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