Re: [viro-vfs:work.dcache2] [__dentry_kill()] 1b738f196e: stress-ng.sysinfo.ops_per_sec -27.2% regression

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

On Wed, Dec 06, 2023 at 05:42:34PM +0100, Mateusz Guzik wrote:

> That is to say your patchset is probably an improvement, but this
> benchmark uses kernfs which is a total crapper, with code like this in
> kernfs_iop_permission:
> 
>         root = kernfs_root(kn);
> 
>         down_read(&root->kernfs_iattr_rwsem);
>         kernfs_refresh_inode(kn, inode);
>         ret = generic_permission(&nop_mnt_idmap, inode, mask);
>         up_read(&root->kernfs_iattr_rwsem);
> 
> 
> Maybe there is an easy way to dodge this, off hand I don't see one.

At a guess - seqcount on kernfs nodes, bumped on metadata changes
and a seqretry loop, not that this was the only problem with kernfs
scalability.

That might account for sysinfo side, but not the unixbench - no kernfs
locks mentioned there.  OTOH, we might be hitting the wall on
->i_rwsem with what it's doing...