Are spaces problematic for devd notifications consumers, typically shell scripts ?

Does it make sense to decode 'reload' flag ?

Both mount and unmount events are invoked while some vnodes are locked. sbuf_new_auto() seems to mean M_WAITOK allocations for the sbuf data by default. Waitable memory allocations while a vnode is locked can cause deadlock if the only way for pagedaemon to make the progress is to flush the locked vnode.

!= 0

I suggest to move all code to calculate sbuf strings from struct mount into kern/vfs_mount.c.
Even if functions like devctl_safe_quite_sb() would need to be exported, or better moved to kern/subr_sbuf.c under more generic name.

Gotcha. I went back and forth. I'll do a NOWAIT malloc for a buffer too big for the stack and just omit the message if that fails (which should be quite rare). That will avoid the deadlock.

Like the whole thing? Or just the stuff that does flags and options?

/dev/devctl uses nowait malloc anyway, AFAIR ?

I would say, move the whole thing. Then it is questionable why do you want to register eventhandlers instead of doing direct calls to the functions.

Yes. It does.

OK. I've moved the whole thing, but need to tweak one or two more things before reposting the review. I'll do that tomorrow.

This needs a smidge more work... I missed it focusing on the kernel.

Hmm, stray now and can be removed.

I was unsure. What do you think?

Not sure this is the right name to use, so would be interested in feedback.

static const

Perhaps devd events stuff could be split in dedicated header, as a followup.

coveredvp is still locked there.

That said, I am not sure that it is reasonable to report mount options for unmount event.

I suggest mount_devd_event()

Won't they communicate how the system was last mounted?

elsewhere we use 'devctl' rather than 'devd' so I'll use mount_devctl_event

this is holding until after we land OpenZFS and won' tbe in this commit.

this won't be in the commit.

Is it useful ? Also the operation of unmounting could mangle options in the process. From very quick view, for instance UFS clears MNT_LOCAL. Why ? I do not know.

buf cannot be null.

Ah, fair point. I'll create a followup.

I'd intended to use M_NOWAIT...

I added to umount, I just realized, to allow us to see if the umount was forced or not, which I know now isn't cleared... should we have a different mechanism to communicate that?

I think that this communication mechanism is fine.