Note that the change as-is isn't complete: after a slab is allocated, a per-CPU bucket will be populated with some items from the cache, and kasan_mark_item_invalid() will overwrite the state bytes, setting them back to KASAN_UMA_FREED even though they are still conceptually still in the NEVERUSED state.

In D41897#954871, @cperciva wrote:

@markj I don't know enough about the UMA code to be 100% certain that this is correct, but I think it's right -- we only call this from keg_alloc_slab, which is assigning never-before-used pages to a zone, right?

I've been able to trigger an sbwait() assertion failure in the receive path when running KTLS tests, but only when running tests in parallel in QEMU. I believe the problem is the following: KTLS introduces a state for received data wherein it is being decrypted but is not yet queued in the socket buffer. In this case (sb_tlsdcc != 0), we can end up calling sbwait() with CANTRECVMORE set because that sleep/wakeup mechanism is used to signal completion of decryption operations.

Approved.

arm64 doesn't need any changes?

The overall idea seems ok to me for what it's worth, my comments are about cosmetic issues.

Fix the in-kernel build.

In D41344#952461, @dougm wrote:

Are there any more issues to address before I commit this?

So far I mostly just looked at the cheat locks.

Looks good to me. Thanks for working on this.

The bhyve manual page needs to be updated as well.

Address review comments.

Perhaps its own man page is warranted in order to avoid polluting dtrace(1)?

Restore PCI b/s/f tuple parameters.

Restore PCI b/s/f tuple parameters.

The main diff looks ok to me, I just don't understand the alignment change.

Please make sure to bump __FreeBSD_version in sys/param.h in this commit. The layout change would affect some out-of-tree kernel modules, requiring them to be recompiled.

In D41740#951159, @jhb wrote:

Hmm, looking back in this series, I only seem vm_raise_msi added for arm64, but not for x86?

I still have a couple of WIP patches: one to update the bhyve man page a bit, and another to update option parsing and usage strings. There are quite a few further TODOs: gdb support, PCI passthrough support, ACPI support. I've started working on the gdb stub in the background.

Update the src.conf.5 description for WITH_BHYVE.

In D40551#951026, @emaste wrote:

Will you combine all of the move-amd64-files into one commit?

Make vmexit_hyp() a stub. Translation faults are handled in the kernel.

In D39951#950771, @kevans wrote:

@markj are you wanting to commit this, or do you mind if I just grab it when I push the other one + bump PORTREVISION?

Seems ok to me, but:

@melifaro does this change seem correct to you?

Include <sys/time.h> instead of <sys/timespec.h>. This causes intentional namespace pollution that mimics Linux.