What's happening is that the assertion in mm/page_alloc.c:move_freepages()
is triggering:
BUG_ON(page_zone(start_page) != page_zone(end_page));
Once I knew this is what was happening, I added some annotations:
if (unlikely(page_zone(start_page) != page_zone(end_page))) {
printk(KERN_ERR "move_freepages: Bogus zones: "
"start_page[%p] end_page[%p] zone[%p]\n",
start_page, end_page, zone);
printk(KERN_ERR "move_freepages: "
"start_zone[%p] end_zone[%p]\n",
page_zone(start_page), page_zone(end_page));
printk(KERN_ERR "move_freepages: "
"start_pfn[0x%lx] end_pfn[0x%lx]\n",
page_to_pfn(start_page), page_to_pfn(end_page));
printk(KERN_ERR "move_freepages: "
"start_nid[%d] end_nid[%d]\n",
page_to_nid(start_page), page_to_nid(end_page));
...
And here's what I got:
move_freepages: Bogus zones: start_page[2207d0000] end_page[2207dffc0] zone[fffff8103effcb00]
move_freepages: start_zone[fffff8103effcb00] end_zone[fffff8003fffeb00]
move_freepages: start_pfn[0x81f600] end_pfn[0x81f7ff]
move_freepages: start_nid[1] end_nid[0]
My memory layout on this box is:
[ 0.000000] Zone PFN ranges:
[ 0.000000] Normal 0x00000000 -> 0x0081ff5d
[ 0.000000] Movable zone start PFN for each node
[ 0.000000] early_node_map[8] active PFN ranges
[ 0.000000] 0: 0x00000000 -> 0x00020000
[ 0.000000] 1: 0x00800000 -> 0x0081f7ff
[ 0.000000] 1: 0x0081f800 -> 0x0081fe50
[ 0.000000] 1: 0x0081fed1 -> 0x0081fed8
[ 0.000000] 1: 0x0081feda -> 0x0081fedb
[ 0.000000] 1: 0x0081fedd -> 0x0081fee5
[ 0.000000] 1: 0x0081fee7 -> 0x0081ff51
[ 0.000000] 1: 0x0081ff59 -> 0x0081ff5d
So it's a block move in that 0x81f600-->0x81f7ff region which triggers
the problem.
This patch:
Declaration of early_pfn_to_nid() is scattered over per-arch include
files, and it seems it's complicated to know when the declaration is used.
I think it makes fix-for-memmap-init not easy.
This patch moves all declaration to include/linux/mm.h
After this,
if !CONFIG_NODES_POPULATES_NODE_MAP && !CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
-> Use static definition in include/linux/mm.h
else if !CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
-> Use generic definition in mm/page_alloc.c
else
-> per-arch back end function will be called.
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Tested-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reported-by: David Miller <davem@davemlloft.net>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: <stable@kernel.org> [2.6.25.x, 2.6.26.x, 2.6.27.x, 2.6.28.x]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Change header guards named "ASM_X86__*" to "_ASM_X86_*" since:
a. the double underscore is ugly and pointless.
b. no leading underscore violates namespace constraints.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
This patch is the result of an automatic script that consolidates the
format of all the headers in include/asm-x86/.
The format:
1. No leading underscore. Names with leading underscores are reserved.
2. Pathname components are separated by two underscores. So we can
distinguish between mm_types.h and mm/types.h.
3. Everything except letters and numbers are turned into single
underscores.
Signed-off-by: Vegard Nossum <vegard.nossum@gmail.com>
I've removed the test from phys_to_nid and made a function from __phys_addr
only when the debugging is enabled (on x86_32).
Signed-off-by: Jiri Slaby <jirislaby@gmail.com>
Cc: tglx@linutronix.de
Cc: hpa@zytor.com
Cc: Mike Travis <travis@sgi.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: <x86@kernel.org>
Cc: linux-mm@kvack.org
Cc: Jiri Slaby <jirislaby@gmail.com>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Add some (configurable) expensive sanity checking to catch wrong address
translations on x86.
- create linux/mmdebug.h file to be able include this file in
asm headers to not get unsolvable loops in header files
- __phys_addr on x86_32 became a function in ioremap.c since
PAGE_OFFSET, is_vmalloc_addr and VMALLOC_* non-constasts are undefined
if declared in page_32.h
- add __phys_addr_const for initializing doublefault_tss.__cr3
Tested on 386, 386pae, x86_64 and x86_64 numa=fake=2.
Contains Andi's enable numa virtual address debug patch.
Signed-off-by: Jiri Slaby <jirislaby@gmail.com>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Change the size of node ids for X86_64 from u8 to s16 to
accomodate more than 32k nodes and allow for NUMA_NO_NODE
(-1) to be sign extended to int.
Cc: David Rientjes <rientjes@google.com>
Cc: Yinghai Lu <yhlu.kernel@gmail.com>
Cc: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: Mike Travis <travis@sgi.com>
Reviewed-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
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.
[ akpm@linux-foundation.org: allnoconfig build fix ]
Acked-by: Andi Kleen <ak@suse.de>
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Move the headers to include/asm-x86 and fixup the
header install make rules
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Extends the numa=fake x86_64 command-line option to allow for configurable
node sizes. These nodes can be used in conjunction with cpusets for coarse
memory resource management.
The old command-line option is still supported:
numa=fake=32 gives 32 fake NUMA nodes, ignoring the NUMA setup of the
actual machine.
But now you may configure your system for the node sizes of your choice:
numa=fake=2*512,1024,2*256
gives two 512M nodes, one 1024M node, two 256M nodes, and
the rest of system memory to a sixth node.
The existing hash function is maintained to support the various node sizes
that are possible with this implementation.
Each node of the same size receives roughly the same amount of available
pages, regardless of any reserved memory with its address range. The total
available pages on the system is calculated and divided by the number of equal
nodes to allocate. These nodes are then dynamically allocated and their
borders extended until such time as their number of available pages reaches
the required size.
Configurable node sizes are recommended when used in conjunction with cpusets
for memory control because it eliminates the overhead associated with scanning
the zonelists of many smaller full nodes on page_alloc().
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Paul Jackson <pj@sgi.com>
Cc: Christoph Lameter <clameter@engr.sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
This patch resolves the issue of running with numa=fake=X on kernel command
line on x86_64 machines that have big IO hole. While calculating the size
of each node now we look at the total hole size in that range.
Previously there were nodes that only had IO holes in them causing kernel
boot problems. We now use the NODE_MIN_SIZE (64MB) as the minimum size of
memory that any node must have. We reduce the number of allocated nodes if
the number of nodes specified on kernel command line results in any node
getting memory smaller than NODE_MIN_SIZE.
This change allows the extra memory to be incremented in NODE_MIN_SIZE
granule and uniformly distribute among as many nodes (called big nodes) as
possible.
[akpm@osdl.org: build fix]
Signed-off-by: David Rientjes <reintjes@google.com>
Signed-off-by: Paul Menage <menage@google.com>
Signed-off-by: Rohit Seth <rohitseth@google.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Remove the statically allocated memory to NUMA node hash map in favor of a
dynamically allocated memory to node hash map (it is cache aligned).
This patch has the nice side effect in that it allows the hash map to grow
for systems with large amounts of memory (256GB - 1TB), but suffer from
having small PCI space tacked onto the boot node (which is somewhere
between 192MB to 512MB on the ES7000).
Signed-off-by: Amul Shah <amul.shah@unisys.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Andi Kleen <ak@suse.de>
Cc: Rohit Seth <rohitseth@google.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Needed for some big Opteron systems to compute a numa hash function
They have more than 12 bits significant address.
TBD switch this over to dynamic allocation or use better hash
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
pfn_to_page() and others need to access both memnode_shift and the very
first bytes of memnodemap[]. If we force memnode_shift to be just before the
memnodemap array, we can reduce the memory footprint to one cache line
instead of two for most setups. This patch introduce a 'memnode' structure
where shift and map[] are carefully placed.
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Saves about ~18K .text in defconfig
There would be more optimization potential, but that's for later.
Suggestion originally from Bill Irwin.
Fix from Andy Whitcroft.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
pfn_to_page really requires pfn_valid to be true now, no question.
Some people stumbled over it, but it was misleading and wrong.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Needed for some newer Opteron systems with E stepping and memory
relocation enabled. The node addresses are different in lower
bits now so the nodemap hash function needs to be enlarged.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
In order to use the alternative sparsemem implmentation for NUMA kernels,
we need to reorganize the config options. This patch effectively abstracts
out the CONFIG_DISCONTIGMEM options to CONFIG_NUMA in most cases. Thus,
the discontigmem implementation may be employed as always, but the
sparsemem implementation may be used alternatively.
Signed-off-by: Matt Tolentino <matthew.e.tolentino@intel.com>
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch effectively eliminates direct use of pgdat->node_mem_map outside
of the DISCONTIG code. On a flat memory system, these fields aren't
currently used, neither are they on a sparsemem system.
There was also a node_mem_map(nid) macro on many architectures. Its use
along with the use of ->node_mem_map itself was not consistent. It has
been removed in favor of two new, more explicit, arch-independent macros:
pgdat_page_nr(pgdat, pagenr)
nid_page_nr(nid, pagenr)
I called them "pgdat" and "nid" because we overload the term "node" to mean
"NUMA node", "DISCONTIG node" or "pg_data_t" in very confusing ways. I
believe the newer names are much clearer.
These macros can be overridden in the sparsemem case with a theoretically
slower operation using node_start_pfn and pfn_to_page(), instead. We could
make this the only behavior if people want, but I don't want to change too
much at once. One thing at a time.
This patch removes more code than it adds.
Compile tested on alpha, alpha discontig, arm, arm-discontig, i386, i386
generic, NUMAQ, Summit, ppc64, ppc64 discontig, and x86_64. Full list
here: http://sr71.net/patches/2.6.12/2.6.12-rc1-mhp2/configs/
Boot tested on NUMAQ, x86 SMP and ppc64 power4/5 LPARs.
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Martin J. Bligh <mbligh@aracnet.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
KAMEZAWA Hiroyuki