freebsd stable - Sep 2016 - Regression with revision 303970 (was kern.proc.pathname failure while booting from zfs)

On Sun, Sep 04, 2016 at 04:11:39PM +0300, Andriy Gapon
wrote:
> On 04/09/2016 11:24, Andriy Gapon wrote:
> > On 27/08/2016 22:09, Frederic Chardon wrote:
> >>> Anybody is able to reproduce this behavior or is it a local
problem?
> >> Reverting 303970 solves this issue. gcore and adb works again, and
I
> >> can start the vboxnet service.
> >> I recreated my boot pool with no properties defined, just to be
sure.
> > 
> > I can not reproduce this issue here.
> 
> I was not trying hard enough.  I've just reproduced the problem using a
> non-default normalization property.  The issue is that 303970 disabled the
use
> of VFS name cache when any name "mangling" (normalization,
case-insensitivity)
> is enabled.  And apparently I misunderstood how vop_stdvptocnp() works. 
So,
> right now zfs_vptocnp() is broken when its argument is a non-directory
vnode.
> That fact is masked when the name cache is used and is exposed otherwise.
It is only masked when name cache has an entry for the vnode.  So sometimes
vn_fullpath() should be broken even if no normalization is applied.

OTOH, classic filesystems like UFS do not have any other means to translate
non-directory inode to name and parent at all, except the namecache hint.

On 04/09/2016 17:51, Konstantin Belousov wrote:
> It is only masked when name cache has an entry for the vnode.  So sometimes
> vn_fullpath() should be broken even if no normalization is applied.
Yes, this is true.
> OTOH, classic filesystems like UFS do not have any other means to translate
> non-directory inode to name and parent at all, except the namecache hint.
In fact, this is true for ZFS as well.  While ZFS znodes have an attribute that
specifies a (single) parent, it's obviously unreliable for files, because a
file
can be linked into multiple directories and then unlinked from a directory
specified by the attribute.

So, at the moment I do not have any good ideas on how to make this work.
Maybe trying to use the parent attribute and failing when it's inconsistent
would be good enough...


-- 
Andriy Gapon