Re: [RFC PATCH] mm: support CONFIG_ZONE_DEVICE + CONFIG_ZONE_DMA

[Dan Williams]

On Tue, Jan 26, 2016 at 5:18 PM, Joonsoo Kim <iamjoonsoo.kim@xxxxxxx> wrote:
> Hello,
>
> On Tue, Jan 26, 2016 at 03:11:36PM -0800, Dan Williams wrote:
>> On Tue, Jan 26, 2016 at 2:51 PM, Andrew Morton
>> <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
>> > On Tue, 26 Jan 2016 14:33:48 -0800 Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
>> >
>> >> >> Towards this end, alias ZONE_DMA and ZONE_DEVICE to work around needing
>> >> >> to maintain a unique zone number for ZONE_DEVICE.  Record the geometry
>> >> >> of ZONE_DMA at init (->init_spanned_pages) and use that information in
>> >> >> is_zone_device_page() to differentiate pages allocated via
>> >> >> devm_memremap_pages() vs true ZONE_DMA pages.  Otherwise, use the
>> >> >> simpler definition of is_zone_device_page() when ZONE_DMA is turned off.
>> >> >>
>> >> >> Note that this also teaches the memory hot remove path that the zone may
>> >> >> not have sections for all pfn spans (->zone_dyn_start_pfn).
>> >> >>
>> >> >> A user visible implication of this change is potentially an unexpectedly
>> >> >> high "spanned" value in /proc/zoneinfo for the DMA zone.
>> >> >
>> >> > Well, all these icky tricks are to avoid increasing ZONES_SHIFT, yes?
>> >> > Is it possible to just use ZONES_SHIFT=3?
>> >>
>> >> Last I tried I hit this warning in mm/memory.c
>> >>
>> >> #warning Unfortunate NUMA and NUMA Balancing config, growing
>> >> page-frame for last_cpupid.
>> >
>> > Well yes, it may take a bit of work - perhaps salvaging a bit from
>> > somewhere else if poss.  But that might provide a better overall
>> > solution so could you please have a think?
>> >
>>
>> Will do, especially since other efforts are feeling the pinch on the
>> MAX_NR_ZONES limitation.
>
> Please refer my previous attempt to add a new zone, ZONE_CMA.
>
> https://lkml.org/lkml/2015/2/12/84
>
> It salvages a bit from SECTION_WIDTH by increasing section size.
> Similarly, I guess we can reduce NODE_WIDTH if needed although
> it could cause to reduce maximum node size.

Dave pointed out to me that LAST__PID_SHIFT might be a better
candidate to reduce to 7 bits.  That field is for storing pids which
are already bigger than 8 bits.  If it is relying on the fact that
pids don't rollover very often then likely the impact of 7-bits
instead of 8 will be minimal.