Re: [PATCH 6/7] work_on_cpu: use on x86/kernel/cpu/intel_cacheinfo.c

[Oleg Nesterov <oleg@tv-sign.ru>]

On 10/23, Rusty Russell wrote:
>
> +static int __cpuinit detect_cache_attributes(unsigned int cpu)
> +{
> +	int			retval;
> +
> +	if (num_cache_leaves == 0)
> +		return -ENOENT;
> +
> +	per_cpu(cpuid4_info, cpu) = kzalloc(
> +	    sizeof(struct _cpuid4_info) * num_cache_leaves, GFP_KERNEL);
> +	if (per_cpu(cpuid4_info, cpu) == NULL)
> +		return -ENOMEM;
> +
> +	retval = work_on_cpu(cpu, get_cpu_leaves, NULL);

This doesn't look right.

cache_add_dev()->cpuid4_cache_sysfs_init()->detect_cache_attributes()
is called by CPU_ONLINE under cpu_hotplug_begin(), this is deadlockable.

Yes, the caller (cpu_hotplug.active_writer) can safely take
get_online_cpus(), but if we have another pending work which needs
get_online_cpus(), we have the deadlock.

Oleg.