lnet/libcfs/tracefile.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/libcfs/tracefile.c
  37  *
  38  * Author: Zach Brown <zab@clusterfs.com>
  39  * Author: Phil Schwan <phil@clusterfs.com>
  40  */
  41
  42
  43 #define DEBUG_SUBSYSTEM S_LNET
  44 #define LUSTRE_TRACEFILE_PRIVATE
  45 #include "tracefile.h"
  46
  47 #include <libcfs/kp30.h>
  48 #include <libcfs/libcfs.h>
  49
  50 /* XXX move things up to the top, comment */
  51 union trace_data_union (*trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  52
  53 char tracefile[TRACEFILE_NAME_SIZE];
  54 long long tracefile_size = TRACEFILE_SIZE;
  55 static struct tracefiled_ctl trace_tctl;
  56 struct semaphore trace_thread_sem;
  57 static int thread_running = 0;
  58
  59 atomic_t tage_allocated = ATOMIC_INIT(0);
  60
  61 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  62                                          struct trace_cpu_data *tcd);
  63
  64 static inline struct trace_page *tage_from_list(struct list_head *list)
  65 {
  66         return list_entry(list, struct trace_page, linkage);
  67 }
  68
  69 static struct trace_page *tage_alloc(int gfp)
  70 {
  71         cfs_page_t        *page;
  72         struct trace_page *tage;
  73
  74         /* My caller is trying to free memory */
  75         if (!cfs_in_interrupt() && libcfs_memory_pressure_get())
  76                 return NULL;
  77
  78         /*
  79          * Don't spam console with allocation failures: they will be reported
  80          * by upper layer anyway.
  81          */
  82         gfp |= CFS_ALLOC_NOWARN;
  83         page = cfs_alloc_page(gfp);
  84         if (page == NULL)
  85                 return NULL;
  86
  87         tage = cfs_alloc(sizeof(*tage), gfp);
  88         if (tage == NULL) {
  89                 cfs_free_page(page);
  90                 return NULL;
  91         }
  92
  93         tage->page = page;
  94         atomic_inc(&tage_allocated);
  95         return tage;
  96 }
  97
  98 static void tage_free(struct trace_page *tage)
  99 {
 100         __LASSERT(tage != NULL);
 101         __LASSERT(tage->page != NULL);
 102
 103         cfs_free_page(tage->page);
 104         cfs_free(tage);
 105         atomic_dec(&tage_allocated);
 106 }
 107
 108 static void tage_to_tail(struct trace_page *tage, struct list_head *queue)
 109 {
 110         __LASSERT(tage != NULL);
 111         __LASSERT(queue != NULL);
 112
 113         list_move_tail(&tage->linkage, queue);
 114 }
 115
 116 int trace_refill_stock(struct trace_cpu_data *tcd, int gfp,
 117                        struct list_head *stock)
 118 {
 119         int i;
 120
 121         /*
 122          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 123          * from here: this will lead to infinite recursion.
 124          */
 125
 126         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++ i) {
 127                 struct trace_page *tage;
 128
 129                 tage = tage_alloc(gfp);
 130                 if (tage == NULL)
 131                         break;
 132                 list_add_tail(&tage->linkage, stock);
 133         }
 134         return i;
 135 }
 136
 137 /* return a page that has 'len' bytes left at the end */
 138 static struct trace_page *trace_get_tage_try(struct trace_cpu_data *tcd,
 139                                              unsigned long len)
 140 {
 141         struct trace_page *tage;
 142
 143         if (tcd->tcd_cur_pages > 0) {
 144                 __LASSERT(!list_empty(&tcd->tcd_pages));
 145                 tage = tage_from_list(tcd->tcd_pages.prev);
 146                 if (tage->used + len <= CFS_PAGE_SIZE)
 147                         return tage;
 148         }
 149
 150         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 151                 if (tcd->tcd_cur_stock_pages > 0) {
 152                         tage = tage_from_list(tcd->tcd_stock_pages.prev);
 153                         -- tcd->tcd_cur_stock_pages;
 154                         list_del_init(&tage->linkage);
 155                 } else {
 156                         tage = tage_alloc(CFS_ALLOC_ATOMIC);
 157                         if (tage == NULL) {
 158                                 printk(KERN_WARNING
 159                                        "failure to allocate a tage (%ld)\n",
 160                                        tcd->tcd_cur_pages);
 161                                 return NULL;
 162                         }
 163                 }
 164
 165                 tage->used = 0;
 166                 tage->cpu = smp_processor_id();
 167                 tage->type = tcd->tcd_type;
 168                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 169                 tcd->tcd_cur_pages++;
 170
 171                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 172                         struct tracefiled_ctl *tctl = &trace_tctl;
 173                         /*
 174                          * wake up tracefiled to process some pages.
 175                          */
 176                         cfs_waitq_signal(&tctl->tctl_waitq);
 177                 }
 178                 return tage;
 179         }
 180         return NULL;
 181 }
 182
 183 static void tcd_shrink(struct trace_cpu_data *tcd)
 184 {
 185         int pgcount = tcd->tcd_cur_pages / 10;
 186         struct page_collection pc;
 187         struct trace_page *tage;
 188         struct trace_page *tmp;
 189
 190         /*
 191          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 192          * from here: this will lead to infinite recursion.
 193          */
 194
 195         if (printk_ratelimit())
 196                 printk(KERN_WARNING "debug daemon buffer overflowed; "
 197                        "discarding  10%% of pages (%d of %ld)\n",
 198                        pgcount + 1, tcd->tcd_cur_pages);
 199
 200         CFS_INIT_LIST_HEAD(&pc.pc_pages);
 201         spin_lock_init(&pc.pc_lock);
 202
 203         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 204                 if (pgcount-- == 0)
 205                         break;
 206
 207                 list_move_tail(&tage->linkage, &pc.pc_pages);
 208                 tcd->tcd_cur_pages--;
 209         }
 210         put_pages_on_tcd_daemon_list(&pc, tcd);
 211 }
 212
 213 /* return a page that has 'len' bytes left at the end */
 214 static struct trace_page *trace_get_tage(struct trace_cpu_data *tcd,
 215                                          unsigned long len)
 216 {
 217         struct trace_page *tage;
 218
 219         /*
 220          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 221          * from here: this will lead to infinite recursion.
 222          */
 223
 224         if (len > CFS_PAGE_SIZE) {
 225                 printk(KERN_ERR
 226                        "cowardly refusing to write %lu bytes in a page\n", len);
 227                 return NULL;
 228         }
 229
 230         tage = trace_get_tage_try(tcd, len);
 231         if (tage != NULL)
 232                 return tage;
 233         if (thread_running)
 234                 tcd_shrink(tcd);
 235         if (tcd->tcd_cur_pages > 0) {
 236                 tage = tage_from_list(tcd->tcd_pages.next);
 237                 tage->used = 0;
 238                 tage_to_tail(tage, &tcd->tcd_pages);
 239         }
 240         return tage;
 241 }
 242
 243 int libcfs_debug_vmsg2(cfs_debug_limit_state_t *cdls, int subsys, int mask,
 244                        const char *file, const char *fn, const int line,
 245                        const char *format1, va_list args,
 246                        const char *format2, ...)
 247 {
 248         struct trace_cpu_data   *tcd = NULL;
 249         struct ptldebug_header   header;
 250         struct trace_page       *tage;
 251         /* string_buf is used only if tcd != NULL, and is always set then */
 252         char                    *string_buf = NULL;
 253         char                    *debug_buf;
 254         int                      known_size;
 255         int                      needed = 85; /* average message length */
 256         int                      max_nob;
 257         va_list                  ap;
 258         int                      depth;
 259         int                      i;
 260         int                      remain;
 261
 262         if (strchr(file, '/'))
 263                 file = strrchr(file, '/') + 1;
 264
 265         tcd = trace_get_tcd();
 266
 267         /* fs_trace_get_tcd() grabs a lock, which disables preemption and
 268          * pins us to a particular CPU.  This avoids an smp_processor_id()
 269          * warning on Linux when debugging is enabled. */
 270         set_ptldebug_header(&header, subsys, mask, line, CDEBUG_STACK());
 271
 272         if (tcd == NULL)                /* arch may not log in IRQ context */
 273                 goto console;
 274
 275         if (tcd->tcd_shutting_down) {
 276                 trace_put_tcd(tcd);
 277                 tcd = NULL;
 278                 goto console;
 279         }
 280
 281         depth = __current_nesting_level();
 282         known_size = strlen(file) + 1 + depth;
 283         if (fn)
 284                 known_size += strlen(fn) + 1;
 285
 286         if (libcfs_debug_binary)
 287                 known_size += sizeof(header);
 288
 289         /*/
 290          * '2' used because vsnprintf return real size required for output
 291          * _without_ terminating NULL.
 292          * if needed is to small for this format.
 293          */
 294         for (i=0;i<2;i++) {
 295                 tage = trace_get_tage(tcd, needed + known_size + 1);
 296                 if (tage == NULL) {
 297                         if (needed + known_size > CFS_PAGE_SIZE)
 298                                 mask |= D_ERROR;
 299
 300                         trace_put_tcd(tcd);
 301                         tcd = NULL;
 302                         goto console;
 303                 }
 304
 305                 string_buf = (char *)cfs_page_address(tage->page)+tage->used+known_size;
 306
 307                 max_nob = CFS_PAGE_SIZE - tage->used - known_size;
 308                 if (max_nob <= 0) {
 309                         printk(KERN_EMERG "negative max_nob: %i\n", max_nob);
 310                         mask |= D_ERROR;
 311                         trace_put_tcd(tcd);
 312                         tcd = NULL;
 313                         goto console;
 314                 }
 315
 316                 needed = 0;
 317                 if (format1) {
 318                         va_copy(ap, args);
 319                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 320                         va_end(ap);
 321                 }
 322
 323                 if (format2) {
 324                         remain = max_nob - needed;
 325                         if (remain < 0)
 326                                 remain = 0;
 327
 328                         va_start(ap, format2);
 329                         needed += vsnprintf(string_buf + needed, remain,
 330                                             format2, ap);
 331                         va_end(ap);
 332                 }
 333
 334                 if (needed < max_nob) /* well. printing ok.. */
 335                         break;
 336         }
 337
 338         if (*(string_buf+needed-1) != '\n')
 339                 printk(KERN_INFO "format at %s:%d:%s doesn't end in newline\n",
 340                        file, line, fn);
 341
 342         header.ph_len = known_size + needed;
 343         debug_buf = (char *)cfs_page_address(tage->page) + tage->used;
 344
 345         if (libcfs_debug_binary) {
 346                 memcpy(debug_buf, &header, sizeof(header));
 347                 tage->used += sizeof(header);
 348                 debug_buf += sizeof(header);
 349         }
 350
 351         /* indent message according to the nesting level */
 352         while (depth-- > 0) {
 353                 *(debug_buf++) = '.';
 354                 ++ tage->used;
 355         }
 356
 357         strcpy(debug_buf, file);
 358         tage->used += strlen(file) + 1;
 359         debug_buf += strlen(file) + 1;
 360
 361         if (fn) {
 362                 strcpy(debug_buf, fn);
 363                 tage->used += strlen(fn) + 1;
 364                 debug_buf += strlen(fn) + 1;
 365         }
 366
 367         __LASSERT(debug_buf == string_buf);
 368
 369         tage->used += needed;
 370         __LASSERT (tage->used <= CFS_PAGE_SIZE);
 371
 372 console:
 373         if ((mask & libcfs_printk) == 0) {
 374                 /* no console output requested */
 375                 if (tcd != NULL)
 376                         trace_put_tcd(tcd);
 377                 return 1;
 378         }
 379
 380         if (cdls != NULL) {
 381                 if (libcfs_console_ratelimit &&
 382                     cdls->cdls_next != 0 &&     /* not first time ever */
 383                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 384                         /* skipping a console message */
 385                         cdls->cdls_count++;
 386                         if (tcd != NULL)
 387                                 trace_put_tcd(tcd);
 388                         return 1;
 389                 }
 390
 391                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 392                                                        libcfs_console_max_delay
 393                                                        + cfs_time_seconds(10))) {
 394                         /* last timeout was a long time ago */
 395                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 396                 } else {
 397                         cdls->cdls_delay *= libcfs_console_backoff;
 398
 399                         if (cdls->cdls_delay < libcfs_console_min_delay)
 400                                 cdls->cdls_delay = libcfs_console_min_delay;
 401                         else if (cdls->cdls_delay > libcfs_console_max_delay)
 402                                 cdls->cdls_delay = libcfs_console_max_delay;
 403                 }
 404
 405                 /* ensure cdls_next is never zero after it's been seen */
 406                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 407         }
 408
 409         if (tcd != NULL) {
 410                 print_to_console(&header, mask, string_buf, needed, file, fn);
 411                 trace_put_tcd(tcd);
 412         } else {
 413                 string_buf = trace_get_console_buffer();
 414
 415                 needed = 0;
 416                 if (format1 != NULL) {
 417                         va_copy(ap, args);
 418                         needed = vsnprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE, format1, ap);
 419                         va_end(ap);
 420                 }
 421                 if (format2 != NULL) {
 422                         remain = TRACE_CONSOLE_BUFFER_SIZE - needed;
 423                         if (remain > 0) {
 424                                 va_start(ap, format2);
 425                                 needed += vsnprintf(string_buf+needed, remain, format2, ap);
 426                                 va_end(ap);
 427                         }
 428                 }
 429                 print_to_console(&header, mask,
 430                                  string_buf, needed, file, fn);
 431
 432                 trace_put_console_buffer(string_buf);
 433         }
 434
 435         if (cdls != NULL && cdls->cdls_count != 0) {
 436                 string_buf = trace_get_console_buffer();
 437
 438                 needed = snprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE,
 439                          "Skipped %d previous similar message%s\n",
 440                          cdls->cdls_count, (cdls->cdls_count > 1) ? "s" : "");
 441
 442                 print_to_console(&header, mask,
 443                                  string_buf, needed, file, fn);
 444
 445                 trace_put_console_buffer(string_buf);
 446                 cdls->cdls_count = 0;
 447         }
 448
 449         return 0;
 450 }
 451 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 452
 453 void
 454 libcfs_assertion_failed(const char *expr, const char *file,
 455                         const char *func, const int line)
 456 {
 457         libcfs_debug_msg(NULL, 0, D_EMERG, file, func, line,
 458                          "ASSERTION(%s) failed\n", expr);
 459         lbug_with_loc(file, func, line);
 460 }
 461 EXPORT_SYMBOL(libcfs_assertion_failed);
 462
 463 void
 464 trace_assertion_failed(const char *str,
 465                        const char *fn, const char *file, int line)
 466 {
 467         struct ptldebug_header hdr;
 468
 469         libcfs_panic_in_progress = 1;
 470         libcfs_catastrophe = 1;
 471         mb();
 472
 473         set_ptldebug_header(&hdr, DEBUG_SUBSYSTEM, D_EMERG, line,
 474                             CDEBUG_STACK());
 475
 476         print_to_console(&hdr, D_EMERG, str, strlen(str), file, fn);
 477
 478         LIBCFS_PANIC("Lustre debug assertion failure\n");
 479
 480         /* not reached */
 481 }
 482
 483 static void
 484 panic_collect_pages(struct page_collection *pc)
 485 {
 486         /* Do the collect_pages job on a single CPU: assumes that all other
 487          * CPUs have been stopped during a panic.  If this isn't true for some
 488          * arch, this will have to be implemented separately in each arch.  */
 489         int                    i;
 490         int                    j;
 491         struct trace_cpu_data *tcd;
 492
 493         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 494
 495         tcd_for_each(tcd, i, j) {
 496                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 497                 tcd->tcd_cur_pages = 0;
 498
 499                 if (pc->pc_want_daemon_pages) {
 500                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 501                         tcd->tcd_cur_daemon_pages = 0;
 502                 }
 503         }
 504 }
 505
 506 static void collect_pages_on_all_cpus(struct page_collection *pc)
 507 {
 508         struct trace_cpu_data *tcd;
 509         int i, cpu;
 510
 511         spin_lock(&pc->pc_lock);
 512         for_each_possible_cpu(cpu) {
 513                 tcd_for_each_type_lock(tcd, i, cpu) {
 514                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 515                         tcd->tcd_cur_pages = 0;
 516                         if (pc->pc_want_daemon_pages) {
 517                                 list_splice_init(&tcd->tcd_daemon_pages,
 518                                                  &pc->pc_pages);
 519                                 tcd->tcd_cur_daemon_pages = 0;
 520                         }
 521                 }
 522         }
 523         spin_unlock(&pc->pc_lock);
 524 }
 525
 526 static void collect_pages(struct page_collection *pc)
 527 {
 528         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 529
 530         if (libcfs_panic_in_progress)
 531                 panic_collect_pages(pc);
 532         else
 533                 collect_pages_on_all_cpus(pc);
 534 }
 535
 536 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 537 {
 538         struct trace_cpu_data *tcd;
 539         struct list_head *cur_head;
 540         struct trace_page *tage;
 541         struct trace_page *tmp;
 542         int i, cpu;
 543
 544         spin_lock(&pc->pc_lock);
 545         for_each_possible_cpu(cpu) {
 546                 tcd_for_each_type_lock(tcd, i, cpu) {
 547                         cur_head = tcd->tcd_pages.next;
 548
 549                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 550                                                  linkage) {
 551
 552                                 __LASSERT_TAGE_INVARIANT(tage);
 553
 554                                 if (tage->cpu != cpu || tage->type != i)
 555                                         continue;
 556
 557                                 tage_to_tail(tage, cur_head);
 558                                 tcd->tcd_cur_pages++;
 559                         }
 560                 }
 561         }
 562         spin_unlock(&pc->pc_lock);
 563 }
 564
 565 static void put_pages_back(struct page_collection *pc)
 566 {
 567         if (!libcfs_panic_in_progress)
 568                 put_pages_back_on_all_cpus(pc);
 569 }
 570
 571 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 572  * we have a good amount of data at all times for dumping during an LBUG, even
 573  * if we have been steadily writing (and otherwise discarding) pages via the
 574  * debug daemon. */
 575 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 576                                          struct trace_cpu_data *tcd)
 577 {
 578         struct trace_page *tage;
 579         struct trace_page *tmp;
 580
 581         spin_lock(&pc->pc_lock);
 582         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 583
 584                 __LASSERT_TAGE_INVARIANT(tage);
 585
 586                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 587                         continue;
 588
 589                 tage_to_tail(tage, &tcd->tcd_daemon_pages);
 590                 tcd->tcd_cur_daemon_pages++;
 591
 592                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 593                         struct trace_page *victim;
 594
 595                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 596                         victim = tage_from_list(tcd->tcd_daemon_pages.next);
 597
 598                         __LASSERT_TAGE_INVARIANT(victim);
 599
 600                         list_del(&victim->linkage);
 601                         tage_free(victim);
 602                         tcd->tcd_cur_daemon_pages--;
 603                 }
 604         }
 605         spin_unlock(&pc->pc_lock);
 606 }
 607
 608 static void put_pages_on_daemon_list(struct page_collection *pc)
 609 {
 610         struct trace_cpu_data *tcd;
 611         int i, cpu;
 612
 613         for_each_possible_cpu(cpu) {
 614                 tcd_for_each_type_lock(tcd, i, cpu)
 615                         put_pages_on_tcd_daemon_list(pc, tcd);
 616         }
 617 }
 618
 619 void trace_debug_print(void)
 620 {
 621         struct page_collection pc;
 622         struct trace_page *tage;
 623         struct trace_page *tmp;
 624
 625         spin_lock_init(&pc.pc_lock);
 626
 627         pc.pc_want_daemon_pages = 1;
 628         collect_pages(&pc);
 629         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 630                 char *p, *file, *fn;
 631                 cfs_page_t *page;
 632
 633                 __LASSERT_TAGE_INVARIANT(tage);
 634
 635                 page = tage->page;
 636                 p = cfs_page_address(page);
 637                 while (p < ((char *)cfs_page_address(page) + tage->used)) {
 638                         struct ptldebug_header *hdr;
 639                         int len;
 640                         hdr = (void *)p;
 641                         p += sizeof(*hdr);
 642                         file = p;
 643                         p += strlen(file) + 1;
 644                         fn = p;
 645                         p += strlen(fn) + 1;
 646                         len = hdr->ph_len - (p - (char *)hdr);
 647
 648                         print_to_console(hdr, D_EMERG, p, len, file, fn);
 649
 650                         p += len;
 651                 }
 652
 653                 list_del(&tage->linkage);
 654                 tage_free(tage);
 655         }
 656 }
 657
 658 int tracefile_dump_all_pages(char *filename)
 659 {
 660         struct page_collection pc;
 661         cfs_file_t *filp;
 662         struct trace_page *tage;
 663         struct trace_page *tmp;
 664         int rc;
 665
 666         CFS_DECL_MMSPACE;
 667
 668         tracefile_write_lock();
 669
 670         filp = cfs_filp_open(filename,
 671                              O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600, &rc);
 672         if (!filp) {
 673                 if (rc != -EEXIST)
 674                         printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 675                                filename, rc);
 676                 goto out;
 677         }
 678
 679         spin_lock_init(&pc.pc_lock);
 680         pc.pc_want_daemon_pages = 1;
 681         collect_pages(&pc);
 682         if (list_empty(&pc.pc_pages)) {
 683                 rc = 0;
 684                 goto close;
 685         }
 686
 687         /* ok, for now, just write the pages.  in the future we'll be building
 688          * iobufs with the pages and calling generic_direct_IO */
 689         CFS_MMSPACE_OPEN;
 690         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 691
 692                 __LASSERT_TAGE_INVARIANT(tage);
 693
 694                 rc = cfs_filp_write(filp, cfs_page_address(tage->page),
 695                                     tage->used, cfs_filp_poff(filp));
 696                 if (rc != (int)tage->used) {
 697                         printk(KERN_WARNING "wanted to write %u but wrote "
 698                                "%d\n", tage->used, rc);
 699                         put_pages_back(&pc);
 700                         __LASSERT(list_empty(&pc.pc_pages));
 701                         break;
 702                 }
 703                 list_del(&tage->linkage);
 704                 tage_free(tage);
 705         }
 706         CFS_MMSPACE_CLOSE;
 707         rc = cfs_filp_fsync(filp);
 708         if (rc)
 709                 printk(KERN_ERR "sync returns %d\n", rc);
 710  close:
 711         cfs_filp_close(filp);
 712  out:
 713         tracefile_write_unlock();
 714         return rc;
 715 }
 716
 717 void trace_flush_pages(void)
 718 {
 719         struct page_collection pc;
 720         struct trace_page *tage;
 721         struct trace_page *tmp;
 722
 723         spin_lock_init(&pc.pc_lock);
 724
 725         pc.pc_want_daemon_pages = 1;
 726         collect_pages(&pc);
 727         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 728
 729                 __LASSERT_TAGE_INVARIANT(tage);
 730
 731                 list_del(&tage->linkage);
 732                 tage_free(tage);
 733         }
 734 }
 735
 736 int trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 737                         const char *usr_buffer, int usr_buffer_nob)
 738 {
 739         int    nob;
 740
 741         if (usr_buffer_nob > knl_buffer_nob)
 742                 return -EOVERFLOW;
 743
 744         if (copy_from_user((void *)knl_buffer,
 745                            (void *)usr_buffer, usr_buffer_nob))
 746                 return -EFAULT;
 747
 748         nob = strnlen(knl_buffer, usr_buffer_nob);
 749         while (nob-- >= 0)                      /* strip trailing whitespace */
 750                 if (!isspace(knl_buffer[nob]))
 751                         break;
 752
 753         if (nob < 0)                            /* empty string */
 754                 return -EINVAL;
 755
 756         if (nob == knl_buffer_nob)              /* no space to terminate */
 757                 return -EOVERFLOW;
 758
 759         knl_buffer[nob + 1] = 0;                /* terminate */
 760         return 0;
 761 }
 762
 763 int trace_copyout_string(char *usr_buffer, int usr_buffer_nob,
 764                          const char *knl_buffer, char *append)
 765 {
 766         /* NB if 'append' != NULL, it's a single character to append to the
 767          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 768          * terminating zero byte) for sysctl entries */
 769         int   nob = strlen(knl_buffer);
 770
 771         if (nob > usr_buffer_nob)
 772                 nob = usr_buffer_nob;
 773
 774         if (copy_to_user(usr_buffer, knl_buffer, nob))
 775                 return -EFAULT;
 776
 777         if (append != NULL && nob < usr_buffer_nob) {
 778                 if (copy_to_user(usr_buffer + nob, append, 1))
 779                         return -EFAULT;
 780
 781                 nob++;
 782         }
 783
 784         return nob;
 785 }
 786 EXPORT_SYMBOL(trace_copyout_string);
 787
 788 int trace_allocate_string_buffer(char **str, int nob)
 789 {
 790         if (nob > 2 * CFS_PAGE_SIZE)            /* string must be "sensible" */
 791                 return -EINVAL;
 792
 793         *str = cfs_alloc(nob, CFS_ALLOC_STD | CFS_ALLOC_ZERO);
 794         if (*str == NULL)
 795                 return -ENOMEM;
 796
 797         return 0;
 798 }
 799
 800 void trace_free_string_buffer(char *str, int nob)
 801 {
 802         cfs_free(str);
 803 }
 804
 805 int trace_dump_debug_buffer_usrstr(void *usr_str, int usr_str_nob)
 806 {
 807         char         *str;
 808         int           rc;
 809
 810         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 811         if (rc != 0)
 812                 return rc;
 813
 814         rc = trace_copyin_string(str, usr_str_nob + 1,
 815                                  usr_str, usr_str_nob);
 816         if (rc != 0)
 817                 goto out;
 818
 819 #if !defined(__WINNT__)
 820         if (str[0] != '/') {
 821                 rc = -EINVAL;
 822                 goto out;
 823         }
 824 #endif
 825         rc = tracefile_dump_all_pages(str);
 826 out:
 827         trace_free_string_buffer(str, usr_str_nob + 1);
 828         return rc;
 829 }
 830
 831 int trace_daemon_command(char *str)
 832 {
 833         int       rc = 0;
 834
 835         tracefile_write_lock();
 836
 837         if (strcmp(str, "stop") == 0) {
 838                 tracefile_write_unlock();
 839                 trace_stop_thread();
 840                 tracefile_write_lock();
 841                 memset(tracefile, 0, sizeof(tracefile));
 842
 843         } else if (strncmp(str, "size=", 5) == 0) {
 844                 tracefile_size = simple_strtoul(str + 5, NULL, 0);
 845                 if (tracefile_size < 10 || tracefile_size > 20480)
 846                         tracefile_size = TRACEFILE_SIZE;
 847                 else
 848                         tracefile_size <<= 20;
 849
 850         } else if (strlen(str) >= sizeof(tracefile)) {
 851                 rc = -ENAMETOOLONG;
 852 #ifndef __WINNT__
 853         } else if (str[0] != '/') {
 854                 rc = -EINVAL;
 855 #endif
 856         } else {
 857                 strcpy(tracefile, str);
 858
 859                 printk(KERN_INFO "Lustre: debug daemon will attempt to start writing "
 860                        "to %s (%lukB max)\n", tracefile,
 861                        (long)(tracefile_size >> 10));
 862
 863                 trace_start_thread();
 864         }
 865
 866         tracefile_write_unlock();
 867         return rc;
 868 }
 869
 870 int trace_daemon_command_usrstr(void *usr_str, int usr_str_nob)
 871 {
 872         char *str;
 873         int   rc;
 874
 875         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 876         if (rc != 0)
 877                 return rc;
 878
 879         rc = trace_copyin_string(str, usr_str_nob + 1,
 880                                  usr_str, usr_str_nob);
 881         if (rc == 0)
 882                 rc = trace_daemon_command(str);
 883
 884         trace_free_string_buffer(str, usr_str_nob + 1);
 885
 886         return rc;
 887 }
 888
 889 int trace_set_debug_mb(int mb)
 890 {
 891         int i;
 892         int j;
 893         int pages;
 894         int limit = trace_max_debug_mb();
 895         struct trace_cpu_data *tcd;
 896
 897         if (mb < num_possible_cpus()) {
 898                 printk(KERN_ERR "Cannot set debug_mb to %d, "
 899                        "the value should be >= %d\n",
 900                        mb, num_possible_cpus());
 901                 return -EINVAL;
 902         }
 903
 904         if (mb > limit) {
 905                 printk(KERN_ERR "Lustre: Refusing to set debug buffer size to "
 906                        "%dMB - limit is %d\n", mb, limit);
 907                 return -EINVAL;
 908         }
 909
 910         mb /= num_possible_cpus();
 911         pages = mb << (20 - CFS_PAGE_SHIFT);
 912
 913         tracefile_write_lock();
 914
 915         tcd_for_each(tcd, i, j)
 916                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 917
 918         tracefile_write_unlock();
 919
 920         return 0;
 921 }
 922
 923 int trace_set_debug_mb_usrstr(void *usr_str, int usr_str_nob)
 924 {
 925         char     str[32];
 926         int      rc;
 927
 928         rc = trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 929         if (rc < 0)
 930                 return rc;
 931
 932         return trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 933 }
 934
 935 int trace_get_debug_mb(void)
 936 {
 937         int i;
 938         int j;
 939         struct trace_cpu_data *tcd;
 940         int total_pages = 0;
 941
 942         tracefile_read_lock();
 943
 944         tcd_for_each(tcd, i, j)
 945                 total_pages += tcd->tcd_max_pages;
 946
 947         tracefile_read_unlock();
 948
 949         return (total_pages >> (20 - CFS_PAGE_SHIFT)) + 1;
 950 }
 951
 952 static int tracefiled(void *arg)
 953 {
 954         struct page_collection pc;
 955         struct tracefiled_ctl *tctl = arg;
 956         struct trace_page *tage;
 957         struct trace_page *tmp;
 958         struct ptldebug_header *hdr;
 959         cfs_file_t *filp;
 960         int last_loop = 0;
 961         int rc;
 962
 963         CFS_DECL_MMSPACE;
 964
 965         /* we're started late enough that we pick up init's fs context */
 966         /* this is so broken in uml?  what on earth is going on? */
 967         cfs_daemonize("ktracefiled");
 968
 969         spin_lock_init(&pc.pc_lock);
 970         complete(&tctl->tctl_start);
 971
 972         while (1) {
 973                 cfs_waitlink_t __wait;
 974
 975                 pc.pc_want_daemon_pages = 0;
 976                 collect_pages(&pc);
 977                 if (list_empty(&pc.pc_pages))
 978                         goto end_loop;
 979
 980                 filp = NULL;
 981                 tracefile_read_lock();
 982                 if (tracefile[0] != 0) {
 983                         filp = cfs_filp_open(tracefile,
 984                                              O_CREAT | O_RDWR | O_LARGEFILE,
 985                                              0600, &rc);
 986                         if (!(filp))
 987                                 printk(KERN_WARNING "couldn't open %s: %d\n",
 988                                        tracefile, rc);
 989                 }
 990                 tracefile_read_unlock();
 991                 if (filp == NULL) {
 992                         put_pages_on_daemon_list(&pc);
 993                         __LASSERT(list_empty(&pc.pc_pages));
 994                         goto end_loop;
 995                 }
 996
 997                 CFS_MMSPACE_OPEN;
 998
 999                 /* mark the first header, so we can sort in chunks */
1000                 tage = tage_from_list(pc.pc_pages.next);
1001                 __LASSERT_TAGE_INVARIANT(tage);
1002
1003                 hdr = cfs_page_address(tage->page);
1004                 hdr->ph_flags |= PH_FLAG_FIRST_RECORD;
1005
1006                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
1007                         static loff_t f_pos;
1008
1009                         __LASSERT_TAGE_INVARIANT(tage);
1010
1011                         if (f_pos >= (off_t)tracefile_size)
1012                                 f_pos = 0;
1013                         else if (f_pos > cfs_filp_size(filp))
1014                                 f_pos = cfs_filp_size(filp);
1015
1016                         rc = cfs_filp_write(filp, cfs_page_address(tage->page),
1017                                             tage->used, &f_pos);
1018                         if (rc != (int)tage->used) {
1019                                 printk(KERN_WARNING "wanted to write %u but "
1020                                        "wrote %d\n", tage->used, rc);
1021                                 put_pages_back(&pc);
1022                                 __LASSERT(list_empty(&pc.pc_pages));
1023                         }
1024                 }
1025                 CFS_MMSPACE_CLOSE;
1026
1027                 cfs_filp_close(filp);
1028                 put_pages_on_daemon_list(&pc);
1029                 if (!list_empty(&pc.pc_pages)) {
1030                         int i;
1031
1032                         printk(KERN_ALERT "Lustre: trace pages aren't empty\n");
1033                         printk(KERN_ALERT "total cpus(%d): ", num_possible_cpus());
1034                         for (i = 0; i < num_possible_cpus(); i++)
1035                                 if (cpu_online(i))
1036                                         printk(KERN_ALERT "%d(on) ", i);
1037                                 else
1038                                         printk(KERN_ALERT "%d(off) ", i);
1039                         printk(KERN_ALERT "\n");
1040
1041                         i = 0;
1042                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1043                                                  linkage)
1044                                 printk(KERN_ALERT "page %d belongs to cpu %d\n",
1045                                        ++i, tage->cpu);
1046                         printk(KERN_ALERT "There are %d pages unwritten\n", i);
1047                 }
1048                 __LASSERT(list_empty(&pc.pc_pages));
1049 end_loop:
1050                 if (atomic_read(&tctl->tctl_shutdown)) {
1051                         if (last_loop == 0) {
1052                                 last_loop = 1;
1053                                 continue;
1054                         } else {
1055                                 break;
1056                         }
1057                 }
1058                 cfs_waitlink_init(&__wait);
1059                 cfs_waitq_add(&tctl->tctl_waitq, &__wait);
1060                 set_current_state(TASK_INTERRUPTIBLE);
1061                 cfs_waitq_timedwait(&__wait, CFS_TASK_INTERRUPTIBLE,
1062                                     cfs_time_seconds(1));
1063                 cfs_waitq_del(&tctl->tctl_waitq, &__wait);
1064         }
1065         complete(&tctl->tctl_stop);
1066         return 0;
1067 }
1068
1069 int trace_start_thread(void)
1070 {
1071         struct tracefiled_ctl *tctl = &trace_tctl;
1072         int rc = 0;
1073
1074         mutex_down(&trace_thread_sem);
1075         if (thread_running)
1076                 goto out;
1077
1078         init_completion(&tctl->tctl_start);
1079         init_completion(&tctl->tctl_stop);
1080         cfs_waitq_init(&tctl->tctl_waitq);
1081         atomic_set(&tctl->tctl_shutdown, 0);
1082
1083         if (cfs_kernel_thread(tracefiled, tctl, 0) < 0) {
1084                 rc = -ECHILD;
1085                 goto out;
1086         }
1087
1088         wait_for_completion(&tctl->tctl_start);
1089         thread_running = 1;
1090 out:
1091         mutex_up(&trace_thread_sem);
1092         return rc;
1093 }
1094
1095 void trace_stop_thread(void)
1096 {
1097         struct tracefiled_ctl *tctl = &trace_tctl;
1098
1099         mutex_down(&trace_thread_sem);
1100         if (thread_running) {
1101                 printk(KERN_INFO "Lustre: shutting down debug daemon thread...\n");
1102                 atomic_set(&tctl->tctl_shutdown, 1);
1103                 wait_for_completion(&tctl->tctl_stop);
1104                 thread_running = 0;
1105         }
1106         mutex_up(&trace_thread_sem);
1107 }
1108
1109 int tracefile_init(int max_pages)
1110 {
1111         struct trace_cpu_data *tcd;
1112         int                    i;
1113         int                    j;
1114         int                    rc;
1115         int                    factor;
1116
1117         rc = tracefile_init_arch();
1118         if (rc != 0)
1119                 return rc;
1120
1121         tcd_for_each(tcd, i, j) {
1122                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1123                 factor = tcd->tcd_pages_factor;
1124                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
1125                 CFS_INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1126                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1127                 tcd->tcd_cur_pages = 0;
1128                 tcd->tcd_cur_stock_pages = 0;
1129                 tcd->tcd_cur_daemon_pages = 0;
1130                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1131                 LASSERT(tcd->tcd_max_pages > 0);
1132                 tcd->tcd_shutting_down = 0;
1133         }
1134
1135         return 0;
1136 }
1137
1138 static void trace_cleanup_on_all_cpus(void)
1139 {
1140         struct trace_cpu_data *tcd;
1141         struct trace_page *tage;
1142         struct trace_page *tmp;
1143         int i, cpu;
1144
1145         for_each_possible_cpu(cpu) {
1146                 tcd_for_each_type_lock(tcd, i, cpu) {
1147                         tcd->tcd_shutting_down = 1;
1148
1149                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages,
1150                                                  linkage) {
1151                                 __LASSERT_TAGE_INVARIANT(tage);
1152
1153                                 list_del(&tage->linkage);
1154                                 tage_free(tage);
1155                         }
1156
1157                         tcd->tcd_cur_pages = 0;
1158                 }
1159         }
1160 }
1161
1162 static void trace_cleanup(void)
1163 {
1164         struct page_collection pc;
1165
1166         CFS_INIT_LIST_HEAD(&pc.pc_pages);
1167         spin_lock_init(&pc.pc_lock);
1168
1169         trace_cleanup_on_all_cpus();
1170
1171         tracefile_fini_arch();
1172 }
1173
1174 void tracefile_exit(void)
1175 {
1176         trace_stop_thread();
1177         trace_cleanup();
1178 }