Re: [rfc-patch, bugfix] x86-microcode

[Max Krasnyansky <maxk@qualcomm.com>]

Dmitry Adamushko wrote:
> 2008/8/4 Max Krasnyansky <maxk@qualcomm.com>:
>> Dmitry Adamushko wrote:
>>> Hi,
>>>
>>>
>>> [ consider it a pre-release and RFC... I'm a bit in hurry now and just send what I have got by this moment.
>>> Although, I expect it to be workable ]
>>>
>>>
>>> this change is supposed to fix bug#11197 (note, its name "Oops in microcode sysfs registration" is misleading)
>>>
>>> The problem description can be found here:
>>> http://www.ussg.iu.edu/hypermail/linux/kernel/0807.3/3791.html
>>> or
>>> http://lkml.org/lkml/2008/7/24/260
>>>
>>> perhaps it does look quite bulky for -rc, although it's mainly move-redesign-some-bits of the code and
>>> I tried to preserve the original logic (even if it looked like a possible optimizations might had been applied)
>>> as much as possible.
>>>
>>> The basic idea is that we introduce another mechanism to run ucode-updates on a target cpu
>>> and replace set_cpus_allowed_ptr() in (1) cpu-hotplug events and (2) module load.
>>>
>>>
>>> [1/2] x86-microcode: generic updates
>>>
>>> Basically, it introduces microcode_update_cpu() which can be run either from start_secondary()
>>> (perhaps via a function pointer) or scheduled via keventd ([2/2]) and reworks the logic of cpu-hotplug events.
>>>
>>> [2/2] x86-microcode: do updates via workqueue
>> Looks good to me. You did not change the old interface which still does
>> set_cpus_allowed_ptr() (ie racy against sched_setaffinity())
> 
> ah, right. I've only fixed reload_store() vs. cpu_down().
> 
> sched_setaffinity() calls get_online_cpus() so a pair of
> cpu_maps_update_begin() + cpu_hotplug_begin() would need to be used by
> the following code:
> 
> old_mask = p->cpus_allowed;
> set_cpus_allowed_ptr(p, new_mask);
> // do_something
> set_cpus_allowed_ptr(p, old_mask);
> 
> or we'd introduce another mutex to be used in both cases...  but I
> think we just can get rid of most of use-cases (if not all) by
> replacing them with schedule_work_on().
Agree. That's I suggested we do it later (ie convert those to use
schedule_work_on()).

Max