Re: geom labels as aliases

From: Andriy Gapon <avg_at_FreeBSD.org>
Date: Wed, 29 Sep 2021 16:36:27 +0300
On 29/09/2021 16:20, Warner Losh wrote:
> 
> 
> On Wed, Sep 29, 2021 at 2:37 AM Andriy Gapon <avg_at_freebsd.org 
> <mailto:avg_at_freebsd.org>> wrote:
> 
> 
>     For an introduction, I never liked consequences of GEOM labels being
>     implemented
>     as geoms.  And I specifically do not mean explicitly created labels via glabel
>     create / label.  I mean GEOM labels based on partition and filesystem labels,
>     disk IDs, etc,  And I don't like things such as opening a device via one label
>     spoiling other labels.
> 
>     So, I've been reading through some recent-ish changes by Warner for GEOM
>     aliases
>     and I've got this (maybe not so) bright idea, what if those labels were
>     implemented as aliases.  Would that work?  Would that require a lot of work?
> 
>     I did a quick hack as a proof of concept.
>     The change is quite small.
> 
>     diff --git a/sys/geom/label/g_label.c b/sys/geom/label/g_label.c
>     index 1df7e799b014..72b97d314de4 100644
>     --- a/sys/geom/label/g_label.c
>     +++ b/sys/geom/label/g_label.c
>     _at__at_ -418,12 +418,15 _at__at_ g_label_taste(struct g_class *mp, struct g_provider *pp,
>     int flags __unused)
>                      if (label[0] == '\0')
>                              continue;
>                      if (g_labels[i] != &g_label_generic) {
>     -                       mediasize = pp->mediasize;
>     +                       if (g_label_is_name_ok(label)) {
>     +                               g_provider_add_alias(pp, "%s%s",
>     +                                   g_labels[i]->ld_dirprefix, label);
>     +                       }
>                      } else {
>                              mediasize = pp->mediasize - pp->sectorsize;
>     +                       g_label_create(NULL, mp, pp, label,
>     +                           g_labels[i]->ld_dirprefix, mediasize);
>                      }
>     -               g_label_create(NULL, mp, pp, label,
>     -                   g_labels[i]->ld_dirprefix, mediasize);
>              }
>              g_access(cp, -1, 0, 0);
>        end:
> 
>     This seems to work in a basic smoke test of just booting up with the change.
[snip]
> 
>     Of course, the change can break existing scripts / tools that count on labels
>     having their own geoms.
> 
>     Also, it can be argued that it would be better to create such symbolic links in
>     userland.  It should be easy to port the label tasting code to userland and
>     hook
>     it to devd events.  That should cover all use cases except for the root
>     filesystem.
> 
>     Finally, I haven't tested glabel create / label yet, so the change may need
>     some
>     more work in that area.
> 
>     What is your opinion?
>     Is this something worth pursuing?
> 
> 
> Conrad did something similar in 5b9b571cb and wound up reverting it. Too many
> unforeseen issues, IIRC.

Whoosh, I completely missed that commit.
Unfortunately, the revert message it too terse:

     Revert r361838

     Reported by:    delphij

I wonder if the reported problem was indeed too hard to overcome.

However, I already see a number of issues in my simplistic change.
First of all, removing obsolete aliases.

I am thinking that maybe each "alias label" should still have a geom attached to 
the corresponding provider but without a (downstream) provider of its own.  This 
way we can still receive events for spoiling, etc without introducing an 
alternative provider.  We then can manipulate aliases in response to the 
upstream changes.

Yeah, I see, a lot more work is needed.

-- 
Andriy Gapon
Received on Wed Sep 29 2021 - 13:36:27 UTC

Original text of this message