On Thu, 2012-03-08 at 14:52 +0530, Archit Taneja wrote: > On Thursday 08 March 2012 02:16 PM, Tomi Valkeinen wrote: > > On Thu, 2012-03-08 at 14:04 +0530, Archit Taneja wrote: > >> On Wednesday 07 March 2012 06:14 PM, Tomi Valkeinen wrote: > > > >>> - r = hdmi_init_platform_driver(); > >>> - if (r) { > >>> - DSSERR("Failed to initialize hdmi\n"); > >>> - goto err_hdmi; > >>> + /* > >>> + * It's ok if the output-driver register fails. It happens, for example, > >>> + * when there is no output-device (e.g. SDI for OMAP4). > >>> + */ > >> > >> Suppose we do a omap2plus_defconfig, CONFIG_OMAP2_DSS_SDI would be > >> selected, and sdi.c would be built, if we boot on OMAP4, why would a sdi > >> driver register cause a failure? Wouldn't the sdi driver just get > >> registered, and wait till eternity for the corresponding sdi platform > >> device to get registered? > > > > No. Well, yes. > > > > Currently we use platform_driver_register() to register the drivers, and > > it does just what you described. But a few patches later I change > > platform_driver_register() to platform_driver_probe(), which will return > > ENODEV if there are no matching devices for the driver. > > > > I originally had the platform_driver_probe() patch before this patch, > > and thus the comment above made sense. Now the patch is after this > > patch, so the comment is not exactly right until the probe patch is also > > applied. > > Oh okay. But the comment after the patch set still says "It's ok if the > output-driver register fails.", we could change it to "It's ok if the > output-driver probe fails." Well, I guess this goes into nitpicking area, but if there are no devices, probe is not called at all. So I think it's the driver register that fails in that case. If there is a device, and it is probed, and that fails, then it's probe which fails. > > The point with platform_driver_probe() is that it can be used with > > non-removable devices which are created at boot time, like the DSS > > components. With platform_driver_probe() the probe function is called > > only at that one time, and never afterwards. So probe can be in __init > > section, and thrown away after init. > > So platform_driver_probe() is like a driver_register() + probe(). Yes. Well, when platform_driver_register() is called, and the devices are already present, it will call the probe also. So in that sense they are similar. The difference is that for platform_driver_register() the probe pointer must be in the driver struct, and it stays there even after init. For platform_driver_probe(), the probe pointer is given as an argument to the function, and thus it's not stored anywhere and can be thrown away afterwards. > Okay, in our case, all the devices are created at boot time, and if > omapdss were a module, the probes would have been thrown away after > module_init(), right? Yes. If omapdss is a module, the functions marked with __init are discarded after the module_init is done. If omapdss is built-in, the __init funcs are thrown away after the kernel's init done. > > One side effect of using platform_driver_probe() is that it returns > > ENODEV is there are no devices. In a simple module, the error can be > > then returned from module_init, thus causing the whole module to be > > unloaded. Our case is a bit more complex as we have multiple drivers in > > the same module. > > > > A downside with that is that we don't really know if the ENODEV error > > happened because there were no devices (which is ok), or if it came from > > probe function (which is not so ok). However, I thought that it doesn't > > matter if an output driver has failed. We can still continue with the > > other output drivers just fine. > > If we ensure that none of our probes return ENODEV(even though it may > make sense to return it if a func within probe fails), we could > differentiate between the 2 cases, right? True, I thought about that. But we can never be sure that the functions called by the probe (clk_get, or whatever) won't return ENODEV. Of course, we could check what they return, and change the error to something else, but I'm not sure if that's good either. > > > > Actually, there is a small problem. If, for example, DSI driver fails to > > load, and DPI driver tries to use DSI PLL... > > If we could differentiate between an error occuring because the device > doesn't exist and an error occuring because the probe failed, we could > bail out if any of the probes fail, right? Yes, but it feels a bit hackish to try to use the error as I pointed out above. So I'd rather go the other way: the drivers should somehow register the stuff they offer, so for example when the DSI1 is probed, it should register the DSI1 PLL to dss core. And DPI would have to ask for the DSI1 PLL from dss core. That, of course, is not a trivial change, so for the moment this stuff is slightly broken in error cases. Perhaps we could figure out some kind of clean hack for that... Alternatively, if the platform driver code was changed to tell us clearly if it was the probe that failed or if there just weren't any devices, we could also use that. Tomi
Attachment:
signature.asc
Description: This is a digitally signed message part