On Tue, Sep 1, 2020 at 2:27 AM Germano Veit Michel <germano(a)redhat.com> wrote:
On Mon, Aug 31, 2020 at 5:00 PM Nir Soffer <nsoffer(a)redhat.com> wrote:
>
> On Mon, Aug 31, 2020 at 4:48 AM Germano Veit Michel <germano(a)redhat.com>
wrote:
> >
> >
> >
> > On Sun, Aug 30, 2020 at 8:46 PM Nir Soffer <nsoffer(a)redhat.com> wrote:
> >>
> >>
> >>
> >> On Fri, Aug 28, 2020, 08:36 Germano Veit Michel <germano(a)redhat.com>
wrote:
> >>>
> >>>
> >>>
> >>> On Fri, Aug 28, 2020 at 9:29 AM Nir Soffer <nsoffer(a)redhat.com>
wrote:
> >>>>
> >>>>
> >>>>
> >>>> On Thu, Aug 27, 2020, 16:38 Tal Nisan <tnisan(a)redhat.com>
wrote:
> >>>>>
> >>>>>
> >>>>>
> >>>>> On Fri, Aug 21, 2020 at 4:34 AM Germano Veit Michel
<germano(a)redhat.com> wrote:
> >>>>>>
> >>>>>> Hi,
> >>>>>>
> >>>>>> Is there a reliable way to figure out if a snapshot is in
preview only using information obtained from the storage domain metadata?
> >>>>>> I'm trying to find a way to distinguish a problematic
snapshot chain (double parent) from a snapshot in preview in order to improve dump-volume
chains.
> >>>>>>
> >>>>>> Currently dump-volume-chains throws an error
(DuplicateParentError) if a snapshot is in preview for the image, as there is a
'Y' shape split in the chain
> >>>>>> with 2 volumes (previous chain + preview) pointing to a
common parent:
> >>>>>>
> >>>>>> image: dff0a0c0-b731-4e5b-9f32-d97310ca40de
> >>>>>>
> >>>>>> Error: more than one volume pointing to the
same parent volume e.g: (_BLANK_UUID<-a), (a<-b), (a<-c)
> >>>>>>
> >>>>>> Unordered volumes and children:
> >>>>>>
> >>>>>> - e6c7bec0-53c6-4729-a4a0-a9b3ef2b8c38 <-
5eb2b29d-82d6-4337-8511-3c86705d566e
> >>>>>> status: OK, voltype: LEAF, format: COW,
legality: LEGAL, type: SPARSE, capacity: 1073741824, truesize: 1073741824
> >>>>>>
> >>>>>> - e0475853-4514-4464-99e7-b185cce9b67d <-
deceff83-9d88-4f87-8304-d5bf74d119b1
> >>>>>> status: OK, voltype: LEAF, format: COW,
legality: LEGAL, type: SPARSE, capacity: 1073741824, truesize: 1073741824
> >>>>>>
> >>>>>> - e6c7bec0-53c6-4729-a4a0-a9b3ef2b8c38 <-
e0475853-4514-4464-99e7-b185cce9b67d
> >>>>>> status: OK, voltype: INTERNAL, format: COW,
legality: LEGAL, type: SPARSE, capacity: 1073741824, truesize: 1073741824
> >>>>>>
> >>>>>> - 00000000-0000-0000-0000-000000000000 <-
e6c7bec0-53c6-4729-a4a0-a9b3ef2b8c38
> >>>>>> status: OK, voltype: INTERNAL, format: RAW,
legality: LEGAL, type: PREALLOCATED, capacity: 1073741824, truesize: 1073741824
> >>>>>>
> >>>>>> From the engine side it's easy, but I'd need to
solve this problem using only metadata from the storage.
> >>>>>>
> >>>>>> The only thing I could think of is that one of the volumes
pointing to the common parent has voltype LEAF. Any better ideas?
> >>>>>
> >>>>> don't think that there is any, Engine is the orchestrator
and due to that the info is only in the database
> >>>>
> >>>>
> >>>> There is no good way, but you can look at the length of the chain,
and the "ctime" value.
> >>>>
> >>>> For example if this was the original chain:
> >>>>
> >>>> a <- b <- c
> >>>>
> >>>> if we preview a:
> >>>>
> >>>> a <- b <- c
> >>>> a <- d
> >>>>
> >>>> You know that d is a preview volume.
> >>>>
> >>>> If we preview b, we will have two chains of same length:
> >>>>
> >>>> a <- b <- c
> >>>> a <- b <- d
> >>>>
> >>>> But the ctime value of d will be larger, since preview is created
after
> >>>> the leaf was created.
> >>>>
> >>>> ctime is using time.time() so it is not affected by time zone
changes
> >>>> but it may be wrong due to host time changes, so it is not
reliable.
> >>>>
> >>>> Can you open a bug for this?
> >>>
> >>>
https://bugzilla.redhat.com/show_bug.cgi?id=1873382
> >>>
> >>> I have a prototype working with some code I pasted in the bugzilla, but
I don't think it's reliable and an overcomplication of what should be simple.
> >>
> >>
> >> I don't think the code in the bug is the way to handle this.
> >>
> >> It will be simpler and more useful to:
> >> 1. Find leaves
> >> 2. Follow the chain from each leaf, until the base (volume with no parent).
> >> 3. Display a tree instead of list, like lsblk.
> >>
> >> For example:
> >>
> >> dbf1e90c-41d5-4c2d-a8d2-15f2f04d3561
> >> ├─ea6af566-922c-4ca2-af17-67f7cd08826c
> >> └─aa5643ef-8c74-4b28-91e0-8d45d6ee426b
> >> └─30c4f6d1-7f1d-470b-96ae-7594cf367dfa
> >
> > I like the idea of this visual representation, but it does not fix the problem.
> >
> > The problem is dump-volume-chains throwing incorrect errors in case there is a
snapshot in preview.
> >
> > Error: more than one volume pointing to the same parent volume e.g:
(_BLANK_UUID<-a), (a<-b), (a<-c)
>
> This error is wrong, you should remove it, and instead show the tree.
>
> > There is still a double parent on the representation above. So if the analysis
is done (text output), there will
> > be an error detected no matter how we print it. If there is no way to
distinguish a preview from a double parent
> > problem without leaving any doubt based only on storage metadata only then we
can improve the
> > representation but ultimately the problem remains there.
> >
> > Ideally I'd like to keep DoubleParentError logic and detect Previews to
eliminate the false errors.
>
> This is not possible now.
>
> > The analysis should be done in the image discrepancy tool on the engine, which
has dump-volume-chains
> > output (json - no analysis) and the engine db. And we are already doing some
basic checks there. Maybe
> > we should even move the entire analysis logic there and make dump-volume-chains
just print and dump
> > data without doing analysis if the analysis cannot be done based on partial
data.
> >
> > The main idea here was to simply stop false errors for those who look for them
in dump-volume-chains
> > text output.
> >
> >>
> >>
> >> Users of the tool will have to check engine db to understand how to fix the
disk.
> >>
> >> Even if it was easy to detect a volume in preview, how do you know which
chain
> >> should be kept? Did it fail just after the user asked to commit the
preview?
> >
> >
> > This tool is not used to diagnose and correct issues on its own. It is used for
2 things, but mainly the first:
> > a) Nice readable way to see volumes and their metadata, plus chain
> > b) Any obvious errors
> >
> > The duplicate parent is printing false problems during preview, breaking the
tool for B.
> >
> > The main use is still A, use of dump-volume-chains is to stop collecting
/dev/VG/metadata LV or *.meta files
> > and have this info for the volumes in the sosreport.
> >
> > I'm not aware of anyone using just the output of the tool to perform chain
changes, every failure
> > also requires checking the DB too and most importantly the logs (unless
rotated).
> >
> >>
> >> Storage format does not have a way to store info about the state of the
disk, or make atomic
> >> changes like remove one chain when committing after a preview. This is also
the reason we
> >> have trouble with removing snapshots.
> >
> >
> > Which means we cannot know for sure what is happening in the chain, right?
> > With this in mind, any suggestion to stop the false errors?
>
> Change the code to handle a tree instead of a list of volumes, error is gone.
But then part of the validation is gone too. We open the possibility of validating trees,
which are all invalid
except for the very specific case of a preview, which we have no data to determine for
sure anyway.
But the validation is incorrect. Trees are actually supported using preview, so
failing and not showing the tree in dump-volume-chain is a bug.
There is not much that can be done as there is no reliable way to
determine if the chain has a snapshot
in preview without several changes on engine and vdsm. And it's not worth
implementing this, I'll close
the bug too.
Thanks for your help!
>
> >
> > Since we cannot be sure of this based just on SD metadata, maybe the simplest is
to remove the
> > duplicate parent error string and/or add some warning that it could be a
snapshot in preview and just
> > print the unordered volumes.
> >
> > The improved visual representation could be handled separate from this. I've
thought of something
> > similar in the past but found hard to print the volume metadata in a nice way
(and we need to handle
> > big chains of several dozen snapshots).
> >
> > Thanks,
> > Germano
> >
> >>
> >>
> >> Nir
> >>
> >>>
> >>>
>