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         cfs_list_for_each_entry_safe_typed(tage, tmp, &tcd->tcd_pages,
 200                                            struct trace_page, linkage) {
 201                 if (pgcount-- == 0)
 202                         break;
 203
 204                 list_move_tail(&tage->linkage, &pc.pc_pages);
 205                 tcd->tcd_cur_pages--;
 206         }
 207         put_pages_on_tcd_daemon_list(&pc, tcd);
 208 }
 209
 210 /* return a page that has 'len' bytes left at the end */
 211 static struct trace_page *trace_get_tage(struct trace_cpu_data *tcd,
 212                                          unsigned long len)
 213 {
 214         struct trace_page *tage;
 215
 216         /*
 217          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 218          * from here: this will lead to infinite recursion.
 219          */
 220
 221         if (len > CFS_PAGE_SIZE) {
 222                 printk(KERN_ERR
 223                        "cowardly refusing to write %lu bytes in a page\n", len);
 224                 return NULL;
 225         }
 226
 227         tage = trace_get_tage_try(tcd, len);
 228         if (tage != NULL)
 229                 return tage;
 230         if (thread_running)
 231                 tcd_shrink(tcd);
 232         if (tcd->tcd_cur_pages > 0) {
 233                 tage = tage_from_list(tcd->tcd_pages.next);
 234                 tage->used = 0;
 235                 tage_to_tail(tage, &tcd->tcd_pages);
 236         }
 237         return tage;
 238 }
 239
 240 int libcfs_debug_vmsg2(cfs_debug_limit_state_t *cdls, int subsys, int mask,
 241                        const char *file, const char *fn, const int line,
 242                        const char *format1, va_list args,
 243                        const char *format2, ...)
 244 {
 245         struct trace_cpu_data   *tcd = NULL;
 246         struct ptldebug_header   header;
 247         struct trace_page       *tage;
 248         /* string_buf is used only if tcd != NULL, and is always set then */
 249         char                    *string_buf = NULL;
 250         char                    *debug_buf;
 251         int                      known_size;
 252         int                      needed = 85; /* average message length */
 253         int                      max_nob;
 254         va_list                  ap;
 255         int                      depth;
 256         int                      i;
 257         int                      remain;
 258
 259         if (strchr(file, '/'))
 260                 file = strrchr(file, '/') + 1;
 261
 262
 263         set_ptldebug_header(&header, subsys, mask, line, CDEBUG_STACK());
 264
 265         tcd = trace_get_tcd();
 266         if (tcd == NULL)                /* arch may not log in IRQ context */
 267                 goto console;
 268
 269         if (tcd->tcd_shutting_down) {
 270                 trace_put_tcd(tcd);
 271                 tcd = NULL;
 272                 goto console;
 273         }
 274
 275         depth = __current_nesting_level();
 276         known_size = strlen(file) + 1 + depth;
 277         if (fn)
 278                 known_size += strlen(fn) + 1;
 279
 280         if (libcfs_debug_binary)
 281                 known_size += sizeof(header);
 282
 283         /*/
 284          * '2' used because vsnprintf return real size required for output
 285          * _without_ terminating NULL.
 286          * if needed is to small for this format.
 287          */
 288         for (i = 0; i < 2; i++) {
 289                 tage = trace_get_tage(tcd, needed + known_size + 1);
 290                 if (tage == NULL) {
 291                         if (needed + known_size > CFS_PAGE_SIZE)
 292                                 mask |= D_ERROR;
 293
 294                         trace_put_tcd(tcd);
 295                         tcd = NULL;
 296                         goto console;
 297                 }
 298
 299                 string_buf = (char *)cfs_page_address(tage->page) +
 300                                         tage->used + known_size;
 301
 302                 max_nob = CFS_PAGE_SIZE - tage->used - known_size;
 303                 if (max_nob <= 0) {
 304                         printk(KERN_EMERG "negative max_nob: %i\n", max_nob);
 305                         mask |= D_ERROR;
 306                         trace_put_tcd(tcd);
 307                         tcd = NULL;
 308                         goto console;
 309                 }
 310
 311                 needed = 0;
 312                 if (format1) {
 313                         va_copy(ap, args);
 314                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 315                         va_end(ap);
 316                 }
 317
 318                 if (format2) {
 319                         remain = max_nob - needed;
 320                         if (remain < 0)
 321                                 remain = 0;
 322
 323                         va_start(ap, format2);
 324                         needed += vsnprintf(string_buf + needed, remain,
 325                                             format2, ap);
 326                         va_end(ap);
 327                 }
 328
 329                 if (needed < max_nob) /* well. printing ok.. */
 330                         break;
 331         }
 332
 333         if (*(string_buf+needed-1) != '\n')
 334                 printk(KERN_INFO "format at %s:%d:%s doesn't end in newline\n",
 335                        file, line, fn);
 336
 337         header.ph_len = known_size + needed;
 338         debug_buf = (char *)cfs_page_address(tage->page) + tage->used;
 339
 340         if (libcfs_debug_binary) {
 341                 memcpy(debug_buf, &header, sizeof(header));
 342                 tage->used += sizeof(header);
 343                 debug_buf += sizeof(header);
 344         }
 345
 346         /* indent message according to the nesting level */
 347         while (depth-- > 0) {
 348                 *(debug_buf++) = '.';
 349                 ++ tage->used;
 350         }
 351
 352         strcpy(debug_buf, file);
 353         tage->used += strlen(file) + 1;
 354         debug_buf += strlen(file) + 1;
 355
 356         if (fn) {
 357                 strcpy(debug_buf, fn);
 358                 tage->used += strlen(fn) + 1;
 359                 debug_buf += strlen(fn) + 1;
 360         }
 361
 362         __LASSERT(debug_buf == string_buf);
 363
 364         tage->used += needed;
 365         __LASSERT (tage->used <= CFS_PAGE_SIZE);
 366
 367 console:
 368         if ((mask & libcfs_printk) == 0) {
 369                 /* no console output requested */
 370                 if (tcd != NULL)
 371                         trace_put_tcd(tcd);
 372                 return 1;
 373         }
 374
 375         if (cdls != NULL) {
 376                 if (libcfs_console_ratelimit &&
 377                     cdls->cdls_next != 0 &&     /* not first time ever */
 378                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 379                         /* skipping a console message */
 380                         cdls->cdls_count++;
 381                         if (tcd != NULL)
 382                                 trace_put_tcd(tcd);
 383                         return 1;
 384                 }
 385
 386                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 387                                                        libcfs_console_max_delay
 388                                                        + cfs_time_seconds(10))) {
 389                         /* last timeout was a long time ago */
 390                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 391                 } else {
 392                         cdls->cdls_delay *= libcfs_console_backoff;
 393
 394                         if (cdls->cdls_delay < libcfs_console_min_delay)
 395                                 cdls->cdls_delay = libcfs_console_min_delay;
 396                         else if (cdls->cdls_delay > libcfs_console_max_delay)
 397                                 cdls->cdls_delay = libcfs_console_max_delay;
 398                 }
 399
 400                 /* ensure cdls_next is never zero after it's been seen */
 401                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 402         }
 403
 404         if (tcd != NULL) {
 405                 print_to_console(&header, mask, string_buf, needed, file, fn);
 406                 trace_put_tcd(tcd);
 407         } else {
 408                 string_buf = trace_get_console_buffer();
 409
 410                 needed = 0;
 411                 if (format1 != NULL) {
 412                         va_copy(ap, args);
 413                         needed = vsnprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE, format1, ap);
 414                         va_end(ap);
 415                 }
 416                 if (format2 != NULL) {
 417                         remain = TRACE_CONSOLE_BUFFER_SIZE - needed;
 418                         if (remain > 0) {
 419                                 va_start(ap, format2);
 420                                 needed += vsnprintf(string_buf+needed, remain, format2, ap);
 421                                 va_end(ap);
 422                         }
 423                 }
 424                 print_to_console(&header, mask,
 425                                  string_buf, needed, file, fn);
 426
 427                 trace_put_console_buffer(string_buf);
 428         }
 429
 430         if (cdls != NULL && cdls->cdls_count != 0) {
 431                 string_buf = trace_get_console_buffer();
 432
 433                 needed = snprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE,
 434                          "Skipped %d previous similar message%s\n",
 435                          cdls->cdls_count, (cdls->cdls_count > 1) ? "s" : "");
 436
 437                 print_to_console(&header, mask,
 438                                  string_buf, needed, file, fn);
 439
 440                 trace_put_console_buffer(string_buf);
 441                 cdls->cdls_count = 0;
 442         }
 443
 444         return 0;
 445 }
 446 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 447
 448 void
 449 libcfs_assertion_failed(const char *expr, const char *file,
 450                         const char *func, const int line)
 451 {
 452         libcfs_debug_msg(NULL, 0, D_EMERG, file, func, line,
 453                          "ASSERTION(%s) failed\n", expr);
 454         lbug_with_loc(file, func, line);
 455 }
 456 EXPORT_SYMBOL(libcfs_assertion_failed);
 457
 458 void
 459 trace_assertion_failed(const char *str,
 460                        const char *fn, const char *file, int line)
 461 {
 462         struct ptldebug_header hdr;
 463
 464         libcfs_panic_in_progress = 1;
 465         libcfs_catastrophe = 1;
 466         mb();
 467
 468         set_ptldebug_header(&hdr, DEBUG_SUBSYSTEM, D_EMERG, line,
 469                             CDEBUG_STACK());
 470
 471         print_to_console(&hdr, D_EMERG, str, strlen(str), file, fn);
 472
 473         LIBCFS_PANIC("Lustre debug assertion failure\n");
 474
 475         /* not reached */
 476 }
 477
 478 static void
 479 panic_collect_pages(struct page_collection *pc)
 480 {
 481         /* Do the collect_pages job on a single CPU: assumes that all other
 482          * CPUs have been stopped during a panic.  If this isn't true for some
 483          * arch, this will have to be implemented separately in each arch.  */
 484         int                    i;
 485         int                    j;
 486         struct trace_cpu_data *tcd;
 487
 488         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 489
 490         tcd_for_each(tcd, i, j) {
 491                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 492                 tcd->tcd_cur_pages = 0;
 493
 494                 if (pc->pc_want_daemon_pages) {
 495                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 496                         tcd->tcd_cur_daemon_pages = 0;
 497                 }
 498         }
 499 }
 500
 501 static void collect_pages_on_cpu(void *info)
 502 {
 503         struct trace_cpu_data *tcd;
 504         struct page_collection *pc = info;
 505         int i;
 506
 507         spin_lock(&pc->pc_lock);
 508         tcd_for_each_type_lock(tcd, i) {
 509                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 510                 tcd->tcd_cur_pages = 0;
 511                 if (pc->pc_want_daemon_pages) {
 512                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 513                         tcd->tcd_cur_daemon_pages = 0;
 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                 trace_call_on_all_cpus(collect_pages_on_cpu, pc);
 527 }
 528
 529 static void put_pages_back_on_cpu(void *info)
 530 {
 531         struct page_collection *pc = info;
 532         struct trace_cpu_data *tcd;
 533         struct list_head *cur_head;
 534         struct trace_page *tage;
 535         struct trace_page *tmp;
 536         int i;
 537
 538         spin_lock(&pc->pc_lock);
 539         tcd_for_each_type_lock(tcd, i) {
 540                 cur_head = tcd->tcd_pages.next;
 541
 542                 cfs_list_for_each_entry_safe_typed(tage, tmp, &pc->pc_pages,
 543                                                    struct trace_page, linkage) {
 544
 545                         __LASSERT_TAGE_INVARIANT(tage);
 546
 547                         if (tage->cpu != smp_processor_id() || tage->type != i)
 548                                 continue;
 549
 550                         tage_to_tail(tage, cur_head);
 551                         tcd->tcd_cur_pages++;
 552                 }
 553         }
 554         spin_unlock(&pc->pc_lock);
 555 }
 556
 557 static void put_pages_back(struct page_collection *pc)
 558 {
 559         if (!libcfs_panic_in_progress)
 560                 trace_call_on_all_cpus(put_pages_back_on_cpu, pc);
 561 }
 562
 563 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 564  * we have a good amount of data at all times for dumping during an LBUG, even
 565  * if we have been steadily writing (and otherwise discarding) pages via the
 566  * debug daemon. */
 567 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 568                                          struct trace_cpu_data *tcd)
 569 {
 570         struct trace_page *tage;
 571         struct trace_page *tmp;
 572
 573         spin_lock(&pc->pc_lock);
 574         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc->pc_pages,
 575                                            struct trace_page, linkage) {
 576
 577                 __LASSERT_TAGE_INVARIANT(tage);
 578
 579                 if (tage->cpu != smp_processor_id() ||
 580                     tage->type != tcd->tcd_type)
 581                         continue;
 582
 583                 tage_to_tail(tage, &tcd->tcd_daemon_pages);
 584                 tcd->tcd_cur_daemon_pages++;
 585
 586                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 587                         struct trace_page *victim;
 588
 589                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 590                         victim = tage_from_list(tcd->tcd_daemon_pages.next);
 591
 592                         __LASSERT_TAGE_INVARIANT(victim);
 593
 594                         list_del(&victim->linkage);
 595                         tage_free(victim);
 596                         tcd->tcd_cur_daemon_pages--;
 597                 }
 598         }
 599         spin_unlock(&pc->pc_lock);
 600 }
 601
 602 static void put_pages_on_daemon_list_on_cpu(void *info)
 603 {
 604         struct trace_cpu_data *tcd;
 605         int i;
 606
 607         tcd_for_each_type_lock(tcd, i)
 608                 put_pages_on_tcd_daemon_list(info, tcd);
 609 }
 610
 611 static void put_pages_on_daemon_list(struct page_collection *pc)
 612 {
 613         trace_call_on_all_cpus(put_pages_on_daemon_list_on_cpu, pc);
 614 }
 615
 616 void trace_debug_print(void)
 617 {
 618         struct page_collection pc;
 619         struct trace_page *tage;
 620         struct trace_page *tmp;
 621
 622         spin_lock_init(&pc.pc_lock);
 623
 624         pc.pc_want_daemon_pages = 1;
 625         collect_pages(&pc);
 626         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 627                                            struct trace_page, linkage) {
 628                 char *p, *file, *fn;
 629                 cfs_page_t *page;
 630
 631                 __LASSERT_TAGE_INVARIANT(tage);
 632
 633                 page = tage->page;
 634                 p = cfs_page_address(page);
 635                 while (p < ((char *)cfs_page_address(page) + tage->used)) {
 636                         struct ptldebug_header *hdr;
 637                         int len;
 638                         hdr = (void *)p;
 639                         p += sizeof(*hdr);
 640                         file = p;
 641                         p += strlen(file) + 1;
 642                         fn = p;
 643                         p += strlen(fn) + 1;
 644                         len = hdr->ph_len - (int)(p - (char *)hdr);
 645
 646                         print_to_console(hdr, D_EMERG, p, len, file, fn);
 647
 648                         p += len;
 649                 }
 650
 651                 list_del(&tage->linkage);
 652                 tage_free(tage);
 653         }
 654 }
 655
 656 int tracefile_dump_all_pages(char *filename)
 657 {
 658         struct page_collection pc;
 659         cfs_file_t *filp;
 660         struct trace_page *tage;
 661         struct trace_page *tmp;
 662         int rc;
 663
 664         CFS_DECL_MMSPACE;
 665
 666         tracefile_write_lock();
 667
 668         filp = cfs_filp_open(filename,
 669                              O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600, &rc);
 670         if (!filp) {
 671                 if (rc != -EEXIST)
 672                         printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 673                                filename, rc);
 674                 goto out;
 675         }
 676
 677         spin_lock_init(&pc.pc_lock);
 678         pc.pc_want_daemon_pages = 1;
 679         collect_pages(&pc);
 680         if (list_empty(&pc.pc_pages)) {
 681                 rc = 0;
 682                 goto close;
 683         }
 684
 685         /* ok, for now, just write the pages.  in the future we'll be building
 686          * iobufs with the pages and calling generic_direct_IO */
 687         CFS_MMSPACE_OPEN;
 688         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 689                                            struct trace_page, linkage) {
 690
 691                 __LASSERT_TAGE_INVARIANT(tage);
 692
 693                 rc = cfs_filp_write(filp, cfs_page_address(tage->page),
 694                                     tage->used, cfs_filp_poff(filp));
 695                 if (rc != (int)tage->used) {
 696                         printk(KERN_WARNING "wanted to write %u but wrote "
 697                                "%d\n", tage->used, rc);
 698                         put_pages_back(&pc);
 699                         __LASSERT(list_empty(&pc.pc_pages));
 700                         break;
 701                 }
 702                 list_del(&tage->linkage);
 703                 tage_free(tage);
 704         }
 705         CFS_MMSPACE_CLOSE;
 706         rc = cfs_filp_fsync(filp);
 707         if (rc)
 708                 printk(KERN_ERR "sync returns %d\n", rc);
 709  close:
 710         cfs_filp_close(filp);
 711  out:
 712         tracefile_write_unlock();
 713         return rc;
 714 }
 715
 716 void trace_flush_pages(void)
 717 {
 718         struct page_collection pc;
 719         struct trace_page *tage;
 720         struct trace_page *tmp;
 721
 722         spin_lock_init(&pc.pc_lock);
 723
 724         pc.pc_want_daemon_pages = 1;
 725         collect_pages(&pc);
 726         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 727                                            struct trace_page, linkage) {
 728
 729                 __LASSERT_TAGE_INVARIANT(tage);
 730
 731                 list_del(&tage->linkage);
 732                 tage_free(tage);
 733         }
 734 }
 735
 736 int trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 737                         const char *usr_buffer, int usr_buffer_nob)
 738 {
 739         int    nob;
 740
 741         if (usr_buffer_nob > knl_buffer_nob)
 742                 return -EOVERFLOW;
 743
 744         if (copy_from_user((void *)knl_buffer,
 745                            (void *)usr_buffer, usr_buffer_nob))
 746                 return -EFAULT;
 747
 748         nob = strnlen(knl_buffer, usr_buffer_nob);
 749         while (nob-- >= 0)                      /* strip trailing whitespace */
 750                 if (!isspace(knl_buffer[nob]))
 751                         break;
 752
 753         if (nob < 0)                            /* empty string */
 754                 return -EINVAL;
 755
 756         if (nob == knl_buffer_nob)              /* no space to terminate */
 757                 return -EOVERFLOW;
 758
 759         knl_buffer[nob + 1] = 0;                /* terminate */
 760         return 0;
 761 }
 762
 763 int trace_copyout_string(char *usr_buffer, int usr_buffer_nob,
 764                          const char *knl_buffer, char *append)
 765 {
 766         /* NB if 'append' != NULL, it's a single character to append to the
 767          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 768          * terminating zero byte) for sysctl entries */
 769         int   nob = strlen(knl_buffer);
 770
 771         if (nob > usr_buffer_nob)
 772                 nob = usr_buffer_nob;
 773
 774         if (copy_to_user(usr_buffer, knl_buffer, nob))
 775                 return -EFAULT;
 776
 777         if (append != NULL && nob < usr_buffer_nob) {
 778                 if (copy_to_user(usr_buffer + nob, append, 1))
 779                         return -EFAULT;
 780
 781                 nob++;
 782         }
 783
 784         return nob;
 785 }
 786
 787 int trace_allocate_string_buffer(char **str, int nob)
 788 {
 789         if (nob > 2 * CFS_PAGE_SIZE)            /* string must be "sensible" */
 790                 return -EINVAL;
 791
 792         *str = cfs_alloc(nob, CFS_ALLOC_STD | CFS_ALLOC_ZERO);
 793         if (*str == NULL)
 794                 return -ENOMEM;
 795
 796         return 0;
 797 }
 798
 799 void trace_free_string_buffer(char *str, int nob)
 800 {
 801         cfs_free(str);
 802 }
 803
 804 int trace_dump_debug_buffer_usrstr(void *usr_str, int usr_str_nob)
 805 {
 806         char         *str;
 807         int           rc;
 808
 809         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 810         if (rc != 0)
 811                 return rc;
 812
 813         rc = trace_copyin_string(str, usr_str_nob + 1,
 814                                  usr_str, usr_str_nob);
 815         if (rc != 0)
 816                 goto out;
 817
 818 #if !defined(__WINNT__)
 819         if (str[0] != '/') {
 820                 rc = -EINVAL;
 821                 goto out;
 822         }
 823 #endif
 824         rc = tracefile_dump_all_pages(str);
 825 out:
 826         trace_free_string_buffer(str, usr_str_nob + 1);
 827         return rc;
 828 }
 829
 830 int trace_daemon_command(char *str)
 831 {
 832         int       rc = 0;
 833
 834         tracefile_write_lock();
 835
 836         if (strcmp(str, "stop") == 0) {
 837                 trace_stop_thread();
 838                 memset(tracefile, 0, sizeof(tracefile));
 839
 840         } else if (strncmp(str, "size=", 5) == 0) {
 841                 tracefile_size = simple_strtoul(str + 5, NULL, 0);
 842                 if (tracefile_size < 10 || tracefile_size > 20480)
 843                         tracefile_size = TRACEFILE_SIZE;
 844                 else
 845                         tracefile_size <<= 20;
 846
 847         } else if (strlen(str) >= sizeof(tracefile)) {
 848                 rc = -ENAMETOOLONG;
 849 #ifndef __WINNT__
 850         } else if (str[0] != '/') {
 851                 rc = -EINVAL;
 852 #endif
 853         } else {
 854                 strcpy(tracefile, str);
 855
 856                 printk(KERN_INFO "Lustre: debug daemon will attempt to start writing "
 857                        "to %s (%lukB max)\n", tracefile,
 858                        (long)(tracefile_size >> 10));
 859
 860                 trace_start_thread();
 861         }
 862
 863         tracefile_write_unlock();
 864         return rc;
 865 }
 866
 867 int trace_daemon_command_usrstr(void *usr_str, int usr_str_nob)
 868 {
 869         char *str;
 870         int   rc;
 871
 872         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 873         if (rc != 0)
 874                 return rc;
 875
 876         rc = trace_copyin_string(str, usr_str_nob + 1,
 877                                  usr_str, usr_str_nob);
 878         if (rc == 0)
 879                 rc = trace_daemon_command(str);
 880
 881         trace_free_string_buffer(str, usr_str_nob + 1);
 882         return rc;
 883 }
 884
 885 int trace_set_debug_mb(int mb)
 886 {
 887         int i;
 888         int j;
 889         int pages;
 890         int limit = trace_max_debug_mb();
 891         struct trace_cpu_data *tcd;
 892
 893         if (mb < num_possible_cpus())
 894                 return -EINVAL;
 895
 896         if (mb > limit) {
 897                 printk(KERN_ERR "Lustre: Refusing to set debug buffer size to "
 898                        "%dMB - limit is %d\n", mb, limit);
 899                 return -EINVAL;
 900         }
 901
 902         mb /= num_possible_cpus();
 903         pages = mb << (20 - CFS_PAGE_SHIFT);
 904
 905         tracefile_write_lock();
 906
 907         tcd_for_each(tcd, i, j)
 908                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 909
 910         tracefile_write_unlock();
 911
 912         return 0;
 913 }
 914
 915 int trace_set_debug_mb_usrstr(void *usr_str, int usr_str_nob)
 916 {
 917         char     str[32];
 918         int      rc;
 919
 920         rc = trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 921         if (rc < 0)
 922                 return rc;
 923
 924         return trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 925 }
 926
 927 int trace_get_debug_mb(void)
 928 {
 929         int i;
 930         int j;
 931         struct trace_cpu_data *tcd;
 932         int total_pages = 0;
 933
 934         tracefile_read_lock();
 935
 936         tcd_for_each(tcd, i, j)
 937                 total_pages += tcd->tcd_max_pages;
 938
 939         tracefile_read_unlock();
 940
 941         return (total_pages >> (20 - CFS_PAGE_SHIFT)) + 1;
 942 }
 943
 944 static int tracefiled(void *arg)
 945 {
 946         struct page_collection pc;
 947         struct tracefiled_ctl *tctl = arg;
 948         struct trace_page *tage;
 949         struct trace_page *tmp;
 950         struct ptldebug_header *hdr;
 951         cfs_file_t *filp;
 952         int rc;
 953
 954         CFS_DECL_MMSPACE;
 955
 956         /* we're started late enough that we pick up init's fs context */
 957         /* this is so broken in uml?  what on earth is going on? */
 958         cfs_daemonize("ktracefiled");
 959
 960         spin_lock_init(&pc.pc_lock);
 961         complete(&tctl->tctl_start);
 962
 963         while (1) {
 964                 cfs_waitlink_t __wait;
 965
 966                 cfs_waitlink_init(&__wait);
 967                 cfs_waitq_add(&tctl->tctl_waitq, &__wait);
 968                 set_current_state(TASK_INTERRUPTIBLE);
 969                 cfs_waitq_timedwait(&__wait, CFS_TASK_INTERRUPTIBLE,
 970                                     cfs_time_seconds(1));
 971                 cfs_waitq_del(&tctl->tctl_waitq, &__wait);
 972
 973                 if (atomic_read(&tctl->tctl_shutdown))
 974                         break;
 975
 976                 pc.pc_want_daemon_pages = 0;
 977                 collect_pages(&pc);
 978                 if (list_empty(&pc.pc_pages))
 979                         continue;
 980
 981                 filp = NULL;
 982                 tracefile_read_lock();
 983                 if (tracefile[0] != 0) {
 984                         filp = cfs_filp_open(tracefile,
 985                                              O_CREAT | O_RDWR | O_LARGEFILE,
 986                                              0600, &rc);
 987                         if (!(filp))
 988                                 printk(KERN_WARNING "couldn't open %s: %d\n",
 989                                        tracefile, rc);
 990                 }
 991                 tracefile_read_unlock();
 992                 if (filp == NULL) {
 993                         put_pages_on_daemon_list(&pc);
 994                         __LASSERT(list_empty(&pc.pc_pages));
 995                         continue;
 996                 }
 997
 998                 CFS_MMSPACE_OPEN;
 999
1000                 /* mark the first header, so we can sort in chunks */
1001                 tage = tage_from_list(pc.pc_pages.next);
1002                 __LASSERT_TAGE_INVARIANT(tage);
1003
1004                 hdr = cfs_page_address(tage->page);
1005                 hdr->ph_flags |= PH_FLAG_FIRST_RECORD;
1006
1007                 cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
1008                                                    struct trace_page, linkage) {
1009                         static loff_t f_pos;
1010
1011                         __LASSERT_TAGE_INVARIANT(tage);
1012
1013                         if (f_pos >= (off_t)tracefile_size)
1014                                 f_pos = 0;
1015                         else if (f_pos > (off_t)cfs_filp_size(filp))
1016                                 f_pos = cfs_filp_size(filp);
1017
1018                         rc = cfs_filp_write(filp, cfs_page_address(tage->page),
1019                                             tage->used, &f_pos);
1020                         if (rc != (int)tage->used) {
1021                                 printk(KERN_WARNING "wanted to write %u but "
1022                                        "wrote %d\n", tage->used, rc);
1023                                 put_pages_back(&pc);
1024                                 __LASSERT(list_empty(&pc.pc_pages));
1025                         }
1026                 }
1027                 CFS_MMSPACE_CLOSE;
1028
1029                 cfs_filp_close(filp);
1030                 put_pages_on_daemon_list(&pc);
1031                 __LASSERT(list_empty(&pc.pc_pages));
1032         }
1033         complete(&tctl->tctl_stop);
1034         return 0;
1035 }
1036
1037 int trace_start_thread(void)
1038 {
1039         struct tracefiled_ctl *tctl = &trace_tctl;
1040         int rc = 0;
1041
1042         mutex_down(&trace_thread_sem);
1043         if (thread_running)
1044                 goto out;
1045
1046         init_completion(&tctl->tctl_start);
1047         init_completion(&tctl->tctl_stop);
1048         cfs_waitq_init(&tctl->tctl_waitq);
1049         atomic_set(&tctl->tctl_shutdown, 0);
1050
1051         if (cfs_kernel_thread(tracefiled, tctl, 0) < 0) {
1052                 rc = -ECHILD;
1053                 goto out;
1054         }
1055
1056         wait_for_completion(&tctl->tctl_start);
1057         thread_running = 1;
1058 out:
1059         mutex_up(&trace_thread_sem);
1060         return rc;
1061 }
1062
1063 void trace_stop_thread(void)
1064 {
1065         struct tracefiled_ctl *tctl = &trace_tctl;
1066
1067         mutex_down(&trace_thread_sem);
1068         if (thread_running) {
1069                 printk(KERN_INFO "Lustre: shutting down debug daemon thread...\n");
1070                 atomic_set(&tctl->tctl_shutdown, 1);
1071                 wait_for_completion(&tctl->tctl_stop);
1072                 thread_running = 0;
1073         }
1074         mutex_up(&trace_thread_sem);
1075 }
1076
1077 int tracefile_init(int max_pages)
1078 {
1079         struct trace_cpu_data *tcd;
1080         int                    i;
1081         int                    j;
1082         int                    rc;
1083         int                    factor;
1084
1085         rc = tracefile_init_arch();
1086         if (rc != 0)
1087                 return rc;
1088
1089         tcd_for_each(tcd, i, j) {
1090                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1091                 factor = tcd->tcd_pages_factor;
1092                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
1093                 CFS_INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1094                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1095                 tcd->tcd_cur_pages = 0;
1096                 tcd->tcd_cur_stock_pages = 0;
1097                 tcd->tcd_cur_daemon_pages = 0;
1098                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1099                 LASSERT(tcd->tcd_max_pages > 0);
1100                 tcd->tcd_shutting_down = 0;
1101         }
1102
1103         return 0;
1104 }
1105
1106 static void trace_cleanup_on_cpu(void *info)
1107 {
1108         struct trace_cpu_data *tcd;
1109         struct trace_page *tage;
1110         struct trace_page *tmp;
1111         int i;
1112
1113         tcd_for_each_type_lock(tcd, i) {
1114                 tcd->tcd_shutting_down = 1;
1115
1116                 cfs_list_for_each_entry_safe_typed(tage, tmp, &tcd->tcd_pages,
1117                                                    struct trace_page, linkage) {
1118                         __LASSERT_TAGE_INVARIANT(tage);
1119
1120                         list_del(&tage->linkage);
1121                         tage_free(tage);
1122                 }
1123                 tcd->tcd_cur_pages = 0;
1124         }
1125 }
1126
1127 static void trace_cleanup(void)
1128 {
1129         struct page_collection pc;
1130
1131         CFS_INIT_LIST_HEAD(&pc.pc_pages);
1132         spin_lock_init(&pc.pc_lock);
1133
1134         trace_call_on_all_cpus(trace_cleanup_on_cpu, &pc);
1135
1136         tracefile_fini_arch();
1137 }
1138
1139 void tracefile_exit(void)
1140 {
1141         trace_stop_thread();
1142         trace_cleanup();
1143 }