Hi,

[Intermediate release - indicator of problem with previous]

Since I am responsible for messing up some aspects of vm
(when fixing others)
here is a patch that tries to solve the introduced problems.

* no more periodic wake up of kswapd - not needed anymore
* no more freeing all zones to (free_pages > pages_high)
* always wakes kswapd up after try_to_free_pages
* kswapd starts when all zones gets zone_wake_kswapd
  (runs once for each zone that hits zone_wake_kswapd)
* removed test for more than pages_high in alloc_pages,
  zones will mostly be in the range [pages_high...pages_low]
* Up to 10% better throughput than 2.4.0-test4, YMMV
* Tested mostly with streaming tests. On a non HIGHMEM config.

+ New: zone_wake_kswapd == 2 indicates a situation where
  free_pages < pages_low another alloc is done.
+ New-BUGFIX: runs kswapd while zone_wake_kswapd == 2 or
  low_on_memory (Quintela)
+ New-BUGFIX? checks if __GFP_IO before rescheduling
  (will become separate patch if correct, thanks Quintela)


- Kills mmap002, did not in the previously released that
  should have this problem (but did not) - why? Will
  investigate further...
- Since kswapd does not wake up periodic anymore, the
  latencies might be worse... Will investigate it
  further when other stuff works.

Note: logic of function keep_kswapd_awake has changed.

/RogerL


--
Home page:
  http://www.norran.net/nra02596/