)
From: Ingo Molnar <mingo@elte.hu>
This is the latest version of the scheduler cache-hot-auto-tune patch.
The first problem was that detection time scaled with O(N^2), which is
unacceptable on larger SMP and NUMA systems. To solve this:
- I've added a 'domain distance' function, which is used to cache
measurement results. Each distance is only measured once. This means
that e.g. on NUMA distances of 0, 1 and 2 might be measured, on HT
distances 0 and 1, and on SMP distance 0 is measured. The code walks
the domain tree to determine the distance, so it automatically follows
whatever hierarchy an architecture sets up. This cuts down on the boot
time significantly and removes the O(N^2) limit. The only assumption
is that migration costs can be expressed as a function of domain
distance - this covers the overwhelming majority of existing systems,
and is a good guess even for more assymetric systems.
[ People hacking systems that have assymetries that break this
assumption (e.g. different CPU speeds) should experiment a bit with
the cpu_distance() function. Adding a ->migration_distance factor to
the domain structure would be one possible solution - but lets first
see the problem systems, if they exist at all. Lets not overdesign. ]
Another problem was that only a single cache-size was used for measuring
the cost of migration, and most architectures didnt set that variable
up. Furthermore, a single cache-size does not fit NUMA hierarchies with
L3 caches and does not fit HT setups, where different CPUs will often
have different 'effective cache sizes'. To solve this problem:
- Instead of relying on a single cache-size provided by the platform and
sticking to it, the code now auto-detects the 'effective migration
cost' between two measured CPUs, via iterating through a wide range of
cachesizes. The code searches for the maximum migration cost, which
occurs when the working set of the test-workload falls just below the
'effective cache size'. I.e. real-life optimized search is done for
the maximum migration cost, between two real CPUs.
This, amongst other things, has the positive effect hat if e.g. two
CPUs share a L2/L3 cache, a different (and accurate) migration cost
will be found than between two CPUs on the same system that dont share
any caches.
(The reliable measurement of migration costs is tricky - see the source
for details.)
Furthermore i've added various boot-time options to override/tune
migration behavior.
Firstly, there's a blanket override for autodetection:
migration_cost=1000,2000,3000
will override the depth 0/1/2 values with 1msec/2msec/3msec values.
Secondly, there's a global factor that can be used to increase (or
decrease) the autodetected values:
migration_factor=120
will increase the autodetected values by 20%. This option is useful to
tune things in a workload-dependent way - e.g. if a workload is
cache-insensitive then CPU utilization can be maximized by specifying
migration_factor=0.
I've tested the autodetection code quite extensively on x86, on 3
P3/Xeon/2MB, and the autodetected values look pretty good:
Dual Celeron (128K L2 cache):
---------------------
migration cost matrix (max_cache_size: 131072, cpu: 467 MHz):
---------------------
[00] [01]
[00]: - 1.7(1)
[01]: 1.7(1) -
---------------------
cacheflush times [2]: 0.0 (0) 1.7 (1784008)
---------------------
Here the slow memory subsystem dominates system performance, and even
though caches are small, the migration cost is 1.7 msecs.
Dual HT P4 (512K L2 cache):
---------------------
migration cost matrix (max_cache_size: 524288, cpu: 2379 MHz):
---------------------
[00] [01] [02] [03]
[00]: - 0.4(1) 0.0(0) 0.4(1)
[01]: 0.4(1) - 0.4(1) 0.0(0)
[02]: 0.0(0) 0.4(1) - 0.4(1)
[03]: 0.4(1) 0.0(0) 0.4(1) -
---------------------
cacheflush times [2]: 0.0 (33900) 0.4 (448514)
---------------------
Here it can be seen that there is no migration cost between two HT
siblings (CPU#0/2 and CPU#1/3 are separate physical CPUs). A fast memory
system makes inter-physical-CPU migration pretty cheap: 0.4 msecs.
8-way P3/Xeon [2MB L2 cache]:
---------------------
migration cost matrix (max_cache_size: 2097152, cpu: 700 MHz):
---------------------
[00] [01] [02] [03] [04] [05] [06] [07]
[00]: - 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1)
[01]: 19.2(1) - 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1)
[02]: 19.2(1) 19.2(1) - 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1)
[03]: 19.2(1) 19.2(1) 19.2(1) - 19.2(1) 19.2(1) 19.2(1) 19.2(1)
[04]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) - 19.2(1) 19.2(1) 19.2(1)
[05]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) - 19.2(1) 19.2(1)
[06]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) - 19.2(1)
[07]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) -
---------------------
cacheflush times [2]: 0.0 (0) 19.2 (19281756)
---------------------
This one has huge caches and a relatively slow memory subsystem - so the
migration cost is 19 msecs.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Ashok Raj <ashok.raj@intel.com>
Signed-off-by: Ken Chen <kenneth.w.chen@intel.com>
Cc: <wilder@us.ibm.com>
Signed-off-by: John Hawkes <hawkes@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Add per-arch sched_cacheflush() which is a write-back cacheflush used by
the migration-cost calibration code at bootup time.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Remove support for obsolete hardware and cleanup.
- Remove checks for non integrated APICs
- Replace apic_write_around with apic_write.
- Remove apic_read_around
- Remove APIC version reads used by old workarounds
- Remove old workaround for Simics
- Fix indentation
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
When building in a separate objtree, file names produced by BUG() & Co. can
get fairly long; printing only the first 50 characters may thus result in
(almost) no useful information. The following change makes it so that rather
the last 50 characters of the filename get printed.
Signed-Off-By: Jan Beulich <jbeulich@novell.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
X86_FEATURE_K8_C was a synthetic Linux CPUID flag that was used for some
code optimizations in Opteron C stepping or later. But support for pre C
stepping optimizations has been removed, so this isn't needed anymore.
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>
They used to be used by the reboot code, but not anymore.
Noticed by Jan Beulich
Cc: JBeulich@novell.com
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Introduce vSMP arch to the kernel.
This patch:
1. Adds CONFIG_X86_VSMP
2. Adds machine specific macros for local_irq_disabled, local_irq_enabled
and irqs_disabled
3. Writes to the vSMP CTL device to indicate kernel compiled with CONFIG_VSMP
Signed-off-by: Ravikiran Thirumalai <kiran@scalemp.com>
Signed-off-by: Shai Fultheim <shai@scalemp.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
vSMP specific alignment patch to
1. Define INTERNODE_CACHE_SHIFT for vSMP
2. Use this for alignment of critical structures
3. Use INTERNODE_CACHE_SHIFT for ARCH_MIN_TASKALIGN,
and let the slab align task_struct allocations to the internode cacheline size
4. Introduce and use ARCH_MIN_MMSTRUCT_ALIGN for mm_struct slab allocations.
Signed-off-by: Ravikiran Thirumalai <kiran@scalemp.com>
Signed-off-by: Shai Fultheim <shai@scalemp.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Fixes
CC fs/nfsctl.o
In file included from include2/asm/atomic.h:427,
from /home/lsrc/quilt/linux/include/linux/file.h:8,
from /home/lsrc/quilt/linux/fs/nfsctl.c:8:
/home/lsrc/quilt/linux/include/asm-generic/atomic.h:20:5: warning: "BITS_PER_LONG" is not defined
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Move the #ifdef into the function body.
Signed-off-by: Brian Gerst <bgerst@didntduck.org>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
It was set as an NMI, but the NMI bit always forces an interrupt
to end up at vector 2. So it was never used. Remove.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
It looks like the new scalable TLB flush code for x86_64 is claiming
one more IRQ vector than it actually uses.
Signed-off-by: Jason Uhlenkott <jasonuhl@sgi.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Patch uses a static PDA array early at boot and reallocates processor PDA
with node local memory when kmalloc is ready, just before pda_init.
The boot_cpu_pda is needed since the cpu_pda is used even before pda_init for
that cpu is called (to set the static per-cpu areas offset table etc)
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>
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>
Passing random input values in eax to cpuid is not a good idea
because the CPU will GPF for unknown ones.
Use the correct x86-64 version that exists for a longer time too.
This also adds a memory barrier to prevent the optimizer from
reordering.
Cc: tigran@veritas.com
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
They already do this in hardware and the Linux algorithm
actually adds errors.
Cc: mingo@elte.hu
Cc: rohit.seth@intel.com
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
cpumask.h wasn't included implicitely into proto.h in this case.
Just move it over to smp.h
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
o Apic id is in most significant 8 bits of APIC_ID register. Current code
is trying to write apic id to least significant 8 bits. This patch fixes
it.
o This fix enables booting uni kdump capture kernel on a cpu with non-zero
apic id.
Signed-off-by: Vivek Goyal <vgoyal@in.ibm.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This function is never used for x86_64.
Signed-off-by: Brian Gerst <bgerst@didntduck.org>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
As suggested by Linus.
This catches driver bugs that could cause corruption on IOMMU architectures.
Also I converted the BUGs to out_of_line_bug()s to save a bit
of text space.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
AK: I hacked Muli's original patch a lot and there were a lot
of changes - all bugs are probably to blame on me now.
There were also some changes in the fall back behaviour
for swiotlb - in particular it doesn't try to use GFP_DMA
now anymore. Also all DMA mapping operations use the
same core dma_alloc_coherent code with proper fallbacks now.
And various other changes and cleanups.
Known problems: iommu=force swiotlb=force together breaks
needs more testing.
This patch cleans up x86_64's DMA mapping dispatching code. Right now
we have three possible IOMMU types: AGP GART, swiotlb and nommu, and
in the future we will also have Xen's x86_64 swiotlb and other HW
IOMMUs for x86_64. In order to support all of them cleanly, this
patch:
- introduces a struct dma_mapping_ops with function pointers for each
of the DMA mapping operations of gart (AMD HW IOMMU), swiotlb
(software IOMMU) and nommu (no IOMMU).
- gets rid of:
if (swiotlb)
return swiotlb_xxx();
- PCI_DMA_BUS_IS_PHYS is now checked against the dma_ops being set
This makes swiotlb faster by avoiding double copying in some cases.
Signed-Off-By: Muli Ben-Yehuda <mulix@mulix.org>
Signed-Off-By: Jon D. Mason <jdmason@us.ibm.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This adds a new notifier chain that is called with IDLE_START
when a CPU goes idle and IDLE_END when it goes out of idle.
The context can be idle thread or interrupt context.
Since we cannot rely on MONITOR/MWAIT existing the idle
end check currently has to be done in all interrupt
handlers.
They were originally inspired by the similar s390 implementation.
They have a variety of applications:
- They will be needed for CONFIG_NO_IDLE_HZ
- They can be used for oprofile to fix up the missing time
in idle when performance counters don't tick.
- They can be used for better C state management in ACPI
- They could be used for microstate accounting.
This is just infrastructure so far, no users.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Whenever we see that a CPU is capable of C3 (during ACPI cstate init), we
disable local APIC timer and switch to using a broadcast from external timer
interrupt (IRQ 0).
Patch below adds the code for x86_64.
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Whenever we see that a CPU is capable of C3 (during ACPI cstate init), we
disable local APIC timer and switch to using a broadcast from external timer
interrupt (IRQ 0). This is needed because Intel CPUs stop the local
APIC timer in C3. This is currently only enabled for Intel CPUs.
Patch below adds the code for i386 and also the ACPI hunk.
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch is on the same lines as Zachary Amsden's i386 GDT page alignemnt
patch in -mm, but for x86_64.
Patch to align and pad x86_64 GDT on page boundries.
[AK: some minor cleanups and fixed incorrect TLS initialization
in CPU init.]
Signed-off-by: Nippun Goel <nippung@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>
Following kmalloc_node.
Needed for another patch to return -1 for unknown nodes in x86-64.
Cc: Christoph Lameter <clameter@engr.sgi.com>
Cc: kiran@scalex86.org
Signed-off-by: Andi Kleen <ak@suse.de>
[ Changed 0 to numa_node_id() on suggestion by Christoph Lameter ]
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
The separation of the rex64 prefix (on fxsave/fxrstor) by way of using
a semicolon resulted in the prefix not always taking effect (because
when extended registers are needed for addressing, another rex prefix
would have been generated by the compiler), thus (depending on the
build) resulting in eventually getting 32-bit saves and/or restores.
Signed-Off-By: Jan Beulich <jbeulich@novell.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Some people need it now on 64bit so reuse the i386 code for
x86-64. This will be also useful for future bug workarounds.
It is a bit simplified there because there is no need
to do it very early on x86-64. This means it doesn't need
early ioremap et.al. We run it as a core initcall right now.
I hope it's not needed for early setup.
I added a general CONFIG_DMI symbol in case IA64 or someone
else wants to reuse the code later too.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Use single instruction for find largest set bit on x86_64.
[Updated by Jan Beulich to fix wrong asm constraints in original
patch -AK]
Cc: jbeulich@novell.com
Signed-off-by: Stephen Hemminger <shemminger@osdl.org>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
As discussed, the flags register on x86-64 is saved and restored by the
assembly code which sets up struct pt_regs, so we do not need to save
and restore it in the inline assembler which already informs gcc that
we're clobbering the flags. This patch has been sanity booted and works
okay here.
Signed-off-by: Benjamin LaHaise <benjamin.c.lahaise@intel.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
I noticed that some lowlevel send_IPI_mask helpers had a hotplug/preempt
race whereupon the cpu_online_map was read before disabling preemption;
...
cpumask_t mask = cpu_online_map;
int cpu = get_cpu();
cpu_clear(cpu, mask);
...
But then i realised that there is no need for these lowlevel functions to
be going through all this trouble when all the callers are already made
hotplug/preempt safe.
Signed-off-by: Zwane Mwaikambo <zwane@arm.linux.org.uk>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This
- switches the INT3 handler to run on an IST stack (to cope with
breakpoints set by a kernel debugger on places where the kernel's
%gs base hasn't been set up, yet); the IST stack used is shared with
the INT1 handler's
[AK: this also allows setting a kprobe on the interrupt/exception entry
points]
- allows nesting of INT1/INT3 handlers so that one can, with a kernel
debugger, debug (at least) the user-mode portions of the INT1/INT3
handling; the nesting isn't actively enabled here since a kernel-
debugger-free kernel doesn't need it
Signed-Off-By: Jan Beulich <jbeulich@novell.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Print bits for RDTSCP, SVM, CR8-LEGACY.
Also now print power flags on i386 like x86-64 always did.
This will add a new line in the 386 cpuinfo, but that shouldn't
be an issue - did that in the past too and I haven't heard
of any breakage.
I shrunk some of the fields in the i386 cpuinfo_x86 to chars
to make up for the new int "x86_power" field. Overall it's
smaller than before.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Define it for i386 too.
This is a synthetic flag that signifies that the CPU's TSC runs
at a constant P state invariant frequency.
Fix up the logic on x86-64/i386 to set it on all known CPUs.
Use the AMD defined bit to set it on future AMD CPUs.
Cc: venkatesh.pallipadi@intel.com
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Was only used by the floppy driver to work around some ancient
hardware bug that should never occur on any 64bit system.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This adjusts things so that handlers of the die() notifier will have
sufficient information about the trap currently being handled. It also
adjusts the notify_die() prototype to (again) match that of i386.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
As a follow-up to the introduction of CONFIG_UNWIND_INFO, this
separates the generation of frame unwind information for x86-64 from
that of full debug information.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Frame unwind information was still incorrect for ia32_ptregs_common
(sorry, my fault), and could be improved for some of the other entry
points.
Signed-Off-By: Jan Beulich <jbeulich@novell.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
- Move capable() from sched.h to capability.h;
- Use <linux/capability.h> where capable() is used
(in include/, block/, ipc/, kernel/, a few drivers/,
mm/, security/, & sound/;
many more drivers/ to go)
Signed-off-by: Randy Dunlap <rdunlap@xenotime.net>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Uninline capable(). Saves 2K of kernel text on a generic .config, and 1K on a
tiny config. In addition it makes the use of capable more consistent between
CONFIG_SECURITY and !CONFIG_SECURITY
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
