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  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  * 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
 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                 if (format2) {
 317                         remain = max_nob - needed;
 318                         if (remain < 0)
 319                                 remain = 0;
 320
 321                         va_start(ap, format2);
 322                         needed += vsnprintf(string_buf + needed, remain,
 323                                             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         lbug_with_loc(file, func, line);
 453 }
 454 EXPORT_SYMBOL(libcfs_assertion_failed);
 455
 456 void
 457 trace_assertion_failed(const char *str,
 458                        const char *fn, const char *file, int line)
 459 {
 460         struct ptldebug_header hdr;
 461
 462         libcfs_panic_in_progress = 1;
 463         libcfs_catastrophe = 1;
 464         mb();
 465
 466         set_ptldebug_header(&hdr, DEBUG_SUBSYSTEM, D_EMERG, line,
 467                             CDEBUG_STACK());
 468
 469         print_to_console(&hdr, D_EMERG, str, strlen(str), file, fn);
 470
 471         LIBCFS_PANIC("Lustre debug assertion failure\n");
 472
 473         /* not reached */
 474 }
 475
 476 static void
 477 panic_collect_pages(struct page_collection *pc)
 478 {
 479         /* Do the collect_pages job on a single CPU: assumes that all other
 480          * CPUs have been stopped during a panic.  If this isn't true for some
 481          * arch, this will have to be implemented separately in each arch.  */
 482         int                    i;
 483         int                    j;
 484         struct trace_cpu_data *tcd;
 485
 486         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 487
 488         tcd_for_each(tcd, i, j) {
 489                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 490                 tcd->tcd_cur_pages = 0;
 491
 492                 if (pc->pc_want_daemon_pages) {
 493                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 494                         tcd->tcd_cur_daemon_pages = 0;
 495                 }
 496         }
 497 }
 498
 499 static void collect_pages_on_all_cpus(struct page_collection *pc)
 500 {
 501         struct trace_cpu_data *tcd;
 502         int i, cpu;
 503
 504         spin_lock(&pc->pc_lock);
 505         for_each_possible_cpu(cpu) {
 506                 tcd_for_each_type_lock(tcd, i, cpu) {
 507                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 508                         tcd->tcd_cur_pages = 0;
 509                         if (pc->pc_want_daemon_pages) {
 510                                 list_splice_init(&tcd->tcd_daemon_pages,
 511                                                  &pc->pc_pages);
 512                                 tcd->tcd_cur_daemon_pages = 0;
 513                         }
 514                 }
 515         }
 516         spin_unlock(&pc->pc_lock);
 517 }
 518
 519 static void collect_pages(struct page_collection *pc)
 520 {
 521         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 522
 523         if (libcfs_panic_in_progress)
 524                 panic_collect_pages(pc);
 525         else
 526                 collect_pages_on_all_cpus(pc);
 527 }
 528
 529 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 530 {
 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, cpu;
 536
 537         spin_lock(&pc->pc_lock);
 538         for_each_possible_cpu(cpu) {
 539                 tcd_for_each_type_lock(tcd, i, cpu) {
 540                         cur_head = tcd->tcd_pages.next;
 541
 542                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 543                                                  linkage) {
 544
 545                                 __LASSERT_TAGE_INVARIANT(tage);
 546
 547                                 if (tage->cpu != cpu || tage->type != i)
 548                                         continue;
 549
 550                                 tage_to_tail(tage, cur_head);
 551                                 tcd->tcd_cur_pages++;
 552                         }
 553                 }
 554         }
 555         spin_unlock(&pc->pc_lock);
 556 }
 557
 558 static void put_pages_back(struct page_collection *pc)
 559 {
 560         if (!libcfs_panic_in_progress)
 561                 put_pages_back_on_all_cpus(pc);
 562 }
 563
 564 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 565  * we have a good amount of data at all times for dumping during an LBUG, even
 566  * if we have been steadily writing (and otherwise discarding) pages via the
 567  * debug daemon. */
 568 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 569                                          struct trace_cpu_data *tcd)
 570 {
 571         struct trace_page *tage;
 572         struct trace_page *tmp;
 573
 574         spin_lock(&pc->pc_lock);
 575         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 576
 577                 __LASSERT_TAGE_INVARIANT(tage);
 578
 579                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 580                         continue;
 581
 582                 tage_to_tail(tage, &tcd->tcd_daemon_pages);
 583                 tcd->tcd_cur_daemon_pages++;
 584
 585                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 586                         struct trace_page *victim;
 587
 588                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 589                         victim = tage_from_list(tcd->tcd_daemon_pages.next);
 590
 591                         __LASSERT_TAGE_INVARIANT(victim);
 592
 593                         list_del(&victim->linkage);
 594                         tage_free(victim);
 595                         tcd->tcd_cur_daemon_pages--;
 596                 }
 597         }
 598         spin_unlock(&pc->pc_lock);
 599 }
 600
 601 static void put_pages_on_daemon_list(struct page_collection *pc)
 602 {
 603         struct trace_cpu_data *tcd;
 604         int i, cpu;
 605
 606         for_each_possible_cpu(cpu) {
 607                 tcd_for_each_type_lock(tcd, i, cpu)
 608                         put_pages_on_tcd_daemon_list(pc, tcd);
 609         }
 610 }
 611
 612 void trace_debug_print(void)
 613 {
 614         struct page_collection pc;
 615         struct trace_page *tage;
 616         struct trace_page *tmp;
 617
 618         spin_lock_init(&pc.pc_lock);
 619
 620         pc.pc_want_daemon_pages = 1;
 621         collect_pages(&pc);
 622         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 623                 char *p, *file, *fn;
 624                 cfs_page_t *page;
 625
 626                 __LASSERT_TAGE_INVARIANT(tage);
 627
 628                 page = tage->page;
 629                 p = cfs_page_address(page);
 630                 while (p < ((char *)cfs_page_address(page) + tage->used)) {
 631                         struct ptldebug_header *hdr;
 632                         int len;
 633                         hdr = (void *)p;
 634                         p += sizeof(*hdr);
 635                         file = p;
 636                         p += strlen(file) + 1;
 637                         fn = p;
 638                         p += strlen(fn) + 1;
 639                         len = hdr->ph_len - (p - (char *)hdr);
 640
 641                         print_to_console(hdr, D_EMERG, p, len, file, fn);
 642
 643                         p += len;
 644                 }
 645
 646                 list_del(&tage->linkage);
 647                 tage_free(tage);
 648         }
 649 }
 650
 651 int tracefile_dump_all_pages(char *filename)
 652 {
 653         struct page_collection pc;
 654         cfs_file_t *filp;
 655         struct trace_page *tage;
 656         struct trace_page *tmp;
 657         int rc;
 658
 659         CFS_DECL_MMSPACE;
 660
 661         tracefile_write_lock();
 662
 663         filp = cfs_filp_open(filename,
 664                              O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600, &rc);
 665         if (!filp) {
 666                 if (rc != -EEXIST)
 667                         printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 668                                filename, rc);
 669                 goto out;
 670         }
 671
 672         spin_lock_init(&pc.pc_lock);
 673         pc.pc_want_daemon_pages = 1;
 674         collect_pages(&pc);
 675         if (list_empty(&pc.pc_pages)) {
 676                 rc = 0;
 677                 goto close;
 678         }
 679
 680         /* ok, for now, just write the pages.  in the future we'll be building
 681          * iobufs with the pages and calling generic_direct_IO */
 682         CFS_MMSPACE_OPEN;
 683         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 684
 685                 __LASSERT_TAGE_INVARIANT(tage);
 686
 687                 rc = cfs_filp_write(filp, cfs_page_address(tage->page),
 688                                     tage->used, cfs_filp_poff(filp));
 689                 if (rc != (int)tage->used) {
 690                         printk(KERN_WARNING "wanted to write %u but wrote "
 691                                "%d\n", tage->used, rc);
 692                         put_pages_back(&pc);
 693                         __LASSERT(list_empty(&pc.pc_pages));
 694                         break;
 695                 }
 696                 list_del(&tage->linkage);
 697                 tage_free(tage);
 698         }
 699         CFS_MMSPACE_CLOSE;
 700         rc = cfs_filp_fsync(filp);
 701         if (rc)
 702                 printk(KERN_ERR "sync returns %d\n", rc);
 703  close:
 704         cfs_filp_close(filp);
 705  out:
 706         tracefile_write_unlock();
 707         return rc;
 708 }
 709
 710 void trace_flush_pages(void)
 711 {
 712         struct page_collection pc;
 713         struct trace_page *tage;
 714         struct trace_page *tmp;
 715
 716         spin_lock_init(&pc.pc_lock);
 717
 718         pc.pc_want_daemon_pages = 1;
 719         collect_pages(&pc);
 720         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 721
 722                 __LASSERT_TAGE_INVARIANT(tage);
 723
 724                 list_del(&tage->linkage);
 725                 tage_free(tage);
 726         }
 727 }
 728
 729 int trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 730                         const char *usr_buffer, int usr_buffer_nob)
 731 {
 732         int    nob;
 733
 734         if (usr_buffer_nob > knl_buffer_nob)
 735                 return -EOVERFLOW;
 736
 737         if (copy_from_user((void *)knl_buffer,
 738                            (void *)usr_buffer, usr_buffer_nob))
 739                 return -EFAULT;
 740
 741         nob = strnlen(knl_buffer, usr_buffer_nob);
 742         while (nob-- >= 0)                      /* strip trailing whitespace */
 743                 if (!isspace(knl_buffer[nob]))
 744                         break;
 745
 746         if (nob < 0)                            /* empty string */
 747                 return -EINVAL;
 748
 749         if (nob == knl_buffer_nob)              /* no space to terminate */
 750                 return -EOVERFLOW;
 751
 752         knl_buffer[nob + 1] = 0;                /* terminate */
 753         return 0;
 754 }
 755
 756 int trace_copyout_string(char *usr_buffer, int usr_buffer_nob,
 757                          const char *knl_buffer, char *append)
 758 {
 759         /* NB if 'append' != NULL, it's a single character to append to the
 760          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 761          * terminating zero byte) for sysctl entries */
 762         int   nob = strlen(knl_buffer);
 763
 764         if (nob > usr_buffer_nob)
 765                 nob = usr_buffer_nob;
 766
 767         if (copy_to_user(usr_buffer, knl_buffer, nob))
 768                 return -EFAULT;
 769
 770         if (append != NULL && nob < usr_buffer_nob) {
 771                 if (copy_to_user(usr_buffer + nob, append, 1))
 772                         return -EFAULT;
 773
 774                 nob++;
 775         }
 776
 777         return nob;
 778 }
 779
 780 int trace_allocate_string_buffer(char **str, int nob)
 781 {
 782         if (nob > 2 * CFS_PAGE_SIZE)            /* string must be "sensible" */
 783                 return -EINVAL;
 784
 785         *str = cfs_alloc(nob, CFS_ALLOC_STD | CFS_ALLOC_ZERO);
 786         if (*str == NULL)
 787                 return -ENOMEM;
 788
 789         return 0;
 790 }
 791
 792 void trace_free_string_buffer(char *str, int nob)
 793 {
 794         cfs_free(str);
 795 }
 796
 797 int trace_dump_debug_buffer_usrstr(void *usr_str, int usr_str_nob)
 798 {
 799         char         *str;
 800         int           rc;
 801
 802         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 803         if (rc != 0)
 804                 return rc;
 805
 806         rc = trace_copyin_string(str, usr_str_nob + 1,
 807                                  usr_str, usr_str_nob);
 808         if (rc != 0)
 809                 goto out;
 810
 811 #if !defined(__WINNT__)
 812         if (str[0] != '/') {
 813                 rc = -EINVAL;
 814                 goto out;
 815         }
 816 #endif
 817         rc = tracefile_dump_all_pages(str);
 818 out:
 819         trace_free_string_buffer(str, usr_str_nob + 1);
 820         return rc;
 821 }
 822
 823 int trace_daemon_command(char *str)
 824 {
 825         int       rc = 0;
 826
 827         tracefile_write_lock();
 828
 829         if (strcmp(str, "stop") == 0) {
 830                 tracefile_write_unlock();
 831                 trace_stop_thread();
 832                 tracefile_write_lock();
 833                 memset(tracefile, 0, sizeof(tracefile));
 834
 835         } else if (strncmp(str, "size=", 5) == 0) {
 836                 tracefile_size = simple_strtoul(str + 5, NULL, 0);
 837                 if (tracefile_size < 10 || tracefile_size > 20480)
 838                         tracefile_size = TRACEFILE_SIZE;
 839                 else
 840                         tracefile_size <<= 20;
 841
 842         } else if (strlen(str) >= sizeof(tracefile)) {
 843                 rc = -ENAMETOOLONG;
 844 #ifndef __WINNT__
 845         } else if (str[0] != '/') {
 846                 rc = -EINVAL;
 847 #endif
 848         } else {
 849                 strcpy(tracefile, str);
 850
 851                 printk(KERN_INFO "Lustre: debug daemon will attempt to start writing "
 852                        "to %s (%lukB max)\n", tracefile,
 853                        (long)(tracefile_size >> 10));
 854
 855                 trace_start_thread();
 856         }
 857
 858         tracefile_write_unlock();
 859         return rc;
 860 }
 861
 862 int trace_daemon_command_usrstr(void *usr_str, int usr_str_nob)
 863 {
 864         char *str;
 865         int   rc;
 866
 867         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 868         if (rc != 0)
 869                 return rc;
 870
 871         rc = trace_copyin_string(str, usr_str_nob + 1,
 872                                  usr_str, usr_str_nob);
 873         if (rc == 0)
 874                 rc = trace_daemon_command(str);
 875
 876         trace_free_string_buffer(str, usr_str_nob + 1);
 877
 878         return rc;
 879 }
 880
 881 int trace_set_debug_mb(int mb)
 882 {
 883         int i;
 884         int j;
 885         int pages;
 886         int limit = trace_max_debug_mb();
 887         struct trace_cpu_data *tcd;
 888
 889         if (mb < num_possible_cpus())
 890                 return -EINVAL;
 891
 892         if (mb > limit) {
 893                 printk(KERN_ERR "Lustre: Refusing to set debug buffer size to "
 894                        "%dMB - limit is %d\n", mb, limit);
 895                 return -EINVAL;
 896         }
 897
 898         mb /= num_possible_cpus();
 899         pages = mb << (20 - CFS_PAGE_SHIFT);
 900
 901         tracefile_write_lock();
 902
 903         tcd_for_each(tcd, i, j)
 904                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 905
 906         tracefile_write_unlock();
 907
 908         return 0;
 909 }
 910
 911 int trace_set_debug_mb_usrstr(void *usr_str, int usr_str_nob)
 912 {
 913         char     str[32];
 914         int      rc;
 915
 916         rc = trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 917         if (rc < 0)
 918                 return rc;
 919
 920         return trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 921 }
 922
 923 int trace_get_debug_mb(void)
 924 {
 925         int i;
 926         int j;
 927         struct trace_cpu_data *tcd;
 928         int total_pages = 0;
 929
 930         tracefile_read_lock();
 931
 932         tcd_for_each(tcd, i, j)
 933                 total_pages += tcd->tcd_max_pages;
 934
 935         tracefile_read_unlock();
 936
 937         return (total_pages >> (20 - CFS_PAGE_SHIFT)) + 1;
 938 }
 939
 940 static int tracefiled(void *arg)
 941 {
 942         struct page_collection pc;
 943         struct tracefiled_ctl *tctl = arg;
 944         struct trace_page *tage;
 945         struct trace_page *tmp;
 946         struct ptldebug_header *hdr;
 947         cfs_file_t *filp;
 948         int rc;
 949
 950         CFS_DECL_MMSPACE;
 951
 952         /* we're started late enough that we pick up init's fs context */
 953         /* this is so broken in uml?  what on earth is going on? */
 954         cfs_daemonize("ktracefiled");
 955
 956         spin_lock_init(&pc.pc_lock);
 957         complete(&tctl->tctl_start);
 958
 959         while (1) {
 960                 cfs_waitlink_t __wait;
 961
 962                 pc.pc_want_daemon_pages = 0;
 963                 collect_pages(&pc);
 964                 if (list_empty(&pc.pc_pages))
 965                         continue;
 966
 967                 filp = NULL;
 968                 tracefile_read_lock();
 969                 if (tracefile[0] != 0) {
 970                         filp = cfs_filp_open(tracefile,
 971                                              O_CREAT | O_RDWR | O_LARGEFILE,
 972                                              0600, &rc);
 973                         if (!(filp))
 974                                 printk(KERN_WARNING "couldn't open %s: %d\n",
 975                                        tracefile, rc);
 976                 }
 977                 tracefile_read_unlock();
 978                 if (filp == NULL) {
 979                         put_pages_on_daemon_list(&pc);
 980                         __LASSERT(list_empty(&pc.pc_pages));
 981                         continue;
 982                 }
 983
 984                 CFS_MMSPACE_OPEN;
 985
 986                 /* mark the first header, so we can sort in chunks */
 987                 tage = tage_from_list(pc.pc_pages.next);
 988                 __LASSERT_TAGE_INVARIANT(tage);
 989
 990                 hdr = cfs_page_address(tage->page);
 991                 hdr->ph_flags |= PH_FLAG_FIRST_RECORD;
 992
 993                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 994                         static loff_t f_pos;
 995
 996                         __LASSERT_TAGE_INVARIANT(tage);
 997
 998                         if (f_pos >= (off_t)tracefile_size)
 999                                 f_pos = 0;
1000                         else if (f_pos > cfs_filp_size(filp))
1001                                 f_pos = cfs_filp_size(filp);
1002
1003                         rc = cfs_filp_write(filp, cfs_page_address(tage->page),
1004                                             tage->used, &f_pos);
1005                         if (rc != (int)tage->used) {
1006                                 printk(KERN_WARNING "wanted to write %u but "
1007                                        "wrote %d\n", tage->used, rc);
1008                                 put_pages_back(&pc);
1009                                 __LASSERT(list_empty(&pc.pc_pages));
1010                         }
1011                 }
1012                 CFS_MMSPACE_CLOSE;
1013
1014                 cfs_filp_close(filp);
1015                 put_pages_on_daemon_list(&pc);
1016                 if (!list_empty(&pc.pc_pages)) {
1017                         int i;
1018
1019                         printk(KERN_ALERT "Lustre: trace pages aren't empty\n");
1020                         printk(KERN_ALERT "total cpus(%d): ", num_possible_cpus());
1021                         for (i = 0; i < num_possible_cpus(); i++)
1022                                 if (cpu_online(i))
1023                                         printk(KERN_ALERT "%d(on) ", i);
1024                                 else
1025                                         printk(KERN_ALERT "%d(off) ", i);
1026                         printk(KERN_ALERT "\n");
1027
1028                         i = 0;
1029                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1030                                                  linkage)
1031                                 printk(KERN_ALERT "page %d belongs to cpu %d\n",
1032                                        ++i, tage->cpu);
1033                         printk(KERN_ALERT "There are %d pages unwritten\n", i);
1034                 }
1035                 __LASSERT(list_empty(&pc.pc_pages));
1036
1037                 cfs_waitlink_init(&__wait);
1038                 cfs_waitq_add(&tctl->tctl_waitq, &__wait);
1039                 set_current_state(TASK_INTERRUPTIBLE);
1040                 cfs_waitq_timedwait(&__wait, CFS_TASK_INTERRUPTIBLE,
1041                                     cfs_time_seconds(1));
1042                 cfs_waitq_del(&tctl->tctl_waitq, &__wait);
1043
1044                 if (atomic_read(&tctl->tctl_shutdown))
1045                         break;
1046         }
1047         complete(&tctl->tctl_stop);
1048         return 0;
1049 }
1050
1051 int trace_start_thread(void)
1052 {
1053         struct tracefiled_ctl *tctl = &trace_tctl;
1054         int rc = 0;
1055
1056         mutex_down(&trace_thread_sem);
1057         if (thread_running)
1058                 goto out;
1059
1060         init_completion(&tctl->tctl_start);
1061         init_completion(&tctl->tctl_stop);
1062         cfs_waitq_init(&tctl->tctl_waitq);
1063         atomic_set(&tctl->tctl_shutdown, 0);
1064
1065         if (cfs_kernel_thread(tracefiled, tctl, 0) < 0) {
1066                 rc = -ECHILD;
1067                 goto out;
1068         }
1069
1070         wait_for_completion(&tctl->tctl_start);
1071         thread_running = 1;
1072 out:
1073         mutex_up(&trace_thread_sem);
1074         return rc;
1075 }
1076
1077 void trace_stop_thread(void)
1078 {
1079         struct tracefiled_ctl *tctl = &trace_tctl;
1080
1081         mutex_down(&trace_thread_sem);
1082         if (thread_running) {
1083                 printk(KERN_INFO "Lustre: shutting down debug daemon thread...\n");
1084                 atomic_set(&tctl->tctl_shutdown, 1);
1085                 wait_for_completion(&tctl->tctl_stop);
1086                 thread_running = 0;
1087         }
1088         mutex_up(&trace_thread_sem);
1089 }
1090
1091 int tracefile_init(int max_pages)
1092 {
1093         struct trace_cpu_data *tcd;
1094         int                    i;
1095         int                    j;
1096         int                    rc;
1097         int                    factor;
1098
1099         rc = tracefile_init_arch();
1100         if (rc != 0)
1101                 return rc;
1102
1103         tcd_for_each(tcd, i, j) {
1104                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1105                 factor = tcd->tcd_pages_factor;
1106                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
1107                 CFS_INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1108                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1109                 tcd->tcd_cur_pages = 0;
1110                 tcd->tcd_cur_stock_pages = 0;
1111                 tcd->tcd_cur_daemon_pages = 0;
1112                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1113                 LASSERT(tcd->tcd_max_pages > 0);
1114                 tcd->tcd_shutting_down = 0;
1115         }
1116
1117         return 0;
1118 }
1119
1120 static void trace_cleanup_on_all_cpus(void)
1121 {
1122         struct trace_cpu_data *tcd;
1123         struct trace_page *tage;
1124         struct trace_page *tmp;
1125         int i, cpu;
1126
1127         for_each_possible_cpu(cpu) {
1128                 tcd_for_each_type_lock(tcd, i, cpu) {
1129                         tcd->tcd_shutting_down = 1;
1130
1131                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages,
1132                                                  linkage) {
1133                                 __LASSERT_TAGE_INVARIANT(tage);
1134
1135                                 list_del(&tage->linkage);
1136                                 tage_free(tage);
1137                         }
1138
1139                         tcd->tcd_cur_pages = 0;
1140                 }
1141         }
1142 }
1143
1144 static void trace_cleanup(void)
1145 {
1146         struct page_collection pc;
1147
1148         CFS_INIT_LIST_HEAD(&pc.pc_pages);
1149         spin_lock_init(&pc.pc_lock);
1150
1151         trace_cleanup_on_all_cpus();
1152
1153         tracefile_fini_arch();
1154 }
1155
1156 void tracefile_exit(void)
1157 {
1158         trace_stop_thread();
1159         trace_cleanup();
1160 }