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

[Eric Dumazet <dada1@cosmosbay.com>]

Linus Torvalds a écrit :
> 
> I assume that the *only* reason for having multiple dentries is really 
> just the output in /proc/<pid>/fd/, right? Or is there any other reason to 
> have separate dentries for these pseudo-files?
> 
> It's a bit sad to waste that much memory (and time) on something like 
> that. I bet that the dentry setup is a noticeable part of the whole 
> sigfd()/timerfd() setup. It's likely also a big part of any memory 
> footprint if you have lots of them.
> 
> So how about just doing:
>  - do a single dentry
>  - make a "struct file_operations" member function that prints out the 
>    name of the thing in /proc/<pid>/fd/, and which *defaults* to just 
>    doing the d_path() on the dentry, but special filesystems like this 
>    could do something else (like print out a fake inode number from the 
>    "file->f_private_data" information)
> 
> There seems to really be no downsides to that approach. No existing 
> filesystem will even notice (they'll all have NULL in the new f_op 
> member), and it would allow pipes etc to be sped up and use less memory.
> 

I would definitly *love* saving dentries for pipes (and sockets too), but how 
are you going to get the inode ?

pipes()/sockets() can use read()/write()/rw_verify_area() and thus need 
file->f_path.dentry->d_inode (so each pipe needs a separate dentry)

Are you suggesting adding a new "struct file_operations" member to get the inode ?
Or re-intoducing an inode pointer in struct file ?