This changeset fixes a common problem for busy networking kthreads.
These threads, e.g. NAPI threads, typically will do:

* polling a batch of packets
* if there are more work, call cond_resched() to allow scheduling
* continue to poll more packets when rx queue is not empty

We observed this being a problem in production, since it can block RCU
tasks from making progress under heavy load. Investigation indicates
that just calling cond_resched() is insufficient for RCU tasks to reach
quiescent states. This also has the side effect of frequently clearing
the TIF_NEED_RESCHED flag on voluntary preempt kernels. As a result,
schedule() will not be called in these circumstances, despite schedule()
in fact provides required quiescent states. This at least affects NAPI
threads, napi_busy_loop, and also cpumap kthread.

By reporting RCU QSes in these kthreads periodically before cond_resched, the
blocked RCU waiters can correctly progress. Instead of just reporting QS for
RCU tasks, these code share the same concern as noted in the commit
d28139c4e967 ("rcu: Apply RCU-bh QSes to RCU-sched and RCU-preempt when safe").
So report a consolidated QS for safety.

It is worth noting that, although this problem is reproducible in
napi_busy_loop, it only shows up when setting the polling interval to as high
as 2ms, which is far larger than recommended 50us-100us in the documentation.
So napi_busy_loop is left untouched.

Lastly, this does not affect RT kernels, which does not enter the scheduler
through cond_resched(). Without the mentioned side effect, schedule() will
be called time by time, and clear the RCU task holdouts.

V4: https://lore.kernel.org/bpf/cover.1710525524.git.yan@cloudflare.com/
V3: https://lore.kernel.org/lkml/20240314145459.7b3aedf1@kernel.org/t/
V2: https://lore.kernel.org/bpf/ZeFPz4D121TgvCje@debian.debian/
V1: https://lore.kernel.org/lkml/Zd4DXTyCf17lcTfq@debian.debian/#t

changes since v4:
 * polished comments and docs for the RCU helper as Paul McKenney suggested

changes since v3:
 * fixed kernel-doc errors

changes since v2:
 * created a helper in rcu header to abstract the behavior
 * fixed cpumap kthread in addition

changes since v1:
 * disable preemption first as Paul McKenney suggested

Yan Zhai (3):
  rcu: add a helper to report consolidated flavor QS
  net: report RCU QS on threaded NAPI repolling
  bpf: report RCU QS in cpumap kthread

 include/linux/rcupdate.h | 31 +++++++++++++++++++++++++++++++
 kernel/bpf/cpumap.c      |  3 +++
 net/core/dev.c           |  3 +++
 3 files changed, 37 insertions(+)

-- 
2.30.2

On 19/03/2024 21.44, Yan Zhai wrote:
> This changeset fixes a common problem for busy networking kthreads.
> These threads, e.g. NAPI threads, typically will do:
> 
> * polling a batch of packets
> * if there are more work, call cond_resched() to allow scheduling
> * continue to poll more packets when rx queue is not empty
> 
> We observed this being a problem in production, since it can block RCU
> tasks from making progress under heavy load. Investigation indicates
> that just calling cond_resched() is insufficient for RCU tasks to reach
> quiescent states. This also has the side effect of frequently clearing
> the TIF_NEED_RESCHED flag on voluntary preempt kernels. As a result,
> schedule() will not be called in these circumstances, despite schedule()
> in fact provides required quiescent states. This at least affects NAPI
> threads, napi_busy_loop, and also cpumap kthread.
> 
> By reporting RCU QSes in these kthreads periodically before cond_resched, the
> blocked RCU waiters can correctly progress. Instead of just reporting QS for
> RCU tasks, these code share the same concern as noted in the commit
> d28139c4e967 ("rcu: Apply RCU-bh QSes to RCU-sched and RCU-preempt when safe").
> So report a consolidated QS for safety.
> 
> It is worth noting that, although this problem is reproducible in
> napi_busy_loop, it only shows up when setting the polling interval to as high
> as 2ms, which is far larger than recommended 50us-100us in the documentation.
> So napi_busy_loop is left untouched.
> 
> Lastly, this does not affect RT kernels, which does not enter the scheduler
> through cond_resched(). Without the mentioned side effect, schedule() will
> be called time by time, and clear the RCU task holdouts.
> 
> V4: https://lore.kernel.org/bpf/cover.1710525524.git.yan@cloudflare.com/
> V3: https://lore.kernel.org/lkml/20240314145459.7b3aedf1@kernel.org/t/
> V2: https://lore.kernel.org/bpf/ZeFPz4D121TgvCje@debian.debian/
> V1: https://lore.kernel.org/lkml/Zd4DXTyCf17lcTfq@debian.debian/#t
> 
> changes since v4:
>   * polished comments and docs for the RCU helper as Paul McKenney suggested
> 
> changes since v3:
>   * fixed kernel-doc errors
> 
> changes since v2:
>   * created a helper in rcu header to abstract the behavior
>   * fixed cpumap kthread in addition
> 
> changes since v1:
>   * disable preemption first as Paul McKenney suggested
> 
> Yan Zhai (3):
>    rcu: add a helper to report consolidated flavor QS
>    net: report RCU QS on threaded NAPI repolling
>    bpf: report RCU QS in cpumap kthread
> 
>   include/linux/rcupdate.h | 31 +++++++++++++++++++++++++++++++
>   kernel/bpf/cpumap.c      |  3 +++
>   net/core/dev.c           |  3 +++
>   3 files changed, 37 insertions(+)
> 

Acked-by: Jesper Dangaard Brouer <hawk@kernel.org>