lnet/libcfs/tracefile.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/libcfs/tracefile.c
  37  *
  38  * Author: Zach Brown <zab@clusterfs.com>
  39  * Author: Phil Schwan <phil@clusterfs.com>
  40  */
  41
  42
  43 #define DEBUG_SUBSYSTEM S_LNET
  44 #define LUSTRE_TRACEFILE_PRIVATE
  45 #include "tracefile.h"
  46
  47 #include <libcfs/kp30.h>
  48 #include <libcfs/libcfs.h>
  49
  50 /* XXX move things up to the top, comment */
  51 union trace_data_union (*trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  52
  53 char tracefile[TRACEFILE_NAME_SIZE];
  54 long long tracefile_size = TRACEFILE_SIZE;
  55 static struct tracefiled_ctl trace_tctl;
  56 struct semaphore trace_thread_sem;
  57 static int thread_running = 0;
  58
  59 atomic_t tage_allocated = ATOMIC_INIT(0);
  60
  61 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  62                                          struct trace_cpu_data *tcd);
  63
  64 static inline struct trace_page *tage_from_list(struct list_head *list)
  65 {
  66         return list_entry(list, struct trace_page, linkage);
  67 }
  68
  69 static struct trace_page *tage_alloc(int gfp)
  70 {
  71         cfs_page_t        *page;
  72         struct trace_page *tage;
  73
  74         /*
  75          * Don't spam console with allocation failures: they will be reported
  76          * by upper layer anyway.
  77          */
  78         gfp |= CFS_ALLOC_NOWARN;
  79         page = cfs_alloc_page(gfp);
  80         if (page == NULL)
  81                 return NULL;
  82
  83         tage = cfs_alloc(sizeof(*tage), gfp);
  84         if (tage == NULL) {
  85                 cfs_free_page(page);
  86                 return NULL;
  87         }
  88
  89         tage->page = page;
  90         atomic_inc(&tage_allocated);
  91         return tage;
  92 }
  93
  94 static void tage_free(struct trace_page *tage)
  95 {
  96         __LASSERT(tage != NULL);
  97         __LASSERT(tage->page != NULL);
  98
  99         cfs_free_page(tage->page);
 100         cfs_free(tage);
 101         atomic_dec(&tage_allocated);
 102 }
 103
 104 static void tage_to_tail(struct trace_page *tage, struct list_head *queue)
 105 {
 106         __LASSERT(tage != NULL);
 107         __LASSERT(queue != NULL);
 108
 109         list_move_tail(&tage->linkage, queue);
 110 }
 111
 112 int trace_refill_stock(struct trace_cpu_data *tcd, int gfp,
 113                        struct list_head *stock)
 114 {
 115         int i;
 116
 117         /*
 118          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 119          * from here: this will lead to infinite recursion.
 120          */
 121
 122         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++ i) {
 123                 struct trace_page *tage;
 124
 125                 tage = tage_alloc(gfp);
 126                 if (tage == NULL)
 127                         break;
 128                 list_add_tail(&tage->linkage, stock);
 129         }
 130         return i;
 131 }
 132
 133 /* return a page that has 'len' bytes left at the end */
 134 static struct trace_page *trace_get_tage_try(struct trace_cpu_data *tcd,
 135                                              unsigned long len)
 136 {
 137         struct trace_page *tage;
 138
 139         if (tcd->tcd_cur_pages > 0) {
 140                 __LASSERT(!list_empty(&tcd->tcd_pages));
 141                 tage = tage_from_list(tcd->tcd_pages.prev);
 142                 if (tage->used + len <= CFS_PAGE_SIZE)
 143                         return tage;
 144         }
 145
 146         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 147                 if (tcd->tcd_cur_stock_pages > 0) {
 148                         tage = tage_from_list(tcd->tcd_stock_pages.prev);
 149                         -- tcd->tcd_cur_stock_pages;
 150                         list_del_init(&tage->linkage);
 151                 } else {
 152                         tage = tage_alloc(CFS_ALLOC_ATOMIC);
 153                         if (tage == NULL) {
 154                                 printk(KERN_WARNING
 155                                        "failure to allocate a tage (%ld)\n",
 156                                        tcd->tcd_cur_pages);
 157                                 return NULL;
 158                         }
 159                 }
 160
 161                 tage->used = 0;
 162                 tage->cpu = smp_processor_id();
 163                 tage->type = tcd->tcd_type;
 164                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 165                 tcd->tcd_cur_pages++;
 166
 167                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 168                         struct tracefiled_ctl *tctl = &trace_tctl;
 169                         /*
 170                          * wake up tracefiled to process some pages.
 171                          */
 172                         cfs_waitq_signal(&tctl->tctl_waitq);
 173                 }
 174                 return tage;
 175         }
 176         return NULL;
 177 }
 178
 179 static void tcd_shrink(struct trace_cpu_data *tcd)
 180 {
 181         int pgcount = tcd->tcd_cur_pages / 10;
 182         struct page_collection pc;
 183         struct trace_page *tage;
 184         struct trace_page *tmp;
 185
 186         /*
 187          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 188          * from here: this will lead to infinite recursion.
 189          */
 190
 191         if (printk_ratelimit())
 192                 printk(KERN_WARNING "debug daemon buffer overflowed; "
 193                        "discarding  10%% of pages (%d of %ld)\n",
 194                        pgcount + 1, tcd->tcd_cur_pages);
 195
 196         CFS_INIT_LIST_HEAD(&pc.pc_pages);
 197         spin_lock_init(&pc.pc_lock);
 198
 199         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 200                 if (pgcount-- == 0)
 201                         break;
 202
 203                 list_move_tail(&tage->linkage, &pc.pc_pages);
 204                 tcd->tcd_cur_pages--;
 205         }
 206         put_pages_on_tcd_daemon_list(&pc, tcd);
 207 }
 208
 209 /* return a page that has 'len' bytes left at the end */
 210 static struct trace_page *trace_get_tage(struct trace_cpu_data *tcd,
 211                                          unsigned long len)
 212 {
 213         struct trace_page *tage;
 214
 215         /*
 216          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 217          * from here: this will lead to infinite recursion.
 218          */
 219
 220         if (len > CFS_PAGE_SIZE) {
 221                 printk(KERN_ERR
 222                        "cowardly refusing to write %lu bytes in a page\n", len);
 223                 return NULL;
 224         }
 225
 226         tage = trace_get_tage_try(tcd, len);
 227         if (tage != NULL)
 228                 return tage;
 229         if (thread_running)
 230                 tcd_shrink(tcd);
 231         if (tcd->tcd_cur_pages > 0) {
 232                 tage = tage_from_list(tcd->tcd_pages.next);
 233                 tage->used = 0;
 234                 tage_to_tail(tage, &tcd->tcd_pages);
 235         }
 236         return tage;
 237 }
 238
 239 int libcfs_debug_vmsg2(cfs_debug_limit_state_t *cdls, int subsys, int mask,
 240                        const char *file, const char *fn, const int line,
 241                        const char *format1, va_list args,
 242                        const char *format2, ...)
 243 {
 244         struct trace_cpu_data   *tcd = NULL;
 245         struct ptldebug_header   header;
 246         struct trace_page       *tage;
 247         /* string_buf is used only if tcd != NULL, and is always set then */
 248         char                    *string_buf = NULL;
 249         char                    *debug_buf;
 250         int                      known_size;
 251         int                      needed = 85; /* average message length */
 252         int                      max_nob;
 253         va_list                  ap;
 254         int                      depth;
 255         int                      i;
 256         int                      remain;
 257
 258         if (strchr(file, '/'))
 259                 file = strrchr(file, '/') + 1;
 260
 261         tcd = trace_get_tcd();
 262
 263         /* fs_trace_get_tcd() grabs a lock, which disables preemption and
 264          * pins us to a particular CPU.  This avoids an smp_processor_id()
 265          * warning on Linux when debugging is enabled. */
 266         set_ptldebug_header(&header, subsys, mask, line, CDEBUG_STACK());
 267
 268         if (tcd == NULL)                /* arch may not log in IRQ context */
 269                 goto console;
 270
 271         if (tcd->tcd_shutting_down) {
 272                 trace_put_tcd(tcd);
 273                 tcd = NULL;
 274                 goto console;
 275         }
 276
 277         depth = __current_nesting_level();
 278         known_size = strlen(file) + 1 + depth;
 279         if (fn)
 280                 known_size += strlen(fn) + 1;
 281
 282         if (libcfs_debug_binary)
 283                 known_size += sizeof(header);
 284
 285         /*/
 286          * '2' used because vsnprintf return real size required for output
 287          * _without_ terminating NULL.
 288          * if needed is to small for this format.
 289          */
 290         for (i=0;i<2;i++) {
 291                 tage = trace_get_tage(tcd, needed + known_size + 1);
 292                 if (tage == NULL) {
 293                         if (needed + known_size > CFS_PAGE_SIZE)
 294                                 mask |= D_ERROR;
 295
 296                         trace_put_tcd(tcd);
 297                         tcd = NULL;
 298                         goto console;
 299                 }
 300
 301                 string_buf = (char *)cfs_page_address(tage->page)+tage->used+known_size;
 302
 303                 max_nob = CFS_PAGE_SIZE - tage->used - known_size;
 304                 if (max_nob <= 0) {
 305                         printk(KERN_EMERG "negative max_nob: %i\n", max_nob);
 306                         mask |= D_ERROR;
 307                         trace_put_tcd(tcd);
 308                         tcd = NULL;
 309                         goto console;
 310                 }
 311
 312                 needed = 0;
 313                 if (format1) {
 314                         va_copy(ap, args);
 315                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 316                         va_end(ap);
 317                 }
 318
 319                 if (format2) {
 320                         remain = max_nob - needed;
 321                         if (remain < 0)
 322                                 remain = 0;
 323
 324                         va_start(ap, format2);
 325                         needed += vsnprintf(string_buf + needed, remain,
 326                                             format2, ap);
 327                         va_end(ap);
 328                 }
 329
 330                 if (needed < max_nob) /* well. printing ok.. */
 331                         break;
 332         }
 333
 334         if (*(string_buf+needed-1) != '\n')
 335                 printk(KERN_INFO "format at %s:%d:%s doesn't end in newline\n",
 336                        file, line, fn);
 337
 338         header.ph_len = known_size + needed;
 339         debug_buf = (char *)cfs_page_address(tage->page) + tage->used;
 340
 341         if (libcfs_debug_binary) {
 342                 memcpy(debug_buf, &header, sizeof(header));
 343                 tage->used += sizeof(header);
 344                 debug_buf += sizeof(header);
 345         }
 346
 347         /* indent message according to the nesting level */
 348         while (depth-- > 0) {
 349                 *(debug_buf++) = '.';
 350                 ++ tage->used;
 351         }
 352
 353         strcpy(debug_buf, file);
 354         tage->used += strlen(file) + 1;
 355         debug_buf += strlen(file) + 1;
 356
 357         if (fn) {
 358                 strcpy(debug_buf, fn);
 359                 tage->used += strlen(fn) + 1;
 360                 debug_buf += strlen(fn) + 1;
 361         }
 362
 363         __LASSERT(debug_buf == string_buf);
 364
 365         tage->used += needed;
 366         __LASSERT (tage->used <= CFS_PAGE_SIZE);
 367
 368 console:
 369         if ((mask & libcfs_printk) == 0) {
 370                 /* no console output requested */
 371                 if (tcd != NULL)
 372                         trace_put_tcd(tcd);
 373                 return 1;
 374         }
 375
 376         if (cdls != NULL) {
 377                 if (libcfs_console_ratelimit &&
 378                     cdls->cdls_next != 0 &&     /* not first time ever */
 379                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 380                         /* skipping a console message */
 381                         cdls->cdls_count++;
 382                         if (tcd != NULL)
 383                                 trace_put_tcd(tcd);
 384                         return 1;
 385                 }
 386
 387                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 388                                                        libcfs_console_max_delay
 389                                                        + cfs_time_seconds(10))) {
 390                         /* last timeout was a long time ago */
 391                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 392                 } else {
 393                         cdls->cdls_delay *= libcfs_console_backoff;
 394
 395                         if (cdls->cdls_delay < libcfs_console_min_delay)
 396                                 cdls->cdls_delay = libcfs_console_min_delay;
 397                         else if (cdls->cdls_delay > libcfs_console_max_delay)
 398                                 cdls->cdls_delay = libcfs_console_max_delay;
 399                 }
 400
 401                 /* ensure cdls_next is never zero after it's been seen */
 402                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 403         }
 404
 405         if (tcd != NULL) {
 406                 print_to_console(&header, mask, string_buf, needed, file, fn);
 407                 trace_put_tcd(tcd);
 408         } else {
 409                 string_buf = trace_get_console_buffer();
 410
 411                 needed = 0;
 412                 if (format1 != NULL) {
 413                         va_copy(ap, args);
 414                         needed = vsnprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE, format1, ap);
 415                         va_end(ap);
 416                 }
 417                 if (format2 != NULL) {
 418                         remain = TRACE_CONSOLE_BUFFER_SIZE - needed;
 419                         if (remain > 0) {
 420                                 va_start(ap, format2);
 421                                 needed += vsnprintf(string_buf+needed, remain, format2, ap);
 422                                 va_end(ap);
 423                         }
 424                 }
 425                 print_to_console(&header, mask,
 426                                  string_buf, needed, file, fn);
 427
 428                 trace_put_console_buffer(string_buf);
 429         }
 430
 431         if (cdls != NULL && cdls->cdls_count != 0) {
 432                 string_buf = trace_get_console_buffer();
 433
 434                 needed = snprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE,
 435                          "Skipped %d previous similar message%s\n",
 436                          cdls->cdls_count, (cdls->cdls_count > 1) ? "s" : "");
 437
 438                 print_to_console(&header, mask,
 439                                  string_buf, needed, file, fn);
 440
 441                 trace_put_console_buffer(string_buf);
 442                 cdls->cdls_count = 0;
 443         }
 444
 445         return 0;
 446 }
 447 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 448
 449 void
 450 libcfs_assertion_failed(const char *expr, const char *file,
 451                         const char *func, const int line)
 452 {
 453         libcfs_debug_msg(NULL, 0, D_EMERG, file, func, line,
 454                          "ASSERTION(%s) failed\n", expr);
 455         lbug_with_loc(file, func, line);
 456 }
 457 EXPORT_SYMBOL(libcfs_assertion_failed);
 458
 459 void
 460 trace_assertion_failed(const char *str,
 461                        const char *fn, const char *file, int line)
 462 {
 463         struct ptldebug_header hdr;
 464
 465         libcfs_panic_in_progress = 1;
 466         libcfs_catastrophe = 1;
 467         mb();
 468
 469         set_ptldebug_header(&hdr, DEBUG_SUBSYSTEM, D_EMERG, line,
 470                             CDEBUG_STACK());
 471
 472         print_to_console(&hdr, D_EMERG, str, strlen(str), file, fn);
 473
 474         LIBCFS_PANIC("Lustre debug assertion failure\n");
 475
 476         /* not reached */
 477 }
 478
 479 static void
 480 panic_collect_pages(struct page_collection *pc)
 481 {
 482         /* Do the collect_pages job on a single CPU: assumes that all other
 483          * CPUs have been stopped during a panic.  If this isn't true for some
 484          * arch, this will have to be implemented separately in each arch.  */
 485         int                    i;
 486         int                    j;
 487         struct trace_cpu_data *tcd;
 488
 489         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 490
 491         tcd_for_each(tcd, i, j) {
 492                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 493                 tcd->tcd_cur_pages = 0;
 494
 495                 if (pc->pc_want_daemon_pages) {
 496                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 497                         tcd->tcd_cur_daemon_pages = 0;
 498                 }
 499         }
 500 }
 501
 502 static void collect_pages_on_all_cpus(struct page_collection *pc)
 503 {
 504         struct trace_cpu_data *tcd;
 505         int i, cpu;
 506
 507         spin_lock(&pc->pc_lock);
 508         for_each_possible_cpu(cpu) {
 509                 tcd_for_each_type_lock(tcd, i, cpu) {
 510                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 511                         tcd->tcd_cur_pages = 0;
 512                         if (pc->pc_want_daemon_pages) {
 513                                 list_splice_init(&tcd->tcd_daemon_pages,
 514                                                  &pc->pc_pages);
 515                                 tcd->tcd_cur_daemon_pages = 0;
 516                         }
 517                 }
 518         }
 519         spin_unlock(&pc->pc_lock);
 520 }
 521
 522 static void collect_pages(struct page_collection *pc)
 523 {
 524         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 525
 526         if (libcfs_panic_in_progress)
 527                 panic_collect_pages(pc);
 528         else
 529                 collect_pages_on_all_cpus(pc);
 530 }
 531
 532 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 533 {
 534         struct trace_cpu_data *tcd;
 535         struct list_head *cur_head;
 536         struct trace_page *tage;
 537         struct trace_page *tmp;
 538         int i, cpu;
 539
 540         spin_lock(&pc->pc_lock);
 541         for_each_possible_cpu(cpu) {
 542                 tcd_for_each_type_lock(tcd, i, cpu) {
 543                         cur_head = tcd->tcd_pages.next;
 544
 545                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 546                                                  linkage) {
 547
 548                                 __LASSERT_TAGE_INVARIANT(tage);
 549
 550                                 if (tage->cpu != cpu || tage->type != i)
 551                                         continue;
 552
 553                                 tage_to_tail(tage, cur_head);
 554                                 tcd->tcd_cur_pages++;
 555                         }
 556                 }
 557         }
 558         spin_unlock(&pc->pc_lock);
 559 }
 560
 561 static void put_pages_back(struct page_collection *pc)
 562 {
 563         if (!libcfs_panic_in_progress)
 564                 put_pages_back_on_all_cpus(pc);
 565 }
 566
 567 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 568  * we have a good amount of data at all times for dumping during an LBUG, even
 569  * if we have been steadily writing (and otherwise discarding) pages via the
 570  * debug daemon. */
 571 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 572                                          struct trace_cpu_data *tcd)
 573 {
 574         struct trace_page *tage;
 575         struct trace_page *tmp;
 576
 577         spin_lock(&pc->pc_lock);
 578         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 579
 580                 __LASSERT_TAGE_INVARIANT(tage);
 581
 582                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 583                         continue;
 584
 585                 tage_to_tail(tage, &tcd->tcd_daemon_pages);
 586                 tcd->tcd_cur_daemon_pages++;
 587
 588                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 589                         struct trace_page *victim;
 590
 591                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 592                         victim = tage_from_list(tcd->tcd_daemon_pages.next);
 593
 594                         __LASSERT_TAGE_INVARIANT(victim);
 595
 596                         list_del(&victim->linkage);
 597                         tage_free(victim);
 598                         tcd->tcd_cur_daemon_pages--;
 599                 }
 600         }
 601         spin_unlock(&pc->pc_lock);
 602 }
 603
 604 static void put_pages_on_daemon_list(struct page_collection *pc)
 605 {
 606         struct trace_cpu_data *tcd;
 607         int i, cpu;
 608
 609         for_each_possible_cpu(cpu) {
 610                 tcd_for_each_type_lock(tcd, i, cpu)
 611                         put_pages_on_tcd_daemon_list(pc, tcd);
 612         }
 613 }
 614
 615 void trace_debug_print(void)
 616 {
 617         struct page_collection pc;
 618         struct trace_page *tage;
 619         struct trace_page *tmp;
 620
 621         spin_lock_init(&pc.pc_lock);
 622
 623         pc.pc_want_daemon_pages = 1;
 624         collect_pages(&pc);
 625         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 626                 char *p, *file, *fn;
 627                 cfs_page_t *page;
 628
 629                 __LASSERT_TAGE_INVARIANT(tage);
 630
 631                 page = tage->page;
 632                 p = cfs_page_address(page);
 633                 while (p < ((char *)cfs_page_address(page) + tage->used)) {
 634                         struct ptldebug_header *hdr;
 635                         int len;
 636                         hdr = (void *)p;
 637                         p += sizeof(*hdr);
 638                         file = p;
 639                         p += strlen(file) + 1;
 640                         fn = p;
 641                         p += strlen(fn) + 1;
 642                         len = hdr->ph_len - (p - (char *)hdr);
 643
 644                         print_to_console(hdr, D_EMERG, p, len, file, fn);
 645
 646                         p += len;
 647                 }
 648
 649                 list_del(&tage->linkage);
 650                 tage_free(tage);
 651         }
 652 }
 653
 654 int tracefile_dump_all_pages(char *filename)
 655 {
 656         struct page_collection pc;
 657         cfs_file_t *filp;
 658         struct trace_page *tage;
 659         struct trace_page *tmp;
 660         int rc;
 661
 662         CFS_DECL_MMSPACE;
 663
 664         tracefile_write_lock();
 665
 666         filp = cfs_filp_open(filename,
 667                              O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600, &rc);
 668         if (!filp) {
 669                 if (rc != -EEXIST)
 670                         printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 671                                filename, rc);
 672                 goto out;
 673         }
 674
 675         spin_lock_init(&pc.pc_lock);
 676         pc.pc_want_daemon_pages = 1;
 677         collect_pages(&pc);
 678         if (list_empty(&pc.pc_pages)) {
 679                 rc = 0;
 680                 goto close;
 681         }
 682
 683         /* ok, for now, just write the pages.  in the future we'll be building
 684          * iobufs with the pages and calling generic_direct_IO */
 685         CFS_MMSPACE_OPEN;
 686         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 687
 688                 __LASSERT_TAGE_INVARIANT(tage);
 689
 690                 rc = cfs_filp_write(filp, cfs_page_address(tage->page),
 691                                     tage->used, cfs_filp_poff(filp));
 692                 if (rc != (int)tage->used) {
 693                         printk(KERN_WARNING "wanted to write %u but wrote "
 694                                "%d\n", tage->used, rc);
 695                         put_pages_back(&pc);
 696                         __LASSERT(list_empty(&pc.pc_pages));
 697                         break;
 698                 }
 699                 list_del(&tage->linkage);
 700                 tage_free(tage);
 701         }
 702         CFS_MMSPACE_CLOSE;
 703         rc = cfs_filp_fsync(filp);
 704         if (rc)
 705                 printk(KERN_ERR "sync returns %d\n", rc);
 706  close:
 707         cfs_filp_close(filp);
 708  out:
 709         tracefile_write_unlock();
 710         return rc;
 711 }
 712
 713 void trace_flush_pages(void)
 714 {
 715         struct page_collection pc;
 716         struct trace_page *tage;
 717         struct trace_page *tmp;
 718
 719         spin_lock_init(&pc.pc_lock);
 720
 721         pc.pc_want_daemon_pages = 1;
 722         collect_pages(&pc);
 723         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 724
 725                 __LASSERT_TAGE_INVARIANT(tage);
 726
 727                 list_del(&tage->linkage);
 728                 tage_free(tage);
 729         }
 730 }
 731
 732 int trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 733                         const char *usr_buffer, int usr_buffer_nob)
 734 {
 735         int    nob;
 736
 737         if (usr_buffer_nob > knl_buffer_nob)
 738                 return -EOVERFLOW;
 739
 740         if (copy_from_user((void *)knl_buffer,
 741                            (void *)usr_buffer, usr_buffer_nob))
 742                 return -EFAULT;
 743
 744         nob = strnlen(knl_buffer, usr_buffer_nob);
 745         while (nob-- >= 0)                      /* strip trailing whitespace */
 746                 if (!isspace(knl_buffer[nob]))
 747                         break;
 748
 749         if (nob < 0)                            /* empty string */
 750                 return -EINVAL;
 751
 752         if (nob == knl_buffer_nob)              /* no space to terminate */
 753                 return -EOVERFLOW;
 754
 755         knl_buffer[nob + 1] = 0;                /* terminate */
 756         return 0;
 757 }
 758
 759 int trace_copyout_string(char *usr_buffer, int usr_buffer_nob,
 760                          const char *knl_buffer, char *append)
 761 {
 762         /* NB if 'append' != NULL, it's a single character to append to the
 763          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 764          * terminating zero byte) for sysctl entries */
 765         int   nob = strlen(knl_buffer);
 766
 767         if (nob > usr_buffer_nob)
 768                 nob = usr_buffer_nob;
 769
 770         if (copy_to_user(usr_buffer, knl_buffer, nob))
 771                 return -EFAULT;
 772
 773         if (append != NULL && nob < usr_buffer_nob) {
 774                 if (copy_to_user(usr_buffer + nob, append, 1))
 775                         return -EFAULT;
 776
 777                 nob++;
 778         }
 779
 780         return nob;
 781 }
 782 EXPORT_SYMBOL(trace_copyout_string);
 783
 784 int trace_allocate_string_buffer(char **str, int nob)
 785 {
 786         if (nob > 2 * CFS_PAGE_SIZE)            /* string must be "sensible" */
 787                 return -EINVAL;
 788
 789         *str = cfs_alloc(nob, CFS_ALLOC_STD | CFS_ALLOC_ZERO);
 790         if (*str == NULL)
 791                 return -ENOMEM;
 792
 793         return 0;
 794 }
 795
 796 void trace_free_string_buffer(char *str, int nob)
 797 {
 798         cfs_free(str);
 799 }
 800
 801 int trace_dump_debug_buffer_usrstr(void *usr_str, int usr_str_nob)
 802 {
 803         char         *str;
 804         int           rc;
 805
 806         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 807         if (rc != 0)
 808                 return rc;
 809
 810         rc = trace_copyin_string(str, usr_str_nob + 1,
 811                                  usr_str, usr_str_nob);
 812         if (rc != 0)
 813                 goto out;
 814
 815 #if !defined(__WINNT__)
 816         if (str[0] != '/') {
 817                 rc = -EINVAL;
 818                 goto out;
 819         }
 820 #endif
 821         rc = tracefile_dump_all_pages(str);
 822 out:
 823         trace_free_string_buffer(str, usr_str_nob + 1);
 824         return rc;
 825 }
 826
 827 int trace_daemon_command(char *str)
 828 {
 829         int       rc = 0;
 830
 831         tracefile_write_lock();
 832
 833         if (strcmp(str, "stop") == 0) {
 834                 tracefile_write_unlock();
 835                 trace_stop_thread();
 836                 tracefile_write_lock();
 837                 memset(tracefile, 0, sizeof(tracefile));
 838
 839         } else if (strncmp(str, "size=", 5) == 0) {
 840                 tracefile_size = simple_strtoul(str + 5, NULL, 0);
 841                 if (tracefile_size < 10 || tracefile_size > 20480)
 842                         tracefile_size = TRACEFILE_SIZE;
 843                 else
 844                         tracefile_size <<= 20;
 845
 846         } else if (strlen(str) >= sizeof(tracefile)) {
 847                 rc = -ENAMETOOLONG;
 848 #ifndef __WINNT__
 849         } else if (str[0] != '/') {
 850                 rc = -EINVAL;
 851 #endif
 852         } else {
 853                 strcpy(tracefile, str);
 854
 855                 printk(KERN_INFO "Lustre: debug daemon will attempt to start writing "
 856                        "to %s (%lukB max)\n", tracefile,
 857                        (long)(tracefile_size >> 10));
 858
 859                 trace_start_thread();
 860         }
 861
 862         tracefile_write_unlock();
 863         return rc;
 864 }
 865
 866 int trace_daemon_command_usrstr(void *usr_str, int usr_str_nob)
 867 {
 868         char *str;
 869         int   rc;
 870
 871         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 872         if (rc != 0)
 873                 return rc;
 874
 875         rc = trace_copyin_string(str, usr_str_nob + 1,
 876                                  usr_str, usr_str_nob);
 877         if (rc == 0)
 878                 rc = trace_daemon_command(str);
 879
 880         trace_free_string_buffer(str, usr_str_nob + 1);
 881
 882         return rc;
 883 }
 884
 885 int trace_set_debug_mb(int mb)
 886 {
 887         int i;
 888         int j;
 889         int pages;
 890         int limit = trace_max_debug_mb();
 891         struct trace_cpu_data *tcd;
 892
 893         if (mb < num_possible_cpus())
 894                 return -EINVAL;
 895
 896         if (mb > limit) {
 897                 printk(KERN_ERR "Lustre: Refusing to set debug buffer size to "
 898                        "%dMB - limit is %d\n", mb, limit);
 899                 return -EINVAL;
 900         }
 901
 902         mb /= num_possible_cpus();
 903         pages = mb << (20 - CFS_PAGE_SHIFT);
 904
 905         tracefile_write_lock();
 906
 907         tcd_for_each(tcd, i, j)
 908                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 909
 910         tracefile_write_unlock();
 911
 912         return 0;
 913 }
 914
 915 int trace_set_debug_mb_usrstr(void *usr_str, int usr_str_nob)
 916 {
 917         char     str[32];
 918         int      rc;
 919
 920         rc = trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 921         if (rc < 0)
 922                 return rc;
 923
 924         return trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 925 }
 926
 927 int trace_get_debug_mb(void)
 928 {
 929         int i;
 930         int j;
 931         struct trace_cpu_data *tcd;
 932         int total_pages = 0;
 933
 934         tracefile_read_lock();
 935
 936         tcd_for_each(tcd, i, j)
 937                 total_pages += tcd->tcd_max_pages;
 938
 939         tracefile_read_unlock();
 940
 941         return (total_pages >> (20 - CFS_PAGE_SHIFT)) + 1;
 942 }
 943
 944 static int tracefiled(void *arg)
 945 {
 946         struct page_collection pc;
 947         struct tracefiled_ctl *tctl = arg;
 948         struct trace_page *tage;
 949         struct trace_page *tmp;
 950         struct ptldebug_header *hdr;
 951         cfs_file_t *filp;
 952         int last_loop = 0;
 953         int rc;
 954
 955         CFS_DECL_MMSPACE;
 956
 957         /* we're started late enough that we pick up init's fs context */
 958         /* this is so broken in uml?  what on earth is going on? */
 959         cfs_daemonize("ktracefiled");
 960
 961         spin_lock_init(&pc.pc_lock);
 962         complete(&tctl->tctl_start);
 963
 964         while (1) {
 965                 cfs_waitlink_t __wait;
 966
 967                 pc.pc_want_daemon_pages = 0;
 968                 collect_pages(&pc);
 969                 if (list_empty(&pc.pc_pages))
 970                         goto end_loop;
 971
 972                 filp = NULL;
 973                 tracefile_read_lock();
 974                 if (tracefile[0] != 0) {
 975                         filp = cfs_filp_open(tracefile,
 976                                              O_CREAT | O_RDWR | O_LARGEFILE,
 977                                              0600, &rc);
 978                         if (!(filp))
 979                                 printk(KERN_WARNING "couldn't open %s: %d\n",
 980                                        tracefile, rc);
 981                 }
 982                 tracefile_read_unlock();
 983                 if (filp == NULL) {
 984                         put_pages_on_daemon_list(&pc);
 985                         __LASSERT(list_empty(&pc.pc_pages));
 986                         goto end_loop;
 987                 }
 988
 989                 CFS_MMSPACE_OPEN;
 990
 991                 /* mark the first header, so we can sort in chunks */
 992                 tage = tage_from_list(pc.pc_pages.next);
 993                 __LASSERT_TAGE_INVARIANT(tage);
 994
 995                 hdr = cfs_page_address(tage->page);
 996                 hdr->ph_flags |= PH_FLAG_FIRST_RECORD;
 997
 998                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 999                         static loff_t f_pos;
1000
1001                         __LASSERT_TAGE_INVARIANT(tage);
1002
1003                         if (f_pos >= (off_t)tracefile_size)
1004                                 f_pos = 0;
1005                         else if (f_pos > cfs_filp_size(filp))
1006                                 f_pos = cfs_filp_size(filp);
1007
1008                         rc = cfs_filp_write(filp, cfs_page_address(tage->page),
1009                                             tage->used, &f_pos);
1010                         if (rc != (int)tage->used) {
1011                                 printk(KERN_WARNING "wanted to write %u but "
1012                                        "wrote %d\n", tage->used, rc);
1013                                 put_pages_back(&pc);
1014                                 __LASSERT(list_empty(&pc.pc_pages));
1015                         }
1016                 }
1017                 CFS_MMSPACE_CLOSE;
1018
1019                 cfs_filp_close(filp);
1020                 put_pages_on_daemon_list(&pc);
1021                 if (!list_empty(&pc.pc_pages)) {
1022                         int i;
1023
1024                         printk(KERN_ALERT "Lustre: trace pages aren't empty\n");
1025                         printk(KERN_ALERT "total cpus(%d): ", num_possible_cpus());
1026                         for (i = 0; i < num_possible_cpus(); i++)
1027                                 if (cpu_online(i))
1028                                         printk(KERN_ALERT "%d(on) ", i);
1029                                 else
1030                                         printk(KERN_ALERT "%d(off) ", i);
1031                         printk(KERN_ALERT "\n");
1032
1033                         i = 0;
1034                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1035                                                  linkage)
1036                                 printk(KERN_ALERT "page %d belongs to cpu %d\n",
1037                                        ++i, tage->cpu);
1038                         printk(KERN_ALERT "There are %d pages unwritten\n", i);
1039                 }
1040                 __LASSERT(list_empty(&pc.pc_pages));
1041 end_loop:
1042                 if (atomic_read(&tctl->tctl_shutdown)) {
1043                         if (last_loop == 0) {
1044                                 last_loop = 1;
1045                                 continue;
1046                         } else {
1047                                 break;
1048                         }
1049                 }
1050                 cfs_waitlink_init(&__wait);
1051                 cfs_waitq_add(&tctl->tctl_waitq, &__wait);
1052                 set_current_state(TASK_INTERRUPTIBLE);
1053                 cfs_waitq_timedwait(&__wait, CFS_TASK_INTERRUPTIBLE,
1054                                     cfs_time_seconds(1));
1055                 cfs_waitq_del(&tctl->tctl_waitq, &__wait);
1056         }
1057         complete(&tctl->tctl_stop);
1058         return 0;
1059 }
1060
1061 int trace_start_thread(void)
1062 {
1063         struct tracefiled_ctl *tctl = &trace_tctl;
1064         int rc = 0;
1065
1066         mutex_down(&trace_thread_sem);
1067         if (thread_running)
1068                 goto out;
1069
1070         init_completion(&tctl->tctl_start);
1071         init_completion(&tctl->tctl_stop);
1072         cfs_waitq_init(&tctl->tctl_waitq);
1073         atomic_set(&tctl->tctl_shutdown, 0);
1074
1075         if (cfs_kernel_thread(tracefiled, tctl, 0) < 0) {
1076                 rc = -ECHILD;
1077                 goto out;
1078         }
1079
1080         wait_for_completion(&tctl->tctl_start);
1081         thread_running = 1;
1082 out:
1083         mutex_up(&trace_thread_sem);
1084         return rc;
1085 }
1086
1087 void trace_stop_thread(void)
1088 {
1089         struct tracefiled_ctl *tctl = &trace_tctl;
1090
1091         mutex_down(&trace_thread_sem);
1092         if (thread_running) {
1093                 printk(KERN_INFO "Lustre: shutting down debug daemon thread...\n");
1094                 atomic_set(&tctl->tctl_shutdown, 1);
1095                 wait_for_completion(&tctl->tctl_stop);
1096                 thread_running = 0;
1097         }
1098         mutex_up(&trace_thread_sem);
1099 }
1100
1101 int tracefile_init(int max_pages)
1102 {
1103         struct trace_cpu_data *tcd;
1104         int                    i;
1105         int                    j;
1106         int                    rc;
1107         int                    factor;
1108
1109         rc = tracefile_init_arch();
1110         if (rc != 0)
1111                 return rc;
1112
1113         tcd_for_each(tcd, i, j) {
1114                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1115                 factor = tcd->tcd_pages_factor;
1116                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
1117                 CFS_INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1118                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1119                 tcd->tcd_cur_pages = 0;
1120                 tcd->tcd_cur_stock_pages = 0;
1121                 tcd->tcd_cur_daemon_pages = 0;
1122                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1123                 LASSERT(tcd->tcd_max_pages > 0);
1124                 tcd->tcd_shutting_down = 0;
1125         }
1126
1127         return 0;
1128 }
1129
1130 static void trace_cleanup_on_all_cpus(void)
1131 {
1132         struct trace_cpu_data *tcd;
1133         struct trace_page *tage;
1134         struct trace_page *tmp;
1135         int i, cpu;
1136
1137         for_each_possible_cpu(cpu) {
1138                 tcd_for_each_type_lock(tcd, i, cpu) {
1139                         tcd->tcd_shutting_down = 1;
1140
1141                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages,
1142                                                  linkage) {
1143                                 __LASSERT_TAGE_INVARIANT(tage);
1144
1145                                 list_del(&tage->linkage);
1146                                 tage_free(tage);
1147                         }
1148
1149                         tcd->tcd_cur_pages = 0;
1150                 }
1151         }
1152 }
1153
1154 static void trace_cleanup(void)
1155 {
1156         struct page_collection pc;
1157
1158         CFS_INIT_LIST_HEAD(&pc.pc_pages);
1159         spin_lock_init(&pc.pc_lock);
1160
1161         trace_cleanup_on_all_cpus();
1162
1163         tracefile_fini_arch();
1164 }
1165
1166 void tracefile_exit(void)
1167 {
1168         trace_stop_thread();
1169         trace_cleanup();
1170 }