[PATCH v3 4/4] sched_clock: Avoid deadlock during read from NMI

[Daniel Thompson <daniel.thompson@linaro.org>]

Currently it is possible for an NMI (or FIQ on ARM) to come in and
read sched_clock() whilst update_sched_clock() has locked the seqcount
for writing. This results in the NMI handler locking up when it calls
raw_read_seqcount_begin().

This patch fixes the NMI safety issues by providing banked clock data.
This is a similar approach to the one used in Thomas Gleixner's
4396e058c52e("timekeeping: Provide fast and NMI safe access to
CLOCK_MONOTONIC").

Suggested-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org>
Cc: Russell King <linux@arm.linux.org.uk>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
---
 kernel/time/sched_clock.c | 91 ++++++++++++++++++++++++++++++-----------------
 1 file changed, 58 insertions(+), 33 deletions(-)

diff --git a/kernel/time/sched_clock.c b/kernel/time/sched_clock.c
index 638c765131fa..2b1a465f1c00 100644
--- a/kernel/time/sched_clock.c
+++ b/kernel/time/sched_clock.c
@@ -40,12 +40,12 @@ struct clock_read_data {
  *                     registration of a new clock source)
  *
  * The ordering of this structure has been chosen to optimize cache
- * performance. In particular seq and read_data (combined) should fit
+ * performance. In particular seq and read_data[0] (combined) should fit
  * into a single 64 byte cache line.
  */
 struct clock_data {
 	seqcount_t seq;
-	struct clock_read_data read_data;
+	struct clock_read_data read_data[2];
 	ktime_t wrap_kt;
 	unsigned long rate;
 	u64 (*actual_read_sched_clock)(void);
@@ -66,10 +66,9 @@ static u64 notrace jiffy_sched_clock_read(void)
 }
 
 static struct clock_data cd ____cacheline_aligned = {
-	.read_data = { .mult = NSEC_PER_SEC / HZ,
-		       .read_sched_clock = jiffy_sched_clock_read, },
+	.read_data[0] = { .mult = NSEC_PER_SEC / HZ,
+			  .read_sched_clock = jiffy_sched_clock_read, },
 	.actual_read_sched_clock = jiffy_sched_clock_read,
-
 };
 
 static inline u64 notrace cyc_to_ns(u64 cyc, u32 mult, u32 shift)
@@ -81,10 +80,11 @@ unsigned long long notrace sched_clock(void)
 {
 	u64 cyc, res;
 	unsigned long seq;
-	struct clock_read_data *rd = &cd.read_data;
+	struct clock_read_data *rd;
 
 	do {
-		seq = raw_read_seqcount_begin(&cd.seq);
+		seq = raw_read_seqcount(&cd.seq);
+		rd = cd.read_data + (seq & 1);
 
 		res = rd->epoch_ns;
 		if (rd->read_sched_clock) {
@@ -98,26 +98,50 @@ unsigned long long notrace sched_clock(void)
 }
 
 /*
+ * Updating the data required to read the clock.
+ *
+ * sched_clock will never observe mis-matched data even if called from
+ * an NMI. We do this by maintaining an odd/even copy of the data and
+ * steering sched_clock to one or the other using a sequence counter.
+ * In order to preserve the data cache profile of sched_clock as much
+ * as possible the system reverts back to the even copy when the update
+ * completes; the odd copy is used *only* during an update.
+ */
+static void update_clock_read_data(struct clock_read_data *rd)
+{
+	/* update the backup (odd) copy with the new data */
+	cd.read_data[1] = *rd;
+
+	/* steer readers towards the new data */
+	raw_write_seqcount_latch(&cd.seq);
+
+	/* now its safe for us to update the normal (even) copy */
+	cd.read_data[0] = *rd;
+
+	/* switch readers back to the even copy */
+	raw_write_seqcount_latch(&cd.seq);
+}
+
+/*
  * Atomically update the sched_clock epoch.
  */
 static void notrace update_sched_clock(void)
 {
-	unsigned long flags;
 	u64 cyc;
 	u64 ns;
-	struct clock_read_data *rd = &cd.read_data;
+	struct clock_read_data rd;
+
+	rd = cd.read_data[0];
 
 	cyc = cd.actual_read_sched_clock();
-	ns = rd->epoch_ns +
-	     cyc_to_ns((cyc - rd->epoch_cyc) & rd->sched_clock_mask,
-		       rd->mult, rd->shift);
-
-	raw_local_irq_save(flags);
-	raw_write_seqcount_begin(&cd.seq);
-	rd->epoch_ns = ns;
-	rd->epoch_cyc = cyc;
-	raw_write_seqcount_end(&cd.seq);
-	raw_local_irq_restore(flags);
+	ns = rd.epoch_ns +
+	     cyc_to_ns((cyc - rd.epoch_cyc) & rd.sched_clock_mask,
+		       rd.mult, rd.shift);
+
+	rd.epoch_ns = ns;
+	rd.epoch_cyc = cyc;
+
+	update_clock_read_data(&rd);
 }
 
 static enum hrtimer_restart sched_clock_poll(struct hrtimer *hrt)
@@ -134,7 +158,7 @@ void __init sched_clock_register(u64 (*read)(void), int bits,
 	u32 new_mult, new_shift;
 	unsigned long r;
 	char r_unit;
-	struct clock_read_data *rd = &cd.read_data;
+	struct clock_read_data rd;
 
 	if (cd.rate > rate)
 		return;
@@ -151,22 +175,23 @@ void __init sched_clock_register(u64 (*read)(void), int bits,
 	wrap = clocks_calc_max_nsecs(new_mult, new_shift, 0, new_mask);
 	cd.wrap_kt = ns_to_ktime(wrap - (wrap >> 3));
 
+	rd = cd.read_data[0];
+
 	/* update epoch for new counter and update epoch_ns from old counter*/
 	new_epoch = read();
 	cyc = cd.actual_read_sched_clock();
-	ns = rd->epoch_ns +
-	     cyc_to_ns((cyc - rd->epoch_cyc) & rd->sched_clock_mask,
-		       rd->mult, rd->shift);
+	ns = rd.epoch_ns +
+	     cyc_to_ns((cyc - rd.epoch_cyc) & rd.sched_clock_mask,
+		       rd.mult, rd.shift);
 	cd.actual_read_sched_clock = read;
 
-	raw_write_seqcount_begin(&cd.seq);
-	rd->read_sched_clock = read;
-	rd->sched_clock_mask = new_mask;
-	rd->mult = new_mult;
-	rd->shift = new_shift;
-	rd->epoch_cyc = new_epoch;
-	rd->epoch_ns = ns;
-	raw_write_seqcount_end(&cd.seq);
+	rd.read_sched_clock = read;
+	rd.sched_clock_mask = new_mask;
+	rd.mult = new_mult;
+	rd.shift = new_shift;
+	rd.epoch_cyc = new_epoch;
+	rd.epoch_ns = ns;
+	update_clock_read_data(&rd);
 
 	r = rate;
 	if (r >= 4000000) {
@@ -213,7 +238,7 @@ void __init sched_clock_postinit(void)
 
 static int sched_clock_suspend(void)
 {
-	struct clock_read_data *rd = &cd.read_data;
+	struct clock_read_data *rd = &cd.read_data[0];
 
 	update_sched_clock();
 	hrtimer_cancel(&sched_clock_timer);
@@ -223,7 +248,7 @@ static int sched_clock_suspend(void)
 
 static void sched_clock_resume(void)
 {
-	struct clock_read_data *rd = &cd.read_data;
+	struct clock_read_data *rd = &cd.read_data[0];
 
 	rd->epoch_cyc = cd.actual_read_sched_clock();
 	hrtimer_start(&sched_clock_timer, cd.wrap_kt, HRTIMER_MODE_REL);
-- 
1.9.3