[pve-devel] [PATCH container] lxc start: warn in case of conflicting lxc.idmap entries

[Wolfgang Bumiller]

On Thu, Mar 16, 2023 at 04:07:34PM +0100, Friedrich Weber wrote:
> Thanks for the review!
> 
> On 16/03/2023 14:59, Wolfgang Bumiller wrote:
> > Both seem a bit excessive to me.
> > 
> > Let's look at the data:
> > We have a set of ranges consisting of a type, 2 starts and a count.
> > The types are uids and gids, so we can view those as 2 separate
> > instances of sets of [ct_start, host_start, count].
> > Since neither the container nor the host sides must overlap we can -
> > again - view these as separate sets of container side [start, count] and
> > host side [start, count].
> > In other words, we can see the entire id map as just 4 sets of [start,
> > count] ranges which must not overlap.
> > 
> > So I think all we need to do is sort these by the 'start' value, and for
> > each element make sure that
> > 
> >      prevous_start + previous_count <= current_start
> > 
> > And yes, that means we need to sort $id_maps twice, once by ct id, once
> > by host id, and then iterate and do the above check.
> > 
> > Should be much shorter (and faster).
> 
> Yeah, good point, splitting $id_maps into separate uid/gid maps, and then
> sorting+iterating twice (I'll call this the "sorting algorithm" below) does
> sound more understandable than the current ad-hoc approach, and faster too.
> 
> However, one small benefit of iterating over $id_maps in its original order
> (instead of sorting) is that the error message always references the *first*
> invalid map entry in the config, e.g. (omitting host uids for clarity)
> 
>   1) u 1000 <...> 100
>   2) u 950 <...> 100
>   3) u 900 <...> 100
>   4) u 850 <...> 100
> 
> The sorting algorithm would error on entry 3, which might suggest to users
> that entries 1-2 are okay (which they are not). The current algorithm errors
> on line 2 already. Similar things would happen with interleaved uid/gid
> mappings, I guess.
> 
> But I'm not sure if this really matters to users. What do you think?

Since it's about helping out users, even better would be to collect all
the errors together and than die() with a message containing all of
them.
And then the order doesn't matter again ;-)