Re: [RFC, PATCH] avoid some atomics in open()/close() for monothreaded processes

[Eric Dumazet <dada1@cosmosbay.com>]

Andrew Morton a écrit :
> Eric Dumazet <dada1@cosmosbay.com> wrote:
>> Goal : Avoid some locking/unlocking 'struct files_struct'->file_lock for mono 
>> threaded processes.
>>
>> We define files_multithreaded() function .
>>
>> static inline int files_multithreaded(const struct files_struct *files)
>> {
>>         return sizeof(files->file_lock) > 0 && atomic_read(&files->count) > 1;
>> }
> 
> That's bascially sizeof(spinlock_t).  That's architecture dependent and
> varies wildly according to the day of week.

I used sizeof(files->file_lock) instead of sizeof(spinlock_t) because I found 
it more explicit , while not using ugly ifdefs.

> 
> It _might_ work in all situations - probably you checked that.  But I still
> wouldn't do it because it might break in the future.  Let's be explicit and
> stick the appropriate ifdefs in there.
> 
> I'd also consider renaming it to files_shared() - processes are
> multithreaded, not data structures.

Thanks for the feedback, I will redo the patch and test it on various 
platforms before resubmit (including performance data :) )

> 
> Once you're done with that we should change fget_light() and fput_light() to
> use this helper.  Separate patch.

Hum... this discussion is not relevant with fget_light() unless I mistaken.

Nowadays, this function doesnt take spinlock thanks to RCU

Eric