Quoting Maulik Shah (2020-11-26 02:18:18)
> lockdep_assert_irqs_disabled() was added to check rpmh_flush()
> can only be invoked when irqs are disabled, this is true for
> APPS RSC as the last CPU going to deepest low power mode is
> writing sleep and wake TCSes.
>
> However for RSCs that support solver mode, drivers can invoke
> rpmh_write_sleep_and_wake() to immediately write cached sleep
> and wake sets to TCSes from any CPU. Conditionally check if RSC
> controller supports 'HW solver' mode then do not check for irqs
> disabled as such RSCs can write sleepand wake TCSes at any point.
>
> Signed-off-by: Maulik Shah <mkshah@xxxxxxxxxxxxxx>
> ---
> drivers/soc/qcom/rpmh-internal.h | 5 +++++
> drivers/soc/qcom/rpmh-rsc.c | 3 +++
> drivers/soc/qcom/rpmh.c | 26 ++++++++++++++++++++++----
> 3 files changed, 30 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/soc/qcom/rpmh-internal.h b/drivers/soc/qcom/rpmh-internal.h
> index 79486d6..39fa3c5 100644
> --- a/drivers/soc/qcom/rpmh-internal.h
> +++ b/drivers/soc/qcom/rpmh-internal.h
> @@ -17,6 +17,9 @@
> #define MAX_TCS_NR (MAX_TCS_PER_TYPE * TCS_TYPE_NR)
> #define MAX_TCS_SLOTS (MAX_CMDS_PER_TCS * MAX_TCS_PER_TYPE)
>
> +/* CTRLR specific flags */
> +#define SOLVER_PRESENT 1
> +
> struct rsc_drv;
>
> /**
> @@ -78,6 +81,7 @@ struct rpmh_request {
> * @cache_lock: synchronize access to the cache data
> * @dirty: was the cache updated since flush
> * @in_solver_mode: Controller is busy in solver mode
> + * @flags: Controller specific flags
> * @batch_cache: Cache sleep and wake requests sent as batch
> */
> struct rpmh_ctrlr {
> @@ -85,6 +89,7 @@ struct rpmh_ctrlr {
> spinlock_t cache_lock;
> bool dirty;
> bool in_solver_mode;
> + u32 flags;
Maybe unsigned long is more appropriate? Do we rely on 32-bits vs.
64-bits?
> struct list_head batch_cache;
> };
>
> diff --git a/drivers/soc/qcom/rpmh-rsc.c b/drivers/soc/qcom/rpmh-rsc.c
> index ffb4ca7..4caaddf 100644
> --- a/drivers/soc/qcom/rpmh-rsc.c
> +++ b/drivers/soc/qcom/rpmh-rsc.c
> @@ -1031,6 +1031,9 @@ static int rpmh_rsc_probe(struct platform_device *pdev)
> if (!solver_config) {
> drv->rsc_pm.notifier_call = rpmh_rsc_cpu_pm_callback;
> cpu_pm_register_notifier(&drv->rsc_pm);
> + drv->client.flags &= ~SOLVER_PRESENT;
> + } else {
> + drv->client.flags |= SOLVER_PRESENT;
It looks like this could be tested by checking for
drv->rsc_pm.notifier_call being non-NULL?
> }
>
> /* Enable the active TCS to send requests immediately */
> diff --git a/drivers/soc/qcom/rpmh.c b/drivers/soc/qcom/rpmh.c
> index 725b8f0..604d511 100644
> --- a/drivers/soc/qcom/rpmh.c
> +++ b/drivers/soc/qcom/rpmh.c
> @@ -83,6 +83,9 @@ static int check_ctrlr_state(struct rpmh_ctrlr *ctrlr, enum rpmh_state state)
> if (state != RPMH_ACTIVE_ONLY_STATE)
> return ret;
>
> + if (!(ctrlr->flags & SOLVER_PRESENT))
> + return ret;
> +
> /* Do not allow sending active votes when in solver mode */
> spin_lock(&ctrlr->cache_lock);
> if (ctrlr->in_solver_mode)
> @@ -468,12 +471,24 @@ int rpmh_flush(struct rpmh_ctrlr *ctrlr)
> struct cache_req *p;
> int ret = 0;
>
> - lockdep_assert_irqs_disabled();
> + /*
> + * For RSC that don't have solver mode,
> + * rpmh_flush() is only called when we think we're running
> + * on the last CPU with irqs_disabled.
> + *
> + * For RSC that have solver mode,
> + * rpmh_flush() can be invoked with irqs enabled by any CPU.
> + *
> + * Conditionally check for irqs_disabled only when solver mode
> + * is not available.
> + */
> +
> + if (!(ctrlr->flags & SOLVER_PRESENT))
> + lockdep_assert_irqs_disabled();
Can we have a different function that is called for the case where
solver mode is present and where solver mode isn't present? It would be
good to clearly show that rpmh_flush() thinks it is being called from
the last CPU vs. from some other random place because the code is
assuming solver vs. non-solver enabled state. It would be clearer from
the call site too.
>
> /*
> - * Currently rpmh_flush() is only called when we think we're running
> - * on the last processor. If the lock is busy it means another
> - * processor is up and it's better to abort than spin.
> + * If the lock is busy it means another transaction is on going,
> + * in such case it's better to abort than spin.
> */
> if (!spin_trylock(&ctrlr->cache_lock))
> return -EBUSY;
