Hi Russell,

On Mon, Aug 23, 2021 at 02:39:18PM +0100, Russell King (Oracle) wrote:

On Mon, Aug 23, 2021 at 09:30:43PM +0800, Leo Yan wrote:

...

On Mon, Aug 23, 2021 at 01:23:42PM +0100, James Clark wrote:

[...]

...

...

For x86, it's possible to include tools/include/asm/atomic.h, but that doesn't include arch/arm/include/asm/atomic.h and there are some other #ifdefs that might make it not so easy for Arm. Just wondering if you considered trying to include the existing one? Or decided that it was easier to duplicate it?

Good finding!

With you reminding, I recognized that the atomic operations for arm/arm64 should be improved for user space program. So far, perf tool simply uses the compiler's atomic implementations (from asm-generic/atomic-gcc.h) for arm/arm64; but for a more reliable implementation, I think we should improve the user space program with architecture's atomic instructions.

No we should not. Sometimes, what's in the kernel is for the kernel's use only, and not for userspace's use. That may be because what works in kernel space does not work in userspace.

For example, the ARMv6+ atomic operations can be executed in userspace _provided_ they are only used on memory which has an exclusive monitor. They can't be used on anything that is not "normal memory".

Okay, IIUC, the requirement for "normal memory" and exclusive monitor should also apply on aarch64 for ldrex/strex, Load-Acquire and Store-Release instructions, etc. Otherwise, it's heavily dependent on the exclusive monitors outside the cache coherency domain (but this is out of the scopes for CPU).

perf tool is very likely to map memory with "normal memory" but we cannot say it's always true.

So I agree there have risk for exporting the aarch32/aarch64 atomic headers to user space.

Prior to ARMv6, the atomic operations rely on disabling interrupts. That facility is simply not available to userspace, so these must not be made available to userspace.

The same applies to bitops.

We've been here before in the past, when the kernel headers were not separated from the user ABI headers, and people would write programs that included e.g. bitops.h on x86 because they had optimised bitops code. This made the userspace programs very non-portable - without re-implementing userspace versions of this stuff in every userspace program that did this stuff.

So no, having experienced the effects of this kind of thing in the past, the kernel should _not_ export architecture specific code in header files to userspace.

Also, it should be pointed out that by doing so, you create a licensing issue. If the code is GPLv2, and you build your program such that it incorporates GPLv2 code, then if the userspace program is not GPLv2 compliant, you have a licensing problem, and in effect the program can be distributed.

Please do not go down this route.

Thanks a lot for the suggestion and quick response.

Leo