On Tue, Jan 17, 2023 at 08:51:51PM -0600, Rob Herring wrote: > On Thu, Jan 12, 2023 at 5:43 PM Simon Glass wrote: > > > > Hi Rob, > > > > On Wed, 11 Jan 2023 at 19:10, Rob Herring wrote: > > > > > > On Mon, Dec 19, 2022 at 1:21 PM Simon Glass wrote: > > > > > > > > Hi Abdellatif, > > > > > > > > On Mon, 19 Dec 2022 at 04:12, Abdellatif El Khlifi > > > > wrote: > > > > > > > > > > On Mon, Dec 05, 2022 at 09:49:30AM -0600, Rob Herring wrote: > > > > > > On Sun, Dec 4, 2022 at 1:22 PM Simon Glass wrote: > > > > > > > > > > > > > > Hi Rob, > > > > > > > > > > > > > > On Tue, 29 Nov 2022 at 05:22, Rob Herring wrote: > > > > > > > > > > > > > > > > On Fri, Nov 25, 2022 at 3:18 PM Simon Glass wrote: > > > > > > > > > > > > > > > > > > Hi Abdellatif, > > > > > > > > > > > > > > > > > > On Thu, 24 Nov 2022 at 06:21, Abdellatif El Khlifi wrote: > > > > > > > > > > > > > > > > > > > > On Tue, Nov 22, 2022 at 07:09:16PM -0700, Simon Glass wrote: > > > > > > > > > > > should be called 'priov' and should beHi Abdellatif, > > > > > > > > > > > > > > > > > > > > > > > > > > > > > [..] > > > > > > > > > > > > > > > > > > > > > +/** > > > > > > > > > > > > + * ffa_device_get - create, bind and probe the arm_ffa device > > > > > > > > > > > > + * @pdev: the address of a device pointer (to be filled when the arm_ffa bus device is created > > > > > > > > > > > > + * successfully) > > > > > > > > > > > > + * > > > > > > > > > > > > + * This function makes sure the arm_ffa device is > > > > > > > > > > > > + * created, bound to this driver, probed and ready to use. > > > > > > > > > > > > + * Arm FF-A transport is implemented through a single U-Boot > > > > > > > > > > > > + * device managing the FF-A bus (arm_ffa). > > > > > > > > > > > > + * > > > > > > > > > > > > + * Return: > > > > > > > > > > > > + * > > > > > > > > > > > > + * 0 on success. Otherwise, failure > > > > > > > > > > > > + */ > > > > > > > > > > > > +int ffa_device_get(struct udevice **pdev) > > > > > > > > > > > > +{ > > > > > > > > > > > > + int ret; > > > > > > > > > > > > + struct udevice *dev = NULL; > > > > > > > > > > > > + > > > > > > > > > > > > + ret = device_bind(dm_root(), DM_DRIVER_GET(arm_ffa), FFA_DRV_NAME, NULL, ofnode_null(), > > > > > > > > > > > > + &dev); > > > > > > > > > > > > > > > > > > > > > > Please add a DT binding. Even if only temporary, we need something for this. > > > > > > > > > > > > > > > > > > > > Thanks for the feedback. I'm happy to address all the comments. > > > > > > > > > > > > > > > > > > > > Regarding DT binding and FF-A discovery. We agreed with Linaro and Rob Herring > > > > > > > > > > about the following: > > > > > > > > > > > > > > > > > > > > - DT is only for what we failed to make discoverable. For hardware, we're stuck > > > > > > > > > > with it. We shouldn't repeat that for software interfaces. This approach is > > > > > > > > > > already applied in the FF-A kernel driver which comes with no DT support and > > > > > > > > > > discovers the bus with bus_register() API [1]. > > > > > > > > > > > > > > > > > > This may be the UEFI view, but it is not how U-Boot works. This is not something we are 'stuck' with. It is how we define what is present on a device. This is how the PCI bus works in U-Boot. It is best practice in U-Boot to use the device tree to make this things visible and configurable. Unlike with Linux there is no other way to provide configuration needed by these devices. > > > > > > > > > > > > > > > > Where do you get UEFI out of this? > > > > > > > > > > > > > > I assume it was UEFI as there was no discussion about this in U-Boot. > > > > > > > Which firmware project was consulted about this? > > > > > > > > > > > > > > > > > > > > > > > It is the discoverability of hardware that is fixed (and we are stuck > > > > > > > > with). We can't change hardware. The disoverability may be PCI > > > > > > > > VID/PID, USB device descriptors, or nothing. We only use DT when those > > > > > > > > are not sufficient. For a software interface, there is no reason to > > > > > > > > make them non-discoverable as the interface can be fixed (at least for > > > > > > > > new things like FF-A). > > > > > > > > > > > > > > Here I am talking about the controller itself, the top-level node in > > > > > > > the device tree. For PCI this is a device tree node and it should be > > > > > > > the same here. So I am not saying that the devices on the bus need to > > > > > > > be in the device tree (that can be optional, but may be useful in some > > > > > > > situations where it is status and known). > > > > > > > > > > > > Sure, the PCI host bridges are not discoverable, have a bunch of > > > > > > resources, and do need to be in DT. The downstream devices only do if > > > > > > they have extra resources such as when a device is soldered down on a > > > > > > board rather than a standard slot. > > > > > > > > > > > > > We need something like: > > > > > > > > > > > > > > ff-a { > > > > > > > compatible = "something"; > > > > > > > }; > > > > > > > > > > > > > > I don't know what mechanism is actually used to communicate with it, > > > > > > > but that will be enough to get the top-level driver started. > > > > > > > > > > > > There's discovery of FF-A itself and then discovery of FF-A features > > > > > > (e.g. partitions). Both of those are discoverable without DT. The > > > > > > first is done by checking the SMCCC version, then checking for FF-A > > > > > > presence and features. Putting this into DT is redundant. Worse, what > > > > > > if they disagree? > > > > > > > > > > Hi Simon, > > > > > > > > > > Do you agree with Rob, Ilias and myself that it makes more sense > > > > > FF-A bus is discovered without a DT node and following the same approach as > > > > > Linux ? (FF-A bus doesn't have a HW controller and is a purely SW bus, > > > > > no configuration/description needed at DT level). > > > > > > > > > > Your suggestions are always welcome. > > > > > > > > I'm sorry I don't agree with that. It does need a compatible string, > > > > like PCI has. You can just add it in U-Boot if Linux won't accept the > > > > binding. > > > > > > It's not like PCI as the host side of PCI has non-discoverable resources. > > > > OK I see. It is certainly an edge case. > > > > > > > > This all could have been designed better, but hindsight is 20/20 and > > > things evolved step by step. There are a bunch of firmware services > > > that are all behind SMCCC. The first (upstream) was PSCI. IIRC, SMCCC > > > was invented a bit after that, but generalized PSCI for other > > > services. Since then more have been added. More services get added one > > > by one and yes we added bindings for them. Because what's one more... > > > But that really needs to stop. We're stuck with h/w that's not > > > discoverable, there's zero reason to do that with s/w interfaces. If > > > we could redo everything, we'd have a node for SMCCC and that's it > > > unless there's h/w resources provided to the rest of DT. But we can't, > > > so SMCCC is discovered by the presence of PSCI. > > > > I understand the background here, but if we don't take a stand on > > this, this sort of thing will continue. > > I agree completely (but I think what you mean for 'this' is > different). Using DT for s/w features has to stop. > > Every new SMCCC firmware interface we get a new binding for it. FF-A > originally had a bunch of crap in the binding. So did Op-tee. Why does > each new firmware feature need a DT node? As software people, can't we > design software interfaces which are discoverable on their own rather > than using a hardware description for discovering them? The only part > we can't discover is whether we have SMCCC or not, but we already have > that in DT with the PSCI node because if we have PSCI, we have SMCCC. > > It is the same problem with that UEFI SPI protocol binding on the list > last week. I guess the problem comes down to, can we have one discovery method that everyone shares, or do we have to let everyone invent a new discovery method every time? FF-A, Op-tee, U-Boot, coreboot, barebox (and everyone else I'm unintentionally forgetting) could just discover these things via device tree. Or, we could all write our own code to perform the discovery. And when RISC-V comes along with similar functionality, we could probe their device tree and see they've implemented the same concept, but a little differently, but still have the discovery portion be in the device tree. To which it sounds like your answer is "not in the device tree". > > Just because something works > > in Linux does not mean that the binding (or lack of it) is good. > > > > The reasons to do this are: > > - avoids needing to manually call device_bind() > > - avoids extra plumbing in U-Boot > > - provides visibility into what is in the system, by looking at the > > DT, like documentation > > - DT is how devices are bound in U-Boot > > > > You can see the problem if you look at ffa_device_get(). It is called > > from ffa_bus_discover() which is a new addition into the board_init > > list. We are trying to remove this list and certainly don't want new > > things added!! > > That seems less than ideal. How do you init PSCI? Or u-boot doesn't > touch it? The low-level code for both should be shared if not already. > > > We don't need to change this in the Linux implementation, just add a > > top-level DT node for U-Boot. I don't understand why that is such a > > big problem? > > It's a line in the sand. Yeah, 1 compatible is not a big deal, but > then it is an invitation to add to the binding and add a compatible > for the next SMCCC firmware feature. Well, I suppose then that this is not the line in the sand we have to have a larger debate about. For this, we'll just follow along with what everyone else has agreed on re bindings, and that just leaves Simon's other feedback about sandbox testing, etc. -- Tom