kernel/trace/trace_ksym.c

   1 /*
   2  * trace_ksym.c - Kernel Symbol Tracer
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write to the Free Software
  16  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  17  *
  18  * Copyright (C) IBM Corporation, 2009
  19  */
  20
  21 #include <linux/kallsyms.h>
  22 #include <linux/uaccess.h>
  23 #include <linux/debugfs.h>
  24 #include <linux/ftrace.h>
  25 #include <linux/module.h>
  26 #include <linux/fs.h>
  27
  28 #include "trace_output.h"
  29 #include "trace_stat.h"
  30 #include "trace.h"
  31
  32 #include <linux/hw_breakpoint.h>
  33 #include <asm/hw_breakpoint.h>
  34
  35 /*
  36  * For now, let us restrict the no. of symbols traced simultaneously to number
  37  * of available hardware breakpoint registers.
  38  */
  39 #define KSYM_TRACER_MAX HBP_NUM
  40
  41 #define KSYM_TRACER_OP_LEN 3 /* rw- */
  42
  43 struct trace_ksym {
  44         struct perf_event       **ksym_hbp;
  45         struct perf_event_attr  attr;
  46 #ifdef CONFIG_PROFILE_KSYM_TRACER
  47         unsigned long           counter;
  48 #endif
  49         struct hlist_node       ksym_hlist;
  50 };
  51
  52 static struct trace_array *ksym_trace_array;
  53
  54 static unsigned int ksym_filter_entry_count;
  55 static unsigned int ksym_tracing_enabled;
  56
  57 static HLIST_HEAD(ksym_filter_head);
  58
  59 static DEFINE_MUTEX(ksym_tracer_mutex);
  60
  61 #ifdef CONFIG_PROFILE_KSYM_TRACER
  62
  63 #define MAX_UL_INT 0xffffffff
  64
  65 void ksym_collect_stats(unsigned long hbp_hit_addr)
  66 {
  67         struct hlist_node *node;
  68         struct trace_ksym *entry;
  69
  70         rcu_read_lock();
  71         hlist_for_each_entry_rcu(entry, node, &ksym_filter_head, ksym_hlist) {
  72                 if ((entry->attr.bp_addr == hbp_hit_addr) &&
  73                     (entry->counter <= MAX_UL_INT)) {
  74                         entry->counter++;
  75                         break;
  76                 }
  77         }
  78         rcu_read_unlock();
  79 }
  80 #endif /* CONFIG_PROFILE_KSYM_TRACER */
  81
  82 void ksym_hbp_handler(struct perf_event *hbp, int nmi,
  83                       struct perf_sample_data *data,
  84                       struct pt_regs *regs)
  85 {
  86         struct ring_buffer_event *event;
  87         struct ksym_trace_entry *entry;
  88         struct ring_buffer *buffer;
  89         int pc;
  90
  91         if (!ksym_tracing_enabled)
  92                 return;
  93
  94         buffer = ksym_trace_array->buffer;
  95
  96         pc = preempt_count();
  97
  98         event = trace_buffer_lock_reserve(buffer, TRACE_KSYM,
  99                                                         sizeof(*entry), 0, pc);
 100         if (!event)
 101                 return;
 102
 103         entry           = ring_buffer_event_data(event);
 104         entry->ip       = instruction_pointer(regs);
 105         entry->type     = hw_breakpoint_type(hbp);
 106         entry->addr     = hw_breakpoint_addr(hbp);
 107         strlcpy(entry->cmd, current->comm, TASK_COMM_LEN);
 108
 109 #ifdef CONFIG_PROFILE_KSYM_TRACER
 110         ksym_collect_stats(hw_breakpoint_addr(hbp));
 111 #endif /* CONFIG_PROFILE_KSYM_TRACER */
 112
 113         trace_buffer_unlock_commit(buffer, event, 0, pc);
 114 }
 115
 116 /* Valid access types are represented as
 117  *
 118  * rw- : Set Read/Write Access Breakpoint
 119  * -w- : Set Write Access Breakpoint
 120  * --- : Clear Breakpoints
 121  * --x : Set Execution Break points (Not available yet)
 122  *
 123  */
 124 static int ksym_trace_get_access_type(char *str)
 125 {
 126         int access = 0;
 127
 128         if (str[0] == 'r')
 129                 access |= HW_BREAKPOINT_R;
 130
 131         if (str[1] == 'w')
 132                 access |= HW_BREAKPOINT_W;
 133
 134         if (str[2] == 'x')
 135                 access |= HW_BREAKPOINT_X;
 136
 137         switch (access) {
 138         case HW_BREAKPOINT_R:
 139         case HW_BREAKPOINT_W:
 140         case HW_BREAKPOINT_W | HW_BREAKPOINT_R:
 141                 return access;
 142         default:
 143                 return -EINVAL;
 144         }
 145 }
 146
 147 /*
 148  * There can be several possible malformed requests and we attempt to capture
 149  * all of them. We enumerate some of the rules
 150  * 1. We will not allow kernel symbols with ':' since it is used as a delimiter.
 151  *    i.e. multiple ':' symbols disallowed. Possible uses are of the form
 152  *    <module>:<ksym_name>:<op>.
 153  * 2. No delimiter symbol ':' in the input string
 154  * 3. Spurious operator symbols or symbols not in their respective positions
 155  * 4. <ksym_name>:--- i.e. clear breakpoint request when ksym_name not in file
 156  * 5. Kernel symbol not a part of /proc/kallsyms
 157  * 6. Duplicate requests
 158  */
 159 static int parse_ksym_trace_str(char *input_string, char **ksymname,
 160                                                         unsigned long *addr)
 161 {
 162         int ret;
 163
 164         *ksymname = strsep(&input_string, ":");
 165         *addr = kallsyms_lookup_name(*ksymname);
 166
 167         /* Check for malformed request: (2), (1) and (5) */
 168         if ((!input_string) ||
 169             (strlen(input_string) != KSYM_TRACER_OP_LEN) ||
 170             (*addr == 0))
 171                 return -EINVAL;;
 172
 173         ret = ksym_trace_get_access_type(input_string);
 174
 175         return ret;
 176 }
 177
 178 int process_new_ksym_entry(char *ksymname, int op, unsigned long addr)
 179 {
 180         struct trace_ksym *entry;
 181         int ret = -ENOMEM;
 182
 183         if (ksym_filter_entry_count >= KSYM_TRACER_MAX) {
 184                 printk(KERN_ERR "ksym_tracer: Maximum limit:(%d) reached. No"
 185                 " new requests for tracing can be accepted now.\n",
 186                         KSYM_TRACER_MAX);
 187                 return -ENOSPC;
 188         }
 189
 190         entry = kzalloc(sizeof(struct trace_ksym), GFP_KERNEL);
 191         if (!entry)
 192                 return -ENOMEM;
 193
 194         hw_breakpoint_init(&entry->attr);
 195
 196         entry->attr.bp_type = op;
 197         entry->attr.bp_addr = addr;
 198         entry->attr.bp_len = HW_BREAKPOINT_LEN_4;
 199
 200         ret = -EAGAIN;
 201         entry->ksym_hbp = register_wide_hw_breakpoint(&entry->attr,
 202                                         ksym_hbp_handler);
 203
 204         if (IS_ERR(entry->ksym_hbp)) {
 205                 ret = PTR_ERR(entry->ksym_hbp);
 206                 printk(KERN_INFO "ksym_tracer request failed. Try again"
 207                                         " later!!\n");
 208                 goto err;
 209         }
 210
 211         hlist_add_head_rcu(&(entry->ksym_hlist), &ksym_filter_head);
 212         ksym_filter_entry_count++;
 213
 214         return 0;
 215
 216 err:
 217         kfree(entry);
 218
 219         return ret;
 220 }
 221
 222 static ssize_t ksym_trace_filter_read(struct file *filp, char __user *ubuf,
 223                                                 size_t count, loff_t *ppos)
 224 {
 225         struct trace_ksym *entry;
 226         struct hlist_node *node;
 227         struct trace_seq *s;
 228         ssize_t cnt = 0;
 229         int ret;
 230
 231         s = kmalloc(sizeof(*s), GFP_KERNEL);
 232         if (!s)
 233                 return -ENOMEM;
 234         trace_seq_init(s);
 235
 236         mutex_lock(&ksym_tracer_mutex);
 237
 238         hlist_for_each_entry(entry, node, &ksym_filter_head, ksym_hlist) {
 239                 ret = trace_seq_printf(s, "%pS:", (void *)entry->attr.bp_addr);
 240                 if (entry->attr.bp_type == HW_BREAKPOINT_R)
 241                         ret = trace_seq_puts(s, "r--\n");
 242                 else if (entry->attr.bp_type == HW_BREAKPOINT_W)
 243                         ret = trace_seq_puts(s, "-w-\n");
 244                 else if (entry->attr.bp_type == (HW_BREAKPOINT_W | HW_BREAKPOINT_R))
 245                         ret = trace_seq_puts(s, "rw-\n");
 246                 WARN_ON_ONCE(!ret);
 247         }
 248
 249         cnt = simple_read_from_buffer(ubuf, count, ppos, s->buffer, s->len);
 250
 251         mutex_unlock(&ksym_tracer_mutex);
 252
 253         kfree(s);
 254
 255         return cnt;
 256 }
 257
 258 static void __ksym_trace_reset(void)
 259 {
 260         struct trace_ksym *entry;
 261         struct hlist_node *node, *node1;
 262
 263         mutex_lock(&ksym_tracer_mutex);
 264         hlist_for_each_entry_safe(entry, node, node1, &ksym_filter_head,
 265                                                                 ksym_hlist) {
 266                 unregister_wide_hw_breakpoint(entry->ksym_hbp);
 267                 ksym_filter_entry_count--;
 268                 hlist_del_rcu(&(entry->ksym_hlist));
 269                 synchronize_rcu();
 270                 kfree(entry);
 271         }
 272         mutex_unlock(&ksym_tracer_mutex);
 273 }
 274
 275 static ssize_t ksym_trace_filter_write(struct file *file,
 276                                         const char __user *buffer,
 277                                                 size_t count, loff_t *ppos)
 278 {
 279         struct trace_ksym *entry;
 280         struct hlist_node *node;
 281         char *input_string, *ksymname = NULL;
 282         unsigned long ksym_addr = 0;
 283         int ret, op, changed = 0;
 284
 285         input_string = kzalloc(count + 1, GFP_KERNEL);
 286         if (!input_string)
 287                 return -ENOMEM;
 288
 289         if (copy_from_user(input_string, buffer, count)) {
 290                 kfree(input_string);
 291                 return -EFAULT;
 292         }
 293         input_string[count] = '\0';
 294
 295         strstrip(input_string);
 296
 297         /*
 298          * Clear all breakpoints if:
 299          * 1: echo > ksym_trace_filter
 300          * 2: echo 0 > ksym_trace_filter
 301          * 3: echo "*:---" > ksym_trace_filter
 302          */
 303         if (!input_string[0] || !strcmp(input_string, "0") ||
 304             !strcmp(input_string, "*:---")) {
 305                 __ksym_trace_reset();
 306                 kfree(input_string);
 307                 return count;
 308         }
 309
 310         ret = op = parse_ksym_trace_str(input_string, &ksymname, &ksym_addr);
 311         if (ret < 0) {
 312                 kfree(input_string);
 313                 return ret;
 314         }
 315
 316         mutex_lock(&ksym_tracer_mutex);
 317
 318         ret = -EINVAL;
 319         hlist_for_each_entry(entry, node, &ksym_filter_head, ksym_hlist) {
 320                 if (entry->attr.bp_addr == ksym_addr) {
 321                         /* Check for malformed request: (6) */
 322                         if (entry->attr.bp_type != op)
 323                                 changed = 1;
 324                         else
 325                                 goto out;
 326                         break;
 327                 }
 328         }
 329         if (changed) {
 330                 unregister_wide_hw_breakpoint(entry->ksym_hbp);
 331                 entry->attr.bp_type = op;
 332                 ret = 0;
 333                 if (op > 0) {
 334                         entry->ksym_hbp =
 335                                 register_wide_hw_breakpoint(&entry->attr,
 336                                         ksym_hbp_handler);
 337                         if (IS_ERR(entry->ksym_hbp))
 338                                 ret = PTR_ERR(entry->ksym_hbp);
 339                         else
 340                                 goto out;
 341                 }
 342                 /* Error or "symbol:---" case: drop it */
 343                 ksym_filter_entry_count--;
 344                 hlist_del_rcu(&(entry->ksym_hlist));
 345                 synchronize_rcu();
 346                 kfree(entry);
 347                 goto out;
 348         } else {
 349                 /* Check for malformed request: (4) */
 350                 if (op == 0)
 351                         goto out;
 352                 ret = process_new_ksym_entry(ksymname, op, ksym_addr);
 353         }
 354 out:
 355         mutex_unlock(&ksym_tracer_mutex);
 356
 357         kfree(input_string);
 358
 359         if (!ret)
 360                 ret = count;
 361         return ret;
 362 }
 363
 364 static const struct file_operations ksym_tracing_fops = {
 365         .open           = tracing_open_generic,
 366         .read           = ksym_trace_filter_read,
 367         .write          = ksym_trace_filter_write,
 368 };
 369
 370 static void ksym_trace_reset(struct trace_array *tr)
 371 {
 372         ksym_tracing_enabled = 0;
 373         __ksym_trace_reset();
 374 }
 375
 376 static int ksym_trace_init(struct trace_array *tr)
 377 {
 378         int cpu, ret = 0;
 379
 380         for_each_online_cpu(cpu)
 381                 tracing_reset(tr, cpu);
 382         ksym_tracing_enabled = 1;
 383         ksym_trace_array = tr;
 384
 385         return ret;
 386 }
 387
 388 static void ksym_trace_print_header(struct seq_file *m)
 389 {
 390         seq_puts(m,
 391                  "#       TASK-PID   CPU#      Symbol                    "
 392                  "Type    Function\n");
 393         seq_puts(m,
 394                  "#          |        |          |                       "
 395                  " |         |\n");
 396 }
 397
 398 static enum print_line_t ksym_trace_output(struct trace_iterator *iter)
 399 {
 400         struct trace_entry *entry = iter->ent;
 401         struct trace_seq *s = &iter->seq;
 402         struct ksym_trace_entry *field;
 403         char str[KSYM_SYMBOL_LEN];
 404         int ret;
 405
 406         if (entry->type != TRACE_KSYM)
 407                 return TRACE_TYPE_UNHANDLED;
 408
 409         trace_assign_type(field, entry);
 410
 411         ret = trace_seq_printf(s, "%11s-%-5d [%03d] %pS", field->cmd,
 412                                 entry->pid, iter->cpu, (char *)field->addr);
 413         if (!ret)
 414                 return TRACE_TYPE_PARTIAL_LINE;
 415
 416         switch (field->type) {
 417         case HW_BREAKPOINT_R:
 418                 ret = trace_seq_printf(s, " R  ");
 419                 break;
 420         case HW_BREAKPOINT_W:
 421                 ret = trace_seq_printf(s, " W  ");
 422                 break;
 423         case HW_BREAKPOINT_R | HW_BREAKPOINT_W:
 424                 ret = trace_seq_printf(s, " RW ");
 425                 break;
 426         default:
 427                 return TRACE_TYPE_PARTIAL_LINE;
 428         }
 429
 430         if (!ret)
 431                 return TRACE_TYPE_PARTIAL_LINE;
 432
 433         sprint_symbol(str, field->ip);
 434         ret = trace_seq_printf(s, "%s\n", str);
 435         if (!ret)
 436                 return TRACE_TYPE_PARTIAL_LINE;
 437
 438         return TRACE_TYPE_HANDLED;
 439 }
 440
 441 struct tracer ksym_tracer __read_mostly =
 442 {
 443         .name           = "ksym_tracer",
 444         .init           = ksym_trace_init,
 445         .reset          = ksym_trace_reset,
 446 #ifdef CONFIG_FTRACE_SELFTEST
 447         .selftest       = trace_selftest_startup_ksym,
 448 #endif
 449         .print_header   = ksym_trace_print_header,
 450         .print_line     = ksym_trace_output
 451 };
 452
 453 __init static int init_ksym_trace(void)
 454 {
 455         struct dentry *d_tracer;
 456         struct dentry *entry;
 457
 458         d_tracer = tracing_init_dentry();
 459         ksym_filter_entry_count = 0;
 460
 461         entry = debugfs_create_file("ksym_trace_filter", 0644, d_tracer,
 462                                     NULL, &ksym_tracing_fops);
 463         if (!entry)
 464                 pr_warning("Could not create debugfs "
 465                            "'ksym_trace_filter' file\n");
 466
 467         return register_tracer(&ksym_tracer);
 468 }
 469 device_initcall(init_ksym_trace);
 470
 471
 472 #ifdef CONFIG_PROFILE_KSYM_TRACER
 473 static int ksym_tracer_stat_headers(struct seq_file *m)
 474 {
 475         seq_puts(m, "  Access Type ");
 476         seq_puts(m, "  Symbol                                       Counter\n");
 477         seq_puts(m, "  ----------- ");
 478         seq_puts(m, "  ------                                       -------\n");
 479         return 0;
 480 }
 481
 482 static int ksym_tracer_stat_show(struct seq_file *m, void *v)
 483 {
 484         struct hlist_node *stat = v;
 485         struct trace_ksym *entry;
 486         int access_type = 0;
 487         char fn_name[KSYM_NAME_LEN];
 488
 489         entry = hlist_entry(stat, struct trace_ksym, ksym_hlist);
 490
 491         access_type = entry->attr.bp_type;
 492
 493         switch (access_type) {
 494         case HW_BREAKPOINT_R:
 495                 seq_puts(m, "  R           ");
 496                 break;
 497         case HW_BREAKPOINT_W:
 498                 seq_puts(m, "  W           ");
 499                 break;
 500         case HW_BREAKPOINT_R | HW_BREAKPOINT_W:
 501                 seq_puts(m, "  RW          ");
 502                 break;
 503         default:
 504                 seq_puts(m, "  NA          ");
 505         }
 506
 507         if (lookup_symbol_name(entry->attr.bp_addr, fn_name) >= 0)
 508                 seq_printf(m, "  %-36s", fn_name);
 509         else
 510                 seq_printf(m, "  %-36s", "<NA>");
 511         seq_printf(m, " %15lu\n", entry->counter);
 512
 513         return 0;
 514 }
 515
 516 static void *ksym_tracer_stat_start(struct tracer_stat *trace)
 517 {
 518         return ksym_filter_head.first;
 519 }
 520
 521 static void *
 522 ksym_tracer_stat_next(void *v, int idx)
 523 {
 524         struct hlist_node *stat = v;
 525
 526         return stat->next;
 527 }
 528
 529 static struct tracer_stat ksym_tracer_stats = {
 530         .name = "ksym_tracer",
 531         .stat_start = ksym_tracer_stat_start,
 532         .stat_next = ksym_tracer_stat_next,
 533         .stat_headers = ksym_tracer_stat_headers,
 534         .stat_show = ksym_tracer_stat_show
 535 };
 536
 537 __init static int ksym_tracer_stat_init(void)
 538 {
 539         int ret;
 540
 541         ret = register_stat_tracer(&ksym_tracer_stats);
 542         if (ret) {
 543                 printk(KERN_WARNING "Warning: could not register "
 544                                     "ksym tracer stats\n");
 545                 return 1;
 546         }
 547
 548         return 0;
 549 }
 550 fs_initcall(ksym_tracer_stat_init);
 551 #endif /* CONFIG_PROFILE_KSYM_TRACER */