On Mon, Mar 16, 2020 at 10:55:00AM +0000, Vincenzo Frascino wrote:

On 3/16/20 10:34 AM, Catalin Marinas wrote:

quoted

quoted

I tried to fine grain the headers as much as I could in order to avoid
unneeded/unwanted inclusions:
 * TASK_SIZE_32 is used to verify ABI consistency on vdso32 (please refer to
   arch/arm64/kernel/vdso32/vgettimeofday.c).

I see. But the test is probably useless. With 4K pages, TASK_SIZE_32 is
1UL << 32, so you can't have a u32 greater than this. So I'd argue that
the ABI compatibility here doesn't matter.

With 16K or 64K pages, TASK_SIZE_32 is slightly smaller but arm32 never
supported it.

What's the side-effect of dropping this check altogether?

The main side-effect is that arm32 and arm64 compat have a different behavior,
that it is what we want to avoid.

The vdsotest [1] I am using, verifies all the side conditions with respect to
the ABI, which we are now compatible with. Removing those checks would break
this condition.

As I said above, I don't see how removing 'if ((u32)ts >= (1UL << 32))'
makes any difference. This check was likely removed by the compiler
already.

Also, userspace doesn't have a trivial way to figure out TASK_SIZE and I
can't see anything that tests this in the vdsotest (though I haven't
spent that much time looking). If it's hard-coded, note that arm32
TASK_SIZE is different from TASK_SIZE_32 on arm64.

Can you tell what actually is failing in vdsotest if you remove the
TASK_SIZE_32 checks in the arm64 compat vdso?

-- 
Catalin

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel