On Tue, Sep 3, 2024 at 4:36 PM Kuninori Morimoto kuninori.morimoto.gx@renesas.com wrote:
Hi Saravana, Rob Cc Mark/ALSA-ML, Geert, Laurent
I got dead-lock issue on drivers/of/property.c to call .probe() I'm now using v6.11-rc5.
I'm using 3 devices, and these are connected via OF-Graph. These are connected like this so far. Let's call it as [Style-A]
[Style-A] +---+ +---+ |(A)+--+ | +-+-+ | | | |(B)| +-+-+ | | |(C)+--+ | +---+ +---+
These are using port/endpoint to connect. It works well for now, no issue I have so far.
(B) is connector for (A)-(C) data, thus, (A) and (C) should be probed first, and (B) needs probe later (For ALSA member, this (B) is Audio Graph Card2).
Here, (A)-(C) can connect directly for some data. Now it is using both data connection (= (A)-(B)-(C) and (A)-(C))
Now, I want to disconnect (A)-(C) connection, like below Let's call it as [Style-B]
[Style-B] +---+ +---+ |(A)+--+ | +---+ | | |(B)| +---+ | | |(C)+--+ | +---+ +---+
Then, it seems dead-lock happen.
In my debug, it seems... - (B) is handled as supplier for (A). - (B) probe() is called, but it needs (A) info which is not yet probed. So it returns -EPROBE_DEFER. - Because (B) is not probed, (A) probe() never called
In [Style-A], it seems __fwnode_link_cycle() is called for (A)-(B) connection, so the dead-lock will be solved. But it is not called for [Style-B]. Because of that (B) is always handled as supplier for (A).
If I used below patch, and use "non-supplier" property on (B), this dead-lock issue was solved. But I know this is not a good solution.
I think it is very normal connection, not super special. How can I solve this issue on correct way ? Or how can I indicate you my issue more detail ? I can add debug patch and test it if you can indicate it to me.
Kinda swamped, so skimmed your diagram. Looks like you are making up a "non-supplier" property?
Anyway, I solved this recently in a general way. Use "post-init-providers" property in the node of "A" and point it to "B". This tells fw_devlink that B is not needed for A to probe.
So, NACK to this patch, but hope my response helps.
-Saravana
diff --git a/drivers/of/property.c b/drivers/of/property.c index def152c61049..2f08210c2ea4 100644 --- a/drivers/of/property.c +++ b/drivers/of/property.c @@ -1502,11 +1502,22 @@ static struct device_node *parse_remote_endpoint(struct device_node *np, const char *prop_name, int index) {
struct device_node *node;/* Return NULL for index > 0 to signify end of remote-endpoints. */ if (index > 0 || strcmp(prop_name, "remote-endpoint")) return NULL;
return of_graph_get_remote_port_parent(np);
node = of_graph_get_remote_port_parent(np);/** There is clearly non-supplier node which is connected via "remote-endpoint".* Ignore it, otherwise dead-lock might occur*/if (of_property_present(node, "non-supplier"))return NULL;return node;}
Thank you for your help !!
Best regards
Kuninori Morimoto