libcfs/libcfs/tracefile.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2004 Cluster File Systems, Inc.
   5  *   Author: Zach Brown <zab@clusterfs.com>
   6  *   Author: Phil Schwan <phil@clusterfs.com>
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24
  25 #define DEBUG_SUBSYSTEM S_LNET
  26 #define LUSTRE_TRACEFILE_PRIVATE
  27 #include "tracefile.h"
  28
  29 #include <libcfs/libcfs.h>
  30
  31 /* XXX move things up to the top, comment */
  32 union trace_data_union (*trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  33
  34 char tracefile[TRACEFILE_NAME_SIZE];
  35 long long tracefile_size = TRACEFILE_SIZE;
  36 static struct tracefiled_ctl trace_tctl;
  37 struct semaphore trace_thread_sem;
  38 static int thread_running = 0;
  39
  40 atomic_t tage_allocated = ATOMIC_INIT(0);
  41
  42 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  43                                          struct trace_cpu_data *tcd);
  44
  45 static inline struct trace_page *tage_from_list(struct list_head *list)
  46 {
  47         return list_entry(list, struct trace_page, linkage);
  48 }
  49
  50 static struct trace_page *tage_alloc(int gfp)
  51 {
  52         cfs_page_t        *page;
  53         struct trace_page *tage;
  54
  55         /*
  56          * Don't spam console with allocation failures: they will be reported
  57          * by upper layer anyway.
  58          */
  59         gfp |= CFS_ALLOC_NOWARN;
  60         page = cfs_alloc_page(gfp);
  61         if (page == NULL)
  62                 return NULL;
  63
  64         tage = cfs_alloc(sizeof(*tage), gfp);
  65         if (tage == NULL) {
  66                 cfs_free_page(page);
  67                 return NULL;
  68         }
  69
  70         tage->page = page;
  71         atomic_inc(&tage_allocated);
  72         return tage;
  73 }
  74
  75 static void tage_free(struct trace_page *tage)
  76 {
  77         __LASSERT(tage != NULL);
  78         __LASSERT(tage->page != NULL);
  79
  80         cfs_free_page(tage->page);
  81         cfs_free(tage);
  82         atomic_dec(&tage_allocated);
  83 }
  84
  85 static void tage_to_tail(struct trace_page *tage, struct list_head *queue)
  86 {
  87         __LASSERT(tage != NULL);
  88         __LASSERT(queue != NULL);
  89
  90         list_move_tail(&tage->linkage, queue);
  91 }
  92
  93 int trace_refill_stock(struct trace_cpu_data *tcd, int gfp,
  94                        struct list_head *stock)
  95 {
  96         int i;
  97
  98         /*
  99          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 100          * from here: this will lead to infinite recursion.
 101          */
 102
 103         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++ i) {
 104                 struct trace_page *tage;
 105
 106                 tage = tage_alloc(gfp);
 107                 if (tage == NULL)
 108                         break;
 109                 list_add_tail(&tage->linkage, stock);
 110         }
 111         return i;
 112 }
 113
 114 /* return a page that has 'len' bytes left at the end */
 115 static struct trace_page *trace_get_tage_try(struct trace_cpu_data *tcd,
 116                                              unsigned long len)
 117 {
 118         struct trace_page *tage;
 119
 120         if (tcd->tcd_cur_pages > 0) {
 121                 __LASSERT(!list_empty(&tcd->tcd_pages));
 122                 tage = tage_from_list(tcd->tcd_pages.prev);
 123                 if (tage->used + len <= CFS_PAGE_SIZE)
 124                         return tage;
 125         }
 126
 127         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 128                 if (tcd->tcd_cur_stock_pages > 0) {
 129                         tage = tage_from_list(tcd->tcd_stock_pages.prev);
 130                         -- tcd->tcd_cur_stock_pages;
 131                         list_del_init(&tage->linkage);
 132                 } else {
 133                         tage = tage_alloc(CFS_ALLOC_ATOMIC);
 134                         if (tage == NULL) {
 135                                 printk(KERN_WARNING
 136                                        "failure to allocate a tage (%ld)\n",
 137                                        tcd->tcd_cur_pages);
 138                                 return NULL;
 139                         }
 140                 }
 141
 142                 tage->used = 0;
 143                 tage->cpu = smp_processor_id();
 144                 tage->type = tcd->tcd_type;
 145                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 146                 tcd->tcd_cur_pages++;
 147
 148                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 149                         struct tracefiled_ctl *tctl = &trace_tctl;
 150                         /*
 151                          * wake up tracefiled to process some pages.
 152                          */
 153                         cfs_waitq_signal(&tctl->tctl_waitq);
 154                 }
 155                 return tage;
 156         }
 157         return NULL;
 158 }
 159
 160 static void tcd_shrink(struct trace_cpu_data *tcd)
 161 {
 162         int pgcount = tcd->tcd_cur_pages / 10;
 163         struct page_collection pc;
 164         struct trace_page *tage;
 165         struct trace_page *tmp;
 166
 167         /*
 168          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 169          * from here: this will lead to infinite recursion.
 170          */
 171
 172         printk(KERN_WARNING "debug daemon buffer overflowed; discarding"
 173                " 10%% of pages (%d of %ld)\n", pgcount + 1, tcd->tcd_cur_pages);
 174
 175         CFS_INIT_LIST_HEAD(&pc.pc_pages);
 176         spin_lock_init(&pc.pc_lock);
 177
 178         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 179                 if (pgcount-- == 0)
 180                         break;
 181
 182                 list_move_tail(&tage->linkage, &pc.pc_pages);
 183                 tcd->tcd_cur_pages--;
 184         }
 185         put_pages_on_tcd_daemon_list(&pc, tcd);
 186 }
 187
 188 /* return a page that has 'len' bytes left at the end */
 189 static struct trace_page *trace_get_tage(struct trace_cpu_data *tcd,
 190                                          unsigned long len)
 191 {
 192         struct trace_page *tage;
 193
 194         /*
 195          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 196          * from here: this will lead to infinite recursion.
 197          */
 198
 199         if (len > CFS_PAGE_SIZE) {
 200                 printk(KERN_ERR
 201                        "cowardly refusing to write %lu bytes in a page\n", len);
 202                 return NULL;
 203         }
 204
 205         tage = trace_get_tage_try(tcd, len);
 206         if (tage != NULL)
 207                 return tage;
 208         if (thread_running)
 209                 tcd_shrink(tcd);
 210         if (tcd->tcd_cur_pages > 0) {
 211                 tage = tage_from_list(tcd->tcd_pages.next);
 212                 tage->used = 0;
 213                 tage_to_tail(tage, &tcd->tcd_pages);
 214         }
 215         return tage;
 216 }
 217
 218 int libcfs_debug_vmsg2(cfs_debug_limit_state_t *cdls, int subsys, int mask,
 219                        const char *file, const char *fn, const int line,
 220                        const char *format1, va_list args,
 221                        const char *format2, ...)
 222 {
 223         struct trace_cpu_data   *tcd = NULL;
 224         struct ptldebug_header   header;
 225         struct trace_page       *tage;
 226         /* string_buf is used only if tcd != NULL, and is always set then */
 227         char                    *string_buf = NULL;
 228         char                    *debug_buf;
 229         int                      known_size;
 230         int                      needed = 85; /* average message length */
 231         int                      max_nob;
 232         va_list                  ap;
 233         int                      depth;
 234         int                      i;
 235         int                      remain;
 236
 237         if (strchr(file, '/'))
 238                 file = strrchr(file, '/') + 1;
 239
 240
 241         set_ptldebug_header(&header, subsys, mask, line, CDEBUG_STACK());
 242
 243         tcd = trace_get_tcd();
 244         if (tcd == NULL)                /* arch may not log in IRQ context */
 245                 goto console;
 246
 247         if (tcd->tcd_shutting_down) {
 248                 trace_put_tcd(tcd);
 249                 tcd = NULL;
 250                 goto console;
 251         }
 252
 253         depth = __current_nesting_level();
 254         known_size = strlen(file) + 1 + depth;
 255         if (fn)
 256                 known_size += strlen(fn) + 1;
 257
 258         if (libcfs_debug_binary)
 259                 known_size += sizeof(header);
 260
 261         /*/
 262          * '2' used because vsnprintf return real size required for output
 263          * _without_ terminating NULL.
 264          * if needed is to small for this format.
 265          */
 266         for (i=0;i<2;i++) {
 267                 tage = trace_get_tage(tcd, needed + known_size + 1);
 268                 if (tage == NULL) {
 269                         if (needed + known_size > CFS_PAGE_SIZE)
 270                                 mask |= D_ERROR;
 271
 272                         trace_put_tcd(tcd);
 273                         tcd = NULL;
 274                         goto console;
 275                 }
 276
 277                 string_buf = (char *)cfs_page_address(tage->page)+tage->used+known_size;
 278
 279                 max_nob = CFS_PAGE_SIZE - tage->used - known_size;
 280                 if (max_nob <= 0) {
 281                         printk(KERN_EMERG "negative max_nob: %i\n", max_nob);
 282                         mask |= D_ERROR;
 283                         trace_put_tcd(tcd);
 284                         tcd = NULL;
 285                         goto console;
 286                 }
 287
 288                 needed = 0;
 289                 if (format1) {
 290                         va_copy(ap, args);
 291                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 292                         va_end(ap);
 293                 }
 294
 295
 296                 if (format2) {
 297                         remain = max_nob - needed;
 298                         if (remain < 0)
 299                                 remain = 0;
 300
 301                         va_start(ap, format2);
 302                         needed += vsnprintf(string_buf+needed, remain, format2, ap);
 303                         va_end(ap);
 304                 }
 305
 306                 if (needed < max_nob) /* well. printing ok.. */
 307                         break;
 308         }
 309
 310         if (*(string_buf+needed-1) != '\n')
 311                 printk(KERN_INFO "format at %s:%d:%s doesn't end in newline\n",
 312                        file, line, fn);
 313
 314         header.ph_len = known_size + needed;
 315         debug_buf = (char *)cfs_page_address(tage->page) + tage->used;
 316
 317         if (libcfs_debug_binary) {
 318                 memcpy(debug_buf, &header, sizeof(header));
 319                 tage->used += sizeof(header);
 320                 debug_buf += sizeof(header);
 321         }
 322
 323         /* indent message according to the nesting level */
 324         while (depth-- > 0) {
 325                 *(debug_buf++) = '.';
 326                 ++ tage->used;
 327         }
 328
 329         strcpy(debug_buf, file);
 330         tage->used += strlen(file) + 1;
 331         debug_buf += strlen(file) + 1;
 332
 333         if (fn) {
 334                 strcpy(debug_buf, fn);
 335                 tage->used += strlen(fn) + 1;
 336                 debug_buf += strlen(fn) + 1;
 337         }
 338
 339         __LASSERT(debug_buf == string_buf);
 340
 341         tage->used += needed;
 342         __LASSERT (tage->used <= CFS_PAGE_SIZE);
 343
 344 console:
 345         if ((mask & libcfs_printk) == 0) {
 346                 /* no console output requested */
 347                 if (tcd != NULL)
 348                         trace_put_tcd(tcd);
 349                 return 1;
 350         }
 351
 352         if (cdls != NULL) {
 353                 if (libcfs_console_ratelimit &&
 354                     cdls->cdls_next != 0 &&     /* not first time ever */
 355                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 356                         /* skipping a console message */
 357                         cdls->cdls_count++;
 358                         if (tcd != NULL)
 359                                 trace_put_tcd(tcd);
 360                         return 1;
 361                 }
 362
 363                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 364                                                        libcfs_console_max_delay
 365                                                        + cfs_time_seconds(10))) {
 366                         /* last timeout was a long time ago */
 367                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 368                 } else {
 369                         cdls->cdls_delay *= libcfs_console_backoff;
 370
 371                         if (cdls->cdls_delay < libcfs_console_min_delay)
 372                                 cdls->cdls_delay = libcfs_console_min_delay;
 373                         else if (cdls->cdls_delay > libcfs_console_max_delay)
 374                                 cdls->cdls_delay = libcfs_console_max_delay;
 375                 }
 376
 377                 /* ensure cdls_next is never zero after it's been seen */
 378                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 379         }
 380
 381         if (tcd != NULL) {
 382                 print_to_console(&header, mask, string_buf, needed, file, fn);
 383                 trace_put_tcd(tcd);
 384         } else {
 385                 string_buf = trace_get_console_buffer();
 386
 387                 needed = 0;
 388                 if (format1 != NULL) {
 389                         va_copy(ap, args);
 390                         needed = vsnprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE, format1, ap);
 391                         va_end(ap);
 392                 }
 393                 if (format2 != NULL) {
 394                         remain = TRACE_CONSOLE_BUFFER_SIZE - needed;
 395                         if (remain > 0) {
 396                                 va_start(ap, format2);
 397                                 needed += vsnprintf(string_buf+needed, remain, format2, ap);
 398                                 va_end(ap);
 399                         }
 400                 }
 401                 print_to_console(&header, mask,
 402                                  string_buf, needed, file, fn);
 403
 404                 trace_put_console_buffer(string_buf);
 405         }
 406
 407         if (cdls != NULL && cdls->cdls_count != 0) {
 408                 string_buf = trace_get_console_buffer();
 409
 410                 needed = snprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE,
 411                          "Skipped %d previous similar message%s\n",
 412                          cdls->cdls_count, (cdls->cdls_count > 1) ? "s" : "");
 413
 414                 print_to_console(&header, mask,
 415                                  string_buf, needed, file, fn);
 416
 417                 trace_put_console_buffer(string_buf);
 418                 cdls->cdls_count = 0;
 419         }
 420
 421         return 0;
 422 }
 423 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 424
 425 void
 426 libcfs_assertion_failed(const char *expr, const char *file,
 427                         const char *func, const int line)
 428 {
 429         libcfs_debug_msg(NULL, 0, D_EMERG, file, func, line,
 430                          "ASSERTION(%s) failed\n", expr);
 431         LBUG();
 432 }
 433 EXPORT_SYMBOL(libcfs_assertion_failed);
 434
 435 void
 436 trace_assertion_failed(const char *str,
 437                        const char *fn, const char *file, int line)
 438 {
 439         struct ptldebug_header hdr;
 440
 441         libcfs_panic_in_progress = 1;
 442         libcfs_catastrophe = 1;
 443         mb();
 444
 445         set_ptldebug_header(&hdr, DEBUG_SUBSYSTEM, D_EMERG, line,
 446                             CDEBUG_STACK());
 447
 448         print_to_console(&hdr, D_EMERG, str, strlen(str), file, fn);
 449
 450         LIBCFS_PANIC("Lustre debug assertion failure\n");
 451
 452         /* not reached */
 453 }
 454
 455 static void
 456 panic_collect_pages(struct page_collection *pc)
 457 {
 458         /* Do the collect_pages job on a single CPU: assumes that all other
 459          * CPUs have been stopped during a panic.  If this isn't true for some
 460          * arch, this will have to be implemented separately in each arch.  */
 461         int                    i;
 462         int                    j;
 463         struct trace_cpu_data *tcd;
 464
 465         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 466
 467         tcd_for_each(tcd, i, j) {
 468                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 469                 tcd->tcd_cur_pages = 0;
 470
 471                 if (pc->pc_want_daemon_pages) {
 472                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 473                         tcd->tcd_cur_daemon_pages = 0;
 474                 }
 475         }
 476 }
 477
 478 static void collect_pages_on_cpu(void *info)
 479 {
 480         struct trace_cpu_data *tcd;
 481         struct page_collection *pc = info;
 482         int i;
 483
 484         spin_lock(&pc->pc_lock);
 485         tcd_for_each_type_lock(tcd, i) {
 486                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 487                 tcd->tcd_cur_pages = 0;
 488                 if (pc->pc_want_daemon_pages) {
 489                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 490                         tcd->tcd_cur_daemon_pages = 0;
 491                 }
 492         }
 493         spin_unlock(&pc->pc_lock);
 494 }
 495
 496 static void collect_pages(struct page_collection *pc)
 497 {
 498         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 499
 500         if (libcfs_panic_in_progress)
 501                 panic_collect_pages(pc);
 502         else
 503                 trace_call_on_all_cpus(collect_pages_on_cpu, pc);
 504 }
 505
 506 static void put_pages_back_on_cpu(void *info)
 507 {
 508         struct page_collection *pc = info;
 509         struct trace_cpu_data *tcd;
 510         struct list_head *cur_head;
 511         struct trace_page *tage;
 512         struct trace_page *tmp;
 513         int i;
 514
 515         spin_lock(&pc->pc_lock);
 516         tcd_for_each_type_lock(tcd, i) {
 517                 cur_head = tcd->tcd_pages.next;
 518
 519                 list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 520
 521                         __LASSERT_TAGE_INVARIANT(tage);
 522
 523                         if (tage->cpu != smp_processor_id() || tage->type != i)
 524                                 continue;
 525
 526                         tage_to_tail(tage, cur_head);
 527                         tcd->tcd_cur_pages++;
 528                 }
 529         }
 530         spin_unlock(&pc->pc_lock);
 531 }
 532
 533 static void put_pages_back(struct page_collection *pc)
 534 {
 535         if (!libcfs_panic_in_progress)
 536                 trace_call_on_all_cpus(put_pages_back_on_cpu, pc);
 537 }
 538
 539 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 540  * we have a good amount of data at all times for dumping during an LBUG, even
 541  * if we have been steadily writing (and otherwise discarding) pages via the
 542  * debug daemon. */
 543 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 544                                          struct trace_cpu_data *tcd)
 545 {
 546         struct trace_page *tage;
 547         struct trace_page *tmp;
 548
 549         spin_lock(&pc->pc_lock);
 550         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 551
 552                 __LASSERT_TAGE_INVARIANT(tage);
 553
 554                 if (tage->cpu != smp_processor_id() ||
 555                     tage->type != tcd->tcd_type)
 556                         continue;
 557
 558                 tage_to_tail(tage, &tcd->tcd_daemon_pages);
 559                 tcd->tcd_cur_daemon_pages++;
 560
 561                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 562                         struct trace_page *victim;
 563
 564                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 565                         victim = tage_from_list(tcd->tcd_daemon_pages.next);
 566
 567                         __LASSERT_TAGE_INVARIANT(victim);
 568
 569                         list_del(&victim->linkage);
 570                         tage_free(victim);
 571                         tcd->tcd_cur_daemon_pages--;
 572                 }
 573         }
 574         spin_unlock(&pc->pc_lock);
 575 }
 576
 577 static void put_pages_on_daemon_list_on_cpu(void *info)
 578 {
 579         struct trace_cpu_data *tcd;
 580         int i;
 581
 582         tcd_for_each_type_lock(tcd, i)
 583                 put_pages_on_tcd_daemon_list(info, tcd);
 584 }
 585
 586 static void put_pages_on_daemon_list(struct page_collection *pc)
 587 {
 588         trace_call_on_all_cpus(put_pages_on_daemon_list_on_cpu, pc);
 589 }
 590
 591 void trace_debug_print(void)
 592 {
 593         struct page_collection pc;
 594         struct trace_page *tage;
 595         struct trace_page *tmp;
 596
 597         spin_lock_init(&pc.pc_lock);
 598
 599         pc.pc_want_daemon_pages = 1;
 600         collect_pages(&pc);
 601         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 602                 char *p, *file, *fn;
 603                 cfs_page_t *page;
 604
 605                 __LASSERT_TAGE_INVARIANT(tage);
 606
 607                 page = tage->page;
 608                 p = cfs_page_address(page);
 609                 while (p < ((char *)cfs_page_address(page) + tage->used)) {
 610                         struct ptldebug_header *hdr;
 611                         int len;
 612                         hdr = (void *)p;
 613                         p += sizeof(*hdr);
 614                         file = p;
 615                         p += strlen(file) + 1;
 616                         fn = p;
 617                         p += strlen(fn) + 1;
 618                         len = hdr->ph_len - (p - (char *)hdr);
 619
 620                         print_to_console(hdr, D_EMERG, p, len, file, fn);
 621
 622                         p += len;
 623                 }
 624
 625                 list_del(&tage->linkage);
 626                 tage_free(tage);
 627         }
 628 }
 629
 630 int tracefile_dump_all_pages(char *filename)
 631 {
 632         struct page_collection pc;
 633         cfs_file_t *filp;
 634         struct trace_page *tage;
 635         struct trace_page *tmp;
 636         int rc;
 637
 638         CFS_DECL_MMSPACE;
 639
 640         tracefile_write_lock();
 641
 642         filp = cfs_filp_open(filename,
 643                              O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600, &rc);
 644         if (!filp) {
 645                 if (rc != -EEXIST)
 646                         printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 647                                filename, rc);
 648                 goto out;
 649         }
 650
 651         spin_lock_init(&pc.pc_lock);
 652         pc.pc_want_daemon_pages = 1;
 653         collect_pages(&pc);
 654         if (list_empty(&pc.pc_pages)) {
 655                 rc = 0;
 656                 goto close;
 657         }
 658
 659         /* ok, for now, just write the pages.  in the future we'll be building
 660          * iobufs with the pages and calling generic_direct_IO */
 661         CFS_MMSPACE_OPEN;
 662         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 663
 664                 __LASSERT_TAGE_INVARIANT(tage);
 665
 666                 rc = cfs_filp_write(filp, cfs_page_address(tage->page),
 667                                     tage->used, cfs_filp_poff(filp));
 668                 if (rc != (int)tage->used) {
 669                         printk(KERN_WARNING "wanted to write %u but wrote "
 670                                "%d\n", tage->used, rc);
 671                         put_pages_back(&pc);
 672                         __LASSERT(list_empty(&pc.pc_pages));
 673                         break;
 674                 }
 675                 list_del(&tage->linkage);
 676                 tage_free(tage);
 677         }
 678         CFS_MMSPACE_CLOSE;
 679         rc = cfs_filp_fsync(filp);
 680         if (rc)
 681                 printk(KERN_ERR "sync returns %d\n", rc);
 682  close:
 683         cfs_filp_close(filp);
 684  out:
 685         tracefile_write_unlock();
 686         return rc;
 687 }
 688
 689 void trace_flush_pages(void)
 690 {
 691         struct page_collection pc;
 692         struct trace_page *tage;
 693         struct trace_page *tmp;
 694
 695         spin_lock_init(&pc.pc_lock);
 696
 697         pc.pc_want_daemon_pages = 1;
 698         collect_pages(&pc);
 699         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 700
 701                 __LASSERT_TAGE_INVARIANT(tage);
 702
 703                 list_del(&tage->linkage);
 704                 tage_free(tage);
 705         }
 706 }
 707
 708 int trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 709                         const char *usr_buffer, int usr_buffer_nob)
 710 {
 711         int    nob;
 712
 713         if (usr_buffer_nob > knl_buffer_nob)
 714                 return -EOVERFLOW;
 715
 716         if (copy_from_user((void *)knl_buffer,
 717                            (void *)usr_buffer, usr_buffer_nob))
 718                 return -EFAULT;
 719
 720         nob = strnlen(knl_buffer, usr_buffer_nob);
 721         while (nob-- >= 0)                      /* strip trailing whitespace */
 722                 if (!isspace(knl_buffer[nob]))
 723                         break;
 724
 725         if (nob < 0)                            /* empty string */
 726                 return -EINVAL;
 727
 728         if (nob == knl_buffer_nob)              /* no space to terminate */
 729                 return -EOVERFLOW;
 730
 731         knl_buffer[nob + 1] = 0;                /* terminate */
 732         return 0;
 733 }
 734
 735 int trace_copyout_string(char *usr_buffer, int usr_buffer_nob,
 736                          const char *knl_buffer, char *append)
 737 {
 738         /* NB if 'append' != NULL, it's a single character to append to the
 739          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 740          * terminating zero byte) for sysctl entries */
 741         int   nob = strlen(knl_buffer);
 742
 743         if (nob > usr_buffer_nob)
 744                 nob = usr_buffer_nob;
 745
 746         if (copy_to_user(usr_buffer, knl_buffer, nob))
 747                 return -EFAULT;
 748
 749         if (append != NULL && nob < usr_buffer_nob) {
 750                 if (copy_to_user(usr_buffer + nob, append, 1))
 751                         return -EFAULT;
 752
 753                 nob++;
 754         }
 755
 756         return nob;
 757 }
 758
 759 int trace_allocate_string_buffer(char **str, int nob)
 760 {
 761         if (nob > 2 * CFS_PAGE_SIZE)            /* string must be "sensible" */
 762                 return -EINVAL;
 763
 764         *str = cfs_alloc(nob, CFS_ALLOC_STD | CFS_ALLOC_ZERO);
 765         if (*str == NULL)
 766                 return -ENOMEM;
 767
 768         return 0;
 769 }
 770
 771 void trace_free_string_buffer(char *str, int nob)
 772 {
 773         cfs_free(str);
 774 }
 775
 776 int trace_dump_debug_buffer_usrstr(void *usr_str, int usr_str_nob)
 777 {
 778         char         *str;
 779         int           rc;
 780
 781         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 782         if (rc != 0)
 783                 return rc;
 784
 785         rc = trace_copyin_string(str, usr_str_nob + 1,
 786                                  usr_str, usr_str_nob);
 787         if (rc != 0)
 788                 goto out;
 789
 790 #if !defined(__WINNT__)
 791         if (str[0] != '/') {
 792                 rc = -EINVAL;
 793                 goto out;
 794         }
 795 #endif
 796         rc = tracefile_dump_all_pages(str);
 797 out:
 798         trace_free_string_buffer(str, usr_str_nob + 1);
 799         return rc;
 800 }
 801
 802 int trace_daemon_command(char *str)
 803 {
 804         int       rc = 0;
 805
 806         tracefile_write_lock();
 807
 808         if (strcmp(str, "stop") == 0) {
 809                 trace_stop_thread();
 810                 memset(tracefile, 0, sizeof(tracefile));
 811
 812         } else if (strncmp(str, "size=", 5) == 0) {
 813                 tracefile_size = simple_strtoul(str + 5, NULL, 0);
 814                 if (tracefile_size < 10 || tracefile_size > 20480)
 815                         tracefile_size = TRACEFILE_SIZE;
 816                 else
 817                         tracefile_size <<= 20;
 818
 819         } else if (strlen(str) >= sizeof(tracefile)) {
 820                 rc = -ENAMETOOLONG;
 821 #ifndef __WINNT__
 822         } else if (str[0] != '/') {
 823                 rc = -EINVAL;
 824 #endif
 825         } else {
 826                 strcpy(tracefile, str);
 827
 828                 printk(KERN_INFO "Lustre: debug daemon will attempt to start writing "
 829                        "to %s (%lukB max)\n", tracefile,
 830                        (long)(tracefile_size >> 10));
 831
 832                 trace_start_thread();
 833         }
 834
 835         tracefile_write_unlock();
 836         return rc;
 837 }
 838
 839 int trace_daemon_command_usrstr(void *usr_str, int usr_str_nob)
 840 {
 841         char *str;
 842         int   rc;
 843
 844         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 845         if (rc != 0)
 846                 return rc;
 847
 848         rc = trace_copyin_string(str, usr_str_nob + 1,
 849                                  usr_str, usr_str_nob);
 850         if (rc == 0)
 851                 rc = trace_daemon_command(str);
 852
 853         trace_free_string_buffer(str, usr_str_nob + 1);
 854         return rc;
 855 }
 856
 857 int trace_set_debug_mb(int mb)
 858 {
 859         int i;
 860         int j;
 861         int pages;
 862         int limit = trace_max_debug_mb();
 863         struct trace_cpu_data *tcd;
 864
 865         if (mb < num_possible_cpus())
 866                 return -EINVAL;
 867
 868         if (mb > limit) {
 869                 printk(KERN_ERR "Lustre: Refusing to set debug buffer size to "
 870                        "%dMB - limit is %d\n", mb, limit);
 871                 return -EINVAL;
 872         }
 873
 874         mb /= num_possible_cpus();
 875         pages = mb << (20 - CFS_PAGE_SHIFT);
 876
 877         tracefile_write_lock();
 878
 879         tcd_for_each(tcd, i, j)
 880                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 881
 882         tracefile_write_unlock();
 883
 884         return 0;
 885 }
 886
 887 int trace_set_debug_mb_usrstr(void *usr_str, int usr_str_nob)
 888 {
 889         char     str[32];
 890         int      rc;
 891
 892         rc = trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 893         if (rc < 0)
 894                 return rc;
 895
 896         return trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 897 }
 898
 899 int trace_get_debug_mb(void)
 900 {
 901         int i;
 902         int j;
 903         struct trace_cpu_data *tcd;
 904         int total_pages = 0;
 905
 906         tracefile_read_lock();
 907
 908         tcd_for_each(tcd, i, j)
 909                 total_pages += tcd->tcd_max_pages;
 910
 911         tracefile_read_unlock();
 912
 913         return (total_pages >> (20 - CFS_PAGE_SHIFT)) + 1;
 914 }
 915
 916 static int tracefiled(void *arg)
 917 {
 918         struct page_collection pc;
 919         struct tracefiled_ctl *tctl = arg;
 920         struct trace_page *tage;
 921         struct trace_page *tmp;
 922         struct ptldebug_header *hdr;
 923         cfs_file_t *filp;
 924         int rc;
 925
 926         CFS_DECL_MMSPACE;
 927
 928         /* we're started late enough that we pick up init's fs context */
 929         /* this is so broken in uml?  what on earth is going on? */
 930         cfs_daemonize("ktracefiled");
 931
 932         spin_lock_init(&pc.pc_lock);
 933         complete(&tctl->tctl_start);
 934
 935         while (1) {
 936                 cfs_waitlink_t __wait;
 937
 938                 cfs_waitlink_init(&__wait);
 939                 cfs_waitq_add(&tctl->tctl_waitq, &__wait);
 940                 set_current_state(TASK_INTERRUPTIBLE);
 941                 cfs_waitq_timedwait(&__wait, CFS_TASK_INTERRUPTIBLE,
 942                                     cfs_time_seconds(1));
 943                 cfs_waitq_del(&tctl->tctl_waitq, &__wait);
 944
 945                 if (atomic_read(&tctl->tctl_shutdown))
 946                         break;
 947
 948                 pc.pc_want_daemon_pages = 0;
 949                 collect_pages(&pc);
 950                 if (list_empty(&pc.pc_pages))
 951                         continue;
 952
 953                 filp = NULL;
 954                 tracefile_read_lock();
 955                 if (tracefile[0] != 0) {
 956                         filp = cfs_filp_open(tracefile,
 957                                              O_CREAT | O_RDWR | O_LARGEFILE,
 958                                              0600, &rc);
 959                         if (!(filp))
 960                                 printk(KERN_WARNING "couldn't open %s: %d\n",
 961                                        tracefile, rc);
 962                 }
 963                 tracefile_read_unlock();
 964                 if (filp == NULL) {
 965                         put_pages_on_daemon_list(&pc);
 966                         __LASSERT(list_empty(&pc.pc_pages));
 967                         continue;
 968                 }
 969
 970                 CFS_MMSPACE_OPEN;
 971
 972                 /* mark the first header, so we can sort in chunks */
 973                 tage = tage_from_list(pc.pc_pages.next);
 974                 __LASSERT_TAGE_INVARIANT(tage);
 975
 976                 hdr = cfs_page_address(tage->page);
 977                 hdr->ph_flags |= PH_FLAG_FIRST_RECORD;
 978
 979                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 980                         static loff_t f_pos;
 981
 982                         __LASSERT_TAGE_INVARIANT(tage);
 983
 984                         if (f_pos >= (off_t)tracefile_size)
 985                                 f_pos = 0;
 986                         else if (f_pos > cfs_filp_size(filp))
 987                                 f_pos = cfs_filp_size(filp);
 988
 989                         rc = cfs_filp_write(filp, cfs_page_address(tage->page),
 990                                             tage->used, &f_pos);
 991                         if (rc != (int)tage->used) {
 992                                 printk(KERN_WARNING "wanted to write %u but "
 993                                        "wrote %d\n", tage->used, rc);
 994                                 put_pages_back(&pc);
 995                                 __LASSERT(list_empty(&pc.pc_pages));
 996                         }
 997                 }
 998                 CFS_MMSPACE_CLOSE;
 999
1000                 cfs_filp_close(filp);
1001                 put_pages_on_daemon_list(&pc);
1002                 __LASSERT(list_empty(&pc.pc_pages));
1003         }
1004         complete(&tctl->tctl_stop);
1005         return 0;
1006 }
1007
1008 int trace_start_thread(void)
1009 {
1010         struct tracefiled_ctl *tctl = &trace_tctl;
1011         int rc = 0;
1012
1013         mutex_down(&trace_thread_sem);
1014         if (thread_running)
1015                 goto out;
1016
1017         init_completion(&tctl->tctl_start);
1018         init_completion(&tctl->tctl_stop);
1019         cfs_waitq_init(&tctl->tctl_waitq);
1020         atomic_set(&tctl->tctl_shutdown, 0);
1021
1022         if (cfs_kernel_thread(tracefiled, tctl, 0) < 0) {
1023                 rc = -ECHILD;
1024                 goto out;
1025         }
1026
1027         wait_for_completion(&tctl->tctl_start);
1028         thread_running = 1;
1029 out:
1030         mutex_up(&trace_thread_sem);
1031         return rc;
1032 }
1033
1034 void trace_stop_thread(void)
1035 {
1036         struct tracefiled_ctl *tctl = &trace_tctl;
1037
1038         mutex_down(&trace_thread_sem);
1039         if (thread_running) {
1040                 printk(KERN_INFO "Lustre: shutting down debug daemon thread...\n");
1041                 atomic_set(&tctl->tctl_shutdown, 1);
1042                 wait_for_completion(&tctl->tctl_stop);
1043                 thread_running = 0;
1044         }
1045         mutex_up(&trace_thread_sem);
1046 }
1047
1048 int tracefile_init(int max_pages)
1049 {
1050         struct trace_cpu_data *tcd;
1051         int                    i;
1052         int                    j;
1053         int                    rc;
1054         int                    factor;
1055
1056         rc = tracefile_init_arch();
1057         if (rc != 0)
1058                 return rc;
1059
1060         tcd_for_each(tcd, i, j) {
1061                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1062                 factor = tcd->tcd_pages_factor;
1063                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
1064                 CFS_INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1065                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1066                 tcd->tcd_cur_pages = 0;
1067                 tcd->tcd_cur_stock_pages = 0;
1068                 tcd->tcd_cur_daemon_pages = 0;
1069                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1070                 LASSERT(tcd->tcd_max_pages > 0);
1071                 tcd->tcd_shutting_down = 0;
1072         }
1073
1074         return 0;
1075 }
1076
1077 static void trace_cleanup_on_cpu(void *info)
1078 {
1079         struct trace_cpu_data *tcd;
1080         struct trace_page *tage;
1081         struct trace_page *tmp;
1082         int i;
1083
1084         tcd_for_each_type_lock(tcd, i) {
1085                 tcd->tcd_shutting_down = 1;
1086
1087                 list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1088                         __LASSERT_TAGE_INVARIANT(tage);
1089
1090                         list_del(&tage->linkage);
1091                         tage_free(tage);
1092                 }
1093                 tcd->tcd_cur_pages = 0;
1094         }
1095 }
1096
1097 static void trace_cleanup(void)
1098 {
1099         struct page_collection pc;
1100
1101         CFS_INIT_LIST_HEAD(&pc.pc_pages);
1102         spin_lock_init(&pc.pc_lock);
1103
1104         trace_call_on_all_cpus(trace_cleanup_on_cpu, &pc);
1105
1106         tracefile_fini_arch();
1107 }
1108
1109 void tracefile_exit(void)
1110 {
1111         trace_stop_thread();
1112         trace_cleanup();
1113 }