On 01/27/2013 06:03 PM, Mats Kärrman wrote:
If watchdog is enabled, the arch/powerpc/lib/ticks.S::wait_ticks()
function
calls the function specified by the WATCHDOG_RESET macro. The wait_ticks function depends on the registers r0, r6 and r7 being preserved however that is not guaranteed, e.g. if the reset function
is a
C function this will probably overwrite r0 and cause an endless loop.
The following patch changes to using r14+r15 instead of r6+r7 (to
resemble
what would have been generated by a C compiler) and saves all
necessary
registers on the stack.
The patch has been tested on a custom MPC5125 based machine using the
512x
powerpc architecture.
Signed-off-by: Mats Karrman mats.karrman@tritech.se Cc: Wolfgang Denk wd@denx.de Acked-by: Joakim Tjernlund joakim.tjernlund@transmode.se
Thanks for this fix. I just reproduced this problem on another MPC5200 based board (a4m2k) while implementing watchdog support. And this patch fixes the problem (hangup in wait_ticks).
So:
Tested-by: Stefan Roese sr@denx.de
If someone wants to play with a C only impl. I THINK it this is a good starting point:
#define WATCHDOG_RESET
/*static inline */unsigned long long get_ticks(void) { union xx { unsigned long long r3_r4; struct yx { unsigned long r3; unsigned long r4; } xy; } res; register unsigned long tmp;
do { asm ("mftbu %0\n\t" "mftb %1\n\t" "mftbu %2" : "=r"(res.xy.r3), "=r"(res.xy.r4), "=r"(tmp)); } while (res.xy.r3 != tmp); return res.r3_r4; }
wait_ticks(unsigned long wt) { unsigned long long ticks_end;
ticks_end = get_ticks() + wt; WATCHDOG_RESET; do {
} while((signed long long)ticks_end - (signed long long)get_ticks() >= 0); }