lkml.org 
[lkml]   [2018]   [Feb]   [14]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH V3 04/10] sched/core: Prevent race condition between cpuset and __sched_setscheduler()
On 14/02/18 11:49, Juri Lelli wrote:
> On 14/02/18 11:36, Juri Lelli wrote:
> > Hi Mathieu,
> >
> > On 13/02/18 13:32, Mathieu Poirier wrote:
> > > No synchronisation mechanism exist between the cpuset subsystem and calls
> > > to function __sched_setscheduler(). As such it is possible that new root
> > > domains are created on the cpuset side while a deadline acceptance test
> > > is carried out in __sched_setscheduler(), leading to a potential oversell
> > > of CPU bandwidth.
> > >
> > > By making available the cpuset_mutex to the core scheduler it is possible
> > > to prevent situations such as the one described above from happening.
> > >
> > > Signed-off-by: Mathieu Poirier <mathieu.poirier@linaro.org>
> > > ---
> >
> > [...]
> >
> > > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > > index f727c3d0064c..0d8badcf1f0f 100644
> > > --- a/kernel/sched/core.c
> > > +++ b/kernel/sched/core.c
> > > @@ -4176,6 +4176,13 @@ static int __sched_setscheduler(struct task_struct *p,
> > > }
> > >
> > > /*
> > > + * Make sure we don't race with the cpuset subsystem where root
> > > + * domains can be rebuilt or modified while operations like DL
> > > + * admission checks are carried out.
> > > + */
> > > + cpuset_lock();
> > > +
> > > + /*
> >
> > Mmm, I'm afraid we can't do this. __sched_setscheduler might be called
> > from interrupt contex by normalize_rt_tasks().
>
> Maybe conditionally grabbing it if pi is true could do? I guess we don't
> care much about domains when sysrq.

Ops.. just got this. :/

--->8---
[ 0.020203] ======================================================
[ 0.020946] WARNING: possible circular locking dependency detected
[ 0.021000] 4.16.0-rc1+ #64 Not tainted
[ 0.021000] ------------------------------------------------------
[ 0.021000] swapper/0/1 is trying to acquire lock:
[ 0.021000] (cpu_hotplug_lock.rw_sem){.+.+}, at: [<000000007164d41d>] smpboot_register_percpu_thread_cpumask+0x2d/0x100
[ 0.021000]
[ 0.021000] but task is already holding lock:
[ 0.021000] (cpuset_mutex){+.+.}, at: [<000000008529a52c>] __sched_setscheduler+0xb5/0x8b0
[ 0.021000]
[ 0.021000] which lock already depends on the new lock.
[ 0.021000]
[ 0.021000]
[ 0.021000] the existing dependency chain (in reverse order) is:
[ 0.021000]
[ 0.021000] -> #2 (cpuset_mutex){+.+.}:
[ 0.021000] __sched_setscheduler+0xb5/0x8b0
[ 0.021000] _sched_setscheduler+0x6c/0x80
[ 0.021000] __kthread_create_on_node+0x10e/0x170
[ 0.021000] kthread_create_on_node+0x37/0x40
[ 0.021000] kthread_create_on_cpu+0x27/0x90
[ 0.021000] __smpboot_create_thread.part.3+0x64/0xe0
[ 0.021000] smpboot_register_percpu_thread_cpumask+0x91/0x100
[ 0.021000] spawn_ksoftirqd+0x37/0x40
[ 0.021000] do_one_initcall+0x3b/0x160
[ 0.021000] kernel_init_freeable+0x118/0x258
[ 0.021000] kernel_init+0xa/0x100
[ 0.021000] ret_from_fork+0x3a/0x50
[ 0.021000]
[ 0.021000] -> #1 (smpboot_threads_lock){+.+.}:
[ 0.021000] smpboot_register_percpu_thread_cpumask+0x3b/0x100
[ 0.021000] spawn_ksoftirqd+0x37/0x40
[ 0.021000] do_one_initcall+0x3b/0x160
[ 0.021000] kernel_init_freeable+0x118/0x258
[ 0.021000] kernel_init+0xa/0x100
[ 0.021000] ret_from_fork+0x3a/0x50
[ 0.021000]
[ 0.021000] -> #0 (cpu_hotplug_lock.rw_sem){.+.+}:
[ 0.021000] cpus_read_lock+0x3e/0x80
[ 0.021000] smpboot_register_percpu_thread_cpumask+0x2d/0x100
[ 0.021000] lockup_detector_init+0x3e/0x74
[ 0.021000] kernel_init_freeable+0x146/0x258
[ 0.021000] kernel_init+0xa/0x100
[ 0.021000] ret_from_fork+0x3a/0x50
[ 0.021000]
[ 0.021000] other info that might help us debug this:
[ 0.021000]
[ 0.021000] Chain exists of:
[ 0.021000] cpu_hotplug_lock.rw_sem --> smpboot_threads_lock --> cpuset_mutex
[ 0.021000]
[ 0.021000] Possible unsafe locking scenario:
[ 0.021000]
[ 0.021000] CPU0 CPU1
[ 0.021000] ---- ----
[ 0.021000] lock(cpuset_mutex);
[ 0.021000] lock(smpboot_threads_lock);
[ 0.021000] lock(cpuset_mutex);
[ 0.021000] lock(cpu_hotplug_lock.rw_sem);
[ 0.021000]
[ 0.021000] *** DEADLOCK ***
[ 0.021000]
[ 0.021000] 1 lock held by swapper/0/1:
[ 0.021000] #0: (cpuset_mutex){+.+.}, at: [<000000008529a52c>] __sched_setscheduler+0xb5/0x8b0
[ 0.021000]
[ 0.021000] stack backtrace:
[ 0.021000] CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.16.0-rc1+ #64
[ 0.021000] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-2.fc27 04/01/2014
[ 0.021000] Call Trace:
[ 0.021000] dump_stack+0x85/0xc5
[ 0.021000] print_circular_bug.isra.38+0x1ce/0x1db
[ 0.021000] __lock_acquire+0x1278/0x1320
[ 0.021000] ? sched_clock_local+0x12/0x80
[ 0.021000] ? lock_acquire+0x9f/0x1f0
[ 0.021000] lock_acquire+0x9f/0x1f0
[ 0.021000] ? smpboot_register_percpu_thread_cpumask+0x2d/0x100
[ 0.021000] cpus_read_lock+0x3e/0x80
[ 0.021000] ? smpboot_register_percpu_thread_cpumask+0x2d/0x100
[ 0.021000] smpboot_register_percpu_thread_cpumask+0x2d/0x100
[ 0.021000] ? set_debug_rodata+0x11/0x11
[ 0.021000] lockup_detector_init+0x3e/0x74
[ 0.021000] kernel_init_freeable+0x146/0x258
[ 0.021000] ? rest_init+0xd0/0xd0
[ 0.021000] kernel_init+0xa/0x100
[ 0.021000] ret_from_fork+0x3a/0x50

\
 
 \ /
  Last update: 2018-02-14 12:28    [W:0.086 / U:5.556 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site