On Tue, Dec 06, 2022 at 08:23:13AM -0800, Ben Levinsky wrote:
> Good Morning Mathieu,
>
>
> I did some testing and replied inline.
>
>
> On 12/2/22 9:00 AM, Mathieu Poirier wrote:
> > On Wed, Nov 30, 2022 at 09:39:33PM +0000, Levinsky, Ben wrote:
> > > Hi Mathieu,
> > >
> > > Thank you for your review. Please see my reply inline.
> > >
> > > Thanks
> > > Ben
> > >
> > > On 11/25/22, 10:05 AM, "Mathieu Poirier"<mathieu.poirier@xxxxxxxxxx> wrote:
> > >
> > > CAUTION: This message has originated from an External Source. Please use proper judgment and caution when opening attachments, clicking links, or responding to this email.
> > >
> > >
> > > Hi Ben,
> > >
> > > On Tue, Nov 15, 2022 at 07:37:53AM -0800, Ben Levinsky wrote:
> > > > Allow users of remoteproc the ability to get a handle to an rproc by
> > > > passing in node that has parent rproc device and an ID that matches
> > > > an expected rproc struct's index field.
> > > >
> > > > This enables to get rproc structure for remoteproc drivers that manage
> > > > more than 1 remote processor (e.g. TI and Xilinx R5 drivers).
> > > >
> > > > Signed-off-by: Ben Levinsky<ben.levinsky@xxxxxxxxxx>
> > > > ---
> > > > drivers/remoteproc/remoteproc_core.c | 64 +++++++++++++++++++++++++++-
> > > > include/linux/remoteproc.h | 1 +
> > > > 2 files changed, 64 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/drivers/remoteproc/remoteproc_core.c b/drivers/remoteproc/remoteproc_core.c
> > > > index 775df165eb45..6f7058bcc80c 100644
> > > > --- a/drivers/remoteproc/remoteproc_core.c
> > > > +++ b/drivers/remoteproc/remoteproc_core.c
> > > > @@ -40,6 +40,7 @@
> > > > #include <linux/virtio_ring.h>
> > > > #include <asm/byteorder.h>
> > > > #include <linux/platform_device.h>
> > > > +#include <linux/of_platform.h>
> > > >
> > > > #include "remoteproc_internal.h"
> > > >
> > > > @@ -2203,13 +2204,74 @@ struct rproc *rproc_get_by_phandle(phandle phandle)
> > > >
> > > > return rproc;
> > > > }
> > > > +
> > > > +/**
> > > > + * rproc_get_by_id() - find a remote processor by ID
> > > > + * @phandle: phandle to the rproc
> > > > + * @id: Index into rproc list that uniquely identifies the rproc struct
> > > > + *
> > > > + * Finds an rproc handle using the remote processor's index, and then
> > > > + * return a handle to the rproc. Before returning, ensure that the
> > > > + * parent node's driver is still loaded.
> > > > + *
> > > > + * This function increments the remote processor's refcount, so always
> > > > + * use rproc_put() to decrement it back once rproc isn't needed anymore.
> > > > + *
> > > > + * Return: rproc handle on success, and NULL on failure
> > > > + */
> > > > +
> > > > +struct rproc *rproc_get_by_id(phandle phandle, unsigned int id)
> > > > +{
> > > > + struct rproc *rproc = NULL, *r;
> > > > + struct platform_device *parent_pdev;
> > > > + struct device_node *np;
> > > > +
> > > > + np = of_find_node_by_phandle(phandle);
> > > > + if (!np)
> > > > + return NULL;
> > > > +
> > > > + parent_pdev = of_find_device_by_node(np->parent);
> > > > + if (!parent_pdev) {
> > > > + dev_err(&parent_pdev->dev,
> > > > + "no platform device for node %pOF\n", np);
> > > > + of_node_put(np);
> > > > + return NULL;
> > > > + }
> > > > +
> > > > + /* prevent underlying implementation from being removed */
> > > > + if (!try_module_get(parent_pdev->dev.driver->owner)) {
> > > > + dev_err(&parent_pdev->dev, "can't get owner\n");
> > > > + of_node_put(np);
> > > > + return NULL;
> > > > + }
> > > > +
> > > > + rcu_read_lock();
> > > > + list_for_each_entry_rcu(r, &rproc_list, node) {
> > > > + if (r->index == id) {
> > > > + rproc = r;
> > > > + get_device(&rproc->dev);
> > > > + break;
> > > > + }
> > > > + }
> > >
> > > This won't work because several remote processors can be on the list. If
> > > another remote processor was discovered before the one @phandle is associated
> > > with, the remote processor pertaining to that previous one will returned.
> > >
> > > I didn't understand. From my point of view passing in the phandle of the child-platform device here will work because each child-platform will have its own entry in the remoteproc list.
> > You are correct, each child platform device will have its own entry in
> > @rproc_list. The problem is that r->index may not match @id that is passed as a
> > parameter.
> >
> > > Also " If another remote processor was discovered before the one" Here this prevented from what I can see because the remoteproc_list is protected by a mutex_lock. Seehttps://github.com/torvalds/linux/blob/master/drivers/remoteproc/remoteproc_core.c#L2288 for the mutex_lock.
> > >
> > > Additionally the calls to zynqmp_r5_add_rproc_core() are called sequentially so this also prevents the race condition.
> > >
> > > I think I am missing something in your paragraph above. Can you expand on this issue?
> > As explained above, the issue is not about race conditions but the value of
> > r->index and @id.
> >
> > > Do you mean to say that if we use the cluster platform device you think using one of the existing APIs will work? For example rproc_get_by_child() or rproc_get_by_phandle()
> > >
> > > Athttps://git.kernel.org/pub/scm/linux/kernel/git/remoteproc/linux.git/tree/drivers/remoteproc/xlnx_r5_remoteproc.c?h=rproc-next#n923 " zynqmp_r5_add_rproc_core(&child_pdev->dev);" Here if we use cluster->dev this will work? To dig deeper into this for both the Xilinx and TI R5 remoteproc drivers, I think this proposed solution will create an issue in that for Split modes, the existing getter APIs will not be able to return one of the corresponding rproc instances because both cores will refer to the same platform-device structure.
> > >
> > > I can bring up the above in the community call.
> > >
> > > There is also an issue with rproc_put().
> > >
> > > If passing the cluster platform device works for the above then rproc_put() should work correct? We can test this on our side as well. That being said I can bring this up in the community call
> > Yes, using the cluster platform device will work with rproc_put().
> >
> > >
> > > I think your description of the problem is mostly correct. The intermediate
> > > devices created by the cascading entries for individual remote processors in the
> > > device tree are causing an issue. The "compatible" string for each remote
> > > processor can't be handled by any platform drivers (as it should be), which
> > > makes try_module_get() fail because r->dev.parent->driver is not bound to
> > > anything.
> > >
> > > Looking at the code for Xilinx's R5F support that I just queued[1], the simplest
> > > solution may be to pass @dev, which is in fact @cluster->dev, to
> > > zynqmp_r5_add_rproc_core() rather than the device associated with the
> > > intermediate platform device.
> > >
> > > That _should_ work. It is hard for me to know for sure since I don't have a
> > > platform that has dual core remote processor to test with.
> > >
> > > Get back to me with how that turned out and we'll go from there.
> > >
> > > Thanks,
> > > Mathieu
> > >
> > >
> > >
> > >
> > > [1].https://git.kernel.org/pub/scm/linux/kernel/git/remoteproc/linux.git/tree/drivers/remoteproc/xlnx_r5_remoteproc.c?h=rproc-next#n923
>
> I have an update on this.
>
>
>
> I tested the following using the RPU-cluster platform device:
>
> test 1: RPU split with 2 core
>
> test 2: RPU split with 1 core
>
> test 3: lockstep RPU
>
>
> I tested with the zynqmp-r5-remoteproc platform probe using the (RPU)
> cluster platform device instead of the core/child platform device. When I
> used this I was unable to properly use the API rproc_get_by_phandle() and
> there was _only_ an issue for test 1. This was because each core will have
> its own call to rproc_alloc(), rproc_add() and each core's remoteproc
> structure has the same parent device.
You haven't specified if my proposal worked with test 2 and 3. I'm guessing
that it does.
>
> This results in the later call to rproc_get_by_phandle() not behaving
> properly because the function will return whichever core had its entries
> added to the list first.
>
That is a valid observation, but at least we are getting closer. The next step
is to find the right remote processor and I think we should look at np->name and
rproc->name. They should be quite close because rproc_alloc() is called with
dev_name(cdev).
I will look into this further tomorrow morning if I have time, but I encourage
you to do the same on your side.
>
> For reference I placed the logic for API rproc_get_by_phandle() that loops
> through device and the rproc_alloc() line where the dev parent is set:
>
>
> Here is the getter API where the loop checking the remoteproc dev parent is:
>
> https://github.com/torvalds/linux/blob/master/drivers/remoteproc/remoteproc_core.c#L2109
>
>
> if(r->dev.parent&& r->dev.parent->of_node== np) {
>
>
> Here is the rproc_alloc() call where they set remoteproc dev parent:
>
> https://github.com/torvalds/linux/blob/master/drivers/remoteproc/remoteproc_core.c#L2448
>
>
> rproc->dev.parent= dev;
>
> Thanks,
>
> Ben
>
