dump_stack: allow specifying printk log level

This patchset lays the foundation work to allow using dump_stack() with a specified printk log level. Currently each architecture uses a different log level in show_stack() and it's not possible to control it without calling directly architecture specific functions.

The motivation behind this work is to limit the amount of kernel messages printed in the console when a process is killed by the OOM killer. In some scenarios (lots of containers running different customers' workloads) OOMs are way more common and don't require the console to be flooded by stack traces when the OOM killer probably did the right choice. During a recent discussion it was determined that a knob to control when dump_stack() is called is a bad idea and instead we should tune the log level in dump_stack() which prompted this work.

This patchset introduces two new functions:
dump_stack_lvl(char *log_lvl)
show_stack_lvl(struct task_struct *task, unsigned long *sp, char *log_lvl)

and both can be reimplemented by each architecture but only the second is expected. The idea is to initially implement show_stack_lvl() in all architectures then simply have show_stack() to require log_lvl as parameter. While that happens, dump_stack() uses can be changed to dump_stack_lvl() and once everything is in place, dump_stack() will require the log_level as well.

I have a draft patch for every architecture but for this patchset I'm only including x86 to get some feedback while I try to get a cross compiler working for each one of them (which is being harder than I thought).

Aristeu

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

This is in preparation for dump_stack_lvl() which will allow the log level to be passed.

---
 lib/dump_stack.c |   16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)

@@ -9,9 +9,9 @@
 #include <linux/smp.h>
 #include <linux/atomic.h>
 
-static void __dump_stack(void)
+static void __dump_stack(char *log_lvl)
 {
-	dump_stack_print_info(KERN_DEFAULT);
+	dump_stack_print_info(log_lvl);
 	show_stack(NULL, NULL);
 }
 

@@ -23,7 +23,7 @@ static void __dump_stack(void)
 #ifdef CONFIG_SMP
 static atomic_t dump_lock = ATOMIC_INIT(-1);
 
-asmlinkage __visible void dump_stack(void)
+static void _dump_stack(char *log_lvl)
 {
 	int was_locked;
 	int old;

@@ -47,17 +47,23 @@ was_locked = 0;
 		goto retry;
 	}
 
-	__dump_stack();
+	__dump_stack(log_lvl);
 
 	if (!was_locked)
 		atomic_set(&dump_lock, -1);
 
 	preempt_enable();
 }
+
+asmlinkage __visible void dump_stack(void)
+{
+	_dump_stack(KERN_DEFAULT);
+}
+
 #else
 asmlinkage __visible void dump_stack(void)
 {
-	__dump_stack();
+	__dump_stack(KERN_DEFAULT);
 }
 #endif
 EXPORT_SYMBOL(dump_stack);

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

dump_stack_lvl() allows passing the log level down to the architecture specific code.

---
 include/linux/printk.h |    1 +
 lib/dump_stack.c       |   11 +++++++++++
 2 files changed, 12 insertions(+)

@@ -55,6 +55,11 @@ was_locked = 0;
 	preempt_enable();
 }
 
+asmlinkage __visible void dump_stack_lvl(char *log_lvl)
+{
+	_dump_stack(log_lvl);
+}
+
 asmlinkage __visible void dump_stack(void)
 {
 	_dump_stack(KERN_DEFAULT);

@@ -65,5 +70,11 @@ asmlinkage __visible void dump_stack(voi
 {
 	__dump_stack(KERN_DEFAULT);
 }
+
+asmlinkage __visible void dump_stack_lvl(char *log_lvl)
+{
+	__dump_stack(log_lvl);
+}
 #endif
 EXPORT_SYMBOL(dump_stack);
+EXPORT_SYMBOL(dump_stack_lvl);

@@ -231,6 +231,7 @@ static inline void show_regs_print_info(
 #endif
 
 extern asmlinkage void dump_stack(void) __cold;
+extern asmlinkage void dump_stack_lvl(char *log_lvl);
 
 #ifndef pr_fmt
 #define pr_fmt(fmt) fmt

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

show_stack_lvl() works like show_stack() but allows also passing the log level. The default implementation that should be overrided by architecture specific code simply will call the existing show_stack().

---
 include/linux/sched.h |    2 ++
 lib/dump_stack.c      |    8 +++++++-
 2 files changed, 9 insertions(+), 1 deletion(-)

@@ -9,10 +9,16 @@
 #include <linux/smp.h>
 #include <linux/atomic.h>
 
+void __weak show_stack_lvl(struct task_struct *task, unsigned long *sp,
+			       char *log_lvl)
+{
+	return show_stack(task, sp);
+}
+
 static void __dump_stack(char *log_lvl)
 {
 	dump_stack_print_info(log_lvl);
-	show_stack(NULL, NULL);
+	show_stack_lvl(NULL, NULL, log_lvl);
 }
 
 /**

@@ -368,6 +368,8 @@ extern void show_regs(struct pt_regs *);
  * trace (or NULL if the entire call-chain of the task should be shown).
  */
 extern void show_stack(struct task_struct *task, unsigned long *sp);
+extern void show_stack_lvl(struct task_struct *task, unsigned long *sp,
+			   char *log_lvl);
 
 extern void cpu_init (void);
 extern void trap_init(void);

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

show_stack_lvl() allows passing the log level and is used by dump_stack_lvl().

---
 arch/x86/kernel/dumpstack.c |    9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

@@ -180,7 +180,7 @@ void show_trace(struct task_struct *task
 	show_trace_log_lvl(task, regs, stack, bp, "");
 }
 
-void show_stack(struct task_struct *task, unsigned long *sp)
+void show_stack_lvl(struct task_struct *task, unsigned long *sp, char *log_lvl)
 {
 	unsigned long bp = 0;
 	unsigned long stack;

@@ -194,7 +194,12 @@ unsigned long bp = 0;
 		bp = stack_frame(current, NULL);
 	}
 
-	show_stack_log_lvl(task, NULL, sp, bp, "");
+	show_stack_log_lvl(task, NULL, sp, bp, log_lvl);
+}
+
+void show_stack(struct task_struct *task, unsigned long *sp)
+{
+	show_stack_lvl(task, sp, KERN_DEFAULT);
 }
 
 static arch_spinlock_t die_lock = __ARCH_SPIN_LOCK_UNLOCKED;

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

dump_stack() isn't always useful and in some scenarios OOMs can be quite common and there's no need to flood the console with dump_stack()'s output.

---
 mm/oom_kill.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

@@ -384,7 +384,7 @@ pr_warning("%s invoked oom-killer: gfp_m
 		current->signal->oom_score_adj);
 	cpuset_print_task_mems_allowed(current);
 	task_unlock(current);
-	dump_stack();
+	dump_stack_lvl(KERN_DEBUG);
 	if (memcg)
 		mem_cgroup_print_oom_info(memcg, p);
 	else

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

[Hmm, this got stuck in my outgoing queue for some reason - sending again]

I think we want to revisit loglevel of other parts of the oom report as well but this is a good start.

Michal Hocko
SUSE Labs

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

This looks good to me FWIW.

Michal Hocko
SUSE Labs

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Seems reasonable to me as well and yes, there will be extra fill-in work to do.

The "lvl" thing stands out - kernel code doesn't do this arbitrary vowelicide to make identifiers shorter. s/lvl/level/g?

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

I have some of the archs already converted, will wait until this initial patchset is in to submit.

There are show_regs_log_lvl(), show_trace_log_lvl(), show_stack_log_lvl() (avr32, x86) and I just wanted to keep the naming consistent. If you're strong about this I can change it and resubmit.

Aristeu

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>