4 * @remark Copyright 2002-2009 OProfile authors
5 * @remark Read the file COPYING
7 * @author John Levon <levon@movementarian.org>
8 * @author Barry Kasindorf
9 * @author Robert Richter <robert.richter@amd.com>
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.
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.
20 * See fs/dcookies.c for a description of the dentry/offset
24 #include <linux/file.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>
32 #include <linux/oprofile.h>
33 #include <linux/sched.h>
34 #include <linux/sched/mm.h>
35 #include <linux/gfp.h>
37 #include "oprofile_stats.h"
38 #include "event_buffer.h"
39 #include "cpu_buffer.h"
40 #include "buffer_sync.h"
42 static LIST_HEAD(dying_tasks
);
43 static LIST_HEAD(dead_tasks
);
44 static cpumask_var_t marked_cpus
;
45 static DEFINE_SPINLOCK(task_mortuary
);
46 static void process_task_mortuary(void);
48 /* Take ownership of the task struct and place it on the
49 * list for processing. Only after two full buffer syncs
50 * does the task eventually get freed, because by then
51 * we are sure we will not reference it again.
52 * Can be invoked from softirq via RCU callback due to
53 * call_rcu() of the task struct, hence the _irqsave.
56 task_free_notify(struct notifier_block
*self
, unsigned long val
, void *data
)
59 struct task_struct
*task
= data
;
60 spin_lock_irqsave(&task_mortuary
, flags
);
61 list_add(&task
->tasks
, &dying_tasks
);
62 spin_unlock_irqrestore(&task_mortuary
, flags
);
67 /* The task is on its way out. A sync of the buffer means we can catch
68 * any remaining samples for this task.
71 task_exit_notify(struct notifier_block
*self
, unsigned long val
, void *data
)
73 /* To avoid latency problems, we only process the current CPU,
74 * hoping that most samples for the task are on this CPU
76 sync_buffer(raw_smp_processor_id());
81 /* The task is about to try a do_munmap(). We peek at what it's going to
82 * do, and if it's an executable region, process the samples first, so
83 * we don't lose any. This does not have to be exact, it's a QoI issue
87 munmap_notify(struct notifier_block
*self
, unsigned long val
, void *data
)
89 unsigned long addr
= (unsigned long)data
;
90 struct mm_struct
*mm
= current
->mm
;
91 struct vm_area_struct
*mpnt
;
93 down_read(&mm
->mmap_sem
);
95 mpnt
= find_vma(mm
, addr
);
96 if (mpnt
&& mpnt
->vm_file
&& (mpnt
->vm_flags
& VM_EXEC
)) {
97 up_read(&mm
->mmap_sem
);
98 /* To avoid latency problems, we only process the current CPU,
99 * hoping that most samples for the task are on this CPU
101 sync_buffer(raw_smp_processor_id());
105 up_read(&mm
->mmap_sem
);
110 /* We need to be told about new modules so we don't attribute to a previously
111 * loaded module, or drop the samples on the floor.
114 module_load_notify(struct notifier_block
*self
, unsigned long val
, void *data
)
116 #ifdef CONFIG_MODULES
117 if (val
!= MODULE_STATE_COMING
)
120 /* FIXME: should we process all CPU buffers ? */
121 mutex_lock(&buffer_mutex
);
122 add_event_entry(ESCAPE_CODE
);
123 add_event_entry(MODULE_LOADED_CODE
);
124 mutex_unlock(&buffer_mutex
);
130 static struct notifier_block task_free_nb
= {
131 .notifier_call
= task_free_notify
,
134 static struct notifier_block task_exit_nb
= {
135 .notifier_call
= task_exit_notify
,
138 static struct notifier_block munmap_nb
= {
139 .notifier_call
= munmap_notify
,
142 static struct notifier_block module_load_nb
= {
143 .notifier_call
= module_load_notify
,
146 static void free_all_tasks(void)
148 /* make sure we don't leak task structs */
149 process_task_mortuary();
150 process_task_mortuary();
157 if (!zalloc_cpumask_var(&marked_cpus
, GFP_KERNEL
))
160 err
= task_handoff_register(&task_free_nb
);
163 err
= profile_event_register(PROFILE_TASK_EXIT
, &task_exit_nb
);
166 err
= profile_event_register(PROFILE_MUNMAP
, &munmap_nb
);
169 err
= register_module_notifier(&module_load_nb
);
178 profile_event_unregister(PROFILE_MUNMAP
, &munmap_nb
);
180 profile_event_unregister(PROFILE_TASK_EXIT
, &task_exit_nb
);
182 task_handoff_unregister(&task_free_nb
);
185 free_cpumask_var(marked_cpus
);
193 unregister_module_notifier(&module_load_nb
);
194 profile_event_unregister(PROFILE_MUNMAP
, &munmap_nb
);
195 profile_event_unregister(PROFILE_TASK_EXIT
, &task_exit_nb
);
196 task_handoff_unregister(&task_free_nb
);
197 barrier(); /* do all of the above first */
202 free_cpumask_var(marked_cpus
);
206 /* Optimisation. We can manage without taking the dcookie sem
207 * because we cannot reach this code without at least one
208 * dcookie user still being registered (namely, the reader
209 * of the event buffer). */
210 static inline unsigned long fast_get_dcookie(const struct path
*path
)
212 unsigned long cookie
;
214 if (path
->dentry
->d_flags
& DCACHE_COOKIE
)
215 return (unsigned long)path
->dentry
;
216 get_dcookie(path
, &cookie
);
221 /* Look up the dcookie for the task's mm->exe_file,
222 * which corresponds loosely to "application name". This is
223 * not strictly necessary but allows oprofile to associate
224 * shared-library samples with particular applications
226 static unsigned long get_exec_dcookie(struct mm_struct
*mm
)
228 unsigned long cookie
= NO_COOKIE
;
229 struct file
*exe_file
;
234 exe_file
= get_mm_exe_file(mm
);
238 cookie
= fast_get_dcookie(&exe_file
->f_path
);
245 /* Convert the EIP value of a sample into a persistent dentry/offset
246 * pair that can then be added to the global event buffer. We make
247 * sure to do this lookup before a mm->mmap modification happens so
248 * we don't lose track.
250 * The caller must ensure the mm is not nil (ie: not a kernel thread).
253 lookup_dcookie(struct mm_struct
*mm
, unsigned long addr
, off_t
*offset
)
255 unsigned long cookie
= NO_COOKIE
;
256 struct vm_area_struct
*vma
;
258 down_read(&mm
->mmap_sem
);
259 for (vma
= find_vma(mm
, addr
); vma
; vma
= vma
->vm_next
) {
261 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
265 cookie
= fast_get_dcookie(&vma
->vm_file
->f_path
);
266 *offset
= (vma
->vm_pgoff
<< PAGE_SHIFT
) + addr
-
269 /* must be an anonymous map */
277 cookie
= INVALID_COOKIE
;
278 up_read(&mm
->mmap_sem
);
283 static unsigned long last_cookie
= INVALID_COOKIE
;
285 static void add_cpu_switch(int i
)
287 add_event_entry(ESCAPE_CODE
);
288 add_event_entry(CPU_SWITCH_CODE
);
290 last_cookie
= INVALID_COOKIE
;
293 static void add_kernel_ctx_switch(unsigned int in_kernel
)
295 add_event_entry(ESCAPE_CODE
);
297 add_event_entry(KERNEL_ENTER_SWITCH_CODE
);
299 add_event_entry(KERNEL_EXIT_SWITCH_CODE
);
303 add_user_ctx_switch(struct task_struct
const *task
, unsigned long cookie
)
305 add_event_entry(ESCAPE_CODE
);
306 add_event_entry(CTX_SWITCH_CODE
);
307 add_event_entry(task
->pid
);
308 add_event_entry(cookie
);
309 /* Another code for daemon back-compat */
310 add_event_entry(ESCAPE_CODE
);
311 add_event_entry(CTX_TGID_CODE
);
312 add_event_entry(task
->tgid
);
316 static void add_cookie_switch(unsigned long cookie
)
318 add_event_entry(ESCAPE_CODE
);
319 add_event_entry(COOKIE_SWITCH_CODE
);
320 add_event_entry(cookie
);
324 static void add_trace_begin(void)
326 add_event_entry(ESCAPE_CODE
);
327 add_event_entry(TRACE_BEGIN_CODE
);
330 static void add_data(struct op_entry
*entry
, struct mm_struct
*mm
)
332 unsigned long code
, pc
, val
;
333 unsigned long cookie
;
336 if (!op_cpu_buffer_get_data(entry
, &code
))
338 if (!op_cpu_buffer_get_data(entry
, &pc
))
340 if (!op_cpu_buffer_get_size(entry
))
344 cookie
= lookup_dcookie(mm
, pc
, &offset
);
346 if (cookie
== NO_COOKIE
)
348 if (cookie
== INVALID_COOKIE
) {
349 atomic_inc(&oprofile_stats
.sample_lost_no_mapping
);
352 if (cookie
!= last_cookie
) {
353 add_cookie_switch(cookie
);
354 last_cookie
= cookie
;
359 add_event_entry(ESCAPE_CODE
);
360 add_event_entry(code
);
361 add_event_entry(offset
); /* Offset from Dcookie */
363 while (op_cpu_buffer_get_data(entry
, &val
))
364 add_event_entry(val
);
367 static inline void add_sample_entry(unsigned long offset
, unsigned long event
)
369 add_event_entry(offset
);
370 add_event_entry(event
);
375 * Add a sample to the global event buffer. If possible the
376 * sample is converted into a persistent dentry/offset pair
377 * for later lookup from userspace. Return 0 on failure.
380 add_sample(struct mm_struct
*mm
, struct op_sample
*s
, int in_kernel
)
382 unsigned long cookie
;
386 add_sample_entry(s
->eip
, s
->event
);
390 /* add userspace sample */
393 atomic_inc(&oprofile_stats
.sample_lost_no_mm
);
397 cookie
= lookup_dcookie(mm
, s
->eip
, &offset
);
399 if (cookie
== INVALID_COOKIE
) {
400 atomic_inc(&oprofile_stats
.sample_lost_no_mapping
);
404 if (cookie
!= last_cookie
) {
405 add_cookie_switch(cookie
);
406 last_cookie
= cookie
;
409 add_sample_entry(offset
, s
->event
);
415 static void release_mm(struct mm_struct
*mm
)
422 static inline int is_code(unsigned long val
)
424 return val
== ESCAPE_CODE
;
428 /* Move tasks along towards death. Any tasks on dead_tasks
429 * will definitely have no remaining references in any
430 * CPU buffers at this point, because we use two lists,
431 * and to have reached the list, it must have gone through
432 * one full sync already.
434 static void process_task_mortuary(void)
437 LIST_HEAD(local_dead_tasks
);
438 struct task_struct
*task
;
439 struct task_struct
*ttask
;
441 spin_lock_irqsave(&task_mortuary
, flags
);
443 list_splice_init(&dead_tasks
, &local_dead_tasks
);
444 list_splice_init(&dying_tasks
, &dead_tasks
);
446 spin_unlock_irqrestore(&task_mortuary
, flags
);
448 list_for_each_entry_safe(task
, ttask
, &local_dead_tasks
, tasks
) {
449 list_del(&task
->tasks
);
455 static void mark_done(int cpu
)
459 cpumask_set_cpu(cpu
, marked_cpus
);
461 for_each_online_cpu(i
) {
462 if (!cpumask_test_cpu(i
, marked_cpus
))
466 /* All CPUs have been processed at least once,
467 * we can process the mortuary once
469 process_task_mortuary();
471 cpumask_clear(marked_cpus
);
475 /* FIXME: this is not sufficient if we implement syscall barrier backtrace
476 * traversal, the code switch to sb_sample_start at first kernel enter/exit
477 * switch so we need a fifth state and some special handling in sync_buffer()
486 /* Sync one of the CPU's buffers into the global event buffer.
487 * Here we need to go through each batch of samples punctuated
488 * by context switch notes, taking the task's mmap_sem and doing
489 * lookup in task->mm->mmap to convert EIP into dcookie/offset
492 void sync_buffer(int cpu
)
494 struct mm_struct
*mm
= NULL
;
495 struct mm_struct
*oldmm
;
497 struct task_struct
*new;
498 unsigned long cookie
= 0;
500 sync_buffer_state state
= sb_buffer_start
;
502 unsigned long available
;
504 struct op_entry entry
;
505 struct op_sample
*sample
;
507 mutex_lock(&buffer_mutex
);
511 op_cpu_buffer_reset(cpu
);
512 available
= op_cpu_buffer_entries(cpu
);
514 for (i
= 0; i
< available
; ++i
) {
515 sample
= op_cpu_buffer_read_entry(&entry
, cpu
);
519 if (is_code(sample
->eip
)) {
520 flags
= sample
->event
;
521 if (flags
& TRACE_BEGIN
) {
525 if (flags
& KERNEL_CTX_SWITCH
) {
526 /* kernel/userspace switch */
527 in_kernel
= flags
& IS_KERNEL
;
528 if (state
== sb_buffer_start
)
529 state
= sb_sample_start
;
530 add_kernel_ctx_switch(flags
& IS_KERNEL
);
532 if (flags
& USER_CTX_SWITCH
533 && op_cpu_buffer_get_data(&entry
, &val
)) {
534 /* userspace context switch */
535 new = (struct task_struct
*)val
;
538 mm
= get_task_mm(new);
540 cookie
= get_exec_dcookie(mm
);
541 add_user_ctx_switch(new, cookie
);
543 if (op_cpu_buffer_get_size(&entry
))
544 add_data(&entry
, mm
);
548 if (state
< sb_bt_start
)
552 if (add_sample(mm
, sample
, in_kernel
))
555 /* ignore backtraces if failed to add a sample */
556 if (state
== sb_bt_start
) {
557 state
= sb_bt_ignore
;
558 atomic_inc(&oprofile_stats
.bt_lost_no_mapping
);
565 mutex_unlock(&buffer_mutex
);
568 /* The function can be used to add a buffer worth of data directly to
569 * the kernel buffer. The buffer is assumed to be a circular buffer.
570 * Take the entries from index start and end at index end, wrapping
573 void oprofile_put_buff(unsigned long *buf
, unsigned int start
,
574 unsigned int stop
, unsigned int max
)
580 mutex_lock(&buffer_mutex
);
582 add_event_entry(buf
[i
++]);
588 mutex_unlock(&buffer_mutex
);