[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/profile.h>
#include <linux/highmem.h>
#include <linux/mutex.h>
#include <asm/sections.h>
#include <asm/semaphore.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 */
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 */

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

	if (!strncmp(str, sleepstr, strlen(sleepstr))) {
		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 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);


void __init profile_init(void)
{
	if (!prof_on) 
		return;
 
	/* only text is profiled */
	prof_len = (_etext - _stext) >> prof_shift;
	prof_buffer = alloc_bootmem(prof_len*sizeof(atomic_t));
}

/* Profile event notifications */
 
#ifdef CONFIG_PROFILING
 
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);
}

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

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;
}

 
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;
}

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

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(register_timer_hook);
EXPORT_SYMBOL_GPL(unregister_timer_hook);
EXPORT_SYMBOL_GPL(task_handoff_register);
EXPORT_SYMBOL_GPL(task_handoff_unregister);

#endif /* CONFIG_PROFILING */

EXPORT_SYMBOL_GPL(profile_event_register);
EXPORT_SYMBOL_GPL(profile_event_unregister);

#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, 0, 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, 0, 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 __devinit 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 | GFP_THISNODE,
					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 | GFP_THISNODE,
					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 */
	if (!(entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir)))
		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 void __init profile_nop(void *unused)
{
}

static int __init 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, 0, 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

static int __init create_proc_profile(void)
{
	struct proc_dir_entry *entry;

	if (!prof_on)
		return 0;
	if (create_hash_tables())
		return -1;
	if (!(entry = create_proc_entry("profile", S_IWUSR | S_IRUGO, NULL)))
		return 0;
	entry->proc_fops = &proc_profile_operations;
	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>
	23	#include <linux/profile.h>
	24	#include <linux/highmem.h>
97d1f15b	25	#include <linux/mutex.h>
1da177e4 LT	26	#include <asm/sections.h>
1da177e4 LT	27	#include <asm/semaphore.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 */
6c036527	40	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 */
	54
	55	static int __init profile_setup(char * str)
	56	{
dfaa9c94	57	static char __initdata schedstr[] = "schedule";
ece8a684	58	static char __initdata sleepstr[] = "sleep";
07031e14	59	static char __initdata kvmstr[] = "kvm";
1da177e4 LT	60	int par;
1da177e4 LT	61
ece8a684 IM	62	if (!strncmp(str, sleepstr, strlen(sleepstr))) {
	63	prof_on = SLEEP_PROFILING;
	64	if (str[strlen(sleepstr)] == ',')
	65	str += strlen(sleepstr) + 1;
	66	if (get_option(&str, &par))
	67	prof_shift = par;
	68	printk(KERN_INFO
	69	"kernel sleep profiling enabled (shift: %ld)\n",
	70	prof_shift);
a75acf85	71	} else if (!strncmp(str, schedstr, strlen(schedstr))) {
1da177e4	72	prof_on = SCHED_PROFILING;
dfaa9c94 NYC	73	if (str[strlen(schedstr)] == ',')
	74	str += strlen(schedstr) + 1;
	75	if (get_option(&str, &par))
	76	prof_shift = par;
	77	printk(KERN_INFO
	78	"kernel schedule profiling enabled (shift: %ld)\n",
	79	prof_shift);
07031e14 IM	80	} else if (!strncmp(str, kvmstr, strlen(kvmstr))) {
	81	prof_on = KVM_PROFILING;
	82	if (str[strlen(kvmstr)] == ',')
	83	str += strlen(kvmstr) + 1;
	84	if (get_option(&str, &par))
	85	prof_shift = par;
	86	printk(KERN_INFO
	87	"kernel KVM profiling enabled (shift: %ld)\n",
	88	prof_shift);
dfaa9c94	89	} else if (get_option(&str, &par)) {
1da177e4 LT	90	prof_shift = par;
	91	prof_on = CPU_PROFILING;
	92	printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n",
	93	prof_shift);
	94	}
	95	return 1;
	96	}
	97	__setup("profile=", profile_setup);
	98
	99
	100	void __init profile_init(void)
	101	{
	102	if (!prof_on)
	103	return;
	104
	105	/* only text is profiled */
	106	prof_len = (_etext - _stext) >> prof_shift;
	107	prof_buffer = alloc_bootmem(prof_len*sizeof(atomic_t));
	108	}
	109
	110	/* Profile event notifications */
	111
	112	#ifdef CONFIG_PROFILING
	113
e041c683 AS	114	static BLOCKING_NOTIFIER_HEAD(task_exit_notifier);
	115	static ATOMIC_NOTIFIER_HEAD(task_free_notifier);
	116	static BLOCKING_NOTIFIER_HEAD(munmap_notifier);
1da177e4 LT	117
	118	void profile_task_exit(struct task_struct * task)
	119	{
e041c683	120	blocking_notifier_call_chain(&task_exit_notifier, 0, task);
1da177e4 LT	121	}
	122
	123	int profile_handoff_task(struct task_struct * task)
	124	{
	125	int ret;
e041c683	126	ret = atomic_notifier_call_chain(&task_free_notifier, 0, task);
1da177e4 LT	127	return (ret == NOTIFY_OK) ? 1 : 0;
	128	}
	129
	130	void profile_munmap(unsigned long addr)
	131	{
e041c683	132	blocking_notifier_call_chain(&munmap_notifier, 0, (void *)addr);
1da177e4 LT	133	}
	134
	135	int task_handoff_register(struct notifier_block * n)
	136	{
e041c683	137	return atomic_notifier_chain_register(&task_free_notifier, n);
1da177e4 LT	138	}
	139
	140	int task_handoff_unregister(struct notifier_block * n)
	141	{
e041c683	142	return atomic_notifier_chain_unregister(&task_free_notifier, n);
1da177e4 LT	143	}
	144
	145	int profile_event_register(enum profile_type type, struct notifier_block * n)
	146	{
	147	int err = -EINVAL;
	148
1da177e4 LT	149	switch (type) {
1da177e4 LT	150	case PROFILE_TASK_EXIT:
e041c683 AS	151	err = blocking_notifier_chain_register(
e041c683 AS	152	&task_exit_notifier, n);
1da177e4 LT	153	break;
1da177e4 LT	154	case PROFILE_MUNMAP:
e041c683 AS	155	err = blocking_notifier_chain_register(
e041c683 AS	156	&munmap_notifier, n);
1da177e4 LT	157	break;
	158	}
	159
1da177e4 LT	160	return err;
	161	}
	162
	163
	164	int profile_event_unregister(enum profile_type type, struct notifier_block * n)
	165	{
	166	int err = -EINVAL;
	167
1da177e4 LT	168	switch (type) {
1da177e4 LT	169	case PROFILE_TASK_EXIT:
e041c683 AS	170	err = blocking_notifier_chain_unregister(
e041c683 AS	171	&task_exit_notifier, n);
1da177e4 LT	172	break;
1da177e4 LT	173	case PROFILE_MUNMAP:
e041c683 AS	174	err = blocking_notifier_chain_unregister(
e041c683 AS	175	&munmap_notifier, n);
1da177e4 LT	176	break;
	177	}
	178
1da177e4 LT	179	return err;
	180	}
	181
	182	int register_timer_hook(int (hook)(struct pt_regs ))
	183	{
	184	if (timer_hook)
	185	return -EBUSY;
	186	timer_hook = hook;
	187	return 0;
	188	}
	189
	190	void unregister_timer_hook(int (hook)(struct pt_regs ))
	191	{
	192	WARN_ON(hook != timer_hook);
	193	timer_hook = NULL;
	194	/* make sure all CPUs see the NULL hook */
fbd568a3	195	synchronize_sched(); /* Allow ongoing interrupts to complete. */
1da177e4 LT	196	}
	197
	198	EXPORT_SYMBOL_GPL(register_timer_hook);
	199	EXPORT_SYMBOL_GPL(unregister_timer_hook);
	200	EXPORT_SYMBOL_GPL(task_handoff_register);
	201	EXPORT_SYMBOL_GPL(task_handoff_unregister);
	202
	203	#endif /* CONFIG_PROFILING */
	204
	205	EXPORT_SYMBOL_GPL(profile_event_register);
	206	EXPORT_SYMBOL_GPL(profile_event_unregister);
	207
	208	#ifdef CONFIG_SMP
	209	/*
	210	* Each cpu has a pair of open-addressed hashtables for pending
	211	* profile hits. read_profile() IPI's all cpus to request them
	212	* to flip buffers and flushes their contents to prof_buffer itself.
	213	* Flip requests are serialized by the profile_flip_mutex. The sole
	214	* use of having a second hashtable is for avoiding cacheline
	215	* contention that would otherwise happen during flushes of pending
	216	* profile hits required for the accuracy of reported profile hits
	217	* and so resurrect the interrupt livelock issue.
	218	*
	219	* The open-addressed hashtables are indexed by profile buffer slot
	220	* and hold the number of pending hits to that profile buffer slot on
	221	* a cpu in an entry. When the hashtable overflows, all pending hits
	222	* are accounted to their corresponding profile buffer slots with
	223	* atomic_add() and the hashtable emptied. As numerous pending hits
	224	* may be accounted to a profile buffer slot in a hashtable entry,
	225	* this amortizes a number of atomic profile buffer increments likely
	226	* to be far larger than the number of entries in the hashtable,
	227	* particularly given that the number of distinct profile buffer
	228	* positions to which hits are accounted during short intervals (e.g.
	229	* several seconds) is usually very small. Exclusion from buffer
	230	* flipping is provided by interrupt disablement (note that for
ece8a684 IM	231	* SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
ece8a684 IM	232	* process context).
1da177e4 LT	233	* The hash function is meant to be lightweight as opposed to strong,
	234	* and was vaguely inspired by ppc64 firmware-supported inverted
	235	* pagetable hash functions, but uses a full hashtable full of finite
	236	* collision chains, not just pairs of them.
	237	*
	238	* -- wli
	239	*/
	240	static void __profile_flip_buffers(void *unused)
	241	{
	242	int cpu = smp_processor_id();
	243
	244	per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
	245	}
	246
	247	static void profile_flip_buffers(void)
	248	{
	249	int i, j, cpu;
	250
97d1f15b	251	mutex_lock(&profile_flip_mutex);
1da177e4 LT	252	j = per_cpu(cpu_profile_flip, get_cpu());
	253	put_cpu();
	254	on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
	255	for_each_online_cpu(cpu) {
	256	struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
	257	for (i = 0; i < NR_PROFILE_HIT; ++i) {
	258	if (!hits[i].hits) {
	259	if (hits[i].pc)
	260	hits[i].pc = 0;
	261	continue;
	262	}
	263	atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
	264	hits[i].hits = hits[i].pc = 0;
	265	}
	266	}
97d1f15b	267	mutex_unlock(&profile_flip_mutex);
1da177e4 LT	268	}
	269
	270	static void profile_discard_flip_buffers(void)
	271	{
	272	int i, cpu;
	273
97d1f15b	274	mutex_lock(&profile_flip_mutex);
1da177e4 LT	275	i = per_cpu(cpu_profile_flip, get_cpu());
	276	put_cpu();
	277	on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
	278	for_each_online_cpu(cpu) {
	279	struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
	280	memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
	281	}
97d1f15b	282	mutex_unlock(&profile_flip_mutex);
1da177e4 LT	283	}
1da177e4 LT	284
ece8a684	285	void profile_hits(int type, void *__pc, unsigned int nr_hits)
1da177e4 LT	286	{
	287	unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
	288	int i, j, cpu;
	289	struct profile_hit *hits;
	290
	291	if (prof_on != type \|\| !prof_buffer)
	292	return;
	293	pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
	294	i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
	295	secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
	296	cpu = get_cpu();
	297	hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
	298	if (!hits) {
	299	put_cpu();
	300	return;
	301	}
ece8a684 IM	302	/*
	303	* We buffer the global profiler buffer into a per-CPU
	304	* queue and thus reduce the number of global (and possibly
	305	* NUMA-alien) accesses. The write-queue is self-coalescing:
	306	*/
1da177e4 LT	307	local_irq_save(flags);
	308	do {
	309	for (j = 0; j < PROFILE_GRPSZ; ++j) {
	310	if (hits[i + j].pc == pc) {
ece8a684	311	hits[i + j].hits += nr_hits;
1da177e4 LT	312	goto out;
	313	} else if (!hits[i + j].hits) {
	314	hits[i + j].pc = pc;
ece8a684	315	hits[i + j].hits = nr_hits;
1da177e4 LT	316	goto out;
	317	}
	318	}
	319	i = (i + secondary) & (NR_PROFILE_HIT - 1);
	320	} while (i != primary);
ece8a684 IM	321
	322	/*
	323	* Add the current hit(s) and flush the write-queue out
	324	* to the global buffer:
	325	*/
	326	atomic_add(nr_hits, &prof_buffer[pc]);
1da177e4 LT	327	for (i = 0; i < NR_PROFILE_HIT; ++i) {
	328	atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
	329	hits[i].pc = hits[i].hits = 0;
	330	}
	331	out:
	332	local_irq_restore(flags);
	333	put_cpu();
	334	}
	335
9c7b216d	336	static int __devinit profile_cpu_callback(struct notifier_block *info,
1da177e4 LT	337	unsigned long action, void *__cpu)
	338	{
	339	int node, cpu = (unsigned long)__cpu;
	340	struct page *page;
	341
	342	switch (action) {
	343	case CPU_UP_PREPARE:
8bb78442	344	case CPU_UP_PREPARE_FROZEN:
1da177e4 LT	345	node = cpu_to_node(cpu);
	346	per_cpu(cpu_profile_flip, cpu) = 0;
	347	if (!per_cpu(cpu_profile_hits, cpu)[1]) {
fbd98167 CL	348	page = alloc_pages_node(node,
	349	GFP_KERNEL \| __GFP_ZERO \| GFP_THISNODE,
	350	0);
1da177e4 LT	351	if (!page)
	352	return NOTIFY_BAD;
	353	per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
	354	}
	355	if (!per_cpu(cpu_profile_hits, cpu)[0]) {
fbd98167 CL	356	page = alloc_pages_node(node,
	357	GFP_KERNEL \| __GFP_ZERO \| GFP_THISNODE,
	358	0);
1da177e4 LT	359	if (!page)
	360	goto out_free;
	361	per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
	362	}
	363	break;
	364	out_free:
	365	page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
	366	per_cpu(cpu_profile_hits, cpu)[1] = NULL;
	367	__free_page(page);
	368	return NOTIFY_BAD;
	369	case CPU_ONLINE:
8bb78442	370	case CPU_ONLINE_FROZEN:
1da177e4 LT	371	cpu_set(cpu, prof_cpu_mask);
	372	break;
	373	case CPU_UP_CANCELED:
8bb78442	374	case CPU_UP_CANCELED_FROZEN:
1da177e4	375	case CPU_DEAD:
8bb78442	376	case CPU_DEAD_FROZEN:
1da177e4 LT	377	cpu_clear(cpu, prof_cpu_mask);
	378	if (per_cpu(cpu_profile_hits, cpu)[0]) {
	379	page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
	380	per_cpu(cpu_profile_hits, cpu)[0] = NULL;
	381	__free_page(page);
	382	}
	383	if (per_cpu(cpu_profile_hits, cpu)[1]) {
	384	page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
	385	per_cpu(cpu_profile_hits, cpu)[1] = NULL;
	386	__free_page(page);
	387	}
	388	break;
	389	}
	390	return NOTIFY_OK;
	391	}
1da177e4 LT	392	#else /* !CONFIG_SMP */
	393	#define profile_flip_buffers() do { } while (0)
	394	#define profile_discard_flip_buffers() do { } while (0)
02316067	395	#define profile_cpu_callback NULL
1da177e4	396
ece8a684	397	void profile_hits(int type, void *__pc, unsigned int nr_hits)
1da177e4 LT	398	{
	399	unsigned long pc;
	400
	401	if (prof_on != type \|\| !prof_buffer)
	402	return;
	403	pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
ece8a684	404	atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
1da177e4 LT	405	}
	406	#endif /* !CONFIG_SMP */
	407
bbe1a59b AM	408	EXPORT_SYMBOL_GPL(profile_hits);
bbe1a59b AM	409
7d12e780	410	void profile_tick(int type)
1da177e4	411	{
7d12e780 DH	412	struct pt_regs *regs = get_irq_regs();
7d12e780 DH	413
1da177e4 LT	414	if (type == CPU_PROFILING && timer_hook)
	415	timer_hook(regs);
	416	if (!user_mode(regs) && cpu_isset(smp_processor_id(), prof_cpu_mask))
	417	profile_hit(type, (void *)profile_pc(regs));
	418	}
	419
	420	#ifdef CONFIG_PROC_FS
	421	#include <linux/proc_fs.h>
	422	#include <asm/uaccess.h>
	423	#include <asm/ptrace.h>
	424
	425	static int prof_cpu_mask_read_proc (char page, char *start, off_t off,
	426	int count, int eof, void data)
	427	{
	428	int len = cpumask_scnprintf(page, count, (cpumask_t )data);
	429	if (count - len < 2)
	430	return -EINVAL;
	431	len += sprintf(page + len, "\n");
	432	return len;
	433	}
	434
	435	static int prof_cpu_mask_write_proc (struct file file, const char __user buffer,
	436	unsigned long count, void *data)
	437	{
	438	cpumask_t mask = (cpumask_t )data;
	439	unsigned long full_count = count, err;
	440	cpumask_t new_value;
	441
01a3ee2b	442	err = cpumask_parse_user(buffer, count, new_value);
1da177e4 LT	443	if (err)
	444	return err;
	445
	446	*mask = new_value;
	447	return full_count;
	448	}
	449
	450	void create_prof_cpu_mask(struct proc_dir_entry *root_irq_dir)
	451	{
	452	struct proc_dir_entry *entry;
	453
	454	/* create /proc/irq/prof_cpu_mask */
	455	if (!(entry = create_proc_entry("prof_cpu_mask", 0600, root_irq_dir)))
	456	return;
1da177e4 LT	457	entry->data = (void *)&prof_cpu_mask;
	458	entry->read_proc = prof_cpu_mask_read_proc;
	459	entry->write_proc = prof_cpu_mask_write_proc;
	460	}
	461
	462	/*
	463	* This function accesses profiling information. The returned data is
	464	* binary: the sampling step and the actual contents of the profile
	465	* buffer. Use of the program readprofile is recommended in order to
	466	* get meaningful info out of these data.
	467	*/
	468	static ssize_t
	469	read_profile(struct file file, char __user buf, size_t count, loff_t *ppos)
	470	{
	471	unsigned long p = *ppos;
	472	ssize_t read;
	473	char * pnt;
	474	unsigned int sample_step = 1 << prof_shift;
	475
	476	profile_flip_buffers();
	477	if (p >= (prof_len+1)*sizeof(unsigned int))
	478	return 0;
	479	if (count > (prof_len+1)*sizeof(unsigned int) - p)
	480	count = (prof_len+1)*sizeof(unsigned int) - p;
	481	read = 0;
	482
	483	while (p < sizeof(unsigned int) && count > 0) {
064b022c HC	484	if (put_user(((char )(&sample_step)+p),buf))
064b022c HC	485	return -EFAULT;
1da177e4 LT	486	buf++; p++; count--; read++;
	487	}
	488	pnt = (char *)prof_buffer + p - sizeof(atomic_t);
	489	if (copy_to_user(buf,(void *)pnt,count))
	490	return -EFAULT;
	491	read += count;
	492	*ppos += read;
	493	return read;
	494	}
	495
	496	/*
	497	* Writing to /proc/profile resets the counters
	498	*
	499	* Writing a 'profiling multiplier' value into it also re-sets the profiling
	500	* interrupt frequency, on architectures that support this.
	501	*/
	502	static ssize_t write_profile(struct file file, const char __user buf,
	503	size_t count, loff_t *ppos)
	504	{
	505	#ifdef CONFIG_SMP
	506	extern int setup_profiling_timer (unsigned int multiplier);
	507
	508	if (count == sizeof(int)) {
	509	unsigned int multiplier;
	510
	511	if (copy_from_user(&multiplier, buf, sizeof(int)))
	512	return -EFAULT;
	513
	514	if (setup_profiling_timer(multiplier))
	515	return -EINVAL;
	516	}
	517	#endif
	518	profile_discard_flip_buffers();
	519	memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
	520	return count;
	521	}
	522
15ad7cdc	523	static const struct file_operations proc_profile_operations = {
1da177e4 LT	524	.read = read_profile,
	525	.write = write_profile,
	526	};
	527
	528	#ifdef CONFIG_SMP
	529	static void __init profile_nop(void *unused)
	530	{
	531	}
	532
	533	static int __init create_hash_tables(void)
	534	{
	535	int cpu;
	536
	537	for_each_online_cpu(cpu) {
	538	int node = cpu_to_node(cpu);
	539	struct page *page;
	540
fbd98167 CL	541	page = alloc_pages_node(node,
	542	GFP_KERNEL \| __GFP_ZERO \| GFP_THISNODE,
	543	0);
1da177e4 LT	544	if (!page)
	545	goto out_cleanup;
	546	per_cpu(cpu_profile_hits, cpu)[1]
	547	= (struct profile_hit *)page_address(page);
fbd98167 CL	548	page = alloc_pages_node(node,
	549	GFP_KERNEL \| __GFP_ZERO \| GFP_THISNODE,
	550	0);
1da177e4 LT	551	if (!page)
	552	goto out_cleanup;
	553	per_cpu(cpu_profile_hits, cpu)[0]
	554	= (struct profile_hit *)page_address(page);
	555	}
	556	return 0;
	557	out_cleanup:
	558	prof_on = 0;
d59dd462	559	smp_mb();
1da177e4 LT	560	on_each_cpu(profile_nop, NULL, 0, 1);
	561	for_each_online_cpu(cpu) {
	562	struct page *page;
	563
	564	if (per_cpu(cpu_profile_hits, cpu)[0]) {
	565	page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
	566	per_cpu(cpu_profile_hits, cpu)[0] = NULL;
	567	__free_page(page);
	568	}
	569	if (per_cpu(cpu_profile_hits, cpu)[1]) {
	570	page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
	571	per_cpu(cpu_profile_hits, cpu)[1] = NULL;
	572	__free_page(page);
	573	}
	574	}
	575	return -1;
	576	}
	577	#else
	578	#define create_hash_tables() ({ 0; })
	579	#endif
	580
	581	static int __init create_proc_profile(void)
	582	{
	583	struct proc_dir_entry *entry;
	584
	585	if (!prof_on)
	586	return 0;
	587	if (create_hash_tables())
	588	return -1;
	589	if (!(entry = create_proc_entry("profile", S_IWUSR \| S_IRUGO, NULL)))
	590	return 0;
	591	entry->proc_fops = &proc_profile_operations;
	592	entry->size = (1+prof_len) * sizeof(atomic_t);
	593	hotcpu_notifier(profile_cpu_callback, 0);
	594	return 0;
	595	}
	596	module_init(create_proc_profile);
	597	#endif /* CONFIG_PROC_FS */