With HIGHRES enabled tick_sched_timer() is programmed every jiffy to
expire the timer_list timers. This timer is programmed accurate in
respect to CLOCK_MONOTONIC so that 0 seconds and nanoseconds is the
first tick and the next one is 1000/CONFIG_HZ ms later. For HZ=250 it is
every 4 ms and so based on the current time the next tick can be
computed.

This accuracy broke since the commit mentioned below because the jiffy
based clocksource is initialized with higher accuracy in
read_persistent_wall_and_boot_offset(). This higher accuracy is
inherited during the setup in tick_setup_device(). The timer still fires
every 4ms with HZ=250 but timer is no longer aligned with
CLOCK_MONOTONIC with 0 as it origin but has an offset in the us/ns part
of the timestamp. The offset differs with every boot and makes it
impossible for user land to align with the tick.

Align the tick timer with CLOCK_MONOTONIC ensuring that it is always a
multiple of 1000/CONFIG_HZ ms.

Reported-by: Gusenleitner Klaus <gus@keba.com>
Link: https://lore.kernel.org/20230406095735.0_14edn3@linutronix.de
Fixes: 857baa87b6422 ("sched/clock: Enable sched clock early")
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
 kernel/time/tick-common.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

diff --git a/kernel/time/tick-common.c b/kernel/time/tick-common.c
index 46789356f856e..65b8658da829e 100644
--- a/kernel/time/tick-common.c
+++ b/kernel/time/tick-common.c
@@ -218,9 +218,19 @@ static void tick_setup_device(struct tick_device *td,
 		 * this cpu:
 		 */
 		if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
+			ktime_t next_p;
+			u32 rem;
+
 			tick_do_timer_cpu = cpu;
 
-			tick_next_period = ktime_get();
+			next_p = ktime_get();
+			div_u64_rem(next_p, TICK_NSEC, &rem);
+			if (rem) {
+				next_p -= rem;
+				next_p += TICK_NSEC;
+			}
+
+			tick_next_period = next_p;
 #ifdef CONFIG_NO_HZ_FULL
 			/*
 			 * The boot CPU may be nohz_full, in which case set
-- 
2.40.0

Hi Sebastian,

On 18.04.23 14:26, Sebastian Andrzej Siewior wrote:
> [...]. The timer still fires
> every 4ms with HZ=250 but timer is no longer aligned with
> CLOCK_MONOTONIC with 0 as it origin but has an offset in the us/ns part
> of the timestamp. The offset differs with every boot and makes it
> impossible for user land to align with the tick.

I can observe these per-boot offsets too, but...

> Align the tick timer with CLOCK_MONOTONIC ensuring that it is always a
> multiple of 1000/CONFIG_HZ ms.

this change doesn't seem to achieve that goal, unfortunately. Quite the 
opposite. It makes the (boot) clock run faster and, because of the per-
boot different offset, differently fast for each boot. Up to the point 
where it's running too fast to make any progress at all.

This patch causes VM boot hangs for us. It took a while to identify as 
the boot hangs were only ~1 out of 30 but it's clearly it. Reverting 
the commit got me 100 boots in a row without any issue.

Instrumenting the kernel a little gave me a clue what the bug is. When 
switching from the boot timer tick device (which is 'hpet' in my setup) 
to 'lapic-deadline', the mode of the timer isn't changed and kept at 
TICKDEV_MODE_PERIODIC. As that device doesn't support this mode, 
tick_setup_periodic() will switch over to CLOCK_EVT_STATE_ONESHOT mode 
and program the next expire event based on tick_next_period.

clockevents_program_event() will calculate the delta of that timestamp 
and ktime_get() and pass that value on to dev->set_next_event() (which 
is lapic_next_deadline()) which will write it to the IA32_TSC_DEADLINE 
MSR.

That delta value -- which is still the per-boot different offset to 
ktime_get() your patch introduces -- now gets stuck and is taken as the 
new *jiffies tick time*. That's because tick_handle_periodic() -> 
tick_periodic() will advance tick_next_period by TICK_NSEC, make 
do_timer() increment jiffies_64 by one and then program the next event 
to be in TICK_NSEC ns based on the device's old expiry time, i.e. keep 
the offset intact. This is followed by re-arming the event by a call to 
clockevents_program_event() which does the already-know delta 
calculation and writes, again, the too little value to 
IA32_TSC_DEADLINE.

This effectively makes the jiffies based clock go too fast as the timer 
IRQ comes too early (less than TICK_NSEC ns). Sometimes it's barely 
noticeable, but sometimes it's so fast that the kernel is overloaded 
with only handling the local timer IRQ without making any further 
progress, especially in (nested) VM setups.

Without commit e9523a0d8189 ("tick/common: Align tick period with the 
HZ tick."), which was backported to many stable and LTS kernels (v6.3.2 
(571c3b46c9b3), v6.2.15 (f0cb827199ec), v6.1.28 (290e26ec0d01), 
v5.15.111 (a55050c7989c), v5.10.180 (c4013689269d) and v5.4.243 
(a3e7a3d472c2)) this clock drift is gone and my VMs boot again.

Before that commit, the delta between tick_next_period and ktime_get() 
was initially zero, so tick_handle_period() had to loop, as 
clockevents_program_event() will return with -ETIME. The next attempt 
would be done with a delta of TICK_NSEC which will make 
clockevents_program_event() succeed and ensure that future events don't 
need the additional loop iteration, as the delta got stuck at TICK_NSEC 
-- exactly where it should be.

We observed the bug first on the v6.3, v6.1 and v5.15 stable branch 
updates from May 11th and then, a week later, on v5.4 too. All first 
occurrences were coinciding with the bad commit going into the 
corresponding stable and LTS kernel releases.

The issue manifests itself as a fast running clock only during boot, 
when the clock source is still jiffies based. That'll eventually lead 
to a boot hang as the timer IRQs are firing too fast.

To reproduce this you can either boot loop a VM and try to get "lucky" 
to hit a big enough 'rem' value or just apply this little diff instead:

---8<---
diff --git a/kernel/time/tick-common.c b/kernel/time/tick-common.c
index 65b8658da829..b01cf18a5d42 100644
--- a/kernel/time/tick-common.c
+++ b/kernel/time/tick-common.c
@@ -225,6 +225,7 @@ static void tick_setup_device(struct tick_device *td,
 
 			next_p = ktime_get();
 			div_u64_rem(next_p, TICK_NSEC, &rem);
+			rem = TICK_NSEC - 123;
 			if (rem) {
 				next_p -= rem;
 				next_p += TICK_NSEC;
--->8---

This should make the kernel get stuck with only handling timer ticks 
but not making any further progress.

Change the subtrahend to 1234 to get a system that boots but has an 
unrealistically fast clock during kernel initialization.

As reverting that commit fixes the issue for us but it seemingly fixes 
another bug for Klaus (or at least attempted to), the now uncovered bug 
should be fixed instead.

The fundamental issue is that the jiffies based clock source cannot be 
trusted and shouldn't be used to calculate offsets to timestamps in the 
future when tick_next_period mod ktime_get() != 0.

Can we defer the offset adjustment of tick_next_period to a later point 
in time when a stable clock source gets used, like 'tsc'?

Thanks,
Mathias

On 2023-06-14 00:59:46 [+0200], Mathias Krause wrote:
> Hi Sebastian,
Hi Mathias,

> this change doesn't seem to achieve that goal, unfortunately. Quite the 
> opposite. It makes the (boot) clock run faster and, because of the per-
> boot different offset, differently fast for each boot. Up to the point 
> where it's running too fast to make any progress at all.

I meant to reply to this thread but took the other by accident. Sorry.
This should be addressed by
	https://lore.kernel.org/20230615091830.RxMV2xf_@linutronix.de

Thank you for the report.

> Thanks,
> Mathias

Sebastian

On Wed, Jun 14, 2023 at 12:59:46AM +0200, Mathias Krause wrote:
> This patch causes VM boot hangs for us. It took a while to identify as 
> the boot hangs were only ~1 out of 30 but it's clearly it. Reverting 
> the commit got me 100 boots in a row without any issue.

FWIW I have quite a nice test program for catching these sorts of boot
hangs, see attached.  You need to change the VMLINUX define to point
to your vmlinux or vmlinuz file.

Rich.

-- 
Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones
Read my programming and virtualization blog: http://rwmj.wordpress.com
virt-builder quickly builds VMs from scratch
http://libguestfs.org/virt-builder.1.html
#include <stdio.h>
#include <stdlib.h>
#include <errno.h>
#include <error.h>
#include <unistd.h>
#include <sys/wait.h>
#include <pthread.h>

#define NR_ITERATIONS 10000
#define MAX_THREADS 8
#define MAX_TIME 60 /* max time to wait for qemu to complete */

//#define VMLINUX "/home/rjones/d/linux/vmlinux"
#define VMLINUX "vmlinux"
//#define QEMU "/home/rjones/d/qemu/build/qemu-system-x86_64"
#define QEMU "qemu-system-x86_64"
#define QEMU_COMMAND \
  QEMU " -no-user-config -nodefaults -display none " \
  "-machine accel=kvm:tcg,graphics=off -cpu max,la57=off -m 1280 " \
  "-no-reboot " \
  "-rtc driftfix=slew -no-hpet -global kvm-pit.lost_tick_policy=discard " \
  "-kernel " VMLINUX " " \
  "-serial stdio " \
  "-append \"panic=1 console=ttyS0 edd=off udevtimeout=6000 udev.event-timeout=6000 no_timer_check printk.time=1 cgroup_disable=memory usbcore.nousb cryptomgr.notests tsc=reliable 8250.nr_uarts=1 selinux=0 TERM=xterm-256color\""

static pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER;
static unsigned iterations;

static void *start_thread (void *);

int
main (int argc, char *argv[])
{
	long n, i;
	int r;
	pthread_t thread[MAX_THREADS];

	n = sysconf (_SC_NPROCESSORS_ONLN);
	if (n == -1) error (EXIT_FAILURE, errno, "sysconf");

	if (n > MAX_THREADS)
		n = MAX_THREADS;

	for (i = 0; i < n; ++i) {
		r = pthread_create (&thread[i], NULL, start_thread, NULL);
		if (r != 0) error (EXIT_FAILURE, r, "pthread_create");
	}
	for (i = 0; i < n; ++i) {
		r = pthread_join (thread[i], NULL);
		if (r != 0) error (EXIT_FAILURE, r, "pthread_join");
	}
	printf ("\n");
	printf ("\n");
	printf ("test ok\n");
	exit (EXIT_SUCCESS);
}

static void *
start_thread (void *vp)
{
	pid_t pid;
	char tmp[] = "/tmp/kernel.out.XXXXXX";
	char cmd[1024];
	int i, r, status;

	if (mkstemp (tmp) == -1)
		error (EXIT_FAILURE, errno, "mkstemp: %s", tmp);

	snprintf (cmd, sizeof cmd, QEMU_COMMAND " >& %s", tmp);

	/* This basically runs a loop starting qemu. */
	for (;;) {
		pthread_mutex_lock (&lock);
		if (iterations >= NR_ITERATIONS) {
			pthread_mutex_unlock (&lock);
			return NULL;
		}
		if (iterations <= MAX_THREADS) { // stagger the start times
			pthread_mutex_unlock (&lock);
			usleep (rand () % 3000000);
			pthread_mutex_lock (&lock);
		}
		iterations++;
		printf ("%d... ", iterations); fflush (stdout);
		pthread_mutex_unlock (&lock);

		pid = fork ();
		if (pid == -1) error (EXIT_FAILURE, errno, "fork");
		if (pid == 0) {
			/* Child process, run qemu and wait. */
			if (system (cmd) != 0)
				_exit (EXIT_FAILURE);
			else
				_exit (EXIT_SUCCESS);
		}

		/* In the parent wait up to MAX_TIME seconds. */
		for (i = 0; i < MAX_TIME; ++i) {
			r = waitpid (pid, &status, WNOHANG);
			if (r == -1) error (EXIT_FAILURE, errno, "waitpid");
			if (r > 0) break;
			sleep (1);
		}

		if (i == MAX_TIME || status != 0) {
			/* Something failed in qemu (or it didn't
			 * exit), dump the whole log and exit with
			 * error.
			 */
			printf ("\n");
			printf ("\n");
			snprintf (cmd, sizeof cmd, "tail -20 %s", tmp);
			system (cmd);
			fprintf (stderr, "*** ERROR OR HANG ***\n");
			exit (EXIT_FAILURE);
		}
	}

	unlink (tmp);
}

On 15.06.23 11:03, Richard W.M. Jones wrote:
> On Wed, Jun 14, 2023 at 12:59:46AM +0200, Mathias Krause wrote:
>> This patch causes VM boot hangs for us. It took a while to identify as 
>> the boot hangs were only ~1 out of 30 but it's clearly it. Reverting 
>> the commit got me 100 boots in a row without any issue.
> 
> FWIW I have quite a nice test program for catching these sorts of boot
> hangs, see attached.  You need to change the VMLINUX define to point
> to your vmlinux or vmlinuz file.

Heh, nice. For this specific case, though, a simple boot loop in a
single VM was sufficient:

  $ kvm -cpu host -smp 8 -m 8G -display none -serial stdio -s \
        -kernel path/to/bzImage -append 'console=ttyS0 panic=1'

This will reboot the VM in an endless loop until it hangs which is
visually noticeable as there's no more output generated to the terminal.
At that time one can attach to it via gdb and investigate further:

  $ gdb -ex 'target remote :1234' path/to/vmlinux

Yet another advantage is that one can filter the output, e.g. via:

  $ kvm ... | grep clock

Thanks,
Mathias

> 
> Rich.
>

The following commit has been merged into the timers/urgent branch of tip:

Commit-ID:     13bb06f8dd42071cb9a49f6e21099eea05d4b856
Gitweb:        https://git.kernel.org/tip/13bb06f8dd42071cb9a49f6e21099eea05d4b856
Author:        Thomas Gleixner <tglx@linutronix.de>
AuthorDate:    Thu, 15 Jun 2023 11:18:30 +02:00
Committer:     Thomas Gleixner <tglx@linutronix.de>
CommitterDate: Fri, 16 Jun 2023 20:45:28 +02:00

tick/common: Align tick period during sched_timer setup

The tick period is aligned very early while the first clock_event_device is
registered. At that point the system runs in periodic mode and switches
later to one-shot mode if possible.

The next wake-up event is programmed based on the aligned value
(tick_next_period) but the delta value, that is used to program the
clock_event_device, is computed based on ktime_get().

With the subtracted offset, the device fires earlier than the exact time
frame. With a large enough offset the system programs the timer for the
next wake-up and the remaining time left is too small to make any boot
progress. The system hangs.

Move the alignment later to the setup of tick_sched timer. At this point
the system switches to oneshot mode and a high resolution clocksource is
available. At this point it is safe to align tick_next_period because
ktime_get() will now return accurate (not jiffies based) time.

[bigeasy: Patch description + testing].

Fixes: e9523a0d81899 ("tick/common: Align tick period with the HZ tick.")
Reported-by: Mathias Krause <minipli@grsecurity.net>
Reported-by: "Bhatnagar, Rishabh" <risbhat@amazon.com>
Suggested-by: Mathias Krause <minipli@grsecurity.net>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Richard W.M. Jones <rjones@redhat.com>
Tested-by: Mathias Krause <minipli@grsecurity.net>
Acked-by: SeongJae Park <sj@kernel.org>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/5a56290d-806e-b9a5-f37c-f21958b5a8c0@grsecurity.net
Link: https://lore.kernel.org/12c6f9a3-d087-b824-0d05-0d18c9bc1bf3@amazon.com
Link: https://lore.kernel.org/r/20230615091830.RxMV2xf_@linutronix.de
---
 kernel/time/tick-common.c | 13 +------------
 kernel/time/tick-sched.c  | 13 ++++++++++++-
 2 files changed, 13 insertions(+), 13 deletions(-)

diff --git a/kernel/time/tick-common.c b/kernel/time/tick-common.c
index 65b8658..e9138cd 100644
--- a/kernel/time/tick-common.c
+++ b/kernel/time/tick-common.c
@@ -218,19 +218,8 @@ static void tick_setup_device(struct tick_device *td,
 		 * this cpu:
 		 */
 		if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
-			ktime_t next_p;
-			u32 rem;
-
 			tick_do_timer_cpu = cpu;
-
-			next_p = ktime_get();
-			div_u64_rem(next_p, TICK_NSEC, &rem);
-			if (rem) {
-				next_p -= rem;
-				next_p += TICK_NSEC;
-			}
-
-			tick_next_period = next_p;
+			tick_next_period = ktime_get();
 #ifdef CONFIG_NO_HZ_FULL
 			/*
 			 * The boot CPU may be nohz_full, in which case set
diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
index 5225467..42c0be3 100644
--- a/kernel/time/tick-sched.c
+++ b/kernel/time/tick-sched.c
@@ -161,8 +161,19 @@ static ktime_t tick_init_jiffy_update(void)
 	raw_spin_lock(&jiffies_lock);
 	write_seqcount_begin(&jiffies_seq);
 	/* Did we start the jiffies update yet ? */
-	if (last_jiffies_update == 0)
+	if (last_jiffies_update == 0) {
+		u32 rem;
+
+		/*
+		 * Ensure that the tick is aligned to a multiple of
+		 * TICK_NSEC.
+		 */
+		div_u64_rem(tick_next_period, TICK_NSEC, &rem);
+		if (rem)
+			tick_next_period += TICK_NSEC - rem;
+
 		last_jiffies_update = tick_next_period;
+	}
 	period = last_jiffies_update;
 	write_seqcount_end(&jiffies_seq);
 	raw_spin_unlock(&jiffies_lock);