Re: [RFC 0/6]mm: add new LRU list for MADV_FREE pages

From: Johannes Weiner
Date: Tue Jan 31 2017 - 16:38:29 EST

On Tue, Jan 31, 2017 at 11:45:47AM -0800, Shaohua Li wrote:
> On Tue, Jan 31, 2017 at 01:59:49PM -0500, Johannes Weiner wrote:
> > Hi Shaohua,
> >
> > On Sun, Jan 29, 2017 at 09:51:17PM -0800, Shaohua Li wrote:
> > > We are trying to use MADV_FREE in jemalloc. Several issues are found. Without
> > > solving the issues, jemalloc can't use the MADV_FREE feature.
> > > - Doesn't support system without swap enabled. Because if swap is off, we can't
> > > or can't efficiently age anonymous pages. And since MADV_FREE pages are mixed
> > > with other anonymous pages, we can't reclaim MADV_FREE pages. In current
> > > implementation, MADV_FREE will fallback to MADV_DONTNEED without swap enabled.
> > > But in our environment, a lot of machines don't enable swap. This will prevent
> > > our setup using MADV_FREE.
> > > - Increases memory pressure. page reclaim bias file pages reclaim against
> > > anonymous pages. This doesn't make sense for MADV_FREE pages, because those
> > > pages could be freed easily and refilled with very slight penality. Even page
> > > reclaim doesn't bias file pages, there is still an issue, because MADV_FREE
> > > pages and other anonymous pages are mixed together. To reclaim a MADV_FREE
> > > page, we probably must scan a lot of other anonymous pages, which is
> > > inefficient. In our test, we usually see oom with MADV_FREE enabled and nothing
> > > without it.
> >
> > Fully agreed, the anon LRU is a bad place for these pages.
> >
> > > For the first two issues, introducing a new LRU list for MADV_FREE pages could
> > > solve the issues. We can directly reclaim MADV_FREE pages without writting them
> > > out to swap, so the first issue could be fixed. If only MADV_FREE pages are in
> > > the new list, page reclaim can easily reclaim such pages without interference
> > > of file or anonymous pages. The memory pressure issue will disappear.
> >
> > Do we actually need a new page flag and a special LRU for them? These
> > pages are basically like clean cache pages at that point. What do you
> > think about clearing their PG_swapbacked flag on MADV_FREE and moving
> > them to the inactive file list? The way isolate+putback works should
> > not even need much modification, something like clear_page_mlock().
> >
> > When the reclaim scanner finds anon && dirty && !swapbacked, it can
> > again set PG_swapbacked and goto keep_locked to move the page back
> > into the anon LRU to get reclaimed according to swapping rules.
> Interesting idea! Not sure though, the MADV_FREE pages are actually anonymous
> pages, this will introduce confusion. On the other hand, if the MADV_FREE pages
> are mixed with inactive file pages, page reclaim need to reclaim a lot of file
> pages first before reclaim the MADV_FREE pages. This doesn't look good. The
> point of a separate LRU is to avoid scan other anon/file pages.

The LRU code and the rest of VM already use independent page type
distinctions. That's because shmem pages are !PageAnon - they have a
page->mapping that points to a real address space, not an anon_vma -
but they are swapbacked and thus go through the anon LRU. This would
just do the reverse: put PageAnon pages on the file LRU when they
don't contain valid data and are thus not swapbacked.

As far as mixing with inactive file pages goes, it'd be possible to
link the MADV_FREE pages to the tail of the inactive list, rather than
the head. That said, I'm not sure reclaiming use-once filesystem cache
before MADV_FREE is such a bad policy. MADV_FREE retains the vmas for
the sole purpose of reusing them in the (near) future. That is
actually a stronger reuse signal than we have for use-once file pages.
If somebody does continuous writes to a logfile or a one-off search
through one or more files, we should actually reclaim that cache
before we go after MADV_FREE pages that are temporarily invalidated.

> > > For the third issue, we can add a separate RSS count for MADV_FREE pages. The
> > > count will be increased in madvise syscall and decreased in page reclaim (eg,
> > > unmap). One issue is activate_page(). A MADV_FREE page can be promoted to
> > > active page there. But there isn't mm_struct context at that place. Iterating
> > > vma there sounds too silly. The patchset don't fix this issue yet. Hopefully
> > > somebody can share a hint how to fix this issue.
> >
> > This problem also goes away if we use the file LRUs.
> Can you elaborate this please? Maybe you mean charge them to MM_FILEPAGES? But
> that doesn't solve the problem. 'statm' proc file will still report a big RSS.

Sorry, I was just referring to the activate_page(). If we use the file
LRUs, then page activation has a clear target. And we wouldn't have to
adjust any RSS counters when a lazyfreed page is activated.

If we have MM context everywhere else, can we add MM_LAZYPAGES or
something and exclude them from MM_ANONPAGES? The total RSS count will
still include everything (including mapped clean cache, which is also
easily reclaimable btw), but /proc/foo/status could provide a detailed
breakdown and allow the user to look at only RssAnon.