When a partition is transferred, once it arrives at the destination node,
the partition is active but much of its memory must be transferred from the
start node.
It depends on the activity in the partition, but the more CPU the partition
has, the more memory to be transferred is likely to be. This causes latency
when accessing pages that need to be transferred, and often, for large
partitions, it triggers the NMI watchdog.
The NMI watchdog causes the CPU stack to dump where it appears to be
stuck. In this case, it does not bring much information since it can happen
during any memory access of the kernel.
In addition, the NMI interrupt mechanism is not secure and can generate a
dump system in the event that the interruption is taken while MSR[RI]=0.
Depending on the LPAR size and load, it may be interesting to extend the
NMI watchdog timer during the LPM.
That's configurable through sysctl with the new introduced variable
(specific to powerpc) nmi_watchdog_factor. This value represents the
percentage added to watchdog_tresh to set the NMI watchdog timeout during a
LPM.
Changes in V5 (no functional changes in this version):
Patch 4/4:
- fixing typos and grammar issues reported by Randy.
- Renaming sysctl value from nmi_watchdog_factor to nmi_wd_lpm_factor as
per Nick request
V4:
https://lore.kernel.org/linuxppc-dev/[email protected]/
Laurent Dufour (4):
powerpc/mobility: wait for memory transfer to complete
watchdog: export lockup_detector_reconfigure
powerpc/watchdog: introduce a NMI watchdog's factor
pseries/mobility: set NMI watchdog factor during an LPM
Documentation/admin-guide/sysctl/kernel.rst | 12 +++
arch/powerpc/include/asm/nmi.h | 2 +
arch/powerpc/kernel/watchdog.c | 21 ++++-
arch/powerpc/platforms/pseries/mobility.c | 91 ++++++++++++++++++++-
include/linux/nmi.h | 2 +
kernel/watchdog.c | 21 +++--
6 files changed, 141 insertions(+), 8 deletions(-)
--
2.37.0
Introduce a factor which would apply to the NMI watchdog timeout.
This factor is a percentage added to the watchdog_tresh value. The value is
set under the watchdog_mutex protection and lockup_detector_reconfigure()
is called to recompute wd_panic_timeout_tb.
Once the factor is set, it remains until it is set back to 0, which means
no impact.
Reviewed-by: Nicholas Piggin <[email protected]>
Signed-off-by: Laurent Dufour <[email protected]>
---
arch/powerpc/include/asm/nmi.h | 2 ++
arch/powerpc/kernel/watchdog.c | 21 ++++++++++++++++++++-
2 files changed, 22 insertions(+), 1 deletion(-)
diff --git a/arch/powerpc/include/asm/nmi.h b/arch/powerpc/include/asm/nmi.h
index ea0e487f87b1..c3c7adef74de 100644
--- a/arch/powerpc/include/asm/nmi.h
+++ b/arch/powerpc/include/asm/nmi.h
@@ -5,8 +5,10 @@
#ifdef CONFIG_PPC_WATCHDOG
extern void arch_touch_nmi_watchdog(void);
long soft_nmi_interrupt(struct pt_regs *regs);
+void watchdog_nmi_set_timeout_pct(u64 pct);
#else
static inline void arch_touch_nmi_watchdog(void) {}
+static inline void watchdog_nmi_set_timeout_pct(u64 pct) {}
#endif
#ifdef CONFIG_NMI_IPI
diff --git a/arch/powerpc/kernel/watchdog.c b/arch/powerpc/kernel/watchdog.c
index 7d28b9553654..5d903e63f932 100644
--- a/arch/powerpc/kernel/watchdog.c
+++ b/arch/powerpc/kernel/watchdog.c
@@ -91,6 +91,10 @@ static cpumask_t wd_smp_cpus_pending;
static cpumask_t wd_smp_cpus_stuck;
static u64 wd_smp_last_reset_tb;
+#ifdef CONFIG_PPC_PSERIES
+static u64 wd_timeout_pct;
+#endif
+
/*
* Try to take the exclusive watchdog action / NMI IPI / printing lock.
* wd_smp_lock must be held. If this fails, we should return and wait
@@ -527,7 +531,13 @@ static int stop_watchdog_on_cpu(unsigned int cpu)
static void watchdog_calc_timeouts(void)
{
- wd_panic_timeout_tb = watchdog_thresh * ppc_tb_freq;
+ u64 threshold = watchdog_thresh;
+
+#ifdef CONFIG_PPC_PSERIES
+ threshold += (READ_ONCE(wd_timeout_pct) * threshold) / 100;
+#endif
+
+ wd_panic_timeout_tb = threshold * ppc_tb_freq;
/* Have the SMP detector trigger a bit later */
wd_smp_panic_timeout_tb = wd_panic_timeout_tb * 3 / 2;
@@ -570,3 +580,12 @@ int __init watchdog_nmi_probe(void)
}
return 0;
}
+
+#ifdef CONFIG_PPC_PSERIES
+void watchdog_nmi_set_timeout_pct(u64 pct)
+{
+ pr_info("Set the NMI watchdog timeout factor to %llu%%\n", pct);
+ WRITE_ONCE(wd_timeout_pct, pct);
+ lockup_detector_reconfigure();
+}
+#endif
--
2.37.0
In pseries_migration_partition(), loop until the memory transfer is
complete. This way the calling drmgr process will not exit earlier,
allowing callbacks to be run only once the migration is fully completed.
If reading the VASI state is done after the hypervisor has completed the
migration, the HCALL is returning H_PARAMETER. We can safely assume that
the memory transfer is achieved if this happens.
This will also allow to manage the NMI watchdog state in the next commits.
Reviewed-by: Nathan Lynch <[email protected]>
Reviewed-by: Nicholas Piggin <[email protected]>
Signed-off-by: Laurent Dufour <[email protected]>
---
arch/powerpc/platforms/pseries/mobility.c | 48 ++++++++++++++++++++++-
1 file changed, 46 insertions(+), 2 deletions(-)
diff --git a/arch/powerpc/platforms/pseries/mobility.c b/arch/powerpc/platforms/pseries/mobility.c
index 78f3f74c7056..6297467072e6 100644
--- a/arch/powerpc/platforms/pseries/mobility.c
+++ b/arch/powerpc/platforms/pseries/mobility.c
@@ -427,6 +427,43 @@ static int wait_for_vasi_session_suspending(u64 handle)
return ret;
}
+static void wait_for_vasi_session_completed(u64 handle)
+{
+ unsigned long state = 0;
+ int ret;
+
+ pr_info("waiting for memory transfer to complete...\n");
+
+ /*
+ * Wait for transition from H_VASI_RESUMED to H_VASI_COMPLETED.
+ */
+ while (true) {
+ ret = poll_vasi_state(handle, &state);
+
+ /*
+ * If the memory transfer is already complete and the migration
+ * has been cleaned up by the hypervisor, H_PARAMETER is return,
+ * which is translate in EINVAL by poll_vasi_state().
+ */
+ if (ret == -EINVAL || (!ret && state == H_VASI_COMPLETED)) {
+ pr_info("memory transfer completed.\n");
+ break;
+ }
+
+ if (ret) {
+ pr_err("H_VASI_STATE return error (%d)\n", ret);
+ break;
+ }
+
+ if (state != H_VASI_RESUMED) {
+ pr_err("unexpected H_VASI_STATE result %lu\n", state);
+ break;
+ }
+
+ msleep(500);
+ }
+}
+
static void prod_single(unsigned int target_cpu)
{
long hvrc;
@@ -673,9 +710,16 @@ static int pseries_migrate_partition(u64 handle)
vas_migration_handler(VAS_SUSPEND);
ret = pseries_suspend(handle);
- if (ret == 0)
+ if (ret == 0) {
post_mobility_fixup();
- else
+ /*
+ * Wait until the memory transfer is complete, so that the user
+ * space process returns from the syscall after the transfer is
+ * complete. This allows the user hooks to be executed at the
+ * right time.
+ */
+ wait_for_vasi_session_completed(handle);
+ } else
pseries_cancel_migration(handle, ret);
vas_migration_handler(VAS_RESUME);
--
2.37.0
During an LPM, while the memory transfer is in progress on the arrival
side, some latencies are generated when accessing not yet transferred pages
on the arrival side. Thus, the NMI watchdog may be triggered too
frequently, which increases the risk to hit an NMI interrupt in a bad place
in the kernel, leading to a kernel panic.
Disabling the Hard Lockup Watchdog until the memory transfer could be a too
strong work around, some users would want this timeout to be eventually
triggered if the system is hanging even during an LPM.
Introduce a new sysctl variable nmi_watchdog_factor. It allows to apply a
factor to the NMI watchdog timeout during an LPM. Just before the CPUs are
stopped for the switchover sequence, the NMI watchdog timer is set to
watchdog_thresh + factor%
A value of 0 has no effect. The default value is 200, meaning that the NMI
watchdog is set to 30s during LPM (based on a 10s watchdog_thresh value).
Once the memory transfer is achieved, the factor is reset to 0.
Setting this value to a high number is like disabling the NMI watchdog
during an LPM.
Reviewed-by: Nicholas Piggin <[email protected]>
Signed-off-by: Laurent Dufour <[email protected]>
---
Documentation/admin-guide/sysctl/kernel.rst | 12 ++++++
arch/powerpc/platforms/pseries/mobility.c | 43 +++++++++++++++++++++
2 files changed, 55 insertions(+)
diff --git a/Documentation/admin-guide/sysctl/kernel.rst b/Documentation/admin-guide/sysctl/kernel.rst
index ddccd1077462..d73faa619c15 100644
--- a/Documentation/admin-guide/sysctl/kernel.rst
+++ b/Documentation/admin-guide/sysctl/kernel.rst
@@ -592,6 +592,18 @@ to the guest kernel command line (see
Documentation/admin-guide/kernel-parameters.rst).
+nmi_wd_lpm_factor (PPC only)
+============================
+
+Factor apply to the NMI watchdog timeout (only when ``nmi_watchdog`` is
+set to 1). This factor represents the percentage added to
+``watchdog_thresh`` when calculating the NMI watchdog timeout during an
+LPM. The soft lockup timeout is not impacted.
+
+A value of 0 means no change. The default value is 200 meaning the NMI
+watchdog is set to 30s (based on ``watchdog_thresh`` equal to 10).
+
+
numa_balancing
==============
diff --git a/arch/powerpc/platforms/pseries/mobility.c b/arch/powerpc/platforms/pseries/mobility.c
index 6297467072e6..3d36a8955eaf 100644
--- a/arch/powerpc/platforms/pseries/mobility.c
+++ b/arch/powerpc/platforms/pseries/mobility.c
@@ -48,6 +48,39 @@ struct update_props_workarea {
#define MIGRATION_SCOPE (1)
#define PRRN_SCOPE -2
+#ifdef CONFIG_PPC_WATCHDOG
+static unsigned int nmi_wd_lpm_factor = 200;
+
+#ifdef CONFIG_SYSCTL
+static struct ctl_table nmi_wd_lpm_factor_ctl_table[] = {
+ {
+ .procname = "nmi_wd_lpm_factor",
+ .data = &nmi_wd_lpm_factor,
+ .maxlen = sizeof(int),
+ .mode = 0644,
+ .proc_handler = proc_douintvec_minmax,
+ },
+ {}
+};
+static struct ctl_table nmi_wd_lpm_factor_sysctl_root[] = {
+ {
+ .procname = "kernel",
+ .mode = 0555,
+ .child = nmi_wd_lpm_factor_ctl_table,
+ },
+ {}
+};
+
+static int __init register_nmi_wd_lpm_factor_sysctl(void)
+{
+ register_sysctl_table(nmi_wd_lpm_factor_sysctl_root);
+
+ return 0;
+}
+device_initcall(register_nmi_wd_lpm_factor_sysctl);
+#endif /* CONFIG_SYSCTL */
+#endif /* CONFIG_PPC_WATCHDOG */
+
static int mobility_rtas_call(int token, char *buf, s32 scope)
{
int rc;
@@ -702,13 +735,20 @@ static int pseries_suspend(u64 handle)
static int pseries_migrate_partition(u64 handle)
{
int ret;
+ unsigned int factor = 0;
+#ifdef CONFIG_PPC_WATCHDOG
+ factor = nmi_wd_lpm_factor;
+#endif
ret = wait_for_vasi_session_suspending(handle);
if (ret)
return ret;
vas_migration_handler(VAS_SUSPEND);
+ if (factor)
+ watchdog_nmi_set_timeout_pct(factor);
+
ret = pseries_suspend(handle);
if (ret == 0) {
post_mobility_fixup();
@@ -722,6 +762,9 @@ static int pseries_migrate_partition(u64 handle)
} else
pseries_cancel_migration(handle, ret);
+ if (factor)
+ watchdog_nmi_set_timeout_pct(0);
+
vas_migration_handler(VAS_RESUME);
return ret;
--
2.37.0
In some circumstances it may be interesting to reconfigure the watchdog
from inside the kernel.
On PowerPC, this may helpful before and after a LPAR migration (LPM) is
initiated, because it implies some latencies, watchdog, and especially NMI
watchdog is expected to be triggered during this operation. Reconfiguring
the watchdog with a factor, would prevent it to happen too frequently
during LPM.
Rename lockup_detector_reconfigure() as __lockup_detector_reconfigure() and
create a new function lockup_detector_reconfigure() calling
__lockup_detector_reconfigure() under the protection of watchdog_mutex.
Cc: Christoph Hellwig <[email protected]>
Signed-off-by: Laurent Dufour <[email protected]>
---
include/linux/nmi.h | 2 ++
kernel/watchdog.c | 21 ++++++++++++++++-----
2 files changed, 18 insertions(+), 5 deletions(-)
diff --git a/include/linux/nmi.h b/include/linux/nmi.h
index 750c7f395ca9..f700ff2df074 100644
--- a/include/linux/nmi.h
+++ b/include/linux/nmi.h
@@ -122,6 +122,8 @@ int watchdog_nmi_probe(void);
int watchdog_nmi_enable(unsigned int cpu);
void watchdog_nmi_disable(unsigned int cpu);
+void lockup_detector_reconfigure(void);
+
/**
* touch_nmi_watchdog - restart NMI watchdog timeout.
*
diff --git a/kernel/watchdog.c b/kernel/watchdog.c
index 20a7a55e62b6..90e6c41d5e33 100644
--- a/kernel/watchdog.c
+++ b/kernel/watchdog.c
@@ -541,7 +541,7 @@ int lockup_detector_offline_cpu(unsigned int cpu)
return 0;
}
-static void lockup_detector_reconfigure(void)
+static void __lockup_detector_reconfigure(void)
{
cpus_read_lock();
watchdog_nmi_stop();
@@ -561,6 +561,13 @@ static void lockup_detector_reconfigure(void)
__lockup_detector_cleanup();
}
+void lockup_detector_reconfigure(void)
+{
+ mutex_lock(&watchdog_mutex);
+ __lockup_detector_reconfigure();
+ mutex_unlock(&watchdog_mutex);
+}
+
/*
* Create the watchdog infrastructure and configure the detector(s).
*/
@@ -577,13 +584,13 @@ static __init void lockup_detector_setup(void)
return;
mutex_lock(&watchdog_mutex);
- lockup_detector_reconfigure();
+ __lockup_detector_reconfigure();
softlockup_initialized = true;
mutex_unlock(&watchdog_mutex);
}
#else /* CONFIG_SOFTLOCKUP_DETECTOR */
-static void lockup_detector_reconfigure(void)
+void __lockup_detector_reconfigure(void)
{
cpus_read_lock();
watchdog_nmi_stop();
@@ -591,9 +598,13 @@ static void lockup_detector_reconfigure(void)
watchdog_nmi_start();
cpus_read_unlock();
}
+static inline void lockup_detector_reconfigure(void)
+{
+ __lockup_detector_reconfigure();
+}
static inline void lockup_detector_setup(void)
{
- lockup_detector_reconfigure();
+ __lockup_detector_reconfigure();
}
#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
@@ -633,7 +644,7 @@ static void proc_watchdog_update(void)
{
/* Remove impossible cpus to keep sysctl output clean. */
cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
- lockup_detector_reconfigure();
+ __lockup_detector_reconfigure();
}
/*
--
2.37.0
Hi Laurent,
On 7/13/22 08:47, Laurent Dufour wrote:
> diff --git a/Documentation/admin-guide/sysctl/kernel.rst b/Documentation/admin-guide/sysctl/kernel.rst
> index ddccd1077462..d73faa619c15 100644
> --- a/Documentation/admin-guide/sysctl/kernel.rst
> +++ b/Documentation/admin-guide/sysctl/kernel.rst
> @@ -592,6 +592,18 @@ to the guest kernel command line (see
> Documentation/admin-guide/kernel-parameters.rst).
>
>
> +nmi_wd_lpm_factor (PPC only)
> +============================
> +
> +Factor apply to the NMI watchdog timeout (only when ``nmi_watchdog`` is
Factor to apply to
> +set to 1). This factor represents the percentage added to
> +``watchdog_thresh`` when calculating the NMI watchdog timeout during an
> +LPM. The soft lockup timeout is not impacted.
> +
> +A value of 0 means no change. The default value is 200 meaning the NMI
> +watchdog is set to 30s (based on ``watchdog_thresh`` equal to 10).
--
~Randy
Le 13/07/2022 à 22:17, Randy Dunlap a écrit :
> Hi Laurent,
>
> On 7/13/22 08:47, Laurent Dufour wrote:
>> diff --git a/Documentation/admin-guide/sysctl/kernel.rst b/Documentation/admin-guide/sysctl/kernel.rst
>> index ddccd1077462..d73faa619c15 100644
>> --- a/Documentation/admin-guide/sysctl/kernel.rst
>> +++ b/Documentation/admin-guide/sysctl/kernel.rst
>> @@ -592,6 +592,18 @@ to the guest kernel command line (see
>> Documentation/admin-guide/kernel-parameters.rst).
>>
>>
>> +nmi_wd_lpm_factor (PPC only)
>> +============================
>> +
>> +Factor apply to the NMI watchdog timeout (only when ``nmi_watchdog`` is
>
> Factor to apply to
Thanks, Randy.
Michael, could you fix that when applying the series?
Cheers,
Laurent
>
>> +set to 1). This factor represents the percentage added to
>> +``watchdog_thresh`` when calculating the NMI watchdog timeout during an
>> +LPM. The soft lockup timeout is not impacted.
>> +
>> +A value of 0 means no change. The default value is 200 meaning the NMI
>> +watchdog is set to 30s (based on ``watchdog_thresh`` equal to 10).
>
Laurent Dufour <[email protected]> writes:
> Le 13/07/2022 à 22:17, Randy Dunlap a écrit :
>> Hi Laurent,
>>
>> On 7/13/22 08:47, Laurent Dufour wrote:
>>> diff --git a/Documentation/admin-guide/sysctl/kernel.rst b/Documentation/admin-guide/sysctl/kernel.rst
>>> index ddccd1077462..d73faa619c15 100644
>>> --- a/Documentation/admin-guide/sysctl/kernel.rst
>>> +++ b/Documentation/admin-guide/sysctl/kernel.rst
>>> @@ -592,6 +592,18 @@ to the guest kernel command line (see
>>> Documentation/admin-guide/kernel-parameters.rst).
>>>
>>>
>>> +nmi_wd_lpm_factor (PPC only)
>>> +============================
>>> +
>>> +Factor apply to the NMI watchdog timeout (only when ``nmi_watchdog`` is
>>
>> Factor to apply to
>
> Thanks, Randy.
>
> Michael, could you fix that when applying the series?
Yes, I did.
cheers
On Wed, 13 Jul 2022 17:47:25 +0200, Laurent Dufour wrote:
> When a partition is transferred, once it arrives at the destination node,
> the partition is active but much of its memory must be transferred from the
> start node.
>
> It depends on the activity in the partition, but the more CPU the partition
> has, the more memory to be transferred is likely to be. This causes latency
> when accessing pages that need to be transferred, and often, for large
> partitions, it triggers the NMI watchdog.
>
> [...]
Applied to powerpc/next.
[1/4] powerpc/mobility: wait for memory transfer to complete
https://git.kernel.org/powerpc/c/882c0d1704cf61df13f01933269202d51e74b9f3
[2/4] watchdog: export lockup_detector_reconfigure
https://git.kernel.org/powerpc/c/7c56a8733d0a2a4be2438a7512566e5ce552fccf
[3/4] powerpc/watchdog: introduce a NMI watchdog's factor
https://git.kernel.org/powerpc/c/f5e74e836097d1004077390717d4bd95d4a2c27a
[4/4] pseries/mobility: set NMI watchdog factor during an LPM
https://git.kernel.org/powerpc/c/118b1366930c8c833b8b36abef657f40d4e26610
cheers