Spin more agressively. This is less fair but also markedly faster.
The numbers:
* dbench 50 (higher is better):
spin 1282MB/s
v10 548MB/s
v10 no wait 1868MB/s
* 4k creates (numbers in files/second higher is better):
spin avg 200.60 median 193.20 std 19.71 high 305.93 low 186.82
v10 avg 180.94 median 175.28 std 13.91 high 229.31 low 168.73
v10 no wait avg 232.18 median 222.38 std 22.91 high 314.66 low 209.12
* File stats (numbers in seconds, lower is better):
spin 2.27s
v10 5.1s
v10 no wait 1.6s
( The source changes are smaller than they look, I just moved the
need_resched checks in __mutex_lock_common after the cmpxchg. )
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Change mutex contention behaviour such that it will sometimes busy wait on
acquisition - moving its behaviour closer to that of spinlocks.
This concept got ported to mainline from the -rt tree, where it was originally
implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins.
Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50)
gave a 345% boost for VFS scalability on my testbox:
# ./test-mutex-shm V 16 10 | grep "^avg ops"
avg ops/sec: 296604
# ./test-mutex-shm V 16 10 | grep "^avg ops"
avg ops/sec: 85870
The key criteria for the busy wait is that the lock owner has to be running on
a (different) cpu. The idea is that as long as the owner is running, there is a
fair chance it'll release the lock soon, and thus we'll be better off spinning
instead of blocking/scheduling.
Since regular mutexes (as opposed to rtmutexes) do not atomically track the
owner, we add the owner in a non-atomic fashion and deal with the races in
the slowpath.
Furthermore, to ease the testing of the performance impact of this new code,
there is means to disable this behaviour runtime (without having to reboot
the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y),
by issuing the following command:
# echo NO_OWNER_SPIN > /debug/sched_features
This command re-enables spinning again (this is also the default):
# echo OWNER_SPIN > /debug/sched_features
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The problem is that dropping the spinlock right before schedule is a voluntary
preemption point and can cause a schedule, right after which we schedule again.
Fix this inefficiency by keeping preemption disabled until we schedule, do this
by explicity disabling preemption and providing a schedule() variant that
assumes preemption is already disabled.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Remove a local variable by combining an assingment and test in one.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Ingo found a build error in the scheduler when RT_GROUP_SCHED was
enabled, but SMP was not. This patch rearranges the code such
that it is a little more streamlined and compiles under all permutations
of SMP, UP and RT_GROUP_SCHED. It was boot tested on my 4-way x86_64
and it still passes preempt-test.
Signed-off-by: Gregory Haskins <ghaskins@novell.com>
Remove the -ENOSYS implementation for !CONFIG_PRINTK and use
the cond_syscall infrastructure instead.
Acked-by: Kyle McMartin <kyle@redhat.com>
Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Convert all system calls to return a long. This should be a NOP since all
converted types should have the same size anyway.
With the exception of sys_exit_group which returned void. But that doesn't
matter since the system call doesn't return.
Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Impact: new tracer
The workqueue tracer provides some statistical informations
about each cpu workqueue thread such as the number of the
works inserted and executed since their creation. It can help
to evaluate the amount of work each of them have to perform.
For example it can help a developer to decide whether he should
choose a per cpu workqueue instead of a singlethreaded one.
It only traces statistical informations for now but it will probably later
provide event tracing too.
Such a tracer could help too, and be improved, to help rt priority sorted
workqueue development.
To have a snapshot of the workqueues state at any time, just do
cat /debugfs/tracing/trace_stat/workqueues
Ie:
1 125 125 reiserfs/1
1 0 0 scsi_tgtd/1
1 0 0 aio/1
1 0 0 ata/1
1 114 114 kblockd/1
1 0 0 kintegrityd/1
1 2147 2147 events/1
0 0 0 kpsmoused
0 105 105 reiserfs/0
0 0 0 scsi_tgtd/0
0 0 0 aio/0
0 0 0 ata_aux
0 0 0 ata/0
0 0 0 cqueue
0 0 0 kacpi_notify
0 0 0 kacpid
0 149 149 kblockd/0
0 0 0 kintegrityd/0
0 1000 1000 khelper
0 2270 2270 events/0
Changes in V2:
_ Drop the static array based on NR_CPU and dynamically allocate the stat array
with num_possible_cpus() and other cpu mask facilities....
_ Trace workqueue insertion at a bit lower level (insert_work instead of queue_work) to handle
even the workqueue barriers.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: tracing's Api change
Currently, the stat tracing depends on the events tracing.
When you switch to a new tracer, the stats files of the previous tracer
will disappear. But it's more scalable to separate those two engines.
This way, we can keep the stat files of one or several tracers when we
want, without bothering of multiple tracer stat files or tracer switching.
To build/destroys its stats files, a tracer just have to call
register_stat_tracer/unregister_stat_tracer everytimes it wants to.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Define FTRACE_ADDR. In IA64, a function pointer isn't a 'unsigned long' but a
'struct {unsigned long ip, unsigned long gp}'.
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
In IA64, a function pointer isn't a 'unsigned long' but a
'struct {unsigned long ip, unsigned long gp}'. MCOUNT_ADDR is determined
at link time not compile time, so explictly ignore kernel/trace/ftrace.o
in recordmcount.pl.
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
At run-time, if softlockup_thresh is changed to a much lower value,
touch_timestamp is likely to be much older than the new softlock_thresh.
This will cause a false softlockup to be detected. If softlockup_panic
is enabled, the system will panic.
The fix is to touch all watchdogs before changing softlockup_thresh.
Signed-off-by: Mandeep Singh Baines <msb@google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: reduce memory footprint
add __cpuinit to rcu_init_percpu_data(), and this function's text
will be discarded after boot when !CONFIG_HOTPLUG_CPU.
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: New perf_counter features
A pinned counter group is one that the user wants to have on the CPU
whenever possible, i.e. whenever the associated task is running, for
a per-task group, or always for a per-cpu group. If the system
cannot satisfy that, it puts the group into an error state where
it is not scheduled any more and reads from it return EOF (i.e. 0
bytes read). The group can be released from error state and made
readable again using prctl(PR_TASK_PERF_COUNTERS_ENABLE). When we
have finer-grained enable/disable controls on counters we'll be able
to reset the error state on individual groups.
An exclusive group is one that the user wants to be the only group
using the CPU performance monitor hardware whenever it is on. The
counter group scheduler will not schedule an exclusive group if there
are already other groups on the CPU and will not schedule other groups
onto the CPU if there is an exclusive group scheduled (that statement
does not apply to groups containing only software counters, which can
always go on and which do not prevent an exclusive group from going on).
With an exclusive group, we will be able to let users program PMU
registers at a low level without the concern that those settings will
perturb other measurements.
Along the way this reorganizes things a little:
- is_software_counter() is moved to perf_counter.h.
- cpuctx->active_oncpu now records the number of hardware counters on
the CPU, i.e. it now excludes software counters. Nothing was reading
cpuctx->active_oncpu before, so this change is harmless.
- A new cpuctx->exclusive field records whether we currently have an
exclusive group on the CPU.
- counter_sched_out moves higher up in perf_counter.c and gets called
from __perf_counter_remove_from_context and __perf_counter_exit_task,
where we used to have essentially the same code.
- __perf_counter_sched_in now goes through the counter list twice, doing
the pinned counters in the first loop and the non-pinned counters in
the second loop, in order to give the pinned counters the best chance
to be scheduled in.
Note that only a group leader can be exclusive or pinned, and that
attribute applies to the whole group. This avoids some awkwardness in
some corner cases (e.g. where a group leader is closed and the other
group members get added to the context list). If we want to relax that
restriction later, we can, and it is easier to relax a restriction than
to apply a new one.
This doesn't yet handle the case where a pinned counter is inherited
and goes into error state in the child - the error state is not
propagated up to the parent when the child exits, and arguably it
should.
Signed-off-by: Paul Mackerras <paulus@samba.org>
This makes sure that we call the platform-specific ppc_md.enable_pmcs
function on each CPU before we try to use the PMU on that CPU. If the
CPU goes off-line and then on-line, we need to do the enable_pmcs call
again, so we use the hw_perf_counter_setup hook to ensure that. It gets
called as each CPU comes online, but it isn't called on the CPU that is
coming up, so this adds the CPU number as an argument to it (there were
no non-empty instances of hw_perf_counter_setup before).
This also arranges to set the pmcregs_in_use field of the lppaca (data
structure shared with the hypervisor) on each CPU when we are using the
PMU and clear it when we are not. This allows the hypervisor to optimize
partition switches by not saving/restoring the PMU registers when we
aren't using the PMU.
Signed-off-by: Paul Mackerras <paulus@samba.org>
* 'core-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
smp_call_function_single(): be slightly less stupid, fix#2
lockdep, mm: fix might_fault() annotation
At 37000 feet somewhere near Greenland I woke up from a half-sleep with the
realisation that __lowest_in_progress() is buggy. After landing I checked
and there were indeed 2 problems with it; this patch fixes both:
* The order of the list checks was wrong
* The locking was not correct.
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This reverts commit 7317d7b87e.
This has been reported (and bisected) by Alexey Zaytsev and
Kamalesh Babulal to produce annoying warnings during bootup
on both x86 and powerpc.
kernel_locked() is not a valid test in IRQ context (we update the
BKL's ->lock_depth and the preempt count separately and non-atomicalyy),
so we cannot put it into the generic preempt debugging checks which
can run in IRQ contexts too.
Reported-and-bisected-by: Alexey Zaytsev <alexey.zaytsev@gmail.com>
Reported-and-bisected-by: Kamalesh Babulal <kamalesh@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Software counters aren't subject to the limitations imposed by the
fixed number of hardware counter registers, so there is no reason not
to enable them all in __perf_counter_sched_in. Previously we used to
break out of the loop when we got to a group that wouldn't fit on the
PMU; with this we continue through the list but only schedule in
software counters (or groups containing only software counters) from
there on.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Ingo Molnar wrote:
> All non-x86 architectures fail to build:
>
> In file included from /home/mingo/tip/include/linux/random.h:11,
> from /home/mingo/tip/include/linux/stackprotector.h:6,
> from /home/mingo/tip/init/main.c:17:
> /home/mingo/tip/include/linux/irqnr.h:26:63: error: asm/irq_vectors.h: No such file or directory
Do not include asm/irq_vectors.h in generic code - it's not available
on all architectures.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: reduce memory usage.
Allocate kstat_irqs_legacy based on nr_cpu_ids to deal with this
memory usage bump when NR_CPUS bumped from 128 to 4096:
8192 +253952 262144 +3100% kstat_irqs_legacy(.bss)
This is only when CONFIG_SPARSE_IRQS=y.
Signed-off-by: Mike Travis <travis@sgi.com>
Impact: Reduce memory usage.
This is the second half of the changes to make the irq_desc_ptrs be
variable sized based on nr_cpu_ids. This is done by adding a new
"max_nr_irqs" macro to irq_vectors.h (and a dummy in irqnr.h) to
return a max NR_IRQS value based on NR_CPUS or nr_cpu_ids.
This necessitated moving the define of MAX_IO_APICS to a separate
file (asm/apicnum.h) so it could be included without the baggage
of the other asm/apicdef.h declarations.
Signed-off-by: Mike Travis <travis@sgi.com>
Impact: allocate irq_desc_ptrs in preparation for making it variable-sized.
This addresses this memory usage bump when NR_CPUS bumped from 128 to 4096:
34816 +229376 264192 +658% irq_desc_ptrs(.data.read_mostly)
The patch is split into two parts, the first simply allocates the
irq_desc_ptrs array. Then next will deal with making it variable.
This is only when CONFIG_SPARSE_IRQS=y.
Signed-off-by: Mike Travis <travis@sgi.com>
