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.
8 * Scheduler profiling support, Arjan van de Ven and Ingo Molnar,
10 * Consolidation of architecture support code for profiling,
11 * Nadia Yvette Chambers, Oracle, July 2004
12 * Amortized hit count accounting via per-cpu open-addressed hashtables
13 * to resolve timer interrupt livelocks, Nadia Yvette Chambers,
17 #include <linux/export.h>
18 #include <linux/profile.h>
19 #include <linux/bootmem.h>
20 #include <linux/notifier.h>
22 #include <linux/cpumask.h>
23 #include <linux/cpu.h>
24 #include <linux/highmem.h>
25 #include <linux/mutex.h>
26 #include <linux/slab.h>
27 #include <linux/vmalloc.h>
28 #include <asm/sections.h>
29 #include <asm/irq_regs.h>
30 #include <asm/ptrace.h>
35 #define PROFILE_GRPSHIFT 3
36 #define PROFILE_GRPSZ (1 << PROFILE_GRPSHIFT)
37 #define NR_PROFILE_HIT (PAGE_SIZE/sizeof(struct profile_hit))
38 #define NR_PROFILE_GRP (NR_PROFILE_HIT/PROFILE_GRPSZ)
40 /* Oprofile timer tick hook */
41 static int (*timer_hook
)(struct pt_regs
*) __read_mostly
;
43 static atomic_t
*prof_buffer
;
44 static unsigned long prof_len
, prof_shift
;
46 int prof_on __read_mostly
;
47 EXPORT_SYMBOL_GPL(prof_on
);
49 static cpumask_var_t prof_cpu_mask
;
51 static DEFINE_PER_CPU(struct profile_hit
*[2], cpu_profile_hits
);
52 static DEFINE_PER_CPU(int, cpu_profile_flip
);
53 static DEFINE_MUTEX(profile_flip_mutex
);
54 #endif /* CONFIG_SMP */
56 int profile_setup(char *str
)
58 static char schedstr
[] = "schedule";
59 static char sleepstr
[] = "sleep";
60 static char kvmstr
[] = "kvm";
63 if (!strncmp(str
, sleepstr
, strlen(sleepstr
))) {
64 #ifdef CONFIG_SCHEDSTATS
65 prof_on
= SLEEP_PROFILING
;
66 if (str
[strlen(sleepstr
)] == ',')
67 str
+= strlen(sleepstr
) + 1;
68 if (get_option(&str
, &par
))
71 "kernel sleep profiling enabled (shift: %ld)\n",
75 "kernel sleep profiling requires CONFIG_SCHEDSTATS\n");
76 #endif /* CONFIG_SCHEDSTATS */
77 } else if (!strncmp(str
, schedstr
, strlen(schedstr
))) {
78 prof_on
= SCHED_PROFILING
;
79 if (str
[strlen(schedstr
)] == ',')
80 str
+= strlen(schedstr
) + 1;
81 if (get_option(&str
, &par
))
84 "kernel schedule profiling enabled (shift: %ld)\n",
86 } else if (!strncmp(str
, kvmstr
, strlen(kvmstr
))) {
87 prof_on
= KVM_PROFILING
;
88 if (str
[strlen(kvmstr
)] == ',')
89 str
+= strlen(kvmstr
) + 1;
90 if (get_option(&str
, &par
))
93 "kernel KVM profiling enabled (shift: %ld)\n",
95 } else if (get_option(&str
, &par
)) {
97 prof_on
= CPU_PROFILING
;
98 printk(KERN_INFO
"kernel profiling enabled (shift: %ld)\n",
103 __setup("profile=", profile_setup
);
106 int __ref
profile_init(void)
112 /* only text is profiled */
113 prof_len
= (_etext
- _stext
) >> prof_shift
;
114 buffer_bytes
= prof_len
*sizeof(atomic_t
);
116 if (!alloc_cpumask_var(&prof_cpu_mask
, GFP_KERNEL
))
119 cpumask_copy(prof_cpu_mask
, cpu_possible_mask
);
121 prof_buffer
= kzalloc(buffer_bytes
, GFP_KERNEL
|__GFP_NOWARN
);
125 prof_buffer
= alloc_pages_exact(buffer_bytes
,
126 GFP_KERNEL
|__GFP_ZERO
|__GFP_NOWARN
);
130 prof_buffer
= vzalloc(buffer_bytes
);
134 free_cpumask_var(prof_cpu_mask
);
138 /* Profile event notifications */
140 static BLOCKING_NOTIFIER_HEAD(task_exit_notifier
);
141 static ATOMIC_NOTIFIER_HEAD(task_free_notifier
);
142 static BLOCKING_NOTIFIER_HEAD(munmap_notifier
);
144 void profile_task_exit(struct task_struct
*task
)
146 blocking_notifier_call_chain(&task_exit_notifier
, 0, task
);
149 int profile_handoff_task(struct task_struct
*task
)
152 ret
= atomic_notifier_call_chain(&task_free_notifier
, 0, task
);
153 return (ret
== NOTIFY_OK
) ? 1 : 0;
156 void profile_munmap(unsigned long addr
)
158 blocking_notifier_call_chain(&munmap_notifier
, 0, (void *)addr
);
161 int task_handoff_register(struct notifier_block
*n
)
163 return atomic_notifier_chain_register(&task_free_notifier
, n
);
165 EXPORT_SYMBOL_GPL(task_handoff_register
);
167 int task_handoff_unregister(struct notifier_block
*n
)
169 return atomic_notifier_chain_unregister(&task_free_notifier
, n
);
171 EXPORT_SYMBOL_GPL(task_handoff_unregister
);
173 int profile_event_register(enum profile_type type
, struct notifier_block
*n
)
178 case PROFILE_TASK_EXIT
:
179 err
= blocking_notifier_chain_register(
180 &task_exit_notifier
, n
);
183 err
= blocking_notifier_chain_register(
184 &munmap_notifier
, n
);
190 EXPORT_SYMBOL_GPL(profile_event_register
);
192 int profile_event_unregister(enum profile_type type
, struct notifier_block
*n
)
197 case PROFILE_TASK_EXIT
:
198 err
= blocking_notifier_chain_unregister(
199 &task_exit_notifier
, n
);
202 err
= blocking_notifier_chain_unregister(
203 &munmap_notifier
, n
);
209 EXPORT_SYMBOL_GPL(profile_event_unregister
);
211 int register_timer_hook(int (*hook
)(struct pt_regs
*))
218 EXPORT_SYMBOL_GPL(register_timer_hook
);
220 void unregister_timer_hook(int (*hook
)(struct pt_regs
*))
222 WARN_ON(hook
!= timer_hook
);
224 /* make sure all CPUs see the NULL hook */
225 synchronize_sched(); /* Allow ongoing interrupts to complete. */
227 EXPORT_SYMBOL_GPL(unregister_timer_hook
);
232 * Each cpu has a pair of open-addressed hashtables for pending
233 * profile hits. read_profile() IPI's all cpus to request them
234 * to flip buffers and flushes their contents to prof_buffer itself.
235 * Flip requests are serialized by the profile_flip_mutex. The sole
236 * use of having a second hashtable is for avoiding cacheline
237 * contention that would otherwise happen during flushes of pending
238 * profile hits required for the accuracy of reported profile hits
239 * and so resurrect the interrupt livelock issue.
241 * The open-addressed hashtables are indexed by profile buffer slot
242 * and hold the number of pending hits to that profile buffer slot on
243 * a cpu in an entry. When the hashtable overflows, all pending hits
244 * are accounted to their corresponding profile buffer slots with
245 * atomic_add() and the hashtable emptied. As numerous pending hits
246 * may be accounted to a profile buffer slot in a hashtable entry,
247 * this amortizes a number of atomic profile buffer increments likely
248 * to be far larger than the number of entries in the hashtable,
249 * particularly given that the number of distinct profile buffer
250 * positions to which hits are accounted during short intervals (e.g.
251 * several seconds) is usually very small. Exclusion from buffer
252 * flipping is provided by interrupt disablement (note that for
253 * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
255 * The hash function is meant to be lightweight as opposed to strong,
256 * and was vaguely inspired by ppc64 firmware-supported inverted
257 * pagetable hash functions, but uses a full hashtable full of finite
258 * collision chains, not just pairs of them.
262 static void __profile_flip_buffers(void *unused
)
264 int cpu
= smp_processor_id();
266 per_cpu(cpu_profile_flip
, cpu
) = !per_cpu(cpu_profile_flip
, cpu
);
269 static void profile_flip_buffers(void)
273 mutex_lock(&profile_flip_mutex
);
274 j
= per_cpu(cpu_profile_flip
, get_cpu());
276 on_each_cpu(__profile_flip_buffers
, NULL
, 1);
277 for_each_online_cpu(cpu
) {
278 struct profile_hit
*hits
= per_cpu(cpu_profile_hits
, cpu
)[j
];
279 for (i
= 0; i
< NR_PROFILE_HIT
; ++i
) {
285 atomic_add(hits
[i
].hits
, &prof_buffer
[hits
[i
].pc
]);
286 hits
[i
].hits
= hits
[i
].pc
= 0;
289 mutex_unlock(&profile_flip_mutex
);
292 static void profile_discard_flip_buffers(void)
296 mutex_lock(&profile_flip_mutex
);
297 i
= per_cpu(cpu_profile_flip
, get_cpu());
299 on_each_cpu(__profile_flip_buffers
, NULL
, 1);
300 for_each_online_cpu(cpu
) {
301 struct profile_hit
*hits
= per_cpu(cpu_profile_hits
, cpu
)[i
];
302 memset(hits
, 0, NR_PROFILE_HIT
*sizeof(struct profile_hit
));
304 mutex_unlock(&profile_flip_mutex
);
307 static void do_profile_hits(int type
, void *__pc
, unsigned int nr_hits
)
309 unsigned long primary
, secondary
, flags
, pc
= (unsigned long)__pc
;
311 struct profile_hit
*hits
;
313 pc
= min((pc
- (unsigned long)_stext
) >> prof_shift
, prof_len
- 1);
314 i
= primary
= (pc
& (NR_PROFILE_GRP
- 1)) << PROFILE_GRPSHIFT
;
315 secondary
= (~(pc
<< 1) & (NR_PROFILE_GRP
- 1)) << PROFILE_GRPSHIFT
;
317 hits
= per_cpu(cpu_profile_hits
, cpu
)[per_cpu(cpu_profile_flip
, cpu
)];
323 * We buffer the global profiler buffer into a per-CPU
324 * queue and thus reduce the number of global (and possibly
325 * NUMA-alien) accesses. The write-queue is self-coalescing:
327 local_irq_save(flags
);
329 for (j
= 0; j
< PROFILE_GRPSZ
; ++j
) {
330 if (hits
[i
+ j
].pc
== pc
) {
331 hits
[i
+ j
].hits
+= nr_hits
;
333 } else if (!hits
[i
+ j
].hits
) {
335 hits
[i
+ j
].hits
= nr_hits
;
339 i
= (i
+ secondary
) & (NR_PROFILE_HIT
- 1);
340 } while (i
!= primary
);
343 * Add the current hit(s) and flush the write-queue out
344 * to the global buffer:
346 atomic_add(nr_hits
, &prof_buffer
[pc
]);
347 for (i
= 0; i
< NR_PROFILE_HIT
; ++i
) {
348 atomic_add(hits
[i
].hits
, &prof_buffer
[hits
[i
].pc
]);
349 hits
[i
].pc
= hits
[i
].hits
= 0;
352 local_irq_restore(flags
);
356 static int __cpuinit
profile_cpu_callback(struct notifier_block
*info
,
357 unsigned long action
, void *__cpu
)
359 int node
, cpu
= (unsigned long)__cpu
;
364 case CPU_UP_PREPARE_FROZEN
:
365 node
= cpu_to_mem(cpu
);
366 per_cpu(cpu_profile_flip
, cpu
) = 0;
367 if (!per_cpu(cpu_profile_hits
, cpu
)[1]) {
368 page
= alloc_pages_exact_node(node
,
369 GFP_KERNEL
| __GFP_ZERO
,
372 return notifier_from_errno(-ENOMEM
);
373 per_cpu(cpu_profile_hits
, cpu
)[1] = page_address(page
);
375 if (!per_cpu(cpu_profile_hits
, cpu
)[0]) {
376 page
= alloc_pages_exact_node(node
,
377 GFP_KERNEL
| __GFP_ZERO
,
381 per_cpu(cpu_profile_hits
, cpu
)[0] = page_address(page
);
385 page
= virt_to_page(per_cpu(cpu_profile_hits
, cpu
)[1]);
386 per_cpu(cpu_profile_hits
, cpu
)[1] = NULL
;
388 return notifier_from_errno(-ENOMEM
);
390 case CPU_ONLINE_FROZEN
:
391 if (prof_cpu_mask
!= NULL
)
392 cpumask_set_cpu(cpu
, prof_cpu_mask
);
394 case CPU_UP_CANCELED
:
395 case CPU_UP_CANCELED_FROZEN
:
397 case CPU_DEAD_FROZEN
:
398 if (prof_cpu_mask
!= NULL
)
399 cpumask_clear_cpu(cpu
, prof_cpu_mask
);
400 if (per_cpu(cpu_profile_hits
, cpu
)[0]) {
401 page
= virt_to_page(per_cpu(cpu_profile_hits
, cpu
)[0]);
402 per_cpu(cpu_profile_hits
, cpu
)[0] = NULL
;
405 if (per_cpu(cpu_profile_hits
, cpu
)[1]) {
406 page
= virt_to_page(per_cpu(cpu_profile_hits
, cpu
)[1]);
407 per_cpu(cpu_profile_hits
, cpu
)[1] = NULL
;
414 #else /* !CONFIG_SMP */
415 #define profile_flip_buffers() do { } while (0)
416 #define profile_discard_flip_buffers() do { } while (0)
417 #define profile_cpu_callback NULL
419 static void do_profile_hits(int type
, void *__pc
, unsigned int nr_hits
)
422 pc
= ((unsigned long)__pc
- (unsigned long)_stext
) >> prof_shift
;
423 atomic_add(nr_hits
, &prof_buffer
[min(pc
, prof_len
- 1)]);
425 #endif /* !CONFIG_SMP */
427 void profile_hits(int type
, void *__pc
, unsigned int nr_hits
)
429 if (prof_on
!= type
|| !prof_buffer
)
431 do_profile_hits(type
, __pc
, nr_hits
);
433 EXPORT_SYMBOL_GPL(profile_hits
);
435 void profile_tick(int type
)
437 struct pt_regs
*regs
= get_irq_regs();
439 if (type
== CPU_PROFILING
&& timer_hook
)
441 if (!user_mode(regs
) && prof_cpu_mask
!= NULL
&&
442 cpumask_test_cpu(smp_processor_id(), prof_cpu_mask
))
443 profile_hit(type
, (void *)profile_pc(regs
));
446 #ifdef CONFIG_PROC_FS
447 #include <linux/proc_fs.h>
448 #include <linux/seq_file.h>
449 #include <asm/uaccess.h>
451 static int prof_cpu_mask_proc_show(struct seq_file
*m
, void *v
)
453 seq_cpumask(m
, prof_cpu_mask
);
458 static int prof_cpu_mask_proc_open(struct inode
*inode
, struct file
*file
)
460 return single_open(file
, prof_cpu_mask_proc_show
, NULL
);
463 static ssize_t
prof_cpu_mask_proc_write(struct file
*file
,
464 const char __user
*buffer
, size_t count
, loff_t
*pos
)
466 cpumask_var_t new_value
;
469 if (!alloc_cpumask_var(&new_value
, GFP_KERNEL
))
472 err
= cpumask_parse_user(buffer
, count
, new_value
);
474 cpumask_copy(prof_cpu_mask
, new_value
);
477 free_cpumask_var(new_value
);
481 static const struct file_operations prof_cpu_mask_proc_fops
= {
482 .open
= prof_cpu_mask_proc_open
,
485 .release
= single_release
,
486 .write
= prof_cpu_mask_proc_write
,
489 void create_prof_cpu_mask(struct proc_dir_entry
*root_irq_dir
)
491 /* create /proc/irq/prof_cpu_mask */
492 proc_create("prof_cpu_mask", 0600, root_irq_dir
, &prof_cpu_mask_proc_fops
);
496 * This function accesses profiling information. The returned data is
497 * binary: the sampling step and the actual contents of the profile
498 * buffer. Use of the program readprofile is recommended in order to
499 * get meaningful info out of these data.
502 read_profile(struct file
*file
, char __user
*buf
, size_t count
, loff_t
*ppos
)
504 unsigned long p
= *ppos
;
507 unsigned int sample_step
= 1 << prof_shift
;
509 profile_flip_buffers();
510 if (p
>= (prof_len
+1)*sizeof(unsigned int))
512 if (count
> (prof_len
+1)*sizeof(unsigned int) - p
)
513 count
= (prof_len
+1)*sizeof(unsigned int) - p
;
516 while (p
< sizeof(unsigned int) && count
> 0) {
517 if (put_user(*((char *)(&sample_step
)+p
), buf
))
519 buf
++; p
++; count
--; read
++;
521 pnt
= (char *)prof_buffer
+ p
- sizeof(atomic_t
);
522 if (copy_to_user(buf
, (void *)pnt
, count
))
530 * Writing to /proc/profile resets the counters
532 * Writing a 'profiling multiplier' value into it also re-sets the profiling
533 * interrupt frequency, on architectures that support this.
535 static ssize_t
write_profile(struct file
*file
, const char __user
*buf
,
536 size_t count
, loff_t
*ppos
)
539 extern int setup_profiling_timer(unsigned int multiplier
);
541 if (count
== sizeof(int)) {
542 unsigned int multiplier
;
544 if (copy_from_user(&multiplier
, buf
, sizeof(int)))
547 if (setup_profiling_timer(multiplier
))
551 profile_discard_flip_buffers();
552 memset(prof_buffer
, 0, prof_len
* sizeof(atomic_t
));
556 static const struct file_operations proc_profile_operations
= {
557 .read
= read_profile
,
558 .write
= write_profile
,
559 .llseek
= default_llseek
,
563 static void profile_nop(void *unused
)
567 static int create_hash_tables(void)
571 for_each_online_cpu(cpu
) {
572 int node
= cpu_to_mem(cpu
);
575 page
= alloc_pages_exact_node(node
,
576 GFP_KERNEL
| __GFP_ZERO
| GFP_THISNODE
,
580 per_cpu(cpu_profile_hits
, cpu
)[1]
581 = (struct profile_hit
*)page_address(page
);
582 page
= alloc_pages_exact_node(node
,
583 GFP_KERNEL
| __GFP_ZERO
| GFP_THISNODE
,
587 per_cpu(cpu_profile_hits
, cpu
)[0]
588 = (struct profile_hit
*)page_address(page
);
594 on_each_cpu(profile_nop
, NULL
, 1);
595 for_each_online_cpu(cpu
) {
598 if (per_cpu(cpu_profile_hits
, cpu
)[0]) {
599 page
= virt_to_page(per_cpu(cpu_profile_hits
, cpu
)[0]);
600 per_cpu(cpu_profile_hits
, cpu
)[0] = NULL
;
603 if (per_cpu(cpu_profile_hits
, cpu
)[1]) {
604 page
= virt_to_page(per_cpu(cpu_profile_hits
, cpu
)[1]);
605 per_cpu(cpu_profile_hits
, cpu
)[1] = NULL
;
612 #define create_hash_tables() ({ 0; })
615 int __ref
create_proc_profile(void) /* false positive from hotcpu_notifier */
617 struct proc_dir_entry
*entry
;
621 if (create_hash_tables())
623 entry
= proc_create("profile", S_IWUSR
| S_IRUGO
,
624 NULL
, &proc_profile_operations
);
627 entry
->size
= (1+prof_len
) * sizeof(atomic_t
);
628 hotcpu_notifier(profile_cpu_callback
, 0);
631 module_init(create_proc_profile
);
632 #endif /* CONFIG_PROC_FS */