kernel/trace/trace_functions_graph.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *
   4  * Function graph tracer.
   5  * Copyright (c) 2008-2009 Frederic Weisbecker <fweisbec@gmail.com>
   6  * Mostly borrowed from function tracer which
   7  * is Copyright (c) Steven Rostedt <srostedt@redhat.com>
   8  *
   9  */
  10 #include <linux/uaccess.h>
  11 #include <linux/ftrace.h>
  12 #include <linux/interrupt.h>
  13 #include <linux/slab.h>
  14 #include <linux/fs.h>
  15
  16 #include "trace.h"
  17 #include "trace_output.h"
  18
  19 static bool kill_ftrace_graph;
  20
  21 /**
  22  * ftrace_graph_is_dead - returns true if ftrace_graph_stop() was called
  23  *
  24  * ftrace_graph_stop() is called when a severe error is detected in
  25  * the function graph tracing. This function is called by the critical
  26  * paths of function graph to keep those paths from doing any more harm.
  27  */
  28 bool ftrace_graph_is_dead(void)
  29 {
  30         return kill_ftrace_graph;
  31 }
  32
  33 /**
  34  * ftrace_graph_stop - set to permanently disable function graph tracincg
  35  *
  36  * In case of an error int function graph tracing, this is called
  37  * to try to keep function graph tracing from causing any more harm.
  38  * Usually this is pretty severe and this is called to try to at least
  39  * get a warning out to the user.
  40  */
  41 void ftrace_graph_stop(void)
  42 {
  43         kill_ftrace_graph = true;
  44 }
  45
  46 /* When set, irq functions will be ignored */
  47 static int ftrace_graph_skip_irqs;
  48
  49 struct fgraph_cpu_data {
  50         pid_t           last_pid;
  51         int             depth;
  52         int             depth_irq;
  53         int             ignore;
  54         unsigned long   enter_funcs[FTRACE_RETFUNC_DEPTH];
  55 };
  56
  57 struct fgraph_data {
  58         struct fgraph_cpu_data __percpu *cpu_data;
  59
  60         /* Place to preserve last processed entry. */
  61         struct ftrace_graph_ent_entry   ent;
  62         struct ftrace_graph_ret_entry   ret;
  63         int                             failed;
  64         int                             cpu;
  65 };
  66
  67 #define TRACE_GRAPH_INDENT      2
  68
  69 unsigned int fgraph_max_depth;
  70
  71 static struct tracer_opt trace_opts[] = {
  72         /* Display overruns? (for self-debug purpose) */
  73         { TRACER_OPT(funcgraph-overrun, TRACE_GRAPH_PRINT_OVERRUN) },
  74         /* Display CPU ? */
  75         { TRACER_OPT(funcgraph-cpu, TRACE_GRAPH_PRINT_CPU) },
  76         /* Display Overhead ? */
  77         { TRACER_OPT(funcgraph-overhead, TRACE_GRAPH_PRINT_OVERHEAD) },
  78         /* Display proc name/pid */
  79         { TRACER_OPT(funcgraph-proc, TRACE_GRAPH_PRINT_PROC) },
  80         /* Display duration of execution */
  81         { TRACER_OPT(funcgraph-duration, TRACE_GRAPH_PRINT_DURATION) },
  82         /* Display absolute time of an entry */
  83         { TRACER_OPT(funcgraph-abstime, TRACE_GRAPH_PRINT_ABS_TIME) },
  84         /* Display interrupts */
  85         { TRACER_OPT(funcgraph-irqs, TRACE_GRAPH_PRINT_IRQS) },
  86         /* Display function name after trailing } */
  87         { TRACER_OPT(funcgraph-tail, TRACE_GRAPH_PRINT_TAIL) },
  88         /* Include sleep time (scheduled out) between entry and return */
  89         { TRACER_OPT(sleep-time, TRACE_GRAPH_SLEEP_TIME) },
  90         /* Include time within nested functions */
  91         { TRACER_OPT(graph-time, TRACE_GRAPH_GRAPH_TIME) },
  92         { } /* Empty entry */
  93 };
  94
  95 static struct tracer_flags tracer_flags = {
  96         /* Don't display overruns, proc, or tail by default */
  97         .val = TRACE_GRAPH_PRINT_CPU | TRACE_GRAPH_PRINT_OVERHEAD |
  98                TRACE_GRAPH_PRINT_DURATION | TRACE_GRAPH_PRINT_IRQS |
  99                TRACE_GRAPH_SLEEP_TIME | TRACE_GRAPH_GRAPH_TIME,
 100         .opts = trace_opts
 101 };
 102
 103 static struct trace_array *graph_array;
 104
 105 /*
 106  * DURATION column is being also used to display IRQ signs,
 107  * following values are used by print_graph_irq and others
 108  * to fill in space into DURATION column.
 109  */
 110 enum {
 111         FLAGS_FILL_FULL  = 1 << TRACE_GRAPH_PRINT_FILL_SHIFT,
 112         FLAGS_FILL_START = 2 << TRACE_GRAPH_PRINT_FILL_SHIFT,
 113         FLAGS_FILL_END   = 3 << TRACE_GRAPH_PRINT_FILL_SHIFT,
 114 };
 115
 116 static void
 117 print_graph_duration(struct trace_array *tr, unsigned long long duration,
 118                      struct trace_seq *s, u32 flags);
 119
 120 /* Add a function return address to the trace stack on thread info.*/
 121 static int
 122 ftrace_push_return_trace(unsigned long ret, unsigned long func,
 123                          unsigned long frame_pointer, unsigned long *retp)
 124 {
 125         unsigned long long calltime;
 126         int index;
 127
 128         if (unlikely(ftrace_graph_is_dead()))
 129                 return -EBUSY;
 130
 131         if (!current->ret_stack)
 132                 return -EBUSY;
 133
 134         /*
 135          * We must make sure the ret_stack is tested before we read
 136          * anything else.
 137          */
 138         smp_rmb();
 139
 140         /* The return trace stack is full */
 141         if (current->curr_ret_stack == FTRACE_RETFUNC_DEPTH - 1) {
 142                 atomic_inc(&current->trace_overrun);
 143                 return -EBUSY;
 144         }
 145
 146         /*
 147          * The curr_ret_stack is an index to ftrace return stack of
 148          * current task.  Its value should be in [0, FTRACE_RETFUNC_
 149          * DEPTH) when the function graph tracer is used.  To support
 150          * filtering out specific functions, it makes the index
 151          * negative by subtracting huge value (FTRACE_NOTRACE_DEPTH)
 152          * so when it sees a negative index the ftrace will ignore
 153          * the record.  And the index gets recovered when returning
 154          * from the filtered function by adding the FTRACE_NOTRACE_
 155          * DEPTH and then it'll continue to record functions normally.
 156          *
 157          * The curr_ret_stack is initialized to -1 and get increased
 158          * in this function.  So it can be less than -1 only if it was
 159          * filtered out via ftrace_graph_notrace_addr() which can be
 160          * set from set_graph_notrace file in tracefs by user.
 161          */
 162         if (current->curr_ret_stack < -1)
 163                 return -EBUSY;
 164
 165         calltime = trace_clock_local();
 166
 167         index = ++current->curr_ret_stack;
 168         if (ftrace_graph_notrace_addr(func))
 169                 current->curr_ret_stack -= FTRACE_NOTRACE_DEPTH;
 170         barrier();
 171         current->ret_stack[index].ret = ret;
 172         current->ret_stack[index].func = func;
 173         current->ret_stack[index].calltime = calltime;
 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         return 0;
 181 }
 182
 183 int function_graph_enter(unsigned long ret, unsigned long func,
 184                          unsigned long frame_pointer, unsigned long *retp)
 185 {
 186         struct ftrace_graph_ent trace;
 187
 188         trace.func = func;
 189         trace.depth = ++current->curr_ret_depth;
 190
 191         if (ftrace_push_return_trace(ret, func,
 192                                      frame_pointer, retp))
 193                 goto out;
 194
 195         /* Only trace if the calling function expects to */
 196         if (!ftrace_graph_entry(&trace))
 197                 goto out_ret;
 198
 199         return 0;
 200  out_ret:
 201         current->curr_ret_stack--;
 202  out:
 203         current->curr_ret_depth--;
 204         return -EBUSY;
 205 }
 206
 207 /* Retrieve a function return address to the trace stack on thread info.*/
 208 static void
 209 ftrace_pop_return_trace(struct ftrace_graph_ret *trace, unsigned long *ret,
 210                         unsigned long frame_pointer)
 211 {
 212         int index;
 213
 214         index = current->curr_ret_stack;
 215
 216         /*
 217          * A negative index here means that it's just returned from a
 218          * notrace'd function.  Recover index to get an original
 219          * return address.  See ftrace_push_return_trace().
 220          *
 221          * TODO: Need to check whether the stack gets corrupted.
 222          */
 223         if (index < 0)
 224                 index += FTRACE_NOTRACE_DEPTH;
 225
 226         if (unlikely(index < 0 || index >= FTRACE_RETFUNC_DEPTH)) {
 227                 ftrace_graph_stop();
 228                 WARN_ON(1);
 229                 /* Might as well panic, otherwise we have no where to go */
 230                 *ret = (unsigned long)panic;
 231                 return;
 232         }
 233
 234 #ifdef HAVE_FUNCTION_GRAPH_FP_TEST
 235         /*
 236          * The arch may choose to record the frame pointer used
 237          * and check it here to make sure that it is what we expect it
 238          * to be. If gcc does not set the place holder of the return
 239          * address in the frame pointer, and does a copy instead, then
 240          * the function graph trace will fail. This test detects this
 241          * case.
 242          *
 243          * Currently, x86_32 with optimize for size (-Os) makes the latest
 244          * gcc do the above.
 245          *
 246          * Note, -mfentry does not use frame pointers, and this test
 247          *  is not needed if CC_USING_FENTRY is set.
 248          */
 249         if (unlikely(current->ret_stack[index].fp != frame_pointer)) {
 250                 ftrace_graph_stop();
 251                 WARN(1, "Bad frame pointer: expected %lx, received %lx\n"
 252                      "  from func %ps return to %lx\n",
 253                      current->ret_stack[index].fp,
 254                      frame_pointer,
 255                      (void *)current->ret_stack[index].func,
 256                      current->ret_stack[index].ret);
 257                 *ret = (unsigned long)panic;
 258                 return;
 259         }
 260 #endif
 261
 262         *ret = current->ret_stack[index].ret;
 263         trace->func = current->ret_stack[index].func;
 264         trace->calltime = current->ret_stack[index].calltime;
 265         trace->overrun = atomic_read(&current->trace_overrun);
 266         trace->depth = current->curr_ret_depth--;
 267         /*
 268          * We still want to trace interrupts coming in if
 269          * max_depth is set to 1. Make sure the decrement is
 270          * seen before ftrace_graph_return.
 271          */
 272         barrier();
 273 }
 274
 275 /*
 276  * Send the trace to the ring-buffer.
 277  * @return the original return address.
 278  */
 279 unsigned long ftrace_return_to_handler(unsigned long frame_pointer)
 280 {
 281         struct ftrace_graph_ret trace;
 282         unsigned long ret;
 283
 284         ftrace_pop_return_trace(&trace, &ret, frame_pointer);
 285         trace.rettime = trace_clock_local();
 286         ftrace_graph_return(&trace);
 287         /*
 288          * The ftrace_graph_return() may still access the current
 289          * ret_stack structure, we need to make sure the update of
 290          * curr_ret_stack is after that.
 291          */
 292         barrier();
 293         current->curr_ret_stack--;
 294         /*
 295          * The curr_ret_stack can be less than -1 only if it was
 296          * filtered out and it's about to return from the function.
 297          * Recover the index and continue to trace normal functions.
 298          */
 299         if (current->curr_ret_stack < -1) {
 300                 current->curr_ret_stack += FTRACE_NOTRACE_DEPTH;
 301                 return ret;
 302         }
 303
 304         if (unlikely(!ret)) {
 305                 ftrace_graph_stop();
 306                 WARN_ON(1);
 307                 /* Might as well panic. What else to do? */
 308                 ret = (unsigned long)panic;
 309         }
 310
 311         return ret;
 312 }
 313
 314 /**
 315  * ftrace_graph_ret_addr - convert a potentially modified stack return address
 316  *                         to its original value
 317  *
 318  * This function can be called by stack unwinding code to convert a found stack
 319  * return address ('ret') to its original value, in case the function graph
 320  * tracer has modified it to be 'return_to_handler'.  If the address hasn't
 321  * been modified, the unchanged value of 'ret' is returned.
 322  *
 323  * 'idx' is a state variable which should be initialized by the caller to zero
 324  * before the first call.
 325  *
 326  * 'retp' is a pointer to the return address on the stack.  It's ignored if
 327  * the arch doesn't have HAVE_FUNCTION_GRAPH_RET_ADDR_PTR defined.
 328  */
 329 #ifdef HAVE_FUNCTION_GRAPH_RET_ADDR_PTR
 330 unsigned long ftrace_graph_ret_addr(struct task_struct *task, int *idx,
 331                                     unsigned long ret, unsigned long *retp)
 332 {
 333         int index = task->curr_ret_stack;
 334         int i;
 335
 336         if (ret != (unsigned long)return_to_handler)
 337                 return ret;
 338
 339         if (index < -1)
 340                 index += FTRACE_NOTRACE_DEPTH;
 341
 342         if (index < 0)
 343                 return ret;
 344
 345         for (i = 0; i <= index; i++)
 346                 if (task->ret_stack[i].retp == retp)
 347                         return task->ret_stack[i].ret;
 348
 349         return ret;
 350 }
 351 #else /* !HAVE_FUNCTION_GRAPH_RET_ADDR_PTR */
 352 unsigned long ftrace_graph_ret_addr(struct task_struct *task, int *idx,
 353                                     unsigned long ret, unsigned long *retp)
 354 {
 355         int task_idx;
 356
 357         if (ret != (unsigned long)return_to_handler)
 358                 return ret;
 359
 360         task_idx = task->curr_ret_stack;
 361
 362         if (!task->ret_stack || task_idx < *idx)
 363                 return ret;
 364
 365         task_idx -= *idx;
 366         (*idx)++;
 367
 368         return task->ret_stack[task_idx].ret;
 369 }
 370 #endif /* HAVE_FUNCTION_GRAPH_RET_ADDR_PTR */
 371
 372 int __trace_graph_entry(struct trace_array *tr,
 373                                 struct ftrace_graph_ent *trace,
 374                                 unsigned long flags,
 375                                 int pc)
 376 {
 377         struct trace_event_call *call = &event_funcgraph_entry;
 378         struct ring_buffer_event *event;
 379         struct ring_buffer *buffer = tr->trace_buffer.buffer;
 380         struct ftrace_graph_ent_entry *entry;
 381
 382         event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_ENT,
 383                                           sizeof(*entry), flags, pc);
 384         if (!event)
 385                 return 0;
 386         entry   = ring_buffer_event_data(event);
 387         entry->graph_ent                        = *trace;
 388         if (!call_filter_check_discard(call, entry, buffer, event))
 389                 trace_buffer_unlock_commit_nostack(buffer, event);
 390
 391         return 1;
 392 }
 393
 394 static inline int ftrace_graph_ignore_irqs(void)
 395 {
 396         if (!ftrace_graph_skip_irqs || trace_recursion_test(TRACE_IRQ_BIT))
 397                 return 0;
 398
 399         return in_irq();
 400 }
 401
 402 int trace_graph_entry(struct ftrace_graph_ent *trace)
 403 {
 404         struct trace_array *tr = graph_array;
 405         struct trace_array_cpu *data;
 406         unsigned long flags;
 407         long disabled;
 408         int ret;
 409         int cpu;
 410         int pc;
 411
 412         if (!ftrace_trace_task(tr))
 413                 return 0;
 414
 415         if (ftrace_graph_ignore_func(trace))
 416                 return 0;
 417
 418         if (ftrace_graph_ignore_irqs())
 419                 return 0;
 420
 421         /*
 422          * Do not trace a function if it's filtered by set_graph_notrace.
 423          * Make the index of ret stack negative to indicate that it should
 424          * ignore further functions.  But it needs its own ret stack entry
 425          * to recover the original index in order to continue tracing after
 426          * returning from the function.
 427          */
 428         if (ftrace_graph_notrace_addr(trace->func))
 429                 return 1;
 430
 431         /*
 432          * Stop here if tracing_threshold is set. We only write function return
 433          * events to the ring buffer.
 434          */
 435         if (tracing_thresh)
 436                 return 1;
 437
 438         local_irq_save(flags);
 439         cpu = raw_smp_processor_id();
 440         data = per_cpu_ptr(tr->trace_buffer.data, cpu);
 441         disabled = atomic_inc_return(&data->disabled);
 442         if (likely(disabled == 1)) {
 443                 pc = preempt_count();
 444                 ret = __trace_graph_entry(tr, trace, flags, pc);
 445         } else {
 446                 ret = 0;
 447         }
 448
 449         atomic_dec(&data->disabled);
 450         local_irq_restore(flags);
 451
 452         return ret;
 453 }
 454
 455 static void
 456 __trace_graph_function(struct trace_array *tr,
 457                 unsigned long ip, unsigned long flags, int pc)
 458 {
 459         u64 time = trace_clock_local();
 460         struct ftrace_graph_ent ent = {
 461                 .func  = ip,
 462                 .depth = 0,
 463         };
 464         struct ftrace_graph_ret ret = {
 465                 .func     = ip,
 466                 .depth    = 0,
 467                 .calltime = time,
 468                 .rettime  = time,
 469         };
 470
 471         __trace_graph_entry(tr, &ent, flags, pc);
 472         __trace_graph_return(tr, &ret, flags, pc);
 473 }
 474
 475 void
 476 trace_graph_function(struct trace_array *tr,
 477                 unsigned long ip, unsigned long parent_ip,
 478                 unsigned long flags, int pc)
 479 {
 480         __trace_graph_function(tr, ip, flags, pc);
 481 }
 482
 483 void __trace_graph_return(struct trace_array *tr,
 484                                 struct ftrace_graph_ret *trace,
 485                                 unsigned long flags,
 486                                 int pc)
 487 {
 488         struct trace_event_call *call = &event_funcgraph_exit;
 489         struct ring_buffer_event *event;
 490         struct ring_buffer *buffer = tr->trace_buffer.buffer;
 491         struct ftrace_graph_ret_entry *entry;
 492
 493         event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_RET,
 494                                           sizeof(*entry), flags, pc);
 495         if (!event)
 496                 return;
 497         entry   = ring_buffer_event_data(event);
 498         entry->ret                              = *trace;
 499         if (!call_filter_check_discard(call, entry, buffer, event))
 500                 trace_buffer_unlock_commit_nostack(buffer, event);
 501 }
 502
 503 void trace_graph_return(struct ftrace_graph_ret *trace)
 504 {
 505         struct trace_array *tr = graph_array;
 506         struct trace_array_cpu *data;
 507         unsigned long flags;
 508         long disabled;
 509         int cpu;
 510         int pc;
 511
 512         ftrace_graph_addr_finish(trace);
 513
 514         local_irq_save(flags);
 515         cpu = raw_smp_processor_id();
 516         data = per_cpu_ptr(tr->trace_buffer.data, cpu);
 517         disabled = atomic_inc_return(&data->disabled);
 518         if (likely(disabled == 1)) {
 519                 pc = preempt_count();
 520                 __trace_graph_return(tr, trace, flags, pc);
 521         }
 522         atomic_dec(&data->disabled);
 523         local_irq_restore(flags);
 524 }
 525
 526 void set_graph_array(struct trace_array *tr)
 527 {
 528         graph_array = tr;
 529
 530         /* Make graph_array visible before we start tracing */
 531
 532         smp_mb();
 533 }
 534
 535 static void trace_graph_thresh_return(struct ftrace_graph_ret *trace)
 536 {
 537         ftrace_graph_addr_finish(trace);
 538
 539         if (tracing_thresh &&
 540             (trace->rettime - trace->calltime < tracing_thresh))
 541                 return;
 542         else
 543                 trace_graph_return(trace);
 544 }
 545
 546 static int graph_trace_init(struct trace_array *tr)
 547 {
 548         int ret;
 549
 550         set_graph_array(tr);
 551         if (tracing_thresh)
 552                 ret = register_ftrace_graph(&trace_graph_thresh_return,
 553                                             &trace_graph_entry);
 554         else
 555                 ret = register_ftrace_graph(&trace_graph_return,
 556                                             &trace_graph_entry);
 557         if (ret)
 558                 return ret;
 559         tracing_start_cmdline_record();
 560
 561         return 0;
 562 }
 563
 564 static void graph_trace_reset(struct trace_array *tr)
 565 {
 566         tracing_stop_cmdline_record();
 567         unregister_ftrace_graph();
 568 }
 569
 570 static int graph_trace_update_thresh(struct trace_array *tr)
 571 {
 572         graph_trace_reset(tr);
 573         return graph_trace_init(tr);
 574 }
 575
 576 static int max_bytes_for_cpu;
 577
 578 static void print_graph_cpu(struct trace_seq *s, int cpu)
 579 {
 580         /*
 581          * Start with a space character - to make it stand out
 582          * to the right a bit when trace output is pasted into
 583          * email:
 584          */
 585         trace_seq_printf(s, " %*d) ", max_bytes_for_cpu, cpu);
 586 }
 587
 588 #define TRACE_GRAPH_PROCINFO_LENGTH     14
 589
 590 static void print_graph_proc(struct trace_seq *s, pid_t pid)
 591 {
 592         char comm[TASK_COMM_LEN];
 593         /* sign + log10(MAX_INT) + '\0' */
 594         char pid_str[11];
 595         int spaces = 0;
 596         int len;
 597         int i;
 598
 599         trace_find_cmdline(pid, comm);
 600         comm[7] = '\0';
 601         sprintf(pid_str, "%d", pid);
 602
 603         /* 1 stands for the "-" character */
 604         len = strlen(comm) + strlen(pid_str) + 1;
 605
 606         if (len < TRACE_GRAPH_PROCINFO_LENGTH)
 607                 spaces = TRACE_GRAPH_PROCINFO_LENGTH - len;
 608
 609         /* First spaces to align center */
 610         for (i = 0; i < spaces / 2; i++)
 611                 trace_seq_putc(s, ' ');
 612
 613         trace_seq_printf(s, "%s-%s", comm, pid_str);
 614
 615         /* Last spaces to align center */
 616         for (i = 0; i < spaces - (spaces / 2); i++)
 617                 trace_seq_putc(s, ' ');
 618 }
 619
 620
 621 static void print_graph_lat_fmt(struct trace_seq *s, struct trace_entry *entry)
 622 {
 623         trace_seq_putc(s, ' ');
 624         trace_print_lat_fmt(s, entry);
 625 }
 626
 627 /* If the pid changed since the last trace, output this event */
 628 static void
 629 verif_pid(struct trace_seq *s, pid_t pid, int cpu, struct fgraph_data *data)
 630 {
 631         pid_t prev_pid;
 632         pid_t *last_pid;
 633
 634         if (!data)
 635                 return;
 636
 637         last_pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
 638
 639         if (*last_pid == pid)
 640                 return;
 641
 642         prev_pid = *last_pid;
 643         *last_pid = pid;
 644
 645         if (prev_pid == -1)
 646                 return;
 647 /*
 648  * Context-switch trace line:
 649
 650  ------------------------------------------
 651  | 1)  migration/0--1  =>  sshd-1755
 652  ------------------------------------------
 653
 654  */
 655         trace_seq_puts(s, " ------------------------------------------\n");
 656         print_graph_cpu(s, cpu);
 657         print_graph_proc(s, prev_pid);
 658         trace_seq_puts(s, " => ");
 659         print_graph_proc(s, pid);
 660         trace_seq_puts(s, "\n ------------------------------------------\n\n");
 661 }
 662
 663 static struct ftrace_graph_ret_entry *
 664 get_return_for_leaf(struct trace_iterator *iter,
 665                 struct ftrace_graph_ent_entry *curr)
 666 {
 667         struct fgraph_data *data = iter->private;
 668         struct ring_buffer_iter *ring_iter = NULL;
 669         struct ring_buffer_event *event;
 670         struct ftrace_graph_ret_entry *next;
 671
 672         /*
 673          * If the previous output failed to write to the seq buffer,
 674          * then we just reuse the data from before.
 675          */
 676         if (data && data->failed) {
 677                 curr = &data->ent;
 678                 next = &data->ret;
 679         } else {
 680
 681                 ring_iter = trace_buffer_iter(iter, iter->cpu);
 682
 683                 /* First peek to compare current entry and the next one */
 684                 if (ring_iter)
 685                         event = ring_buffer_iter_peek(ring_iter, NULL);
 686                 else {
 687                         /*
 688                          * We need to consume the current entry to see
 689                          * the next one.
 690                          */
 691                         ring_buffer_consume(iter->trace_buffer->buffer, iter->cpu,
 692                                             NULL, NULL);
 693                         event = ring_buffer_peek(iter->trace_buffer->buffer, iter->cpu,
 694                                                  NULL, NULL);
 695                 }
 696
 697                 if (!event)
 698                         return NULL;
 699
 700                 next = ring_buffer_event_data(event);
 701
 702                 if (data) {
 703                         /*
 704                          * Save current and next entries for later reference
 705                          * if the output fails.
 706                          */
 707                         data->ent = *curr;
 708                         /*
 709                          * If the next event is not a return type, then
 710                          * we only care about what type it is. Otherwise we can
 711                          * safely copy the entire event.
 712                          */
 713                         if (next->ent.type == TRACE_GRAPH_RET)
 714                                 data->ret = *next;
 715                         else
 716                                 data->ret.ent.type = next->ent.type;
 717                 }
 718         }
 719
 720         if (next->ent.type != TRACE_GRAPH_RET)
 721                 return NULL;
 722
 723         if (curr->ent.pid != next->ent.pid ||
 724                         curr->graph_ent.func != next->ret.func)
 725                 return NULL;
 726
 727         /* this is a leaf, now advance the iterator */
 728         if (ring_iter)
 729                 ring_buffer_read(ring_iter, NULL);
 730
 731         return next;
 732 }
 733
 734 static void print_graph_abs_time(u64 t, struct trace_seq *s)
 735 {
 736         unsigned long usecs_rem;
 737
 738         usecs_rem = do_div(t, NSEC_PER_SEC);
 739         usecs_rem /= 1000;
 740
 741         trace_seq_printf(s, "%5lu.%06lu |  ",
 742                          (unsigned long)t, usecs_rem);
 743 }
 744
 745 static void
 746 print_graph_irq(struct trace_iterator *iter, unsigned long addr,
 747                 enum trace_type type, int cpu, pid_t pid, u32 flags)
 748 {
 749         struct trace_array *tr = iter->tr;
 750         struct trace_seq *s = &iter->seq;
 751         struct trace_entry *ent = iter->ent;
 752
 753         if (addr < (unsigned long)__irqentry_text_start ||
 754                 addr >= (unsigned long)__irqentry_text_end)
 755                 return;
 756
 757         if (tr->trace_flags & TRACE_ITER_CONTEXT_INFO) {
 758                 /* Absolute time */
 759                 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
 760                         print_graph_abs_time(iter->ts, s);
 761
 762                 /* Cpu */
 763                 if (flags & TRACE_GRAPH_PRINT_CPU)
 764                         print_graph_cpu(s, cpu);
 765
 766                 /* Proc */
 767                 if (flags & TRACE_GRAPH_PRINT_PROC) {
 768                         print_graph_proc(s, pid);
 769                         trace_seq_puts(s, " | ");
 770                 }
 771
 772                 /* Latency format */
 773                 if (tr->trace_flags & TRACE_ITER_LATENCY_FMT)
 774                         print_graph_lat_fmt(s, ent);
 775         }
 776
 777         /* No overhead */
 778         print_graph_duration(tr, 0, s, flags | FLAGS_FILL_START);
 779
 780         if (type == TRACE_GRAPH_ENT)
 781                 trace_seq_puts(s, "==========>");
 782         else
 783                 trace_seq_puts(s, "<==========");
 784
 785         print_graph_duration(tr, 0, s, flags | FLAGS_FILL_END);
 786         trace_seq_putc(s, '\n');
 787 }
 788
 789 void
 790 trace_print_graph_duration(unsigned long long duration, struct trace_seq *s)
 791 {
 792         unsigned long nsecs_rem = do_div(duration, 1000);
 793         /* log10(ULONG_MAX) + '\0' */
 794         char usecs_str[21];
 795         char nsecs_str[5];
 796         int len;
 797         int i;
 798
 799         sprintf(usecs_str, "%lu", (unsigned long) duration);
 800
 801         /* Print msecs */
 802         trace_seq_printf(s, "%s", usecs_str);
 803
 804         len = strlen(usecs_str);
 805
 806         /* Print nsecs (we don't want to exceed 7 numbers) */
 807         if (len < 7) {
 808                 size_t slen = min_t(size_t, sizeof(nsecs_str), 8UL - len);
 809
 810                 snprintf(nsecs_str, slen, "%03lu", nsecs_rem);
 811                 trace_seq_printf(s, ".%s", nsecs_str);
 812                 len += strlen(nsecs_str) + 1;
 813         }
 814
 815         trace_seq_puts(s, " us ");
 816
 817         /* Print remaining spaces to fit the row's width */
 818         for (i = len; i < 8; i++)
 819                 trace_seq_putc(s, ' ');
 820 }
 821
 822 static void
 823 print_graph_duration(struct trace_array *tr, unsigned long long duration,
 824                      struct trace_seq *s, u32 flags)
 825 {
 826         if (!(flags & TRACE_GRAPH_PRINT_DURATION) ||
 827             !(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
 828                 return;
 829
 830         /* No real adata, just filling the column with spaces */
 831         switch (flags & TRACE_GRAPH_PRINT_FILL_MASK) {
 832         case FLAGS_FILL_FULL:
 833                 trace_seq_puts(s, "              |  ");
 834                 return;
 835         case FLAGS_FILL_START:
 836                 trace_seq_puts(s, "  ");
 837                 return;
 838         case FLAGS_FILL_END:
 839                 trace_seq_puts(s, " |");
 840                 return;
 841         }
 842
 843         /* Signal a overhead of time execution to the output */
 844         if (flags & TRACE_GRAPH_PRINT_OVERHEAD)
 845                 trace_seq_printf(s, "%c ", trace_find_mark(duration));
 846         else
 847                 trace_seq_puts(s, "  ");
 848
 849         trace_print_graph_duration(duration, s);
 850         trace_seq_puts(s, "|  ");
 851 }
 852
 853 /* Case of a leaf function on its call entry */
 854 static enum print_line_t
 855 print_graph_entry_leaf(struct trace_iterator *iter,
 856                 struct ftrace_graph_ent_entry *entry,
 857                 struct ftrace_graph_ret_entry *ret_entry,
 858                 struct trace_seq *s, u32 flags)
 859 {
 860         struct fgraph_data *data = iter->private;
 861         struct trace_array *tr = iter->tr;
 862         struct ftrace_graph_ret *graph_ret;
 863         struct ftrace_graph_ent *call;
 864         unsigned long long duration;
 865         int cpu = iter->cpu;
 866         int i;
 867
 868         graph_ret = &ret_entry->ret;
 869         call = &entry->graph_ent;
 870         duration = graph_ret->rettime - graph_ret->calltime;
 871
 872         if (data) {
 873                 struct fgraph_cpu_data *cpu_data;
 874
 875                 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
 876
 877                 /* If a graph tracer ignored set_graph_notrace */
 878                 if (call->depth < -1)
 879                         call->depth += FTRACE_NOTRACE_DEPTH;
 880
 881                 /*
 882                  * Comments display at + 1 to depth. Since
 883                  * this is a leaf function, keep the comments
 884                  * equal to this depth.
 885                  */
 886                 cpu_data->depth = call->depth - 1;
 887
 888                 /* No need to keep this function around for this depth */
 889                 if (call->depth < FTRACE_RETFUNC_DEPTH &&
 890                     !WARN_ON_ONCE(call->depth < 0))
 891                         cpu_data->enter_funcs[call->depth] = 0;
 892         }
 893
 894         /* Overhead and duration */
 895         print_graph_duration(tr, duration, s, flags);
 896
 897         /* Function */
 898         for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
 899                 trace_seq_putc(s, ' ');
 900
 901         trace_seq_printf(s, "%ps();\n", (void *)call->func);
 902
 903         print_graph_irq(iter, graph_ret->func, TRACE_GRAPH_RET,
 904                         cpu, iter->ent->pid, flags);
 905
 906         return trace_handle_return(s);
 907 }
 908
 909 static enum print_line_t
 910 print_graph_entry_nested(struct trace_iterator *iter,
 911                          struct ftrace_graph_ent_entry *entry,
 912                          struct trace_seq *s, int cpu, u32 flags)
 913 {
 914         struct ftrace_graph_ent *call = &entry->graph_ent;
 915         struct fgraph_data *data = iter->private;
 916         struct trace_array *tr = iter->tr;
 917         int i;
 918
 919         if (data) {
 920                 struct fgraph_cpu_data *cpu_data;
 921                 int cpu = iter->cpu;
 922
 923                 /* If a graph tracer ignored set_graph_notrace */
 924                 if (call->depth < -1)
 925                         call->depth += FTRACE_NOTRACE_DEPTH;
 926
 927                 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
 928                 cpu_data->depth = call->depth;
 929
 930                 /* Save this function pointer to see if the exit matches */
 931                 if (call->depth < FTRACE_RETFUNC_DEPTH &&
 932                     !WARN_ON_ONCE(call->depth < 0))
 933                         cpu_data->enter_funcs[call->depth] = call->func;
 934         }
 935
 936         /* No time */
 937         print_graph_duration(tr, 0, s, flags | FLAGS_FILL_FULL);
 938
 939         /* Function */
 940         for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
 941                 trace_seq_putc(s, ' ');
 942
 943         trace_seq_printf(s, "%ps() {\n", (void *)call->func);
 944
 945         if (trace_seq_has_overflowed(s))
 946                 return TRACE_TYPE_PARTIAL_LINE;
 947
 948         /*
 949          * we already consumed the current entry to check the next one
 950          * and see if this is a leaf.
 951          */
 952         return TRACE_TYPE_NO_CONSUME;
 953 }
 954
 955 static void
 956 print_graph_prologue(struct trace_iterator *iter, struct trace_seq *s,
 957                      int type, unsigned long addr, u32 flags)
 958 {
 959         struct fgraph_data *data = iter->private;
 960         struct trace_entry *ent = iter->ent;
 961         struct trace_array *tr = iter->tr;
 962         int cpu = iter->cpu;
 963
 964         /* Pid */
 965         verif_pid(s, ent->pid, cpu, data);
 966
 967         if (type)
 968                 /* Interrupt */
 969                 print_graph_irq(iter, addr, type, cpu, ent->pid, flags);
 970
 971         if (!(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
 972                 return;
 973
 974         /* Absolute time */
 975         if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
 976                 print_graph_abs_time(iter->ts, s);
 977
 978         /* Cpu */
 979         if (flags & TRACE_GRAPH_PRINT_CPU)
 980                 print_graph_cpu(s, cpu);
 981
 982         /* Proc */
 983         if (flags & TRACE_GRAPH_PRINT_PROC) {
 984                 print_graph_proc(s, ent->pid);
 985                 trace_seq_puts(s, " | ");
 986         }
 987
 988         /* Latency format */
 989         if (tr->trace_flags & TRACE_ITER_LATENCY_FMT)
 990                 print_graph_lat_fmt(s, ent);
 991
 992         return;
 993 }
 994
 995 /*
 996  * Entry check for irq code
 997  *
 998  * returns 1 if
 999  *  - we are inside irq code
1000  *  - we just entered irq code
1001  *
1002  * retunns 0 if
1003  *  - funcgraph-interrupts option is set
1004  *  - we are not inside irq code
1005  */
1006 static int
1007 check_irq_entry(struct trace_iterator *iter, u32 flags,
1008                 unsigned long addr, int depth)
1009 {
1010         int cpu = iter->cpu;
1011         int *depth_irq;
1012         struct fgraph_data *data = iter->private;
1013
1014         /*
1015          * If we are either displaying irqs, or we got called as
1016          * a graph event and private data does not exist,
1017          * then we bypass the irq check.
1018          */
1019         if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
1020             (!data))
1021                 return 0;
1022
1023         depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
1024
1025         /*
1026          * We are inside the irq code
1027          */
1028         if (*depth_irq >= 0)
1029                 return 1;
1030
1031         if ((addr < (unsigned long)__irqentry_text_start) ||
1032             (addr >= (unsigned long)__irqentry_text_end))
1033                 return 0;
1034
1035         /*
1036          * We are entering irq code.
1037          */
1038         *depth_irq = depth;
1039         return 1;
1040 }
1041
1042 /*
1043  * Return check for irq code
1044  *
1045  * returns 1 if
1046  *  - we are inside irq code
1047  *  - we just left irq code
1048  *
1049  * returns 0 if
1050  *  - funcgraph-interrupts option is set
1051  *  - we are not inside irq code
1052  */
1053 static int
1054 check_irq_return(struct trace_iterator *iter, u32 flags, int depth)
1055 {
1056         int cpu = iter->cpu;
1057         int *depth_irq;
1058         struct fgraph_data *data = iter->private;
1059
1060         /*
1061          * If we are either displaying irqs, or we got called as
1062          * a graph event and private data does not exist,
1063          * then we bypass the irq check.
1064          */
1065         if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
1066             (!data))
1067                 return 0;
1068
1069         depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
1070
1071         /*
1072          * We are not inside the irq code.
1073          */
1074         if (*depth_irq == -1)
1075                 return 0;
1076
1077         /*
1078          * We are inside the irq code, and this is returning entry.
1079          * Let's not trace it and clear the entry depth, since
1080          * we are out of irq code.
1081          *
1082          * This condition ensures that we 'leave the irq code' once
1083          * we are out of the entry depth. Thus protecting us from
1084          * the RETURN entry loss.
1085          */
1086         if (*depth_irq >= depth) {
1087                 *depth_irq = -1;
1088                 return 1;
1089         }
1090
1091         /*
1092          * We are inside the irq code, and this is not the entry.
1093          */
1094         return 1;
1095 }
1096
1097 static enum print_line_t
1098 print_graph_entry(struct ftrace_graph_ent_entry *field, struct trace_seq *s,
1099                         struct trace_iterator *iter, u32 flags)
1100 {
1101         struct fgraph_data *data = iter->private;
1102         struct ftrace_graph_ent *call = &field->graph_ent;
1103         struct ftrace_graph_ret_entry *leaf_ret;
1104         static enum print_line_t ret;
1105         int cpu = iter->cpu;
1106
1107         if (check_irq_entry(iter, flags, call->func, call->depth))
1108                 return TRACE_TYPE_HANDLED;
1109
1110         print_graph_prologue(iter, s, TRACE_GRAPH_ENT, call->func, flags);
1111
1112         leaf_ret = get_return_for_leaf(iter, field);
1113         if (leaf_ret)
1114                 ret = print_graph_entry_leaf(iter, field, leaf_ret, s, flags);
1115         else
1116                 ret = print_graph_entry_nested(iter, field, s, cpu, flags);
1117
1118         if (data) {
1119                 /*
1120                  * If we failed to write our output, then we need to make
1121                  * note of it. Because we already consumed our entry.
1122                  */
1123                 if (s->full) {
1124                         data->failed = 1;
1125                         data->cpu = cpu;
1126                 } else
1127                         data->failed = 0;
1128         }
1129
1130         return ret;
1131 }
1132
1133 static enum print_line_t
1134 print_graph_return(struct ftrace_graph_ret *trace, struct trace_seq *s,
1135                    struct trace_entry *ent, struct trace_iterator *iter,
1136                    u32 flags)
1137 {
1138         unsigned long long duration = trace->rettime - trace->calltime;
1139         struct fgraph_data *data = iter->private;
1140         struct trace_array *tr = iter->tr;
1141         pid_t pid = ent->pid;
1142         int cpu = iter->cpu;
1143         int func_match = 1;
1144         int i;
1145
1146         if (check_irq_return(iter, flags, trace->depth))
1147                 return TRACE_TYPE_HANDLED;
1148
1149         if (data) {
1150                 struct fgraph_cpu_data *cpu_data;
1151                 int cpu = iter->cpu;
1152
1153                 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
1154
1155                 /*
1156                  * Comments display at + 1 to depth. This is the
1157                  * return from a function, we now want the comments
1158                  * to display at the same level of the bracket.
1159                  */
1160                 cpu_data->depth = trace->depth - 1;
1161
1162                 if (trace->depth < FTRACE_RETFUNC_DEPTH &&
1163                     !WARN_ON_ONCE(trace->depth < 0)) {
1164                         if (cpu_data->enter_funcs[trace->depth] != trace->func)
1165                                 func_match = 0;
1166                         cpu_data->enter_funcs[trace->depth] = 0;
1167                 }
1168         }
1169
1170         print_graph_prologue(iter, s, 0, 0, flags);
1171
1172         /* Overhead and duration */
1173         print_graph_duration(tr, duration, s, flags);
1174
1175         /* Closing brace */
1176         for (i = 0; i < trace->depth * TRACE_GRAPH_INDENT; i++)
1177                 trace_seq_putc(s, ' ');
1178
1179         /*
1180          * If the return function does not have a matching entry,
1181          * then the entry was lost. Instead of just printing
1182          * the '}' and letting the user guess what function this
1183          * belongs to, write out the function name. Always do
1184          * that if the funcgraph-tail option is enabled.
1185          */
1186         if (func_match && !(flags & TRACE_GRAPH_PRINT_TAIL))
1187                 trace_seq_puts(s, "}\n");
1188         else
1189                 trace_seq_printf(s, "} /* %ps */\n", (void *)trace->func);
1190
1191         /* Overrun */
1192         if (flags & TRACE_GRAPH_PRINT_OVERRUN)
1193                 trace_seq_printf(s, " (Overruns: %lu)\n",
1194                                  trace->overrun);
1195
1196         print_graph_irq(iter, trace->func, TRACE_GRAPH_RET,
1197                         cpu, pid, flags);
1198
1199         return trace_handle_return(s);
1200 }
1201
1202 static enum print_line_t
1203 print_graph_comment(struct trace_seq *s, struct trace_entry *ent,
1204                     struct trace_iterator *iter, u32 flags)
1205 {
1206         struct trace_array *tr = iter->tr;
1207         unsigned long sym_flags = (tr->trace_flags & TRACE_ITER_SYM_MASK);
1208         struct fgraph_data *data = iter->private;
1209         struct trace_event *event;
1210         int depth = 0;
1211         int ret;
1212         int i;
1213
1214         if (data)
1215                 depth = per_cpu_ptr(data->cpu_data, iter->cpu)->depth;
1216
1217         print_graph_prologue(iter, s, 0, 0, flags);
1218
1219         /* No time */
1220         print_graph_duration(tr, 0, s, flags | FLAGS_FILL_FULL);
1221
1222         /* Indentation */
1223         if (depth > 0)
1224                 for (i = 0; i < (depth + 1) * TRACE_GRAPH_INDENT; i++)
1225                         trace_seq_putc(s, ' ');
1226
1227         /* The comment */
1228         trace_seq_puts(s, "/* ");
1229
1230         switch (iter->ent->type) {
1231         case TRACE_BPUTS:
1232                 ret = trace_print_bputs_msg_only(iter);
1233                 if (ret != TRACE_TYPE_HANDLED)
1234                         return ret;
1235                 break;
1236         case TRACE_BPRINT:
1237                 ret = trace_print_bprintk_msg_only(iter);
1238                 if (ret != TRACE_TYPE_HANDLED)
1239                         return ret;
1240                 break;
1241         case TRACE_PRINT:
1242                 ret = trace_print_printk_msg_only(iter);
1243                 if (ret != TRACE_TYPE_HANDLED)
1244                         return ret;
1245                 break;
1246         default:
1247                 event = ftrace_find_event(ent->type);
1248                 if (!event)
1249                         return TRACE_TYPE_UNHANDLED;
1250
1251                 ret = event->funcs->trace(iter, sym_flags, event);
1252                 if (ret != TRACE_TYPE_HANDLED)
1253                         return ret;
1254         }
1255
1256         if (trace_seq_has_overflowed(s))
1257                 goto out;
1258
1259         /* Strip ending newline */
1260         if (s->buffer[s->seq.len - 1] == '\n') {
1261                 s->buffer[s->seq.len - 1] = '\0';
1262                 s->seq.len--;
1263         }
1264
1265         trace_seq_puts(s, " */\n");
1266  out:
1267         return trace_handle_return(s);
1268 }
1269
1270
1271 enum print_line_t
1272 print_graph_function_flags(struct trace_iterator *iter, u32 flags)
1273 {
1274         struct ftrace_graph_ent_entry *field;
1275         struct fgraph_data *data = iter->private;
1276         struct trace_entry *entry = iter->ent;
1277         struct trace_seq *s = &iter->seq;
1278         int cpu = iter->cpu;
1279         int ret;
1280
1281         if (data && per_cpu_ptr(data->cpu_data, cpu)->ignore) {
1282                 per_cpu_ptr(data->cpu_data, cpu)->ignore = 0;
1283                 return TRACE_TYPE_HANDLED;
1284         }
1285
1286         /*
1287          * If the last output failed, there's a possibility we need
1288          * to print out the missing entry which would never go out.
1289          */
1290         if (data && data->failed) {
1291                 field = &data->ent;
1292                 iter->cpu = data->cpu;
1293                 ret = print_graph_entry(field, s, iter, flags);
1294                 if (ret == TRACE_TYPE_HANDLED && iter->cpu != cpu) {
1295                         per_cpu_ptr(data->cpu_data, iter->cpu)->ignore = 1;
1296                         ret = TRACE_TYPE_NO_CONSUME;
1297                 }
1298                 iter->cpu = cpu;
1299                 return ret;
1300         }
1301
1302         switch (entry->type) {
1303         case TRACE_GRAPH_ENT: {
1304                 /*
1305                  * print_graph_entry() may consume the current event,
1306                  * thus @field may become invalid, so we need to save it.
1307                  * sizeof(struct ftrace_graph_ent_entry) is very small,
1308                  * it can be safely saved at the stack.
1309                  */
1310                 struct ftrace_graph_ent_entry saved;
1311                 trace_assign_type(field, entry);
1312                 saved = *field;
1313                 return print_graph_entry(&saved, s, iter, flags);
1314         }
1315         case TRACE_GRAPH_RET: {
1316                 struct ftrace_graph_ret_entry *field;
1317                 trace_assign_type(field, entry);
1318                 return print_graph_return(&field->ret, s, entry, iter, flags);
1319         }
1320         case TRACE_STACK:
1321         case TRACE_FN:
1322                 /* dont trace stack and functions as comments */
1323                 return TRACE_TYPE_UNHANDLED;
1324
1325         default:
1326                 return print_graph_comment(s, entry, iter, flags);
1327         }
1328
1329         return TRACE_TYPE_HANDLED;
1330 }
1331
1332 static enum print_line_t
1333 print_graph_function(struct trace_iterator *iter)
1334 {
1335         return print_graph_function_flags(iter, tracer_flags.val);
1336 }
1337
1338 static enum print_line_t
1339 print_graph_function_event(struct trace_iterator *iter, int flags,
1340                            struct trace_event *event)
1341 {
1342         return print_graph_function(iter);
1343 }
1344
1345 static void print_lat_header(struct seq_file *s, u32 flags)
1346 {
1347         static const char spaces[] = "                " /* 16 spaces */
1348                 "    "                                  /* 4 spaces */
1349                 "                 ";                    /* 17 spaces */
1350         int size = 0;
1351
1352         if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1353                 size += 16;
1354         if (flags & TRACE_GRAPH_PRINT_CPU)
1355                 size += 4;
1356         if (flags & TRACE_GRAPH_PRINT_PROC)
1357                 size += 17;
1358
1359         seq_printf(s, "#%.*s  _-----=> irqs-off        \n", size, spaces);
1360         seq_printf(s, "#%.*s / _----=> need-resched    \n", size, spaces);
1361         seq_printf(s, "#%.*s| / _---=> hardirq/softirq \n", size, spaces);
1362         seq_printf(s, "#%.*s|| / _--=> preempt-depth   \n", size, spaces);
1363         seq_printf(s, "#%.*s||| /                      \n", size, spaces);
1364 }
1365
1366 static void __print_graph_headers_flags(struct trace_array *tr,
1367                                         struct seq_file *s, u32 flags)
1368 {
1369         int lat = tr->trace_flags & TRACE_ITER_LATENCY_FMT;
1370
1371         if (lat)
1372                 print_lat_header(s, flags);
1373
1374         /* 1st line */
1375         seq_putc(s, '#');
1376         if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1377                 seq_puts(s, "     TIME       ");
1378         if (flags & TRACE_GRAPH_PRINT_CPU)
1379                 seq_puts(s, " CPU");
1380         if (flags & TRACE_GRAPH_PRINT_PROC)
1381                 seq_puts(s, "  TASK/PID       ");
1382         if (lat)
1383                 seq_puts(s, "||||");
1384         if (flags & TRACE_GRAPH_PRINT_DURATION)
1385                 seq_puts(s, "  DURATION   ");
1386         seq_puts(s, "               FUNCTION CALLS\n");
1387
1388         /* 2nd line */
1389         seq_putc(s, '#');
1390         if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1391                 seq_puts(s, "      |         ");
1392         if (flags & TRACE_GRAPH_PRINT_CPU)
1393                 seq_puts(s, " |  ");
1394         if (flags & TRACE_GRAPH_PRINT_PROC)
1395                 seq_puts(s, "   |    |        ");
1396         if (lat)
1397                 seq_puts(s, "||||");
1398         if (flags & TRACE_GRAPH_PRINT_DURATION)
1399                 seq_puts(s, "   |   |      ");
1400         seq_puts(s, "               |   |   |   |\n");
1401 }
1402
1403 static void print_graph_headers(struct seq_file *s)
1404 {
1405         print_graph_headers_flags(s, tracer_flags.val);
1406 }
1407
1408 void print_graph_headers_flags(struct seq_file *s, u32 flags)
1409 {
1410         struct trace_iterator *iter = s->private;
1411         struct trace_array *tr = iter->tr;
1412
1413         if (!(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
1414                 return;
1415
1416         if (tr->trace_flags & TRACE_ITER_LATENCY_FMT) {
1417                 /* print nothing if the buffers are empty */
1418                 if (trace_empty(iter))
1419                         return;
1420
1421                 print_trace_header(s, iter);
1422         }
1423
1424         __print_graph_headers_flags(tr, s, flags);
1425 }
1426
1427 void graph_trace_open(struct trace_iterator *iter)
1428 {
1429         /* pid and depth on the last trace processed */
1430         struct fgraph_data *data;
1431         gfp_t gfpflags;
1432         int cpu;
1433
1434         iter->private = NULL;
1435
1436         /* We can be called in atomic context via ftrace_dump() */
1437         gfpflags = (in_atomic() || irqs_disabled()) ? GFP_ATOMIC : GFP_KERNEL;
1438
1439         data = kzalloc(sizeof(*data), gfpflags);
1440         if (!data)
1441                 goto out_err;
1442
1443         data->cpu_data = alloc_percpu_gfp(struct fgraph_cpu_data, gfpflags);
1444         if (!data->cpu_data)
1445                 goto out_err_free;
1446
1447         for_each_possible_cpu(cpu) {
1448                 pid_t *pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
1449                 int *depth = &(per_cpu_ptr(data->cpu_data, cpu)->depth);
1450                 int *ignore = &(per_cpu_ptr(data->cpu_data, cpu)->ignore);
1451                 int *depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
1452
1453                 *pid = -1;
1454                 *depth = 0;
1455                 *ignore = 0;
1456                 *depth_irq = -1;
1457         }
1458
1459         iter->private = data;
1460
1461         return;
1462
1463  out_err_free:
1464         kfree(data);
1465  out_err:
1466         pr_warn("function graph tracer: not enough memory\n");
1467 }
1468
1469 void graph_trace_close(struct trace_iterator *iter)
1470 {
1471         struct fgraph_data *data = iter->private;
1472
1473         if (data) {
1474                 free_percpu(data->cpu_data);
1475                 kfree(data);
1476         }
1477 }
1478
1479 static int
1480 func_graph_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set)
1481 {
1482         if (bit == TRACE_GRAPH_PRINT_IRQS)
1483                 ftrace_graph_skip_irqs = !set;
1484
1485         if (bit == TRACE_GRAPH_SLEEP_TIME)
1486                 ftrace_graph_sleep_time_control(set);
1487
1488         if (bit == TRACE_GRAPH_GRAPH_TIME)
1489                 ftrace_graph_graph_time_control(set);
1490
1491         return 0;
1492 }
1493
1494 static struct trace_event_functions graph_functions = {
1495         .trace          = print_graph_function_event,
1496 };
1497
1498 static struct trace_event graph_trace_entry_event = {
1499         .type           = TRACE_GRAPH_ENT,
1500         .funcs          = &graph_functions,
1501 };
1502
1503 static struct trace_event graph_trace_ret_event = {
1504         .type           = TRACE_GRAPH_RET,
1505         .funcs          = &graph_functions
1506 };
1507
1508 static struct tracer graph_trace __tracer_data = {
1509         .name           = "function_graph",
1510         .update_thresh  = graph_trace_update_thresh,
1511         .open           = graph_trace_open,
1512         .pipe_open      = graph_trace_open,
1513         .close          = graph_trace_close,
1514         .pipe_close     = graph_trace_close,
1515         .init           = graph_trace_init,
1516         .reset          = graph_trace_reset,
1517         .print_line     = print_graph_function,
1518         .print_header   = print_graph_headers,
1519         .flags          = &tracer_flags,
1520         .set_flag       = func_graph_set_flag,
1521 #ifdef CONFIG_FTRACE_SELFTEST
1522         .selftest       = trace_selftest_startup_function_graph,
1523 #endif
1524 };
1525
1526
1527 static ssize_t
1528 graph_depth_write(struct file *filp, const char __user *ubuf, size_t cnt,
1529                   loff_t *ppos)
1530 {
1531         unsigned long val;
1532         int ret;
1533
1534         ret = kstrtoul_from_user(ubuf, cnt, 10, &val);
1535         if (ret)
1536                 return ret;
1537
1538         fgraph_max_depth = val;
1539
1540         *ppos += cnt;
1541
1542         return cnt;
1543 }
1544
1545 static ssize_t
1546 graph_depth_read(struct file *filp, char __user *ubuf, size_t cnt,
1547                  loff_t *ppos)
1548 {
1549         char buf[15]; /* More than enough to hold UINT_MAX + "\n"*/
1550         int n;
1551
1552         n = sprintf(buf, "%d\n", fgraph_max_depth);
1553
1554         return simple_read_from_buffer(ubuf, cnt, ppos, buf, n);
1555 }
1556
1557 static const struct file_operations graph_depth_fops = {
1558         .open           = tracing_open_generic,
1559         .write          = graph_depth_write,
1560         .read           = graph_depth_read,
1561         .llseek         = generic_file_llseek,
1562 };
1563
1564 static __init int init_graph_tracefs(void)
1565 {
1566         struct dentry *d_tracer;
1567
1568         d_tracer = tracing_init_dentry();
1569         if (IS_ERR(d_tracer))
1570                 return 0;
1571
1572         trace_create_file("max_graph_depth", 0644, d_tracer,
1573                           NULL, &graph_depth_fops);
1574
1575         return 0;
1576 }
1577 fs_initcall(init_graph_tracefs);
1578
1579 static __init int init_graph_trace(void)
1580 {
1581         max_bytes_for_cpu = snprintf(NULL, 0, "%u", nr_cpu_ids - 1);
1582
1583         if (!register_trace_event(&graph_trace_entry_event)) {
1584                 pr_warn("Warning: could not register graph trace events\n");
1585                 return 1;
1586         }
1587
1588         if (!register_trace_event(&graph_trace_ret_event)) {
1589                 pr_warn("Warning: could not register graph trace events\n");
1590                 return 1;
1591         }
1592
1593         return register_tracer(&graph_trace);
1594 }
1595
1596 core_initcall(init_graph_trace);