Re: [PATCH 3/3] mm/mmu_gather: send tlb_remove_table_smp_sync IPI only to CPUs in kernel mode

[Valentin Schneider <vschneid@redhat.com>]

On 05/04/23 14:05, Frederic Weisbecker wrote:
>  static void smp_call_function_many_cond(const struct cpumask *mask,
>                                       smp_call_func_t func, void *info,
> @@ -946,10 +948,13 @@ static void smp_call_function_many_cond(const struct cpumask *mask,
>  #endif
>                       cfd_seq_store(pcpu->seq_queue, this_cpu, cpu, CFD_SEQ_QUEUE);
>                       if (llist_add(&csd->node.llist, &per_cpu(call_single_queue, cpu))) {
> -				__cpumask_set_cpu(cpu, cfd->cpumask_ipi);
> -				nr_cpus++;
> -				last_cpu = cpu;
> -
> +				if (!(scf_flags & SCF_NO_USER) ||
> +				    !IS_ENABLED(CONFIG_GENERIC_ENTRY) ||
> +				     ct_state_cpu(cpu) != CONTEXT_USER) {
> +					__cpumask_set_cpu(cpu, cfd->cpumask_ipi);
> +					nr_cpus++;
> +					last_cpu = cpu;
> +				}

I've been hacking on something like this (CSD deferral for NOHZ-full),
and unfortunately this uses the CPU-local cfd_data storage thing, which
means any further smp_call_function() from the same CPU to the same
destination will spin on csd_lock_wait(), waiting for the target CPU to
come out of userspace and flush the queue - and we've just spent extra
effort into *not* disturbing it, so that'll take a while :(

I don't have much that is in a shareable state yet (though I'm supposed to
talk some more about it at OSPM in <2 weeks, so I'll have to get there),
but ATM I'm playing with
o a bitmask (like in [1]) for coalescable stuff such as do_sync_core() for
  x86 instruction patching
o a CSD-like queue for things that need to pass data around, using
  statically-allocated storage (so with a limit on how much it can be used) - the
  alternative being allocating a struct on sending, since you don't have a
  bound on how much crap you can queue on an undisturbed NOHZ-full CPU...

[1]: https://lore.kernel.org/all/20210929152429.067060646@infradead.org/