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         /* cfs_enter_debugger(); */
 455         lbug_with_loc(file, func, line);
 456 }
 457 EXPORT_SYMBOL(libcfs_assertion_failed);
 458
 459 void
 460 trace_assertion_failed(const char *str,
 461                        const char *fn, const char *file, int line)
 462 {
 463         struct ptldebug_header hdr;
 464
 465         libcfs_panic_in_progress = 1;
 466         libcfs_catastrophe = 1;
 467         mb();
 468
 469         set_ptldebug_header(&hdr, DEBUG_SUBSYSTEM, D_EMERG, line,
 470                             CDEBUG_STACK());
 471
 472         print_to_console(&hdr, D_EMERG, str, strlen(str), file, fn);
 473
 474         LIBCFS_PANIC("Lustre debug assertion failure\n");
 475
 476         /* not reached */
 477 }
 478
 479 static void
 480 panic_collect_pages(struct page_collection *pc)
 481 {
 482         /* Do the collect_pages job on a single CPU: assumes that all other
 483          * CPUs have been stopped during a panic.  If this isn't true for some
 484          * arch, this will have to be implemented separately in each arch.  */
 485         int                    i;
 486         int                    j;
 487         struct trace_cpu_data *tcd;
 488
 489         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 490
 491         tcd_for_each(tcd, i, j) {
 492                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 493                 tcd->tcd_cur_pages = 0;
 494
 495                 if (pc->pc_want_daemon_pages) {
 496                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 497                         tcd->tcd_cur_daemon_pages = 0;
 498                 }
 499         }
 500 }
 501
 502 static void collect_pages_on_cpu(void *info)
 503 {
 504         struct trace_cpu_data *tcd;
 505         struct page_collection *pc = info;
 506         int i;
 507
 508         spin_lock(&pc->pc_lock);
 509         tcd_for_each_type_lock(tcd, i) {
 510                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 511                 tcd->tcd_cur_pages = 0;
 512                 if (pc->pc_want_daemon_pages) {
 513                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 514                         tcd->tcd_cur_daemon_pages = 0;
 515                 }
 516         }
 517         spin_unlock(&pc->pc_lock);
 518 }
 519
 520 static void collect_pages(struct page_collection *pc)
 521 {
 522         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 523
 524         if (libcfs_panic_in_progress)
 525                 panic_collect_pages(pc);
 526         else
 527                 trace_call_on_all_cpus(collect_pages_on_cpu, pc);
 528 }
 529
 530 static void put_pages_back_on_cpu(void *info)
 531 {
 532         struct page_collection *pc = info;
 533         struct trace_cpu_data *tcd;
 534         struct list_head *cur_head;
 535         struct trace_page *tage;
 536         struct trace_page *tmp;
 537         int i;
 538
 539         spin_lock(&pc->pc_lock);
 540         tcd_for_each_type_lock(tcd, i) {
 541                 cur_head = tcd->tcd_pages.next;
 542
 543                 cfs_list_for_each_entry_safe_typed(tage, tmp, &pc->pc_pages,
 544                                                    struct trace_page, linkage) {
 545
 546                         __LASSERT_TAGE_INVARIANT(tage);
 547
 548                         if (tage->cpu != smp_processor_id() || tage->type != i)
 549                                 continue;
 550
 551                         tage_to_tail(tage, cur_head);
 552                         tcd->tcd_cur_pages++;
 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                 trace_call_on_all_cpus(put_pages_back_on_cpu, 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         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc->pc_pages,
 576                                            struct trace_page, linkage) {
 577
 578                 __LASSERT_TAGE_INVARIANT(tage);
 579
 580                 if (tage->cpu != smp_processor_id() ||
 581                     tage->type != tcd->tcd_type)
 582                         continue;
 583
 584                 tage_to_tail(tage, &tcd->tcd_daemon_pages);
 585                 tcd->tcd_cur_daemon_pages++;
 586
 587                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 588                         struct trace_page *victim;
 589
 590                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 591                         victim = tage_from_list(tcd->tcd_daemon_pages.next);
 592
 593                         __LASSERT_TAGE_INVARIANT(victim);
 594
 595                         list_del(&victim->linkage);
 596                         tage_free(victim);
 597                         tcd->tcd_cur_daemon_pages--;
 598                 }
 599         }
 600         spin_unlock(&pc->pc_lock);
 601 }
 602
 603 static void put_pages_on_daemon_list_on_cpu(void *info)
 604 {
 605         struct trace_cpu_data *tcd;
 606         int i;
 607
 608         tcd_for_each_type_lock(tcd, i)
 609                 put_pages_on_tcd_daemon_list(info, tcd);
 610 }
 611
 612 static void put_pages_on_daemon_list(struct page_collection *pc)
 613 {
 614         trace_call_on_all_cpus(put_pages_on_daemon_list_on_cpu, pc);
 615 }
 616
 617 void trace_debug_print(void)
 618 {
 619         struct page_collection pc;
 620         struct trace_page *tage;
 621         struct trace_page *tmp;
 622
 623         spin_lock_init(&pc.pc_lock);
 624
 625         pc.pc_want_daemon_pages = 1;
 626         collect_pages(&pc);
 627         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 628                                            struct trace_page, linkage) {
 629                 char *p, *file, *fn;
 630                 cfs_page_t *page;
 631
 632                 __LASSERT_TAGE_INVARIANT(tage);
 633
 634                 page = tage->page;
 635                 p = cfs_page_address(page);
 636                 while (p < ((char *)cfs_page_address(page) + tage->used)) {
 637                         struct ptldebug_header *hdr;
 638                         int len;
 639                         hdr = (void *)p;
 640                         p += sizeof(*hdr);
 641                         file = p;
 642                         p += strlen(file) + 1;
 643                         fn = p;
 644                         p += strlen(fn) + 1;
 645                         len = hdr->ph_len - (int)(p - (char *)hdr);
 646
 647                         print_to_console(hdr, D_EMERG, p, len, file, fn);
 648
 649                         p += len;
 650                 }
 651
 652                 list_del(&tage->linkage);
 653                 tage_free(tage);
 654         }
 655 }
 656
 657 int tracefile_dump_all_pages(char *filename)
 658 {
 659         struct page_collection pc;
 660         cfs_file_t *filp;
 661         struct trace_page *tage;
 662         struct trace_page *tmp;
 663         int rc;
 664
 665         CFS_DECL_MMSPACE;
 666
 667         tracefile_write_lock();
 668
 669         filp = cfs_filp_open(filename,
 670                              O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600, &rc);
 671         if (!filp) {
 672                 if (rc != -EEXIST)
 673                         printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 674                                filename, rc);
 675                 goto out;
 676         }
 677
 678         spin_lock_init(&pc.pc_lock);
 679         pc.pc_want_daemon_pages = 1;
 680         collect_pages(&pc);
 681         if (list_empty(&pc.pc_pages)) {
 682                 rc = 0;
 683                 goto close;
 684         }
 685
 686         /* ok, for now, just write the pages.  in the future we'll be building
 687          * iobufs with the pages and calling generic_direct_IO */
 688         CFS_MMSPACE_OPEN;
 689         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 690                                            struct trace_page, linkage) {
 691
 692                 __LASSERT_TAGE_INVARIANT(tage);
 693
 694                 rc = cfs_filp_write(filp, cfs_page_address(tage->page),
 695                                     tage->used, cfs_filp_poff(filp));
 696                 if (rc != (int)tage->used) {
 697                         printk(KERN_WARNING "wanted to write %u but wrote "
 698                                "%d\n", tage->used, rc);
 699                         put_pages_back(&pc);
 700                         __LASSERT(list_empty(&pc.pc_pages));
 701                         break;
 702                 }
 703                 list_del(&tage->linkage);
 704                 tage_free(tage);
 705         }
 706         CFS_MMSPACE_CLOSE;
 707         rc = cfs_filp_fsync(filp);
 708         if (rc)
 709                 printk(KERN_ERR "sync returns %d\n", rc);
 710  close:
 711         cfs_filp_close(filp);
 712  out:
 713         tracefile_write_unlock();
 714         return rc;
 715 }
 716
 717 void trace_flush_pages(void)
 718 {
 719         struct page_collection pc;
 720         struct trace_page *tage;
 721         struct trace_page *tmp;
 722
 723         spin_lock_init(&pc.pc_lock);
 724
 725         pc.pc_want_daemon_pages = 1;
 726         collect_pages(&pc);
 727         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 728                                            struct trace_page, linkage) {
 729
 730                 __LASSERT_TAGE_INVARIANT(tage);
 731
 732                 list_del(&tage->linkage);
 733                 tage_free(tage);
 734         }
 735 }
 736
 737 int trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 738                         const char *usr_buffer, int usr_buffer_nob)
 739 {
 740         int    nob;
 741
 742         if (usr_buffer_nob > knl_buffer_nob)
 743                 return -EOVERFLOW;
 744
 745         if (copy_from_user((void *)knl_buffer,
 746                            (void *)usr_buffer, usr_buffer_nob))
 747                 return -EFAULT;
 748
 749         nob = strnlen(knl_buffer, usr_buffer_nob);
 750         while (nob-- >= 0)                      /* strip trailing whitespace */
 751                 if (!isspace(knl_buffer[nob]))
 752                         break;
 753
 754         if (nob < 0)                            /* empty string */
 755                 return -EINVAL;
 756
 757         if (nob == knl_buffer_nob)              /* no space to terminate */
 758                 return -EOVERFLOW;
 759
 760         knl_buffer[nob + 1] = 0;                /* terminate */
 761         return 0;
 762 }
 763
 764 int trace_copyout_string(char *usr_buffer, int usr_buffer_nob,
 765                          const char *knl_buffer, char *append)
 766 {
 767         /* NB if 'append' != NULL, it's a single character to append to the
 768          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 769          * terminating zero byte) for sysctl entries */
 770         int   nob = strlen(knl_buffer);
 771
 772         if (nob > usr_buffer_nob)
 773                 nob = usr_buffer_nob;
 774
 775         if (copy_to_user(usr_buffer, knl_buffer, nob))
 776                 return -EFAULT;
 777
 778         if (append != NULL && nob < usr_buffer_nob) {
 779                 if (copy_to_user(usr_buffer + nob, append, 1))
 780                         return -EFAULT;
 781
 782                 nob++;
 783         }
 784
 785         return nob;
 786 }
 787
 788 int trace_allocate_string_buffer(char **str, int nob)
 789 {
 790         if (nob > 2 * CFS_PAGE_SIZE)            /* string must be "sensible" */
 791                 return -EINVAL;
 792
 793         *str = cfs_alloc(nob, CFS_ALLOC_STD | CFS_ALLOC_ZERO);
 794         if (*str == NULL)
 795                 return -ENOMEM;
 796
 797         return 0;
 798 }
 799
 800 void trace_free_string_buffer(char *str, int nob)
 801 {
 802         cfs_free(str);
 803 }
 804
 805 int trace_dump_debug_buffer_usrstr(void *usr_str, int usr_str_nob)
 806 {
 807         char         *str;
 808         int           rc;
 809
 810         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 811         if (rc != 0)
 812                 return rc;
 813
 814         rc = trace_copyin_string(str, usr_str_nob + 1,
 815                                  usr_str, usr_str_nob);
 816         if (rc != 0)
 817                 goto out;
 818
 819 #if !defined(__WINNT__)
 820         if (str[0] != '/') {
 821                 rc = -EINVAL;
 822                 goto out;
 823         }
 824 #endif
 825         rc = tracefile_dump_all_pages(str);
 826 out:
 827         trace_free_string_buffer(str, usr_str_nob + 1);
 828         return rc;
 829 }
 830
 831 int trace_daemon_command(char *str)
 832 {
 833         int       rc = 0;
 834
 835         tracefile_write_lock();
 836
 837         if (strcmp(str, "stop") == 0) {
 838                 tracefile_write_unlock();
 839                 trace_stop_thread();
 840                 tracefile_write_lock();
 841                 memset(tracefile, 0, sizeof(tracefile));
 842
 843         } else if (strncmp(str, "size=", 5) == 0) {
 844                 tracefile_size = simple_strtoul(str + 5, NULL, 0);
 845                 if (tracefile_size < 10 || tracefile_size > 20480)
 846                         tracefile_size = TRACEFILE_SIZE;
 847                 else
 848                         tracefile_size <<= 20;
 849
 850         } else if (strlen(str) >= sizeof(tracefile)) {
 851                 rc = -ENAMETOOLONG;
 852 #ifndef __WINNT__
 853         } else if (str[0] != '/') {
 854                 rc = -EINVAL;
 855 #endif
 856         } else {
 857                 strcpy(tracefile, str);
 858
 859                 printk(KERN_INFO "Lustre: debug daemon will attempt to start writing "
 860                        "to %s (%lukB max)\n", tracefile,
 861                        (long)(tracefile_size >> 10));
 862
 863                 trace_start_thread();
 864         }
 865
 866         tracefile_write_unlock();
 867         return rc;
 868 }
 869
 870 int trace_daemon_command_usrstr(void *usr_str, int usr_str_nob)
 871 {
 872         char *str;
 873         int   rc;
 874
 875         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 876         if (rc != 0)
 877                 return rc;
 878
 879         rc = trace_copyin_string(str, usr_str_nob + 1,
 880                                  usr_str, usr_str_nob);
 881         if (rc == 0)
 882                 rc = trace_daemon_command(str);
 883
 884         trace_free_string_buffer(str, usr_str_nob + 1);
 885         return rc;
 886 }
 887
 888 int trace_set_debug_mb(int mb)
 889 {
 890         int i;
 891         int j;
 892         int pages;
 893         int limit = trace_max_debug_mb();
 894         struct trace_cpu_data *tcd;
 895
 896         if (mb < num_possible_cpus())
 897                 return -EINVAL;
 898
 899         if (mb > limit) {
 900                 printk(KERN_ERR "Lustre: Refusing to set debug buffer size to "
 901                        "%dMB - limit is %d\n", mb, limit);
 902                 return -EINVAL;
 903         }
 904
 905         mb /= num_possible_cpus();
 906         pages = mb << (20 - CFS_PAGE_SHIFT);
 907
 908         tracefile_write_lock();
 909
 910         tcd_for_each(tcd, i, j)
 911                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 912
 913         tracefile_write_unlock();
 914
 915         return 0;
 916 }
 917
 918 int trace_set_debug_mb_usrstr(void *usr_str, int usr_str_nob)
 919 {
 920         char     str[32];
 921         int      rc;
 922
 923         rc = trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 924         if (rc < 0)
 925                 return rc;
 926
 927         return trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 928 }
 929
 930 int trace_get_debug_mb(void)
 931 {
 932         int i;
 933         int j;
 934         struct trace_cpu_data *tcd;
 935         int total_pages = 0;
 936
 937         tracefile_read_lock();
 938
 939         tcd_for_each(tcd, i, j)
 940                 total_pages += tcd->tcd_max_pages;
 941
 942         tracefile_read_unlock();
 943
 944         return (total_pages >> (20 - CFS_PAGE_SHIFT)) + 1;
 945 }
 946
 947 static int tracefiled(void *arg)
 948 {
 949         struct page_collection pc;
 950         struct tracefiled_ctl *tctl = arg;
 951         struct trace_page *tage;
 952         struct trace_page *tmp;
 953         struct ptldebug_header *hdr;
 954         cfs_file_t *filp;
 955         int rc;
 956
 957         CFS_DECL_MMSPACE;
 958
 959         /* we're started late enough that we pick up init's fs context */
 960         /* this is so broken in uml?  what on earth is going on? */
 961         cfs_daemonize("ktracefiled");
 962
 963         spin_lock_init(&pc.pc_lock);
 964         complete(&tctl->tctl_start);
 965
 966         while (1) {
 967                 cfs_waitlink_t __wait;
 968
 969                 cfs_waitlink_init(&__wait);
 970                 cfs_waitq_add(&tctl->tctl_waitq, &__wait);
 971                 set_current_state(TASK_INTERRUPTIBLE);
 972                 cfs_waitq_timedwait(&__wait, CFS_TASK_INTERRUPTIBLE,
 973                                     cfs_time_seconds(1));
 974                 cfs_waitq_del(&tctl->tctl_waitq, &__wait);
 975
 976                 if (atomic_read(&tctl->tctl_shutdown))
 977                         break;
 978
 979                 pc.pc_want_daemon_pages = 0;
 980                 collect_pages(&pc);
 981                 if (list_empty(&pc.pc_pages))
 982                         continue;
 983
 984                 filp = NULL;
 985                 tracefile_read_lock();
 986                 if (tracefile[0] != 0) {
 987                         filp = cfs_filp_open(tracefile,
 988                                              O_CREAT | O_RDWR | O_LARGEFILE,
 989                                              0600, &rc);
 990                         if (!(filp))
 991                                 printk(KERN_WARNING "couldn't open %s: %d\n",
 992                                        tracefile, rc);
 993                 }
 994                 tracefile_read_unlock();
 995                 if (filp == NULL) {
 996                         put_pages_on_daemon_list(&pc);
 997                         __LASSERT(list_empty(&pc.pc_pages));
 998                         continue;
 999                 }
1000
1001                 CFS_MMSPACE_OPEN;
1002
1003                 /* mark the first header, so we can sort in chunks */
1004                 tage = tage_from_list(pc.pc_pages.next);
1005                 __LASSERT_TAGE_INVARIANT(tage);
1006
1007                 hdr = cfs_page_address(tage->page);
1008                 hdr->ph_flags |= PH_FLAG_FIRST_RECORD;
1009
1010                 cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
1011                                                    struct trace_page, linkage) {
1012                         static loff_t f_pos;
1013
1014                         __LASSERT_TAGE_INVARIANT(tage);
1015
1016                         if (f_pos >= (off_t)tracefile_size)
1017                                 f_pos = 0;
1018                         else if (f_pos > (off_t)cfs_filp_size(filp))
1019                                 f_pos = cfs_filp_size(filp);
1020
1021                         rc = cfs_filp_write(filp, cfs_page_address(tage->page),
1022                                             tage->used, &f_pos);
1023                         if (rc != (int)tage->used) {
1024                                 printk(KERN_WARNING "wanted to write %u but "
1025                                        "wrote %d\n", tage->used, rc);
1026                                 put_pages_back(&pc);
1027                                 __LASSERT(list_empty(&pc.pc_pages));
1028                         }
1029                 }
1030                 CFS_MMSPACE_CLOSE;
1031
1032                 cfs_filp_close(filp);
1033                 put_pages_on_daemon_list(&pc);
1034                 if (!list_empty(&pc.pc_pages)) {
1035                         int i;
1036
1037                         printk(KERN_ALERT "Lustre: trace pages aren't empty\n");
1038                         printk(KERN_ERR "total cpus(%d): ", num_possible_cpus());
1039                         for (i = 0; i < num_possible_cpus(); i++)
1040                                 if (cpu_online(i))
1041                                         printk(KERN_ERR "%d(on) ", i);
1042                                 else
1043                                         printk(KERN_ERR "%d(off) ", i);
1044                         printk(KERN_ERR "\n");
1045
1046                         i = 0;
1047                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1048                                                  linkage)
1049                                 printk(KERN_ERR "page %d belongs to cpu %d\n",
1050                                        ++i, tage->cpu);
1051                         printk(KERN_ERR "There are %d pages unwritten\n", i);
1052                 }
1053                 __LASSERT(list_empty(&pc.pc_pages));
1054         }
1055         complete(&tctl->tctl_stop);
1056         return 0;
1057 }
1058
1059 int trace_start_thread(void)
1060 {
1061         struct tracefiled_ctl *tctl = &trace_tctl;
1062         int rc = 0;
1063
1064         mutex_down(&trace_thread_sem);
1065         if (thread_running)
1066                 goto out;
1067
1068         init_completion(&tctl->tctl_start);
1069         init_completion(&tctl->tctl_stop);
1070         cfs_waitq_init(&tctl->tctl_waitq);
1071         atomic_set(&tctl->tctl_shutdown, 0);
1072
1073         if (cfs_kernel_thread(tracefiled, tctl, 0) < 0) {
1074                 rc = -ECHILD;
1075                 goto out;
1076         }
1077
1078         wait_for_completion(&tctl->tctl_start);
1079         thread_running = 1;
1080 out:
1081         mutex_up(&trace_thread_sem);
1082         return rc;
1083 }
1084
1085 void trace_stop_thread(void)
1086 {
1087         struct tracefiled_ctl *tctl = &trace_tctl;
1088
1089         mutex_down(&trace_thread_sem);
1090         if (thread_running) {
1091                 printk(KERN_INFO "Lustre: shutting down debug daemon thread...\n");
1092                 atomic_set(&tctl->tctl_shutdown, 1);
1093                 wait_for_completion(&tctl->tctl_stop);
1094                 thread_running = 0;
1095         }
1096         mutex_up(&trace_thread_sem);
1097 }
1098
1099 int tracefile_init(int max_pages)
1100 {
1101         struct trace_cpu_data *tcd;
1102         int                    i;
1103         int                    j;
1104         int                    rc;
1105         int                    factor;
1106
1107         rc = tracefile_init_arch();
1108         if (rc != 0)
1109                 return rc;
1110
1111         tcd_for_each(tcd, i, j) {
1112                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1113                 factor = tcd->tcd_pages_factor;
1114                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
1115                 CFS_INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1116                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1117                 tcd->tcd_cur_pages = 0;
1118                 tcd->tcd_cur_stock_pages = 0;
1119                 tcd->tcd_cur_daemon_pages = 0;
1120                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1121                 LASSERT(tcd->tcd_max_pages > 0);
1122                 tcd->tcd_shutting_down = 0;
1123         }
1124
1125         return 0;
1126 }
1127
1128 static void trace_cleanup_on_cpu(void *info)
1129 {
1130         struct trace_cpu_data *tcd;
1131         struct trace_page *tage;
1132         struct trace_page *tmp;
1133         int i;
1134
1135         tcd_for_each_type_lock(tcd, i) {
1136                 tcd->tcd_shutting_down = 1;
1137
1138                 cfs_list_for_each_entry_safe_typed(tage, tmp, &tcd->tcd_pages,
1139                                                    struct trace_page, linkage) {
1140                         __LASSERT_TAGE_INVARIANT(tage);
1141
1142                         list_del(&tage->linkage);
1143                         tage_free(tage);
1144                 }
1145                 tcd->tcd_cur_pages = 0;
1146         }
1147 }
1148
1149 static void trace_cleanup(void)
1150 {
1151         struct page_collection pc;
1152
1153         CFS_INIT_LIST_HEAD(&pc.pc_pages);
1154         spin_lock_init(&pc.pc_lock);
1155
1156         trace_call_on_all_cpus(trace_cleanup_on_cpu, &pc);
1157
1158         tracefile_fini_arch();
1159 }
1160
1161 void tracefile_exit(void)
1162 {
1163         trace_stop_thread();
1164         trace_cleanup();
1165 }