Re: [PATCH] removes filp_count_lock and changes nr_files type to atomic_t

[Nick Piggin <nickpiggin@yahoo.com.au>]

Eric Dumazet wrote:
> Nick Piggin a écrit :
> 

>> Would you just be able to add the atomic sysctl handler that
>> Christoph suggested?
>>
> 
> Quite a lot of work indeed, and it would force to convert 3 int 
> (nr_files, nr_free_files, max_files) to 3 atomic_t. I feel bad 
> introducing a lot of sysctl rework for a tiny change (removing 
> filp_count_lock)
> 

True, I didn't notice that.

>> This introduces lost update problems. 2 CPUs may store to nr_files
>> in the opposite order that they incremented atomic_nr_files.
>>
> 
> That's true, and the difference can be relatively important in case of 
> preemption.
> 
> Each time the true and correct value  (atomic_nr_files) is updated, a 
> copy is done on nr_files : as nr_files is only used to be a guard value 
> against too many file allocations, a somewhat 'lazy' value has no impact 
> at all.
> 

OK, well I would prefer you do the proper atomic operations throughout
where it "really matters" in file_table.c, and do your lazy synchronize
with just the sysctl exported value.

Unless the fs people had a problem with that.

And you may as well get rid of the atomic_inc_return which can be more
expensive on some platforms and doesn't buy you much.
   atomic_inc;
   atomic_read;
Should be enough if you don't care about lost updates here, yeah?

Nick

-- 
SUSE Labs, Novell Inc.

Send instant messages to your online friends http://au.messenger.yahoo.com