On 02/27/2021 02:29 PM, Huang Pei wrote:
It can make some metadata of MM, like pglist_data and zone
NUMA-aware
Signed-off-by: Huang Pei <huangpei@xxxxxxxxxxx>
---
arch/mips/loongson64/numa.c | 18 +++++++++++++++---
1 file changed, 15 insertions(+), 3 deletions(-)
diff --git a/arch/mips/loongson64/numa.c b/arch/mips/loongson64/numa.c
index cf9459f79f9b..5912b2e7b10c 100644
--- a/arch/mips/loongson64/numa.c
+++ b/arch/mips/loongson64/numa.c
@@ -26,7 +26,6 @@
#include <asm/wbflush.h>
#include <boot_param.h>
-static struct pglist_data prealloc__node_data[MAX_NUMNODES];
unsigned char __node_distances[MAX_NUMNODES][MAX_NUMNODES];
EXPORT_SYMBOL(__node_distances);
struct pglist_data *__node_data[MAX_NUMNODES];
@@ -151,8 +150,12 @@ static void __init szmem(unsigned int node)
static void __init node_mem_init(unsigned int node)
{
+ struct pglist_data *nd;
unsigned long node_addrspace_offset;
unsigned long start_pfn, end_pfn;
+ unsigned long nd_pa;
+ int tnid;
+ const size_t nd_size = roundup(sizeof(pg_data_t), SMP_CACHE_BYTES);
node_addrspace_offset = nid_to_addrbase(node);
pr_info("Node%d's addrspace_offset is 0x%lx\n",
@@ -162,8 +165,16 @@ static void __init node_mem_init(unsigned int node)
pr_info("Node%d: start_pfn=0x%lx, end_pfn=0x%lx\n",
node, start_pfn, end_pfn);
- __node_data[node] = prealloc__node_data + node;
-
+ nd_pa = memblock_phys_alloc_try_nid(nd_size, SMP_CACHE_BYTES, node);
[...]
Hi, all,
Few related to this patch. About memblock, I have a question.
In my own understanding, 3 stages at previous part of arch_mem_init().
stage1: Add memory to memblock.memory at plat_mem_setup().
stage2: parse_early_param() parses parameter about memroy, such as "mem"
"memmap".
check_kernel_sections_mem() checks whether the current
memblock.memory contains the kernel.
At this stage, user can defined memblock.memory by themselves.
Also it is the final stage of determining memblock.memory.
stage3: others. use memblock.memory and add them to memblock.reserve.
Calling to memblock_alloc*() should be after the reserve kernel(), see:
memblock_set_current_limit(PFN_PHYS(max_low_pfn))
(
Here few about this patch: It works no problem, memblock_dump shows:
[ 0.000000] memory.cnt = 0x2
[ 0.000000] memory[0x0] [0x0000000000200000-0x000000000effffff],
0x000000000ee00000 bytes on node 0 flags: 0x0
[ 0.000000] memory[0x1] [0x0000000090200000-0x000000025fffffff],
0x00000001cfe00000 bytes on node 0 flags: 0x0
[ 0.000000] reserved.cnt = 0x3
[ 0.000000] reserved[0x0] [0x0000000000200000-0x0000000001ed7fff],
0x0000000001cd8000 bytes flags: 0x0
[ 0.000000] reserved[0x1] [0x00000000fe000000-0x00000000ffffffff],
0x0000000002000000 bytes flags: 0x0
--->>> [ 0.000000] reserved[0x2]
[0x000000025fffd6c0-0x000000025fffffff], 0x0000000000002940 bytes flags:
0x0 <<<---
memblock.bottom_up is not enabled here, and it is not destroy kernel,
although looks strange.
)
Morever, about "mem" parameter.
When parsing the mem parameter for the first time, all memblock.memory
is removed.
For NUMA, memblock.memory contain important node information. These
information are
imported at plat_mem_setup(). Without these node information, the NUMA
platform
may not be able to use memory correctly.
The mem parameter is rarely used, but it has meaning. For example,
kdump. I have
done some fixes before, but it looks fool.
Huacai suggested me that use pa_to_nid().
memblock_add(start, size) -> memblock_add_node(start, size,
pa_to_nid(start))
I think this is a good way. Does anyone have other suggestions?
At last, should the NUMA platform reserve the kernel area after parse
"mem" rather than before it?
Thanks,
Jinyang
