Q:

I have a NAS running FreeBSD 10.1 with 3 disks: ada2 is the boot device, ada0 and ada1 are a ZFS mirror.

dmesg shows this for the ZFS mirror:

GEOM: ada0: the primary GPT table is corrupt or invalid.

How can I recover the primary GPT tables for ada0 and ada1?

A:

Please post the output of zpool status. If ZFS uses the whole disks there won't be a GPT table (or any other partition scheme) at all.

A:

yes, ZFS uses the whole disks.

A:

In that case both ada0 and ada1 do not have a partition table.

You create a pool of directly to device ada0 ada1 - this led to the fact that the pool was created using the entire device (ignore partitions!) not the partition on it (to use every device possible! it is necessary for others to organize the loading of the OS!) and, of course, information about GPT was destroyed.

You should have created a zfs pool:
(for striped) zpool create tank ada0p3 ada1p3
(for mirror) zpool create tank mirror ada0p3 ada1p3
If You set the label in the command gpart add ... -l zdisk0 ada0 , so:
(for mirror) zpool create tank mirror /dev/gpt/zdisk0 /dev/gpt/zdisk1

Should be you case:
zpool status -v

AA:

Because disks can vary in exact size, ZFS leaves some space unused at the end of the disk. (I don't know of an easy way to find out how much. Somebody pointed me at the source once, but I can't find that now. I think it would have to be at least a megabyte to allow for disk variance, but that is an estimate.)

The backup copy of the GPT is stored at the very end of the disk. The boot code tries to verify GPT tables, and is likely finding that leftover backup GPT at the end of the disk.

The trick is clearing that backup GPT without damaging the ZFS data. Do not attempt to do that without a full, verified backup of that ZFS mirror. After that, use diskinfo -v ada0 to get the mediasize in sectors. The standard backup GPT is 33 blocks long, so erasing the last 33 blocks on the disk with dd(1) should be enough to avoid the error without interfering with the ZFS data. dd(1) does not have a way to say "the last n blocks", so the seek= option has to be used to seek to (mediasize in blocks - 33).

WARNING: make a full, verified backup of everything on the disk first!
....
Repeat the procedure for ada1. Do not just reuse the same dd(1) command because the two disks might not have identical block counts.

In the future, the easy way is to erase GPT metadata before reusing the disk. That can be done with gpart destroy (see gpart(8)).

https://forums.freebsd.org/threads/gpt-table-corrupt.52102/post-292341

At the very least make sure there's no partition scheme on the disks before adding them. If there is a gpart destroy adaX should clear it.