kernel/hung_task.c: Introduce sysctl to print all traces when a hung task is detected
Subsystems (5)
PATCH v2 1/1 kernel/hung_task.c: Introduce sysctl to print all traces when a hung task is detected
Commit 401c636a0eeb ("kernel/hung_task.c: show all hung tasks before panic") introduced a change in that we started to show all CPUs backtraces when a hung task is detected _and_ the sysctl/kernel parameter "hung_task_panic" is set. The idea is good, because usually when observing deadlocks (that may lead to hung tasks), the culprit is another task holding a lock and not necessarily the task detected as hung.
The problem with this approach is that dumping backtraces is a slightly expensive task, specially printing that on console (and specially in many CPU machines, as servers commonly found nowadays). So, users that plan to collect a kdump to investigate the hung tasks and narrow down the deadlock definitely don't need the CPUs backtrace on dmesg/console, which will delay the panic and pollute the log (crash tool would easily grab all CPUs traces with 'bt -a' command).
Also, there's the reciprocal scenario: some users may be interested in seeing the CPUs backtraces but not have the system panic when a hung task is detected. The current approach hence is almost as embedding a policy in the kernel, by forcing the CPUs backtraces' dump (only) on hung_task_panic.
This patch decouples the panic event on hung task from the CPUs backtraces dump, by creating (and documenting) a new sysctl/kernel parameter called "hung_task_all_cpu_backtrace", analog to the approach taken on soft/hard lockups, that have both a panic and an "all_cpu_backtrace" sysctl to allow individual control. The new mechanism for dumping the CPUs backtraces on hung task detection respects "hung_task_warnings" by not dumping the traces in case there's no warnings left.
---
V2: Followed suggestions from Kees and Tetsuo (and other grammar improvements). Also, followed Tetsuo suggestion to itereate kernel testing community - but I don't really know a ML for that, so I've CCed Coccinelle community and kernel-api ML.
Also, Tetsuo suggested that this option could be default to 1 - I'm open to it, but given it is only available if hung_task panic is set as of now and the goal of this patch is give users more flexibility, I vote to keep default as 0. I can respin a V3 in case more people want to see it enabled by default. Thanks in advance for the review! Cheers,
Guilherme
.../admin-guide/kernel-parameters.txt | 6 ++++ Documentation/admin-guide/sysctl/kernel.rst | 15 ++++++++++ include/linux/sched/sysctl.h | 7 +++++ kernel/hung_task.c | 30 +++++++++++++++++-- kernel/sysctl.c | 11 +++++++ 5 files changed, 67 insertions(+), 2 deletions(-)
Documentation/admin-guide/kernel-parameters.txt +6 -0
@@ -1453,6 +1453,12 @@ x86-64 are 2M (when the CPU supports "pse") and 1G (when the CPU supports the "pdpe1gb" cpuinfo flag). + hung_task_all_cpu_backtrace= + [KNL] Should kernel generate backtraces on all cpus + when a hung task is detected. Defaults to 0 and can + be controlled by hung_task_all_cpu_backtrace sysctl. + Format: <integer> + hung_task_panic= [KNL] Should the hung task detector generate panics. Format: <integer>
Documentation/admin-guide/sysctl/kernel.rst +15 -0
@@ -40,6 +40,7 @@ show up in /proc/sys/kernel: - hotplug - hardlockup_all_cpu_backtrace - hardlockup_panic +- hung_task_all_cpu_backtrace - hung_task_panic - hung_task_check_count - hung_task_timeout_secs
@@ -338,6 +339,20 @@ Path for the hotplug policy agent. Default value is "/sbin/hotplug". +hung_task_all_cpu_backtrace: +================ + +If this option is set, the kernel will send an NMI to all CPUs to dump +their backtraces when a hung task is detected. This file shows up if +CONFIG_DETECT_HUNG_TASK and CONFIG_SMP are enabled. + +0: Won't show all CPUs backtraces when a hung task is detected. +This is the default behavior. + +1: Will non-maskably interrupt all CPUs and dump their backtraces when +a hung task is detected. + + hung_task_panic: ================
include/linux/sched/sysctl.h +7 -0
@@ -7,6 +7,13 @@struct ctl_table ;#ifdef CONFIG_DETECT_HUNG_TASK + +#ifdef CONFIG_SMP +extern unsigned int sysctl_hung_task_all_cpu_backtrace; +#else +#define sysctl_hung_task_all_cpu_backtrace 0 +#endif /* CONFIG_SMP */ +extern int sysctl_hung_task_check_count;extern unsigned int sysctl_hung_task_panic;extern unsigned long sysctl_hung_task_timeout_secs;
kernel/hung_task.c +28 -2
@@ -53,9 +53,28 @@ int __read_mostly sysctl_hung_task_warnings = 10;static int __read_mostly did_panic;static bool hung_task_show_lock;static bool hung_task_call_panic; +static bool hung_task_show_all_bt;static struct task_struct *watchdog_task; +#ifdef CONFIG_SMP +/* +* Should we dump all CPUs backtraces in a hung task event? +* Defaults to 0, can be changed either via cmdline or sysctl. +*/ +unsigned int __read_mostly sysctl_hung_task_all_cpu_backtrace; + +static int __init hung_task_backtrace_setup (char *str) +{ +int rc =kstrtouint (str,0 , &sysctl_hung_task_all_cpu_backtrace); + +if (rc) +return rc; +return 1 ; +} +__setup ("hung_task_all_cpu_backtrace=" , hung_task_backtrace_setup); +#endif /* CONFIG_SMP */ +/* * Should we panic (and reboot, if panic_timeout= is set) when a * hung task is detected:
@@ -137,6 +156,9 @@ static void check_hung_task(struct task_struct *t, unsigned long timeout) " disables this message.\n");sched_show_task (t); hung_task_show_lock = true; + +if (sysctl_hung_task_all_cpu_backtrace) + hung_task_show_all_bt = true; }touch_nmi_watchdog ();
@@ -201,10 +223,14 @@ static void check_hung_uninterruptible_tasks(unsigned long timeout)rcu_read_unlock ();if (hung_task_show_lock)debug_show_all_locks (); -if (hung_task_call_panic) { + +if (hung_task_show_all_bt) { + hung_task_show_all_bt = false;trigger_all_cpu_backtrace (); + } + +if (hung_task_call_panic)panic ("hung_task: blocked tasks" ); - } } static long hung_timeout_jiffies(unsigned long last_checked,
kernel/sysctl.c +11 -0
@@ -1098,6 +1098,17 @@ static struct ctl_table kern_table[] = { },#endif #ifdef CONFIG_DETECT_HUNG_TASK +#ifdef CONFIG_SMP + { + .procname = "hung_task_all_cpu_backtrace", + .data = &sysctl_hung_task_all_cpu_backtrace, + .maxlen = sizeof(int), + .mode = 0644, + .proc_handler = proc_dointvec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = SYSCTL_ONE, + }, +#endif /* CONFIG_SMP */ { .procname = "hung_task_panic", .data = &sysctl_hung_task_panic,
Signature · 1 line
2.25.1
Quoted 37 lines · Guilherme G. Piccoli
Commit
401c636a0eeb("kernel/hung_task.c: show all hung tasks before panic") introduced a change in that we started to show all CPUs backtraces when a hung task is detected _and_ the sysctl/kernel parameter "hung_task_panic" is set. The idea is good, because usually when observing deadlocks (that may lead to hung tasks), the culprit is another task holding a lock and not necessarily the task detected as hung.The problem with this approach is that dumping backtraces is a slightly expensive task, specially printing that on console (and specially in many CPU machines, as servers commonly found nowadays). So, users that plan to collect a kdump to investigate the hung tasks and narrow down the deadlock definitely don't need the CPUs backtrace on dmesg/console, which will delay the panic and pollute the log (crash tool would easily grab all CPUs traces with 'bt -a' command).
Also, there's the reciprocal scenario: some users may be interested in seeing the CPUs backtraces but not have the system panic when a hung task is detected. The current approach hence is almost as embedding a policy in the kernel, by forcing the CPUs backtraces' dump (only) onhung_task_panic.This patch decouples the panic event on hung task from the CPUs backtraces dump, by creating (and documenting) a new sysctl/kernel parameter called "
hung_task_all_cpu_backtrace", analog to the approach taken on soft/hard lockups, that have both a panic and an "all_cpu_backtrace" sysctl to allow individual control. The new mechanism for dumping the CPUs backtraces on hung task detection respects "hung_task_warnings" by not dumping the traces in case there's no warnings left.Cc Tetsuo Handa <[email protected]>Signed-off-by Guilherme G. Piccoli <[email protected]>---
V2: Followed suggestions from Kees and Tetsuo (and other grammar improvements). Also, followed Tetsuo suggestion to itereate kernel testing community - but I don't really know a ML for that, so I've CCed Coccinelle community and kernel-api ML.
Quoted 5 lines
Also, Tetsuo suggested that this option could be default to 1 - I'm open to it, but given it is only available if
hung_taskpanic is set as of now and the goal of this patch is give users more flexibility, I vote to keep default as 0. I can respin a V3 in case more people want to see it enabled by default. Thanks in advance for the review!
Yeah, most things like this we've tried to be conservative. I'd like to see it stay zero too.
-Kees
Quoted 165 lines
Cheers,
Guilherme
.../admin-guide/kernel-parameters.txt | 6 ++++ Documentation/admin-guide/sysctl/kernel.rst | 15 ++++++++++ include/linux/sched/sysctl.h | 7 +++++ kernel/hung_task.c | 30 +++++++++++++++++-- kernel/sysctl.c | 11 +++++++ 5 files changed, 67 insertions(+), 2 deletions(-)Documentation/admin-guide/kernel-parameters.txt +6
-0@@ -1453,6 +1453,12 @@ x86-64 are 2M (when the CPU supports "pse") and 1G (when the CPU supports the "pdpe1gb" cpuinfo flag). + hung_task_all_cpu_backtrace= + [KNL] Should kernel generate backtraces on all cpus + when a hung task is detected. Defaults to 0 and can + be controlled by hung_task_all_cpu_backtrace sysctl. + Format: <integer> + hung_task_panic= [KNL] Should the hung task detector generate panics. Format: <integer>Documentation/admin-guide/sysctl/kernel.rst +15
-0@@ -40,6 +40,7 @@ show up in /proc/sys/kernel: - hotplug - hardlockup_all_cpu_backtrace - hardlockup_panic +- hung_task_all_cpu_backtrace - hung_task_panic - hung_task_check_count - hung_task_timeout_secs@@ -338,6 +339,20 @@ Path for the hotplug policy agent. Default value is "/sbin/hotplug". +hung_task_all_cpu_backtrace: +================ + +If this option is set, the kernel will send an NMI to all CPUs to dump +their backtraces when a hung task is detected. This file shows up if +CONFIG_DETECT_HUNG_TASK and CONFIG_SMP are enabled. + +0: Won't show all CPUs backtraces when a hung task is detected. +This is the default behavior. + +1: Will non-maskably interrupt all CPUs and dump their backtraces when +a hung task is detected. + + hung_task_panic: ================include/linux/sched/sysctl.h +7
-0@@ -7,6 +7,13 @@struct ctl_table ;#ifdef CONFIG_DETECT_HUNG_TASK + +#ifdef CONFIG_SMP +extern unsigned int sysctl_hung_task_all_cpu_backtrace; +#else +#define sysctl_hung_task_all_cpu_backtrace 0 +#endif /* CONFIG_SMP */ +extern int sysctl_hung_task_check_count;extern unsigned int sysctl_hung_task_panic;extern unsigned long sysctl_hung_task_timeout_secs;kernel/hung_task.c +28
-2@@ -53,9 +53,28 @@ int __read_mostly sysctl_hung_task_warnings = 10;static int __read_mostly did_panic;static bool hung_task_show_lock;static bool hung_task_call_panic; +static bool hung_task_show_all_bt;static struct task_struct *watchdog_task; +#ifdef CONFIG_SMP +/* +* Should we dump all CPUs backtraces in a hung task event? +* Defaults to 0, can be changed either via cmdline or sysctl. +*/ +unsigned int __read_mostly sysctl_hung_task_all_cpu_backtrace; + +static int __init hung_task_backtrace_setup (char *str) +{ +int rc =kstrtouint (str,0 , &sysctl_hung_task_all_cpu_backtrace); + +if (rc) +return rc; +return 1 ; +} +__setup ("hung_task_all_cpu_backtrace=" , hung_task_backtrace_setup); +#endif /* CONFIG_SMP */ +/* * Should we panic (and reboot, if panic_timeout= is set) when a * hung task is detected: @@ -137,6 +156,9 @@ static void check_hung_task(struct task_struct *t, unsigned long timeout) " disables this message.\n");sched_show_task (t); hung_task_show_lock = true; + +if (sysctl_hung_task_all_cpu_backtrace) + hung_task_show_all_bt = true; }touch_nmi_watchdog ();@@ -201,10 +223,14 @@ static void check_hung_uninterruptible_tasks(unsigned long timeout)rcu_read_unlock ();if (hung_task_show_lock)debug_show_all_locks (); -if (hung_task_call_panic) { + +if (hung_task_show_all_bt) { + hung_task_show_all_bt = false;trigger_all_cpu_backtrace (); + } + +if (hung_task_call_panic)panic ("hung_task: blocked tasks" ); - } } static long hung_timeout_jiffies(unsigned long last_checked,kernel/sysctl.c +11
-0@@ -1098,6 +1098,17 @@ static struct ctl_table kern_table[] = { },#endif #ifdef CONFIG_DETECT_HUNG_TASK +#ifdef CONFIG_SMP + { + .procname = "hung_task_all_cpu_backtrace", + .data = &sysctl_hung_task_all_cpu_backtrace, + .maxlen = sizeof(int), + .mode = 0644, + .proc_handler = proc_dointvec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = SYSCTL_ONE, + }, +#endif /* CONFIG_SMP */ { .procname = "hung_task_panic", .data = &sysctl_hung_task_panic,Signature · 1 line
2.25.1
Signature · 1 line
Kees Cook
Randy Dunlap
reviewerHi,
Quoted 8 lines · Guilherme G. Piccoli
.../admin-guide/kernel-parameters.txt | 6 ++++ Documentation/admin-guide/sysctl/kernel.rst | 15 ++++++++++ include/linux/sched/sysctl.h | 7 +++++ kernel/hung_task.c | 30 +++++++++++++++++-- kernel/sysctl.c | 11 +++++++ 5 files changed, 67 insertions(+), 2 deletions(-)
admin-guide/kernel-parameters.txt predominantly uses "CPUs" for plural CPUs when not part of a cmdline keyword etc., so please adjust below:
Documentation/admin-guide/kernel-parameters.txt line 1457
Quoted 10 lines
Documentation/admin-guide/kernel-parameters.txt +2
-0@@ -1453,6 +1453,12 @@ x86-64 are 2M (when the CPU supports "pse") and 1G (when the CPU supports the "pdpe1gb" cpuinfo flag). + hung_task_all_cpu_backtrace= + [KNL] Should kernel generate backtraces on all cpus
CPUs
Quoted 30 lines
Documentation/admin-guide/kernel-parameters.txt +4
-0+ when a hung task is detected. Defaults to 0 and can + be controlled by hung_task_all_cpu_backtrace sysctl. + Format: <integer> + hung_task_panic= [KNL] Should the hung task detector generate panics. Format: <integer>Documentation/admin-guide/sysctl/kernel.rst +8
-0@@ -40,6 +40,7 @@ show up in /proc/sys/kernel: - hotplug - hardlockup_all_cpu_backtrace - hardlockup_panic +- hung_task_all_cpu_backtrace - hung_task_panic - hung_task_check_count - hung_task_timeout_secs@@ -338,6 +339,20 @@ Path for the hotplug policy agent. Default value is "/sbin/hotplug". +hung_task_all_cpu_backtrace: +================ + +If this option is set, the kernel will send an NMI to all CPUs to dump +their backtraces when a hung task is detected. This file shows up if +CONFIG_DETECT_HUNG_TASK and CONFIG_SMP are enabled. +
Documentation/admin-guide/sysctl/kernel.rst Path for the hotplug policy agent. · line 358
Quoted 10 lines
Documentation/admin-guide/sysctl/kernel.rst +7
-0+0: Won't show all CPUs backtraces when a hung task is detected. +This is the default behavior. + +1: Will non-maskably interrupt all CPUs and dump their backtraces when +a hung task is detected. + + hung_task_panic: ================
thanks.
Signature · 1 line
~Randy
Quoted 59 lines · Guilherme G. Piccoli
Commit
401c636a0eeb("kernel/hung_task.c: show all hung tasks before panic") introduced a change in that we started to show all CPUs backtraces when a hung task is detected _and_ the sysctl/kernel parameter "hung_task_panic" is set. The idea is good, because usually when observing deadlocks (that may lead to hung tasks), the culprit is another task holding a lock and not necessarily the task detected as hung.The problem with this approach is that dumping backtraces is a slightly expensive task, specially printing that on console (and specially in many CPU machines, as servers commonly found nowadays). So, users that plan to collect a kdump to investigate the hung tasks and narrow down the deadlock definitely don't need the CPUs backtrace on dmesg/console, which will delay the panic and pollute the log (crash tool would easily grab all CPUs traces with 'bt -a' command).
Also, there's the reciprocal scenario: some users may be interested in seeing the CPUs backtraces but not have the system panic when a hung task is detected. The current approach hence is almost as embedding a policy in the kernel, by forcing the CPUs backtraces' dump (only) onhung_task_panic.This patch decouples the panic event on hung task from the CPUs backtraces dump, by creating (and documenting) a new sysctl/kernel parameter called "
hung_task_all_cpu_backtrace", analog to the approach taken on soft/hard lockups, that have both a panic and an "all_cpu_backtrace" sysctl to allow individual control. The new mechanism for dumping the CPUs backtraces on hung task detection respects "hung_task_warnings" by not dumping the traces in case there's no warnings left.Cc Tetsuo Handa <[email protected]>Signed-off-by Guilherme G. Piccoli <[email protected]>---
V2: Followed suggestions from Kees and Tetsuo (and other grammar improvements). Also, followed Tetsuo suggestion to itereate kernel testing community - but I don't really know a ML for that, so I've CCed Coccinelle community and kernel-api ML.
Also, Tetsuo suggested that this option could be default to 1 - I'm open to it, but given it is only available if
hung_taskpanic is set as of now and the goal of this patch is give users more flexibility, I vote to keep default as 0. I can respin a V3 in case more people want to see it enabled by default. Thanks in advance for the review! Cheers,Guilherme
.../admin-guide/kernel-parameters.txt | 6 ++++ Documentation/admin-guide/sysctl/kernel.rst | 15 ++++++++++ include/linux/sched/sysctl.h | 7 +++++ kernel/hung_task.c | 30 +++++++++++++++++-- kernel/sysctl.c | 11 +++++++ 5 files changed, 67 insertions(+), 2 deletions(-)Documentation/admin-guide/kernel-parameters.txt +0
-0
Documentation/admin-guide/kernel-parameters.txt line 1461
Quoted 9 lines
Documentation/admin-guide/kernel-parameters.txt +6
-0@@ -1453,6 +1453,12 @@ x86-64 are 2M (when the CPU supports "pse") and 1G (when the CPU supports the "pdpe1gb" cpuinfo flag). + hung_task_all_cpu_backtrace= + [KNL] Should kernel generate backtraces on all cpus + when a hung task is detected. Defaults to 0 and can + be controlled by hung_task_all_cpu_backtrace sysctl. + Format: <integer> +
Before adding a new thing as both kernel parameter and sysctl, could we perhaps not add the kernel parameter, in favor of the generic sysctl parameter solution?
[1] There were no objections and some support from Kees, so I will try to send a new version ASAP that will work properly with all "static" sysctls - we don't need to be blocked by a full solution for dynamically registered sysctls yet, I guess?
Thanks,
Vlastimil
Quoted 8 lines · Vlastimil Babka
[...]
Before adding a new thing as both kernel parameter and sysctl, could we perhaps not add the kernel parameter, in favor of the generic sysctl parameter solution?
[1] There were no objections and some support from Kees, so I will try to send a new version ASAP that will work properly with all "static" sysctls - we don't need to be blocked by a full solution for dynamically registered sysctls yet, I guess?
Quoted 5 lines
Thanks,
Vlastimil
Thanks Randy and Vlastimil for the comments. I really liked your approach Vlastimil, I agree that we have no reason to not have a generic sysctl setting via cmdline mechanism - I'll rework this patch removing the kernel parameter (same for other patch I just submitted).
If you can CC me on the new iterations of the generic sysctl patches Vlastimil, I appreciate - I can maybe test that, I'd like to see it in kernel.
Thanks,
Guilherme
Quoted 4 lines · Guilherme G. Piccoli
Thanks Randy and Vlastimil for the comments. I really liked your approach Vlastimil, I agree that we have no reason to not have a generic sysctl setting via cmdline mechanism - I'll rework this patch removing the kernel parameter (same for other patch I just submitted).
I've been thinking we'll likely want to have a big patch series that removes all the old "duplicate" boot params and adds some kind of "alias" mechanism.
Vlastimil, have you happened to keep a list of other "redundant" boot params you've noticed in the kernel? I bet there are a lot. :)
Signature · 1 line
Kees Cook
Quoted 6 lines · Kees Cook
Quoted 4 lines · Guilherme G. Piccoli
Thanks Randy and Vlastimil for the comments. I really liked your approach Vlastimil, I agree that we have no reason to not have a generic sysctl setting via cmdline mechanism - I'll rework this patch removing the kernel parameter (same for other patch I just submitted).
Quoted 6 lines
I've been thinking we'll likely want to have a big patch series that removes all the old "duplicate" boot params and adds some kind of "alias" mechanism.
Vlastimil, have you happened to keep a list of other "redundant" boot params you've noticed in the kernel? I bet there are a lot. :)
Well, I found about 4 that mentioned sysctl in Documentation/admin-guide/kernel-parameters.txt
I suspect there will be more, but won't be trivial to identify them.