On Thu, 26 Nov 2015, Joonsoo Kim wrote: > I understand design decision, but, it is better to get value as much > as accurate if there is no performance problem. My patch would not > cause much performance degradation because it is just adding one > this_cpu_read(). > > Consider about following example. Current implementation returns > interesting output if someone do following things. > > v1 = zone_page_state(XXX); > mod_zone_page_state(XXX, 1); > v2 = zone_page_state(XXX); > > v2 would be same with v1 in most of cases even if we already update > it. > > This situation could occurs in page allocation path and others. If > some task try to allocate many pages, then watermark check returns > same values until updating vmstat even if some freepage are allocated. > There are some adjustments for this imprecision but why not do it become > accurate? I think that this change is reasonable trade-off. >
I'm not sure that NR_ISOLATED_* should be vmstats in the first place. The most important callers that depend on its accuracy is zone_reclaimable_pages() and the too_many_isolated() loop in both shrink_inactive_list() and memory compaction. If zlc's are updated every 1s, the HZ/10 in those loops don't really matter, they may as well be HZ/2. I think memory compaction updates the counters in the most appropriate way, by incrementing a counter and then finally doing mod_zone_page_state() for the counter. The other updaters are thp collapse and page migration. I discount user-visible vmstats here because the trade-off has already been made that they may be stale for up to 1s and userspace isn't affected. So what happens if we simply convert NR_ISOLATED_* into per-zone atomic64_t?
