Source-Changes-HG archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

[src/trunk]: src/sys/arch/sparc introduce a hack to workaround the savefpstat...



details:   https://anonhg.NetBSD.org/src/rev/fee7b3750927
branches:  trunk
changeset: 760795:fee7b3750927
user:      mrg <mrg%NetBSD.org@localhost>
date:      Thu Jan 13 05:20:27 2011 +0000

description:
introduce a hack to workaround the savefpstate IPI get's NULL
struct fpstate *.

add a new entry point ipi_savefpstate() that is used for this IPI
instead of the plain ipi_savefpstate().  check for %o0 being NULL,
and if so, just return.

add event counters for the savefpstate IPI, and another one for
when it detects the NULL problem.

make the cpu_info struct xmpsg be aligned to a single cache line.

with this change applied (and another minor change in testing) my
SS20 with dual SM75's has survived for 17 hours on build.sh -j5,
and has reported 7 NULL savefps.  (the minor change seems to have
a significant effect at reducing this number, but more testing is
needed for it.)


XXX:  this is horrible and we really need to find the real problem
XXX:  but this should let people use sparc smp again, and fixes
XXX:  for the real problem can be tested by seeing if any of the
XXX:  savefp IPI null counter becomes non-zero.

diffstat:

 sys/arch/sparc/include/cpu.h      |   3 +-
 sys/arch/sparc/sparc/cpu.c        |  25 +++++++++++++++++++++-
 sys/arch/sparc/sparc/cpuvar.h     |  14 +++++++++---
 sys/arch/sparc/sparc/genassym.cf  |   4 ++-
 sys/arch/sparc/sparc/locore.s     |  42 +++++++++++++++++++++++++++++++++++++-
 sys/arch/sparc/sparc/machdep.c    |   6 ++--
 sys/arch/sparc/sparc/trap.c       |   6 ++--
 sys/arch/sparc/sparc/vm_machdep.c |   8 +++---
 8 files changed, 88 insertions(+), 20 deletions(-)

diffs (truncated from 307 to 300 lines):

diff -r 1f40df2c9b77 -r fee7b3750927 sys/arch/sparc/include/cpu.h
--- a/sys/arch/sparc/include/cpu.h      Thu Jan 13 05:14:48 2011 +0000
+++ b/sys/arch/sparc/include/cpu.h      Thu Jan 13 05:20:27 2011 +0000
@@ -1,4 +1,4 @@
-/*     $NetBSD: cpu.h,v 1.89 2010/01/03 23:03:21 mrg Exp $ */
+/*     $NetBSD: cpu.h,v 1.90 2011/01/13 05:20:27 mrg Exp $ */
 
 /*
  * Copyright (c) 1992, 1993
@@ -193,6 +193,7 @@
 
 /* locore.s */
 struct fpstate;
+void   ipi_savefpstate(struct fpstate *);
 void   savefpstate(struct fpstate *);
 void   loadfpstate(struct fpstate *);
 int    probeget(void *, int);
diff -r 1f40df2c9b77 -r fee7b3750927 sys/arch/sparc/sparc/cpu.c
--- a/sys/arch/sparc/sparc/cpu.c        Thu Jan 13 05:14:48 2011 +0000
+++ b/sys/arch/sparc/sparc/cpu.c        Thu Jan 13 05:20:27 2011 +0000
@@ -1,4 +1,4 @@
-/*     $NetBSD: cpu.c,v 1.223 2010/06/22 18:29:02 rmind Exp $ */
+/*     $NetBSD: cpu.c,v 1.224 2011/01/13 05:20:27 mrg Exp $ */
 
 /*
  * Copyright (c) 1996
@@ -52,7 +52,7 @@
  */
 
 #include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: cpu.c,v 1.223 2010/06/22 18:29:02 rmind Exp $");
+__KERNEL_RCSID(0, "$NetBSD: cpu.c,v 1.224 2011/01/13 05:20:27 mrg Exp $");
 
 #include "opt_multiprocessor.h"
 #include "opt_lockdebug.h"
@@ -64,7 +64,9 @@
 #include <sys/device.h>
 #include <sys/malloc.h>
 #include <sys/kernel.h>
+#include <sys/evcnt.h>
 #include <sys/xcall.h>
+#include <sys/cpu.h>
 
 #include <uvm/uvm.h>
 
@@ -119,6 +121,7 @@
     cpu_cpuunit_match, cpu_cpuunit_attach, NULL, NULL);
 #endif /* SUN4D */
 
+static void cpu_init_evcnt(struct cpu_info *cpi);
 static void cpu_attach(struct cpu_softc *, int, int);
 
 static const char *fsrtoname(int, int, int);
@@ -293,6 +296,21 @@
 }
 #endif /* SUN4D */
 
+static void
+cpu_init_evcnt(struct cpu_info *cpi)
+{
+
+       /*
+        * Setup the per-cpu savefpstate counters.  The "savefp null"
+        * counter should go away when the NULL struct fpstate * bug
+        * is fixed.
+        */
+       evcnt_attach_dynamic(&cpi->ci_savefpstate, EVCNT_TYPE_MISC,
+                            NULL, cpu_name(cpi), "savefp ipi");
+       evcnt_attach_dynamic(&cpi->ci_savefpstate_null, EVCNT_TYPE_MISC,
+                            NULL, cpu_name(cpi), "savefp null ipi");
+}
+
 /*
  * Attach the CPU.
  * Discover interesting goop about the virtual address cache
@@ -341,10 +359,13 @@
 #if defined(MULTIPROCESSOR)
        if (cpu_attach_count > 1) {
                cpu_attach_non_boot(sc, cpi, node);
+               cpu_init_evcnt(cpi);
                return;
        }
 #endif /* MULTIPROCESSOR */
 
+       cpu_init_evcnt(cpi);
+
        /* Stuff to only run on the boot CPU */
        cpu_setup();
        snprintf(buf, sizeof buf, "%s @ %s MHz, %s FPU",
diff -r 1f40df2c9b77 -r fee7b3750927 sys/arch/sparc/sparc/cpuvar.h
--- a/sys/arch/sparc/sparc/cpuvar.h     Thu Jan 13 05:14:48 2011 +0000
+++ b/sys/arch/sparc/sparc/cpuvar.h     Thu Jan 13 05:20:27 2011 +0000
@@ -1,4 +1,4 @@
-/*     $NetBSD: cpuvar.h,v 1.83 2010/01/23 16:06:57 mrg Exp $ */
+/*     $NetBSD: cpuvar.h,v 1.84 2011/01/13 05:20:27 mrg Exp $ */
 
 /*
  *  Copyright (c) 1996 The NetBSD Foundation, Inc.
@@ -127,6 +127,13 @@
 struct cpu_info {
        struct cpu_data ci_data;        /* MI per-cpu data */
 
+       /*
+        * Primary Inter-processor message area.  Keep this aligned
+        * to a cache line boundary if possible, as the structure
+        * itself is one (normal 32 byte) cache-line.
+        */
+       struct xpmsg    msg __aligned(32);
+
        /* Scheduler flags */
        int     ci_want_ast;
        int     ci_want_resched;
@@ -142,9 +149,6 @@
         */
        struct cpu_info * volatile ci_self;
 
-       /* Primary Inter-processor message area */
-       struct xpmsg    msg;
-
        int             ci_cpuid;       /* CPU index (see cpus[] array) */
 
        /* Context administration */
@@ -334,6 +338,8 @@
 
        struct evcnt ci_lev10;
        struct evcnt ci_lev14;
+       struct evcnt ci_savefpstate;
+       struct evcnt ci_savefpstate_null;
 };
 
 /*
diff -r 1f40df2c9b77 -r fee7b3750927 sys/arch/sparc/sparc/genassym.cf
--- a/sys/arch/sparc/sparc/genassym.cf  Thu Jan 13 05:14:48 2011 +0000
+++ b/sys/arch/sparc/sparc/genassym.cf  Thu Jan 13 05:20:27 2011 +0000
@@ -1,4 +1,4 @@
-#      $NetBSD: genassym.cf,v 1.61 2010/12/20 00:25:44 matt Exp $
+#      $NetBSD: genassym.cf,v 1.62 2011/01/13 05:20:27 mrg Exp $
 
 #
 # Copyright (c) 1998 The NetBSD Foundation, Inc.
@@ -166,6 +166,8 @@
 define CPUINFO_MTX_COUNT       offsetof(struct cpu_info, ci_mtx_count)
 define CPUINFO_MTX_OLDSPL      offsetof(struct cpu_info, ci_mtx_oldspl)
 define CPUINFO_IDEPTH          offsetof(struct cpu_info, ci_idepth)
+define CPUINFO_SAVEFPSTATE     offsetof(struct cpu_info, ci_savefpstate)
+define CPUINFO_SAVEFPSTATE_NULL offsetof(struct cpu_info, ci_savefpstate_null)
 
 # PTE bits and related information
 define PG_W            PG_W
diff -r 1f40df2c9b77 -r fee7b3750927 sys/arch/sparc/sparc/locore.s
--- a/sys/arch/sparc/sparc/locore.s     Thu Jan 13 05:14:48 2011 +0000
+++ b/sys/arch/sparc/sparc/locore.s     Thu Jan 13 05:20:27 2011 +0000
@@ -1,4 +1,4 @@
-/*     $NetBSD: locore.s,v 1.258 2010/12/20 00:25:44 matt Exp $        */
+/*     $NetBSD: locore.s,v 1.259 2011/01/13 05:20:27 mrg Exp $ */
 
 /*
  * Copyright (c) 1996 Paul Kranenburg
@@ -5841,12 +5841,29 @@
        /* NOTREACHED */
 
 /*
- * savefpstate(f) struct fpstate *f;
+ * savefpstate(struct fpstate *f);
+ * ipi_savefpstate(struct fpstate *f);
  *
  * Store the current FPU state.  The first `st %fsr' may cause a trap;
  * our trap handler knows how to recover (by `returning' to savefpcont).
+ *
+ * The IPI version just deals with updating event counters first.
  */
+Lpanic_savefpstate:
+       .asciz  "cpu%d: NULL fpstate"
+       _ALIGN
+
+ENTRY(ipi_savefpstate)
+       sethi   %hi(CPUINFO_VA), %o5
+       ldd     [%o5 + CPUINFO_SAVEFPSTATE], %o2
+       inccc   %o3
+       addx    %o2, 0, %o2
+       std     %o2, [%o5 + CPUINFO_SAVEFPSTATE]
+
 ENTRY(savefpstate)
+       cmp     %o0, 0
+       bz      Lfp_null_fpstate
+        nop
        rd      %psr, %o1               ! enable FP before we begin
        set     PSR_EF, %o2
        or      %o1, %o2, %o1
@@ -5889,6 +5906,27 @@
         std    %f30, [%o0 + FS_REGS + (4*30)]
 
 /*
+ * We really should panic here but while we figure out what the bug is
+ * that a remote CPU gets a NULL struct fpstate *, this lets the system
+ * work at least seemingly stably.
+ */
+Lfp_null_fpstate:
+#if 1
+       sethi   %hi(CPUINFO_VA), %o5
+       ldd     [%o5 + CPUINFO_SAVEFPSTATE_NULL], %o2
+       inccc   %o3
+       addx    %o2, 0, %o2
+       retl
+        std    %o2, [%o5 + CPUINFO_SAVEFPSTATE_NULL]
+#else
+       ld      [%o5 + CPUINFO_CPUNO], %o1
+       sethi   %hi(Lpanic_savefpstate), %o0
+       call    _C_LABEL(panic)
+        or     %o0, %lo(Lpanic_savefpstate), %o0
+#endif
+1:
+
+/*
  * Store the (now known nonempty) FP queue.
  * We have to reread the fsr each time in order to get the new QNE bit.
  */
diff -r 1f40df2c9b77 -r fee7b3750927 sys/arch/sparc/sparc/machdep.c
--- a/sys/arch/sparc/sparc/machdep.c    Thu Jan 13 05:14:48 2011 +0000
+++ b/sys/arch/sparc/sparc/machdep.c    Thu Jan 13 05:20:27 2011 +0000
@@ -1,4 +1,4 @@
-/*     $NetBSD: machdep.c,v 1.303 2010/12/20 00:25:44 matt Exp $ */
+/*     $NetBSD: machdep.c,v 1.304 2011/01/13 05:20:27 mrg Exp $ */
 
 /*-
  * Copyright (c) 1996, 1997, 1998 The NetBSD Foundation, Inc.
@@ -71,7 +71,7 @@
  */
 
 #include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: machdep.c,v 1.303 2010/12/20 00:25:44 matt Exp $");
+__KERNEL_RCSID(0, "$NetBSD: machdep.c,v 1.304 2011/01/13 05:20:27 mrg Exp $");
 
 #include "opt_compat_netbsd.h"
 #include "opt_compat_sunos.h"
@@ -376,7 +376,7 @@
                                savefpstate(fs);
 #if defined(MULTIPROCESSOR)
                        else
-                               XCALL1(savefpstate, fs, 1 << cpi->ci_cpuid);
+                               XCALL1(ipi_savefpstate, fs, 1 << cpi->ci_cpuid);
 #endif
                        cpi->fplwp = NULL;
                }
diff -r 1f40df2c9b77 -r fee7b3750927 sys/arch/sparc/sparc/trap.c
--- a/sys/arch/sparc/sparc/trap.c       Thu Jan 13 05:14:48 2011 +0000
+++ b/sys/arch/sparc/sparc/trap.c       Thu Jan 13 05:20:27 2011 +0000
@@ -1,4 +1,4 @@
-/*     $NetBSD: trap.c,v 1.184 2010/12/20 00:25:44 matt Exp $ */
+/*     $NetBSD: trap.c,v 1.185 2011/01/13 05:20:27 mrg Exp $ */
 
 /*
  * Copyright (c) 1996
@@ -49,7 +49,7 @@
  */
 
 #include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: trap.c,v 1.184 2010/12/20 00:25:44 matt Exp $");
+__KERNEL_RCSID(0, "$NetBSD: trap.c,v 1.185 2011/01/13 05:20:27 mrg Exp $");
 
 #include "opt_ddb.h"
 #include "opt_compat_svr4.h"
@@ -477,7 +477,7 @@
                                        panic("FPU(%d): state for %p",
                                                        cpi->ci_cpuid, l);
 #if defined(MULTIPROCESSOR)
-                               XCALL1(savefpstate, fs, 1 << cpi->ci_cpuid);
+                               XCALL1(ipi_savefpstate, fs, 1 << cpi->ci_cpuid);
 #endif
                                cpi->fplwp = NULL;
                        }
diff -r 1f40df2c9b77 -r fee7b3750927 sys/arch/sparc/sparc/vm_machdep.c
--- a/sys/arch/sparc/sparc/vm_machdep.c Thu Jan 13 05:14:48 2011 +0000
+++ b/sys/arch/sparc/sparc/vm_machdep.c Thu Jan 13 05:20:27 2011 +0000
@@ -1,4 +1,4 @@
-/*     $NetBSD: vm_machdep.c,v 1.101 2009/11/21 04:16:52 rmind Exp $ */
+/*     $NetBSD: vm_machdep.c,v 1.102 2011/01/13 05:20:27 mrg Exp $ */
 
 /*
  * Copyright (c) 1996
@@ -49,7 +49,7 @@
  */
 
 #include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: vm_machdep.c,v 1.101 2009/11/21 04:16:52 rmind Exp $");
+__KERNEL_RCSID(0, "$NetBSD: vm_machdep.c,v 1.102 2011/01/13 05:20:27 mrg Exp $");
 
 #include "opt_multiprocessor.h"
 
@@ -228,7 +228,7 @@
                                savefpstate(l1->l_md.md_fpstate);
 #if defined(MULTIPROCESSOR)
                        else
-                               XCALL1(savefpstate, l1->l_md.md_fpstate,
+                               XCALL1(ipi_savefpstate, l1->l_md.md_fpstate,
                                        1 << cpi->ci_cpuid);
 #endif
                }
@@ -306,7 +306,7 @@
                                savefpstate(fs);



Home | Main Index | Thread Index | Old Index