Re: [PATCH 00/10] fix freepage count problems due to memory isolation

[Joonsoo Kim]

On Wed, Jul 16, 2014 at 01:14:26PM +0200, Vlastimil Babka wrote:
> On 07/16/2014 10:43 AM, Joonsoo Kim wrote:
> >> I think your plan of multiple parallel CMA allocations (and thus
> >> multiple parallel isolations) is also possible. The isolate pcplists
> >> can be shared by pages coming from multiple parallel isolations. But
> >> the flush operation needs a pfn start/end parameters to only flush
> >> pages belonging to the given isolation. That might mean a bit of
> >> inefficient list traversing, but I don't think it's a problem.
> > 
> > I think that special pcplist would cause a problem if we should check
> > pfn range. If there are too many pages on this pcplist, move pages from
> > this pcplist to isolate freelist takes too long time in irq context and
> > system could be broken. This operation cannot be easily stopped because
> > it is initiated by IPI on other cpu and starter of this IPI expect that
> > all pages on other cpus' pcplist are moved properly when returning
> > from on_each_cpu().
> > 
> > And, if there are so many pages, serious lock contention would happen
> > in this case.
> 
> Hm I see. So what if it wasn't a special pcplist, but a special "free list"
> where the pages would be just linked together as on pcplist, regardless of
> order, and would not merge until the CPU that drives the memory isolation
> process decides it is safe to flush them away. That would remove the need for
> IPI's and provide the same guarantees I think.

Looks good. It would work. I think that your solution is better than mine.
I will implement it and test. 

> 
> > Anyway, my idea's key point is using PageIsolated() to distinguish
> > isolated page, instead of using PageBuddy(). If page is PageIsolated(),
> 
> Is PageIsolated a completely new page flag? Those are a limited resource so I
> would expect some resistance to such approach. Or a new special page->_mapcount
> value? That could maybe work.

Yes, it is new special page->_mapcount.

> > it isn't handled as freepage although it is in buddy allocator. During free,
> > page with MIGRATETYPE_ISOLATE will be marked as PageIsolated() and
> > won't be merged and counted for freepage.
> 
> OK. Preventing wrong merging is the key point and this should work.
> 
> > When we move pages from normal buddy list to isolate buddy
> > list, we check PageBuddy() and subtract number of PageBuddy() pages
> 
> Do we really need to check PageBuddy()? Could a page get marked as PageIsolate()
> but still go to normal list instead of isolate list?

Checking PageBuddy() is used for identifying page linked in normal
list.

Thanks.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/