I'm testing swapout workload in a two-socket Xeon machine. The workload
has 10 threads, each thread sequentially accesses separate memory
region. TLB flush overhead is very big in the workload. For each page,
page reclaim need move it from active lru list and then unmap it. Both
need a TLB flush. And this is a multthread workload, TLB flush happens
in 10 CPUs. In X86, TLB flush uses generic smp_call)function. So this
workload stress smp_call_function_many heavily.
Without patch, perf shows:
+ 24.49% [k] generic_smp_call_function_interrupt
- 21.72% [k] _raw_spin_lock
- _raw_spin_lock
+ 79.80% __page_check_address
+ 6.42% generic_smp_call_function_interrupt
+ 3.31% get_swap_page
+ 2.37% free_pcppages_bulk
+ 1.75% handle_pte_fault
+ 1.54% put_super
+ 1.41% grab_super_passive
+ 1.36% __swap_duplicate
+ 0.68% blk_flush_plug_list
+ 0.62% swap_info_get
+ 6.55% [k] flush_tlb_func
+ 6.46% [k] smp_call_function_many
+ 5.09% [k] call_function_interrupt
+ 4.75% [k] default_send_IPI_mask_sequence_phys
+ 2.18% [k] find_next_bit
swapout throughput is around 1300M/s.
With the patch, perf shows:
- 27.23% [k] _raw_spin_lock
- _raw_spin_lock
+ 80.53% __page_check_address
+ 8.39% generic_smp_call_function_single_interrupt
+ 2.44% get_swap_page
+ 1.76% free_pcppages_bulk
+ 1.40% handle_pte_fault
+ 1.15% __swap_duplicate
+ 1.05% put_super
+ 0.98% grab_super_passive
+ 0.86% blk_flush_plug_list
+ 0.57% swap_info_get
+ 8.25% [k] default_send_IPI_mask_sequence_phys
+ 7.55% [k] call_function_interrupt
+ 7.47% [k] smp_call_function_many
+ 7.25% [k] flush_tlb_func
+ 3.81% [k] _raw_spin_lock_irqsave
+ 3.78% [k] generic_smp_call_function_single_interrupt
swapout throughput is around 1400M/s. So there is around a 7%
improvement, and total cpu utilization doesn't change.
Without the patch, cfd_data is shared by all CPUs.
generic_smp_call_function_interrupt does read/write cfd_data several times
which will create a lot of cache ping-pong. With the patch, the data
becomes per-cpu. The ping-pong is avoided. And from the perf data, this
doesn't make call_single_queue lock contend.
Next step is to remove generic_smp_call_function_interrupt() from arch
code.
Signed-off-by: Shaohua Li <shli@fusionio.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Add the on_each_cpu_cond() function that wraps on_each_cpu_mask() and
calculates the cpumask of cpus to IPI by calling a function supplied as a
parameter in order to determine whether to IPI each specific cpu.
The function works around allocation failure of cpumask variable in
CONFIG_CPUMASK_OFFSTACK=y by itereating over cpus sending an IPI a time
via smp_call_function_single().
The function is useful since it allows to seperate the specific code that
decided in each case whether to IPI a specific cpu for a specific request
from the common boilerplate code of handling creating the mask, handling
failures etc.
[akpm@linux-foundation.org: s/gfpflags/gfp_flags/]
[akpm@linux-foundation.org: avoid double-evaluation of `info' (per Michal), parenthesise evaluation of `cond_func']
[akpm@linux-foundation.org: s/CPU/CPUs, use all 80 cols in comment]
Signed-off-by: Gilad Ben-Yossef <gilad@benyossef.com>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Russell King <linux@arm.linux.org.uk>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Matt Mackall <mpm@selenic.com>
Cc: Sasha Levin <levinsasha928@gmail.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Avi Kivity <avi@redhat.com>
Acked-by: Michal Nazarewicz <mina86@mina86.org>
Cc: Kosaki Motohiro <kosaki.motohiro@gmail.com>
Cc: Milton Miller <miltonm@bga.com>
Reviewed-by: "Srivatsa S. Bhat" <srivatsa.bhat@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We have lots of infrastructure in place to partition multi-core systems
such that we have a group of CPUs that are dedicated to specific task:
cgroups, scheduler and interrupt affinity, and cpuisol= boot parameter.
Still, kernel code will at times interrupt all CPUs in the system via IPIs
for various needs. These IPIs are useful and cannot be avoided
altogether, but in certain cases it is possible to interrupt only specific
CPUs that have useful work to do and not the entire system.
This patch set, inspired by discussions with Peter Zijlstra and Frederic
Weisbecker when testing the nohz task patch set, is a first stab at trying
to explore doing this by locating the places where such global IPI calls
are being made and turning the global IPI into an IPI for a specific group
of CPUs. The purpose of the patch set is to get feedback if this is the
right way to go for dealing with this issue and indeed, if the issue is
even worth dealing with at all. Based on the feedback from this patch set
I plan to offer further patches that address similar issue in other code
paths.
This patch creates an on_each_cpu_mask() and on_each_cpu_cond()
infrastructure API (the former derived from existing arch specific
versions in Tile and Arm) and uses them to turn several global IPI
invocation to per CPU group invocations.
Core kernel:
on_each_cpu_mask() calls a function on processors specified by cpumask,
which may or may not include the local processor.
You must not call this function with disabled interrupts or from a
hardware interrupt handler or from a bottom half handler.
arch/arm:
Note that the generic version is a little different then the Arm one:
1. It has the mask as first parameter
2. It calls the function on the calling CPU with interrupts disabled,
but this should be OK since the function is called on the other CPUs
with interrupts disabled anyway.
arch/tile:
The API is the same as the tile private one, but the generic version
also calls the function on the with interrupts disabled in UP case
This is OK since the function is called on the other CPUs
with interrupts disabled.
Signed-off-by: Gilad Ben-Yossef <gilad@benyossef.com>
Reviewed-by: Christoph Lameter <cl@linux.com>
Acked-by: Chris Metcalf <cmetcalf@tilera.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Russell King <linux@arm.linux.org.uk>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Matt Mackall <mpm@selenic.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Sasha Levin <levinsasha928@gmail.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Avi Kivity <avi@redhat.com>
Acked-by: Michal Nazarewicz <mina86@mina86.org>
Cc: Kosaki Motohiro <kosaki.motohiro@gmail.com>
Cc: Milton Miller <miltonm@bga.com>
Cc: Russell King <linux@arm.linux.org.uk>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
There is a problem that kdump(2nd kernel) sometimes hangs up due
to a pending IPI from 1st kernel. Kernel panic occurs because IPI
comes before call_single_queue is initialized.
To fix the crash, rename init_call_single_data() to call_function_init()
and call it in start_kernel() so that call_single_queue can be
initialized before enabling interrupts.
The details of the crash are:
(1) 2nd kernel boots up
(2) A pending IPI from 1st kernel comes when irqs are first enabled
in start_kernel().
(3) Kernel tries to handle the interrupt, but call_single_queue
is not initialized yet at this point. As a result, in the
generic_smp_call_function_single_interrupt(), NULL pointer
dereference occurs when list_replace_init() tries to access
&q->list.next.
Therefore this patch changes the name of init_call_single_data()
to call_function_init() and calls it before local_irq_enable()
in start_kernel().
Signed-off-by: Takao Indoh <indou.takao@jp.fujitsu.com>
Reviewed-by: WANG Cong <xiyou.wangcong@gmail.com>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Milton Miller <miltonm@bga.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: kexec@lists.infradead.org
Link: http://lkml.kernel.org/r/D6CBEE2F420741indou.takao@jp.fujitsu.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Now that powerpc has removed its use of MSG_ALL_BUT_SELF and MSG_ALL
all these MSG_ flags are unused.
Signed-off-by: Milton Miller <miltonm@bga.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Commit 34db18a054 ("smp: move smp setup functions to kernel/smp.c")
causes this build error on s390 because of a missing init.h include:
CC arch/s390/kernel/asm-offsets.s
In file included from /home2/heicarst/linux-2.6/arch/s390/include/asm/spinlock.h:14:0,
from include/linux/spinlock.h:87,
from include/linux/seqlock.h:29,
from include/linux/time.h:8,
from include/linux/timex.h:56,
from include/linux/sched.h:57,
from arch/s390/kernel/asm-offsets.c:10:
include/linux/smp.h:117:20: error: expected '=', ',', ';', 'asm' or '__attribute__' before 'setup_nr_cpu_ids'
include/linux/smp.h:118:20: error: expected '=', ',', ';', 'asm' or '__attribute__' before 'smp_init'
Fix it by adding the include statement.
Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Acked-by: WANG Cong <amwang@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Typedef the pointer to the function to be called by smp_call_function() and
friends:
typedef void (*smp_call_func_t)(void *info);
as it is used in a fair number of places.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: linux-arch@vger.kernel.org
put_cpu_no_resched() is an optimization of put_cpu() which unfortunately
can cause high latencies.
The nfs iostats code uses put_cpu_no_resched() in a code sequence where a
reschedule request caused by an interrupt between the get_cpu() and the
put_cpu_no_resched() can delay the reschedule for at least HZ.
The other users of put_cpu_no_resched() optimize correctly in interrupt
code, but there is no real harm in using the put_cpu() function which is
an alias for preempt_enable(). The extra check of the preemmpt count is
not as critical as the potential source of missing a reschedule.
Debugged in the preempt-rt tree and verified in mainline.
Impact: remove a high latency source
[akpm@linux-foundation.org: build fix]
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: Tony Luck <tony.luck@intel.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Cc: "J. Bruce Fields" <bfields@fieldses.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Impact: cleanup, no code changed
Remove an ugly #ifdef CONFIG_SMP from panic(), by providing
an smp_send_stop() wrapper on UP too.
LKML-Reference: <49B91A7E.76E4.0078.0@novell.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: cleanup
disable_ioapic_setup() in init/main.c is ugly as the function is
x86-specific. The #ifdef inline prototype there is ugly too.
Replace it with a generic arch_disable_smp_support() function - which
has a weak alias for non-x86 architectures and for non-ioapic x86 builds.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
If you do
smp_call_function_single(expression-with-side-effects, ...)
then expression-with-side-effects never gets evaluated on UP builds.
As always, implementing it in C is the correct thing to do.
While we're there, uninline it for size and possible header dependency
reasons.
And create a new kernel/up.c, as a place in which to put
uniprocessor-specific code and storage. It should mirror kernel/smp.c.
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: Implementation change to remove cpumask_t from stack.
Actually change smp_call_function_mask() to smp_call_function_many().
We avoid cpumasks on the stack in this version.
(S390 has its own version, but that's going away apparently).
We have to do some dancing to figure out if 0 or 1 other cpus are in
the mask supplied and the online mask without allocating a tmp
cpumask. It's still fairly cheap.
We allocate the cpumask at the end of the call_function_data
structure: if allocation fails we fallback to smp_call_function_single
rather than using the baroque quiescing code (which needs a cpumask on
stack).
(Thanks to Hiroshi Shimamoto for spotting several bugs in previous versions!)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Mike Travis <travis@sgi.com>
Cc: Hiroshi Shimamoto <h-shimamoto@ct.jp.nec.com>
Cc: npiggin@suse.de
Cc: axboe@kernel.dk
