Re: [PATCH] s390, sched: Fix thread_struct move fallout to __switch_to()

[Martin Schwidefsky <schwidefsky@de.ibm.com>]

On Mon, 20 Jul 2015 10:20:04 +0200
Ingo Molnar <mingo@kernel.org> wrote:

> 
> * Heiko Carstens <heiko.carstens@de.ibm.com> wrote:
> 
> > Unfortunately not true. It breaks the build on s390 since a couple of
> > displacements used in asm code now get too large:
> > 
> > arch/s390/kernel/entry.S:181: Error: operand out of range (0x00000000000018a8 is not between 0x0000000000000000 and 0x0000000000000fff)
> > arch/s390/kernel/entry.S:191: Error: operand out of range (0x00000000000018a8 is not between 0x0000000000000000 and 0x0000000000000fff)
> > arch/s390/kernel/entry.S:423: Error: operand out of range (0x0000000000001924 is not between 0x0000000000000000 and 0x0000000000000fff)
> > arch/s390/kernel/entry.S:437: Error: operand out of range (0x00000000000018e8 is not between 0x0000000000000000 and 0x0000000000000fff)
> > arch/s390/kernel/entry.S:438: Error: operand out of range (0x00000000000018e0 is not between 0x0000000000000000 and 0x0000000000000fff)
> > arch/s390/kernel/entry.S:439: Error: operand out of range (0x00000000000018f0 is not between 0x0000000000000000 and 0x0000000000000fff)
> > make[1]: *** [arch/s390/kernel/entry.o] Error 1
> > 
> > Let's see how we can fix this.
> 
> Sorry about this!
> 
> So I looked at the __switch_to() assembly, and the main complication appears to be 
> that we have two uses of 'prev':
> 
>         lg      %r4,__THREAD_info(%r2)          # get thread_info of prev
> 
> 
>         stg     %r15,__THREAD_ksp(%r2)          # store kernel stack of prev
> 
> the __THREAD_info offset is best expressed relative to task_struct - the 
> __THREAD_ksp offset is best expressed relative to thread_struct.
> 
> I.e. I think the best fix would be to extend the signature of s390's __switch_to() 
> from (prev,next) to (prev,next,prev_thread,next_thread).
> 
> Is the C parameter mapping r2,r3,r4,r5? If yes then the patch below should do the 
> trick. (Utterly untested.)
> 
> Note:
> 
>   - I renamed __THREAD_info to __TASK_thread_info, to better separate task_struct
>     and thread_struct offsets syntactically and visually.
> 
>   - I removed __THREAD_mm_segment which is unused
> 
> I also looked at fixing pgm_check_handler(), but my s390-fu gave up completely on 
> that one: task_struct is in 'r14', but this is a hardware entry function it 
> appears. So to fix it we'd have to pick a temporary register, put thread_struct 
> pointer into it, and fix up these offsets:
> 
> arch/s390/kernel/asm-offsets.c: DEFINE(__THREAD_per_address, offsetof(struct task_struct, thread.per_event.address));
> arch/s390/kernel/asm-offsets.c: DEFINE(__THREAD_per_paid, offsetof(struct task_struct, thread.per_event.paid));
> arch/s390/kernel/asm-offsets.c: DEFINE(__THREAD_trap_tdb, offsetof(struct task_struct, thread.trap_tdb));
> 
> to be thread_struct relative.

Just use the patch I've sent. No worries :-)

-- 
blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.