Monotonic clock code uses reader/writer lock which is prone to same
starvation problems as we saw with xtime. This patch changes it to seq_lock
which is faster and won't starve writers in face of lots of readers.
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 Mon Sep 15 15:48:01 2003
+++ b/arch/i386/kernel/timers/timer_tsc.c Mon Sep 15 15:48:01 2003
@@ -39,7 +39,7 @@
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;
-static rwlock_t monotonic_lock = RW_LOCK_UNLOCKED;
+static seqlock_t monotonic_lock = SEQLOCK_UNLOCKED;
/* convert from cycles(64bits) => nanoseconds (64bits)
* basic equation:
@@ -111,12 +111,14 @@
static unsigned long long monotonic_clock_tsc(void)
{
unsigned long long last_offset, this_offset, base;
+ unsigned seq;
/* atomically read monotonic base & last_offset */
- read_lock_irq(&monotonic_lock);
- last_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
- base = monotonic_base;
- read_unlock_irq(&monotonic_lock);
+ do {
+ seq = read_seqbegin(&monotonic_lock);
+ last_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
+ base = monotonic_base;
+ } while (read_seqretry(&monotonic_lock, seq));
/* Read the Time Stamp Counter */
rdtscll(this_offset);
@@ -135,7 +137,7 @@
unsigned long long this_offset, last_offset;
static int lost_count = 0;
- write_lock(&monotonic_lock);
+ write_seqlock(&monotonic_lock);
last_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
/*
* It is important that these two operations happen almost at
@@ -204,7 +206,7 @@
/* 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);
- write_unlock(&monotonic_lock);
+ write_sequnlock(&monotonic_lock);
/* calculate delay_at_last_interrupt */
count = ((LATCH-1) - count) * TICK_SIZE;
@@ -236,7 +238,7 @@
unsigned long long this_offset, last_offset;
unsigned long offset, temp, hpet_current;
- write_lock(&monotonic_lock);
+ write_seqlock(&monotonic_lock);
last_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
/*
* It is important that these two operations happen almost at
@@ -264,7 +266,7 @@
/* 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);
- write_unlock(&monotonic_lock);
+ write_sequnlock(&monotonic_lock);
/* calculate delay_at_last_interrupt */
/*
On Tue, 16 Sep 2003 11:59:35 -0700
Andrew Morton <[email protected]> wrote:
>
> So timer_cyclone and timer_hpet need the same change?
Yes.
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 Tue Sep 16 13:59:54 2003
+++ b/arch/i386/kernel/timers/timer_cyclone.c Tue Sep 16 13:59:54 2003
@@ -35,7 +35,7 @@
static u32 last_cyclone_low;
static u32 last_cyclone_high;
static unsigned long long monotonic_base;
-static rwlock_t monotonic_lock = RW_LOCK_UNLOCKED;
+static seqlock_t monotonic_lock = SEQLOCK_UNLOCKED;
/* helper macro to atomically read both cyclone counter registers */
#define read_cyclone_counter(low,high) \
@@ -51,7 +51,7 @@
int count;
unsigned long long this_offset, last_offset;
- write_lock(&monotonic_lock);
+ write_seqlock(&monotonic_lock);
last_offset = ((unsigned long long)last_cyclone_high<<32)|last_cyclone_low;
spin_lock(&i8253_lock);
@@ -76,7 +76,7 @@
/* 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;
- write_unlock(&monotonic_lock);
+ write_sequnlock(&monotonic_lock);
/* calculate delay_at_last_interrupt */
count = ((LATCH-1) - count) * TICK_SIZE;
@@ -117,12 +117,15 @@
u32 now_low, now_high;
unsigned long long last_offset, this_offset, base;
unsigned long long ret;
+ unsigned seq;
/* atomically read monotonic base & last_offset */
- read_lock_irq(&monotonic_lock);
- last_offset = ((unsigned long long)last_cyclone_high<<32)|last_cyclone_low;
- base = monotonic_base;
- read_unlock_irq(&monotonic_lock);
+ do {
+ seq = read_seqbegin(&monotonic_lock);
+ last_offset = ((unsigned long long)last_cyclone_high<<32)|last_cyclone_low;
+ base = monotonic_base;
+ } while (read_seqretry(&monotonic_lock, seq));
+
/* Read the cyclone counter */
read_cyclone_counter(now_low,now_high);
diff -Nru a/arch/i386/kernel/timers/timer_hpet.c b/arch/i386/kernel/timers/timer_hpet.c
--- a/arch/i386/kernel/timers/timer_hpet.c Tue Sep 16 13:59:54 2003
+++ b/arch/i386/kernel/timers/timer_hpet.c Tue Sep 16 13:59:54 2003
@@ -24,7 +24,7 @@
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;
-static rwlock_t monotonic_lock = RW_LOCK_UNLOCKED;
+static seqlock_t monotonic_lock = SEQLOCK_UNLOCKED;
/* convert from cycles(64bits) => nanoseconds (64bits)
* basic equation:
@@ -57,12 +57,14 @@
static unsigned long long monotonic_clock_hpet(void)
{
unsigned long long last_offset, this_offset, base;
+ unsigned seq;
/* atomically read monotonic base & last_offset */
- read_lock_irq(&monotonic_lock);
- last_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
- base = monotonic_base;
- read_unlock_irq(&monotonic_lock);
+ do {
+ seq = read_seqbegin(&monotonic_lock);
+ last_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
+ base = monotonic_base;
+ } while (read_seqretry(&monotonic_lock, seq));
/* Read the Time Stamp Counter */
rdtscll(this_offset);
@@ -99,7 +101,7 @@
unsigned long long this_offset, last_offset;
unsigned long offset;
- write_lock(&monotonic_lock);
+ write_seqlock(&monotonic_lock);
last_offset = ((unsigned long long)last_tsc_high<<32)|last_tsc_low;
rdtsc(last_tsc_low, last_tsc_high);
@@ -113,7 +115,7 @@
/* 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);
- write_unlock(&monotonic_lock);
+ write_sequnlock(&monotonic_lock);
}
void delay_hpet(unsigned long loops)
On Tue, 2003-09-16 at 14:00, Stephen Hemminger wrote:
> On Tue, 16 Sep 2003 11:59:35 -0700
> Andrew Morton <[email protected]> wrote:
>
> >
> > So timer_cyclone and timer_hpet need the same change?
>
> Yes.
The cyclone bits match those I was testing to send.
Looks good.
thanks
-john