Hi Laurent, On Wed, Jun 24, 2020 at 03:12:53AM +0300, Laurent Pinchart wrote: > Hi Sakari, > > On Wed, Jun 24, 2020 at 02:12:05AM +0300, Sakari Ailus wrote: > > On Wed, Jun 24, 2020 at 12:22:41AM +0300, Laurent Pinchart wrote: > > > On Wed, Jun 24, 2020 at 12:13:06AM +0300, Sakari Ailus wrote: > > >> On Sun, Jun 21, 2020 at 03:00:26AM +0300, Laurent Pinchart wrote: > > >>> fwnode matching was designed to match on nodes corresponding to a > > >>> device. Some drivers, however, needed to match on endpoints, and have > > >>> passed endpoint fwnodes to v4l2-async. This works when both the subdev > > >>> and the notifier use the same fwnode types (endpoint or device), but > > >>> makes drivers that use different types incompatible. > > >>> > > >>> Fix this by extending the fwnode match to handle fwnodes of different > > >>> types. When the types (deduced from the presence of remote endpoints) > > >>> are different, retrieve the device fwnode for the side that provides an > > >>> endpoint fwnode, and compare it with the device fwnode provided by the > > >>> other side. This allows interoperability between all drivers, regardless > > >>> of which type of fwnode they use for matching. > > >>> > > >>> Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@xxxxxxxxxxxxxxxx> > > >>> Tested-by: Lad Prabhakar <prabhakar.mahadev-lad.rj@xxxxxxxxxxxxxx> > > >>> Reviewed-by: Kieran Bingham <kieran.bingham+renesas@xxxxxxxxxxxxxxxx> > > >>> Reviewed-by: Jacopo Mondi <jacopo+renesas@xxxxxxxxxx> > > >>> --- > > >>> Changes since v2: > > >>> > > >>> - Add comment to explain that we're matching connecting endpoints > > >>> - Don't check fwnode name to detect endpoint > > >>> --- > > >>> drivers/media/v4l2-core/v4l2-async.c | 45 +++++++++++++++++++++++++++- > > >>> 1 file changed, 44 insertions(+), 1 deletion(-) > > >>> > > >>> diff --git a/drivers/media/v4l2-core/v4l2-async.c b/drivers/media/v4l2-core/v4l2-async.c > > >>> index 8bde33c21ce4..f82e0a32647d 100644 > > >>> --- a/drivers/media/v4l2-core/v4l2-async.c > > >>> +++ b/drivers/media/v4l2-core/v4l2-async.c > > >>> @@ -71,7 +71,50 @@ static bool match_devname(struct v4l2_subdev *sd, > > >>> > > >>> static bool match_fwnode(struct v4l2_subdev *sd, struct v4l2_async_subdev *asd) > > >>> { > > >>> - return sd->fwnode == asd->match.fwnode; > > >>> + struct fwnode_handle *other_fwnode; > > >>> + struct fwnode_handle *dev_fwnode; > > >>> + bool asd_fwnode_is_ep; > > >>> + bool sd_fwnode_is_ep; > > >>> + > > >>> + /* > > >>> + * Both the subdev and the async subdev can provide either an endpoint > > >>> + * fwnode or a device fwnode. Start with the simple case of direct > > >>> + * fwnode matching. > > >>> + */ > > >>> + if (sd->fwnode == asd->match.fwnode) > > >>> + return true; > > >>> + > > >>> + /* > > >>> + * Otherwise, check if the sd fwnode and the asd fwnode refer to an > > >>> + * endpoint or a device. If they're of the same type, there's no match. > > >>> + * Technically speaking this checks if the nodes refer to a connected > > >>> + * endpoint, which is the simplest check that works for both OF and > > >>> + * ACPI. This won't make a difference, as drivers should not try to > > >>> + * match unconnected endpoints. > > >>> + */ > > >>> + sd_fwnode_is_ep = fwnode_property_present(sd->fwnode, > > >>> + "remote-endpoint"); > > >>> + asd_fwnode_is_ep = fwnode_property_present(asd->match.fwnode, > > >>> + "remote-endpoint"); > > >> > > >> Please don't try parsing graph bindings outside the main parsers. > > > > > > Why is that ? On the DT side, bindings are considered to be stable, so > > > isolating their parsing in helpers would not help with ABI compatibility > > > anyway. Maybe it would be useful if you could explain how graph parsing > > > works in ACPI ? The fact that there's a remote-endpoint property without > > > endpoints is a the minimum quite puzzling. > > > > No other drivers (or frameworks to my knowledge) work with the graphs > > directly anymore. There was a staging driver (IMX) that did but that has > > been fixed now. It's easier to ensure the code is correct --- this is > > because the data structure is hard to parse, especially while taking > > firmware type differences into account but the functions that access it are > > relatively simple to use. > > > > The fwnode property API has operations callbacks that are specific to the > > type of the node. Most access functions have a firmware specific backend. > > > > With the presence of the "remote-endpoint" property there's no variation > > across the firmware types, at least not right now. But still putting it > > here right now looks like technical debt to me: much of the code parsing > > graph data structure outside the main parser has been buggy in the past. > > For my information, could you still briefly explain how remote-endpoint > works on ACPI, without any fwnode named "endpoint" ? There have been a different versions of the ACPI graph definitions, and firmware using both exists. See e.g. is_acpi_graph_node() and functions below that in drivers/acpi/property.c . > > > >> There's no API function to do just this, but you could go and check for the > > >> port parent right away. The code might be even more simple that way. > > > > > > How will that help ? With OF at least, fwnode_graph_get_port_parent() > > > will return the grand-parent if the passed node isn't an endpoint, not > > > much can be deduced from that. > > > > I meant to say fwnode_graph_get_remote_endpoint(). You'd need to release > > the fwnode reference, too. > > That makes more sense :-) > > > >> Alternatively, I guess we could add fwnode_graph_is_endpoint() or something > > >> but I wonder if it'd be worth it just for this. > > Would > > static inline bool fwnode_graph_is_endpoint(struct fwnode_handle *fwnode) > { > return fwnode_property_present(fwnode, "remote-endpoint"); > } > > in include/linux/property.h be acceptable to you ? I think that'd be fine. If there's a need to change the implementation in the future, it'll be easy. -- Sakari Ailus