drivers/oprofile/buffer_sync.c

   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>
   8  *
   9  * This is the core of the buffer management. Each
  10  * CPU buffer is processed and entered into the
  11  * global event buffer. Such processing is necessary
  12  * in several circumstances, mentioned below.
  13  *
  14  * The processing does the job of converting the
  15  * transitory EIP value into a persistent dentry/offset
  16  * value that the profiler can record at its leisure.
  17  *
  18  * See fs/dcookies.c for a description of the dentry/offset
  19  * objects.
  20  */
  21
  22 #include <linux/mm.h>
  23 #include <linux/workqueue.h>
  24 #include <linux/notifier.h>
  25 #include <linux/dcookies.h>
  26 #include <linux/profile.h>
  27 #include <linux/module.h>
  28 #include <linux/fs.h>
  29
  30 #include "oprofile_stats.h"
  31 #include "event_buffer.h"
  32 #include "cpu_buffer.h"
  33 #include "buffer_sync.h"
  34
  35 static LIST_HEAD(dying_tasks);
  36 static LIST_HEAD(dead_tasks);
  37 static cpumask_t marked_cpus = CPU_MASK_NONE;
  38 static DEFINE_SPINLOCK(task_mortuary);
  39 static void process_task_mortuary(void);
  40
  41
  42 /* Take ownership of the task struct and place it on the
  43  * list for processing. Only after two full buffer syncs
  44  * does the task eventually get freed, because by then
  45  * we are sure we will not reference it again.
  46  * Can be invoked from softirq via RCU callback due to
  47  * call_rcu() of the task struct, hence the _irqsave.
  48  */
  49 static int task_free_notify(struct notifier_block * self, unsigned long val, void * data)
  50 {
  51         unsigned long flags;
  52         struct task_struct * task = data;
  53         spin_lock_irqsave(&task_mortuary, flags);
  54         list_add(&task->tasks, &dying_tasks);
  55         spin_unlock_irqrestore(&task_mortuary, flags);
  56         return NOTIFY_OK;
  57 }
  58
  59
  60 /* The task is on its way out. A sync of the buffer means we can catch
  61  * any remaining samples for this task.
  62  */
  63 static int task_exit_notify(struct notifier_block * self, unsigned long val, void * data)
  64 {
  65         /* To avoid latency problems, we only process the current CPU,
  66          * hoping that most samples for the task are on this CPU
  67          */
  68         sync_buffer(raw_smp_processor_id());
  69         return 0;
  70 }
  71
  72
  73 /* The task is about to try a do_munmap(). We peek at what it's going to
  74  * do, and if it's an executable region, process the samples first, so
  75  * we don't lose any. This does not have to be exact, it's a QoI issue
  76  * only.
  77  */
  78 static int munmap_notify(struct notifier_block * self, unsigned long val, void * data)
  79 {
  80         unsigned long addr = (unsigned long)data;
  81         struct mm_struct * mm = current->mm;
  82         struct vm_area_struct * mpnt;
  83
  84         down_read(&mm->mmap_sem);
  85
  86         mpnt = find_vma(mm, addr);
  87         if (mpnt && mpnt->vm_file && (mpnt->vm_flags & VM_EXEC)) {
  88                 up_read(&mm->mmap_sem);
  89                 /* To avoid latency problems, we only process the current CPU,
  90                  * hoping that most samples for the task are on this CPU
  91                  */
  92                 sync_buffer(raw_smp_processor_id());
  93                 return 0;
  94         }
  95
  96         up_read(&mm->mmap_sem);
  97         return 0;
  98 }
  99
 100
 101 /* We need to be told about new modules so we don't attribute to a previously
 102  * loaded module, or drop the samples on the floor.
 103  */
 104 static int module_load_notify(struct notifier_block * self, unsigned long val, void * data)
 105 {
 106 #ifdef CONFIG_MODULES
 107         if (val != MODULE_STATE_COMING)
 108                 return 0;
 109
 110         /* FIXME: should we process all CPU buffers ? */
 111         down(&buffer_sem);
 112         add_event_entry(ESCAPE_CODE);
 113         add_event_entry(MODULE_LOADED_CODE);
 114         up(&buffer_sem);
 115 #endif
 116         return 0;
 117 }
 118
 119
 120 static struct notifier_block task_free_nb = {
 121         .notifier_call  = task_free_notify,
 122 };
 123
 124 static struct notifier_block task_exit_nb = {
 125         .notifier_call  = task_exit_notify,
 126 };
 127
 128 static struct notifier_block munmap_nb = {
 129         .notifier_call  = munmap_notify,
 130 };
 131
 132 static struct notifier_block module_load_nb = {
 133         .notifier_call = module_load_notify,
 134 };
 135
 136
 137 static void end_sync(void)
 138 {
 139         end_cpu_work();
 140         /* make sure we don't leak task structs */
 141         process_task_mortuary();
 142         process_task_mortuary();
 143 }
 144
 145
 146 int sync_start(void)
 147 {
 148         int err;
 149
 150         start_cpu_work();
 151
 152         err = task_handoff_register(&task_free_nb);
 153         if (err)
 154                 goto out1;
 155         err = profile_event_register(PROFILE_TASK_EXIT, &task_exit_nb);
 156         if (err)
 157                 goto out2;
 158         err = profile_event_register(PROFILE_MUNMAP, &munmap_nb);
 159         if (err)
 160                 goto out3;
 161         err = register_module_notifier(&module_load_nb);
 162         if (err)
 163                 goto out4;
 164
 165 out:
 166         return err;
 167 out4:
 168         profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
 169 out3:
 170         profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
 171 out2:
 172         task_handoff_unregister(&task_free_nb);
 173 out1:
 174         end_sync();
 175         goto out;
 176 }
 177
 178
 179 void sync_stop(void)
 180 {
 181         unregister_module_notifier(&module_load_nb);
 182         profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
 183         profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
 184         task_handoff_unregister(&task_free_nb);
 185         end_sync();
 186 }
 187
 188
 189 /* Optimisation. We can manage without taking the dcookie sem
 190  * because we cannot reach this code without at least one
 191  * dcookie user still being registered (namely, the reader
 192  * of the event buffer). */
 193 static inline unsigned long fast_get_dcookie(struct dentry * dentry,
 194         struct vfsmount * vfsmnt)
 195 {
 196         unsigned long cookie;
 197
 198         if (dentry->d_cookie)
 199                 return (unsigned long)dentry;
 200         get_dcookie(dentry, vfsmnt, &cookie);
 201         return cookie;
 202 }
 203
 204
 205 /* Look up the dcookie for the task's first VM_EXECUTABLE mapping,
 206  * which corresponds loosely to "application name". This is
 207  * not strictly necessary but allows oprofile to associate
 208  * shared-library samples with particular applications
 209  */
 210 static unsigned long get_exec_dcookie(struct mm_struct * mm)
 211 {
 212         unsigned long cookie = NO_COOKIE;
 213         struct vm_area_struct * vma;
 214
 215         if (!mm)
 216                 goto out;
 217
 218         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 219                 if (!vma->vm_file)
 220                         continue;
 221                 if (!(vma->vm_flags & VM_EXECUTABLE))
 222                         continue;
 223                 cookie = fast_get_dcookie(vma->vm_file->f_dentry,
 224                         vma->vm_file->f_vfsmnt);
 225                 break;
 226         }
 227
 228 out:
 229         return cookie;
 230 }
 231
 232
 233 /* Convert the EIP value of a sample into a persistent dentry/offset
 234  * pair that can then be added to the global event buffer. We make
 235  * sure to do this lookup before a mm->mmap modification happens so
 236  * we don't lose track.
 237  */
 238 static unsigned long lookup_dcookie(struct mm_struct * mm, unsigned long addr, off_t * offset)
 239 {
 240         unsigned long cookie = NO_COOKIE;
 241         struct vm_area_struct * vma;
 242
 243         for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
 244
 245                 if (addr < vma->vm_start || addr >= vma->vm_end)
 246                         continue;
 247
 248                 if (vma->vm_file) {
 249                         cookie = fast_get_dcookie(vma->vm_file->f_dentry,
 250                                 vma->vm_file->f_vfsmnt);
 251                         *offset = (vma->vm_pgoff << PAGE_SHIFT) + addr -
 252                                 vma->vm_start;
 253                 } else {
 254                         /* must be an anonymous map */
 255                         *offset = addr;
 256                 }
 257
 258                 break;
 259         }
 260
 261         if (!vma)
 262                 cookie = INVALID_COOKIE;
 263
 264         return cookie;
 265 }
 266
 267
 268 static unsigned long last_cookie = INVALID_COOKIE;
 269
 270 static void add_cpu_switch(int i)
 271 {
 272         add_event_entry(ESCAPE_CODE);
 273         add_event_entry(CPU_SWITCH_CODE);
 274         add_event_entry(i);
 275         last_cookie = INVALID_COOKIE;
 276 }
 277
 278 static void add_kernel_ctx_switch(unsigned int in_kernel)
 279 {
 280         add_event_entry(ESCAPE_CODE);
 281         if (in_kernel)
 282                 add_event_entry(KERNEL_ENTER_SWITCH_CODE);
 283         else
 284                 add_event_entry(KERNEL_EXIT_SWITCH_CODE);
 285 }
 286
 287 static void
 288 add_user_ctx_switch(struct task_struct const * task, unsigned long cookie)
 289 {
 290         add_event_entry(ESCAPE_CODE);
 291         add_event_entry(CTX_SWITCH_CODE);
 292         add_event_entry(task->pid);
 293         add_event_entry(cookie);
 294         /* Another code for daemon back-compat */
 295         add_event_entry(ESCAPE_CODE);
 296         add_event_entry(CTX_TGID_CODE);
 297         add_event_entry(task->tgid);
 298 }
 299
 300
 301 static void add_cookie_switch(unsigned long cookie)
 302 {
 303         add_event_entry(ESCAPE_CODE);
 304         add_event_entry(COOKIE_SWITCH_CODE);
 305         add_event_entry(cookie);
 306 }
 307
 308
 309 static void add_trace_begin(void)
 310 {
 311         add_event_entry(ESCAPE_CODE);
 312         add_event_entry(TRACE_BEGIN_CODE);
 313 }
 314
 315
 316 static void add_sample_entry(unsigned long offset, unsigned long event)
 317 {
 318         add_event_entry(offset);
 319         add_event_entry(event);
 320 }
 321
 322
 323 static int add_us_sample(struct mm_struct * mm, struct op_sample * s)
 324 {
 325         unsigned long cookie;
 326         off_t offset;
 327
 328         cookie = lookup_dcookie(mm, s->eip, &offset);
 329
 330         if (cookie == INVALID_COOKIE) {
 331                 atomic_inc(&oprofile_stats.sample_lost_no_mapping);
 332                 return 0;
 333         }
 334
 335         if (cookie != last_cookie) {
 336                 add_cookie_switch(cookie);
 337                 last_cookie = cookie;
 338         }
 339
 340         add_sample_entry(offset, s->event);
 341
 342         return 1;
 343 }
 344
 345
 346 /* Add a sample to the global event buffer. If possible the
 347  * sample is converted into a persistent dentry/offset pair
 348  * for later lookup from userspace.
 349  */
 350 static int
 351 add_sample(struct mm_struct * mm, struct op_sample * s, int in_kernel)
 352 {
 353         if (in_kernel) {
 354                 add_sample_entry(s->eip, s->event);
 355                 return 1;
 356         } else if (mm) {
 357                 return add_us_sample(mm, s);
 358         } else {
 359                 atomic_inc(&oprofile_stats.sample_lost_no_mm);
 360         }
 361         return 0;
 362 }
 363
 364
 365 static void release_mm(struct mm_struct * mm)
 366 {
 367         if (!mm)
 368                 return;
 369         up_read(&mm->mmap_sem);
 370         mmput(mm);
 371 }
 372
 373
 374 static struct mm_struct * take_tasks_mm(struct task_struct * task)
 375 {
 376         struct mm_struct * mm = get_task_mm(task);
 377         if (mm)
 378                 down_read(&mm->mmap_sem);
 379         return mm;
 380 }
 381
 382
 383 static inline int is_code(unsigned long val)
 384 {
 385         return val == ESCAPE_CODE;
 386 }
 387
 388
 389 /* "acquire" as many cpu buffer slots as we can */
 390 static unsigned long get_slots(struct oprofile_cpu_buffer * b)
 391 {
 392         unsigned long head = b->head_pos;
 393         unsigned long tail = b->tail_pos;
 394
 395         /*
 396          * Subtle. This resets the persistent last_task
 397          * and in_kernel values used for switching notes.
 398          * BUT, there is a small window between reading
 399          * head_pos, and this call, that means samples
 400          * can appear at the new head position, but not
 401          * be prefixed with the notes for switching
 402          * kernel mode or a task switch. This small hole
 403          * can lead to mis-attribution or samples where
 404          * we don't know if it's in the kernel or not,
 405          * at the start of an event buffer.
 406          */
 407         cpu_buffer_reset(b);
 408
 409         if (head >= tail)
 410                 return head - tail;
 411
 412         return head + (b->buffer_size - tail);
 413 }
 414
 415
 416 static void increment_tail(struct oprofile_cpu_buffer * b)
 417 {
 418         unsigned long new_tail = b->tail_pos + 1;
 419
 420         rmb();
 421
 422         if (new_tail < b->buffer_size)
 423                 b->tail_pos = new_tail;
 424         else
 425                 b->tail_pos = 0;
 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  */
 435 static 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
 456 static void mark_done(int cpu)
 457 {
 458         int i;
 459
 460         cpu_set(cpu, marked_cpus);
 461
 462         for_each_online_cpu(i) {
 463                 if (!cpu_isset(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         cpus_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  */
 480 typedef 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_sem and doing
 490  * lookup in task->mm->mmap to convert EIP into dcookie/offset
 491  * value.
 492  */
 493 void sync_buffer(int cpu)
 494 {
 495         struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[cpu];
 496         struct mm_struct *mm = NULL;
 497         struct task_struct * new;
 498         unsigned long cookie = 0;
 499         int in_kernel = 1;
 500         unsigned int i;
 501         sync_buffer_state state = sb_buffer_start;
 502         unsigned long available;
 503
 504         down(&buffer_sem);
 505
 506         add_cpu_switch(cpu);
 507
 508         /* Remember, only we can modify tail_pos */
 509
 510         available = get_slots(cpu_buf);
 511
 512         for (i = 0; i < available; ++i) {
 513                 struct op_sample * s = &cpu_buf->buffer[cpu_buf->tail_pos];
 514
 515                 if (is_code(s->eip)) {
 516                         if (s->event <= CPU_IS_KERNEL) {
 517                                 /* kernel/userspace switch */
 518                                 in_kernel = s->event;
 519                                 if (state == sb_buffer_start)
 520                                         state = sb_sample_start;
 521                                 add_kernel_ctx_switch(s->event);
 522                         } else if (s->event == CPU_TRACE_BEGIN) {
 523                                 state = sb_bt_start;
 524                                 add_trace_begin();
 525                         } else {
 526                                 struct mm_struct * oldmm = mm;
 527
 528                                 /* userspace context switch */
 529                                 new = (struct task_struct *)s->event;
 530
 531                                 release_mm(oldmm);
 532                                 mm = take_tasks_mm(new);
 533                                 if (mm != oldmm)
 534                                         cookie = get_exec_dcookie(mm);
 535                                 add_user_ctx_switch(new, cookie);
 536                         }
 537                 } else {
 538                         if (state >= sb_bt_start &&
 539                             !add_sample(mm, s, in_kernel)) {
 540                                 if (state == sb_bt_start) {
 541                                         state = sb_bt_ignore;
 542                                         atomic_inc(&oprofile_stats.bt_lost_no_mapping);
 543                                 }
 544                         }
 545                 }
 546
 547                 increment_tail(cpu_buf);
 548         }
 549         release_mm(mm);
 550
 551         mark_done(cpu);
 552
 553         up(&buffer_sem);
 554 }