[PATCH v1 09/10] capabilities: Expose NUMA interconnects

Michal Prívozník mprivozn at redhat.com
Mon Jun 7 12:31:10 UTC 2021


On 6/4/21 2:50 PM, Martin Kletzander wrote:
> On Mon, May 31, 2021 at 10:36:10AM +0200, Michal Privoznik wrote:
>> Links between NUMA nodes can have different latencies and
>> bandwidths. This info is newly defined in ACPI 6.2 under
>> Heterogeneous Memory Attribute Table (HMAT) table. Linux kernel
>> learned how to report these values under sysfs and thus we can
>> expose them in our capabilities XML. The sysfs interface is
>> documented in kernel's Documentation/admin-guide/mm/numaperf.rst.
>>
>> Long story short, two nodes can be in initiator-target
>> relationship. A node can be initiator if it has a CPU or a device
>> that's capable of initiating memory transfer. Therefore a node
>> that has just memory can only be target. An initiator-target link
>> can then have any combination of {bandwidth, latency} - {access,
>> read, write} attribute (6 in total). However, the standard says
>> access is applicable iff read and write values are the same.
>> Therefore, we really have just four combinations of attributes:
>> bandwidth-read, bandwidth-write, latency-read, latency-write.
>>
>> This is the combination that kernel reports anyway.
>>
>> Then, under /sys/system/devices/node/nodeX/acccess0/initiators we
>> find values for those 4 attributes and also symlinks named
>> "nodeN" which then represent initiators to nodeX. For instance:
>>
>>  /sys/system/node/node1/access0/initiators/node0 -> ../../node0
>>  /sys/system/node/node1/access0/initiators/read_bandwidth
>>  /sys/system/node/node1/access0/initiators/read_latency
>>  /sys/system/node/node1/access0/initiators/write_bandwidth
>>  /sys/system/node/node1/access0/initiators/write_latency
>>
>> This means that node0 is initiator and node1 is target and values
>> of the interconnect can be read.
>>
>> In theory, there can be separate links to memory side caches too
>> (e.g. one link from node X to node Y's main memory, another from
>> node X to node Y's L1 cache, another one to L2 cache and so on).
>> But sysfs does not express this relationship just yet.
>>
>> Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1786309
>> Signed-off-by: Michal Privoznik <mprivozn at redhat.com>
>> ---
>> docs/schemas/capability.rng |   3 +
>> src/conf/capabilities.c     | 181 +++++++++++++++++++++++++++++++++++-
>> src/conf/capabilities.h     |   1 +
>> 3 files changed, 184 insertions(+), 1 deletion(-)
>>
>> diff --git a/src/conf/capabilities.c b/src/conf/capabilities.c
>> index 7b60676070..8f2d7b75d7 100644
>> --- a/src/conf/capabilities.c
>> +++ b/src/conf/capabilities.c
>> @@ -1735,6 +1743,174 @@
>> virCapabilitiesHostNUMAInitFake(virCapsHostNUMA *caps)
>> }
>>
>>
>> +static void
>> +virCapabilitiesHostInsertHMAT(GArray *interconnects,
>> +                              int initiator,
>> +                              int target,
>> +                              unsigned int read_bandwidth,
>> +                              unsigned int write_bandwidth,
>> +                              unsigned int read_latency,
>> +                              unsigned int write_latency)
>> +{
>> +    virNumaInterconnect ni;
>> +
>> +    ni = (virNumaInterconnect) { VIR_NUMA_INTERCONNECT_TYPE_BANDWIDTH,
>> +        initiator, target, 0, VIR_MEMORY_LATENCY_READ, read_bandwidth};
>> +    g_array_append_val(interconnects, ni);
>> +
>> +    ni = (virNumaInterconnect) { VIR_NUMA_INTERCONNECT_TYPE_BANDWIDTH,
>> +        initiator, target, 0, VIR_MEMORY_LATENCY_WRITE,
>> write_bandwidth};
>> +    g_array_append_val(interconnects, ni);
>> +
>> +    ni = (virNumaInterconnect) { VIR_NUMA_INTERCONNECT_TYPE_LATENCY,
>> +        initiator, target, 0, VIR_MEMORY_LATENCY_READ, read_latency};
>> +    g_array_append_val(interconnects, ni);
>> +
>> +    ni = (virNumaInterconnect) { VIR_NUMA_INTERCONNECT_TYPE_LATENCY,
>> +        initiator, target, 0, VIR_MEMORY_LATENCY_WRITE, write_latency};
>> +    g_array_append_val(interconnects, ni);
>> +}
>> +
>> +
>> +static int
>> +virCapabilitiesHostNUMAInitInterconnectsNode(GArray *interconnects,
>> +                                             int node)
>> +{
>> +    g_autofree char *path = NULL;
>> +    g_autofree char *initPath = NULL;
>> +    g_autoptr(DIR) dir = NULL;
>> +    int direrr = 0;
>> +    struct dirent *entry;
>> +    unsigned int read_bandwidth;
>> +    unsigned int write_bandwidth;
>> +    unsigned int read_latency;
>> +    unsigned int write_latency;
>> +
>> +    path = g_strdup_printf(SYSFS_SYSTEM_PATH "/node/node%d/access0",
>> node);
>> +
> 
> How come you are not checking the relationships for the other access
> classes?  Missing or forgotten code change?  Or do I misunderstand the
> documentation?

That's mostly because I don't quite understand what different access
classes are. I mean, my initial hunch was that it's for different
caches, as in "access0" would be for generic memory, "access1" would be
for L1 cache and so on. But that did not correspond with my findings
when introspecting sysfs with the XML config from 10/10. I mean,
"access1" had the same values as "access0".

And I don't understand it any better after reading kernel docs either.

> 
>> +    if (!virFileExists(path))
>> +        return 0;
>> +
>> +    if (virCapabilitiesGetNodeCacheReadFile(path, "initiators",
>> +                                            "read_bandwidth",
>> +                                            &read_bandwidth) < 0)
>> +        return -1;
>> +    if (virCapabilitiesGetNodeCacheReadFile(path, "initiators",
>> +                                            "write_bandwidth",
>> +                                            &write_bandwidth) < 0)
>> +        return -1;
>> +
>> +    /* Bandwidths are read in MiB but stored in KiB */
>> +    read_bandwidth <<= 10;
>> +    write_bandwidth <<= 10;
>> +
>> +    if (virCapabilitiesGetNodeCacheReadFile(path, "initiators",
>> +                                            "read_latency",
>> +                                            &read_latency) < 0)
>> +        return -1;
>> +    if (virCapabilitiesGetNodeCacheReadFile(path, "initiators",
>> +                                            "write_latency",
>> +                                            &write_latency) < 0)
>> +        return -1;
>> +
>> +    initPath = g_strdup_printf("%s/initiators", path);
>> +
>> +    if (virDirOpen(&dir, initPath) < 0)
>> +        return -1;
>> +
>> +    while ((direrr = virDirRead(dir, &entry, path)) > 0) {
>> +        const char *dname = STRSKIP(entry->d_name, "node");
>> +        int initNode;
>> +
>> +        if (!dname)
>> +            continue;
>> +
>> +        if (virStrToLong_i(dname, NULL, 10, &initNode) < 0) {
> 
> I do not see the value having an "unused" state, e.g. with `-1`, so I
> think it should be unsigned.  Similarly below [1]

Fair enough. I though I saw "node%d" used in kernel so I went with
signed int. But it doesn't really matter - nobody has a machine with 2
billions of NUMA nodes :-)

Michal




More information about the libvir-list mailing list