1/**
  2 * @file buffer_sync.c
  3 *
  4 * @remark Copyright 2002-2009 OProfile authors
  5 * @remark Read the file COPYING
  6 *
  7 * @author John Levon <levon@movementarian.org>
  8 * @author Barry Kasindorf
  9 * @author Robert Richter <robert.richter@amd.com>
 10 *
 11 * This is the core of the buffer management. Each
 12 * CPU buffer is processed and entered into the
 13 * global event buffer. Such processing is necessary
 14 * in several circumstances, mentioned below.
 15 *
 16 * The processing does the job of converting the
 17 * transitory EIP value into a persistent dentry/offset
 18 * value that the profiler can record at its leisure.
 19 *
 20 * See fs/dcookies.c for a description of the dentry/offset
 21 * objects.
 22 */
 23
 24#include <linux/mm.h>
 25#include <linux/workqueue.h>
 26#include <linux/notifier.h>
 27#include <linux/dcookies.h>
 28#include <linux/profile.h>
 29#include <linux/module.h>
 30#include <linux/fs.h>
 31#include <linux/oprofile.h>
 32#include <linux/sched.h>
 33#include <linux/gfp.h>
 34
 35#include "oprofile_stats.h"
 36#include "event_buffer.h"
 37#include "cpu_buffer.h"
 38#include "buffer_sync.h"
 39
 40static LIST_HEAD(dying_tasks);
 41static LIST_HEAD(dead_tasks);
 42static cpumask_var_t marked_cpus;
 43static DEFINE_SPINLOCK(task_mortuary);
 44static void process_task_mortuary(void);
 45
 46/* Take ownership of the task struct and place it on the
 47 * list for processing. Only after two full buffer syncs
 48 * does the task eventually get freed, because by then
 49 * we are sure we will not reference it again.
 50 * Can be invoked from softirq via RCU callback due to
 51 * call_rcu() of the task struct, hence the _irqsave.
 52 */
 53static int
 54task_free_notify(struct notifier_block *self, unsigned long val, void *data)
 55{
 56	unsigned long flags;
 57	struct task_struct *task = data;
 58	spin_lock_irqsave(&task_mortuary, flags);
 59	list_add(&task->tasks, &dying_tasks);
 60	spin_unlock_irqrestore(&task_mortuary, flags);
 61	return NOTIFY_OK;
 62}
 63
 64
 65/* The task is on its way out. A sync of the buffer means we can catch
 66 * any remaining samples for this task.
 67 */
 68static int
 69task_exit_notify(struct notifier_block *self, unsigned long val, void *data)
 70{
 71	/* To avoid latency problems, we only process the current CPU,
 72	 * hoping that most samples for the task are on this CPU
 73	 */
 74	sync_buffer(raw_smp_processor_id());
 75	return 0;
 76}
 77
 78
 79/* The task is about to try a do_munmap(). We peek at what it's going to
 80 * do, and if it's an executable region, process the samples first, so
 81 * we don't lose any. This does not have to be exact, it's a QoI issue
 82 * only.
 83 */
 84static int
 85munmap_notify(struct notifier_block *self, unsigned long val, void *data)
 86{
 87	unsigned long addr = (unsigned long)data;
 88	struct mm_struct *mm = current->mm;
 89	struct vm_area_struct *mpnt;
 90
 91	down_read(&mm->mmap_sem);
 92
 93	mpnt = find_vma(mm, addr);
 94	if (mpnt && mpnt->vm_file && (mpnt->vm_flags & VM_EXEC)) {
 95		up_read(&mm->mmap_sem);
 96		/* To avoid latency problems, we only process the current CPU,
 97		 * hoping that most samples for the task are on this CPU
 98		 */
 99		sync_buffer(raw_smp_processor_id());
100		return 0;
101	}
102
103	up_read(&mm->mmap_sem);
104	return 0;
105}
106
107
108/* We need to be told about new modules so we don't attribute to a previously
109 * loaded module, or drop the samples on the floor.
110 */
111static int
112module_load_notify(struct notifier_block *self, unsigned long val, void *data)
113{
114#ifdef CONFIG_MODULES
115	if (val != MODULE_STATE_COMING)
116		return 0;
117
118	/* FIXME: should we process all CPU buffers ? */
119	mutex_lock(&buffer_mutex);
120	add_event_entry(ESCAPE_CODE);
121	add_event_entry(MODULE_LOADED_CODE);
122	mutex_unlock(&buffer_mutex);
123#endif
124	return 0;
125}
126
127
128static struct notifier_block task_free_nb = {
129	.notifier_call	= task_free_notify,
130};
131
132static struct notifier_block task_exit_nb = {
133	.notifier_call	= task_exit_notify,
134};
135
136static struct notifier_block munmap_nb = {
137	.notifier_call	= munmap_notify,
138};
139
140static struct notifier_block module_load_nb = {
141	.notifier_call = module_load_notify,
142};
143
144static void free_all_tasks(void)
145{
146	/* make sure we don't leak task structs */
147	process_task_mortuary();
148	process_task_mortuary();
149}
150
151int sync_start(void)
152{
153	int err;
154
155	if (!zalloc_cpumask_var(&marked_cpus, GFP_KERNEL))
156		return -ENOMEM;
157
158	err = task_handoff_register(&task_free_nb);
159	if (err)
160		goto out1;
161	err = profile_event_register(PROFILE_TASK_EXIT, &task_exit_nb);
162	if (err)
163		goto out2;
164	err = profile_event_register(PROFILE_MUNMAP, &munmap_nb);
165	if (err)
166		goto out3;
167	err = register_module_notifier(&module_load_nb);
168	if (err)
169		goto out4;
170
171	start_cpu_work();
172
173out:
174	return err;
175out4:
176	profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
177out3:
178	profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
179out2:
180	task_handoff_unregister(&task_free_nb);
181	free_all_tasks();
182out1:
183	free_cpumask_var(marked_cpus);
184	goto out;
185}
186
187
188void sync_stop(void)
189{
190	end_cpu_work();
191	unregister_module_notifier(&module_load_nb);
192	profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
193	profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
194	task_handoff_unregister(&task_free_nb);
195	barrier();			/* do all of the above first */
196
197	flush_cpu_work();
198
199	free_all_tasks();
200	free_cpumask_var(marked_cpus);
201}
202
203
204/* Optimisation. We can manage without taking the dcookie sem
205 * because we cannot reach this code without at least one
206 * dcookie user still being registered (namely, the reader
207 * of the event buffer). */
208static inline unsigned long fast_get_dcookie(struct path *path)
209{
210	unsigned long cookie;
211
212	if (path->dentry->d_flags & DCACHE_COOKIE)
213		return (unsigned long)path->dentry;
214	get_dcookie(path, &cookie);
215	return cookie;
216}
217
218
219/* Look up the dcookie for the task's first VM_EXECUTABLE mapping,
220 * which corresponds loosely to "application name". This is
221 * not strictly necessary but allows oprofile to associate
222 * shared-library samples with particular applications
223 */
224static unsigned long get_exec_dcookie(struct mm_struct *mm)
225{
226	unsigned long cookie = NO_COOKIE;
227	struct vm_area_struct *vma;
228
229	if (!mm)
230		goto out;
231
232	for (vma = mm->mmap; vma; vma = vma->vm_next) {
233		if (!vma->vm_file)
234			continue;
235		if (!(vma->vm_flags & VM_EXECUTABLE))
236			continue;
237		cookie = fast_get_dcookie(&vma->vm_file->f_path);
238		break;
239	}
240
241out:
242	return cookie;
243}
244
245
246/* Convert the EIP value of a sample into a persistent dentry/offset
247 * pair that can then be added to the global event buffer. We make
248 * sure to do this lookup before a mm->mmap modification happens so
249 * we don't lose track.
250 */
251static unsigned long
252lookup_dcookie(struct mm_struct *mm, unsigned long addr, off_t *offset)
253{
254	unsigned long cookie = NO_COOKIE;
255	struct vm_area_struct *vma;
256
257	for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
258
259		if (addr < vma->vm_start || addr >= vma->vm_end)
260			continue;
261
262		if (vma->vm_file) {
263			cookie = fast_get_dcookie(&vma->vm_file->f_path);
264			*offset = (vma->vm_pgoff << PAGE_SHIFT) + addr -
265				vma->vm_start;
266		} else {
267			/* must be an anonymous map */
268			*offset = addr;
269		}
270
271		break;
272	}
273
274	if (!vma)
275		cookie = INVALID_COOKIE;
276
277	return cookie;
278}
279
280static unsigned long last_cookie = INVALID_COOKIE;
281
282static void add_cpu_switch(int i)
283{
284	add_event_entry(ESCAPE_CODE);
285	add_event_entry(CPU_SWITCH_CODE);
286	add_event_entry(i);
287	last_cookie = INVALID_COOKIE;
288}
289
290static void add_kernel_ctx_switch(unsigned int in_kernel)
291{
292	add_event_entry(ESCAPE_CODE);
293	if (in_kernel)
294		add_event_entry(KERNEL_ENTER_SWITCH_CODE);
295	else
296		add_event_entry(KERNEL_EXIT_SWITCH_CODE);
297}
298
299static void
300add_user_ctx_switch(struct task_struct const *task, unsigned long cookie)
301{
302	add_event_entry(ESCAPE_CODE);
303	add_event_entry(CTX_SWITCH_CODE);
304	add_event_entry(task->pid);
305	add_event_entry(cookie);
306	/* Another code for daemon back-compat */
307	add_event_entry(ESCAPE_CODE);
308	add_event_entry(CTX_TGID_CODE);
309	add_event_entry(task->tgid);
310}
311
312
313static void add_cookie_switch(unsigned long cookie)
314{
315	add_event_entry(ESCAPE_CODE);
316	add_event_entry(COOKIE_SWITCH_CODE);
317	add_event_entry(cookie);
318}
319
320
321static void add_trace_begin(void)
322{
323	add_event_entry(ESCAPE_CODE);
324	add_event_entry(TRACE_BEGIN_CODE);
325}
326
327static void add_data(struct op_entry *entry, struct mm_struct *mm)
328{
329	unsigned long code, pc, val;
330	unsigned long cookie;
331	off_t offset;
332
333	if (!op_cpu_buffer_get_data(entry, &code))
334		return;
335	if (!op_cpu_buffer_get_data(entry, &pc))
336		return;
337	if (!op_cpu_buffer_get_size(entry))
338		return;
339
340	if (mm) {
341		cookie = lookup_dcookie(mm, pc, &offset);
342
343		if (cookie == NO_COOKIE)
344			offset = pc;
345		if (cookie == INVALID_COOKIE) {
346			atomic_inc(&oprofile_stats.sample_lost_no_mapping);
347			offset = pc;
348		}
349		if (cookie != last_cookie) {
350			add_cookie_switch(cookie);
351			last_cookie = cookie;
352		}
353	} else
354		offset = pc;
355
356	add_event_entry(ESCAPE_CODE);
357	add_event_entry(code);
358	add_event_entry(offset);	/* Offset from Dcookie */
359
360	while (op_cpu_buffer_get_data(entry, &val))
361		add_event_entry(val);
362}
363
364static inline void add_sample_entry(unsigned long offset, unsigned long event)
365{
366	add_event_entry(offset);
367	add_event_entry(event);
368}
369
370
371/*
372 * Add a sample to the global event buffer. If possible the
373 * sample is converted into a persistent dentry/offset pair
374 * for later lookup from userspace. Return 0 on failure.
375 */
376static int
377add_sample(struct mm_struct *mm, struct op_sample *s, int in_kernel)
378{
379	unsigned long cookie;
380	off_t offset;
381
382	if (in_kernel) {
383		add_sample_entry(s->eip, s->event);
384		return 1;
385	}
386
387	/* add userspace sample */
388
389	if (!mm) {
390		atomic_inc(&oprofile_stats.sample_lost_no_mm);
391		return 0;
392	}
393
394	cookie = lookup_dcookie(mm, s->eip, &offset);
395
396	if (cookie == INVALID_COOKIE) {
397		atomic_inc(&oprofile_stats.sample_lost_no_mapping);
398		return 0;
399	}
400
401	if (cookie != last_cookie) {
402		add_cookie_switch(cookie);
403		last_cookie = cookie;
404	}
405
406	add_sample_entry(offset, s->event);
407
408	return 1;
409}
410
411
412static void release_mm(struct mm_struct *mm)
413{
414	if (!mm)
415		return;
416	up_read(&mm->mmap_sem);
417	mmput(mm);
418}
419
420
421static struct mm_struct *take_tasks_mm(struct task_struct *task)
422{
423	struct mm_struct *mm = get_task_mm(task);
424	if (mm)
425		down_read(&mm->mmap_sem);
426	return mm;
427}
428
429
430static inline int is_code(unsigned long val)
431{
432	return val == ESCAPE_CODE;
433}
434
435
436/* Move tasks along towards death. Any tasks on dead_tasks
437 * will definitely have no remaining references in any
438 * CPU buffers at this point, because we use two lists,
439 * and to have reached the list, it must have gone through
440 * one full sync already.
441 */
442static void process_task_mortuary(void)
443{
444	unsigned long flags;
445	LIST_HEAD(local_dead_tasks);
446	struct task_struct *task;
447	struct task_struct *ttask;
448
449	spin_lock_irqsave(&task_mortuary, flags);
450
451	list_splice_init(&dead_tasks, &local_dead_tasks);
452	list_splice_init(&dying_tasks, &dead_tasks);
453
454	spin_unlock_irqrestore(&task_mortuary, flags);
455
456	list_for_each_entry_safe(task, ttask, &local_dead_tasks, tasks) {
457		list_del(&task->tasks);
458		free_task(task);
459	}
460}
461
462
463static void mark_done(int cpu)
464{
465	int i;
466
467	cpumask_set_cpu(cpu, marked_cpus);
468
469	for_each_online_cpu(i) {
470		if (!cpumask_test_cpu(i, marked_cpus))
471			return;
472	}
473
474	/* All CPUs have been processed at least once,
475	 * we can process the mortuary once
476	 */
477	process_task_mortuary();
478
479	cpumask_clear(marked_cpus);
480}
481
482
483/* FIXME: this is not sufficient if we implement syscall barrier backtrace
484 * traversal, the code switch to sb_sample_start at first kernel enter/exit
485 * switch so we need a fifth state and some special handling in sync_buffer()
486 */
487typedef enum {
488	sb_bt_ignore = -2,
489	sb_buffer_start,
490	sb_bt_start,
491	sb_sample_start,
492} sync_buffer_state;
493
494/* Sync one of the CPU's buffers into the global event buffer.
495 * Here we need to go through each batch of samples punctuated
496 * by context switch notes, taking the task's mmap_sem and doing
497 * lookup in task->mm->mmap to convert EIP into dcookie/offset
498 * value.
499 */
500void sync_buffer(int cpu)
501{
502	struct mm_struct *mm = NULL;
503	struct mm_struct *oldmm;
504	unsigned long val;
505	struct task_struct *new;
506	unsigned long cookie = 0;
507	int in_kernel = 1;
508	sync_buffer_state state = sb_buffer_start;
509	unsigned int i;
510	unsigned long available;
511	unsigned long flags;
512	struct op_entry entry;
513	struct op_sample *sample;
514
515	mutex_lock(&buffer_mutex);
516
517	add_cpu_switch(cpu);
518
519	op_cpu_buffer_reset(cpu);
520	available = op_cpu_buffer_entries(cpu);
521
522	for (i = 0; i < available; ++i) {
523		sample = op_cpu_buffer_read_entry(&entry, cpu);
524		if (!sample)
525			break;
526
527		if (is_code(sample->eip)) {
528			flags = sample->event;
529			if (flags & TRACE_BEGIN) {
530				state = sb_bt_start;
531				add_trace_begin();
532			}
533			if (flags & KERNEL_CTX_SWITCH) {
534				/* kernel/userspace switch */
535				in_kernel = flags & IS_KERNEL;
536				if (state == sb_buffer_start)
537					state = sb_sample_start;
538				add_kernel_ctx_switch(flags & IS_KERNEL);
539			}
540			if (flags & USER_CTX_SWITCH
541			    && op_cpu_buffer_get_data(&entry, &val)) {
542				/* userspace context switch */
543				new = (struct task_struct *)val;
544				oldmm = mm;
545				release_mm(oldmm);
546				mm = take_tasks_mm(new);
547				if (mm != oldmm)
548					cookie = get_exec_dcookie(mm);
549				add_user_ctx_switch(new, cookie);
550			}
551			if (op_cpu_buffer_get_size(&entry))
552				add_data(&entry, mm);
553			continue;
554		}
555
556		if (state < sb_bt_start)
557			/* ignore sample */
558			continue;
559
560		if (add_sample(mm, sample, in_kernel))
561			continue;
562
563		/* ignore backtraces if failed to add a sample */
564		if (state == sb_bt_start) {
565			state = sb_bt_ignore;
566			atomic_inc(&oprofile_stats.bt_lost_no_mapping);
567		}
568	}
569	release_mm(mm);
570
571	mark_done(cpu);
572
573	mutex_unlock(&buffer_mutex);
574}
575
576/* The function can be used to add a buffer worth of data directly to
577 * the kernel buffer. The buffer is assumed to be a circular buffer.
578 * Take the entries from index start and end at index end, wrapping
579 * at max_entries.
580 */
581void oprofile_put_buff(unsigned long *buf, unsigned int start,
582		       unsigned int stop, unsigned int max)
583{
584	int i;
585
586	i = start;
587
588	mutex_lock(&buffer_mutex);
589	while (i != stop) {
590		add_event_entry(buf[i++]);
591
592		if (i >= max)
593			i = 0;
594	}
595
596	mutex_unlock(&buffer_mutex);
597}
598