Thanks for bringing these back for me.

Double check that there aren't any other registers needed saving for HTM, too.

@breno.leitao_gmail.com Can you pull out the pcb.h and pcpu.h changes into a separate diff? That way we can get that in before the KBI freeze, and wait until 12.1 to get the actual code in. Once KBI is frozen it's frozen for the full 12.0 life, and any KBI breakage we want to perform will have to wait for 13.0.

Survives a buildworld, something it couldn't do before.

I like this. It cleans things up, and drops some ugly macros.

Approved. A cleaner solution might be to add another SI_SUB_ node, or reuse an existing one for all cases. But this unblocks us, and is sufficient. SI_SUB_INIT_IF looks like a good one (no need to do it now).

Looks good

@jimharris is this good to commit as-is now, or should I look at more changes?

I just created https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=230334 for adding a bus_dmamap_sync_range() KPI.

There is no prp_dma_map. The prp_bus_addr exists in the qpair->queuemem_map, so does get synced. Or am I reading it wrong?

I just reverted this patch to test as well. I ran 2 'make buildworld -j72 -s' and 1 'make buildworld buildkernel -j72 -s'. The performance difference I saw was not as great as I originally wrote, something else had improved performance elsewhere in the last couple weeks. I did, however, see a 5-10% improvement, so it's not entirely insignificant. I also did see "Missing interrupt" messages with the patch, they just did not occur as frequently as previously without.

Bring back the wmb() for !powerpc. powerpc's bus_dmamap_sync already includes
an explicit hwsync, but other archs do not.

Address jimharris's comments.

Hi Jim,

Remove x86 names (rdrand, ivy) from the function names. No other changes.

I should add, from the ISA reference:

In D16351#348390, @jhb wrote:

This looks ok to me, but I'm not a ports committer. Did you use poudriere testport or 'make check-plist' to verify the packing lists?

Thanks for the review, guys. It's quite obvious now that my patch is more WIP than complete and ready, but I think the bulk is still good.

Revert a no-longer-necessary diff against libvpx. There may be more, as I did
the work on one tree and submitted the diff from another tree.

Just one minor nit, but otherwise good to go.

Great work tracking this down!

This approach is wrong. The problem I was trying to solve is the interrupt handler accessing a bad sc. The proper fix will be to instead make the interrupt handler use a sc, instead of a consdev, and move more state data into the sc.

Reduce the diff, by sharing the same VT driver between the early attach and delayed attach. This required moving some config setting into the probe routine, but that should be fine.

I should probably caffeinate before making such trivial changes.

Change to nwhitehorn's preference. Either is fine with me.