The kernel test robot has reported that the use of __this_cpu_add()
causes bug messages like:
BUG: using __this_cpu_add() in preemptible [00000000] code: ...
This is only an issue on preempt kernel where preemption can happen
in the middle of the multi-instruction percpu operation. It is not an
issue on x86 as the percpu operation is a single instruction. The lock
events code is updated to use the slower this_cpu_add() for non-x86
preempt kernel or when CONFIG_DEBUG_PREEMPT is defined.
Fixes: a8654596f0371 ("locking/rwsem: Enable lock event counting")
Signed-off-by: Waiman Long <[email protected]>
---
kernel/locking/lock_events.h | 27 +++++++++++++++++++++++++--
1 file changed, 25 insertions(+), 2 deletions(-)
diff --git a/kernel/locking/lock_events.h b/kernel/locking/lock_events.h
index feb1acc54611..2b6c8b7588dc 100644
--- a/kernel/locking/lock_events.h
+++ b/kernel/locking/lock_events.h
@@ -30,13 +30,36 @@ enum lock_events {
*/
DECLARE_PER_CPU(unsigned long, lockevents[lockevent_num]);
+/*
+ * The purpose of the lock event counting subsystem is to provide a low
+ * overhead way to record the number of specific locking events by using
+ * percpu counters. It is the percpu sum that matters, not specifically
+ * how many of them happens in each cpu.
+ *
+ * In !preempt kernel, we can just use __this_cpu_{inc|add}() as preemption
+ * won't happen in the middle of the percpu operation. In preempt kernel,
+ * it depends on whether the percpu operation is atomic (1 instruction)
+ * or not. We know x86 generates a single instruction to do percpu op, but
+ * we can't guarantee that for other architectures. We also need to use
+ * the slower this_cpu_{inc|add}() when CONFIG_DEBUG_PREEMPT is defined
+ * to make the checking code happy.
+ */
+#if defined(CONFIG_PREEMPT) && \
+ (defined(CONFIG_DEBUG_PREEMPT) || !defined(CONFIG_X86))
+#define lockevent_percpu_inc(x) this_cpu_inc(x)
+#define lockevent_percpu_add(x, v) this_cpu_add(x, v)
+#else
+#define lockevent_percpu_inc(x) __this_cpu_inc(x)
+#define lockevent_percpu_add(x, v) __this_cpu_add(x, v)
+#endif
+
/*
* Increment the PV qspinlock statistical counters
*/
static inline void __lockevent_inc(enum lock_events event, bool cond)
{
if (cond)
- __this_cpu_inc(lockevents[event]);
+ lockevent_percpu_inc(lockevents[event]);
}
#define lockevent_inc(ev) __lockevent_inc(LOCKEVENT_ ##ev, true)
@@ -44,7 +67,7 @@ static inline void __lockevent_inc(enum lock_events event, bool cond)
static inline void __lockevent_add(enum lock_events event, int inc)
{
- __this_cpu_add(lockevents[event], inc);
+ lockevent_percpu_add(lockevents[event], inc);
}
#define lockevent_add(ev, c) __lockevent_add(LOCKEVENT_ ##ev, c)
--
2.18.1
On Wed, May 22, 2019 at 8:40 AM Waiman Long <[email protected]> wrote:
>
> +#if defined(CONFIG_PREEMPT) && \
> + (defined(CONFIG_DEBUG_PREEMPT) || !defined(CONFIG_X86))
> +#define lockevent_percpu_inc(x) this_cpu_inc(x)
> +#define lockevent_percpu_add(x, v) this_cpu_add(x, v)
Why that CONFIG_X86 special case?
On x86, the regular non-underscore versionm is perfectly fine, and the
underscore is no faster or simpler.
So just make it be
#if defined(CONFIG_PREEMPT)
.. non-underscore versions..
#else
.. underscore versions ..
#endif
and realize that x86 simply doesn't _care_. On x86, it will be one
single instruction regardless.
Non-x86 may prefer the underscore versions for the non-preempt case.
Linus
On 5/22/19 3:54 PM, Linus Torvalds wrote:
> On Wed, May 22, 2019 at 8:40 AM Waiman Long <[email protected]> wrote:
>> +#if defined(CONFIG_PREEMPT) && \
>> + (defined(CONFIG_DEBUG_PREEMPT) || !defined(CONFIG_X86))
>> +#define lockevent_percpu_inc(x) this_cpu_inc(x)
>> +#define lockevent_percpu_add(x, v) this_cpu_add(x, v)
> Why that CONFIG_X86 special case?
>
> On x86, the regular non-underscore versionm is perfectly fine, and the
> underscore is no faster or simpler.
The condition is to use non-underscore version only when
1) It is a preempt kernel; AND
2) It either have CONFIG_DEBUG_PREEMPT on, OR it is a non-x86 system.
> So just make it be
>
> #if defined(CONFIG_PREEMPT)
> .. non-underscore versions..
> #else
> .. underscore versions ..
> #endif
>
> and realize that x86 simply doesn't _care_. On x86, it will be one
> single instruction regardless.
>
> Non-x86 may prefer the underscore versions for the non-preempt case.
I was thinking of doing that originally, but then change it so x86
preempt kernel will also use the underscore version as long as
CONFIG_DEBUG_PREEMPT is not set.
I can change it back if that makes it less confusing.
Cheers,
Longman
On Wed, May 22, 2019 at 12:54:13PM -0700, Linus Torvalds wrote:
> On Wed, May 22, 2019 at 8:40 AM Waiman Long <[email protected]> wrote:
> >
> > +#if defined(CONFIG_PREEMPT) && \
> > + (defined(CONFIG_DEBUG_PREEMPT) || !defined(CONFIG_X86))
> > +#define lockevent_percpu_inc(x) this_cpu_inc(x)
> > +#define lockevent_percpu_add(x, v) this_cpu_add(x, v)
>
> Why that CONFIG_X86 special case?
>
> On x86, the regular non-underscore versionm is perfectly fine, and the
> underscore is no faster or simpler.
>
> So just make it be
>
> #if defined(CONFIG_PREEMPT)
> .. non-underscore versions..
> #else
> .. underscore versions ..
> #endif
>
> and realize that x86 simply doesn't _care_. On x86, it will be one
> single instruction regardless.
>
> Non-x86 may prefer the underscore versions for the non-preempt case.
To be honest, given this depends on LOCK_EVENT_COUNTS, I'd be inclined to
keep things simple and drop the underscore versions entirely. Saves having
to worry about things like "could I take an interrupt during the add?".
Will
On 5/23/19 10:58 AM, Will Deacon wrote:
> On Wed, May 22, 2019 at 12:54:13PM -0700, Linus Torvalds wrote:
>> On Wed, May 22, 2019 at 8:40 AM Waiman Long <[email protected]> wrote:
>>> +#if defined(CONFIG_PREEMPT) && \
>>> + (defined(CONFIG_DEBUG_PREEMPT) || !defined(CONFIG_X86))
>>> +#define lockevent_percpu_inc(x) this_cpu_inc(x)
>>> +#define lockevent_percpu_add(x, v) this_cpu_add(x, v)
>> Why that CONFIG_X86 special case?
>>
>> On x86, the regular non-underscore versionm is perfectly fine, and the
>> underscore is no faster or simpler.
>>
>> So just make it be
>>
>> #if defined(CONFIG_PREEMPT)
>> .. non-underscore versions..
>> #else
>> .. underscore versions ..
>> #endif
>>
>> and realize that x86 simply doesn't _care_. On x86, it will be one
>> single instruction regardless.
>>
>> Non-x86 may prefer the underscore versions for the non-preempt case.
> To be honest, given this depends on LOCK_EVENT_COUNTS, I'd be inclined to
> keep things simple and drop the underscore versions entirely. Saves having
> to worry about things like "could I take an interrupt during the add?".
>
I have sent out the v2 patch that simplifies the condition. Now the
underscore versions will be used for !preempt kernel and non-underscore
version used in preempt kernel. The non-underscore versions may generate
a lot more unnecessary code when CONFIG_PREEMPT_COUNT is defined.
Cheers,
Longman