IMO high memory should not be balanced. Stock pre7-9 tried to balance high
memory once it got below the treshold (causing very bad VM behavior and
high kswapd usage) - this is incorrect because there is nothing special
about the highmem zone, it's more like an 'extension' of the normal zone,
from which specific caches can turn. (patch attached)

another problem is that even during a mild test the DMA zone gets emptied
easily - but on a big RAM box kswapd has to work _alot_ to fill it up. In
fact on an 8GB box it's completely futile to fill up the DMA zone. What
worked for me is this zone-chainlist trick in the zone setup code:

                        case ZONE_NORMAL:
                                zone = pgdat->node_zones + ZONE_NORMAL;
                                if (zone->size)
                                        zonelist->zones[j++] = zone;
++                              break;
                        case ZONE_DMA:
                                zone = pgdat->node_zones + ZONE_DMA;
                                if (zone->size)
                                        zonelist->zones[j++] = zone;

no 'normal' allocation chain leads to the ZONE_DMA zone, except GFP_DMA
and GFP_ATOMIC - both of them rightfully access the DMA zone.

this is a RL problem, without the above a 8GB box under load crashes
pretty quickly due to failed SCSI-layer DMA allocations. (i think those
allocations are silly in the first place.)

the above is suboptimal on boxes which have total RAM within one order of
magnitude of 16MB (the DMA zone stays empty most of the time and is
unaccessible to various caches) - so maybe the following (not yet
implemented) solution would be generic and acceptable:

allocate 5% of total RAM or 16MB to the DMA zone (via fixing up zone sizes
on bootup), whichever is smaller, in 2MB increments. Disadvantage of this
method: eg. it wastes 2MB RAM on a 8MB box. We could probably live with
64kb increments (there are 64kb ISA DMA constraints the sound drivers and
some SCSI drivers are hitting) - is this really true? If nobody objects
i'll implement this later one (together with the assymetric allocation
chain trick) - there will be a 64kb DMA pool allocated on the smallest
boxes, which should be acceptable even on a 4MB box. We could turn off the
DMA zone altogether on most boxes, if it wasnt for the SCSI layer
allocating DMA pages even for PCI drivers ...

Comments?

	Ingo