kernel/trace/trace_functions_graph.c

   1 /*
   2  *
   3  * Function graph tracer.
   4  * Copyright (c) 2008-2009 Frederic Weisbecker <fweisbec@gmail.com>
   5  * Mostly borrowed from function tracer which
   6  * is Copyright (c) Steven Rostedt <srostedt@redhat.com>
   7  *
   8  */
   9 #include <linux/uaccess.h>
  10 #include <linux/ftrace.h>
  11 #include <linux/slab.h>
  12 #include <linux/fs.h>
  13
  14 #include "trace.h"
  15 #include "trace_output.h"
  16
  17 static bool kill_ftrace_graph;
  18
  19 /**
  20  * ftrace_graph_is_dead - returns true if ftrace_graph_stop() was called
  21  *
  22  * ftrace_graph_stop() is called when a severe error is detected in
  23  * the function graph tracing. This function is called by the critical
  24  * paths of function graph to keep those paths from doing any more harm.
  25  */
  26 bool ftrace_graph_is_dead(void)
  27 {
  28         return kill_ftrace_graph;
  29 }
  30
  31 /**
  32  * ftrace_graph_stop - set to permanently disable function graph tracincg
  33  *
  34  * In case of an error int function graph tracing, this is called
  35  * to try to keep function graph tracing from causing any more harm.
  36  * Usually this is pretty severe and this is called to try to at least
  37  * get a warning out to the user.
  38  */
  39 void ftrace_graph_stop(void)
  40 {
  41         kill_ftrace_graph = true;
  42 }
  43
  44 /* When set, irq functions will be ignored */
  45 static int ftrace_graph_skip_irqs;
  46
  47 struct fgraph_cpu_data {
  48         pid_t           last_pid;
  49         int             depth;
  50         int             depth_irq;
  51         int             ignore;
  52         unsigned long   enter_funcs[FTRACE_RETFUNC_DEPTH];
  53 };
  54
  55 struct fgraph_data {
  56         struct fgraph_cpu_data __percpu *cpu_data;
  57
  58         /* Place to preserve last processed entry. */
  59         struct ftrace_graph_ent_entry   ent;
  60         struct ftrace_graph_ret_entry   ret;
  61         int                             failed;
  62         int                             cpu;
  63 };
  64
  65 #define TRACE_GRAPH_INDENT      2
  66
  67 /* Flag options */
  68 #define TRACE_GRAPH_PRINT_FLAT          0x80
  69
  70 static unsigned int max_depth;
  71
  72 static struct tracer_opt trace_opts[] = {
  73         /* Display overruns? (for self-debug purpose) */
  74         { TRACER_OPT(funcgraph-overrun, TRACE_GRAPH_PRINT_OVERRUN) },
  75         /* Display CPU ? */
  76         { TRACER_OPT(funcgraph-cpu, TRACE_GRAPH_PRINT_CPU) },
  77         /* Display Overhead ? */
  78         { TRACER_OPT(funcgraph-overhead, TRACE_GRAPH_PRINT_OVERHEAD) },
  79         /* Display proc name/pid */
  80         { TRACER_OPT(funcgraph-proc, TRACE_GRAPH_PRINT_PROC) },
  81         /* Display duration of execution */
  82         { TRACER_OPT(funcgraph-duration, TRACE_GRAPH_PRINT_DURATION) },
  83         /* Display absolute time of an entry */
  84         { TRACER_OPT(funcgraph-abstime, TRACE_GRAPH_PRINT_ABS_TIME) },
  85         /* Display interrupts */
  86         { TRACER_OPT(funcgraph-irqs, TRACE_GRAPH_PRINT_IRQS) },
  87         /* Display function name after trailing } */
  88         { TRACER_OPT(funcgraph-tail, TRACE_GRAPH_PRINT_TAIL) },
  89         /* Include sleep time (scheduled out) between entry and return */
  90         { TRACER_OPT(sleep-time, TRACE_GRAPH_SLEEP_TIME) },
  91         /* Include time within nested functions */
  92         { TRACER_OPT(graph-time, TRACE_GRAPH_GRAPH_TIME) },
  93         /* Use standard trace formatting rather than hierarchical */
  94         { TRACER_OPT(funcgraph-flat, TRACE_GRAPH_PRINT_FLAT) },
  95         { } /* Empty entry */
  96 };
  97
  98 static struct tracer_flags tracer_flags = {
  99         /* Don't display overruns, proc, or tail by default */
 100         .val = TRACE_GRAPH_PRINT_CPU | TRACE_GRAPH_PRINT_OVERHEAD |
 101                TRACE_GRAPH_PRINT_DURATION | TRACE_GRAPH_PRINT_IRQS |
 102                TRACE_GRAPH_SLEEP_TIME | TRACE_GRAPH_GRAPH_TIME,
 103         .opts = trace_opts
 104 };
 105
 106 static struct trace_array *graph_array;
 107
 108 /*
 109  * DURATION column is being also used to display IRQ signs,
 110  * following values are used by print_graph_irq and others
 111  * to fill in space into DURATION column.
 112  */
 113 enum {
 114         FLAGS_FILL_FULL  = 1 << TRACE_GRAPH_PRINT_FILL_SHIFT,
 115         FLAGS_FILL_START = 2 << TRACE_GRAPH_PRINT_FILL_SHIFT,
 116         FLAGS_FILL_END   = 3 << TRACE_GRAPH_PRINT_FILL_SHIFT,
 117 };
 118
 119 static void
 120 print_graph_duration(struct trace_array *tr, unsigned long long duration,
 121                      struct trace_seq *s, u32 flags);
 122
 123 /* Add a function return address to the trace stack on thread info.*/
 124 int
 125 ftrace_push_return_trace(unsigned long ret, unsigned long func, int *depth,
 126                          unsigned long frame_pointer)
 127 {
 128         unsigned long long calltime;
 129         int index;
 130
 131         if (unlikely(ftrace_graph_is_dead()))
 132                 return -EBUSY;
 133
 134         if (!current->ret_stack)
 135                 return -EBUSY;
 136
 137         /*
 138          * We must make sure the ret_stack is tested before we read
 139          * anything else.
 140          */
 141         smp_rmb();
 142
 143         /* The return trace stack is full */
 144         if (current->curr_ret_stack == FTRACE_RETFUNC_DEPTH - 1) {
 145                 atomic_inc(&current->trace_overrun);
 146                 return -EBUSY;
 147         }
 148
 149         /*
 150          * The curr_ret_stack is an index to ftrace return stack of
 151          * current task.  Its value should be in [0, FTRACE_RETFUNC_
 152          * DEPTH) when the function graph tracer is used.  To support
 153          * filtering out specific functions, it makes the index
 154          * negative by subtracting huge value (FTRACE_NOTRACE_DEPTH)
 155          * so when it sees a negative index the ftrace will ignore
 156          * the record.  And the index gets recovered when returning
 157          * from the filtered function by adding the FTRACE_NOTRACE_
 158          * DEPTH and then it'll continue to record functions normally.
 159          *
 160          * The curr_ret_stack is initialized to -1 and get increased
 161          * in this function.  So it can be less than -1 only if it was
 162          * filtered out via ftrace_graph_notrace_addr() which can be
 163          * set from set_graph_notrace file in tracefs by user.
 164          */
 165         if (current->curr_ret_stack < -1)
 166                 return -EBUSY;
 167
 168         calltime = trace_clock_local();
 169
 170         index = ++current->curr_ret_stack;
 171         if (ftrace_graph_notrace_addr(func))
 172                 current->curr_ret_stack -= FTRACE_NOTRACE_DEPTH;
 173         barrier();
 174         current->ret_stack[index].ret = ret;
 175         current->ret_stack[index].func = func;
 176         current->ret_stack[index].calltime = calltime;
 177         current->ret_stack[index].subtime = 0;
 178         current->ret_stack[index].fp = frame_pointer;
 179         *depth = current->curr_ret_stack;
 180
 181         return 0;
 182 }
 183
 184 /* Retrieve a function return address to the trace stack on thread info.*/
 185 static void
 186 ftrace_pop_return_trace(struct ftrace_graph_ret *trace, unsigned long *ret,
 187                         unsigned long frame_pointer)
 188 {
 189         int index;
 190
 191         index = current->curr_ret_stack;
 192
 193         /*
 194          * A negative index here means that it's just returned from a
 195          * notrace'd function.  Recover index to get an original
 196          * return address.  See ftrace_push_return_trace().
 197          *
 198          * TODO: Need to check whether the stack gets corrupted.
 199          */
 200         if (index < 0)
 201                 index += FTRACE_NOTRACE_DEPTH;
 202
 203         if (unlikely(index < 0 || index >= FTRACE_RETFUNC_DEPTH)) {
 204                 ftrace_graph_stop();
 205                 WARN_ON(1);
 206                 /* Might as well panic, otherwise we have no where to go */
 207                 *ret = (unsigned long)panic;
 208                 return;
 209         }
 210
 211 #if defined(CONFIG_HAVE_FUNCTION_GRAPH_FP_TEST) && !defined(CC_USING_FENTRY)
 212         /*
 213          * The arch may choose to record the frame pointer used
 214          * and check it here to make sure that it is what we expect it
 215          * to be. If gcc does not set the place holder of the return
 216          * address in the frame pointer, and does a copy instead, then
 217          * the function graph trace will fail. This test detects this
 218          * case.
 219          *
 220          * Currently, x86_32 with optimize for size (-Os) makes the latest
 221          * gcc do the above.
 222          *
 223          * Note, -mfentry does not use frame pointers, and this test
 224          *  is not needed if CC_USING_FENTRY is set.
 225          */
 226         if (unlikely(current->ret_stack[index].fp != frame_pointer)) {
 227                 ftrace_graph_stop();
 228                 WARN(1, "Bad frame pointer: expected %lx, received %lx\n"
 229                      "  from func %ps return to %lx\n",
 230                      current->ret_stack[index].fp,
 231                      frame_pointer,
 232                      (void *)current->ret_stack[index].func,
 233                      current->ret_stack[index].ret);
 234                 *ret = (unsigned long)panic;
 235                 return;
 236         }
 237 #endif
 238
 239         *ret = current->ret_stack[index].ret;
 240         trace->func = current->ret_stack[index].func;
 241         trace->calltime = current->ret_stack[index].calltime;
 242         trace->overrun = atomic_read(&current->trace_overrun);
 243         trace->depth = index;
 244 }
 245
 246 /*
 247  * Send the trace to the ring-buffer.
 248  * @return the original return address.
 249  */
 250 unsigned long ftrace_return_to_handler(unsigned long frame_pointer)
 251 {
 252         struct ftrace_graph_ret trace;
 253         unsigned long ret;
 254
 255         ftrace_pop_return_trace(&trace, &ret, frame_pointer);
 256         trace.rettime = trace_clock_local();
 257         barrier();
 258         current->curr_ret_stack--;
 259         /*
 260          * The curr_ret_stack can be less than -1 only if it was
 261          * filtered out and it's about to return from the function.
 262          * Recover the index and continue to trace normal functions.
 263          */
 264         if (current->curr_ret_stack < -1) {
 265                 current->curr_ret_stack += FTRACE_NOTRACE_DEPTH;
 266                 return ret;
 267         }
 268
 269         /*
 270          * The trace should run after decrementing the ret counter
 271          * in case an interrupt were to come in. We don't want to
 272          * lose the interrupt if max_depth is set.
 273          */
 274         ftrace_graph_return(&trace);
 275
 276         if (unlikely(!ret)) {
 277                 ftrace_graph_stop();
 278                 WARN_ON(1);
 279                 /* Might as well panic. What else to do? */
 280                 ret = (unsigned long)panic;
 281         }
 282
 283         return ret;
 284 }
 285
 286 int __trace_graph_entry(struct trace_array *tr,
 287                                 struct ftrace_graph_ent *trace,
 288                                 unsigned long flags,
 289                                 int pc)
 290 {
 291         struct trace_event_call *call = &event_funcgraph_entry;
 292         struct ring_buffer_event *event;
 293         struct ring_buffer *buffer = tr->trace_buffer.buffer;
 294         struct ftrace_graph_ent_entry *entry;
 295
 296         event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_ENT,
 297                                           sizeof(*entry), flags, pc);
 298         if (!event)
 299                 return 0;
 300         entry   = ring_buffer_event_data(event);
 301         entry->graph_ent                        = *trace;
 302         if (!call_filter_check_discard(call, entry, buffer, event))
 303                 __buffer_unlock_commit(buffer, event);
 304
 305         return 1;
 306 }
 307
 308 static inline int ftrace_graph_ignore_irqs(void)
 309 {
 310         if (!ftrace_graph_skip_irqs || trace_recursion_test(TRACE_IRQ_BIT))
 311                 return 0;
 312
 313         return in_irq();
 314 }
 315
 316 int trace_graph_entry(struct ftrace_graph_ent *trace)
 317 {
 318         struct trace_array *tr = graph_array;
 319         struct trace_array_cpu *data;
 320         unsigned long flags;
 321         long disabled;
 322         int ret;
 323         int cpu;
 324         int pc;
 325
 326         if (!ftrace_trace_task(current))
 327                 return 0;
 328
 329         /* trace it when it is-nested-in or is a function enabled. */
 330         if ((!(trace->depth || ftrace_graph_addr(trace->func)) ||
 331              ftrace_graph_ignore_irqs()) || (trace->depth < 0) ||
 332             (max_depth && trace->depth >= max_depth))
 333                 return 0;
 334
 335         /*
 336          * Do not trace a function if it's filtered by set_graph_notrace.
 337          * Make the index of ret stack negative to indicate that it should
 338          * ignore further functions.  But it needs its own ret stack entry
 339          * to recover the original index in order to continue tracing after
 340          * returning from the function.
 341          */
 342         if (ftrace_graph_notrace_addr(trace->func))
 343                 return 1;
 344
 345         local_irq_save(flags);
 346         cpu = raw_smp_processor_id();
 347         data = per_cpu_ptr(tr->trace_buffer.data, cpu);
 348         disabled = atomic_inc_return(&data->disabled);
 349         if (likely(disabled == 1)) {
 350                 pc = preempt_count();
 351                 ret = __trace_graph_entry(tr, trace, flags, pc);
 352         } else {
 353                 ret = 0;
 354         }
 355
 356         atomic_dec(&data->disabled);
 357         local_irq_restore(flags);
 358
 359         return ret;
 360 }
 361
 362 static int trace_graph_thresh_entry(struct ftrace_graph_ent *trace)
 363 {
 364         if (tracing_thresh)
 365                 return 1;
 366         else
 367                 return trace_graph_entry(trace);
 368 }
 369
 370 static void
 371 __trace_graph_function(struct trace_array *tr,
 372                 unsigned long ip, unsigned long flags, int pc)
 373 {
 374         u64 time = trace_clock_local();
 375         struct ftrace_graph_ent ent = {
 376                 .func  = ip,
 377                 .depth = 0,
 378         };
 379         struct ftrace_graph_ret ret = {
 380                 .func     = ip,
 381                 .depth    = 0,
 382                 .calltime = time,
 383                 .rettime  = time,
 384         };
 385
 386         __trace_graph_entry(tr, &ent, flags, pc);
 387         __trace_graph_return(tr, &ret, flags, pc);
 388 }
 389
 390 void
 391 trace_graph_function(struct trace_array *tr,
 392                 unsigned long ip, unsigned long parent_ip,
 393                 unsigned long flags, int pc)
 394 {
 395         __trace_graph_function(tr, ip, flags, pc);
 396 }
 397
 398 void __trace_graph_return(struct trace_array *tr,
 399                                 struct ftrace_graph_ret *trace,
 400                                 unsigned long flags,
 401                                 int pc)
 402 {
 403         struct trace_event_call *call = &event_funcgraph_exit;
 404         struct ring_buffer_event *event;
 405         struct ring_buffer *buffer = tr->trace_buffer.buffer;
 406         struct ftrace_graph_ret_entry *entry;
 407
 408         event = trace_buffer_lock_reserve(buffer, TRACE_GRAPH_RET,
 409                                           sizeof(*entry), flags, pc);
 410         if (!event)
 411                 return;
 412         entry   = ring_buffer_event_data(event);
 413         entry->ret                              = *trace;
 414         if (!call_filter_check_discard(call, entry, buffer, event))
 415                 __buffer_unlock_commit(buffer, event);
 416 }
 417
 418 void trace_graph_return(struct ftrace_graph_ret *trace)
 419 {
 420         struct trace_array *tr = graph_array;
 421         struct trace_array_cpu *data;
 422         unsigned long flags;
 423         long disabled;
 424         int cpu;
 425         int pc;
 426
 427         local_irq_save(flags);
 428         cpu = raw_smp_processor_id();
 429         data = per_cpu_ptr(tr->trace_buffer.data, cpu);
 430         disabled = atomic_inc_return(&data->disabled);
 431         if (likely(disabled == 1)) {
 432                 pc = preempt_count();
 433                 __trace_graph_return(tr, trace, flags, pc);
 434         }
 435         atomic_dec(&data->disabled);
 436         local_irq_restore(flags);
 437 }
 438
 439 void set_graph_array(struct trace_array *tr)
 440 {
 441         graph_array = tr;
 442
 443         /* Make graph_array visible before we start tracing */
 444
 445         smp_mb();
 446 }
 447
 448 static void trace_graph_thresh_return(struct ftrace_graph_ret *trace)
 449 {
 450         if (tracing_thresh &&
 451             (trace->rettime - trace->calltime < tracing_thresh))
 452                 return;
 453         else
 454                 trace_graph_return(trace);
 455 }
 456
 457 static int graph_trace_init(struct trace_array *tr)
 458 {
 459         int ret;
 460
 461         set_graph_array(tr);
 462         if (tracing_thresh)
 463                 ret = register_ftrace_graph(&trace_graph_thresh_return,
 464                                             &trace_graph_thresh_entry);
 465         else
 466                 ret = register_ftrace_graph(&trace_graph_return,
 467                                             &trace_graph_entry);
 468         if (ret)
 469                 return ret;
 470         tracing_start_cmdline_record();
 471
 472         return 0;
 473 }
 474
 475 static void graph_trace_reset(struct trace_array *tr)
 476 {
 477         tracing_stop_cmdline_record();
 478         unregister_ftrace_graph();
 479 }
 480
 481 static int graph_trace_update_thresh(struct trace_array *tr)
 482 {
 483         graph_trace_reset(tr);
 484         return graph_trace_init(tr);
 485 }
 486
 487 static int max_bytes_for_cpu;
 488
 489 static void print_graph_cpu(struct trace_seq *s, int cpu)
 490 {
 491         /*
 492          * Start with a space character - to make it stand out
 493          * to the right a bit when trace output is pasted into
 494          * email:
 495          */
 496         trace_seq_printf(s, " %*d) ", max_bytes_for_cpu, cpu);
 497 }
 498
 499 #define TRACE_GRAPH_PROCINFO_LENGTH     14
 500
 501 static void print_graph_proc(struct trace_seq *s, pid_t pid)
 502 {
 503         char comm[TASK_COMM_LEN];
 504         /* sign + log10(MAX_INT) + '\0' */
 505         char pid_str[11];
 506         int spaces = 0;
 507         int len;
 508         int i;
 509
 510         trace_find_cmdline(pid, comm);
 511         comm[7] = '\0';
 512         sprintf(pid_str, "%d", pid);
 513
 514         /* 1 stands for the "-" character */
 515         len = strlen(comm) + strlen(pid_str) + 1;
 516
 517         if (len < TRACE_GRAPH_PROCINFO_LENGTH)
 518                 spaces = TRACE_GRAPH_PROCINFO_LENGTH - len;
 519
 520         /* First spaces to align center */
 521         for (i = 0; i < spaces / 2; i++)
 522                 trace_seq_putc(s, ' ');
 523
 524         trace_seq_printf(s, "%s-%s", comm, pid_str);
 525
 526         /* Last spaces to align center */
 527         for (i = 0; i < spaces - (spaces / 2); i++)
 528                 trace_seq_putc(s, ' ');
 529 }
 530
 531
 532 static void print_graph_lat_fmt(struct trace_seq *s, struct trace_entry *entry)
 533 {
 534         trace_seq_putc(s, ' ');
 535         trace_print_lat_fmt(s, entry);
 536 }
 537
 538 /* If the pid changed since the last trace, output this event */
 539 static void
 540 verif_pid(struct trace_seq *s, pid_t pid, int cpu, struct fgraph_data *data)
 541 {
 542         pid_t prev_pid;
 543         pid_t *last_pid;
 544
 545         if (!data)
 546                 return;
 547
 548         last_pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
 549
 550         if (*last_pid == pid)
 551                 return;
 552
 553         prev_pid = *last_pid;
 554         *last_pid = pid;
 555
 556         if (prev_pid == -1)
 557                 return;
 558 /*
 559  * Context-switch trace line:
 560
 561  ------------------------------------------
 562  | 1)  migration/0--1  =>  sshd-1755
 563  ------------------------------------------
 564
 565  */
 566         trace_seq_puts(s, " ------------------------------------------\n");
 567         print_graph_cpu(s, cpu);
 568         print_graph_proc(s, prev_pid);
 569         trace_seq_puts(s, " => ");
 570         print_graph_proc(s, pid);
 571         trace_seq_puts(s, "\n ------------------------------------------\n\n");
 572 }
 573
 574 static struct ftrace_graph_ret_entry *
 575 get_return_for_leaf(struct trace_iterator *iter,
 576                 struct ftrace_graph_ent_entry *curr)
 577 {
 578         struct fgraph_data *data = iter->private;
 579         struct ring_buffer_iter *ring_iter = NULL;
 580         struct ring_buffer_event *event;
 581         struct ftrace_graph_ret_entry *next;
 582
 583         /*
 584          * If the previous output failed to write to the seq buffer,
 585          * then we just reuse the data from before.
 586          */
 587         if (data && data->failed) {
 588                 curr = &data->ent;
 589                 next = &data->ret;
 590         } else {
 591
 592                 ring_iter = trace_buffer_iter(iter, iter->cpu);
 593
 594                 /* First peek to compare current entry and the next one */
 595                 if (ring_iter)
 596                         event = ring_buffer_iter_peek(ring_iter, NULL);
 597                 else {
 598                         /*
 599                          * We need to consume the current entry to see
 600                          * the next one.
 601                          */
 602                         ring_buffer_consume(iter->trace_buffer->buffer, iter->cpu,
 603                                             NULL, NULL);
 604                         event = ring_buffer_peek(iter->trace_buffer->buffer, iter->cpu,
 605                                                  NULL, NULL);
 606                 }
 607
 608                 if (!event)
 609                         return NULL;
 610
 611                 next = ring_buffer_event_data(event);
 612
 613                 if (data) {
 614                         /*
 615                          * Save current and next entries for later reference
 616                          * if the output fails.
 617                          */
 618                         data->ent = *curr;
 619                         /*
 620                          * If the next event is not a return type, then
 621                          * we only care about what type it is. Otherwise we can
 622                          * safely copy the entire event.
 623                          */
 624                         if (next->ent.type == TRACE_GRAPH_RET)
 625                                 data->ret = *next;
 626                         else
 627                                 data->ret.ent.type = next->ent.type;
 628                 }
 629         }
 630
 631         if (next->ent.type != TRACE_GRAPH_RET)
 632                 return NULL;
 633
 634         if (curr->ent.pid != next->ent.pid ||
 635                         curr->graph_ent.func != next->ret.func)
 636                 return NULL;
 637
 638         /* this is a leaf, now advance the iterator */
 639         if (ring_iter)
 640                 ring_buffer_read(ring_iter, NULL);
 641
 642         return next;
 643 }
 644
 645 static void print_graph_abs_time(u64 t, struct trace_seq *s)
 646 {
 647         unsigned long usecs_rem;
 648
 649         usecs_rem = do_div(t, NSEC_PER_SEC);
 650         usecs_rem /= 1000;
 651
 652         trace_seq_printf(s, "%5lu.%06lu |  ",
 653                          (unsigned long)t, usecs_rem);
 654 }
 655
 656 static void
 657 print_graph_irq(struct trace_iterator *iter, unsigned long addr,
 658                 enum trace_type type, int cpu, pid_t pid, u32 flags)
 659 {
 660         struct trace_array *tr = iter->tr;
 661         struct trace_seq *s = &iter->seq;
 662         struct trace_entry *ent = iter->ent;
 663
 664         if (addr < (unsigned long)__irqentry_text_start ||
 665                 addr >= (unsigned long)__irqentry_text_end)
 666                 return;
 667
 668         if (tr->trace_flags & TRACE_ITER_CONTEXT_INFO) {
 669                 /* Absolute time */
 670                 if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
 671                         print_graph_abs_time(iter->ts, s);
 672
 673                 /* Cpu */
 674                 if (flags & TRACE_GRAPH_PRINT_CPU)
 675                         print_graph_cpu(s, cpu);
 676
 677                 /* Proc */
 678                 if (flags & TRACE_GRAPH_PRINT_PROC) {
 679                         print_graph_proc(s, pid);
 680                         trace_seq_puts(s, " | ");
 681                 }
 682
 683                 /* Latency format */
 684                 if (tr->trace_flags & TRACE_ITER_LATENCY_FMT)
 685                         print_graph_lat_fmt(s, ent);
 686         }
 687
 688         /* No overhead */
 689         print_graph_duration(tr, 0, s, flags | FLAGS_FILL_START);
 690
 691         if (type == TRACE_GRAPH_ENT)
 692                 trace_seq_puts(s, "==========>");
 693         else
 694                 trace_seq_puts(s, "<==========");
 695
 696         print_graph_duration(tr, 0, s, flags | FLAGS_FILL_END);
 697         trace_seq_putc(s, '\n');
 698 }
 699
 700 void
 701 trace_print_graph_duration(unsigned long long duration, struct trace_seq *s)
 702 {
 703         unsigned long nsecs_rem = do_div(duration, 1000);
 704         /* log10(ULONG_MAX) + '\0' */
 705         char usecs_str[21];
 706         char nsecs_str[5];
 707         int len;
 708         int i;
 709
 710         sprintf(usecs_str, "%lu", (unsigned long) duration);
 711
 712         /* Print msecs */
 713         trace_seq_printf(s, "%s", usecs_str);
 714
 715         len = strlen(usecs_str);
 716
 717         /* Print nsecs (we don't want to exceed 7 numbers) */
 718         if (len < 7) {
 719                 size_t slen = min_t(size_t, sizeof(nsecs_str), 8UL - len);
 720
 721                 snprintf(nsecs_str, slen, "%03lu", nsecs_rem);
 722                 trace_seq_printf(s, ".%s", nsecs_str);
 723                 len += strlen(nsecs_str) + 1;
 724         }
 725
 726         trace_seq_puts(s, " us ");
 727
 728         /* Print remaining spaces to fit the row's width */
 729         for (i = len; i < 8; i++)
 730                 trace_seq_putc(s, ' ');
 731 }
 732
 733 static void
 734 print_graph_duration(struct trace_array *tr, unsigned long long duration,
 735                      struct trace_seq *s, u32 flags)
 736 {
 737         if (!(flags & TRACE_GRAPH_PRINT_DURATION) ||
 738             !(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
 739                 return;
 740
 741         /* No real adata, just filling the column with spaces */
 742         switch (flags & TRACE_GRAPH_PRINT_FILL_MASK) {
 743         case FLAGS_FILL_FULL:
 744                 trace_seq_puts(s, "              |  ");
 745                 return;
 746         case FLAGS_FILL_START:
 747                 trace_seq_puts(s, "  ");
 748                 return;
 749         case FLAGS_FILL_END:
 750                 trace_seq_puts(s, " |");
 751                 return;
 752         }
 753
 754         /* Signal a overhead of time execution to the output */
 755         if (flags & TRACE_GRAPH_PRINT_OVERHEAD)
 756                 trace_seq_printf(s, "%c ", trace_find_mark(duration));
 757         else
 758                 trace_seq_puts(s, "  ");
 759
 760         trace_print_graph_duration(duration, s);
 761         trace_seq_puts(s, "|  ");
 762 }
 763
 764 /* Case of a leaf function on its call entry */
 765 static enum print_line_t
 766 print_graph_entry_leaf(struct trace_iterator *iter,
 767                 struct ftrace_graph_ent_entry *entry,
 768                 struct ftrace_graph_ret_entry *ret_entry,
 769                 struct trace_seq *s, u32 flags)
 770 {
 771         struct fgraph_data *data = iter->private;
 772         struct trace_array *tr = iter->tr;
 773         struct ftrace_graph_ret *graph_ret;
 774         struct ftrace_graph_ent *call;
 775         unsigned long long duration;
 776         int i;
 777
 778         graph_ret = &ret_entry->ret;
 779         call = &entry->graph_ent;
 780         duration = graph_ret->rettime - graph_ret->calltime;
 781
 782         if (data) {
 783                 struct fgraph_cpu_data *cpu_data;
 784                 int cpu = iter->cpu;
 785
 786                 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
 787
 788                 /* If a graph tracer ignored set_graph_notrace */
 789                 if (call->depth < -1)
 790                         call->depth += FTRACE_NOTRACE_DEPTH;
 791
 792                 /*
 793                  * Comments display at + 1 to depth. Since
 794                  * this is a leaf function, keep the comments
 795                  * equal to this depth.
 796                  */
 797                 cpu_data->depth = call->depth - 1;
 798
 799                 /* No need to keep this function around for this depth */
 800                 if (call->depth < FTRACE_RETFUNC_DEPTH &&
 801                     !WARN_ON_ONCE(call->depth < 0))
 802                         cpu_data->enter_funcs[call->depth] = 0;
 803         }
 804
 805         /* Overhead and duration */
 806         print_graph_duration(tr, duration, s, flags);
 807
 808         /* Function */
 809         for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
 810                 trace_seq_putc(s, ' ');
 811
 812         trace_seq_printf(s, "%ps();\n", (void *)call->func);
 813
 814         return trace_handle_return(s);
 815 }
 816
 817 static enum print_line_t
 818 print_graph_entry_nested(struct trace_iterator *iter,
 819                          struct ftrace_graph_ent_entry *entry,
 820                          struct trace_seq *s, int cpu, u32 flags)
 821 {
 822         struct ftrace_graph_ent *call = &entry->graph_ent;
 823         struct fgraph_data *data = iter->private;
 824         struct trace_array *tr = iter->tr;
 825         int i;
 826
 827         if (data) {
 828                 struct fgraph_cpu_data *cpu_data;
 829                 int cpu = iter->cpu;
 830
 831                 /* If a graph tracer ignored set_graph_notrace */
 832                 if (call->depth < -1)
 833                         call->depth += FTRACE_NOTRACE_DEPTH;
 834
 835                 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
 836                 cpu_data->depth = call->depth;
 837
 838                 /* Save this function pointer to see if the exit matches */
 839                 if (call->depth < FTRACE_RETFUNC_DEPTH &&
 840                     !WARN_ON_ONCE(call->depth < 0))
 841                         cpu_data->enter_funcs[call->depth] = call->func;
 842         }
 843
 844         /* No time */
 845         print_graph_duration(tr, 0, s, flags | FLAGS_FILL_FULL);
 846
 847         /* Function */
 848         for (i = 0; i < call->depth * TRACE_GRAPH_INDENT; i++)
 849                 trace_seq_putc(s, ' ');
 850
 851         trace_seq_printf(s, "%ps() {\n", (void *)call->func);
 852
 853         if (trace_seq_has_overflowed(s))
 854                 return TRACE_TYPE_PARTIAL_LINE;
 855
 856         /*
 857          * we already consumed the current entry to check the next one
 858          * and see if this is a leaf.
 859          */
 860         return TRACE_TYPE_NO_CONSUME;
 861 }
 862
 863 static void
 864 print_graph_prologue(struct trace_iterator *iter, struct trace_seq *s,
 865                      int type, unsigned long addr, u32 flags)
 866 {
 867         struct fgraph_data *data = iter->private;
 868         struct trace_entry *ent = iter->ent;
 869         struct trace_array *tr = iter->tr;
 870         int cpu = iter->cpu;
 871
 872         /* Pid */
 873         verif_pid(s, ent->pid, cpu, data);
 874
 875         if (type)
 876                 /* Interrupt */
 877                 print_graph_irq(iter, addr, type, cpu, ent->pid, flags);
 878
 879         if (!(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
 880                 return;
 881
 882         /* Absolute time */
 883         if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
 884                 print_graph_abs_time(iter->ts, s);
 885
 886         /* Cpu */
 887         if (flags & TRACE_GRAPH_PRINT_CPU)
 888                 print_graph_cpu(s, cpu);
 889
 890         /* Proc */
 891         if (flags & TRACE_GRAPH_PRINT_PROC) {
 892                 print_graph_proc(s, ent->pid);
 893                 trace_seq_puts(s, " | ");
 894         }
 895
 896         /* Latency format */
 897         if (tr->trace_flags & TRACE_ITER_LATENCY_FMT)
 898                 print_graph_lat_fmt(s, ent);
 899
 900         return;
 901 }
 902
 903 /*
 904  * Entry check for irq code
 905  *
 906  * returns 1 if
 907  *  - we are inside irq code
 908  *  - we just entered irq code
 909  *
 910  * retunns 0 if
 911  *  - funcgraph-interrupts option is set
 912  *  - we are not inside irq code
 913  */
 914 static int
 915 check_irq_entry(struct trace_iterator *iter, u32 flags,
 916                 unsigned long addr, int depth)
 917 {
 918         int cpu = iter->cpu;
 919         int *depth_irq;
 920         struct fgraph_data *data = iter->private;
 921
 922         /*
 923          * If we are either displaying irqs, or we got called as
 924          * a graph event and private data does not exist,
 925          * then we bypass the irq check.
 926          */
 927         if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
 928             (!data))
 929                 return 0;
 930
 931         depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
 932
 933         /*
 934          * We are inside the irq code
 935          */
 936         if (*depth_irq >= 0)
 937                 return 1;
 938
 939         if ((addr < (unsigned long)__irqentry_text_start) ||
 940             (addr >= (unsigned long)__irqentry_text_end))
 941                 return 0;
 942
 943         /*
 944          * We are entering irq code.
 945          */
 946         *depth_irq = depth;
 947         return 1;
 948 }
 949
 950 /*
 951  * Return check for irq code
 952  *
 953  * returns 1 if
 954  *  - we are inside irq code
 955  *  - we just left irq code
 956  *
 957  * returns 0 if
 958  *  - funcgraph-interrupts option is set
 959  *  - we are not inside irq code
 960  */
 961 static int
 962 check_irq_return(struct trace_iterator *iter, u32 flags, int depth)
 963 {
 964         int cpu = iter->cpu;
 965         int *depth_irq;
 966         struct fgraph_data *data = iter->private;
 967
 968         /*
 969          * If we are either displaying irqs, or we got called as
 970          * a graph event and private data does not exist,
 971          * then we bypass the irq check.
 972          */
 973         if ((flags & TRACE_GRAPH_PRINT_IRQS) ||
 974             (!data))
 975                 return 0;
 976
 977         depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
 978
 979         /*
 980          * We are not inside the irq code.
 981          */
 982         if (*depth_irq == -1)
 983                 return 0;
 984
 985         /*
 986          * We are inside the irq code, and this is returning entry.
 987          * Let's not trace it and clear the entry depth, since
 988          * we are out of irq code.
 989          *
 990          * This condition ensures that we 'leave the irq code' once
 991          * we are out of the entry depth. Thus protecting us from
 992          * the RETURN entry loss.
 993          */
 994         if (*depth_irq >= depth) {
 995                 *depth_irq = -1;
 996                 return 1;
 997         }
 998
 999         /*
1000          * We are inside the irq code, and this is not the entry.
1001          */
1002         return 1;
1003 }
1004
1005 static enum print_line_t
1006 print_graph_entry(struct ftrace_graph_ent_entry *field, struct trace_seq *s,
1007                         struct trace_iterator *iter, u32 flags)
1008 {
1009         struct fgraph_data *data = iter->private;
1010         struct ftrace_graph_ent *call = &field->graph_ent;
1011         struct ftrace_graph_ret_entry *leaf_ret;
1012         static enum print_line_t ret;
1013         int cpu = iter->cpu;
1014
1015         if (check_irq_entry(iter, flags, call->func, call->depth))
1016                 return TRACE_TYPE_HANDLED;
1017
1018         print_graph_prologue(iter, s, TRACE_GRAPH_ENT, call->func, flags);
1019
1020         leaf_ret = get_return_for_leaf(iter, field);
1021         if (leaf_ret)
1022                 ret = print_graph_entry_leaf(iter, field, leaf_ret, s, flags);
1023         else
1024                 ret = print_graph_entry_nested(iter, field, s, cpu, flags);
1025
1026         if (data) {
1027                 /*
1028                  * If we failed to write our output, then we need to make
1029                  * note of it. Because we already consumed our entry.
1030                  */
1031                 if (s->full) {
1032                         data->failed = 1;
1033                         data->cpu = cpu;
1034                 } else
1035                         data->failed = 0;
1036         }
1037
1038         return ret;
1039 }
1040
1041 static enum print_line_t
1042 print_graph_return(struct ftrace_graph_ret *trace, struct trace_seq *s,
1043                    struct trace_entry *ent, struct trace_iterator *iter,
1044                    u32 flags)
1045 {
1046         unsigned long long duration = trace->rettime - trace->calltime;
1047         struct fgraph_data *data = iter->private;
1048         struct trace_array *tr = iter->tr;
1049         pid_t pid = ent->pid;
1050         int cpu = iter->cpu;
1051         int func_match = 1;
1052         int i;
1053
1054         if (check_irq_return(iter, flags, trace->depth))
1055                 return TRACE_TYPE_HANDLED;
1056
1057         if (data) {
1058                 struct fgraph_cpu_data *cpu_data;
1059                 int cpu = iter->cpu;
1060
1061                 cpu_data = per_cpu_ptr(data->cpu_data, cpu);
1062
1063                 /*
1064                  * Comments display at + 1 to depth. This is the
1065                  * return from a function, we now want the comments
1066                  * to display at the same level of the bracket.
1067                  */
1068                 cpu_data->depth = trace->depth - 1;
1069
1070                 if (trace->depth < FTRACE_RETFUNC_DEPTH &&
1071                     !WARN_ON_ONCE(trace->depth < 0)) {
1072                         if (cpu_data->enter_funcs[trace->depth] != trace->func)
1073                                 func_match = 0;
1074                         cpu_data->enter_funcs[trace->depth] = 0;
1075                 }
1076         }
1077
1078         print_graph_prologue(iter, s, 0, 0, flags);
1079
1080         /* Overhead and duration */
1081         print_graph_duration(tr, duration, s, flags);
1082
1083         /* Closing brace */
1084         for (i = 0; i < trace->depth * TRACE_GRAPH_INDENT; i++)
1085                 trace_seq_putc(s, ' ');
1086
1087         /*
1088          * If the return function does not have a matching entry,
1089          * then the entry was lost. Instead of just printing
1090          * the '}' and letting the user guess what function this
1091          * belongs to, write out the function name. Always do
1092          * that if the funcgraph-tail option is enabled.
1093          */
1094         if (func_match && !(flags & TRACE_GRAPH_PRINT_TAIL))
1095                 trace_seq_puts(s, "}\n");
1096         else
1097                 trace_seq_printf(s, "} /* %ps */\n", (void *)trace->func);
1098
1099         /* Overrun */
1100         if (flags & TRACE_GRAPH_PRINT_OVERRUN)
1101                 trace_seq_printf(s, " (Overruns: %lu)\n",
1102                                  trace->overrun);
1103
1104         print_graph_irq(iter, trace->func, TRACE_GRAPH_RET,
1105                         cpu, pid, flags);
1106
1107         return trace_handle_return(s);
1108 }
1109
1110 static enum print_line_t
1111 print_graph_comment(struct trace_seq *s, struct trace_entry *ent,
1112                     struct trace_iterator *iter, u32 flags)
1113 {
1114         struct trace_array *tr = iter->tr;
1115         unsigned long sym_flags = (tr->trace_flags & TRACE_ITER_SYM_MASK);
1116         struct fgraph_data *data = iter->private;
1117         struct trace_event *event;
1118         int depth = 0;
1119         int ret;
1120         int i;
1121
1122         if (data)
1123                 depth = per_cpu_ptr(data->cpu_data, iter->cpu)->depth;
1124
1125         print_graph_prologue(iter, s, 0, 0, flags);
1126
1127         /* No time */
1128         print_graph_duration(tr, 0, s, flags | FLAGS_FILL_FULL);
1129
1130         /* Indentation */
1131         if (depth > 0)
1132                 for (i = 0; i < (depth + 1) * TRACE_GRAPH_INDENT; i++)
1133                         trace_seq_putc(s, ' ');
1134
1135         /* The comment */
1136         trace_seq_puts(s, "/* ");
1137
1138         switch (iter->ent->type) {
1139         case TRACE_BPRINT:
1140                 ret = trace_print_bprintk_msg_only(iter);
1141                 if (ret != TRACE_TYPE_HANDLED)
1142                         return ret;
1143                 break;
1144         case TRACE_PRINT:
1145                 ret = trace_print_printk_msg_only(iter);
1146                 if (ret != TRACE_TYPE_HANDLED)
1147                         return ret;
1148                 break;
1149         default:
1150                 event = ftrace_find_event(ent->type);
1151                 if (!event)
1152                         return TRACE_TYPE_UNHANDLED;
1153
1154                 ret = event->funcs->trace(iter, sym_flags, event);
1155                 if (ret != TRACE_TYPE_HANDLED)
1156                         return ret;
1157         }
1158
1159         if (trace_seq_has_overflowed(s))
1160                 goto out;
1161
1162         /* Strip ending newline */
1163         if (s->buffer[s->seq.len - 1] == '\n') {
1164                 s->buffer[s->seq.len - 1] = '\0';
1165                 s->seq.len--;
1166         }
1167
1168         trace_seq_puts(s, " */\n");
1169  out:
1170         return trace_handle_return(s);
1171 }
1172
1173
1174 enum print_line_t
1175 print_graph_function_flags(struct trace_iterator *iter, u32 flags)
1176 {
1177         struct ftrace_graph_ent_entry *field;
1178         struct fgraph_data *data = iter->private;
1179         struct trace_entry *entry = iter->ent;
1180         struct trace_seq *s = &iter->seq;
1181         int cpu = iter->cpu;
1182         int ret;
1183
1184         if (flags & TRACE_GRAPH_PRINT_FLAT)
1185                 return TRACE_TYPE_UNHANDLED;
1186
1187         if (data && per_cpu_ptr(data->cpu_data, cpu)->ignore) {
1188                 per_cpu_ptr(data->cpu_data, cpu)->ignore = 0;
1189                 return TRACE_TYPE_HANDLED;
1190         }
1191
1192         /*
1193          * If the last output failed, there's a possibility we need
1194          * to print out the missing entry which would never go out.
1195          */
1196         if (data && data->failed) {
1197                 field = &data->ent;
1198                 iter->cpu = data->cpu;
1199                 ret = print_graph_entry(field, s, iter, flags);
1200                 if (ret == TRACE_TYPE_HANDLED && iter->cpu != cpu) {
1201                         per_cpu_ptr(data->cpu_data, iter->cpu)->ignore = 1;
1202                         ret = TRACE_TYPE_NO_CONSUME;
1203                 }
1204                 iter->cpu = cpu;
1205                 return ret;
1206         }
1207
1208         switch (entry->type) {
1209         case TRACE_GRAPH_ENT: {
1210                 /*
1211                  * print_graph_entry() may consume the current event,
1212                  * thus @field may become invalid, so we need to save it.
1213                  * sizeof(struct ftrace_graph_ent_entry) is very small,
1214                  * it can be safely saved at the stack.
1215                  */
1216                 struct ftrace_graph_ent_entry saved;
1217                 trace_assign_type(field, entry);
1218                 saved = *field;
1219                 return print_graph_entry(&saved, s, iter, flags);
1220         }
1221         case TRACE_GRAPH_RET: {
1222                 struct ftrace_graph_ret_entry *field;
1223                 trace_assign_type(field, entry);
1224                 return print_graph_return(&field->ret, s, entry, iter, flags);
1225         }
1226         case TRACE_STACK:
1227         case TRACE_FN:
1228                 /* dont trace stack and functions as comments */
1229                 return TRACE_TYPE_UNHANDLED;
1230
1231         default:
1232                 return print_graph_comment(s, entry, iter, flags);
1233         }
1234
1235         return TRACE_TYPE_HANDLED;
1236 }
1237
1238 static enum print_line_t
1239 print_graph_function(struct trace_iterator *iter)
1240 {
1241         return print_graph_function_flags(iter, tracer_flags.val);
1242 }
1243
1244 static void print_lat_header(struct seq_file *s, u32 flags)
1245 {
1246         static const char spaces[] = "                " /* 16 spaces */
1247                 "    "                                  /* 4 spaces */
1248                 "                 ";                    /* 17 spaces */
1249         int size = 0;
1250
1251         if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1252                 size += 16;
1253         if (flags & TRACE_GRAPH_PRINT_CPU)
1254                 size += 4;
1255         if (flags & TRACE_GRAPH_PRINT_PROC)
1256                 size += 17;
1257
1258         seq_printf(s, "#%.*s  _-----=> irqs-off        \n", size, spaces);
1259         seq_printf(s, "#%.*s / _----=> need-resched    \n", size, spaces);
1260         seq_printf(s, "#%.*s| / _---=> hardirq/softirq \n", size, spaces);
1261         seq_printf(s, "#%.*s|| / _--=> preempt-depth   \n", size, spaces);
1262         seq_printf(s, "#%.*s||| /                      \n", size, spaces);
1263 }
1264
1265 static void __print_graph_headers_flags(struct trace_array *tr,
1266                                         struct seq_file *s, u32 flags)
1267 {
1268         int lat = tr->trace_flags & TRACE_ITER_LATENCY_FMT;
1269
1270         if (lat)
1271                 print_lat_header(s, flags);
1272
1273         /* 1st line */
1274         seq_putc(s, '#');
1275         if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1276                 seq_puts(s, "     TIME       ");
1277         if (flags & TRACE_GRAPH_PRINT_CPU)
1278                 seq_puts(s, " CPU");
1279         if (flags & TRACE_GRAPH_PRINT_PROC)
1280                 seq_puts(s, "  TASK/PID       ");
1281         if (lat)
1282                 seq_puts(s, "||||");
1283         if (flags & TRACE_GRAPH_PRINT_DURATION)
1284                 seq_puts(s, "  DURATION   ");
1285         seq_puts(s, "               FUNCTION CALLS\n");
1286
1287         /* 2nd line */
1288         seq_putc(s, '#');
1289         if (flags & TRACE_GRAPH_PRINT_ABS_TIME)
1290                 seq_puts(s, "      |         ");
1291         if (flags & TRACE_GRAPH_PRINT_CPU)
1292                 seq_puts(s, " |  ");
1293         if (flags & TRACE_GRAPH_PRINT_PROC)
1294                 seq_puts(s, "   |    |        ");
1295         if (lat)
1296                 seq_puts(s, "||||");
1297         if (flags & TRACE_GRAPH_PRINT_DURATION)
1298                 seq_puts(s, "   |   |      ");
1299         seq_puts(s, "               |   |   |   |\n");
1300 }
1301
1302 static void print_graph_headers(struct seq_file *s)
1303 {
1304         print_graph_headers_flags(s, tracer_flags.val);
1305 }
1306
1307 void print_graph_headers_flags(struct seq_file *s, u32 flags)
1308 {
1309         struct trace_iterator *iter = s->private;
1310         struct trace_array *tr = iter->tr;
1311
1312         if (flags & TRACE_GRAPH_PRINT_FLAT) {
1313                 trace_default_header(s);
1314                 return;
1315         }
1316
1317         if (!(tr->trace_flags & TRACE_ITER_CONTEXT_INFO))
1318                 return;
1319
1320         if (tr->trace_flags & TRACE_ITER_LATENCY_FMT) {
1321                 /* print nothing if the buffers are empty */
1322                 if (trace_empty(iter))
1323                         return;
1324
1325                 print_trace_header(s, iter);
1326         }
1327
1328         __print_graph_headers_flags(tr, s, flags);
1329 }
1330
1331 void graph_trace_open(struct trace_iterator *iter)
1332 {
1333         /* pid and depth on the last trace processed */
1334         struct fgraph_data *data;
1335         gfp_t gfpflags;
1336         int cpu;
1337
1338         iter->private = NULL;
1339
1340         /* We can be called in atomic context via ftrace_dump() */
1341         gfpflags = (in_atomic() || irqs_disabled()) ? GFP_ATOMIC : GFP_KERNEL;
1342
1343         data = kzalloc(sizeof(*data), gfpflags);
1344         if (!data)
1345                 goto out_err;
1346
1347         data->cpu_data = alloc_percpu_gfp(struct fgraph_cpu_data, gfpflags);
1348         if (!data->cpu_data)
1349                 goto out_err_free;
1350
1351         for_each_possible_cpu(cpu) {
1352                 pid_t *pid = &(per_cpu_ptr(data->cpu_data, cpu)->last_pid);
1353                 int *depth = &(per_cpu_ptr(data->cpu_data, cpu)->depth);
1354                 int *ignore = &(per_cpu_ptr(data->cpu_data, cpu)->ignore);
1355                 int *depth_irq = &(per_cpu_ptr(data->cpu_data, cpu)->depth_irq);
1356
1357                 *pid = -1;
1358                 *depth = 0;
1359                 *ignore = 0;
1360                 *depth_irq = -1;
1361         }
1362
1363         iter->private = data;
1364
1365         return;
1366
1367  out_err_free:
1368         kfree(data);
1369  out_err:
1370         pr_warning("function graph tracer: not enough memory\n");
1371 }
1372
1373 void graph_trace_close(struct trace_iterator *iter)
1374 {
1375         struct fgraph_data *data = iter->private;
1376
1377         if (data) {
1378                 free_percpu(data->cpu_data);
1379                 kfree(data);
1380         }
1381 }
1382
1383 static int
1384 func_graph_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set)
1385 {
1386         if (bit == TRACE_GRAPH_PRINT_IRQS)
1387                 ftrace_graph_skip_irqs = !set;
1388
1389         if (bit == TRACE_GRAPH_SLEEP_TIME)
1390                 ftrace_graph_sleep_time_control(set);
1391
1392         if (bit == TRACE_GRAPH_GRAPH_TIME)
1393                 ftrace_graph_graph_time_control(set);
1394
1395         return 0;
1396 }
1397
1398
1399 static struct tracer graph_trace __tracer_data = {
1400         .name           = "function_graph",
1401         .update_thresh  = graph_trace_update_thresh,
1402         .open           = graph_trace_open,
1403         .pipe_open      = graph_trace_open,
1404         .close          = graph_trace_close,
1405         .pipe_close     = graph_trace_close,
1406         .init           = graph_trace_init,
1407         .reset          = graph_trace_reset,
1408         .print_line     = print_graph_function,
1409         .print_header   = print_graph_headers,
1410         .flags          = &tracer_flags,
1411         .set_flag       = func_graph_set_flag,
1412 #ifdef CONFIG_FTRACE_SELFTEST
1413         .selftest       = trace_selftest_startup_function_graph,
1414 #endif
1415 };
1416
1417
1418 static ssize_t
1419 graph_depth_write(struct file *filp, const char __user *ubuf, size_t cnt,
1420                   loff_t *ppos)
1421 {
1422         unsigned long val;
1423         int ret;
1424
1425         ret = kstrtoul_from_user(ubuf, cnt, 10, &val);
1426         if (ret)
1427                 return ret;
1428
1429         max_depth = val;
1430
1431         *ppos += cnt;
1432
1433         return cnt;
1434 }
1435
1436 static ssize_t
1437 graph_depth_read(struct file *filp, char __user *ubuf, size_t cnt,
1438                  loff_t *ppos)
1439 {
1440         char buf[15]; /* More than enough to hold UINT_MAX + "\n"*/
1441         int n;
1442
1443         n = sprintf(buf, "%d\n", max_depth);
1444
1445         return simple_read_from_buffer(ubuf, cnt, ppos, buf, n);
1446 }
1447
1448 static const struct file_operations graph_depth_fops = {
1449         .open           = tracing_open_generic,
1450         .write          = graph_depth_write,
1451         .read           = graph_depth_read,
1452         .llseek         = generic_file_llseek,
1453 };
1454
1455 static __init int init_graph_tracefs(void)
1456 {
1457         struct dentry *d_tracer;
1458
1459         d_tracer = tracing_init_dentry();
1460         if (IS_ERR(d_tracer))
1461                 return 0;
1462
1463         trace_create_file("max_graph_depth", 0644, d_tracer,
1464                           NULL, &graph_depth_fops);
1465
1466         return 0;
1467 }
1468 fs_initcall(init_graph_tracefs);
1469
1470 static __init int init_graph_trace(void)
1471 {
1472         max_bytes_for_cpu = snprintf(NULL, 0, "%d", nr_cpu_ids - 1);
1473
1474         return register_tracer(&graph_trace);
1475 }
1476
1477 core_initcall(init_graph_trace);