hrtimer preemption bug fixes
cover hrtimer preemption bug fixes
Here are a couple of hrtimer bug fixes related to preemption.
Michael Bohan (2):
hrtimer: Consider preemption when migrating hrtimer cpu_bases
hrtimer: Prevent enqueue of hrtimer on dead CPU
kernel/hrtimer.c | 7 +++++-- 1 files changed, 5 insertions(+), 2 deletions(-)
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, hosted by The Linux Foundation
PATCH 1/2 hrtimer: Consider preemption when migrating hrtimer cpu_bases
When switching to a new cpu_base in switch_hrtimer_base(), we briefly enable preemption by unlocking the cpu_base lock in two places. During this interval it's possible for the running thread to be swapped to a different CPU.
Consider the following example:
CPU #0 CPU #1 ---- ---- hrtimer_start() ...
lock_hrtimer_base()
switch_hrtimer_base()
this_cpu = 0;
target_cpu_base = 0;
raw_spin_unlock(&cpu_base->lock)
<migrate to CPU 1>
... this_cpu == 0
cpu == this_cpu
timer->base = CPU #0
timer->base != LOCAL_CPUSince the cached this_cpu is no longer accurate, we'll skip the hrtimer_check_target() check. Once we eventually go to program the hardware, we'll decide not to do so since it knows the real CPU that we're running on is not the same as the chosen base. As a consequence, we may end up missing the hrtimer's deadline.
Fix this by updating the local CPU number each time we retake a cpu_base lock in switch_hrtimer_base().
Another possibility is to disable preemption across the whole of switch_hrtimer_base. This looks suboptimal since preemption would be disabled while waiting for lock(s).
--- kernel/hrtimer.c | 4 +++- 1 files changed, 3 insertions(+), 1 deletions(-)
kernel/hrtimer.c +3 -1
@@ -225,10 +225,12 @@ again:raw_spin_unlock (&base->cpu_base->lock);raw_spin_lock (&new_base->cpu_base->lock); + this_cpu =smp_processor_id (); +if (cpu != this_cpu &&hrtimer_check_target (timer, new_base)) { - cpu = this_cpu;raw_spin_unlock (&new_base->cpu_base->lock);raw_spin_lock (&base->cpu_base->lock); + cpu =smp_processor_id (); timer->base = base; goto again; }
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, hosted by The Linux Foundation
Re: [PATCH 1/2] hrtimer: Consider preemption when migrating hrtimer cpu_bases
Quoted 10 lines · Michael Bohan
When switching to a new
cpu_baseinswitch_hrtimer_base(), we briefly enable preemption by unlocking thecpu_baselock in two places. During this interval it's possible for the running thread to be swapped to a different CPU.Consider the following example:
CPU #0 CPU #1 ---- ---- hrtimer_start() ...
Quoted 6 lines
lock_hrtimer_base()
switch_hrtimer_base()
this_cpu= 0;
target_cpu_base= 0;
raw_spin_unlock(&cpu_base->lock)
<migrate to CPU 1>
Errm. switch_hrtimer_base() is called with interrupts disabled and they stay disabled, so how exactly is the task going to be migrated?
Thanks,
tglx
Re: [PATCH 1/2] hrtimer: Consider preemption when migrating hrtimer cpu_bases
Quoted 19 lines · Thomas Gleixner
Quoted 15 lines · Michael Bohan
When switching to a new
cpu_baseinswitch_hrtimer_base(), we briefly enable preemption by unlocking thecpu_baselock in two places. During this interval it's possible for the running thread to be swapped to a different CPU.Consider the following example:
CPU #0 CPU #1 ---- ---- hrtimer_start() ... lock_hrtimer_base() switch_hrtimer_base() this_cpu = 0; target_cpu_base = 0; raw_spin_unlock(&cpu_base->lock)Quoted 1 line
<migrate to CPU 1>
Quoted 2 lines
Errm.
switch_hrtimer_base()is called with interrupts disabled and they stay disabled, so how exactly is the task going to be migrated?
My mistake - I missed that important fact. Please disregard this.
Thanks,
Mike
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, hosted by The Linux Foundation
PATCH 2/2 hrtimer: Prevent enqueue of hrtimer on dead CPU
In reply to Michael BohanWhen switching the hrtimer cpu_base, we briefly allow for preemption to become enabled by unlocking the cpu_base lock. During this time, the CPU corresponding to the new cpu_base that was selected may in fact go offline. In this scenario, the hrtimer is enqueued to a CPU that's not online, and therefore it never fires.
As an example, consider this example:
CPU #0 CPU #1
---- ----
... hrtimer_start()
lock_hrtimer_base()
switch_hrtimer_base()
cpu = hrtimer_get_target() -> 1
spin_unlock(&cpu_base->lock)
<migrate thread to CPU #0>
<offline>spin_lock(&new_base->lock)this_cpu = 0
cpu != this_cpuenqueue_hrtimer(cpu_base #1)
To prevent this scenario, verify that the CPU corresponding to the new cpu_base is indeed online before selecting it in hrtimer_switch_base(). If it's not online, fallback to using the base of the current CPU.
--- kernel/hrtimer.c | 3 ++- 1 files changed, 2 insertions(+), 1 deletions(-)
kernel/hrtimer.c +2 -1
@@ -227,7 +227,8 @@ again: this_cpu =smp_processor_id (); -if (cpu != this_cpu &&hrtimer_check_target (timer, new_base)) { +if (cpu != this_cpu && (hrtimer_check_target (timer, new_base) + || !cpu_online (cpu))) {raw_spin_unlock (&new_base->cpu_base->lock);raw_spin_lock (&base->cpu_base->lock); cpu =smp_processor_id ();
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, hosted by The Linux Foundation
Re: [PATCH 2/2] hrtimer: Prevent enqueue of hrtimer on dead CPU
Quoted 2 lines · Michael Bohan
When switching the hrtimer
cpu_base, we briefly allow for preemption to become enabled by unlocking thecpu_baselock.
No, interrupts still stay disabled and there is no way to preempt here. Can you please explain, what you are trying to do and what makes you think there is a problem in that code?
Thanks,
tglx