Re: kern.timecounter.smp_tsc_adjust

>In any case, at least one, and maybe two, routines need their
>types changed (and maybe the TSC_READ macro as well).  Is it OK
>to just assume the upper 32 bits are in sync?

For whatever it may be worth, the following patch (to change
everything to 64 bits) is boot-tested and no longer crashes the
system that did crash with the unpatched version.  It also
successfully synchronized TSCs at least once (Intel based board
with the "invariant TSC" bit set, 40 CPUs).

Chris

    x86/tsc: fix SMP TSC adjustment code
    
    On SMP systems, if kern.timecounter.smp_tsc is set, the kernel
    attempts to determine whether the TSCs on the processors are
    sufficiently in-sync to be used for time counting.  If not, and
    kern.timecounter.smp_tsc_adjust is set, we then attempt to
    adjust all the TSCs.
    
    The adjustment code assumed we kept a full 64-bit value for each
    TSC, but the data-gathering and comparison code used 32-bit
    values.  As a result, if the timecounters were out of sync, the
    adjustment code could crash (depending on the number of CPUs).
    
    This converts everything to full 64-bit values.

diff --git a/x86/x86/tsc.c b/x86/x86/tsc.c
index 30b8a30..cf05e94 100644
--- a/x86/x86/tsc.c
+++ b/x86/x86/tsc.c
@@ -373,11 +373,11 @@ init_TSC(void)
 static void								\
 tsc_read_##x(void *arg)							\
 {									\
-	uint32_t *tsc = arg;						\
+	uint64_t *tsc = arg;						\
 	u_int cpu = PCPU_GET(cpuid);					\
 									\
 	__asm __volatile("cpuid" : : : "eax", "ebx", "ecx", "edx");	\
-	tsc[cpu * 3 + x] = rdtsc32();					\
+	tsc[cpu * 3 + x] = rdtsc();					\
 }
 TSC_READ(0)
 TSC_READ(1)
@@ -389,8 +389,8 @@ TSC_READ(2)
 static void
 comp_smp_tsc(void *arg)
 {
-	uint32_t *tsc;
-	int32_t d1, d2;
+	uint64_t *tsc;
+	int64_t d1, d2;
 	u_int cpu = PCPU_GET(cpuid);
 	u_int i, j, size;
 
@@ -454,7 +454,7 @@ adj_smp_tsc(void *arg)
 static int
 test_tsc(void)
 {
-	uint32_t *data, *tsc;
+	uint64_t *data, *tsc;
 	u_int i, size, adj;
 
 	if ((!smp_tsc && !tsc_is_invariant) || vm_guest)
_______________________________________________
freebsd-hackers@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-hackers
To unsubscribe, send any mail to "freebsd-hackers-unsubscribe@freebsd.org"

