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 int
 122 ftrace_push_return_trace(unsigned long ret, unsigned long func, int *depth,
 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         *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                 trace_buffer_unlock_commit_nostack(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         if (ftrace_graph_ignore_func(trace))
 389                 return 0;
 390
 391         if (ftrace_graph_ignore_irqs())
 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                 trace_buffer_unlock_commit_nostack(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         ftrace_graph_addr_finish(trace);
 486
 487         local_irq_save(flags);
 488         cpu = raw_smp_processor_id();
 489         data = per_cpu_ptr(tr->trace_buffer.data, cpu);
 490         disabled = atomic_inc_return(&data->disabled);
 491         if (likely(disabled == 1)) {
 492                 pc = preempt_count();
 493                 __trace_graph_return(tr, trace, flags, pc);
 494         }
 495         atomic_dec(&data->disabled);
 496         local_irq_restore(flags);
 497 }
 498
 499 void set_graph_array(struct trace_array *tr)
 500 {
 501         graph_array = tr;
 502
 503         /* Make graph_array visible before we start tracing */
 504
 505         smp_mb();
 506 }
 507
 508 static void trace_graph_thresh_return(struct ftrace_graph_ret *trace)
 509 {
 510         ftrace_graph_addr_finish(trace);
 511
 512         if (tracing_thresh &&
 513             (trace->rettime - trace->calltime < tracing_thresh))
 514                 return;
 515         else
 516                 trace_graph_return(trace);
 517 }
 518
 519 static int graph_trace_init(struct trace_array *tr)
 520 {
 521         int ret;
 522
 523         set_graph_array(tr);
 524         if (tracing_thresh)
 525                 ret = register_ftrace_graph(&trace_graph_thresh_return,
 526                                             &trace_graph_entry);
 527         else
 528                 ret = register_ftrace_graph(&trace_graph_return,
 529                                             &trace_graph_entry);
 530         if (ret)
 531                 return ret;
 532         tracing_start_cmdline_record();
 533
 534         return 0;
 535 }
 536
 537 static void graph_trace_reset(struct trace_array *tr)
 538 {
 539         tracing_stop_cmdline_record();
 540         unregister_ftrace_graph();
 541 }
 542
 543 static int graph_trace_update_thresh(struct trace_array *tr)
 544 {
 545         graph_trace_reset(tr);
 546         return graph_trace_init(tr);
 547 }
 548
 549 static int max_bytes_for_cpu;
 550
 551 static void print_graph_cpu(struct trace_seq *s, int cpu)
 552 {
 553         /*
 554          * Start with a space character - to make it stand out
 555          * to the right a bit when trace output is pasted into
 556          * email:
 557          */
 558         trace_seq_printf(s, " %*d) ", max_bytes_for_cpu, cpu);
 559 }
 560
 561 #define TRACE_GRAPH_PROCINFO_LENGTH     14
 562
 563 static void print_graph_proc(struct trace_seq *s, pid_t pid)
 564 {
 565         char comm[TASK_COMM_LEN];
 566         /* sign + log10(MAX_INT) + '\0' */
 567         char pid_str[11];
 568         int spaces = 0;
 569         int len;
 570         int i;
 571
 572         trace_find_cmdline(pid, comm);
 573         comm[7] = '\0';
 574         sprintf(pid_str, "%d", pid);
 575
 576         /* 1 stands for the "-" character */
 577         len = strlen(comm) + strlen(pid_str) + 1;
 578
 579         if (len < TRACE_GRAPH_PROCINFO_LENGTH)
 580                 spaces = TRACE_GRAPH_PROCINFO_LENGTH - len;
 581
 582         /* First spaces to align center */
 583         for (i = 0; i < spaces / 2; i++)
 584                 trace_seq_putc(s, ' ');
 585
 586         trace_seq_printf(s, "%s-%s", comm, pid_str);
 587
 588         /* Last spaces to align center */
 589         for (i = 0; i < spaces - (spaces / 2); i++)
 590                 trace_seq_putc(s, ' ');
 591 }
 592
 593
 594 static void print_graph_lat_fmt(struct trace_seq *s, struct trace_entry *entry)
 595 {
 596         trace_seq_putc(s, ' ');
 597         trace_print_lat_fmt(s, entry);
 598 }
 599
 600 /* If the pid changed since the last trace, output this event */
 601 static void
 602 verif_pid(struct trace_seq *s, pid_t pid, int cpu, struct fgraph_data *data)
 603 {
 604         pid_t prev_pid;
 605         pid_t *last_pid;
 606
 607         if (!data)
 608                 return;
 609
 610         last_pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
 611
 612         if (*last_pid == pid)
 613                 return;
 614
 615         prev_pid = *last_pid;
 616         *last_pid = pid;
 617
 618         if (prev_pid == -1)
 619                 return;
 620 /*
 621  * Context-switch trace line:
 622
 623  ------------------------------------------
 624  | 1)  migration/0--1  =>  sshd-1755
 625  ------------------------------------------
 626
 627  */
 628         trace_seq_puts(s, " ------------------------------------------\n");
 629         print_graph_cpu(s, cpu);
 630         print_graph_proc(s, prev_pid);
 631         trace_seq_puts(s, " => ");
 632         print_graph_proc(s, pid);
 633         trace_seq_puts(s, "\n ------------------------------------------\n\n");
 634 }
 635
 636 static struct ftrace_graph_ret_entry *
 637 get_return_for_leaf(struct trace_iterator *iter,
 638                 struct ftrace_graph_ent_entry *curr)
 639 {
 640         struct fgraph_data *data = iter->private;
 641         struct ring_buffer_iter *ring_iter = NULL;
 642         struct ring_buffer_event *event;
 643         struct ftrace_graph_ret_entry *next;
 644
 645         /*
 646          * If the previous output failed to write to the seq buffer,
 647          * then we just reuse the data from before.
 648          */
 649         if (data && data->failed) {
 650                 curr = &data->ent;
 651                 next = &data->ret;
 652         } else {
 653
 654                 ring_iter = trace_buffer_iter(iter, iter->cpu);
 655
 656                 /* First peek to compare current entry and the next one */
 657                 if (ring_iter)
 658                         event = ring_buffer_iter_peek(ring_iter, NULL);
 659                 else {
 660                         /*
 661                          * We need to consume the current entry to see
 662                          * the next one.
 663                          */
 664                         ring_buffer_consume(iter->trace_buffer->buffer, iter->cpu,
 665                                             NULL, NULL);
 666                         event = ring_buffer_peek(iter->trace_buffer->buffer, iter->cpu,
 667                                                  NULL, NULL);
 668                 }
 669
 670                 if (!event)
 671                         return NULL;
 672
 673                 next = ring_buffer_event_data(event);
 674
 675                 if (data) {
 676                         /*
 677                          * Save current and next entries for later reference
 678                          * if the output fails.
 679                          */
 680                         data->ent = *curr;
 681                         /*
 682                          * If the next event is not a return type, then
 683                          * we only care about what type it is. Otherwise we can
 684                          * safely copy the entire event.
 685                          */
 686                         if (next->ent.type == TRACE_GRAPH_RET)
 687                                 data->ret = *next;
 688                         else
 689                                 data->ret.ent.type = next->ent.type;
 690                 }
 691         }
 692
 693         if (next->ent.type != TRACE_GRAPH_RET)
 694                 return NULL;
 695
 696         if (curr->ent.pid != next->ent.pid ||
 697                         curr->graph_ent.func != next->ret.func)
 698                 return NULL;
 699
 700         /* this is a leaf, now advance the iterator */
 701         if (ring_iter)
 702                 ring_buffer_read(ring_iter, NULL);
 703
 704         return next;
 705 }
 706
 707 static void print_graph_abs_time(u64 t, struct trace_seq *s)
 708 {
 709         unsigned long usecs_rem;
 710
 711         usecs_rem = do_div(t, NSEC_PER_SEC);
 712         usecs_rem /= 1000;
 713
 714         trace_seq_printf(s, "%5lu.%06lu |  ",
 715                          (unsigned long)t, usecs_rem);
 716 }
 717
 718 static void
 719 print_graph_irq(struct trace_iterator *iter, unsigned long addr,
 720                 enum trace_type type, int cpu, pid_t pid, u32 flags)
 721 {
 722         struct trace_array *tr = iter->tr;
 723         struct trace_seq *s = &iter->seq;
 724         struct trace_entry *ent = iter->ent;
 725
 726         if (addr < (unsigned long)__irqentry_text_start ||
 727                 addr >= (unsigned long)__irqentry_text_end)
 728                 return;
 729
 730         if (tr->trace_flags & TRACE_ITER_CONTEXT_INFO) {
 731                 /* Absolute time */
 732                 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
 733                         print_graph_abs_time(iter->ts, s);
 734
 735                 /* Cpu */
 736                 if (flags & TRACE_GRAPH_PRINT_CPU)
 737                         print_graph_cpu(s, cpu);
 738
 739                 /* Proc */
 740                 if (flags & TRACE_GRAPH_PRINT_PROC) {
 741                         print_graph_proc(s, pid);
 742                         trace_seq_puts(s, " | ");
 743                 }
 744
 745                 /* Latency format */
 746                 if (tr->trace_flags & TRACE_ITER_LATENCY_FMT)
 747                         print_graph_lat_fmt(s, ent);
 748         }
 749
 750         /* No overhead */
 751         print_graph_duration(tr, 0, s, flags | FLAGS_FILL_START);
 752
 753         if (type == TRACE_GRAPH_ENT)
 754                 trace_seq_puts(s, "==========>");
 755         else
 756                 trace_seq_puts(s, "<==========");
 757
 758         print_graph_duration(tr, 0, s, flags | FLAGS_FILL_END);
 759         trace_seq_putc(s, '\n');
 760 }
 761
 762 void
 763 trace_print_graph_duration(unsigned long long duration, struct trace_seq *s)
 764 {
 765         unsigned long nsecs_rem = do_div(duration, 1000);
 766         /* log10(ULONG_MAX) + '\0' */
 767         char usecs_str[21];
 768         char nsecs_str[5];
 769         int len;
 770         int i;
 771
 772         sprintf(usecs_str, "%lu", (unsigned long) duration);
 773
 774         /* Print msecs */
 775         trace_seq_printf(s, "%s", usecs_str);
 776
 777         len = strlen(usecs_str);
 778
 779         /* Print nsecs (we don't want to exceed 7 numbers) */
 780         if (len < 7) {
 781                 size_t slen = min_t(size_t, sizeof(nsecs_str), 8UL - len);
 782
 783                 snprintf(nsecs_str, slen, "%03lu", nsecs_rem);
 784                 trace_seq_printf(s, ".%s", nsecs_str);
 785                 len += strlen(nsecs_str) + 1;
 786         }
 787
 788         trace_seq_puts(s, " us ");
 789
 790         /* Print remaining spaces to fit the row's width */
 791         for (i = len; i < 8; i++)
 792                 trace_seq_putc(s, ' ');
 793 }
 794
 795 static void
 796 print_graph_duration(struct trace_array *tr, unsigned long long duration,
 797                      struct trace_seq *s, u32 flags)
 798 {
 799         if (!(flags & TRACE_GRAPH_PRINT_DURATION) ||
 800             !(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
 801                 return;
 802
 803         /* No real adata, just filling the column with spaces */
 804         switch (flags & TRACE_GRAPH_PRINT_FILL_MASK) {
 805         case FLAGS_FILL_FULL:
 806                 trace_seq_puts(s, "              |  ");
 807                 return;
 808         case FLAGS_FILL_START:
 809                 trace_seq_puts(s, "  ");
 810                 return;
 811         case FLAGS_FILL_END:
 812                 trace_seq_puts(s, " |");
 813                 return;
 814         }
 815
 816         /* Signal a overhead of time execution to the output */
 817         if (flags & TRACE_GRAPH_PRINT_OVERHEAD)
 818                 trace_seq_printf(s, "%c ", trace_find_mark(duration));
 819         else
 820                 trace_seq_puts(s, "  ");
 821
 822         trace_print_graph_duration(duration, s);
 823         trace_seq_puts(s, "|  ");
 824 }
 825
 826 /* Case of a leaf function on its call entry */
 827 static enum print_line_t
 828 print_graph_entry_leaf(struct trace_iterator *iter,
 829                 struct ftrace_graph_ent_entry *entry,
 830                 struct ftrace_graph_ret_entry *ret_entry,
 831                 struct trace_seq *s, u32 flags)
 832 {
 833         struct fgraph_data *data = iter->private;
 834         struct trace_array *tr = iter->tr;
 835         struct ftrace_graph_ret *graph_ret;
 836         struct ftrace_graph_ent *call;
 837         unsigned long long duration;
 838         int cpu = iter->cpu;
 839         int i;
 840
 841         graph_ret = &ret_entry->ret;
 842         call = &entry->graph_ent;
 843         duration = graph_ret->rettime - graph_ret->calltime;
 844
 845         if (data) {
 846                 struct fgraph_cpu_data *cpu_data;
 847
 848                 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
 849
 850                 /* If a graph tracer ignored set_graph_notrace */
 851                 if (call->depth < -1)
 852                         call->depth += FTRACE_NOTRACE_DEPTH;
 853
 854                 /*
 855                  * Comments display at + 1 to depth. Since
 856                  * this is a leaf function, keep the comments
 857                  * equal to this depth.
 858                  */
 859                 cpu_data->depth = call->depth - 1;
 860
 861                 /* No need to keep this function around for this depth */
 862                 if (call->depth < FTRACE_RETFUNC_DEPTH &&
 863                     !WARN_ON_ONCE(call->depth < 0))
 864                         cpu_data->enter_funcs[call->depth] = 0;
 865         }
 866
 867         /* Overhead and duration */
 868         print_graph_duration(tr, duration, s, flags);
 869
 870         /* Function */
 871         for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
 872                 trace_seq_putc(s, ' ');
 873
 874         trace_seq_printf(s, "%ps();\n", (void *)call->func);
 875
 876         print_graph_irq(iter, graph_ret->func, TRACE_GRAPH_RET,
 877                         cpu, iter->ent->pid, flags);
 878
 879         return trace_handle_return(s);
 880 }
 881
 882 static enum print_line_t
 883 print_graph_entry_nested(struct trace_iterator *iter,
 884                          struct ftrace_graph_ent_entry *entry,
 885                          struct trace_seq *s, int cpu, u32 flags)
 886 {
 887         struct ftrace_graph_ent *call = &entry->graph_ent;
 888         struct fgraph_data *data = iter->private;
 889         struct trace_array *tr = iter->tr;
 890         int i;
 891
 892         if (data) {
 893                 struct fgraph_cpu_data *cpu_data;
 894                 int cpu = iter->cpu;
 895
 896                 /* If a graph tracer ignored set_graph_notrace */
 897                 if (call->depth < -1)
 898                         call->depth += FTRACE_NOTRACE_DEPTH;
 899
 900                 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
 901                 cpu_data->depth = call->depth;
 902
 903                 /* Save this function pointer to see if the exit matches */
 904                 if (call->depth < FTRACE_RETFUNC_DEPTH &&
 905                     !WARN_ON_ONCE(call->depth < 0))
 906                         cpu_data->enter_funcs[call->depth] = call->func;
 907         }
 908
 909         /* No time */
 910         print_graph_duration(tr, 0, s, flags | FLAGS_FILL_FULL);
 911
 912         /* Function */
 913         for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
 914                 trace_seq_putc(s, ' ');
 915
 916         trace_seq_printf(s, "%ps() {\n", (void *)call->func);
 917
 918         if (trace_seq_has_overflowed(s))
 919                 return TRACE_TYPE_PARTIAL_LINE;
 920
 921         /*
 922          * we already consumed the current entry to check the next one
 923          * and see if this is a leaf.
 924          */
 925         return TRACE_TYPE_NO_CONSUME;
 926 }
 927
 928 static void
 929 print_graph_prologue(struct trace_iterator *iter, struct trace_seq *s,
 930                      int type, unsigned long addr, u32 flags)
 931 {
 932         struct fgraph_data *data = iter->private;
 933         struct trace_entry *ent = iter->ent;
 934         struct trace_array *tr = iter->tr;
 935         int cpu = iter->cpu;
 936
 937         /* Pid */
 938         verif_pid(s, ent->pid, cpu, data);
 939
 940         if (type)
 941                 /* Interrupt */
 942                 print_graph_irq(iter, addr, type, cpu, ent->pid, flags);
 943
 944         if (!(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
 945                 return;
 946
 947         /* Absolute time */
 948         if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
 949                 print_graph_abs_time(iter->ts, s);
 950
 951         /* Cpu */
 952         if (flags & TRACE_GRAPH_PRINT_CPU)
 953                 print_graph_cpu(s, cpu);
 954
 955         /* Proc */
 956         if (flags & TRACE_GRAPH_PRINT_PROC) {
 957                 print_graph_proc(s, ent->pid);
 958                 trace_seq_puts(s, " | ");
 959         }
 960
 961         /* Latency format */
 962         if (tr->trace_flags & TRACE_ITER_LATENCY_FMT)
 963                 print_graph_lat_fmt(s, ent);
 964
 965         return;
 966 }
 967
 968 /*
 969  * Entry check for irq code
 970  *
 971  * returns 1 if
 972  *  - we are inside irq code
 973  *  - we just entered irq code
 974  *
 975  * retunns 0 if
 976  *  - funcgraph-interrupts option is set
 977  *  - we are not inside irq code
 978  */
 979 static int
 980 check_irq_entry(struct trace_iterator *iter, u32 flags,
 981                 unsigned long addr, int depth)
 982 {
 983         int cpu = iter->cpu;
 984         int *depth_irq;
 985         struct fgraph_data *data = iter->private;
 986
 987         /*
 988          * If we are either displaying irqs, or we got called as
 989          * a graph event and private data does not exist,
 990          * then we bypass the irq check.
 991          */
 992         if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
 993             (!data))
 994                 return 0;
 995
 996         depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
 997
 998         /*
 999          * We are inside the irq code
1000          */
1001         if (*depth_irq >= 0)
1002                 return 1;
1003
1004         if ((addr < (unsigned long)__irqentry_text_start) ||
1005             (addr >= (unsigned long)__irqentry_text_end))
1006                 return 0;
1007
1008         /*
1009          * We are entering irq code.
1010          */
1011         *depth_irq = depth;
1012         return 1;
1013 }
1014
1015 /*
1016  * Return check for irq code
1017  *
1018  * returns 1 if
1019  *  - we are inside irq code
1020  *  - we just left irq code
1021  *
1022  * returns 0 if
1023  *  - funcgraph-interrupts option is set
1024  *  - we are not inside irq code
1025  */
1026 static int
1027 check_irq_return(struct trace_iterator *iter, u32 flags, int depth)
1028 {
1029         int cpu = iter->cpu;
1030         int *depth_irq;
1031         struct fgraph_data *data = iter->private;
1032
1033         /*
1034          * If we are either displaying irqs, or we got called as
1035          * a graph event and private data does not exist,
1036          * then we bypass the irq check.
1037          */
1038         if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
1039             (!data))
1040                 return 0;
1041
1042         depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
1043
1044         /*
1045          * We are not inside the irq code.
1046          */
1047         if (*depth_irq == -1)
1048                 return 0;
1049
1050         /*
1051          * We are inside the irq code, and this is returning entry.
1052          * Let's not trace it and clear the entry depth, since
1053          * we are out of irq code.
1054          *
1055          * This condition ensures that we 'leave the irq code' once
1056          * we are out of the entry depth. Thus protecting us from
1057          * the RETURN entry loss.
1058          */
1059         if (*depth_irq >= depth) {
1060                 *depth_irq = -1;
1061                 return 1;
1062         }
1063
1064         /*
1065          * We are inside the irq code, and this is not the entry.
1066          */
1067         return 1;
1068 }
1069
1070 static enum print_line_t
1071 print_graph_entry(struct ftrace_graph_ent_entry *field, struct trace_seq *s,
1072                         struct trace_iterator *iter, u32 flags)
1073 {
1074         struct fgraph_data *data = iter->private;
1075         struct ftrace_graph_ent *call = &field->graph_ent;
1076         struct ftrace_graph_ret_entry *leaf_ret;
1077         static enum print_line_t ret;
1078         int cpu = iter->cpu;
1079
1080         if (check_irq_entry(iter, flags, call->func, call->depth))
1081                 return TRACE_TYPE_HANDLED;
1082
1083         print_graph_prologue(iter, s, TRACE_GRAPH_ENT, call->func, flags);
1084
1085         leaf_ret = get_return_for_leaf(iter, field);
1086         if (leaf_ret)
1087                 ret = print_graph_entry_leaf(iter, field, leaf_ret, s, flags);
1088         else
1089                 ret = print_graph_entry_nested(iter, field, s, cpu, flags);
1090
1091         if (data) {
1092                 /*
1093                  * If we failed to write our output, then we need to make
1094                  * note of it. Because we already consumed our entry.
1095                  */
1096                 if (s->full) {
1097                         data->failed = 1;
1098                         data->cpu = cpu;
1099                 } else
1100                         data->failed = 0;
1101         }
1102
1103         return ret;
1104 }
1105
1106 static enum print_line_t
1107 print_graph_return(struct ftrace_graph_ret *trace, struct trace_seq *s,
1108                    struct trace_entry *ent, struct trace_iterator *iter,
1109                    u32 flags)
1110 {
1111         unsigned long long duration = trace->rettime - trace->calltime;
1112         struct fgraph_data *data = iter->private;
1113         struct trace_array *tr = iter->tr;
1114         pid_t pid = ent->pid;
1115         int cpu = iter->cpu;
1116         int func_match = 1;
1117         int i;
1118
1119         if (check_irq_return(iter, flags, trace->depth))
1120                 return TRACE_TYPE_HANDLED;
1121
1122         if (data) {
1123                 struct fgraph_cpu_data *cpu_data;
1124                 int cpu = iter->cpu;
1125
1126                 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
1127
1128                 /*
1129                  * Comments display at + 1 to depth. This is the
1130                  * return from a function, we now want the comments
1131                  * to display at the same level of the bracket.
1132                  */
1133                 cpu_data->depth = trace->depth - 1;
1134
1135                 if (trace->depth < FTRACE_RETFUNC_DEPTH &&
1136                     !WARN_ON_ONCE(trace->depth < 0)) {
1137                         if (cpu_data->enter_funcs[trace->depth] != trace->func)
1138                                 func_match = 0;
1139                         cpu_data->enter_funcs[trace->depth] = 0;
1140                 }
1141         }
1142
1143         print_graph_prologue(iter, s, 0, 0, flags);
1144
1145         /* Overhead and duration */
1146         print_graph_duration(tr, duration, s, flags);
1147
1148         /* Closing brace */
1149         for (i = 0; i < trace->depth * TRACE_GRAPH_INDENT; i++)
1150                 trace_seq_putc(s, ' ');
1151
1152         /*
1153          * If the return function does not have a matching entry,
1154          * then the entry was lost. Instead of just printing
1155          * the '}' and letting the user guess what function this
1156          * belongs to, write out the function name. Always do
1157          * that if the funcgraph-tail option is enabled.
1158          */
1159         if (func_match && !(flags & TRACE_GRAPH_PRINT_TAIL))
1160                 trace_seq_puts(s, "}\n");
1161         else
1162                 trace_seq_printf(s, "} /* %ps */\n", (void *)trace->func);
1163
1164         /* Overrun */
1165         if (flags & TRACE_GRAPH_PRINT_OVERRUN)
1166                 trace_seq_printf(s, " (Overruns: %lu)\n",
1167                                  trace->overrun);
1168
1169         print_graph_irq(iter, trace->func, TRACE_GRAPH_RET,
1170                         cpu, pid, flags);
1171
1172         return trace_handle_return(s);
1173 }
1174
1175 static enum print_line_t
1176 print_graph_comment(struct trace_seq *s, struct trace_entry *ent,
1177                     struct trace_iterator *iter, u32 flags)
1178 {
1179         struct trace_array *tr = iter->tr;
1180         unsigned long sym_flags = (tr->trace_flags & TRACE_ITER_SYM_MASK);
1181         struct fgraph_data *data = iter->private;
1182         struct trace_event *event;
1183         int depth = 0;
1184         int ret;
1185         int i;
1186
1187         if (data)
1188                 depth = per_cpu_ptr(data->cpu_data, iter->cpu)->depth;
1189
1190         print_graph_prologue(iter, s, 0, 0, flags);
1191
1192         /* No time */
1193         print_graph_duration(tr, 0, s, flags | FLAGS_FILL_FULL);
1194
1195         /* Indentation */
1196         if (depth > 0)
1197                 for (i = 0; i < (depth + 1) * TRACE_GRAPH_INDENT; i++)
1198                         trace_seq_putc(s, ' ');
1199
1200         /* The comment */
1201         trace_seq_puts(s, "/* ");
1202
1203         switch (iter->ent->type) {
1204         case TRACE_BPUTS:
1205                 ret = trace_print_bputs_msg_only(iter);
1206                 if (ret != TRACE_TYPE_HANDLED)
1207                         return ret;
1208                 break;
1209         case TRACE_BPRINT:
1210                 ret = trace_print_bprintk_msg_only(iter);
1211                 if (ret != TRACE_TYPE_HANDLED)
1212                         return ret;
1213                 break;
1214         case TRACE_PRINT:
1215                 ret = trace_print_printk_msg_only(iter);
1216                 if (ret != TRACE_TYPE_HANDLED)
1217                         return ret;
1218                 break;
1219         default:
1220                 event = ftrace_find_event(ent->type);
1221                 if (!event)
1222                         return TRACE_TYPE_UNHANDLED;
1223
1224                 ret = event->funcs->trace(iter, sym_flags, event);
1225                 if (ret != TRACE_TYPE_HANDLED)
1226                         return ret;
1227         }
1228
1229         if (trace_seq_has_overflowed(s))
1230                 goto out;
1231
1232         /* Strip ending newline */
1233         if (s->buffer[s->seq.len - 1] == '\n') {
1234                 s->buffer[s->seq.len - 1] = '\0';
1235                 s->seq.len--;
1236         }
1237
1238         trace_seq_puts(s, " */\n");
1239  out:
1240         return trace_handle_return(s);
1241 }
1242
1243
1244 enum print_line_t
1245 print_graph_function_flags(struct trace_iterator *iter, u32 flags)
1246 {
1247         struct ftrace_graph_ent_entry *field;
1248         struct fgraph_data *data = iter->private;
1249         struct trace_entry *entry = iter->ent;
1250         struct trace_seq *s = &iter->seq;
1251         int cpu = iter->cpu;
1252         int ret;
1253
1254         if (data && per_cpu_ptr(data->cpu_data, cpu)->ignore) {
1255                 per_cpu_ptr(data->cpu_data, cpu)->ignore = 0;
1256                 return TRACE_TYPE_HANDLED;
1257         }
1258
1259         /*
1260          * If the last output failed, there's a possibility we need
1261          * to print out the missing entry which would never go out.
1262          */
1263         if (data && data->failed) {
1264                 field = &data->ent;
1265                 iter->cpu = data->cpu;
1266                 ret = print_graph_entry(field, s, iter, flags);
1267                 if (ret == TRACE_TYPE_HANDLED && iter->cpu != cpu) {
1268                         per_cpu_ptr(data->cpu_data, iter->cpu)->ignore = 1;
1269                         ret = TRACE_TYPE_NO_CONSUME;
1270                 }
1271                 iter->cpu = cpu;
1272                 return ret;
1273         }
1274
1275         switch (entry->type) {
1276         case TRACE_GRAPH_ENT: {
1277                 /*
1278                  * print_graph_entry() may consume the current event,
1279                  * thus @field may become invalid, so we need to save it.
1280                  * sizeof(struct ftrace_graph_ent_entry) is very small,
1281                  * it can be safely saved at the stack.
1282                  */
1283                 struct ftrace_graph_ent_entry saved;
1284                 trace_assign_type(field, entry);
1285                 saved = *field;
1286                 return print_graph_entry(&saved, s, iter, flags);
1287         }
1288         case TRACE_GRAPH_RET: {
1289                 struct ftrace_graph_ret_entry *field;
1290                 trace_assign_type(field, entry);
1291                 return print_graph_return(&field->ret, s, entry, iter, flags);
1292         }
1293         case TRACE_STACK:
1294         case TRACE_FN:
1295                 /* dont trace stack and functions as comments */
1296                 return TRACE_TYPE_UNHANDLED;
1297
1298         default:
1299                 return print_graph_comment(s, entry, iter, flags);
1300         }
1301
1302         return TRACE_TYPE_HANDLED;
1303 }
1304
1305 static enum print_line_t
1306 print_graph_function(struct trace_iterator *iter)
1307 {
1308         return print_graph_function_flags(iter, tracer_flags.val);
1309 }
1310
1311 static enum print_line_t
1312 print_graph_function_event(struct trace_iterator *iter, int flags,
1313                            struct trace_event *event)
1314 {
1315         return print_graph_function(iter);
1316 }
1317
1318 static void print_lat_header(struct seq_file *s, u32 flags)
1319 {
1320         static const char spaces[] = "                " /* 16 spaces */
1321                 "    "                                  /* 4 spaces */
1322                 "                 ";                    /* 17 spaces */
1323         int size = 0;
1324
1325         if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1326                 size += 16;
1327         if (flags & TRACE_GRAPH_PRINT_CPU)
1328                 size += 4;
1329         if (flags & TRACE_GRAPH_PRINT_PROC)
1330                 size += 17;
1331
1332         seq_printf(s, "#%.*s  _-----=> irqs-off        \n", size, spaces);
1333         seq_printf(s, "#%.*s / _----=> need-resched    \n", size, spaces);
1334         seq_printf(s, "#%.*s| / _---=> hardirq/softirq \n", size, spaces);
1335         seq_printf(s, "#%.*s|| / _--=> preempt-depth   \n", size, spaces);
1336         seq_printf(s, "#%.*s||| /                      \n", size, spaces);
1337 }
1338
1339 static void __print_graph_headers_flags(struct trace_array *tr,
1340                                         struct seq_file *s, u32 flags)
1341 {
1342         int lat = tr->trace_flags & TRACE_ITER_LATENCY_FMT;
1343
1344         if (lat)
1345                 print_lat_header(s, flags);
1346
1347         /* 1st line */
1348         seq_putc(s, '#');
1349         if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1350                 seq_puts(s, "     TIME       ");
1351         if (flags & TRACE_GRAPH_PRINT_CPU)
1352                 seq_puts(s, " CPU");
1353         if (flags & TRACE_GRAPH_PRINT_PROC)
1354                 seq_puts(s, "  TASK/PID       ");
1355         if (lat)
1356                 seq_puts(s, "||||");
1357         if (flags & TRACE_GRAPH_PRINT_DURATION)
1358                 seq_puts(s, "  DURATION   ");
1359         seq_puts(s, "               FUNCTION CALLS\n");
1360
1361         /* 2nd line */
1362         seq_putc(s, '#');
1363         if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1364                 seq_puts(s, "      |         ");
1365         if (flags & TRACE_GRAPH_PRINT_CPU)
1366                 seq_puts(s, " |  ");
1367         if (flags & TRACE_GRAPH_PRINT_PROC)
1368                 seq_puts(s, "   |    |        ");
1369         if (lat)
1370                 seq_puts(s, "||||");
1371         if (flags & TRACE_GRAPH_PRINT_DURATION)
1372                 seq_puts(s, "   |   |      ");
1373         seq_puts(s, "               |   |   |   |\n");
1374 }
1375
1376 static void print_graph_headers(struct seq_file *s)
1377 {
1378         print_graph_headers_flags(s, tracer_flags.val);
1379 }
1380
1381 void print_graph_headers_flags(struct seq_file *s, u32 flags)
1382 {
1383         struct trace_iterator *iter = s->private;
1384         struct trace_array *tr = iter->tr;
1385
1386         if (!(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
1387                 return;
1388
1389         if (tr->trace_flags & TRACE_ITER_LATENCY_FMT) {
1390                 /* print nothing if the buffers are empty */
1391                 if (trace_empty(iter))
1392                         return;
1393
1394                 print_trace_header(s, iter);
1395         }
1396
1397         __print_graph_headers_flags(tr, s, flags);
1398 }
1399
1400 void graph_trace_open(struct trace_iterator *iter)
1401 {
1402         /* pid and depth on the last trace processed */
1403         struct fgraph_data *data;
1404         gfp_t gfpflags;
1405         int cpu;
1406
1407         iter->private = NULL;
1408
1409         /* We can be called in atomic context via ftrace_dump() */
1410         gfpflags = (in_atomic() || irqs_disabled()) ? GFP_ATOMIC : GFP_KERNEL;
1411
1412         data = kzalloc(sizeof(*data), gfpflags);
1413         if (!data)
1414                 goto out_err;
1415
1416         data->cpu_data = alloc_percpu_gfp(struct fgraph_cpu_data, gfpflags);
1417         if (!data->cpu_data)
1418                 goto out_err_free;
1419
1420         for_each_possible_cpu(cpu) {
1421                 pid_t *pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
1422                 int *depth = &(per_cpu_ptr(data->cpu_data, cpu)->depth);
1423                 int *ignore = &(per_cpu_ptr(data->cpu_data, cpu)->ignore);
1424                 int *depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
1425
1426                 *pid = -1;
1427                 *depth = 0;
1428                 *ignore = 0;
1429                 *depth_irq = -1;
1430         }
1431
1432         iter->private = data;
1433
1434         return;
1435
1436  out_err_free:
1437         kfree(data);
1438  out_err:
1439         pr_warn("function graph tracer: not enough memory\n");
1440 }
1441
1442 void graph_trace_close(struct trace_iterator *iter)
1443 {
1444         struct fgraph_data *data = iter->private;
1445
1446         if (data) {
1447                 free_percpu(data->cpu_data);
1448                 kfree(data);
1449         }
1450 }
1451
1452 static int
1453 func_graph_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set)
1454 {
1455         if (bit == TRACE_GRAPH_PRINT_IRQS)
1456                 ftrace_graph_skip_irqs = !set;
1457
1458         if (bit == TRACE_GRAPH_SLEEP_TIME)
1459                 ftrace_graph_sleep_time_control(set);
1460
1461         if (bit == TRACE_GRAPH_GRAPH_TIME)
1462                 ftrace_graph_graph_time_control(set);
1463
1464         return 0;
1465 }
1466
1467 static struct trace_event_functions graph_functions = {
1468         .trace          = print_graph_function_event,
1469 };
1470
1471 static struct trace_event graph_trace_entry_event = {
1472         .type           = TRACE_GRAPH_ENT,
1473         .funcs          = &graph_functions,
1474 };
1475
1476 static struct trace_event graph_trace_ret_event = {
1477         .type           = TRACE_GRAPH_RET,
1478         .funcs          = &graph_functions
1479 };
1480
1481 static struct tracer graph_trace __tracer_data = {
1482         .name           = "function_graph",
1483         .update_thresh  = graph_trace_update_thresh,
1484         .open           = graph_trace_open,
1485         .pipe_open      = graph_trace_open,
1486         .close          = graph_trace_close,
1487         .pipe_close     = graph_trace_close,
1488         .init           = graph_trace_init,
1489         .reset          = graph_trace_reset,
1490         .print_line     = print_graph_function,
1491         .print_header   = print_graph_headers,
1492         .flags          = &tracer_flags,
1493         .set_flag       = func_graph_set_flag,
1494 #ifdef CONFIG_FTRACE_SELFTEST
1495         .selftest       = trace_selftest_startup_function_graph,
1496 #endif
1497 };
1498
1499
1500 static ssize_t
1501 graph_depth_write(struct file *filp, const char __user *ubuf, size_t cnt,
1502                   loff_t *ppos)
1503 {
1504         unsigned long val;
1505         int ret;
1506
1507         ret = kstrtoul_from_user(ubuf, cnt, 10, &val);
1508         if (ret)
1509                 return ret;
1510
1511         fgraph_max_depth = val;
1512
1513         *ppos += cnt;
1514
1515         return cnt;
1516 }
1517
1518 static ssize_t
1519 graph_depth_read(struct file *filp, char __user *ubuf, size_t cnt,
1520                  loff_t *ppos)
1521 {
1522         char buf[15]; /* More than enough to hold UINT_MAX + "\n"*/
1523         int n;
1524
1525         n = sprintf(buf, "%d\n", fgraph_max_depth);
1526
1527         return simple_read_from_buffer(ubuf, cnt, ppos, buf, n);
1528 }
1529
1530 static const struct file_operations graph_depth_fops = {
1531         .open           = tracing_open_generic,
1532         .write          = graph_depth_write,
1533         .read           = graph_depth_read,
1534         .llseek         = generic_file_llseek,
1535 };
1536
1537 static __init int init_graph_tracefs(void)
1538 {
1539         struct dentry *d_tracer;
1540
1541         d_tracer = tracing_init_dentry();
1542         if (IS_ERR(d_tracer))
1543                 return 0;
1544
1545         trace_create_file("max_graph_depth", 0644, d_tracer,
1546                           NULL, &graph_depth_fops);
1547
1548         return 0;
1549 }
1550 fs_initcall(init_graph_tracefs);
1551
1552 static __init int init_graph_trace(void)
1553 {
1554         max_bytes_for_cpu = snprintf(NULL, 0, "%u", nr_cpu_ids - 1);
1555
1556         if (!register_trace_event(&graph_trace_entry_event)) {
1557                 pr_warn("Warning: could not register graph trace events\n");
1558                 return 1;
1559         }
1560
1561         if (!register_trace_event(&graph_trace_ret_event)) {
1562                 pr_warn("Warning: could not register graph trace events\n");
1563                 return 1;
1564         }
1565
1566         return register_tracer(&graph_trace);
1567 }
1568
1569 core_initcall(init_graph_trace);