locking/lock_events: Use this_cpu_add() when necessary
Waiman Long
reviewerPATCH v4 1/1 locking/lock_events: Use this_cpu_add() when necessary
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: ...
Given the imprecise nature of the count and the possibility of resetting the count and doing the measurement again, this is not really a big problem to use the unprotected __this_cpu_*() functions.
To make the preemption checking code happy, the this_cpu_*() functions will be used if CONFIG_DEBUG_PREEMPT is defined.
The imprecise nature of the locking counts are also documented with the suggestion that we should run the measurement a few times with the counts reset in between to get a better picture of what is going on under the hood.
--- kernel/locking/lock_events.h | 42 ++++++++++++++++++++++++++++++++++-- 1 file changed, 40 insertions(+), 2 deletions(-)
kernel/locking/lock_events.h +40 -2
@@ -30,13 +30,51 @@ 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. +* +* It is possible that the same percpu counter may be modified in both +* the process and interrupt contexts. For architectures that perform +* percpu operation with multiple instructions, it is possible to lose +* count if a process context percpu update is interrupted in the middle +* and the same counter is updated in the interrupt context. Therefore, +* the generated percpu sum may not be precise. The error, if any, should +* be small and insignificant. +* +* For those architectures that do multi-instruction percpu operation, +* preemption in the middle and moving the task to another cpu may cause +* a larger error in the count. Again, this will be few and far between. +* Given the imprecise nature of the count and the possibility of resetting +* the count and doing the measurement again, this is not really a big +* problem. +* +* To get a better picture of what is happening under the hood, it is +* suggested that a few measurements should be taken with the counts +* reset in between to stamp out outliner because of these possible +* error conditions. +* +* To minimize overhead, we use __this_cpu_*() in all cases except when +* CONFIG_DEBUG_PREEMPT is defined. In this particular case, this_cpu_*() +* will be used to avoid the appearance of unwanted BUG messages. +*/ +#ifdef CONFIG_DEBUG_PREEMPT +#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 +82,7 @@ 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)
Signature · 1 line
2.18.1
Quoted 2 lines · Waiman Long
Fixes a8654596f0371 ("locking/rwsem: Enable lock event counting")
Applied directly,
Linus
Peter Zijlstra
maintainerReview
Quoted 7 lines · Waiman Long
(quoted diff) +7
-0+#ifdef CONFIG_DEBUG_PREEMPT +#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
That's disguisting... I see Linus already applied it, but yuck. That's what we have raw_cpu_*() for.
Something like the below perhaps.
---
kernel/locking/lock_events.h +4 -41
@@ -31,50 +31,13 @@ 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. -* -* It is possible that the same percpu counter may be modified in both -* the process and interrupt contexts. For architectures that perform -* percpu operation with multiple instructions, it is possible to lose -* count if a process context percpu update is interrupted in the middle -* and the same counter is updated in the interrupt context. Therefore, -* the generated percpu sum may not be precise. The error, if any, should -* be small and insignificant. -* -* For those architectures that do multi-instruction percpu operation, -* preemption in the middle and moving the task to another cpu may cause -* a larger error in the count. Again, this will be few and far between. -* Given the imprecise nature of the count and the possibility of resetting -* the count and doing the measurement again, this is not really a big -* problem. -* -* To get a better picture of what is happening under the hood, it is -* suggested that a few measurements should be taken with the counts -* reset in between to stamp out outliner because of these possible -* error conditions. -* -* To minimize overhead, we use __this_cpu_*() in all cases except when -* CONFIG_DEBUG_PREEMPT is defined. In this particular case, this_cpu_*() -* will be used to avoid the appearance of unwanted BUG messages. -*/ -#ifdef CONFIG_DEBUG_PREEMPT -#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 +* Increment the statistical counters. use raw_cpu_inc() because of lower +* overhead and we don't care if we loose the occasional update. */ static inline void __lockevent_inc (enum lock_events event,bool cond) {if (cond) -lockevent_percpu_inc (lockevents[event]); +raw_cpu_inc (lockevents[event]); }#define lockevent_inc (ev) __lockevent_inc(LOCKEVENT_ ##ev, true)
@@ -82,7 +45,7 @@ static inline void __lockevent_inc(enum lock_events event, bool cond)static inline void __lockevent_add (enum lock_events event,int inc) { -lockevent_percpu_add (lockevents[event], inc); +raw_cpu_add (lockevents[event], inc); }#define lockevent_add (ev, c) __lockevent_add(LOCKEVENT_ ##ev, c)
Quoted 3 lines · Peter Zijlstra
That's disguisting... I see Linus already applied it, but yuck. That's what we have raw_cpu_*() for.
Ahh, I tried to look for that, but there was enough indirection and confusion that I wasn't sure they were generically available.
And the "raw_cpu_*()" functions are rare enough that I'd never encountered them enough to really be aware of them. In fact, we seem to have exactly _one_ user of "raw_cpu_add()" in the whole kernel, and a handful of "raw_cpu_inc()".
But ack on your patch, and a heartfelt "yeah, that's the right thing". Thanks,
Linus
Peter Zijlstra
maintainerQuoted 5 lines · Linus Torvalds
Quoted 3 lines · Peter Zijlstra
That's disguisting... I see Linus already applied it, but yuck. That's what we have raw_cpu_*() for.
Quoted 7 lines
Ahh, I tried to look for that, but there was enough indirection and confusion that I wasn't sure they were generically available.
And the "raw_cpu_*()" functions are rare enough that I'd never encountered them enough to really be aware of them. In fact, we seem to have exactly _one_ user of "
raw_cpu_add()" in the whole kernel, and a handful of "raw_cpu_inc()".
Yeah, not having many is good. From a correctness PoV they're basically always the wrong thing to use, except for this one usecase where we prefer speed over correctness.
Quoted 1 line
But ack on your patch, and a heartfelt "yeah, that's the right thing". Thanks,
Thanks, I'll go write me a Changelog then ;-)
tip-bot for Peter Zijlstra
merge bot[tip:locking/core] locking/lock_events: Use raw_cpu_{add,inc}() for stats
In reply to Peter Zijlstralocking/lock_events: Use raw_cpu_{add,inc}() for stats
Commit-ID: 24811637dbfd07c69da7e9db586d35d17e6afca3 Gitweb: https://git.kernel.org/tip/24811637dbfd07c69da7e9db586d35d17e6afca3 Author: Peter Zijlstra <[email protected]> AuthorDate: Mon, 27 May 2019 10:23:26 +0200 Committer: Ingo Molnar <[email protected]> CommitDate: Mon, 3 Jun 2019 12:32:56 +0200
locking/lock_events: Use raw_cpu_{add,inc}() for stats
Instead of playing silly games with CONFIG_DEBUG_PREEMPT toggling between this_cpu_*() and __this_cpu_*() use raw_cpu_*(), which is exactly what we want here.
--- kernel/locking/lock_events.h | 45 ++++---------------------------------------- 1 file changed, 4 insertions(+), 41 deletions(-)
kernel/locking/lock_events.h +4 -41
@@ -31,50 +31,13 @@ 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. -* -* It is possible that the same percpu counter may be modified in both -* the process and interrupt contexts. For architectures that perform -* percpu operation with multiple instructions, it is possible to lose -* count if a process context percpu update is interrupted in the middle -* and the same counter is updated in the interrupt context. Therefore, -* the generated percpu sum may not be precise. The error, if any, should -* be small and insignificant. -* -* For those architectures that do multi-instruction percpu operation, -* preemption in the middle and moving the task to another cpu may cause -* a larger error in the count. Again, this will be few and far between. -* Given the imprecise nature of the count and the possibility of resetting -* the count and doing the measurement again, this is not really a big -* problem. -* -* To get a better picture of what is happening under the hood, it is -* suggested that a few measurements should be taken with the counts -* reset in between to stamp out outliner because of these possible -* error conditions. -* -* To minimize overhead, we use __this_cpu_*() in all cases except when -* CONFIG_DEBUG_PREEMPT is defined. In this particular case, this_cpu_*() -* will be used to avoid the appearance of unwanted BUG messages. -*/ -#ifdef CONFIG_DEBUG_PREEMPT -#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 +* Increment the statistical counters. use raw_cpu_inc() because of lower +* overhead and we don't care if we loose the occasional update. */ static inline void __lockevent_inc (enum lock_events event,bool cond) {if (cond) -lockevent_percpu_inc (lockevents[event]); +raw_cpu_inc (lockevents[event]); }#define lockevent_inc (ev) __lockevent_inc(LOCKEVENT_ ##ev, true)
@@ -82,7 +45,7 @@ static inline void __lockevent_inc(enum lock_events event, bool cond)static inline void __lockevent_add (enum lock_events event,int inc) { -lockevent_percpu_add (lockevents[event], inc); +raw_cpu_add (lockevents[event], inc); }#define lockevent_add (ev, c) __lockevent_add(LOCKEVENT_ ##ev, c)