Where is this optimization actually used ? Almost all calls to vm_pager_has_page() pass NULLs for behind/ahead.

shadow_count > 0 implies OBJ_ANON.

Kind of obvious idea is to switch to periodic scans for collapses, and make the scan to walk the whole shadow chain for each map entry that has anon object on top. I.e. instead of trying to do collapses at precise points, we would have yet another daemon.

This should be fine IMO.

I am somewhat curious how much does the upgrade succeed. Did you measured it ? E.g. the counters of total vm_pager_page_unswapped calls during buildworld vs. the number of times the lock needs to be upgraded vs. the number of successfull upgrades.

x86_var.h was missed.

Make execution submode numeric.

Add MSR_OP_RENDEZVOUS. Assert that exactly one command mode flag is passed.

Rename to x86_msr_op().
Restore sched_bind() mode of broadcast to keep the ssb toggle algorithm unchanged. I do think that my switch to smp_rendezvous for ssb is actually fine because our use of mwait does not re-enable interrupts, but let do that as separate change.

In D22434#490451, @mav wrote:

I am not completely certain in what part of the driver it is cleaner to put this line. I though about the beginning, as I understand it should be just before the first bus_dma tag access or interrupt setup. But in case of attach failure I guess it should better be reverted somehow? Since there is no such method I put it when most of probe process is done.

There are two requirements for the call to bus_dma_dmar_set_buswide():

1. the call must occur before the first busdma tag allocation for anything on this bus
2. the call must not occur before you are certain that the device identified has the bus-wide property.

Weaken asserts when pages are unbusy in context of io completion handlers.

Alexander, I obviously cannot test this, I only ensured that machine with DMAR enabled still boots.

One more 'p'.

Only use imgp after it is initialized, and use p later.

You also need to reset the taa state after resume, at least if it was set to disable TSX. See how ibrs and sbb are reset in acpi_sleep_machdep(). Or you need to re-apply TSX disabling command after resume.

That said, would be it more straightforward to add logic to hw_mds_recalculate() instead, to enable the workaround if either of mds_disable or tsx_disable is forced. I mean, get rid of hw_tsx_recalculate(), check TAA in hw_mds_recalculate.

If you want to disable ASLR, it is enough to issue procctl(), checking global system settings gives no value and is in fact racy.

Handle avg' suggestions.

In D22303#487963, @jhb wrote:

Note that gdb also knows which symbols have which behavior, so changing the stack frame layout will break kgdb, and then kgdb would have to somehow detect which i386 kernels had which behavior. We've done that once before by peeking at instructions [1], but I think it's not worth spending that much effort on i386 and easier to just let the frame layouts stay as-is. The reason for the enums was to handle the separate cases as you note for args. I don't mind changing this to use 'narg'.
1: https://github.com/bsdjhb/gdb/blob/kgdb-8.3/gdb/i386fbsd-kern.c#L82
For reference: here is how we currently decode frame offsets in kgdb now:
https://github.com/bsdjhb/gdb/blob/kgdb-8.3/gdb/i386fbsd-kern.c#L337

Ok, I think I am convinced enough. As result, I do not like this code even more.

The rest of the change, now I limit MFENCE only to non-Intel vendors.