libcfs/libcfs/tracefile.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * libcfs/libcfs/tracefile.c
  37  *
  38  * Author: Zach Brown <zab@clusterfs.com>
  39  * Author: Phil Schwan <phil@clusterfs.com>
  40  */
  41
  42
  43 #define DEBUG_SUBSYSTEM S_LNET
  44 #define LUSTRE_TRACEFILE_PRIVATE
  45 #include "tracefile.h"
  46
  47 #include <libcfs/libcfs.h>
  48
  49 /* XXX move things up to the top, comment */
  50 union trace_data_union (*trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  51
  52 char tracefile[TRACEFILE_NAME_SIZE];
  53 long long tracefile_size = TRACEFILE_SIZE;
  54 static struct tracefiled_ctl trace_tctl;
  55 struct semaphore trace_thread_sem;
  56 static int thread_running = 0;
  57
  58 atomic_t tage_allocated = ATOMIC_INIT(0);
  59
  60 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  61                                          struct trace_cpu_data *tcd);
  62
  63 static inline struct trace_page *tage_from_list(struct list_head *list)
  64 {
  65         return list_entry(list, struct trace_page, linkage);
  66 }
  67
  68 static struct trace_page *tage_alloc(int gfp)
  69 {
  70         cfs_page_t        *page;
  71         struct trace_page *tage;
  72
  73         /*
  74          * Don't spam console with allocation failures: they will be reported
  75          * by upper layer anyway.
  76          */
  77         gfp |= CFS_ALLOC_NOWARN;
  78         page = cfs_alloc_page(gfp);
  79         if (page == NULL)
  80                 return NULL;
  81
  82         tage = cfs_alloc(sizeof(*tage), gfp);
  83         if (tage == NULL) {
  84                 cfs_free_page(page);
  85                 return NULL;
  86         }
  87
  88         tage->page = page;
  89         atomic_inc(&tage_allocated);
  90         return tage;
  91 }
  92
  93 static void tage_free(struct trace_page *tage)
  94 {
  95         __LASSERT(tage != NULL);
  96         __LASSERT(tage->page != NULL);
  97
  98         cfs_free_page(tage->page);
  99         cfs_free(tage);
 100         atomic_dec(&tage_allocated);
 101 }
 102
 103 static void tage_to_tail(struct trace_page *tage, struct list_head *queue)
 104 {
 105         __LASSERT(tage != NULL);
 106         __LASSERT(queue != NULL);
 107
 108         list_move_tail(&tage->linkage, queue);
 109 }
 110
 111 int trace_refill_stock(struct trace_cpu_data *tcd, int gfp,
 112                        struct list_head *stock)
 113 {
 114         int i;
 115
 116         /*
 117          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 118          * from here: this will lead to infinite recursion.
 119          */
 120
 121         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++ i) {
 122                 struct trace_page *tage;
 123
 124                 tage = tage_alloc(gfp);
 125                 if (tage == NULL)
 126                         break;
 127                 list_add_tail(&tage->linkage, stock);
 128         }
 129         return i;
 130 }
 131
 132 /* return a page that has 'len' bytes left at the end */
 133 static struct trace_page *trace_get_tage_try(struct trace_cpu_data *tcd,
 134                                              unsigned long len)
 135 {
 136         struct trace_page *tage;
 137
 138         if (tcd->tcd_cur_pages > 0) {
 139                 __LASSERT(!list_empty(&tcd->tcd_pages));
 140                 tage = tage_from_list(tcd->tcd_pages.prev);
 141                 if (tage->used + len <= CFS_PAGE_SIZE)
 142                         return tage;
 143         }
 144
 145         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 146                 if (tcd->tcd_cur_stock_pages > 0) {
 147                         tage = tage_from_list(tcd->tcd_stock_pages.prev);
 148                         -- tcd->tcd_cur_stock_pages;
 149                         list_del_init(&tage->linkage);
 150                 } else {
 151                         tage = tage_alloc(CFS_ALLOC_ATOMIC);
 152                         if (tage == NULL) {
 153                                 printk(KERN_WARNING
 154                                        "failure to allocate a tage (%ld)\n",
 155                                        tcd->tcd_cur_pages);
 156                                 return NULL;
 157                         }
 158                 }
 159
 160                 tage->used = 0;
 161                 tage->cpu = smp_processor_id();
 162                 tage->type = tcd->tcd_type;
 163                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 164                 tcd->tcd_cur_pages++;
 165
 166                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 167                         struct tracefiled_ctl *tctl = &trace_tctl;
 168                         /*
 169                          * wake up tracefiled to process some pages.
 170                          */
 171                         cfs_waitq_signal(&tctl->tctl_waitq);
 172                 }
 173                 return tage;
 174         }
 175         return NULL;
 176 }
 177
 178 static void tcd_shrink(struct trace_cpu_data *tcd)
 179 {
 180         int pgcount = tcd->tcd_cur_pages / 10;
 181         struct page_collection pc;
 182         struct trace_page *tage;
 183         struct trace_page *tmp;
 184
 185         /*
 186          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 187          * from here: this will lead to infinite recursion.
 188          */
 189
 190         if (printk_ratelimit())
 191                 printk(KERN_WARNING "debug daemon buffer overflowed; "
 192                        "discarding 10%% of pages (%d of %ld)\n",
 193                        pgcount + 1, tcd->tcd_cur_pages);
 194
 195         CFS_INIT_LIST_HEAD(&pc.pc_pages);
 196         spin_lock_init(&pc.pc_lock);
 197
 198         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 199                 if (pgcount-- == 0)
 200                         break;
 201
 202                 list_move_tail(&tage->linkage, &pc.pc_pages);
 203                 tcd->tcd_cur_pages--;
 204         }
 205         put_pages_on_tcd_daemon_list(&pc, tcd);
 206 }
 207
 208 /* return a page that has 'len' bytes left at the end */
 209 static struct trace_page *trace_get_tage(struct trace_cpu_data *tcd,
 210                                          unsigned long len)
 211 {
 212         struct trace_page *tage;
 213
 214         /*
 215          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 216          * from here: this will lead to infinite recursion.
 217          */
 218
 219         if (len > CFS_PAGE_SIZE) {
 220                 printk(KERN_ERR
 221                        "cowardly refusing to write %lu bytes in a page\n", len);
 222                 return NULL;
 223         }
 224
 225         tage = trace_get_tage_try(tcd, len);
 226         if (tage != NULL)
 227                 return tage;
 228         if (thread_running)
 229                 tcd_shrink(tcd);
 230         if (tcd->tcd_cur_pages > 0) {
 231                 tage = tage_from_list(tcd->tcd_pages.next);
 232                 tage->used = 0;
 233                 tage_to_tail(tage, &tcd->tcd_pages);
 234         }
 235         return tage;
 236 }
 237
 238 int libcfs_debug_vmsg2(cfs_debug_limit_state_t *cdls, int subsys, int mask,
 239                        const char *file, const char *fn, const int line,
 240                        const char *format1, va_list args,
 241                        const char *format2, ...)
 242 {
 243         struct trace_cpu_data   *tcd = NULL;
 244         struct ptldebug_header   header;
 245         struct trace_page       *tage;
 246         /* string_buf is used only if tcd != NULL, and is always set then */
 247         char                    *string_buf = NULL;
 248         char                    *debug_buf;
 249         int                      known_size;
 250         int                      needed = 85; /* average message length */
 251         int                      max_nob;
 252         va_list                  ap;
 253         int                      depth;
 254         int                      i;
 255         int                      remain;
 256
 257         if (strchr(file, '/'))
 258                 file = strrchr(file, '/') + 1;
 259
 260
 261         set_ptldebug_header(&header, subsys, mask, line, CDEBUG_STACK());
 262
 263         tcd = trace_get_tcd();
 264         if (tcd == NULL)                /* arch may not log in IRQ context */
 265                 goto console;
 266
 267         if (tcd->tcd_shutting_down) {
 268                 trace_put_tcd(tcd);
 269                 tcd = NULL;
 270                 goto console;
 271         }
 272
 273         depth = __current_nesting_level();
 274         known_size = strlen(file) + 1 + depth;
 275         if (fn)
 276                 known_size += strlen(fn) + 1;
 277
 278         if (libcfs_debug_binary)
 279                 known_size += sizeof(header);
 280
 281         /*/
 282          * '2' used because vsnprintf return real size required for output
 283          * _without_ terminating NULL.
 284          * if needed is to small for this format.
 285          */
 286         for (i=0;i<2;i++) {
 287                 tage = trace_get_tage(tcd, needed + known_size + 1);
 288                 if (tage == NULL) {
 289                         if (needed + known_size > CFS_PAGE_SIZE)
 290                                 mask |= D_ERROR;
 291
 292                         trace_put_tcd(tcd);
 293                         tcd = NULL;
 294                         goto console;
 295                 }
 296
 297                 string_buf = (char *)cfs_page_address(tage->page)+tage->used+known_size;
 298
 299                 max_nob = CFS_PAGE_SIZE - tage->used - known_size;
 300                 if (max_nob <= 0) {
 301                         printk(KERN_EMERG "negative max_nob: %i\n", max_nob);
 302                         mask |= D_ERROR;
 303                         trace_put_tcd(tcd);
 304                         tcd = NULL;
 305                         goto console;
 306                 }
 307
 308                 needed = 0;
 309                 if (format1) {
 310                         va_copy(ap, args);
 311                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 312                         va_end(ap);
 313                 }
 314
 315
 316                 if (format2) {
 317                         remain = max_nob - needed;
 318                         if (remain < 0)
 319                                 remain = 0;
 320
 321                         va_start(ap, format2);
 322                         needed += vsnprintf(string_buf+needed, remain, format2, ap);
 323                         va_end(ap);
 324                 }
 325
 326                 if (needed < max_nob) /* well. printing ok.. */
 327                         break;
 328         }
 329
 330         if (*(string_buf+needed-1) != '\n')
 331                 printk(KERN_INFO "format at %s:%d:%s doesn't end in newline\n",
 332                        file, line, fn);
 333
 334         header.ph_len = known_size + needed;
 335         debug_buf = (char *)cfs_page_address(tage->page) + tage->used;
 336
 337         if (libcfs_debug_binary) {
 338                 memcpy(debug_buf, &header, sizeof(header));
 339                 tage->used += sizeof(header);
 340                 debug_buf += sizeof(header);
 341         }
 342
 343         /* indent message according to the nesting level */
 344         while (depth-- > 0) {
 345                 *(debug_buf++) = '.';
 346                 ++ tage->used;
 347         }
 348
 349         strcpy(debug_buf, file);
 350         tage->used += strlen(file) + 1;
 351         debug_buf += strlen(file) + 1;
 352
 353         if (fn) {
 354                 strcpy(debug_buf, fn);
 355                 tage->used += strlen(fn) + 1;
 356                 debug_buf += strlen(fn) + 1;
 357         }
 358
 359         __LASSERT(debug_buf == string_buf);
 360
 361         tage->used += needed;
 362         __LASSERT (tage->used <= CFS_PAGE_SIZE);
 363
 364 console:
 365         if ((mask & libcfs_printk) == 0) {
 366                 /* no console output requested */
 367                 if (tcd != NULL)
 368                         trace_put_tcd(tcd);
 369                 return 1;
 370         }
 371
 372         if (cdls != NULL) {
 373                 if (libcfs_console_ratelimit &&
 374                     cdls->cdls_next != 0 &&     /* not first time ever */
 375                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 376                         /* skipping a console message */
 377                         cdls->cdls_count++;
 378                         if (tcd != NULL)
 379                                 trace_put_tcd(tcd);
 380                         return 1;
 381                 }
 382
 383                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 384                                                        libcfs_console_max_delay
 385                                                        + cfs_time_seconds(10))) {
 386                         /* last timeout was a long time ago */
 387                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 388                 } else {
 389                         cdls->cdls_delay *= libcfs_console_backoff;
 390
 391                         if (cdls->cdls_delay < libcfs_console_min_delay)
 392                                 cdls->cdls_delay = libcfs_console_min_delay;
 393                         else if (cdls->cdls_delay > libcfs_console_max_delay)
 394                                 cdls->cdls_delay = libcfs_console_max_delay;
 395                 }
 396
 397                 /* ensure cdls_next is never zero after it's been seen */
 398                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 399         }
 400
 401         if (tcd != NULL) {
 402                 print_to_console(&header, mask, string_buf, needed, file, fn);
 403                 trace_put_tcd(tcd);
 404         } else {
 405                 string_buf = trace_get_console_buffer();
 406
 407                 needed = 0;
 408                 if (format1 != NULL) {
 409                         va_copy(ap, args);
 410                         needed = vsnprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE, format1, ap);
 411                         va_end(ap);
 412                 }
 413                 if (format2 != NULL) {
 414                         remain = TRACE_CONSOLE_BUFFER_SIZE - needed;
 415                         if (remain > 0) {
 416                                 va_start(ap, format2);
 417                                 needed += vsnprintf(string_buf+needed, remain, format2, ap);
 418                                 va_end(ap);
 419                         }
 420                 }
 421                 print_to_console(&header, mask,
 422                                  string_buf, needed, file, fn);
 423
 424                 trace_put_console_buffer(string_buf);
 425         }
 426
 427         if (cdls != NULL && cdls->cdls_count != 0) {
 428                 string_buf = trace_get_console_buffer();
 429
 430                 needed = snprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE,
 431                          "Skipped %d previous similar message%s\n",
 432                          cdls->cdls_count, (cdls->cdls_count > 1) ? "s" : "");
 433
 434                 print_to_console(&header, mask,
 435                                  string_buf, needed, file, fn);
 436
 437                 trace_put_console_buffer(string_buf);
 438                 cdls->cdls_count = 0;
 439         }
 440
 441         return 0;
 442 }
 443 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 444
 445 void
 446 libcfs_assertion_failed(const char *expr, const char *file,
 447                         const char *func, const int line)
 448 {
 449         libcfs_debug_msg(NULL, 0, D_EMERG, file, func, line,
 450                          "ASSERTION(%s) failed\n", expr);
 451         LBUG();
 452 }
 453 EXPORT_SYMBOL(libcfs_assertion_failed);
 454
 455 void
 456 trace_assertion_failed(const char *str,
 457                        const char *fn, const char *file, int line)
 458 {
 459         struct ptldebug_header hdr;
 460
 461         libcfs_panic_in_progress = 1;
 462         libcfs_catastrophe = 1;
 463         mb();
 464
 465         set_ptldebug_header(&hdr, DEBUG_SUBSYSTEM, D_EMERG, line,
 466                             CDEBUG_STACK());
 467
 468         print_to_console(&hdr, D_EMERG, str, strlen(str), file, fn);
 469
 470         LIBCFS_PANIC("Lustre debug assertion failure\n");
 471
 472         /* not reached */
 473 }
 474
 475 static void
 476 panic_collect_pages(struct page_collection *pc)
 477 {
 478         /* Do the collect_pages job on a single CPU: assumes that all other
 479          * CPUs have been stopped during a panic.  If this isn't true for some
 480          * arch, this will have to be implemented separately in each arch.  */
 481         int                    i;
 482         int                    j;
 483         struct trace_cpu_data *tcd;
 484
 485         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 486
 487         tcd_for_each(tcd, i, j) {
 488                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 489                 tcd->tcd_cur_pages = 0;
 490
 491                 if (pc->pc_want_daemon_pages) {
 492                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 493                         tcd->tcd_cur_daemon_pages = 0;
 494                 }
 495         }
 496 }
 497
 498 static void collect_pages_on_cpu(void *info)
 499 {
 500         struct trace_cpu_data *tcd;
 501         struct page_collection *pc = info;
 502         int i;
 503
 504         spin_lock(&pc->pc_lock);
 505         tcd_for_each_type_lock(tcd, i) {
 506                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 507                 tcd->tcd_cur_pages = 0;
 508                 if (pc->pc_want_daemon_pages) {
 509                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 510                         tcd->tcd_cur_daemon_pages = 0;
 511                 }
 512         }
 513         spin_unlock(&pc->pc_lock);
 514 }
 515
 516 static void collect_pages(struct page_collection *pc)
 517 {
 518         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 519
 520         if (libcfs_panic_in_progress)
 521                 panic_collect_pages(pc);
 522         else
 523                 trace_call_on_all_cpus(collect_pages_on_cpu, pc);
 524 }
 525
 526 static void put_pages_back_on_cpu(void *info)
 527 {
 528         struct page_collection *pc = info;
 529         struct trace_cpu_data *tcd;
 530         struct list_head *cur_head;
 531         struct trace_page *tage;
 532         struct trace_page *tmp;
 533         int i;
 534
 535         spin_lock(&pc->pc_lock);
 536         tcd_for_each_type_lock(tcd, i) {
 537                 cur_head = tcd->tcd_pages.next;
 538
 539                 list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 540
 541                         __LASSERT_TAGE_INVARIANT(tage);
 542
 543                         if (tage->cpu != smp_processor_id() || tage->type != i)
 544                                 continue;
 545
 546                         tage_to_tail(tage, cur_head);
 547                         tcd->tcd_cur_pages++;
 548                 }
 549         }
 550         spin_unlock(&pc->pc_lock);
 551 }
 552
 553 static void put_pages_back(struct page_collection *pc)
 554 {
 555         if (!libcfs_panic_in_progress)
 556                 trace_call_on_all_cpus(put_pages_back_on_cpu, pc);
 557 }
 558
 559 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 560  * we have a good amount of data at all times for dumping during an LBUG, even
 561  * if we have been steadily writing (and otherwise discarding) pages via the
 562  * debug daemon. */
 563 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 564                                          struct trace_cpu_data *tcd)
 565 {
 566         struct trace_page *tage;
 567         struct trace_page *tmp;
 568
 569         spin_lock(&pc->pc_lock);
 570         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 571
 572                 __LASSERT_TAGE_INVARIANT(tage);
 573
 574                 if (tage->cpu != smp_processor_id() ||
 575                     tage->type != tcd->tcd_type)
 576                         continue;
 577
 578                 tage_to_tail(tage, &tcd->tcd_daemon_pages);
 579                 tcd->tcd_cur_daemon_pages++;
 580
 581                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 582                         struct trace_page *victim;
 583
 584                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 585                         victim = tage_from_list(tcd->tcd_daemon_pages.next);
 586
 587                         __LASSERT_TAGE_INVARIANT(victim);
 588
 589                         list_del(&victim->linkage);
 590                         tage_free(victim);
 591                         tcd->tcd_cur_daemon_pages--;
 592                 }
 593         }
 594         spin_unlock(&pc->pc_lock);
 595 }
 596
 597 static void put_pages_on_daemon_list_on_cpu(void *info)
 598 {
 599         struct trace_cpu_data *tcd;
 600         int i;
 601
 602         tcd_for_each_type_lock(tcd, i)
 603                 put_pages_on_tcd_daemon_list(info, tcd);
 604 }
 605
 606 static void put_pages_on_daemon_list(struct page_collection *pc)
 607 {
 608         trace_call_on_all_cpus(put_pages_on_daemon_list_on_cpu, pc);
 609 }
 610
 611 void trace_debug_print(void)
 612 {
 613         struct page_collection pc;
 614         struct trace_page *tage;
 615         struct trace_page *tmp;
 616
 617         spin_lock_init(&pc.pc_lock);
 618
 619         pc.pc_want_daemon_pages = 1;
 620         collect_pages(&pc);
 621         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 622                 char *p, *file, *fn;
 623                 cfs_page_t *page;
 624
 625                 __LASSERT_TAGE_INVARIANT(tage);
 626
 627                 page = tage->page;
 628                 p = cfs_page_address(page);
 629                 while (p < ((char *)cfs_page_address(page) + tage->used)) {
 630                         struct ptldebug_header *hdr;
 631                         int len;
 632                         hdr = (void *)p;
 633                         p += sizeof(*hdr);
 634                         file = p;
 635                         p += strlen(file) + 1;
 636                         fn = p;
 637                         p += strlen(fn) + 1;
 638                         len = hdr->ph_len - (p - (char *)hdr);
 639
 640                         print_to_console(hdr, D_EMERG, p, len, file, fn);
 641
 642                         p += len;
 643                 }
 644
 645                 list_del(&tage->linkage);
 646                 tage_free(tage);
 647         }
 648 }
 649
 650 int tracefile_dump_all_pages(char *filename)
 651 {
 652         struct page_collection pc;
 653         cfs_file_t *filp;
 654         struct trace_page *tage;
 655         struct trace_page *tmp;
 656         int rc;
 657
 658         CFS_DECL_MMSPACE;
 659
 660         tracefile_write_lock();
 661
 662         filp = cfs_filp_open(filename,
 663                              O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600, &rc);
 664         if (!filp) {
 665                 if (rc != -EEXIST)
 666                         printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 667                                filename, rc);
 668                 goto out;
 669         }
 670
 671         spin_lock_init(&pc.pc_lock);
 672         pc.pc_want_daemon_pages = 1;
 673         collect_pages(&pc);
 674         if (list_empty(&pc.pc_pages)) {
 675                 rc = 0;
 676                 goto close;
 677         }
 678
 679         /* ok, for now, just write the pages.  in the future we'll be building
 680          * iobufs with the pages and calling generic_direct_IO */
 681         CFS_MMSPACE_OPEN;
 682         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 683
 684                 __LASSERT_TAGE_INVARIANT(tage);
 685
 686                 rc = cfs_filp_write(filp, cfs_page_address(tage->page),
 687                                     tage->used, cfs_filp_poff(filp));
 688                 if (rc != (int)tage->used) {
 689                         printk(KERN_WARNING "wanted to write %u but wrote "
 690                                "%d\n", tage->used, rc);
 691                         put_pages_back(&pc);
 692                         __LASSERT(list_empty(&pc.pc_pages));
 693                         break;
 694                 }
 695                 list_del(&tage->linkage);
 696                 tage_free(tage);
 697         }
 698         CFS_MMSPACE_CLOSE;
 699         rc = cfs_filp_fsync(filp);
 700         if (rc)
 701                 printk(KERN_ERR "sync returns %d\n", rc);
 702  close:
 703         cfs_filp_close(filp);
 704  out:
 705         tracefile_write_unlock();
 706         return rc;
 707 }
 708
 709 void trace_flush_pages(void)
 710 {
 711         struct page_collection pc;
 712         struct trace_page *tage;
 713         struct trace_page *tmp;
 714
 715         spin_lock_init(&pc.pc_lock);
 716
 717         pc.pc_want_daemon_pages = 1;
 718         collect_pages(&pc);
 719         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 720
 721                 __LASSERT_TAGE_INVARIANT(tage);
 722
 723                 list_del(&tage->linkage);
 724                 tage_free(tage);
 725         }
 726 }
 727
 728 int trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 729                         const char *usr_buffer, int usr_buffer_nob)
 730 {
 731         int    nob;
 732
 733         if (usr_buffer_nob > knl_buffer_nob)
 734                 return -EOVERFLOW;
 735
 736         if (copy_from_user((void *)knl_buffer,
 737                            (void *)usr_buffer, usr_buffer_nob))
 738                 return -EFAULT;
 739
 740         nob = strnlen(knl_buffer, usr_buffer_nob);
 741         while (nob-- >= 0)                      /* strip trailing whitespace */
 742                 if (!isspace(knl_buffer[nob]))
 743                         break;
 744
 745         if (nob < 0)                            /* empty string */
 746                 return -EINVAL;
 747
 748         if (nob == knl_buffer_nob)              /* no space to terminate */
 749                 return -EOVERFLOW;
 750
 751         knl_buffer[nob + 1] = 0;                /* terminate */
 752         return 0;
 753 }
 754
 755 int trace_copyout_string(char *usr_buffer, int usr_buffer_nob,
 756                          const char *knl_buffer, char *append)
 757 {
 758         /* NB if 'append' != NULL, it's a single character to append to the
 759          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 760          * terminating zero byte) for sysctl entries */
 761         int   nob = strlen(knl_buffer);
 762
 763         if (nob > usr_buffer_nob)
 764                 nob = usr_buffer_nob;
 765
 766         if (copy_to_user(usr_buffer, knl_buffer, nob))
 767                 return -EFAULT;
 768
 769         if (append != NULL && nob < usr_buffer_nob) {
 770                 if (copy_to_user(usr_buffer + nob, append, 1))
 771                         return -EFAULT;
 772
 773                 nob++;
 774         }
 775
 776         return nob;
 777 }
 778
 779 int trace_allocate_string_buffer(char **str, int nob)
 780 {
 781         if (nob > 2 * CFS_PAGE_SIZE)            /* string must be "sensible" */
 782                 return -EINVAL;
 783
 784         *str = cfs_alloc(nob, CFS_ALLOC_STD | CFS_ALLOC_ZERO);
 785         if (*str == NULL)
 786                 return -ENOMEM;
 787
 788         return 0;
 789 }
 790
 791 void trace_free_string_buffer(char *str, int nob)
 792 {
 793         cfs_free(str);
 794 }
 795
 796 int trace_dump_debug_buffer_usrstr(void *usr_str, int usr_str_nob)
 797 {
 798         char         *str;
 799         int           rc;
 800
 801         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 802         if (rc != 0)
 803                 return rc;
 804
 805         rc = trace_copyin_string(str, usr_str_nob + 1,
 806                                  usr_str, usr_str_nob);
 807         if (rc != 0)
 808                 goto out;
 809
 810 #if !defined(__WINNT__)
 811         if (str[0] != '/') {
 812                 rc = -EINVAL;
 813                 goto out;
 814         }
 815 #endif
 816         rc = tracefile_dump_all_pages(str);
 817 out:
 818         trace_free_string_buffer(str, usr_str_nob + 1);
 819         return rc;
 820 }
 821
 822 int trace_daemon_command(char *str)
 823 {
 824         int       rc = 0;
 825
 826         tracefile_write_lock();
 827
 828         if (strcmp(str, "stop") == 0) {
 829                 trace_stop_thread();
 830                 memset(tracefile, 0, sizeof(tracefile));
 831
 832         } else if (strncmp(str, "size=", 5) == 0) {
 833                 tracefile_size = simple_strtoul(str + 5, NULL, 0);
 834                 if (tracefile_size < 10 || tracefile_size > 20480)
 835                         tracefile_size = TRACEFILE_SIZE;
 836                 else
 837                         tracefile_size <<= 20;
 838
 839         } else if (strlen(str) >= sizeof(tracefile)) {
 840                 rc = -ENAMETOOLONG;
 841 #ifndef __WINNT__
 842         } else if (str[0] != '/') {
 843                 rc = -EINVAL;
 844 #endif
 845         } else {
 846                 strcpy(tracefile, str);
 847
 848                 printk(KERN_INFO "Lustre: debug daemon will attempt to start writing "
 849                        "to %s (%lukB max)\n", tracefile,
 850                        (long)(tracefile_size >> 10));
 851
 852                 trace_start_thread();
 853         }
 854
 855         tracefile_write_unlock();
 856         return rc;
 857 }
 858
 859 int trace_daemon_command_usrstr(void *usr_str, int usr_str_nob)
 860 {
 861         char *str;
 862         int   rc;
 863
 864         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 865         if (rc != 0)
 866                 return rc;
 867
 868         rc = trace_copyin_string(str, usr_str_nob + 1,
 869                                  usr_str, usr_str_nob);
 870         if (rc == 0)
 871                 rc = trace_daemon_command(str);
 872
 873         trace_free_string_buffer(str, usr_str_nob + 1);
 874         return rc;
 875 }
 876
 877 int trace_set_debug_mb(int mb)
 878 {
 879         int i;
 880         int j;
 881         int pages;
 882         int limit = trace_max_debug_mb();
 883         struct trace_cpu_data *tcd;
 884
 885         if (mb < num_possible_cpus())
 886                 return -EINVAL;
 887
 888         if (mb > limit) {
 889                 printk(KERN_ERR "Lustre: Refusing to set debug buffer size to "
 890                        "%dMB - limit is %d\n", mb, limit);
 891                 return -EINVAL;
 892         }
 893
 894         mb /= num_possible_cpus();
 895         pages = mb << (20 - CFS_PAGE_SHIFT);
 896
 897         tracefile_write_lock();
 898
 899         tcd_for_each(tcd, i, j)
 900                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 901
 902         tracefile_write_unlock();
 903
 904         return 0;
 905 }
 906
 907 int trace_set_debug_mb_usrstr(void *usr_str, int usr_str_nob)
 908 {
 909         char     str[32];
 910         int      rc;
 911
 912         rc = trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 913         if (rc < 0)
 914                 return rc;
 915
 916         return trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 917 }
 918
 919 int trace_get_debug_mb(void)
 920 {
 921         int i;
 922         int j;
 923         struct trace_cpu_data *tcd;
 924         int total_pages = 0;
 925
 926         tracefile_read_lock();
 927
 928         tcd_for_each(tcd, i, j)
 929                 total_pages += tcd->tcd_max_pages;
 930
 931         tracefile_read_unlock();
 932
 933         return (total_pages >> (20 - CFS_PAGE_SHIFT)) + 1;
 934 }
 935
 936 static int tracefiled(void *arg)
 937 {
 938         struct page_collection pc;
 939         struct tracefiled_ctl *tctl = arg;
 940         struct trace_page *tage;
 941         struct trace_page *tmp;
 942         struct ptldebug_header *hdr;
 943         cfs_file_t *filp;
 944         int rc;
 945
 946         CFS_DECL_MMSPACE;
 947
 948         /* we're started late enough that we pick up init's fs context */
 949         /* this is so broken in uml?  what on earth is going on? */
 950         cfs_daemonize("ktracefiled");
 951
 952         spin_lock_init(&pc.pc_lock);
 953         complete(&tctl->tctl_start);
 954
 955         while (1) {
 956                 cfs_waitlink_t __wait;
 957
 958                 cfs_waitlink_init(&__wait);
 959                 cfs_waitq_add(&tctl->tctl_waitq, &__wait);
 960                 set_current_state(TASK_INTERRUPTIBLE);
 961                 cfs_waitq_timedwait(&__wait, CFS_TASK_INTERRUPTIBLE,
 962                                     cfs_time_seconds(1));
 963                 cfs_waitq_del(&tctl->tctl_waitq, &__wait);
 964
 965                 if (atomic_read(&tctl->tctl_shutdown))
 966                         break;
 967
 968                 pc.pc_want_daemon_pages = 0;
 969                 collect_pages(&pc);
 970                 if (list_empty(&pc.pc_pages))
 971                         continue;
 972
 973                 filp = NULL;
 974                 tracefile_read_lock();
 975                 if (tracefile[0] != 0) {
 976                         filp = cfs_filp_open(tracefile,
 977                                              O_CREAT | O_RDWR | O_LARGEFILE,
 978                                              0600, &rc);
 979                         if (!(filp))
 980                                 printk(KERN_WARNING "couldn't open %s: %d\n",
 981                                        tracefile, rc);
 982                 }
 983                 tracefile_read_unlock();
 984                 if (filp == NULL) {
 985                         put_pages_on_daemon_list(&pc);
 986                         __LASSERT(list_empty(&pc.pc_pages));
 987                         continue;
 988                 }
 989
 990                 CFS_MMSPACE_OPEN;
 991
 992                 /* mark the first header, so we can sort in chunks */
 993                 tage = tage_from_list(pc.pc_pages.next);
 994                 __LASSERT_TAGE_INVARIANT(tage);
 995
 996                 hdr = cfs_page_address(tage->page);
 997                 hdr->ph_flags |= PH_FLAG_FIRST_RECORD;
 998
 999                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
1000                         static loff_t f_pos;
1001
1002                         __LASSERT_TAGE_INVARIANT(tage);
1003
1004                         if (f_pos >= (off_t)tracefile_size)
1005                                 f_pos = 0;
1006                         else if (f_pos > cfs_filp_size(filp))
1007                                 f_pos = cfs_filp_size(filp);
1008
1009                         rc = cfs_filp_write(filp, cfs_page_address(tage->page),
1010                                             tage->used, &f_pos);
1011                         if (rc != (int)tage->used) {
1012                                 printk(KERN_WARNING "wanted to write %u but "
1013                                        "wrote %d\n", tage->used, rc);
1014                                 put_pages_back(&pc);
1015                                 __LASSERT(list_empty(&pc.pc_pages));
1016                         }
1017                 }
1018                 CFS_MMSPACE_CLOSE;
1019
1020                 cfs_filp_close(filp);
1021                 put_pages_on_daemon_list(&pc);
1022                 __LASSERT(list_empty(&pc.pc_pages));
1023         }
1024         complete(&tctl->tctl_stop);
1025         return 0;
1026 }
1027
1028 int trace_start_thread(void)
1029 {
1030         struct tracefiled_ctl *tctl = &trace_tctl;
1031         int rc = 0;
1032
1033         mutex_down(&trace_thread_sem);
1034         if (thread_running)
1035                 goto out;
1036
1037         init_completion(&tctl->tctl_start);
1038         init_completion(&tctl->tctl_stop);
1039         cfs_waitq_init(&tctl->tctl_waitq);
1040         atomic_set(&tctl->tctl_shutdown, 0);
1041
1042         if (cfs_kernel_thread(tracefiled, tctl, 0) < 0) {
1043                 rc = -ECHILD;
1044                 goto out;
1045         }
1046
1047         wait_for_completion(&tctl->tctl_start);
1048         thread_running = 1;
1049 out:
1050         mutex_up(&trace_thread_sem);
1051         return rc;
1052 }
1053
1054 void trace_stop_thread(void)
1055 {
1056         struct tracefiled_ctl *tctl = &trace_tctl;
1057
1058         mutex_down(&trace_thread_sem);
1059         if (thread_running) {
1060                 printk(KERN_INFO "Lustre: shutting down debug daemon thread...\n");
1061                 atomic_set(&tctl->tctl_shutdown, 1);
1062                 wait_for_completion(&tctl->tctl_stop);
1063                 thread_running = 0;
1064         }
1065         mutex_up(&trace_thread_sem);
1066 }
1067
1068 int tracefile_init(int max_pages)
1069 {
1070         struct trace_cpu_data *tcd;
1071         int                    i;
1072         int                    j;
1073         int                    rc;
1074         int                    factor;
1075
1076         rc = tracefile_init_arch();
1077         if (rc != 0)
1078                 return rc;
1079
1080         tcd_for_each(tcd, i, j) {
1081                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1082                 factor = tcd->tcd_pages_factor;
1083                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
1084                 CFS_INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1085                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1086                 tcd->tcd_cur_pages = 0;
1087                 tcd->tcd_cur_stock_pages = 0;
1088                 tcd->tcd_cur_daemon_pages = 0;
1089                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1090                 LASSERT(tcd->tcd_max_pages > 0);
1091                 tcd->tcd_shutting_down = 0;
1092         }
1093
1094         return 0;
1095 }
1096
1097 static void trace_cleanup_on_cpu(void *info)
1098 {
1099         struct trace_cpu_data *tcd;
1100         struct trace_page *tage;
1101         struct trace_page *tmp;
1102         int i;
1103
1104         tcd_for_each_type_lock(tcd, i) {
1105                 tcd->tcd_shutting_down = 1;
1106
1107                 list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1108                         __LASSERT_TAGE_INVARIANT(tage);
1109
1110                         list_del(&tage->linkage);
1111                         tage_free(tage);
1112                 }
1113                 tcd->tcd_cur_pages = 0;
1114         }
1115 }
1116
1117 static void trace_cleanup(void)
1118 {
1119         struct page_collection pc;
1120
1121         CFS_INIT_LIST_HEAD(&pc.pc_pages);
1122         spin_lock_init(&pc.pc_lock);
1123
1124         trace_call_on_all_cpus(trace_cleanup_on_cpu, &pc);
1125
1126         tracefile_fini_arch();
1127 }
1128
1129 void tracefile_exit(void)
1130 {
1131         trace_stop_thread();
1132         trace_cleanup();
1133 }