Sorry to nitpick a bit more. Please let me know if you'd like me to commit the change.

I don't really like this interface. The name is too general and the interface is kind of simplistic, for example because it can only describe a single contiguous memory region and can't refer to different instances of a device. I'm not what a better interface would look like, I haven't yet read more of the patch series. But I think this deserves some more thought.

Looks ok. I'd explicitly initialize MXPS and NOPS to zero as well.

I believe this is superseded by commit 1d9eae9fb2e2253ca3d3764a5cc7f124b10e358b .

Could you please upload a diff with context?

Make do_config() take a size_t instead.

In D33133#748895, @kp wrote:

Wouldn't it make more sense to change l to be a size_t? sopt->sopt_valsize is a size_t, and the arguments to sootcopyin() are also size_t.
It's only do_config() which takes an int, and that may want to be changed to take a size_t as well.

size_t -> unsigned long

a -> an
Use .Vt for referring to an Elf(32|64)_Chdr

In D33120#748480, @kib wrote:

This is definitely an improvement because before, global ifunc symbols were not resolved at all.

But ultimately, it does not matter if the symbol is global or not, ifuncs generally should be processed after other types of relocations, global or local.

This looks good!

Attempt to create the alias iff the device unit number is 0.

In D33068#747710, @imp wrote:

In D33068#747699, @markj wrote:

In D33068#747080, @kib wrote:

IMO if we pretend that the driver is useful, it should create units as mentioned in summary. Or disable attach of any but first unit.

What is the canonical way to disable attach of all but the first unit? Is there some newbus mechanism for this?

I think 'if (device_get_unit(dev) != 0) return ENXIO;' in the attach routine. That produces an ugly error though,
so often times I've seen in private code 'return 0' and have no resources created for the device, perhaps apart
from activating the resources automatically assigned to the device.

Go a bit further and create /dev/agpgart%d for all devices, and create the alias
for the device with unit number 0.

• Fix ip_output()
• Move the mb_unmapped_to_ext() call out of ip6_output_delayed_csum()

P.S., I spent some time trying to replace this implementation with an updated one from NetBSD. Theirs is easier to read and is licensed under BSD 2-clause. It got somewhat messy, though, and wasn't important to my immediate goal, so I dropped it. I think there's also some room for micro-optimizations here, for instance in_pseudo() and in_addword() should probably be inlines. in_pseudo() in particular is dead simple, it's just a ones complement sum:

In D33068#747080, @kib wrote:

IMO if we pretend that the driver is useful, it should create units as mentioned in summary. Or disable attach of any but first unit.

Is it sufficient to instead set be_state = IDENT_WAIT in fwctl_inl(), or handle be_state == REQ the same as be_state == IDENT_WAIT in fwctl_outw()?

In D28278#747302, @c.koehne_beckhoff.com wrote:

In D28278#747021, @markj wrote:

I think D27220 was intended to solve the same problem? I'm not sure if a bhyve patch to use libuvmem is available.

This seems like a reasonable interim solution.

I'm unfamiliar with libuvmem. I'm unsure whether libuvmem is able to solve the problem too or not.