-ID: a0c4acd2c220376b4e9690e75782d0c0afdaab9f
> > Gitweb: http://git.kernel.org/tip/a0c4acd2c220376b4e9690e75782d0c0afdaab9f
> > Author: Kirill Tkhai <ktkhai@xxxxxxxxxxxxx>
> > AuthorDate: Fri, 16 Jun 2017 16:44:34 +0300
> > Committer: Ingo Molnar <mingo@xxxxxxxxxx>
> > CommitDate: Wed, 5 Jul 2017 12:26:29 +0200
> > locking/rwsem-spinlock: Fix EINTR branch in __down_write_common()
> > If a writer could been woken up, the above branch
> > if (sem->count == 0)
> > break;
> > would have moved us to taking the sem. So, it's
> > not the time to wake a writer now, and only readers
> > are allowed now. Thus, 0 must be passed to __rwsem_do_wake().
> > Next, __rwsem_do_wake() wakes readers unconditionally.
> > But we mustn't do that if the sem is owned by writer
> > in the moment. Otherwise, writer and reader own the sem
> > the same time, which leads to memory corruption in
> > callers.
> > rwsem-xadd.c does not need that, as:
> > 1) the similar check is made lockless there,
> > 2) in __rwsem_mark_wake::try_reader_grant we test,
> > that sem is not owned by writer.
> > Signed-off-by: Kirill Tkhai <ktkhai@xxxxxxxxxxxxx>
> > Acked-by: Peter Zijlstra <a.p.zijlstra@xxxxxxxxx>
> > Cc: <stable@xxxxxxxxxxxxxxx>
> > Cc: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
> > Cc: Niklas Cassel <niklas.cassel@xxxxxxxx>
> > Cc: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> > Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> > Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > Fixes: 17fcbd590d0c "locking/rwsem: Fix down_write_killable() for CONFIG_RWSEM_GENERIC_SPINLOCK=y"
> > Link: http://lkml.kernel.org/r/149762063282.19811.9129615532201147826.stgit@xxxxxxxxxxxxxxxxxxxxx
> > Signed-off-by: Ingo Molnar <mingo@xxxxxxxxxx>
> > ---
> > kernel/locking/rwsem-spinlock.c | 4 ++--
> > 1 file changed, 2 insertions(+), 2 deletions(-)
> > diff --git a/kernel/locking/rwsem-spinlock.c b/kernel/locking/rwsem-spinlock.c
> > index c65f798..20819df 100644
> > --- a/kernel/locking/rwsem-spinlock.c
> > +++ b/kernel/locking/rwsem-spinlock.c
> > @@ -231,8 +231,8 @@ int __sched __down_write_common(struct rw_semaphore *sem, int state)
> > out_nolock:
> > list_del(&waiter.list);
> > - if (!list_empty(&sem->wait_list))
> > - __rwsem_do_wake(sem, 1);
> > + if (!list_empty(&sem->wait_list) && sem->count >= 0)
> > + __rwsem_do_wake(sem, 0);
> > raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
> > return -EINTR;
> For the record, there is actually a v2 of this:
Hm, so I missed that because it was within the discussion - please post v2 patches
with a new subject line next time around.
But I also disagree with -v2 mildly: in practice a >= test has the same CPU
overhead as a > test, and if we rely on the earlier "sem->count == 0" test then we
should also comment on that.
It's more straightforward to just do the canonical sem->count >= 0 test that we do
elsewhere in the rwsem-spinlock code.
PeterZ, what's your preference?