Hi Arnd,
On 21 June 2014 05:34, Arnd Bergmann arnd@arndb.de wrote:
linux/mailbox_client.h
18 * struct mbox_client - User of a mailbox 19 * @dev: The client device 20 * @chan_name: The "controller:channel" this client wants
Please correct me if I'm mistaken, but I think this comment in the header is confusing. It gives the impression that the user is expected to fill in this field as "controller name: channel id". But, looking at an example of a DT based mbox client [1] , that doesnt seem to be the case. And "chan_name" is compared with "mbox-names", which seems to contain a list of Channel names. The mailbox is then identified by a separate DT binding : "mbox", which has the mailbox name and the channel id. So shouldnt this comment not say anything about the "controller" and the DT binding should be changed to "channel-names", instead of "mbox-names" to keep things consistent?
Instead of dev, I added a name string to the mbox controller structure. So now the client gets its channel by requesting "controller:channel" where controller should match with mbox->name and channel becomes an index into mbox->chans[].
Right, I looked at the wrong version, sorry about that.
No problem. Many thanks for the review.
However, it seems you still make the same mistake here: The name that gets passed as chan_name in the mailbox API is a local identifier that is supposed to be interpreted for the client device and used to look up a pointer to the mailbox device and channel. If you require drivers to put global data (e.g. the mbox->name, or the channel number) in there, it's impossible to write a driver that works on both DT and ACPI. If you want to use the mbox_request_channel() interface from a driver, you need some form of lookup table in the ACPI data to do the conversion.
Fair point. The more I think about this, it seems that if we want to use the mailbox framework for ACPI kernels, we should have a PCC specific bypass, something like the one you suggested below. The ACPI spec defines PCC as the only "mailbox" like mechanism. There are 3 PCC clients defined as well; CPPC, MPST and RASF. Each of these have their own ACPI tables and so they dont require special DSDT entries. Moreover, these PCC client drivers will be very ACPI specific anyway. So, trying to emulate DT like mbox controller-client matching in ACPI at this point is rather pointless. It will require creating dummy DSDT entries for the PCC mailbox controller and PCC mailbox clients which have their own well defined ACPI tables (and so dont belong in the OS agnostic DSDT) and then coming up with customized Device Specific Methods (DSMs) for mbox clients to refer to mbox controllers.
The other alternative is to skip the mailbox framework altogether. One thing to note is that the PCC driver and its clients should work on X86, ARMv8 and any other platform that has ACPI support. Currently the Mailbox framework looks platform agnostic but is tied to DT, so it may not work well for everyone. But like I mentioned early on, the framework provides for async notification and queuing which is useful for PCC, so I'd prefer the PCC specific bypass option.
The alternative would be not to use mbox_request_channel() at all for now, but to add a new interface that can only be used PCC and that matches by ID but is independent of the use of ACPI or DT, something like:
struct mbox_chan *pcc_mbox_get_channel(struct mbox_client *cl, char *name, unsigned chan_id, struct mbox_chan **chan) { struct mbox_controller *mbox; mbox = mbox_find_pcc_controller(name, ...);
*chan = &mbox->chans[chan_id]; return init_channel(*chan, cl);}
This would mean that we'd have to special-case "pcc" users, which is not very nice, but at least it would work on both DT and ACPI, and a future ACPI version could still add support for the mailbox API later.
I'll play around with this idea a bit and see how it looks.
Cheers, Ashwin
[1] - https://github.com/hackerspace/rpi-linux/commit/cd0b9584cbedf46812cfd220ba47...