Grabbing the spinlock for every bucket even if it's empty, was causing
significant perfomance cost when traversing htab maps that have only a
few entries. This patch addresses the issue by checking first the
bucket_cnt, if the bucket has some entries then we go and grab the
spinlock and proceed with the batching.
Tested with a htab of size 50K and different value of populated entries.
Before:
Benchmark Time(ns) CPU(ns)
---------------------------------------------
BM_DumpHashMap/1 2759655 2752033
BM_DumpHashMap/10 2933722 2930825
BM_DumpHashMap/200 3171680 3170265
BM_DumpHashMap/500 3639607 3635511
BM_DumpHashMap/1000 4369008 4364981
BM_DumpHashMap/5k 11171919 11134028
BM_DumpHashMap/20k 69150080 69033496
BM_DumpHashMap/39k 190501036 190226162
After:
Benchmark Time(ns) CPU(ns)
---------------------------------------------
BM_DumpHashMap/1 202707 200109
BM_DumpHashMap/10 213441 210569
BM_DumpHashMap/200 478641 472350
BM_DumpHashMap/500 980061 967102
BM_DumpHashMap/1000 1863835 1839575
BM_DumpHashMap/5k 8961836 8902540
BM_DumpHashMap/20k 69761497 69322756
BM_DumpHashMap/39k 187437830 186551111
Fixes: 057996380a42 ("bpf: Add batch ops to all htab bpf map")
Cc: Yonghong Song <[email protected]>
Signed-off-by: Brian Vazquez <[email protected]>
---
kernel/bpf/hashtab.c | 21 +++++++++++++++++++--
1 file changed, 19 insertions(+), 2 deletions(-)
diff --git a/kernel/bpf/hashtab.c b/kernel/bpf/hashtab.c
index 2d182c4ee9d99..fdbde28b0fe06 100644
--- a/kernel/bpf/hashtab.c
+++ b/kernel/bpf/hashtab.c
@@ -1260,6 +1260,7 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map,
struct hlist_nulls_head *head;
struct hlist_nulls_node *n;
unsigned long flags;
+ bool locked = false;
struct htab_elem *l;
struct bucket *b;
int ret = 0;
@@ -1319,15 +1320,25 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map,
dst_val = values;
b = &htab->buckets[batch];
head = &b->head;
- raw_spin_lock_irqsave(&b->lock, flags);
+ /* do not grab the lock unless need it (bucket_cnt > 0). */
+ if (locked)
+ raw_spin_lock_irqsave(&b->lock, flags);
bucket_cnt = 0;
hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
bucket_cnt++;
+ if (bucket_cnt && !locked) {
+ locked = true;
+ goto again_nocopy;
+ }
+
if (bucket_cnt > (max_count - total)) {
if (total == 0)
ret = -ENOSPC;
+ /* Note that since bucket_cnt > 0 here, it is implicit
+ * that the locked was grabbed, so release it.
+ */
raw_spin_unlock_irqrestore(&b->lock, flags);
rcu_read_unlock();
this_cpu_dec(bpf_prog_active);
@@ -1337,6 +1348,9 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map,
if (bucket_cnt > bucket_size) {
bucket_size = bucket_cnt;
+ /* Note that since bucket_cnt > 0 here, it is implicit
+ * that the locked was grabbed, so release it.
+ */
raw_spin_unlock_irqrestore(&b->lock, flags);
rcu_read_unlock();
this_cpu_dec(bpf_prog_active);
@@ -1379,7 +1393,10 @@ __htab_map_lookup_and_delete_batch(struct bpf_map *map,
dst_val += value_size;
}
- raw_spin_unlock_irqrestore(&b->lock, flags);
+ if (locked) {
+ raw_spin_unlock_irqrestore(&b->lock, flags);
+ locked = false;
+ }
/* If we are not copying data, we can go to next bucket and avoid
* unlocking the rcu.
*/
--
2.25.0.265.gbab2e86ba0-goog
On 2/14/20 2:43 PM, Brian Vazquez wrote:
> Grabbing the spinlock for every bucket even if it's empty, was causing
> significant perfomance cost when traversing htab maps that have only a
> few entries. This patch addresses the issue by checking first the
> bucket_cnt, if the bucket has some entries then we go and grab the
> spinlock and proceed with the batching.
>
> Tested with a htab of size 50K and different value of populated entries.
>
> Before:
> Benchmark Time(ns) CPU(ns)
> ---------------------------------------------
> BM_DumpHashMap/1 2759655 2752033
> BM_DumpHashMap/10 2933722 2930825
> BM_DumpHashMap/200 3171680 3170265
> BM_DumpHashMap/500 3639607 3635511
> BM_DumpHashMap/1000 4369008 4364981
> BM_DumpHashMap/5k 11171919 11134028
> BM_DumpHashMap/20k 69150080 69033496
> BM_DumpHashMap/39k 190501036 190226162
>
> After:
> Benchmark Time(ns) CPU(ns)
> ---------------------------------------------
> BM_DumpHashMap/1 202707 200109
> BM_DumpHashMap/10 213441 210569
> BM_DumpHashMap/200 478641 472350
> BM_DumpHashMap/500 980061 967102
> BM_DumpHashMap/1000 1863835 1839575
> BM_DumpHashMap/5k 8961836 8902540
> BM_DumpHashMap/20k 69761497 69322756
> BM_DumpHashMap/39k 187437830 186551111
>
> Fixes: 057996380a42 ("bpf: Add batch ops to all htab bpf map")
> Cc: Yonghong Song <[email protected]>
> Signed-off-by: Brian Vazquez <[email protected]>
Acked-by: Yonghong Song <[email protected]>
On 2/18/20 4:43 PM, Yonghong Song wrote:
> On 2/14/20 2:43 PM, Brian Vazquez wrote:
>> Grabbing the spinlock for every bucket even if it's empty, was causing
>> significant perfomance cost when traversing htab maps that have only a
>> few entries. This patch addresses the issue by checking first the
>> bucket_cnt, if the bucket has some entries then we go and grab the
>> spinlock and proceed with the batching.
>>
>> Tested with a htab of size 50K and different value of populated entries.
>>
>> Before:
>> Benchmark Time(ns) CPU(ns)
>> ---------------------------------------------
>> BM_DumpHashMap/1 2759655 2752033
>> BM_DumpHashMap/10 2933722 2930825
>> BM_DumpHashMap/200 3171680 3170265
>> BM_DumpHashMap/500 3639607 3635511
>> BM_DumpHashMap/1000 4369008 4364981
>> BM_DumpHashMap/5k 11171919 11134028
>> BM_DumpHashMap/20k 69150080 69033496
>> BM_DumpHashMap/39k 190501036 190226162
>>
>> After:
>> Benchmark Time(ns) CPU(ns)
>> ---------------------------------------------
>> BM_DumpHashMap/1 202707 200109
>> BM_DumpHashMap/10 213441 210569
>> BM_DumpHashMap/200 478641 472350
>> BM_DumpHashMap/500 980061 967102
>> BM_DumpHashMap/1000 1863835 1839575
>> BM_DumpHashMap/5k 8961836 8902540
>> BM_DumpHashMap/20k 69761497 69322756
>> BM_DumpHashMap/39k 187437830 186551111
>>
>> Fixes: 057996380a42 ("bpf: Add batch ops to all htab bpf map")
>> Cc: Yonghong Song <[email protected]>
>> Signed-off-by: Brian Vazquez <[email protected]>
>
> Acked-by: Yonghong Song <[email protected]>
I must probably be missing something, but how is this safe? Presume we
traverse in the walk with bucket_cnt = 0. Meanwhile a different CPU added
entries to this bucket since not locked. Same reader on the other CPU with
bucket_cnt = 0 then starts to traverse the second
hlist_nulls_for_each_entry_safe() unlocked e.g. deleting entries?
On 2/18/20 7:56 AM, Daniel Borkmann wrote:
> On 2/18/20 4:43 PM, Yonghong Song wrote:
>> On 2/14/20 2:43 PM, Brian Vazquez wrote:
>>> Grabbing the spinlock for every bucket even if it's empty, was causing
>>> significant perfomance cost when traversing htab maps that have only a
>>> few entries. This patch addresses the issue by checking first the
>>> bucket_cnt, if the bucket has some entries then we go and grab the
>>> spinlock and proceed with the batching.
>>>
>>> Tested with a htab of size 50K and different value of populated entries.
>>>
>>> Before:
>>> Benchmark Time(ns) CPU(ns)
>>> ---------------------------------------------
>>> BM_DumpHashMap/1 2759655 2752033
>>> BM_DumpHashMap/10 2933722 2930825
>>> BM_DumpHashMap/200 3171680 3170265
>>> BM_DumpHashMap/500 3639607 3635511
>>> BM_DumpHashMap/1000 4369008 4364981
>>> BM_DumpHashMap/5k 11171919 11134028
>>> BM_DumpHashMap/20k 69150080 69033496
>>> BM_DumpHashMap/39k 190501036 190226162
>>>
>>> After:
>>> Benchmark Time(ns) CPU(ns)
>>> ---------------------------------------------
>>> BM_DumpHashMap/1 202707 200109
>>> BM_DumpHashMap/10 213441 210569
>>> BM_DumpHashMap/200 478641 472350
>>> BM_DumpHashMap/500 980061 967102
>>> BM_DumpHashMap/1000 1863835 1839575
>>> BM_DumpHashMap/5k 8961836 8902540
>>> BM_DumpHashMap/20k 69761497 69322756
>>> BM_DumpHashMap/39k 187437830 186551111
>>>
>>> Fixes: 057996380a42 ("bpf: Add batch ops to all htab bpf map")
>>> Cc: Yonghong Song <[email protected]>
>>> Signed-off-by: Brian Vazquez <[email protected]>
>>
>> Acked-by: Yonghong Song <[email protected]>
>
> I must probably be missing something, but how is this safe? Presume we
> traverse in the walk with bucket_cnt = 0. Meanwhile a different CPU added
> entries to this bucket since not locked. Same reader on the other CPU with
> bucket_cnt = 0 then starts to traverse the second
> hlist_nulls_for_each_entry_safe() unlocked e.g. deleting entries?
Thanks for pointing this out. Yes, you are correct. If bucket_cnt is 0
and buck->lock is not held, we should skip the
hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
...
}
as another cpu may traverse the bucket in parallel by adding/deleting
the elements.
On Tue, Feb 18, 2020 at 8:34 AM Yonghong Song <[email protected]> wrote:
>
>
>
> On 2/18/20 7:56 AM, Daniel Borkmann wrote:
> > On 2/18/20 4:43 PM, Yonghong Song wrote:
> >> On 2/14/20 2:43 PM, Brian Vazquez wrote:
> >>> Grabbing the spinlock for every bucket even if it's empty, was causing
> >>> significant perfomance cost when traversing htab maps that have only a
> >>> few entries. This patch addresses the issue by checking first the
> >>> bucket_cnt, if the bucket has some entries then we go and grab the
> >>> spinlock and proceed with the batching.
> >>>
> >>> Tested with a htab of size 50K and different value of populated entries.
> >>>
> >>> Before:
> >>> Benchmark Time(ns) CPU(ns)
> >>> ---------------------------------------------
> >>> BM_DumpHashMap/1 2759655 2752033
> >>> BM_DumpHashMap/10 2933722 2930825
> >>> BM_DumpHashMap/200 3171680 3170265
> >>> BM_DumpHashMap/500 3639607 3635511
> >>> BM_DumpHashMap/1000 4369008 4364981
> >>> BM_DumpHashMap/5k 11171919 11134028
> >>> BM_DumpHashMap/20k 69150080 69033496
> >>> BM_DumpHashMap/39k 190501036 190226162
> >>>
> >>> After:
> >>> Benchmark Time(ns) CPU(ns)
> >>> ---------------------------------------------
> >>> BM_DumpHashMap/1 202707 200109
> >>> BM_DumpHashMap/10 213441 210569
> >>> BM_DumpHashMap/200 478641 472350
> >>> BM_DumpHashMap/500 980061 967102
> >>> BM_DumpHashMap/1000 1863835 1839575
> >>> BM_DumpHashMap/5k 8961836 8902540
> >>> BM_DumpHashMap/20k 69761497 69322756
> >>> BM_DumpHashMap/39k 187437830 186551111
> >>>
> >>> Fixes: 057996380a42 ("bpf: Add batch ops to all htab bpf map")
> >>> Cc: Yonghong Song <[email protected]>
> >>> Signed-off-by: Brian Vazquez <[email protected]>
> >>
> >> Acked-by: Yonghong Song <[email protected]>
> >
> > I must probably be missing something, but how is this safe? Presume we
> > traverse in the walk with bucket_cnt = 0. Meanwhile a different CPU added
> > entries to this bucket since not locked. Same reader on the other CPU with
> > bucket_cnt = 0 then starts to traverse the second
> > hlist_nulls_for_each_entry_safe() unlocked e.g. deleting entries?
>
> Thanks for pointing this out. Yes, you are correct. If bucket_cnt is 0
> and buck->lock is not held, we should skip the
> hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
> ...
> }
> as another cpu may traverse the bucket in parallel by adding/deleting
> the elements.
Makes sense. Let me fix it in the next version, thanks for reviewing it!