lkml.org 
[lkml]   [2015]   [May]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC v1 07/11] net/mlx4: Cache irq_desc->affinity instead of irq_desc
On 5/5/2015 5:53 PM, Thomas Gleixner wrote:
> On Tue, 5 May 2015, Amir Vadai wrote:
>
>> On Mon, May 4, 2015 at 6:10 PM, Thomas Gleixner <tglx@linutronix.de> wrote:
>>> On Mon, 4 May 2015, Amir Vadai wrote:
>>>
>>>> On Mon, May 4, 2015 at 6:15 AM, Jiang Liu <jiang.liu@linux.intel.com> wrote:
>>>>> The field 'affinity' in irq_desc won't change once the irq_desc data
>>>>> structure is created. So cache irq_desc->affinity instead of irq_desc.
>>>>> This also helps to hide struct irq_desc from device drivers.
>>>>
>>>> Hi Jiang,
>>>>
>>>> I might not understand the new changes irq core, but up until now
>>>> affinity was changed when the user changed it through
>>>> /proc/irq/<IRQ>/smp_affinity.
>>>> This code is monitoring the affinity from the napi_poll context to
>>>> detect affinity changes, and prevent napi from keep running on the
>>>> wrong CPU.
>>>> Therefore, the affinity can't be cached at the beginning. Please
>>>> revert this caching.
>>>
>>> Sigh. All of this is completely wrong. Polling in the internals of
>>> irq_desc is just crap.
>> Hi,
>>
>>>
>>> irq_set_affinity_notifier() has been added for exactly this purpose.
>> rmap is already registered on irq_set_affinity_notifier().
>> Actually, I did post an extension to the affinity notifier, to use
>> affinity chain which you rejected [1] (and BTW, convinced me that this
>> is not the way to do it).
>> Current code is according to your suggestion.
>>
>> [1] - https://lkml.org/lkml/2014/5/26/222
>
> No, it's not. I wrote:
>
> "So what's the point of adding notifier call chain complexity,
> ordering problems etc., if you can simply note the fact that the
> affinity changed in the rmap itself and check that in the RX path?"
>
> I did not tell you to fiddle in irqdesc, really.
>
> Thanks,
>
> tglx
>

It is problematic, since I didn't want to link the napi polling affinity
issue and CONFIG_CPU_RMAP being selected. That's why I CC'd you when
submitted the patch [1]. Guess it was my bad that I didn't do it in a
better way that would give you a chance to fully review it.

Anyway, I understand now that I shouldn't access irqdesc attributes from
a device driver.
I will see if I can live with disabling this whole feature (migrate a
currently looping NAPI when affinity is changed), when CONFIG_CPU_RMAP
is not selected. If not, I might modify the rmap code to give me this
information even when it is not selected. I assume that rmap notifier
callback can access this information.

Thanks,
Amir



\
 
 \ /
  Last update: 2015-05-07 13:21    [W:0.049 / U:0.192 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site