Re: [RFC PATCH] vfs: dodge strlen() in vfs_readlink() when ->i_link is populated
From: Jan Kara
Date: Mon Nov 18 2024 - 06:54:12 EST
On Mon 18-11-24 09:53:57, Mateusz Guzik wrote:
> This gives me about 1.5% speed up when issuing readlink on /initrd.img
> on ext4.
>
> Signed-off-by: Mateusz Guzik <mjguzik@xxxxxxxxx>
> ---
>
> I had this running with the following debug:
>
> if (strlen(link) != inode->i_size)
> printk(KERN_CRIT "mismatch [%s] %l %l\n", link,
> strlen(link), inode->i_size);
>
> nothing popped up
Then you didn't run with UDF I guess ;). There inode->i_size is the length
of on-disk symlink encoding which is definitely different from the length
of the string we return to VFS (it uses weird standards-defined cross OS
compatible encoding of a path and I'm not even mentioning its own special
encoding of character sets somewhat resembling UCS-2).
> I would leave something of that sort in if it was not defeating the
> point of the change.
>
> However, I'm a little worried some crap fs *does not* fill this in
> despite populating i_link.
>
> Perhaps it would make sense to keep the above with the patch hanging out
> in next and remove later?
>
> Anyhow, worst case, should it turn out i_size does not work there are at
> least two 4-byte holes which can be used to store the length (and
> chances are some existing field can be converted into a union instead).
I'm not sure I completely follow your proposal here...
Honza
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR