Hello,
After some of the more urgent regressions XenServer testing has found
have been fixed, I finally have time to get back to this one.
CA-113568 - Dumping interesting state:
local_irq_count was 8
num_hpets_used = 8
HPET[00] idx 0, cpu 3, flags 0x00000001, cpumask 00000008
MSI: HPET 25 vec=41 fixed edge assert phys cpu dest=00000002
mask=1/0/?
HPET[01] idx 1, cpu 4294967295, flags 0x00000000, cpumask 00000000
MSI: HPET 26 vec=c0 fixed edge assert phys cpu dest=00000002
mask=1/0/?
HPET[02] idx 2, cpu 4294967295, flags 0x00000000, cpumask 00000000
MSI: HPET 27 vec=c8 fixed edge assert phys cpu dest=00000002
mask=1/0/?
HPET[03] idx 3, cpu 4294967295, flags 0x00000000, cpumask 00000000
MSI: HPET 28 vec=d0 fixed edge assert phys cpu dest=00000006
mask=1/0/?
HPET[04] idx 4, cpu 4294967295, flags 0x00000000, cpumask 00000000
MSI: HPET 29 vec=21 fixed edge assert phys cpu dest=00000006
mask=1/0/?
HPET[05] idx 5, cpu 2, flags 0x00000001, cpumask 00000004
MSI: HPET 30 vec=29 fixed edge assert phys cpu dest=00000006
mask=1/0/?
HPET[06] idx 6, cpu 4294967295, flags 0x00000000, cpumask 00000000
MSI: HPET 31 vec=31 fixed edge assert phys cpu dest=00000002
mask=1/0/?
HPET[07] idx 7, cpu 1, flags 0x00000001, cpumask 00000002
MSI: HPET 32 vec=39 fixed edge assert phys cpu dest=00000006
mask=1/0/?
This debugging is taken after nmi_shootdown_cpus(), when all other pcpus
are stopped (In due course, I shall get around to submitting the
debugging infrastructure patches, as I suspect others upstream might
find them useful). This is one of the more interesting traces, but
other times have been seen every HPET with cpu set to -1.
local_irq_count proves that we have taken 8 nested interrupts, and the
full stack debugging (not included here for brevity) shows that they
were all HPET interrupts with different vectors.
On this particular hardware, using "maxcpus=4" does work around the
issue, but is not a valid fix. Playing with the position of
ack_APIC_irq() does appear to affect the problem (as suspected), but I
am not convinced it is the correct fix either.
Having looked at the implementation, which uses regular irqs and irq
migration, I have to admit to being surprised it even works in the
slightest. For safety reasons, the irq migration code requires a irq to
arrive at the old cpu, at which point state gets updates to point it at
the new cpu. It also means that the vector is essentially random
between 0x21 and 0xef, which plays havoc with priorities of other
interrupts, including line level interrupts (where a high priority line
level interrupt with its ack sitting on the pending EOI stack can block
a a lower priority HPET timer interrupt for an indefinite period of
time). From my reading of the code, a pcpu which calls
hpet_get_channel() and needs to share channels will cause an early HPET
interrupt to occur on the old pcpu, which appears to then go out of its
way to wake up the cpu which should have received the interrupt.
It occurs to me that a substantially more sane method of doing this is
to install a high priority handler with the hpet handler directly wired
in. This way, hpet_get_channel() becomes vastly more simple and just
involves rewriting the MSI to change the destination. It also allows
for correct use of ack_APIC_irq() (i.e. prevent reentrancy), and frees
up some space in the dynamic range.
Jan: as the author of the current code, is there anything I have overlooked?
~Andrew