Re: generic_permission() optimization

["Theodore Ts'o" <tytso@mit.edu>]

On Sun, Apr 13, 2025 at 02:52:32PM +0200, Mateusz Guzik wrote:
> On Sun, Apr 13, 2025 at 2:40 PM Theodore Ts'o <tytso@mit.edu> wrote:
> >
> > On Sun, Apr 13, 2025 at 11:41:47AM +0200, Mateusz Guzik wrote:
> > > This is the rootfs of the thing, so I tried it out with merely
> > > printing it. I got 70 entries at boot time. I don't think figuring out
> > > what this is specifically is warranted (it is on debian though).
> >
> > Well, can you run:
> >
> > debugfs -R "stat <INO>" /dev/ROOT_DEV
> >
> 
> attached full list after boot

So it looks like the test is working corretly.  Most of the inodes
either (a) have a Posix ACL defined, so we were definitely doing the
right thing, or (b) had a user.crtime_usec xattr.  My personal opinion
is that crtime is fairly pointless, and having microsecond accuracy on
the creation time is *completely* pointless, but we can't stop
programs from doing that.  (Therre was also a single xattr field that
contained the xattr user.random-seed-creditable.)

So it will ultimately come down to how much user think performance
compares to microsecond-level accuracy on crtime, which as far as I
know, no Linux programs other than samba / CIFS servers who want
Microsoft feature-for-feature compatibility care about.

(Or SELinux when it sets security ID's, but if you are using SELinux a
few extra branch instructions are the *least* of your performace
headaches....)

						- Ted