Re: [PATCH net] can: isotp: isotp_rcv_cf(): fix so->rx race problem

Hi Marc,

On 07.02.22 09:11, Marc Kleine-Budde wrote:

On 28.01.2022 15:48:05, Oliver Hartkopp wrote:

quoted

Hello Marc, hello William,

On 28.01.22 09:46, Marc Kleine-Budde wrote:

quoted

On 28.01.2022 09:32:40, Oliver Hartkopp wrote:

quoted

On 28.01.22 09:07, Marc Kleine-Budde wrote:

quoted

On 28.01.2022 08:56:19, Oliver Hartkopp wrote:

quoted

I've seen the frame processing sometimes freezes for one second when
stressing the isotp_rcv() from multiple sources. This finally freezes
the entire softirq which is either not good and not needed as we only
need to fix this race for stress tests - and not for real world usage
that does not create this case.

Hmmm, this doesn't sound good. Can you test with LOCKDEP enabled?

quoted

quoted

#
# Lock Debugging (spinlocks, mutexes, etc...)
#
CONFIG_LOCK_DEBUGGING_SUPPORT=y
# CONFIG_PROVE_LOCKING is not set

CONFIG_PROVE_LOCKING=y

Now enabled even more locking (seen relevant kernel config at the end).

It turns out that there is no visible difference when using spin_lock() or
spin_trylock().

I only got some of these kernel log entries

Jan 28 11:13:14 silver kernel: [ 2396.323211] perf: interrupt took too long
(2549 > 2500), lowering kernel.perf_event_max_sample_rate to 78250
Jan 28 11:25:49 silver kernel: [ 3151.172773] perf: interrupt took too long
(3188 > 3186), lowering kernel.perf_event_max_sample_rate to 62500
Jan 28 11:45:24 silver kernel: [ 4325.583328] perf: interrupt took too long
(4009 > 3985), lowering kernel.perf_event_max_sample_rate to 49750
Jan 28 12:15:46 silver kernel: [ 6148.238246] perf: interrupt took too long
(5021 > 5011), lowering kernel.perf_event_max_sample_rate to 39750
Jan 28 13:01:45 silver kernel: [ 8907.303715] perf: interrupt took too long
(6285 > 6276), lowering kernel.perf_event_max_sample_rate to 31750

But I get these sporadically anyway. No other LOCKDEP splat.

At least the issue reported by William should be fixed now - but I'm still
unclear whether spin_lock() or spin_trylock() is the best approach here in
the NET_RX softirq?!?

With the !spin_trylock() -> return you are saying if something
concurrent happens, drop it. This doesn't sound correct.

Yes, I had the same feeling and did some extensive load tests using both 
variants.

It turned out the standard spin_lock() works excellent to fix the issue.

Thanks for taking it for upstream here:
https://lore.kernel.org/linux-can/20220209074818.3ylfz4zmuhit7orc@pengutronix.de/T/#t (local)

Best regards,
Oliver