On 12/04/2013 01:47 AM, Eduardo Habkost wrote:
> On Tue, Dec 03, 2013 at 02:30:48PM +0100, Andreas Färber wrote:
>> Am 03.12.2013 00:03, schrieb Alexey Kardashevskiy:
>>> On 12/03/2013 09:09 AM, Andreas Färber wrote:
>>>> Am 02.12.2013 18:06, schrieb Michael Tokarev:
>>>>> 25.11.2013 07:39, Alexey Kardashevskiy wrote:
>>>>>> Since modern POWER7/POWER8 chips can have more that 256 CPU threads
>>>>>> (>2000 actually), remove this check from smp_parse.
>>>>>>
>>>>>> The CPUs number is still checked against machine->max_cpus and this check
>>>>>> should be enough not to break other archs.
>>>>
>>>> "should be" is not exactly the highest level of confidence for a
>>>> "trivial" patch... :/
>> [...]
>>>> Alexey, did you actually check that, e.g., x86 machines don't break with
>>>> 256 or 257 CPUs now?
>>>
>>> PC_DEFAULT_MACHINE_OPTIONS sets it to 255. And I cannot find any machine
>>> which would not define max_cpus, have I missed any?
>>
>> If you've actually *checked* the other machines' code then fine with me,
>> just say so in the commit message. :)
>
> I just grepped for "max_cpus" and checked every match. The largest
> values I found were:
>
> hw/ppc/spapr.c: 256
> s390: 255
> pc: 255
>
> All the rest had values <= 32.
>
> Machines with missing max_cpus value shouldn't be a problem, as
> max_cpus==0 is interpreted as 1 by the vl.c code.
>
> But we still need to add a check for max_cpus > machine->max_cpus to
> vl.c, before we eliminate the smp_parse() check.
Since smp_parse() checks if (max_cpus >= smp_cpus), this should just work:
diff --git a/vl.c b/vl.c
index e6ed260..544165a 100644
--- a/vl.c
+++ b/vl.c
@@ -3882,9 +3882,9 @@ int main(int argc, char **argv, char **envp)
smp_parse(qemu_opts_find(qemu_find_opts("smp-opts"), NULL));
machine->max_cpus = machine->max_cpus ?: 1; /* Default to UP */
- if (smp_cpus > machine->max_cpus) {
+ if (max_cpus > machine->max_cpus) {
fprintf(stderr, "Number of SMP cpus requested (%d), exceeds max cpus "
- "supported by machine `%s' (%d)\n", smp_cpus, machine->name,
+ "supported by machine `%s' (%d)\n", max_cpus, machine->name,
machine->max_cpus);
exit(1);
}
> There's also this, at main():
>
> if (i == nb_numa_nodes) {
> for (i = 0; i < max_cpus; i++) {
> set_bit(i, node_cpumask[i % nb_numa_nodes]);
> }
> }
>
> node_cpumask[] is initialized using bitmap_new(MAX_CPUMASK_BITS), and
> MAX_CPUMASK_BITS is 255. To fix this, we can initialize node_cpumask[] using
> max_cpus instead, if we initialize it after smp_parse().
Nope. At the moment when we parse -numa in vl.c, we may not know yet what
machine is going to be used and machines can have different max_cpus.
For now, I would simply change MAX_CPUMASK_BITS to something crazy, like
16384 (2KB per numa node), I hope QEMU can survive such a memory waste :)
Ok?
--
Alexey
