[deliverable/linux.git] / drivers / oprofile / buffer_sync.c

/**
 * @file buffer_sync.c
 *
 * @remark Copyright 2002 OProfile authors
 * @remark Read the file COPYING
 *
 * @author John Levon <levon@movementarian.org>
 * @author Barry Kasindorf
 *
 * This is the core of the buffer management. Each
 * CPU buffer is processed and entered into the
 * global event buffer. Such processing is necessary
 * in several circumstances, mentioned below.
 *
 * The processing does the job of converting the
 * transitory EIP value into a persistent dentry/offset
 * value that the profiler can record at its leisure.
 *
 * See fs/dcookies.c for a description of the dentry/offset
 * objects.
 */

#include <linux/mm.h>
#include <linux/workqueue.h>
#include <linux/notifier.h>
#include <linux/dcookies.h>
#include <linux/profile.h>
#include <linux/module.h>
#include <linux/fs.h>
#include <linux/oprofile.h>
#include <linux/sched.h>

#include "oprofile_stats.h"
#include "event_buffer.h"
#include "cpu_buffer.h"
#include "buffer_sync.h"

static LIST_HEAD(dying_tasks);
static LIST_HEAD(dead_tasks);
static cpumask_t marked_cpus = CPU_MASK_NONE;
static DEFINE_SPINLOCK(task_mortuary);
static void process_task_mortuary(void);


/* Take ownership of the task struct and place it on the
 * list for processing. Only after two full buffer syncs
 * does the task eventually get freed, because by then
 * we are sure we will not reference it again.
 * Can be invoked from softirq via RCU callback due to
 * call_rcu() of the task struct, hence the _irqsave.
 */
static int
task_free_notify(struct notifier_block *self, unsigned long val, void *data)
{
	unsigned long flags;
	struct task_struct *task = data;
	spin_lock_irqsave(&task_mortuary, flags);
	list_add(&task->tasks, &dying_tasks);
	spin_unlock_irqrestore(&task_mortuary, flags);
	return NOTIFY_OK;
}


/* The task is on its way out. A sync of the buffer means we can catch
 * any remaining samples for this task.
 */
static int
task_exit_notify(struct notifier_block *self, unsigned long val, void *data)
{
	/* To avoid latency problems, we only process the current CPU,
	 * hoping that most samples for the task are on this CPU
	 */
	sync_buffer(raw_smp_processor_id());
	return 0;
}


/* The task is about to try a do_munmap(). We peek at what it's going to
 * do, and if it's an executable region, process the samples first, so
 * we don't lose any. This does not have to be exact, it's a QoI issue
 * only.
 */
static int
munmap_notify(struct notifier_block *self, unsigned long val, void *data)
{
	unsigned long addr = (unsigned long)data;
	struct mm_struct *mm = current->mm;
	struct vm_area_struct *mpnt;

	down_read(&mm->mmap_sem);

	mpnt = find_vma(mm, addr);
	if (mpnt && mpnt->vm_file && (mpnt->vm_flags & VM_EXEC)) {
		up_read(&mm->mmap_sem);
		/* To avoid latency problems, we only process the current CPU,
		 * hoping that most samples for the task are on this CPU
		 */
		sync_buffer(raw_smp_processor_id());
		return 0;
	}

	up_read(&mm->mmap_sem);
	return 0;
}


/* We need to be told about new modules so we don't attribute to a previously
 * loaded module, or drop the samples on the floor.
 */
static int
module_load_notify(struct notifier_block *self, unsigned long val, void *data)
{
#ifdef CONFIG_MODULES
	if (val != MODULE_STATE_COMING)
		return 0;

	/* FIXME: should we process all CPU buffers ? */
	mutex_lock(&buffer_mutex);
	add_event_entry(ESCAPE_CODE);
	add_event_entry(MODULE_LOADED_CODE);
	mutex_unlock(&buffer_mutex);
#endif
	return 0;
}


static struct notifier_block task_free_nb = {
	.notifier_call	= task_free_notify,
};

static struct notifier_block task_exit_nb = {
	.notifier_call	= task_exit_notify,
};

static struct notifier_block munmap_nb = {
	.notifier_call	= munmap_notify,
};

static struct notifier_block module_load_nb = {
	.notifier_call = module_load_notify,
};


static void end_sync(void)
{
	end_cpu_work();
	/* make sure we don't leak task structs */
	process_task_mortuary();
	process_task_mortuary();
}


int sync_start(void)
{
	int err;

	start_cpu_work();

	err = task_handoff_register(&task_free_nb);
	if (err)
		goto out1;
	err = profile_event_register(PROFILE_TASK_EXIT, &task_exit_nb);
	if (err)
		goto out2;
	err = profile_event_register(PROFILE_MUNMAP, &munmap_nb);
	if (err)
		goto out3;
	err = register_module_notifier(&module_load_nb);
	if (err)
		goto out4;

out:
	return err;
out4:
	profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
out3:
	profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
out2:
	task_handoff_unregister(&task_free_nb);
out1:
	end_sync();
	goto out;
}


void sync_stop(void)
{
	unregister_module_notifier(&module_load_nb);
	profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
	profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
	task_handoff_unregister(&task_free_nb);
	end_sync();
}


/* Optimisation. We can manage without taking the dcookie sem
 * because we cannot reach this code without at least one
 * dcookie user still being registered (namely, the reader
 * of the event buffer). */
static inline unsigned long fast_get_dcookie(struct path *path)
{
	unsigned long cookie;

	if (path->dentry->d_cookie)
		return (unsigned long)path->dentry;
	get_dcookie(path, &cookie);
	return cookie;
}


/* Look up the dcookie for the task's first VM_EXECUTABLE mapping,
 * which corresponds loosely to "application name". This is
 * not strictly necessary but allows oprofile to associate
 * shared-library samples with particular applications
 */
static unsigned long get_exec_dcookie(struct mm_struct *mm)
{
	unsigned long cookie = NO_COOKIE;
	struct vm_area_struct *vma;

	if (!mm)
		goto out;

	for (vma = mm->mmap; vma; vma = vma->vm_next) {
		if (!vma->vm_file)
			continue;
		if (!(vma->vm_flags & VM_EXECUTABLE))
			continue;
		cookie = fast_get_dcookie(&vma->vm_file->f_path);
		break;
	}

out:
	return cookie;
}


/* Convert the EIP value of a sample into a persistent dentry/offset
 * pair that can then be added to the global event buffer. We make
 * sure to do this lookup before a mm->mmap modification happens so
 * we don't lose track.
 */
static unsigned long
lookup_dcookie(struct mm_struct *mm, unsigned long addr, off_t *offset)
{
	unsigned long cookie = NO_COOKIE;
	struct vm_area_struct *vma;

	for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {

		if (addr < vma->vm_start || addr >= vma->vm_end)
			continue;

		if (vma->vm_file) {
			cookie = fast_get_dcookie(&vma->vm_file->f_path);
			*offset = (vma->vm_pgoff << PAGE_SHIFT) + addr -
				vma->vm_start;
		} else {
			/* must be an anonymous map */
			*offset = addr;
		}

		break;
	}

	if (!vma)
		cookie = INVALID_COOKIE;

	return cookie;
}

static void increment_tail(struct oprofile_cpu_buffer *b)
{
	unsigned long new_tail = b->tail_pos + 1;

	rmb();	/* be sure fifo pointers are synchromized */

	if (new_tail < b->buffer_size)
		b->tail_pos = new_tail;
	else
		b->tail_pos = 0;
}

static unsigned long last_cookie = INVALID_COOKIE;

static void add_cpu_switch(int i)
{
	add_event_entry(ESCAPE_CODE);
	add_event_entry(CPU_SWITCH_CODE);
	add_event_entry(i);
	last_cookie = INVALID_COOKIE;
}

static void add_kernel_ctx_switch(unsigned int in_kernel)
{
	add_event_entry(ESCAPE_CODE);
	if (in_kernel)
		add_event_entry(KERNEL_ENTER_SWITCH_CODE);
	else
		add_event_entry(KERNEL_EXIT_SWITCH_CODE);
}

static void
add_user_ctx_switch(struct task_struct const *task, unsigned long cookie)
{
	add_event_entry(ESCAPE_CODE);
	add_event_entry(CTX_SWITCH_CODE);
	add_event_entry(task->pid);
	add_event_entry(cookie);
	/* Another code for daemon back-compat */
	add_event_entry(ESCAPE_CODE);
	add_event_entry(CTX_TGID_CODE);
	add_event_entry(task->tgid);
}


static void add_cookie_switch(unsigned long cookie)
{
	add_event_entry(ESCAPE_CODE);
	add_event_entry(COOKIE_SWITCH_CODE);
	add_event_entry(cookie);
}


static void add_trace_begin(void)
{
	add_event_entry(ESCAPE_CODE);
	add_event_entry(TRACE_BEGIN_CODE);
}

#ifdef CONFIG_OPROFILE_IBS

#define IBS_FETCH_CODE_SIZE	2
#define IBS_OP_CODE_SIZE	5
#define IBS_EIP(offset)				\
	(((struct op_sample *)&cpu_buf->buffer[(offset)])->eip)
#define IBS_EVENT(offset)				\
	(((struct op_sample *)&cpu_buf->buffer[(offset)])->event)

/*
 * Add IBS fetch and op entries to event buffer
 */
static void add_ibs_begin(struct oprofile_cpu_buffer *cpu_buf, int code,
	int in_kernel, struct mm_struct *mm)
{
	unsigned long rip;
	int i, count;
	unsigned long ibs_cookie = 0;
	off_t offset;

	increment_tail(cpu_buf);	/* move to RIP entry */

	rip = IBS_EIP(cpu_buf->tail_pos);

#ifdef __LP64__
	rip += IBS_EVENT(cpu_buf->tail_pos) << 32;
#endif

	if (mm) {
		ibs_cookie = lookup_dcookie(mm, rip, &offset);

		if (ibs_cookie == NO_COOKIE)
			offset = rip;
		if (ibs_cookie == INVALID_COOKIE) {
			atomic_inc(&oprofile_stats.sample_lost_no_mapping);
			offset = rip;
		}
		if (ibs_cookie != last_cookie) {
			add_cookie_switch(ibs_cookie);
			last_cookie = ibs_cookie;
		}
	} else
		offset = rip;

	add_event_entry(ESCAPE_CODE);
	add_event_entry(code);
	add_event_entry(offset);	/* Offset from Dcookie */

	/* we send the Dcookie offset, but send the raw Linear Add also*/
	add_event_entry(IBS_EIP(cpu_buf->tail_pos));
	add_event_entry(IBS_EVENT(cpu_buf->tail_pos));

	if (code == IBS_FETCH_CODE)
		count = IBS_FETCH_CODE_SIZE;	/*IBS FETCH is 2 int64s*/
	else
		count = IBS_OP_CODE_SIZE;	/*IBS OP is 5 int64s*/

	for (i = 0; i < count; i++) {
		increment_tail(cpu_buf);
		add_event_entry(IBS_EIP(cpu_buf->tail_pos));
		add_event_entry(IBS_EVENT(cpu_buf->tail_pos));
	}
}

#endif

static void add_sample_entry(unsigned long offset, unsigned long event)
{
	add_event_entry(offset);
	add_event_entry(event);
}


static int add_us_sample(struct mm_struct *mm, struct op_sample *s)
{
	unsigned long cookie;
	off_t offset;

	cookie = lookup_dcookie(mm, s->eip, &offset);

	if (cookie == INVALID_COOKIE) {
		atomic_inc(&oprofile_stats.sample_lost_no_mapping);
		return 0;
	}

	if (cookie != last_cookie) {
		add_cookie_switch(cookie);
		last_cookie = cookie;
	}

	add_sample_entry(offset, s->event);

	return 1;
}


/* Add a sample to the global event buffer. If possible the
 * sample is converted into a persistent dentry/offset pair
 * for later lookup from userspace.
 */
static int
add_sample(struct mm_struct *mm, struct op_sample *s, int in_kernel)
{
	if (in_kernel) {
		add_sample_entry(s->eip, s->event);
		return 1;
	} else if (mm) {
		return add_us_sample(mm, s);
	} else {
		atomic_inc(&oprofile_stats.sample_lost_no_mm);
	}
	return 0;
}


static void release_mm(struct mm_struct *mm)
{
	if (!mm)
		return;
	up_read(&mm->mmap_sem);
	mmput(mm);
}


static struct mm_struct *take_tasks_mm(struct task_struct *task)
{
	struct mm_struct *mm = get_task_mm(task);
	if (mm)
		down_read(&mm->mmap_sem);
	return mm;
}


static inline int is_code(unsigned long val)
{
	return val == ESCAPE_CODE;
}


/* "acquire" as many cpu buffer slots as we can */
static unsigned long get_slots(struct oprofile_cpu_buffer *b)
{
	unsigned long head = b->head_pos;
	unsigned long tail = b->tail_pos;

	/*
	 * Subtle. This resets the persistent last_task
	 * and in_kernel values used for switching notes.
	 * BUT, there is a small window between reading
	 * head_pos, and this call, that means samples
	 * can appear at the new head position, but not
	 * be prefixed with the notes for switching
	 * kernel mode or a task switch. This small hole
	 * can lead to mis-attribution or samples where
	 * we don't know if it's in the kernel or not,
	 * at the start of an event buffer.
	 */
	cpu_buffer_reset(b);

	if (head >= tail)
		return head - tail;

	return head + (b->buffer_size - tail);
}


/* Move tasks along towards death. Any tasks on dead_tasks
 * will definitely have no remaining references in any
 * CPU buffers at this point, because we use two lists,
 * and to have reached the list, it must have gone through
 * one full sync already.
 */
static void process_task_mortuary(void)
{
	unsigned long flags;
	LIST_HEAD(local_dead_tasks);
	struct task_struct *task;
	struct task_struct *ttask;

	spin_lock_irqsave(&task_mortuary, flags);

	list_splice_init(&dead_tasks, &local_dead_tasks);
	list_splice_init(&dying_tasks, &dead_tasks);

	spin_unlock_irqrestore(&task_mortuary, flags);

	list_for_each_entry_safe(task, ttask, &local_dead_tasks, tasks) {
		list_del(&task->tasks);
		free_task(task);
	}
}


static void mark_done(int cpu)
{
	int i;

	cpu_set(cpu, marked_cpus);

	for_each_online_cpu(i) {
		if (!cpu_isset(i, marked_cpus))
			return;
	}

	/* All CPUs have been processed at least once,
	 * we can process the mortuary once
	 */
	process_task_mortuary();

	cpus_clear(marked_cpus);
}


/* FIXME: this is not sufficient if we implement syscall barrier backtrace
 * traversal, the code switch to sb_sample_start at first kernel enter/exit
 * switch so we need a fifth state and some special handling in sync_buffer()
 */
typedef enum {
	sb_bt_ignore = -2,
	sb_buffer_start,
	sb_bt_start,
	sb_sample_start,
} sync_buffer_state;

/* Sync one of the CPU's buffers into the global event buffer.
 * Here we need to go through each batch of samples punctuated
 * by context switch notes, taking the task's mmap_sem and doing
 * lookup in task->mm->mmap to convert EIP into dcookie/offset
 * value.
 */
void sync_buffer(int cpu)
{
	struct oprofile_cpu_buffer *cpu_buf = &per_cpu(cpu_buffer, cpu);
	struct mm_struct *mm = NULL;
	struct task_struct *new;
	unsigned long cookie = 0;
	int in_kernel = 1;
	unsigned int i;
	sync_buffer_state state = sb_buffer_start;
	unsigned long available;

	mutex_lock(&buffer_mutex);

	add_cpu_switch(cpu);

	/* Remember, only we can modify tail_pos */

	available = get_slots(cpu_buf);

	for (i = 0; i < available; ++i) {
		struct op_sample *s = &cpu_buf->buffer[cpu_buf->tail_pos];

		if (is_code(s->eip)) {
			if (s->event <= CPU_IS_KERNEL) {
				/* kernel/userspace switch */
				in_kernel = s->event;
				if (state == sb_buffer_start)
					state = sb_sample_start;
				add_kernel_ctx_switch(s->event);
			} else if (s->event == CPU_TRACE_BEGIN) {
				state = sb_bt_start;
				add_trace_begin();
#ifdef CONFIG_OPROFILE_IBS
			} else if (s->event == IBS_FETCH_BEGIN) {
				state = sb_bt_start;
				add_ibs_begin(cpu_buf,
					IBS_FETCH_CODE, in_kernel, mm);
			} else if (s->event == IBS_OP_BEGIN) {
				state = sb_bt_start;
				add_ibs_begin(cpu_buf,
					IBS_OP_CODE, in_kernel, mm);
#endif
			} else {
				struct mm_struct *oldmm = mm;

				/* userspace context switch */
				new = (struct task_struct *)s->event;

				release_mm(oldmm);
				mm = take_tasks_mm(new);
				if (mm != oldmm)
					cookie = get_exec_dcookie(mm);
				add_user_ctx_switch(new, cookie);
			}
		} else if (state >= sb_bt_start &&
			   !add_sample(mm, s, in_kernel)) {
			if (state == sb_bt_start) {
				state = sb_bt_ignore;
				atomic_inc(&oprofile_stats.bt_lost_no_mapping);
			}
		}

		increment_tail(cpu_buf);
	}
	release_mm(mm);

	mark_done(cpu);

	mutex_unlock(&buffer_mutex);
}
Commit	Line	Data
1da177e4 LT	1	/**
	2	* @file buffer_sync.c
	3	*
	4	* @remark Copyright 2002 OProfile authors
	5	* @remark Read the file COPYING
	6	*
	7	* @author John Levon <levon@movementarian.org>
345c2573	8	* @author Barry Kasindorf
1da177e4 LT	9	*
	10	* This is the core of the buffer management. Each
	11	* CPU buffer is processed and entered into the
	12	* global event buffer. Such processing is necessary
	13	* in several circumstances, mentioned below.
	14	*
	15	* The processing does the job of converting the
	16	* transitory EIP value into a persistent dentry/offset
	17	* value that the profiler can record at its leisure.
	18	*
	19	* See fs/dcookies.c for a description of the dentry/offset
	20	* objects.
	21	*/
	22
	23	#include <linux/mm.h>
	24	#include <linux/workqueue.h>
	25	#include <linux/notifier.h>
	26	#include <linux/dcookies.h>
	27	#include <linux/profile.h>
	28	#include <linux/module.h>
	29	#include <linux/fs.h>
1474855d	30	#include <linux/oprofile.h>
e8edc6e0	31	#include <linux/sched.h>
1474855d	32
1da177e4 LT	33	#include "oprofile_stats.h"
	34	#include "event_buffer.h"
	35	#include "cpu_buffer.h"
	36	#include "buffer_sync.h"
73185e0a	37
1da177e4 LT	38	static LIST_HEAD(dying_tasks);
	39	static LIST_HEAD(dead_tasks);
	40	static cpumask_t marked_cpus = CPU_MASK_NONE;
	41	static DEFINE_SPINLOCK(task_mortuary);
	42	static void process_task_mortuary(void);
	43
	44
	45	/* Take ownership of the task struct and place it on the
	46	* list for processing. Only after two full buffer syncs
	47	* does the task eventually get freed, because by then
	48	* we are sure we will not reference it again.
4369ef3c PM	49	* Can be invoked from softirq via RCU callback due to
4369ef3c PM	50	* call_rcu() of the task struct, hence the _irqsave.
1da177e4	51	*/
73185e0a RR	52	static int
73185e0a RR	53	task_free_notify(struct notifier_block self, unsigned long val, void data)
1da177e4	54	{
4369ef3c	55	unsigned long flags;
73185e0a	56	struct task_struct *task = data;
4369ef3c	57	spin_lock_irqsave(&task_mortuary, flags);
1da177e4	58	list_add(&task->tasks, &dying_tasks);
4369ef3c	59	spin_unlock_irqrestore(&task_mortuary, flags);
1da177e4 LT	60	return NOTIFY_OK;
	61	}
	62
	63
	64	/* The task is on its way out. A sync of the buffer means we can catch
	65	* any remaining samples for this task.
	66	*/
73185e0a RR	67	static int
73185e0a RR	68	task_exit_notify(struct notifier_block self, unsigned long val, void data)
1da177e4 LT	69	{
	70	/* To avoid latency problems, we only process the current CPU,
	71	* hoping that most samples for the task are on this CPU
	72	*/
39c715b7	73	sync_buffer(raw_smp_processor_id());
73185e0a	74	return 0;
1da177e4 LT	75	}
	76
	77
	78	/* The task is about to try a do_munmap(). We peek at what it's going to
	79	* do, and if it's an executable region, process the samples first, so
	80	* we don't lose any. This does not have to be exact, it's a QoI issue
	81	* only.
	82	*/
73185e0a RR	83	static int
73185e0a RR	84	munmap_notify(struct notifier_block self, unsigned long val, void data)
1da177e4 LT	85	{
1da177e4 LT	86	unsigned long addr = (unsigned long)data;
73185e0a RR	87	struct mm_struct *mm = current->mm;
73185e0a RR	88	struct vm_area_struct *mpnt;
1da177e4 LT	89
	90	down_read(&mm->mmap_sem);
	91
	92	mpnt = find_vma(mm, addr);
	93	if (mpnt && mpnt->vm_file && (mpnt->vm_flags & VM_EXEC)) {
	94	up_read(&mm->mmap_sem);
	95	/* To avoid latency problems, we only process the current CPU,
	96	* hoping that most samples for the task are on this CPU
	97	*/
39c715b7	98	sync_buffer(raw_smp_processor_id());
1da177e4 LT	99	return 0;
	100	}
	101
	102	up_read(&mm->mmap_sem);
	103	return 0;
	104	}
	105
73185e0a	106
1da177e4 LT	107	/* We need to be told about new modules so we don't attribute to a previously
	108	* loaded module, or drop the samples on the floor.
	109	*/
73185e0a RR	110	static int
73185e0a RR	111	module_load_notify(struct notifier_block self, unsigned long val, void data)
1da177e4 LT	112	{
	113	#ifdef CONFIG_MODULES
	114	if (val != MODULE_STATE_COMING)
	115	return 0;
	116
	117	/* FIXME: should we process all CPU buffers ? */
59cc185a	118	mutex_lock(&buffer_mutex);
1da177e4 LT	119	add_event_entry(ESCAPE_CODE);
1da177e4 LT	120	add_event_entry(MODULE_LOADED_CODE);
59cc185a	121	mutex_unlock(&buffer_mutex);
1da177e4 LT	122	#endif
	123	return 0;
	124	}
	125
73185e0a	126
1da177e4 LT	127	static struct notifier_block task_free_nb = {
	128	.notifier_call = task_free_notify,
	129	};
	130
	131	static struct notifier_block task_exit_nb = {
	132	.notifier_call = task_exit_notify,
	133	};
	134
	135	static struct notifier_block munmap_nb = {
	136	.notifier_call = munmap_notify,
	137	};
	138
	139	static struct notifier_block module_load_nb = {
	140	.notifier_call = module_load_notify,
	141	};
	142
73185e0a	143
1da177e4 LT	144	static void end_sync(void)
	145	{
	146	end_cpu_work();
	147	/* make sure we don't leak task structs */
	148	process_task_mortuary();
	149	process_task_mortuary();
	150	}
	151
	152
	153	int sync_start(void)
	154	{
	155	int err;
	156
	157	start_cpu_work();
	158
	159	err = task_handoff_register(&task_free_nb);
	160	if (err)
	161	goto out1;
	162	err = profile_event_register(PROFILE_TASK_EXIT, &task_exit_nb);
	163	if (err)
	164	goto out2;
	165	err = profile_event_register(PROFILE_MUNMAP, &munmap_nb);
	166	if (err)
	167	goto out3;
	168	err = register_module_notifier(&module_load_nb);
	169	if (err)
	170	goto out4;
	171
	172	out:
	173	return err;
	174	out4:
	175	profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
	176	out3:
	177	profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
	178	out2:
	179	task_handoff_unregister(&task_free_nb);
	180	out1:
	181	end_sync();
	182	goto out;
	183	}
	184
	185
	186	void sync_stop(void)
	187	{
	188	unregister_module_notifier(&module_load_nb);
	189	profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
	190	profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
	191	task_handoff_unregister(&task_free_nb);
	192	end_sync();
	193	}
	194
448678a0	195
1da177e4 LT	196	/* Optimisation. We can manage without taking the dcookie sem
	197	* because we cannot reach this code without at least one
	198	* dcookie user still being registered (namely, the reader
	199	* of the event buffer). */
448678a0	200	static inline unsigned long fast_get_dcookie(struct path *path)
1da177e4 LT	201	{
1da177e4 LT	202	unsigned long cookie;
448678a0 JB	203
	204	if (path->dentry->d_cookie)
	205	return (unsigned long)path->dentry;
	206	get_dcookie(path, &cookie);
1da177e4 LT	207	return cookie;
	208	}
	209
448678a0	210
1da177e4 LT	211	/* Look up the dcookie for the task's first VM_EXECUTABLE mapping,
	212	* which corresponds loosely to "application name". This is
	213	* not strictly necessary but allows oprofile to associate
	214	* shared-library samples with particular applications
	215	*/
73185e0a	216	static unsigned long get_exec_dcookie(struct mm_struct *mm)
1da177e4	217	{
0c0a400d	218	unsigned long cookie = NO_COOKIE;
73185e0a RR	219	struct vm_area_struct *vma;
73185e0a RR	220
1da177e4 LT	221	if (!mm)
1da177e4 LT	222	goto out;
73185e0a	223
1da177e4 LT	224	for (vma = mm->mmap; vma; vma = vma->vm_next) {
	225	if (!vma->vm_file)
	226	continue;
	227	if (!(vma->vm_flags & VM_EXECUTABLE))
	228	continue;
448678a0	229	cookie = fast_get_dcookie(&vma->vm_file->f_path);
1da177e4 LT	230	break;
	231	}
	232
	233	out:
	234	return cookie;
	235	}
	236
	237
	238	/* Convert the EIP value of a sample into a persistent dentry/offset
	239	* pair that can then be added to the global event buffer. We make
	240	* sure to do this lookup before a mm->mmap modification happens so
	241	* we don't lose track.
	242	*/
73185e0a RR	243	static unsigned long
73185e0a RR	244	lookup_dcookie(struct mm_struct mm, unsigned long addr, off_t offset)
1da177e4	245	{
0c0a400d	246	unsigned long cookie = NO_COOKIE;
73185e0a	247	struct vm_area_struct *vma;
1da177e4 LT	248
1da177e4 LT	249	for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
73185e0a	250
1da177e4 LT	251	if (addr < vma->vm_start \|\| addr >= vma->vm_end)
	252	continue;
	253
0c0a400d	254	if (vma->vm_file) {
448678a0	255	cookie = fast_get_dcookie(&vma->vm_file->f_path);
0c0a400d JL	256	*offset = (vma->vm_pgoff << PAGE_SHIFT) + addr -
	257	vma->vm_start;
	258	} else {
	259	/* must be an anonymous map */
	260	*offset = addr;
	261	}
	262
1da177e4 LT	263	break;
	264	}
	265
0c0a400d JL	266	if (!vma)
	267	cookie = INVALID_COOKIE;
	268
1da177e4 LT	269	return cookie;
	270	}
	271
5e11f98d RR	272	static void increment_tail(struct oprofile_cpu_buffer *b)
	273	{
	274	unsigned long new_tail = b->tail_pos + 1;
	275
345c2573	276	rmb(); /* be sure fifo pointers are synchromized */
5e11f98d RR	277
	278	if (new_tail < b->buffer_size)
	279	b->tail_pos = new_tail;
	280	else
	281	b->tail_pos = 0;
	282	}
1da177e4	283
0c0a400d	284	static unsigned long last_cookie = INVALID_COOKIE;
73185e0a	285
1da177e4 LT	286	static void add_cpu_switch(int i)
	287	{
	288	add_event_entry(ESCAPE_CODE);
	289	add_event_entry(CPU_SWITCH_CODE);
	290	add_event_entry(i);
0c0a400d	291	last_cookie = INVALID_COOKIE;
1da177e4 LT	292	}
	293
	294	static void add_kernel_ctx_switch(unsigned int in_kernel)
	295	{
	296	add_event_entry(ESCAPE_CODE);
	297	if (in_kernel)
73185e0a	298	add_event_entry(KERNEL_ENTER_SWITCH_CODE);
1da177e4	299	else
73185e0a	300	add_event_entry(KERNEL_EXIT_SWITCH_CODE);
1da177e4	301	}
73185e0a	302
1da177e4	303	static void
73185e0a	304	add_user_ctx_switch(struct task_struct const *task, unsigned long cookie)
1da177e4 LT	305	{
1da177e4 LT	306	add_event_entry(ESCAPE_CODE);
73185e0a	307	add_event_entry(CTX_SWITCH_CODE);
1da177e4 LT	308	add_event_entry(task->pid);
	309	add_event_entry(cookie);
	310	/* Another code for daemon back-compat */
	311	add_event_entry(ESCAPE_CODE);
	312	add_event_entry(CTX_TGID_CODE);
	313	add_event_entry(task->tgid);
	314	}
	315
73185e0a	316
1da177e4 LT	317	static void add_cookie_switch(unsigned long cookie)
	318	{
	319	add_event_entry(ESCAPE_CODE);
	320	add_event_entry(COOKIE_SWITCH_CODE);
	321	add_event_entry(cookie);
	322	}
	323
73185e0a	324
1da177e4 LT	325	static void add_trace_begin(void)
	326	{
	327	add_event_entry(ESCAPE_CODE);
	328	add_event_entry(TRACE_BEGIN_CODE);
	329	}
	330
852402cc RR	331	#ifdef CONFIG_OPROFILE_IBS
852402cc RR	332
345c2573 BK	333	#define IBS_FETCH_CODE_SIZE 2
	334	#define IBS_OP_CODE_SIZE 5
	335	#define IBS_EIP(offset) \
	336	(((struct op_sample *)&cpu_buf->buffer[(offset)])->eip)
	337	#define IBS_EVENT(offset) \
	338	(((struct op_sample *)&cpu_buf->buffer[(offset)])->event)
	339
	340	/*
	341	* Add IBS fetch and op entries to event buffer
	342	*/
	343	static void add_ibs_begin(struct oprofile_cpu_buffer *cpu_buf, int code,
	344	int in_kernel, struct mm_struct *mm)
	345	{
	346	unsigned long rip;
	347	int i, count;
	348	unsigned long ibs_cookie = 0;
	349	off_t offset;
	350
	351	increment_tail(cpu_buf); /* move to RIP entry */
	352
	353	rip = IBS_EIP(cpu_buf->tail_pos);
	354
	355	#ifdef __LP64__
	356	rip += IBS_EVENT(cpu_buf->tail_pos) << 32;
	357	#endif
	358
	359	if (mm) {
	360	ibs_cookie = lookup_dcookie(mm, rip, &offset);
	361
	362	if (ibs_cookie == NO_COOKIE)
	363	offset = rip;
	364	if (ibs_cookie == INVALID_COOKIE) {
	365	atomic_inc(&oprofile_stats.sample_lost_no_mapping);
	366	offset = rip;
	367	}
	368	if (ibs_cookie != last_cookie) {
	369	add_cookie_switch(ibs_cookie);
	370	last_cookie = ibs_cookie;
	371	}
	372	} else
	373	offset = rip;
	374
	375	add_event_entry(ESCAPE_CODE);
	376	add_event_entry(code);
	377	add_event_entry(offset); /* Offset from Dcookie */
	378
	379	/* we send the Dcookie offset, but send the raw Linear Add also*/
	380	add_event_entry(IBS_EIP(cpu_buf->tail_pos));
	381	add_event_entry(IBS_EVENT(cpu_buf->tail_pos));
	382
	383	if (code == IBS_FETCH_CODE)
	384	count = IBS_FETCH_CODE_SIZE; /IBS FETCH is 2 int64s/
	385	else
	386	count = IBS_OP_CODE_SIZE; /IBS OP is 5 int64s/
	387
	388	for (i = 0; i < count; i++) {
	389	increment_tail(cpu_buf);
	390	add_event_entry(IBS_EIP(cpu_buf->tail_pos));
	391	add_event_entry(IBS_EVENT(cpu_buf->tail_pos));
	392	}
	393	}
1da177e4	394
852402cc RR	395	#endif
852402cc RR	396
1da177e4 LT	397	static void add_sample_entry(unsigned long offset, unsigned long event)
	398	{
	399	add_event_entry(offset);
	400	add_event_entry(event);
	401	}
	402
	403
73185e0a	404	static int add_us_sample(struct mm_struct mm, struct op_sample s)
1da177e4 LT	405	{
	406	unsigned long cookie;
	407	off_t offset;
73185e0a RR	408
	409	cookie = lookup_dcookie(mm, s->eip, &offset);
	410
0c0a400d	411	if (cookie == INVALID_COOKIE) {
1da177e4 LT	412	atomic_inc(&oprofile_stats.sample_lost_no_mapping);
	413	return 0;
	414	}
	415
	416	if (cookie != last_cookie) {
	417	add_cookie_switch(cookie);
	418	last_cookie = cookie;
	419	}
	420
	421	add_sample_entry(offset, s->event);
	422
	423	return 1;
	424	}
	425
73185e0a	426
1da177e4 LT	427	/* Add a sample to the global event buffer. If possible the
	428	* sample is converted into a persistent dentry/offset pair
	429	* for later lookup from userspace.
	430	*/
	431	static int
73185e0a	432	add_sample(struct mm_struct mm, struct op_sample s, int in_kernel)
1da177e4 LT	433	{
	434	if (in_kernel) {
	435	add_sample_entry(s->eip, s->event);
	436	return 1;
	437	} else if (mm) {
	438	return add_us_sample(mm, s);
	439	} else {
	440	atomic_inc(&oprofile_stats.sample_lost_no_mm);
	441	}
	442	return 0;
	443	}
1da177e4	444
73185e0a RR	445
73185e0a RR	446	static void release_mm(struct mm_struct *mm)
1da177e4 LT	447	{
	448	if (!mm)
	449	return;
	450	up_read(&mm->mmap_sem);
	451	mmput(mm);
	452	}
	453
	454
73185e0a	455	static struct mm_struct take_tasks_mm(struct task_struct task)
1da177e4	456	{
73185e0a	457	struct mm_struct *mm = get_task_mm(task);
1da177e4 LT	458	if (mm)
	459	down_read(&mm->mmap_sem);
	460	return mm;
	461	}
	462
	463
	464	static inline int is_code(unsigned long val)
	465	{
	466	return val == ESCAPE_CODE;
	467	}
73185e0a	468
1da177e4 LT	469
1da177e4 LT	470	/* "acquire" as many cpu buffer slots as we can */
73185e0a	471	static unsigned long get_slots(struct oprofile_cpu_buffer *b)
1da177e4 LT	472	{
	473	unsigned long head = b->head_pos;
	474	unsigned long tail = b->tail_pos;
	475
	476	/*
	477	* Subtle. This resets the persistent last_task
	478	* and in_kernel values used for switching notes.
	479	* BUT, there is a small window between reading
	480	* head_pos, and this call, that means samples
	481	* can appear at the new head position, but not
	482	* be prefixed with the notes for switching
	483	* kernel mode or a task switch. This small hole
	484	* can lead to mis-attribution or samples where
	485	* we don't know if it's in the kernel or not,
	486	* at the start of an event buffer.
	487	*/
	488	cpu_buffer_reset(b);
	489
	490	if (head >= tail)
	491	return head - tail;
	492
	493	return head + (b->buffer_size - tail);
	494	}
	495
	496
1da177e4 LT	497	/* Move tasks along towards death. Any tasks on dead_tasks
	498	* will definitely have no remaining references in any
	499	* CPU buffers at this point, because we use two lists,
	500	* and to have reached the list, it must have gone through
	501	* one full sync already.
	502	*/
	503	static void process_task_mortuary(void)
	504	{
4369ef3c PM	505	unsigned long flags;
4369ef3c PM	506	LIST_HEAD(local_dead_tasks);
73185e0a RR	507	struct task_struct *task;
73185e0a RR	508	struct task_struct *ttask;
1da177e4	509
4369ef3c	510	spin_lock_irqsave(&task_mortuary, flags);
1da177e4	511
4369ef3c PM	512	list_splice_init(&dead_tasks, &local_dead_tasks);
4369ef3c PM	513	list_splice_init(&dying_tasks, &dead_tasks);
1da177e4	514
4369ef3c PM	515	spin_unlock_irqrestore(&task_mortuary, flags);
	516
	517	list_for_each_entry_safe(task, ttask, &local_dead_tasks, tasks) {
1da177e4	518	list_del(&task->tasks);
4369ef3c	519	free_task(task);
1da177e4	520	}
1da177e4 LT	521	}
	522
	523
	524	static void mark_done(int cpu)
	525	{
	526	int i;
	527
	528	cpu_set(cpu, marked_cpus);
	529
	530	for_each_online_cpu(i) {
	531	if (!cpu_isset(i, marked_cpus))
	532	return;
	533	}
	534
	535	/* All CPUs have been processed at least once,
	536	* we can process the mortuary once
	537	*/
	538	process_task_mortuary();
	539
	540	cpus_clear(marked_cpus);
	541	}
	542
	543
	544	/* FIXME: this is not sufficient if we implement syscall barrier backtrace
	545	* traversal, the code switch to sb_sample_start at first kernel enter/exit
	546	* switch so we need a fifth state and some special handling in sync_buffer()
	547	*/
	548	typedef enum {
	549	sb_bt_ignore = -2,
	550	sb_buffer_start,
	551	sb_bt_start,
	552	sb_sample_start,
	553	} sync_buffer_state;
	554
	555	/* Sync one of the CPU's buffers into the global event buffer.
	556	* Here we need to go through each batch of samples punctuated
	557	* by context switch notes, taking the task's mmap_sem and doing
	558	* lookup in task->mm->mmap to convert EIP into dcookie/offset
	559	* value.
	560	*/
	561	void sync_buffer(int cpu)
	562	{
608dfddd	563	struct oprofile_cpu_buffer *cpu_buf = &per_cpu(cpu_buffer, cpu);
1da177e4	564	struct mm_struct *mm = NULL;
73185e0a	565	struct task_struct *new;
1da177e4 LT	566	unsigned long cookie = 0;
	567	int in_kernel = 1;
	568	unsigned int i;
	569	sync_buffer_state state = sb_buffer_start;
	570	unsigned long available;
	571
59cc185a	572	mutex_lock(&buffer_mutex);
73185e0a	573
1da177e4 LT	574	add_cpu_switch(cpu);
	575
	576	/* Remember, only we can modify tail_pos */
	577
	578	available = get_slots(cpu_buf);
	579
	580	for (i = 0; i < available; ++i) {
73185e0a RR	581	struct op_sample *s = &cpu_buf->buffer[cpu_buf->tail_pos];
73185e0a RR	582
1da177e4 LT	583	if (is_code(s->eip)) {
	584	if (s->event <= CPU_IS_KERNEL) {
	585	/* kernel/userspace switch */
	586	in_kernel = s->event;
	587	if (state == sb_buffer_start)
	588	state = sb_sample_start;
	589	add_kernel_ctx_switch(s->event);
	590	} else if (s->event == CPU_TRACE_BEGIN) {
	591	state = sb_bt_start;
	592	add_trace_begin();
852402cc	593	#ifdef CONFIG_OPROFILE_IBS
345c2573 BK	594	} else if (s->event == IBS_FETCH_BEGIN) {
	595	state = sb_bt_start;
	596	add_ibs_begin(cpu_buf,
	597	IBS_FETCH_CODE, in_kernel, mm);
	598	} else if (s->event == IBS_OP_BEGIN) {
	599	state = sb_bt_start;
	600	add_ibs_begin(cpu_buf,
	601	IBS_OP_CODE, in_kernel, mm);
852402cc	602	#endif
1da177e4	603	} else {
73185e0a	604	struct mm_struct *oldmm = mm;
1da177e4 LT	605
	606	/* userspace context switch */
	607	new = (struct task_struct *)s->event;
	608
	609	release_mm(oldmm);
	610	mm = take_tasks_mm(new);
	611	if (mm != oldmm)
	612	cookie = get_exec_dcookie(mm);
	613	add_user_ctx_switch(new, cookie);
	614	}
73185e0a RR	615	} else if (state >= sb_bt_start &&
	616	!add_sample(mm, s, in_kernel)) {
	617	if (state == sb_bt_start) {
	618	state = sb_bt_ignore;
	619	atomic_inc(&oprofile_stats.bt_lost_no_mapping);
1da177e4 LT	620	}
	621	}
	622
	623	increment_tail(cpu_buf);
	624	}
	625	release_mm(mm);
	626
	627	mark_done(cpu);
	628
59cc185a	629	mutex_unlock(&buffer_mutex);
1da177e4	630	}