Re: [PATCH] sysfs: Cache the last sysfs_dirent to improve readdir scalability

[Linus Torvalds <torvalds@linux-foundation.org>]

On Fri, 1 Jan 2010, Eric W. Biederman wrote:
> 
> When sysfs_readdir stops short we now cache the next sysfs_dirent to
> return to user space in filp->private_data.  There is no impact on the
> rest of sysfs by doing this and in the common case it allows us to
> pick up exactly where we left off with no seeking.
> 
> Additionally I drop and regrab the sysfs_mutex around filldir to avoid
> a page fault arbitrarily increasing the hold time on the sysfs_mutex.

Ok, looks mostly sane, but a few things look odd.

>  
> -	if (filp->f_pos == 0) {
> +	if (!pos && filp->f_pos == 0) {
>  		ino = parent_sd->s_ino;
>  		if (filldir(dirent, ".", 1, filp->f_pos, ino, DT_DIR) == 0)
>  			filp->f_pos++;
>  	}
> -	if (filp->f_pos == 1) {
> +	if (!pos && filp->f_pos == 1) {
>  		if (parent_sd->s_parent)
>  			ino = parent_sd->s_parent->s_ino;
>  		else
> @@ -847,29 +879,35 @@ static int sysfs_readdir(struct file * filp, void * dirent, filldir_t filldir)
>  		if (filldir(dirent, "..", 2, filp->f_pos, ino, DT_DIR) == 0)
>  			filp->f_pos++;
>  	}
> -	if ((filp->f_pos > 1) && (filp->f_pos < INT_MAX)) {
> -		mutex_lock(&sysfs_mutex);
> -
> -		/* Skip the dentries we have already reported */
> -		pos = parent_sd->s_dir.children;
> -		while (pos && (filp->f_pos > pos->s_ino))
> -			pos = pos->s_sibling;
> +	/* EOF? */
> +	if (!pos && filp->f_pos > 2)
> +		return 0;

These are all incorrect in the presense of 'lseek'. You can't do that

	if (!pos && "test filp->f_pos")

thing, because you get all the wrong results for both the case of an lseek 
before doing any readdir (which is undefined behavior, so I guess that's 
technically ok) _and_ for the 'lseek back to zero _after_ doing a readdir' 
case (which is _not_ undefined behavior!

It looks like it might be easy to fix by making a sysfs_llseek() function 
that does something like

	.. sysfs_llseek(..)
	{
		mutex_lock(&sysfs_mutex);
		sysfs_release();
		filp->private_data = NULL;
		mutex_unlock(&sysfs_mutex);

		return generic_file_llseek(..);
	}

or similar.  Except themn you'll need to change the EOF condition testing 
and turn it into a re-validation event. Or maybe do the re-validation in 
sysfs_llseek() itself, rather than just dropping the cached data.

Hmm? I haven't thought it through very deeply, so maybe I'm missing 
something.

		Linus