Hi,

This patch tries to improve/fix some problems with the current vm,
most notably out of memory problems.

* page_alloc.c, hunk 1
Since the balancing point is lower than on earlier kernels I think
there are reasons to keep more pages around on the special DMA zone...
(most computers has 16 MB DMA memory without this patch this gives an
 actual max of 32 pages, with this patch it gets 125 - probably more
 balanced against other zones)

* page_alloc.c, hunk 2
 More info in Sysreq-M (debug code)

* vmscan.c, hunk 1
The implemented balancing breaks stops when any zone gets out
of zone_wake_kswapd... but what about when only one zone is
really low on memory ( < pages_min ) - kswapd won't run.
This fixes that.
(I have tried to make the limit < pages_low but that disturbs
 the alternating balancing too much.)

* vmscan.c, hunk 2
After sleeping - pages might be gone, restart count.
(does not fix a resembling problem when waiting for a buffer in
shrink_mmap)

* vmscan.c, hunk 3
Actually it is not unlikely to have less than 'pages_low' free
due to the implemented balancing.


/RogerL

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