Re: [PATCH 1/9] sched/balancing: Switch the 'DEFINE_SPINLOCK(balancing)' spinlock into an 'atomic_t sched_balance_running' flag

[Valentin Schneider]

On 04/03/24 10:48, Ingo Molnar wrote:
> The 'balancing' spinlock added in:
>
>   08c183f31bdb ("[PATCH] sched: add option to serialize load balancing")
>
> ... is taken when the SD_SERIALIZE flag is set in a domain, but in reality it
> is a glorified global atomic flag serializing the load-balancing of
> those domains.
>
> It doesn't have any explicit locking semantics per se: we just
> spin_trylock() it.
>
> Turn it into a ... global atomic flag. This makes it more
> clear what is going on here, and reduces overhead and code
> size a bit:
>
>   # kernel/sched/fair.o: [x86-64 defconfig]
>
>      text	   data	    bss	    dec	    hex	filename
>     60730	   2721	    104	  63555	   f843	fair.o.before
>     60718	   2721	    104	  63543	   f837	fair.o.after
>
> Also document the flag a bit.
>
> No change in functionality intended.
>
> Signed-off-by: Ingo Molnar <mingo@xxxxxxxxxx>
> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Cc: Vincent Guittot <vincent.guittot@xxxxxxxxxx>
> Cc: Dietmar Eggemann <dietmar.eggemann@xxxxxxx>
> Cc: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
> Cc: Valentin Schneider <vschneid@xxxxxxxxxx>

Few comment nits, otherwise:

Reviewed-by: Valentin Schneider <vschneid@xxxxxxxxxx>

> ---
>  kernel/sched/fair.c | 19 ++++++++++++++++---
>  1 file changed, 16 insertions(+), 3 deletions(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 6a16129f9a5c..64ae3d8dc93b 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -11633,7 +11633,20 @@ static int active_load_balance_cpu_stop(void *data)
>       return 0;
>  }
>
> -static DEFINE_SPINLOCK(balancing);
> +/*
> + * This flag serializes load-balancing passes over large domains
> + * (such as SD_NUMA) - only once load-balancing instance may run
                               ^^^^
                           s/once/one/

Also, currently the flag is only set for domains above the NODE topology
level, sd_init() will reject an architecture that forces SD_SERIALIZE in a
topology level's ->sd_flags(), so what about:

s/(such as SD_NUMA)/(above the NODE topology level)

> + * at a time, to reduce overhead on very large systems with lots
> + * of CPUs and large NUMA distances.
> + *
> + * - Note that load-balancing passes triggered while another one
> + *   is executing are skipped and not re-tried.
> + *
> + * - Also note that this does not serialize sched_balance_domains()
                                               ^^^^^^^^^^^^^^^^^^^^^
Did you mean rebalance_domains()?

> + *   execution, as non-SD_SERIALIZE domains will still be
> + *   load-balanced in parallel.
> + */
> +static atomic_t sched_balance_running = ATOMIC_INIT(0);
>
>  /*
>   * Scale the max load_balance interval with the number of CPUs in the system.
> @@ -11711,7 +11724,7 @@ static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
>
>               need_serialize = sd->flags & SD_SERIALIZE;
>               if (need_serialize) {
> -			if (!spin_trylock(&balancing))
> +			if (atomic_cmpxchg_acquire(&sched_balance_running, 0, 1))
>                               goto out;
>               }
>
> @@ -11729,7 +11742,7 @@ static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
>                       interval = get_sd_balance_interval(sd, busy);
>               }
>               if (need_serialize)
> -			spin_unlock(&balancing);
> +			atomic_set_release(&sched_balance_running, 0);
>  out:
>               if (time_after(next_balance, sd->last_balance + interval)) {
>                       next_balance = sd->last_balance + interval;
> --
> 2.40.1