Re: [CFT] Re: VFS deadlock ?

[Al Viro <viro@ZenIV.linux.org.uk>]

On Thu, Mar 21, 2013 at 09:55:10PM -0700, Linus Torvalds wrote:
> However, I do wonder if we could take another approach... There's
> really no reason why we should look up an old inode with iget_locked()
> in proc_get_inode() and only fill it in if it is new. We might as well
> just always create a new inode. The "iget_locked()" logic really comes
> from the bad old days when the inode was the primary data structure,
> but it's really the dentry that is the important data structure, and
> the inode might as well follow the dentry, instead of the other way
> around.
> 
> So why not just use "new_inode_pseudo()" instead? IOW, something like
> this (totally untested, natch) patch? This way, if you have a new
> dentry, you are guaranteed to always have a new inode. None of the
> silly inode alias issues..
> 
> This seems too simple, but I don't see why iget_locked() would be any
> better. It just wastes time hashing the inode that we aren't really
> interested in hashing. The inode is always filled by the caller
> anyway, so we migth as well just get a fresh pseudo-filesystem inode
> without any crazy hashing..

Umm...
static int proc_delete_dentry(const struct dentry * dentry)
{
        return 1;
}

static const struct dentry_operations proc_dentry_operations =
{
        .d_delete       = proc_delete_dentry,
};

IOW, dcache retention in procfs is inexistent and the damn thing tries
to cut down on the amount of inode allocation/freeing/filling.

I agree that we could get rid of icache retention there and everything
ought to keep working.  Hell knows...  It applies only to the stuff that
isn't per-process, so it's probably not particulary hot anyway, but it
does need profiling...  OTOH, we probably could mark "stable" dentries
in some way and let proc_delete_dentry() check that flag in proc_dir_entry -
I seriously suspect that really hot non-per-process ones are of the
"never become invalid" variety.