But please wait for a node from @mav.

Other than the NULL pointer check, looks good to me.
Thank you!

LGTM. Thanks!

a better solution to sbintime_t vs hrtime_t

no need to get a root vnode of the mp to check if it is rootvnode

use _NOW suffix instead of more cryptic _IM

LGTM

In D12330#255757, @cem wrote:

It seems that ZFS (which we don't use) has some ABI dependency on these constants via zfs_cmd, and this change breaks a CTASSERT on line 202 of zfs_ioctl.c. For some reason passing 3 MAXPATHLENs and 1 MAXNAMELEN via ioctl is too much. Someone would need to figure out how to fix ZFS to use this patch.

@marcel I never intended for "persistent" to mean "can never change". I only meant it to mean that, using Andrey's example, opening md0p1 for writing can not result in its label changing.
However, I did miss the fact that gpart control interface operates by explicitly spoiling consumers attached to modified partitions. I thought that the partitions themselves were spoiled and the consumers became orphaned.
So, this was a hole in my original design that Andrey pointed out.
But I think that it does not invalidate the general idea.
It's unfortunate that both the opening of md0p1 for writing and the re-labeling of md0p1 use the same mechanism, the spoiling, to convey the actions to consumers.

@ae initially I was a little bit confused about your example, but now I see what you mean.
It seems that "GEOM::media" attribute change is posted for consumers that do not get spoiled.
Maybe the g_label code can make use of that fact.
Here is a very rough change that seems to do the job:

static void
g_label_attrchanged(struct g_consumer *cp, const char *attr)
{
        struct g_class *mp;
        struct g_provider *pp;

LGTM.

In D12330#255757, @cem wrote:

It seems that ZFS (which we don't use) has some ABI dependency on these constants via zfs_cmd, and this change breaks a CTASSERT on line 202 of zfs_ioctl.c. For some reason passing 3 MAXPATHLENs and 1 MAXNAMELEN via ioctl is too much. Someone would need to figure out how to fix ZFS to use this patch.

immediately bump sync-id on a write failure

clear mnt_vnodecovered before calling vfs_mount_destroy in case of the mount failure,
that's safe to do because the new struct mount is not visible to the outside world

move hardware probing to the device probe method

LGTM

The reason I originally implemented the feature only for Bank 4 is that in the public BKDG for family 10h MC0_MISC to MC3_MISC are specified as read-only all zero-s. And MC5_MISC is documented to have an unspecified value. So, I thought why bother...
For example:

MSR0000_0403 DC Machine Check Miscellaneous Register (MC0_MISC)
This register is read-only, reset: 0000 0000 0000 0000h.

LGTM