On Wed, May 01, 2019 at 09:17:37AM +0200, Esben Haabendal wrote: > Andy Shevchenko <andriy.shevchenko@xxxxxxxxxxxxxxx> writes: > > Hmm... Currently it's done inside individual port drivers, like 8250_dw.c. > > Each of the drivers can do it differently, for example 8250_lpss.c or > > 8250_pnp.c. > > So, you would prefer to create a new "specialized" port driver that uses > platform resources? I am not doing anything else different from > the generic port driver here in 8250_core.c. If it's required and using serial8250 directly is not enough. > >> + if (!(port->flags & UPF_DEV_RESOURCES)) > >> + release_mem_region(port->mapbase, size); > > > > This is again same issue. The parent should not request resource it > > doesn't use. > > Yes, this is same issue. > > But the last part is not true. A parent mfd driver might "use" a memory > resource for the sole purpose of splitting it up for it's mfd child > devices. This is a core part of mfd framework, and not something I am > inventing with this patch. I am just trying to make it possible to use > 8250 driver in that context. > > > I think I understand what is a confusion here. > > > > For the IO resources we have two operations: > > - mapping / re-mapping (may be shared) > > - requesting (exclusive) > > > > In the parenthesis I put a level of access to it. While many device > > drivers can *share* same resource (mapped or unmapped), the only one > > can actually request it. > > Mostly true. But there is an important twist to the exclusive restriction. > > The exclusive part of the request is limited to the the same root/parent > resource. > > When you request a memory resource from the root resource > (iomem_resource), the resource returned can be used as a new parent > resource. This new parent can then be used to give exclusive access to > slices of that resource. When used like that, I expect that the parent > resource is not supposed to be used for anything else than honoring > resource requests. > > And this is exactly what mfd-core uses the mem_base argument > in mfd_add_devices(). > > > So, the parent can take an slice resources as it would be > > appropriated, but not requesting them. > > The parent is not and should not be doing that by itself. The request > is done on by mfd-core when mfd_add_devices() is called. No, MFD *does not* (and actually *may not* in order to allow standalone drivers to be used as children w/o modifications) request resources. It just passes them to children as parent suggested. > > OTOH, it's possible to have a (weird) MFD case where parent *requested* > > resources, and *all* of its children are aware of that. > > I am not sure what you mean with this, but mfd drivers should not pass > along it's intire requested memory resource(s) to child devices. The > child devices will get the requested resource slices, as implemented by > mfd_add_devices(). > > I hope you can see that I am not violating any fundamental design > decissions here, but actually try adhere to them (resource management, > platform_device resource management, and mfd-core). -- With Best Regards, Andy Shevchenko
