Export CPU Scheduler Tunables via DebugFS
PATCH 1/1 Export CPU Scheduler Tunables via DebugFS
This patch exports the CPU scheduler tunables via DebugFS.
philips@plankton:~$ ls /debug/cpu-scheduler/ child_penalty interactive_delta min_timeslice prio_bonus_ratio def_timeslice max_bonus on_runqueue_weight starvation_limit exit_weight max_sleep_avg parent_penalty
It is similiar to the patches by Robert Love and Ingo Molnar that were available for the 2.5 series: http://kerneltrap.org/node/525/1938
The goal remains the same: offer a simple way for _developers_ to tune and debug the scheduler.
Perhaps someone could even automate the testing of different values[1] under different workloads[2].
[1] http://meikon.homeip.net/extras/papers/linux-2.5-scheduler-analysis.pdf
[2] http://test.kernel.org/autotest/QuickStart
--- kernel/sched.c | 115 ++++++++++++++++++++++++++++++++++++++++++++++++------ lib/Kconfig.debug | 8 +++ 2 files changed, 112 insertions(+), 11 deletions(-)
kernel/sched.c +104 -11
@@ -52,6 +52,7 @@#include <linux/acct.h> #include <linux/kprobes.h> #include <linux/delayacct.h> +#include <linux/debugfs.h> #include <asm/tlb.h> #include <asm/unistd.h>
@@ -87,17 +88,60 @@* default timeslice is 100 msecs, maximum timeslice is 800 msecs. * Timeslices get refilled after they expire. */ -#define MIN_TIMESLICE max(5 * HZ / 1000, 1) -#define DEF_TIMESLICE (100 * HZ / 1000) -#define ON_RUNQUEUE_WEIGHT 30 -#define CHILD_PENALTY 95 -#define PARENT_PENALTY 100 -#define EXIT_WEIGHT 3 -#define PRIO_BONUS_RATIO 25 -#define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100) -#define INTERACTIVE_DELTA 2 -#define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS) -#define STARVATION_LIMIT (MAX_SLEEP_AVG) +#define DEFAULT_MIN_TIMESLICE max(5 * HZ / 1000, 1) +#define DEFAULT_DEF_TIMESLICE (100 * HZ / 1000) +#define DEFAULT_ON_RUNQUEUE_WEIGHT 30 +#define DEFAULT_CHILD_PENALTY 95 +#define DEFAULT_PARENT_PENALTY 100 +#define DEFAULT_EXIT_WEIGHT 3 +#define DEFAULT_PRIO_BONUS_RATIO 25 +#define DEFAULT_MAX_BONUS (MAX_USER_PRIO * DEFAULT_PRIO_BONUS_RATIO / 100) +#define DEFAULT_INTERACTIVE_DELTA 2 +#define DEFAULT_MAX_SLEEP_AVG (DEFAULT_DEF_TIMESLICE * DEFAULT_MAX_BONUS) +#define DEFAULT_STARVATION_LIMIT (DEFAULT_MAX_SLEEP_AVG) +#define DEFAULT_NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(DEFAULT_MAX_SLEEP_AVG)) + +/* +* Simply use the static definitions if we are not exporting tunables via +* DebugFS for runtime tuning. +*/ +#ifndef CONFIG_DEBUGFS_SCHED +#define MIN_TIMESLICE DEFAULT_MIN_TIMESLICE +#define DEF_TIMESLICE DEFAULT_DEF_TIMESLICE +#define ON_RUNQUEUE_WEIGHT DEFAULT_ON_RUNQUEUE_WEIGHT +#define CHILD_PENALTY DEFAULT_CHILD_PENALTY +#define PARENT_PENALTY DEFAULT_PARENT_PENALTY +#define EXIT_WEIGHT DEFAULT_EXIT_WEIGHT +#define PRIO_BONUS_RATIO DEFAULT_PRIO_BONUS_RATIO +#define MAX_BONUS DEFAULT_MAX_BONUS +#define INTERACTIVE_DELTA DEFAULT_INTERACTIVE_DELTA +#define MAX_SLEEP_AVG DEFAULT_MAX_SLEEP_AVG +#define STARVATION_LIMIT DEFAULT_STARVATION_LIMIT +#else +int min_timeslice; +#define MIN_TIMESLICE (min_timeslice) +int def_timeslice; +#define DEF_TIMESLICE (def_timeslice) +int on_runqueue_weight; +#define ON_RUNQUEUE_WEIGHT (on_runqueue_weight) +int child_penalty; +#define CHILD_PENALTY (child_penalty) +int parent_penalty; +#define PARENT_PENALTY (parent_penalty) +int exit_weight; +#define EXIT_WEIGHT (exit_weight) +int prio_bonus_ratio; +#define PRIO_BONUS_RATIO (prio_bonus_ratio) +int max_bonus; +#define MAX_BONUS (max_bonus) +int interactive_delta; +#define INTERACTIVE_DELTA (interactive_delta) +int max_sleep_avg; +#define MAX_SLEEP_AVG (max_sleep_avg) +int starvation_limit; +#define STARVATION_LIMIT (starvation_limit) +#endif +#define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))/*
@@ -6725,10 +6769,59 @@ int in_sched_functions(unsigned long add && addr < (unsigned long)__sched_text_end); } +#ifdef CONFIG_DEBUGFS_SCHED + +#define debugfs_sched_create (name, parent) \ +debugfs_create_u32(#name, 0644, parent, &name) + +int __init init_debugfs (void ) +{ +struct dentry * root; + +if ((root =debugfs_create_dir ("cpu-scheduler" ,NULL )) ==NULL ) + goto err_root; + +debugfs_sched_create (min_timeslice, root); +debugfs_sched_create (def_timeslice, root); +debugfs_sched_create (on_runqueue_weight, root); +debugfs_sched_create (child_penalty, root); +debugfs_sched_create (parent_penalty, root); +debugfs_sched_create (exit_weight, root); +debugfs_sched_create (prio_bonus_ratio, root); +debugfs_sched_create (max_bonus, root); +debugfs_sched_create (interactive_delta , root); +debugfs_sched_create (max_sleep_avg, root); +debugfs_sched_create (starvation_limit, root); + +err_root: +return 0 ; +} +postcore_initcall (init_debugfs); + +void __init init_tunables (void ) +{ + min_timeslice =DEFAULT_MIN_TIMESLICE ; + def_timeslice =DEFAULT_DEF_TIMESLICE ; + on_runqueue_weight =DEFAULT_ON_RUNQUEUE_WEIGHT ; + child_penalty =DEFAULT_CHILD_PENALTY ; + parent_penalty =DEFAULT_PARENT_PENALTY ; + exit_weight =DEFAULT_EXIT_WEIGHT ; + prio_bonus_ratio =DEFAULT_PRIO_BONUS_RATIO ; + max_bonus =DEFAULT_MAX_BONUS ; + interactive_delta =DEFAULT_INTERACTIVE_DELTA ; + max_sleep_avg =DEFAULT_MAX_SLEEP_AVG ; + starvation_limit =DEFAULT_STARVATION_LIMIT ; +} +#endif +void __init sched_init (void ) {int i, j, k; +#ifdef CONFIG_DEBUGFS_SCHED +init_tunables (); +#endif +for_each_possible_cpu (i) {struct prio_array *array;struct rq *rq;
lib/Kconfig.debug +8 -0
@@ -93,6 +93,14 @@ config SCHEDSTATS application, you can say N to avoid the very slight overhead this adds. +config DEBUGFS_SCHED + bool "Export CPU Scheduler Tunables" + depends on DEBUG_KERNEL && DEBUG_FS + help + If you say Y here, CPU Scheduler tuning knobs will be available via + debugfs in /debug/cpu-scheduler/. Note: No sanity checking is done + on the values. + config DEBUG_SLAB bool "Debug slab memory allocations" depends on DEBUG_KERNEL && SLAB
--
Signature · 1 line
VGER BF report: H 1.49591e-07
Hi,
Quoted 15 lines · Brandon Philips
This patch exports the CPU scheduler tunables via DebugFS.
kernel/sched.c +12
-0+int min_timeslice; +#define MIN_TIMESLICE (min_timeslice) +int def_timeslice; +#define DEF_TIMESLICE (def_timeslice) +int on_runqueue_weight; +#define ON_RUNQUEUE_WEIGHT (on_runqueue_weight) +int child_penalty; +#define CHILD_PENALTY (child_penalty) +int parent_penalty; +#define PARENT_PENALTY (parent_penalty) +int exit_weight; +#define EXIT_WEIGHT (exit_weight)
kernel/sched.c line 142
Quoted 10 lines
kernel/sched.c +10
-0+int prio_bonus_ratio; +#define PRIO_BONUS_RATIO (prio_bonus_ratio) +int max_bonus; +#define MAX_BONUS (max_bonus) +int interactive_delta; +#define INTERACTIVE_DELTA (interactive_delta) +int max_sleep_avg; +#define MAX_SLEEP_AVG (max_sleep_avg) +int starvation_limit; +#define STARVATION_LIMIT (starvation_limit)
__read_mostly? Except for those variables which are being tweaked quite often by live scheduler code... (probably none)
This only concerns the case of CONFIG_DEBUGFS_SCHED being activated, but it may still be nice to have, I think, since scheduler stuff should rather be fast than slow.
But since it is a debugging-only option omitting __read_mostly may make sense...
Andreas Mohr
Signature · 1 line
VGER BF report: H 0.00129204