Skip to contentSkip to conversation

Generic sched_clock locking fix for tip/timers/urgent

PATCH New John Stultz 3 participants 6 messages Feb 17, 2014 – Feb 19, 2014

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Generic sched_clock locking fix for tip/timers/urgent

Hey Ingo, Thomas,
Wanted to send along this sched_clock fix from Stephen for
tip/timers/urgent. Let me know if you have any issues or concerns.

thanks
-john

Cc Peter Zijlstra <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Stephen Boyd <[email protected]>
Cc Will Deacon <[email protected]>
Cc Josh Cartwright <[email protected]>

Stephen Boyd (1):
sched_clock: Prevent callers from seeing half-updated data

 kernel/time/sched_clock.c | 46 +++++++++++++++++++++++++++++-----------------
 1 file changed, 29 insertions(+), 17 deletions(-)
Signature · 1 line
1.8.3.2
#1 lore raw

PATCH 1/1 sched_clock: Prevent callers from seeing half-updated data

From: Stephen Boyd <[email protected]>

The generic sched_clock registration function was previously done lockless, due to the fact that it was expected to be called only once. However, now there are systems that may register multiple sched_clock sources, for which the lack of locking has casued problems:

If two sched_clock sources are registered we may end up in a situation where a call to sched_clock() may be accessing the epoch cycle count for the old counter and the cycle count for the new counter. This can lead to confusing results where sched_clock() values jump and then are reset to 0 (due to the way the registration function forces the epoch_ns to be 0).

Fix this by reorganizing the registration function to hold the seqlock for as short a time as possible while we update the clock_data structure for a new counter. We also put any accumulated time into epoch_ns instead of resetting the time to 0 so that the clock doesn't reset after each successful registration.

Cc Peter Zijlstra <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Stephen Boyd <[email protected]>
Cc Will Deacon <[email protected]>
Cc Josh Cartwright <[email protected]>
[jstultz: Added extra context to the commit message]
---
 kernel/time/sched_clock.c | 46 +++++++++++++++++++++++++++++-----------------
 1 file changed, 29 insertions(+), 17 deletions(-)
kernel/time/sched_clock.c +29 -17
index 0abb364..4d23dc4 100644
@@ -116,20 +116,42 @@ static enum hrtimer_restart sched_clock_poll(struct hrtimer *hrt)
 void __init sched_clock_register(u64 (*read)(void), int bits,
 				 unsigned long rate)
 {
+	u64 res, wrap, new_mask, new_epoch, cyc, ns;
+	u32 new_mult, new_shift;
+	ktime_t new_wrap_kt;
 	unsigned long r;
-	u64 res, wrap;
 	char r_unit;
 
 	if (cd.rate > rate)
 		return;
 
 	WARN_ON(!irqs_disabled());
-	read_sched_clock = read;
-	sched_clock_mask = CLOCKSOURCE_MASK(bits);
-	cd.rate = rate;
 
 	/* calculate the mult/shift to convert counter ticks to ns. */
-	clocks_calc_mult_shift(&cd.mult, &cd.shift, rate, NSEC_PER_SEC, 3600);
+	clocks_calc_mult_shift(&new_mult, &new_shift, rate, NSEC_PER_SEC, 3600);
+
+	new_mask = CLOCKSOURCE_MASK(bits);
+
+	/* calculate how many ns until we wrap */
+	wrap = clocks_calc_max_nsecs(new_mult, new_shift, 0, new_mask);
+	new_wrap_kt = ns_to_ktime(wrap - (wrap >> 3));
+
+	/* update epoch for new counter and update epoch_ns from old counter*/
+	new_epoch = read();
+	cyc = read_sched_clock();
+	ns = cd.epoch_ns + cyc_to_ns((cyc - cd.epoch_cyc) & sched_clock_mask,
+			  cd.mult, cd.shift);
+
+	raw_write_seqcount_begin(&cd.seq);
+	read_sched_clock = read;
+	sched_clock_mask = new_mask;
+	cd.rate = rate;
+	cd.wrap_kt = new_wrap_kt;
+	cd.mult = new_mult;
+	cd.shift = new_shift;
+	cd.epoch_cyc = new_epoch;
+	cd.epoch_ns = ns;
+	raw_write_seqcount_end(&cd.seq);
 
 	r = rate;
 	if (r >= 4000000) {
@@ -141,22 +163,12 @@ void __init sched_clock_register(u64 (*read)(void), int bits,
 	} else
 		r_unit = ' ';
 
-	/* calculate how many ns until we wrap */
-	wrap = clocks_calc_max_nsecs(cd.mult, cd.shift, 0, sched_clock_mask);
-	cd.wrap_kt = ns_to_ktime(wrap - (wrap >> 3));
-
 	/* calculate the ns resolution of this counter */
-	res = cyc_to_ns(1ULL, cd.mult, cd.shift);
+	res = cyc_to_ns(1ULL, new_mult, new_shift);
+
 	pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lluns\n",
 		bits, r, r_unit, res, wrap);
 
-	update_sched_clock();
-
-	/*
-	 * Ensure that sched_clock() starts off at 0ns
-	 */
-	cd.epoch_ns = 0;
-
 	/* Enable IRQ time accounting if we have a fast enough sched_clock */
 	if (irqtime > 0 || (irqtime == -1 && rate >= 1000000))
 		enable_sched_clock_irqtime();
Signature · 1 line
1.8.3.2
#2 lore raw

Re: [PATCH] sched_clock: Prevent callers from seeing half-updated data

Quoted 15 lines · John Stultz

From: Stephen Boyd <[email protected]>

The generic sched_clock registration function was previously done lockless, due to the fact that it was expected to be called only once. However, now there are systems that may register multiple sched_clock sources, for which the lack of locking has casued problems:

If two sched_clock sources are registered we may end up in a situation where a call to sched_clock() may be accessing the epoch cycle count for the old counter and the cycle count for the new counter. This can lead to confusing results where sched_clock() values jump and then are reset to 0 (due to the way the registration function forces the epoch_ns to be 0).

Quoted 8 lines

Fix this by reorganizing the registration function to hold the seqlock for as short a time as possible while we update the clock_data structure for a new counter. We also put any accumulated time into epoch_ns instead of resetting the time to 0 so that the clock doesn't reset after each successful registration.

Cc Peter Zijlstra <[email protected]>

Peter ???

Quoted 94 lines
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Stephen Boyd <[email protected]>
Cc Will Deacon <[email protected]>
Cc Josh Cartwright <[email protected]>
[jstultz: Added extra context to the commit message]
---
 kernel/time/sched_clock.c | 46 +++++++++++++++++++++++++++++-----------------
 1 file changed, 29 insertions(+), 17 deletions(-)
kernel/time/sched_clock.c +29 -17
index 0abb364..4d23dc4 100644
@@ -116,20 +116,42 @@ static enum hrtimer_restart sched_clock_poll(struct hrtimer *hrt)
 void __init sched_clock_register(u64 (*read)(void), int bits,
 				 unsigned long rate)
 {
+	u64 res, wrap, new_mask, new_epoch, cyc, ns;
+	u32 new_mult, new_shift;
+	ktime_t new_wrap_kt;
 	unsigned long r;
-	u64 res, wrap;
 	char r_unit;
 
 	if (cd.rate > rate)
 		return;
 
 	WARN_ON(!irqs_disabled());
-	read_sched_clock = read;
-	sched_clock_mask = CLOCKSOURCE_MASK(bits);
-	cd.rate = rate;
 
 	/* calculate the mult/shift to convert counter ticks to ns. */
-	clocks_calc_mult_shift(&cd.mult, &cd.shift, rate, NSEC_PER_SEC, 3600);
+	clocks_calc_mult_shift(&new_mult, &new_shift, rate, NSEC_PER_SEC, 3600);
+
+	new_mask = CLOCKSOURCE_MASK(bits);
+
+	/* calculate how many ns until we wrap */
+	wrap = clocks_calc_max_nsecs(new_mult, new_shift, 0, new_mask);
+	new_wrap_kt = ns_to_ktime(wrap - (wrap >> 3));
+
+	/* update epoch for new counter and update epoch_ns from old counter*/
+	new_epoch = read();
+	cyc = read_sched_clock();
+	ns = cd.epoch_ns + cyc_to_ns((cyc - cd.epoch_cyc) & sched_clock_mask,
+			  cd.mult, cd.shift);
+
+	raw_write_seqcount_begin(&cd.seq);
+	read_sched_clock = read;
+	sched_clock_mask = new_mask;
+	cd.rate = rate;
+	cd.wrap_kt = new_wrap_kt;
+	cd.mult = new_mult;
+	cd.shift = new_shift;
+	cd.epoch_cyc = new_epoch;
+	cd.epoch_ns = ns;
+	raw_write_seqcount_end(&cd.seq);
 
 	r = rate;
 	if (r >= 4000000) {
@@ -141,22 +163,12 @@ void __init sched_clock_register(u64 (*read)(void), int bits,
 	} else
 		r_unit = ' ';
 
-	/* calculate how many ns until we wrap */
-	wrap = clocks_calc_max_nsecs(cd.mult, cd.shift, 0, sched_clock_mask);
-	cd.wrap_kt = ns_to_ktime(wrap - (wrap >> 3));
-
 	/* calculate the ns resolution of this counter */
-	res = cyc_to_ns(1ULL, cd.mult, cd.shift);
+	res = cyc_to_ns(1ULL, new_mult, new_shift);
+
 	pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lluns\n",
 		bits, r, r_unit, res, wrap);
 
-	update_sched_clock();
-
-	/*
-	 * Ensure that sched_clock() starts off at 0ns
-	 */
-	cd.epoch_ns = 0;
-
 	/* Enable IRQ time accounting if we have a fast enough sched_clock */
 	if (irqtime > 0 || (irqtime == -1 && rate >= 1000000))
 		enable_sched_clock_irqtime();
Signature · 1 line
1.8.3.2
#3 lore raw

Re: [PATCH] sched_clock: Prevent callers from seeing half-updated data

Quoted 24 lines · Thomas Gleixner
Quoted 22 lines · John Stultz

From: Stephen Boyd <[email protected]>

The generic sched_clock registration function was previously done lockless, due to the fact that it was expected to be called only once. However, now there are systems that may register multiple sched_clock sources, for which the lack of locking has casued problems:

If two sched_clock sources are registered we may end up in a situation where a call to sched_clock() may be accessing the epoch cycle count for the old counter and the cycle count for the new counter. This can lead to confusing results where sched_clock() values jump and then are reset to 0 (due to the way the registration function forces the epoch_ns to be 0).

Fix this by reorganizing the registration function to hold the seqlock for as short a time as possible while we update the clock_data structure for a new counter. We also put any accumulated time into epoch_ns instead of resetting the time to 0 so that the clock doesn't reset after each successful registration.

Quoted 3 lines
Quoted 1 line
Cc Peter Zijlstra <[email protected]>

Peter ???

So this is the generic sched_clock work, which Peter really hasn't had much involvement with yet (mostly because its not yet generic enough to work with more then a few arches). But I included him in the CC since I think it would be good to have him following along.

thanks
-john

#4 lore raw

Re: [PATCH] sched_clock: Prevent callers from seeing half-updated data

Quoted 25 lines · John Stultz
Quoted 24 lines · Thomas Gleixner
Quoted 22 lines · John Stultz

From: Stephen Boyd <[email protected]>

The generic sched_clock registration function was previously done lockless, due to the fact that it was expected to be called only once. However, now there are systems that may register multiple sched_clock sources, for which the lack of locking has casued problems:

If two sched_clock sources are registered we may end up in a situation where a call to sched_clock() may be accessing the epoch cycle count for the old counter and the cycle count for the new counter. This can lead to confusing results where sched_clock() values jump and then are reset to 0 (due to the way the registration function forces the epoch_ns to be 0).

Fix this by reorganizing the registration function to hold the seqlock for as short a time as possible while we update the clock_data structure for a new counter. We also put any accumulated time into epoch_ns instead of resetting the time to 0 so that the clock doesn't reset after each successful registration.

Quoted 8 lines
Quoted 3 lines
Quoted 1 line
Cc Peter Zijlstra <[email protected]>

Peter ???

So this is the generic sched_clock work, which Peter really hasn't had much involvement with yet (mostly because its not yet generic enough to work with more then a few arches). But I included him in the CC since I think it would be good to have him following along.

Fair enough. I'll pick it up.

Thanks,

tglx

#5 lore raw

[tip:timers/urgent] sched_clock: Prevent callers from seeing half-updated data

In reply to John Stultz
sched_clock: Prevent callers from seeing half-updated data
Commit-ID:  5ae8aabeaec3fe69c4fb21cbe5b17b72b35b5892
Gitweb:     http://git.kernel.org/tip/5ae8aabeaec3fe69c4fb21cbe5b17b72b35b5892
Author:     Stephen Boyd <[email protected]>
AuthorDate: Mon, 17 Feb 2014 10:45:36 -0800
Committer:  Thomas Gleixner <[email protected]>
CommitDate: Wed, 19 Feb 2014 17:07:22 +0100

sched_clock: Prevent callers from seeing half-updated data

The generic sched_clock registration function was previously done lockless, due to the fact that it was expected to be called only once. However, now there are systems that may register multiple sched_clock sources, for which the lack of locking has casued problems:

If two sched_clock sources are registered we may end up in a situation where a call to sched_clock() may be accessing the epoch cycle count for the old counter and the cycle count for the new counter. This can lead to confusing results where sched_clock() values jump and then are reset to 0 (due to the way the registration function forces the epoch_ns to be 0).

Fix this by reorganizing the registration function to hold the seqlock for as short a time as possible while we update the clock_data structure for a new counter. We also put any accumulated time into epoch_ns instead of resetting the time to 0 so that the clock doesn't reset after each successful registration.

[jstultz: Added extra context to the commit message]

---
 kernel/time/sched_clock.c | 46 +++++++++++++++++++++++++++++-----------------
 1 file changed, 29 insertions(+), 17 deletions(-)
kernel/time/sched_clock.c +29 -17
index 0abb364..4d23dc4 100644
@@ -116,20 +116,42 @@ static enum hrtimer_restart sched_clock_poll(struct hrtimer *hrt)
 void __init sched_clock_register(u64 (*read)(void), int bits,
 				 unsigned long rate)
 {
+	u64 res, wrap, new_mask, new_epoch, cyc, ns;
+	u32 new_mult, new_shift;
+	ktime_t new_wrap_kt;
 	unsigned long r;
-	u64 res, wrap;
 	char r_unit;
 
 	if (cd.rate > rate)
 		return;
 
 	WARN_ON(!irqs_disabled());
-	read_sched_clock = read;
-	sched_clock_mask = CLOCKSOURCE_MASK(bits);
-	cd.rate = rate;
 
 	/* calculate the mult/shift to convert counter ticks to ns. */
-	clocks_calc_mult_shift(&cd.mult, &cd.shift, rate, NSEC_PER_SEC, 3600);
+	clocks_calc_mult_shift(&new_mult, &new_shift, rate, NSEC_PER_SEC, 3600);
+
+	new_mask = CLOCKSOURCE_MASK(bits);
+
+	/* calculate how many ns until we wrap */
+	wrap = clocks_calc_max_nsecs(new_mult, new_shift, 0, new_mask);
+	new_wrap_kt = ns_to_ktime(wrap - (wrap >> 3));
+
+	/* update epoch for new counter and update epoch_ns from old counter*/
+	new_epoch = read();
+	cyc = read_sched_clock();
+	ns = cd.epoch_ns + cyc_to_ns((cyc - cd.epoch_cyc) & sched_clock_mask,
+			  cd.mult, cd.shift);
+
+	raw_write_seqcount_begin(&cd.seq);
+	read_sched_clock = read;
+	sched_clock_mask = new_mask;
+	cd.rate = rate;
+	cd.wrap_kt = new_wrap_kt;
+	cd.mult = new_mult;
+	cd.shift = new_shift;
+	cd.epoch_cyc = new_epoch;
+	cd.epoch_ns = ns;
+	raw_write_seqcount_end(&cd.seq);
 
 	r = rate;
 	if (r >= 4000000) {
@@ -141,22 +163,12 @@ void __init sched_clock_register(u64 (*read)(void), int bits,
 	} else
 		r_unit = ' ';
 
-	/* calculate how many ns until we wrap */
-	wrap = clocks_calc_max_nsecs(cd.mult, cd.shift, 0, sched_clock_mask);
-	cd.wrap_kt = ns_to_ktime(wrap - (wrap >> 3));
-
 	/* calculate the ns resolution of this counter */
-	res = cyc_to_ns(1ULL, cd.mult, cd.shift);
+	res = cyc_to_ns(1ULL, new_mult, new_shift);
+
 	pr_info("sched_clock: %u bits at %lu%cHz, resolution %lluns, wraps every %lluns\n",
 		bits, r, r_unit, res, wrap);
 
-	update_sched_clock();
-
-	/*
-	 * Ensure that sched_clock() starts off at 0ns
-	 */
-	cd.epoch_ns = 0;
-
 	/* Enable IRQ time accounting if we have a fast enough sched_clock */
 	if (irqtime > 0 || (irqtime == -1 && rate >= 1000000))
 		enable_sched_clock_irqtime();