On Thu 18 Jun 16:00 PDT 2020, Alex Elder wrote:
> On 5/27/20 10:34 PM, Rishabh Bhatnagar wrote:
> > Currently there is a single 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 a global
> > list of remoteproc notification info data structures. This will hold the
> > name and notifier_list information for a particular remoteproc. The API
> > to register for notifications will use name argument to retrieve the
> > notification info data structure and the notifier block will be added to
> > that data structure's notification chain. Also move from blocking notifier
> > to srcu notifer based implementation to support dynamic notifier head
> > creation.
>
> I'm looking at these patches now, without having reviewed the
> previous versions. Forgive me if I contradict or duplicate
> previous feedback.
>
> I have a number of suggestions, below.
>
> -Alex
>
Thanks for your review Alex, some feedback on the patch and your
response below.
> > Signed-off-by: Siddharth Gupta <sidgup@xxxxxxxxxxxxxx>
> > Signed-off-by: Rishabh Bhatnagar <rishabhb@xxxxxxxxxxxxxx>
> > ---
> > drivers/remoteproc/qcom_common.c | 84 ++++++++++++++++++++++++++++++-----
> > drivers/remoteproc/qcom_common.h | 5 ++-
> > include/linux/remoteproc/qcom_rproc.h | 20 ++++++---
> > 3 files changed, 90 insertions(+), 19 deletions(-)
> >
> > diff --git a/drivers/remoteproc/qcom_common.c b/drivers/remoteproc/qcom_common.c
> > index 9028cea..61ff2dd 100644
> > --- a/drivers/remoteproc/qcom_common.c
> > +++ b/drivers/remoteproc/qcom_common.c
> > @@ -12,6 +12,7 @@
> > #include <linux/module.h>
> > #include <linux/notifier.h>
> > #include <linux/remoteproc.h>
> > +#include <linux/remoteproc/qcom_rproc.h>
> > #include <linux/rpmsg/qcom_glink.h>
> > #include <linux/rpmsg/qcom_smd.h>
> > #include <linux/soc/qcom/mdt_loader.h>
> > @@ -23,7 +24,14 @@
> > #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);
> > +struct qcom_ssr_subsystem {
> > + const char *name;
> > + struct srcu_notifier_head notifier_list;
> > + struct list_head list;
> > +};
> > +
> > +static LIST_HEAD(qcom_ssr_subsystem_list);
> > +DEFINE_MUTEX(qcom_ssr_subsys_lock);
>
> There is no need for this mutex to be global.
>
> > static int glink_subdev_start(struct rproc_subdev *subdev)
> > {
> > @@ -189,39 +197,79 @@ void qcom_remove_smd_subdev(struct rproc *rproc, struct qcom_rproc_subdev *smd)
> > }
> > EXPORT_SYMBOL_GPL(qcom_remove_smd_subdev);
> > +struct qcom_ssr_subsystem *qcom_ssr_get_subsys(const char *name)
>
> This function should be made private (static).
>
Yes.
> I think the mutex should be taken in this function rather than
> the caller (more on this below). But if you leave it this way,
> please mention something in a comment that indicates the caller
> must hold the qcom_ssr_subsys_lock mutex.
>
I agree, that would simplify reasoning about the lock.
> > +{
> > + struct qcom_ssr_subsystem *info;
> > +
> > + /* Match in the global qcom_ssr_subsystem_list with name */
> > + list_for_each_entry(info, &qcom_ssr_subsystem_list, list) {
> > + if (!strcmp(info->name, name))
> > + return info;
>
> This probably isn't strictly necessary, because you are
> returning a void pointer, but you could do this here:
> return ERR_CAST(info);
Info is a struct qcom_ssr_subsystem * and that's the function's return
type as well, so Rishabh's approach looks correct to me.
>
> > + }
>
> This is purely a style thing, but the curly braces around
> the loop body aren't necessary.
>
> > + info = kzalloc(sizeof(*info), GFP_KERNEL);
> > + if (!info)
> > + return ERR_PTR(-ENOMEM);
> > + info->name = kstrdup_const(name, GFP_KERNEL);
> > + srcu_init_notifier_head(&info->notifier_list);
> > +
> > + /* Add to global notif list */
>
> s/notif/notification/
>
> > + INIT_LIST_HEAD(&info->list);
>
> No need to initialize the list element when adding it
> to a list. Both uts fields will be overwritten anyway.
>
> > + list_add_tail(&info->list, &qcom_ssr_subsystem_list);
> > +
> > + return info;
> > +}
> > +
> > /**
> > * qcom_register_ssr_notifier() - register SSR notification handler
> > + * @name: name that will be searched in global ssr subsystem list
>
> Maybe just "SSR subsystem name".
>
> > * @nb: notifier_block to notify for restart notifications
>
> Drop or modify "restart" in the comment line above.
>
> > *
> > - * Returns 0 on success, negative errno on failure.
> > + * Returns a subsystem cookie on success, ERR_PTR on failure.
>
> Maybe make the above a @Return: comment.
>
No @ in that, but "Return: foo" is the appropriate format.
> > *
> > - * 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 @nb notifier block as part the notifier chain for a
> > + * remoteproc associated with @name. The notifier block's callback
> > + * will be invoked when the particular remote processor is stopped.
>
> It's not just for stopping, right? Maybe something
> more like:
> Register to receive notification callbacks when
> remoteproc SSR events occur (pre- and post-startup
> and pre- and post-shutdown).
>
And this description of the function should go above the Return:
See https://www.kernel.org/doc/html/latest/doc-guide/kernel-doc.html#function-documentation
> > */
> > -int qcom_register_ssr_notifier(struct notifier_block *nb)
> > +void *qcom_register_ssr_notifier(const char *name, struct notifier_block *nb)
> > {
> > - return blocking_notifier_chain_register(&ssr_notifiers, nb);
> > + struct qcom_ssr_subsystem *info;
> > +
> > + mutex_lock(&qcom_ssr_subsys_lock);
>
> Can you explain why the mutex is taken here (and in
> qcom_add_ssr_subdev()), rather than having the mutex
> logic be isolated in qcom_ssr_get_subsys()?
>
> > + info = qcom_ssr_get_subsys(name);
> > + if (IS_ERR(info)) {
> > + mutex_unlock(&qcom_ssr_subsys_lock);
> > + return info;
> > + }
>
> I don't think there's any need to have the next function
> call be protected by the mutex, but maybe I'm mistaken.
>
> > + srcu_notifier_chain_register(&info->notifier_list, nb);
> > + mutex_unlock(&qcom_ssr_subsys_lock);
> > + return &info->notifier_list;
> > }
> > EXPORT_SYMBOL_GPL(qcom_register_ssr_notifier);
> > /**
> > * qcom_unregister_ssr_notifier() - unregister SSR notification handler
> > + * @notify: subsystem coookie returned from qcom_register_ssr_notifier
> > * @nb: notifier_block to unregister
>
> Add a @Return comment (0 on success, %ENOENT otherwise).
>
> > */
> > -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);
> > + return srcu_notifier_chain_unregister(notify, nb);
> > }
> > EXPORT_SYMBOL_GPL(qcom_unregister_ssr_notifier);
> > static void ssr_notify_unprepare(struct rproc_subdev *subdev)
> > {
> > struct qcom_rproc_ssr *ssr = to_ssr_subdev(subdev);
> > + struct qcom_ssr_notif_data data = {
> > + .name = ssr->info->name,
> > + .crashed = false,
> > + };
> > - blocking_notifier_call_chain(&ssr_notifiers, 0, (void *)ssr->name);
> > + srcu_notifier_call_chain(&ssr->info->notifier_list, 0, &data);
> > }
> > +
> > /**
> > * qcom_add_ssr_subdev() - register subdevice as restart notification source
> > * @rproc: rproc handle
> > @@ -229,12 +277,23 @@ static void ssr_notify_unprepare(struct rproc_subdev *subdev)
> > * @ssr_name: identifier to use for notifications originating from @rproc
> > *
> > * As the @ssr is registered with the @rproc SSR events will be sent to all
> > - * registered listeners in the system as the remoteproc is shut down.
> > + * registered listeners for the particular remoteproc when it is shutdown.
> > */
> > void qcom_add_ssr_subdev(struct rproc *rproc, struct qcom_rproc_ssr *ssr,
> > const char *ssr_name)
> > {
> > - ssr->name = ssr_name;
> > + struct qcom_ssr_subsystem *info;
> > +
> > + mutex_lock(&qcom_ssr_subsys_lock);
> > + info = qcom_ssr_get_subsys(ssr_name);
>
> If there already exists an SSR subsystem having the given
> name, its info structure is returned here. Is that OK?
> In practice, I don't expect this to be a problem, but it
> would be better to return an error if
>
You're right...that shouldn't happen. So a WARN_ON() and early return
would be in order.
> > + if (IS_ERR(info)) {
> > + dev_err(&rproc->dev, "Failed to add ssr subdevice\n");
> > + mutex_unlock(&qcom_ssr_subsys_lock);
> > + return;
> > + }
> > +
> > + mutex_unlock(&qcom_ssr_subsys_lock);
> > + ssr->info = info;
> > ssr->subdev.unprepare = ssr_notify_unprepare;
> > rproc_add_subdev(rproc, &ssr->subdev);
> > @@ -249,6 +308,7 @@ void qcom_add_ssr_subdev(struct rproc *rproc, struct qcom_rproc_ssr *ssr,
> > void qcom_remove_ssr_subdev(struct rproc *rproc, struct qcom_rproc_ssr *ssr)
> > {
> > rproc_remove_subdev(rproc, &ssr->subdev);
> > + ssr->info = NULL;
> > }
> > EXPORT_SYMBOL_GPL(qcom_remove_ssr_subdev);
> > diff --git a/drivers/remoteproc/qcom_common.h b/drivers/remoteproc/qcom_common.h
> > index 34e5188..dfc641c 100644
> > --- a/drivers/remoteproc/qcom_common.h
> > +++ b/drivers/remoteproc/qcom_common.h
> > @@ -26,10 +26,11 @@ struct qcom_rproc_subdev {
> > struct qcom_smd_edge *edge;
> > };
> > +struct qcom_ssr_subsystem;
> > +
> > struct qcom_rproc_ssr {
> > struct rproc_subdev subdev;
> > -
> > - const char *name;
> > + struct qcom_ssr_subsystem *info;
> > };
> > void qcom_add_glink_subdev(struct rproc *rproc, struct qcom_rproc_glink *glink,
> > diff --git a/include/linux/remoteproc/qcom_rproc.h b/include/linux/remoteproc/qcom_rproc.h
> > index fa8e386..58422b1 100644
> > --- a/include/linux/remoteproc/qcom_rproc.h
> > +++ b/include/linux/remoteproc/qcom_rproc.h
> > @@ -5,17 +5,27 @@
> > #if IS_ENABLED(CONFIG_QCOM_RPROC_COMMON)
> > -int qcom_register_ssr_notifier(struct notifier_block *nb);
> > -void qcom_unregister_ssr_notifier(struct notifier_block *nb);
> > +struct qcom_ssr_notif_data {
> > + const char *name;
> > + bool crashed;
>
> Is the crashed field strictly necessary? Could we instead have
> a QCOM_SSR_CRASHED event (in place of QCOM_SSR_BEFORE_SHUTDOWN)?
> I don't know, it's possible doing it the way you do ultimately
> simplifies the logic... So I'm asking, but not suggesting.
>
I looked at something similar for the subdev callbacks, but concluded
that most cases I could find was cleaner if I just passed a bool crashed
than having two separate functions to deal with in the client drivers.
So I'm okay with this.
Regards,
Bjorn
> > +};
> > +
> > +void *qcom_register_ssr_notifier(const char *name, struct notifier_block *nb);
> > +int qcom_unregister_ssr_notifier(void *notify, struct notifier_block *nb);
> > #else
> > -static inline int qcom_register_ssr_notifier(struct notifier_block *nb)
> > +static inline void *qcom_register_ssr_notifier(const char *name,
> > + struct notifier_block *nb)
> > {
> > - return 0;
> > + return NULL;
> > }
> > -static inline void qcom_unregister_ssr_notifier(struct notifier_block *nb) {}
> > +static inline int qcom_unregister_ssr_notifier(void *notify,
> > + struct notifier_block *nb)
> > +{
> > + return 0;
> > +}
> > #endif
> >
>
