On Thu, Nov 05, 2020 at 09:10:12PM +0800, Yafang Shao wrote:
> The memory utilization (Used / Total) is used to monitor the memory
> pressure by us. If it is too high, it means the system may be OOM sooner
> or later when swap is off, then we will make adjustment on this system.
>
> However, this method is broken since MADV_FREE is introduced, because
> these lazily free anonymous can be reclaimed under memory pressure while
> they are still accounted in NR_ANON_MAPPED.
>
> Furthermore, since commit f7ad2a6cb9f7 ("mm: move MADV_FREE pages into
> LRU_INACTIVE_FILE list"), these lazily free anonymous pages are moved
> from anon lru list into file lru list. That means
> (Inactive(file) + Active(file)) may be much larger than Cached in
> /proc/meminfo. That makes our users confused.
>
> So we'd better account the lazily freed anonoymous pages in
> NR_FILE_PAGES as well.
What about the share of pages that have been reused? After all, the
idea behind deferred reclaim is cheap reuse of already allocated and
faulted in pages.
Anywhere between 0% and 100% of MADV_FREEd pages may be dirty and need
swap-out to reclaim. That means even after this patch, your formula
would still have an error margin of 100%.
The tradeoff with saving the reuse fault and relying on the MMU is
that the kernel simply *cannot do* lazy free accounting. Userspace
needs to do it. E.g. if a malloc implementation or similar uses
MADV_FREE, it has to keep track of what is and isn't used and make
those stats available.
If that's not practical, I don't see an alternative to trapping minor
faults upon page reuse, eating the additional TLB flush, and doing the
accounting properly inside the kernel.
> @@ -1312,8 +1312,13 @@ static void page_remove_anon_compound_rmap(struct page *page)
> if (unlikely(PageMlocked(page)))
> clear_page_mlock(page);
>
> - if (nr)
> - __mod_lruvec_page_state(page, NR_ANON_MAPPED, -nr);
> + if (nr) {
> + if (PageLRU(page) && PageAnon(page) && !PageSwapBacked(page) &&
> + !PageSwapCache(page) && !PageUnevictable(page))
> + __mod_lruvec_page_state(page, NR_FILE_PAGES, -nr);
> + else
> + __mod_lruvec_page_state(page, NR_ANON_MAPPED, -nr);
I don't think this would work. The page can be temporarily off-LRU for
compaction, migration, reclaim etc. and then you'd misaccount it here.
