11:12 a.m.
On 03/26/2015 12:08 PM, Wei Huang wrote:
On 03/26/2015 10:49 AM, Don Dutile wrote:
> On 03/26/2015 07:03 AM, Ján Tomko wrote:
>> On Thu, Mar 26, 2015 at 12:48:13AM -0400, Wei Huang wrote:
>>> Current libvirt can only handle up to 1024 thread siblings when it
>>> reads Linux sysfs topology/thread_siblings. This isn't enough for
>>> Linux distributions that support a large value. This patch fixes
>>> the problem by using VIR_ALLOC()/VIR_FREE(), instead of using a
>>> fixed-size (1024) local char array. In the meanwhile
>>> SYSFS_THREAD_SIBLINGS_LIST_LENGTH_MAX is increased to 8192 which
>>> should be large enough for a foreseeable future.
>>>
>>> Signed-off-by: Wei Huang <wei(a)redhat.com>
>>> ---
>>> src/nodeinfo.c | 10 +++++++---
>>> 1 file changed, 7 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/src/nodeinfo.c b/src/nodeinfo.c
>>> index 34d27a6..66dc7ef 100644
>>> --- a/src/nodeinfo.c
>>> +++ b/src/nodeinfo.c
>>> @@ -287,7 +287,7 @@ freebsdNodeGetMemoryStats(virNodeMemoryStatsPtr
>>> params,
>>> # define PROCSTAT_PATH "/proc/stat"
>>> # define MEMINFO_PATH "/proc/meminfo"
>>> # define SYSFS_MEMORY_SHARED_PATH "/sys/kernel/mm/ksm"
>>> -# define SYSFS_THREAD_SIBLINGS_LIST_LENGTH_MAX 1024
>>> +# define SYSFS_THREAD_SIBLINGS_LIST_LENGTH_MAX 8192
>>
>> There is thread_siblings_list, which contains a range:
>> 22-23
>> and thread_siblings file has all the bits set:
>> 00c00000
>>
>> For the second one, the 1024-byte buffer should be enough for 16368
>> possible siblings.
>>
> a 4096 siblings file will generate a (cpumask_t -based) output of :
> 00000000,00000000,00000000,00000000,00000000,00000000,00000000,00000000,
> 00000000,00000000,00000000,00000000,00000000,00000000,00000000,00000000,
> 00000000,00000000,00000000,00000000,00000000,00000000,00000000,00000000,
> 00000000,00000000,00000000,00000000,00000000,00000000,00000000,00000000,
> 00000000,00000000,00000000,00000000,00000000,00000000,00000000,00000000,
> 00000000,00000000,00000000,00000000,00000000,00000000,00000000,00000000,
> 00000000,00000000,00000000,00000000,00000000,00000000,00000000,00000000,
> 00000000,00000000,00000000,00000000,00000000,00000000,00000000,00000000,
> 00000000,00000000,00000000,00000000,00000000,00000000,00000000,00000000,
> 00000000,00000000,00000000,00000000,00000000,00000000,00000000,00000000,
> 00000000,00000000,00000000,00000000,00000000,00000000,00000000,00000000,
> 00000000,00000000,00000000,00000000,00000000,00000000,00000000,00000000,
> 00000000,00000000,00000000,00000000,00000000,00000000,00000000,00000000,
> 00000000,00000000,00000000,00000000,00000000,00000000,00000000,00000000,
> 00000000,00000000,00000000,00000000,00000000,00000000,00000000,00000000,
> 00000000,00000000,00000000,00000000,00000000,00000000,00000000,00000080
> 9(characters per 32-bit mask, including the comma)*8(masks/row)*16(rows)
> -1(last entry doesn't have a comma) = 1152
>
> Other releases/arch's avoid this issue by using cpumask_var_t vs
> cpumask_t for siblings
> so it's reflective of actual cpu count a system (not operating system)
> could provide/support.
Don, could ARM kernel use cpumask_var_t as well? Or this will require
lots of change on top of existing code?
Yes. Working on that (kernel) patch now.
It was simple/fast to use cpumask_t b/c historically,
the counts (& kernel NR_CPUS value) were low.
On x86, they were ACPI-driven. On arm64, need ACPI & DT-based solution,
and arm64-acpi looks like it was based more on ia64 then x86, so need
to create/support some new globals on arm64 that cpumask_var_t depend on,
and have to roll DT to do the same.
> cpumask_t objects are NR_CPUS -sized.
> In the not so distant future, though, real systems will have 1024 cpus,
> so might as well accomodate for a couple years after that.
>
So we agree that such fix would be necessary, because: i) it will fail
on cpumask_t based kernel (like Red Hat ARM); ii) eventually we might
need to revisit this issue when a currently working system reaches the
tipping point of CPU count (>1000).
Yes.
>> For the first one, the results depend on the topology - if
the sibling
>> ranges are contiguous, even million CPUs should fit there.
> The _list files(core_siblings_list, thread_siblings_list) have ranges;
> the non _list (core_siblings, thread_siblings) files have mask like above.
>
>> For the worst case, when every other cpu is a sibling, the second file
>> is more space-efficient.
>>
>>
>> I'm OK with using the same limit for both (8k seems sufficiently large),
>> but I would like to know:
>>
>> Which one is the file that failed to parse in your case?
>>
> /sys/devices/system/cpu/cpu*/topology/thread_siblings
>
>> I think both virNodeCountThreadSiblings and virNodeGetSiblingsList could
>> be rewritten to share some code and only look at one of the sysfs files.
>> The question is - which one?
>>
>> Jan
>>
>