sched/clock, x86: Rewrite cyc2ns() to avoid the need to disable IRQs
Use a ring-buffer like multi-version object structure which allows
always having a coherent object; we use this to avoid having to
disable IRQs while reading sched_clock() and avoids a problem when
getting an NMI while changing the cyc2ns data.
MAINLINE PRE POST
sched_clock_stable: 1 1 1
(cold) sched_clock: 329841 331312 257223
(cold) local_clock: 301773 310296 309889
(warm) sched_clock: 38375 38247 25280
(warm) local_clock: 100371 102713 85268
(warm) rdtsc: 27340 27289 24247
sched_clock_stable: 0 0 0
(cold) sched_clock: 382634 372706 301224
(cold) local_clock: 396890 399275 399870
(warm) sched_clock: 38194 38124 25630
(warm) local_clock: 143452 148698 129629
(warm) rdtsc: 27345 27365 24307
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Link: http://lkml.kernel.org/n/tip-s567in1e5ekq2nlyhn8f987r@git.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
diff --git a/arch/x86/kernel/tsc.c b/arch/x86/kernel/tsc.c
index b4a04ac..92b090b 100644
--- a/arch/x86/kernel/tsc.c
+++ b/arch/x86/kernel/tsc.c
@@ -39,7 +39,119 @@
int tsc_clocksource_reliable;
-/* Accelerators for sched_clock()
+/*
+ * Use a ring-buffer like data structure, where a writer advances the head by
+ * writing a new data entry and a reader advances the tail when it observes a
+ * new entry.
+ *
+ * Writers are made to wait on readers until there's space to write a new
+ * entry.
+ *
+ * This means that we can always use an {offset, mul} pair to compute a ns
+ * value that is 'roughly' in the right direction, even if we're writing a new
+ * {offset, mul} pair during the clock read.
+ *
+ * The down-side is that we can no longer guarantee strict monotonicity anymore
+ * (assuming the TSC was that to begin with), because while we compute the
+ * intersection point of the two clock slopes and make sure the time is
+ * continuous at the point of switching; we can no longer guarantee a reader is
+ * strictly before or after the switch point.
+ *
+ * It does mean a reader no longer needs to disable IRQs in order to avoid
+ * CPU-Freq updates messing with his times, and similarly an NMI reader will
+ * no longer run the risk of hitting half-written state.
+ */
+
+struct cyc2ns {
+ struct cyc2ns_data data[2]; /* 0 + 2*24 = 48 */
+ struct cyc2ns_data *head; /* 48 + 8 = 56 */
+ struct cyc2ns_data *tail; /* 56 + 8 = 64 */
+}; /* exactly fits one cacheline */
+
+static DEFINE_PER_CPU_ALIGNED(struct cyc2ns, cyc2ns);
+
+struct cyc2ns_data *cyc2ns_read_begin(void)
+{
+ struct cyc2ns_data *head;
+
+ preempt_disable();
+
+ head = this_cpu_read(cyc2ns.head);
+ /*
+ * Ensure we observe the entry when we observe the pointer to it.
+ * matches the wmb from cyc2ns_write_end().
+ */
+ smp_read_barrier_depends();
+ head->__count++;
+ barrier();
+
+ return head;
+}
+
+void cyc2ns_read_end(struct cyc2ns_data *head)
+{
+ barrier();
+ /*
+ * If we're the outer most nested read; update the tail pointer
+ * when we're done. This notifies possible pending writers
+ * that we've observed the head pointer and that the other
+ * entry is now free.
+ */
+ if (!--head->__count) {
+ /*
+ * x86-TSO does not reorder writes with older reads;
+ * therefore once this write becomes visible to another
+ * cpu, we must be finished reading the cyc2ns_data.
+ *
+ * matches with cyc2ns_write_begin().
+ */
+ this_cpu_write(cyc2ns.tail, head);
+ }
+ preempt_enable();
+}
+
+/*
+ * Begin writing a new @data entry for @cpu.
+ *
+ * Assumes some sort of write side lock; currently 'provided' by the assumption
+ * that cpufreq will call its notifiers sequentially.
+ */
+static struct cyc2ns_data *cyc2ns_write_begin(int cpu)
+{
+ struct cyc2ns *c2n = &per_cpu(cyc2ns, cpu);
+ struct cyc2ns_data *data = c2n->data;
+
+ if (data == c2n->head)
+ data++;
+
+ /* XXX send an IPI to @cpu in order to guarantee a read? */
+
+ /*
+ * When we observe the tail write from cyc2ns_read_end(),
+ * the cpu must be done with that entry and its safe
+ * to start writing to it.
+ */
+ while (c2n->tail == data)
+ cpu_relax();
+
+ return data;
+}
+
+static void cyc2ns_write_end(int cpu, struct cyc2ns_data *data)
+{
+ struct cyc2ns *c2n = &per_cpu(cyc2ns, cpu);
+
+ /*
+ * Ensure the @data writes are visible before we publish the
+ * entry. Matches the data-depencency in cyc2ns_read_begin().
+ */
+ smp_wmb();
+
+ ACCESS_ONCE(c2n->head) = data;
+}
+
+/*
+ * Accelerators for sched_clock()
* convert from cycles(64bits) => nanoseconds (64bits)
* basic equation:
* ns = cycles / (freq / ns_per_sec)
@@ -61,49 +173,106 @@
* -johnstul@us.ibm.com "math is hard, lets go shopping!"
*/
-DEFINE_PER_CPU(unsigned long, cyc2ns);
-DEFINE_PER_CPU(unsigned long long, cyc2ns_offset);
-
#define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
+static void cyc2ns_data_init(struct cyc2ns_data *data)
+{
+ data->cyc2ns_mul = 1U << CYC2NS_SCALE_FACTOR;
+ data->cyc2ns_shift = CYC2NS_SCALE_FACTOR;
+ data->cyc2ns_offset = 0;
+ data->__count = 0;
+}
+
+static void cyc2ns_init(int cpu)
+{
+ struct cyc2ns *c2n = &per_cpu(cyc2ns, cpu);
+
+ cyc2ns_data_init(&c2n->data[0]);
+ cyc2ns_data_init(&c2n->data[1]);
+
+ c2n->head = c2n->data;
+ c2n->tail = c2n->data;
+}
+
static inline unsigned long long cycles_2_ns(unsigned long long cyc)
{
- unsigned long long ns = this_cpu_read(cyc2ns_offset);
- ns += mul_u64_u32_shr(cyc, this_cpu_read(cyc2ns), CYC2NS_SCALE_FACTOR);
+ struct cyc2ns_data *data, *tail;
+ unsigned long long ns;
+
+ /*
+ * See cyc2ns_read_*() for details; replicated in order to avoid
+ * an extra few instructions that came with the abstraction.
+ * Notable, it allows us to only do the __count and tail update
+ * dance when its actually needed.
+ */
+
+ preempt_disable();
+ data = this_cpu_read(cyc2ns.head);
+ tail = this_cpu_read(cyc2ns.tail);
+
+ if (likely(data == tail)) {
+ ns = data->cyc2ns_offset;
+ ns += mul_u64_u32_shr(cyc, data->cyc2ns_mul, CYC2NS_SCALE_FACTOR);
+ } else {
+ data->__count++;
+
+ barrier();
+
+ ns = data->cyc2ns_offset;
+ ns += mul_u64_u32_shr(cyc, data->cyc2ns_mul, CYC2NS_SCALE_FACTOR);
+
+ barrier();
+
+ if (!--data->__count)
+ this_cpu_write(cyc2ns.tail, data);
+ }
+ preempt_enable();
+
return ns;
}
+/* XXX surely we already have this someplace in the kernel?! */
+#define DIV_ROUND(n, d) (((n) + ((d) / 2)) / (d))
+
static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
{
- unsigned long long tsc_now, ns_now, *offset;
- unsigned long flags, *scale;
+ unsigned long long tsc_now, ns_now;
+ struct cyc2ns_data *data;
+ unsigned long flags;
local_irq_save(flags);
sched_clock_idle_sleep_event();
- scale = &per_cpu(cyc2ns, cpu);
- offset = &per_cpu(cyc2ns_offset, cpu);
+ if (!cpu_khz)
+ goto done;
+
+ data = cyc2ns_write_begin(cpu);
rdtscll(tsc_now);
ns_now = cycles_2_ns(tsc_now);
- if (cpu_khz) {
- *scale = ((NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR) +
- cpu_khz / 2) / cpu_khz;
- *offset = ns_now - mult_frac(tsc_now, *scale,
- (1UL << CYC2NS_SCALE_FACTOR));
- }
+ /*
+ * Compute a new multiplier as per the above comment and ensure our
+ * time function is continuous; see the comment near struct
+ * cyc2ns_data.
+ */
+ data->cyc2ns_mul = DIV_ROUND(NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR, cpu_khz);
+ data->cyc2ns_shift = CYC2NS_SCALE_FACTOR;
+ data->cyc2ns_offset = ns_now -
+ mul_u64_u32_shr(tsc_now, data->cyc2ns_mul, CYC2NS_SCALE_FACTOR);
+ cyc2ns_write_end(cpu, data);
+
+done:
sched_clock_idle_wakeup_event(0);
local_irq_restore(flags);
}
-
/*
* Scheduler clock - returns current time in nanosec units.
*/
u64 native_sched_clock(void)
{
- u64 this_offset;
+ u64 tsc_now;
/*
* Fall back to jiffies if there's no TSC available:
@@ -119,10 +288,10 @@
}
/* read the Time Stamp Counter: */
- rdtscll(this_offset);
+ rdtscll(tsc_now);
/* return the value in ns */
- return cycles_2_ns(this_offset);
+ return cycles_2_ns(tsc_now);
}
/* We need to define a real function for sched_clock, to override the
@@ -678,11 +847,21 @@
local_irq_save(flags);
- __this_cpu_write(cyc2ns_offset, 0);
+ /*
+ * We're comming out of suspend, there's no concurrency yet; don't
+ * bother being nice about the RCU stuff, just write to both
+ * data fields.
+ */
+
+ this_cpu_write(cyc2ns.data[0].cyc2ns_offset, 0);
+ this_cpu_write(cyc2ns.data[1].cyc2ns_offset, 0);
+
offset = cyc2ns_suspend - sched_clock();
- for_each_possible_cpu(cpu)
- per_cpu(cyc2ns_offset, cpu) = offset;
+ for_each_possible_cpu(cpu) {
+ per_cpu(cyc2ns.data[0].cyc2ns_offset, cpu) = offset;
+ per_cpu(cyc2ns.data[1].cyc2ns_offset, cpu) = offset;
+ }
local_irq_restore(flags);
}
@@ -1005,8 +1184,10 @@
* speed as the bootup CPU. (cpufreq notifiers will fix this
* up if their speed diverges)
*/
- for_each_possible_cpu(cpu)
+ for_each_possible_cpu(cpu) {
+ cyc2ns_init(cpu);
set_cyc2ns_scale(cpu_khz, cpu);
+ }
if (tsc_disabled > 0)
return;