On 3/13/26 12:34 AM, Vlastimil Babka (SUSE) wrote:
On 3/13/26 07:14, JP Kobryn (Meta) wrote:
On 3/12/26 10:07 PM, Huang, Ying wrote:
"JP Kobryn (Meta)" <[email protected]> writes:
On 3/12/26 6:40 AM, Vlastimil Babka (SUSE) wrote:
How about I change from per-policy hit/miss/foreign triplets to a single
aggregated policy triplet (i.e. just 3 new counters which account for
all policies)? They would follow the same hit/miss/foreign semantics
already proposed (visible in quoted text above). This would still
provide the otherwise missing signal of whether policy-driven
allocations to a node are intentional or fallback.
Note that I am also planning on moving the stats off of the memcg so the
3 new counters will be global per-node in response to similar feedback.
Emm, what's the difference between these newly added counters and the
existing numa_hit/miss/foreign counters?
The existing counters don't account for node masks in the policies that
make use of them. An allocation can land on a node in the mask and still
be considered a miss because it wasn't the preferred node.
That sounds like we could just a new counter e.g. numa_hit_preferred and
adjust definitions accordingly? Or some other variant that fills the gap?
It's an interesting thought. Looking into these existing counters more,
the in-kernel direct node allocations, which don't fall under any
mempolicy, are also included in these stats. One good example might be
include/linux/skbuff.h, where __dev_alloc_pages() calls
alloc_pages_node_noprof(NUMA_NO_NODE, ...) which eventually reaches
zone_statistics() and increments the stats. So if we applied the
hit/miss/foreign semantics in this patch to the existing counters we
would be mixing allocations that are in and out of policy, losing the
accuracy.
The new 3 counters I last proposed (in an effort to reduce the amount of
new counters as much as possible) would isolate mempolicy allocs and be
named to reflect that: numa_mpol_{hit,miss,foreign}.
