granicus.if.org Git - zfs/blob - module/spl/spl-debug.c

   1 /*****************************************************************************\
   2  *  Copyright (C) 2007-2010 Lawrence Livermore National Security, LLC.
   3  *  Copyright (C) 2007 The Regents of the University of California.
   4  *  Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER).
   5  *  Written by Brian Behlendorf <behlendorf1@llnl.gov>.
   6  *  UCRL-CODE-235197
   7  *
   8  *  This file is part of the SPL, Solaris Porting Layer.
   9  *  For details, see <http://github.com/behlendorf/spl/>.
  10  *
  11  *  The SPL is free software; you can redistribute it and/or modify it
  12  *  under the terms of the GNU General Public License as published by the
  13  *  Free Software Foundation; either version 2 of the License, or (at your
  14  *  option) any later version.
  15  *
  16  *  The SPL is distributed in the hope that it will be useful, but WITHOUT
  17  *  ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  18  *  FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  19  *  for more details.
  20  *
  21  *  You should have received a copy of the GNU General Public License along
  22  *  with the SPL.  If not, see <http://www.gnu.org/licenses/>.
  23  *****************************************************************************
  24  *  Solaris Porting Layer (SPL) Debug Implementation.
  25 \*****************************************************************************/
  26
  27 #include <linux/kmod.h>
  28 #include <linux/mm.h>
  29 #include <linux/vmalloc.h>
  30 #include <linux/pagemap.h>
  31 #include <linux/slab.h>
  32 #include <linux/ctype.h>
  33 #include <linux/kthread.h>
  34 #include <linux/hardirq.h>
  35 #include <linux/interrupt.h>
  36 #include <linux/spinlock.h>
  37 #include <linux/proc_compat.h>
  38 #include <linux/file_compat.h>
  39 #include <sys/sysmacros.h>
  40 #include <spl-debug.h>
  41 #include <spl-trace.h>
  42 #include <spl-ctl.h>
  43
  44 #ifdef SS_DEBUG_SUBSYS
  45 #undef SS_DEBUG_SUBSYS
  46 #endif
  47
  48 #define SS_DEBUG_SUBSYS SS_DEBUG
  49
  50 unsigned long spl_debug_subsys = ~0;
  51 EXPORT_SYMBOL(spl_debug_subsys);
  52 module_param(spl_debug_subsys, ulong, 0644);
  53 MODULE_PARM_DESC(spl_debug_subsys, "Subsystem debugging level mask.");
  54
  55 unsigned long spl_debug_mask = SD_CANTMASK;
  56 EXPORT_SYMBOL(spl_debug_mask);
  57 module_param(spl_debug_mask, ulong, 0644);
  58 MODULE_PARM_DESC(spl_debug_mask, "Debugging level mask.");
  59
  60 unsigned long spl_debug_printk = SD_CANTMASK;
  61 EXPORT_SYMBOL(spl_debug_printk);
  62 module_param(spl_debug_printk, ulong, 0644);
  63 MODULE_PARM_DESC(spl_debug_printk, "Console printk level mask.");
  64
  65 int spl_debug_mb = -1;
  66 EXPORT_SYMBOL(spl_debug_mb);
  67 module_param(spl_debug_mb, int, 0644);
  68 MODULE_PARM_DESC(spl_debug_mb, "Total debug buffer size.");
  69
  70 unsigned int spl_debug_binary = 1;
  71 EXPORT_SYMBOL(spl_debug_binary);
  72
  73 unsigned int spl_debug_catastrophe;
  74 EXPORT_SYMBOL(spl_debug_catastrophe);
  75
  76 unsigned int spl_debug_panic_on_bug = 0;
  77 EXPORT_SYMBOL(spl_debug_panic_on_bug);
  78 module_param(spl_debug_panic_on_bug, uint, 0644);
  79 MODULE_PARM_DESC(spl_debug_panic_on_bug, "Panic on BUG");
  80
  81 static char spl_debug_file_name[PATH_MAX];
  82 char spl_debug_file_path[PATH_MAX] = "/tmp/spl-log";
  83
  84 unsigned int spl_console_ratelimit = 1;
  85 EXPORT_SYMBOL(spl_console_ratelimit);
  86
  87 long spl_console_max_delay;
  88 EXPORT_SYMBOL(spl_console_max_delay);
  89
  90 long spl_console_min_delay;
  91 EXPORT_SYMBOL(spl_console_min_delay);
  92
  93 unsigned int spl_console_backoff = SPL_DEFAULT_BACKOFF;
  94 EXPORT_SYMBOL(spl_console_backoff);
  95
  96 unsigned int spl_debug_stack;
  97 EXPORT_SYMBOL(spl_debug_stack);
  98
  99 static int spl_panic_in_progress;
 100
 101 union trace_data_union (*trace_data[TCD_TYPE_MAX])[NR_CPUS] __cacheline_aligned;
 102 char *trace_console_buffers[NR_CPUS][3];
 103 struct rw_semaphore trace_sem;
 104 atomic_t trace_tage_allocated = ATOMIC_INIT(0);
 105
 106 static int spl_debug_dump_all_pages(dumplog_priv_t *dp, char *);
 107 static void trace_fini(void);
 108
 109
 110 /* Memory percentage breakdown by type */
 111 static unsigned int pages_factor[TCD_TYPE_MAX] = {
 112        80,  /* 80% pages for TCD_TYPE_PROC */
 113        10,  /* 10% pages for TCD_TYPE_SOFTIRQ */
 114        10   /* 10% pages for TCD_TYPE_IRQ */
 115 };
 116
 117 const char *
 118 spl_debug_subsys2str(int subsys)
 119 {
 120         switch (subsys) {
 121         default:
 122                 return NULL;
 123         case SS_UNDEFINED:
 124                 return "undefined";
 125         case SS_ATOMIC:
 126                 return "atomic";
 127         case SS_KOBJ:
 128                 return "kobj";
 129         case SS_VNODE:
 130                 return "vnode";
 131         case SS_TIME:
 132                 return "time";
 133         case SS_RWLOCK:
 134                 return "rwlock";
 135         case SS_THREAD:
 136                 return "thread";
 137         case SS_CONDVAR:
 138                 return "condvar";
 139         case SS_MUTEX:
 140                 return "mutex";
 141         case SS_RNG:
 142                 return "rng";
 143         case SS_TASKQ:
 144                 return "taskq";
 145         case SS_KMEM:
 146                 return "kmem";
 147         case SS_DEBUG:
 148                 return "debug";
 149         case SS_GENERIC:
 150                 return "generic";
 151         case SS_PROC:
 152                 return "proc";
 153         case SS_MODULE:
 154                 return "module";
 155         case SS_CRED:
 156                 return "cred";
 157         case SS_KSTAT:
 158                 return "kstat";
 159         case SS_XDR:
 160                 return "xdr";
 161         case SS_TSD:
 162                 return "tsd";
 163         case SS_USER1:
 164                 return "user1";
 165         case SS_USER2:
 166                 return "user2";
 167         case SS_USER3:
 168                 return "user3";
 169         case SS_USER4:
 170                 return "user4";
 171         case SS_USER5:
 172                 return "user5";
 173         case SS_USER6:
 174                 return "user6";
 175         case SS_USER7:
 176                 return "user7";
 177         case SS_USER8:
 178                 return "user8";
 179         }
 180 }
 181
 182 const char *
 183 spl_debug_dbg2str(int debug)
 184 {
 185         switch (debug) {
 186         default:
 187                 return NULL;
 188         case SD_TRACE:
 189                 return "trace";
 190         case SD_INFO:
 191                 return "info";
 192         case SD_WARNING:
 193                 return "warning";
 194         case SD_ERROR:
 195                 return "error";
 196         case SD_EMERG:
 197                 return "emerg";
 198         case SD_CONSOLE:
 199                 return "console";
 200         case SD_IOCTL:
 201                 return "ioctl";
 202         case SD_DPRINTF:
 203                 return "dprintf";
 204         case SD_OTHER:
 205                 return "other";
 206         }
 207 }
 208
 209 int
 210 spl_debug_mask2str(char *str, int size, unsigned long mask, int is_subsys)
 211 {
 212         const char *(*fn)(int bit) = is_subsys ? spl_debug_subsys2str :
 213                                                  spl_debug_dbg2str;
 214         const char *token;
 215         int i, bit, len = 0;
 216
 217         if (mask == 0) {                        /* "0" */
 218                 if (size > 0)
 219                         str[0] = '0';
 220                 len = 1;
 221         } else {                                /* space-separated tokens */
 222                 for (i = 0; i < 32; i++) {
 223                         bit = 1 << i;
 224
 225                         if ((mask & bit) == 0)
 226                                 continue;
 227
 228                         token = fn(bit);
 229                         if (token == NULL)              /* unused bit */
 230                                 continue;
 231
 232                         if (len > 0) {                  /* separator? */
 233                                 if (len < size)
 234                                         str[len] = ' ';
 235                                 len++;
 236                         }
 237
 238                         while (*token != 0) {
 239                                 if (len < size)
 240                                         str[len] = *token;
 241                                 token++;
 242                                 len++;
 243                         }
 244                 }
 245         }
 246
 247         /* terminate 'str' */
 248         if (len < size)
 249                 str[len] = 0;
 250         else
 251                 str[size - 1] = 0;
 252
 253         return len;
 254 }
 255
 256 static int
 257 spl_debug_token2mask(int *mask, const char *str, int len, int is_subsys)
 258 {
 259         const char *(*fn)(int bit) = is_subsys ? spl_debug_subsys2str :
 260                                                  spl_debug_dbg2str;
 261         const char   *token;
 262         int i, j, bit;
 263
 264         /* match against known tokens */
 265         for (i = 0; i < 32; i++) {
 266                 bit = 1 << i;
 267
 268                 token = fn(bit);
 269                 if (token == NULL)              /* unused? */
 270                         continue;
 271
 272                 /* strcasecmp */
 273                 for (j = 0; ; j++) {
 274                         if (j == len) {         /* end of token */
 275                                 if (token[j] == 0) {
 276                                         *mask = bit;
 277                                         return 0;
 278                                 }
 279                                 break;
 280                         }
 281
 282                         if (token[j] == 0)
 283                                 break;
 284
 285                         if (str[j] == token[j])
 286                                 continue;
 287
 288                         if (str[j] < 'A' || 'Z' < str[j])
 289                                 break;
 290
 291                         if (str[j] - 'A' + 'a' != token[j])
 292                                 break;
 293                 }
 294         }
 295
 296         return -EINVAL;                         /* no match */
 297 }
 298
 299 int
 300 spl_debug_str2mask(unsigned long *mask, const char *str, int is_subsys)
 301 {
 302         char op = 0;
 303         int m = 0, matched, n, t;
 304
 305         /* Allow a number for backwards compatibility */
 306         for (n = strlen(str); n > 0; n--)
 307                 if (!isspace(str[n-1]))
 308                         break;
 309         matched = n;
 310
 311         if ((t = sscanf(str, "%i%n", &m, &matched)) >= 1 && matched == n) {
 312                 *mask = m;
 313                 return 0;
 314         }
 315
 316         /* <str> must be a list of debug tokens or numbers separated by
 317          * whitespace and optionally an operator ('+' or '-').  If an operator
 318          * appears first in <str>, '*mask' is used as the starting point
 319          * (relative), otherwise 0 is used (absolute).  An operator applies to
 320          * all following tokens up to the next operator. */
 321         matched = 0;
 322         while (*str != 0) {
 323                 while (isspace(*str)) /* skip whitespace */
 324                         str++;
 325
 326                 if (*str == 0)
 327                         break;
 328
 329                 if (*str == '+' || *str == '-') {
 330                         op = *str++;
 331
 332                         /* op on first token == relative */
 333                         if (!matched)
 334                                 m = *mask;
 335
 336                         while (isspace(*str)) /* skip whitespace */
 337                                 str++;
 338
 339                         if (*str == 0)          /* trailing op */
 340                                 return -EINVAL;
 341                 }
 342
 343                 /* find token length */
 344                 for (n = 0; str[n] != 0 && !isspace(str[n]); n++);
 345
 346                 /* match token */
 347                 if (spl_debug_token2mask(&t, str, n, is_subsys) != 0)
 348                         return -EINVAL;
 349
 350                 matched = 1;
 351                 if (op == '-')
 352                         m &= ~t;
 353                 else
 354                         m |= t;
 355
 356                 str += n;
 357         }
 358
 359         if (!matched)
 360                 return -EINVAL;
 361
 362         *mask = m;
 363         return 0;
 364 }
 365
 366 static void
 367 spl_debug_dumplog_internal(dumplog_priv_t *dp)
 368 {
 369         void *journal_info;
 370
 371         journal_info = current->journal_info;
 372         current->journal_info = NULL;
 373
 374         snprintf(spl_debug_file_name, sizeof(spl_debug_file_path) - 1,
 375                  "%s.%ld.%ld", spl_debug_file_path,
 376                  get_seconds(), (long)dp->dp_pid);
 377         printk("SPL: Dumping log to %s\n", spl_debug_file_name);
 378         spl_debug_dump_all_pages(dp, spl_debug_file_name);
 379
 380         current->journal_info = journal_info;
 381 }
 382
 383 static int
 384 spl_debug_dumplog_thread(void *arg)
 385 {
 386         dumplog_priv_t *dp = (dumplog_priv_t *)arg;
 387
 388         spl_debug_dumplog_internal(dp);
 389         atomic_set(&dp->dp_done, 1);
 390         wake_up(&dp->dp_waitq);
 391         complete_and_exit(NULL, 0);
 392
 393         return 0; /* Unreachable */
 394 }
 395
 396 /* When flag is set do not use a new thread for the debug dump */
 397 int
 398 spl_debug_dumplog(int flags)
 399 {
 400         struct task_struct *tsk;
 401         dumplog_priv_t dp;
 402
 403         init_waitqueue_head(&dp.dp_waitq);
 404         dp.dp_pid = current->pid;
 405         dp.dp_flags = flags;
 406         atomic_set(&dp.dp_done, 0);
 407
 408         if (dp.dp_flags & DL_NOTHREAD) {
 409                 spl_debug_dumplog_internal(&dp);
 410         } else {
 411
 412                 tsk = kthread_create(spl_debug_dumplog_thread,(void *)&dp,"spl_debug");
 413                 if (tsk == NULL)
 414                         return -ENOMEM;
 415
 416                 wake_up_process(tsk);
 417                 wait_event(dp.dp_waitq, atomic_read(&dp.dp_done));
 418         }
 419
 420         return 0;
 421 }
 422 EXPORT_SYMBOL(spl_debug_dumplog);
 423
 424 static char *
 425 trace_get_console_buffer(void)
 426 {
 427         int  cpu = get_cpu();
 428         int  idx;
 429
 430         if (in_irq()) {
 431                 idx = 0;
 432         } else if (in_softirq()) {
 433                 idx = 1;
 434         } else {
 435                 idx = 2;
 436         }
 437
 438         return trace_console_buffers[cpu][idx];
 439 }
 440
 441 static void
 442 trace_put_console_buffer(char *buffer)
 443 {
 444         put_cpu();
 445 }
 446
 447 static int
 448 trace_lock_tcd(struct trace_cpu_data *tcd)
 449 {
 450         __ASSERT(tcd->tcd_type < TCD_TYPE_MAX);
 451
 452         spin_lock_irqsave(&tcd->tcd_lock, tcd->tcd_lock_flags);
 453
 454         return 1;
 455 }
 456
 457 static void
 458 trace_unlock_tcd(struct trace_cpu_data *tcd)
 459 {
 460         __ASSERT(tcd->tcd_type < TCD_TYPE_MAX);
 461
 462         spin_unlock_irqrestore(&tcd->tcd_lock, tcd->tcd_lock_flags);
 463 }
 464
 465 static struct trace_cpu_data *
 466 trace_get_tcd(void)
 467 {
 468         int cpu;
 469         struct trace_cpu_data *tcd;
 470
 471         cpu = get_cpu();
 472         if (in_irq())
 473                 tcd = &(*trace_data[TCD_TYPE_IRQ])[cpu].tcd;
 474         else if (in_softirq())
 475                 tcd = &(*trace_data[TCD_TYPE_SOFTIRQ])[cpu].tcd;
 476         else
 477                 tcd = &(*trace_data[TCD_TYPE_PROC])[cpu].tcd;
 478
 479         trace_lock_tcd(tcd);
 480
 481         return tcd;
 482 }
 483
 484 static void
 485 trace_put_tcd (struct trace_cpu_data *tcd)
 486 {
 487         trace_unlock_tcd(tcd);
 488
 489         put_cpu();
 490 }
 491
 492 static void
 493 trace_set_debug_header(struct spl_debug_header *header, int subsys,
 494                        int mask, const int line, unsigned long stack)
 495 {
 496         struct timeval tv;
 497
 498         do_gettimeofday(&tv);
 499
 500         header->ph_subsys = subsys;
 501         header->ph_mask = mask;
 502         header->ph_cpu_id = smp_processor_id();
 503         header->ph_sec = (__u32)tv.tv_sec;
 504         header->ph_usec = tv.tv_usec;
 505         header->ph_stack = stack;
 506         header->ph_pid = current->pid;
 507         header->ph_line_num = line;
 508
 509         return;
 510 }
 511
 512 static void
 513 trace_print_to_console(struct spl_debug_header *hdr, int mask, const char *buf,
 514                        int len, const char *file, const char *fn)
 515 {
 516         char *prefix = "SPL", *ptype = NULL;
 517
 518         if ((mask & SD_EMERG) != 0) {
 519                 prefix = "SPLError";
 520                 ptype = KERN_EMERG;
 521         } else if ((mask & SD_ERROR) != 0) {
 522                 prefix = "SPLError";
 523                 ptype = KERN_ERR;
 524         } else if ((mask & SD_WARNING) != 0) {
 525                 prefix = "SPL";
 526                 ptype = KERN_WARNING;
 527         } else if ((mask & (SD_CONSOLE | spl_debug_printk)) != 0) {
 528                 prefix = "SPL";
 529                 ptype = KERN_INFO;
 530         }
 531
 532         if ((mask & SD_CONSOLE) != 0) {
 533                 printk("%s%s: %.*s", ptype, prefix, len, buf);
 534         } else {
 535                 printk("%s%s: %d:%d:(%s:%d:%s()) %.*s", ptype, prefix,
 536                        hdr->ph_pid, hdr->ph_stack, file,
 537                        hdr->ph_line_num, fn, len, buf);
 538         }
 539
 540         return;
 541 }
 542
 543 static int
 544 trace_max_debug_mb(void)
 545 {
 546         return MAX(512, ((num_physpages >> (20 - PAGE_SHIFT)) * 80) / 100);
 547 }
 548
 549 static struct trace_page *
 550 tage_alloc(int gfp)
 551 {
 552         struct page *page;
 553         struct trace_page *tage;
 554
 555         page = alloc_pages(gfp | __GFP_NOWARN, 0);
 556         if (page == NULL)
 557                 return NULL;
 558
 559         tage = kmalloc(sizeof(*tage), gfp);
 560         if (tage == NULL) {
 561                 __free_pages(page, 0);
 562                 return NULL;
 563         }
 564
 565         tage->page = page;
 566         atomic_inc(&trace_tage_allocated);
 567
 568         return tage;
 569 }
 570
 571 static void
 572 tage_free(struct trace_page *tage)
 573 {
 574         __ASSERT(tage != NULL);
 575         __ASSERT(tage->page != NULL);
 576
 577         __free_pages(tage->page, 0);
 578         kfree(tage);
 579         atomic_dec(&trace_tage_allocated);
 580 }
 581
 582 static struct trace_page *
 583 tage_from_list(struct list_head *list)
 584 {
 585         return list_entry(list, struct trace_page, linkage);
 586 }
 587
 588 static void
 589 tage_to_tail(struct trace_page *tage, struct list_head *queue)
 590 {
 591         __ASSERT(tage != NULL);
 592         __ASSERT(queue != NULL);
 593
 594         list_move_tail(&tage->linkage, queue);
 595 }
 596
 597 /* try to return a page that has 'len' bytes left at the end */
 598 static struct trace_page *
 599 trace_get_tage_try(struct trace_cpu_data *tcd, unsigned long len)
 600 {
 601         struct trace_page *tage;
 602
 603         if (tcd->tcd_cur_pages > 0) {
 604                 __ASSERT(!list_empty(&tcd->tcd_pages));
 605                 tage = tage_from_list(tcd->tcd_pages.prev);
 606                 if (tage->used + len <= PAGE_SIZE)
 607                         return tage;
 608         }
 609
 610         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 611                 if (tcd->tcd_cur_stock_pages > 0) {
 612                         tage = tage_from_list(tcd->tcd_stock_pages.prev);
 613                         tcd->tcd_cur_stock_pages--;
 614                         list_del_init(&tage->linkage);
 615                 } else {
 616                         tage = tage_alloc(GFP_ATOMIC);
 617                         if (tage == NULL) {
 618                                 printk(KERN_WARNING
 619                                        "failure to allocate a tage (%ld)\n",
 620                                        tcd->tcd_cur_pages);
 621                                 return NULL;
 622                         }
 623                 }
 624
 625                 tage->used = 0;
 626                 tage->cpu = smp_processor_id();
 627                 tage->type = tcd->tcd_type;
 628                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 629                 tcd->tcd_cur_pages++;
 630
 631                 return tage;
 632         }
 633
 634         return NULL;
 635 }
 636
 637 /* return a page that has 'len' bytes left at the end */
 638 static struct trace_page *
 639 trace_get_tage(struct trace_cpu_data *tcd, unsigned long len)
 640 {
 641         struct trace_page *tage;
 642
 643         __ASSERT(len <= PAGE_SIZE);
 644
 645         tage = trace_get_tage_try(tcd, len);
 646         if (tage)
 647                 return tage;
 648
 649         if (tcd->tcd_cur_pages > 0) {
 650                 tage = tage_from_list(tcd->tcd_pages.next);
 651                 tage->used = 0;
 652                 tage_to_tail(tage, &tcd->tcd_pages);
 653         }
 654
 655         return tage;
 656 }
 657
 658 int
 659 spl_debug_msg(void *arg, int subsys, int mask, const char *file,
 660     const char *fn, const int line, const char *format, ...)
 661 {
 662         spl_debug_limit_state_t *cdls = arg;
 663         struct trace_cpu_data   *tcd = NULL;
 664         struct spl_debug_header header = { 0, };
 665         struct trace_page       *tage;
 666         /* string_buf is used only if tcd != NULL, and is always set then */
 667         char                    *string_buf = NULL;
 668         char                    *debug_buf;
 669         int                      known_size;
 670         int                      needed = 85; /* average message length */
 671         int                      max_nob;
 672         va_list                  ap;
 673         int                      i;
 674
 675         if (subsys == 0)
 676                 subsys = SS_DEBUG_SUBSYS;
 677
 678         if (mask == 0)
 679                 mask = SD_EMERG;
 680
 681         if (strchr(file, '/'))
 682                 file = strrchr(file, '/') + 1;
 683
 684         trace_set_debug_header(&header, subsys, mask, line, 0);
 685
 686         tcd = trace_get_tcd();
 687         if (tcd == NULL)
 688                 goto console;
 689
 690         if (tcd->tcd_shutting_down) {
 691                 trace_put_tcd(tcd);
 692                 tcd = NULL;
 693                 goto console;
 694         }
 695
 696         known_size = strlen(file) + 1;
 697         if (fn)
 698                 known_size += strlen(fn) + 1;
 699
 700         if (spl_debug_binary)
 701                 known_size += sizeof(header);
 702
 703         /* '2' used because vsnprintf returns real size required for output
 704          * _without_ terminating NULL. */
 705         for (i = 0; i < 2; i++) {
 706                 tage = trace_get_tage(tcd, needed + known_size + 1);
 707                 if (tage == NULL) {
 708                         if (needed + known_size > PAGE_SIZE)
 709                                 mask |= SD_ERROR;
 710
 711                         trace_put_tcd(tcd);
 712                         tcd = NULL;
 713                         goto console;
 714                 }
 715
 716                 string_buf = (char *)page_address(tage->page) +
 717                              tage->used + known_size;
 718
 719                 max_nob = PAGE_SIZE - tage->used - known_size;
 720                 if (max_nob <= 0) {
 721                         printk(KERN_EMERG "negative max_nob: %i\n", max_nob);
 722                         mask |= SD_ERROR;
 723                         trace_put_tcd(tcd);
 724                         tcd = NULL;
 725                         goto console;
 726                 }
 727
 728                 needed = 0;
 729                 if (format) {
 730                         va_start(ap, format);
 731                         needed += vsnprintf(string_buf, max_nob, format, ap);
 732                         va_end(ap);
 733                 }
 734
 735                 if (needed < max_nob)
 736                         break;
 737         }
 738
 739         header.ph_len = known_size + needed;
 740         debug_buf = (char *)page_address(tage->page) + tage->used;
 741
 742         if (spl_debug_binary) {
 743                 memcpy(debug_buf, &header, sizeof(header));
 744                 tage->used += sizeof(header);
 745                 debug_buf += sizeof(header);
 746         }
 747
 748         strcpy(debug_buf, file);
 749         tage->used += strlen(file) + 1;
 750         debug_buf += strlen(file) + 1;
 751
 752         if (fn) {
 753                 strcpy(debug_buf, fn);
 754                 tage->used += strlen(fn) + 1;
 755                 debug_buf += strlen(fn) + 1;
 756         }
 757
 758         __ASSERT(debug_buf == string_buf);
 759
 760         tage->used += needed;
 761         __ASSERT (tage->used <= PAGE_SIZE);
 762
 763 console:
 764         if ((mask & spl_debug_printk) == 0) {
 765                 /* no console output requested */
 766                 if (tcd != NULL)
 767                         trace_put_tcd(tcd);
 768                 return 1;
 769         }
 770
 771         if (cdls != NULL) {
 772                 if (spl_console_ratelimit && cdls->cdls_next != 0 &&
 773                     !time_before(cdls->cdls_next, jiffies)) {
 774                         /* skipping a console message */
 775                         cdls->cdls_count++;
 776                         if (tcd != NULL)
 777                                 trace_put_tcd(tcd);
 778                         return 1;
 779                 }
 780
 781                 if (time_before(cdls->cdls_next + spl_console_max_delay +
 782                                 (10 * HZ), jiffies)) {
 783                         /* last timeout was a long time ago */
 784                         cdls->cdls_delay /= spl_console_backoff * 4;
 785                 } else {
 786                         cdls->cdls_delay *= spl_console_backoff;
 787
 788                         if (cdls->cdls_delay < spl_console_min_delay)
 789                                 cdls->cdls_delay = spl_console_min_delay;
 790                         else if (cdls->cdls_delay > spl_console_max_delay)
 791                                 cdls->cdls_delay = spl_console_max_delay;
 792                 }
 793
 794                 /* ensure cdls_next is never zero after it's been seen */
 795                 cdls->cdls_next = (jiffies + cdls->cdls_delay) | 1;
 796         }
 797
 798         if (tcd != NULL) {
 799                 trace_print_to_console(&header, mask, string_buf, needed, file, fn);
 800                 trace_put_tcd(tcd);
 801         } else {
 802                 string_buf = trace_get_console_buffer();
 803
 804                 needed = 0;
 805                 if (format != NULL) {
 806                         va_start(ap, format);
 807                         needed += vsnprintf(string_buf,
 808                             TRACE_CONSOLE_BUFFER_SIZE, format, ap);
 809                         va_end(ap);
 810                 }
 811                 trace_print_to_console(&header, mask,
 812                                  string_buf, needed, file, fn);
 813
 814                 trace_put_console_buffer(string_buf);
 815         }
 816
 817         if (cdls != NULL && cdls->cdls_count != 0) {
 818                 string_buf = trace_get_console_buffer();
 819
 820                 needed = snprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE,
 821                          "Skipped %d previous similar message%s\n",
 822                          cdls->cdls_count, (cdls->cdls_count > 1) ? "s" : "");
 823
 824                 trace_print_to_console(&header, mask,
 825                                  string_buf, needed, file, fn);
 826
 827                 trace_put_console_buffer(string_buf);
 828                 cdls->cdls_count = 0;
 829         }
 830
 831         return 0;
 832 }
 833 EXPORT_SYMBOL(spl_debug_msg);
 834
 835 /* Do the collect_pages job on a single CPU: assumes that all other
 836  * CPUs have been stopped during a panic.  If this isn't true for
 837  * some arch, this will have to be implemented separately in each arch.
 838  */
 839 static void
 840 collect_pages_from_single_cpu(struct page_collection *pc)
 841 {
 842         struct trace_cpu_data *tcd;
 843         int i, j;
 844
 845         tcd_for_each(tcd, i, j) {
 846                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 847                 tcd->tcd_cur_pages = 0;
 848         }
 849 }
 850
 851 static void
 852 collect_pages_on_all_cpus(struct page_collection *pc)
 853 {
 854         struct trace_cpu_data *tcd;
 855         int i, cpu;
 856
 857         spin_lock(&pc->pc_lock);
 858         for_each_possible_cpu(cpu) {
 859                 tcd_for_each_type_lock(tcd, i, cpu) {
 860                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 861                         tcd->tcd_cur_pages = 0;
 862                 }
 863         }
 864         spin_unlock(&pc->pc_lock);
 865 }
 866
 867 static void
 868 collect_pages(dumplog_priv_t *dp, struct page_collection *pc)
 869 {
 870         INIT_LIST_HEAD(&pc->pc_pages);
 871
 872         if (spl_panic_in_progress || dp->dp_flags & DL_SINGLE_CPU)
 873                 collect_pages_from_single_cpu(pc);
 874         else
 875                 collect_pages_on_all_cpus(pc);
 876 }
 877
 878 static void
 879 put_pages_back_on_all_cpus(struct page_collection *pc)
 880 {
 881         struct trace_cpu_data *tcd;
 882         struct list_head *cur_head;
 883         struct trace_page *tage;
 884         struct trace_page *tmp;
 885         int i, cpu;
 886
 887         spin_lock(&pc->pc_lock);
 888
 889         for_each_possible_cpu(cpu) {
 890                 tcd_for_each_type_lock(tcd, i, cpu) {
 891                         cur_head = tcd->tcd_pages.next;
 892
 893                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 894                                                  linkage) {
 895                                 if (tage->cpu != cpu || tage->type != i)
 896                                         continue;
 897
 898                                 tage_to_tail(tage, cur_head);
 899                                 tcd->tcd_cur_pages++;
 900                         }
 901                 }
 902         }
 903
 904         spin_unlock(&pc->pc_lock);
 905 }
 906
 907 static void
 908 put_pages_back(struct page_collection *pc)
 909 {
 910         if (!spl_panic_in_progress)
 911                 put_pages_back_on_all_cpus(pc);
 912 }
 913
 914 static int
 915 spl_debug_dump_all_pages(dumplog_priv_t *dp, char *filename)
 916 {
 917         struct page_collection pc;
 918         struct file *filp;
 919         struct trace_page *tage;
 920         struct trace_page *tmp;
 921         mm_segment_t oldfs;
 922         int rc = 0;
 923
 924         down_write(&trace_sem);
 925
 926         filp = spl_filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE,
 927                                0600, &rc);
 928         if (filp == NULL) {
 929                 if (rc != -EEXIST)
 930                         printk(KERN_ERR "SPL: Can't open %s for dump: %d\n",
 931                                filename, rc);
 932                 goto out;
 933         }
 934
 935         spin_lock_init(&pc.pc_lock);
 936         collect_pages(dp, &pc);
 937         if (list_empty(&pc.pc_pages)) {
 938                 rc = 0;
 939                 goto close;
 940         }
 941
 942         oldfs = get_fs();
 943         set_fs(get_ds());
 944
 945         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 946                 rc = spl_filp_write(filp, page_address(tage->page),
 947                                     tage->used, spl_filp_poff(filp));
 948                 if (rc != (int)tage->used) {
 949                         printk(KERN_WARNING "SPL: Wanted to write %u "
 950                                "but wrote %d\n", tage->used, rc);
 951                         put_pages_back(&pc);
 952                         __ASSERT(list_empty(&pc.pc_pages));
 953                         break;
 954                 }
 955                 list_del(&tage->linkage);
 956                 tage_free(tage);
 957         }
 958
 959         set_fs(oldfs);
 960
 961         rc = spl_filp_fsync(filp, 1);
 962         if (rc)
 963                 printk(KERN_ERR "SPL: Unable to sync: %d\n", rc);
 964  close:
 965         spl_filp_close(filp);
 966  out:
 967         up_write(&trace_sem);
 968
 969         return rc;
 970 }
 971
 972 static void
 973 spl_debug_flush_pages(void)
 974 {
 975         dumplog_priv_t dp;
 976         struct page_collection pc;
 977         struct trace_page *tage;
 978         struct trace_page *tmp;
 979
 980         spin_lock_init(&pc.pc_lock);
 981         init_waitqueue_head(&dp.dp_waitq);
 982         dp.dp_pid = current->pid;
 983         dp.dp_flags = 0;
 984         atomic_set(&dp.dp_done, 0);
 985
 986         collect_pages(&dp, &pc);
 987         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 988                 list_del(&tage->linkage);
 989                 tage_free(tage);
 990         }
 991 }
 992
 993 unsigned long
 994 spl_debug_set_mask(unsigned long mask) {
 995         spl_debug_mask = mask;
 996         return 0;
 997 }
 998 EXPORT_SYMBOL(spl_debug_set_mask);
 999
1000 unsigned long
1001 spl_debug_get_mask(void) {
1002         return spl_debug_mask;
1003 }
1004 EXPORT_SYMBOL(spl_debug_get_mask);
1005
1006 unsigned long
1007 spl_debug_set_subsys(unsigned long subsys) {
1008         spl_debug_subsys = subsys;
1009         return 0;
1010 }
1011 EXPORT_SYMBOL(spl_debug_set_subsys);
1012
1013 unsigned long
1014 spl_debug_get_subsys(void) {
1015         return spl_debug_subsys;
1016 }
1017 EXPORT_SYMBOL(spl_debug_get_subsys);
1018
1019 int
1020 spl_debug_set_mb(int mb)
1021 {
1022         int i, j, pages;
1023         int limit = trace_max_debug_mb();
1024         struct trace_cpu_data *tcd;
1025
1026         if (mb < num_possible_cpus()) {
1027                 printk(KERN_ERR "SPL: Refusing to set debug buffer size to "
1028                        "%dMB - lower limit is %d\n", mb, num_possible_cpus());
1029                 return -EINVAL;
1030         }
1031
1032         if (mb > limit) {
1033                 printk(KERN_ERR "SPL: Refusing to set debug buffer size to "
1034                        "%dMB - upper limit is %d\n", mb, limit);
1035                 return -EINVAL;
1036         }
1037
1038         mb /= num_possible_cpus();
1039         pages = mb << (20 - PAGE_SHIFT);
1040
1041         down_write(&trace_sem);
1042
1043         tcd_for_each(tcd, i, j)
1044                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
1045
1046         up_write(&trace_sem);
1047
1048         return 0;
1049 }
1050 EXPORT_SYMBOL(spl_debug_set_mb);
1051
1052 int
1053 spl_debug_get_mb(void)
1054 {
1055         int i, j;
1056         struct trace_cpu_data *tcd;
1057         int total_pages = 0;
1058
1059         down_read(&trace_sem);
1060
1061         tcd_for_each(tcd, i, j)
1062                 total_pages += tcd->tcd_max_pages;
1063
1064         up_read(&trace_sem);
1065
1066         return (total_pages >> (20 - PAGE_SHIFT)) + 1;
1067 }
1068 EXPORT_SYMBOL(spl_debug_get_mb);
1069
1070 void spl_debug_dumpstack(struct task_struct *tsk)
1071 {
1072         extern void show_task(struct task_struct *);
1073
1074         if (tsk == NULL)
1075                 tsk = current;
1076
1077         printk("SPL: Showing stack for process %d\n", tsk->pid);
1078         dump_stack();
1079 }
1080 EXPORT_SYMBOL(spl_debug_dumpstack);
1081
1082 void spl_debug_bug(char *file, const char *func, const int line, int flags)
1083 {
1084         spl_debug_catastrophe = 1;
1085         spl_debug_msg(NULL, 0, SD_EMERG, file, func, line, "SPL PANIC\n");
1086
1087         if (in_interrupt())
1088                 panic("SPL PANIC in interrupt.\n");
1089
1090         if (in_atomic() || irqs_disabled())
1091                 flags |= DL_NOTHREAD;
1092
1093         /* Ensure all debug pages and dumped by current cpu */
1094          if (spl_debug_panic_on_bug)
1095                 spl_panic_in_progress = 1;
1096
1097         spl_debug_dumpstack(NULL);
1098         spl_debug_dumplog(flags);
1099
1100         if (spl_debug_panic_on_bug)
1101                 panic("SPL PANIC");
1102
1103         set_task_state(current, TASK_UNINTERRUPTIBLE);
1104         while (1)
1105                 schedule();
1106 }
1107 EXPORT_SYMBOL(spl_debug_bug);
1108
1109 int
1110 spl_debug_clear_buffer(void)
1111 {
1112         spl_debug_flush_pages();
1113         return 0;
1114 }
1115 EXPORT_SYMBOL(spl_debug_clear_buffer);
1116
1117 int
1118 spl_debug_mark_buffer(char *text)
1119 {
1120         SDEBUG(SD_WARNING, "*************************************\n");
1121         SDEBUG(SD_WARNING, "DEBUG MARKER: %s\n", text);
1122         SDEBUG(SD_WARNING, "*************************************\n");
1123
1124         return 0;
1125 }
1126 EXPORT_SYMBOL(spl_debug_mark_buffer);
1127
1128 static int
1129 trace_init(int max_pages)
1130 {
1131         struct trace_cpu_data *tcd;
1132         int i, j;
1133
1134         init_rwsem(&trace_sem);
1135
1136         /* initialize trace_data */
1137         memset(trace_data, 0, sizeof(trace_data));
1138         for (i = 0; i < TCD_TYPE_MAX; i++) {
1139                 trace_data[i] = kmalloc(sizeof(union trace_data_union) *
1140                                         NR_CPUS, GFP_KERNEL);
1141                 if (trace_data[i] == NULL)
1142                         goto out;
1143         }
1144
1145         tcd_for_each(tcd, i, j) {
1146                 spin_lock_init(&tcd->tcd_lock);
1147                 tcd->tcd_pages_factor = pages_factor[i];
1148                 tcd->tcd_type = i;
1149                 tcd->tcd_cpu = j;
1150                 INIT_LIST_HEAD(&tcd->tcd_pages);
1151                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1152                 tcd->tcd_cur_pages = 0;
1153                 tcd->tcd_cur_stock_pages = 0;
1154                 tcd->tcd_max_pages = (max_pages * pages_factor[i]) / 100;
1155                 tcd->tcd_shutting_down = 0;
1156         }
1157
1158         for (i = 0; i < num_possible_cpus(); i++) {
1159                 for (j = 0; j < 3; j++) {
1160                         trace_console_buffers[i][j] =
1161                                 kmalloc(TRACE_CONSOLE_BUFFER_SIZE,
1162                                         GFP_KERNEL);
1163
1164                         if (trace_console_buffers[i][j] == NULL)
1165                                 goto out;
1166                 }
1167        }
1168
1169         return 0;
1170 out:
1171         trace_fini();
1172         printk(KERN_ERR "SPL: Insufficient memory for debug logs\n");
1173         return -ENOMEM;
1174 }
1175
1176 int
1177 debug_init(void)
1178 {
1179         int rc, max = spl_debug_mb;
1180
1181         spl_console_max_delay = SPL_DEFAULT_MAX_DELAY;
1182         spl_console_min_delay = SPL_DEFAULT_MIN_DELAY;
1183
1184         /* If spl_debug_mb is set to an invalid value or uninitialized
1185          * then just make the total buffers smp_num_cpus TCD_MAX_PAGES */
1186         if (max > (num_physpages >> (20 - 2 - PAGE_SHIFT)) / 5 ||
1187             max >= 512 || max < 0) {
1188                 max = TCD_MAX_PAGES;
1189         } else {
1190                 max = (max / num_online_cpus()) << (20 - PAGE_SHIFT);
1191         }
1192
1193         rc = trace_init(max);
1194         if (rc)
1195                 return rc;
1196
1197         return rc;
1198 }
1199
1200 static void
1201 trace_cleanup_on_all_cpus(void)
1202 {
1203         struct trace_cpu_data *tcd;
1204         struct trace_page *tage;
1205         struct trace_page *tmp;
1206         int i, cpu;
1207
1208         for_each_possible_cpu(cpu) {
1209                 tcd_for_each_type_lock(tcd, i, cpu) {
1210                         tcd->tcd_shutting_down = 1;
1211
1212                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages,
1213                                                  linkage) {
1214                                 list_del(&tage->linkage);
1215                                 tage_free(tage);
1216                         }
1217                         tcd->tcd_cur_pages = 0;
1218                 }
1219         }
1220 }
1221
1222 static void
1223 trace_fini(void)
1224 {
1225         int i, j;
1226
1227         trace_cleanup_on_all_cpus();
1228
1229         for (i = 0; i < num_possible_cpus(); i++) {
1230                 for (j = 0; j < 3; j++) {
1231                         if (trace_console_buffers[i][j] != NULL) {
1232                                 kfree(trace_console_buffers[i][j]);
1233                                 trace_console_buffers[i][j] = NULL;
1234                         }
1235                 }
1236         }
1237
1238         for (i = 0; i < TCD_TYPE_MAX && trace_data[i] != NULL; i++) {
1239                 kfree(trace_data[i]);
1240                 trace_data[i] = NULL;
1241         }
1242 }
1243
1244 void
1245 debug_fini(void)
1246 {
1247         trace_fini();
1248 }