[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Tue, 04 Oct 2022 19:01:04 +0300
From: Alexey Izbyshev <izbyshev@...ras.ru>
To: musl@...ts.openwall.com
Subject: Re: Illegal killlock skipping when transitioning to
single-threaded state
On 2022-10-04 16:50, Alexey Izbyshev wrote:
> On 2022-10-04 02:05, Rich Felker wrote:
>> On Mon, Oct 03, 2022 at 06:54:17PM -0400, Rich Felker wrote:
>>> On Mon, Oct 03, 2022 at 11:27:05PM +0200, Szabolcs Nagy wrote:
>>> > * Szabolcs Nagy <nsz@...t70.net> [2022-10-03 15:26:15 +0200]:
>>> >
>>> > > * Alexey Izbyshev <izbyshev@...ras.ru> [2022-10-03 09:16:03 +0300]:
>>> > > > On 2022-09-19 18:29, Rich Felker wrote:
>>> > > > > On Wed, Sep 07, 2022 at 03:46:53AM +0300, Alexey Izbyshev wrote:
>>> > > ...
>>> > > > > > Reordering the "libc.need_locks = -1" assignment and
>>> > > > > > UNLOCK(E->killlock) and providing a store barrier between them
>>> > > > > > should fix the issue.
>>> > > > >
>>> > > > > I think this all sounds correct. I'm not sure what you mean by a store
>>> > > > > barrier between them, since all lock and unlock operations are already
>>> > > > > full barriers.
>>> > > > >
>>> > > >
>>> > > > Before sending the report I tried to infer the intended ordering semantics
>>> > > > of LOCK/UNLOCK by looking at their implementations. For AArch64, I didn't
>>> > > > see why they would provide a full barrier (my reasoning is below), so I
>>> > > > concluded that probably acquire/release semantics was intended in general
>>> > > > and suggested an extra store barrier to prevent hoisting of "libc.need_locks
>>> > > > = -1" store spelled after UNLOCK(E->killlock) back into the critical
>>> > > > section.
>>> > > >
>>> > > > UNLOCK is implemented via a_fetch_add(). On AArch64, it is a simple
>>> > > > a_ll()/a_sc() loop without extra barriers, and a_ll()/a_sc() are implemented
>>> > > > via load-acquire/store-release instructions. Therefore, if we consider a
>>> > > > LOCK/UNLOCK critical section containing only plain loads and stores, (a) any
>>> > > > such memory access can be reordered with the initial ldaxr in UNLOCK, and
>>> > > > (b) any plain load following UNLOCK can be reordered with stlxr (assuming
>>> > > > the processor predicts that stlxr succeeds), and further, due to (a), with
>>> > > > any memory access inside the critical section. Therefore, UNLOCK is not full
>>> > > > barrier. Is this right?
>>> > >
>>> > > i dont think this is right.
>>> >
>>> >
>>> > i think i was wrong and you are right.
>>> >
>>> > so with your suggested swap of UNLOCK(killlock) and need_locks=-1 and
>>> > starting with 'something == 0' the exiting E and remaining R threads:
>>> >
>>> > E:something=1 // protected by killlock
>>> > E:UNLOCK(killlock)
>>> > E:need_locks=-1
>>> >
>>> > R:LOCK(unrelated) // reads need_locks == -1
>>> > R:need_locks=0
>>> > R:UNLOCK(unrelated)
>>> > R:LOCK(killlock) // does not lock
>>> > R:read something // can it be 0 ?
>>> >
>>> > and here something can be 0 (ie. not protected by killlock) on aarch64
>>> > because
>>> >
>>> > T1
>>> > something=1
>>> > ldaxr ... killlock
>>> > stlxr ... killlock
>>> > need_locks=-1
>>> >
>>> > T2
>>> > x=need_locks
>>> > ldaxr ... unrelated
>>> > stlxr ... unrelated
>>> > y=something
>>> >
>>> > can end with x==-1 and y==0.
>>> >
>>> > and to fix it, both a_fetch_add and a_cas need an a_barrier.
>>> >
>>> > i need to think how to support such lock usage on aarch64
>>> > without adding too many dmb.
>>>
>>> I don't really understand this, but FWIW gcc emits
>>>
>>> ldxr
>>> ...
>>> stlxr
>>> ...
>>> dmb ish
>>>
>>> for __sync_val_compare_and_swap. So this is probably the right thing
>>> we should have. And it seems to match what the kernel folks discussed
>>> here:
>>>
>>> http://lists.infradead.org/pipermail/linux-arm-kernel/2014-February/229588.html
>>>
>>> I wondered if there are similar issues for any others archs which
>>> need
>>> review, but it looks like all the other llsc archs have explicit
>>> pre/post barriers defined.
>>
>> Actually I don't understand what's going on with cmpxchg there. The
>> patch I linked has it using ldxr/stxr (not stlxr) for cmpxchg. There's
>> some follow-up in the thread I don't understand, about the case where
>> the cas fails, but we already handle that by doing an explicit barrier
>> in that case.
>>
> I think in that follow-up[1] they mean the following case (in musl
> terms):
>
> volatile int x, flag;
>
> T1:
> x = 1;
> a_store(&flag, 1);
>
> T2:
> while (!flag);
> a_cas(&x, 0, 1); // can this fail?
>
I made a mistake here, this should be a_cas(&x, 1, 2). Everything else
stands.
> They want it to never fail. But if a_cas() is implemented as
> ldrx/stlrx/dmb, this is not guaranteed because ldxr can be reordered
> with the load of flag.
>
> Note that musl does *not* handle this now, because a_barrier() in the
> failure path is after a_ll().
>
> [1]
> https://lists.infradead.org/pipermail/linux-arm-kernel/2014-February/229693.html
>
> Alexey
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
