percpu_counter: add percpu_counter_sum_all interface
Commit Message
The percpu_counter is used for scenarios where performance is more
important than the accuracy. For percpu_counter users, who want more
accurate information in their slowpath, percpu_counter_sum is provided
which traverses all the online CPUs to accumulate the data. The reason
it only needs to traverse online CPUs is because percpu_counter does
implement CPU offline callback which syncs the local data of the
offlined CPU.
However there is a small race window between the online CPUs traversal
of percpu_counter_sum and the CPU offline callback. The offline callback
has to traverse all the percpu_counters on the system to flush the CPU
local data which can be a lot. During that time, the CPU which is going
offline has already been published as offline to all the readers. So, as
the offline callback is running, percpu_counter_sum can be called for
one counter which has some state on the CPU going offline. Since
percpu_counter_sum only traverses online CPUs, it will skip that
specific CPU and the offline callback might not have flushed the state
for that specific percpu_counter on that offlined CPU.
Normally this is not an issue because percpu_counter users can deal with
some inaccuracy for small time window. However a new user i.e. mm_struct
on the cleanup path wants to check the exact state of the percpu_counter
through check_mm(). For such users, this patch introduces
percpu_counter_sum_all() which traverses all possible CPUs.
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Reported-by: Marek Szyprowski <m.szyprowski@samsung.com>
Tested-by: Marek Szyprowski <m.szyprowski@samsung.com>
---
include/linux/percpu_counter.h | 6 ++++++
kernel/fork.c | 2 +-
lib/percpu_counter.c | 29 +++++++++++++++++++++++------
3 files changed, 30 insertions(+), 7 deletions(-)
Comments
On Sat, 5 Nov 2022 01:40:13 +0000 Shakeel Butt <shakeelb@google.com> wrote:
> The percpu_counter is used for scenarios where performance is more
> important than the accuracy. For percpu_counter users, who want more
> accurate information in their slowpath, percpu_counter_sum is provided
> which traverses all the online CPUs to accumulate the data. The reason
> it only needs to traverse online CPUs is because percpu_counter does
> implement CPU offline callback which syncs the local data of the
> offlined CPU.
>
> However there is a small race window between the online CPUs traversal
> of percpu_counter_sum and the CPU offline callback. The offline callback
> has to traverse all the percpu_counters on the system to flush the CPU
> local data which can be a lot. During that time, the CPU which is going
> offline has already been published as offline to all the readers. So, as
> the offline callback is running, percpu_counter_sum can be called for
> one counter which has some state on the CPU going offline. Since
> percpu_counter_sum only traverses online CPUs, it will skip that
> specific CPU and the offline callback might not have flushed the state
> for that specific percpu_counter on that offlined CPU.
OK, got it, thanks.
> Normally this is not an issue because percpu_counter users can deal with
> some inaccuracy for small time window. However a new user i.e. mm_struct
> on the cleanup path wants to check the exact state of the percpu_counter
> through check_mm(). For such users, this patch introduces
> percpu_counter_sum_all() which traverses all possible CPUs.
And uses it in fork.c:check_mm()!
> --- a/kernel/fork.c
> +++ b/kernel/fork.c
> @@ -756,7 +756,7 @@ static void check_mm(struct mm_struct *mm)
> "Please make sure 'struct resident_page_types[]' is updated as well");
>
> for (i = 0; i < NR_MM_COUNTERS; i++) {
> - long x = percpu_counter_sum(&mm->rss_stat[i]);
> + long x = percpu_counter_sum_all(&mm->rss_stat[i]);
check_mm() just became more expensive in some cases. nr_possible_cpus
* 4. I wonder if this is enough for people to start caring about.
check_mm() is presently non-optional and I'd be reluctant to change
this, given how commonly we see the "BUG: Bad rss-counter state"
getting reported (22 million hits in a google search!).
We could save a ton of that cost by running percpu_counter_sum() first,
then trying percpu_counter_sum_all() if percpu_counter_sum() indicated
an error. This is only worth bothering about if the new check_mm()
cost is a concern.
On Mon, Nov 7, 2022 at 1:05 PM Andrew Morton <akpm@linux-foundation.org> wrote:
>
> On Sat, 5 Nov 2022 01:40:13 +0000 Shakeel Butt <shakeelb@google.com> wrote:
>
> > The percpu_counter is used for scenarios where performance is more
> > important than the accuracy. For percpu_counter users, who want more
> > accurate information in their slowpath, percpu_counter_sum is provided
> > which traverses all the online CPUs to accumulate the data. The reason
> > it only needs to traverse online CPUs is because percpu_counter does
> > implement CPU offline callback which syncs the local data of the
> > offlined CPU.
> >
> > However there is a small race window between the online CPUs traversal
> > of percpu_counter_sum and the CPU offline callback. The offline callback
> > has to traverse all the percpu_counters on the system to flush the CPU
> > local data which can be a lot. During that time, the CPU which is going
> > offline has already been published as offline to all the readers. So, as
> > the offline callback is running, percpu_counter_sum can be called for
> > one counter which has some state on the CPU going offline. Since
> > percpu_counter_sum only traverses online CPUs, it will skip that
> > specific CPU and the offline callback might not have flushed the state
> > for that specific percpu_counter on that offlined CPU.
>
> OK, got it, thanks.
>
> > Normally this is not an issue because percpu_counter users can deal with
> > some inaccuracy for small time window. However a new user i.e. mm_struct
> > on the cleanup path wants to check the exact state of the percpu_counter
> > through check_mm(). For such users, this patch introduces
> > percpu_counter_sum_all() which traverses all possible CPUs.
>
> And uses it in fork.c:check_mm()!
>
> > --- a/kernel/fork.c
> > +++ b/kernel/fork.c
> > @@ -756,7 +756,7 @@ static void check_mm(struct mm_struct *mm)
> > "Please make sure 'struct resident_page_types[]' is updated as well");
> >
> > for (i = 0; i < NR_MM_COUNTERS; i++) {
> > - long x = percpu_counter_sum(&mm->rss_stat[i]);
> > + long x = percpu_counter_sum_all(&mm->rss_stat[i]);
>
> check_mm() just became more expensive in some cases. nr_possible_cpus
> * 4. I wonder if this is enough for people to start caring about.
>
> check_mm() is presently non-optional and I'd be reluctant to change
> this, given how commonly we see the "BUG: Bad rss-counter state"
> getting reported (22 million hits in a google search!).
>
> We could save a ton of that cost by running percpu_counter_sum() first,
> then trying percpu_counter_sum_all() if percpu_counter_sum() indicated
> an error. This is only worth bothering about if the new check_mm()
> cost is a concern.
>
Yes, this makes much more sense. I had run hackbench on the original
patch and didn't see any significant difference. I will update this
and run some more perf benchmarks to make sure there is no regression
due to this change.
thanks,
Shakeel
@@ -45,6 +45,7 @@ void percpu_counter_set(struct percpu_counter *fbc, s64 amount);
void percpu_counter_add_batch(struct percpu_counter *fbc, s64 amount,
s32 batch);
s64 __percpu_counter_sum(struct percpu_counter *fbc);
+s64 percpu_counter_sum_all(struct percpu_counter *fbc);
int __percpu_counter_compare(struct percpu_counter *fbc, s64 rhs, s32 batch);
void percpu_counter_sync(struct percpu_counter *fbc);
@@ -193,6 +194,11 @@ static inline s64 percpu_counter_sum(struct percpu_counter *fbc)
return percpu_counter_read(fbc);
}
+static inline s64 percpu_counter_sum_all(struct percpu_counter *fbc)
+{
+ return percpu_counter_read(fbc);
+}
+
static inline bool percpu_counter_initialized(struct percpu_counter *fbc)
{
return true;
@@ -756,7 +756,7 @@ static void check_mm(struct mm_struct *mm)
"Please make sure 'struct resident_page_types[]' is updated as well");
for (i = 0; i < NR_MM_COUNTERS; i++) {
- long x = percpu_counter_sum(&mm->rss_stat[i]);
+ long x = percpu_counter_sum_all(&mm->rss_stat[i]);
if (unlikely(x))
pr_alert("BUG: Bad rss-counter state mm:%p type:%s val:%ld\n",
@@ -117,11 +117,8 @@ void percpu_counter_sync(struct percpu_counter *fbc)
}
EXPORT_SYMBOL(percpu_counter_sync);
-/*
- * Add up all the per-cpu counts, return the result. This is a more accurate
- * but much slower version of percpu_counter_read_positive()
- */
-s64 __percpu_counter_sum(struct percpu_counter *fbc)
+static s64 __percpu_counter_sum_mask(struct percpu_counter *fbc,
+ const struct cpumask *cpu_mask)
{
s64 ret;
int cpu;
@@ -129,15 +126,35 @@ s64 __percpu_counter_sum(struct percpu_counter *fbc)
raw_spin_lock_irqsave(&fbc->lock, flags);
ret = fbc->count;
- for_each_online_cpu(cpu) {
+ for_each_cpu(cpu, cpu_mask) {
s32 *pcount = per_cpu_ptr(fbc->counters, cpu);
ret += *pcount;
}
raw_spin_unlock_irqrestore(&fbc->lock, flags);
return ret;
}
+
+/*
+ * Add up all the per-cpu counts, return the result. This is a more accurate
+ * but much slower version of percpu_counter_read_positive()
+ */
+s64 __percpu_counter_sum(struct percpu_counter *fbc)
+{
+ return __percpu_counter_sum_mask(fbc, cpu_online_mask);
+}
EXPORT_SYMBOL(__percpu_counter_sum);
+/*
+ * This is slower version of percpu_counter_sum as it traverses all possible
+ * cpus. Use this only in the cases where accurate data is needed in the
+ * presense of CPUs getting offlined.
+ */
+s64 percpu_counter_sum_all(struct percpu_counter *fbc)
+{
+ return __percpu_counter_sum_mask(fbc, cpu_possible_mask);
+}
+EXPORT_SYMBOL(percpu_counter_sum_all);
+
int __percpu_counter_init(struct percpu_counter *fbc, s64 amount, gfp_t gfp,
struct lock_class_key *key)
{