> will actually set zone_start_pfn/zone_end_pfn to the values from node0's
> ZONE_NORMAL?
> So we use clamp to actually check if such values fall within what node1's
> memory spans, and ignore them otherwise?
That's right.
Normally, zone_start_pfn/zone_end_pfn has the same value for all nodes.
Let's look at another example, which is obtained by adding some debugging
information.
e.g.
Zone ranges:
DMA [mem 0x0000000000001000-0x0000000000ffffff]
DMA32 [mem 0x0000000001000000-0x00000000ffffffff]
Normal [mem 0x0000000100000000-0x0000000792ffffff]
Movable zone start for each node
Node 0: 0x0000000100000000
Node 1: 0x00000002b1000000
Node 2: 0x0000000522000000
Early memory node ranges
node 0: [mem 0x0000000000001000-0x000000000009efff]
node 0: [mem 0x0000000000100000-0x00000000bffdefff]
node 0: [mem 0x0000000100000000-0x00000002b0ffffff]
node 1: [mem 0x00000002b1000000-0x0000000521ffffff]
node 2: [mem 0x0000000522000000-0x0000000792ffffff]
Node 0:
node_start_pfn=1 node_end_pfn=2822144
DMA zone_low=1 zone_high=4096
DMA32 zone_low=4096 zone_high=1048576
Normal zone_low=1048576 zone_high=7942144
Movable zone_low=0 zone_high=0
Node 1:
node_start_pfn=2822144 node_end_pfn=5382144
DMA zone_low=1 zone_high=4096
DMA32 zone_low=4096 zone_high=1048576
Normal zone_low=1048576 zone_high=7942144
Movable zone_low=0 zone_high=0
Node 2:
node_start_pfn=5382144 node_end_pfn=7942144
DMA zone_low=1 zone_high=4096
DMA32 zone_low=4096 zone_high=1048576
Normal zone_low=1048576 zone_high=7942144
Movable zone_low=0 zone_high=0
Before this patch, zone_start_pfn/zone_end_pfn in node 0,1,2 is the same:
DMA zone_start_pfn:1 zone_end_pfn:4096
DMA32 zone_start_pfn:4096 zone_end_pfn:1048576
Normal zone_start_pfn:1048576 zone_end_pfn:7942144
Movable zone_start_pfn:0 zone_end_pfn:0
spaned pages resuelt:
node 0:
DMA spanned:4095
DMA32 spanned:1044480
Normal spanned:0
Movable spanned:1773568
totalpages:2559869
node 1:
DMA spanned:0
DMA32 spanned:0
Normal spanned:2560000
Movable spanned:2560000
totalpages:5120000
node 2:
DMA spanned:0
DMA32 spanned:0
Normal spanned:2560000
Movable spanned:2560000
totalpages:5120000
After this patch:
node 0:
DMA zone_start_pfn:1 zone_end_pfn:4096 spanned:4095
DMA32 zone_start_pfn:4096 zone_end_pfn:1048576 spanned:1044480
Normal zone_start_pfn:1048576 zone_end_pfn:2822144 spanned:0
Movable zone_start_pfn:0 zone_end_pfn:0 spanned:1773568
totalpages:2559869
node 1:
DMA zone_start_pfn:4096 zone_end_pfn:4096 spanned:0
DMA32 zone_start_pfn:1048576 zone_end_pfn:1048576 spanned:0
Normal zone_start_pfn:2822144 zone_end_pfn:5382144 spanned:0
Movable zone_start_pfn:0 zone_end_pfn:0 spanned:2560000
totalpages:2560000
node 2:
DMA zone_start_pfn:4096 zone_end_pfn:4096 spanned:0
DMA32 zone_start_pfn:1048576 zone_end_pfn:1048576 spanned:0
Normal zone_start_pfn:5382144 zone_end_pfn:7942144 spanned:0
Movable zone_start_pfn:0 zone_end_pfn:0 spanned:2560000
totalpages:2560000
It is easy to construct such a scenario by configuring kernelcore=mirror
in a multi-NUMA machine without full mirrored memory.
Of course, it can be a machine without any mirrored memory.
A great difference can be observed by startup information and viewing
/proc/pagetypeinfo.
On earlier kernel versions, such BUGs will directly double the memory of
some nodes.
Although these redundant memory exists in the form of reserved memory,
this should not be expected.
