[RFC,2/8] thread_info: allow custom in-task thread_info

Currently, task_struct is defined in <linux/sched.h>, which (indirectly)
pulls in a number of low-level arch headers such as <asm/preempt.h>
through a number of other headers. Thus, code and structures in these
headers can't rely on the definition of task_struct. Some of these
headers are necessary for the definition of task_struct, so moving
task_struct into its own header is insufficient tio avoid circular
includes.

With CONFIG_THREAD_INFO_IN_TASK, arch code needs to implement its own
get_current() for the generic get_thread_info(). To avoid header
dependency issues, this relies on thread_info being the first member of
task_struct. This can be used by low-level arch code, as it doesn't
depend on the definition of task_struct.

For architectures without preempt-safe this_cpu ops, some data required
by low-level arch code (e.g. preempt_count) has to be stored per-thread,
and for the reasons above, we cannot place this in task_struct (or
thread_struct, since we cannot know its offset from within task_struct).
The only practical location for these is thread_info.

This patch allows architectures with CONFIG_ARCH_HAS_OWN_THREAD_INFO to
define their own thread_info, avoiding the problems described above.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>

Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Kees Cook <keescook@chromium.org>
Cc: Will Deacon <will.deacon@arm.com>
Cc: linux-kernel@vger.kernel.org
---
 include/linux/thread_info.h | 3 ++-
 init/Kconfig                | 3 +++
 2 files changed, 5 insertions(+), 1 deletion(-)

-- 
1.9.1

On Thu, Sep 15, 2016 at 11:37:47AM -0700, Andy Lutomirski wrote:
> On Thu, Sep 15, 2016 at 6:49 AM, Mark Rutland <mark.rutland@arm.com> wrote:

> > Currently, task_struct is defined in <linux/sched.h>, which (indirectly)

> > pulls in a number of low-level arch headers such as <asm/preempt.h>

> > through a number of other headers. Thus, code and structures in these

> > headers can't rely on the definition of task_struct. Some of these

> > headers are necessary for the definition of task_struct, so moving

> > task_struct into its own header is insufficient tio avoid circular

> > includes.

> 

> The flippant answer is to fix the headers, but I tried that myself and

> gave up :(


Agreed; likewise (though I gave up quicker, I suspect). :(

Longer-term I'd still hope that we can do this.

> But how about this slightly less duplicative alternative:

> 

> struct thread_info {

> #ifdef arch_thread_info

>   struct arch_thread_info arch_ti;

> #endif

> };


I'm happy to have an arch_thread_info.

Just to check, what do you mean to happen with the flags field? Should
that always be in the generic thread_info? e.g.

struct thread_info {
	u32 flags;
#ifdef arch_thread_info
	struct arch_thread_info arch_ti;
#endif
};

Thanks,
Mark,

Hi Andy,

On Fri, Sep 16, 2016 at 08:11:14AM -0700, Andy Lutomirski wrote:
> > On Thu, Sep 15, 2016 at 11:37:47AM -0700, Andy Lutomirski wrote:

> > Just to check, what do you mean to happen with the flags field? Should

> > that always be in the generic thread_info? e.g.

> >

> > struct thread_info {

> >         u32 flags;

> > #ifdef arch_thread_info

> >         struct arch_thread_info arch_ti;

> > #endif

> > };

> 

> Exactly.  Possibly with a comment that using thread_struct should be

> preferred and that arch_thread_info should be used only if some header

> file requires access via current_thread_info() or task_thread_info().


While fixing up these patches, I realised that I'm somewhat concerned by 
flags becoming a u32 (where it was previously an unsigned long for
arm64).

The generic {test,set,*}_ti_thread_flag() helpers use the usual bitops,
which perform accesses of sizeof(unsigned long) at a time, and for arm64
these need to be naturally-aligned.

We happen to get that alignment from subsequent fields in task_struct
and/or thread_info, and for arm64 we don't seem to have a problem with
tearing, but it feels somewhat fragile, and leaves me uneasy.

Looking at the git log, it seems that x86 also use unsigned long until
commit affa219b60a11b32 ("x86: change thread_info's flag field back to
32 bits"), where if I'm reading correctly, this was done to get rid of
unnecessary padding. With THREAD_INFO_IN_STACK, thread_info::flags is
immediately followed by a long on x86, so we save no padding.

Given all that, can we make the generic thread_info::flags an unsigned
long, matching what the thread flag helpers implicitly assume?

Thanks,
Mark.

On Thu, Sep 22, 2016 at 03:23:59PM -0700, Andy Lutomirski wrote:
> On Sep 21, 2016 12:28 AM, "Mark Rutland" <mark.rutland@arm.com> wrote:


> > Given all that, can we make the generic thread_info::flags an unsigned

> > long, matching what the thread flag helpers implicitly assume?

> 

> Yes.  Want to send the patch or should I?


I've sent a patch out to LKML [1], though if that doesn't look right I'm
more than happy for you to send a correct one. ;)

Thanks,
Mark.

[1] http://lkml.kernel.org/r/1474651447-30447-1-git-send-email-mark.rutland@arm.com