Re: [PATCH] [RFQ] dm-integrity: Add a lazy commit mode for journal

[Mikulas Patocka <mpatocka@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 3871 bytes --]



On Fri, 23 Feb 2024, Weiß, Simone wrote:

> On Tue, 2024-02-20 at 19:52 +0100, Mikulas Patocka wrote:
> > CAUTION: This email originated from outside of the Elektrobit organization. Do
> > not click links or open attachments unless you recognize the sender and know
> > the content is safe.
> > 
> > 
> > On Fri, 9 Feb 2024, Simone Weiß wrote:
> > 
> > > Extend the dm-integrity driver to omit writing unused journal data sectors.
> > > Instead of filling up the whole journal section, mark the last used
> > > sector with a special commit ID. The commit ID still uses the same base
> > > value,
> > > but section number and sector number are inverted. At replay when commit IDs
> > > are analyzed this special commit ID is detected as end of valid data for
> > > this
> > > section. The main goal is to prolong the live times of e.g. eMMCs by
> > > avoiding
> > > to write the whole journal data sectors.
> > > 
> > > The change is right now to be seen as experimental and gets applied if
> > > CONFIG_DMINT_LAZY_COMMIT is set to y. Note please that this is NOT
> > > planned for a final version of the changes. I would make it configurable
> > > via flags passed e.g. via dmsetup and stored in the superblock.
> > > 
> > > Architectural Limitations:
> > > - A dm-integrity partition, that was previously used with lazy commit,
> > >  can't be replayed with a dm-integrity driver not using lazy commit.
> > > - A dm-integrity driver that uses lazy commit is expected
> > >  to be able to cope with a partition that was created and used without
> > >  lazy commit.
> > > - With dm-integrity lazy commit, a partially written journal (e.g. due to a
> > >  power cut) can cause a tag mismatch during replay if the journal entry
> > > marking
> > >  the end of the journal section is missing. Due to lazy commit, older
> > > journal
> > >  entries are not erased and might be processed if they have the same commit
> > > ID
> > >  as adjacent newer journal entries.
> > 
> > Hi
> > 
> > I was thinking about it and I think that this problem is a showstopper.
> > 
> > Suppose that a journal section contains these commit IDs:
> > 
> >         2       2       2       2(EOF)  3       3       3       3
> > 
> > The IDs "3" are left over from previous iterations. The IDs "2" contain
> > the current data. And now, the journal rolls over and we attempt to write
> > all 8 pages with the ID "3". However, a power failure happens and we only
> > write 4 pages with the ID "3". So, the journal will look like:
> > 
> >         3(new)  3(new)  3(new)  3(new)  3(old)  3(old)  3(old)  3(old)
> > 
> > After a reboot, the journal-replay logic will falsely believe that the
> > whole journal section is consistent and it will attempt to replay it.
> > 
> > This could be fixed by having always increasing commit IDs - the commit
> > IDs have 8 bytes, so we can assume that they never roll-over and it would
> > prevent us from mixing old IDs into the current transaction.
> Hi
> 
> Thanks for the review of the concept. I was out this week and could only think
> about it now. I understood it right, that the proposal is to add an extra value
> to the commit ID, that is e.g. incremented when integrity_commit is executed?
> 
> If so, I tried this quickly and looks good on first glance. Will check and test
> further next.
> 
> Simone

I propose to use the commit ID 0 when writing the journal for the first 
time, then 1 when the journal rolls over, 2 when it rolls over again, 3 
when it rolls over again, 4 on another roll over and so on up to 
0x7fffffffffffffff (which will be never reached in practice).

And use the top bit as an end-of-section marker. As the commit IDs will 
never roll over, it won't happen that an old transaction would be mixed 
into a new transaction on partial journal write.

Mikulas