On Thu, May 7, 2020 at 7:00 PM Eric Dumazet <eduma...@google.com> wrote:
>
> percpu_counter_add() uses a default batch size which is quite big
> on platforms with 256 cpus. (2*256 -> 512)
>
> This means dst_entries_get_fast() can be off by +/- 2*(nr_cpus^2)
> (131072 on servers with 256 cpus)
>
> Reduce the batch size to something more reasonable, and
> add logic to ip6_dst_gc() to call dst_entries_get_slow()
> before calling the _very_ expensive fib6_run_gc() function.
>
> Signed-off-by: Eric Dumazet <eduma...@google.com>
> ---
> include/net/dst_ops.h | 4 +++-
> net/core/dst.c | 8 ++++----
> net/ipv6/route.c | 3 +++
> 3 files changed, 10 insertions(+), 5 deletions(-)
>
> diff --git a/include/net/dst_ops.h b/include/net/dst_ops.h
> index
> 443863c7b8da362476c15fd290ac2a32a8aa86e3..88ff7bb2bb9bd950cc54fd5e0ae4573d4c66873d
> 100644
> --- a/include/net/dst_ops.h
> +++ b/include/net/dst_ops.h
> @@ -53,9 +53,11 @@ static inline int dst_entries_get_slow(struct dst_ops *dst)
> return percpu_counter_sum_positive(&dst->pcpuc_entries);
> }
>
> +#define DST_PERCPU_COUNTER_BATCH 32
> static inline void dst_entries_add(struct dst_ops *dst, int val)
> {
> - percpu_counter_add(&dst->pcpuc_entries, val);
> + percpu_counter_add_batch(&dst->pcpuc_entries, val,
> + DST_PERCPU_COUNTER_BATCH);
> }
>
> static inline int dst_entries_init(struct dst_ops *dst)
> diff --git a/net/core/dst.c b/net/core/dst.c
> index
> 193af526e908afa4b868cf128470f0fbc3850698..d6b6ced0d451a39c0ccb88ae39dba225ea9f5705
> 100644
> --- a/net/core/dst.c
> +++ b/net/core/dst.c
> @@ -81,11 +81,11 @@ void *dst_alloc(struct dst_ops *ops, struct net_device
> *dev,
> {
> struct dst_entry *dst;
>
> - if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
> + if (ops->gc &&
> + !(flags & DST_NOCOUNT) &&
> + dst_entries_get_fast(ops) > ops->gc_thresh) {
> if (ops->gc(ops)) {
> - printk_ratelimited(KERN_NOTICE "Route cache is full: "
> - "consider increasing sysctl "
> - "net.ipv[4|6].route.max_size.\n");
> + pr_notice_ratelimited("Route cache is full: consider
> increasing sysctl net.ipv6.route.max_size.\n");
> return NULL;
> }
> }
> diff --git a/net/ipv6/route.c b/net/ipv6/route.c
> index
> 1ff142393c768f85c495474a1d05e1ae1642301c..a9072dba00f4fb0b61bce1fc0f44a3a81ba702fa
> 100644
> --- a/net/ipv6/route.c
> +++ b/net/ipv6/route.c
> @@ -3195,6 +3195,9 @@ static int ip6_dst_gc(struct dst_ops *ops)
> int entries;
>
> entries = dst_entries_get_fast(ops);
> + if (entries > rt_max_size)
> + entries = dst_entries_get_slow(ops);
> +
> if (time_after(rt_last_gc + rt_min_interval, jiffies) &&
if this part of the condition is not satisfied, you are going to call
fib6_run_gc anyways and after that you will update the entries. So I
was wondering if code here could be something like:
--- a/net/ipv6/route.c
+++ b/net/ipv6/route.c
@@ -3197,11 +3197,16 @@ static int ip6_dst_gc(struct dst_ops *ops)
unsigned long rt_last_gc = net->ipv6.ip6_rt_last_gc;
int entries;
+ if (time_before(rt_last_gc + rt_min_interval, jiffies)
+ goto run_gc;
+
entries = dst_entries_get_fast(ops);
- if (time_after(rt_last_gc + rt_min_interval, jiffies) &&
- entries <= rt_max_size)
+ if (entries > rt_max_size)
+ entries = dst_entries_get_slow(ops);
+ if (entries <= rt_max_size)
goto out;
+run_gc:
net->ipv6.ip6_rt_gc_expire++;
fib6_run_gc(net->ipv6.ip6_rt_gc_expire, net, true);
entries = dst_entries_get_slow(ops);
That way you could potentially avoid an extra call to
dst_entries_get_slow when you know for sure that fib6_run_gc will be
run. WDYT?
> entries <= rt_max_size)
> goto out;
> --
> 2.26.2.645.ge9eca65c58-goog
>
