kernel/trace/trace_functions_graph.c

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