netdev - Re: [PATCH net-next] net/dst: use a smaller percpu_counter batch for dst entries accounting

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <CABCgpaUqymfoGGyExvKv65UDvLfHnw2cavVCr1Pq8coz21ujKA@mail.gmail.com>
Date:   Fri, 8 May 2020 10:30:17 -0700
From:   Brian Vazquez <brianvv.kernel@...il.com>
To:     Eric Dumazet <edumazet@...gle.com>
Cc:     "David S . Miller" <davem@...emloft.net>,
        netdev <netdev@...r.kernel.org>,
        Eric Dumazet <eric.dumazet@...il.com>
Subject: Re: [PATCH net-next] net/dst: use a smaller percpu_counter batch for
 dst entries accounting

On Thu, May 7, 2020 at 7:00 PM Eric Dumazet <edumazet@...gle.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 <edumazet@...gle.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
>