Re: [RFC v8 PATCH 00/20] memory-hotplug: hot-remove physical memory

Hi,
On Tue, Sep 18, 2012 at 05:39:37PM +0800, Wen Congyang wrote:

At 09/13/2012 01:18 AM, Vasilis Liaskovitis Wrote:

quoted

Hi,

On Wed, Sep 12, 2012 at 01:20:28PM +0800, Wen Congyang wrote:

quoted

Hmm, seabios doesn't support ACPI table SLIT. We can specify node it for dimm
device, so I think we should support SLIT in seabios. Otherwise we may meet
the following kernel messages:
[  325.016769] init_memory_mapping: [mem 0x40000000-0x5fffffff]
[  325.018060]  [mem 0x40000000-0x5fffffff] page 2M
[  325.019168] [ffffea0001000000-ffffea00011fffff] potential offnode page_structs
[  325.024172] [ffffea0001200000-ffffea00013fffff] potential offnode page_structs
[  325.028596]  [ffffea0001400000-ffffea00017fffff] PMD -> [ffff880035000000-ffff8800353fffff] on node 1
[  325.031775] [ffffea0001600000-ffffea00017fffff] potential offnode page_structs

Do you have plan to do it?

thanks for testing.

commit 5294828 from https://github.com/vliaskov/seabios/commits/memhp-v2
implements a SLIT table for the given numa nodes.

Hmm, why do you set node_distance(i, j) to REMOTE_DISTANCE if i != j?

What's the alternative?

Afaik SLIT[i][j] shows the distance between proximity domains (_PXM) i and j. It
doesn't correspond to individual SRAT entries. So i and j here are not memory
ranges associated with 2 different dimms. They denote domains i and j, which map
to 2 different logical nodeids in the kernel.

A default setting would be to set the entry to REMOTE_DISTANCE for all different
domains (i!=j). So this SLIT implementation is not useful, since it results
in the same numa_distance values as the non-SLIT kernel calculation in
include/linux/topology.h

quoted

However I am not sure the SLIT is the problem. The kernel builds a default
numa_distance table in arch/x86/mm/numa.c: numa_alloc_distance(). If the BIOS
doesn't present a SLIT, this should take effect (numactl --hardware should
report this table)

If the BIOS doesn't present a SLIT, numa_distance_cnt is set to 0 in the
function numa_reset_distance(). So node_distance(i, j) is REMOTE_DISTANCE(i != j).

quoted

Do you have more details on how to reproduce the warning? e.g. how many dimms
are present in the system? Does this happen on the first dimm hot-plugged?
Are all SRAT entries parsed correctly at boot-time or do you see any other
warnings at boot-time?

I can't reproduce it again. IIRC, I only do the following things:
hotplug a memory device, online the pages, offline the pages and hot remove
the memory device.

Is the sparse_vmemmap allocation supposed to guarantee no off-node allocations?
If not, then the warning could be valid.

thanks,

- Vasilis