Created attachment 261438 [details]
Minimal but changes struct layout with unknown impact

In review D12230, support for GEOM::ident was added to vnode-backed md(4) devices. The message is saying, "This is useful for gmountver(8) regression tests."

At first, the identifier was generated as random, then, in an effort to improve uniqueness, it was changed to MD-DEV<fsid>-INO<fileid>, and merged in this state.

It's worth to notice that the buffer for this value is only 32 bytes long. Unfortunately, on a 64-bit system with ZFS, such identifier immediately gets truncated around the first digit of <fileid>, making its uniqueness much worse than random:

# ls -i /root/disk*
29580 /root/disk1
29828 /root/disk2
# mdconfig /root/disk1
md0
# mdconfig /root/disk2
md1
# geom md list | grep -E 'name|file|ident'
Geom name: md0
   file: /root/disk1
   ident: MD-DEV11968766667392717182-INO2
Geom name: md1
   file: /root/disk2
   ident: MD-DEV11968766667392717182-INO2

[Hmm, I'd like to have a look at those gmountver(8) regression tests, are they public?]

So, I see a number of solutions here.

1) md-ident-fix-minimal-but-changes-struct-layout.patch: Updates the buffer size to fit two UINT64_MAX decimal representations (51 bytes, rounded up to 64 bytes). Seems to work well, but I am not familiar enough with this code to suggest changing the struct layout: 

# geom md list | grep -E 'name|file|ident'
Geom name: md0
   file: /root/disk1
   ident: MD-DEV11968766667392717182-INO29580
Geom name: md1
   file: /root/disk2
   ident: MD-DEV11968766667392717182-INO29828

2) md-ident-fix-simple.patch: A little bit ugly. Via kvprintf's %r conversion specifier, uses base r36 to fit two 64-bit values in 26 bytes and have 5 more bytes left for our namespacing pleasure:

# geom md list | grep -E 'name|file|ident'
Geom name: md0
   file: /root/disk1
   ident: MD-2IXL9OT1ZWGAM-MTO
Geom name: md1
   file: /root/disk2
   ident: MD-2IXL9OT1ZWGAM-N0K

3) md-ident-fix-invasive.patch: A little less ugly, but adds another undocumented conversion specifier (%R) to kvprintf() to skip an extra uppercasing loop.

4) md-ident-fix-fast.patch: An ugly but speed-optimized solution which also has the benefit of touching less code parts than 2) and 3), but serializes the numbers in little-endian order:

# geom md list | grep -E 'name|file|ident'
Geom name: md0
   file: /root/disk1
   ident: MD-MAGWZ1TO9LXI2-OTM
Geom name: md1
   file: /root/disk2
   ident: MD-MAGWZ1TO9LXI2-K0N

I hope this helps.