Re: segmentation fault on power system

[Han Pingtian <hanpt@linux.vnet.ibm.com>]

On Tue, Jul 23, 2013 at 09:15:40AM -0600, Jens Axboe wrote:
> On 07/23/2013 04:23 AM, Erwan Velu wrote:
> > On 23/07/2013 12:13, Han Pingtian wrote:
> >> Hey there,
> >>
> >> When trying to run fio on one of our power system, segmentation fault
> >>
> > Can you give us the kind of cpu you are using ? (/proc/cpuinfo)
> > I'm not used with PPC but maybe your processor doesn't support ATBU call
> > on mfspr.
> 
> That is definitely the problem, so CPU info would help. In the mean
> time, you can use clocksource=clock_gettime to get rid of the illegal
> instruction.

This is the contents of /proc/cpuinfo:

processor       : 0
cpu             : POWER7 (architected), altivec supported
clock           : 3550.000000MHz
revision        : 2.3 (pvr 003f 0203)

processor       : 1
cpu             : POWER7 (architected), altivec supported
clock           : 3550.000000MHz
revision        : 2.3 (pvr 003f 0203)

processor       : 2
cpu             : POWER7 (architected), altivec supported
clock           : 3550.000000MHz
revision        : 2.3 (pvr 003f 0203)

processor       : 3
cpu             : POWER7 (architected), altivec supported
clock           : 3550.000000MHz
revision        : 2.3 (pvr 003f 0203)

timebase        : 512000000
platform        : pSeries
model           : IBM,8231-E2C
machine         : CHRP IBM,8231-E2C

But 'clocksource=clock_gettime' doesn't fix the fault. I can also get
the same two cores. 

If I changed the code like this:

================================================================================
diff --git a/arch/arch-ppc.h b/arch/arch-ppc.h
index 65e6b74..30c315c 100644
--- a/arch/arch-ppc.h
+++ b/arch/arch-ppc.h
@@ -67,15 +67,15 @@ static inline unsigned long long get_cpu_clock(void)
        unsigned long long ret;

        do {
-               if (arch_flags & ARCH_FLAG_1) {
-                       tbu0 = mfspr(SPRN_ATBU);
-                       tbl = mfspr(SPRN_ATBL);
-                       tbu1 = mfspr(SPRN_ATBU);
-               } else {
+               //if (arch_flags & ARCH_FLAG_1) {
+               //      tbu0 = mfspr(SPRN_ATBU);
+               //      tbl = mfspr(SPRN_ATBL);
+               //      tbu1 = mfspr(SPRN_ATBU);
+               //} else {
                        tbu0 = mfspr(SPRN_TBRU);
                        tbl = mfspr(SPRN_TBRL);
                        tbu1 = mfspr(SPRN_TBRU);
-               }
+               //}
        } while (tbu0 != tbu1);

        ret = (((unsigned long long)tbu0) << 32) | tbl;
================================================================================

then only one core dumpped which has this backtrace:

================================================================================
Core was generated by `./fio/fio --debug=parse fio-jobs/randomw.fio '.
Program terminated with signal 11, Segmentation fault.
#0  0x000000001006288c in init_disk_util (td=0xfff9a730000) at diskutil.c:481
481             if (!td->o.do_disk_util ||
(gdb) bt
#0  0x000000001006288c in init_disk_util (td=0xfff9a730000) at diskutil.c:481
#1  0x0000000010050d30 in run_threads () at backend.c:1691
#2  0x000000001005159c in fio_backend () at backend.c:1911
#3  0x00000000100669a4 in main (argc=<value optimized out>, argv=0xfffd20c9ec8, envp=<value optimized out>)
    at fio.c:50
(gdb) p td
$1 = (struct thread_data *) 0xfff9a730000
(gdb) p threads
$2 = (struct thread_data *) 0xfff9a730000
(gdb) p td->io_ops
$3 = (struct ioengine_ops *) 0x0
(gdb) p threads->io_ops
$4 = (struct ioengine_ops *) 0x1000e262720
(gdb)
================================================================================