On Wed, Aug 30, 2017 at 10:58:54AM +0100, Mark Rutland wrote:
> On Wed, Aug 30, 2017 at 12:22:49PM +0300, Yury Norov wrote:
> > Hi Mark, all.
>
> Hi Yury,
>
> > In patch 'dbc9344a68e506f19f8 ("arm64: clean up THREAD_* definitions")'
> > you move THREAD_* definitions from arch/arm64/include/asm/thread_info.h
> > to asm/memory.h. After that asm/thread_info.h starts to depend on
> > asm/memory.h.
> >
> > When I try to apply ilp32 series on top of it [1], it causes circular
> > dependencies like this one:
> >
> > In file included from ./arch/arm64/include/asm/memory.h:30:0,
> > from ./arch/arm64/include/asm/thread_info.h:30,
> > from ./include/linux/thread_bits.h:20,
> > from ./include/linux/thread_info.h:13,
> > from ./include/asm-generic/preempt.h:4,
> > from ./arch/arm64/include/generated/asm/preempt.h:1,
> > from ./include/linux/preempt.h:80,
> > from ./include/linux/rcupdate.h:40,
> > from ./include/linux/rculist.h:10,
> > from ./include/linux/pid.h:4,
> > from ./include/linux/sched.h:13,
> > from arch/arm64/kernel/asm-offsets.c:21:
> > ./arch/arm64/include/asm/is_compat.h: In function ‘is_a32_compat_task’:
> > ./arch/arm64/include/asm/is_compat.h:25:9: error: implicit declaration of
> > function ‘test_thread_flag’ [-Werror=implicit-function-declaration]
> > return test_thread_flag(TIF_32BIT);
> > ^~~~~~~~~~~~~~~~
> >
> > The problem is that asm/memory.h depends on asm/is_compat.h to define
> > TASK_SIZE, which in turn requires asm/thread_info.h.
>
> ... and <asm/thread_info.h> include <asm/memory.h>, giving a circular
> dependency.
>
> In other architectures, TASK_SIZE is defined in processor.h. Can we not
> move TASK_SIZE instead of THREAD_SIZE, given that TASK_SIZE is what
> causes the dependency?
>
> We'd need a new __ASSEMBLY__ guard, but otherwise it looks like that
> would solve the issue.
>
> > The most obvious solution for it is to create is_compat.c file and make
> > is_*_compat() real functions. The other option is to move THREAD_*
> > definitions to separated macro. I would prefer 2nd one because of following
> > reasons:
> > - TASK_SIZE macro is used many times in kernel, including hot paths;
> > - asm/memory.h is included widely, as well as asm/thread_info.h, and it's
> > better not to make them depend one from another;
> > - THREAD_SIZE etc are not memory-related definitions.
>
> I disagree with this last point. THREAD_SIZE is in <asm/memory.h>
> because it affects the kernel's memory layout, as with other definitions
> at the top of <asm/memory.h>.
>
> I would much prefer keeping those definitions together.
>
> > In this patch THREAD_* definitions moved to separated asm/thread_size.h
> > header. It's enough to resolve dependency above.
> >
> > If you find this approach useful, I can prepare other patch that moves
> > TASK_* definitions from asm/memory.h to new header to remove the dependency
> > from asm/is_compat.h.
>
> I'd prefer that we moved the TASK_SIZE* definitions first, but if
> necessary, I'm ok with seeing the THREAD_SIZE* definitions move.
OK, I will try this, and if it's be enough will send new patch. If not -
I'll resend this patch and one that moves TASK_SIZE* together in series.
> [...]
>
> > diff --git a/arch/arm64/include/asm/thread_info.h
> > b/arch/arm64/include/asm/thread_info.h
>
> > -#include <asm/memory.h>
> > +#include <asm/thread_size.h>
> > #include <asm/stack_pointer.h>
> > #include <asm/types.h>
>
> Nit: please keep includes ordered alphabetically.
OK
> [...]
>
> > diff --git a/arch/arm64/include/asm/thread_size.h
> > b/arch/arm64/include/asm/thread_size.h
>
> > +#ifdef __KERNEL__
>
> This can go. We haven't needed __KERNEL__ ifdefs for a long time now...
OK
> [...]
>
> > diff --git a/arch/arm64/kernel/hibernate-asm.S
> > b/arch/arm64/kernel/hibernate-asm.S
> > index e56d848b6466..46b91702ea26 100644
> > --- a/arch/arm64/kernel/hibernate-asm.S
> > +++ b/arch/arm64/kernel/hibernate-asm.S
> > @@ -22,7 +22,6 @@
> > #include <asm/asm-offsets.h>
> > #include <asm/assembler.h>
> > #include <asm/cputype.h>
> > -#include <asm/memory.h>
> > #include <asm/page.h>
> > #include <asm/virt.h>
>
> AFAICT, we can also lose <linux/errno.h>, <asm/cputype.h>, and
> <asm/virt.h>.
>
> Changes to this file might be better as a preparatory cleanup.
OK.
Yury
