On Thu, May 03, 2012 at 11:10:12AM +0300, Pekka Enberg wrote:
On Tue, May 1, 2012 at 4:24 PM, Anton Vorontsov anton.vorontsov@linaro.org wrote:
Accounting only free pages is very inaccurate for low memory handling, so we have to be smarter here.
Can you elaborate on what kind of problems there are with tracking free pages?
Well, there's no problem with tracking itself, the word 'inaccurate' was probably misleading. Tracking just free pages is inaccurate for our "low memory" notification needs, but NR_FREE_PAGES tracking itself is fine.
The thing is that NR_FREE_PAGES accounts only completely unused (wasted) pages. Most of the time we have very low NR_FREE_PAGES, and lots of page cache and block buffers (i.e. NR_FILE_PAGES).
The file pages are easily reclaimable (except shmem/tmpfs and locked pages), so file pages may be considered as "somewhat free" pages.
The cache might contain very stale data (or not), so we have to maneuver between the two strategies: sacrifice caches, or start freeing memory (which prevents caches draining).
The strategy is described in the third patch in the series. It might be not ideal, but the logic itself is not part of the ABI (this is very similar "not ABI" rules as we have for OOM scoring logic), and is subject for changes.
Thanks,