4.8 muqss

Documentation/scheduler/sched-MuQSS.txt

@@ -0,0 +1,78 @@
+MuQSS - The Multiple Queue Skiplist Scheduler by Con Kolivas.
+
+See sched-BFS.txt for basic design; MuQSS is a per-cpu runqueue variant with
+one 8 level skiplist per runqueue, and fine grained locking for much more
+scalability.
+
+Goals.
+
+The goal of the Multiple Queue Skiplist Scheduler, referred to as MuQSS from
+here on (pronounced mux) is to completely do away with the complex designs of
+the past for the cpu process scheduler and instead implement one that is very
+simple in basic design. The main focus of MuQSS is to achieve excellent desktop
+interactivity and responsiveness without heuristics and tuning knobs that are
+difficult to understand, impossible to model and predict the effect of, and when
+tuned to one workload cause massive detriment to another, while still being
+scalable to many CPUs and processes.
+
+
+Design summary.
+
+MuQSS is best described as per-cpu multiple runqueue, O(log n) insertion, O(1)
+lookup, earliest effective virtual deadline first design, loosely based on EEVDF
+(earliest eligible virtual deadline first) and my previous Staircase Deadline
+scheduler, and evolved from the single runqueue O(n) BFS scheduler. Each
+component shall be described in order to understand the significance of, and
+reasoning for it.
+
+
+Design reasoning.
+
+In BFS, the use of a single runqueue across all CPUs meant that each CPU would
+need to scan the entire runqueue looking for the process with the earliest
+deadline and schedule that next, regardless of which CPU it originally came
+from. This made BFS deterministic with respect to latency and provided
+guaranteed latencies dependent on number of processes and CPUs. The single
+runqueue, however, meant that all CPUs would compete for the single lock
+protecting it, which would lead to increasing lock contention as the number of
+CPUs rose and appeared to limit scalability of common workloads beyond 16
+logical CPUs. Additionally, the O(n) lookup of the runqueue list obviously
+increased overhead proportionate to the number of queued proecesses and led to
+cache thrashing while iterating over the linked list.
+
+MuQSS is an evolution of BFS, designed to maintain the same scheduling
+decision mechanism and be virtually deterministic without relying on the
+constrained design of the single runqueue by splitting out the single runqueue
+to be per-CPU and use skiplists instead of linked lists.
+
+The original reason for going back to a single runqueue design for BFS was that
+once multiple runqueues are introduced, per-CPU or otherwise, there will be
+complex interactions as each runqueue will be responsible for the scheduling
+latency and fairness of the tasks only on its own runqueue, and to achieve
+fairness and low latency across multiple CPUs, any advantage in throughput of
+having CPU local tasks causes other disadvantages. This is due to requiring a
+very complex balancing system to at best achieve some semblance of fairness
+across CPUs and can only maintain relatively low latency for tasks bound to the
+same CPUs, not across them. To increase said fairness and latency across CPUs,
+the advantage of local runqueue locking, which makes for better scalability, is
+lost due to having to grab multiple locks.
+
+MuQSS works around the problems inherent in multiple runqueue designs by
+making its skip lists priority ordered and through novel use of lockless
+examination of each other runqueue it can decide if it should take the earliest
+deadline task from another runqueue for latency reasons, or for CPU balancing
+reasons. It still does not have a balancing system, choosing to allow the
+next task scheduling decision and task wakeup CPU choice to allow balancing to
+happen by virtue of its choices.
+
+
+Design:
+
+MuQSS is an 8 level skip list per runqueue variant of BFS.
+
+See sched-BFS.txt for some of the shared design details.
+
+Documentation yet to be completed.
+
+
+Con Kolivas <kernel@kolivas.org> Sun, 2nd October 2016

Documentation/sysctl/kernel.txt

@@ -39,6 +39,7 @@ show up in /proc/sys/kernel:
 - hung_task_timeout_secs
 - hung_task_warnings
 - kexec_load_disabled
+- iso_cpu
 - kptr_restrict
 - kstack_depth_to_print       [ X86 only ]
 - l2cr                        [ PPC only ]
@@ -73,6 +74,7 @@ show up in /proc/sys/kernel:
 - randomize_va_space
 - real-root-dev               ==> Documentation/initrd.txt
 - reboot-cmd                  [ SPARC only ]
+- rr_interval
 - rtsig-max
 - rtsig-nr
 - sem
@@ -402,6 +404,16 @@ kernel stack.
 
 ==============================================================
 
+iso_cpu: (MuQSS CPU scheduler only).
+
+This sets the percentage cpu that the unprivileged SCHED_ISO tasks can
+run effectively at realtime priority, averaged over a rolling five
+seconds over the -whole- system, meaning all cpus.
+
+Set to 70 (percent) by default.
+
+==============================================================
+
 l2cr: (PPC only)
 
 This flag controls the L2 cache of G3 processor boards. If
@@ -818,6 +830,20 @@ rebooting. ???
 
 ==============================================================
 
+rr_interval: (MuQSS CPU scheduler only)
+
+This is the smallest duration that any cpu process scheduling unit
+will run for. Increasing this value can increase throughput of cpu
+bound tasks substantially but at the expense of increased latencies
+overall. Conversely decreasing it will decrease average and maximum
+latencies but at the expense of throughput. This value is in
+milliseconds and the default value chosen depends on the number of
+cpus available at scheduler initialisation with a minimum of 6.
+
+Valid values are from 1-1000.
+
+==============================================================
+
 rtsig-max & rtsig-nr:
 
 The file rtsig-max can be used to tune the maximum number

arch/powerpc/platforms/cell/spufs/sched.c

@@ -63,11 +63,6 @@ static struct task_struct *spusched_task;
 static struct timer_list spusched_timer;
 static struct timer_list spuloadavg_timer;
 
-/*
- * Priority of a normal, non-rt, non-niced'd process (aka nice level 0).
- */
-#define NORMAL_PRIO		120
-
 /*
  * Frequency of the spu scheduler tick.  By default we do one SPU scheduler
  * tick for every 10 CPU scheduler ticks.

arch/x86/Kconfig

@@ -914,10 +914,26 @@ config SCHED_SMT
 	depends on SMP
 	---help---
 	  SMT scheduler support improves the CPU scheduler's decision making
-	  when dealing with Intel Pentium 4 chips with HyperThreading at a
+	  when dealing with Intel P4/Core 2 chips with HyperThreading at a
 	  cost of slightly increased overhead in some places. If unsure say
 	  N here.
 
+config SMT_NICE
+	bool "SMT (Hyperthreading) aware nice priority and policy support"
+	depends on SCHED_MUQSS && SCHED_SMT
+	default y
+	---help---
+	  Enabling Hyperthreading on Intel CPUs decreases the effectiveness
+	  of the use of 'nice' levels and different scheduling policies
+	  (e.g. realtime) due to sharing of CPU power between hyperthreads.
+	  SMT nice support makes each logical CPU aware of what is running on
+	  its hyperthread siblings, maintaining appropriate distribution of
+	  CPU according to nice levels and scheduling policies at the expense
+	  of slightly increased overhead.
+
+	  If unsure say Y here.
+
+
 config SCHED_MC
 	def_bool y
 	prompt "Multi-core scheduler support"
@@ -2036,7 +2052,7 @@ config HOTPLUG_CPU
 config BOOTPARAM_HOTPLUG_CPU0
 	bool "Set default setting of cpu0_hotpluggable"
 	default n
-	depends on HOTPLUG_CPU
+	depends on HOTPLUG_CPU && !SCHED_MUQSS
 	---help---
 	  Set whether default state of cpu0_hotpluggable is on or off.
 
@@ -2065,7 +2081,7 @@ config BOOTPARAM_HOTPLUG_CPU0
 config DEBUG_HOTPLUG_CPU0
 	def_bool n
 	prompt "Debug CPU0 hotplug"
-	depends on HOTPLUG_CPU
+	depends on HOTPLUG_CPU && !SCHED_MUQSS
 	---help---
 	  Enabling this option offlines CPU0 (if CPU0 can be offlined) as
 	  soon as possible and boots up userspace with CPU0 offlined. User

drivers/cpufreq/cpufreq_conservative.c

@@ -30,8 +30,8 @@ struct cs_dbs_tuners {
 };
 
 /* Conservative governor macros */
-#define DEF_FREQUENCY_UP_THRESHOLD		(80)
-#define DEF_FREQUENCY_DOWN_THRESHOLD		(20)
+#define DEF_FREQUENCY_UP_THRESHOLD		(63)
+#define DEF_FREQUENCY_DOWN_THRESHOLD		(26)
 #define DEF_FREQUENCY_STEP			(5)
 #define DEF_SAMPLING_DOWN_FACTOR		(1)
 #define MAX_SAMPLING_DOWN_FACTOR		(10)

drivers/cpufreq/cpufreq_ondemand.c

@@ -20,7 +20,7 @@
 #include "cpufreq_ondemand.h"
 
 /* On-demand governor macros */
-#define DEF_FREQUENCY_UP_THRESHOLD		(80)
+#define DEF_FREQUENCY_UP_THRESHOLD		(63)
 #define DEF_SAMPLING_DOWN_FACTOR		(1)
 #define MAX_SAMPLING_DOWN_FACTOR		(100000)
 #define MICRO_FREQUENCY_UP_THRESHOLD		(95)
@@ -129,7 +129,7 @@ static void dbs_freq_increase(struct cpufreq_policy *policy, unsigned int freq)
 }
 
 /*
- * Every sampling_rate, we check, if current idle time is less than 20%
+ * Every sampling_rate, we check, if current idle time is less than 37%
  * (default), then we try to increase frequency. Else, we adjust the frequency
  * proportional to load.
  */

fs/proc/base.c

@@ -505,7 +505,7 @@ static int proc_pid_schedstat(struct seq_file *m, struct pid_namespace *ns,
 		seq_printf(m, "0 0 0\n");
 	else
 		seq_printf(m, "%llu %llu %lu\n",
-		   (unsigned long long)task->se.sum_exec_runtime,
+		   (unsigned long long)tsk_seruntime(task),
 		   (unsigned long long)task->sched_info.run_delay,
 		   task->sched_info.pcount);
 

include/linux/init_task.h

@@ -157,8 +157,6 @@ extern struct task_group root_task_group;
 # define INIT_VTIME(tsk)
 #endif
 
-#define INIT_TASK_COMM "swapper"
-
 #ifdef CONFIG_RT_MUTEXES
 # define INIT_RT_MUTEXES(tsk)						\
 	.pi_waiters = RB_ROOT,						\
@@ -187,6 +185,77 @@ extern struct task_group root_task_group;
  *  INIT_TASK is used to set up the first task table, touch at
  * your own risk!. Base=0, limit=0x1fffff (=2MB)
  */
+#ifdef CONFIG_SCHED_MUQSS
+#define INIT_TASK_COMM "MuQSS"
+#define INIT_TASK(tsk)	\
+{									\
+	.state		= 0,						\
+	.stack		= &init_thread_info,				\
+	.usage		= ATOMIC_INIT(2),				\
+	.flags		= PF_KTHREAD,					\
+	.prio		= NORMAL_PRIO,					\
+	.static_prio	= MAX_PRIO-20,					\
+	.normal_prio	= NORMAL_PRIO,					\
+	.deadline	= 0,						\
+	.policy		= SCHED_NORMAL,					\
+	.cpus_allowed	= CPU_MASK_ALL,					\
+	.mm		= NULL,						\
+	.active_mm	= &init_mm,					\
+	.restart_block = {						\
+		.fn = do_no_restart_syscall,				\
+	},								\
+	.time_slice	= 1000000,					\
+	.tasks		= LIST_HEAD_INIT(tsk.tasks),			\
+	INIT_PUSHABLE_TASKS(tsk)					\
+	.ptraced	= LIST_HEAD_INIT(tsk.ptraced),			\
+	.ptrace_entry	= LIST_HEAD_INIT(tsk.ptrace_entry),		\
+	.real_parent	= &tsk,						\
+	.parent		= &tsk,						\
+	.children	= LIST_HEAD_INIT(tsk.children),			\
+	.sibling	= LIST_HEAD_INIT(tsk.sibling),			\
+	.group_leader	= &tsk,						\
+	RCU_POINTER_INITIALIZER(real_cred, &init_cred),			\
+	RCU_POINTER_INITIALIZER(cred, &init_cred),			\
+	.comm		= INIT_TASK_COMM,				\
+	.thread		= INIT_THREAD,					\
+	.fs		= &init_fs,					\
+	.files		= &init_files,					\
+	.signal		= &init_signals,				\
+	.sighand	= &init_sighand,				\
+	.nsproxy	= &init_nsproxy,				\
+	.pending	= {						\
+		.list = LIST_HEAD_INIT(tsk.pending.list),		\
+		.signal = {{0}}},					\
+	.blocked	= {{0}},					\
+	.alloc_lock	= __SPIN_LOCK_UNLOCKED(tsk.alloc_lock),		\
+	.journal_info	= NULL,						\
+	.cpu_timers	= INIT_CPU_TIMERS(tsk.cpu_timers),		\
+	.pi_lock	= __RAW_SPIN_LOCK_UNLOCKED(tsk.pi_lock),		\
+	.timer_slack_ns = 50000, /* 50 usec default slack */		\
+	.pids = {							\
+		[PIDTYPE_PID]  = INIT_PID_LINK(PIDTYPE_PID),		\
+		[PIDTYPE_PGID] = INIT_PID_LINK(PIDTYPE_PGID),		\
+		[PIDTYPE_SID]  = INIT_PID_LINK(PIDTYPE_SID),		\
+	},								\
+	.thread_group	= LIST_HEAD_INIT(tsk.thread_group),		\
+	.thread_node	= LIST_HEAD_INIT(init_signals.thread_head),	\
+	INIT_IDS							\
+	INIT_PERF_EVENTS(tsk)						\
+	INIT_TRACE_IRQFLAGS						\
+	INIT_LOCKDEP							\
+	INIT_FTRACE_GRAPH						\
+	INIT_TRACE_RECURSION						\
+	INIT_TASK_RCU_PREEMPT(tsk)					\
+	INIT_TASK_RCU_TASKS(tsk)					\
+	INIT_CPUSET_SEQ(tsk)						\
+	INIT_RT_MUTEXES(tsk)						\
+	INIT_PREV_CPUTIME(tsk)						\
+	INIT_VTIME(tsk)							\
+	INIT_NUMA_BALANCING(tsk)					\
+	INIT_KASAN(tsk)							\
+}
+#else /* CONFIG_SCHED_MUQSS */
+#define INIT_TASK_COMM "swapper"
 #define INIT_TASK(tsk)	\
 {									\
 	.state		= 0,						\
@@ -261,7 +330,7 @@ extern struct task_group root_task_group;
 	INIT_NUMA_BALANCING(tsk)					\
 	INIT_KASAN(tsk)							\
 }
-
+#endif /* CONFIG_SCHED_MUQSS */
 
 #define INIT_CPU_TIMERS(cpu_timers)					\
 {									\

include/linux/ioprio.h

@@ -52,6 +52,8 @@ enum {
  */
 static inline int task_nice_ioprio(struct task_struct *task)
 {
+	if (iso_task(task))
+		return 0;
 	return (task_nice(task) + 20) / 5;
 }
 

include/linux/jiffies.h

@@ -164,7 +164,7 @@ static inline u64 get_jiffies_64(void)
  * Have the 32 bit jiffies value wrap 5 minutes after boot
  * so jiffies wrap bugs show up earlier.
  */
-#define INITIAL_JIFFIES ((unsigned long)(unsigned int) (-300*HZ))
+#define INITIAL_JIFFIES ((unsigned long)(unsigned int) (-10*HZ))
 
 /*
  * Change timeval to jiffies, trying to avoid the