lkml.org 
[lkml]   [2013]   [Oct]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH 0/6] Optimize the cpu hotplug locking -v2
On Thu, Oct 10, 2013 at 11:49:15AM -0700, Linus Torvalds wrote:
> On Thu, Oct 10, 2013 at 11:34 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> >
> > But my point is that even though there aren't many of these today; with
> > the growing number of cpus in 'commodity' hardware you want to move away
> > from using preempt_disable() as hotplug lock.
>
> Umm.
>
> Wasn't this pretty much the argument for the COMPLETELY F*CKED UP
> change to make the vma locking use a semaphore?

Not entirely; but fair enough, I did screw up there.

> The whole "it's more convenient to use sleeping locks" argument is
> PURE AND UTTER SHIT when it comes to really core code. We are *much*
> better off saying "this is core, we care so deeply about performance
> that you had better use your brain before you do this".
>
> Seriously. Your argument is bad, but more importantly, it is
> *dangerously* bad. It's crap that results in bad code: and the bad
> code is almost impossible to fix up later, because once you encourage
> people to do the easy thing, they'll do so.

Right, I fell face first into this trap. The existence of
{get,put}_online_cpus() made me stop thinking and use it.

As a penance I'll start by removing all get_online_cpus() usage from the
scheduler.

---
Subject: sched: Remove get_online_cpus() usage

Remove get_online_cpus() usage from the scheduler; there's 4 sites that
use it:

- sched_init_smp(); where its completely superfluous since we're in
'early' boot and there simply cannot be any hotplugging.

- sched_getaffinity(); we already take a raw spinlock to protect the
task cpus_allowed mask, this disables preemption and therefore
also stabilizes cpu_online_mask as that's modified using
stop_machine. However switch to active mask for symmetry with
sched_setaffinity()/set_cpus_allowed_ptr(). We guarantee active
mask stability by inserting sync_rcu/sched() into _cpu_down.

- sched_setaffinity(); we don't appear to need get_online_cpus()
either, there's two sites where hotplug appears relevant:
* cpuset_cpus_allowed(); for the !cpuset case we use possible_mask,
for the cpuset case we hold task_lock, which is a spinlock and
thus for mainline disables preemption (might cause pain on RT).
* set_cpus_allowed_ptr(); Holds all scheduler locks and thus has
preemption properly disabled; also it already deals with hotplug
races explicitly where it releases them.

- migrate_swap(); we can make stop_two_cpus() do the heavy lifting for
us with a little trickery. By adding a sync_sched/rcu() after the
CPU_DOWN_PREPARE notifier we can provide preempt/rcu guarantees for
cpu_active_mask. Use these to validate that both our cpus are active
when queueing the stop work before we queue the stop_machine works
for take_cpu_down().

Signed-off-by: Peter Zijlstra <peterz@infradead.org>
---
kernel/cpu.c | 17 +++++++++++++++++
kernel/sched/core.c | 20 ++++++++++----------
kernel/stop_machine.c | 26 +++++++++++++++++++++-----
3 files changed, 48 insertions(+), 15 deletions(-)

diff --git a/kernel/cpu.c b/kernel/cpu.c
index d7f07a2da5a6..63aa50d7ce1e 100644
--- a/kernel/cpu.c
+++ b/kernel/cpu.c
@@ -308,6 +308,23 @@ static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
}
smpboot_park_threads(cpu);

+ /*
+ * By now we've cleared cpu_active_mask, wait for all preempt-disabled
+ * and RCU users of this state to go away such that all new such users
+ * will observe it.
+ *
+ * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
+ * not imply sync_sched(), so explicitly call both.
+ */
+#ifdef CONFIG_PREEMPT
+ synchronize_sched();
+#endif
+ synchronize_rcu();
+
+ /*
+ * So now all preempt/rcu users must observe !cpu_active().
+ */
+
err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
if (err) {
/* CPU didn't die: tell everyone. Can't complain. */
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 0c3feebcf112..498a5e5a53f5 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -1081,8 +1081,6 @@ int migrate_swap(struct task_struct *cur, struct task_struct *p)
struct migration_swap_arg arg;
int ret = -EINVAL;

- get_online_cpus();
-
arg = (struct migration_swap_arg){
.src_task = cur,
.src_cpu = task_cpu(cur),
@@ -1093,6 +1091,10 @@ int migrate_swap(struct task_struct *cur, struct task_struct *p)
if (arg.src_cpu == arg.dst_cpu)
goto out;

+ /*
+ * These three tests are all lockless; this is OK since all of them
+ * will be re-checked with proper locks held further down the line.
+ */
if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
goto out;

@@ -1105,7 +1107,6 @@ int migrate_swap(struct task_struct *cur, struct task_struct *p)
ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);

out:
- put_online_cpus();
return ret;
}

@@ -3706,7 +3707,6 @@ long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
struct task_struct *p;
int retval;

- get_online_cpus();
rcu_read_lock();

p = find_process_by_pid(pid);
@@ -3769,7 +3769,6 @@ long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
free_cpumask_var(cpus_allowed);
out_put_task:
put_task_struct(p);
- put_online_cpus();
return retval;
}

@@ -3814,7 +3813,6 @@ long sched_getaffinity(pid_t pid, struct cpumask *mask)
unsigned long flags;
int retval;

- get_online_cpus();
rcu_read_lock();

retval = -ESRCH;
@@ -3827,12 +3825,11 @@ long sched_getaffinity(pid_t pid, struct cpumask *mask)
goto out_unlock;

raw_spin_lock_irqsave(&p->pi_lock, flags);
- cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
+ cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
raw_spin_unlock_irqrestore(&p->pi_lock, flags);

out_unlock:
rcu_read_unlock();
- put_online_cpus();

return retval;
}
@@ -6490,14 +6487,17 @@ void __init sched_init_smp(void)

sched_init_numa();

- get_online_cpus();
+ /*
+ * There's no userspace yet to cause hotplug operations; hence all the
+ * cpu masks are stable and all blatant races in the below code cannot
+ * happen.
+ */
mutex_lock(&sched_domains_mutex);
init_sched_domains(cpu_active_mask);
cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
if (cpumask_empty(non_isolated_cpus))
cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
mutex_unlock(&sched_domains_mutex);
- put_online_cpus();

hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
diff --git a/kernel/stop_machine.c b/kernel/stop_machine.c
index 32a6c44d8f78..a6eb6d519284 100644
--- a/kernel/stop_machine.c
+++ b/kernel/stop_machine.c
@@ -234,11 +234,13 @@ static void irq_cpu_stop_queue_work(void *arg)
*/
int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
{
- int call_cpu;
struct cpu_stop_done done;
struct cpu_stop_work work1, work2;
struct irq_cpu_stop_queue_work_info call_args;
- struct multi_stop_data msdata = {
+ struct multi_stop_data msdata;
+
+ preempt_disable();
+ msdata = (struct multi_stop_date){
.fn = fn,
.data = arg,
.num_threads = 2,
@@ -262,16 +264,30 @@ int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *
set_state(&msdata, MULTI_STOP_PREPARE);

/*
+ * If we observe both CPUs active we know _cpu_down() cannot yet have
+ * queued its stop_machine works and therefore ours will get executed
+ * first. Or its not either one of our CPUs that's getting unplugged,
+ * in which case we don't care.
+ *
+ * This relies on the stopper workqueues to be FIFO.
+ */
+ if (!cpu_active(cpu1) || !cpu_active(cpu2)) {
+ preempt_enable();
+ return -ENOENT;
+ }
+
+ /*
* Queuing needs to be done by the lowest numbered CPU, to ensure
* that works are always queued in the same order on every CPU.
* This prevents deadlocks.
*/
- call_cpu = min(cpu1, cpu2);
-
- smp_call_function_single(call_cpu, &irq_cpu_stop_queue_work,
+ smp_call_function_single(min(cpu1, cpu2),
+ &irq_cpu_stop_queue_work,
&call_args, 0);
+ preempt_enable();

wait_for_completion(&done.completion);
+
return done.executed ? done.ret : -ENOENT;
}


\
 
 \ /
  Last update: 2013-10-11 15:01    [W:0.115 / U:0.372 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site