Re: [PATCH V3] x86, UV: BAU performance and error recovery

[JD <jd1008@gmail.com>]

On 05/28/2010 04:36 PM, Cliff Wickman wrote:
> On Fri, May 28, 2010 at 03:23:21PM -0700, Greg KH wrote:
>    
>> On Fri, May 28, 2010 at 02:43:35PM -0500, Cliff Wickman wrote:
>>      
>>> On Fri, May 28, 2010 at 09:47:25AM -0700, Greg KH wrote:
>>>        
>>>> On Fri, May 28, 2010 at 09:30:25AM -0700, H. Peter Anvin wrote:
>>>>          
>>>>> On 05/28/2010 06:33 AM, Cliff Wickman wrote:
>>>>>            
>>>>>> - adds modification of tuning variables through /proc/sgi_uv
>>>>>>              
>>>>> Adding new directories in /proc for a proprietary architecture is
>>>>> frowned upon, to put it mildly.  At the very least try to find a place
>>>>> in sysfs for it.
>>>>>
>>>>> [Cc: gregkh in order to find a place in sysfs]
>>>>>            
>>>> Hm, what type of files are you needing here?  Do they corrispond with
>>>> any specific hardware devices?  If so, just put them on the hardware
>>>> devices in sysfs.
>>>>
>>>> thanks,
>>>>
>>>> greg k-h
>>>>          
>>> There is an SGI-specific directory in sysfs: /sys/firmware/sgi_uv
>>> though the tuning variables for the handling of the hardware Broadcast
>>> Assist Unit don't fit there very logically.
>>>
>>> The BAU's statistics are already displayed through the UV-only
>>> /proc/sgi_uv/ptc_statistics.  This was deemed necessary because of the
>>> potentially very large size of that file --- it is still true that a /sys file
>>> is limited to one page, is it not?
>>>        
>> It is limited to one single value, which would never be larger than a
>> page, so yes, it is limited to one page.
>>
>> What kind of information are you showing here?  Should this thing just
>> be in debugfs instead?  You can do whatever you want there.
>>
>> thanks,
>>
>> greg k-h
>>      
> /proc/sgi_uv/bau_tunables would be a read/write file to display and change
> nine threshold and delay values for tuning the BAU driver.
>
> I like debugfs, except that a distro may not build the kernel with it
> configured on.  The tunables should be available as administrative
> options on a customer kernel, not just as a development tool.
>
> And in our case the distros are already building with other such writable
> options in /proc/sgi_uv.  We'd like to postpone a wholesale move of such
> options (assuming there will be some better place) and stay with the existing
> location for this release.
>
> I know we (the community) would like to move non-process info out of /proc.
> It seems to me that we need a similar filesystem for large and/or
> administrative files.  That's my perspective.
>
> -Cliff
>    
Would the similar filesystem you propose be backed by permanent storage, 
unlike /proc?