On 06/09/11 17:02, Kenneth R Westerback wrote:
On Thu, Jun 09, 2011 at 12:18:06AM +0200, Mark Kettenis wrote:
The current amd64 code allows for a fairly limited number of device
interrupts on the primary CPU. The exact number is a bit fuzzy, but
is somewhere between 11 and 27. If you go beyond this limit, the
interrupts will be sent to a secondary CPU if you have one and are
running GENERIC.MP. Otherwise any additional devices won't get any
interrupts at all, and typically won't function. And having USB or
block device interrupts on a secondary CPU causes problems as well.
On some big servers and many modern laprops we're almost out. There
may even be machines that don't boot because they already need more
interrupts than we can handle on the primary CPU. And MSI is
increasing the pressure.
The diff below increases the number of interrupts that can be handled
on a CPU to 64. I've sucessfully baked a muild and built xenocara on
my amd64 box running GENERIC. Further testing, especially on
GENERIC.MP is needed.
Thanks,
Mark
I have installed a locally build snapshot containing the diff on 3
bsd.mp amd64 boxes so far. I am attempting to put it on a fourth.
Of the four systems three have re(4) interfaces and all three emitted
'reN watchdog timeout' messages which I can't recall them doing
before. All the re are on msi interrupts. The only box w/o re(4)
(it had alc(4)) had no issues, and it was also on msi due to a local
diff.
The fourth box I'm attempting to put this on seems to have terminal
re(4) problems. It stalls in base49.tgz or comp49.tgz. It is an hp
g56 laptop.
I can see those re watchdog timeouts without that patch. I can't tell
when exactly they appeared, last time I rebuild the system was on May
14th, now I am running with source from yesterday.
HTH
Markus
