[PATCH net] tcp/dccp: do not block bh too long in inet_twdr_twkill_work()

[PATCH net] tcp/dccp: do not block bh too long in inet_twdr_twkill_work()

[Eric Dumazet]

From: Eric Dumazet <edumazet@google.com>

I have seen inet_twdr_twkill_work() blocking softirq for
periods up to 1.5 seconds, depending on number of timewait sockets.

This is an unacceptable source of latency.

Note that inet_twdr_do_twkill_work() releases death_lock spinlock
for every tw handled, but does not take care of bh enabling.

Signed-off-by: Eric Dumazet <edumazet@google.com>
---
 net/ipv4/inet_timewait_sock.c |    8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)

diff --git a/net/ipv4/inet_timewait_sock.c b/net/ipv4/inet_timewait_sock.c
index 6d592f8555fb..9c7f480d6ad8 100644
--- a/net/ipv4/inet_timewait_sock.c
+++ b/net/ipv4/inet_timewait_sock.c
@@ -303,11 +303,9 @@ void inet_twdr_twkill_work(struct work_struct *work)
 				continue;
 
 			while (inet_twdr_do_twkill_work(twdr, i) != 0) {
-				if (need_resched()) {
-					spin_unlock_bh(&twdr->death_lock);
-					schedule();
-					spin_lock_bh(&twdr->death_lock);
-				}
+				spin_unlock_bh(&twdr->death_lock);
+				cond_resched();
+				spin_lock_bh(&twdr->death_lock);
 			}
 
 			twdr->thread_slots &= ~(1 << i);

Re: [PATCH net] tcp/dccp: do not block bh too long in inet_twdr_twkill_work()

[Eric Dumazet]

On Fri, 2015-04-10 at 04:19 -0700, Eric Dumazet wrote:
> From: Eric Dumazet <edumazet@google.com>
> 
> I have seen inet_twdr_twkill_work() blocking softirq for
> periods up to 1.5 seconds, depending on number of timewait sockets.
> 
> This is an unacceptable source of latency.
> 
> Note that inet_twdr_do_twkill_work() releases death_lock spinlock
> for every tw handled, but does not take care of bh enabling.
> 
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> ---

BTW, inet_twdr_twcal_tick() suffers from similar problem, with latencies
of ~25ms, when /proc/sys/net/ipv4/tcp_tw_recycle is enabled.

Since it holds death_lock for the whole run, all other cpus are spinning
on it.

I believe we simply should add a timer per timewait, as I did for
request sockets.

This would simplify the code a lot, and would get rid of this awful non
scalable timer wheel.

Re: [PATCH net] tcp/dccp: do not block bh too long in inet_twdr_twkill_work()

[David Miller]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Fri, 10 Apr 2015 04:19:07 -0700

> From: Eric Dumazet <edumazet@google.com>
> 
> I have seen inet_twdr_twkill_work() blocking softirq for
> periods up to 1.5 seconds, depending on number of timewait sockets.
> 
> This is an unacceptable source of latency.
> 
> Note that inet_twdr_do_twkill_work() releases death_lock spinlock
> for every tw handled, but does not take care of bh enabling.
> 
> Signed-off-by: Eric Dumazet <edumazet@google.com>

I think it makes sense to use local_softirq_pending() here rather
than flip the lock unconditionally.

Re: [PATCH net] tcp/dccp: do not block bh too long in inet_twdr_twkill_work()

[Eric Dumazet]

On Sun, 2015-04-12 at 21:04 -0400, David Miller wrote:
> From: Eric Dumazet <eric.dumazet@gmail.com>
> Date: Fri, 10 Apr 2015 04:19:07 -0700
> 
> > From: Eric Dumazet <edumazet@google.com>
> > 
> > I have seen inet_twdr_twkill_work() blocking softirq for
> > periods up to 1.5 seconds, depending on number of timewait sockets.
> > 
> > This is an unacceptable source of latency.
> > 
> > Note that inet_twdr_do_twkill_work() releases death_lock spinlock
> > for every tw handled, but does not take care of bh enabling.
> > 
> > Signed-off-by: Eric Dumazet <edumazet@google.com>
> 
> I think it makes sense to use local_softirq_pending() here rather
> than flip the lock unconditionally.

I thought about that, but this meant having to test 2 conditions.

Not sure we need this patch anyway, if we merge the 'tcp/dccp: get rid
of central timewait timer' in 4.1 

Its not like it is a new bug ...

diff --git a/net/ipv4/inet_timewait_sock.c b/net/ipv4/inet_timewait_sock.c
index 6d592f8555fb..942b582d044e 100644
--- a/net/ipv4/inet_timewait_sock.c
+++ b/net/ipv4/inet_timewait_sock.c
@@ -303,9 +303,10 @@ void inet_twdr_twkill_work(struct work_struct *work)
 				continue;
 
 			while (inet_twdr_do_twkill_work(twdr, i) != 0) {
-				if (need_resched()) {
+				if (local_softirq_pending() ||
+				    need_resched()) {
 					spin_unlock_bh(&twdr->death_lock);
-					schedule();
+					cond_resched();
 					spin_lock_bh(&twdr->death_lock);
 				}
 			}

Re: [PATCH net] tcp/dccp: do not block bh too long in inet_twdr_twkill_work()

[David Miller]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Sun, 12 Apr 2015 18:52:37 -0700

> On Sun, 2015-04-12 at 21:04 -0400, David Miller wrote:
>> From: Eric Dumazet <eric.dumazet@gmail.com>
>> Date: Fri, 10 Apr 2015 04:19:07 -0700
>> 
>> > From: Eric Dumazet <edumazet@google.com>
>> > 
>> > I have seen inet_twdr_twkill_work() blocking softirq for
>> > periods up to 1.5 seconds, depending on number of timewait sockets.
>> > 
>> > This is an unacceptable source of latency.
>> > 
>> > Note that inet_twdr_do_twkill_work() releases death_lock spinlock
>> > for every tw handled, but does not take care of bh enabling.
>> > 
>> > Signed-off-by: Eric Dumazet <edumazet@google.com>
>> 
>> I think it makes sense to use local_softirq_pending() here rather
>> than flip the lock unconditionally.
> 
> I thought about that, but this meant having to test 2 conditions.
> 
> Not sure we need this patch anyway, if we merge the 'tcp/dccp: get rid
> of central timewait timer' in 4.1 
> 
> Its not like it is a new bug ...

Yeah, since 4.0 is out already, probably just pass on this patch.