Messages in this thread | | | Date | Thu, 22 Sep 2022 12:52:36 +0200 | From | "Arnd Bergmann" <> | Subject | Re: [mtd:nand/next 11/31] drivers/mtd/nand/raw/cadence-nand-controller.c:1893:4: error: implicit declaration of function 'ioread64_rep' is invalid in C99 |
| |
On Thu, Sep 22, 2022, at 11:36 AM, Miquel Raynal wrote: > vkorenblit@sequans.com wrote on Thu, 22 Sep 2022 10:18:46 +0200: >> >> Correct, this was my initial idea. However, this driver should work >> with every architecture or do we limit the scope to arm/arm64/x86_64? > > The driver should work on ARM and aarch64, I'm not aware of other > architectures with this IP. > > The driver should compile when COMPILE_TEST=y.
It should also be written in a way that makes it plausible to use elsewhere. Since this is just a licensed IP core, there is a good chance that someone reused it on mips or riscv, or anything else.
>> >> I believe what Valentin wanted to achieve in the first place, was to >> >> use 64-bit accesses when relevant (otherwise it does not work). >> > The width is read from a device specific register at >> > runtime, it is not related to the architecture you are >> > running on, presumably this is hardwired during the >> > design of an SoC, based on the capabilities of the DMA >> > engine: > > Well, yes, but in the mean time 64-bit DMA width will never be > used on 32-bit platforms.
Why? Most architectures (including x86 and arm) allow you to run a 32-bit kernel on a 64-bit SoC. While this is almost always a bad idea to actually do, a driver should be written to work correctly in this setup.
>> > This usually means the largest access that is valid for >> > reading from the FIFO, but usually smaller accesses work >> > as well, just slower. > > Mmh, ok, that's interesting, thanks for the pointer. > > But in the mean time I am only half satisfied, because we plan to do > twice more accesses than needed _just_ because of a the COMPILE_TEST > constraint.
In my example, I had an #ifdef so it would only fall back to 32-bit accesses on the 64-bit register when running an actual 32-bit kernel, but leaving the 64-bit case efficient.
Arnd
| |