When I did this work originally, I had hoped Spectra would find time to finish memory descriptors and embed those in BIOs instead. The BIO_VLIST thing was an expedient "hack".

Please fix the extra whitespace and test both with and without a roundup2 defined prior to including the header before landing.

I'm just not sure about the new dependency on sys/param.h. I leave that for you to decide.

Should we inline roundup2? Prior to this change, at least in user space, bitstring.h only relied on functionality provided in the C standard. It might be nice to keep that.

If I want to dynamically allocate a bitstr, I use bitstr_size() to tell me how much to allocate. Without the roundup, I'll potentially allocate too little memory.

I see how this can over allocate on platforms with long > 8bits, but I'm missing how 32bit long platforms are special and can get too little space allocated. Can you provide an example in the checkin comment?

Just a few minor nits. Otherwise LGTM.

In the description for this change, under the subr_unit.c heading, I think you mean bit_count, not bit_alloc.

check_unhdr() still must validate the count in the header. The current brute force "bitset_count" in check_unhdr() works, but bitset itself could do better (e.g. word at a time accumulation using mask, plus, and shift).

Given the small size of the bitmap, and the improved performance of the bitset_ffc/bitset_ffs calls you'd need, i think that would be fine for the non-diagnostic code. For the diganostic code (which isn't performance critical), you'd need to add a bitset_count() API. Doing a proper count of the bitmap seems a more robust check than relying on the locally maintained count anyway (no chance that two bugs in the subr_unit code cancel each other out and cause the diagnostic to fail to catch a bug).

Is this the same as D1408? I didn't push that revision because I wanted to update subr_unit.c to avoid the janky union stuff. The accessors should just use the last 8 bits in the bitmap for the count and assemble/decompose the value via mask and shift.

Improve man pages. Fix compilation errors in blkback.c.

Fix spelling errors in comments.

Improve man page documentation for taskqueue_drain_all().

Dec -> December.

Bump .Dd in tasqueue.9.

Ignore the second diff. It was pulled from the wrong tree.

Fix typo in man page: .Pr -> .Pp.

Disallowing unaligned buffers from userland would be a huge POLA violation. If we want to go down this route, I would be okay with detecting unaligned buffers upon entry to the kernel (in the syscall handlers or a common top-level API all of these syscalls invoke) and thunking them there. We'd want to track these alignment faults and make them visible via sysctl and tools like ktrace.