net: raise RCU qs after each threaded NAPI poll

We noticed task RCUs being blocked when threaded NAPIs are very busy at
workloads: detaching any BPF tracing programs, i.e. removing a ftrace trampoline, will simply block for very long in rcu_tasks_wait_gp. This ranges from hundreds of seconds to even an hour, severely harming any observability tools that rely on BPF tracing programs. It can be easily reproduced locally with following setup:

ip netns add test1
ip netns add test2

ip -n test1 link add veth1 type veth peer name veth2 netns test2

ip -n test1 link set veth1 up
ip -n test1 link set lo up
ip -n test2 link set veth2 up
ip -n test2 link set lo up

ip -n test1 addr add 192.168.1.2/31 dev veth1
ip -n test1 addr add 1.1.1.1/32 dev lo
ip -n test2 addr add 192.168.1.3/31 dev veth2
ip -n test2 addr add 2.2.2.2/31 dev lo

ip -n test1 route add default via 192.168.1.3
ip -n test2 route add default via 192.168.1.2

for i in seq 10 210; do
for j in seq 10 210; do

    ip netns exec test2 iptables -I INPUT -s 3.3.$i.$j -p udp --dport 5201

done
done

ip netns exec test2 ethtool -K veth2 gro on
ip netns exec test2 bash -c 'echo 1 > /sys/class/net/veth2/threaded' ip netns exec test1 ethtool -K veth1 tso off

Then run an iperf3 client/server and a bpftrace script can trigger it:

ip netns exec test2 iperf3 -s -B 2.2.2.2 >/dev/null&
ip netns exec test1 iperf3 -c 2.2.2.2 -B 1.1.1.1 -u -l 1500 -b 3g -t 100 >/dev/null& bpftrace -e 'kfunc:__napi_poll{@=count();} interval:s:1{exit();}'

Above reproduce for net-next kernel with following RCU and preempt configuraitons:

# RCU Subsystem
CONFIG_TREE_RCU=y
CONFIG_PREEMPT_RCU=y
# CONFIG_RCU_EXPERT is not set
CONFIG_SRCU=y
CONFIG_TREE_SRCU=y
CONFIG_TASKS_RCU_GENERIC=y
CONFIG_TASKS_RCU=y
CONFIG_TASKS_RUDE_RCU=y
CONFIG_TASKS_TRACE_RCU=y
CONFIG_RCU_STALL_COMMON=y
CONFIG_RCU_NEED_SEGCBLIST=y

# end of RCU Subsystem
# RCU Debugging
# CONFIG_RCU_SCALE_TEST is not set
# CONFIG_RCU_TORTURE_TEST is not set
# CONFIG_RCU_REF_SCALE_TEST is not set
CONFIG_RCU_CPU_STALL_TIMEOUT=21
CONFIG_RCU_EXP_CPU_STALL_TIMEOUT=0
# CONFIG_RCU_TRACE is not set
# CONFIG_RCU_EQS_DEBUG is not set
# end of RCU Debugging

CONFIG_PREEMPT_BUILD=y
# CONFIG_PREEMPT_NONE is not set
CONFIG_PREEMPT_VOLUNTARY=y
# CONFIG_PREEMPT is not set
CONFIG_PREEMPT_COUNT=y
CONFIG_PREEMPTION=y
CONFIG_PREEMPT_DYNAMIC=y
CONFIG_PREEMPT_RCU=y
CONFIG_HAVE_PREEMPT_DYNAMIC=y
CONFIG_HAVE_PREEMPT_DYNAMIC_CALL=y
CONFIG_PREEMPT_NOTIFIERS=y
# CONFIG_DEBUG_PREEMPT is not set
# CONFIG_PREEMPT_TRACER is not set
# CONFIG_PREEMPTIRQ_DELAY_TEST is not set

An interesting observation is that, while tasks RCUs are blocked, related NAPI thread is still being scheduled (even across cores) regularly. Looking at the gp conditions, I am inclining to cond_resched after each __napi_poll being the problem: cond_resched enters the scheduler with PREEMPT bit, which does not account as a gp for tasks RCUs. Meanwhile, since the thread has been frequently resched, the normal scheduling point (no PREEMPT bit, accounted as a task RCU gp) seems to have very little chance to kick in. Given the nature of "busy polling" program, such NAPI thread won't have task->nvcsw or task->on_rq updated (other gp conditions), the result is that such NAPI thread is put on RCU holdouts list for indefinitely long time.

This is simply fixed by adapting similar behavior of ksoftirqd: after the thread repolls for a while, raise a RCU QS to help expedite the tasks RCU grace period. No more blocking afterwards.

Some brief iperf3 throughput testing in my VM with net-next kernel shows no noteable perf difference with 1500 byte MTU for 10 repeat runs each:

Before:
UDP: 3.073Gbps (+-0.070Gbps)
TCP: 37.850Gbps (+-1.947Gbps)

After:
UDP: 3.077Gbps (+-0.121 Gbps)
TCP: 38.120Gbps (+-2.272 Gbps)

Note I didn't enable GRO for UDP so its throughput is lower than TCP.

Fixes 29863d41bb6e ("net: implement threaded-able napi poll loop support")

Suggested-by Paul E. McKenney <[email protected]>

Reviewed-by Joel Fernandes (Google) <[email protected]>

Signed-off-by Yan Zhai <[email protected]>

---
v1->v2: moved rcu_softirq_qs out from bh critical section, and only raise it after a second of repolling. Added some brief perf test result.

---
 net/core/dev.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)

net/core/dev.c +18 -1

index 275fd5259a4a..76cff3849e1f 100644
@@ -6751,9 +6751,12 @@ static int napi_threaded_poll(void *data)
 {
        struct napi_struct *napi = data;
        struct softnet_data *sd;
+       unsigned long next_qs;
        void *have;
 
        while (!napi_thread_wait(napi)) {
+               next_qs = jiffies + HZ;
+
                for (;;) {
                        bool repoll = false;
 

@@ -6778,6 +6781,21 @@ static int napi_threaded_poll(void *data)
                        if (!repoll)
                                break;
 
+                       /* cond_resched cannot unblock tasks RCU writers, so it
+                        * is necessary to relax periodically and raise a QS to
+                        * avoid starving writers under frequent repoll, e.g.
+                        * ftrace trampoline clean up work. When not repoll,
+                        * napi_thread_wait will enter sleep and have the same
+                        * QS effect.
+                        */
+                       if (!IS_ENABLED(CONFIG_PREEMPT_RT) &&
+                           time_after(jiffies, next_qs)) {
+                               preempt_disable();
+                               rcu_softirq_qs();
+                               preempt_enable();
+                               next_qs = jiffies + HZ;
+                       }
+
                        cond_resched();
                }
        }
--

2.30.2

Quoted 5 lines · Yan Zhai

We noticed task RCUs being blocked when threaded NAPIs are very busy at
workloads: detaching any BPF tracing programs, i.e. removing a ftrace trampoline, will simply block for very long in rcu_tasks_wait_gp. This ranges from hundreds of seconds to even an hour, severely harming any

Quoted 9 lines

Fixes 29863d41bb6e ("net: implement threaded-able napi poll loop support")

Suggested-by Paul E. McKenney <[email protected]>

Reviewed-by Joel Fernandes (Google) <[email protected]>

Signed-off-by Yan Zhai <[email protected]>

---
v1->v2: moved rcu_softirq_qs out from bh critical section, and only raise it after a second of repolling. Added some brief perf test result.

Link to v1: https://lore.kernel.org/netdev/[email protected]/T/#u
And I apparently forgot to rename the subject since it's not raising after every poll (let me know if it is prefered to send a V3 to fix it)

I could not see the reason for 1sec (HZ) delays.

Would calling rcu_softirq_qs() every ~10ms instead be a serious issue ?

In anycase, if this all about rcu_tasks, I would prefer using a macro defined in kernel/rcu/tasks.h
instead of having a hidden constant in a networking core function.

Thanks.

Hi Eric,

Quoted 5 lines · Eric Dumazet

I could not see the reason for 1sec (HZ) delays.

Would calling rcu_softirq_qs() every ~10ms instead be a serious issue ?

The trouble scenarios are often when we need to detach an ad-hoc BPF tracing program, or restart a monitoring service. It is fine as long as they do not block for 10+ seconds or even completely stall under heavy traffic. Raising a QS every few ms or HZ both work in such cases.

Quoted 3 lines

In anycase, if this all about rcu_tasks, I would prefer using a macro defined in kernel/rcu/tasks.h
instead of having a hidden constant in a networking core function.

Paul E. McKenney was suggesting either current form or

         local_bh_enable();
         if (!IS_ENABLED(CONFIG_PREEMPT_RT))
                 rcu_softirq_qs_enable(local_bh_enable());
         else
                 local_bh_enable();

With an interval it might have to be "rcu_softirq_qs_enable(local_bh_enable(), &next_qs);" to avoid an unnecessary extern/static var. Will it make more sense to you?

Quoted 26 lines · Yan Zhai

Hi Eric,

Quoted 5 lines · Eric Dumazet

I could not see the reason for 1sec (HZ) delays.

Would calling rcu_softirq_qs() every ~10ms instead be a serious issue ?

The trouble scenarios are often when we need to detach an ad-hoc BPF tracing program, or restart a monitoring service. It is fine as long as they do not block for 10+ seconds or even completely stall under heavy traffic. Raising a QS every few ms or HZ both work in such cases.

Quoted 3 lines

In anycase, if this all about rcu_tasks, I would prefer using a macro defined in kernel/rcu/tasks.h
instead of having a hidden constant in a networking core function.

Paul E. McKenney was suggesting either current form or

         local_bh_enable();
         if (!IS_ENABLED(CONFIG_PREEMPT_RT))
                 rcu_softirq_qs_enable(local_bh_enable());
         else
                 local_bh_enable();

Quoted 3 lines

With an interval it might have to be "rcu_softirq_qs_enable(local_bh_enable(), &next_qs);" to avoid an unnecessary extern/static var. Will it make more sense to you?

I was thinking in terms of something like this (untested):

        #define rcu_softirq_qs_enable(enable_stmt, oldj) \
        do { \
                if (!IS_ENABLED(CONFIG_PREEMPT_RT) && \
                    time_after(oldj + HZ / 10, jiffies) { \
                        rcu_softirq_qs(); \
                        (oldj) = jiffies; \
                } \
                do  { enable_stmt; } while (0) \
        } while (0)

Then the call could be "rcu_softirq_qs_enable(local_bh_enable(), last_qs)", where last_qs is initialized by the caller to jiffies.

The reason for putting "enable_stmt;" into anothor do-while loop is in case someone typos an "else" as the first part of the "enable_stmt" argument.

Would that work?

Thanx, Paul

Quoted 51 lines · Paul E. McKenney

Quoted 26 lines · Yan Zhai

Hi Eric,

Quoted 5 lines · Eric Dumazet

I could not see the reason for 1sec (HZ) delays.

Would calling rcu_softirq_qs() every ~10ms instead be a serious issue ?

The trouble scenarios are often when we need to detach an ad-hoc BPF tracing program, or restart a monitoring service. It is fine as long as they do not block for 10+ seconds or even completely stall under heavy traffic. Raising a QS every few ms or HZ both work in such cases.

Quoted 3 lines

In anycase, if this all about rcu_tasks, I would prefer using a macro defined in kernel/rcu/tasks.h
instead of having a hidden constant in a networking core function.

Paul E. McKenney was suggesting either current form or

         local_bh_enable();
         if (!IS_ENABLED(CONFIG_PREEMPT_RT))
                 rcu_softirq_qs_enable(local_bh_enable());
         else
                 local_bh_enable();

Quoted 3 lines

With an interval it might have to be "rcu_softirq_qs_enable(local_bh_enable(), &next_qs);" to avoid an unnecessary extern/static var. Will it make more sense to you?

I was thinking in terms of something like this (untested):

        #define rcu_softirq_qs_enable(enable_stmt, oldj) \
        do { \
                if (!IS_ENABLED(CONFIG_PREEMPT_RT) && \
                    time_after(oldj + HZ / 10, jiffies) { \
                        rcu_softirq_qs(); \
                        (oldj) = jiffies; \
                } \
                do  { enable_stmt; } while (0) \
        } while (0)

Then the call could be "rcu_softirq_qs_enable(local_bh_enable(), last_qs)", where last_qs is initialized by the caller to jiffies.

The reason for putting "enable_stmt;" into anothor do-while loop is in case someone typos an "else" as the first part of the "enable_stmt" argument.

Quoted 2 lines

Would that work?

Thanks Paul, just got time to continue this thread as I was travelling. I think it is probably better to move preempt_disable/enable into the macro to avoid the friction. And also since this can affect NAPI thread, NAPI busy loop and XDP cpu map thread (+Jesper who reminded me about this), let me send a v3 later to cover all of those places.