08:13 <mani_s> arnd, is the {get/put}_unaligned macros supposed to be working fine on unaligned device memory on ARM64?

08:14 <mani_s> I thought it would but this macro boils down to ldr that generates alignment fault

08:14 <mani_s> ldrx3, [x1, #20]!

08:15 <mani_s> So x1 has the address of the packed struct and 20 is the offset of the u64 field that is not 8byte aligned

08:57 <arnd> mani_s: for device memory, you have to use readb()/writeb()

08:57 <mani_s> arnd, okay, so memcpy_{from/to}io should be fine?

08:58 <arnd> get_unaligned()/put_unaligned() is only meaningful on normal memory, where it will use whatever the architecture allows, e.g. ldr on arm/arm64 but not ldrd on arm

08:58 <arnd> memcpy_fromio works fine, that should fall back to byte accesses where necessary, but may use larger alignment if it can

08:59 <arnd> not sure if the arm64 implementation makes use of the optimized path

08:59 <mani_s> arnd, yes, the arm64 implementation will use readb/writeb for non-aligned addresses

08:59 <mani_s> and readl/writel/ for 8byte aligned ones

09:00 <mani_s> arnd, I think this limitation should be documented in doc/Documentation/unaligned-memory-access.txt

09:00 <mani_s> what do you think?

09:01 <arnd> good idea. If you send a patch for Documentation/core-api/unaligned-memory-access.rst to the documentation maintainers, I'll review it

09:02 <mani_s> arnd, thanks

09:02 <mani_s> will do

09:04 <arnd> There is a bit in Documentation/driver-api/device-io.rst that indicates that it's wrong to access __iomem pointers with anything other than helper functions expecting an __iomem pointer, but it's not explicit the effects on alignment

09:09 <arnd> maybe add a pointer to the device-io.rst in unaligned-memory-access.rst when mentioning __iomem, that should keep it nice and concise

09:12 <mani_s> arnd, sounds good!

10:17 <aka_[m]> Do anyone have idea why touchscreen driver fails to probe with timeout?... (full message at https://matrix.org/_matrix/media/r0/download/matrix.org/tmNpwTcgUoalNhhhLdzUrhxo)

15:29 <Tooniis[m]> aka_: maybe you need to increase the timeout period a bit? if it works after a reload then maybe whatever timed out before became ready before the second probe