Re: port-macppc/54331: macppc MP kernels fail to boot successfully in -current (8.99.49)

[Michael <macallan%netbsd.org@localhost>]

The following reply was made to PR port-macppc/54331; it has been noted by GNATS.

From: Michael <macallan%netbsd.org@localhost>
To: gnats-bugs%netbsd.org@localhost
Cc: 
Subject: Re: port-macppc/54331: macppc MP kernels fail to boot successfully
 in -current (8.99.49)
Date: Thu, 25 Jul 2019 13:12:31 -0400

 I finally had a look at the code and the logs - looks like a race
 between cpu_setup() on the newly hatched CPU and delay(200000) in
 cpu_spinup(), where we just wait a fixed time and then bail out if the
 other CPU isn't ready. This would explain why the 2nd CPU still wakes
 up later on but then things go south because we didn't install the IPI
 handler.
 
 The following patch adds a period of polling h->hatch_running for a
 while instead of just blindly increasing the delay().
 
 Index: cpu_subr.c
 ===================================================================
 RCS file: /cvsroot/src/sys/arch/powerpc/oea/cpu_subr.c,v
 retrieving revision 1.99
 diff -u -w -r1.99 cpu_subr.c
 --- cpu_subr.c  6 Feb 2019 07:32:50 -0000       1.99
 +++ cpu_subr.c  25 Jul 2019 17:03:14 -0000
 @@ -1381,6 +1385,17 @@
         __asm volatile ("dcbst 0,%0"::"r"(&h->hatch_running):"memory");
         __asm volatile ("sync; isync");
  #endif
 +       int hatch_bail = 0;
 +       while ((h->hatch_running < 1) && (hatch_bail < 100000)) {
 +               delay(1);
 +               hatch_bail++;
 +#ifdef CACHE_PROTO_MEI
 +               __asm volatile ("dcbi 0,%0"::"r"(&h->hatch_running):"memory");
 +               __asm volatile ("sync; isync");
 +               __asm volatile ("dcbst 0,%0"::"r"(&h->hatch_running):"memory");
 +               __asm volatile ("sync; isync");
 +#endif
 +       }
         if (h->hatch_running < 1) {
  #ifdef CACHE_PROTO_MEI
                 __asm volatile ("dcbi
                 0,%0"::"r"(&cpu_spinstart_ack):"memory");