Re: [patch v2] epoll use a single inode ...

["H. Peter Anvin" <hpa@zytor.com>]

Eric Dumazet wrote:
> Linus Torvalds a écrit :
>>
>> On Tue, 6 Mar 2007, Eric Dumazet wrote:
>>> I did a user space program, attached to this mail.
>>>
>>> I rewrote the reciprocal_div() for i386 so that one multiply is used.
>>
>> Ok, this is definitely faster on Core 2 as well, so "numbers talk, 
>> bullshit walks". No more objections.
> 
> And the numbers were ? :)
> 
>>
>> (That said, I bet you could do even better for octal and hex numbers, 
>> so if you *really* want to speed things up, you should just make a 
>> special-case routine for each base (there's just three of them), and 
>> you can then also optimize the base-10 thing much better (you can do 
>> two digits at a time by dividing by 100, etc)
> 
> Well, given that sprintf() is frequently called only for pipe/sockets 
> creation, we probably better :
> 
> 1) wait a very clever idea to suppress individual dentry per 
> pipe/sockets (no more sprintf() at pipe/socket setup)
> 
> 2) delay the sprintf() only if needed as you mentioned in a previous 
> mail (when someone wants ls -l /proc/pid/fd/....), since their dentries 
> are not anymore inserted in the global dcache hash, they could stay with 
> a (nul) dname.

Yes, the right thing to do is probably to only generate these strings 
when someone tries to list them, not on every socket/pipe/epoll 
creation.  One can assign a counter and keep it as a binary value at the 
start, but create the strings when necessary.

	-hpa