[PATCH v3] net: use exponential backoff in netdev_wait_allrefs

[PATCH v3] net: use exponential backoff in netdev_wait_allrefs

[Francesco Ruggeri]

The combination of aca_free_rcu, introduced in commit 2384d02520ff
("net/ipv6: Add anycast addresses to a global hashtable"), and
fib6_info_destroy_rcu, introduced in commit 9b0a8da8c4c6 ("net/ipv6:
respect rcu grace period before freeing fib6_info"), can result in
an extra rcu grace period being needed when deleting an interface,
with the result that netdev_wait_allrefs ends up hitting the msleep(250),
which is considerably longer than the required grace period.
This can result in long delays when deleting a large number of interfaces,
and it can be observed with this script:

ns=dummy-ns
NIFS=100

ip netns add $ns
ip netns exec $ns ip link set lo up
ip netns exec $ns sysctl net.ipv6.conf.default.disable_ipv6=0
ip netns exec $ns sysctl net.ipv6.conf.default.forwarding=1

for ((i=0; i<$NIFS; i++))
do
        if=eth$i
        ip netns exec $ns ip link add $if type dummy
        ip netns exec $ns ip link set $if up
        ip netns exec $ns ip -6 addr add 2021:$i::1/120 dev $if
done

for ((i=0; i<$NIFS; i++))
do
        if=eth$i
        ip netns exec $ns ip link del $if
done

ip netns del $ns

This patch uses exponential backoff instead of the fixed msleep(250)
to get out of the loop faster.

Time with this patch on a 5.4 kernel:

real	0m8.199s
user	0m0.402s
sys	0m1.213s

Time without this patch:

real	0m31.522s
user	0m0.438s
sys	0m1.156s

v2: use exponential backoff instead of trying to wake up
    netdev_wait_allrefs.
v3: preserve reverse christmas tree ordering of local variables

Signed-off-by: Francesco Ruggeri <fruggeri@arista.com>
---
 net/core/dev.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/net/core/dev.c b/net/core/dev.c
index 4086d335978c..e5fa60cb8832 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -9986,9 +9986,12 @@ EXPORT_SYMBOL(netdev_refcnt_read);
  * We can get stuck here if buggy protocols don't correctly
  * call dev_put.
  */
+#define MIN_MSLEEP	((unsigned int)16)
+#define MAX_MSLEEP	((unsigned int)250)
 static void netdev_wait_allrefs(struct net_device *dev)
 {
 	unsigned long rebroadcast_time, warning_time;
+	unsigned int wait = MIN_MSLEEP;
 	int refcnt;
 
 	linkwatch_forget_dev(dev);
@@ -10023,7 +10026,8 @@ static void netdev_wait_allrefs(struct net_device *dev)
 			rebroadcast_time = jiffies;
 		}
 
-		msleep(250);
+		msleep(wait);
+		wait = min(wait << 1, MAX_MSLEEP);
 
 		refcnt = netdev_refcnt_read(dev);
 
-- 
2.28.0

Re: [PATCH v3] net: use exponential backoff in netdev_wait_allrefs

[Stephen Hemminger]

On Thu, 17 Sep 2020 16:49:53 -0700
fruggeri@arista.com (Francesco Ruggeri) wrote:

> The combination of aca_free_rcu, introduced in commit 2384d02520ff
> ("net/ipv6: Add anycast addresses to a global hashtable"), and
> fib6_info_destroy_rcu, introduced in commit 9b0a8da8c4c6 ("net/ipv6:
> respect rcu grace period before freeing fib6_info"), can result in
> an extra rcu grace period being needed when deleting an interface,
> with the result that netdev_wait_allrefs ends up hitting the msleep(250),
> which is considerably longer than the required grace period.
> This can result in long delays when deleting a large number of interfaces,
> and it can be observed with this script:
> 

Is there anyway to make RCU trigger faster?

Re: [PATCH v3] net: use exponential backoff in netdev_wait_allrefs

[Francesco Ruggeri]

On Thu, Sep 17, 2020 at 5:02 PM Stephen Hemminger
<stephen@networkplumber.org> wrote:
> Is there anyway to make RCU trigger faster?

This is a case of the networking code requiring multiple cascading grace periods
(functions executing at the end of a period scheduling more functions
for the end
of the next period), so it's a matter of how many steps are required,
rather than
how fast each step is, if that is what you are suggesting. I don't
think that expediting
rcu periods would help in this case, but I will defer to people with better
knowledge of the networking code than me to comment.

Re: [PATCH v3] net: use exponential backoff in netdev_wait_allrefs

[Eric Dumazet]

On Fri, Sep 18, 2020 at 1:49 AM Francesco Ruggeri <fruggeri@arista.com> wrote:
>
> The combination of aca_free_rcu, introduced in commit 2384d02520ff
> ("net/ipv6: Add anycast addresses to a global hashtable"), and
> fib6_info_destroy_rcu, introduced in commit 9b0a8da8c4c6 ("net/ipv6:
> respect rcu grace period before freeing fib6_info"), can result in
> an extra rcu grace period being needed when deleting an interface,
> with the result that netdev_wait_allrefs ends up hitting the msleep(250),
> which is considerably longer than the required grace period.
> This can result in long delays when deleting a large number of interfaces,
> and it can be observed with this script:
>
> ns=dummy-ns
> NIFS=100
>
> ip netns add $ns
> ip netns exec $ns ip link set lo up
> ip netns exec $ns sysctl net.ipv6.conf.default.disable_ipv6=0
> ip netns exec $ns sysctl net.ipv6.conf.default.forwarding=1
>
> for ((i=0; i<$NIFS; i++))
> do
>         if=eth$i
>         ip netns exec $ns ip link add $if type dummy
>         ip netns exec $ns ip link set $if up
>         ip netns exec $ns ip -6 addr add 2021:$i::1/120 dev $if
> done
>
> for ((i=0; i<$NIFS; i++))
> do
>         if=eth$i
>         ip netns exec $ns ip link del $if
> done
>
> ip netns del $ns
>
> This patch uses exponential backoff instead of the fixed msleep(250)
> to get out of the loop faster.
>
> Time with this patch on a 5.4 kernel:
>
> real    0m8.199s
> user    0m0.402s
> sys     0m1.213s
>
> Time without this patch:
>
> real    0m31.522s
> user    0m0.438s
> sys     0m1.156s
>
> v2: use exponential backoff instead of trying to wake up
>     netdev_wait_allrefs.
> v3: preserve reverse christmas tree ordering of local variables
>
> Signed-off-by: Francesco Ruggeri <fruggeri@arista.com>
> ---
>  net/core/dev.c | 6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)
>
> diff --git a/net/core/dev.c b/net/core/dev.c
> index 4086d335978c..e5fa60cb8832 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -9986,9 +9986,12 @@ EXPORT_SYMBOL(netdev_refcnt_read);
>   * We can get stuck here if buggy protocols don't correctly
>   * call dev_put.
>   */
> +#define MIN_MSLEEP     ((unsigned int)16)
> +#define MAX_MSLEEP     ((unsigned int)250)


No need for a cast, also I would use names less likely to collide with
include files, and I would start at 1 ms.

#define WAIT_REFS_MIN_MSECS 1
#define WAIT_REFS_MAX_MSECS 250

>
>  static void netdev_wait_allrefs(struct net_device *dev)
>  {
>         unsigned long rebroadcast_time, warning_time;
> +       unsigned int wait = MIN_MSLEEP;


int wait =  WAIT_REFS_MIN_MSECS;

>
>         int refcnt;
>
>         linkwatch_forget_dev(dev);
> @@ -10023,7 +10026,8 @@ static void netdev_wait_allrefs(struct net_device *dev)
>                         rebroadcast_time = jiffies;
>                 }
>
> -               msleep(250);
> +               msleep(wait);
> +               wait = min(wait << 1, MAX_MSLEEP);



wait = min(wait << 1,  WAIT_REFS_MAX_MSECS);

>
>
>                 refcnt = netdev_refcnt_read(dev);
>
> --
> 2.28.0



Also, I would try using synchronize_rcu() instead of the first
msleep(), this might avoid all msleep() calls in your case.

Patch without the macros to see the general idea :

diff --git a/net/core/dev.c b/net/core/dev.c
index 266073e300b5fc21440ea8f8ffc9306a1fc9f370..2d3b65034bc0dd99017dea846e6c0a966f1207ee
100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -9989,7 +9989,7 @@ EXPORT_SYMBOL(netdev_refcnt_read);
 static void netdev_wait_allrefs(struct net_device *dev)
 {
        unsigned long rebroadcast_time, warning_time;
-       int refcnt;
+       int wait = 0, refcnt;

        linkwatch_forget_dev(dev);

@@ -10023,8 +10023,13 @@ static void netdev_wait_allrefs(struct net_device *dev)
                        rebroadcast_time = jiffies;
                }

-               msleep(250);
-
+               if (!wait) {
+                       synchronize_rcu();
+                       wait = 1;
+               } else {
+                       msleep(wait);
+                       wait = min(wait << 1, 250);
+               }
                refcnt = netdev_refcnt_read(dev);

                if (refcnt && time_after(jiffies, warning_time + 10 * HZ)) {

Re: [PATCH v3] net: use exponential backoff in netdev_wait_allrefs

[Eric Dumazet]

On Fri, Sep 18, 2020 at 10:48 AM Eric Dumazet <edumazet@google.com> wrote:

>
>
> Also, I would try using synchronize_rcu() instead of the first

s/synchronize_rcu/rcu_barrier/  of course :/

> msleep(), this might avoid all msleep() calls in your case.
>
> Patch without the macros to see the general idea :
>
> diff --git a/net/core/dev.c b/net/core/dev.c
> index 266073e300b5fc21440ea8f8ffc9306a1fc9f370..2d3b65034bc0dd99017dea846e6c0a966f1207ee
> 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -9989,7 +9989,7 @@ EXPORT_SYMBOL(netdev_refcnt_read);
>  static void netdev_wait_allrefs(struct net_device *dev)
>  {
>         unsigned long rebroadcast_time, warning_time;
> -       int refcnt;
> +       int wait = 0, refcnt;
>
>         linkwatch_forget_dev(dev);
>
> @@ -10023,8 +10023,13 @@ static void netdev_wait_allrefs(struct net_device *dev)
>                         rebroadcast_time = jiffies;
>                 }
>
> -               msleep(250);
> -
> +               if (!wait) {
> +                       synchronize_rcu();

rcu_barrier();

> +                       wait = 1;
> +               } else {
> +                       msleep(wait);
> +                       wait = min(wait << 1, 250);
> +               }
>                 refcnt = netdev_refcnt_read(dev);
>
>                 if (refcnt && time_after(jiffies, warning_time + 10 * HZ)) {