On 1/17/2023 4:44 PM, David Hildenbrand wrote:
On 16.01.23 03:25, Raghavendra K T wrote:
During the Numa scanning make sure only relevant vmas of the
tasks are scanned.
Logic:
1) For the first two time allow unconditional scanning of vmas
2) Store recent 4 unique tasks (last 8bits of PIDs) accessed the vma.
False negetives in case of collison should be fine here.
3) If more than 4 pids exist assume task indeed accessed vma to
to avoid false negetives
Co-developed-by: Bharata B Rao <bharata@xxxxxxx>
(initial patch to store pid information)
Suggested-by: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>
Signed-off-by: Bharata B Rao <bharata@xxxxxxx>
Signed-off-by: Raghavendra K T <raghavendra.kt@xxxxxxx>
---
include/linux/mm_types.h | 2 ++
kernel/sched/fair.c | 32 ++++++++++++++++++++++++++++++++
mm/memory.c | 21 +++++++++++++++++++++
3 files changed, 55 insertions(+)
diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 500e536796ca..07feae37b8e6 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -506,6 +506,8 @@ struct vm_area_struct {
struct mempolicy *vm_policy; /* NUMA policy for the VMA */
#endif
struct vm_userfaultfd_ctx vm_userfaultfd_ctx;
+ unsigned int accessing_pids;
+ int next_pid_slot;
} __randomize_layout;
What immediately jumps at me is the unconditional grow of a VMA by 8
bytes. A process with 64k mappings consumes 512 KiB more of memory,
possibly completely unnecessarily.
This at least needs to be fenced by CONFIG_NUMA_BALANCING.
Thanks for the review David. Good point.. I do agree. I see I will have
to fence further in memory.c only since fair.c is already taken care.
