cpumask_and() only initializes nr_cpu_ids bits, so the (deprecated)
first_cpu() might find one of those uninitialized bits if nr_cpu_ids
is less than NR_CPUS (as it can be for CONFIG_CPUMASK_OFFSTACK).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: prevents panic from stack overflow on numa-capable machines.
Some of the "removal of stack hogs" changes in kernel/sched.c by using
node_to_cpumask_ptr were undone by the early cpumask API updates, and
causes a panic due to stack overflow. This patch undoes those changes
by using cpumask_of_node() which returns a 'const struct cpumask *'.
In addition, cpu_coregoup_map is replaced with cpu_coregroup_mask further
reducing stack usage. (Both of these updates removed 9 FIXME's!)
Also:
Pick up some remaining changes from the old 'cpumask_t' functions to
the new 'struct cpumask *' functions.
Optimize memory traffic by allocating each percpu local_cpu_mask on the
same node as the referring cpu.
Signed-off-by: Mike Travis <travis@sgi.com>
Acked-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: fix potential of rare crash
for_each_leaf_rt_rq() walks an RCU protected list (rq->leaf_rt_rq_list),
but doesn't use list_for_each_entry_rcu(). Fix this.
Signed-off-by: Bharata B Rao <bharata@linux.vnet.ibm.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Move double_lock_balance()/double_unlock_balance() higher to fix the following
with gcc-3.4.6:
CC kernel/sched.o
In file included from kernel/sched.c:1605:
kernel/sched_rt.c: In function `find_lock_lowest_rq':
kernel/sched_rt.c:914: sorry, unimplemented: inlining failed in call to 'double_unlock_balance': function body not available
kernel/sched_rt.c:1077: sorry, unimplemented: called from here
make[2]: *** [kernel/sched.o] Error 1
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: Trivial API conversion
NR_CPUS -> nr_cpu_ids
cpumask_t -> struct cpumask
sizeof(cpumask_t) -> cpumask_size()
cpumask_a = cpumask_b -> cpumask_copy(&cpumask_a, &cpumask_b)
cpu_set() -> cpumask_set_cpu()
first_cpu() -> cpumask_first()
cpumask_of_cpu() -> cpumask_of()
cpus_* -> cpumask_*
There are some FIXMEs where we all archs to complete infrastructure
(patches have been sent):
cpu_coregroup_map -> cpu_coregroup_mask
node_to_cpumask* -> cpumask_of_node
There is also one FIXME where we pass an array of cpumasks to
partition_sched_domains(): this implies knowing the definition of
'struct cpumask' and the size of a cpumask. This will be fixed in a
future patch.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: (future) size reduction for large NR_CPUS.
Dynamically allocating cpumasks (when CONFIG_CPUMASK_OFFSTACK) saves
space for small nr_cpu_ids but big CONFIG_NR_CPUS. cpumask_var_t
is just a struct cpumask for !CONFIG_CPUMASK_OFFSTACK.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: stack reduction for large NR_CPUS
Dynamically allocating cpumasks (when CONFIG_CPUMASK_OFFSTACK) saves
stack space.
We simply return if the allocation fails: since we don't use it we
could just pass NULL to cpupri_find and have it handle that.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: (future) size reduction for large NR_CPUS.
Dynamically allocating cpumasks (when CONFIG_CPUMASK_OFFSTACK) saves
space for small nr_cpu_ids but big CONFIG_NR_CPUS. cpumask_var_t
is just a struct cpumask for !CONFIG_CPUMASK_OFFSTACK.
def_root_domain is static, and so its masks are initialized with
alloc_bootmem_cpumask_var. After that, alloc_cpumask_var is used.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: trivial wrap of member accesses
This eases the transition in the next patch.
We also get rid of a temporary cpumask in find_idlest_cpu() thanks to
for_each_cpu_and, and sched_balance_self() due to getting weight before
setting sd to NULL.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
We have a test case which measures the variation in the amount of time
needed to perform a fixed amount of work on the preempt_rt kernel. We
started seeing deterioration in it's performance recently. The test
should never take more than 10 microseconds, but we started 5-10%
failure rate.
Using elimination method, we traced the problem to commit
1b12bbc747 (lockdep: re-annotate
scheduler runqueues).
When LOCKDEP is disabled, this patch only adds an additional function
call to double_unlock_balance(). Hence I inlined double_unlock_balance()
and the problem went away. Here is a patch to make this change.
Signed-off-by: Sripathi Kodi <sripathik@in.ibm.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: micro-optimization to SCHED_FIFO/RR scheduling
A very minor improvement, but might it be better to check sched_rt_runtime(rt_rq)
before taking the rt_runtime_lock?
Peter Zijlstra observes:
> Yes, I think its ok to do so.
>
> Like pointed out in the other thread, there are two races:
>
> - sched_rt_runtime() going to RUNTIME_INF, and that will be handled
> properly by sched_rt_runtime_exceeded()
>
> - sched_rt_runtime() going to !RUNTIME_INF, and here we can miss an
> accounting cycle, but I don't think that is something to worry too
> much about.
Signed-off-by: Dimitri Sivanich <sivanich@sgi.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
--
kernel/sched_rt.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
a patch from Henrik Austad did this:
>> Do not declare select_task_rq as part of sched_class when CONFIG_SMP is
>> not set.
Peter observed:
> While a proper cleanup, could you do it by re-arranging the methods so
> as to not create an additional ifdef?
Do not declare select_task_rq and some other methods as part of sched_class
when CONFIG_SMP is not set.
Also gather those methods to avoid CONFIG_SMP mess.
Idea-by: Henrik Austad <henrik.austad@gmail.com>
Signed-off-by: Li Zefan <lizf@cn.fujitsu.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: Henrik Austad <henrik@austad.us>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
While working on the new version of the code for SCHED_SPORADIC I
noticed something strange in the present throttling mechanism. More
specifically in the throttling timer handler in sched_rt.c
(do_sched_rt_period_timer()) and in rt_rq_enqueue().
The problem is that, when unthrottling a runqueue, rt_rq_enqueue() only
asks for rescheduling if the runqueue has a sched_entity associated to
it (i.e., rt_rq->rt_se != NULL).
Now, if the runqueue is the root rq (which has a rt_se = NULL)
rescheduling does not take place, and it is delayed to some undefined
instant in the future.
This imply some random bandwidth usage by the RT tasks under throttling.
For instance, setting rt_runtime_us/rt_period_us = 950ms/1000ms an RT
task will get less than 95%. In our tests we got something varying
between 70% to 95%.
Using smaller time values, e.g., 95ms/100ms, things are even worse, and
I can see values also going down to 20-25%!!
The tests we performed are simply running 'yes' as a SCHED_FIFO task,
and checking the CPU usage with top, but we can investigate thoroughly
if you think it is needed.
Things go much better, for us, with the attached patch... Don't know if
it is the best approach, but it solved the issue for us.
Signed-off-by: Dario Faggioli <raistlin@linux.it>
Signed-off-by: Michael Trimarchi <trimarchimichael@yahoo.it>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: <stable@kernel.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Lin Ming reported a 10% OLTP regression against 2.6.27-rc4.
The difference seems to come from different preemption agressiveness,
which affects the cache footprint of the workload and its effective
cache trashing.
Aggresively preempt a task if its avg overlap is very small, this should
avoid the task going to sleep and find it still running when we schedule
back to it - saving a wakeup.
Reported-by: Lin Ming <ming.m.lin@intel.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
On my tulsa x86-64 machine, kernel 2.6.25-rc5 couldn't boot randomly.
Basically, function __enable_runtime forgets to reset rt_rq->rt_throttled
to 0. When every cpu is up, per-cpu migration_thread is created and it runs
very fast, sometimes to mark the corresponding rt_rq->rt_throttled to 1 very
quickly. After all cpus are up, with below calling chain:
sched_init_smp => arch_init_sched_domains => build_sched_domains => ...
=> cpu_attach_domain => rq_attach_root => set_rq_online => ...
=> _enable_runtime
_enable_runtime is called against every rt_rq again, so rt_rq->rt_time is
reset to 0, but rt_rq->rt_throttled might be still 1. Later on function
do_sched_rt_period_timer couldn't reset it, and all RT tasks couldn't be
scheduled to run on that cpu. here is RT task migration_thread which is
woken up when a task is migrated to another cpu.
Below patch fixes it against 2.6.27-rc5.
Signed-off-by: Zhang Yanmin <yanmin_zhang@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
It fixes an accounting bug where we would continue accumulating runtime
even though the bandwidth control is disabled. This would lead to very long
throttle periods once bandwidth control gets turned on again.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
When sysctl_sched_rt_runtime is set to something other than -1 and the
CONFIG_RT_GROUP_SCHED kernel parameter is NOT enabled, we get into a state
where we see one or more CPUs idling forvever even though there are
real-time
tasks in their rt runqueue that are able to run (no longer throttled).
The sequence is:
- A real-time task is running when the timer sets the rt runqueue
to throttled, and the rt task is resched_task()ed and switched
out, and idle is switched in since there are no non-rt tasks to
run on that cpu.
- Eventually the do_sched_rt_period_timer() runs and un-throttles
the rt runqueue, but we just exit the timer interrupt and go back
to executing the idle task in the idle loop forever.
If we change the sched_rt_rq_enqueue() routine to use some of the code
from the CONFIG_RT_GROUP_SCHED enabled version of this same routine and
resched_task() the currently executing task (idle in our case) if it is
a lower priority task than the higher rt task in the now un-throttled
runqueue, the problem is no longer observed.
Signed-off-by: John Blackwood <john.blackwood@ccur.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
More extensive disable of bandwidth control. It allows sysctl_sched_rt_runtime
to disable full group bandwidth control.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
It fixes an accounting bug where we would continue accumulating runtime
even though the bandwidth control is disabled. This would lead to very long
throttle periods once bandwidth control gets turned on again.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
When we hot-unplug a cpu and rebuild the sched-domain, all cpus will be
detatched. Alex observed the case where a runqueue was stealing bandwidth
from an already disabled runqueue to satisfy its own needs.
Stop this by skipping over already disabled runqueues.
Reported-by: Alex Nixon <alex.nixon@citrix.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Tested-by: Alex Nixon <alex.nixon@citrix.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Instead of using a per-rq lock class, use the regular nesting operations.
However, take extra care with double_lock_balance() as it can release the
already held rq->lock (and therefore change its nesting class).
So what can happen is:
spin_lock(rq->lock); // this rq subclass 0
double_lock_balance(rq, other_rq);
// release rq
// acquire other_rq->lock subclass 0
// acquire rq->lock subclass 1
spin_unlock(other_rq->lock);
leaving you with rq->lock in subclass 1
So a subsequent double_lock_balance() call can try to nest a subclass 1
lock while already holding a subclass 1 lock.
Fix this by introducing double_unlock_balance() which releases the other
rq's lock, but also re-sets the subclass for this rq's lock to 0.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Fix inc_rt_tasks() to not declare variable 'rq' if it's not needed. It is
declared if CONFIG_SMP or CONFIG_RT_GROUP_SCHED, but only used if CONFIG_SMP.
This is a consequence of patch 1f11eb6a8b plus
patch 1100ac91b6.
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This is based on Linus' idea of creating cpu_active_map that prevents
scheduler load balancer from migrating tasks to the cpu that is going
down.
It allows us to simplify domain management code and avoid unecessary
domain rebuilds during cpu hotplug event handling.
Please ignore the cpusets part for now. It needs some more work in order
to avoid crazy lock nesting. Although I did simplfy and unify domain
reinitialization logic. We now simply call partition_sched_domains() in
all the cases. This means that we're using exact same code paths as in
cpusets case and hence the test below cover cpusets too.
Cpuset changes to make rebuild_sched_domains() callable from various
contexts are in the separate patch (right next after this one).
This not only boots but also easily handles
while true; do make clean; make -j 8; done
and
while true; do on-off-cpu 1; done
at the same time.
(on-off-cpu 1 simple does echo 0/1 > /sys/.../cpu1/online thing).
Suprisingly the box (dual-core Core2) is quite usable. In fact I'm typing
this on right now in gnome-terminal and things are moving just fine.
Also this is running with most of the debug features enabled (lockdep,
mutex, etc) no BUG_ONs or lockdep complaints so far.
I believe I addressed all of the Dmitry's comments for original Linus'
version. I changed both fair and rt balancer to mask out non-active cpus.
And replaced cpu_is_offline() with !cpu_active() in the main scheduler
code where it made sense (to me).
Signed-off-by: Max Krasnyanskiy <maxk@qualcomm.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Gregory Haskins <ghaskins@novell.com>
Cc: dmitry.adamushko@gmail.com
Cc: pj@sgi.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
(1) handle in a generic way all cases when a newly woken-up task is
not migratable (not just a corner case when "rt_se->nr_cpus_allowed ==
1")
(2) if current is to be preempted, then make sure "p" will be picked
up by pick_next_task_rt().
i.e. move task's group at the head of its list as well.
currently, it's not a case for the group-scheduling case as described
here: http://www.ussg.iu.edu/hypermail/linux/kernel/0807.0/0134.html
Signed-off-by: Dmitry Adamushko <dmitry.adamushko@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Gregory Haskins <ghaskins@novell.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
In file included from /mnt/build/linux-2.6/kernel/sched.c:1496:
/mnt/build/linux-2.6/kernel/sched_rt.c: In function '__enable_runtime':
/mnt/build/linux-2.6/kernel/sched_rt.c:339: warning: unused variable 'rd'
/mnt/build/linux-2.6/kernel/sched_rt.c: In function 'requeue_rt_entity':
/mnt/build/linux-2.6/kernel/sched_rt.c:692: warning: unused variable 'queue'
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
So if the group ever gets throttled, it will never wake up again.
Reported-by: "Daniel K." <dk@uw.no>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Tested-by: Daniel K. <dk@uw.no>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
So if the group ever gets throttled, it will never wake up again.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: "Daniel K." <dk@uw.no>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Reported-by: "Daniel K." <dk@uw.no>
Now we exceed the runtime and get throttled - the period rollover tick
will subtract the cpu quota from the runtime and check if we're below
quota. However with this cpu having a very small portion of the runtime
it will not refresh as fast as it should.
Therefore, also rebalance the runtime when we're throttled.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: "Daniel K." <dk@uw.no>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
In tick_task_rt() we first call update_curr_rt() which can dequeue a runqueue
due to it running out of runtime, and then we try to requeue it, of it also
having exhausted its RR quota. Obviously requeueing something that is no longer
on the runqueue will not have the expected result.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Tested-by: Daniel K. <dk@uw.no>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The bandwidth throttle code dequeues a group when it runs out of quota, and
re-queues it once the period rolls over and the quota gets refreshed.
Sadly it failed to take the hierarchy into consideration. Share more of the
enqueue/dequeue code with regular task opterations.
Also, some operations like sched_setscheduler() can dequeue/enqueue tasks that
are in throttled runqueues, we should not inadvertly re-enqueue empty runqueues
so check for that.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Tested-by: Daniel K. <dk@uw.no>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
regarding this commit: 45c01e8249
I think we can do it simpler. Please take a look at the patch below.
Instead of having 2 separate arrays (which is + ~800 bytes on x86_32 and
twice so on x86_64), let's add "exclusive" (the ones that are bound to
this CPU) tasks to the head of the queue and "shared" ones -- to the
end.
In case of a few newly woken up "exclusive" tasks, they are 'stacked'
(not queued as now), meaning that a task {i+1} is being placed in front
of the previously woken up task {i}. But I don't think that this
behavior may cause any realistic problems.
There are a couple of changes on top of this one.
(1) in check_preempt_curr_rt()
I don't think there is a need for the "pick_next_rt_entity(rq, &rq->rt)
!= &rq->curr->rt" check.
enqueue_task_rt(p) and check_preempt_curr_rt() are always called one
after another with rq->lock being held so the following check
"p->rt.nr_cpus_allowed == 1 && rq->curr->rt.nr_cpus_allowed != 1" should
be enough (well, just its left part) to guarantee that 'p' has been
queued in front of the 'curr'.
(2) in set_cpus_allowed_rt()
I don't thinks there is a need for requeue_task_rt() here.
Perhaps, the only case when 'requeue' (+ reschedule) might be useful is
as follows:
i) weight == 1 && cpu_isset(task_cpu(p), *new_mask)
i.e. a task is being bound to this CPU);
ii) 'p' != rq->curr
but here, 'p' has already been on this CPU for a while and was not
migrated. i.e. it's possible that 'rq->curr' would not have high chances
to be migrated right at this particular moment (although, has chance in
a bit longer term), should we allow it to be preempted.
Anyway, I think we should not perhaps make it more complex trying to
address some rare corner cases. For instance, that's why a single queue
approach would be preferable. Unless I'm missing something obvious, this
approach gives us similar functionality at lower cost.
Verified only compilation-wise.
(Almost)-Signed-off-by: Dmitry Adamushko <dmitry.adamushko@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Cliff Wickman wrote:
> I built an ia64 kernel from Andrew's tree (2.6.26-rc2-mm1)
> and get a very predictable hotplug cpu problem.
> billberry1:/tmp/cpw # ./dis
> disabled cpu 17
> enabled cpu 17
> billberry1:/tmp/cpw # ./dis
> disabled cpu 17
> enabled cpu 17
> billberry1:/tmp/cpw # ./dis
>
> The script that disables the cpu always hangs (unkillable)
> on the 3rd attempt.
>
> And a bit further:
> The kstopmachine thread always sits on the run queue (real time) for about
> 30 minutes before running.
this fix solves some (but not all) issues between CPU hotplug and
RT bandwidth throttling.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
this patch was not built on !SMP:
kernel/sched_rt.c: In function 'inc_rt_tasks':
kernel/sched_rt.c:404: error: 'struct rq' has no member named 'online'
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The RT folks over at RedHat found an issue w.r.t. hotplug support which
was traced to problems with the cpupri infrastructure in the scheduler:
https://bugzilla.redhat.com/show_bug.cgi?id=449676
This bug affects 23-rt12+, 24-rtX, 25-rtX, and sched-devel. This patch
applies to 25.4-rt4, though it should trivially apply to most cpupri enabled
kernels mentioned above.
It turned out that the issue was that offline cpus could get inadvertently
registered with cpupri so that they were erroneously selected during
migration decisions. The end result would be an OOPS as the offline cpu
had tasks routed to it.
This patch generalizes the old join/leave domain interface into an
online/offline interface, and adjusts the root-domain/hotplug code to
utilize it.
I was able to easily reproduce the issue prior to this patch, and am no
longer able to reproduce it after this patch. I can offline cpus
indefinately and everything seems to be in working order.
Thanks to Arnaldo (acme), Thomas, and Peter for doing the legwork to point
me in the right direction. Also thank you to Peter for reviewing the
early iterations of this patch.
Signed-off-by: Gregory Haskins <ghaskins@novell.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
The current code use a linear algorithm which causes scaling issues
on larger SMP machines. This patch replaces that algorithm with a
2-dimensional bitmap to reduce latencies in the wake-up path.
Signed-off-by: Gregory Haskins <ghaskins@novell.com>
Acked-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Dmitry Adamushko pointed out a known flaw in the rt-balancing algorithm
that could allow suboptimal balancing if a non-migratable task gets
queued behind a running migratable one. It is discussed in this thread:
http://lkml.org/lkml/2008/4/22/296
This issue has been further exacerbated by a recent checkin to
sched-devel (git-id 5eee63a5ebc19a870ac40055c0be49457f3a89a3).
>From a pure priority standpoint, the run-queue is doing the "right"
thing. Using Dmitry's nomenclature, if T0 is on cpu1 first, and T1
wakes up at equal or lower priority (affined only to cpu1) later, it
*should* wait for T0 to finish. However, in reality that is likely
suboptimal from a system perspective if there are other cores that
could allow T0 and T1 to run concurrently. Since T1 can not migrate,
the only choice for higher concurrency is to try to move T0. This is
not something we addessed in the recent rt-balancing re-work.
This patch tries to enhance the balancing algorithm by accomodating this
scenario. It accomplishes this by incorporating the migratability of a
task into its priority calculation. Within a numerical tsk->prio, a
non-migratable task is logically higher than a migratable one. We
maintain this by introducing a new per-priority queue (xqueue, or
exclusive-queue) for holding non-migratable tasks. The scheduler will
draw from the xqueue over the standard shared-queue (squeue) when
available.
There are several details for utilizing this properly.
1) During task-wake-up, we not only need to check if the priority
preempts the current task, but we also need to check for this
non-migratable condition. Therefore, if a non-migratable task wakes
up and sees an equal priority migratable task already running, it
will attempt to preempt it *if* there is a likelyhood that the
current task will find an immediate home.
2) Tasks only get this non-migratable "priority boost" on wake-up. Any
requeuing will result in the non-migratable task being queued to the
end of the shared queue. This is an attempt to prevent the system
from being completely unfair to migratable tasks during things like
SCHED_RR timeslicing.
I am sure this patch introduces potentially "odd" behavior if you
concoct a scenario where a bunch of non-migratable threads could starve
migratable ones given the right pattern. I am not yet convinced that
this is a problem since we are talking about tasks of equal RT priority
anyway, and there never is much in the way of guarantees against
starvation under that scenario anyway. (e.g. you could come up with a
similar scenario with a specific timing environment verses an affinity
environment). I can be convinced otherwise, but for now I think this is
"ok".
Signed-off-by: Gregory Haskins <ghaskins@novell.com>
CC: Dmitry Adamushko <dmitry.adamushko@gmail.com>
CC: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Yanmin Zhang reported:
Comparing with 2.6.25, volanoMark has big regression with kernel 2.6.26-rc1.
It's about 50% on my 8-core stoakley, 16-core tigerton, and Itanium Montecito.
With bisect, I located the following patch:
| 18d95a2832 is first bad commit
| commit 18d95a2832
| Author: Peter Zijlstra <a.p.zijlstra@chello.nl>
| Date: Sat Apr 19 19:45:00 2008 +0200
|
| sched: fair-group: SMP-nice for group scheduling
Revert it so that we get v2.6.25 behavior.
Bisected-by: Yanmin Zhang <yanmin_zhang@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Change references from for_each_cpu_mask to for_each_cpu_mask_nr
where appropriate
Reviewed-by: Paul Jackson <pj@sgi.com>
Reviewed-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Mike Travis <travis@sgi.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Dmitry Adamushko pointed out a logic error in task_wake_up_rt() where we
will always evaluate to "true". You can find the thread here:
http://lkml.org/lkml/2008/4/22/296
In reality, we only want to try to push tasks away when a wake up request is
not going to preempt the current task. So lets fix it.
Note: We introduce test_tsk_need_resched() instead of open-coding the flag
check so that the merge-conflict with -rt should help remind us that we
may need to support NEEDS_RESCHED_DELAYED in the future, too.
Signed-off-by: Gregory Haskins <ghaskins@novell.com>
CC: Dmitry Adamushko <dmitry.adamushko@gmail.com>
CC: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The C files are included directly in sched.c, so they are
effectively static.
Signed-off-by: Harvey Harrison <harvey.harrison@gmail.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Now that the group hierarchy can have an arbitrary depth the O(n^2) nature
of RT task dequeues will really hurt. Optimize this by providing space to
store the tree path, so we can walk it the other way.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Implement SMP nice support for the full group hierarchy.
On each load-balance action, compile a sched_domain wide view of the full
task_group tree. We compute the domain wide view when walking down the
hierarchy, and readjust the weights when walking back up.
After collecting and readjusting the domain wide view, we try to balance the
tasks within the task_groups. The current approach is a naively balance each
task group until we've moved the targeted amount of load.
Inspired by Srivatsa Vaddsgiri's previous code and Abhishek Chandra's H-SMP
paper.
XXX: there will be some numerical issues due to the limited nature of
SCHED_LOAD_SCALE wrt to representing a task_groups influence on the
total weight. When the tree is deep enough, or the task weight small
enough, we'll run out of bits.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
CC: Abhishek Chandra <chandra@cs.umn.edu>
CC: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Rusty Russell