lnet/libcfs/tracefile.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2004 Cluster File Systems, Inc.
   5  *   Author: Zach Brown <zab@clusterfs.com>
   6  *   Author: Phil Schwan <phil@clusterfs.com>
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24
  25 #define DEBUG_SUBSYSTEM S_LNET
  26 #define LUSTRE_TRACEFILE_PRIVATE
  27 #include "tracefile.h"
  28
  29 #include <libcfs/kp30.h>
  30 #include <libcfs/libcfs.h>
  31
  32 /* XXX move things up to the top, comment */
  33 union trace_data_union trace_data[NR_CPUS] __cacheline_aligned;
  34
  35 char tracefile[TRACEFILE_NAME_SIZE];
  36 int64_t tracefile_size = TRACEFILE_SIZE;
  37 static struct tracefiled_ctl trace_tctl;
  38 struct semaphore trace_thread_sem;
  39 static int thread_running = 0;
  40
  41 atomic_t tage_allocated = ATOMIC_INIT(0);
  42
  43 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  44                                          struct trace_cpu_data *tcd);
  45
  46 static inline struct trace_page *tage_from_list(struct list_head *list)
  47 {
  48         return list_entry(list, struct trace_page, linkage);
  49 }
  50
  51 static struct trace_page *tage_alloc(int gfp)
  52 {
  53         cfs_page_t        *page;
  54         struct trace_page *tage;
  55
  56         /*
  57          * Don't spam console with allocation failures: they will be reported
  58          * by upper layer anyway.
  59          */
  60         gfp |= CFS_ALLOC_NOWARN;
  61         page = cfs_alloc_page(gfp);
  62         if (page == NULL)
  63                 return NULL;
  64
  65         tage = cfs_alloc(sizeof(*tage), gfp);
  66         if (tage == NULL) {
  67                 cfs_free_page(page);
  68                 return NULL;
  69         }
  70
  71         tage->page = page;
  72         atomic_inc(&tage_allocated);
  73         return tage;
  74 }
  75
  76 static void tage_free(struct trace_page *tage)
  77 {
  78         __LASSERT(tage != NULL);
  79         __LASSERT(tage->page != NULL);
  80
  81         cfs_free_page(tage->page);
  82         cfs_free(tage);
  83         atomic_dec(&tage_allocated);
  84 }
  85
  86 static void tage_to_tail(struct trace_page *tage, struct list_head *queue)
  87 {
  88         __LASSERT(tage != NULL);
  89         __LASSERT(queue != NULL);
  90
  91         list_move_tail(&tage->linkage, queue);
  92 }
  93
  94 int trace_refill_stock(struct trace_cpu_data *tcd, int gfp,
  95                        struct list_head *stock)
  96 {
  97         int i;
  98
  99         /*
 100          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 101          * from here: this will lead to infinite recursion.
 102          */
 103
 104         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++ i) {
 105                 struct trace_page *tage;
 106
 107                 tage = tage_alloc(gfp);
 108                 if (tage == NULL)
 109                         break;
 110                 list_add_tail(&tage->linkage, stock);
 111         }
 112         return i;
 113 }
 114
 115 /* return a page that has 'len' bytes left at the end */
 116 static struct trace_page *trace_get_tage_try(struct trace_cpu_data *tcd,
 117                                              unsigned long len)
 118 {
 119         struct trace_page *tage;
 120
 121         if (tcd->tcd_cur_pages > 0) {
 122                 __LASSERT(!list_empty(&tcd->tcd_pages));
 123                 tage = tage_from_list(tcd->tcd_pages.prev);
 124                 if (tage->used + len <= CFS_PAGE_SIZE)
 125                         return tage;
 126         }
 127
 128         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 129                 if (tcd->tcd_cur_stock_pages > 0) {
 130                         tage = tage_from_list(tcd->tcd_stock_pages.prev);
 131                         -- tcd->tcd_cur_stock_pages;
 132                         list_del_init(&tage->linkage);
 133                 } else {
 134                         tage = tage_alloc(CFS_ALLOC_ATOMIC);
 135                         if (tage == NULL) {
 136                                 printk(KERN_WARNING
 137                                        "failure to allocate a tage (%ld)\n",
 138                                        tcd->tcd_cur_pages);
 139                                 return NULL;
 140                         }
 141                 }
 142
 143                 tage->used = 0;
 144                 tage->cpu = smp_processor_id();
 145                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 146                 tcd->tcd_cur_pages++;
 147
 148                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 149                         struct tracefiled_ctl *tctl = &trace_tctl;
 150                         /*
 151                          * wake up tracefiled to process some pages.
 152                          */
 153                         cfs_waitq_signal(&tctl->tctl_waitq);
 154                 }
 155                 return tage;
 156         }
 157         return NULL;
 158 }
 159
 160 static void tcd_shrink(struct trace_cpu_data *tcd)
 161 {
 162         int pgcount = tcd->tcd_cur_pages / 10;
 163         struct page_collection pc;
 164         struct trace_page *tage;
 165         struct trace_page *tmp;
 166
 167         /*
 168          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 169          * from here: this will lead to infinite recursion.
 170          */
 171
 172         printk(KERN_WARNING "debug daemon buffer overflowed; discarding"
 173                " 10%% of pages (%d of %ld)\n", pgcount + 1, tcd->tcd_cur_pages);
 174
 175         CFS_INIT_LIST_HEAD(&pc.pc_pages);
 176         spin_lock_init(&pc.pc_lock);
 177
 178         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 179                 if (pgcount-- == 0)
 180                         break;
 181
 182                 list_move_tail(&tage->linkage, &pc.pc_pages);
 183                 tcd->tcd_cur_pages--;
 184         }
 185         put_pages_on_tcd_daemon_list(&pc, tcd);
 186 }
 187
 188 /* return a page that has 'len' bytes left at the end */
 189 static struct trace_page *trace_get_tage(struct trace_cpu_data *tcd,
 190                                          unsigned long len)
 191 {
 192         struct trace_page *tage;
 193
 194         /*
 195          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 196          * from here: this will lead to infinite recursion.
 197          */
 198
 199         if (len > CFS_PAGE_SIZE) {
 200                 printk(KERN_ERR
 201                        "cowardly refusing to write %lu bytes in a page\n", len);
 202                 return NULL;
 203         }
 204
 205         tage = trace_get_tage_try(tcd, len);
 206         if (tage != NULL)
 207                 return tage;
 208         if (thread_running)
 209                 tcd_shrink(tcd);
 210         if (tcd->tcd_cur_pages > 0) {
 211                 tage = tage_from_list(tcd->tcd_pages.next);
 212                 tage->used = 0;
 213                 tage_to_tail(tage, &tcd->tcd_pages);
 214         }
 215         return tage;
 216 }
 217
 218 int libcfs_debug_vmsg2(cfs_debug_limit_state_t *cdls, int subsys, int mask,
 219                        const char *file, const char *fn, const int line,
 220                        const char *format1, va_list args,
 221                        const char *format2, ...)
 222 {
 223         struct trace_cpu_data   *tcd = NULL;
 224         struct ptldebug_header   header;
 225         struct trace_page       *tage;
 226         /* string_buf is used only if tcd != NULL, and is always set then */
 227         char                    *string_buf = NULL;
 228         char                    *debug_buf;
 229         int                      known_size;
 230         int                      needed = 85; /* average message length */
 231         int                      max_nob;
 232         va_list                  ap;
 233         int                      depth;
 234         int                      i;
 235         int                      remain;
 236
 237         if (strchr(file, '/'))
 238                 file = strrchr(file, '/') + 1;
 239
 240
 241         set_ptldebug_header(&header, subsys, mask, line, CDEBUG_STACK());
 242
 243         tcd = trace_get_tcd();
 244         if (tcd == NULL)                /* arch may not log in IRQ context */
 245                 goto console;
 246
 247         if (tcd->tcd_shutting_down) {
 248                 trace_put_tcd(tcd);
 249                 tcd = NULL;
 250                 goto console;
 251         }
 252
 253         depth = __current_nesting_level();
 254         known_size = strlen(file) + 1 + depth;
 255         if (fn)
 256                 known_size += strlen(fn) + 1;
 257
 258         if (libcfs_debug_binary)
 259                 known_size += sizeof(header);
 260
 261         /*/
 262          * '2' used because vsnprintf return real size required for output
 263          * _without_ terminating NULL.
 264          * if needed is to small for this format.
 265          */
 266         for (i=0;i<2;i++) {
 267                 tage = trace_get_tage(tcd, needed + known_size + 1);
 268                 if (tage == NULL) {
 269                         if (needed + known_size > CFS_PAGE_SIZE)
 270                                 mask |= D_ERROR;
 271
 272                         trace_put_tcd(tcd);
 273                         tcd = NULL;
 274                         goto console;
 275                 }
 276
 277                 string_buf = (char *)cfs_page_address(tage->page)+tage->used+known_size;
 278
 279                 max_nob = CFS_PAGE_SIZE - tage->used - known_size;
 280                 if (max_nob <= 0) {
 281                         printk(KERN_EMERG "negative max_nob: %i\n", max_nob);
 282                         mask |= D_ERROR;
 283                         trace_put_tcd(tcd);
 284                         tcd = NULL;
 285                         goto console;
 286                 }
 287
 288                 needed = 0;
 289                 if (format1) {
 290                         va_copy(ap, args);
 291                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 292                         va_end(ap);
 293                 }
 294
 295
 296                 if (format2) {
 297                         remain = max_nob - needed;
 298                         if (remain < 0)
 299                                 remain = 0;
 300
 301                         va_start(ap, format2);
 302                         needed += vsnprintf(string_buf+needed, remain, format2, ap);
 303                         va_end(ap);
 304                 }
 305
 306                 if (needed < max_nob) /* well. printing ok.. */
 307                         break;
 308         }
 309
 310         if (*(string_buf+needed-1) != '\n')
 311                 printk(KERN_INFO "format at %s:%d:%s doesn't end in newline\n",
 312                        file, line, fn);
 313
 314         header.ph_len = known_size + needed;
 315         debug_buf = (char *)cfs_page_address(tage->page) + tage->used;
 316
 317         if (libcfs_debug_binary) {
 318                 memcpy(debug_buf, &header, sizeof(header));
 319                 tage->used += sizeof(header);
 320                 debug_buf += sizeof(header);
 321         }
 322
 323         /* indent message according to the nesting level */
 324         while (depth-- > 0) {
 325                 *(debug_buf++) = '.';
 326                 ++ tage->used;
 327         }
 328
 329         strcpy(debug_buf, file);
 330         tage->used += strlen(file) + 1;
 331         debug_buf += strlen(file) + 1;
 332
 333         if (fn) {
 334                 strcpy(debug_buf, fn);
 335                 tage->used += strlen(fn) + 1;
 336                 debug_buf += strlen(fn) + 1;
 337         }
 338
 339         __LASSERT(debug_buf == string_buf);
 340
 341         tage->used += needed;
 342         __LASSERT (tage->used <= CFS_PAGE_SIZE);
 343
 344 console:
 345         if (!((mask & D_CANTMASK) != 0 || (mask & libcfs_printk) != 0)) {
 346                 /* no console output requested */
 347                 if (tcd != NULL)
 348                         trace_put_tcd(tcd);
 349                 return 1;
 350         }
 351
 352         if (cdls != NULL) {
 353                 cfs_time_t      t = cdls->cdls_next +
 354                                     cfs_time_seconds(CDEBUG_MAX_LIMIT + 10);
 355                 cfs_duration_t  dmax = cfs_time_seconds(CDEBUG_MAX_LIMIT);
 356
 357                 if (libcfs_console_ratelimit &&
 358                     cdls->cdls_next != 0 &&     /* not first time ever */
 359                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 360                         /* skipping a console message */
 361                         cdls->cdls_count++;
 362                         if (tcd != NULL)
 363                                 trace_put_tcd(tcd);
 364                         return 1;
 365                 }
 366
 367                 if (cfs_time_after(cfs_time_current(), t)) {
 368                         /* last timeout was a long time ago */
 369                         cdls->cdls_delay /= 8;
 370                 } else {
 371                         cdls->cdls_delay *= 2;
 372
 373                         if (cdls->cdls_delay < CFS_TICK)
 374                                 cdls->cdls_delay = CFS_TICK;
 375                         else if (cdls->cdls_delay > dmax)
 376                                 cdls->cdls_delay = dmax;
 377                 }
 378
 379                 /* ensure cdls_next is never zero after it's been seen */
 380                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 381         }
 382
 383         if (tcd != NULL) {
 384                 print_to_console(&header, mask, string_buf, needed, file, fn);
 385                 trace_put_tcd(tcd);
 386         } else {
 387                 string_buf = trace_get_console_buffer();
 388
 389                 needed = 0;
 390                 if (format1 != NULL) {
 391                         va_copy(ap, args);
 392                         needed = vsnprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE, format1, ap);
 393                         va_end(ap);
 394                 }
 395                 if (format2 != NULL) {
 396                         remain = TRACE_CONSOLE_BUFFER_SIZE - needed;
 397                         if (remain > 0) {
 398                                 va_start(ap, format2);
 399                                 needed += vsnprintf(string_buf+needed, remain, format2, ap);
 400                                 va_end(ap);
 401                         }
 402                 }
 403                 print_to_console(&header, mask,
 404                                  string_buf, needed, file, fn);
 405
 406                 trace_put_console_buffer(string_buf);
 407         }
 408
 409         if (cdls != NULL && cdls->cdls_count != 0) {
 410                 string_buf = trace_get_console_buffer();
 411
 412                 needed = snprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE,
 413                          "Skipped %d previous similar message%s\n",
 414                          cdls->cdls_count, (cdls->cdls_count > 1) ? "s" : "");
 415
 416                 print_to_console(&header, mask,
 417                                  string_buf, needed, file, fn);
 418
 419                 trace_put_console_buffer(string_buf);
 420                 cdls->cdls_count = 0;
 421         }
 422
 423         return 0;
 424 }
 425 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 426
 427 void
 428 libcfs_assertion_failed(const char *expr, const char *file,
 429                         const char *func, const int line)
 430 {
 431         libcfs_debug_msg(NULL, 0, D_EMERG, file, func, line,
 432                          "ASSERTION(%s) failed\n", expr);
 433         LBUG();
 434 }
 435 EXPORT_SYMBOL(libcfs_assertion_failed);
 436
 437 void
 438 trace_assertion_failed(const char *str,
 439                        const char *fn, const char *file, int line)
 440 {
 441         struct ptldebug_header hdr;
 442
 443         libcfs_panic_in_progress = 1;
 444         libcfs_catastrophe = 1;
 445         mb();
 446
 447         set_ptldebug_header(&hdr, DEBUG_SUBSYSTEM, D_EMERG, line,
 448                             CDEBUG_STACK());
 449
 450         print_to_console(&hdr, D_EMERG, str, strlen(str), file, fn);
 451
 452         LIBCFS_PANIC("Lustre debug assertion failure\n");
 453
 454         /* not reached */
 455 }
 456
 457 static void
 458 panic_collect_pages(struct page_collection *pc)
 459 {
 460         /* Do the collect_pages job on a single CPU: assumes that all other
 461          * CPUs have been stopped during a panic.  If this isn't true for some
 462          * arch, this will have to be implemented separately in each arch.  */
 463         int                    i;
 464         struct trace_cpu_data *tcd;
 465
 466         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 467
 468         for (i = 0; i < NR_CPUS; i++) {
 469                 tcd = &trace_data[i].tcd;
 470
 471                 list_splice(&tcd->tcd_pages, &pc->pc_pages);
 472                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
 473                 tcd->tcd_cur_pages = 0;
 474
 475                 if (pc->pc_want_daemon_pages) {
 476                         list_splice(&tcd->tcd_daemon_pages, &pc->pc_pages);
 477                         CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
 478                         tcd->tcd_cur_daemon_pages = 0;
 479                 }
 480         }
 481 }
 482
 483 static void collect_pages_on_cpu(void *info)
 484 {
 485         struct trace_cpu_data *tcd;
 486         struct page_collection *pc = info;
 487
 488         tcd = trace_get_tcd();
 489         __LASSERT (tcd != NULL);
 490
 491         spin_lock(&pc->pc_lock);
 492         list_splice(&tcd->tcd_pages, &pc->pc_pages);
 493         CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
 494         tcd->tcd_cur_pages = 0;
 495         if (pc->pc_want_daemon_pages) {
 496                 list_splice(&tcd->tcd_daemon_pages, &pc->pc_pages);
 497                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
 498                 tcd->tcd_cur_daemon_pages = 0;
 499         }
 500         spin_unlock(&pc->pc_lock);
 501
 502         trace_put_tcd(tcd);
 503 }
 504
 505 static void collect_pages(struct page_collection *pc)
 506 {
 507         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 508
 509         if (libcfs_panic_in_progress)
 510                 panic_collect_pages(pc);
 511         else
 512                 trace_call_on_all_cpus(collect_pages_on_cpu, pc);
 513 }
 514
 515 static void put_pages_back_on_cpu(void *info)
 516 {
 517         struct page_collection *pc = info;
 518         struct trace_cpu_data *tcd;
 519         struct list_head *cur_head;
 520         struct trace_page *tage;
 521         struct trace_page *tmp;
 522
 523         tcd = trace_get_tcd();
 524         __LASSERT (tcd != NULL);
 525
 526         cur_head = tcd->tcd_pages.next;
 527
 528         spin_lock(&pc->pc_lock);
 529         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 530
 531                 __LASSERT_TAGE_INVARIANT(tage);
 532
 533                 if (tage->cpu != smp_processor_id())
 534                         continue;
 535
 536                 tage_to_tail(tage, cur_head);
 537                 tcd->tcd_cur_pages++;
 538         }
 539         spin_unlock(&pc->pc_lock);
 540
 541         trace_put_tcd(tcd);
 542 }
 543
 544 static void put_pages_back(struct page_collection *pc)
 545 {
 546         if (!libcfs_panic_in_progress)
 547                 trace_call_on_all_cpus(put_pages_back_on_cpu, pc);
 548 }
 549
 550 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 551  * we have a good amount of data at all times for dumping during an LBUG, even
 552  * if we have been steadily writing (and otherwise discarding) pages via the
 553  * debug daemon. */
 554 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 555                                          struct trace_cpu_data *tcd)
 556 {
 557         struct trace_page *tage;
 558         struct trace_page *tmp;
 559
 560         spin_lock(&pc->pc_lock);
 561         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 562
 563                 __LASSERT_TAGE_INVARIANT(tage);
 564
 565                 if (tage->cpu != smp_processor_id())
 566                         continue;
 567
 568                 tage_to_tail(tage, &tcd->tcd_daemon_pages);
 569                 tcd->tcd_cur_daemon_pages++;
 570
 571                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 572                         struct trace_page *victim;
 573
 574                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 575                         victim = tage_from_list(tcd->tcd_daemon_pages.next);
 576
 577                         __LASSERT_TAGE_INVARIANT(victim);
 578
 579                         list_del(&victim->linkage);
 580                         tage_free(victim);
 581                         tcd->tcd_cur_daemon_pages--;
 582                 }
 583         }
 584         spin_unlock(&pc->pc_lock);
 585 }
 586
 587 static void put_pages_on_daemon_list_on_cpu(void *info)
 588 {
 589         struct trace_cpu_data *tcd;
 590
 591         tcd = trace_get_tcd();
 592         __LASSERT (tcd != NULL);
 593
 594         put_pages_on_tcd_daemon_list(info, tcd);
 595
 596         trace_put_tcd(tcd);
 597 }
 598
 599 static void put_pages_on_daemon_list(struct page_collection *pc)
 600 {
 601         trace_call_on_all_cpus(put_pages_on_daemon_list_on_cpu, pc);
 602 }
 603
 604 void trace_debug_print(void)
 605 {
 606         struct page_collection pc;
 607         struct trace_page *tage;
 608         struct trace_page *tmp;
 609
 610         spin_lock_init(&pc.pc_lock);
 611
 612         pc.pc_want_daemon_pages = 1;
 613         collect_pages(&pc);
 614         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 615                 char *p, *file, *fn;
 616                 cfs_page_t *page;
 617
 618                 __LASSERT_TAGE_INVARIANT(tage);
 619
 620                 page = tage->page;
 621                 p = cfs_page_address(page);
 622                 while (p < ((char *)cfs_page_address(page) + tage->used)) {
 623                         struct ptldebug_header *hdr;
 624                         int len;
 625                         hdr = (void *)p;
 626                         p += sizeof(*hdr);
 627                         file = p;
 628                         p += strlen(file) + 1;
 629                         fn = p;
 630                         p += strlen(fn) + 1;
 631                         len = hdr->ph_len - (p - (char *)hdr);
 632
 633                         print_to_console(hdr, D_EMERG, p, len, file, fn);
 634
 635                         p += len;
 636                 }
 637
 638                 list_del(&tage->linkage);
 639                 tage_free(tage);
 640         }
 641 }
 642
 643 int tracefile_dump_all_pages(char *filename)
 644 {
 645         struct page_collection pc;
 646         cfs_file_t *filp;
 647         struct trace_page *tage;
 648         struct trace_page *tmp;
 649         int rc;
 650
 651         CFS_DECL_MMSPACE;
 652
 653         tracefile_write_lock();
 654
 655         filp = cfs_filp_open(filename,
 656                              O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600, &rc);
 657         if (!filp) {
 658                 if (rc != -EEXIST)
 659                         printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 660                                filename, rc);
 661                 goto out;
 662         }
 663
 664         spin_lock_init(&pc.pc_lock);
 665         pc.pc_want_daemon_pages = 1;
 666         collect_pages(&pc);
 667         if (list_empty(&pc.pc_pages)) {
 668                 rc = 0;
 669                 goto close;
 670         }
 671
 672         /* ok, for now, just write the pages.  in the future we'll be building
 673          * iobufs with the pages and calling generic_direct_IO */
 674         CFS_MMSPACE_OPEN;
 675         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 676
 677                 __LASSERT_TAGE_INVARIANT(tage);
 678
 679                 rc = cfs_filp_write(filp, cfs_page_address(tage->page),
 680                                     tage->used, cfs_filp_poff(filp));
 681                 if (rc != (int)tage->used) {
 682                         printk(KERN_WARNING "wanted to write %u but wrote "
 683                                "%d\n", tage->used, rc);
 684                         put_pages_back(&pc);
 685                         __LASSERT(list_empty(&pc.pc_pages));
 686                         break;
 687                 }
 688                 list_del(&tage->linkage);
 689                 tage_free(tage);
 690         }
 691         CFS_MMSPACE_CLOSE;
 692         rc = cfs_filp_fsync(filp);
 693         if (rc)
 694                 printk(KERN_ERR "sync returns %d\n", rc);
 695  close:
 696         cfs_filp_close(filp);
 697  out:
 698         tracefile_write_unlock();
 699         return rc;
 700 }
 701
 702 void trace_flush_pages(void)
 703 {
 704         struct page_collection pc;
 705         struct trace_page *tage;
 706         struct trace_page *tmp;
 707
 708         spin_lock_init(&pc.pc_lock);
 709
 710         pc.pc_want_daemon_pages = 1;
 711         collect_pages(&pc);
 712         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 713
 714                 __LASSERT_TAGE_INVARIANT(tage);
 715
 716                 list_del(&tage->linkage);
 717                 tage_free(tage);
 718         }
 719 }
 720
 721 int trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 722                         const char *usr_buffer, int usr_buffer_nob)
 723 {
 724         int    nob;
 725
 726         if (usr_buffer_nob > knl_buffer_nob)
 727                 return -EOVERFLOW;
 728
 729         if (copy_from_user((void *)knl_buffer,
 730                            (void *)usr_buffer, usr_buffer_nob))
 731                 return -EFAULT;
 732
 733         nob = strnlen(knl_buffer, usr_buffer_nob);
 734         while (nob-- >= 0)                      /* strip trailing whitespace */
 735                 if (!isspace(knl_buffer[nob]))
 736                         break;
 737
 738         if (nob < 0)                            /* empty string */
 739                 return -EINVAL;
 740
 741         if (nob == knl_buffer_nob)              /* no space to terminate */
 742                 return -EOVERFLOW;
 743
 744         knl_buffer[nob + 1] = 0;                /* terminate */
 745         return 0;
 746 }
 747
 748 int trace_copyout_string(char *usr_buffer, int usr_buffer_nob,
 749                          const char *knl_buffer, char *append)
 750 {
 751         /* NB if 'append' != NULL, it's a single character to append to the
 752          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 753          * terminating zero byte) for sysctl entries */
 754         int   nob = strlen(knl_buffer);
 755
 756         if (nob > usr_buffer_nob)
 757                 nob = usr_buffer_nob;
 758
 759         if (copy_to_user(usr_buffer, knl_buffer, nob))
 760                 return -EFAULT;
 761
 762         if (append != NULL && nob < usr_buffer_nob) {
 763                 if (copy_to_user(usr_buffer + nob, append, 1))
 764                         return -EFAULT;
 765
 766                 nob++;
 767         }
 768
 769         return nob;
 770 }
 771
 772 int trace_allocate_string_buffer(char **str, int nob)
 773 {
 774         if (nob > 2 * CFS_PAGE_SIZE)            /* string must be "sensible" */
 775                 return -EINVAL;
 776
 777         *str = cfs_alloc(nob, CFS_ALLOC_STD | CFS_ALLOC_ZERO);
 778         if (*str == NULL)
 779                 return -ENOMEM;
 780
 781         return 0;
 782 }
 783
 784 void trace_free_string_buffer(char *str, int nob)
 785 {
 786         cfs_free(str);
 787 }
 788
 789 int trace_dump_debug_buffer_usrstr(void *usr_str, int usr_str_nob)
 790 {
 791         char         *str;
 792         int           rc;
 793
 794         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 795         if (rc != 0)
 796                 return rc;
 797
 798         rc = trace_copyin_string(str, usr_str_nob + 1,
 799                                  usr_str, usr_str_nob);
 800         if (rc != 0)
 801                 goto out;
 802
 803 #if !defined(__WINNT__)
 804         if (str[0] != '/') {
 805                 rc = -EINVAL;
 806                 goto out;
 807         }
 808 #endif
 809         rc = tracefile_dump_all_pages(str);
 810 out:
 811         trace_free_string_buffer(str, usr_str_nob + 1);
 812         return rc;
 813 }
 814
 815 int trace_daemon_command(char *str)
 816 {
 817         int       rc = 0;
 818
 819         tracefile_write_lock();
 820
 821         if (strcmp(str, "stop") == 0) {
 822                 trace_stop_thread();
 823                 memset(tracefile, 0, sizeof(tracefile));
 824
 825         } else if (strncmp(str, "size=", 5) == 0) {
 826                 tracefile_size = simple_strtoul(str + 5, NULL, 0);
 827                 if (tracefile_size < 10 || tracefile_size > 20480)
 828                         tracefile_size = TRACEFILE_SIZE;
 829                 else
 830                         tracefile_size <<= 20;
 831
 832         } else if (strlen(str) >= sizeof(tracefile)) {
 833                 rc = -ENAMETOOLONG;
 834 #ifndef __WINNT__
 835         } else if (str[0] != '/') {
 836                 rc = -EINVAL;
 837 #endif
 838         } else {
 839                 strcpy(tracefile, str);
 840
 841                 printk(KERN_INFO "Lustre: debug daemon will attempt to start writing "
 842                        "to %s (%lukB max)\n", tracefile,
 843                        (long)(tracefile_size >> 10));
 844
 845                 trace_start_thread();
 846         }
 847
 848         tracefile_write_unlock();
 849         return rc;
 850 }
 851
 852 int trace_daemon_command_usrstr(void *usr_str, int usr_str_nob)
 853 {
 854         char *str;
 855         int   rc;
 856
 857         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 858         if (rc != 0)
 859                 return rc;
 860
 861         rc = trace_copyin_string(str, usr_str_nob + 1,
 862                                  usr_str, usr_str_nob);
 863         if (rc == 0)
 864                 rc = trace_daemon_command(str);
 865
 866         trace_free_string_buffer(str, usr_str_nob + 1);
 867         return rc;
 868 }
 869
 870 int trace_set_debug_mb(int mb)
 871 {
 872         int i;
 873         int limit = trace_max_debug_mb();
 874
 875         if (mb <= 0)
 876                 return -EINVAL;
 877
 878         if (mb > limit) {
 879                 printk(KERN_ERR "Lustre: Refusing to set debug buffer size to "
 880                        "%dMB - limit is %d\n", mb, limit);
 881                 return -EINVAL;
 882         }
 883
 884         mb /= smp_num_cpus;
 885
 886         tracefile_write_lock();
 887
 888         for (i = 0; i < NR_CPUS; i++) {
 889                 struct trace_cpu_data *tcd = &trace_data[i].tcd;
 890
 891                 tcd->tcd_max_pages = mb << (20 - CFS_PAGE_SHIFT);
 892         }
 893
 894         tracefile_write_unlock();
 895
 896         return 0;
 897 }
 898
 899 int trace_set_debug_mb_usrstr(void *usr_str, int usr_str_nob)
 900 {
 901         char     str[32];
 902         int      rc;
 903
 904         rc = trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 905         if (rc < 0)
 906                 return rc;
 907
 908         return trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 909 }
 910
 911 int trace_get_debug_mb(void)
 912 {
 913         int i;
 914         int total_pages = 0;
 915
 916         tracefile_read_lock();
 917
 918         for (i = 0; i < NR_CPUS; i++) {
 919                 struct trace_cpu_data *tcd = &trace_data[i].tcd;
 920
 921                 total_pages += tcd->tcd_max_pages;
 922         }
 923
 924         tracefile_read_unlock();
 925
 926         return total_pages >> (20 - CFS_PAGE_SHIFT);
 927 }
 928
 929 static int tracefiled(void *arg)
 930 {
 931         struct page_collection pc;
 932         struct tracefiled_ctl *tctl = arg;
 933         struct trace_page *tage;
 934         struct trace_page *tmp;
 935         struct ptldebug_header *hdr;
 936         cfs_file_t *filp;
 937         int rc;
 938
 939         CFS_DECL_MMSPACE;
 940
 941         /* we're started late enough that we pick up init's fs context */
 942         /* this is so broken in uml?  what on earth is going on? */
 943         cfs_daemonize("ktracefiled");
 944
 945         spin_lock_init(&pc.pc_lock);
 946         complete(&tctl->tctl_start);
 947
 948         while (1) {
 949                 cfs_waitlink_t __wait;
 950
 951                 cfs_waitlink_init(&__wait);
 952                 cfs_waitq_add(&tctl->tctl_waitq, &__wait);
 953                 set_current_state(TASK_INTERRUPTIBLE);
 954                 cfs_waitq_timedwait(&__wait, CFS_TASK_INTERRUPTIBLE,
 955                                     cfs_time_seconds(1));
 956                 cfs_waitq_del(&tctl->tctl_waitq, &__wait);
 957
 958                 if (atomic_read(&tctl->tctl_shutdown))
 959                         break;
 960
 961                 pc.pc_want_daemon_pages = 0;
 962                 collect_pages(&pc);
 963                 if (list_empty(&pc.pc_pages))
 964                         continue;
 965
 966                 filp = NULL;
 967                 tracefile_read_lock();
 968                 if (tracefile[0] != 0) {
 969                         filp = cfs_filp_open(tracefile,
 970                                              O_CREAT | O_RDWR | O_LARGEFILE,
 971                                              0600, &rc);
 972                         if (!(filp))
 973                                 printk(KERN_WARNING "couldn't open %s: %d\n",
 974                                        tracefile, rc);
 975                 }
 976                 tracefile_read_unlock();
 977                 if (filp == NULL) {
 978                         put_pages_on_daemon_list(&pc);
 979                         __LASSERT(list_empty(&pc.pc_pages));
 980                         continue;
 981                 }
 982
 983                 CFS_MMSPACE_OPEN;
 984
 985                 /* mark the first header, so we can sort in chunks */
 986                 tage = tage_from_list(pc.pc_pages.next);
 987                 __LASSERT_TAGE_INVARIANT(tage);
 988
 989                 hdr = cfs_page_address(tage->page);
 990                 hdr->ph_flags |= PH_FLAG_FIRST_RECORD;
 991
 992                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 993                         static loff_t f_pos;
 994
 995                         __LASSERT_TAGE_INVARIANT(tage);
 996
 997                         if (f_pos >= (off_t)tracefile_size)
 998                                 f_pos = 0;
 999                         else if (f_pos > cfs_filp_size(filp))
1000                                 f_pos = cfs_filp_size(filp);
1001
1002                         rc = cfs_filp_write(filp, cfs_page_address(tage->page),
1003                                             tage->used, &f_pos);
1004                         if (rc != (int)tage->used) {
1005                                 printk(KERN_WARNING "wanted to write %u but "
1006                                        "wrote %d\n", tage->used, rc);
1007                                 put_pages_back(&pc);
1008                                 __LASSERT(list_empty(&pc.pc_pages));
1009                         }
1010                 }
1011                 CFS_MMSPACE_CLOSE;
1012
1013                 cfs_filp_close(filp);
1014                 put_pages_on_daemon_list(&pc);
1015                 __LASSERT(list_empty(&pc.pc_pages));
1016         }
1017         complete(&tctl->tctl_stop);
1018         return 0;
1019 }
1020
1021 int trace_start_thread(void)
1022 {
1023         struct tracefiled_ctl *tctl = &trace_tctl;
1024         int rc = 0;
1025
1026         mutex_down(&trace_thread_sem);
1027         if (thread_running)
1028                 goto out;
1029
1030         init_completion(&tctl->tctl_start);
1031         init_completion(&tctl->tctl_stop);
1032         cfs_waitq_init(&tctl->tctl_waitq);
1033         atomic_set(&tctl->tctl_shutdown, 0);
1034
1035         if (cfs_kernel_thread(tracefiled, tctl, 0) < 0) {
1036                 rc = -ECHILD;
1037                 goto out;
1038         }
1039
1040         wait_for_completion(&tctl->tctl_start);
1041         thread_running = 1;
1042 out:
1043         mutex_up(&trace_thread_sem);
1044         return rc;
1045 }
1046
1047 void trace_stop_thread(void)
1048 {
1049         struct tracefiled_ctl *tctl = &trace_tctl;
1050
1051         mutex_down(&trace_thread_sem);
1052         if (thread_running) {
1053                 printk(KERN_INFO "Lustre: shutting down debug daemon thread...\n");
1054                 atomic_set(&tctl->tctl_shutdown, 1);
1055                 wait_for_completion(&tctl->tctl_stop);
1056                 thread_running = 0;
1057         }
1058         mutex_up(&trace_thread_sem);
1059 }
1060
1061 int tracefile_init(void)
1062 {
1063         struct trace_cpu_data *tcd;
1064         int                    i;
1065         int                    rc;
1066
1067         rc = tracefile_init_arch();
1068         if (rc != 0)
1069                 return rc;
1070
1071         for (i = 0; i < NR_CPUS; i++) {
1072                 tcd = &trace_data[i].tcd;
1073                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
1074                 CFS_INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1075                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1076                 tcd->tcd_cur_pages = 0;
1077                 tcd->tcd_cur_stock_pages = 0;
1078                 tcd->tcd_cur_daemon_pages = 0;
1079                 tcd->tcd_max_pages = TCD_MAX_PAGES;
1080                 tcd->tcd_shutting_down = 0;
1081                 tcd->tcd_cpu = i;
1082         }
1083
1084         return 0;
1085 }
1086
1087 static void trace_cleanup_on_cpu(void *info)
1088 {
1089         struct trace_cpu_data *tcd;
1090         struct trace_page *tage;
1091         struct trace_page *tmp;
1092
1093         tcd = trace_get_tcd();
1094         __LASSERT (tcd != NULL);
1095
1096         tcd->tcd_shutting_down = 1;
1097
1098         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1099                 __LASSERT_TAGE_INVARIANT(tage);
1100
1101                 list_del(&tage->linkage);
1102                 tage_free(tage);
1103         }
1104         tcd->tcd_cur_pages = 0;
1105
1106         trace_put_tcd(tcd);
1107 }
1108
1109 static void trace_cleanup(void)
1110 {
1111         struct page_collection pc;
1112
1113         CFS_INIT_LIST_HEAD(&pc.pc_pages);
1114         spin_lock_init(&pc.pc_lock);
1115
1116         trace_call_on_all_cpus(trace_cleanup_on_cpu, &pc);
1117
1118         tracefile_fini_arch();
1119 }
1120
1121 void tracefile_exit(void)
1122 {
1123         trace_stop_thread();
1124         trace_cleanup();
1125 }