Re: [PATCH 07/16] smp: optimize smp_call_function_many_cond()

[Andy Shevchenko]

On Mon, Jul 18, 2022 at 12:28:35PM -0700, Yury Norov wrote:
> smp_call_function_many_cond() is often passed with cpu_online_mask.
> If this is the case, we can use num_online_cpus(), which is O(1)
> instead of cpumask_{first,next}(), which is O(N).
> 
> It can be optimized further: if cpu_online_mask has 0 or single bit
> set (depending on cpu_online(this_cpu), we can return result without
> AND'ing with user's mask.

> Caught with CONFIG_DEBUG_BITMAP:
> [    7.830337] Call trace:
> [    7.830397]  __bitmap_check_params+0x1d8/0x260
> [    7.830499]  smp_call_function_many_cond+0x1e8/0x45c
> [    7.830607]  kick_all_cpus_sync+0x44/0x80
> [    7.830698]  bpf_int_jit_compile+0x34c/0x5cc
> [    7.830796]  bpf_prog_select_runtime+0x118/0x190
> [    7.830900]  bpf_prepare_filter+0x3dc/0x51c
> [    7.830995]  __get_filter+0xd4/0x170
> [    7.831145]  sk_attach_filter+0x18/0xb0
> [    7.831236]  sock_setsockopt+0x5b0/0x1214
> [    7.831330]  __sys_setsockopt+0x144/0x170
> [    7.831431]  __arm64_sys_setsockopt+0x2c/0x40
> [    7.831541]  invoke_syscall+0x48/0x114
> [    7.831634]  el0_svc_common.constprop.0+0x44/0xfc
> [    7.831745]  do_el0_svc+0x30/0xc0
> [    7.831825]  el0_svc+0x2c/0x84
> [    7.831899]  el0t_64_sync_handler+0xbc/0x140
> [    7.831999]  el0t_64_sync+0x18c/0x190
> [    7.832086] ---[ end trace 0000000000000000 ]---
> [    7.832375] b1:		ffff24d1ffd98a48
> [    7.832385] b2:		ffffa65533a29a38
> [    7.832393] b3:		ffffa65533a29a38
> [    7.832400] nbits:	256
> [    7.832407] start:	0
> [    7.832412] off:	0
> [    7.832418] smp: Bitmap: parameters check failed
> [    7.832432] smp: include/linux/bitmap.h [363]: bitmap_and

Documentation specifically says:

https://www.kernel.org/doc/html/latest/process/submitting-patches.html#backtraces-in-commit-mesages

...

> +	default:
> +		if (mask == cpu_online_mask)
> +			return true;

Instead, put (missed) break; here and do "default" case together below.

> +	cpu = cpumask_first_and(mask, cpu_online_mask);
> +	if (cpu == this_cpu)
> +		cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
> +
> +	return cpu < nr_cpu_ids;

...

> +	run_remote = __need_remote_exec(mask, this_cpu);

>  

Now you may remove this blank line.

>  	if (run_remote) {

-- 
With Best Regards,
Andy Shevchenko