[linux.git] / kernel / profile.c

/*
 *  linux/kernel/profile.c
 *  Simple profiling. Manages a direct-mapped profile hit count buffer,
 *  with configurable resolution, support for restricting the cpus on
 *  which profiling is done, and switching between cpu time and
 *  schedule() calls via kernel command line parameters passed at boot.
 *
 *  Scheduler profiling support, Arjan van de Ven and Ingo Molnar,
 *	Red Hat, July 2004
 *  Consolidation of architecture support code for profiling,
 *	William Irwin, Oracle, July 2004
 *  Amortized hit count accounting via per-cpu open-addressed hashtables
 *	to resolve timer interrupt livelocks, William Irwin, Oracle, 2004
 */

#include <linux/module.h>
#include <linux/profile.h>
#include <linux/bootmem.h>
#include <linux/notifier.h>
#include <linux/mm.h>
#include <linux/cpumask.h>
#include <linux/cpu.h>
#include <linux/highmem.h>
#include <linux/mutex.h>
#include <linux/slab.h>
#include <linux/vmalloc.h>
#include <asm/sections.h>
#include <asm/irq_regs.h>
#include <asm/ptrace.h>

struct profile_hit {
	u32 pc, hits;
};
#define PROFILE_GRPSHIFT	3
#define PROFILE_GRPSZ		(1 << PROFILE_GRPSHIFT)
#define NR_PROFILE_HIT		(PAGE_SIZE/sizeof(struct profile_hit))
#define NR_PROFILE_GRP		(NR_PROFILE_HIT/PROFILE_GRPSZ)

/* Oprofile timer tick hook */
static int (*timer_hook)(struct pt_regs *) __read_mostly;

static atomic_t *prof_buffer;
static unsigned long prof_len, prof_shift;

int prof_on __read_mostly;
EXPORT_SYMBOL_GPL(prof_on);

static cpumask_t prof_cpu_mask = CPU_MASK_ALL;
#ifdef CONFIG_SMP
static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);
static DEFINE_PER_CPU(int, cpu_profile_flip);
static DEFINE_MUTEX(profile_flip_mutex);
#endif /* CONFIG_SMP */

int profile_setup(char *str)
{
	static char schedstr[] = "schedule";
	static char sleepstr[] = "sleep";
	static char kvmstr[] = "kvm";
	int par;

	if (!strncmp(str, sleepstr, strlen(sleepstr))) {
#ifdef CONFIG_SCHEDSTATS
		prof_on = SLEEP_PROFILING;
		if (str[strlen(sleepstr)] == ',')
			str += strlen(sleepstr) + 1;
		if (get_option(&str, &par))
			prof_shift = par;
		printk(KERN_INFO
			"kernel sleep profiling enabled (shift: %ld)\n",
			prof_shift);
#else
		printk(KERN_WARNING
			"kernel sleep profiling requires CONFIG_SCHEDSTATS\n");
#endif /* CONFIG_SCHEDSTATS */
	} else if (!strncmp(str, schedstr, strlen(schedstr))) {
		prof_on = SCHED_PROFILING;
		if (str[strlen(schedstr)] == ',')
			str += strlen(schedstr) + 1;
		if (get_option(&str, &par))
			prof_shift = par;
		printk(KERN_INFO
			"kernel schedule profiling enabled (shift: %ld)\n",
			prof_shift);
	} else if (!strncmp(str, kvmstr, strlen(kvmstr))) {
		prof_on = KVM_PROFILING;
		if (str[strlen(kvmstr)] == ',')
			str += strlen(kvmstr) + 1;
		if (get_option(&str, &par))
			prof_shift = par;
		printk(KERN_INFO
			"kernel KVM profiling enabled (shift: %ld)\n",
			prof_shift);
	} else if (get_option(&str, &par)) {
		prof_shift = par;
		prof_on = CPU_PROFILING;
		printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n",
			prof_shift);
	}
	return 1;
}
__setup("profile=", profile_setup);


int __ref profile_init(void)
{
	int buffer_bytes;
	if (!prof_on)
		return 0;

	/* only text is profiled */
	prof_len = (_etext - _stext) >> prof_shift;
	buffer_bytes = prof_len*sizeof(atomic_t);
	if (!slab_is_available()) {
		prof_buffer = alloc_bootmem(buffer_bytes);
		return 0;
	}

	prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL);
	if (prof_buffer)
		return 0;

	prof_buffer = alloc_pages_exact(buffer_bytes, GFP_KERNEL|__GFP_ZERO);
	if (prof_buffer)
		return 0;

	prof_buffer = vmalloc(buffer_bytes);
	if (prof_buffer)
		return 0;

	return -ENOMEM;
}

/* Profile event notifications */

static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
static BLOCKING_NOTIFIER_HEAD(munmap_notifier);

void profile_task_exit(struct task_struct *task)
{
	blocking_notifier_call_chain(&task_exit_notifier, 0, task);
}

int profile_handoff_task(struct task_struct *task)
{
	int ret;
	ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
	return (ret == NOTIFY_OK) ? 1 : 0;
}

void profile_munmap(unsigned long addr)
{
	blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
}

int task_handoff_register(struct notifier_block *n)
{
	return atomic_notifier_chain_register(&task_free_notifier, n);
}
EXPORT_SYMBOL_GPL(task_handoff_register);

int task_handoff_unregister(struct notifier_block *n)
{
	return atomic_notifier_chain_unregister(&task_free_notifier, n);
}
EXPORT_SYMBOL_GPL(task_handoff_unregister);

int profile_event_register(enum profile_type type, struct notifier_block *n)
{
	int err = -EINVAL;

	switch (type) {
	case PROFILE_TASK_EXIT:
		err = blocking_notifier_chain_register(
				&task_exit_notifier, n);
		break;
	case PROFILE_MUNMAP:
		err = blocking_notifier_chain_register(
				&munmap_notifier, n);
		break;
	}

	return err;
}
EXPORT_SYMBOL_GPL(profile_event_register);

int profile_event_unregister(enum profile_type type, struct notifier_block *n)
{
	int err = -EINVAL;

	switch (type) {
	case PROFILE_TASK_EXIT:
		err = blocking_notifier_chain_unregister(
				&task_exit_notifier, n);
		break;
	case PROFILE_MUNMAP:
		err = blocking_notifier_chain_unregister(
				&munmap_notifier, n);
		break;
	}

	return err;
}
EXPORT_SYMBOL_GPL(profile_event_unregister);

int register_timer_hook(int (*hook)(struct pt_regs *))
{
	if (timer_hook)
		return -EBUSY;
	timer_hook = hook;
	return 0;
}
EXPORT_SYMBOL_GPL(register_timer_hook);

void unregister_timer_hook(int (*hook)(struct pt_regs *))
{
	WARN_ON(hook != timer_hook);
	timer_hook = NULL;
	/* make sure all CPUs see the NULL hook */
	synchronize_sched();  /* Allow ongoing interrupts to complete. */
}
EXPORT_SYMBOL_GPL(unregister_timer_hook);


#ifdef CONFIG_SMP
/*
 * Each cpu has a pair of open-addressed hashtables for pending
 * profile hits. read_profile() IPI's all cpus to request them
 * to flip buffers and flushes their contents to prof_buffer itself.
 * Flip requests are serialized by the profile_flip_mutex. The sole
 * use of having a second hashtable is for avoiding cacheline
 * contention that would otherwise happen during flushes of pending
 * profile hits required for the accuracy of reported profile hits
 * and so resurrect the interrupt livelock issue.
 *
 * The open-addressed hashtables are indexed by profile buffer slot
 * and hold the number of pending hits to that profile buffer slot on
 * a cpu in an entry. When the hashtable overflows, all pending hits
 * are accounted to their corresponding profile buffer slots with
 * atomic_add() and the hashtable emptied. As numerous pending hits
 * may be accounted to a profile buffer slot in a hashtable entry,
 * this amortizes a number of atomic profile buffer increments likely
 * to be far larger than the number of entries in the hashtable,
 * particularly given that the number of distinct profile buffer
 * positions to which hits are accounted during short intervals (e.g.
 * several seconds) is usually very small. Exclusion from buffer
 * flipping is provided by interrupt disablement (note that for
 * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
 * process context).
 * The hash function is meant to be lightweight as opposed to strong,
 * and was vaguely inspired by ppc64 firmware-supported inverted
 * pagetable hash functions, but uses a full hashtable full of finite
 * collision chains, not just pairs of them.
 *
 * -- wli
 */
static void __profile_flip_buffers(void *unused)
{
	int cpu = smp_processor_id();

	per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
}

static void profile_flip_buffers(void)
{
	int i, j, cpu;

	mutex_lock(&profile_flip_mutex);
	j = per_cpu(cpu_profile_flip, get_cpu());
	put_cpu();
	on_each_cpu(__profile_flip_buffers, NULL, 1);
	for_each_online_cpu(cpu) {
		struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
		for (i = 0; i < NR_PROFILE_HIT; ++i) {
			if (!hits[i].hits) {
				if (hits[i].pc)
					hits[i].pc = 0;
				continue;
			}
			atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
			hits[i].hits = hits[i].pc = 0;
		}
	}
	mutex_unlock(&profile_flip_mutex);
}

static void profile_discard_flip_buffers(void)
{
	int i, cpu;

	mutex_lock(&profile_flip_mutex);
	i = per_cpu(cpu_profile_flip, get_cpu());
	put_cpu();
	on_each_cpu(__profile_flip_buffers, NULL, 1);
	for_each_online_cpu(cpu) {
		struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
		memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
	}
	mutex_unlock(&profile_flip_mutex);
}

void profile_hits(int type, void *__pc, unsigned int nr_hits)
{
	unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
	int i, j, cpu;
	struct profile_hit *hits;

	if (prof_on != type || !prof_buffer)
		return;
	pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
	i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
	secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
	cpu = get_cpu();
	hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
	if (!hits) {
		put_cpu();
		return;
	}
	/*
	 * We buffer the global profiler buffer into a per-CPU
	 * queue and thus reduce the number of global (and possibly
	 * NUMA-alien) accesses. The write-queue is self-coalescing:
	 */
	local_irq_save(flags);
	do {
		for (j = 0; j < PROFILE_GRPSZ; ++j) {
			if (hits[i + j].pc == pc) {
				hits[i + j].hits += nr_hits;
				goto out;
			} else if (!hits[i + j].hits) {
				hits[i + j].pc = pc;
				hits[i + j].hits = nr_hits;
				goto out;
			}
		}
		i = (i + secondary) & (NR_PROFILE_HIT - 1);
	} while (i != primary);

	/*
	 * Add the current hit(s) and flush the write-queue out
	 * to the global buffer:
	 */
	atomic_add(nr_hits, &prof_buffer[pc]);
	for (i = 0; i < NR_PROFILE_HIT; ++i) {
		atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
		hits[i].pc = hits[i].hits = 0;
	}
out:
	local_irq_restore(flags);
	put_cpu();
}

static int __cpuinit profile_cpu_callback(struct notifier_block *info,
					unsigned long action, void *__cpu)
{
	int node, cpu = (unsigned long)__cpu;
	struct page *page;

	switch (action) {
	case CPU_UP_PREPARE:
	case CPU_UP_PREPARE_FROZEN:
		node = cpu_to_node(cpu);
		per_cpu(cpu_profile_flip, cpu) = 0;
		if (!per_cpu(cpu_profile_hits, cpu)[1]) {
			page = alloc_pages_node(node,
					GFP_KERNEL | __GFP_ZERO,
					0);
			if (!page)
				return NOTIFY_BAD;
			per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
		}
		if (!per_cpu(cpu_profile_hits, cpu)[0]) {
			page = alloc_pages_node(node,
					GFP_KERNEL | __GFP_ZERO,
					0);
			if (!page)
				goto out_free;
			per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
		}
		break;
out_free:
		page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
		per_cpu(cpu_profile_hits, cpu)[1] = NULL;
		__free_page(page);
		return NOTIFY_BAD;
	case CPU_ONLINE:
	case CPU_ONLINE_FROZEN:
		cpu_set(cpu, prof_cpu_mask);
		break;
	case CPU_UP_CANCELED:
	case CPU_UP_CANCELED_FROZEN:
	case CPU_DEAD:
	case CPU_DEAD_FROZEN:
		cpu_clear(cpu, prof_cpu_mask);
		if (per_cpu(cpu_profile_hits, cpu)[0]) {
			page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
			per_cpu(cpu_profile_hits, cpu)[0] = NULL;
			__free_page(page);
		}
		if (per_cpu(cpu_profile_hits, cpu)[1]) {
			page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
			per_cpu(cpu_profile_hits, cpu)[1] = NULL;
			__free_page(page);
		}
		break;
	}
	return NOTIFY_OK;
}
#else /* !CONFIG_SMP */
#define profile_flip_buffers()		do { } while (0)
#define profile_discard_flip_buffers()	do { } while (0)
#define profile_cpu_callback		NULL

void profile_hits(int type, void *__pc, unsigned int nr_hits)
{
	unsigned long pc;

	if (prof_on != type || !prof_buffer)
		return;
	pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
	atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
}
#endif /* !CONFIG_SMP */
EXPORT_SYMBOL_GPL(profile_hits);

void profile_tick(int type)
{
	struct pt_regs *regs = get_irq_regs();

	if (type == CPU_PROFILING && timer_hook)
		timer_hook(regs);
	if (!user_mode(regs) && cpu_isset(smp_processor_id(), prof_cpu_mask))
		profile_hit(type, (void *)profile_pc(regs));
}

#ifdef CONFIG_PROC_FS
#include <linux/proc_fs.h>
#include <asm/uaccess.h>
#include <asm/ptrace.h>

static int prof_cpu_mask_read_proc(char *page, char **start, off_t off,
			int count, int *eof, void *data)
{
	int len = cpumask_scnprintf(page, count, *(cpumask_t *)data);
	if (count - len < 2)
		return -EINVAL;
	len += sprintf(page + len, "\n");
	return len;
}

static int prof_cpu_mask_write_proc(struct file *file,
	const char __user *buffer,  unsigned long count, void *data)
{
	cpumask_t *mask = (cpumask_t *)data;
	unsigned long full_count = count, err;
	cpumask_t new_value;

	err = cpumask_parse_user(buffer, count, new_value);
	if (err)
		return err;

	*mask = new_value;
	return full_count;
}

void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir)
{
	struct proc_dir_entry *entry;

	/* create /proc/irq/prof_cpu_mask */
	entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir);
	if (!entry)
		return;
	entry->data = (void *)&prof_cpu_mask;
	entry->read_proc = prof_cpu_mask_read_proc;
	entry->write_proc = prof_cpu_mask_write_proc;
}

/*
 * This function accesses profiling information. The returned data is
 * binary: the sampling step and the actual contents of the profile
 * buffer. Use of the program readprofile is recommended in order to
 * get meaningful info out of these data.
 */
static ssize_t
read_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
{
	unsigned long p = *ppos;
	ssize_t read;
	char *pnt;
	unsigned int sample_step = 1 << prof_shift;

	profile_flip_buffers();
	if (p >= (prof_len+1)*sizeof(unsigned int))
		return 0;
	if (count > (prof_len+1)*sizeof(unsigned int) - p)
		count = (prof_len+1)*sizeof(unsigned int) - p;
	read = 0;

	while (p < sizeof(unsigned int) && count > 0) {
		if (put_user(*((char *)(&sample_step)+p), buf))
			return -EFAULT;
		buf++; p++; count--; read++;
	}
	pnt = (char *)prof_buffer + p - sizeof(atomic_t);
	if (copy_to_user(buf, (void *)pnt, count))
		return -EFAULT;
	read += count;
	*ppos += read;
	return read;
}

/*
 * Writing to /proc/profile resets the counters
 *
 * Writing a 'profiling multiplier' value into it also re-sets the profiling
 * interrupt frequency, on architectures that support this.
 */
static ssize_t write_profile(struct file *file, const char __user *buf,
			     size_t count, loff_t *ppos)
{
#ifdef CONFIG_SMP
	extern int setup_profiling_timer(unsigned int multiplier);

	if (count == sizeof(int)) {
		unsigned int multiplier;

		if (copy_from_user(&multiplier, buf, sizeof(int)))
			return -EFAULT;

		if (setup_profiling_timer(multiplier))
			return -EINVAL;
	}
#endif
	profile_discard_flip_buffers();
	memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
	return count;
}

static const struct file_operations proc_profile_operations = {
	.read		= read_profile,
	.write		= write_profile,
};

#ifdef CONFIG_SMP
static inline void profile_nop(void *unused)
{
}

static int create_hash_tables(void)
{
	int cpu;

	for_each_online_cpu(cpu) {
		int node = cpu_to_node(cpu);
		struct page *page;

		page = alloc_pages_node(node,
				GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
				0);
		if (!page)
			goto out_cleanup;
		per_cpu(cpu_profile_hits, cpu)[1]
				= (struct profile_hit *)page_address(page);
		page = alloc_pages_node(node,
				GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
				0);
		if (!page)
			goto out_cleanup;
		per_cpu(cpu_profile_hits, cpu)[0]
				= (struct profile_hit *)page_address(page);
	}
	return 0;
out_cleanup:
	prof_on = 0;
	smp_mb();
	on_each_cpu(profile_nop, NULL, 1);
	for_each_online_cpu(cpu) {
		struct page *page;

		if (per_cpu(cpu_profile_hits, cpu)[0]) {
			page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
			per_cpu(cpu_profile_hits, cpu)[0] = NULL;
			__free_page(page);
		}
		if (per_cpu(cpu_profile_hits, cpu)[1]) {
			page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
			per_cpu(cpu_profile_hits, cpu)[1] = NULL;
			__free_page(page);
		}
	}
	return -1;
}
#else
#define create_hash_tables()			({ 0; })
#endif

int __ref create_proc_profile(void) /* false positive from hotcpu_notifier */
{
	struct proc_dir_entry *entry;

	if (!prof_on)
		return 0;
	if (create_hash_tables())
		return -ENOMEM;
	entry = proc_create("profile", S_IWUSR | S_IRUGO,
			    NULL, &proc_profile_operations);
	if (!entry)
		return 0;
	entry->size = (1+prof_len) * sizeof(atomic_t);
	hotcpu_notifier(profile_cpu_callback, 0);
	return 0;
}
module_init(create_proc_profile);
#endif /* CONFIG_PROC_FS */
Commit	Line	Data
1da177e4 LT	1	/*
	2	* linux/kernel/profile.c
	3	* Simple profiling. Manages a direct-mapped profile hit count buffer,
	4	* with configurable resolution, support for restricting the cpus on
	5	* which profiling is done, and switching between cpu time and
	6	* schedule() calls via kernel command line parameters passed at boot.
	7	*
	8	* Scheduler profiling support, Arjan van de Ven and Ingo Molnar,
	9	* Red Hat, July 2004
	10	* Consolidation of architecture support code for profiling,
	11	* William Irwin, Oracle, July 2004
	12	* Amortized hit count accounting via per-cpu open-addressed hashtables
	13	* to resolve timer interrupt livelocks, William Irwin, Oracle, 2004
	14	*/
	15
1da177e4 LT	16	#include <linux/module.h>
	17	#include <linux/profile.h>
	18	#include <linux/bootmem.h>
	19	#include <linux/notifier.h>
	20	#include <linux/mm.h>
	21	#include <linux/cpumask.h>
	22	#include <linux/cpu.h>
1da177e4	23	#include <linux/highmem.h>
97d1f15b	24	#include <linux/mutex.h>
22b8ce94 DH	25	#include <linux/slab.h>
22b8ce94 DH	26	#include <linux/vmalloc.h>
1da177e4	27	#include <asm/sections.h>
7d12e780	28	#include <asm/irq_regs.h>
e8edc6e0	29	#include <asm/ptrace.h>
1da177e4 LT	30
	31	struct profile_hit {
	32	u32 pc, hits;
	33	};
	34	#define PROFILE_GRPSHIFT 3
	35	#define PROFILE_GRPSZ (1 << PROFILE_GRPSHIFT)
	36	#define NR_PROFILE_HIT (PAGE_SIZE/sizeof(struct profile_hit))
	37	#define NR_PROFILE_GRP (NR_PROFILE_HIT/PROFILE_GRPSZ)
	38
	39	/* Oprofile timer tick hook */
b012d346	40	static int (timer_hook)(struct pt_regs ) __read_mostly;
1da177e4 LT	41
	42	static atomic_t *prof_buffer;
	43	static unsigned long prof_len, prof_shift;
07031e14	44
ece8a684	45	int prof_on __read_mostly;
07031e14 IM	46	EXPORT_SYMBOL_GPL(prof_on);
07031e14 IM	47
1da177e4 LT	48	static cpumask_t prof_cpu_mask = CPU_MASK_ALL;
	49	#ifdef CONFIG_SMP
	50	static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);
	51	static DEFINE_PER_CPU(int, cpu_profile_flip);
97d1f15b	52	static DEFINE_MUTEX(profile_flip_mutex);
1da177e4 LT	53	#endif /* CONFIG_SMP */
1da177e4 LT	54
22b8ce94	55	int profile_setup(char *str)
1da177e4	56	{
22b8ce94 DH	57	static char schedstr[] = "schedule";
	58	static char sleepstr[] = "sleep";
	59	static char kvmstr[] = "kvm";
1da177e4 LT	60	int par;
1da177e4 LT	61
ece8a684	62	if (!strncmp(str, sleepstr, strlen(sleepstr))) {
b3da2a73	63	#ifdef CONFIG_SCHEDSTATS
ece8a684 IM	64	prof_on = SLEEP_PROFILING;
	65	if (str[strlen(sleepstr)] == ',')
	66	str += strlen(sleepstr) + 1;
	67	if (get_option(&str, &par))
	68	prof_shift = par;
	69	printk(KERN_INFO
	70	"kernel sleep profiling enabled (shift: %ld)\n",
	71	prof_shift);
b3da2a73 MG	72	#else
	73	printk(KERN_WARNING
	74	"kernel sleep profiling requires CONFIG_SCHEDSTATS\n");
	75	#endif /* CONFIG_SCHEDSTATS */
a75acf85	76	} else if (!strncmp(str, schedstr, strlen(schedstr))) {
1da177e4	77	prof_on = SCHED_PROFILING;
dfaa9c94 NYC	78	if (str[strlen(schedstr)] == ',')
	79	str += strlen(schedstr) + 1;
	80	if (get_option(&str, &par))
	81	prof_shift = par;
	82	printk(KERN_INFO
	83	"kernel schedule profiling enabled (shift: %ld)\n",
	84	prof_shift);
07031e14 IM	85	} else if (!strncmp(str, kvmstr, strlen(kvmstr))) {
	86	prof_on = KVM_PROFILING;
	87	if (str[strlen(kvmstr)] == ',')
	88	str += strlen(kvmstr) + 1;
	89	if (get_option(&str, &par))
	90	prof_shift = par;
	91	printk(KERN_INFO
	92	"kernel KVM profiling enabled (shift: %ld)\n",
	93	prof_shift);
dfaa9c94	94	} else if (get_option(&str, &par)) {
1da177e4 LT	95	prof_shift = par;
	96	prof_on = CPU_PROFILING;
	97	printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n",
	98	prof_shift);
	99	}
	100	return 1;
	101	}
	102	__setup("profile=", profile_setup);
	103
	104
ce05fcc3	105	int __ref profile_init(void)
1da177e4	106	{
22b8ce94	107	int buffer_bytes;
1ad82fd5	108	if (!prof_on)
22b8ce94	109	return 0;
1ad82fd5	110
1da177e4 LT	111	/* only text is profiled */
1da177e4 LT	112	prof_len = (_etext - _stext) >> prof_shift;
22b8ce94 DH	113	buffer_bytes = prof_len*sizeof(atomic_t);
	114	if (!slab_is_available()) {
	115	prof_buffer = alloc_bootmem(buffer_bytes);
	116	return 0;
	117	}
	118
	119	prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL);
	120	if (prof_buffer)
	121	return 0;
	122
	123	prof_buffer = alloc_pages_exact(buffer_bytes, GFP_KERNEL\|__GFP_ZERO);
	124	if (prof_buffer)
	125	return 0;
	126
	127	prof_buffer = vmalloc(buffer_bytes);
	128	if (prof_buffer)
	129	return 0;
	130
	131	return -ENOMEM;
1da177e4 LT	132	}
	133
	134	/* Profile event notifications */
1ad82fd5	135
e041c683 AS	136	static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
	137	static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
	138	static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
1ad82fd5 PC	139
1ad82fd5 PC	140	void profile_task_exit(struct task_struct *task)
1da177e4	141	{
e041c683	142	blocking_notifier_call_chain(&task_exit_notifier, 0, task);
1da177e4	143	}
1ad82fd5 PC	144
1ad82fd5 PC	145	int profile_handoff_task(struct task_struct *task)
1da177e4 LT	146	{
1da177e4 LT	147	int ret;
e041c683	148	ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
1da177e4 LT	149	return (ret == NOTIFY_OK) ? 1 : 0;
	150	}
	151
	152	void profile_munmap(unsigned long addr)
	153	{
e041c683	154	blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
1da177e4 LT	155	}
1da177e4 LT	156
1ad82fd5	157	int task_handoff_register(struct notifier_block *n)
1da177e4	158	{
e041c683	159	return atomic_notifier_chain_register(&task_free_notifier, n);
1da177e4	160	}
1ad82fd5	161	EXPORT_SYMBOL_GPL(task_handoff_register);
1da177e4	162
1ad82fd5	163	int task_handoff_unregister(struct notifier_block *n)
1da177e4	164	{
e041c683	165	return atomic_notifier_chain_unregister(&task_free_notifier, n);
1da177e4	166	}
1ad82fd5	167	EXPORT_SYMBOL_GPL(task_handoff_unregister);
1da177e4	168
1ad82fd5	169	int profile_event_register(enum profile_type type, struct notifier_block *n)
1da177e4 LT	170	{
1da177e4 LT	171	int err = -EINVAL;
1ad82fd5	172
1da177e4	173	switch (type) {
1ad82fd5 PC	174	case PROFILE_TASK_EXIT:
	175	err = blocking_notifier_chain_register(
	176	&task_exit_notifier, n);
	177	break;
	178	case PROFILE_MUNMAP:
	179	err = blocking_notifier_chain_register(
	180	&munmap_notifier, n);
	181	break;
1da177e4	182	}
1ad82fd5	183
1da177e4 LT	184	return err;
1da177e4 LT	185	}
1ad82fd5	186	EXPORT_SYMBOL_GPL(profile_event_register);
1da177e4	187
1ad82fd5	188	int profile_event_unregister(enum profile_type type, struct notifier_block *n)
1da177e4 LT	189	{
1da177e4 LT	190	int err = -EINVAL;
1ad82fd5	191
1da177e4	192	switch (type) {
1ad82fd5 PC	193	case PROFILE_TASK_EXIT:
	194	err = blocking_notifier_chain_unregister(
	195	&task_exit_notifier, n);
	196	break;
	197	case PROFILE_MUNMAP:
	198	err = blocking_notifier_chain_unregister(
	199	&munmap_notifier, n);
	200	break;
1da177e4 LT	201	}
1da177e4 LT	202
1da177e4 LT	203	return err;
1da177e4 LT	204	}
1ad82fd5	205	EXPORT_SYMBOL_GPL(profile_event_unregister);
1da177e4 LT	206
	207	int register_timer_hook(int (hook)(struct pt_regs ))
	208	{
	209	if (timer_hook)
	210	return -EBUSY;
	211	timer_hook = hook;
	212	return 0;
	213	}
1ad82fd5	214	EXPORT_SYMBOL_GPL(register_timer_hook);
1da177e4 LT	215
	216	void unregister_timer_hook(int (hook)(struct pt_regs ))
	217	{
	218	WARN_ON(hook != timer_hook);
	219	timer_hook = NULL;
	220	/* make sure all CPUs see the NULL hook */
fbd568a3	221	synchronize_sched(); /* Allow ongoing interrupts to complete. */
1da177e4	222	}
1da177e4	223	EXPORT_SYMBOL_GPL(unregister_timer_hook);
1da177e4	224
1da177e4 LT	225
	226	#ifdef CONFIG_SMP
	227	/*
	228	* Each cpu has a pair of open-addressed hashtables for pending
	229	* profile hits. read_profile() IPI's all cpus to request them
	230	* to flip buffers and flushes their contents to prof_buffer itself.
	231	* Flip requests are serialized by the profile_flip_mutex. The sole
	232	* use of having a second hashtable is for avoiding cacheline
	233	* contention that would otherwise happen during flushes of pending
	234	* profile hits required for the accuracy of reported profile hits
	235	* and so resurrect the interrupt livelock issue.
	236	*
	237	* The open-addressed hashtables are indexed by profile buffer slot
	238	* and hold the number of pending hits to that profile buffer slot on
	239	* a cpu in an entry. When the hashtable overflows, all pending hits
	240	* are accounted to their corresponding profile buffer slots with
	241	* atomic_add() and the hashtable emptied. As numerous pending hits
	242	* may be accounted to a profile buffer slot in a hashtable entry,
	243	* this amortizes a number of atomic profile buffer increments likely
	244	* to be far larger than the number of entries in the hashtable,
	245	* particularly given that the number of distinct profile buffer
	246	* positions to which hits are accounted during short intervals (e.g.
	247	* several seconds) is usually very small. Exclusion from buffer
	248	* flipping is provided by interrupt disablement (note that for
ece8a684 IM	249	* SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
ece8a684 IM	250	* process context).
1da177e4 LT	251	* The hash function is meant to be lightweight as opposed to strong,
	252	* and was vaguely inspired by ppc64 firmware-supported inverted
	253	* pagetable hash functions, but uses a full hashtable full of finite
	254	* collision chains, not just pairs of them.
	255	*
	256	* -- wli
	257	*/
	258	static void __profile_flip_buffers(void *unused)
	259	{
	260	int cpu = smp_processor_id();
	261
	262	per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
	263	}
	264
	265	static void profile_flip_buffers(void)
	266	{
	267	int i, j, cpu;
	268
97d1f15b	269	mutex_lock(&profile_flip_mutex);
1da177e4 LT	270	j = per_cpu(cpu_profile_flip, get_cpu());
1da177e4 LT	271	put_cpu();
15c8b6c1	272	on_each_cpu(__profile_flip_buffers, NULL, 1);
1da177e4 LT	273	for_each_online_cpu(cpu) {
	274	struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
	275	for (i = 0; i < NR_PROFILE_HIT; ++i) {
	276	if (!hits[i].hits) {
	277	if (hits[i].pc)
	278	hits[i].pc = 0;
	279	continue;
	280	}
	281	atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
	282	hits[i].hits = hits[i].pc = 0;
	283	}
	284	}
97d1f15b	285	mutex_unlock(&profile_flip_mutex);
1da177e4 LT	286	}
	287
	288	static void profile_discard_flip_buffers(void)
	289	{
	290	int i, cpu;
	291
97d1f15b	292	mutex_lock(&profile_flip_mutex);
1da177e4 LT	293	i = per_cpu(cpu_profile_flip, get_cpu());
1da177e4 LT	294	put_cpu();
15c8b6c1	295	on_each_cpu(__profile_flip_buffers, NULL, 1);
1da177e4 LT	296	for_each_online_cpu(cpu) {
	297	struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
	298	memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
	299	}
97d1f15b	300	mutex_unlock(&profile_flip_mutex);
1da177e4 LT	301	}
1da177e4 LT	302
ece8a684	303	void profile_hits(int type, void *__pc, unsigned int nr_hits)
1da177e4 LT	304	{
	305	unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
	306	int i, j, cpu;
	307	struct profile_hit *hits;
	308
	309	if (prof_on != type \|\| !prof_buffer)
	310	return;
	311	pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
	312	i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
	313	secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
	314	cpu = get_cpu();
	315	hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
	316	if (!hits) {
	317	put_cpu();
	318	return;
	319	}
ece8a684 IM	320	/*
	321	* We buffer the global profiler buffer into a per-CPU
	322	* queue and thus reduce the number of global (and possibly
	323	* NUMA-alien) accesses. The write-queue is self-coalescing:
	324	*/
1da177e4 LT	325	local_irq_save(flags);
	326	do {
	327	for (j = 0; j < PROFILE_GRPSZ; ++j) {
	328	if (hits[i + j].pc == pc) {
ece8a684	329	hits[i + j].hits += nr_hits;
1da177e4 LT	330	goto out;
	331	} else if (!hits[i + j].hits) {
	332	hits[i + j].pc = pc;
ece8a684	333	hits[i + j].hits = nr_hits;
1da177e4 LT	334	goto out;
	335	}
	336	}
	337	i = (i + secondary) & (NR_PROFILE_HIT - 1);
	338	} while (i != primary);
ece8a684 IM	339
	340	/*
	341	* Add the current hit(s) and flush the write-queue out
	342	* to the global buffer:
	343	*/
	344	atomic_add(nr_hits, &prof_buffer[pc]);
1da177e4 LT	345	for (i = 0; i < NR_PROFILE_HIT; ++i) {
	346	atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
	347	hits[i].pc = hits[i].hits = 0;
	348	}
	349	out:
	350	local_irq_restore(flags);
	351	put_cpu();
	352	}
	353
84196414	354	static int __cpuinit profile_cpu_callback(struct notifier_block *info,
1da177e4 LT	355	unsigned long action, void *__cpu)
	356	{
	357	int node, cpu = (unsigned long)__cpu;
	358	struct page *page;
	359
	360	switch (action) {
	361	case CPU_UP_PREPARE:
8bb78442	362	case CPU_UP_PREPARE_FROZEN:
1da177e4 LT	363	node = cpu_to_node(cpu);
	364	per_cpu(cpu_profile_flip, cpu) = 0;
	365	if (!per_cpu(cpu_profile_hits, cpu)[1]) {
fbd98167	366	page = alloc_pages_node(node,
4199cfa0	367	GFP_KERNEL \| __GFP_ZERO,
fbd98167	368	0);
1da177e4 LT	369	if (!page)
	370	return NOTIFY_BAD;
	371	per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
	372	}
	373	if (!per_cpu(cpu_profile_hits, cpu)[0]) {
fbd98167	374	page = alloc_pages_node(node,
4199cfa0	375	GFP_KERNEL \| __GFP_ZERO,
fbd98167	376	0);
1da177e4 LT	377	if (!page)
	378	goto out_free;
	379	per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
	380	}
	381	break;
1ad82fd5	382	out_free:
1da177e4 LT	383	page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
	384	per_cpu(cpu_profile_hits, cpu)[1] = NULL;
	385	__free_page(page);
	386	return NOTIFY_BAD;
	387	case CPU_ONLINE:
8bb78442	388	case CPU_ONLINE_FROZEN:
1da177e4 LT	389	cpu_set(cpu, prof_cpu_mask);
	390	break;
	391	case CPU_UP_CANCELED:
8bb78442	392	case CPU_UP_CANCELED_FROZEN:
1da177e4	393	case CPU_DEAD:
8bb78442	394	case CPU_DEAD_FROZEN:
1da177e4 LT	395	cpu_clear(cpu, prof_cpu_mask);
	396	if (per_cpu(cpu_profile_hits, cpu)[0]) {
	397	page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
	398	per_cpu(cpu_profile_hits, cpu)[0] = NULL;
	399	__free_page(page);
	400	}
	401	if (per_cpu(cpu_profile_hits, cpu)[1]) {
	402	page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
	403	per_cpu(cpu_profile_hits, cpu)[1] = NULL;
	404	__free_page(page);
	405	}
	406	break;
	407	}
	408	return NOTIFY_OK;
	409	}
1da177e4 LT	410	#else /* !CONFIG_SMP */
	411	#define profile_flip_buffers() do { } while (0)
	412	#define profile_discard_flip_buffers() do { } while (0)
02316067	413	#define profile_cpu_callback NULL
1da177e4	414
ece8a684	415	void profile_hits(int type, void *__pc, unsigned int nr_hits)
1da177e4 LT	416	{
	417	unsigned long pc;
	418
	419	if (prof_on != type \|\| !prof_buffer)
	420	return;
	421	pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
ece8a684	422	atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
1da177e4 LT	423	}
1da177e4 LT	424	#endif /* !CONFIG_SMP */
bbe1a59b AM	425	EXPORT_SYMBOL_GPL(profile_hits);
bbe1a59b AM	426
7d12e780	427	void profile_tick(int type)
1da177e4	428	{
7d12e780 DH	429	struct pt_regs *regs = get_irq_regs();
7d12e780 DH	430
1da177e4 LT	431	if (type == CPU_PROFILING && timer_hook)
	432	timer_hook(regs);
	433	if (!user_mode(regs) && cpu_isset(smp_processor_id(), prof_cpu_mask))
	434	profile_hit(type, (void *)profile_pc(regs));
	435	}
	436
	437	#ifdef CONFIG_PROC_FS
	438	#include <linux/proc_fs.h>
	439	#include <asm/uaccess.h>
	440	#include <asm/ptrace.h>
	441
1ad82fd5	442	static int prof_cpu_mask_read_proc(char page, char *start, off_t off,
1da177e4 LT	443	int count, int eof, void data)
	444	{
	445	int len = cpumask_scnprintf(page, count, (cpumask_t )data);
	446	if (count - len < 2)
	447	return -EINVAL;
	448	len += sprintf(page + len, "\n");
	449	return len;
	450	}
	451
1ad82fd5 PC	452	static int prof_cpu_mask_write_proc(struct file *file,
1ad82fd5 PC	453	const char __user buffer, unsigned long count, void data)
1da177e4 LT	454	{
	455	cpumask_t mask = (cpumask_t )data;
	456	unsigned long full_count = count, err;
	457	cpumask_t new_value;
	458
01a3ee2b	459	err = cpumask_parse_user(buffer, count, new_value);
1da177e4 LT	460	if (err)
	461	return err;
	462
	463	*mask = new_value;
	464	return full_count;
	465	}
	466
	467	void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir)
	468	{
	469	struct proc_dir_entry *entry;
	470
	471	/* create /proc/irq/prof_cpu_mask */
1ad82fd5 PC	472	entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir);
1ad82fd5 PC	473	if (!entry)
1da177e4	474	return;
1da177e4 LT	475	entry->data = (void *)&prof_cpu_mask;
	476	entry->read_proc = prof_cpu_mask_read_proc;
	477	entry->write_proc = prof_cpu_mask_write_proc;
	478	}
	479
	480	/*
	481	* This function accesses profiling information. The returned data is
	482	* binary: the sampling step and the actual contents of the profile
	483	* buffer. Use of the program readprofile is recommended in order to
	484	* get meaningful info out of these data.
	485	*/
	486	static ssize_t
	487	read_profile(struct file file, char __user buf, size_t count, loff_t *ppos)
	488	{
	489	unsigned long p = *ppos;
	490	ssize_t read;
1ad82fd5	491	char *pnt;
1da177e4 LT	492	unsigned int sample_step = 1 << prof_shift;
	493
	494	profile_flip_buffers();
	495	if (p >= (prof_len+1)*sizeof(unsigned int))
	496	return 0;
	497	if (count > (prof_len+1)*sizeof(unsigned int) - p)
	498	count = (prof_len+1)*sizeof(unsigned int) - p;
	499	read = 0;
	500
	501	while (p < sizeof(unsigned int) && count > 0) {
1ad82fd5	502	if (put_user(((char )(&sample_step)+p), buf))
064b022c	503	return -EFAULT;
1da177e4 LT	504	buf++; p++; count--; read++;
	505	}
	506	pnt = (char *)prof_buffer + p - sizeof(atomic_t);
1ad82fd5	507	if (copy_to_user(buf, (void *)pnt, count))
1da177e4 LT	508	return -EFAULT;
	509	read += count;
	510	*ppos += read;
	511	return read;
	512	}
	513
	514	/*
	515	* Writing to /proc/profile resets the counters
	516	*
	517	* Writing a 'profiling multiplier' value into it also re-sets the profiling
	518	* interrupt frequency, on architectures that support this.
	519	*/
	520	static ssize_t write_profile(struct file file, const char __user buf,
	521	size_t count, loff_t *ppos)
	522	{
	523	#ifdef CONFIG_SMP
1ad82fd5	524	extern int setup_profiling_timer(unsigned int multiplier);
1da177e4 LT	525
	526	if (count == sizeof(int)) {
	527	unsigned int multiplier;
	528
	529	if (copy_from_user(&multiplier, buf, sizeof(int)))
	530	return -EFAULT;
	531
	532	if (setup_profiling_timer(multiplier))
	533	return -EINVAL;
	534	}
	535	#endif
	536	profile_discard_flip_buffers();
	537	memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
	538	return count;
	539	}
	540
15ad7cdc	541	static const struct file_operations proc_profile_operations = {
1da177e4 LT	542	.read = read_profile,
	543	.write = write_profile,
	544	};
	545
	546	#ifdef CONFIG_SMP
e270219f	547	static inline void profile_nop(void *unused)
1da177e4 LT	548	{
	549	}
	550
22b8ce94	551	static int create_hash_tables(void)
1da177e4 LT	552	{
	553	int cpu;
	554
	555	for_each_online_cpu(cpu) {
	556	int node = cpu_to_node(cpu);
	557	struct page *page;
	558
fbd98167 CL	559	page = alloc_pages_node(node,
	560	GFP_KERNEL \| __GFP_ZERO \| GFP_THISNODE,
	561	0);
1da177e4 LT	562	if (!page)
	563	goto out_cleanup;
	564	per_cpu(cpu_profile_hits, cpu)[1]
	565	= (struct profile_hit *)page_address(page);
fbd98167 CL	566	page = alloc_pages_node(node,
	567	GFP_KERNEL \| __GFP_ZERO \| GFP_THISNODE,
	568	0);
1da177e4 LT	569	if (!page)
	570	goto out_cleanup;
	571	per_cpu(cpu_profile_hits, cpu)[0]
	572	= (struct profile_hit *)page_address(page);
	573	}
	574	return 0;
	575	out_cleanup:
	576	prof_on = 0;
d59dd462	577	smp_mb();
15c8b6c1	578	on_each_cpu(profile_nop, NULL, 1);
1da177e4 LT	579	for_each_online_cpu(cpu) {
	580	struct page *page;
	581
	582	if (per_cpu(cpu_profile_hits, cpu)[0]) {
	583	page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
	584	per_cpu(cpu_profile_hits, cpu)[0] = NULL;
	585	__free_page(page);
	586	}
	587	if (per_cpu(cpu_profile_hits, cpu)[1]) {
	588	page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
	589	per_cpu(cpu_profile_hits, cpu)[1] = NULL;
	590	__free_page(page);
	591	}
	592	}
	593	return -1;
	594	}
	595	#else
	596	#define create_hash_tables() ({ 0; })
	597	#endif
	598
84196414	599	int __ref create_proc_profile(void) /* false positive from hotcpu_notifier */
1da177e4 LT	600	{
	601	struct proc_dir_entry *entry;
	602
	603	if (!prof_on)
	604	return 0;
	605	if (create_hash_tables())
22b8ce94	606	return -ENOMEM;
c33fff0a DL	607	entry = proc_create("profile", S_IWUSR \| S_IRUGO,
c33fff0a DL	608	NULL, &proc_profile_operations);
1ad82fd5	609	if (!entry)
1da177e4	610	return 0;
1da177e4 LT	611	entry->size = (1+prof_len) * sizeof(atomic_t);
	612	hotcpu_notifier(profile_cpu_callback, 0);
	613	return 0;
	614	}
	615	module_init(create_proc_profile);
	616	#endif /* CONFIG_PROC_FS */