sched: latencytop support

LatencyTOP kernel infrastructure; it measures latencies in the
scheduler and tracks it system wide and per process.

Signed-off-by: Arjan van de Ven <[e-mail]>
Signed-off-by: Ingo Molnar <[e-mail]>

arch/x86/kernel/stacktrace.c

···
 		trace->entries[trace->nr_entries++] = addr;
 }
 
+static void save_stack_address_nosched(void *data, unsigned long addr)
+{
+	struct stack_trace *trace = (struct stack_trace *)data;
+	if (in_sched_functions(addr))
+		return;
+	if (trace->skip > 0) {
+		trace->skip--;
+		return;
+	}
+	if (trace->nr_entries < trace->max_entries)
+		trace->entries[trace->nr_entries++] = addr;
+}
+
 static const struct stacktrace_ops save_stack_ops = {
 	.warning = save_stack_warning,
 	.warning_symbol = save_stack_warning_symbol,
 	.stack = save_stack_stack,
 	.address = save_stack_address,
+};
+
+static const struct stacktrace_ops save_stack_ops_nosched = {
+	.warning = save_stack_warning,
+	.warning_symbol = save_stack_warning_symbol,
+	.stack = save_stack_stack,
+	.address = save_stack_address_nosched,
 };
 
 /*
···
 		trace->entries[trace->nr_entries++] = ULONG_MAX;
 }
 EXPORT_SYMBOL(save_stack_trace);
+
+void save_stack_trace_tsk(struct task_struct *tsk, struct stack_trace *trace)
+{
+	dump_trace(tsk, NULL, NULL, &save_stack_ops_nosched, trace);
+	if (trace->nr_entries < trace->max_entries)
+		trace->entries[trace->nr_entries++] = ULONG_MAX;
+}

fs/proc/base.c

···
 }
 #endif
 
+#ifdef CONFIG_LATENCYTOP
+static int lstats_show_proc(struct seq_file *m, void *v)
+{
+	int i;
+	struct task_struct *task = m->private;
+	seq_puts(m, "Latency Top version : v0.1\n");
+
+	for (i = 0; i < 32; i++) {
+		if (task->latency_record[i].backtrace[0]) {
+			int q;
+			seq_printf(m, "%i %li %li ",
+				task->latency_record[i].count,
+				task->latency_record[i].time,
+				task->latency_record[i].max);
+			for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
+				char sym[KSYM_NAME_LEN];
+				char *c;
+				if (!task->latency_record[i].backtrace[q])
+					break;
+				if (task->latency_record[i].backtrace[q] == ULONG_MAX)
+					break;
+				sprint_symbol(sym, task->latency_record[i].backtrace[q]);
+				c = strchr(sym, '+');
+				if (c)
+					*c = 0;
+				seq_printf(m, "%s ", sym);
+			}
+			seq_printf(m, "\n");
+		}
+
+	}
+	return 0;
+}
+
+static int lstats_open(struct inode *inode, struct file *file)
+{
+	int ret;
+	struct seq_file *m;
+	struct task_struct *task = get_proc_task(inode);
+
+	ret = single_open(file, lstats_show_proc, NULL);
+	if (!ret) {
+		m = file->private_data;
+		m->private = task;
+	}
+	return ret;
+}
+
+static ssize_t lstats_write(struct file *file, const char __user *buf,
+			    size_t count, loff_t *offs)
+{
+	struct seq_file *m;
+	struct task_struct *task;
+
+	m = file->private_data;
+	task = m->private;
+	clear_all_latency_tracing(task);
+
+	return count;
+}
+
+static const struct file_operations proc_lstats_operations = {
+	.open		= lstats_open,
+	.read		= seq_read,
+	.write		= lstats_write,
+	.llseek		= seq_lseek,
+	.release	= single_release,
+};
+
+#endif
+
 /* The badness from the OOM killer */
 unsigned long badness(struct task_struct *p, unsigned long uptime);
 static int proc_oom_score(struct task_struct *task, char *buffer)
···
 	.write		= proc_fault_inject_write,
 };
 #endif
+
 
 #ifdef CONFIG_SCHED_DEBUG
 /*
···
 #ifdef CONFIG_SCHEDSTATS
 	INF("schedstat",  S_IRUGO, pid_schedstat),
 #endif
+#ifdef CONFIG_LATENCYTOP
+	REG("latency",  S_IRUGO, lstats),
+#endif
 #ifdef CONFIG_PROC_PID_CPUSET
 	REG("cpuset",     S_IRUGO, cpuset),
 #endif
···
 #endif
 #ifdef CONFIG_SCHEDSTATS
 	INF("schedstat", S_IRUGO, pid_schedstat),
+#endif
+#ifdef CONFIG_LATENCYTOP
+	REG("latency",  S_IRUGO, lstats),
 #endif
 #ifdef CONFIG_PROC_PID_CPUSET
 	REG("cpuset",    S_IRUGO, cpuset),

include/linux/latencytop.h

···
+/*
+ * latencytop.h: Infrastructure for displaying latency
+ *
+ * (C) Copyright 2008 Intel Corporation
+ * Author: Arjan van de Ven <[e-mail]>
+ *
+ */
+
+#ifndef _INCLUDE_GUARD_LATENCYTOP_H_
+#define _INCLUDE_GUARD_LATENCYTOP_H_
+
+#ifdef CONFIG_LATENCYTOP
+
+#define LT_SAVECOUNT		32
+#define LT_BACKTRACEDEPTH	12
+
+struct latency_record {
+	unsigned long	backtrace[LT_BACKTRACEDEPTH];
+	unsigned int	count;
+	unsigned long	time;
+	unsigned long	max;
+};
+
+
+struct task_struct;
+
+void account_scheduler_latency(struct task_struct *task, int usecs, int inter);
+
+void clear_all_latency_tracing(struct task_struct *p);
+
+#else
+
+static inline void
+account_scheduler_latency(struct task_struct *task, int usecs, int inter)
+{
+}
+
+static inline void clear_all_latency_tracing(struct task_struct *p)
+{
+}
+
+#endif
+
+#endif

include/linux/sched.h

···
 #include <linux/hrtimer.h>
 #include <linux/task_io_accounting.h>
 #include <linux/kobject.h>
+#include <linux/latencytop.h>
 
 #include <asm/processor.h>
 
···
 	int make_it_fail;
 #endif
 	struct prop_local_single dirties;
+#ifdef CONFIG_LATENCYTOP
+	int latency_record_count;
+	struct latency_record latency_record[LT_SAVECOUNT];
+#endif
 };
 
 /*

include/linux/stacktrace.h

···
 };
 
 extern void save_stack_trace(struct stack_trace *trace);
+extern void save_stack_trace_tsk(struct task_struct *tsk,
+				struct stack_trace *trace);
 
 extern void print_stack_trace(struct stack_trace *trace, int spaces);
 #else
 # define save_stack_trace(trace)			do { } while (0)
+# define save_stack_trace_tsk(tsk, trace)		do { } while (0)
 # define print_stack_trace(trace, spaces)		do { } while (0)
 #endif
 

kernel/Makefile

···
 obj-$(CONFIG_TASK_DELAY_ACCT) += delayacct.o
 obj-$(CONFIG_TASKSTATS) += taskstats.o tsacct.o
 obj-$(CONFIG_MARKERS) += marker.o
+obj-$(CONFIG_LATENCYTOP) += latencytop.o
 
 ifneq ($(CONFIG_SCHED_NO_NO_OMIT_FRAME_POINTER),y)
 # According to Alan Modra <[e-mail]>, the -fno-omit-frame-pointer is

kernel/fork.c

···
 #ifdef TIF_SYSCALL_EMU
 	clear_tsk_thread_flag(p, TIF_SYSCALL_EMU);
 #endif
+	clear_all_latency_tracing(p);
 
 	/* Our parent execution domain becomes current domain
 	   These must match for thread signalling to apply */

kernel/latencytop.c

···
+/*
+ * latencytop.c: Latency display infrastructure
+ *
+ * (C) Copyright 2008 Intel Corporation
+ * Author: Arjan van de Ven <[e-mail]>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; version 2
+ * of the License.
+ */
+#include <linux/latencytop.h>
+#include <linux/kallsyms.h>
+#include <linux/seq_file.h>
+#include <linux/notifier.h>
+#include <linux/spinlock.h>
+#include <linux/proc_fs.h>
+#include <linux/module.h>
+#include <linux/sched.h>
+#include <linux/list.h>
+#include <linux/slab.h>
+#include <linux/stacktrace.h>
+
+static DEFINE_SPINLOCK(latency_lock);
+
+#define MAXLR 128
+static struct latency_record latency_record[MAXLR];
+
+int latencytop_enabled;
+
+void clear_all_latency_tracing(struct task_struct *p)
+{
+	unsigned long flags;
+
+	if (!latencytop_enabled)
+		return;
+
+	spin_lock_irqsave(&latency_lock, flags);
+	memset(&p->latency_record, 0, sizeof(p->latency_record));
+	p->latency_record_count = 0;
+	spin_unlock_irqrestore(&latency_lock, flags);
+}
+
+static void clear_global_latency_tracing(void)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&latency_lock, flags);
+	memset(&latency_record, 0, sizeof(latency_record));
+	spin_unlock_irqrestore(&latency_lock, flags);
+}
+
+static void __sched
+account_global_scheduler_latency(struct task_struct *tsk, struct latency_record *lat)
+{
+	int firstnonnull = MAXLR + 1;
+	int i;
+
+	if (!latencytop_enabled)
+		return;
+
+	/* skip kernel threads for now */
+	if (!tsk->mm)
+		return;
+
+	for (i = 0; i < MAXLR; i++) {
+		int q;
+		int same = 1;
+		/* Nothing stored: */
+		if (!latency_record[i].backtrace[0]) {
+			if (firstnonnull > i)
+				firstnonnull = i;
+			continue;
+		}
+		for (q = 0 ; q < LT_BACKTRACEDEPTH ; q++) {
+			if (latency_record[i].backtrace[q] !=
+				lat->backtrace[q])
+				same = 0;
+			if (same && lat->backtrace[q] == 0)
+				break;
+			if (same && lat->backtrace[q] == ULONG_MAX)
+				break;
+		}
+		if (same) {
+			latency_record[i].count++;
+			latency_record[i].time += lat->time;
+			if (lat->time > latency_record[i].max)
+				latency_record[i].max = lat->time;
+			return;
+		}
+	}
+
+	i = firstnonnull;
+	if (i >= MAXLR - 1)
+		return;
+
+	/* Allocted a new one: */
+	memcpy(&latency_record[i], lat, sizeof(struct latency_record));
+}
+
+static inline void store_stacktrace(struct task_struct *tsk, struct latency_record *lat)
+{
+	struct stack_trace trace;
+
+	memset(&trace, 0, sizeof(trace));
+	trace.max_entries = LT_BACKTRACEDEPTH;
+	trace.entries = &lat->backtrace[0];
+	trace.skip = 0;
+	save_stack_trace_tsk(tsk, &trace);
+}
+
+void __sched
+account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
+{
+	unsigned long flags;
+	int i, q;
+	struct latency_record lat;
+
+	if (!latencytop_enabled)
+		return;
+
+	/* Long interruptible waits are generally user requested... */
+	if (inter && usecs > 5000)
+		return;
+
+	memset(&lat, 0, sizeof(lat));
+	lat.count = 1;
+	lat.time = usecs;
+	lat.max = usecs;
+	store_stacktrace(tsk, &lat);
+
+	spin_lock_irqsave(&latency_lock, flags);
+
+	account_global_scheduler_latency(tsk, &lat);
+
+	/*
+	 * short term hack; if we're > 32 we stop; future we recycle:
+	 */
+	tsk->latency_record_count++;
+	if (tsk->latency_record_count >= LT_SAVECOUNT)
+		goto out_unlock;
+
+	for (i = 0; i < LT_SAVECOUNT ; i++) {
+		struct latency_record *mylat;
+		int same = 1;
+		mylat = &tsk->latency_record[i];
+		for (q = 0 ; q < LT_BACKTRACEDEPTH ; q++) {
+			if (mylat->backtrace[q] !=
+				lat.backtrace[q])
+				same = 0;
+			if (same && lat.backtrace[q] == 0)
+				break;
+			if (same && lat.backtrace[q] == ULONG_MAX)
+				break;
+		}
+		if (same) {
+			mylat->count++;
+			mylat->time += lat.time;
+			if (lat.time > mylat->max)
+				mylat->max = lat.time;
+			goto out_unlock;
+		}
+	}
+
+	/* Allocated a new one: */
+	i = tsk->latency_record_count;
+	memcpy(&tsk->latency_record[i], &lat, sizeof(struct latency_record));
+
+out_unlock:
+	spin_unlock_irqrestore(&latency_lock, flags);
+}
+
+static int lstats_show(struct seq_file *m, void *v)
+{
+	int i;
+
+	seq_puts(m, "Latency Top version : v0.1\n");
+
+	for (i = 0; i < MAXLR; i++) {
+		if (latency_record[i].backtrace[0]) {
+			int q;
+			seq_printf(m, "%i %li %li ",
+				latency_record[i].count,
+				latency_record[i].time,
+				latency_record[i].max);
+			for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
+				char sym[KSYM_NAME_LEN];
+				char *c;
+				if (!latency_record[i].backtrace[q])
+					break;
+				if (latency_record[i].backtrace[q] == ULONG_MAX)
+					break;
+				sprint_symbol(sym, latency_record[i].backtrace[q]);
+				c = strchr(sym, '+');
+				if (c)
+					*c = 0;
+				seq_printf(m, "%s ", sym);
+			}
+			seq_printf(m, "\n");
+		}
+	}
+	return 0;
+}
+
+static ssize_t
+lstats_write(struct file *file, const char __user *buf, size_t count,
+	     loff_t *offs)
+{
+	clear_global_latency_tracing();
+
+	return count;
+}
+
+static int lstats_open(struct inode *inode, struct file *filp)
+{
+	return single_open(filp, lstats_show, NULL);
+}
+
+static struct file_operations lstats_fops = {
+	.open		= lstats_open,
+	.read		= seq_read,
+	.write		= lstats_write,
+	.llseek		= seq_lseek,
+	.release	= single_release,
+};
+
+static int __init init_lstats_procfs(void)
+{
+	struct proc_dir_entry *pe;
+
+	pe = create_proc_entry("latency_stats", 0644, NULL);
+	if (!pe)
+		return -ENOMEM;
+
+	pe->proc_fops = &lstats_fops;
+
+	return 0;
+}
+__initcall(init_lstats_procfs);

kernel/sched_fair.c

···
  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <[e-mail]>
  */
 
+#include <linux/latencytop.h>
+
 /*
  * Targeted preemption latency for CPU-bound tasks:
  * (default: 20ms * (1 + ilog(ncpus)), units: nanoseconds)
···
 #ifdef CONFIG_SCHEDSTATS
 	if (se->sleep_start) {
 		u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
+		struct task_struct *tsk = task_of(se);
 
 		if ((s64)delta < 0)
 			delta = 0;
···
 
 		se->sleep_start = 0;
 		se->sum_sleep_runtime += delta;
+
+		account_scheduler_latency(tsk, delta >> 10, 1);
 	}
 	if (se->block_start) {
 		u64 delta = rq_of(cfs_rq)->clock - se->block_start;
+		struct task_struct *tsk = task_of(se);
 
 		if ((s64)delta < 0)
 			delta = 0;
···
 		 * time that the task spent sleeping:
 		 */
 		if (unlikely(prof_on == SLEEP_PROFILING)) {
-			struct task_struct *tsk = task_of(se);
 
 			profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
 				     delta >> 20);
 		}
+		account_scheduler_latency(tsk, delta >> 10, 0);
 	}
 #endif
 }

kernel/sysctl.c

···
 extern int maps_protect;
 extern int sysctl_stat_interval;
 extern int audit_argv_kb;
+extern int latencytop_enabled;
 
 /* Constants used for minimum and  maximum */
 #ifdef CONFIG_DETECT_SOFTLOCKUP
···
 		.maxlen		= sizeof(int),
 		.mode		= 0644,
 		.proc_handler	= &proc_dointvec_taint,
+	},
+#endif
+#ifdef CONFIG_LATENCYTOP
+	{
+		.procname	= "latencytop",
+		.data		= &latencytop_enabled,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
 	},
 #endif
 #ifdef CONFIG_SECURITY_CAPABILITIES

lib/Kconfig.debug

···
 	help
 	  Provide stacktrace filter for fault-injection capabilities
 
+config LATENCYTOP
+	bool "Latency measuring infrastructure"
+	select FRAME_POINTER if !MIPS
+	select KALLSYMS
+	select KALLSYMS_ALL
+	select STACKTRACE
+	select SCHEDSTATS
+	select SCHED_DEBUG
+	depends on X86 || X86_64
+	help
+	  Enable this option if you want to use the LatencyTOP tool
+	  to find out which userspace is blocking on what kernel operations.
+
+
 source "samples/Kconfig"