@@ -52,10 +52,51 @@ static u64 vmevent_attr_swap_pages(struct vmevent_watch *watch,
#endif
}
+/*
+ * In the worst case, this is inaccurate by
+ *
+ * ±(pcp->stat_threshold * zones * online_cpus)
+ *
+ * For say 4-core 2GB setup that would be ~350 KB worst case inaccuracy,
+ * but to reach this inaccuracy, CPUs would all need have to keep
+ * allocating (or freeing) pages from all the zones at the same time, and
+ * all their current vm_stat_diff values would need to be pretty close to
+ * pcp->stat_threshold.
+ *
+ * The larger the system, the more inaccurare vm_stat is (but at the same
+ * time, on large systems we care much less about small chunks of memory).
+ * When a more predicted behaviour is needed, userland can set a desired
+ * accuracy via attr->value2.
+ */
+static ulong vmevent_global_page_state(struct vmevent_attr *attr,
+ enum zone_stat_item si)
+{
+ ulong global = global_page_state(si);
+#ifdef CONFIG_SMP
+ struct zone *zone;
+
+ if (!attr->value2)
+ return global;
+
+ for_each_populated_zone(zone) {
+ uint cpu;
+
+ for_each_online_cpu(cpu) {
+ struct per_cpu_pageset *pcp;
+
+ pcp = per_cpu_ptr(zone->pageset, cpu);
+
+ global += pcp->vm_stat_diff[si];
+ }
+ }
+#endif
+ return global;
+}
+
static u64 vmevent_attr_free_pages(struct vmevent_watch *watch,
struct vmevent_attr *attr)
{
- return global_page_state(NR_FREE_PAGES);
+ return vmevent_global_page_state(attr, NR_FREE_PAGES);
}
static u64 vmevent_attr_avail_pages(struct vmevent_watch *watch,
Currently vmevent relies on the global page state stats, which is updated once per stat_interval (1 second) or when the per CPU pageset stats hit their threshold. We can sum the vm_stat_diff values asynchronously: they will be possibly a bit inconsistent, but overall this should improve accuracy, since with previous scheme we would always use worst-case scenario (i.e. we'd wait for threshold to hit or 1 second delay), but now we use somewhat average accuracy. The idea is very similar to zone_page_state_snapshot(). Note that this might cause more pressure on CPU caches, so we only use this when userland explicitly asks for accuracy, plus since we gather stats outside of any fastpath, this should be OK in general. Signed-off-by: Anton Vorontsov <anton.vorontsov@linaro.org> --- mm/vmevent.c | 43 ++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 42 insertions(+), 1 deletion(-)