Date:   Wed, 22 Apr 2020 17:53:56 -0700
From:   Bjorn Andersson <bjorn.andersson@...aro.org>
To:     Siddharth Gupta <sidgup@...eaurora.org>
Cc:     agross@...nel.org, ohad@...ery.com,
        linux-remoteproc@...r.kernel.org, linux-kernel@...r.kernel.org,
        linux-arm-msm@...r.kernel.org,
        linux-arm-kernel@...ts.infradead.org, tsoni@...eaurora.org,
        psodagud@...eaurora.org, rishabhb@...eaurora.org
Subject: Re: [PATCH v2 5/6] remoteproc: qcom: Add per subsystem SSR
 notification

On Wed 08 Apr 16:36 PDT 2020, Siddharth Gupta wrote:

> Currently there is a global notification chain which is called whenever any
> remoteproc shuts down. This leads to all the listeners being notified, and
> is not an optimal design as kernel drivers might only be interested in
> listening to notifications from a particular remoteproc. Create an
> individual notifier chain for every SSR subdevice, and modify the
> notification registration API to include the remoteproc struct as an
> argument. Update the existing user of the registration API to get the
> phandle of the remoteproc dt node to register for SSR notifications.
> 
> Signed-off-by: Rishabh Bhatnagar <rishabhb@...eaurora.org>
> Signed-off-by: Siddharth Gupta <sidgup@...eaurora.org>
> ---
>  drivers/remoteproc/qcom_common.c      | 49 +++++++++++++++++++++++++++--------
>  drivers/remoteproc/qcom_common.h      |  1 +
>  drivers/soc/qcom/glink_ssr.c          | 20 ++++++++++++--
>  include/linux/remoteproc/qcom_rproc.h | 17 ++++++++----
>  4 files changed, 69 insertions(+), 18 deletions(-)
> 
> diff --git a/drivers/remoteproc/qcom_common.c b/drivers/remoteproc/qcom_common.c
> index 1d2351b..56b0c3e 100644
> --- a/drivers/remoteproc/qcom_common.c
> +++ b/drivers/remoteproc/qcom_common.c
> @@ -23,8 +23,6 @@
>  #define to_smd_subdev(d) container_of(d, struct qcom_rproc_subdev, subdev)
>  #define to_ssr_subdev(d) container_of(d, struct qcom_rproc_ssr, subdev)
>  
> -static BLOCKING_NOTIFIER_HEAD(ssr_notifiers);
> -
>  static int glink_subdev_start(struct rproc_subdev *subdev)
>  {
>  	struct qcom_rproc_glink *glink = to_glink_subdev(subdev);
> @@ -180,27 +178,52 @@ EXPORT_SYMBOL_GPL(qcom_remove_smd_subdev);
>  
>  /**
>   * qcom_register_ssr_notifier() - register SSR notification handler
> + * @rproc:	pointer to the remoteproc structure
>   * @nb:		notifier_block to notify for restart notifications
>   *
> - * Returns 0 on success, negative errno on failure.
> + * Returns pointer to srcu notifier head on success, ERR_PTR on failure.
>   *
> - * This register the @notify function as handler for restart notifications. As
> - * remote processors are stopped this function will be called, with the SSR
> - * name passed as a parameter.
> + * This registers the @notify function as handler for restart notifications. As
> + * remote processors are stopped this function will be called, with the rproc
> + * pointer passed as a parameter.
>   */
> -int qcom_register_ssr_notifier(struct notifier_block *nb)
> +void *qcom_register_ssr_notifier(struct rproc *rproc, struct notifier_block *nb)
>  {
> -	return blocking_notifier_chain_register(&ssr_notifiers, nb);
> +	struct rproc_subdev *subdev;
> +	struct qcom_rproc_ssr *ssr;
> +	int ret;
> +
> +	if (!rproc)
> +		return ERR_PTR(-EINVAL);
> +
> +	mutex_lock(&rproc->lock);
> +	list_for_each_entry(subdev, &rproc->subdevs, node) {

I would prefer that we don't touch the lock or subdevs list outside of
the remoteproc core.

> +		ret = strcmp(subdev->name, "ssr_notifs");
> +		if (!ret)
> +			break;
> +	}
> +	mutex_unlock(&rproc->lock);
> +	if (ret)
> +		return ERR_PTR(-ENOENT);
> +
> +	ssr = to_ssr_subdev(subdev);
> +	srcu_notifier_chain_register(ssr->rproc_notif_list, nb);

Adding the notifier to an existing ssr_subdev means that any client
driver that is interested in notification about a remoteproc coming and
going will need to be registered (typically probed) after the remoteproc
driver.

I presume this would be handled by probe deferring on
rproc_get_by_phandle(), but I'm concerned that this will cause
unnecessary probe deferral. But more importantly, it wouldn't allow for
the remoteproc driver to be unloaded and loaded again (as that would be
a new notifier list).

So I think you should carry a global list of "watchers" and upon subdev
events you can match entries in this list based on either struct
of_node or perhaps by ssr_name?

> +
> +	return ssr->rproc_notif_list;
>  }
>  EXPORT_SYMBOL_GPL(qcom_register_ssr_notifier);
>  
>  /**
>   * qcom_unregister_ssr_notifier() - unregister SSR notification handler
> + * @notify:	pointer to srcu notifier head
>   * @nb:		notifier_block to unregister
>   */
> -void qcom_unregister_ssr_notifier(struct notifier_block *nb)
> +int qcom_unregister_ssr_notifier(void *notify, struct notifier_block *nb)
>  {
> -	blocking_notifier_chain_unregister(&ssr_notifiers, nb);
> +	if (!notify)
> +		return -EINVAL;
> +
> +	return srcu_notifier_chain_unregister(notify, nb);
>  }
>  EXPORT_SYMBOL_GPL(qcom_unregister_ssr_notifier);
>  
> @@ -208,7 +231,7 @@ static void ssr_notify_unprepare(struct rproc_subdev *subdev)
>  {
>  	struct qcom_rproc_ssr *ssr = to_ssr_subdev(subdev);
>  
> -	blocking_notifier_call_chain(&ssr_notifiers, 0, (void *)ssr->name);
> +	srcu_notifier_call_chain(ssr->rproc_notif_list, 0, (void *)ssr->name);
>  }
>  
>  /**
> @@ -226,6 +249,9 @@ void qcom_add_ssr_subdev(struct rproc *rproc, struct qcom_rproc_ssr *ssr,
>  	ssr->name = ssr_name;
>  	ssr->subdev.name = kstrdup("ssr_notifs", GFP_KERNEL);
>  	ssr->subdev.unprepare = ssr_notify_unprepare;
> +	ssr->rproc_notif_list = kzalloc(sizeof(struct srcu_notifier_head),
> +								GFP_KERNEL);
> +	srcu_init_notifier_head(ssr->rproc_notif_list);
>  
>  	rproc_add_subdev(rproc, &ssr->subdev);
>  }
> @@ -239,6 +265,7 @@ EXPORT_SYMBOL_GPL(qcom_add_ssr_subdev);
>  void qcom_remove_ssr_subdev(struct rproc *rproc, struct qcom_rproc_ssr *ssr)
>  {
>  	kfree(ssr->subdev.name);
> +	kfree(ssr->rproc_notif_list);
>  	rproc_remove_subdev(rproc, &ssr->subdev);
>  }
>  EXPORT_SYMBOL_GPL(qcom_remove_ssr_subdev);
> diff --git a/drivers/remoteproc/qcom_common.h b/drivers/remoteproc/qcom_common.h
> index 58de71e..7792691 100644
> --- a/drivers/remoteproc/qcom_common.h
> +++ b/drivers/remoteproc/qcom_common.h
> @@ -27,6 +27,7 @@ struct qcom_rproc_subdev {
>  struct qcom_rproc_ssr {
>  	struct rproc_subdev subdev;
>  
> +	struct srcu_notifier_head *rproc_notif_list;
>  	const char *name;
>  };
>  
> diff --git a/drivers/soc/qcom/glink_ssr.c b/drivers/soc/qcom/glink_ssr.c
> index d7babe3..2b39683 100644
> --- a/drivers/soc/qcom/glink_ssr.c
> +++ b/drivers/soc/qcom/glink_ssr.c
> @@ -7,6 +7,7 @@
>  #include <linux/completion.h>
>  #include <linux/module.h>
>  #include <linux/notifier.h>
> +#include <linux/remoteproc.h>
>  #include <linux/rpmsg.h>
>  #include <linux/remoteproc/qcom_rproc.h>
>  
> @@ -49,6 +50,7 @@ struct glink_ssr {
>  	struct rpmsg_endpoint *ept;
>  
>  	struct notifier_block nb;
> +	void *notifier_head;
>  
>  	u32 seq_num;
>  	struct completion completion;
> @@ -112,6 +114,7 @@ static int qcom_glink_ssr_notify(struct notifier_block *nb, unsigned long event,
>  static int qcom_glink_ssr_probe(struct rpmsg_device *rpdev)
>  {
>  	struct glink_ssr *ssr;
> +	struct rproc *rproc;
>  
>  	ssr = devm_kzalloc(&rpdev->dev, sizeof(*ssr), GFP_KERNEL);
>  	if (!ssr)
> @@ -125,14 +128,27 @@ static int qcom_glink_ssr_probe(struct rpmsg_device *rpdev)
>  
>  	dev_set_drvdata(&rpdev->dev, ssr);
>  
> -	return qcom_register_ssr_notifier(&ssr->nb);
> +	rproc = rproc_get_by_child(&rpdev->dev);

As we discussed in our meeting offline earlier today, not all glink_ssr
instances has a remoteproc ancestor. After going back and forth on how
to handle this I posted below series:

https://lore.kernel.org/linux-arm-msm/20200423003736.2027371-1-bjorn.andersson@linaro.org/T/#t

With this we are flexible to tie the ssr_subdev API to remoteproc
instances...

Regards,
Bjorn

Powered by blists - more mailing lists