Linus, all,
This patch, written with the advice of Joel Becker, addresses a problem
with the hangcheck-timer. The basic problem is that the hangcheck-timer
code (Required for Oracle) needs a accurate hard clock which can be used
to detect OS stalls (due to udelay() or pci bus hangs) that would cause
system time to skew (its sort of a sanity check that insures the
system's notion of time is accurate). However, currently they are using
get_cycles() to fetch the cpu's TSC register, thus this does not work on
systems w/o a synced TSC. As suggested by Andi Kleen (see thread here:
http://www.uwsg.iu.edu/hypermail/linux/kernel/0302.0/1234.html ) I've
worked with Joel and others to implement the monotonic_clock()
interface. Some of the major considerations made when writing this patch
were:
o Needs to be able to return accurate time in the absence of
multiple timer interrupts
o Needs to be abstracted out from the hardware
o Avoids impacting gettimeofday() performance
This interface returns a unsigned long long representing the number of
nanoseconds that has passed since time_init().
Please consider for inclusion.
thanks
-john
Resending w/o the GPG signature to avoid crazy MIME corruption.
Linus, all,
This patch, written with the advice of Joel Becker, addresses a
problem
with the hangcheck-timer. The basic problem is that the hangcheck-timer
code (Required for Oracle) needs a accurate hard clock which can be used
to detect OS stalls (due to udelay() or pci bus hangs) that would cause
system time to skew (its sort of a sanity check that insures the
system's notion of time is accurate). However, currently they are using
get_cycles() to fetch the cpu's TSC register, thus this does not work on
systems w/o a synced TSC. As suggested by Andi Kleen (see thread here:
http://www.uwsg.iu.edu/hypermail/linux/kernel/0302.0/1234.html ) I've
worked with Joel and others to implement the monotonic_clock()
interface. Some of the major considerations made when writing this patch
were:=20
o Needs to be able to return accurate time in the absence of
multiple timer interrupts
o Needs to be abstracted out from the hardware
o Avoids impacting gettimeofday() performance
This interface returns a unsigned long long representing the number of
nanoseconds that has passed since time_init().=20
Please consider for inclusion.
thanks
-john
diff -Nru a/arch/i386/kernel/time.c b/arch/i386/kernel/time.c
-- a/arch/i386/kernel/time.c Fri Mar 28 14:31:18 2003
+++ b/arch/i386/kernel/time.c Fri Mar 28 14:31:18 2003
@@ -138,6 +138,23 @@
clock_was_set();
}
+/* monotonic_clock(): returns # of nanoseconds passed since time_init()
+ * Note: This function is required to return accurate
+ * time even in the absence of multiple timer ticks.
+ */
+unsigned long long monotonic_clock(void)
+{
+ unsigned long long ret;
+ unsigned long seq;
+ do {
+ seq = read_seqbegin(&xtime_lock);
+ ret = timer->monotonic_clock();
+ } while (read_seqretry(&xtime_lock, seq));
+ return ret;
+}
+EXPORT_SYMBOL(monotonic_clock);
+
+
/*
* In order to set the CMOS clock precisely, set_rtc_mmss has to be
* called 500 ms after the second nowtime has started, because when
diff -Nru a/arch/i386/kernel/timers/timer_cyclone.c b/arch/i386/kernel/timers/timer_cyclone.c
-- a/arch/i386/kernel/timers/timer_cyclone.c Fri Mar 28 14:31:18 2003
+++ b/arch/i386/kernel/timers/timer_cyclone.c Fri Mar 28 14:31:18 2003
@@ -27,19 +27,26 @@
#define CYCLONE_MPMC_OFFSET 0x51D0
#define CYCLONE_MPCS_OFFSET 0x51A8
#define CYCLONE_TIMER_FREQ 100000000
-
+#define CYCLONE_TIMER_MASK (((u64)1<<40)-1) /* 40 bit mask */
int use_cyclone = 0;
static u32* volatile cyclone_timer; /* Cyclone MPMC0 register */
-static u32 last_cyclone_timer;
+static u32 last_cyclone_low;
+static u32 last_cyclone_high;
+static unsigned long long monotonic_base;
static void mark_offset_cyclone(void)
{
int count;
+ unsigned long long this_offset, last_offset;
+ last_offset = ((unsigned long long)last_cyclone_high<<32)|last_cyclone_low;
+
spin_lock(&i8253_lock);
/* quickly read the cyclone timer */
- if(cyclone_timer)
- last_cyclone_timer = cyclone_timer[0];
+ do {
+ last_cyclone_high = cyclone_timer[1];
+ last_cyclone_low = cyclone_timer[0];
+ } while (last_cyclone_high != cyclone_timer[1]);
/* calculate delay_at_last_interrupt */
outb_p(0x00, 0x43); /* latch the count ASAP */
@@ -50,6 +57,10 @@
count = ((LATCH-1) - count) * TICK_SIZE;
delay_at_last_interrupt = (count + LATCH/2) / LATCH;
+
+ /* update the monotonic base value */
+ this_offset = ((unsigned long long)last_cyclone_high<<32)|last_cyclone_low;
+ monotonic_base += (this_offset - last_offset) & CYCLONE_TIMER_MASK;
}
static unsigned long get_offset_cyclone(void)
@@ -63,7 +74,7 @@
offset = cyclone_timer[0];
/* .. relative to previous jiffy */
- offset = offset - last_cyclone_timer;
+ offset = offset - last_cyclone_low;
/* convert cyclone ticks to microseconds */
/* XXX slow, can we speed this up? */
@@ -73,6 +84,25 @@
return delay_at_last_interrupt + offset;
}
+static unsigned long long monotonic_clock_cyclone(void)
+{
+
+ u32 now_low, now_high;
+ unsigned long long last_offset, this_offset;
+ unsigned long long ret;
+
+ do {
+ now_high = cyclone_timer[1];
+ now_low = cyclone_timer[0];
+ } while (now_high != cyclone_timer[1]);
+ last_offset = ((unsigned long long)last_cyclone_high<<32)|last_cyclone_low;
+ this_offset = ((unsigned long long)now_high<<32)|now_low;
+
+ ret = monotonic_base + ((this_offset - last_offset)&CYCLONE_TIMER_MASK);
+ ret = ret * (1000000000 / CYCLONE_TIMER_FREQ);
+ return ret;
+}
+
static int init_cyclone(void)
{
u32* reg;
@@ -190,5 +220,6 @@
.init = init_cyclone,
.mark_offset = mark_offset_cyclone,
.get_offset = get_offset_cyclone,
+ .monotonic_clock = monotonic_clock_cyclone,
.delay = delay_cyclone,
};
diff -Nru a/arch/i386/kernel/timers/timer_none.c b/arch/i386/kernel/timers/timer_none.c
-- a/arch/i386/kernel/timers/timer_none.c Fri Mar 28 14:31:18 2003
+++ b/arch/i386/kernel/timers/timer_none.c Fri Mar 28 14:31:18 2003
@@ -15,6 +15,11 @@
return 0;
}
+static unsigned long long monotonic_clock_none(void)
+{
+ return 0;
+}
+
static void delay_none(unsigned long loops)
{
int d0;
@@ -33,5 +38,6 @@
.init = init_none,
.mark_offset = mark_offset_none,
.get_offset = get_offset_none,
+ .monotonic_clock = monotonic_clock_none,
.delay = delay_none,
};
diff -Nru a/arch/i386/kernel/timers/timer_pit.c b/arch/i386/kernel/timers/timer_pit.c
-- a/arch/i386/kernel/timers/timer_pit.c Fri Mar 28 14:31:18 2003
+++ b/arch/i386/kernel/timers/timer_pit.c Fri Mar 28 14:31:18 2003
@@ -27,6 +27,11 @@
/* nothing needed */
}
+static unsigned long long monotonic_clock_pit(void)
+{
+ return 0;
+}
+
static void delay_pit(unsigned long loops)
{
int d0;
@@ -141,5 +146,6 @@
.init = init_pit,
.mark_offset = mark_offset_pit,
.get_offset = get_offset_pit,
+ .monotonic_clock = monotonic_clock_pit,
.delay = delay_pit,
};
diff -Nru a/arch/i386/kernel/timers/timer_tsc.c b/arch/i386/kernel/timers/timer_tsc.c
-- a/arch/i386/kernel/timers/timer_tsc.c Fri Mar 28 14:31:18 2003
+++ b/arch/i386/kernel/timers/timer_tsc.c Fri Mar 28 14:31:18 2003
@@ -23,6 +23,38 @@
static int delay_at_last_interrupt;
static unsigned long last_tsc_low; /* lsb 32 bits of Time Stamp Counter */
+static unsigned long last_tsc_high; /* msb 32 bits of Time Stamp Counter */
+static unsigned long long monotonic_base;
+
+
+/* convert from cycles(64bits) => nanoseconds (64bits)
+ * basic equation:
+ * ns = cycles / (freq / ns_per_sec)
+ * ns = cycles * (ns_per_sec / freq)
+ * ns = cycles * (10^9 / (cpu_mhz * 10^6))
+ * ns = cycles * (10^3 / cpu_mhz)
+ *
+ * Then we use scaling math (suggested by [email protected]) to get:
+ * ns = cycles * (10^3 * SC / cpu_mhz) / SC
+ * ns = cycles * cyc2ns_scale / SC
+ *
+ * And since SC is a constant power of two, we can convert the div
+ * into a shift.
+ * [email protected] "math is hard, lets go shopping!"
+ */
+static unsigned long cyc2ns_scale;
+#define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
+
+static inline void set_cyc2ns_scale(unsigned long cpu_mhz)
+{
+ cyc2ns_scale = (1000 << CYC2NS_SCALE_FACTOR)/cpu_mhz;
+}
+
+static inline unsigned long long cycles_2_ns(unsigned long long cyc)
+{
+ return (cyc * cyc2ns_scale) >> CYC2NS_SCALE_FACTOR;
+}
+
/* Cached *multiplier* to convert TSC counts to microseconds.
* (see the equation below).
@@ -60,11 +92,25 @@
return delay_at_last_interrupt + edx;
}
+static unsigned long long monotonic_clock_tsc(void)
+{
+ unsigned long long last_offset, this_offset;
+ last_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
+
+ /* Read the Time Stamp Counter */
+ rdtscll(this_offset);
+
+ /* return the value in ns */
+ return monotonic_base + cycles_2_ns(this_offset - last_offset);
+}
+
static void mark_offset_tsc(void)
{
int count;
int countmp;
static int count1=0, count2=LATCH;
+ unsigned long long this_offset, last_offset;
+ last_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
/*
* It is important that these two operations happen almost at
* the same time. We do the RDTSC stuff first, since it's
@@ -79,7 +125,7 @@
/* read Pentium cycle counter */
- rdtscl(last_tsc_low);
+ rdtsc(last_tsc_low, last_tsc_high);
spin_lock(&i8253_lock);
outb_p(0x00, 0x43); /* latch the count ASAP */
@@ -104,6 +150,11 @@
count = ((LATCH-1) - count) * TICK_SIZE;
delay_at_last_interrupt = (count + LATCH/2) / LATCH;
+
+ /* update the monotonic base value */
+ this_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
+ monotonic_base += cycles_2_ns(this_offset - last_offset);
+
}
static void delay_tsc(unsigned long loops)
@@ -295,6 +346,7 @@
"0" (eax), "1" (edx));
printk("Detected %lu.%03lu MHz processor.\n", cpu_khz / 1000, cpu_khz % 1000);
}
+ set_cyc2ns_scale(cpu_khz/1000);
return 0;
}
}
@@ -328,5 +380,6 @@
.init = init_tsc,
.mark_offset = mark_offset_tsc,
.get_offset = get_offset_tsc,
+ .monotonic_clock = monotonic_clock_tsc,
.delay = delay_tsc,
};
diff -Nru a/drivers/char/hangcheck-timer.c b/drivers/char/hangcheck-timer.c
-- a/drivers/char/hangcheck-timer.c Fri Mar 28 14:31:18 2003
+++ b/drivers/char/hangcheck-timer.c Fri Mar 28 14:31:18 2003
@@ -78,11 +78,13 @@
static struct timer_list hangcheck_ticktock =
TIMER_INITIALIZER(hangcheck_fire, 0, 0);
+extern unsigned long long monotonic_clock(void);
+
static void hangcheck_fire(unsigned long data)
{
unsigned long long cur_tsc, tsc_diff;
- cur_tsc = get_cycles();
+ cur_tsc = monotonic_clock();
if (cur_tsc > hangcheck_tsc)
tsc_diff = cur_tsc - hangcheck_tsc;
@@ -98,7 +100,7 @@
}
}
mod_timer(&hangcheck_ticktock, jiffies + (hangcheck_tick*HZ));
- hangcheck_tsc = get_cycles();
+ hangcheck_tsc = monotonic_clock();
}
@@ -108,10 +110,10 @@
VERSION_STR, hangcheck_tick, hangcheck_margin);
hangcheck_tsc_margin = hangcheck_margin + hangcheck_tick;
- hangcheck_tsc_margin *= HZ;
- hangcheck_tsc_margin *= current_cpu_data.loops_per_jiffy;
+ hangcheck_tsc_margin *= 1000000000;
+
- hangcheck_tsc = get_cycles();
+ hangcheck_tsc = monotonic_clock();
mod_timer(&hangcheck_ticktock, jiffies + (hangcheck_tick*HZ));
return 0;
diff -Nru a/include/asm-i386/timer.h b/include/asm-i386/timer.h
-- a/include/asm-i386/timer.h Fri Mar 28 14:31:18 2003
+++ b/include/asm-i386/timer.h Fri Mar 28 14:31:18 2003
@@ -14,6 +14,7 @@
int (*init)(void);
void (*mark_offset)(void);
unsigned long (*get_offset)(void);
+ unsigned long long (*monotonic_clock)(void);
void (*delay)(unsigned long);
};
john stultz <[email protected]> writes:
> + do {
> + seq = read_seqbegin(&xtime_lock);
> + ret = timer->monotonic_clock();
> + } while (read_seqretry(&xtime_lock, seq));
Why does it need to check xtime lock ? xtime should be independent
of the monotonic time because it can be changed.
Also doing seqlocks around hardware register reads is quite nasty,
because a hardware register read can be hundreds of cycles and you're
very likely to get retries. If you really need a seqlock I would
move it into the low level function and do it after the hardware access.
But as far as I can see it can be just removed.
-Andi
On Fri, 2003-03-28 at 23:57, Andi Kleen wrote:
> john stultz <[email protected]> writes:
>
> > + do {
> > + seq = read_seqbegin(&xtime_lock);
> > + ret = timer->monotonic_clock();
> > + } while (read_seqretry(&xtime_lock, seq));
>
> Why does it need to check xtime lock ? xtime should be independent
> of the monotonic time because it can be changed.
Ok, fair enough. I was using the xtime lock to protect monotonic_base in
(updated in mark_offset()), but since that isn't obvious I should be
more explicit and use a different lock.
> Also doing seqlocks around hardware register reads is quite nasty,
> because a hardware register read can be hundreds of cycles and you're
> very likely to get retries. If you really need a seqlock I would
> move it into the low level function and do it after the hardware access.
> But as far as I can see it can be just removed.
I should be able to use a regular rwlock w/o troubles. I change that and
resubmit.
thanks for the feedback!
-john