Re: [PATCH RFC v2 3/4] pc-bios: s390x: Save io and external new PSWs before overwriting them

[Thomas Huth]

On 27/08/2020 16.30, Janosch Frank wrote:
> On 8/27/20 2:52 PM, Thomas Huth wrote:
>> On 27/08/2020 11.31, Janosch Frank wrote:
>>> Currently we always overwrite the mentioned exception new PSWs before
>>> loading the enabled wait PSW. Let's save the PSW before overwriting
>>> and restore it right before starting the loaded kernel.
>>>
>>> Signed-off-by: Janosch Frank <frankja@linux.ibm.com>
>>> ---
>>>
>>> Maybe we should rather statically allocate a lowcore so we don't dirty
>>> 0x0 at all.
>>>
>>> ---
>>>  pc-bios/s390-ccw/jump2ipl.c |  3 ++
>>>  pc-bios/s390-ccw/start.S    | 62 +++++++++++++++++++++++++++----------
>>>  2 files changed, 48 insertions(+), 17 deletions(-)
>>>
>>> diff --git a/pc-bios/s390-ccw/jump2ipl.c b/pc-bios/s390-ccw/jump2ipl.c
>>> index 143d027bf7..a44f3ab5b3 100644
>>> --- a/pc-bios/s390-ccw/jump2ipl.c
>>> +++ b/pc-bios/s390-ccw/jump2ipl.c
>>> @@ -13,12 +13,15 @@
>>>  #define KERN_IMAGE_START 0x010000UL
>>>  #define RESET_PSW_MASK (PSW_MASK_SHORTPSW | PSW_MASK_64)
>>>  
>>> +extern uint64_t *psw_save_io, *psw_save_ext;
>>
>> I think that should be
>>
>>  extern uint64_t psw_save_io[], psw_save_ext[];
>>
>> instead ... otherwise you'll end up with some funny bugs here, won't you?
> 
> What kind of bugs are you expecting?

Well, "extern uint64_t var[];" and "extern uint64_t *var;" are two
different kind of things. One is an array, one is a pointer variable.
Looking at your assembler code, you obviously tried to declare an array
there, not a pointer variable.

Have a try with this test program:

 #include <string.h>

 extern unsigned long *var;

 void main(void)
 {
        asm volatile (" nop ; nop ; nop "); /* marker */
        memcpy((void *)0x1f0, var, 16);
        asm volatile (" nop ; nop ; nop "); /* marker */
 }

After compiling that with -O2, and disassembling the corresponding .o
file, I get this code between the nops:

   c:   c4 18 00 00 00 00       lgrl    %r1,c <main+0xc>
                        e: R_390_PC32DBL        var+0x2
  12:   e7 00 10 00 00 06       vl      %v0,0(%r1)
  18:   e7 00 01 f0 00 0e       vst     %v0,496

The "lgrl %r1,var" is likely not what you wanted here.

If you now replace the "*var" with "var[]", you get this instead:

   c:   c0 10 00 00 00 00       larl    %r1,c <main+0xc>
                        e: R_390_PC32DBL        var+0x2
  12:   e7 00 10 00 00 06       vl      %v0,0(%r1)
  18:   e7 00 01 f0 00 0e       vst     %v0,496

"larl" looks better now, doesn't it?

>>
>>>  uint64_t *reset_psw = 0, save_psw, ipl_continue;
>>>  
>>>  static void jump_to_IPL_2(void)
>>>  {
>>>      /* Restore reset PSW and io and external new PSWs */
>>
>> Ok, now the comment makes sense :-)
>>>      *reset_psw = save_psw;
>>> +    memcpy((void *)0x1f0, psw_save_io, 16);
>>> +    memcpy((void *)0x1b0, psw_save_ext, 16);
>>
>> Could you use &lowcore->external_new_psw and &lowcore->io_new_psw
>> instead of the magic numbers?
> 
> I can, but that means that I need to declare lowcore in netmain.c as
> well as including s390-arch.h

If that does not cause any other big hurdles, I think I'd prefer that
instead of using magic numbers.

 Thanks,
  Thomas