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