Re: [patch 1/2]vmscan: correct all_unreclaimable for zone without lru pages

[Minchan Kim <minchan.kim@gmail.com>]

On Fri, Sep 30, 2011 at 10:12:23AM +0800, Shaohua Li wrote:
> On Thu, 2011-09-29 at 17:18 +0800, Minchan Kim wrote:
> > On Thu, Sep 29, 2011 at 09:14:51AM +0800, Shaohua Li wrote:
> > > On Thu, 2011-09-29 at 01:57 +0800, Minchan Kim wrote:
> > > > On Wed, Sep 28, 2011 at 03:08:31PM +0800, Shaohua Li wrote:
> > > > > On Wed, 2011-09-28 at 14:57 +0800, Minchan Kim wrote:
> > > > > > On Tue, Sep 27, 2011 at 03:23:04PM +0800, Shaohua Li wrote:
> > > > > > > I saw DMA zone always has ->all_unreclaimable set. The reason is the high zones
> > > > > > > are big, so zone_watermark_ok/_safe() will always return false with a high
> > > > > > > classzone_idx for DMA zone, because DMA zone's lowmem_reserve is big for a high
> > > > > > > classzone_idx. When kswapd runs into DMA zone, it doesn't scan/reclaim any
> > > > > > > pages(no pages in lru), but mark the zone as all_unreclaimable. This can
> > > > > > > happen in other low zones too.
> > > > > > 
> > > > > > Good catch!
> > > > > > 
> > > > > > > This is confusing and can potentially cause oom. Say a low zone has
> > > > > > > all_unreclaimable when high zone hasn't enough memory. Then allocating
> > > > > > > some pages in low zone(for example reading blkdev with highmem support),
> > > > > > > then run into direct reclaim. Since the zone has all_unreclaimable set,
> > > > > > > direct reclaim might reclaim nothing and an oom reported. If
> > > > > > > all_unreclaimable is unset, the zone can actually reclaim some pages.
> > > > > > > If all_unreclaimable is unset, in the inner loop of balance_pgdat we always have
> > > > > > > all_zones_ok 0 when checking a low zone's watermark. If high zone watermark isn't
> > > > > > > good, there is no problem. Otherwise, we might loop one more time in the outer
> > > > > > > loop, but since high zone watermark is ok, the end_zone will be lower, then low
> > > > > > > zone's watermark check will be ok and the outer loop will break. So looks this
> > > > > > > doesn't bring any problem.
> > > > > > 
> > > > > > I think it would be better to correct zone_reclaimable.
> > > > > > My point is zone_reclaimable should consider zone->pages_scanned.
> > > > > > The point of the function is how many pages scanned VS how many pages remained in LRU.
> > > > > > If reclaimer doesn't scan the zone at all because of no lru pages, it shouldn't tell
> > > > > > the zone is all_unreclaimable.
> > > > > actually this is exact my first version of the patch. The problem is if
> > > > > a zone is true unreclaimable (used by kenrel pages or whatever), we will
> > > > > have zone->pages_scanned 0 too. I thought we should set
> > > > > all_unreclaimable in this case.
> > > > 
> > > > Let's think the problem again.
> > > > Fundamental problem is that why the lower zone's lowmem_reserve for higher zone is huge big
> > > > that might be bigger than the zone's size.
> > > > I think we need the boundary for limiting lowmem_reseve.
> > > > So how about this?
> > > I didn't see a reason why high zone allocation should fallback to low
> > > zone if high zone is big. Changing the lowmem_reserve can cause the
> > > fallback. Has any rationale here?
> > 
> > I try to think better solution than yours but I got failed. :(
> > The why I try to avoid your patch is that kswapd is very complicated these days so
> > I wanted to not add more logic for handling corner cases if we can solve it
> > other ways. But as I said, but I got failed.
> > 
> > It seems that it doesn't make sense that previous my patch that limit lowmem_reserve.
> > Because we can have higher zone which is very big size so that lowmem_zone[higher_zone] of
> > low zone could be bigger freely than the lowmem itself size.
> > It implies the low zone should be not used for higher allocation.
> > It has no reason to limit it. My brain was broken. :(
> > 
> > But I have a question about your patch still.
> > What happens if DMA zone sets zone->all_unreclaimable with 1?
> > 
> > You said as follows,
> > 
> > > This is confusing and can potentially cause oom. Say a low zone has
> > > all_unreclaimable when high zone hasn't enough memory. Then allocating
> > > some pages in low zone(for example reading blkdev with highmem support),
> > > then run into direct reclaim. Since the zone has all_unreclaimable set,
> > 
> > If low zone has enough pages for allocation, it cannot have entered reclaim.
> > It means now low zone doesn't have enough free pages for the order allocation.
> > So it's natural to enter reclaim path.
> > 
> > > direct reclaim might reclaim nothing and an oom reported. If
> > 
> > It's not correct "nothing". At least, it will do something in DEF_PRIORITY.
> it does something, but might not reclaim any pages, for example, it
> starts page write, but page isn't in disk yet in DEF_PRIORITY and it
> skip further reclaiming in !DEF_PRIORITY.
> 
> > > all_unreclaimable is unset, the zone can actually reclaim some pages.
> > 
> > The reason of this problem is that the zone has no lru page, you said.
> > Then how could we reclaim some pages in the zone even if the zone's all_unreclaimable is unset?
> > You expect slab pages?
> The zone could have lru pages. Let's take an example, allocation from
> ZONE_HIGHMEM, then kswapd runs, ZONE_NORMAL gets all_unreclaimable set
> even it has free pages. Then we do write blkdev device, which use
> ZONE_NORMAL for page cache. Some pages in ZONE_NORMAL are in lru, then
> we run into direct page reclaim for ZONE_NORMAL. Since all_unreclaimable
> is set and pages in ZONE_NORMAL lru are dirty, direct reclaim could
> fail. But I'd agree this is a corner case.
> Besides when I saw ZONE_DMA has a lot of free pages and
> all_unreclaimable is set, it's really confusing.

Hi Shaohua,
Sorry for late response and Thanks for your explanation.
It's valuable to fix, I think.
How about this?

I hope other guys have a interest in the problem.
Cced them.