[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Message-ID: <20240819053605.11706-1-neilb@suse.de>
Date: Mon, 19 Aug 2024 15:20:34 +1000
From: NeilBrown <neilb@...e.de>
To: Ingo Molnar <mingo@...hat.com>,
Peter Zijlstra <peterz@...radead.org>,
Linus Torvalds <torvalds@...ux-foundation.org>
Cc: linux-kernel@...r.kernel.org,
linux-fsdevel@...r.kernel.org
Subject: [PATCH 0/9 RFC] Make wake_up_{bit,var} less fragile
I wasn't really sure who to send this too, and get_maintainer.pl
suggested 132 addresses which seemed excessive. So I've focussed on
'sched' maintainers. I'll probably submit individual patches to
relevant maintainers/lists if I get positive feedback at this level.
This series was motivated by
Commit ed0172af5d6f ("SUNRPC: Fix a race to wake a sync task")
which adds smp_mb__after_atomic(). I thought "any API that requires that
sort of thing needs to be fixed".
The main patches here are 7 and 8 which revise wake_up_bit and
wake_up_var respectively. They result in 3 interfaces:
wake_up_{bit,var} includes smp_mb__after_atomic()
wake_up_{bit,var}_relaxed() doesn't have a barrier
wake_up_{bit,var}_mb() includes smb_mb().
I think this set of interfaces should be easier to use correctly. They
are also now documented more clearly.
The preceeding patches clean up various places where the exiting
interfaces weren't used optimally. The final patch uses
clear_and_wake_up_bit() more widely because it seems like a good idea.
I have three questions:
1/ is my understanding of the needed barriers correct.
i.e:
smp_mb__after_atomic() needed after a clear_bit() to atomic_set()
or similar, or a change inside a locked region
smb_mb() needed after any non-locked update
nothing needed after test_and_clear_bit() or atomic_dec_and_test()
or similar.
(I realised while working on this that my previous understanding
of the barrier requires was wrong, so maybe it still is).
2/ How should we handle the "flag day" where a barrier is added to
wake_up_bit() and wake_up_var(). Some options are:
a/ have a big patch for the flag-day as this series does
b/ add the barrier in a new wake_up_atomic_{bit,var} and deprecate
wake_up_{bit,var}
c/ don't worry about the fact that there will be an extra barrier for
a while - just make the change to wake_up_xxx() first, then submit
individual patches to remove barriers as needed.
3/ Who else should I ask to remove this at this high level?
Thanks,
NeilBrown
[PATCH 1/9] i915: remove wake_up on I915_RESET_MODESET.
[PATCH 2/9] Introduce atomic_dec_and_wake_up_var().
[PATCH 3/9] XFS: use wait_var_event() when waiting of i_pincount.
[PATCH 4/9] Use wait_var_event() instead of I_DIO_WAKEUP
[PATCH 5/9] Block: switch bd_prepare_to_claim to use
[PATCH 6/9] block/pktdvd: switch congestion waiting to
[PATCH 7/9] Improve and expand wake_up_bit() interface.
[PATCH 8/9] Improve and extend wake_up_var() interface.
[PATCH 9/9] Use clear_and_wake_up_bit() where appropriate.
Powered by blists - more mailing lists