[Patch] Robustly search for ZFS labels & uberblocks

[Zachary Bedell]

Greetings,

In working with ZFSonLinux, I found myself with a number of somewhat inconstant 
pools which failed either when running grub-probe or at boot time.  In all 
cases, these pools were importable by the ZFS driver and completed a `zpool 
scrub` with no errors reported, but still Grub wasn't able to touch the pools.  
In most (but not all) cases, running `zpool scrub` made the pool accessible to 
Grub again, though obviously that was less than helpful for boot-time failures.

I'm including an (unfortunately rather large) patch to Grub's ZFS code which 
fixes several edge cases where Grub is unable to read a pool which is otherwise 
valid as far as the full ZFS driver is concerned.  Changes include:

ashift:

* Support non-default values of ashift pool attribute.  When ashift is 
increased beyond 10, the size of the uberblocks changes.  Scan to find 
uberblocks must account for this and skip over the extra padding.  Based on 
patch to Grub 0.97 by Hans Rosenfeld at http://www.illumos.org/issues/1303


Label parsing changes:

* Access the two end-device labels at label-size aligned location rather than 
device_end-(2*label_size).  On-disk spec document incorrectly describes 
end-label locations.  Adapted from Grub 0.97 behavior.

* Scan all readable labels for uberblocks and accept the one with the highest 
txg/create date.  Previously UB scan would stop if a valid UB was found in 
Label 0 without checking for newer UB's in the other labels.  All labels must 
be scanned as a more recent block may be found in the other labels in certain 
circumstances.  This fixes a case where Grub would be unable to access a ZFS 
system unless the pool was scrubbed first (but ZFS itself saw no problems & 
scrub reported zero errors).

* Verify that uberblock txg is greater than that of the label before accepting 
the UB so that partially written UB's are not considered.

* Verify that underlying data pointed by uberblock ub_rootbp has valid 
checksums before accepting the UB.  This gives some possibility of booting from 
a technically invalid pool and effectively falls back to older uberblocks in a 
case where the correct uberblock is damaged.

* Store pool uuid found during zfs_mount to grub_zfs_data in order to prevent 
duplicated logic in zfs_uuid.


Data integrity:

* Verify checksum in grub_read_data before accepting block.  Attempt to read 
backup DVA's if checksum on first fails.  Previously backup DVA's were checked 
only if the physical read failed, not if bad data was read without error.  This 
resulted in pools which were valid and readable by ZFS (and transparently 
healed by ZFS' read behavior) being rejected by Grub.


Logging/Debugging:

* Provide more debugging output describing inconsistencies found in pool.

* Remove superflous debugging output to reduce bootup time in verbose mode to a 
mangeable level (~30min down to ~5min to boot w/ 'debug=all' in grub.cfg)



I do apologies for the monolithic patch.  I took a second look at the changes 
to try to pull certain elements apart, but I ended up in several situations 
where the test pools I have snapshotted in VM's each hit two or more of the 
issues above making testing in isolation difficult.  The total change set is 
able to probe and boot from all of the odd (as well as all of the normal) pools 
that I have on hand.  

If remixing this patch in some way would help in integrating it, I'd welcome 
any feedback on how to better present the changes.

Best regards,
Zac Bedell

grub-zfs-ashift-label.patch
Description: Binary data