[Bug 228768] EARLY_AP_STARTUP causes panic on amd machines

[bugzilla-noreply at freebsd.org]

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=228768

John Baldwin <jhb at FreeBSD.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |kib at FreeBSD.org,
                   |                            |mjg at FreeBSD.org

--- Comment #4 from John Baldwin <jhb at FreeBSD.org> ---
Oh, I think I have seen this once before.  The issue is that the
mtx_lock_spin() code tries to use DELAY(1) and DELAY() is trying to use the
"clock_lock" in getit().  In your case, 'tsc_is_invariant' isn't true so
DELAY() in sys/x86/x86/delay.c isn't using the TSC.

I think the error is probably that we shouldn't be using DELAY(1) in our mutex
code as it is too high level and/or we should special case DELAY(1) (vs
DELAY(n)).  Older versions of FreeBSD would map DELAY(1) to just 'inb(0x84)'
without talking to the 8254 at all.  It looks like i8254_delay() already does
that when inside of KDB.

Also, checking for tsc_is_invariant in DELAY is probably overly-correct.  We
aren't going to change P-states while we are spinning since we are actively
spinning on the CPU.  The worst that might happen is that if the CPU were
running at a lower clock, we might wait too long.  Given how DELAY is used,
that's probably better than falling back to a slow timecounter with locks, etc.

-- 
You are receiving this mail because:
You are the assignee for the bug.