On Wed, Aug 23, 2023 at 04:49:05PM +0200, Andi Shyti wrote: > Hi, > > On Tue, Aug 22, 2023 at 03:08:38PM +0300, Andy Shevchenko wrote: > > On Tue, Aug 22, 2023 at 02:00:05PM +0200, Geert Uytterhoeven wrote: > > > On Tue, Aug 22, 2023 at 1:44 PM Andy Shevchenko > > > <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote: > > > > On Tue, Aug 22, 2023 at 09:21:19AM +0200, Geert Uytterhoeven wrote: > > > > > On Mon, Aug 21, 2023 at 7:09 PM Andy Shevchenko > > > > > <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote: > > > > > > On Mon, Aug 21, 2023 at 05:40:05PM +0200, Geert Uytterhoeven wrote: > > > > > > > On Mon, Aug 21, 2023 at 5:25 PM Andy Shevchenko > > > > > > > <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote: > > > > > > > > On Mon, Aug 21, 2023 at 03:27:43PM +0200, Geert Uytterhoeven wrote: > > > > > > > > > On Mon, Aug 21, 2023 at 3:04 PM Andy Shevchenko > > > > > > > > > <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote: > > > > > > > > > > On Sun, Aug 20, 2023 at 07:44:00PM +0100, Biju Das wrote: > > > > ... > > > > > > > > > > > > For all your work likes this as I noted in the reply to Guenter that > > > > > > > > > > the couple of the selling points here are: > > > > > > > > > > 1) avoidance of the pointer abuse in OF table > > > > > > > > > > (we need that to be a valid pointer); > > > > > > > > > > > > > > > > > > There is no pointer abuse: both const void * (in e.g. of_device_id) > > > > > > > > > and kernel_ulong_t (in e.g. i2c_device_id) can be used by drivers > > > > > > > > > to store a magic cookie, being either a pointer, or an integer value. > > > > > > > > > The same is true for the various unsigned long and void * "driver_data" > > > > > > > > > fields in subsystem-specific driver structures. > > > > > > > > > > > > > > > > (void *)5 is the abuse of the pointer. > > > > > > > > We carry something which is not a valid pointer from kernel perspective. > > > > > > > > > > > > > > But the data field is not required to be a valid pointer. > > > > > > > What kind and type of information it represents is specific to the driver. > > > > > > > > > > > > Where to find necessary information which is not always an integer constant. > > > > > > For example, for the driver data that has callbacks it can't be invalid pointer. > > > > > > > > > > If the driver uses it to store callbacks, of course it needs to be a > > > > > valid pointer. But that is internal to the driver. It is not that > > > > > we're passing random integer values to a function that expects a > > > > > pointer that can actually be dereferenced. > > > > > > > > > > > Since OF ID table structure is universal, it uses pointers. Maybe you need to > > > > > > update it to use plain integer instead? > > > > > > > > > > It is fairly common in the kernel to use void * to indicate a > > > > > driver-specific cookie, being either a real pointer or an integral > > > > > value, that is passed verbatim. See also e.g. the "dev" parameter > > > > > of request_irq(). > > > > > > > > Yes, that parameter is void * due to calling kfree(free_irq(...)). > > > > So, that's argument for my concerns. > > > > > > Sorry, I don't understand this comment. > > > (kfree(free_irq(...)) is only called in pci_free_irq()?) > > > > Passing void * for a "driver cookie" makes sense due to possibility of the > > passing it to other functions that want to have void * as your example shows. > > And that supports my idea of having void * over the unsigned long. > > I actually agree with Andy here... not much to add to his > arguments but if a void * is used as an integer then just change > the type. > > I also was quite puzzled when I started seeing this flow of > patches. > > I would rather prefer to store pointers in u64 variables rather > than integers in a pointer. I think pointers should be stored in pointers, and preserve constness. The reason that many legacy device id structures use kernel_ulong_t to store driver "cookie" is shortsightedness on our part long time ago. Back then we just needed "a simple piece of data" to be attached, a true/false flag or maybe a combination of flags. Nowadays we often want to attach a structure describing a chip's parameters there with several flags, maybe some methods, etc. So OF/DT folks did the right thing, and used const void *, and we should try and convert the rest to use const void * as well. I posted RFC for this for I2C here: https://lore.kernel.org/all/20230814-i2c-id-rework-v1-0-3e5bc71c49ee@xxxxxxxxx/ And then we can unify OF and legacy handling if driver tables. Thanks. -- Dmitry
