Memory Add Fixes for ppc64
cover Memory Add Fixes for ppc64
When memory add was merged into mainline in 2.6.14, there were various bits and pieces missing that prevent it from working on ppc64. The following patches are against 2.6.14-git7 and address all but one of the know issues.
1) Create hptes for new sections
2) Clear page count before freeing new pages
3) Kludge to add new memory to node 0
4) Ensure probe file is created for memory add via sysfs
Signature · 1 line
Mike
PATCH 1/4 Memory Add Fixes for ppc64
Add the create_section_mapping() routine to create hptes for memory sections dynamically added after system boot.
arch/powerpc/mm/hash_utils_64.c +9 -0
@@ -176,6 +176,15 @@ static unsigned long get_hashtable_size(return pteg_count <<7 ; } +#ifdef CONFIG_MEMORY_HOTPLUG +void create_section_mapping (unsigned long start,unsigned long end) +{ +create_pte_mapping (start, end, + _PAGE_ACCESSED | _PAGE_COHERENT |PP_RWXX , + cur_cpu_spec->cpu_features &CPU_FTR_16M_PAGE ?1 :0 ); +} +#endif /* CONFIG_MEMORY_HOTPLUG */ +void __init htab_initialize (void ) {unsigned long table, htab_size_bytes;
arch/powerpc/mm/mem.c +3 -0
@@ -124,6 +124,9 @@ int __devinit add_memory(u64 start, u64unsigned long start_pfn = start >>PAGE_SHIFT ;unsigned long nr_pages = size >>PAGE_SHIFT ; + start +=KERNELBASE ; +create_section_mapping (start, start + size); +/* this should work for most non-highmem platforms */ zone = pgdata->node_zones;
include/asm-ppc64/sparsemem.h +4 -0
@@ -11,6 +11,10 @@#define MAX_PHYSADDR_BITS 38#define MAX_PHYSMEM_BITS 36 +#ifdef CONFIG_MEMORY_HOTPLUG +extern void create_section_mapping (unsigned long start,unsigned long end); +#endif /* CONFIG_MEMORY_HOTPLUG */ +#endif /* CONFIG_SPARSEMEM */ #endif /* _ASM_PPC64_SPARSEMEM_H */
Quoted 4 lines · Mike Kravetz
Add the
create_section_mapping()routine to create hptes for memory sections dynamically added after system boot.Signed-off-by Mike Kravetz <[email protected]>
This patch will have to be slightly reworked on top of the 64k pages one. It should be trivial though.
Ben.
Quoted 51 lines
diff -Naupr
linux-2.6.14-git7/arch/powerpc/mm/hash_utils_64.clinux-2.6.14-git7.work/arch/powerpc/mm/hash_utils_64.carch/powerpc/mm/hash_utils_64.c +9
-0@@ -176,6 +176,15 @@ static unsigned long get_hashtable_size(return pteg_count <<7 ; } +#ifdef CONFIG_MEMORY_HOTPLUG +void create_section_mapping (unsigned long start,unsigned long end) +{ +create_pte_mapping (start, end, + _PAGE_ACCESSED | _PAGE_COHERENT |PP_RWXX , + cur_cpu_spec->cpu_features &CPU_FTR_16M_PAGE ?1 :0 ); +} +#endif /* CONFIG_MEMORY_HOTPLUG */ +void __init htab_initialize (void ) {unsigned long table, htab_size_bytes;diff -Naupr
linux-2.6.14-git7/arch/powerpc/mm/mem.clinux-2.6.14-git7.work/arch/powerpc/mm/mem.carch/powerpc/mm/mem.c +3
-0@@ -124,6 +124,9 @@ int __devinit add_memory(u64 start, u64unsigned long start_pfn = start >>PAGE_SHIFT ;unsigned long nr_pages = size >>PAGE_SHIFT ; + start +=KERNELBASE ; +create_section_mapping (start, start + size); +/* this should work for most non-highmem platforms */ zone = pgdata->node_zones;diff -Naupr
linux-2.6.14-git7/include/asm-ppc64/sparsemem.hlinux-2.6.14-git7.work/include/asm-ppc64/sparsemem.hinclude/asm-ppc64/sparsemem.h +4
-1@@ -11,6 +11,10 @@#define MAX_PHYSADDR_BITS 38#define MAX_PHYSMEM_BITS 36 +#ifdef CONFIG_MEMORY_HOTPLUG +extern void create_section_mapping (unsigned long start,unsigned long end); +#endif /* CONFIG_MEMORY_HOTPLUG */ +#endif /* CONFIG_SPARSEMEM */ #endif /* _ASM_PPC64_SPARSEMEM_H */ -To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Quoted 8 lines · Benjamin Herrenschmidt
Quoted 4 lines · Mike Kravetz
Add the
create_section_mapping()routine to create hptes for memory sections dynamically added after system boot.Signed-off-by Mike Kravetz <[email protected]>This patch will have to be slightly reworked on top of the 64k pages one. It should be trivial though.
OK. I'll respin on top of your patch at:
http://gate.crashing.org/~benh/ppc64-64k-pages.diff
Let me know if there is a different version going upstream.
Signature · 1 line
Mike
Quoted 7 lines · Mike Kravetz
Quoted 6 lines · Benjamin Herrenschmidt
Quoted 4 lines · Mike Kravetz
Add the
create_section_mapping()routine to create hptes for memory sections dynamically added after system boot.Signed-off-by Mike Kravetz <[email protected]>
Quoted 8 lines
Quoted 2 lines
This patch will have to be slightly reworked on top of the 64k pages one. It should be trivial though.
OK. I'll respin on top of your patch at:
http://gate.crashing.org/~benh/ppc64-64k-pages.diff
Let me know if there is a different version going upstream
I'll check if it still applied after linus pulls the next round of ppc updates
Ben.
Quoted 2 lines · Benjamin Herrenschmidt
This patch will have to be slightly reworked on top of the 64k pages one. It should be trivial though.
Ran into an issue with the interaction of SPARSEMEM and 64k pages. SPARSEMEM defines the pp64 section size to be 16MB which corresponds to the smallest LMB size. There is a check in the SPARSEMEM code to ensure that MAX_ORDER (actually MAX_ORDER-1) block size is not greater than section size. Within the Kconfig file, there is this:
# We optimistically allocate largepages from the VM, so make the limit
# large enough (16MB). This badly named config option is actually
# max order + 1
config FORCE_MAX_ZONEORDER
int
depends on PPC64
default "13"Just curious if we still want to boost MAX_ORDER like this with 64k pages? Doesn't that make the MAX_ORDER block size 256MB in this case? Also, not quite sure what happens if memory size (a 16 MB multiple) does not align with a MAX_ORDER block size (a 256MB multiple in this case). My 'guess' is that the page allocator would not use it as it would not fit within the buddy system.
cc'ing SPARSEMEM author Andy Whitcroft.
Signature · 1 line
Mike
Quoted 9 lines
Just curious if we still want to boost
MAX_ORDERlike this with 64k pages? Doesn't that make theMAX_ORDERblock size 256MB in this case? Also, not quite sure what happens if memory size (a 16 MB multiple) does not align with aMAX_ORDERblock size (a 256MB multiple in this case). My 'guess' is that the page allocator would not use it as it would not fit within the buddy system.cc'ing SPARSEMEM author Andy Whitcroft.
Yes, the MAX_ORDER should be different indeed. But can Kconfig do that ? That is have the default value be different based on a Kconfig option ? I don't see that ... We may have to do things differently here...
Ben.
Quoted 3 lines · Benjamin Herrenschmidt
Yes, the
MAX_ORDERshould be different indeed. But canKconfigdo that ? That is have the default value be different based on aKconfigoption ? I don't see that ... We may have to do things differently here...
This seems to be done in other parts of the Kconfig file. Using those as an example, this should keep the MAX_ORDER block size at 16MB.
arch/powerpc/Kconfig +1 -0
@@ -463,6 +463,7 @@ source "fs/Kconfig.binfmt" config FORCE_MAX_ZONEORDER int depends on PPC64 + default "9" if PPC_64K_PAGES default "13" config MATH_EMULATION
arch/ppc64/Kconfig +1 -0
@@ -56,6 +56,7 @@ config PPC_STD_MMU # max order + 1 config FORCE_MAX_ZONEORDER int + default "9" if PPC_64K_PAGES default "13" source "init/Kconfig"
Quoted 7 lines · Mike Kravetz
Quoted 3 lines · Benjamin Herrenschmidt
Yes, the
MAX_ORDERshould be different indeed. But canKconfigdo that ? That is have the default value be different based on aKconfigoption ? I don't see that ... We may have to do things differently here...This seems to be done in other parts of the
Kconfigfile. Using those as an example, this should keep theMAX_ORDERblock size at 16MB.
Ok, I verified it does the right thing with Kconfig, thanks.
Paul, can you add to the merge tree too ?
Ben.
Quoted 6 lines · Mike Kravetz
Just curious if we still want to boost
MAX_ORDERlike this with 64k pages? Doesn't that make theMAX_ORDERblock size 256MB in this case? Also, not quite sure what happens if memory size (a 16 MB multiple) does not align with aMAX_ORDERblock size (a 256MB multiple in this case). My 'guess' is that the page allocator would not use it as it would not fit within the buddy system.
The buddy system and the SPARSEMEM mem_map are separate really. The key limitation is the a MAX_ORDER chunk must fit within the SPARSEMEM block size it cannot span two blocks. This is because the algorithm by which the buddy system finds buddies for a returning allocation assumes that mem_map is contigious upto the maximum buddy size (MAX_ORDER); it assumes it can use relative addressing to locate them.
The buddy system doesn't really care about the alignment of any of its blocks. The allocator is built empty and all existant pages are freed back to it. If there is a chunk of memory which can never coalesce back to MAX_ORDER it will simply sit lower in the tree 'waiting' for these non-existant buddies and will never merge. It will still be usable.
-apw
PATCH 1/4 revised Memory Add Fixes for ppc64
In reply to Benjamin HerrenschmidtQuoted 2 lines · Benjamin Herrenschmidt
This patch will have to be slightly reworked on top of the 64k pages one. It should be trivial though.
Here is a new version of the patch on top of 64k page support (actually 2.6.14-git10). One filename also changed due to more merge changes.
Add the create_section_mapping() routine to create hptes for memory sections dynamically added after system boot.
arch/powerpc/mm/hash_utils_64.c +9 -0
@@ -385,6 +385,15 @@ static unsigned long __init htab_get_tabreturn pteg_count <<7 ; } +#ifdef CONFIG_MEMORY_HOTPLUG +void create_section_mapping (unsigned long start,unsigned long end) +{ +BUG_ON (htab_bolt_mapping (start, end, start, + _PAGE_ACCESSED | _PAGE_DIRTY | _PAGE_COHERENT |PP_RWXX , + mmu_linear_psize)); +} +#endif /* CONFIG_MEMORY_HOTPLUG */ +void __init htab_initialize (void ) {unsigned long table, htab_size_bytes;
arch/powerpc/mm/mem.c +3 -0
@@ -127,6 +127,9 @@ int __devinit add_memory(u64 start, u64unsigned long start_pfn = start >>PAGE_SHIFT ;unsigned long nr_pages = size >>PAGE_SHIFT ; + start +=KERNELBASE ; +create_section_mapping (start, start + size); +/* this should work for most non-highmem platforms */ zone = pgdata->node_zones;
include/asm-powerpc/sparsemem.h +4 -0
@@ -11,6 +11,10 @@#define MAX_PHYSADDR_BITS 38#define MAX_PHYSMEM_BITS 36 +#ifdef CONFIG_MEMORY_HOTPLUG +extern void create_section_mapping (unsigned long start,unsigned long end); +#endif /* CONFIG_MEMORY_HOTPLUG */ +#endif /* CONFIG_SPARSEMEM */ #endif /* _ASM_POWERPC_SPARSEMEM_H */
Re: [PATCH 1/4] revised Memory Add Fixes for ppc64
Quoted 2 lines · Mike Kravetz
Here is a new version of the patch on top of 64k page support (actually 2.6.14-git10). One filename also changed due to more merge changes.
So, should I send this on to Linus along with the original 2/4 and 3/4 you posted and the revised 4/4?
Paul.
Re: [PATCH 1/4] revised Memory Add Fixes for ppc64
Quoted 2 lines · Paul Mackerras
So, should I send this on to Linus along with the original 2/4 and 3/4 you posted and the revised 4/4?
Yes, those should provide basic memory add support for ppc64.
Thanks,
Signature · 1 line
Mike
PATCH 2/4 Memory Add Fixes for ppc64
In reply to Mike Kravetzmemmap_init_zone() sets page count to 1. Before 'freeing' the page, we need to clear the count. This is the same that is done on free_all_bootmem_core() for memory discovered at boot time.
arch/powerpc/mm/mem.c +1 -0
@@ -107,6 +107,7 @@ EXPORT_SYMBOL(phys_mem_access_prot);void online_page (struct page *page) {ClearPageReserved (page); +set_page_count (page,0 );free_cold_page (page); totalram_pages++; num_physpages++;
PATCH 3/4 Memory Add Fixes for ppc64
In reply to Mike KravetzThis is a temporary kludge that supports adding all new memory to node 0. I will provide a more complete solution similar to that used for dynamically added CPUs in a few days.
include/asm-ppc64/mmzone.h +8 -0
@@ -33,6 +33,9 @@ extern int numa_cpu_lookup_table[];extern char *numa_memory_lookup_table;extern cpumask_t numa_cpumask_lookup_table[];extern int nr_cpus_in_node[]; +#ifdef CONFIG_MEMORY_HOTPLUG +extern unsigned long max_pfn; +#endif /* 16MB regions */ #define MEMORY_INCREMENT_SHIFT 24
@@ -45,6 +48,11 @@ static inline int pa_to_nid(unsigned lon {int nid; +#ifdef CONFIG_MEMORY_HOTPLUG +/* kludge hot added sections default to node 0 */ +if (pa >= (max_pfn <<PAGE_SHIFT )) +return 0 ; +#endif nid = numa_memory_lookup_table[pa >>MEMORY_INCREMENT_SHIFT ];#ifdef DEBUG_NUMA
PATCH 4/4 Memory Add Fixes for ppc64
In reply to Mike Kravetzppc64 needs a special sysfs probe file for adding new memory.
arch/ppc64/Kconfig +4 -0
@@ -277,6 +277,10 @@ config HAVE_ARCH_EARLY_PFN_TO_NID def_bool y depends on NEED_MULTIPLE_NODES +config ARCH_MEMORY_PROBE + def_bool y + depends on MEMORY_HOTPLUG + # Some NUMA nodes have memory ranges that span # other nodes. Even though a pfn is valid and # between a node's start and end pfns, it may not
Quoted 8 lines · Mike Kravetz
ppc64 needs a special sysfs probe file for adding new memory.
Signed-off-by Mike Kravetz <[email protected]>diff -Naupr linux-2.6.14-git7/arch/ppc64/Kconfig linux-2.6.14-git7.work/arch/ppc64/Kconfig
arch/ppc64/Kconfig config HAVE_ARCH_EARLY_PFN_TO_NID · line 283
Quoted 7 lines
arch/ppc64/Kconfig +4
-0@@ -277,6 +277,10 @@ config HAVE_ARCH_EARLY_PFN_TO_NID def_bool y depends on NEED_MULTIPLE_NODES +config ARCH_MEMORY_PROBE + def_bool y + depends on MEMORY_HOTPLUG +
Does arch/powerpc/Kconfig need a similar fix then?
Paul.
Quoted 4 lines · Paul Mackerras
Quoted 1 line · Mike Kravetz
ppc64 needs a special sysfs probe file for adding new memory.
Does
arch/powerpc/Kconfigneed a similar fix then?
Yes it does. Sorry, I haven't been paying as much attention to the merge as I should. Here is a new version.
ppc64 needs a special sysfs probe file for adding new memory.
arch/powerpc/Kconfig +4 -0
@@ -569,6 +569,10 @@ config HAVE_ARCH_EARLY_PFN_TO_NID def_bool y depends on NEED_MULTIPLE_NODES +config ARCH_MEMORY_PROBE + def_bool y + depends on MEMORY_HOTPLUG + # Some NUMA nodes have memory ranges that span # other nodes. Even though a pfn is valid and # between a node's start and end pfns, it may not
arch/ppc64/Kconfig +4 -0
@@ -277,6 +277,10 @@ config HAVE_ARCH_EARLY_PFN_TO_NID def_bool y depends on NEED_MULTIPLE_NODES +config ARCH_MEMORY_PROBE + def_bool y + depends on MEMORY_HOTPLUG + # Some NUMA nodes have memory ranges that span # other nodes. Even though a pfn is valid and # between a node's start and end pfns, it may not
Hi Mike,
Quoted 9 lines
When memory add was merged into mainline in 2.6.14, there were various bits and pieces missing that prevent it from working on ppc64. The following patches are against 2.6.14-git7 and address all but one of the know issues.
1) Create hptes for new sections
2) Clear page count before freeing new pages
3) Kludge to add new memory to node 0
4) Ensure probe file is created for memory add via sysfs
Ive got a patch that reworks our numa code and it might reject with your stuff. I'll send them out for review this afternoon.
Anton
Quoted 2 lines · Anton Blanchard
Ive got a patch that reworks our numa code and it might reject with your stuff. I'll send them out for review this afternoon.
Interesting in that I was going to start reworking some of the numa code to make it play nice with hot add. Doubt this patch set will impact your changes. This set is not very intelligent WRT numa and doesn't really modify any of the real code.
Signature · 1 line
Mike