[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <39c3c4dc-d852-40b3-a662-6202c5422acf@kernel.dk>
Date: Tue, 12 Mar 2024 15:40:07 -0600
From: Jens Axboe <axboe@...nel.dk>
To: Linus Torvalds <torvalds@...ux-foundation.org>,
Jakub Kicinski <kuba@...nel.org>,
Johannes Thumshirn <johannes.thumshirn@....com>
Cc: davem@...emloft.net, netdev@...r.kernel.org,
linux-kernel@...r.kernel.org, pabeni@...hat.com, bpf@...r.kernel.org,
Tejun Heo <tj@...nel.org>
Subject: Re: [GIT PULL] Networking for v6.9
On 3/12/24 3:11 PM, Linus Torvalds wrote:
> On Tue, 12 Mar 2024 at 13:47, Jakub Kicinski <kuba@...nel.org> wrote:
>>
>> With your tree as of 65d287c7eb1d it gets to prompt but dies soon after
>> when prod services kick in (dunno what rpm Kdump does but says iocost
>> so adding Tejun):
>
> Both of your traces are timers that seem to either lock up in ioc_now():
>
> https://lore.kernel.org/all/20240312133427.1a744844@kernel.org/
>
> and now it looks like ioc_timer_fn():
>
> https://lore.kernel.org/all/20240312134739.248e6bd3@kernel.org/
>
> But in neither case does it actually look like it's a lockup on a *lock*.
>
> IOW, the NMI isn't happening on some spin_lock sequence or anything like that.
>
> Yes, ioc_now() could have been looping on the seq read-lock if the
> sequence number was odd. But the writers do seem to be done with
> interrupts disabled, plus then you wouldn't have this lockup in
> ioc_timer_fn, so it's probably not that.
>
> And yes, ioc_timer_fn() does take locks, but again, that doesn't seem
> to be where it is hanging.
>
> So it smells like it's an endless loop in ioc_timer_fn() to me, or
> perhaps retriggering the timer itself infinitely.
>
> Which would then explain both of those traces (that endless loop would
> call ioc_now() as part of it).
>
> The blk-iocost.c code itself hasn't changed, but the timer code has
> gone through big changes.
>
> That said, there's a more blk-related change: da4c8c3d0975 ("block:
> cache current nsec time in struct blk_plug").
>
> *And* your second dump is from that
>
> period_vtime = now.vnow - ioc->period_at_vtime;
> if (WARN_ON_ONCE(!period_vtime)) {
>
> so it smells like the blk-iocost code is just completely confused by
> the time caching. Jens?
>
> Jakub, it might be worth seeing if just reverting that commit
> da4c8c3d0975 makes the problem go away. Otherwise a bisect might be
> needed...
Hmm, I wonder if the below will fix it. At least from the timer side,
we should not be using the cached clock.
diff --git a/block/blk-iocost.c b/block/blk-iocost.c
index 9a85bfbbc45a..646b50e1c914 100644
--- a/block/blk-iocost.c
+++ b/block/blk-iocost.c
@@ -1044,7 +1044,7 @@ static void ioc_now(struct ioc *ioc, struct ioc_now *now)
unsigned seq;
u64 vrate;
- now->now_ns = blk_time_get_ns();
+ now->now_ns = ktime_get_ns();
now->now = ktime_to_us(now->now_ns);
vrate = atomic64_read(&ioc->vtime_rate);
--
Jens Axboe
Powered by blists - more mailing lists