[Bug 230246] GPT label mishandling

[bugzilla-noreply at freebsd.org]

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=230246

--- Comment #15 from andrew at tao11.riddles.org.uk ---
Just to wrap this up for the record (I explained all this to the OP in another
place):

There's no bug in the code as such, the odd behavior is just an artifact of how
geom tasting works, and the fact that it works rather unpredictably when geoms
are being created and removed while the system is up (rather than at boot
time).

Specifically, the sort of thing that happens is that a partition might become
visible via many different provider paths, such as ada0p1, diskid/DISK-NNNNp1,
gpt/partname, gptid/someuuid, and then if there is existing metadata at the
partition's last sector, in this case for gmirror, the mirror geom might attach
to any of the providers. If it attaches (say) diskid/DISK-NNNNp1, then the
exclusive open of the underlying disk causes the gpart instances providing
ada0p*, gpt/*, etc. to detach, so those labels go away.

This shows up as more of an issue with GPT because the combination of GPT and
glabel produces many providers, often inconveniently many. (For example you can
end up with "gmirror stop" failing to work, because as soon as the mirror
closes, it becomes visible via another provider and immediately starts again.)

So this is ultimately, in my view, a "user experience" issue.

-- 
You are receiving this mail because:
You are the assignee for the bug.