[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20240806220116.GH623957@frogsfrogsfrogs>
Date: Tue, 6 Aug 2024 15:01:16 -0700
From: "Darrick J. Wong" <djwong@...nel.org>
To: Peter Zijlstra <peterz@...radead.org>
Cc: Chandan Babu R <chandanbabu@...nel.org>,
Matthew Wilcox <willy@...radead.org>,
xfs <linux-xfs@...r.kernel.org>,
linux-fsdevel <linux-fsdevel@...r.kernel.org>,
linux-kernel <linux-kernel@...r.kernel.org>, x86@...nel.org,
tglx@...utronix.de
Subject: Re: Are jump labels broken on 6.11-rc1?
On Tue, Aug 06, 2024 at 12:38:08PM +0200, Peter Zijlstra wrote:
> On Tue, Aug 06, 2024 at 11:44:13AM +0200, Peter Zijlstra wrote:
> > On Mon, Aug 05, 2024 at 07:35:22AM -0700, Darrick J. Wong wrote:
> > > On Wed, Jul 31, 2024 at 12:55:57PM +0200, Peter Zijlstra wrote:
> > > > On Tue, Jul 30, 2024 at 10:33:41PM -0700, Darrick J. Wong wrote:
> > > >
> > > > > Sooooo... it turns out that somehow your patch got mismerged on the
> > > > > first go-round, and that worked. The second time, there was no
> > > > > mismerge, which mean that the wrong atomic_cmpxchg() callsite was
> > > > > tested.
> > > > >
> > > > > Looking back at the mismerge, it actually changed
> > > > > __static_key_slow_dec_cpuslocked, which had in 6.10:
> > > > >
> > > > > if (atomic_dec_and_test(&key->enabled))
> > > > > jump_label_update(key);
> > > > >
> > > > > Decrement, then return true if the value was set to zero. With the 6.11
> > > > > code, it looks like we want to exchange a 1 with a 0, and act only if
> > > > > the previous value had been 1.
> > > > >
> > > > > So perhaps we really want this change? I'll send it out to the fleet
> > > > > and we'll see what it reports tomorrow morning.
> > > >
> > > > Bah yes, I missed we had it twice. Definitely both sites want this.
> > > >
> > > > I'll tentatively merge the below patch in tip/locking/urgent. I can
> > > > rebase if there is need.
> > >
> > > Hi Peter,
> > >
> > > This morning, I noticed the splat below with -rc2.
> > >
> > > WARNING: CPU: 0 PID: 8578 at kernel/jump_label.c:295 __static_key_slow_dec_cpuslocked.part.0+0x50/0x60
> > >
> > > Line 295 is the else branch of this code:
> > >
> > > if (atomic_cmpxchg(&key->enabled, 1, 0) == 1)
> > > jump_label_update(key);
> > > else
> > > WARN_ON_ONCE(!static_key_slow_try_dec(key));
> > >
> > > Apparently static_key_slow_try_dec returned false? Looking at that
> > > function, I suppose the atomic_read of key->enabled returned 0, since it
> > > didn't trigger the "WARN_ON_ONCE(v < 0)" code. Does that mean the value
> > > must have dropped from positive N to 0 without anyone ever taking the
> > > jump_label_mutex?
> >
> > One possible scenario I see:
> >
> > slow_dec
> > if (try_dec) // dec_not_one-ish, false
> > // enabled == 1
> > slow_inc
> > if (inc_not_disabled) // inc_not_zero-ish
> > // enabled == 2
> > return
> >
> > guard((mutex)(&jump_label_mutex);
> > if (atomic_cmpxchg(1,0)==1) // false, we're 2
> >
> > slow_dec
> > if (try-dec) // dec_not_one, true
> > // enabled == 1
> > return
> > else
> > try_dec() // dec_not_one, false
> > WARN
> >
> >
> > Let me go play to see how best to cure this.
>
> I've ended up with this, not exactly pretty :/
>
> Thomas?
It seems to survive a short test, will send it out for overnight testing
on the full fleet, thanks.
--D
> ---
> diff --git a/kernel/jump_label.c b/kernel/jump_label.c
> index 6dc76b590703..5fa2c9f094b1 100644
> --- a/kernel/jump_label.c
> +++ b/kernel/jump_label.c
> @@ -168,8 +168,8 @@ bool static_key_slow_inc_cpuslocked(struct static_key *key)
> jump_label_update(key);
> /*
> * Ensure that when static_key_fast_inc_not_disabled() or
> - * static_key_slow_try_dec() observe the positive value,
> - * they must also observe all the text changes.
> + * static_key_dec() observe the positive value, they must also
> + * observe all the text changes.
> */
> atomic_set_release(&key->enabled, 1);
> } else {
> @@ -250,7 +250,7 @@ void static_key_disable(struct static_key *key)
> }
> EXPORT_SYMBOL_GPL(static_key_disable);
>
> -static bool static_key_slow_try_dec(struct static_key *key)
> +static bool static_key_dec(struct static_key *key, bool fast)
> {
> int v;
>
> @@ -268,31 +268,45 @@ static bool static_key_slow_try_dec(struct static_key *key)
> v = atomic_read(&key->enabled);
> do {
> /*
> - * Warn about the '-1' case though; since that means a
> - * decrement is concurrent with a first (0->1) increment. IOW
> - * people are trying to disable something that wasn't yet fully
> - * enabled. This suggests an ordering problem on the user side.
> + * Warn about the '-1' case; since that means a decrement is
> + * concurrent with a first (0->1) increment. IOW people are
> + * trying to disable something that wasn't yet fully enabled.
> + * This suggests an ordering problem on the user side.
> + *
> + * Warn about the '0' case; simple underflow.
> + *
> + * Neither case should succeed and change things.
> + */
> + if (WARN_ON_ONCE(v <= 0))
> + return false;
> +
> + /*
> + * Lockless fast-path, dec-not-one like behaviour.
> */
> - WARN_ON_ONCE(v < 0);
> - if (v <= 1)
> + if (fast && v <= 1)
> return false;
> } while (!likely(atomic_try_cmpxchg(&key->enabled, &v, v - 1)));
>
> - return true;
> + if (fast)
> + return true;
> +
> + /*
> + * Locked slow path, dec-and-test like behaviour.
> + */
> + lockdep_assert_held(&jump_label_mutex);
> + return v == 1;
> }
>
> static void __static_key_slow_dec_cpuslocked(struct static_key *key)
> {
> lockdep_assert_cpus_held();
>
> - if (static_key_slow_try_dec(key))
> + if (static_key_dec(key, true)) // dec-not-one
> return;
>
> guard(mutex)(&jump_label_mutex);
> - if (atomic_cmpxchg(&key->enabled, 1, 0) == 1)
> + if (static_key_dec(key, false)) // dec-and-test
> jump_label_update(key);
> - else
> - WARN_ON_ONCE(!static_key_slow_try_dec(key));
> }
>
> static void __static_key_slow_dec(struct static_key *key)
> @@ -329,7 +343,7 @@ void __static_key_slow_dec_deferred(struct static_key *key,
> {
> STATIC_KEY_CHECK_USE(key);
>
> - if (static_key_slow_try_dec(key))
> + if (static_key_dec(key, true)) // dec-not-one
> return;
>
> schedule_delayed_work(work, timeout);
Powered by blists - more mailing lists