I hate to resuscitate this old thread, sorry but I'm still not sure we resolved it and I want to make sure this issue isn't here as I see it.
On Saturday 08 April 2006 11:25, Nick Piggin wrote:
Con Kolivas wrote:
Ok. I think I presented enough information for why I thought
zone_watermark_ok would fail (for ZONE_DMA). With 16MB ZONE_DMA and a
vmsplit of 3GB we have a lowmem_reserve of 12MB. It's pretty hard to keep
that much ZONE_DMA free, I don't think I've ever seen that much free on
my ZONE_DMA on an ordinary desktop without any particular ZONE_DMA users.
Changing the tunable can make the lowmem_reserve larger than ZONE_DMA is
on any vmsplit too as far as I understand the ratio.
Umm, for ZONE_DMA allocations, ZONE_DMA isn't a lower zone. So that
12MB protection should never come into it (unless it is buggy?).
An i386 pc with a 3GB split will have approx
4000 pages ZONE_DMA
and lowmem reserve will set lowmem reserve to approx
0 0 3000 3000
So if we call zone_watermark_ok with zone of ZONE_DMA and a classzone_idx of a ZONE_NORMAL we will fail a zone_watermark_ok test almost always since it's almost impossible to have 3000 free ZONE_DMA pages. I believe it can happen like this:
In balance_pgdat (vmscan.c:1116) if we end up with end_zone being a ZONE_NORMAL zone, then during the scan below we (vmscan.c:1137) iterate over all zones from 0 to end_zone and (vmscan.c:1147) we end up calling
if (!zone_watermark_ok(zone, order, zone->pages_high, end_zone, 0))
which would now call zone_watermark_ok with zone being a ZONE_DMA, and end_zone being the idx of a ZONE_NORMAL.
So in summary if I'm not mistaken (and I'm good at being mistaken), if we balance pgdat and find that ZONE_NORMAL or higher needs scanning, we'll end up trying to flush the crap out of ZONE_DMA.
On my test case this indeed happens and my ZONE_DMA never goes below 3000
pages free. If I lower the reserve even further my pages free gets stuck at
3208 and can't free any more, and doesn't ever drop below that either.
Here is the patch I was proposing