Use sparsemem as the only memory model for UP, SMP and NUMA.
Measurements indicate that DISCONTIGMEM has a higher
overhead than sparsemem. And FLATMEMs benefits are minimal. So I think its
best to simply standardize on sparsemem.
Results of page allocator tests (test can be had via git from the slab git
tree on git.kernel.org. See the branch tests)
Measurements in cycle counts. 1000 allocations were performed and then the
average cycle count was calculated.
Order FlatMem Discontig SparseMem
0 639 665 641
1 567 647 593
2 679 774 692
3 763 967 781
4 961 1501 962
5 1356 2344 1392
6 2224 3982 2336
7 4869 7225 5074
8 12500 14048 12732
9 27926 28223 28165
10 58578 58714 58682
If this patch is accepted then we can remove the code for discontig and
flatmem support from x86_64.
(Not sure if I got all the config settings right. Andy?)
Signed-off-by: Christoph Lameter <[email protected]>
---
arch/x86/Kconfig.x86_64 | 13 +------------
1 file changed, 1 insertion(+), 12 deletions(-)
Index: linux-2.6/arch/x86/Kconfig.x86_64
===================================================================
--- linux-2.6.orig/arch/x86/Kconfig.x86_64 2007-11-12 15:17:50.721767735 -0800
+++ linux-2.6/arch/x86/Kconfig.x86_64 2007-11-12 15:21:16.659017509 -0800
@@ -390,28 +390,17 @@ config NUMA_EMU
into virtual nodes when booted with "numa=fake=N", where N is the
number of nodes. This is only useful for debugging.
-config ARCH_DISCONTIGMEM_ENABLE
- bool
- depends on NUMA
- default y
-
-config ARCH_DISCONTIGMEM_DEFAULT
+config ARCH_SPARSEMEM_DEFAULT
def_bool y
- depends on NUMA
config ARCH_SPARSEMEM_ENABLE
def_bool y
- depends on (NUMA || EXPERIMENTAL)
select SPARSEMEM_VMEMMAP_ENABLE
config ARCH_MEMORY_PROBE
def_bool y
depends on MEMORY_HOTPLUG
-config ARCH_FLATMEM_ENABLE
- def_bool y
- depends on !NUMA
-
source "mm/Kconfig"
config MEMORY_HOTPLUG_RESERVE
On Tuesday 13 November 2007 00:52:14 Christoph Lameter wrote:
> Use sparsemem as the only memory model for UP, SMP and NUMA.
>
> Measurements indicate that DISCONTIGMEM has a higher
> overhead than sparsemem. And FLATMEMs benefits are minimal. So I think its
> best to simply standardize on sparsemem.
How about the memory overhead? Is it the same too?
And code size vs flatmem?
-Andi
On Tue, 13 Nov 2007, Andi Kleen wrote:
> On Tuesday 13 November 2007 00:52:14 Christoph Lameter wrote:
> > Use sparsemem as the only memory model for UP, SMP and NUMA.
> >
> > Measurements indicate that DISCONTIGMEM has a higher
> > overhead than sparsemem. And FLATMEMs benefits are minimal. So I think its
> > best to simply standardize on sparsemem.
>
> How about the memory overhead? Is it the same too?
> And code size vs flatmem?
SMP Sparsemem
-------------
Kernel size:
text data bss dec hex filename
3849268 397739 1264856 5511863 541ab7 vmlinux
total used free shared buffers cached
Mem: 8242252 41164 8201088 0 352 11512
-/+ buffers/cache: 29300 8212952
Swap: 9775512 0 9775512
SMP Flatmem
-----------
Kernel size:
text data bss dec hex filename
3844612 397739 1264536 5506887 540747 vmlinux
So 4.5k growth in text size vs. FLATMEM.
total used free shared buffers cached
Mem: 8244052 40544 8203508 0 352 11484
-/+ buffers/cache: 28708 8215344
2k growth in overall memory use after boot.
NUMA discontig:
text data bss dec hex filename
3888124 470659 1276504 5635287 55fcd7 vmlinux
total used free shared buffers cached
Mem: 8256256 56908 8199348 0 352 11496
-/+ buffers/cache: 45060 8211196
Swap: 9775512 0 9775512
NUMA sparse:
text data bss dec hex filename
3896428 470659 1276824 5643911 561e87 vmlinux
8k text growth. Given that we fully inline virt_to_page and friends now
that is rather good.
total used free shared buffers cached
Mem: 8264720 57240 8207480 0 352 11516
-/+ buffers/cache: 45372 8219348
Swap: 9775512 0 9775512
Hmmm... More memory free? How did that happen? More pages cached for some
reason. The total available memory is increased by 8k.
> Hmmm... More memory free? How did that happen? More pages cached for some
> reason. The total available memory is increased by 8k.
Nice. Looks all reasonable. Thanks for the numbers.
-Andi
On Tue, 13 Nov 2007, Andi Kleen wrote:
>
> > Hmmm... More memory free? How did that happen? More pages cached for some
> > reason. The total available memory is increased by 8k.
>
> Nice. Looks all reasonable. Thanks for the numbers.
Ok here is the patch to remove DISCONTIG and FLATMEM
x86_64: Make sparsemem/vmemmap the default memory model
Use sparsemem as the only memory model for UP, SMP and NUMA.
Measurements indicate that DISCONTIGMEM has a higher overhead
than sparsemem. And FLATMEMs benefits are minimal. So I think its
best to simply standardize on sparsemem.
Results of page allocator tests (test can be had via git from slab git
tree branch tests)
Measurements in cycle counts. 1000 allocations were performed and then the
average cycle count was calculated.
Order FlatMem Discontig SparseMem
0 639 665 641
1 567 647 593
2 679 774 692
3 763 967 781
4 961 1501 962
5 1356 2344 1392
6 2224 3982 2336
7 4869 7225 5074
8 12500 14048 12732
9 27926 28223 28165
10 58578 58714 58682
Memory use:
SMP Sparsemem
-------------
Kernel size:
text data bss dec hex filename
3849268 397739 1264856 5511863 541ab7 vmlinux
total used free shared buffers cached
Mem: 8242252 41164 8201088 0 352 11512
-/+ buffers/cache: 29300 8212952
Swap: 9775512 0 9775512
SMP Flatmem
-----------
Kernel size:
text data bss dec hex filename
3844612 397739 1264536 5506887 540747 vmlinux
So 4.5k growth in text size vs. FLATMEM.
total used free shared buffers cached
Mem: 8244052 40544 8203508 0 352 11484
-/+ buffers/cache: 28708 8215344
2k growth in overall memory use after boot.
NUMA discontig:
text data bss dec hex filename
3888124 470659 1276504 5635287 55fcd7 vmlinux
total used free shared buffers cached
Mem: 8256256 56908 8199348 0 352 11496
-/+ buffers/cache: 45060 8211196
Swap: 9775512 0 9775512
NUMA sparse:
text data bss dec hex filename
3896428 470659 1276824 5643911 561e87 vmlinux
8k text growth. Given that we fully inline virt_to_page and friends now
that is rather good.
total used free shared buffers cached
Mem: 8264720 57240 8207480 0 352 11516
-/+ buffers/cache: 45372 8219348
Swap: 9775512 0 9775512
The total available memory is increased by 8k.
This patch makes sparsemem the default and removes discontig and
flatmem support from x86.
Acked-by: Andi Kleen <[email protected]>
Signed-off-by: Christoph Lameter <[email protected]>
---
arch/x86/Kconfig.x86_64 | 21 -------------
arch/x86/configs/x86_64_defconfig | 9 -----
arch/x86/kernel/machine_kexec_64.c | 5 ---
arch/x86/mm/init_64.c | 30 -------------------
arch/x86/mm/ioremap_64.c | 17 -----------
arch/x86/mm/numa_64.c | 21 -------------
arch/x86/mm/srat_64.c | 57 -------------------------------------
include/asm-x86/mmzone_64.h | 6 ---
include/asm-x86/page_64.h | 3 -
9 files changed, 1 insertion(+), 168 deletions(-)
Index: linux-2.6/arch/x86/Kconfig.x86_64
===================================================================
--- linux-2.6.orig/arch/x86/Kconfig.x86_64 2007-11-12 19:10:56.697664127 -0800
+++ linux-2.6/arch/x86/Kconfig.x86_64 2007-11-12 19:14:02.437902563 -0800
@@ -411,42 +411,23 @@ config NUMA_EMU
into virtual nodes when booted with "numa=fake=N", where N is the
number of nodes. This is only useful for debugging.
-config ARCH_DISCONTIGMEM_ENABLE
- bool
- depends on NUMA
- default y
-
-config ARCH_DISCONTIGMEM_DEFAULT
+config ARCH_SPARSEMEM_DEFAULT
def_bool y
- depends on NUMA
config ARCH_SPARSEMEM_ENABLE
def_bool y
- depends on (NUMA || EXPERIMENTAL)
select SPARSEMEM_VMEMMAP_ENABLE
config ARCH_MEMORY_PROBE
def_bool y
depends on MEMORY_HOTPLUG
-config ARCH_FLATMEM_ENABLE
- def_bool y
- depends on !NUMA
-
source "mm/Kconfig"
-config MEMORY_HOTPLUG_RESERVE
- def_bool y
- depends on (MEMORY_HOTPLUG && DISCONTIGMEM)
-
config HAVE_ARCH_EARLY_PFN_TO_NID
def_bool y
depends on NUMA
-config OUT_OF_LINE_PFN_TO_PAGE
- def_bool y
- depends on DISCONTIGMEM
-
config NR_CPUS
int "Maximum number of CPUs (2-255)"
range 2 255
Index: linux-2.6/arch/x86/configs/x86_64_defconfig
===================================================================
--- linux-2.6.orig/arch/x86/configs/x86_64_defconfig 2007-11-12 18:27:12.516523443 -0800
+++ linux-2.6/arch/x86/configs/x86_64_defconfig 2007-11-12 19:14:02.441902438 -0800
@@ -145,15 +145,6 @@ CONFIG_K8_NUMA=y
CONFIG_NODES_SHIFT=6
CONFIG_X86_64_ACPI_NUMA=y
CONFIG_NUMA_EMU=y
-CONFIG_ARCH_DISCONTIGMEM_ENABLE=y
-CONFIG_ARCH_DISCONTIGMEM_DEFAULT=y
-CONFIG_ARCH_SPARSEMEM_ENABLE=y
-CONFIG_SELECT_MEMORY_MODEL=y
-# CONFIG_FLATMEM_MANUAL is not set
-CONFIG_DISCONTIGMEM_MANUAL=y
-# CONFIG_SPARSEMEM_MANUAL is not set
-CONFIG_DISCONTIGMEM=y
-CONFIG_FLAT_NODE_MEM_MAP=y
CONFIG_NEED_MULTIPLE_NODES=y
# CONFIG_SPARSEMEM_STATIC is not set
CONFIG_SPLIT_PTLOCK_CPUS=4
Index: linux-2.6/arch/x86/kernel/machine_kexec_64.c
===================================================================
--- linux-2.6.orig/arch/x86/kernel/machine_kexec_64.c 2007-11-12 18:27:12.524523631 -0800
+++ linux-2.6/arch/x86/kernel/machine_kexec_64.c 2007-11-12 19:14:02.453902454 -0800
@@ -234,10 +234,5 @@ NORET_TYPE void machine_kexec(struct kim
void arch_crash_save_vmcoreinfo(void)
{
VMCOREINFO_SYMBOL(init_level4_pgt);
-
-#ifdef CONFIG_ARCH_DISCONTIGMEM_ENABLE
- VMCOREINFO_SYMBOL(node_data);
- VMCOREINFO_LENGTH(node_data, MAX_NUMNODES);
-#endif
}
Index: linux-2.6/arch/x86/mm/ioremap_64.c
===================================================================
--- linux-2.6.orig/arch/x86/mm/ioremap_64.c 2007-11-12 18:27:12.552523359 -0800
+++ linux-2.6/arch/x86/mm/ioremap_64.c 2007-11-12 19:14:02.461902441 -0800
@@ -86,23 +86,6 @@ void __iomem * __ioremap(unsigned long p
if (phys_addr >= ISA_START_ADDRESS && last_addr < ISA_END_ADDRESS)
return (__force void __iomem *)phys_to_virt(phys_addr);
-#ifdef CONFIG_FLATMEM
- /*
- * Don't allow anybody to remap normal RAM that we're using..
- */
- if (last_addr < virt_to_phys(high_memory)) {
- char *t_addr, *t_end;
- struct page *page;
-
- t_addr = __va(phys_addr);
- t_end = t_addr + (size - 1);
-
- for(page = virt_to_page(t_addr); page <= virt_to_page(t_end); page++)
- if(!PageReserved(page))
- return NULL;
- }
-#endif
-
pgprot = __pgprot(_PAGE_PRESENT | _PAGE_RW | _PAGE_GLOBAL
| _PAGE_DIRTY | _PAGE_ACCESSED | flags);
/*
Index: linux-2.6/arch/x86/mm/numa_64.c
===================================================================
--- linux-2.6.orig/arch/x86/mm/numa_64.c 2007-11-12 18:27:12.528523345 -0800
+++ linux-2.6/arch/x86/mm/numa_64.c 2007-11-12 19:24:34.601056905 -0800
@@ -149,12 +149,10 @@ int __init compute_hash_shift(struct boo
return shift;
}
-#ifdef CONFIG_SPARSEMEM
int early_pfn_to_nid(unsigned long pfn)
{
return phys_to_nid(pfn << PAGE_SHIFT);
}
-#endif
static void * __init
early_node_mem(int nodeid, unsigned long start, unsigned long end,
@@ -626,23 +624,4 @@ EXPORT_SYMBOL(node_to_cpumask);
EXPORT_SYMBOL(memnode);
EXPORT_SYMBOL(node_data);
-#ifdef CONFIG_DISCONTIGMEM
-/*
- * Functions to convert PFNs from/to per node page addresses.
- * These are out of line because they are quite big.
- * They could be all tuned by pre caching more state.
- * Should do that.
- */
-int pfn_valid(unsigned long pfn)
-{
- unsigned nid;
- if (pfn >= num_physpages)
- return 0;
- nid = pfn_to_nid(pfn);
- if (nid == 0xff)
- return 0;
- return pfn >= node_start_pfn(nid) && (pfn) < node_end_pfn(nid);
-}
-EXPORT_SYMBOL(pfn_valid);
-#endif
Index: linux-2.6/include/asm-x86/mmzone_64.h
===================================================================
--- linux-2.6.orig/include/asm-x86/mmzone_64.h 2007-11-12 18:27:12.560523466 -0800
+++ linux-2.6/include/asm-x86/mmzone_64.h 2007-11-12 19:14:02.485902361 -0800
@@ -41,12 +41,6 @@ static inline __attribute__((pure)) int
#define node_end_pfn(nid) (NODE_DATA(nid)->node_start_pfn + \
NODE_DATA(nid)->node_spanned_pages)
-#ifdef CONFIG_DISCONTIGMEM
-#define pfn_to_nid(pfn) phys_to_nid((unsigned long)(pfn) << PAGE_SHIFT)
-
-extern int pfn_valid(unsigned long pfn);
-#endif
-
#ifdef CONFIG_NUMA_EMU
#define FAKE_NODE_MIN_SIZE (64*1024*1024)
#define FAKE_NODE_MIN_HASH_MASK (~(FAKE_NODE_MIN_SIZE - 1uL))
Index: linux-2.6/include/asm-x86/page_64.h
===================================================================
--- linux-2.6.orig/include/asm-x86/page_64.h 2007-11-12 18:27:12.568523834 -0800
+++ linux-2.6/include/asm-x86/page_64.h 2007-11-12 19:14:02.497902025 -0800
@@ -121,9 +121,6 @@ extern unsigned long __phys_addr(unsigne
#define __va(x) ((void *)((unsigned long)(x)+PAGE_OFFSET))
#define __boot_va(x) __va(x)
#define __boot_pa(x) __pa(x)
-#ifdef CONFIG_FLATMEM
-#define pfn_valid(pfn) ((pfn) < end_pfn)
-#endif
#define virt_to_page(kaddr) pfn_to_page(__pa(kaddr) >> PAGE_SHIFT)
#define virt_addr_valid(kaddr) pfn_valid(__pa(kaddr) >> PAGE_SHIFT)
Index: linux-2.6/arch/x86/mm/init_64.c
===================================================================
--- linux-2.6.orig/arch/x86/mm/init_64.c 2007-11-12 19:24:08.016577199 -0800
+++ linux-2.6/arch/x86/mm/init_64.c 2007-11-12 19:29:48.689827052 -0800
@@ -485,34 +485,6 @@ EXPORT_SYMBOL_GPL(memory_add_physaddr_to
#endif /* CONFIG_MEMORY_HOTPLUG */
-#ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
-/*
- * Memory Hotadd without sparsemem. The mem_maps have been allocated in advance,
- * just online the pages.
- */
-int __add_pages(struct zone *z, unsigned long start_pfn, unsigned long nr_pages)
-{
- int err = -EIO;
- unsigned long pfn;
- unsigned long total = 0, mem = 0;
- for (pfn = start_pfn; pfn < start_pfn + nr_pages; pfn++) {
- if (pfn_valid(pfn)) {
- online_page(pfn_to_page(pfn));
- err = 0;
- mem++;
- }
- total++;
- }
- if (!err) {
- z->spanned_pages += total;
- z->present_pages += mem;
- z->zone_pgdat->node_spanned_pages += total;
- z->zone_pgdat->node_present_pages += mem;
- }
- return err;
-}
-#endif
-
static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel, kcore_modules,
kcore_vsyscall;
@@ -738,7 +710,6 @@ const char *arch_vma_name(struct vm_area
return NULL;
}
-#ifdef CONFIG_SPARSEMEM_VMEMMAP
/*
* Initialise the sparsemem vmemmap using huge-pages at the PMD level.
*/
@@ -781,7 +752,6 @@ int __meminit vmemmap_populate(struct pa
return 0;
}
-#endif
#ifdef CONFIG_NUMA
int __meminit cpu_area_populate(void *start, unsigned long size,
Index: linux-2.6/arch/x86/mm/srat_64.c
===================================================================
--- linux-2.6.orig/arch/x86/mm/srat_64.c 2007-11-12 19:24:54.344327144 -0800
+++ linux-2.6/arch/x86/mm/srat_64.c 2007-11-12 19:25:36.108577445 -0800
@@ -151,62 +151,6 @@ acpi_numa_processor_affinity_init(struct
pxm, pa->apic_id, node);
}
-#ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
-/*
- * Protect against too large hotadd areas that would fill up memory.
- */
-static int hotadd_enough_memory(struct bootnode *nd)
-{
- static unsigned long allocated;
- static unsigned long last_area_end;
- unsigned long pages = (nd->end - nd->start) >> PAGE_SHIFT;
- long mem = pages * sizeof(struct page);
- unsigned long addr;
- unsigned long allowed;
- unsigned long oldpages = pages;
-
- if (mem < 0)
- return 0;
- allowed = (end_pfn - absent_pages_in_range(0, end_pfn)) * PAGE_SIZE;
- allowed = (allowed / 100) * hotadd_percent;
- if (allocated + mem > allowed) {
- unsigned long range;
- /* Give them at least part of their hotadd memory upto hotadd_percent
- It would be better to spread the limit out
- over multiple hotplug areas, but that is too complicated
- right now */
- if (allocated >= allowed)
- return 0;
- range = allowed - allocated;
- pages = (range / PAGE_SIZE);
- mem = pages * sizeof(struct page);
- nd->end = nd->start + range;
- }
- /* Not completely fool proof, but a good sanity check */
- addr = find_e820_area(last_area_end, end_pfn<<PAGE_SHIFT, mem);
- if (addr == -1UL)
- return 0;
- if (pages != oldpages)
- printk(KERN_NOTICE "SRAT: Hotadd area limited to %lu bytes\n",
- pages << PAGE_SHIFT);
- last_area_end = addr + mem;
- allocated += mem;
- return 1;
-}
-
-static int update_end_of_memory(unsigned long end)
-{
- found_add_area = 1;
- if ((end >> PAGE_SHIFT) > end_pfn)
- end_pfn = end >> PAGE_SHIFT;
- return 1;
-}
-
-static inline int save_add_info(void)
-{
- return hotadd_percent > 0;
-}
-#else
int update_end_of_memory(unsigned long end) {return -1;}
static int hotadd_enough_memory(struct bootnode *nd) {return 1;}
#ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
@@ -214,7 +158,6 @@ static inline int save_add_info(void) {r
#else
static inline int save_add_info(void) {return 0;}
#endif
-#endif
/*
* Update nodes_add and decide if to include add are in the zone.
* Both SPARSE and RESERVE need nodes_add infomation.
On Nov 12, 2007 7:42 PM, Christoph Lameter <[email protected]> wrote:
> Ok here is the patch to remove DISCONTIG and FLATMEM
>
> x86_64: Make sparsemem/vmemmap the default memory model
>
> Use sparsemem as the only memory model for UP, SMP and NUMA.
> Measurements indicate that DISCONTIGMEM has a higher overhead
> than sparsemem. And FLATMEMs benefits are minimal. So I think its
> best to simply standardize on sparsemem.
>
> Results of page allocator tests (test can be had via git from slab git
> tree branch tests)
>
> Measurements in cycle counts. 1000 allocations were performed and then the
> average cycle count was calculated.
>
> Order FlatMem Discontig SparseMem
> 0 639 665 641
> 1 567 647 593
> 2 679 774 692
> 3 763 967 781
> 4 961 1501 962
> 5 1356 2344 1392
> 6 2224 3982 2336
> 7 4869 7225 5074
> 8 12500 14048 12732
> 9 27926 28223 28165
> 10 58578 58714 58682
Discontig obviously needs to die. However, FlatMem is consistently
faster, averaging about 2.1% better overall for your numbers above. Is
the page allocator not, erm, a fast path, where that matters?
Order Flat Sparse % diff
0 639 641 0.3
1 567 593 4.4
2 679 692 1.9
3 763 781 2.3
4 961 962 0.1
5 1356 1392 2.6
6 2224 2336 4.8
7 4869 5074 4.0
8 12500 12732 1.8
9 27926 28165 0.8
10 58578 58682 0.2
On Mon, 12 Nov 2007, Ray Lee wrote:
> Discontig obviously needs to die. However, FlatMem is consistently
> faster, averaging about 2.1% better overall for your numbers above. Is
> the page allocator not, erm, a fast path, where that matters?
>
> Order Flat Sparse % diff
> 0 639 641 0.3
IMHO Order 0 currently matters most and the difference is negligible
there.
On Mon, 12 November 2007 20:41:10 -0800, Christoph Lameter wrote:
> On Mon, 12 Nov 2007, Ray Lee wrote:
>
> > Discontig obviously needs to die. However, FlatMem is consistently
> > faster, averaging about 2.1% better overall for your numbers above. Is
> > the page allocator not, erm, a fast path, where that matters?
> >
> > Order Flat Sparse % diff
> > 0 639 641 0.3
>
> IMHO Order 0 currently matters most and the difference is negligible
> there.
Is it? I am a bit concerned about the non-monotonic distribution.
Difference starts a near-0, grows to 4.4, drops to near-0, grows to 4.9,
drops to near-0.
Order Flat Sparse % diff
0 639 641 0.3
1 567 593 4.4
2 679 692 1.9
3 763 781 2.3
4 961 962 0.1
5 1356 1392 2.6
6 2224 2336 4.8
7 4869 5074 4.0
8 12500 12732 1.8
9 27926 28165 0.8
10 58578 58682 0.2
Is there an explanation for this behaviour? More to the point, could
repeated runs also return 4% difference for order-0?
Jörn
--
It does not require a majority to prevail, but rather an irate,
tireless minority keen to set brush fires in people's minds.
-- Samuel Adams
On Tue, 13 Nov 2007, J?rn Engel wrote:
> On Mon, 12 November 2007 20:41:10 -0800, Christoph Lameter wrote:
> > On Mon, 12 Nov 2007, Ray Lee wrote:
> >
> > > Discontig obviously needs to die. However, FlatMem is consistently
> > > faster, averaging about 2.1% better overall for your numbers above. Is
> > > the page allocator not, erm, a fast path, where that matters?
> > >
> > > Order Flat Sparse % diff
> > > 0 639 641 0.3
> >
> > IMHO Order 0 currently matters most and the difference is negligible
> > there.
>
> Is it? I am a bit concerned about the non-monotonic distribution.
> Difference starts a near-0, grows to 4.4, drops to near-0, grows to 4.9,
> drops to near-0.
The problem also is that the comparison here is between a SMP config for
flatmem vs a NUMA config for sparsemem. There is additional overhead in
the NUMA config.
The effect may also be due to the system being able to place
some pages in the same 2MB section as the memmap with flatmem. However,
that is only feasable immeidately after bootup. In regular operations this
should vanish.
Could you run your own test to verify?
> Is there an explanation for this behaviour? More to the point, could
> repeated runs also return 4% difference for order-0?
I hope I have given some above. The number of the page allocator suggests
that we have far too much fat in the allocation paths. IMHO reasonable
numbers for an order-0 alloc should be ~100 cycles.
On Tue, 13 November 2007 13:52:17 -0800, Christoph Lameter wrote:
>
> Could you run your own test to verify?
You bastard! You know I'm too lazy to do that. ;)
As long as the order-0 number is stable across multiple runs I don't
mind. The numbers just looked suspiciously as if they were not stable.
That's all.
Jörn
--
Why do musicians compose symphonies and poets write poems?
They do it because life wouldn't have any meaning for them if they didn't.
That's why I draw cartoons. It's my life.
-- Charles Shultz
On Mon, 12 Nov 2007 19:42:31 -0800 (PST)
Christoph Lameter <[email protected]> wrote:
> x86_64: Make sparsemem/vmemmap the default memory model
>
> Use sparsemem as the only memory model for UP, SMP and NUMA.
> Measurements indicate that DISCONTIGMEM has a higher overhead
> than sparsemem. And FLATMEMs benefits are minimal. So I think its
> best to simply standardize on sparsemem.
Unfortunately some loon has gone and merged the i386 and x86_64 Kconfig
files. I was fixing that up but I worry what effects these Kconfig changes
might have on, for example, i386 NUMA setups.
So I'll duck this version, sorry.
On Thu, 15 Nov 2007, Andrew Morton wrote:
> Unfortunately some loon has gone and merged the i386 and x86_64 Kconfig
> files. I was fixing that up but I worry what effects these Kconfig changes
> might have on, for example, i386 NUMA setups.
>
> So I'll duck this version, sorry.
Is there a tree that I can rediff the patch against?
On Thu, 15 Nov 2007 18:24:54 -0800 (PST) Christoph Lameter <[email protected]> wrote:
> On Thu, 15 Nov 2007, Andrew Morton wrote:
>
> > Unfortunately some loon has gone and merged the i386 and x86_64 Kconfig
> > files. I was fixing that up but I worry what effects these Kconfig changes
> > might have on, for example, i386 NUMA setups.
> >
> > So I'll duck this version, sorry.
>
> Is there a tree that I can rediff the patch against?
Mainline.
Updated version against mainline. Looks even a bit cleaner.
x86_64: Make sparsemem/vmemmap the default memory model V2
V1->V2:
- Rediff against new upstream x86 code that unifies the Kconfig files.
Use sparsemem as the only memory model for UP, SMP and NUMA.
Measurements indicate that DISCONTIGMEM has a higher overhead
than sparsemem. And FLATMEMs benefits are minimal. So I think its
best to simply standardize on sparsemem.
Results of page allocator tests (test can be had via git from slab git
tree branch tests)
Measurements in cycle counts. 1000 allocations were performed and then the
average cycle count was calculated.
Order FlatMem Discontig SparseMem
0 639 665 641
1 567 647 593
2 679 774 692
3 763 967 781
4 961 1501 962
5 1356 2344 1392
6 2224 3982 2336
7 4869 7225 5074
8 12500 14048 12732
9 27926 28223 28165
10 58578 58714 58682
(Note that FlatMem is an SMP config and the rest NUMA configurations)
Memory use:
SMP Sparsemem
-------------
Kernel size:
text data bss dec hex filename
3849268 397739 1264856 5511863 541ab7 vmlinux
total used free shared buffers cached
Mem: 8242252 41164 8201088 0 352 11512
-/+ buffers/cache: 29300 8212952
Swap: 9775512 0 9775512
SMP Flatmem
-----------
Kernel size:
text data bss dec hex filename
3844612 397739 1264536 5506887 540747 vmlinux
So 4.5k growth in text size vs. FLATMEM.
total used free shared buffers cached
Mem: 8244052 40544 8203508 0 352 11484
-/+ buffers/cache: 28708 8215344
2k growth in overall memory use after boot.
NUMA discontig:
text data bss dec hex filename
3888124 470659 1276504 5635287 55fcd7 vmlinux
total used free shared buffers cached
Mem: 8256256 56908 8199348 0 352 11496
-/+ buffers/cache: 45060 8211196
Swap: 9775512 0 9775512
NUMA sparse:
text data bss dec hex filename
3896428 470659 1276824 5643911 561e87 vmlinux
8k text growth. Given that we fully inline virt_to_page and friends now
that is rather good.
total used free shared buffers cached
Mem: 8264720 57240 8207480 0 352 11516
-/+ buffers/cache: 45372 8219348
Swap: 9775512 0 9775512
The total available memory is increased by 8k.
This patch makes sparsemem the default and removes discontig and
flatmem support from x86.
Acked-by: Andi Kleen <[email protected]>
Signed-off-by: Christoph Lameter <[email protected]>
---
arch/x86/Kconfig | 20 +++++-------
arch/x86/configs/x86_64_defconfig | 9 -----
arch/x86/kernel/machine_kexec_64.c | 5 ---
arch/x86/mm/init_64.c | 30 -------------------
arch/x86/mm/ioremap_64.c | 17 -----------
arch/x86/mm/numa_64.c | 21 -------------
arch/x86/mm/srat_64.c | 57 -------------------------------------
include/asm-x86/mmzone_64.h | 6 ---
include/asm-x86/page_64.h | 3 -
include/asm-x86/sparsemem_64.h | 4 --
10 files changed, 8 insertions(+), 164 deletions(-)
Index: linux-2.6/arch/x86/configs/x86_64_defconfig
===================================================================
--- linux-2.6.orig/arch/x86/configs/x86_64_defconfig 2007-11-13 11:52:34.785633363 -0800
+++ linux-2.6/arch/x86/configs/x86_64_defconfig 2007-11-15 19:39:59.073745644 -0800
@@ -145,15 +145,6 @@ CONFIG_K8_NUMA=y
CONFIG_NODES_SHIFT=6
CONFIG_X86_64_ACPI_NUMA=y
CONFIG_NUMA_EMU=y
-CONFIG_ARCH_DISCONTIGMEM_ENABLE=y
-CONFIG_ARCH_DISCONTIGMEM_DEFAULT=y
-CONFIG_ARCH_SPARSEMEM_ENABLE=y
-CONFIG_SELECT_MEMORY_MODEL=y
-# CONFIG_FLATMEM_MANUAL is not set
-CONFIG_DISCONTIGMEM_MANUAL=y
-# CONFIG_SPARSEMEM_MANUAL is not set
-CONFIG_DISCONTIGMEM=y
-CONFIG_FLAT_NODE_MEM_MAP=y
CONFIG_NEED_MULTIPLE_NODES=y
# CONFIG_SPARSEMEM_STATIC is not set
CONFIG_SPLIT_PTLOCK_CPUS=4
Index: linux-2.6/arch/x86/kernel/machine_kexec_64.c
===================================================================
--- linux-2.6.orig/arch/x86/kernel/machine_kexec_64.c 2007-11-13 11:52:34.793633597 -0800
+++ linux-2.6/arch/x86/kernel/machine_kexec_64.c 2007-11-15 19:39:59.126128308 -0800
@@ -234,10 +234,5 @@ NORET_TYPE void machine_kexec(struct kim
void arch_crash_save_vmcoreinfo(void)
{
VMCOREINFO_SYMBOL(init_level4_pgt);
-
-#ifdef CONFIG_ARCH_DISCONTIGMEM_ENABLE
- VMCOREINFO_SYMBOL(node_data);
- VMCOREINFO_LENGTH(node_data, MAX_NUMNODES);
-#endif
}
Index: linux-2.6/arch/x86/mm/ioremap_64.c
===================================================================
--- linux-2.6.orig/arch/x86/mm/ioremap_64.c 2007-11-13 11:52:34.801633088 -0800
+++ linux-2.6/arch/x86/mm/ioremap_64.c 2007-11-15 19:39:59.149995698 -0800
@@ -86,23 +86,6 @@ void __iomem * __ioremap(unsigned long p
if (phys_addr >= ISA_START_ADDRESS && last_addr < ISA_END_ADDRESS)
return (__force void __iomem *)phys_to_virt(phys_addr);
-#ifdef CONFIG_FLATMEM
- /*
- * Don't allow anybody to remap normal RAM that we're using..
- */
- if (last_addr < virt_to_phys(high_memory)) {
- char *t_addr, *t_end;
- struct page *page;
-
- t_addr = __va(phys_addr);
- t_end = t_addr + (size - 1);
-
- for(page = virt_to_page(t_addr); page <= virt_to_page(t_end); page++)
- if(!PageReserved(page))
- return NULL;
- }
-#endif
-
pgprot = __pgprot(_PAGE_PRESENT | _PAGE_RW | _PAGE_GLOBAL
| _PAGE_DIRTY | _PAGE_ACCESSED | flags);
/*
Index: linux-2.6/arch/x86/mm/numa_64.c
===================================================================
--- linux-2.6.orig/arch/x86/mm/numa_64.c 2007-11-13 11:52:34.805632772 -0800
+++ linux-2.6/arch/x86/mm/numa_64.c 2007-11-15 19:39:59.157245106 -0800
@@ -149,12 +149,10 @@ int __init compute_hash_shift(struct boo
return shift;
}
-#ifdef CONFIG_SPARSEMEM
int early_pfn_to_nid(unsigned long pfn)
{
return phys_to_nid(pfn << PAGE_SHIFT);
}
-#endif
static void * __init
early_node_mem(int nodeid, unsigned long start, unsigned long end,
@@ -626,23 +624,4 @@ EXPORT_SYMBOL(node_to_cpumask);
EXPORT_SYMBOL(memnode);
EXPORT_SYMBOL(node_data);
-#ifdef CONFIG_DISCONTIGMEM
-/*
- * Functions to convert PFNs from/to per node page addresses.
- * These are out of line because they are quite big.
- * They could be all tuned by pre caching more state.
- * Should do that.
- */
-int pfn_valid(unsigned long pfn)
-{
- unsigned nid;
- if (pfn >= num_physpages)
- return 0;
- nid = pfn_to_nid(pfn);
- if (nid == 0xff)
- return 0;
- return pfn >= node_start_pfn(nid) && (pfn) < node_end_pfn(nid);
-}
-EXPORT_SYMBOL(pfn_valid);
-#endif
Index: linux-2.6/include/asm-x86/mmzone_64.h
===================================================================
--- linux-2.6.orig/include/asm-x86/mmzone_64.h 2007-11-13 11:52:34.825632852 -0800
+++ linux-2.6/include/asm-x86/mmzone_64.h 2007-11-15 19:39:59.173245524 -0800
@@ -41,12 +41,6 @@ static inline __attribute__((pure)) int
#define node_end_pfn(nid) (NODE_DATA(nid)->node_start_pfn + \
NODE_DATA(nid)->node_spanned_pages)
-#ifdef CONFIG_DISCONTIGMEM
-#define pfn_to_nid(pfn) phys_to_nid((unsigned long)(pfn) << PAGE_SHIFT)
-
-extern int pfn_valid(unsigned long pfn);
-#endif
-
#ifdef CONFIG_NUMA_EMU
#define FAKE_NODE_MIN_SIZE (64*1024*1024)
#define FAKE_NODE_MIN_HASH_MASK (~(FAKE_NODE_MIN_SIZE - 1uL))
Index: linux-2.6/include/asm-x86/page_64.h
===================================================================
--- linux-2.6.orig/include/asm-x86/page_64.h 2007-11-13 11:52:34.833632806 -0800
+++ linux-2.6/include/asm-x86/page_64.h 2007-11-15 19:39:59.216745079 -0800
@@ -121,9 +121,6 @@ extern unsigned long __phys_addr(unsigne
#define __va(x) ((void *)((unsigned long)(x)+PAGE_OFFSET))
#define __boot_va(x) __va(x)
#define __boot_pa(x) __pa(x)
-#ifdef CONFIG_FLATMEM
-#define pfn_valid(pfn) ((pfn) < end_pfn)
-#endif
#define virt_to_page(kaddr) pfn_to_page(__pa(kaddr) >> PAGE_SHIFT)
#define virt_addr_valid(kaddr) pfn_valid(__pa(kaddr) >> PAGE_SHIFT)
Index: linux-2.6/arch/x86/mm/init_64.c
===================================================================
--- linux-2.6.orig/arch/x86/mm/init_64.c 2007-11-15 12:19:14.820392365 -0800
+++ linux-2.6/arch/x86/mm/init_64.c 2007-11-15 19:39:59.228495304 -0800
@@ -484,34 +484,6 @@ EXPORT_SYMBOL_GPL(memory_add_physaddr_to
#endif /* CONFIG_MEMORY_HOTPLUG */
-#ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
-/*
- * Memory Hotadd without sparsemem. The mem_maps have been allocated in advance,
- * just online the pages.
- */
-int __add_pages(struct zone *z, unsigned long start_pfn, unsigned long nr_pages)
-{
- int err = -EIO;
- unsigned long pfn;
- unsigned long total = 0, mem = 0;
- for (pfn = start_pfn; pfn < start_pfn + nr_pages; pfn++) {
- if (pfn_valid(pfn)) {
- online_page(pfn_to_page(pfn));
- err = 0;
- mem++;
- }
- total++;
- }
- if (!err) {
- z->spanned_pages += total;
- z->present_pages += mem;
- z->zone_pgdat->node_spanned_pages += total;
- z->zone_pgdat->node_present_pages += mem;
- }
- return err;
-}
-#endif
-
static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel, kcore_modules,
kcore_vsyscall;
@@ -737,7 +709,6 @@ const char *arch_vma_name(struct vm_area
return NULL;
}
-#ifdef CONFIG_SPARSEMEM_VMEMMAP
/*
* Initialise the sparsemem vmemmap using huge-pages at the PMD level.
*/
@@ -780,4 +751,3 @@ int __meminit vmemmap_populate(struct pa
return 0;
}
-#endif
Index: linux-2.6/arch/x86/mm/srat_64.c
===================================================================
--- linux-2.6.orig/arch/x86/mm/srat_64.c 2007-11-13 11:52:34.821633366 -0800
+++ linux-2.6/arch/x86/mm/srat_64.c 2007-11-15 19:39:59.228495304 -0800
@@ -151,62 +151,6 @@ acpi_numa_processor_affinity_init(struct
pxm, pa->apic_id, node);
}
-#ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
-/*
- * Protect against too large hotadd areas that would fill up memory.
- */
-static int hotadd_enough_memory(struct bootnode *nd)
-{
- static unsigned long allocated;
- static unsigned long last_area_end;
- unsigned long pages = (nd->end - nd->start) >> PAGE_SHIFT;
- long mem = pages * sizeof(struct page);
- unsigned long addr;
- unsigned long allowed;
- unsigned long oldpages = pages;
-
- if (mem < 0)
- return 0;
- allowed = (end_pfn - absent_pages_in_range(0, end_pfn)) * PAGE_SIZE;
- allowed = (allowed / 100) * hotadd_percent;
- if (allocated + mem > allowed) {
- unsigned long range;
- /* Give them at least part of their hotadd memory upto hotadd_percent
- It would be better to spread the limit out
- over multiple hotplug areas, but that is too complicated
- right now */
- if (allocated >= allowed)
- return 0;
- range = allowed - allocated;
- pages = (range / PAGE_SIZE);
- mem = pages * sizeof(struct page);
- nd->end = nd->start + range;
- }
- /* Not completely fool proof, but a good sanity check */
- addr = find_e820_area(last_area_end, end_pfn<<PAGE_SHIFT, mem);
- if (addr == -1UL)
- return 0;
- if (pages != oldpages)
- printk(KERN_NOTICE "SRAT: Hotadd area limited to %lu bytes\n",
- pages << PAGE_SHIFT);
- last_area_end = addr + mem;
- allocated += mem;
- return 1;
-}
-
-static int update_end_of_memory(unsigned long end)
-{
- found_add_area = 1;
- if ((end >> PAGE_SHIFT) > end_pfn)
- end_pfn = end >> PAGE_SHIFT;
- return 1;
-}
-
-static inline int save_add_info(void)
-{
- return hotadd_percent > 0;
-}
-#else
int update_end_of_memory(unsigned long end) {return -1;}
static int hotadd_enough_memory(struct bootnode *nd) {return 1;}
#ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
@@ -214,7 +158,6 @@ static inline int save_add_info(void) {r
#else
static inline int save_add_info(void) {return 0;}
#endif
-#endif
/*
* Update nodes_add and decide if to include add are in the zone.
* Both SPARSE and RESERVE need nodes_add infomation.
Index: linux-2.6/arch/x86/Kconfig
===================================================================
--- linux-2.6.orig/arch/x86/Kconfig 2007-11-15 12:19:41.950157617 -0800
+++ linux-2.6/arch/x86/Kconfig 2007-11-15 19:39:59.240495606 -0800
@@ -889,19 +889,23 @@ config HAVE_ARCH_ALLOC_REMAP
config ARCH_FLATMEM_ENABLE
def_bool y
- depends on (X86_32 && ARCH_SELECT_MEMORY_MODEL && X86_PC) || (X86_64 && !NUMA)
+ depends on X86_32 && ARCH_SELECT_MEMORY_MODEL && X86_PC
config ARCH_DISCONTIGMEM_ENABLE
def_bool y
- depends on NUMA
+ depends on NUMA && X86_32
config ARCH_DISCONTIGMEM_DEFAULT
def_bool y
- depends on NUMA
+ depends on NUMA && X86_32
+
+config ARCH_SPARSEMEM_DEFAULT
+ def_bool y
+ depends on X86_64
config ARCH_SPARSEMEM_ENABLE
def_bool y
- depends on NUMA || (EXPERIMENTAL && (X86_PC || X86_64))
+ depends on X86_64 || NUMA || (EXPERIMENTAL && X86_PC)
select SPARSEMEM_STATIC if X86_32
select SPARSEMEM_VMEMMAP_ENABLE if X86_64
@@ -1205,18 +1209,10 @@ config ARCH_ENABLE_MEMORY_HOTPLUG
def_bool y
depends on X86_64 || (X86_32 && HIGHMEM)
-config MEMORY_HOTPLUG_RESERVE
- def_bool X86_64
- depends on (MEMORY_HOTPLUG && DISCONTIGMEM)
-
config HAVE_ARCH_EARLY_PFN_TO_NID
def_bool X86_64
depends on NUMA
-config OUT_OF_LINE_PFN_TO_PAGE
- def_bool X86_64
- depends on DISCONTIGMEM
-
menu "Power management options"
depends on !X86_VOYAGER
Index: linux-2.6/include/asm-x86/sparsemem_64.h
===================================================================
--- linux-2.6.orig/include/asm-x86/sparsemem_64.h 2007-11-15 19:50:25.169427205 -0800
+++ linux-2.6/include/asm-x86/sparsemem_64.h 2007-11-15 19:50:32.165046504 -0800
@@ -1,8 +1,6 @@
#ifndef _ASM_X86_64_SPARSEMEM_H
#define _ASM_X86_64_SPARSEMEM_H 1
-#ifdef CONFIG_SPARSEMEM
-
/*
* generic non-linear memory support:
*
@@ -21,6 +19,4 @@
extern int early_pfn_to_nid(unsigned long pfn);
-#endif /* CONFIG_SPARSEMEM */
-
#endif /* _ASM_X86_64_SPARSEMEM_H */