Re: [PATCH 2/2] docs: Update RCU's hotplug requirements with a bit about design

From: Joel Fernandes
Date: Tue Sep 29 2020 - 15:32:53 EST


Hi Paul,

On Tue, Sep 29, 2020 at 03:29:28PM -0400, Joel Fernandes (Google) wrote:
> RCU's hotplug design will help understand the requirements an RCU
> implementation needs to fullfill, such as dead-lock avoidance.
>
> The rcu_barrier() section of the "Hotplug CPU" section already talks
> about deadlocks, however the description of what else can deadlock other
> than rcu_barrier is rather incomplete.
>
> This commit therefore continues the section by describing how RCU's
> design handles CPU hotplug in a deadlock-free way.
>
> Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> ---
> .../RCU/Design/Requirements/Requirements.rst | 30 +++++++++++++++++--
> 1 file changed, 28 insertions(+), 2 deletions(-)
>
> diff --git a/Documentation/RCU/Design/Requirements/Requirements.rst b/Documentation/RCU/Design/Requirements/Requirements.rst
> index 1ae79a10a8de..e0413aa989dd 100644
> --- a/Documentation/RCU/Design/Requirements/Requirements.rst
> +++ b/Documentation/RCU/Design/Requirements/Requirements.rst
> @@ -1929,8 +1929,10 @@ The Linux-kernel CPU-hotplug implementation has notifiers that are used
> to allow the various kernel subsystems (including RCU) to respond
> appropriately to a given CPU-hotplug operation. Most RCU operations may
> be invoked from CPU-hotplug notifiers, including even synchronous
> -grace-period operations such as ``synchronize_rcu()`` and
> -``synchronize_rcu_expedited()``.
> +grace-period operations such as. However, the synchronous variants
> +(``synchronize_rcu()`` and ``synchronize_rcu_expedited()``) should not
> +from notifiers that execute via ``stop_machine()`` -- specifically those

The "should not from notifiers" should be "should not be used from
notifiers" here. Sorry and hope you can fix it up.

thanks,

- Joel



> +between the ``CPUHP_AP_OFFLINE`` and ``CPUHP_AP_ONLINE`` states.
>
> However, all-callback-wait operations such as ``rcu_barrier()`` are also
> not supported, due to the fact that there are phases of CPU-hotplug
> @@ -1940,6 +1942,30 @@ deadlock. Furthermore, ``rcu_barrier()`` blocks CPU-hotplug operations
> during its execution, which results in another type of deadlock when
> invoked from a CPU-hotplug notifier.
>
> +Also, RCU's implementation avoids serious deadlocks which could occur due to
> +interaction between hotplug, timers and grace period processing. It does so by
> +maintaining its own books of every CPU's hotplug state, independent of
> +the existing general-purpose CPU masks and by reporting quiescent states
> +explictly when an online CPU is going down. Due to this design, the force
> +quiescent state loop (FQS) is not required to report quiescent states for
> +offline CPUs, like it does for idle CPUs, but it does splat if offline CPUs are
> +stalling the RCU grace period for too long.
> +
> +For an offline CPU, the quiescent state will be reported in either of:
> +1. During CPU offlining, using RCU's hotplug notifier (``rcu_report_dead()``).
> +2. During grace period initialization (``rcu_gp_init()``) if it detected a race
> + with CPU offlining, or a race with a task unblocking on a node which
> + previously had all of its CPUs offlined.
> +
> +The CPU onlining path (``rcu_cpu_starting()``) does not need to report a
> +quiescent state for an offline CPU; in fact it would trigger a warning if a
> +quiescent state was not already reported for that CPU.
> +
> +During the checking/modification of RCU's hotplug bookkeeping, the
> +corresponding CPU's leaf node lock is held. This avoids race conditions between
> +RCU's hotplug notifier hooks, grace period initialization code and the FQS loop
> +which can concurrently refer to or modify the bookkeeping.
> +
> Scheduler and RCU
> ~~~~~~~~~~~~~~~~~
>
> --
> 2.28.0.709.gb0816b6eb0-goog
>