From: Con Kolivas Here is a patch against the current O1int patch in 2.5.74-mm1. Since the O1int didn't mean anything I thought I'd call this O2int. This one wont blow you away but tames those corner cases. Changes: The child penalty is set on 80% which means that tasks that wait on their children have children forking just on the edge of the interactive delta so they shouldn't starve their own children. The non linear sleep avg boost is scaled down slightly to prevent this particular boost from being capable of making a task highly interactive. This makes very new tasks less likely to have a little spurt of too high priority. Idle tasks now get their static priority over the full time they've been running rather than starting again at 1 second. This makes it harder for idle tasks to suddenly become highly interactive and _then_ fork an interactive bomb. Not sure on this one yet. The sched_exit penalty to parents of cpu hungry children is scaled accordingly (was missed on the original conversion so works better now). Hysteresis on interactive buffer removed (was unecessary). Minor cleanup. Known issue remaining: Mozilla acts just like X in that it is mostly interactive but has bursts of heavy cpu activity so it gets the same bonus as X. However it makes X jerky during it's heavy cpu activity, and might in some circumstances make audio skip. Fixing this kills X smoothness as they seem very similar to the estimator. Still haven't sorted a workaround for this one but I'm working on it. Ingo's original timeslice granularity patch helps a little and may be worth resuscitating (and the desktop only people can change the granularity down to 10ms to satisfy their needs). kernel/sched.c | 21 +++++++++++---------- 1 files changed, 11 insertions(+), 10 deletions(-) diff -puN kernel/sched.c~o2int kernel/sched.c --- 25/kernel/sched.c~o2int 2003-07-03 22:10:21.000000000 -0700 +++ 25-akpm/kernel/sched.c 2003-07-03 22:10:21.000000000 -0700 @@ -68,7 +68,7 @@ */ #define MIN_TIMESLICE ( 10 * HZ / 1000) #define MAX_TIMESLICE (200 * HZ / 1000) -#define CHILD_PENALTY 50 +#define CHILD_PENALTY 80 #define PARENT_PENALTY 100 #define EXIT_WEIGHT 3 #define PRIO_BONUS_RATIO 25 @@ -405,30 +405,30 @@ static inline void activate_task(task_t * from continually getting larger. */ if (runtime < MAX_SLEEP_AVG) - p->sleep_avg += (runtime - p->sleep_avg) * (MAX_SLEEP_AVG - runtime) / MAX_SLEEP_AVG; + p->sleep_avg += (runtime - p->sleep_avg) * (MAX_SLEEP_AVG - runtime) * + (10 - INTERACTIVE_DELTA) / 10 / MAX_SLEEP_AVG; /* - * Keep a buffer of 10-20% bonus sleep_avg with hysteresis + * Keep a buffer of 10% sleep_avg * to prevent short bursts of cpu activity from making * interactive tasks lose their bonus */ - if (p->sleep_avg > MAX_SLEEP_AVG * 12/10) + if (p->sleep_avg > MAX_SLEEP_AVG * 11/10) p->sleep_avg = MAX_SLEEP_AVG * 11/10; /* * Tasks that sleep a long time are categorised as idle and * get their static priority only */ - if (sleep_time > MIN_SLEEP_AVG){ - p->avg_start = jiffies - MIN_SLEEP_AVG; - p->sleep_avg = MIN_SLEEP_AVG / 2; - } + if (sleep_time > MIN_SLEEP_AVG) + p->sleep_avg = runtime / 2; + if (unlikely(p->avg_start > jiffies)){ p->avg_start = jiffies; p->sleep_avg = 0; } - p->prio = effective_prio(p); } + p->prio = effective_prio(p); __activate_task(p, rq); } @@ -605,7 +605,6 @@ void wake_up_forked_process(task_t * p) * from forking tasks that are max-interactive. */ current->sleep_avg = current->sleep_avg * PARENT_PENALTY / 100; - p->avg_start = current->avg_start; normalise_sleep(p); p->sleep_avg = p->sleep_avg * CHILD_PENALTY / 100; p->prio = effective_prio(p); @@ -647,6 +646,8 @@ void sched_exit(task_t * p) * If the child was a (relative-) CPU hog then decrease * the sleep_avg of the parent as well. */ + normalise_sleep(p); + normalise_sleep(p->parent); if (p->sleep_avg < p->parent->sleep_avg) p->parent->sleep_avg = (p->parent->sleep_avg * EXIT_WEIGHT + p->sleep_avg) / (EXIT_WEIGHT + 1); _