drivers/staging/lustre/lustre/libcfs/tracefile.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, Intel Corporation.
  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 "../../include/linux/libcfs/libcfs.h"
  48
  49 /* XXX move things up to the top, comment */
  50 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  51
  52 char cfs_tracefile[TRACEFILE_NAME_SIZE];
  53 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
  54 static struct tracefiled_ctl trace_tctl;
  55 struct mutex cfs_trace_thread_mutex;
  56 static int thread_running = 0;
  57
  58 atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
  59
  60 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  61                                          struct cfs_trace_cpu_data *tcd);
  62
  63 static inline struct cfs_trace_page *
  64 cfs_tage_from_list(struct list_head *list)
  65 {
  66         return list_entry(list, struct cfs_trace_page, linkage);
  67 }
  68
  69 static struct cfs_trace_page *cfs_tage_alloc(gfp_t gfp)
  70 {
  71         struct page         *page;
  72         struct cfs_trace_page *tage;
  73
  74         /* My caller is trying to free memory */
  75         if (!in_interrupt() && 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 |= __GFP_NOWARN;
  83         page = alloc_page(gfp);
  84         if (page == NULL)
  85                 return NULL;
  86
  87         tage = kmalloc(sizeof(*tage), gfp);
  88         if (tage == NULL) {
  89                 __free_page(page);
  90                 return NULL;
  91         }
  92
  93         tage->page = page;
  94         atomic_inc(&cfs_tage_allocated);
  95         return tage;
  96 }
  97
  98 static void cfs_tage_free(struct cfs_trace_page *tage)
  99 {
 100         __LASSERT(tage != NULL);
 101         __LASSERT(tage->page != NULL);
 102
 103         __free_page(tage->page);
 104         kfree(tage);
 105         atomic_dec(&cfs_tage_allocated);
 106 }
 107
 108 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
 109                              struct list_head *queue)
 110 {
 111         __LASSERT(tage != NULL);
 112         __LASSERT(queue != NULL);
 113
 114         list_move_tail(&tage->linkage, queue);
 115 }
 116
 117 int cfs_trace_refill_stock(struct cfs_trace_cpu_data *tcd, gfp_t gfp,
 118                            struct list_head *stock)
 119 {
 120         int i;
 121
 122         /*
 123          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 124          * from here: this will lead to infinite recursion.
 125          */
 126
 127         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++ i) {
 128                 struct cfs_trace_page *tage;
 129
 130                 tage = cfs_tage_alloc(gfp);
 131                 if (tage == NULL)
 132                         break;
 133                 list_add_tail(&tage->linkage, stock);
 134         }
 135         return i;
 136 }
 137
 138 /* return a page that has 'len' bytes left at the end */
 139 static struct cfs_trace_page *
 140 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
 141 {
 142         struct cfs_trace_page *tage;
 143
 144         if (tcd->tcd_cur_pages > 0) {
 145                 __LASSERT(!list_empty(&tcd->tcd_pages));
 146                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
 147                 if (tage->used + len <= PAGE_CACHE_SIZE)
 148                         return tage;
 149         }
 150
 151         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 152                 if (tcd->tcd_cur_stock_pages > 0) {
 153                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
 154                         --tcd->tcd_cur_stock_pages;
 155                         list_del_init(&tage->linkage);
 156                 } else {
 157                         tage = cfs_tage_alloc(GFP_ATOMIC);
 158                         if (unlikely(tage == NULL)) {
 159                                 if ((!memory_pressure_get() ||
 160                                      in_interrupt()) && printk_ratelimit())
 161                                         printk(KERN_WARNING
 162                                                "cannot allocate a tage (%ld)\n",
 163                                                tcd->tcd_cur_pages);
 164                                 return NULL;
 165                         }
 166                 }
 167
 168                 tage->used = 0;
 169                 tage->cpu = smp_processor_id();
 170                 tage->type = tcd->tcd_type;
 171                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 172                 tcd->tcd_cur_pages++;
 173
 174                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 175                         struct tracefiled_ctl *tctl = &trace_tctl;
 176                         /*
 177                          * wake up tracefiled to process some pages.
 178                          */
 179                         wake_up(&tctl->tctl_waitq);
 180                 }
 181                 return tage;
 182         }
 183         return NULL;
 184 }
 185
 186 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
 187 {
 188         int pgcount = tcd->tcd_cur_pages / 10;
 189         struct page_collection pc;
 190         struct cfs_trace_page *tage;
 191         struct cfs_trace_page *tmp;
 192
 193         /*
 194          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 195          * from here: this will lead to infinite recursion.
 196          */
 197
 198         if (printk_ratelimit())
 199                 printk(KERN_WARNING "debug daemon buffer overflowed; discarding 10%% of pages (%d of %ld)\n",
 200                        pgcount + 1, tcd->tcd_cur_pages);
 201
 202         INIT_LIST_HEAD(&pc.pc_pages);
 203         spin_lock_init(&pc.pc_lock);
 204
 205         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 206                 if (pgcount-- == 0)
 207                         break;
 208
 209                 list_move_tail(&tage->linkage, &pc.pc_pages);
 210                 tcd->tcd_cur_pages--;
 211         }
 212         put_pages_on_tcd_daemon_list(&pc, tcd);
 213 }
 214
 215 /* return a page that has 'len' bytes left at the end */
 216 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 217                                                  unsigned long len)
 218 {
 219         struct cfs_trace_page *tage;
 220
 221         /*
 222          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 223          * from here: this will lead to infinite recursion.
 224          */
 225
 226         if (len > PAGE_CACHE_SIZE) {
 227                 printk(KERN_ERR
 228                        "cowardly refusing to write %lu bytes in a page\n", len);
 229                 return NULL;
 230         }
 231
 232         tage = cfs_trace_get_tage_try(tcd, len);
 233         if (tage != NULL)
 234                 return tage;
 235         if (thread_running)
 236                 cfs_tcd_shrink(tcd);
 237         if (tcd->tcd_cur_pages > 0) {
 238                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 239                 tage->used = 0;
 240                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 241         }
 242         return tage;
 243 }
 244
 245 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 246                      const char *format, ...)
 247 {
 248         va_list args;
 249         int     rc;
 250
 251         va_start(args, format);
 252         rc = libcfs_debug_vmsg2(msgdata, format, args, NULL);
 253         va_end(args);
 254
 255         return rc;
 256 }
 257 EXPORT_SYMBOL(libcfs_debug_msg);
 258
 259 int libcfs_debug_vmsg2(struct libcfs_debug_msg_data *msgdata,
 260                        const char *format1, va_list args,
 261                        const char *format2, ...)
 262 {
 263         struct cfs_trace_cpu_data *tcd = NULL;
 264         struct ptldebug_header     header = {0};
 265         struct cfs_trace_page     *tage;
 266         /* string_buf is used only if tcd != NULL, and is always set then */
 267         char                  *string_buf = NULL;
 268         char                  *debug_buf;
 269         int                     known_size;
 270         int                     needed = 85; /* average message length */
 271         int                     max_nob;
 272         va_list             ap;
 273         int                     depth;
 274         int                     i;
 275         int                     remain;
 276         int                     mask = msgdata->msg_mask;
 277         const char              *file = kbasename(msgdata->msg_file);
 278         struct cfs_debug_limit_state   *cdls = msgdata->msg_cdls;
 279
 280         tcd = cfs_trace_get_tcd();
 281
 282         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 283          * pins us to a particular CPU.  This avoids an smp_processor_id()
 284          * warning on Linux when debugging is enabled. */
 285         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 286
 287         if (tcd == NULL)                /* arch may not log in IRQ context */
 288                 goto console;
 289
 290         if (tcd->tcd_cur_pages == 0)
 291                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 292
 293         if (tcd->tcd_shutting_down) {
 294                 cfs_trace_put_tcd(tcd);
 295                 tcd = NULL;
 296                 goto console;
 297         }
 298
 299         depth = __current_nesting_level();
 300         known_size = strlen(file) + 1 + depth;
 301         if (msgdata->msg_fn)
 302                 known_size += strlen(msgdata->msg_fn) + 1;
 303
 304         if (libcfs_debug_binary)
 305                 known_size += sizeof(header);
 306
 307         /*/
 308          * '2' used because vsnprintf return real size required for output
 309          * _without_ terminating NULL.
 310          * if needed is to small for this format.
 311          */
 312         for (i = 0; i < 2; i++) {
 313                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 314                 if (tage == NULL) {
 315                         if (needed + known_size > PAGE_CACHE_SIZE)
 316                                 mask |= D_ERROR;
 317
 318                         cfs_trace_put_tcd(tcd);
 319                         tcd = NULL;
 320                         goto console;
 321                 }
 322
 323                 string_buf = (char *)page_address(tage->page) +
 324                                         tage->used + known_size;
 325
 326                 max_nob = PAGE_CACHE_SIZE - tage->used - known_size;
 327                 if (max_nob <= 0) {
 328                         printk(KERN_EMERG "negative max_nob: %d\n",
 329                                max_nob);
 330                         mask |= D_ERROR;
 331                         cfs_trace_put_tcd(tcd);
 332                         tcd = NULL;
 333                         goto console;
 334                 }
 335
 336                 needed = 0;
 337                 if (format1) {
 338                         va_copy(ap, args);
 339                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 340                         va_end(ap);
 341                 }
 342
 343                 if (format2) {
 344                         remain = max_nob - needed;
 345                         if (remain < 0)
 346                                 remain = 0;
 347
 348                         va_start(ap, format2);
 349                         needed += vsnprintf(string_buf + needed, remain,
 350                                             format2, ap);
 351                         va_end(ap);
 352                 }
 353
 354                 if (needed < max_nob) /* well. printing ok.. */
 355                         break;
 356         }
 357
 358         if (*(string_buf+needed-1) != '\n')
 359                 printk(KERN_INFO "format at %s:%d:%s doesn't end in newline\n",
 360                        file, msgdata->msg_line, msgdata->msg_fn);
 361
 362         header.ph_len = known_size + needed;
 363         debug_buf = (char *)page_address(tage->page) + tage->used;
 364
 365         if (libcfs_debug_binary) {
 366                 memcpy(debug_buf, &header, sizeof(header));
 367                 tage->used += sizeof(header);
 368                 debug_buf += sizeof(header);
 369         }
 370
 371         /* indent message according to the nesting level */
 372         while (depth-- > 0) {
 373                 *(debug_buf++) = '.';
 374                 ++ tage->used;
 375         }
 376
 377         strcpy(debug_buf, file);
 378         tage->used += strlen(file) + 1;
 379         debug_buf += strlen(file) + 1;
 380
 381         if (msgdata->msg_fn) {
 382                 strcpy(debug_buf, msgdata->msg_fn);
 383                 tage->used += strlen(msgdata->msg_fn) + 1;
 384                 debug_buf += strlen(msgdata->msg_fn) + 1;
 385         }
 386
 387         __LASSERT(debug_buf == string_buf);
 388
 389         tage->used += needed;
 390         __LASSERT (tage->used <= PAGE_CACHE_SIZE);
 391
 392 console:
 393         if ((mask & libcfs_printk) == 0) {
 394                 /* no console output requested */
 395                 if (tcd != NULL)
 396                         cfs_trace_put_tcd(tcd);
 397                 return 1;
 398         }
 399
 400         if (cdls != NULL) {
 401                 if (libcfs_console_ratelimit &&
 402                     cdls->cdls_next != 0 &&     /* not first time ever */
 403                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 404                         /* skipping a console message */
 405                         cdls->cdls_count++;
 406                         if (tcd != NULL)
 407                                 cfs_trace_put_tcd(tcd);
 408                         return 1;
 409                 }
 410
 411                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 412                                                        libcfs_console_max_delay
 413                                                        + cfs_time_seconds(10))) {
 414                         /* last timeout was a long time ago */
 415                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 416                 } else {
 417                         cdls->cdls_delay *= libcfs_console_backoff;
 418                 }
 419
 420                 if (cdls->cdls_delay < libcfs_console_min_delay)
 421                         cdls->cdls_delay = libcfs_console_min_delay;
 422                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 423                         cdls->cdls_delay = libcfs_console_max_delay;
 424
 425                 /* ensure cdls_next is never zero after it's been seen */
 426                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 427         }
 428
 429         if (tcd != NULL) {
 430                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 431                                      msgdata->msg_fn);
 432                 cfs_trace_put_tcd(tcd);
 433         } else {
 434                 string_buf = cfs_trace_get_console_buffer();
 435
 436                 needed = 0;
 437                 if (format1 != NULL) {
 438                         va_copy(ap, args);
 439                         needed = vsnprintf(string_buf,
 440                                            CFS_TRACE_CONSOLE_BUFFER_SIZE,
 441                                            format1, ap);
 442                         va_end(ap);
 443                 }
 444                 if (format2 != NULL) {
 445                         remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
 446                         if (remain > 0) {
 447                                 va_start(ap, format2);
 448                                 needed += vsnprintf(string_buf+needed, remain,
 449                                                     format2, ap);
 450                                 va_end(ap);
 451                         }
 452                 }
 453                 cfs_print_to_console(&header, mask,
 454                                      string_buf, needed, file, msgdata->msg_fn);
 455
 456                 cfs_trace_put_console_buffer(string_buf);
 457         }
 458
 459         if (cdls != NULL && cdls->cdls_count != 0) {
 460                 string_buf = cfs_trace_get_console_buffer();
 461
 462                 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 463                                   "Skipped %d previous similar message%s\n",
 464                                   cdls->cdls_count,
 465                                   (cdls->cdls_count > 1) ? "s" : "");
 466
 467                 cfs_print_to_console(&header, mask,
 468                                      string_buf, needed, file, msgdata->msg_fn);
 469
 470                 cfs_trace_put_console_buffer(string_buf);
 471                 cdls->cdls_count = 0;
 472         }
 473
 474         return 0;
 475 }
 476 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 477
 478 void
 479 cfs_trace_assertion_failed(const char *str,
 480                            struct libcfs_debug_msg_data *msgdata)
 481 {
 482         struct ptldebug_header hdr;
 483
 484         libcfs_panic_in_progress = 1;
 485         libcfs_catastrophe = 1;
 486         mb();
 487
 488         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 489
 490         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 491                              msgdata->msg_file, msgdata->msg_fn);
 492
 493         panic("Lustre debug assertion failure\n");
 494
 495         /* not reached */
 496 }
 497
 498 static void
 499 panic_collect_pages(struct page_collection *pc)
 500 {
 501         /* Do the collect_pages job on a single CPU: assumes that all other
 502          * CPUs have been stopped during a panic.  If this isn't true for some
 503          * arch, this will have to be implemented separately in each arch.  */
 504         int                     i;
 505         int                     j;
 506         struct cfs_trace_cpu_data *tcd;
 507
 508         INIT_LIST_HEAD(&pc->pc_pages);
 509
 510         cfs_tcd_for_each(tcd, i, j) {
 511                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 512                 tcd->tcd_cur_pages = 0;
 513
 514                 if (pc->pc_want_daemon_pages) {
 515                         list_splice_init(&tcd->tcd_daemon_pages,
 516                                              &pc->pc_pages);
 517                         tcd->tcd_cur_daemon_pages = 0;
 518                 }
 519         }
 520 }
 521
 522 static void collect_pages_on_all_cpus(struct page_collection *pc)
 523 {
 524         struct cfs_trace_cpu_data *tcd;
 525         int i, cpu;
 526
 527         spin_lock(&pc->pc_lock);
 528         for_each_possible_cpu(cpu) {
 529                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 530                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 531                         tcd->tcd_cur_pages = 0;
 532                         if (pc->pc_want_daemon_pages) {
 533                                 list_splice_init(&tcd->tcd_daemon_pages,
 534                                                      &pc->pc_pages);
 535                                 tcd->tcd_cur_daemon_pages = 0;
 536                         }
 537                 }
 538         }
 539         spin_unlock(&pc->pc_lock);
 540 }
 541
 542 static void collect_pages(struct page_collection *pc)
 543 {
 544         INIT_LIST_HEAD(&pc->pc_pages);
 545
 546         if (libcfs_panic_in_progress)
 547                 panic_collect_pages(pc);
 548         else
 549                 collect_pages_on_all_cpus(pc);
 550 }
 551
 552 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 553 {
 554         struct cfs_trace_cpu_data *tcd;
 555         struct list_head *cur_head;
 556         struct cfs_trace_page *tage;
 557         struct cfs_trace_page *tmp;
 558         int i, cpu;
 559
 560         spin_lock(&pc->pc_lock);
 561         for_each_possible_cpu(cpu) {
 562                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 563                         cur_head = tcd->tcd_pages.next;
 564
 565                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 566                                                  linkage) {
 567
 568                                 __LASSERT_TAGE_INVARIANT(tage);
 569
 570                                 if (tage->cpu != cpu || tage->type != i)
 571                                         continue;
 572
 573                                 cfs_tage_to_tail(tage, cur_head);
 574                                 tcd->tcd_cur_pages++;
 575                         }
 576                 }
 577         }
 578         spin_unlock(&pc->pc_lock);
 579 }
 580
 581 static void put_pages_back(struct page_collection *pc)
 582 {
 583         if (!libcfs_panic_in_progress)
 584                 put_pages_back_on_all_cpus(pc);
 585 }
 586
 587 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 588  * we have a good amount of data at all times for dumping during an LBUG, even
 589  * if we have been steadily writing (and otherwise discarding) pages via the
 590  * debug daemon. */
 591 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 592                                          struct cfs_trace_cpu_data *tcd)
 593 {
 594         struct cfs_trace_page *tage;
 595         struct cfs_trace_page *tmp;
 596
 597         spin_lock(&pc->pc_lock);
 598         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 599
 600                 __LASSERT_TAGE_INVARIANT(tage);
 601
 602                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 603                         continue;
 604
 605                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 606                 tcd->tcd_cur_daemon_pages++;
 607
 608                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 609                         struct cfs_trace_page *victim;
 610
 611                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 612                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 613
 614                         __LASSERT_TAGE_INVARIANT(victim);
 615
 616                         list_del(&victim->linkage);
 617                         cfs_tage_free(victim);
 618                         tcd->tcd_cur_daemon_pages--;
 619                 }
 620         }
 621         spin_unlock(&pc->pc_lock);
 622 }
 623
 624 static void put_pages_on_daemon_list(struct page_collection *pc)
 625 {
 626         struct cfs_trace_cpu_data *tcd;
 627         int i, cpu;
 628
 629         for_each_possible_cpu(cpu) {
 630                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 631                         put_pages_on_tcd_daemon_list(pc, tcd);
 632         }
 633 }
 634
 635 void cfs_trace_debug_print(void)
 636 {
 637         struct page_collection pc;
 638         struct cfs_trace_page *tage;
 639         struct cfs_trace_page *tmp;
 640
 641         spin_lock_init(&pc.pc_lock);
 642
 643         pc.pc_want_daemon_pages = 1;
 644         collect_pages(&pc);
 645         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 646                 char *p, *file, *fn;
 647                 struct page *page;
 648
 649                 __LASSERT_TAGE_INVARIANT(tage);
 650
 651                 page = tage->page;
 652                 p = page_address(page);
 653                 while (p < ((char *)page_address(page) + tage->used)) {
 654                         struct ptldebug_header *hdr;
 655                         int len;
 656                         hdr = (void *)p;
 657                         p += sizeof(*hdr);
 658                         file = p;
 659                         p += strlen(file) + 1;
 660                         fn = p;
 661                         p += strlen(fn) + 1;
 662                         len = hdr->ph_len - (int)(p - (char *)hdr);
 663
 664                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 665
 666                         p += len;
 667                 }
 668
 669                 list_del(&tage->linkage);
 670                 cfs_tage_free(tage);
 671         }
 672 }
 673
 674 int cfs_tracefile_dump_all_pages(char *filename)
 675 {
 676         struct page_collection  pc;
 677         struct file             *filp;
 678         struct cfs_trace_page   *tage;
 679         struct cfs_trace_page   *tmp;
 680         char                    *buf;
 681         int rc;
 682
 683         DECL_MMSPACE;
 684
 685         cfs_tracefile_write_lock();
 686
 687         filp = filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600);
 688         if (IS_ERR(filp)) {
 689                 rc = PTR_ERR(filp);
 690                 filp = NULL;
 691                 printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 692                       filename, rc);
 693                 goto out;
 694         }
 695
 696         spin_lock_init(&pc.pc_lock);
 697         pc.pc_want_daemon_pages = 1;
 698         collect_pages(&pc);
 699         if (list_empty(&pc.pc_pages)) {
 700                 rc = 0;
 701                 goto close;
 702         }
 703
 704         /* ok, for now, just write the pages.  in the future we'll be building
 705          * iobufs with the pages and calling generic_direct_IO */
 706         MMSPACE_OPEN;
 707         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 708
 709                 __LASSERT_TAGE_INVARIANT(tage);
 710
 711                 buf = kmap(tage->page);
 712                 rc = vfs_write(filp, (__force const char __user *)buf,
 713                                tage->used, &filp->f_pos);
 714                 kunmap(tage->page);
 715
 716                 if (rc != (int)tage->used) {
 717                         printk(KERN_WARNING "wanted to write %u but wrote %d\n",
 718                                tage->used, rc);
 719                         put_pages_back(&pc);
 720                         __LASSERT(list_empty(&pc.pc_pages));
 721                         break;
 722                 }
 723                 list_del(&tage->linkage);
 724                 cfs_tage_free(tage);
 725         }
 726         MMSPACE_CLOSE;
 727         rc = vfs_fsync(filp, 1);
 728         if (rc)
 729                 printk(KERN_ERR "sync returns %d\n", rc);
 730 close:
 731         filp_close(filp, NULL);
 732 out:
 733         cfs_tracefile_write_unlock();
 734         return rc;
 735 }
 736
 737 void cfs_trace_flush_pages(void)
 738 {
 739         struct page_collection pc;
 740         struct cfs_trace_page *tage;
 741         struct cfs_trace_page *tmp;
 742
 743         spin_lock_init(&pc.pc_lock);
 744
 745         pc.pc_want_daemon_pages = 1;
 746         collect_pages(&pc);
 747         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 748
 749                 __LASSERT_TAGE_INVARIANT(tage);
 750
 751                 list_del(&tage->linkage);
 752                 cfs_tage_free(tage);
 753         }
 754 }
 755
 756 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 757                             const char __user *usr_buffer, int usr_buffer_nob)
 758 {
 759         int    nob;
 760
 761         if (usr_buffer_nob > knl_buffer_nob)
 762                 return -EOVERFLOW;
 763
 764         if (copy_from_user((void *)knl_buffer,
 765                            usr_buffer, usr_buffer_nob))
 766                 return -EFAULT;
 767
 768         nob = strnlen(knl_buffer, usr_buffer_nob);
 769         while (nob-- >= 0)                    /* strip trailing whitespace */
 770                 if (!isspace(knl_buffer[nob]))
 771                         break;
 772
 773         if (nob < 0)                        /* empty string */
 774                 return -EINVAL;
 775
 776         if (nob == knl_buffer_nob)            /* no space to terminate */
 777                 return -EOVERFLOW;
 778
 779         knl_buffer[nob + 1] = 0;                /* terminate */
 780         return 0;
 781 }
 782 EXPORT_SYMBOL(cfs_trace_copyin_string);
 783
 784 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 785                              const char *knl_buffer, char *append)
 786 {
 787         /* NB if 'append' != NULL, it's a single character to append to the
 788          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 789          * terminating zero byte) for sysctl entries */
 790         int   nob = strlen(knl_buffer);
 791
 792         if (nob > usr_buffer_nob)
 793                 nob = usr_buffer_nob;
 794
 795         if (copy_to_user(usr_buffer, knl_buffer, nob))
 796                 return -EFAULT;
 797
 798         if (append != NULL && nob < usr_buffer_nob) {
 799                 if (copy_to_user(usr_buffer + nob, append, 1))
 800                         return -EFAULT;
 801
 802                 nob++;
 803         }
 804
 805         return nob;
 806 }
 807 EXPORT_SYMBOL(cfs_trace_copyout_string);
 808
 809 int cfs_trace_allocate_string_buffer(char **str, int nob)
 810 {
 811         if (nob > 2 * PAGE_CACHE_SIZE)      /* string must be "sensible" */
 812                 return -EINVAL;
 813
 814         *str = kmalloc(nob, GFP_IOFS | __GFP_ZERO);
 815         if (*str == NULL)
 816                 return -ENOMEM;
 817
 818         return 0;
 819 }
 820
 821 void cfs_trace_free_string_buffer(char *str, int nob)
 822 {
 823         kfree(str);
 824 }
 825
 826 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
 827 {
 828         char     *str;
 829         int        rc;
 830
 831         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 832         if (rc != 0)
 833                 return rc;
 834
 835         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 836                                      usr_str, usr_str_nob);
 837         if (rc != 0)
 838                 goto out;
 839
 840         if (str[0] != '/') {
 841                 rc = -EINVAL;
 842                 goto out;
 843         }
 844         rc = cfs_tracefile_dump_all_pages(str);
 845 out:
 846         cfs_trace_free_string_buffer(str, usr_str_nob + 1);
 847         return rc;
 848 }
 849
 850 int cfs_trace_daemon_command(char *str)
 851 {
 852         int       rc = 0;
 853
 854         cfs_tracefile_write_lock();
 855
 856         if (strcmp(str, "stop") == 0) {
 857                 cfs_tracefile_write_unlock();
 858                 cfs_trace_stop_thread();
 859                 cfs_tracefile_write_lock();
 860                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 861
 862         } else if (strncmp(str, "size=", 5) == 0) {
 863                 cfs_tracefile_size = simple_strtoul(str + 5, NULL, 0);
 864                 if (cfs_tracefile_size < 10 || cfs_tracefile_size > 20480)
 865                         cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 866                 else
 867                         cfs_tracefile_size <<= 20;
 868
 869         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 870                 rc = -ENAMETOOLONG;
 871         } else if (str[0] != '/') {
 872                 rc = -EINVAL;
 873         } else {
 874                 strcpy(cfs_tracefile, str);
 875
 876                 printk(KERN_INFO
 877                        "Lustre: debug daemon will attempt to start writing to %s (%lukB max)\n",
 878                        cfs_tracefile,
 879                        (long)(cfs_tracefile_size >> 10));
 880
 881                 cfs_trace_start_thread();
 882         }
 883
 884         cfs_tracefile_write_unlock();
 885         return rc;
 886 }
 887
 888 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
 889 {
 890         char *str;
 891         int   rc;
 892
 893         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 894         if (rc != 0)
 895                 return rc;
 896
 897         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 898                                  usr_str, usr_str_nob);
 899         if (rc == 0)
 900                 rc = cfs_trace_daemon_command(str);
 901
 902         cfs_trace_free_string_buffer(str, usr_str_nob + 1);
 903         return rc;
 904 }
 905
 906 int cfs_trace_set_debug_mb(int mb)
 907 {
 908         int i;
 909         int j;
 910         int pages;
 911         int limit = cfs_trace_max_debug_mb();
 912         struct cfs_trace_cpu_data *tcd;
 913
 914         if (mb < num_possible_cpus()) {
 915                 printk(KERN_WARNING
 916                        "Lustre: %d MB is too small for debug buffer size, setting it to %d MB.\n",
 917                        mb, num_possible_cpus());
 918                 mb = num_possible_cpus();
 919         }
 920
 921         if (mb > limit) {
 922                 printk(KERN_WARNING
 923                        "Lustre: %d MB is too large for debug buffer size, setting it to %d MB.\n",
 924                        mb, limit);
 925                 mb = limit;
 926         }
 927
 928         mb /= num_possible_cpus();
 929         pages = mb << (20 - PAGE_CACHE_SHIFT);
 930
 931         cfs_tracefile_write_lock();
 932
 933         cfs_tcd_for_each(tcd, i, j)
 934                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 935
 936         cfs_tracefile_write_unlock();
 937
 938         return 0;
 939 }
 940
 941 int cfs_trace_set_debug_mb_usrstr(void __user *usr_str, int usr_str_nob)
 942 {
 943         char     str[32];
 944         int      rc;
 945
 946         rc = cfs_trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 947         if (rc < 0)
 948                 return rc;
 949
 950         return cfs_trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 951 }
 952
 953 int cfs_trace_get_debug_mb(void)
 954 {
 955         int i;
 956         int j;
 957         struct cfs_trace_cpu_data *tcd;
 958         int total_pages = 0;
 959
 960         cfs_tracefile_read_lock();
 961
 962         cfs_tcd_for_each(tcd, i, j)
 963                 total_pages += tcd->tcd_max_pages;
 964
 965         cfs_tracefile_read_unlock();
 966
 967         return (total_pages >> (20 - PAGE_CACHE_SHIFT)) + 1;
 968 }
 969
 970 static int tracefiled(void *arg)
 971 {
 972         struct page_collection pc;
 973         struct tracefiled_ctl *tctl = arg;
 974         struct cfs_trace_page *tage;
 975         struct cfs_trace_page *tmp;
 976         struct file *filp;
 977         char *buf;
 978         int last_loop = 0;
 979         int rc;
 980
 981         DECL_MMSPACE;
 982
 983         /* we're started late enough that we pick up init's fs context */
 984         /* this is so broken in uml?  what on earth is going on? */
 985
 986         spin_lock_init(&pc.pc_lock);
 987         complete(&tctl->tctl_start);
 988
 989         while (1) {
 990                 wait_queue_t __wait;
 991
 992                 pc.pc_want_daemon_pages = 0;
 993                 collect_pages(&pc);
 994                 if (list_empty(&pc.pc_pages))
 995                         goto end_loop;
 996
 997                 filp = NULL;
 998                 cfs_tracefile_read_lock();
 999                 if (cfs_tracefile[0] != 0) {
1000                         filp = filp_open(cfs_tracefile,
1001                                          O_CREAT | O_RDWR | O_LARGEFILE,
1002                                          0600);
1003                         if (IS_ERR(filp)) {
1004                                 rc = PTR_ERR(filp);
1005                                 filp = NULL;
1006                                 printk(KERN_WARNING "couldn't open %s: %d\n",
1007                                        cfs_tracefile, rc);
1008                         }
1009                 }
1010                 cfs_tracefile_read_unlock();
1011                 if (filp == NULL) {
1012                         put_pages_on_daemon_list(&pc);
1013                         __LASSERT(list_empty(&pc.pc_pages));
1014                         goto end_loop;
1015                 }
1016
1017                 MMSPACE_OPEN;
1018
1019                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1020                                                    linkage) {
1021                         static loff_t f_pos;
1022
1023                         __LASSERT_TAGE_INVARIANT(tage);
1024
1025                         if (f_pos >= (off_t)cfs_tracefile_size)
1026                                 f_pos = 0;
1027                         else if (f_pos > i_size_read(filp->f_dentry->d_inode))
1028                                 f_pos = i_size_read(filp->f_dentry->d_inode);
1029
1030                         buf = kmap(tage->page);
1031                         rc = vfs_write(filp, (__force const char __user *)buf,
1032                                        tage->used, &f_pos);
1033                         kunmap(tage->page);
1034
1035                         if (rc != (int)tage->used) {
1036                                 printk(KERN_WARNING "wanted to write %u but wrote %d\n",
1037                                        tage->used, rc);
1038                                 put_pages_back(&pc);
1039                                 __LASSERT(list_empty(&pc.pc_pages));
1040                         }
1041                 }
1042                 MMSPACE_CLOSE;
1043
1044                 filp_close(filp, NULL);
1045                 put_pages_on_daemon_list(&pc);
1046                 if (!list_empty(&pc.pc_pages)) {
1047                         int i;
1048
1049                         printk(KERN_ALERT "Lustre: trace pages aren't empty\n");
1050                         printk(KERN_ERR "total cpus(%d): ",
1051                                num_possible_cpus());
1052                         for (i = 0; i < num_possible_cpus(); i++)
1053                                 if (cpu_online(i))
1054                                         printk(KERN_ERR "%d(on) ", i);
1055                                 else
1056                                         printk(KERN_ERR "%d(off) ", i);
1057                         printk(KERN_ERR "\n");
1058
1059                         i = 0;
1060                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1061                                                      linkage)
1062                                 printk(KERN_ERR "page %d belongs to cpu %d\n",
1063                                        ++i, tage->cpu);
1064                         printk(KERN_ERR "There are %d pages unwritten\n",
1065                                i);
1066                 }
1067                 __LASSERT(list_empty(&pc.pc_pages));
1068 end_loop:
1069                 if (atomic_read(&tctl->tctl_shutdown)) {
1070                         if (last_loop == 0) {
1071                                 last_loop = 1;
1072                                 continue;
1073                         } else {
1074                                 break;
1075                         }
1076                 }
1077                 init_waitqueue_entry(&__wait, current);
1078                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1079                 set_current_state(TASK_INTERRUPTIBLE);
1080                 schedule_timeout(cfs_time_seconds(1));
1081                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1082         }
1083         complete(&tctl->tctl_stop);
1084         return 0;
1085 }
1086
1087 int cfs_trace_start_thread(void)
1088 {
1089         struct tracefiled_ctl *tctl = &trace_tctl;
1090         int rc = 0;
1091
1092         mutex_lock(&cfs_trace_thread_mutex);
1093         if (thread_running)
1094                 goto out;
1095
1096         init_completion(&tctl->tctl_start);
1097         init_completion(&tctl->tctl_stop);
1098         init_waitqueue_head(&tctl->tctl_waitq);
1099         atomic_set(&tctl->tctl_shutdown, 0);
1100
1101         if (IS_ERR(kthread_run(tracefiled, tctl, "ktracefiled"))) {
1102                 rc = -ECHILD;
1103                 goto out;
1104         }
1105
1106         wait_for_completion(&tctl->tctl_start);
1107         thread_running = 1;
1108 out:
1109         mutex_unlock(&cfs_trace_thread_mutex);
1110         return rc;
1111 }
1112
1113 void cfs_trace_stop_thread(void)
1114 {
1115         struct tracefiled_ctl *tctl = &trace_tctl;
1116
1117         mutex_lock(&cfs_trace_thread_mutex);
1118         if (thread_running) {
1119                 printk(KERN_INFO
1120                        "Lustre: shutting down debug daemon thread...\n");
1121                 atomic_set(&tctl->tctl_shutdown, 1);
1122                 wait_for_completion(&tctl->tctl_stop);
1123                 thread_running = 0;
1124         }
1125         mutex_unlock(&cfs_trace_thread_mutex);
1126 }
1127
1128 int cfs_tracefile_init(int max_pages)
1129 {
1130         struct cfs_trace_cpu_data *tcd;
1131         int                 i;
1132         int                 j;
1133         int                 rc;
1134         int                 factor;
1135
1136         rc = cfs_tracefile_init_arch();
1137         if (rc != 0)
1138                 return rc;
1139
1140         cfs_tcd_for_each(tcd, i, j) {
1141                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1142                 factor = tcd->tcd_pages_factor;
1143                 INIT_LIST_HEAD(&tcd->tcd_pages);
1144                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1145                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1146                 tcd->tcd_cur_pages = 0;
1147                 tcd->tcd_cur_stock_pages = 0;
1148                 tcd->tcd_cur_daemon_pages = 0;
1149                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1150                 LASSERT(tcd->tcd_max_pages > 0);
1151                 tcd->tcd_shutting_down = 0;
1152         }
1153
1154         return 0;
1155 }
1156
1157 static void trace_cleanup_on_all_cpus(void)
1158 {
1159         struct cfs_trace_cpu_data *tcd;
1160         struct cfs_trace_page *tage;
1161         struct cfs_trace_page *tmp;
1162         int i, cpu;
1163
1164         for_each_possible_cpu(cpu) {
1165                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1166                         tcd->tcd_shutting_down = 1;
1167
1168                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages,
1169                                                            linkage) {
1170                                 __LASSERT_TAGE_INVARIANT(tage);
1171
1172                                 list_del(&tage->linkage);
1173                                 cfs_tage_free(tage);
1174                         }
1175
1176                         tcd->tcd_cur_pages = 0;
1177                 }
1178         }
1179 }
1180
1181 static void cfs_trace_cleanup(void)
1182 {
1183         struct page_collection pc;
1184
1185         INIT_LIST_HEAD(&pc.pc_pages);
1186         spin_lock_init(&pc.pc_lock);
1187
1188         trace_cleanup_on_all_cpus();
1189
1190         cfs_tracefile_fini_arch();
1191 }
1192
1193 void cfs_tracefile_exit(void)
1194 {
1195         cfs_trace_stop_thread();
1196         cfs_trace_cleanup();
1197 }