From: Keith Mannthey, Andi Kleen
Implement memory hotadd without sparsemem. The memory in the SRAT
hotadd area is just preserved instead and can be activated later.
There are a few restrictions:
- Only one continuous hotadd area allowed per node
The main problem is dealing with the many buggy SRAT tables
that are out there. The strategy here is to reject anything
suspicious.
Originally from Keith Mannthey, with several hacks and changes by AK
and also contributions from Andrew Morton
[ TBD: Problems pointed out by KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>:
1) Goto's rebuild_zonelist patch will not work if CONFIG_MEMORY_HOTPLUG=n.
Rebuilding zonelist is necessary when the system has just memory <
4G at boot, and hot add memory > 4G. because x86_64 has DMA32,
ZONE_NORAML is not included into zonelist at boot time if system
doesn't have memory >4G at boot.
[AK: should just force the higher zones at boot time when SRAT tells us]
2) zone and node's spanned_pages and present_pages are not incremented.
They should be.
For example, our server (ia64/Fujitsu PrimeQuest) can equip memory
from 4G to 1T(maybe 2T in future), and SRAT will *always* say we have
possible 1T +memory. (Microsoft requires "write all possible memory
in SRAT") When we reserve memmap for possible 1T memory, Linux will
not work well in +minimum 4G configuraion ;)
[AK: needs limiting to 5-10% of max memory]
]
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>
This fixes problems with very large nodes (over 128GB) filling up all of
the first 4GB with their mem_map and not leaving enough space for the
swiotlb.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
It conflicts with the struct node in node.h
Actually the x86-64 version was there first, but ..
Suggested by Jan Beulich
Cc: jbeulich@novell.com
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The early initialization of cpu_to_node code as it is now only updates the
cpu_to_node array, and does not update cpu_pda()->nodemember. This will
cause numa_node_id() to return 0 on systems where CPU 0 is not on Node 0.
This leads to a kernel panic in slab.c.
I've tested the patch below on a 16 processor x86_64 ES7000-600 server, and
no longer see the panic I saw with the original 2.6.16-rc3.
Signed-off-by: Dan Yeisley <dan.yeisley@unisys.com>
Acked-by: Andi Kleen <ak@muc.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
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>
Patch enables early intialization of cpu_to_node.
apicid_to_node is built by reading the SRAT table, from acpi_numa_init with
ACPI_NUMA and k8_scan_nodes with K8_NUMA.
x86_cpu_to_apicid is built by parsing the ACPI MADT table, from acpi_boot_init.
We combine these two tables and setup cpu_to_node.
Early intialization helps the static per_cpu_areas in getting pages from
correct node.
Change since last release:
Do not initialize early init_cpu_to_node for faking node cases.
Patch tested on TYAN dual core 4P board with K8 only, ACPI_NUMA.
Tested on EM64T NUMA. Also tested with numa=off, numa=fake, and running
a kernel compiled with NUMA on a regular EM64 2 way SMP.
Signed-off-by: Alok N Kataria <alokk@calsoftinc.com>
Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org>
Signed-off-by: Shai Fultheim <shai@scalex86.org>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
As reported by Keith Mannthey, there are problems in populate_memnodemap()
The bug was that the compute_hash_shift() was returning 31, with incorrect
initialization of memnodemap[]
To correct the bug, we must use (1UL << shift) instead of (1 << shift) to
avoid an integer overflow, and we must check that shift < 64 to avoid an
infinite loop.
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Fix up booting with sparse mem enabled. Otherwise it would just
cause an early PANIC at boot.
Signed-off-by: Bob Picco <bob.picco@hp.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The current x86_64 NUMA memory code is inconsequent when it comes to node
memory ranges. The exact behaviour varies depending on which config option
that is used.
setup_node_bootmem() has start and end as arguments and these are used to
calculate the size of the node like this: (end - start). This is all fine
if end is pointing to the first non-available byte. The problem is that the
current x86_64 code sometimes treats it as the last present byte and sometimes
as the first non-available byte. The result is that some configurations might
lose a page at the end of the range.
This patch tries to fix CONFIG_ACPI_NUMA, CONFIG_K8_NUMA and CONFIG_NUMA_EMU
so they all treat the end variable as the first non-available byte. This is
the same way as the single node code.
The patch is boot tested on dual x86_64 hardware with the above configurations,
but maybe the removed code is needed as some workaround?
Signed-off-by: Magnus Damm <magnus@valinux.co.jp>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Compute the highest possible value for memnode_shift, in order to reduce
footprint of memnodemap[] to the minimum, thus making all users
(phys_to_nid(), kfree()), more cache friendly.
Before the patch :
Node 0 MemBase 0000000000000000 Limit 00000001ffffffff
Node 1 MemBase 0000000200000000 Limit 00000003ffffffff
Using 23 for the hash shift. Max adder is 3ffffffff
After the patch :
Node 0 MemBase 0000000000000000 Limit 00000001ffffffff
Node 1 MemBase 0000000200000000 Limit 00000003ffffffff
Using 33 for the hash shift.
In this case, only 2 bytes of memnodemap[] are used, instead of 2048
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Not go from the CPU number to an mapping array.
Mode number is often used now in fast paths.
This also adds a generic numa_node_id to all the topology includes
Suggested by Eric Dumazet
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Add a new 4GB GFP_DMA32 zone between the GFP_DMA and GFP_NORMAL zones.
As a bit of historical background: when the x86-64 port
was originally designed we had some discussion if we should
use a 16MB DMA zone like i386 or a 4GB DMA zone like IA64 or
both. Both was ruled out at this point because it was in early
2.4 when VM is still quite shakey and had bad troubles even
dealing with one DMA zone. We settled on the 16MB DMA zone mainly
because we worried about older soundcards and the floppy.
But this has always caused problems since then because
device drivers had trouble getting enough DMA able memory. These days
the VM works much better and the wide use of NUMA has proven
it can deal with many zones successfully.
So this patch adds both zones.
This helps drivers who need a lot of memory below 4GB because
their hardware is not accessing more (graphic drivers - proprietary
and free ones, video frame buffer drivers, sound drivers etc.).
Previously they could only use IOMMU+16MB GFP_DMA, which
was not enough memory.
Another common problem is that hardware who has full memory
addressing for >4GB misses it for some control structures in memory
(like transmit rings or other metadata). They tended to allocate memory
in the 16MB GFP_DMA or the IOMMU/swiotlb then using pci_alloc_consistent,
but that can tie up a lot of precious 16MB GFPDMA/IOMMU/swiotlb memory
(even on AMD systems the IOMMU tends to be quite small) especially if you have
many devices. With the new zone pci_alloc_consistent can just put
this stuff into memory below 4GB which works better.
One argument was still if the zone should be 4GB or 2GB. The main
motivation for 2GB would be an unnamed not so unpopular hardware
raid controller (mostly found in older machines from a particular four letter
company) who has a strange 2GB restriction in firmware. But
that one works ok with swiotlb/IOMMU anyways, so it doesn't really
need GFP_DMA32. I chose 4GB to be compatible with IA64 and because
it seems to be the most common restriction.
The new zone is so far added only for x86-64.
For other architectures who don't set up this
new zone nothing changes. Architectures can set a compatibility
define in Kconfig CONFIG_DMA_IS_DMA32 that will define GFP_DMA32
as GFP_DMA. Otherwise it's a nop because on 32bit architectures
it's normally not needed because GFP_NORMAL (=0) is DMA able
enough.
One problem is still that GFP_DMA means different things on different
architectures. e.g. some drivers used to have #ifdef ia64 use GFP_DMA
(trusting it to be 4GB) #elif __x86_64__ (use other hacks like
the swiotlb because 16MB is not enough) ... . This was quite
ugly and is now obsolete.
These should be now converted to use GFP_DMA32 unconditionally. I haven't done
this yet. Or best only use pci_alloc_consistent/dma_alloc_coherent
which will use GFP_DMA32 transparently.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The tests Alok carried out on Petr's box confirmed that cpu_to_node[BP] is
not setup early enough by numa_init_array due to the x86_64 changes in
2.6.14-rc*, and unfortunately set wrongly by the work around code in
numa_init_array(). cpu_to_node[0] gets set with 1 early and later gets set
properly to 0 during identify_cpu() when all cpus are brought up, but
confusing the numa slab in the process.
Here is a quick fix for this. The right fix obviously is to have
cpu_to_node[bsp] setup early for numa_init_array(). The following patch
will fix the problem now, and the code can stay on even when
cpu_to_node{BP] gets fixed early correctly.
Thanks to Petr for access to his box.
Signed off by: Ravikiran Thirumalai <kiran@scalex86.org>
Signed-off-by: Alok N Kataria <alokk@calsoftinc.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Fix the BP node_to_cpumask. 2.6.14-rc* broke the boot cpu bit as the
cpu_to_node(0) is now not setup early enough for numa_init_array.
cpu_to_node[] is setup much later at srat_detect_node on acpi srat based
em64t machines. This seems like a problem on amd machines too, Tested on
em64t though. /sys/devices/system/node/node0/cpumap shows up sanely after
this patch.
Signed off by: Ravikiran Thirumalai <kiran@scalex86.org>
Signed-off-by: Shai Fultheim <shai@scalex86.org>
Cc: Andi Kleen <ak@muc.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
In particular on systems where the local APIC space and node space
is very different from the Linux CPU number space.
Previously the older NUMA setup code directly parsing the K8
northbridge registers had some issues on 8 socket or dual core
systems. This patch fixes them.
This is mainly done by fixing some confusion between Linux
CPU numbers and local APIC ids. We now pass the local APIC IDs
to later code, which avoids mismatches.
Also add some heuristics to detect cases where the Hypertransport
nodeids and the local APIC IDs don't match, but are shifted
by a constant offset.
This is still all quite hackish, hopefully BIOS writers fill
in correct SRATs instead.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Do that later when the CPU boots. SRAT just stores the APIC<->Node
mapping node. This fixes problems on systems where the order
of SRAT entries does not match the MADT.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Mark variables which are usually accessed for reads with __readmostly.
Signed-off-by: Alok N Kataria <alokk@calsoftinc.com>
Signed-off-by: Shai Fultheim <shai@scalex86.org>
Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Some nodes can have large holes on x86-64.
This fixes problems with the VM allowing too many dirty pages because it
overestimates the number of available RAM in a node. In extreme cases you
can end up with all RAM filled with dirty pages which can lead to deadlocks
and other nasty behaviour.
This patch just tells the VM about the known holes from e820. Reserved
(like the kernel text or mem_map) is still not taken into account, but that
should be only a few percent error now.
Small detail is that the flat setup uses the NUMA free_area_init_node() now
too because it offers more flexibility.
(akpm: lotsa thanks to Martin for working this problem out)
Cc: Martin Bligh <mbligh@mbligh.org>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch adds __cpuinit and __cpuinitdata sections that need to exist past
boot to support cpu hotplug.
Caveat: This is done *only* for EM64T CPU Hotplug support, on request from
Andi Kleen. Much of the generic hotplug code in kernel, and none of the other
archs that support CPU hotplug today, i386, ia64, ppc64, s390 and parisc dont
mark sections with __cpuinit, but only mark them as __devinit, and
__devinitdata.
If someone is motivated to change generic code, we need to make sure all
existing hotplug code does not break, on other arch's that dont use __cpuinit,
and __cpudevinit.
Signed-off-by: Ashok Raj <ashok.raj@intel.com>
Acked-by: Andi Kleen <ak@muc.de>
Acked-by: Zwane Mwaikambo <zwane@arm.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch adds in the necessary support for sparsemem such that x86-64
kernels may use sparsemem as an alternative to discontigmem for NUMA
kernels. Note that this does no preclude one from continuing to build NUMA
kernels using discontigmem, but merely allows the option to build NUMA
kernels with sparsemem.
Interestingly, the use of sparsemem in lieu of discontigmem in NUMA kernels
results in reduced text size for otherwise equivalent kernels as shown in
the example builds below:
text data bss dec hex filename
2371036 765884 1237108 4374028 42be0c vmlinux.discontig
2366549 776484 1302772 4445805 43d66d vmlinux.sparse
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>
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!
