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 long long 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                 if (libcfs_console_ratelimit &&
 354                     cdls->cdls_next != 0 &&     /* not first time ever */
 355                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 356                         /* skipping a console message */
 357                         cdls->cdls_count++;
 358                         if (tcd != NULL)
 359                                 trace_put_tcd(tcd);
 360                         return 1;
 361                 }
 362
 363                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 364                                                        libcfs_console_max_delay
 365                                                        + cfs_time_seconds(10))) {
 366                         /* last timeout was a long time ago */
 367                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 368                 } else {
 369                         cdls->cdls_delay *= libcfs_console_backoff;
 370
 371                         if (cdls->cdls_delay < libcfs_console_min_delay)
 372                                 cdls->cdls_delay = libcfs_console_min_delay;
 373                         else if (cdls->cdls_delay > libcfs_console_max_delay)
 374                                 cdls->cdls_delay = libcfs_console_max_delay;
 375                 }
 376
 377                 /* ensure cdls_next is never zero after it's been seen */
 378                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 379         }
 380
 381         if (tcd != NULL) {
 382                 print_to_console(&header, mask, string_buf, needed, file, fn);
 383                 trace_put_tcd(tcd);
 384         } else {
 385                 string_buf = trace_get_console_buffer();
 386
 387                 needed = 0;
 388                 if (format1 != NULL) {
 389                         va_copy(ap, args);
 390                         needed = vsnprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE, format1, ap);
 391                         va_end(ap);
 392                 }
 393                 if (format2 != NULL) {
 394                         remain = TRACE_CONSOLE_BUFFER_SIZE - needed;
 395                         if (remain > 0) {
 396                                 va_start(ap, format2);
 397                                 needed += vsnprintf(string_buf+needed, remain, format2, ap);
 398                                 va_end(ap);
 399                         }
 400                 }
 401                 print_to_console(&header, mask,
 402                                  string_buf, needed, file, fn);
 403
 404                 trace_put_console_buffer(string_buf);
 405         }
 406
 407         if (cdls != NULL && cdls->cdls_count != 0) {
 408                 string_buf = trace_get_console_buffer();
 409
 410                 needed = snprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE,
 411                          "Skipped %d previous similar message%s\n",
 412                          cdls->cdls_count, (cdls->cdls_count > 1) ? "s" : "");
 413
 414                 print_to_console(&header, mask,
 415                                  string_buf, needed, file, fn);
 416
 417                 trace_put_console_buffer(string_buf);
 418                 cdls->cdls_count = 0;
 419         }
 420
 421         return 0;
 422 }
 423 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 424
 425 void
 426 libcfs_assertion_failed(const char *expr, const char *file,
 427                         const char *func, const int line)
 428 {
 429         libcfs_debug_msg(NULL, 0, D_EMERG, file, func, line,
 430                          "ASSERTION(%s) failed\n", expr);
 431         LBUG();
 432 }
 433 EXPORT_SYMBOL(libcfs_assertion_failed);
 434
 435 void
 436 trace_assertion_failed(const char *str,
 437                        const char *fn, const char *file, int line)
 438 {
 439         struct ptldebug_header hdr;
 440
 441         libcfs_panic_in_progress = 1;
 442         libcfs_catastrophe = 1;
 443         mb();
 444
 445         set_ptldebug_header(&hdr, DEBUG_SUBSYSTEM, D_EMERG, line,
 446                             CDEBUG_STACK());
 447
 448         print_to_console(&hdr, D_EMERG, str, strlen(str), file, fn);
 449
 450         LIBCFS_PANIC("Lustre debug assertion failure\n");
 451
 452         /* not reached */
 453 }
 454
 455 static void
 456 panic_collect_pages(struct page_collection *pc)
 457 {
 458         /* Do the collect_pages job on a single CPU: assumes that all other
 459          * CPUs have been stopped during a panic.  If this isn't true for some
 460          * arch, this will have to be implemented separately in each arch.  */
 461         int                    i;
 462         struct trace_cpu_data *tcd;
 463
 464         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 465
 466         for (i = 0; i < num_possible_cpus(); i++) {
 467                 tcd = &trace_data[i].tcd;
 468
 469                 list_splice(&tcd->tcd_pages, &pc->pc_pages);
 470                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
 471                 tcd->tcd_cur_pages = 0;
 472
 473                 if (pc->pc_want_daemon_pages) {
 474                         list_splice(&tcd->tcd_daemon_pages, &pc->pc_pages);
 475                         CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
 476                         tcd->tcd_cur_daemon_pages = 0;
 477                 }
 478         }
 479 }
 480
 481 static void collect_pages_on_cpu(void *info)
 482 {
 483         struct trace_cpu_data *tcd;
 484         struct page_collection *pc = info;
 485
 486         tcd = trace_get_tcd();
 487         __LASSERT (tcd != NULL);
 488
 489         spin_lock(&pc->pc_lock);
 490         list_splice(&tcd->tcd_pages, &pc->pc_pages);
 491         CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
 492         tcd->tcd_cur_pages = 0;
 493         if (pc->pc_want_daemon_pages) {
 494                 list_splice(&tcd->tcd_daemon_pages, &pc->pc_pages);
 495                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
 496                 tcd->tcd_cur_daemon_pages = 0;
 497         }
 498         spin_unlock(&pc->pc_lock);
 499
 500         trace_put_tcd(tcd);
 501 }
 502
 503 static void collect_pages(struct page_collection *pc)
 504 {
 505         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 506
 507         if (libcfs_panic_in_progress)
 508                 panic_collect_pages(pc);
 509         else
 510                 trace_call_on_all_cpus(collect_pages_on_cpu, pc);
 511 }
 512
 513 static void put_pages_back_on_cpu(void *info)
 514 {
 515         struct page_collection *pc = info;
 516         struct trace_cpu_data *tcd;
 517         struct list_head *cur_head;
 518         struct trace_page *tage;
 519         struct trace_page *tmp;
 520
 521         tcd = trace_get_tcd();
 522         __LASSERT (tcd != NULL);
 523
 524         cur_head = tcd->tcd_pages.next;
 525
 526         spin_lock(&pc->pc_lock);
 527         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 528
 529                 __LASSERT_TAGE_INVARIANT(tage);
 530
 531                 if (tage->cpu != smp_processor_id())
 532                         continue;
 533
 534                 tage_to_tail(tage, cur_head);
 535                 tcd->tcd_cur_pages++;
 536         }
 537         spin_unlock(&pc->pc_lock);
 538
 539         trace_put_tcd(tcd);
 540 }
 541
 542 static void put_pages_back(struct page_collection *pc)
 543 {
 544         if (!libcfs_panic_in_progress)
 545                 trace_call_on_all_cpus(put_pages_back_on_cpu, pc);
 546 }
 547
 548 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 549  * we have a good amount of data at all times for dumping during an LBUG, even
 550  * if we have been steadily writing (and otherwise discarding) pages via the
 551  * debug daemon. */
 552 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 553                                          struct trace_cpu_data *tcd)
 554 {
 555         struct trace_page *tage;
 556         struct trace_page *tmp;
 557
 558         spin_lock(&pc->pc_lock);
 559         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 560
 561                 __LASSERT_TAGE_INVARIANT(tage);
 562
 563                 if (tage->cpu != smp_processor_id())
 564                         continue;
 565
 566                 tage_to_tail(tage, &tcd->tcd_daemon_pages);
 567                 tcd->tcd_cur_daemon_pages++;
 568
 569                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 570                         struct trace_page *victim;
 571
 572                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 573                         victim = tage_from_list(tcd->tcd_daemon_pages.next);
 574
 575                         __LASSERT_TAGE_INVARIANT(victim);
 576
 577                         list_del(&victim->linkage);
 578                         tage_free(victim);
 579                         tcd->tcd_cur_daemon_pages--;
 580                 }
 581         }
 582         spin_unlock(&pc->pc_lock);
 583 }
 584
 585 static void put_pages_on_daemon_list_on_cpu(void *info)
 586 {
 587         struct trace_cpu_data *tcd;
 588
 589         tcd = trace_get_tcd();
 590         __LASSERT (tcd != NULL);
 591
 592         put_pages_on_tcd_daemon_list(info, tcd);
 593
 594         trace_put_tcd(tcd);
 595 }
 596
 597 static void put_pages_on_daemon_list(struct page_collection *pc)
 598 {
 599         trace_call_on_all_cpus(put_pages_on_daemon_list_on_cpu, pc);
 600 }
 601
 602 void trace_debug_print(void)
 603 {
 604         struct page_collection pc;
 605         struct trace_page *tage;
 606         struct trace_page *tmp;
 607
 608         spin_lock_init(&pc.pc_lock);
 609
 610         pc.pc_want_daemon_pages = 1;
 611         collect_pages(&pc);
 612         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 613                 char *p, *file, *fn;
 614                 cfs_page_t *page;
 615
 616                 __LASSERT_TAGE_INVARIANT(tage);
 617
 618                 page = tage->page;
 619                 p = cfs_page_address(page);
 620                 while (p < ((char *)cfs_page_address(page) + tage->used)) {
 621                         struct ptldebug_header *hdr;
 622                         int len;
 623                         hdr = (void *)p;
 624                         p += sizeof(*hdr);
 625                         file = p;
 626                         p += strlen(file) + 1;
 627                         fn = p;
 628                         p += strlen(fn) + 1;
 629                         len = hdr->ph_len - (p - (char *)hdr);
 630
 631                         print_to_console(hdr, D_EMERG, p, len, file, fn);
 632
 633                         p += len;
 634                 }
 635
 636                 list_del(&tage->linkage);
 637                 tage_free(tage);
 638         }
 639 }
 640
 641 int tracefile_dump_all_pages(char *filename)
 642 {
 643         struct page_collection pc;
 644         cfs_file_t *filp;
 645         struct trace_page *tage;
 646         struct trace_page *tmp;
 647         int rc;
 648
 649         CFS_DECL_MMSPACE;
 650
 651         tracefile_write_lock();
 652
 653         filp = cfs_filp_open(filename,
 654                              O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600, &rc);
 655         if (!filp) {
 656                 if (rc != -EEXIST)
 657                         printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 658                                filename, rc);
 659                 goto out;
 660         }
 661
 662         spin_lock_init(&pc.pc_lock);
 663         pc.pc_want_daemon_pages = 1;
 664         collect_pages(&pc);
 665         if (list_empty(&pc.pc_pages)) {
 666                 rc = 0;
 667                 goto close;
 668         }
 669
 670         /* ok, for now, just write the pages.  in the future we'll be building
 671          * iobufs with the pages and calling generic_direct_IO */
 672         CFS_MMSPACE_OPEN;
 673         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 674
 675                 __LASSERT_TAGE_INVARIANT(tage);
 676
 677                 rc = cfs_filp_write(filp, cfs_page_address(tage->page),
 678                                     tage->used, cfs_filp_poff(filp));
 679                 if (rc != (int)tage->used) {
 680                         printk(KERN_WARNING "wanted to write %u but wrote "
 681                                "%d\n", tage->used, rc);
 682                         put_pages_back(&pc);
 683                         __LASSERT(list_empty(&pc.pc_pages));
 684                         break;
 685                 }
 686                 list_del(&tage->linkage);
 687                 tage_free(tage);
 688         }
 689         CFS_MMSPACE_CLOSE;
 690         rc = cfs_filp_fsync(filp);
 691         if (rc)
 692                 printk(KERN_ERR "sync returns %d\n", rc);
 693  close:
 694         cfs_filp_close(filp);
 695  out:
 696         tracefile_write_unlock();
 697         return rc;
 698 }
 699
 700 void trace_flush_pages(void)
 701 {
 702         struct page_collection pc;
 703         struct trace_page *tage;
 704         struct trace_page *tmp;
 705
 706         spin_lock_init(&pc.pc_lock);
 707
 708         pc.pc_want_daemon_pages = 1;
 709         collect_pages(&pc);
 710         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 711
 712                 __LASSERT_TAGE_INVARIANT(tage);
 713
 714                 list_del(&tage->linkage);
 715                 tage_free(tage);
 716         }
 717 }
 718
 719 int trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 720                         const char *usr_buffer, int usr_buffer_nob)
 721 {
 722         int    nob;
 723
 724         if (usr_buffer_nob > knl_buffer_nob)
 725                 return -EOVERFLOW;
 726
 727         if (copy_from_user((void *)knl_buffer,
 728                            (void *)usr_buffer, usr_buffer_nob))
 729                 return -EFAULT;
 730
 731         nob = strnlen(knl_buffer, usr_buffer_nob);
 732         while (nob-- >= 0)                      /* strip trailing whitespace */
 733                 if (!isspace(knl_buffer[nob]))
 734                         break;
 735
 736         if (nob < 0)                            /* empty string */
 737                 return -EINVAL;
 738
 739         if (nob == knl_buffer_nob)              /* no space to terminate */
 740                 return -EOVERFLOW;
 741
 742         knl_buffer[nob + 1] = 0;                /* terminate */
 743         return 0;
 744 }
 745
 746 int trace_copyout_string(char *usr_buffer, int usr_buffer_nob,
 747                          const char *knl_buffer, char *append)
 748 {
 749         /* NB if 'append' != NULL, it's a single character to append to the
 750          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 751          * terminating zero byte) for sysctl entries */
 752         int   nob = strlen(knl_buffer);
 753
 754         if (nob > usr_buffer_nob)
 755                 nob = usr_buffer_nob;
 756
 757         if (copy_to_user(usr_buffer, knl_buffer, nob))
 758                 return -EFAULT;
 759
 760         if (append != NULL && nob < usr_buffer_nob) {
 761                 if (copy_to_user(usr_buffer + nob, append, 1))
 762                         return -EFAULT;
 763
 764                 nob++;
 765         }
 766
 767         return nob;
 768 }
 769
 770 int trace_allocate_string_buffer(char **str, int nob)
 771 {
 772         if (nob > 2 * CFS_PAGE_SIZE)            /* string must be "sensible" */
 773                 return -EINVAL;
 774
 775         *str = cfs_alloc(nob, CFS_ALLOC_STD | CFS_ALLOC_ZERO);
 776         if (*str == NULL)
 777                 return -ENOMEM;
 778
 779         return 0;
 780 }
 781
 782 void trace_free_string_buffer(char *str, int nob)
 783 {
 784         cfs_free(str);
 785 }
 786
 787 int trace_dump_debug_buffer_usrstr(void *usr_str, int usr_str_nob)
 788 {
 789         char         *str;
 790         int           rc;
 791
 792         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 793         if (rc != 0)
 794                 return rc;
 795
 796         rc = trace_copyin_string(str, usr_str_nob + 1,
 797                                  usr_str, usr_str_nob);
 798         if (rc != 0)
 799                 goto out;
 800
 801 #if !defined(__WINNT__)
 802         if (str[0] != '/') {
 803                 rc = -EINVAL;
 804                 goto out;
 805         }
 806 #endif
 807         rc = tracefile_dump_all_pages(str);
 808 out:
 809         trace_free_string_buffer(str, usr_str_nob + 1);
 810         return rc;
 811 }
 812
 813 int trace_daemon_command(char *str)
 814 {
 815         int       rc = 0;
 816
 817         tracefile_write_lock();
 818
 819         if (strcmp(str, "stop") == 0) {
 820                 trace_stop_thread();
 821                 memset(tracefile, 0, sizeof(tracefile));
 822
 823         } else if (strncmp(str, "size=", 5) == 0) {
 824                 tracefile_size = simple_strtoul(str + 5, NULL, 0);
 825                 if (tracefile_size < 10 || tracefile_size > 20480)
 826                         tracefile_size = TRACEFILE_SIZE;
 827                 else
 828                         tracefile_size <<= 20;
 829
 830         } else if (strlen(str) >= sizeof(tracefile)) {
 831                 rc = -ENAMETOOLONG;
 832 #ifndef __WINNT__
 833         } else if (str[0] != '/') {
 834                 rc = -EINVAL;
 835 #endif
 836         } else {
 837                 strcpy(tracefile, str);
 838
 839                 printk(KERN_INFO "Lustre: debug daemon will attempt to start writing "
 840                        "to %s (%lukB max)\n", tracefile,
 841                        (long)(tracefile_size >> 10));
 842
 843                 trace_start_thread();
 844         }
 845
 846         tracefile_write_unlock();
 847         return rc;
 848 }
 849
 850 int trace_daemon_command_usrstr(void *usr_str, int usr_str_nob)
 851 {
 852         char *str;
 853         int   rc;
 854
 855         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 856         if (rc != 0)
 857                 return rc;
 858
 859         rc = trace_copyin_string(str, usr_str_nob + 1,
 860                                  usr_str, usr_str_nob);
 861         if (rc == 0)
 862                 rc = trace_daemon_command(str);
 863
 864         trace_free_string_buffer(str, usr_str_nob + 1);
 865         return rc;
 866 }
 867
 868 int trace_set_debug_mb(int mb)
 869 {
 870         int i;
 871         int limit = trace_max_debug_mb();
 872
 873         if (mb <= 0)
 874                 return -EINVAL;
 875
 876         if (mb > limit) {
 877                 printk(KERN_ERR "Lustre: Refusing to set debug buffer size to "
 878                        "%dMB - limit is %d\n", mb, limit);
 879                 return -EINVAL;
 880         }
 881
 882         mb /= num_possible_cpus();
 883
 884         tracefile_write_lock();
 885
 886         for (i = 0; i < num_possible_cpus(); i++) {
 887                 struct trace_cpu_data *tcd = &trace_data[i].tcd;
 888
 889                 tcd->tcd_max_pages = mb << (20 - CFS_PAGE_SHIFT);
 890         }
 891
 892         tracefile_write_unlock();
 893
 894         return 0;
 895 }
 896
 897 int trace_set_debug_mb_usrstr(void *usr_str, int usr_str_nob)
 898 {
 899         char     str[32];
 900         int      rc;
 901
 902         rc = trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 903         if (rc < 0)
 904                 return rc;
 905
 906         return trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 907 }
 908
 909 int trace_get_debug_mb(void)
 910 {
 911         int i;
 912         int total_pages = 0;
 913
 914         tracefile_read_lock();
 915
 916         for (i = 0; i < num_possible_cpus(); i++) {
 917                 struct trace_cpu_data *tcd = &trace_data[i].tcd;
 918
 919                 total_pages += tcd->tcd_max_pages;
 920         }
 921
 922         tracefile_read_unlock();
 923
 924         return total_pages >> (20 - CFS_PAGE_SHIFT);
 925 }
 926
 927 static int tracefiled(void *arg)
 928 {
 929         struct page_collection pc;
 930         struct tracefiled_ctl *tctl = arg;
 931         struct trace_page *tage;
 932         struct trace_page *tmp;
 933         struct ptldebug_header *hdr;
 934         cfs_file_t *filp;
 935         int rc;
 936
 937         CFS_DECL_MMSPACE;
 938
 939         /* we're started late enough that we pick up init's fs context */
 940         /* this is so broken in uml?  what on earth is going on? */
 941         cfs_daemonize("ktracefiled");
 942
 943         spin_lock_init(&pc.pc_lock);
 944         complete(&tctl->tctl_start);
 945
 946         while (1) {
 947                 cfs_waitlink_t __wait;
 948
 949                 cfs_waitlink_init(&__wait);
 950                 cfs_waitq_add(&tctl->tctl_waitq, &__wait);
 951                 set_current_state(TASK_INTERRUPTIBLE);
 952                 cfs_waitq_timedwait(&__wait, CFS_TASK_INTERRUPTIBLE,
 953                                     cfs_time_seconds(1));
 954                 cfs_waitq_del(&tctl->tctl_waitq, &__wait);
 955
 956                 if (atomic_read(&tctl->tctl_shutdown))
 957                         break;
 958
 959                 pc.pc_want_daemon_pages = 0;
 960                 collect_pages(&pc);
 961                 if (list_empty(&pc.pc_pages))
 962                         continue;
 963
 964                 filp = NULL;
 965                 tracefile_read_lock();
 966                 if (tracefile[0] != 0) {
 967                         filp = cfs_filp_open(tracefile,
 968                                              O_CREAT | O_RDWR | O_LARGEFILE,
 969                                              0600, &rc);
 970                         if (!(filp))
 971                                 printk(KERN_WARNING "couldn't open %s: %d\n",
 972                                        tracefile, rc);
 973                 }
 974                 tracefile_read_unlock();
 975                 if (filp == NULL) {
 976                         put_pages_on_daemon_list(&pc);
 977                         __LASSERT(list_empty(&pc.pc_pages));
 978                         continue;
 979                 }
 980
 981                 CFS_MMSPACE_OPEN;
 982
 983                 /* mark the first header, so we can sort in chunks */
 984                 tage = tage_from_list(pc.pc_pages.next);
 985                 __LASSERT_TAGE_INVARIANT(tage);
 986
 987                 hdr = cfs_page_address(tage->page);
 988                 hdr->ph_flags |= PH_FLAG_FIRST_RECORD;
 989
 990                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 991                         static loff_t f_pos;
 992
 993                         __LASSERT_TAGE_INVARIANT(tage);
 994
 995                         if (f_pos >= (off_t)tracefile_size)
 996                                 f_pos = 0;
 997                         else if (f_pos > cfs_filp_size(filp))
 998                                 f_pos = cfs_filp_size(filp);
 999
1000                         rc = cfs_filp_write(filp, cfs_page_address(tage->page),
1001                                             tage->used, &f_pos);
1002                         if (rc != (int)tage->used) {
1003                                 printk(KERN_WARNING "wanted to write %u but "
1004                                        "wrote %d\n", tage->used, rc);
1005                                 put_pages_back(&pc);
1006                                 __LASSERT(list_empty(&pc.pc_pages));
1007                         }
1008                 }
1009                 CFS_MMSPACE_CLOSE;
1010
1011                 cfs_filp_close(filp);
1012                 put_pages_on_daemon_list(&pc);
1013                 __LASSERT(list_empty(&pc.pc_pages));
1014         }
1015         complete(&tctl->tctl_stop);
1016         return 0;
1017 }
1018
1019 int trace_start_thread(void)
1020 {
1021         struct tracefiled_ctl *tctl = &trace_tctl;
1022         int rc = 0;
1023
1024         mutex_down(&trace_thread_sem);
1025         if (thread_running)
1026                 goto out;
1027
1028         init_completion(&tctl->tctl_start);
1029         init_completion(&tctl->tctl_stop);
1030         cfs_waitq_init(&tctl->tctl_waitq);
1031         atomic_set(&tctl->tctl_shutdown, 0);
1032
1033         if (cfs_kernel_thread(tracefiled, tctl, 0) < 0) {
1034                 rc = -ECHILD;
1035                 goto out;
1036         }
1037
1038         wait_for_completion(&tctl->tctl_start);
1039         thread_running = 1;
1040 out:
1041         mutex_up(&trace_thread_sem);
1042         return rc;
1043 }
1044
1045 void trace_stop_thread(void)
1046 {
1047         struct tracefiled_ctl *tctl = &trace_tctl;
1048
1049         mutex_down(&trace_thread_sem);
1050         if (thread_running) {
1051                 printk(KERN_INFO "Lustre: shutting down debug daemon thread...\n");
1052                 atomic_set(&tctl->tctl_shutdown, 1);
1053                 wait_for_completion(&tctl->tctl_stop);
1054                 thread_running = 0;
1055         }
1056         mutex_up(&trace_thread_sem);
1057 }
1058
1059 int tracefile_init(void)
1060 {
1061         struct trace_cpu_data *tcd;
1062         int                    i;
1063         int                    rc;
1064
1065         rc = tracefile_init_arch();
1066         if (rc != 0)
1067                 return rc;
1068
1069         for (i = 0; i < num_possible_cpus(); i++) {
1070                 tcd = &trace_data[i].tcd;
1071                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
1072                 CFS_INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1073                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1074                 tcd->tcd_cur_pages = 0;
1075                 tcd->tcd_cur_stock_pages = 0;
1076                 tcd->tcd_cur_daemon_pages = 0;
1077                 tcd->tcd_max_pages = TCD_MAX_PAGES;
1078                 tcd->tcd_shutting_down = 0;
1079                 tcd->tcd_cpu = i;
1080         }
1081
1082         return 0;
1083 }
1084
1085 static void trace_cleanup_on_cpu(void *info)
1086 {
1087         struct trace_cpu_data *tcd;
1088         struct trace_page *tage;
1089         struct trace_page *tmp;
1090
1091         tcd = trace_get_tcd();
1092         __LASSERT (tcd != NULL);
1093
1094         tcd->tcd_shutting_down = 1;
1095
1096         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1097                 __LASSERT_TAGE_INVARIANT(tage);
1098
1099                 list_del(&tage->linkage);
1100                 tage_free(tage);
1101         }
1102         tcd->tcd_cur_pages = 0;
1103
1104         trace_put_tcd(tcd);
1105 }
1106
1107 static void trace_cleanup(void)
1108 {
1109         struct page_collection pc;
1110
1111         CFS_INIT_LIST_HEAD(&pc.pc_pages);
1112         spin_lock_init(&pc.pc_lock);
1113
1114         trace_call_on_all_cpus(trace_cleanup_on_cpu, &pc);
1115
1116         tracefile_fini_arch();
1117 }
1118
1119 void tracefile_exit(void)
1120 {
1121         trace_stop_thread();
1122         trace_cleanup();
1123 }