Re: [PATCH 2/3 v3] x86/topology: Avoid wasting 128k for package id array

[Prarit Bhargava <prarit@redhat.com>]

On 10/20/2017 05:03 AM, Thomas Gleixner wrote:
> On Thu, 19 Oct 2017, Prarit Bhargava wrote:
>>  static void remove_siblinginfo(int cpu)
>>  {
>> -	int sibling;
>> +	int phys_pkg_id, sibling;
>>  	struct cpuinfo_x86 *c = &cpu_data(cpu);
>>  
>>  	for_each_cpu(sibling, topology_core_cpumask(cpu)) {
>> @@ -1529,6 +1526,12 @@ static void remove_siblinginfo(int cpu)
>>  	cpumask_clear(topology_core_cpumask(cpu));
>>  	c->phys_proc_id = 0;
>>  	c->cpu_core_id = 0;
>> +
>> +	phys_pkg_id = c->phys_pkg_id;
>> +	c->phys_pkg_id = U16_MAX;
> 
> This leaves c->logical_proc_set = 1, which is inconsistent at best. I have
> no idea why we need this logical_proc_set flag at all.
> 
>> +	if (topology_phys_to_logical_pkg(phys_pkg_id) < 0)
>> +		logical_packages--;
> 
> Now this has another issue. Depending on hotplug ordering the logical
> package association can change across hotplug operations. I don't know it
> that's an issue, but this needs to be analyzed before we merge that.
> 

Thanks for making me look at this Thomas.

Andi, it looks like this is unfortunately an issue.  I have a reworked
patchset that fixes by dynamically allocating a u16
logical_to_physical_package array that

a)  maps the logical to physical package IDs,
b) is dynamically sized to the value of logical_packages, and,
c) is of size logical_packages (not MAX_APICS)

This will resolve the problem that Thomas has pointed out, and it
would address the issue of wasting memory.

I'm waiting for a 4S & 8S system to test the LTP hotplug tests on
and then I'll post.  Andi, I'll send it to you in private email for
a quick review.

P.

> Thanks,
> 
> 	tglx
> 
>