Alex,
On Mon, Aug 24 2020 at 19:29, Alexander Graf wrote:
> I'm currently trying to understand a performance regression with
> ScyllaDB on i3en.3xlarge (KVM based VM on Skylake) which we reliably
> bisected down to this commit:
>
> https://github.com/scylladb/scylla/issues/7036
>
> What we're seeing is that syscalls such as membarrier() take forever
> (0-10 µs would be normal):
...
> That again seems to stem from a vastly slowed down
> smp_call_function_many_cond():
>
> Samples: 218K of event 'cpu-clock', 4000 Hz
> Overhead Shared Object Symbol
> 94.51% [kernel] [k] smp_call_function_many_cond
> 0.76% [kernel] [k] __do_softirq
> 0.32% [kernel] [k] native_queued_spin_lock_slowpath
> [...]
>
> which is stuck in
>
> │ csd_lock_wait():
> │ smp_cond_load_acquire(&csd->flags, !(VAL &
> 0.00 │ mov 0x8(%rcx),%edx
> 0.00 │ and $0x1,%edx
> │ ↓ je 2b9
> │ rep_nop():
> 0.70 │2af: pause
> │ csd_lock_wait():
> 92.82 │ mov 0x8(%rcx),%edx
> 6.48 │ and $0x1,%edx
> 0.00 │ ↑ jne 2af
> 0.00 │2b9: ↑ jmp 282
>
>
> Given the patch at hand I was expecting lost IPIs, but I can't quite see
> anything getting lost.
I have no idea how that patch should be related to IPI and smp function
calls. It's changing the way how regular device interrupts and their
spurious counterpart are handled and not the way how IPIs are
handled. They are handled by direct vectors and do not go through
do_IRQ() at all.
Aside of that the commit just changes the way how the interrupt vector
of a regular device interrupt is stored and conveyed. The extra read and
write on the cache hot stack is hardly related to anything IPI.
Thanks,
tglx
