Re: LVM snapshot broke between 4.14 and 4.16
From: Mike Snitzer
Date: Sat Aug 04 2018 - 11:19:59 EST
On Sat, Aug 04 2018 at 1:20am -0400,
Theodore Y. Ts'o <tytso@xxxxxxx> wrote:
> On Fri, Aug 03, 2018 at 03:30:37PM -0400, Mike Snitzer wrote:
> > I was trying to give context for the "best to update lvm2 anyway"
> > disclaimer that was used. Yeah, it was specious.
> Well, it seemed to indicate a certain attitude that both Linus and I
> are concerned about. I tried to use more of a "pursuading" style to
> impress why that attitude was not ideal/correct. Linus used a much
> more assertive style (e.g., "Hell, no!").
[I debated just ignoring this portion of your reply but it needs to be
dealt with directly]
I prefer how Linus handled it (at least he was timely with his
follow-ups). Your initial reply where you joined a fragment of my
initial reply with Zdenek's (we sent simultaneously, each half way
around the world) served to merge Zdenek and myself into one fictional
straw-man you both could attack. If you have something to say to _me_
address me directly; don't put words in my mouth because you thought I
had a complete mind-meld with someone else.
And please don't act like this wasn't already beaten to death yesterday;
which left me (as DM maintainer) initially _unwarrantedly_ compromised.
There was a block regression that I wasn't aware of but someone on my
broader team (Zdenek) papered over it in userspace rather than report it
as a regression.
I did brush off the seriousness of side-effects on readonly dm-snapshot
("Because dm-snapshot"). But that doesn't speak to some systemic
"problem" you seem to be concerned about.
> > And yeah, that isn't a good excuse to ignore it but: dm-snapshot is a
> > steaming pile as compared to dm thin-provisioning...
> On a side note, this is the first that I've heard the assertion that
> dm-thin was better than dm-snapshot.
You don't follow DM much, that's fine. But thinp is considerably more
powerful for modern use-cases.
> My impression was that dm-snapshot was a proven code base, that only
> did one thing and (as far as I could tell) did it well. In contrast,
> dm-thin is much newer code, **far** more complex, with functionality
> and corner cases approaching that of a file system ---
dm-snapshot's scaling is _awful_. This is due to the N-way copy-out
penalty associated with N snapshots. So lots of snapshots perform very
very slowly. Even one snapshot is slow compared to dm-thinp.
dm-thin (2011) certainly is newer than dm-snapshot (well before 2005),
and yes dm-thin is complex, but dm-snapshot's code isn't exactly
"simple". The on-disk layout is but that simplicity contributes to why
it doesn't scale at all.
DM thin is a modern approach to snapshots grounded in the same
btree-based concepts and design used by btrfs. Given dm-thinp's
requirements and how it has been deployed and pushed _hard_ it really is
holding up amazingly well.
> and just to be even more exciting, it [dm-thin] doesn't have an
> fsck/repair tool to deal with corrupted metadata.
That's one definition for "exciting" on your Friday night ... ;)
The documentation was outdated, see this thread:
Where I shared that this Documentation update was staged for 4.19:
That said, thin_repair has shown itself to be hit-or-miss. There are
certain corruptions that it doesn't cope with well (leaving the metadata
"repaired" but the result is an empty thin-pool). Those cases are more
rare but still occur.
So repairing thinp corruption can require escalations through
"enterprise support" (which results in fixes to thin_repair, etc).
> In your opinion, is it because you disagree with the assumption that
> dm-thin is scary? Or is the argument that dm-snapshot is even
Apples and oranges. DM thinp is complex but necessarily so.
dm-snapshot is still complex yet only covers legacy and narrow (read:
now useless) use-cases.
In the same thread I referenced above, see how Drew Hastings is looking
to use DM thinp to host VM guest storage, which implies a scaling
dm-snapshot has _zero_ hope of providing:
> P.S. It could be that my impression is wrong/out-dated, but the
> kernel documentation still says that userspace tools for checking and
> repairing the metadata are "under development". As a file system
> developer, the reaction this inspires is best summed up as:
Already addressed this.