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 [sun.com URL with a
  20  * copy of GPLv2].
  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                                 printk(KERN_WARNING
 154                                        "failure to allocate a tage (%ld)\n",
 155                                        tcd->tcd_cur_pages);
 156                                 return NULL;
 157                         }
 158                 }
 159
 160                 tage->used = 0;
 161                 tage->cpu = smp_processor_id();
 162                 tage->type = tcd->tcd_type;
 163                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 164                 tcd->tcd_cur_pages++;
 165
 166                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 167                         struct tracefiled_ctl *tctl = &trace_tctl;
 168                         /*
 169                          * wake up tracefiled to process some pages.
 170                          */
 171                         cfs_waitq_signal(&tctl->tctl_waitq);
 172                 }
 173                 return tage;
 174         }
 175         return NULL;
 176 }
 177
 178 static void tcd_shrink(struct trace_cpu_data *tcd)
 179 {
 180         int pgcount = tcd->tcd_cur_pages / 10;
 181         struct page_collection pc;
 182         struct trace_page *tage;
 183         struct trace_page *tmp;
 184
 185         /*
 186          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 187          * from here: this will lead to infinite recursion.
 188          */
 189
 190         printk(KERN_WARNING "debug daemon buffer overflowed; discarding"
 191                " 10%% of pages (%d of %ld)\n", pgcount + 1, tcd->tcd_cur_pages);
 192
 193         CFS_INIT_LIST_HEAD(&pc.pc_pages);
 194         spin_lock_init(&pc.pc_lock);
 195
 196         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 197                 if (pgcount-- == 0)
 198                         break;
 199
 200                 list_move_tail(&tage->linkage, &pc.pc_pages);
 201                 tcd->tcd_cur_pages--;
 202         }
 203         put_pages_on_tcd_daemon_list(&pc, tcd);
 204 }
 205
 206 /* return a page that has 'len' bytes left at the end */
 207 static struct trace_page *trace_get_tage(struct trace_cpu_data *tcd,
 208                                          unsigned long len)
 209 {
 210         struct trace_page *tage;
 211
 212         /*
 213          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 214          * from here: this will lead to infinite recursion.
 215          */
 216
 217         if (len > CFS_PAGE_SIZE) {
 218                 printk(KERN_ERR
 219                        "cowardly refusing to write %lu bytes in a page\n", len);
 220                 return NULL;
 221         }
 222
 223         tage = trace_get_tage_try(tcd, len);
 224         if (tage != NULL)
 225                 return tage;
 226         if (thread_running)
 227                 tcd_shrink(tcd);
 228         if (tcd->tcd_cur_pages > 0) {
 229                 tage = tage_from_list(tcd->tcd_pages.next);
 230                 tage->used = 0;
 231                 tage_to_tail(tage, &tcd->tcd_pages);
 232         }
 233         return tage;
 234 }
 235
 236 int libcfs_debug_vmsg2(cfs_debug_limit_state_t *cdls, int subsys, int mask,
 237                        const char *file, const char *fn, const int line,
 238                        const char *format1, va_list args,
 239                        const char *format2, ...)
 240 {
 241         struct trace_cpu_data   *tcd = NULL;
 242         struct ptldebug_header   header;
 243         struct trace_page       *tage;
 244         /* string_buf is used only if tcd != NULL, and is always set then */
 245         char                    *string_buf = NULL;
 246         char                    *debug_buf;
 247         int                      known_size;
 248         int                      needed = 85; /* average message length */
 249         int                      max_nob;
 250         va_list                  ap;
 251         int                      depth;
 252         int                      i;
 253         int                      remain;
 254
 255         if (strchr(file, '/'))
 256                 file = strrchr(file, '/') + 1;
 257
 258
 259         set_ptldebug_header(&header, subsys, mask, line, CDEBUG_STACK());
 260
 261         tcd = trace_get_tcd();
 262         if (tcd == NULL)                /* arch may not log in IRQ context */
 263                 goto console;
 264
 265         if (tcd->tcd_shutting_down) {
 266                 trace_put_tcd(tcd);
 267                 tcd = NULL;
 268                 goto console;
 269         }
 270
 271         depth = __current_nesting_level();
 272         known_size = strlen(file) + 1 + depth;
 273         if (fn)
 274                 known_size += strlen(fn) + 1;
 275
 276         if (libcfs_debug_binary)
 277                 known_size += sizeof(header);
 278
 279         /*/
 280          * '2' used because vsnprintf return real size required for output
 281          * _without_ terminating NULL.
 282          * if needed is to small for this format.
 283          */
 284         for (i=0;i<2;i++) {
 285                 tage = trace_get_tage(tcd, needed + known_size + 1);
 286                 if (tage == NULL) {
 287                         if (needed + known_size > CFS_PAGE_SIZE)
 288                                 mask |= D_ERROR;
 289
 290                         trace_put_tcd(tcd);
 291                         tcd = NULL;
 292                         goto console;
 293                 }
 294
 295                 string_buf = (char *)cfs_page_address(tage->page)+tage->used+known_size;
 296
 297                 max_nob = CFS_PAGE_SIZE - tage->used - known_size;
 298                 if (max_nob <= 0) {
 299                         printk(KERN_EMERG "negative max_nob: %i\n", max_nob);
 300                         mask |= D_ERROR;
 301                         trace_put_tcd(tcd);
 302                         tcd = NULL;
 303                         goto console;
 304                 }
 305
 306                 needed = 0;
 307                 if (format1) {
 308                         va_copy(ap, args);
 309                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 310                         va_end(ap);
 311                 }
 312
 313
 314                 if (format2) {
 315                         remain = max_nob - needed;
 316                         if (remain < 0)
 317                                 remain = 0;
 318
 319                         va_start(ap, format2);
 320                         needed += vsnprintf(string_buf+needed, remain, format2, ap);
 321                         va_end(ap);
 322                 }
 323
 324                 if (needed < max_nob) /* well. printing ok.. */
 325                         break;
 326         }
 327
 328         if (*(string_buf+needed-1) != '\n')
 329                 printk(KERN_INFO "format at %s:%d:%s doesn't end in newline\n",
 330                        file, line, fn);
 331
 332         header.ph_len = known_size + needed;
 333         debug_buf = (char *)cfs_page_address(tage->page) + tage->used;
 334
 335         if (libcfs_debug_binary) {
 336                 memcpy(debug_buf, &header, sizeof(header));
 337                 tage->used += sizeof(header);
 338                 debug_buf += sizeof(header);
 339         }
 340
 341         /* indent message according to the nesting level */
 342         while (depth-- > 0) {
 343                 *(debug_buf++) = '.';
 344                 ++ tage->used;
 345         }
 346
 347         strcpy(debug_buf, file);
 348         tage->used += strlen(file) + 1;
 349         debug_buf += strlen(file) + 1;
 350
 351         if (fn) {
 352                 strcpy(debug_buf, fn);
 353                 tage->used += strlen(fn) + 1;
 354                 debug_buf += strlen(fn) + 1;
 355         }
 356
 357         __LASSERT(debug_buf == string_buf);
 358
 359         tage->used += needed;
 360         __LASSERT (tage->used <= CFS_PAGE_SIZE);
 361
 362 console:
 363         if ((mask & libcfs_printk) == 0) {
 364                 /* no console output requested */
 365                 if (tcd != NULL)
 366                         trace_put_tcd(tcd);
 367                 return 1;
 368         }
 369
 370         if (cdls != NULL) {
 371                 if (libcfs_console_ratelimit &&
 372                     cdls->cdls_next != 0 &&     /* not first time ever */
 373                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 374                         /* skipping a console message */
 375                         cdls->cdls_count++;
 376                         if (tcd != NULL)
 377                                 trace_put_tcd(tcd);
 378                         return 1;
 379                 }
 380
 381                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 382                                                        libcfs_console_max_delay
 383                                                        + cfs_time_seconds(10))) {
 384                         /* last timeout was a long time ago */
 385                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 386                 } else {
 387                         cdls->cdls_delay *= libcfs_console_backoff;
 388
 389                         if (cdls->cdls_delay < libcfs_console_min_delay)
 390                                 cdls->cdls_delay = libcfs_console_min_delay;
 391                         else if (cdls->cdls_delay > libcfs_console_max_delay)
 392                                 cdls->cdls_delay = libcfs_console_max_delay;
 393                 }
 394
 395                 /* ensure cdls_next is never zero after it's been seen */
 396                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 397         }
 398
 399         if (tcd != NULL) {
 400                 print_to_console(&header, mask, string_buf, needed, file, fn);
 401                 trace_put_tcd(tcd);
 402         } else {
 403                 string_buf = trace_get_console_buffer();
 404
 405                 needed = 0;
 406                 if (format1 != NULL) {
 407                         va_copy(ap, args);
 408                         needed = vsnprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE, format1, ap);
 409                         va_end(ap);
 410                 }
 411                 if (format2 != NULL) {
 412                         remain = TRACE_CONSOLE_BUFFER_SIZE - needed;
 413                         if (remain > 0) {
 414                                 va_start(ap, format2);
 415                                 needed += vsnprintf(string_buf+needed, remain, format2, ap);
 416                                 va_end(ap);
 417                         }
 418                 }
 419                 print_to_console(&header, mask,
 420                                  string_buf, needed, file, fn);
 421
 422                 trace_put_console_buffer(string_buf);
 423         }
 424
 425         if (cdls != NULL && cdls->cdls_count != 0) {
 426                 string_buf = trace_get_console_buffer();
 427
 428                 needed = snprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE,
 429                          "Skipped %d previous similar message%s\n",
 430                          cdls->cdls_count, (cdls->cdls_count > 1) ? "s" : "");
 431
 432                 print_to_console(&header, mask,
 433                                  string_buf, needed, file, fn);
 434
 435                 trace_put_console_buffer(string_buf);
 436                 cdls->cdls_count = 0;
 437         }
 438
 439         return 0;
 440 }
 441 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 442
 443 void
 444 libcfs_assertion_failed(const char *expr, const char *file,
 445                         const char *func, const int line)
 446 {
 447         libcfs_debug_msg(NULL, 0, D_EMERG, file, func, line,
 448                          "ASSERTION(%s) failed\n", expr);
 449         LBUG();
 450 }
 451 EXPORT_SYMBOL(libcfs_assertion_failed);
 452
 453 void
 454 trace_assertion_failed(const char *str,
 455                        const char *fn, const char *file, int line)
 456 {
 457         struct ptldebug_header hdr;
 458
 459         libcfs_panic_in_progress = 1;
 460         libcfs_catastrophe = 1;
 461         mb();
 462
 463         set_ptldebug_header(&hdr, DEBUG_SUBSYSTEM, D_EMERG, line,
 464                             CDEBUG_STACK());
 465
 466         print_to_console(&hdr, D_EMERG, str, strlen(str), file, fn);
 467
 468         LIBCFS_PANIC("Lustre debug assertion failure\n");
 469
 470         /* not reached */
 471 }
 472
 473 static void
 474 panic_collect_pages(struct page_collection *pc)
 475 {
 476         /* Do the collect_pages job on a single CPU: assumes that all other
 477          * CPUs have been stopped during a panic.  If this isn't true for some
 478          * arch, this will have to be implemented separately in each arch.  */
 479         int                    i;
 480         int                    j;
 481         struct trace_cpu_data *tcd;
 482
 483         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 484
 485         tcd_for_each(tcd, i, j) {
 486                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 487                 tcd->tcd_cur_pages = 0;
 488
 489                 if (pc->pc_want_daemon_pages) {
 490                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 491                         tcd->tcd_cur_daemon_pages = 0;
 492                 }
 493         }
 494 }
 495
 496 static void collect_pages_on_cpu(void *info)
 497 {
 498         struct trace_cpu_data *tcd;
 499         struct page_collection *pc = info;
 500         int i;
 501
 502         spin_lock(&pc->pc_lock);
 503         tcd_for_each_type_lock(tcd, i) {
 504                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 505                 tcd->tcd_cur_pages = 0;
 506                 if (pc->pc_want_daemon_pages) {
 507                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 508                         tcd->tcd_cur_daemon_pages = 0;
 509                 }
 510         }
 511         spin_unlock(&pc->pc_lock);
 512 }
 513
 514 static void collect_pages(struct page_collection *pc)
 515 {
 516         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 517
 518         if (libcfs_panic_in_progress)
 519                 panic_collect_pages(pc);
 520         else
 521                 trace_call_on_all_cpus(collect_pages_on_cpu, pc);
 522 }
 523
 524 static void put_pages_back_on_cpu(void *info)
 525 {
 526         struct page_collection *pc = info;
 527         struct trace_cpu_data *tcd;
 528         struct list_head *cur_head;
 529         struct trace_page *tage;
 530         struct trace_page *tmp;
 531         int i;
 532
 533         spin_lock(&pc->pc_lock);
 534         tcd_for_each_type_lock(tcd, i) {
 535                 cur_head = tcd->tcd_pages.next;
 536
 537                 list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 538
 539                         __LASSERT_TAGE_INVARIANT(tage);
 540
 541                         if (tage->cpu != smp_processor_id() || tage->type != i)
 542                                 continue;
 543
 544                         tage_to_tail(tage, cur_head);
 545                         tcd->tcd_cur_pages++;
 546                 }
 547         }
 548         spin_unlock(&pc->pc_lock);
 549 }
 550
 551 static void put_pages_back(struct page_collection *pc)
 552 {
 553         if (!libcfs_panic_in_progress)
 554                 trace_call_on_all_cpus(put_pages_back_on_cpu, pc);
 555 }
 556
 557 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 558  * we have a good amount of data at all times for dumping during an LBUG, even
 559  * if we have been steadily writing (and otherwise discarding) pages via the
 560  * debug daemon. */
 561 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 562                                          struct trace_cpu_data *tcd)
 563 {
 564         struct trace_page *tage;
 565         struct trace_page *tmp;
 566
 567         spin_lock(&pc->pc_lock);
 568         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 569
 570                 __LASSERT_TAGE_INVARIANT(tage);
 571
 572                 if (tage->cpu != smp_processor_id() ||
 573                     tage->type != tcd->tcd_type)
 574                         continue;
 575
 576                 tage_to_tail(tage, &tcd->tcd_daemon_pages);
 577                 tcd->tcd_cur_daemon_pages++;
 578
 579                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 580                         struct trace_page *victim;
 581
 582                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 583                         victim = tage_from_list(tcd->tcd_daemon_pages.next);
 584
 585                         __LASSERT_TAGE_INVARIANT(victim);
 586
 587                         list_del(&victim->linkage);
 588                         tage_free(victim);
 589                         tcd->tcd_cur_daemon_pages--;
 590                 }
 591         }
 592         spin_unlock(&pc->pc_lock);
 593 }
 594
 595 static void put_pages_on_daemon_list_on_cpu(void *info)
 596 {
 597         struct trace_cpu_data *tcd;
 598         int i;
 599
 600         tcd_for_each_type_lock(tcd, i)
 601                 put_pages_on_tcd_daemon_list(info, tcd);
 602 }
 603
 604 static void put_pages_on_daemon_list(struct page_collection *pc)
 605 {
 606         trace_call_on_all_cpus(put_pages_on_daemon_list_on_cpu, pc);
 607 }
 608
 609 void trace_debug_print(void)
 610 {
 611         struct page_collection pc;
 612         struct trace_page *tage;
 613         struct trace_page *tmp;
 614
 615         spin_lock_init(&pc.pc_lock);
 616
 617         pc.pc_want_daemon_pages = 1;
 618         collect_pages(&pc);
 619         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 620                 char *p, *file, *fn;
 621                 cfs_page_t *page;
 622
 623                 __LASSERT_TAGE_INVARIANT(tage);
 624
 625                 page = tage->page;
 626                 p = cfs_page_address(page);
 627                 while (p < ((char *)cfs_page_address(page) + tage->used)) {
 628                         struct ptldebug_header *hdr;
 629                         int len;
 630                         hdr = (void *)p;
 631                         p += sizeof(*hdr);
 632                         file = p;
 633                         p += strlen(file) + 1;
 634                         fn = p;
 635                         p += strlen(fn) + 1;
 636                         len = hdr->ph_len - (p - (char *)hdr);
 637
 638                         print_to_console(hdr, D_EMERG, p, len, file, fn);
 639
 640                         p += len;
 641                 }
 642
 643                 list_del(&tage->linkage);
 644                 tage_free(tage);
 645         }
 646 }
 647
 648 int tracefile_dump_all_pages(char *filename)
 649 {
 650         struct page_collection pc;
 651         cfs_file_t *filp;
 652         struct trace_page *tage;
 653         struct trace_page *tmp;
 654         int rc;
 655
 656         CFS_DECL_MMSPACE;
 657
 658         tracefile_write_lock();
 659
 660         filp = cfs_filp_open(filename,
 661                              O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600, &rc);
 662         if (!filp) {
 663                 if (rc != -EEXIST)
 664                         printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 665                                filename, rc);
 666                 goto out;
 667         }
 668
 669         spin_lock_init(&pc.pc_lock);
 670         pc.pc_want_daemon_pages = 1;
 671         collect_pages(&pc);
 672         if (list_empty(&pc.pc_pages)) {
 673                 rc = 0;
 674                 goto close;
 675         }
 676
 677         /* ok, for now, just write the pages.  in the future we'll be building
 678          * iobufs with the pages and calling generic_direct_IO */
 679         CFS_MMSPACE_OPEN;
 680         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 681
 682                 __LASSERT_TAGE_INVARIANT(tage);
 683
 684                 rc = cfs_filp_write(filp, cfs_page_address(tage->page),
 685                                     tage->used, cfs_filp_poff(filp));
 686                 if (rc != (int)tage->used) {
 687                         printk(KERN_WARNING "wanted to write %u but wrote "
 688                                "%d\n", tage->used, rc);
 689                         put_pages_back(&pc);
 690                         __LASSERT(list_empty(&pc.pc_pages));
 691                         break;
 692                 }
 693                 list_del(&tage->linkage);
 694                 tage_free(tage);
 695         }
 696         CFS_MMSPACE_CLOSE;
 697         rc = cfs_filp_fsync(filp);
 698         if (rc)
 699                 printk(KERN_ERR "sync returns %d\n", rc);
 700  close:
 701         cfs_filp_close(filp);
 702  out:
 703         tracefile_write_unlock();
 704         return rc;
 705 }
 706
 707 void trace_flush_pages(void)
 708 {
 709         struct page_collection pc;
 710         struct trace_page *tage;
 711         struct trace_page *tmp;
 712
 713         spin_lock_init(&pc.pc_lock);
 714
 715         pc.pc_want_daemon_pages = 1;
 716         collect_pages(&pc);
 717         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 718
 719                 __LASSERT_TAGE_INVARIANT(tage);
 720
 721                 list_del(&tage->linkage);
 722                 tage_free(tage);
 723         }
 724 }
 725
 726 int trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 727                         const char *usr_buffer, int usr_buffer_nob)
 728 {
 729         int    nob;
 730
 731         if (usr_buffer_nob > knl_buffer_nob)
 732                 return -EOVERFLOW;
 733
 734         if (copy_from_user((void *)knl_buffer,
 735                            (void *)usr_buffer, usr_buffer_nob))
 736                 return -EFAULT;
 737
 738         nob = strnlen(knl_buffer, usr_buffer_nob);
 739         while (nob-- >= 0)                      /* strip trailing whitespace */
 740                 if (!isspace(knl_buffer[nob]))
 741                         break;
 742
 743         if (nob < 0)                            /* empty string */
 744                 return -EINVAL;
 745
 746         if (nob == knl_buffer_nob)              /* no space to terminate */
 747                 return -EOVERFLOW;
 748
 749         knl_buffer[nob + 1] = 0;                /* terminate */
 750         return 0;
 751 }
 752
 753 int trace_copyout_string(char *usr_buffer, int usr_buffer_nob,
 754                          const char *knl_buffer, char *append)
 755 {
 756         /* NB if 'append' != NULL, it's a single character to append to the
 757          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 758          * terminating zero byte) for sysctl entries */
 759         int   nob = strlen(knl_buffer);
 760
 761         if (nob > usr_buffer_nob)
 762                 nob = usr_buffer_nob;
 763
 764         if (copy_to_user(usr_buffer, knl_buffer, nob))
 765                 return -EFAULT;
 766
 767         if (append != NULL && nob < usr_buffer_nob) {
 768                 if (copy_to_user(usr_buffer + nob, append, 1))
 769                         return -EFAULT;
 770
 771                 nob++;
 772         }
 773
 774         return nob;
 775 }
 776
 777 int trace_allocate_string_buffer(char **str, int nob)
 778 {
 779         if (nob > 2 * CFS_PAGE_SIZE)            /* string must be "sensible" */
 780                 return -EINVAL;
 781
 782         *str = cfs_alloc(nob, CFS_ALLOC_STD | CFS_ALLOC_ZERO);
 783         if (*str == NULL)
 784                 return -ENOMEM;
 785
 786         return 0;
 787 }
 788
 789 void trace_free_string_buffer(char *str, int nob)
 790 {
 791         cfs_free(str);
 792 }
 793
 794 int trace_dump_debug_buffer_usrstr(void *usr_str, int usr_str_nob)
 795 {
 796         char         *str;
 797         int           rc;
 798
 799         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 800         if (rc != 0)
 801                 return rc;
 802
 803         rc = trace_copyin_string(str, usr_str_nob + 1,
 804                                  usr_str, usr_str_nob);
 805         if (rc != 0)
 806                 goto out;
 807
 808 #if !defined(__WINNT__)
 809         if (str[0] != '/') {
 810                 rc = -EINVAL;
 811                 goto out;
 812         }
 813 #endif
 814         rc = tracefile_dump_all_pages(str);
 815 out:
 816         trace_free_string_buffer(str, usr_str_nob + 1);
 817         return rc;
 818 }
 819
 820 int trace_daemon_command(char *str)
 821 {
 822         int       rc = 0;
 823
 824         tracefile_write_lock();
 825
 826         if (strcmp(str, "stop") == 0) {
 827                 trace_stop_thread();
 828                 memset(tracefile, 0, sizeof(tracefile));
 829
 830         } else if (strncmp(str, "size=", 5) == 0) {
 831                 tracefile_size = simple_strtoul(str + 5, NULL, 0);
 832                 if (tracefile_size < 10 || tracefile_size > 20480)
 833                         tracefile_size = TRACEFILE_SIZE;
 834                 else
 835                         tracefile_size <<= 20;
 836
 837         } else if (strlen(str) >= sizeof(tracefile)) {
 838                 rc = -ENAMETOOLONG;
 839 #ifndef __WINNT__
 840         } else if (str[0] != '/') {
 841                 rc = -EINVAL;
 842 #endif
 843         } else {
 844                 strcpy(tracefile, str);
 845
 846                 printk(KERN_INFO "Lustre: debug daemon will attempt to start writing "
 847                        "to %s (%lukB max)\n", tracefile,
 848                        (long)(tracefile_size >> 10));
 849
 850                 trace_start_thread();
 851         }
 852
 853         tracefile_write_unlock();
 854         return rc;
 855 }
 856
 857 int trace_daemon_command_usrstr(void *usr_str, int usr_str_nob)
 858 {
 859         char *str;
 860         int   rc;
 861
 862         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 863         if (rc != 0)
 864                 return rc;
 865
 866         rc = trace_copyin_string(str, usr_str_nob + 1,
 867                                  usr_str, usr_str_nob);
 868         if (rc == 0)
 869                 rc = trace_daemon_command(str);
 870
 871         trace_free_string_buffer(str, usr_str_nob + 1);
 872         return rc;
 873 }
 874
 875 int trace_set_debug_mb(int mb)
 876 {
 877         int i;
 878         int j;
 879         int pages;
 880         int limit = trace_max_debug_mb();
 881         struct trace_cpu_data *tcd;
 882
 883         if (mb < num_possible_cpus())
 884                 return -EINVAL;
 885
 886         if (mb > limit) {
 887                 printk(KERN_ERR "Lustre: Refusing to set debug buffer size to "
 888                        "%dMB - limit is %d\n", mb, limit);
 889                 return -EINVAL;
 890         }
 891
 892         mb /= num_possible_cpus();
 893         pages = mb << (20 - CFS_PAGE_SHIFT);
 894
 895         tracefile_write_lock();
 896
 897         tcd_for_each(tcd, i, j)
 898                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 899
 900         tracefile_write_unlock();
 901
 902         return 0;
 903 }
 904
 905 int trace_set_debug_mb_usrstr(void *usr_str, int usr_str_nob)
 906 {
 907         char     str[32];
 908         int      rc;
 909
 910         rc = trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 911         if (rc < 0)
 912                 return rc;
 913
 914         return trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 915 }
 916
 917 int trace_get_debug_mb(void)
 918 {
 919         int i;
 920         int j;
 921         struct trace_cpu_data *tcd;
 922         int total_pages = 0;
 923
 924         tracefile_read_lock();
 925
 926         tcd_for_each(tcd, i, j)
 927                 total_pages += tcd->tcd_max_pages;
 928
 929         tracefile_read_unlock();
 930
 931         return (total_pages >> (20 - CFS_PAGE_SHIFT)) + 1;
 932 }
 933
 934 static int tracefiled(void *arg)
 935 {
 936         struct page_collection pc;
 937         struct tracefiled_ctl *tctl = arg;
 938         struct trace_page *tage;
 939         struct trace_page *tmp;
 940         struct ptldebug_header *hdr;
 941         cfs_file_t *filp;
 942         int rc;
 943
 944         CFS_DECL_MMSPACE;
 945
 946         /* we're started late enough that we pick up init's fs context */
 947         /* this is so broken in uml?  what on earth is going on? */
 948         cfs_daemonize("ktracefiled");
 949
 950         spin_lock_init(&pc.pc_lock);
 951         complete(&tctl->tctl_start);
 952
 953         while (1) {
 954                 cfs_waitlink_t __wait;
 955
 956                 cfs_waitlink_init(&__wait);
 957                 cfs_waitq_add(&tctl->tctl_waitq, &__wait);
 958                 set_current_state(TASK_INTERRUPTIBLE);
 959                 cfs_waitq_timedwait(&__wait, CFS_TASK_INTERRUPTIBLE,
 960                                     cfs_time_seconds(1));
 961                 cfs_waitq_del(&tctl->tctl_waitq, &__wait);
 962
 963                 if (atomic_read(&tctl->tctl_shutdown))
 964                         break;
 965
 966                 pc.pc_want_daemon_pages = 0;
 967                 collect_pages(&pc);
 968                 if (list_empty(&pc.pc_pages))
 969                         continue;
 970
 971                 filp = NULL;
 972                 tracefile_read_lock();
 973                 if (tracefile[0] != 0) {
 974                         filp = cfs_filp_open(tracefile,
 975                                              O_CREAT | O_RDWR | O_LARGEFILE,
 976                                              0600, &rc);
 977                         if (!(filp))
 978                                 printk(KERN_WARNING "couldn't open %s: %d\n",
 979                                        tracefile, rc);
 980                 }
 981                 tracefile_read_unlock();
 982                 if (filp == NULL) {
 983                         put_pages_on_daemon_list(&pc);
 984                         __LASSERT(list_empty(&pc.pc_pages));
 985                         continue;
 986                 }
 987
 988                 CFS_MMSPACE_OPEN;
 989
 990                 /* mark the first header, so we can sort in chunks */
 991                 tage = tage_from_list(pc.pc_pages.next);
 992                 __LASSERT_TAGE_INVARIANT(tage);
 993
 994                 hdr = cfs_page_address(tage->page);
 995                 hdr->ph_flags |= PH_FLAG_FIRST_RECORD;
 996
 997                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 998                         static loff_t f_pos;
 999
1000                         __LASSERT_TAGE_INVARIANT(tage);
1001
1002                         if (f_pos >= (off_t)tracefile_size)
1003                                 f_pos = 0;
1004                         else if (f_pos > cfs_filp_size(filp))
1005                                 f_pos = cfs_filp_size(filp);
1006
1007                         rc = cfs_filp_write(filp, cfs_page_address(tage->page),
1008                                             tage->used, &f_pos);
1009                         if (rc != (int)tage->used) {
1010                                 printk(KERN_WARNING "wanted to write %u but "
1011                                        "wrote %d\n", tage->used, rc);
1012                                 put_pages_back(&pc);
1013                                 __LASSERT(list_empty(&pc.pc_pages));
1014                         }
1015                 }
1016                 CFS_MMSPACE_CLOSE;
1017
1018                 cfs_filp_close(filp);
1019                 put_pages_on_daemon_list(&pc);
1020                 __LASSERT(list_empty(&pc.pc_pages));
1021         }
1022         complete(&tctl->tctl_stop);
1023         return 0;
1024 }
1025
1026 int trace_start_thread(void)
1027 {
1028         struct tracefiled_ctl *tctl = &trace_tctl;
1029         int rc = 0;
1030
1031         mutex_down(&trace_thread_sem);
1032         if (thread_running)
1033                 goto out;
1034
1035         init_completion(&tctl->tctl_start);
1036         init_completion(&tctl->tctl_stop);
1037         cfs_waitq_init(&tctl->tctl_waitq);
1038         atomic_set(&tctl->tctl_shutdown, 0);
1039
1040         if (cfs_kernel_thread(tracefiled, tctl, 0) < 0) {
1041                 rc = -ECHILD;
1042                 goto out;
1043         }
1044
1045         wait_for_completion(&tctl->tctl_start);
1046         thread_running = 1;
1047 out:
1048         mutex_up(&trace_thread_sem);
1049         return rc;
1050 }
1051
1052 void trace_stop_thread(void)
1053 {
1054         struct tracefiled_ctl *tctl = &trace_tctl;
1055
1056         mutex_down(&trace_thread_sem);
1057         if (thread_running) {
1058                 printk(KERN_INFO "Lustre: shutting down debug daemon thread...\n");
1059                 atomic_set(&tctl->tctl_shutdown, 1);
1060                 wait_for_completion(&tctl->tctl_stop);
1061                 thread_running = 0;
1062         }
1063         mutex_up(&trace_thread_sem);
1064 }
1065
1066 int tracefile_init(int max_pages)
1067 {
1068         struct trace_cpu_data *tcd;
1069         int                    i;
1070         int                    j;
1071         int                    rc;
1072         int                    factor;
1073
1074         rc = tracefile_init_arch();
1075         if (rc != 0)
1076                 return rc;
1077
1078         tcd_for_each(tcd, i, j) {
1079                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1080                 factor = tcd->tcd_pages_factor;
1081                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
1082                 CFS_INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1083                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1084                 tcd->tcd_cur_pages = 0;
1085                 tcd->tcd_cur_stock_pages = 0;
1086                 tcd->tcd_cur_daemon_pages = 0;
1087                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1088                 LASSERT(tcd->tcd_max_pages > 0);
1089                 tcd->tcd_shutting_down = 0;
1090         }
1091
1092         return 0;
1093 }
1094
1095 static void trace_cleanup_on_cpu(void *info)
1096 {
1097         struct trace_cpu_data *tcd;
1098         struct trace_page *tage;
1099         struct trace_page *tmp;
1100         int i;
1101
1102         tcd_for_each_type_lock(tcd, i) {
1103                 tcd->tcd_shutting_down = 1;
1104
1105                 list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1106                         __LASSERT_TAGE_INVARIANT(tage);
1107
1108                         list_del(&tage->linkage);
1109                         tage_free(tage);
1110                 }
1111                 tcd->tcd_cur_pages = 0;
1112         }
1113 }
1114
1115 static void trace_cleanup(void)
1116 {
1117         struct page_collection pc;
1118
1119         CFS_INIT_LIST_HEAD(&pc.pc_pages);
1120         spin_lock_init(&pc.pc_lock);
1121
1122         trace_call_on_all_cpus(trace_cleanup_on_cpu, &pc);
1123
1124         tracefile_fini_arch();
1125 }
1126
1127 void tracefile_exit(void)
1128 {
1129         trace_stop_thread();
1130         trace_cleanup();
1131 }