kernel/trace/trace_functions_graph.c

   1 /*
   2  *
   3  * Function graph tracer.
   4  * Copyright (c) 2008-2009 Frederic Weisbecker <[email protected]>
   5  * Mostly borrowed from function tracer which
   6  * is Copyright (c) Steven Rostedt <[email protected]>
   7  *
   8  */
   9 #include <linux/uaccess.h>
  10 #include <linux/ftrace.h>
  11 #include <linux/interrupt.h>
  12 #include <linux/slab.h>
  13 #include <linux/fs.h>
  14
  15 #include "trace.h"
  16 #include "trace_output.h"
  17
  18 static bool kill_ftrace_graph;
  19
  20 /**
  21  * ftrace_graph_is_dead - returns true if ftrace_graph_stop() was called
  22  *
  23  * ftrace_graph_stop() is called when a severe error is detected in
  24  * the function graph tracing. This function is called by the critical
  25  * paths of function graph to keep those paths from doing any more harm.
  26  */
  27 bool ftrace_graph_is_dead(void)
  28 {
  29         return kill_ftrace_graph;
  30 }
  31
  32 /**
  33  * ftrace_graph_stop - set to permanently disable function graph tracincg
  34  *
  35  * In case of an error int function graph tracing, this is called
  36  * to try to keep function graph tracing from causing any more harm.
  37  * Usually this is pretty severe and this is called to try to at least
  38  * get a warning out to the user.
  39  */
  40 void ftrace_graph_stop(void)
  41 {
  42         kill_ftrace_graph = true;
  43 }
  44
  45 /* When set, irq functions will be ignored */
  46 static int ftrace_graph_skip_irqs;
  47
  48 struct fgraph_cpu_data {
  49         pid_t           last_pid;
  50         int             depth;
  51         int             depth_irq;
  52         int             ignore;
  53         unsigned long   enter_funcs[FTRACE_RETFUNC_DEPTH];
  54 };
  55
  56 struct fgraph_data {
  57         struct fgraph_cpu_data __percpu *cpu_data;
  58
  59         /* Place to preserve last processed entry. */
  60         struct ftrace_graph_ent_entry   ent;
  61         struct ftrace_graph_ret_entry   ret;
  62         int                             failed;
  63         int                             cpu;
  64 };
  65
  66 #define TRACE_GRAPH_INDENT      2
  67
  68 unsigned int fgraph_max_depth;
  69
  70 static struct tracer_opt trace_opts[] = {
  71         /* Display overruns? (for self-debug purpose) */
  72         { TRACER_OPT(funcgraph-overrun, TRACE_GRAPH_PRINT_OVERRUN) },
  73         /* Display CPU ? */
  74         { TRACER_OPT(funcgraph-cpu, TRACE_GRAPH_PRINT_CPU) },
  75         /* Display Overhead ? */
  76         { TRACER_OPT(funcgraph-overhead, TRACE_GRAPH_PRINT_OVERHEAD) },
  77         /* Display proc name/pid */
  78         { TRACER_OPT(funcgraph-proc, TRACE_GRAPH_PRINT_PROC) },
  79         /* Display duration of execution */
  80         { TRACER_OPT(funcgraph-duration, TRACE_GRAPH_PRINT_DURATION) },
  81         /* Display absolute time of an entry */
  82         { TRACER_OPT(funcgraph-abstime, TRACE_GRAPH_PRINT_ABS_TIME) },
  83         /* Display interrupts */
  84         { TRACER_OPT(funcgraph-irqs, TRACE_GRAPH_PRINT_IRQS) },
  85         /* Display function name after trailing } */
  86         { TRACER_OPT(funcgraph-tail, TRACE_GRAPH_PRINT_TAIL) },
  87         /* Include sleep time (scheduled out) between entry and return */
  88         { TRACER_OPT(sleep-time, TRACE_GRAPH_SLEEP_TIME) },
  89         /* Include time within nested functions */
  90         { TRACER_OPT(graph-time, TRACE_GRAPH_GRAPH_TIME) },
  91         { } /* Empty entry */
  92 };
  93
  94 static struct tracer_flags tracer_flags = {
  95         /* Don't display overruns, proc, or tail by default */
  96         .val = TRACE_GRAPH_PRINT_CPU | TRACE_GRAPH_PRINT_OVERHEAD |
  97                TRACE_GRAPH_PRINT_DURATION | TRACE_GRAPH_PRINT_IRQS |
  98                TRACE_GRAPH_SLEEP_TIME | TRACE_GRAPH_GRAPH_TIME,
  99         .opts = trace_opts
 100 };
 101
 102 static struct trace_array *graph_array;
 103
 104 /*
 105  * DURATION column is being also used to display IRQ signs,
 106  * following values are used by print_graph_irq and others
 107  * to fill in space into DURATION column.
 108  */
 109 enum {
 110         FLAGS_FILL_FULL  = 1 << TRACE_GRAPH_PRINT_FILL_SHIFT,
 111         FLAGS_FILL_START = 2 << TRACE_GRAPH_PRINT_FILL_SHIFT,
 112         FLAGS_FILL_END   = 3 << TRACE_GRAPH_PRINT_FILL_SHIFT,
 113 };
 114
 115 static void
 116 print_graph_duration(struct trace_array *tr, unsigned long long duration,
 117                      struct trace_seq *s, u32 flags);
 118
 119 /* Add a function return address to the trace stack on thread info.*/
 120 int
 121 ftrace_push_return_trace(unsigned long ret, unsigned long func, int *depth,
 122                          unsigned long frame_pointer, unsigned long *retp)
 123 {
 124         unsigned long long calltime;
 125         int index;
 126
 127         if (unlikely(ftrace_graph_is_dead()))
 128                 return -EBUSY;
 129
 130         if (!current->ret_stack)
 131                 return -EBUSY;
 132
 133         /*
 134          * We must make sure the ret_stack is tested before we read
 135          * anything else.
 136          */
 137         smp_rmb();
 138
 139         /* The return trace stack is full */
 140         if (current->curr_ret_stack == FTRACE_RETFUNC_DEPTH - 1) {
 141                 atomic_inc(&current->trace_overrun);
 142                 return -EBUSY;
 143         }
 144
 145         /*
 146          * The curr_ret_stack is an index to ftrace return stack of
 147          * current task.  Its value should be in [0, FTRACE_RETFUNC_
 148          * DEPTH) when the function graph tracer is used.  To support
 149          * filtering out specific functions, it makes the index
 150          * negative by subtracting huge value (FTRACE_NOTRACE_DEPTH)
 151          * so when it sees a negative index the ftrace will ignore
 152          * the record.  And the index gets recovered when returning
 153          * from the filtered function by adding the FTRACE_NOTRACE_
 154          * DEPTH and then it'll continue to record functions normally.
 155          *
 156          * The curr_ret_stack is initialized to -1 and get increased
 157          * in this function.  So it can be less than -1 only if it was
 158          * filtered out via ftrace_graph_notrace_addr() which can be
 159          * set from set_graph_notrace file in tracefs by user.
 160          */
 161         if (current->curr_ret_stack < -1)
 162                 return -EBUSY;
 163
 164         calltime = trace_clock_local();
 165
 166         index = ++current->curr_ret_stack;
 167         if (ftrace_graph_notrace_addr(func))
 168                 current->curr_ret_stack -= FTRACE_NOTRACE_DEPTH;
 169         barrier();
 170         current->ret_stack[index].ret = ret;
 171         current->ret_stack[index].func = func;
 172         current->ret_stack[index].calltime = calltime;
 173 #ifdef HAVE_FUNCTION_GRAPH_FP_TEST
 174         current->ret_stack[index].fp = frame_pointer;
 175 #endif
 176 #ifdef HAVE_FUNCTION_GRAPH_RET_ADDR_PTR
 177         current->ret_stack[index].retp = retp;
 178 #endif
 179         *depth = current->curr_ret_stack;
 180
 181         return 0;
 182 }
 183
 184 /* Retrieve a function return address to the trace stack on thread info.*/
 185 static void
 186 ftrace_pop_return_trace(struct ftrace_graph_ret *trace, unsigned long *ret,
 187                         unsigned long frame_pointer)
 188 {
 189         int index;
 190
 191         index = current->curr_ret_stack;
 192
 193         /*
 194          * A negative index here means that it's just returned from a
 195          * notrace'd function.  Recover index to get an original
 196          * return address.  See ftrace_push_return_trace().
 197          *
 198          * TODO: Need to check whether the stack gets corrupted.
 199          */
 200         if (index < 0)
 201                 index += FTRACE_NOTRACE_DEPTH;
 202
 203         if (unlikely(index < 0 || index >= FTRACE_RETFUNC_DEPTH)) {
 204                 ftrace_graph_stop();
 205                 WARN_ON(1);
 206                 /* Might as well panic, otherwise we have no where to go */
 207                 *ret = (unsigned long)panic;
 208                 return;
 209         }
 210
 211 #ifdef HAVE_FUNCTION_GRAPH_FP_TEST
 212         /*
 213          * The arch may choose to record the frame pointer used
 214          * and check it here to make sure that it is what we expect it
 215          * to be. If gcc does not set the place holder of the return
 216          * address in the frame pointer, and does a copy instead, then
 217          * the function graph trace will fail. This test detects this
 218          * case.
 219          *
 220          * Currently, x86_32 with optimize for size (-Os) makes the latest
 221          * gcc do the above.
 222          *
 223          * Note, -mfentry does not use frame pointers, and this test
 224          *  is not needed if CC_USING_FENTRY is set.
 225          */
 226         if (unlikely(current->ret_stack[index].fp != frame_pointer)) {
 227                 ftrace_graph_stop();
 228                 WARN(1, "Bad frame pointer: expected %lx, received %lx\n"
 229                      "  from func %ps return to %lx\n",
 230                      current->ret_stack[index].fp,
 231                      frame_pointer,
 232                      (void *)current->ret_stack[index].func,
 233                      current->ret_stack[index].ret);
 234                 *ret = (unsigned long)panic;
 235                 return;
 236         }
 237 #endif
 238
 239         *ret = current->ret_stack[index].ret;
 240         trace->func = current->ret_stack[index].func;
 241         trace->calltime = current->ret_stack[index].calltime;
 242         trace->overrun = atomic_read(&current->trace_overrun);
 243         trace->depth = index;
 244 }
 245
 246 /*
 247  * Send the trace to the ring-buffer.
 248  * @return the original return address.
 249  */
 250 unsigned long ftrace_return_to_handler(unsigned long frame_pointer)
 251 {
 252         struct ftrace_graph_ret trace;
 253         unsigned long ret;
 254
 255         ftrace_pop_return_trace(&trace, &ret, frame_pointer);
 256         trace.rettime = trace_clock_local();
 257         barrier();
 258         current->curr_ret_stack--;
 259         /*
 260          * The curr_ret_stack can be less than -1 only if it was
 261          * filtered out and it's about to return from the function.
 262          * Recover the index and continue to trace normal functions.
 263          */
 264         if (current->curr_ret_stack < -1) {
 265                 current->curr_ret_stack += FTRACE_NOTRACE_DEPTH;
 266                 return ret;
 267         }
 268
 269         /*
 270          * The trace should run after decrementing the ret counter
 271          * in case an interrupt were to come in. We don't want to
 272          * lose the interrupt if max_depth is set.
 273          */
 274         ftrace_graph_return(&trace);
 275
 276         if (unlikely(!ret)) {
 277                 ftrace_graph_stop();
 278                 WARN_ON(1);
 279                 /* Might as well panic. What else to do? */
 280                 ret = (unsigned long)panic;
 281         }
 282
 283         return ret;
 284 }
 285
 286 /**
 287  * ftrace_graph_ret_addr - convert a potentially modified stack return address
 288  *                         to its original value
 289  *
 290  * This function can be called by stack unwinding code to convert a found stack
 291  * return address ('ret') to its original value, in case the function graph
 292  * tracer has modified it to be 'return_to_handler'.  If the address hasn't
 293  * been modified, the unchanged value of 'ret' is returned.
 294  *
 295  * 'idx' is a state variable which should be initialized by the caller to zero
 296  * before the first call.
 297  *
 298  * 'retp' is a pointer to the return address on the stack.  It's ignored if
 299  * the arch doesn't have HAVE_FUNCTION_GRAPH_RET_ADDR_PTR defined.
 300  */
 301 #ifdef HAVE_FUNCTION_GRAPH_RET_ADDR_PTR
 302 unsigned long ftrace_graph_ret_addr(struct task_struct *task, int *idx,
 303                                     unsigned long ret, unsigned long *retp)
 304 {
 305         int index = task->curr_ret_stack;
 306         int i;
 307
 308         if (ret != (unsigned long)return_to_handler)
 309                 return ret;
 310
 311         if (index < -1)
 312                 index += FTRACE_NOTRACE_DEPTH;
 313
 314         if (index < 0)
 315                 return ret;
 316
 317         for (i = 0; i <= index; i++)
 318                 if (task->ret_stack[i].retp == retp)
 319                         return task->ret_stack[i].ret;
 320
 321         return ret;
 322 }
 323 #else /* !HAVE_FUNCTION_GRAPH_RET_ADDR_PTR */
 324 unsigned long ftrace_graph_ret_addr(struct task_struct *task, int *idx,
 325                                     unsigned long ret, unsigned long *retp)
 326 {
 327         int task_idx;
 328
 329         if (ret != (unsigned long)return_to_handler)
 330                 return ret;
 331
 332         task_idx = task->curr_ret_stack;
 333
 334         if (!task->ret_stack || task_idx < *idx)
 335                 return ret;
 336
 337         task_idx -= *idx;
 338         (*idx)++;
 339
 340         return task->ret_stack[task_idx].ret;
 341 }
 342 #endif /* HAVE_FUNCTION_GRAPH_RET_ADDR_PTR */
 343
 344 int __trace_graph_entry(struct trace_array *tr,
 345                                 struct ftrace_graph_ent *trace,
 346                                 unsigned long flags,
 347                                 int pc)
 348 {
 349         struct trace_event_call *call = &event_funcgraph_entry;
 350         struct ring_buffer_event *event;
 351         struct ring_buffer *buffer = tr->trace_buffer.buffer;
 352         struct ftrace_graph_ent_entry *entry;
 353
 354         event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_ENT,
 355                                           sizeof(*entry), flags, pc);
 356         if (!event)
 357                 return 0;
 358         entry   = ring_buffer_event_data(event);
 359         entry->graph_ent                        = *trace;
 360         if (!call_filter_check_discard(call, entry, buffer, event))
 361                 trace_buffer_unlock_commit_nostack(buffer, event);
 362
 363         return 1;
 364 }
 365
 366 static inline int ftrace_graph_ignore_irqs(void)
 367 {
 368         if (!ftrace_graph_skip_irqs || trace_recursion_test(TRACE_IRQ_BIT))
 369                 return 0;
 370
 371         return in_irq();
 372 }
 373
 374 int trace_graph_entry(struct ftrace_graph_ent *trace)
 375 {
 376         struct trace_array *tr = graph_array;
 377         struct trace_array_cpu *data;
 378         unsigned long flags;
 379         long disabled;
 380         int ret;
 381         int cpu;
 382         int pc;
 383
 384         if (!ftrace_trace_task(tr))
 385                 return 0;
 386
 387         if (ftrace_graph_ignore_func(trace))
 388                 return 0;
 389
 390         if (ftrace_graph_ignore_irqs())
 391                 return 0;
 392
 393         /*
 394          * Do not trace a function if it's filtered by set_graph_notrace.
 395          * Make the index of ret stack negative to indicate that it should
 396          * ignore further functions.  But it needs its own ret stack entry
 397          * to recover the original index in order to continue tracing after
 398          * returning from the function.
 399          */
 400         if (ftrace_graph_notrace_addr(trace->func))
 401                 return 1;
 402
 403         /*
 404          * Stop here if tracing_threshold is set. We only write function return
 405          * events to the ring buffer.
 406          */
 407         if (tracing_thresh)
 408                 return 1;
 409
 410         local_irq_save(flags);
 411         cpu = raw_smp_processor_id();
 412         data = per_cpu_ptr(tr->trace_buffer.data, cpu);
 413         disabled = atomic_inc_return(&data->disabled);
 414         if (likely(disabled == 1)) {
 415                 pc = preempt_count();
 416                 ret = __trace_graph_entry(tr, trace, flags, pc);
 417         } else {
 418                 ret = 0;
 419         }
 420
 421         atomic_dec(&data->disabled);
 422         local_irq_restore(flags);
 423
 424         return ret;
 425 }
 426
 427 static void
 428 __trace_graph_function(struct trace_array *tr,
 429                 unsigned long ip, unsigned long flags, int pc)
 430 {
 431         u64 time = trace_clock_local();
 432         struct ftrace_graph_ent ent = {
 433                 .func  = ip,
 434                 .depth = 0,
 435         };
 436         struct ftrace_graph_ret ret = {
 437                 .func     = ip,
 438                 .depth    = 0,
 439                 .calltime = time,
 440                 .rettime  = time,
 441         };
 442
 443         __trace_graph_entry(tr, &ent, flags, pc);
 444         __trace_graph_return(tr, &ret, flags, pc);
 445 }
 446
 447 void
 448 trace_graph_function(struct trace_array *tr,
 449                 unsigned long ip, unsigned long parent_ip,
 450                 unsigned long flags, int pc)
 451 {
 452         __trace_graph_function(tr, ip, flags, pc);
 453 }
 454
 455 void __trace_graph_return(struct trace_array *tr,
 456                                 struct ftrace_graph_ret *trace,
 457                                 unsigned long flags,
 458                                 int pc)
 459 {
 460         struct trace_event_call *call = &event_funcgraph_exit;
 461         struct ring_buffer_event *event;
 462         struct ring_buffer *buffer = tr->trace_buffer.buffer;
 463         struct ftrace_graph_ret_entry *entry;
 464
 465         event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_RET,
 466                                           sizeof(*entry), flags, pc);
 467         if (!event)
 468                 return;
 469         entry   = ring_buffer_event_data(event);
 470         entry->ret                              = *trace;
 471         if (!call_filter_check_discard(call, entry, buffer, event))
 472                 trace_buffer_unlock_commit_nostack(buffer, event);
 473 }
 474
 475 void trace_graph_return(struct ftrace_graph_ret *trace)
 476 {
 477         struct trace_array *tr = graph_array;
 478         struct trace_array_cpu *data;
 479         unsigned long flags;
 480         long disabled;
 481         int cpu;
 482         int pc;
 483
 484         local_irq_save(flags);
 485         cpu = raw_smp_processor_id();
 486         data = per_cpu_ptr(tr->trace_buffer.data, cpu);
 487         disabled = atomic_inc_return(&data->disabled);
 488         if (likely(disabled == 1)) {
 489                 pc = preempt_count();
 490                 __trace_graph_return(tr, trace, flags, pc);
 491         }
 492         atomic_dec(&data->disabled);
 493         local_irq_restore(flags);
 494 }
 495
 496 void set_graph_array(struct trace_array *tr)
 497 {
 498         graph_array = tr;
 499
 500         /* Make graph_array visible before we start tracing */
 501
 502         smp_mb();
 503 }
 504
 505 static void trace_graph_thresh_return(struct ftrace_graph_ret *trace)
 506 {
 507         if (tracing_thresh &&
 508             (trace->rettime - trace->calltime < tracing_thresh))
 509                 return;
 510         else
 511                 trace_graph_return(trace);
 512 }
 513
 514 static int graph_trace_init(struct trace_array *tr)
 515 {
 516         int ret;
 517
 518         set_graph_array(tr);
 519         if (tracing_thresh)
 520                 ret = register_ftrace_graph(&trace_graph_thresh_return,
 521                                             &trace_graph_entry);
 522         else
 523                 ret = register_ftrace_graph(&trace_graph_return,
 524                                             &trace_graph_entry);
 525         if (ret)
 526                 return ret;
 527         tracing_start_cmdline_record();
 528
 529         return 0;
 530 }
 531
 532 static void graph_trace_reset(struct trace_array *tr)
 533 {
 534         tracing_stop_cmdline_record();
 535         unregister_ftrace_graph();
 536 }
 537
 538 static int graph_trace_update_thresh(struct trace_array *tr)
 539 {
 540         graph_trace_reset(tr);
 541         return graph_trace_init(tr);
 542 }
 543
 544 static int max_bytes_for_cpu;
 545
 546 static void print_graph_cpu(struct trace_seq *s, int cpu)
 547 {
 548         /*
 549          * Start with a space character - to make it stand out
 550          * to the right a bit when trace output is pasted into
 551          * email:
 552          */
 553         trace_seq_printf(s, " %*d) ", max_bytes_for_cpu, cpu);
 554 }
 555
 556 #define TRACE_GRAPH_PROCINFO_LENGTH     14
 557
 558 static void print_graph_proc(struct trace_seq *s, pid_t pid)
 559 {
 560         char comm[TASK_COMM_LEN];
 561         /* sign + log10(MAX_INT) + '\0' */
 562         char pid_str[11];
 563         int spaces = 0;
 564         int len;
 565         int i;
 566
 567         trace_find_cmdline(pid, comm);
 568         comm[7] = '\0';
 569         sprintf(pid_str, "%d", pid);
 570
 571         /* 1 stands for the "-" character */
 572         len = strlen(comm) + strlen(pid_str) + 1;
 573
 574         if (len < TRACE_GRAPH_PROCINFO_LENGTH)
 575                 spaces = TRACE_GRAPH_PROCINFO_LENGTH - len;
 576
 577         /* First spaces to align center */
 578         for (i = 0; i < spaces / 2; i++)
 579                 trace_seq_putc(s, ' ');
 580
 581         trace_seq_printf(s, "%s-%s", comm, pid_str);
 582
 583         /* Last spaces to align center */
 584         for (i = 0; i < spaces - (spaces / 2); i++)
 585                 trace_seq_putc(s, ' ');
 586 }
 587
 588
 589 static void print_graph_lat_fmt(struct trace_seq *s, struct trace_entry *entry)
 590 {
 591         trace_seq_putc(s, ' ');
 592         trace_print_lat_fmt(s, entry);
 593 }
 594
 595 /* If the pid changed since the last trace, output this event */
 596 static void
 597 verif_pid(struct trace_seq *s, pid_t pid, int cpu, struct fgraph_data *data)
 598 {
 599         pid_t prev_pid;
 600         pid_t *last_pid;
 601
 602         if (!data)
 603                 return;
 604
 605         last_pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
 606
 607         if (*last_pid == pid)
 608                 return;
 609
 610         prev_pid = *last_pid;
 611         *last_pid = pid;
 612
 613         if (prev_pid == -1)
 614                 return;
 615 /*
 616  * Context-switch trace line:
 617
 618  ------------------------------------------
 619  | 1)  migration/0--1  =>  sshd-1755
 620  ------------------------------------------
 621
 622  */
 623         trace_seq_puts(s, " ------------------------------------------\n");
 624         print_graph_cpu(s, cpu);
 625         print_graph_proc(s, prev_pid);
 626         trace_seq_puts(s, " => ");
 627         print_graph_proc(s, pid);
 628         trace_seq_puts(s, "\n ------------------------------------------\n\n");
 629 }
 630
 631 static struct ftrace_graph_ret_entry *
 632 get_return_for_leaf(struct trace_iterator *iter,
 633                 struct ftrace_graph_ent_entry *curr)
 634 {
 635         struct fgraph_data *data = iter->private;
 636         struct ring_buffer_iter *ring_iter = NULL;
 637         struct ring_buffer_event *event;
 638         struct ftrace_graph_ret_entry *next;
 639
 640         /*
 641          * If the previous output failed to write to the seq buffer,
 642          * then we just reuse the data from before.
 643          */
 644         if (data && data->failed) {
 645                 curr = &data->ent;
 646                 next = &data->ret;
 647         } else {
 648
 649                 ring_iter = trace_buffer_iter(iter, iter->cpu);
 650
 651                 /* First peek to compare current entry and the next one */
 652                 if (ring_iter)
 653                         event = ring_buffer_iter_peek(ring_iter, NULL);
 654                 else {
 655                         /*
 656                          * We need to consume the current entry to see
 657                          * the next one.
 658                          */
 659                         ring_buffer_consume(iter->trace_buffer->buffer, iter->cpu,
 660                                             NULL, NULL);
 661                         event = ring_buffer_peek(iter->trace_buffer->buffer, iter->cpu,
 662                                                  NULL, NULL);
 663                 }
 664
 665                 if (!event)
 666                         return NULL;
 667
 668                 next = ring_buffer_event_data(event);
 669
 670                 if (data) {
 671                         /*
 672                          * Save current and next entries for later reference
 673                          * if the output fails.
 674                          */
 675                         data->ent = *curr;
 676                         /*
 677                          * If the next event is not a return type, then
 678                          * we only care about what type it is. Otherwise we can
 679                          * safely copy the entire event.
 680                          */
 681                         if (next->ent.type == TRACE_GRAPH_RET)
 682                                 data->ret = *next;
 683                         else
 684                                 data->ret.ent.type = next->ent.type;
 685                 }
 686         }
 687
 688         if (next->ent.type != TRACE_GRAPH_RET)
 689                 return NULL;
 690
 691         if (curr->ent.pid != next->ent.pid ||
 692                         curr->graph_ent.func != next->ret.func)
 693                 return NULL;
 694
 695         /* this is a leaf, now advance the iterator */
 696         if (ring_iter)
 697                 ring_buffer_read(ring_iter, NULL);
 698
 699         return next;
 700 }
 701
 702 static void print_graph_abs_time(u64 t, struct trace_seq *s)
 703 {
 704         unsigned long usecs_rem;
 705
 706         usecs_rem = do_div(t, NSEC_PER_SEC);
 707         usecs_rem /= 1000;
 708
 709         trace_seq_printf(s, "%5lu.%06lu |  ",
 710                          (unsigned long)t, usecs_rem);
 711 }
 712
 713 static void
 714 print_graph_irq(struct trace_iterator *iter, unsigned long addr,
 715                 enum trace_type type, int cpu, pid_t pid, u32 flags)
 716 {
 717         struct trace_array *tr = iter->tr;
 718         struct trace_seq *s = &iter->seq;
 719         struct trace_entry *ent = iter->ent;
 720
 721         if (addr < (unsigned long)__irqentry_text_start ||
 722                 addr >= (unsigned long)__irqentry_text_end)
 723                 return;
 724
 725         if (tr->trace_flags & TRACE_ITER_CONTEXT_INFO) {
 726                 /* Absolute time */
 727                 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
 728                         print_graph_abs_time(iter->ts, s);
 729
 730                 /* Cpu */
 731                 if (flags & TRACE_GRAPH_PRINT_CPU)
 732                         print_graph_cpu(s, cpu);
 733
 734                 /* Proc */
 735                 if (flags & TRACE_GRAPH_PRINT_PROC) {
 736                         print_graph_proc(s, pid);
 737                         trace_seq_puts(s, " | ");
 738                 }
 739
 740                 /* Latency format */
 741                 if (tr->trace_flags & TRACE_ITER_LATENCY_FMT)
 742                         print_graph_lat_fmt(s, ent);
 743         }
 744
 745         /* No overhead */
 746         print_graph_duration(tr, 0, s, flags | FLAGS_FILL_START);
 747
 748         if (type == TRACE_GRAPH_ENT)
 749                 trace_seq_puts(s, "==========>");
 750         else
 751                 trace_seq_puts(s, "<==========");
 752
 753         print_graph_duration(tr, 0, s, flags | FLAGS_FILL_END);
 754         trace_seq_putc(s, '\n');
 755 }
 756
 757 void
 758 trace_print_graph_duration(unsigned long long duration, struct trace_seq *s)
 759 {
 760         unsigned long nsecs_rem = do_div(duration, 1000);
 761         /* log10(ULONG_MAX) + '\0' */
 762         char usecs_str[21];
 763         char nsecs_str[5];
 764         int len;
 765         int i;
 766
 767         sprintf(usecs_str, "%lu", (unsigned long) duration);
 768
 769         /* Print msecs */
 770         trace_seq_printf(s, "%s", usecs_str);
 771
 772         len = strlen(usecs_str);
 773
 774         /* Print nsecs (we don't want to exceed 7 numbers) */
 775         if (len < 7) {
 776                 size_t slen = min_t(size_t, sizeof(nsecs_str), 8UL - len);
 777
 778                 snprintf(nsecs_str, slen, "%03lu", nsecs_rem);
 779                 trace_seq_printf(s, ".%s", nsecs_str);
 780                 len += strlen(nsecs_str) + 1;
 781         }
 782
 783         trace_seq_puts(s, " us ");
 784
 785         /* Print remaining spaces to fit the row's width */
 786         for (i = len; i < 8; i++)
 787                 trace_seq_putc(s, ' ');
 788 }
 789
 790 static void
 791 print_graph_duration(struct trace_array *tr, unsigned long long duration,
 792                      struct trace_seq *s, u32 flags)
 793 {
 794         if (!(flags & TRACE_GRAPH_PRINT_DURATION) ||
 795             !(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
 796                 return;
 797
 798         /* No real adata, just filling the column with spaces */
 799         switch (flags & TRACE_GRAPH_PRINT_FILL_MASK) {
 800         case FLAGS_FILL_FULL:
 801                 trace_seq_puts(s, "              |  ");
 802                 return;
 803         case FLAGS_FILL_START:
 804                 trace_seq_puts(s, "  ");
 805                 return;
 806         case FLAGS_FILL_END:
 807                 trace_seq_puts(s, " |");
 808                 return;
 809         }
 810
 811         /* Signal a overhead of time execution to the output */
 812         if (flags & TRACE_GRAPH_PRINT_OVERHEAD)
 813                 trace_seq_printf(s, "%c ", trace_find_mark(duration));
 814         else
 815                 trace_seq_puts(s, "  ");
 816
 817         trace_print_graph_duration(duration, s);
 818         trace_seq_puts(s, "|  ");
 819 }
 820
 821 /* Case of a leaf function on its call entry */
 822 static enum print_line_t
 823 print_graph_entry_leaf(struct trace_iterator *iter,
 824                 struct ftrace_graph_ent_entry *entry,
 825                 struct ftrace_graph_ret_entry *ret_entry,
 826                 struct trace_seq *s, u32 flags)
 827 {
 828         struct fgraph_data *data = iter->private;
 829         struct trace_array *tr = iter->tr;
 830         struct ftrace_graph_ret *graph_ret;
 831         struct ftrace_graph_ent *call;
 832         unsigned long long duration;
 833         int i;
 834
 835         graph_ret = &ret_entry->ret;
 836         call = &entry->graph_ent;
 837         duration = graph_ret->rettime - graph_ret->calltime;
 838
 839         if (data) {
 840                 struct fgraph_cpu_data *cpu_data;
 841                 int cpu = iter->cpu;
 842
 843                 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
 844
 845                 /* If a graph tracer ignored set_graph_notrace */
 846                 if (call->depth < -1)
 847                         call->depth += FTRACE_NOTRACE_DEPTH;
 848
 849                 /*
 850                  * Comments display at + 1 to depth. Since
 851                  * this is a leaf function, keep the comments
 852                  * equal to this depth.
 853                  */
 854                 cpu_data->depth = call->depth - 1;
 855
 856                 /* No need to keep this function around for this depth */
 857                 if (call->depth < FTRACE_RETFUNC_DEPTH &&
 858                     !WARN_ON_ONCE(call->depth < 0))
 859                         cpu_data->enter_funcs[call->depth] = 0;
 860         }
 861
 862         /* Overhead and duration */
 863         print_graph_duration(tr, duration, s, flags);
 864
 865         /* Function */
 866         for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
 867                 trace_seq_putc(s, ' ');
 868
 869         trace_seq_printf(s, "%ps();\n", (void *)call->func);
 870
 871         return trace_handle_return(s);
 872 }
 873
 874 static enum print_line_t
 875 print_graph_entry_nested(struct trace_iterator *iter,
 876                          struct ftrace_graph_ent_entry *entry,
 877                          struct trace_seq *s, int cpu, u32 flags)
 878 {
 879         struct ftrace_graph_ent *call = &entry->graph_ent;
 880         struct fgraph_data *data = iter->private;
 881         struct trace_array *tr = iter->tr;
 882         int i;
 883
 884         if (data) {
 885                 struct fgraph_cpu_data *cpu_data;
 886                 int cpu = iter->cpu;
 887
 888                 /* If a graph tracer ignored set_graph_notrace */
 889                 if (call->depth < -1)
 890                         call->depth += FTRACE_NOTRACE_DEPTH;
 891
 892                 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
 893                 cpu_data->depth = call->depth;
 894
 895                 /* Save this function pointer to see if the exit matches */
 896                 if (call->depth < FTRACE_RETFUNC_DEPTH &&
 897                     !WARN_ON_ONCE(call->depth < 0))
 898                         cpu_data->enter_funcs[call->depth] = call->func;
 899         }
 900
 901         /* No time */
 902         print_graph_duration(tr, 0, s, flags | FLAGS_FILL_FULL);
 903
 904         /* Function */
 905         for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
 906                 trace_seq_putc(s, ' ');
 907
 908         trace_seq_printf(s, "%ps() {\n", (void *)call->func);
 909
 910         if (trace_seq_has_overflowed(s))
 911                 return TRACE_TYPE_PARTIAL_LINE;
 912
 913         /*
 914          * we already consumed the current entry to check the next one
 915          * and see if this is a leaf.
 916          */
 917         return TRACE_TYPE_NO_CONSUME;
 918 }
 919
 920 static void
 921 print_graph_prologue(struct trace_iterator *iter, struct trace_seq *s,
 922                      int type, unsigned long addr, u32 flags)
 923 {
 924         struct fgraph_data *data = iter->private;
 925         struct trace_entry *ent = iter->ent;
 926         struct trace_array *tr = iter->tr;
 927         int cpu = iter->cpu;
 928
 929         /* Pid */
 930         verif_pid(s, ent->pid, cpu, data);
 931
 932         if (type)
 933                 /* Interrupt */
 934                 print_graph_irq(iter, addr, type, cpu, ent->pid, flags);
 935
 936         if (!(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
 937                 return;
 938
 939         /* Absolute time */
 940         if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
 941                 print_graph_abs_time(iter->ts, s);
 942
 943         /* Cpu */
 944         if (flags & TRACE_GRAPH_PRINT_CPU)
 945                 print_graph_cpu(s, cpu);
 946
 947         /* Proc */
 948         if (flags & TRACE_GRAPH_PRINT_PROC) {
 949                 print_graph_proc(s, ent->pid);
 950                 trace_seq_puts(s, " | ");
 951         }
 952
 953         /* Latency format */
 954         if (tr->trace_flags & TRACE_ITER_LATENCY_FMT)
 955                 print_graph_lat_fmt(s, ent);
 956
 957         return;
 958 }
 959
 960 /*
 961  * Entry check for irq code
 962  *
 963  * returns 1 if
 964  *  - we are inside irq code
 965  *  - we just entered irq code
 966  *
 967  * retunns 0 if
 968  *  - funcgraph-interrupts option is set
 969  *  - we are not inside irq code
 970  */
 971 static int
 972 check_irq_entry(struct trace_iterator *iter, u32 flags,
 973                 unsigned long addr, int depth)
 974 {
 975         int cpu = iter->cpu;
 976         int *depth_irq;
 977         struct fgraph_data *data = iter->private;
 978
 979         /*
 980          * If we are either displaying irqs, or we got called as
 981          * a graph event and private data does not exist,
 982          * then we bypass the irq check.
 983          */
 984         if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
 985             (!data))
 986                 return 0;
 987
 988         depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
 989
 990         /*
 991          * We are inside the irq code
 992          */
 993         if (*depth_irq >= 0)
 994                 return 1;
 995
 996         if ((addr < (unsigned long)__irqentry_text_start) ||
 997             (addr >= (unsigned long)__irqentry_text_end))
 998                 return 0;
 999
1000         /*
1001          * We are entering irq code.
1002          */
1003         *depth_irq = depth;
1004         return 1;
1005 }
1006
1007 /*
1008  * Return check for irq code
1009  *
1010  * returns 1 if
1011  *  - we are inside irq code
1012  *  - we just left irq code
1013  *
1014  * returns 0 if
1015  *  - funcgraph-interrupts option is set
1016  *  - we are not inside irq code
1017  */
1018 static int
1019 check_irq_return(struct trace_iterator *iter, u32 flags, int depth)
1020 {
1021         int cpu = iter->cpu;
1022         int *depth_irq;
1023         struct fgraph_data *data = iter->private;
1024
1025         /*
1026          * If we are either displaying irqs, or we got called as
1027          * a graph event and private data does not exist,
1028          * then we bypass the irq check.
1029          */
1030         if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
1031             (!data))
1032                 return 0;
1033
1034         depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
1035
1036         /*
1037          * We are not inside the irq code.
1038          */
1039         if (*depth_irq == -1)
1040                 return 0;
1041
1042         /*
1043          * We are inside the irq code, and this is returning entry.
1044          * Let's not trace it and clear the entry depth, since
1045          * we are out of irq code.
1046          *
1047          * This condition ensures that we 'leave the irq code' once
1048          * we are out of the entry depth. Thus protecting us from
1049          * the RETURN entry loss.
1050          */
1051         if (*depth_irq >= depth) {
1052                 *depth_irq = -1;
1053                 return 1;
1054         }
1055
1056         /*
1057          * We are inside the irq code, and this is not the entry.
1058          */
1059         return 1;
1060 }
1061
1062 static enum print_line_t
1063 print_graph_entry(struct ftrace_graph_ent_entry *field, struct trace_seq *s,
1064                         struct trace_iterator *iter, u32 flags)
1065 {
1066         struct fgraph_data *data = iter->private;
1067         struct ftrace_graph_ent *call = &field->graph_ent;
1068         struct ftrace_graph_ret_entry *leaf_ret;
1069         static enum print_line_t ret;
1070         int cpu = iter->cpu;
1071
1072         if (check_irq_entry(iter, flags, call->func, call->depth))
1073                 return TRACE_TYPE_HANDLED;
1074
1075         print_graph_prologue(iter, s, TRACE_GRAPH_ENT, call->func, flags);
1076
1077         leaf_ret = get_return_for_leaf(iter, field);
1078         if (leaf_ret)
1079                 ret = print_graph_entry_leaf(iter, field, leaf_ret, s, flags);
1080         else
1081                 ret = print_graph_entry_nested(iter, field, s, cpu, flags);
1082
1083         if (data) {
1084                 /*
1085                  * If we failed to write our output, then we need to make
1086                  * note of it. Because we already consumed our entry.
1087                  */
1088                 if (s->full) {
1089                         data->failed = 1;
1090                         data->cpu = cpu;
1091                 } else
1092                         data->failed = 0;
1093         }
1094
1095         return ret;
1096 }
1097
1098 static enum print_line_t
1099 print_graph_return(struct ftrace_graph_ret *trace, struct trace_seq *s,
1100                    struct trace_entry *ent, struct trace_iterator *iter,
1101                    u32 flags)
1102 {
1103         unsigned long long duration = trace->rettime - trace->calltime;
1104         struct fgraph_data *data = iter->private;
1105         struct trace_array *tr = iter->tr;
1106         pid_t pid = ent->pid;
1107         int cpu = iter->cpu;
1108         int func_match = 1;
1109         int i;
1110
1111         if (check_irq_return(iter, flags, trace->depth))
1112                 return TRACE_TYPE_HANDLED;
1113
1114         if (data) {
1115                 struct fgraph_cpu_data *cpu_data;
1116                 int cpu = iter->cpu;
1117
1118                 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
1119
1120                 /*
1121                  * Comments display at + 1 to depth. This is the
1122                  * return from a function, we now want the comments
1123                  * to display at the same level of the bracket.
1124                  */
1125                 cpu_data->depth = trace->depth - 1;
1126
1127                 if (trace->depth < FTRACE_RETFUNC_DEPTH &&
1128                     !WARN_ON_ONCE(trace->depth < 0)) {
1129                         if (cpu_data->enter_funcs[trace->depth] != trace->func)
1130                                 func_match = 0;
1131                         cpu_data->enter_funcs[trace->depth] = 0;
1132                 }
1133         }
1134
1135         print_graph_prologue(iter, s, 0, 0, flags);
1136
1137         /* Overhead and duration */
1138         print_graph_duration(tr, duration, s, flags);
1139
1140         /* Closing brace */
1141         for (i = 0; i < trace->depth * TRACE_GRAPH_INDENT; i++)
1142                 trace_seq_putc(s, ' ');
1143
1144         /*
1145          * If the return function does not have a matching entry,
1146          * then the entry was lost. Instead of just printing
1147          * the '}' and letting the user guess what function this
1148          * belongs to, write out the function name. Always do
1149          * that if the funcgraph-tail option is enabled.
1150          */
1151         if (func_match && !(flags & TRACE_GRAPH_PRINT_TAIL))
1152                 trace_seq_puts(s, "}\n");
1153         else
1154                 trace_seq_printf(s, "} /* %ps */\n", (void *)trace->func);
1155
1156         /* Overrun */
1157         if (flags & TRACE_GRAPH_PRINT_OVERRUN)
1158                 trace_seq_printf(s, " (Overruns: %lu)\n",
1159                                  trace->overrun);
1160
1161         print_graph_irq(iter, trace->func, TRACE_GRAPH_RET,
1162                         cpu, pid, flags);
1163
1164         return trace_handle_return(s);
1165 }
1166
1167 static enum print_line_t
1168 print_graph_comment(struct trace_seq *s, struct trace_entry *ent,
1169                     struct trace_iterator *iter, u32 flags)
1170 {
1171         struct trace_array *tr = iter->tr;
1172         unsigned long sym_flags = (tr->trace_flags & TRACE_ITER_SYM_MASK);
1173         struct fgraph_data *data = iter->private;
1174         struct trace_event *event;
1175         int depth = 0;
1176         int ret;
1177         int i;
1178
1179         if (data)
1180                 depth = per_cpu_ptr(data->cpu_data, iter->cpu)->depth;
1181
1182         print_graph_prologue(iter, s, 0, 0, flags);
1183
1184         /* No time */
1185         print_graph_duration(tr, 0, s, flags | FLAGS_FILL_FULL);
1186
1187         /* Indentation */
1188         if (depth > 0)
1189                 for (i = 0; i < (depth + 1) * TRACE_GRAPH_INDENT; i++)
1190                         trace_seq_putc(s, ' ');
1191
1192         /* The comment */
1193         trace_seq_puts(s, "/* ");
1194
1195         switch (iter->ent->type) {
1196         case TRACE_BPUTS:
1197                 ret = trace_print_bputs_msg_only(iter);
1198                 if (ret != TRACE_TYPE_HANDLED)
1199                         return ret;
1200                 break;
1201         case TRACE_BPRINT:
1202                 ret = trace_print_bprintk_msg_only(iter);
1203                 if (ret != TRACE_TYPE_HANDLED)
1204                         return ret;
1205                 break;
1206         case TRACE_PRINT:
1207                 ret = trace_print_printk_msg_only(iter);
1208                 if (ret != TRACE_TYPE_HANDLED)
1209                         return ret;
1210                 break;
1211         default:
1212                 event = ftrace_find_event(ent->type);
1213                 if (!event)
1214                         return TRACE_TYPE_UNHANDLED;
1215
1216                 ret = event->funcs->trace(iter, sym_flags, event);
1217                 if (ret != TRACE_TYPE_HANDLED)
1218                         return ret;
1219         }
1220
1221         if (trace_seq_has_overflowed(s))
1222                 goto out;
1223
1224         /* Strip ending newline */
1225         if (s->buffer[s->seq.len - 1] == '\n') {
1226                 s->buffer[s->seq.len - 1] = '\0';
1227                 s->seq.len--;
1228         }
1229
1230         trace_seq_puts(s, " */\n");
1231  out:
1232         return trace_handle_return(s);
1233 }
1234
1235
1236 enum print_line_t
1237 print_graph_function_flags(struct trace_iterator *iter, u32 flags)
1238 {
1239         struct ftrace_graph_ent_entry *field;
1240         struct fgraph_data *data = iter->private;
1241         struct trace_entry *entry = iter->ent;
1242         struct trace_seq *s = &iter->seq;
1243         int cpu = iter->cpu;
1244         int ret;
1245
1246         if (data && per_cpu_ptr(data->cpu_data, cpu)->ignore) {
1247                 per_cpu_ptr(data->cpu_data, cpu)->ignore = 0;
1248                 return TRACE_TYPE_HANDLED;
1249         }
1250
1251         /*
1252          * If the last output failed, there's a possibility we need
1253          * to print out the missing entry which would never go out.
1254          */
1255         if (data && data->failed) {
1256                 field = &data->ent;
1257                 iter->cpu = data->cpu;
1258                 ret = print_graph_entry(field, s, iter, flags);
1259                 if (ret == TRACE_TYPE_HANDLED && iter->cpu != cpu) {
1260                         per_cpu_ptr(data->cpu_data, iter->cpu)->ignore = 1;
1261                         ret = TRACE_TYPE_NO_CONSUME;
1262                 }
1263                 iter->cpu = cpu;
1264                 return ret;
1265         }
1266
1267         switch (entry->type) {
1268         case TRACE_GRAPH_ENT: {
1269                 /*
1270                  * print_graph_entry() may consume the current event,
1271                  * thus @field may become invalid, so we need to save it.
1272                  * sizeof(struct ftrace_graph_ent_entry) is very small,
1273                  * it can be safely saved at the stack.
1274                  */
1275                 struct ftrace_graph_ent_entry saved;
1276                 trace_assign_type(field, entry);
1277                 saved = *field;
1278                 return print_graph_entry(&saved, s, iter, flags);
1279         }
1280         case TRACE_GRAPH_RET: {
1281                 struct ftrace_graph_ret_entry *field;
1282                 trace_assign_type(field, entry);
1283                 return print_graph_return(&field->ret, s, entry, iter, flags);
1284         }
1285         case TRACE_STACK:
1286         case TRACE_FN:
1287                 /* dont trace stack and functions as comments */
1288                 return TRACE_TYPE_UNHANDLED;
1289
1290         default:
1291                 return print_graph_comment(s, entry, iter, flags);
1292         }
1293
1294         return TRACE_TYPE_HANDLED;
1295 }
1296
1297 static enum print_line_t
1298 print_graph_function(struct trace_iterator *iter)
1299 {
1300         return print_graph_function_flags(iter, tracer_flags.val);
1301 }
1302
1303 static enum print_line_t
1304 print_graph_function_event(struct trace_iterator *iter, int flags,
1305                            struct trace_event *event)
1306 {
1307         return print_graph_function(iter);
1308 }
1309
1310 static void print_lat_header(struct seq_file *s, u32 flags)
1311 {
1312         static const char spaces[] = "                " /* 16 spaces */
1313                 "    "                                  /* 4 spaces */
1314                 "                 ";                    /* 17 spaces */
1315         int size = 0;
1316
1317         if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1318                 size += 16;
1319         if (flags & TRACE_GRAPH_PRINT_CPU)
1320                 size += 4;
1321         if (flags & TRACE_GRAPH_PRINT_PROC)
1322                 size += 17;
1323
1324         seq_printf(s, "#%.*s  _-----=> irqs-off        \n", size, spaces);
1325         seq_printf(s, "#%.*s / _----=> need-resched    \n", size, spaces);
1326         seq_printf(s, "#%.*s| / _---=> hardirq/softirq \n", size, spaces);
1327         seq_printf(s, "#%.*s|| / _--=> preempt-depth   \n", size, spaces);
1328         seq_printf(s, "#%.*s||| /                      \n", size, spaces);
1329 }
1330
1331 static void __print_graph_headers_flags(struct trace_array *tr,
1332                                         struct seq_file *s, u32 flags)
1333 {
1334         int lat = tr->trace_flags & TRACE_ITER_LATENCY_FMT;
1335
1336         if (lat)
1337                 print_lat_header(s, flags);
1338
1339         /* 1st line */
1340         seq_putc(s, '#');
1341         if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1342                 seq_puts(s, "     TIME       ");
1343         if (flags & TRACE_GRAPH_PRINT_CPU)
1344                 seq_puts(s, " CPU");
1345         if (flags & TRACE_GRAPH_PRINT_PROC)
1346                 seq_puts(s, "  TASK/PID       ");
1347         if (lat)
1348                 seq_puts(s, "||||");
1349         if (flags & TRACE_GRAPH_PRINT_DURATION)
1350                 seq_puts(s, "  DURATION   ");
1351         seq_puts(s, "               FUNCTION CALLS\n");
1352
1353         /* 2nd line */
1354         seq_putc(s, '#');
1355         if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1356                 seq_puts(s, "      |         ");
1357         if (flags & TRACE_GRAPH_PRINT_CPU)
1358                 seq_puts(s, " |  ");
1359         if (flags & TRACE_GRAPH_PRINT_PROC)
1360                 seq_puts(s, "   |    |        ");
1361         if (lat)
1362                 seq_puts(s, "||||");
1363         if (flags & TRACE_GRAPH_PRINT_DURATION)
1364                 seq_puts(s, "   |   |      ");
1365         seq_puts(s, "               |   |   |   |\n");
1366 }
1367
1368 static void print_graph_headers(struct seq_file *s)
1369 {
1370         print_graph_headers_flags(s, tracer_flags.val);
1371 }
1372
1373 void print_graph_headers_flags(struct seq_file *s, u32 flags)
1374 {
1375         struct trace_iterator *iter = s->private;
1376         struct trace_array *tr = iter->tr;
1377
1378         if (!(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
1379                 return;
1380
1381         if (tr->trace_flags & TRACE_ITER_LATENCY_FMT) {
1382                 /* print nothing if the buffers are empty */
1383                 if (trace_empty(iter))
1384                         return;
1385
1386                 print_trace_header(s, iter);
1387         }
1388
1389         __print_graph_headers_flags(tr, s, flags);
1390 }
1391
1392 void graph_trace_open(struct trace_iterator *iter)
1393 {
1394         /* pid and depth on the last trace processed */
1395         struct fgraph_data *data;
1396         gfp_t gfpflags;
1397         int cpu;
1398
1399         iter->private = NULL;
1400
1401         /* We can be called in atomic context via ftrace_dump() */
1402         gfpflags = (in_atomic() || irqs_disabled()) ? GFP_ATOMIC : GFP_KERNEL;
1403
1404         data = kzalloc(sizeof(*data), gfpflags);
1405         if (!data)
1406                 goto out_err;
1407
1408         data->cpu_data = alloc_percpu_gfp(struct fgraph_cpu_data, gfpflags);
1409         if (!data->cpu_data)
1410                 goto out_err_free;
1411
1412         for_each_possible_cpu(cpu) {
1413                 pid_t *pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
1414                 int *depth = &(per_cpu_ptr(data->cpu_data, cpu)->depth);
1415                 int *ignore = &(per_cpu_ptr(data->cpu_data, cpu)->ignore);
1416                 int *depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
1417
1418                 *pid = -1;
1419                 *depth = 0;
1420                 *ignore = 0;
1421                 *depth_irq = -1;
1422         }
1423
1424         iter->private = data;
1425
1426         return;
1427
1428  out_err_free:
1429         kfree(data);
1430  out_err:
1431         pr_warn("function graph tracer: not enough memory\n");
1432 }
1433
1434 void graph_trace_close(struct trace_iterator *iter)
1435 {
1436         struct fgraph_data *data = iter->private;
1437
1438         if (data) {
1439                 free_percpu(data->cpu_data);
1440                 kfree(data);
1441         }
1442 }
1443
1444 static int
1445 func_graph_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set)
1446 {
1447         if (bit == TRACE_GRAPH_PRINT_IRQS)
1448                 ftrace_graph_skip_irqs = !set;
1449
1450         if (bit == TRACE_GRAPH_SLEEP_TIME)
1451                 ftrace_graph_sleep_time_control(set);
1452
1453         if (bit == TRACE_GRAPH_GRAPH_TIME)
1454                 ftrace_graph_graph_time_control(set);
1455
1456         return 0;
1457 }
1458
1459 static struct trace_event_functions graph_functions = {
1460         .trace          = print_graph_function_event,
1461 };
1462
1463 static struct trace_event graph_trace_entry_event = {
1464         .type           = TRACE_GRAPH_ENT,
1465         .funcs          = &graph_functions,
1466 };
1467
1468 static struct trace_event graph_trace_ret_event = {
1469         .type           = TRACE_GRAPH_RET,
1470         .funcs          = &graph_functions
1471 };
1472
1473 static struct tracer graph_trace __tracer_data = {
1474         .name           = "function_graph",
1475         .update_thresh  = graph_trace_update_thresh,
1476         .open           = graph_trace_open,
1477         .pipe_open      = graph_trace_open,
1478         .close          = graph_trace_close,
1479         .pipe_close     = graph_trace_close,
1480         .init           = graph_trace_init,
1481         .reset          = graph_trace_reset,
1482         .print_line     = print_graph_function,
1483         .print_header   = print_graph_headers,
1484         .flags          = &tracer_flags,
1485         .set_flag       = func_graph_set_flag,
1486 #ifdef CONFIG_FTRACE_SELFTEST
1487         .selftest       = trace_selftest_startup_function_graph,
1488 #endif
1489 };
1490
1491
1492 static ssize_t
1493 graph_depth_write(struct file *filp, const char __user *ubuf, size_t cnt,
1494                   loff_t *ppos)
1495 {
1496         unsigned long val;
1497         int ret;
1498
1499         ret = kstrtoul_from_user(ubuf, cnt, 10, &val);
1500         if (ret)
1501                 return ret;
1502
1503         fgraph_max_depth = val;
1504
1505         *ppos += cnt;
1506
1507         return cnt;
1508 }
1509
1510 static ssize_t
1511 graph_depth_read(struct file *filp, char __user *ubuf, size_t cnt,
1512                  loff_t *ppos)
1513 {
1514         char buf[15]; /* More than enough to hold UINT_MAX + "\n"*/
1515         int n;
1516
1517         n = sprintf(buf, "%d\n", fgraph_max_depth);
1518
1519         return simple_read_from_buffer(ubuf, cnt, ppos, buf, n);
1520 }
1521
1522 static const struct file_operations graph_depth_fops = {
1523         .open           = tracing_open_generic,
1524         .write          = graph_depth_write,
1525         .read           = graph_depth_read,
1526         .llseek         = generic_file_llseek,
1527 };
1528
1529 static __init int init_graph_tracefs(void)
1530 {
1531         struct dentry *d_tracer;
1532
1533         d_tracer = tracing_init_dentry();
1534         if (IS_ERR(d_tracer))
1535                 return 0;
1536
1537         trace_create_file("max_graph_depth", 0644, d_tracer,
1538                           NULL, &graph_depth_fops);
1539
1540         return 0;
1541 }
1542 fs_initcall(init_graph_tracefs);
1543
1544 static __init int init_graph_trace(void)
1545 {
1546         max_bytes_for_cpu = snprintf(NULL, 0, "%d", nr_cpu_ids - 1);
1547
1548         if (!register_trace_event(&graph_trace_entry_event)) {
1549                 pr_warn("Warning: could not register graph trace events\n");
1550                 return 1;
1551         }
1552
1553         if (!register_trace_event(&graph_trace_ret_event)) {
1554                 pr_warn("Warning: could not register graph trace events\n");
1555                 return 1;
1556         }
1557
1558         return register_tracer(&graph_trace);
1559 }
1560
1561 core_initcall(init_graph_trace);