libcfs/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  2008 Sun Microsystems, Inc. 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  * libcfs/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/libcfs.h>
  48
  49 /* XXX move things up to the top, comment */
  50 union trace_data_union (*trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  51
  52 char tracefile[TRACEFILE_NAME_SIZE];
  53 long long tracefile_size = TRACEFILE_SIZE;
  54 static struct tracefiled_ctl trace_tctl;
  55 struct semaphore trace_thread_sem;
  56 static int thread_running = 0;
  57
  58 atomic_t tage_allocated = ATOMIC_INIT(0);
  59
  60 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  61                                          struct trace_cpu_data *tcd);
  62
  63 static inline struct trace_page *tage_from_list(struct list_head *list)
  64 {
  65         return list_entry(list, struct trace_page, linkage);
  66 }
  67
  68 static struct trace_page *tage_alloc(int gfp)
  69 {
  70         cfs_page_t        *page;
  71         struct trace_page *tage;
  72
  73         /*
  74          * Don't spam console with allocation failures: they will be reported
  75          * by upper layer anyway.
  76          */
  77         gfp |= CFS_ALLOC_NOWARN;
  78         page = cfs_alloc_page(gfp);
  79         if (page == NULL)
  80                 return NULL;
  81
  82         tage = cfs_alloc(sizeof(*tage), gfp);
  83         if (tage == NULL) {
  84                 cfs_free_page(page);
  85                 return NULL;
  86         }
  87
  88         tage->page = page;
  89         atomic_inc(&tage_allocated);
  90         return tage;
  91 }
  92
  93 static void tage_free(struct trace_page *tage)
  94 {
  95         __LASSERT(tage != NULL);
  96         __LASSERT(tage->page != NULL);
  97
  98         cfs_free_page(tage->page);
  99         cfs_free(tage);
 100         atomic_dec(&tage_allocated);
 101 }
 102
 103 static void tage_to_tail(struct trace_page *tage, struct list_head *queue)
 104 {
 105         __LASSERT(tage != NULL);
 106         __LASSERT(queue != NULL);
 107
 108         list_move_tail(&tage->linkage, queue);
 109 }
 110
 111 int trace_refill_stock(struct trace_cpu_data *tcd, int gfp,
 112                        struct list_head *stock)
 113 {
 114         int i;
 115
 116         /*
 117          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 118          * from here: this will lead to infinite recursion.
 119          */
 120
 121         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++ i) {
 122                 struct trace_page *tage;
 123
 124                 tage = tage_alloc(gfp);
 125                 if (tage == NULL)
 126                         break;
 127                 list_add_tail(&tage->linkage, stock);
 128         }
 129         return i;
 130 }
 131
 132 /* return a page that has 'len' bytes left at the end */
 133 static struct trace_page *trace_get_tage_try(struct trace_cpu_data *tcd,
 134                                              unsigned long len)
 135 {
 136         struct trace_page *tage;
 137
 138         if (tcd->tcd_cur_pages > 0) {
 139                 __LASSERT(!list_empty(&tcd->tcd_pages));
 140                 tage = tage_from_list(tcd->tcd_pages.prev);
 141                 if (tage->used + len <= CFS_PAGE_SIZE)
 142                         return tage;
 143         }
 144
 145         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 146                 if (tcd->tcd_cur_stock_pages > 0) {
 147                         tage = tage_from_list(tcd->tcd_stock_pages.prev);
 148                         -- tcd->tcd_cur_stock_pages;
 149                         list_del_init(&tage->linkage);
 150                 } else {
 151                         tage = tage_alloc(CFS_ALLOC_ATOMIC);
 152                         if (tage == NULL) {
 153                                 if (printk_ratelimit())
 154                                         printk(KERN_WARNING
 155                                                "cannot 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
 262         set_ptldebug_header(&header, subsys, mask, line, CDEBUG_STACK());
 263
 264         tcd = trace_get_tcd();
 265         if (tcd == NULL)                /* arch may not log in IRQ context */
 266                 goto console;
 267
 268         if (tcd->tcd_shutting_down) {
 269                 trace_put_tcd(tcd);
 270                 tcd = NULL;
 271                 goto console;
 272         }
 273
 274         depth = __current_nesting_level();
 275         known_size = strlen(file) + 1 + depth;
 276         if (fn)
 277                 known_size += strlen(fn) + 1;
 278
 279         if (libcfs_debug_binary)
 280                 known_size += sizeof(header);
 281
 282         /*/
 283          * '2' used because vsnprintf return real size required for output
 284          * _without_ terminating NULL.
 285          * if needed is to small for this format.
 286          */
 287         for (i=0;i<2;i++) {
 288                 tage = trace_get_tage(tcd, needed + known_size + 1);
 289                 if (tage == NULL) {
 290                         if (needed + known_size > CFS_PAGE_SIZE)
 291                                 mask |= D_ERROR;
 292
 293                         trace_put_tcd(tcd);
 294                         tcd = NULL;
 295                         goto console;
 296                 }
 297
 298                 string_buf = (char *)cfs_page_address(tage->page)+tage->used+known_size;
 299
 300                 max_nob = CFS_PAGE_SIZE - tage->used - known_size;
 301                 if (max_nob <= 0) {
 302                         printk(KERN_EMERG "negative max_nob: %i\n", max_nob);
 303                         mask |= D_ERROR;
 304                         trace_put_tcd(tcd);
 305                         tcd = NULL;
 306                         goto console;
 307                 }
 308
 309                 needed = 0;
 310                 if (format1) {
 311                         va_copy(ap, args);
 312                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 313                         va_end(ap);
 314                 }
 315
 316
 317                 if (format2) {
 318                         remain = max_nob - needed;
 319                         if (remain < 0)
 320                                 remain = 0;
 321
 322                         va_start(ap, format2);
 323                         needed += vsnprintf(string_buf+needed, remain, format2, ap);
 324                         va_end(ap);
 325                 }
 326
 327                 if (needed < max_nob) /* well. printing ok.. */
 328                         break;
 329         }
 330
 331         if (*(string_buf+needed-1) != '\n')
 332                 printk(KERN_INFO "format at %s:%d:%s doesn't end in newline\n",
 333                        file, line, fn);
 334
 335         header.ph_len = known_size + needed;
 336         debug_buf = (char *)cfs_page_address(tage->page) + tage->used;
 337
 338         if (libcfs_debug_binary) {
 339                 memcpy(debug_buf, &header, sizeof(header));
 340                 tage->used += sizeof(header);
 341                 debug_buf += sizeof(header);
 342         }
 343
 344         /* indent message according to the nesting level */
 345         while (depth-- > 0) {
 346                 *(debug_buf++) = '.';
 347                 ++ tage->used;
 348         }
 349
 350         strcpy(debug_buf, file);
 351         tage->used += strlen(file) + 1;
 352         debug_buf += strlen(file) + 1;
 353
 354         if (fn) {
 355                 strcpy(debug_buf, fn);
 356                 tage->used += strlen(fn) + 1;
 357                 debug_buf += strlen(fn) + 1;
 358         }
 359
 360         __LASSERT(debug_buf == string_buf);
 361
 362         tage->used += needed;
 363         __LASSERT (tage->used <= CFS_PAGE_SIZE);
 364
 365 console:
 366         if ((mask & libcfs_printk) == 0) {
 367                 /* no console output requested */
 368                 if (tcd != NULL)
 369                         trace_put_tcd(tcd);
 370                 return 1;
 371         }
 372
 373         if (cdls != NULL) {
 374                 if (libcfs_console_ratelimit &&
 375                     cdls->cdls_next != 0 &&     /* not first time ever */
 376                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 377                         /* skipping a console message */
 378                         cdls->cdls_count++;
 379                         if (tcd != NULL)
 380                                 trace_put_tcd(tcd);
 381                         return 1;
 382                 }
 383
 384                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 385                                                        libcfs_console_max_delay
 386                                                        + cfs_time_seconds(10))) {
 387                         /* last timeout was a long time ago */
 388                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 389                 } else {
 390                         cdls->cdls_delay *= libcfs_console_backoff;
 391
 392                         if (cdls->cdls_delay < libcfs_console_min_delay)
 393                                 cdls->cdls_delay = libcfs_console_min_delay;
 394                         else if (cdls->cdls_delay > libcfs_console_max_delay)
 395                                 cdls->cdls_delay = libcfs_console_max_delay;
 396                 }
 397
 398                 /* ensure cdls_next is never zero after it's been seen */
 399                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 400         }
 401
 402         if (tcd != NULL) {
 403                 print_to_console(&header, mask, string_buf, needed, file, fn);
 404                 trace_put_tcd(tcd);
 405         } else {
 406                 string_buf = trace_get_console_buffer();
 407
 408                 needed = 0;
 409                 if (format1 != NULL) {
 410                         va_copy(ap, args);
 411                         needed = vsnprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE, format1, ap);
 412                         va_end(ap);
 413                 }
 414                 if (format2 != NULL) {
 415                         remain = TRACE_CONSOLE_BUFFER_SIZE - needed;
 416                         if (remain > 0) {
 417                                 va_start(ap, format2);
 418                                 needed += vsnprintf(string_buf+needed, remain, format2, ap);
 419                                 va_end(ap);
 420                         }
 421                 }
 422                 print_to_console(&header, mask,
 423                                  string_buf, needed, file, fn);
 424
 425                 trace_put_console_buffer(string_buf);
 426         }
 427
 428         if (cdls != NULL && cdls->cdls_count != 0) {
 429                 string_buf = trace_get_console_buffer();
 430
 431                 needed = snprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE,
 432                          "Skipped %d previous similar message%s\n",
 433                          cdls->cdls_count, (cdls->cdls_count > 1) ? "s" : "");
 434
 435                 print_to_console(&header, mask,
 436                                  string_buf, needed, file, fn);
 437
 438                 trace_put_console_buffer(string_buf);
 439                 cdls->cdls_count = 0;
 440         }
 441
 442         return 0;
 443 }
 444 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 445
 446 void
 447 libcfs_assertion_failed(const char *expr, const char *file,
 448                         const char *func, const int line)
 449 {
 450         libcfs_debug_msg(NULL, 0, D_EMERG, file, func, line,
 451                          "ASSERTION(%s) failed\n", expr);
 452         cfs_enter_debugger();
 453         lbug_with_loc(file, func, line);
 454 }
 455 EXPORT_SYMBOL(libcfs_assertion_failed);
 456
 457 void
 458 trace_assertion_failed(const char *str,
 459                        const char *fn, const char *file, int line)
 460 {
 461         struct ptldebug_header hdr;
 462
 463         libcfs_panic_in_progress = 1;
 464         libcfs_catastrophe = 1;
 465         mb();
 466
 467         set_ptldebug_header(&hdr, DEBUG_SUBSYSTEM, D_EMERG, line,
 468                             CDEBUG_STACK());
 469
 470         print_to_console(&hdr, D_EMERG, str, strlen(str), file, fn);
 471
 472         LIBCFS_PANIC("Lustre debug assertion failure\n");
 473
 474         /* not reached */
 475 }
 476
 477 static void
 478 panic_collect_pages(struct page_collection *pc)
 479 {
 480         /* Do the collect_pages job on a single CPU: assumes that all other
 481          * CPUs have been stopped during a panic.  If this isn't true for some
 482          * arch, this will have to be implemented separately in each arch.  */
 483         int                    i;
 484         int                    j;
 485         struct trace_cpu_data *tcd;
 486
 487         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 488
 489         tcd_for_each(tcd, i, j) {
 490                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 491                 tcd->tcd_cur_pages = 0;
 492
 493                 if (pc->pc_want_daemon_pages) {
 494                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 495                         tcd->tcd_cur_daemon_pages = 0;
 496                 }
 497         }
 498 }
 499
 500 static void collect_pages_on_cpu(void *info)
 501 {
 502         struct trace_cpu_data *tcd;
 503         struct page_collection *pc = info;
 504         int i;
 505
 506         spin_lock(&pc->pc_lock);
 507         tcd_for_each_type_lock(tcd, i) {
 508                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 509                 tcd->tcd_cur_pages = 0;
 510                 if (pc->pc_want_daemon_pages) {
 511                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 512                         tcd->tcd_cur_daemon_pages = 0;
 513                 }
 514         }
 515         spin_unlock(&pc->pc_lock);
 516 }
 517
 518 static void collect_pages(struct page_collection *pc)
 519 {
 520         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 521
 522         if (libcfs_panic_in_progress)
 523                 panic_collect_pages(pc);
 524         else
 525                 trace_call_on_all_cpus(collect_pages_on_cpu, pc);
 526 }
 527
 528 static void put_pages_back_on_cpu(void *info)
 529 {
 530         struct page_collection *pc = info;
 531         struct trace_cpu_data *tcd;
 532         struct list_head *cur_head;
 533         struct trace_page *tage;
 534         struct trace_page *tmp;
 535         int i;
 536
 537         spin_lock(&pc->pc_lock);
 538         tcd_for_each_type_lock(tcd, i) {
 539                 cur_head = tcd->tcd_pages.next;
 540
 541                 list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 542
 543                         __LASSERT_TAGE_INVARIANT(tage);
 544
 545                         if (tage->cpu != smp_processor_id() || tage->type != i)
 546                                 continue;
 547
 548                         tage_to_tail(tage, cur_head);
 549                         tcd->tcd_cur_pages++;
 550                 }
 551         }
 552         spin_unlock(&pc->pc_lock);
 553 }
 554
 555 static void put_pages_back(struct page_collection *pc)
 556 {
 557         if (!libcfs_panic_in_progress)
 558                 trace_call_on_all_cpus(put_pages_back_on_cpu, pc);
 559 }
 560
 561 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 562  * we have a good amount of data at all times for dumping during an LBUG, even
 563  * if we have been steadily writing (and otherwise discarding) pages via the
 564  * debug daemon. */
 565 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 566                                          struct trace_cpu_data *tcd)
 567 {
 568         struct trace_page *tage;
 569         struct trace_page *tmp;
 570
 571         spin_lock(&pc->pc_lock);
 572         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 573
 574                 __LASSERT_TAGE_INVARIANT(tage);
 575
 576                 if (tage->cpu != smp_processor_id() ||
 577                     tage->type != tcd->tcd_type)
 578                         continue;
 579
 580                 tage_to_tail(tage, &tcd->tcd_daemon_pages);
 581                 tcd->tcd_cur_daemon_pages++;
 582
 583                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 584                         struct trace_page *victim;
 585
 586                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 587                         victim = tage_from_list(tcd->tcd_daemon_pages.next);
 588
 589                         __LASSERT_TAGE_INVARIANT(victim);
 590
 591                         list_del(&victim->linkage);
 592                         tage_free(victim);
 593                         tcd->tcd_cur_daemon_pages--;
 594                 }
 595         }
 596         spin_unlock(&pc->pc_lock);
 597 }
 598
 599 static void put_pages_on_daemon_list_on_cpu(void *info)
 600 {
 601         struct trace_cpu_data *tcd;
 602         int i;
 603
 604         tcd_for_each_type_lock(tcd, i)
 605                 put_pages_on_tcd_daemon_list(info, tcd);
 606 }
 607
 608 static void put_pages_on_daemon_list(struct page_collection *pc)
 609 {
 610         trace_call_on_all_cpus(put_pages_on_daemon_list_on_cpu, pc);
 611 }
 612
 613 void trace_debug_print(void)
 614 {
 615         struct page_collection pc;
 616         struct trace_page *tage;
 617         struct trace_page *tmp;
 618
 619         spin_lock_init(&pc.pc_lock);
 620
 621         pc.pc_want_daemon_pages = 1;
 622         collect_pages(&pc);
 623         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 624                 char *p, *file, *fn;
 625                 cfs_page_t *page;
 626
 627                 __LASSERT_TAGE_INVARIANT(tage);
 628
 629                 page = tage->page;
 630                 p = cfs_page_address(page);
 631                 while (p < ((char *)cfs_page_address(page) + tage->used)) {
 632                         struct ptldebug_header *hdr;
 633                         int len;
 634                         hdr = (void *)p;
 635                         p += sizeof(*hdr);
 636                         file = p;
 637                         p += strlen(file) + 1;
 638                         fn = p;
 639                         p += strlen(fn) + 1;
 640                         len = hdr->ph_len - (p - (char *)hdr);
 641
 642                         print_to_console(hdr, D_EMERG, p, len, file, fn);
 643
 644                         p += len;
 645                 }
 646
 647                 list_del(&tage->linkage);
 648                 tage_free(tage);
 649         }
 650 }
 651
 652 int tracefile_dump_all_pages(char *filename)
 653 {
 654         struct page_collection pc;
 655         cfs_file_t *filp;
 656         struct trace_page *tage;
 657         struct trace_page *tmp;
 658         int rc;
 659
 660         CFS_DECL_MMSPACE;
 661
 662         tracefile_write_lock();
 663
 664         filp = cfs_filp_open(filename,
 665                              O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600, &rc);
 666         if (!filp) {
 667                 if (rc != -EEXIST)
 668                         printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 669                                filename, rc);
 670                 goto out;
 671         }
 672
 673         spin_lock_init(&pc.pc_lock);
 674         pc.pc_want_daemon_pages = 1;
 675         collect_pages(&pc);
 676         if (list_empty(&pc.pc_pages)) {
 677                 rc = 0;
 678                 goto close;
 679         }
 680
 681         /* ok, for now, just write the pages.  in the future we'll be building
 682          * iobufs with the pages and calling generic_direct_IO */
 683         CFS_MMSPACE_OPEN;
 684         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 685
 686                 __LASSERT_TAGE_INVARIANT(tage);
 687
 688                 rc = cfs_filp_write(filp, cfs_page_address(tage->page),
 689                                     tage->used, cfs_filp_poff(filp));
 690                 if (rc != (int)tage->used) {
 691                         printk(KERN_WARNING "wanted to write %u but wrote "
 692                                "%d\n", tage->used, rc);
 693                         put_pages_back(&pc);
 694                         __LASSERT(list_empty(&pc.pc_pages));
 695                         break;
 696                 }
 697                 list_del(&tage->linkage);
 698                 tage_free(tage);
 699         }
 700         CFS_MMSPACE_CLOSE;
 701         rc = cfs_filp_fsync(filp);
 702         if (rc)
 703                 printk(KERN_ERR "sync returns %d\n", rc);
 704  close:
 705         cfs_filp_close(filp);
 706  out:
 707         tracefile_write_unlock();
 708         return rc;
 709 }
 710
 711 void trace_flush_pages(void)
 712 {
 713         struct page_collection pc;
 714         struct trace_page *tage;
 715         struct trace_page *tmp;
 716
 717         spin_lock_init(&pc.pc_lock);
 718
 719         pc.pc_want_daemon_pages = 1;
 720         collect_pages(&pc);
 721         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 722
 723                 __LASSERT_TAGE_INVARIANT(tage);
 724
 725                 list_del(&tage->linkage);
 726                 tage_free(tage);
 727         }
 728 }
 729
 730 int trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 731                         const char *usr_buffer, int usr_buffer_nob)
 732 {
 733         int    nob;
 734
 735         if (usr_buffer_nob > knl_buffer_nob)
 736                 return -EOVERFLOW;
 737
 738         if (copy_from_user((void *)knl_buffer,
 739                            (void *)usr_buffer, usr_buffer_nob))
 740                 return -EFAULT;
 741
 742         nob = strnlen(knl_buffer, usr_buffer_nob);
 743         while (nob-- >= 0)                      /* strip trailing whitespace */
 744                 if (!isspace(knl_buffer[nob]))
 745                         break;
 746
 747         if (nob < 0)                            /* empty string */
 748                 return -EINVAL;
 749
 750         if (nob == knl_buffer_nob)              /* no space to terminate */
 751                 return -EOVERFLOW;
 752
 753         knl_buffer[nob + 1] = 0;                /* terminate */
 754         return 0;
 755 }
 756
 757 int trace_copyout_string(char *usr_buffer, int usr_buffer_nob,
 758                          const char *knl_buffer, char *append)
 759 {
 760         /* NB if 'append' != NULL, it's a single character to append to the
 761          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 762          * terminating zero byte) for sysctl entries */
 763         int   nob = strlen(knl_buffer);
 764
 765         if (nob > usr_buffer_nob)
 766                 nob = usr_buffer_nob;
 767
 768         if (copy_to_user(usr_buffer, knl_buffer, nob))
 769                 return -EFAULT;
 770
 771         if (append != NULL && nob < usr_buffer_nob) {
 772                 if (copy_to_user(usr_buffer + nob, append, 1))
 773                         return -EFAULT;
 774
 775                 nob++;
 776         }
 777
 778         return nob;
 779 }
 780
 781 int trace_allocate_string_buffer(char **str, int nob)
 782 {
 783         if (nob > 2 * CFS_PAGE_SIZE)            /* string must be "sensible" */
 784                 return -EINVAL;
 785
 786         *str = cfs_alloc(nob, CFS_ALLOC_STD | CFS_ALLOC_ZERO);
 787         if (*str == NULL)
 788                 return -ENOMEM;
 789
 790         return 0;
 791 }
 792
 793 void trace_free_string_buffer(char *str, int nob)
 794 {
 795         cfs_free(str);
 796 }
 797
 798 int trace_dump_debug_buffer_usrstr(void *usr_str, int usr_str_nob)
 799 {
 800         char         *str;
 801         int           rc;
 802
 803         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 804         if (rc != 0)
 805                 return rc;
 806
 807         rc = trace_copyin_string(str, usr_str_nob + 1,
 808                                  usr_str, usr_str_nob);
 809         if (rc != 0)
 810                 goto out;
 811
 812 #if !defined(__WINNT__)
 813         if (str[0] != '/') {
 814                 rc = -EINVAL;
 815                 goto out;
 816         }
 817 #endif
 818         rc = tracefile_dump_all_pages(str);
 819 out:
 820         trace_free_string_buffer(str, usr_str_nob + 1);
 821         return rc;
 822 }
 823
 824 int trace_daemon_command(char *str)
 825 {
 826         int       rc = 0;
 827
 828         tracefile_write_lock();
 829
 830         if (strcmp(str, "stop") == 0) {
 831                 trace_stop_thread();
 832                 memset(tracefile, 0, sizeof(tracefile));
 833
 834         } else if (strncmp(str, "size=", 5) == 0) {
 835                 tracefile_size = simple_strtoul(str + 5, NULL, 0);
 836                 if (tracefile_size < 10 || tracefile_size > 20480)
 837                         tracefile_size = TRACEFILE_SIZE;
 838                 else
 839                         tracefile_size <<= 20;
 840
 841         } else if (strlen(str) >= sizeof(tracefile)) {
 842                 rc = -ENAMETOOLONG;
 843 #ifndef __WINNT__
 844         } else if (str[0] != '/') {
 845                 rc = -EINVAL;
 846 #endif
 847         } else {
 848                 strcpy(tracefile, str);
 849
 850                 printk(KERN_INFO "Lustre: debug daemon will attempt to start writing "
 851                        "to %s (%lukB max)\n", tracefile,
 852                        (long)(tracefile_size >> 10));
 853
 854                 trace_start_thread();
 855         }
 856
 857         tracefile_write_unlock();
 858         return rc;
 859 }
 860
 861 int trace_daemon_command_usrstr(void *usr_str, int usr_str_nob)
 862 {
 863         char *str;
 864         int   rc;
 865
 866         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 867         if (rc != 0)
 868                 return rc;
 869
 870         rc = trace_copyin_string(str, usr_str_nob + 1,
 871                                  usr_str, usr_str_nob);
 872         if (rc == 0)
 873                 rc = trace_daemon_command(str);
 874
 875         trace_free_string_buffer(str, usr_str_nob + 1);
 876         return rc;
 877 }
 878
 879 int trace_set_debug_mb(int mb)
 880 {
 881         int i;
 882         int j;
 883         int pages;
 884         int limit = trace_max_debug_mb();
 885         struct trace_cpu_data *tcd;
 886
 887         if (mb < num_possible_cpus())
 888                 return -EINVAL;
 889
 890         if (mb > limit) {
 891                 printk(KERN_ERR "Lustre: Refusing to set debug buffer size to "
 892                        "%dMB - limit is %d\n", mb, limit);
 893                 return -EINVAL;
 894         }
 895
 896         mb /= num_possible_cpus();
 897         pages = mb << (20 - CFS_PAGE_SHIFT);
 898
 899         tracefile_write_lock();
 900
 901         tcd_for_each(tcd, i, j)
 902                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 903
 904         tracefile_write_unlock();
 905
 906         return 0;
 907 }
 908
 909 int trace_set_debug_mb_usrstr(void *usr_str, int usr_str_nob)
 910 {
 911         char     str[32];
 912         int      rc;
 913
 914         rc = trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 915         if (rc < 0)
 916                 return rc;
 917
 918         return trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 919 }
 920
 921 int trace_get_debug_mb(void)
 922 {
 923         int i;
 924         int j;
 925         struct trace_cpu_data *tcd;
 926         int total_pages = 0;
 927
 928         tracefile_read_lock();
 929
 930         tcd_for_each(tcd, i, j)
 931                 total_pages += tcd->tcd_max_pages;
 932
 933         tracefile_read_unlock();
 934
 935         return (total_pages >> (20 - CFS_PAGE_SHIFT)) + 1;
 936 }
 937
 938 static int tracefiled(void *arg)
 939 {
 940         struct page_collection pc;
 941         struct tracefiled_ctl *tctl = arg;
 942         struct trace_page *tage;
 943         struct trace_page *tmp;
 944         struct ptldebug_header *hdr;
 945         cfs_file_t *filp;
 946         int rc;
 947
 948         CFS_DECL_MMSPACE;
 949
 950         /* we're started late enough that we pick up init's fs context */
 951         /* this is so broken in uml?  what on earth is going on? */
 952         cfs_daemonize("ktracefiled");
 953
 954         spin_lock_init(&pc.pc_lock);
 955         complete(&tctl->tctl_start);
 956
 957         while (1) {
 958                 cfs_waitlink_t __wait;
 959
 960                 cfs_waitlink_init(&__wait);
 961                 cfs_waitq_add(&tctl->tctl_waitq, &__wait);
 962                 set_current_state(TASK_INTERRUPTIBLE);
 963                 cfs_waitq_timedwait(&__wait, CFS_TASK_INTERRUPTIBLE,
 964                                     cfs_time_seconds(1));
 965                 cfs_waitq_del(&tctl->tctl_waitq, &__wait);
 966
 967                 if (atomic_read(&tctl->tctl_shutdown))
 968                         break;
 969
 970                 pc.pc_want_daemon_pages = 0;
 971                 collect_pages(&pc);
 972                 if (list_empty(&pc.pc_pages))
 973                         continue;
 974
 975                 filp = NULL;
 976                 tracefile_read_lock();
 977                 if (tracefile[0] != 0) {
 978                         filp = cfs_filp_open(tracefile,
 979                                              O_CREAT | O_RDWR | O_LARGEFILE,
 980                                              0600, &rc);
 981                         if (!(filp))
 982                                 printk(KERN_WARNING "couldn't open %s: %d\n",
 983                                        tracefile, rc);
 984                 }
 985                 tracefile_read_unlock();
 986                 if (filp == NULL) {
 987                         put_pages_on_daemon_list(&pc);
 988                         __LASSERT(list_empty(&pc.pc_pages));
 989                         continue;
 990                 }
 991
 992                 CFS_MMSPACE_OPEN;
 993
 994                 /* mark the first header, so we can sort in chunks */
 995                 tage = tage_from_list(pc.pc_pages.next);
 996                 __LASSERT_TAGE_INVARIANT(tage);
 997
 998                 hdr = cfs_page_address(tage->page);
 999                 hdr->ph_flags |= PH_FLAG_FIRST_RECORD;
1000
1001                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
1002                         static loff_t f_pos;
1003
1004                         __LASSERT_TAGE_INVARIANT(tage);
1005
1006                         if (f_pos >= (off_t)tracefile_size)
1007                                 f_pos = 0;
1008                         else if (f_pos > cfs_filp_size(filp))
1009                                 f_pos = cfs_filp_size(filp);
1010
1011                         rc = cfs_filp_write(filp, cfs_page_address(tage->page),
1012                                             tage->used, &f_pos);
1013                         if (rc != (int)tage->used) {
1014                                 printk(KERN_WARNING "wanted to write %u but "
1015                                        "wrote %d\n", tage->used, rc);
1016                                 put_pages_back(&pc);
1017                                 __LASSERT(list_empty(&pc.pc_pages));
1018                         }
1019                 }
1020                 CFS_MMSPACE_CLOSE;
1021
1022                 cfs_filp_close(filp);
1023                 put_pages_on_daemon_list(&pc);
1024                 __LASSERT(list_empty(&pc.pc_pages));
1025         }
1026         complete(&tctl->tctl_stop);
1027         return 0;
1028 }
1029
1030 int trace_start_thread(void)
1031 {
1032         struct tracefiled_ctl *tctl = &trace_tctl;
1033         int rc = 0;
1034
1035         mutex_down(&trace_thread_sem);
1036         if (thread_running)
1037                 goto out;
1038
1039         init_completion(&tctl->tctl_start);
1040         init_completion(&tctl->tctl_stop);
1041         cfs_waitq_init(&tctl->tctl_waitq);
1042         atomic_set(&tctl->tctl_shutdown, 0);
1043
1044         if (cfs_kernel_thread(tracefiled, tctl, 0) < 0) {
1045                 rc = -ECHILD;
1046                 goto out;
1047         }
1048
1049         wait_for_completion(&tctl->tctl_start);
1050         thread_running = 1;
1051 out:
1052         mutex_up(&trace_thread_sem);
1053         return rc;
1054 }
1055
1056 void trace_stop_thread(void)
1057 {
1058         struct tracefiled_ctl *tctl = &trace_tctl;
1059
1060         mutex_down(&trace_thread_sem);
1061         if (thread_running) {
1062                 printk(KERN_INFO "Lustre: shutting down debug daemon thread...\n");
1063                 atomic_set(&tctl->tctl_shutdown, 1);
1064                 wait_for_completion(&tctl->tctl_stop);
1065                 thread_running = 0;
1066         }
1067         mutex_up(&trace_thread_sem);
1068 }
1069
1070 int tracefile_init(int max_pages)
1071 {
1072         struct trace_cpu_data *tcd;
1073         int                    i;
1074         int                    j;
1075         int                    rc;
1076         int                    factor;
1077
1078         rc = tracefile_init_arch();
1079         if (rc != 0)
1080                 return rc;
1081
1082         tcd_for_each(tcd, i, j) {
1083                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1084                 factor = tcd->tcd_pages_factor;
1085                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
1086                 CFS_INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1087                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1088                 tcd->tcd_cur_pages = 0;
1089                 tcd->tcd_cur_stock_pages = 0;
1090                 tcd->tcd_cur_daemon_pages = 0;
1091                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1092                 LASSERT(tcd->tcd_max_pages > 0);
1093                 tcd->tcd_shutting_down = 0;
1094         }
1095
1096         return 0;
1097 }
1098
1099 static void trace_cleanup_on_cpu(void *info)
1100 {
1101         struct trace_cpu_data *tcd;
1102         struct trace_page *tage;
1103         struct trace_page *tmp;
1104         int i;
1105
1106         tcd_for_each_type_lock(tcd, i) {
1107                 tcd->tcd_shutting_down = 1;
1108
1109                 list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1110                         __LASSERT_TAGE_INVARIANT(tage);
1111
1112                         list_del(&tage->linkage);
1113                         tage_free(tage);
1114                 }
1115                 tcd->tcd_cur_pages = 0;
1116         }
1117 }
1118
1119 static void trace_cleanup(void)
1120 {
1121         struct page_collection pc;
1122
1123         CFS_INIT_LIST_HEAD(&pc.pc_pages);
1124         spin_lock_init(&pc.pc_lock);
1125
1126         trace_call_on_all_cpus(trace_cleanup_on_cpu, &pc);
1127
1128         tracefile_fini_arch();
1129 }
1130
1131 void tracefile_exit(void)
1132 {
1133         trace_stop_thread();
1134         trace_cleanup();
1135 }