libcfs/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, 2013, 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 <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(cfs_list_t *list)
  65 {
  66         return cfs_list_entry(list, struct cfs_trace_page, linkage);
  67 }
  68
  69 static struct cfs_trace_page *cfs_tage_alloc(int 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                              cfs_list_t *queue)
 110 {
 111         __LASSERT(tage != NULL);
 112         __LASSERT(queue != NULL);
 113
 114         cfs_list_move_tail(&tage->linkage, queue);
 115 }
 116
 117 int cfs_trace_refill_stock(struct cfs_trace_cpu_data *tcd, int gfp,
 118                            cfs_list_t *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                 cfs_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(!cfs_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                         cfs_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                 cfs_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; "
 200                        "discarding 10%% of pages (%d of %ld)\n",
 201                        pgcount + 1, tcd->tcd_cur_pages);
 202
 203         CFS_INIT_LIST_HEAD(&pc.pc_pages);
 204
 205         cfs_list_for_each_entry_safe_typed(tage, tmp, &tcd->tcd_pages,
 206                                            struct cfs_trace_page, linkage) {
 207                 if (pgcount-- == 0)
 208                         break;
 209
 210                 cfs_list_move_tail(&tage->linkage, &pc.pc_pages);
 211                 tcd->tcd_cur_pages--;
 212         }
 213         put_pages_on_tcd_daemon_list(&pc, tcd);
 214 }
 215
 216 /* return a page that has 'len' bytes left at the end */
 217 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 218                                                  unsigned long len)
 219 {
 220         struct cfs_trace_page *tage;
 221
 222         /*
 223          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 224          * from here: this will lead to infinite recursion.
 225          */
 226
 227         if (len > PAGE_CACHE_SIZE) {
 228                 printk(KERN_ERR
 229                        "cowardly refusing to write %lu bytes in a page\n", len);
 230                 return NULL;
 231         }
 232
 233         tage = cfs_trace_get_tage_try(tcd, len);
 234         if (tage != NULL)
 235                 return tage;
 236         if (thread_running)
 237                 cfs_tcd_shrink(tcd);
 238         if (tcd->tcd_cur_pages > 0) {
 239                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 240                 tage->used = 0;
 241                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 242         }
 243         return tage;
 244 }
 245
 246 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 247                      const char *format, ...)
 248 {
 249         va_list args;
 250         int     rc;
 251
 252         va_start(args, format);
 253         rc = libcfs_debug_vmsg2(msgdata, format, args, NULL);
 254         va_end(args);
 255
 256         return rc;
 257 }
 258 EXPORT_SYMBOL(libcfs_debug_msg);
 259
 260 int libcfs_debug_vmsg2(struct libcfs_debug_msg_data *msgdata,
 261                        const char *format1, va_list args,
 262                        const char *format2, ...)
 263 {
 264         struct cfs_trace_cpu_data *tcd = NULL;
 265         struct ptldebug_header     header = {0};
 266         struct cfs_trace_page     *tage;
 267         /* string_buf is used only if tcd != NULL, and is always set then */
 268         char                      *string_buf = NULL;
 269         char                      *debug_buf;
 270         int                        known_size;
 271         int                        needed = 85; /* average message length */
 272         int                        max_nob;
 273         va_list                    ap;
 274         int                        depth;
 275         int                        i;
 276         int                        remain;
 277         int                        mask = msgdata->msg_mask;
 278         char                      *file = (char *)msgdata->msg_file;
 279         cfs_debug_limit_state_t   *cdls = msgdata->msg_cdls;
 280
 281         if (strchr(file, '/'))
 282                 file = strrchr(file, '/') + 1;
 283
 284         tcd = cfs_trace_get_tcd();
 285
 286         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 287          * pins us to a particular CPU.  This avoids an smp_processor_id()
 288          * warning on Linux when debugging is enabled. */
 289         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 290
 291         if (tcd == NULL)                /* arch may not log in IRQ context */
 292                 goto console;
 293
 294         if (tcd->tcd_cur_pages == 0)
 295                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 296
 297         if (tcd->tcd_shutting_down) {
 298                 cfs_trace_put_tcd(tcd);
 299                 tcd = NULL;
 300                 goto console;
 301         }
 302
 303         depth = __current_nesting_level();
 304         known_size = strlen(file) + 1 + depth;
 305         if (msgdata->msg_fn)
 306                 known_size += strlen(msgdata->msg_fn) + 1;
 307
 308         if (libcfs_debug_binary)
 309                 known_size += sizeof(header);
 310
 311         /*/
 312          * '2' used because vsnprintf return real size required for output
 313          * _without_ terminating NULL.
 314          * if needed is to small for this format.
 315          */
 316         for (i = 0; i < 2; i++) {
 317                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 318                 if (tage == NULL) {
 319                         if (needed + known_size > PAGE_CACHE_SIZE)
 320                                 mask |= D_ERROR;
 321
 322                         cfs_trace_put_tcd(tcd);
 323                         tcd = NULL;
 324                         goto console;
 325                 }
 326
 327                 string_buf = (char *)page_address(tage->page) +
 328                                         tage->used + known_size;
 329
 330                 max_nob = PAGE_CACHE_SIZE - tage->used - known_size;
 331                 if (max_nob <= 0) {
 332                         printk(KERN_EMERG "negative max_nob: %d\n",
 333                                max_nob);
 334                         mask |= D_ERROR;
 335                         cfs_trace_put_tcd(tcd);
 336                         tcd = NULL;
 337                         goto console;
 338                 }
 339
 340                 needed = 0;
 341                 if (format1) {
 342                         va_copy(ap, args);
 343                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 344                         va_end(ap);
 345                 }
 346
 347                 if (format2) {
 348                         remain = max_nob - needed;
 349                         if (remain < 0)
 350                                 remain = 0;
 351
 352                         va_start(ap, format2);
 353                         needed += vsnprintf(string_buf + needed, remain,
 354                                             format2, ap);
 355                         va_end(ap);
 356                 }
 357
 358                 if (needed < max_nob) /* well. printing ok.. */
 359                         break;
 360         }
 361
 362         if (*(string_buf+needed-1) != '\n')
 363                 printk(KERN_INFO "format at %s:%d:%s doesn't end in "
 364                        "newline\n", file, msgdata->msg_line, msgdata->msg_fn);
 365
 366         header.ph_len = known_size + needed;
 367         debug_buf = (char *)page_address(tage->page) + tage->used;
 368
 369         if (libcfs_debug_binary) {
 370                 memcpy(debug_buf, &header, sizeof(header));
 371                 tage->used += sizeof(header);
 372                 debug_buf += sizeof(header);
 373         }
 374
 375         /* indent message according to the nesting level */
 376         while (depth-- > 0) {
 377                 *(debug_buf++) = '.';
 378                 ++ tage->used;
 379         }
 380
 381         strcpy(debug_buf, file);
 382         tage->used += strlen(file) + 1;
 383         debug_buf += strlen(file) + 1;
 384
 385         if (msgdata->msg_fn) {
 386                 strcpy(debug_buf, msgdata->msg_fn);
 387                 tage->used += strlen(msgdata->msg_fn) + 1;
 388                 debug_buf += strlen(msgdata->msg_fn) + 1;
 389         }
 390
 391         __LASSERT(debug_buf == string_buf);
 392
 393         tage->used += needed;
 394         __LASSERT(tage->used <= PAGE_CACHE_SIZE);
 395
 396 console:
 397         if ((mask & libcfs_printk) == 0) {
 398                 /* no console output requested */
 399                 if (tcd != NULL)
 400                         cfs_trace_put_tcd(tcd);
 401                 return 1;
 402         }
 403
 404         if (cdls != NULL) {
 405                 if (libcfs_console_ratelimit &&
 406                     cdls->cdls_next != 0 &&     /* not first time ever */
 407                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 408                         /* skipping a console message */
 409                         cdls->cdls_count++;
 410                         if (tcd != NULL)
 411                                 cfs_trace_put_tcd(tcd);
 412                         return 1;
 413                 }
 414
 415                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 416                                                        libcfs_console_max_delay
 417                                                        + cfs_time_seconds(10))) {
 418                         /* last timeout was a long time ago */
 419                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 420                 } else {
 421                         cdls->cdls_delay *= libcfs_console_backoff;
 422
 423                         if (cdls->cdls_delay < libcfs_console_min_delay)
 424                                 cdls->cdls_delay = libcfs_console_min_delay;
 425                         else if (cdls->cdls_delay > libcfs_console_max_delay)
 426                                 cdls->cdls_delay = libcfs_console_max_delay;
 427                 }
 428
 429                 /* ensure cdls_next is never zero after it's been seen */
 430                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 431         }
 432
 433         if (tcd != NULL) {
 434                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 435                                      msgdata->msg_fn);
 436                 cfs_trace_put_tcd(tcd);
 437         } else {
 438                 string_buf = cfs_trace_get_console_buffer();
 439
 440                 needed = 0;
 441                 if (format1 != NULL) {
 442                         va_copy(ap, args);
 443                         needed = vsnprintf(string_buf,
 444                                            CFS_TRACE_CONSOLE_BUFFER_SIZE,
 445                                            format1, ap);
 446                         va_end(ap);
 447                 }
 448                 if (format2 != NULL) {
 449                         remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
 450                         if (remain > 0) {
 451                                 va_start(ap, format2);
 452                                 needed += vsnprintf(string_buf+needed, remain,
 453                                                     format2, ap);
 454                                 va_end(ap);
 455                         }
 456                 }
 457                 cfs_print_to_console(&header, mask,
 458                                      string_buf, needed, file, msgdata->msg_fn);
 459
 460                 cfs_trace_put_console_buffer(string_buf);
 461         }
 462
 463         if (cdls != NULL && cdls->cdls_count != 0) {
 464                 string_buf = cfs_trace_get_console_buffer();
 465
 466                 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 467                                   "Skipped %d previous similar message%s\n",
 468                                   cdls->cdls_count,
 469                                   (cdls->cdls_count > 1) ? "s" : "");
 470
 471                 cfs_print_to_console(&header, mask,
 472                                      string_buf, needed, file, msgdata->msg_fn);
 473
 474                 cfs_trace_put_console_buffer(string_buf);
 475                 cdls->cdls_count = 0;
 476         }
 477
 478         return 0;
 479 }
 480 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 481
 482 void
 483 cfs_trace_assertion_failed(const char *str,
 484                            struct libcfs_debug_msg_data *msgdata)
 485 {
 486         struct ptldebug_header hdr;
 487
 488         libcfs_panic_in_progress = 1;
 489         libcfs_catastrophe = 1;
 490         smp_mb();
 491
 492         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 493
 494         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 495                              msgdata->msg_file, msgdata->msg_fn);
 496
 497         panic("Lustre debug assertion failure\n");
 498
 499         /* not reached */
 500 }
 501
 502 static void
 503 panic_collect_pages(struct page_collection *pc)
 504 {
 505         /* Do the collect_pages job on a single CPU: assumes that all other
 506          * CPUs have been stopped during a panic.  If this isn't true for some
 507          * arch, this will have to be implemented separately in each arch.  */
 508         int                        i;
 509         int                        j;
 510         struct cfs_trace_cpu_data *tcd;
 511
 512         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 513
 514         cfs_tcd_for_each(tcd, i, j) {
 515                 cfs_list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 516                 tcd->tcd_cur_pages = 0;
 517
 518                 if (pc->pc_want_daemon_pages) {
 519                         cfs_list_splice_init(&tcd->tcd_daemon_pages,
 520                                              &pc->pc_pages);
 521                         tcd->tcd_cur_daemon_pages = 0;
 522                 }
 523         }
 524 }
 525
 526 static void collect_pages_on_all_cpus(struct page_collection *pc)
 527 {
 528         struct cfs_trace_cpu_data *tcd;
 529         int i, cpu;
 530
 531         cfs_for_each_possible_cpu(cpu) {
 532                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 533                         cfs_list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 534                         tcd->tcd_cur_pages = 0;
 535                         if (pc->pc_want_daemon_pages) {
 536                                 cfs_list_splice_init(&tcd->tcd_daemon_pages,
 537                                                      &pc->pc_pages);
 538                                 tcd->tcd_cur_daemon_pages = 0;
 539                         }
 540                 }
 541         }
 542 }
 543
 544 static void collect_pages(struct page_collection *pc)
 545 {
 546         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 547
 548         if (libcfs_panic_in_progress)
 549                 panic_collect_pages(pc);
 550         else
 551                 collect_pages_on_all_cpus(pc);
 552 }
 553
 554 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 555 {
 556         struct cfs_trace_cpu_data *tcd;
 557         cfs_list_t *cur_head;
 558         struct cfs_trace_page *tage;
 559         struct cfs_trace_page *tmp;
 560         int i, cpu;
 561
 562         cfs_for_each_possible_cpu(cpu) {
 563                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 564                         cur_head = tcd->tcd_pages.next;
 565
 566                         cfs_list_for_each_entry_safe_typed(tage, tmp,
 567                                                            &pc->pc_pages,
 568                                                            struct cfs_trace_page,
 569                                                            linkage) {
 570
 571                                 __LASSERT_TAGE_INVARIANT(tage);
 572
 573                                 if (tage->cpu != cpu || tage->type != i)
 574                                         continue;
 575
 576                                 cfs_tage_to_tail(tage, cur_head);
 577                                 tcd->tcd_cur_pages++;
 578                         }
 579                 }
 580         }
 581 }
 582
 583 static void put_pages_back(struct page_collection *pc)
 584 {
 585         if (!libcfs_panic_in_progress)
 586                 put_pages_back_on_all_cpus(pc);
 587 }
 588
 589 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 590  * we have a good amount of data at all times for dumping during an LBUG, even
 591  * if we have been steadily writing (and otherwise discarding) pages via the
 592  * debug daemon. */
 593 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 594                                          struct cfs_trace_cpu_data *tcd)
 595 {
 596         struct cfs_trace_page *tage;
 597         struct cfs_trace_page *tmp;
 598
 599         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc->pc_pages,
 600                                            struct cfs_trace_page, linkage) {
 601
 602                 __LASSERT_TAGE_INVARIANT(tage);
 603
 604                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 605                         continue;
 606
 607                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 608                 tcd->tcd_cur_daemon_pages++;
 609
 610                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 611                         struct cfs_trace_page *victim;
 612
 613                         __LASSERT(!cfs_list_empty(&tcd->tcd_daemon_pages));
 614                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 615
 616                         __LASSERT_TAGE_INVARIANT(victim);
 617
 618                         cfs_list_del(&victim->linkage);
 619                         cfs_tage_free(victim);
 620                         tcd->tcd_cur_daemon_pages--;
 621                 }
 622         }
 623 }
 624
 625 static void put_pages_on_daemon_list(struct page_collection *pc)
 626 {
 627         struct cfs_trace_cpu_data *tcd;
 628         int i, cpu;
 629
 630         cfs_for_each_possible_cpu(cpu) {
 631                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 632                         put_pages_on_tcd_daemon_list(pc, tcd);
 633         }
 634 }
 635
 636 void cfs_trace_debug_print(void)
 637 {
 638         struct page_collection pc;
 639         struct cfs_trace_page *tage;
 640         struct cfs_trace_page *tmp;
 641
 642         pc.pc_want_daemon_pages = 1;
 643         collect_pages(&pc);
 644         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 645                                            struct cfs_trace_page, 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                 cfs_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         int rc;
 681
 682         DECL_MMSPACE;
 683
 684         cfs_tracefile_write_lock();
 685
 686         filp = filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600);
 687         if (IS_ERR(filp)) {
 688                 rc = PTR_ERR(filp);
 689                 filp = NULL;
 690                 printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 691                       filename, rc);
 692                 goto out;
 693         }
 694
 695         pc.pc_want_daemon_pages = 1;
 696         collect_pages(&pc);
 697         if (cfs_list_empty(&pc.pc_pages)) {
 698                 rc = 0;
 699                 goto close;
 700         }
 701
 702         /* ok, for now, just write the pages.  in the future we'll be building
 703          * iobufs with the pages and calling generic_direct_IO */
 704         MMSPACE_OPEN;
 705         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 706                                            struct cfs_trace_page, linkage) {
 707
 708                 __LASSERT_TAGE_INVARIANT(tage);
 709
 710                 rc = filp_write(filp, page_address(tage->page),
 711                                 tage->used, filp_poff(filp));
 712                 if (rc != (int)tage->used) {
 713                         printk(KERN_WARNING "wanted to write %u but wrote "
 714                                "%d\n", tage->used, rc);
 715                         put_pages_back(&pc);
 716                         __LASSERT(cfs_list_empty(&pc.pc_pages));
 717                         break;
 718                 }
 719                 cfs_list_del(&tage->linkage);
 720                 cfs_tage_free(tage);
 721         }
 722         MMSPACE_CLOSE;
 723         rc = filp_fsync(filp);
 724         if (rc)
 725                 printk(KERN_ERR "sync returns %d\n", rc);
 726 close:
 727         filp_close(filp, NULL);
 728 out:
 729         cfs_tracefile_write_unlock();
 730         return rc;
 731 }
 732
 733 void cfs_trace_flush_pages(void)
 734 {
 735         struct page_collection pc;
 736         struct cfs_trace_page *tage;
 737         struct cfs_trace_page *tmp;
 738
 739         pc.pc_want_daemon_pages = 1;
 740         collect_pages(&pc);
 741         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 742                                            struct cfs_trace_page, linkage) {
 743
 744                 __LASSERT_TAGE_INVARIANT(tage);
 745
 746                 cfs_list_del(&tage->linkage);
 747                 cfs_tage_free(tage);
 748         }
 749 }
 750
 751 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 752                             const char *usr_buffer, int usr_buffer_nob)
 753 {
 754         int    nob;
 755
 756         if (usr_buffer_nob > knl_buffer_nob)
 757                 return -EOVERFLOW;
 758
 759         if (copy_from_user((void *)knl_buffer,
 760                            (void *)usr_buffer, usr_buffer_nob))
 761                 return -EFAULT;
 762
 763         nob = strnlen(knl_buffer, usr_buffer_nob);
 764         while (nob-- >= 0)                      /* strip trailing whitespace */
 765                 if (!isspace(knl_buffer[nob]))
 766                         break;
 767
 768         if (nob < 0)                            /* empty string */
 769                 return -EINVAL;
 770
 771         if (nob == knl_buffer_nob)              /* no space to terminate */
 772                 return -EOVERFLOW;
 773
 774         knl_buffer[nob + 1] = 0;                /* terminate */
 775         return 0;
 776 }
 777 EXPORT_SYMBOL(cfs_trace_copyin_string);
 778
 779 int cfs_trace_copyout_string(char *usr_buffer, int usr_buffer_nob,
 780                              const char *knl_buffer, char *append)
 781 {
 782         /* NB if 'append' != NULL, it's a single character to append to the
 783          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 784          * terminating zero byte) for sysctl entries */
 785         int   nob = strlen(knl_buffer);
 786
 787         if (nob > usr_buffer_nob)
 788                 nob = usr_buffer_nob;
 789
 790         if (copy_to_user(usr_buffer, knl_buffer, nob))
 791                 return -EFAULT;
 792
 793         if (append != NULL && nob < usr_buffer_nob) {
 794                 if (copy_to_user(usr_buffer + nob, append, 1))
 795                         return -EFAULT;
 796
 797                 nob++;
 798         }
 799
 800         return nob;
 801 }
 802 EXPORT_SYMBOL(cfs_trace_copyout_string);
 803
 804 int cfs_trace_allocate_string_buffer(char **str, int nob)
 805 {
 806         if (nob > 2 * PAGE_CACHE_SIZE)  /* string must be "sensible" */
 807                 return -EINVAL;
 808
 809         *str = kmalloc(nob, GFP_IOFS | __GFP_ZERO);
 810         if (*str == NULL)
 811                 return -ENOMEM;
 812
 813         return 0;
 814 }
 815
 816 void cfs_trace_free_string_buffer(char *str, int nob)
 817 {
 818         kfree(str);
 819 }
 820
 821 int cfs_trace_dump_debug_buffer_usrstr(void *usr_str, int usr_str_nob)
 822 {
 823         char         *str;
 824         int           rc;
 825
 826         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 827         if (rc != 0)
 828                 return rc;
 829
 830         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 831                                      usr_str, usr_str_nob);
 832         if (rc != 0)
 833                 goto out;
 834
 835 #if !defined(__WINNT__)
 836         if (str[0] != '/') {
 837                 rc = -EINVAL;
 838                 goto out;
 839         }
 840 #endif
 841         rc = cfs_tracefile_dump_all_pages(str);
 842 out:
 843         cfs_trace_free_string_buffer(str, usr_str_nob + 1);
 844         return rc;
 845 }
 846
 847 int cfs_trace_daemon_command(char *str)
 848 {
 849         int       rc = 0;
 850
 851         cfs_tracefile_write_lock();
 852
 853         if (strcmp(str, "stop") == 0) {
 854                 cfs_tracefile_write_unlock();
 855                 cfs_trace_stop_thread();
 856                 cfs_tracefile_write_lock();
 857                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 858
 859         } else if (strncmp(str, "size=", 5) == 0) {
 860                 cfs_tracefile_size = simple_strtoul(str + 5, NULL, 0);
 861                 if (cfs_tracefile_size < 10 || cfs_tracefile_size > 20480)
 862                         cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 863                 else
 864                         cfs_tracefile_size <<= 20;
 865
 866         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 867                 rc = -ENAMETOOLONG;
 868 #ifndef __WINNT__
 869         } else if (str[0] != '/') {
 870                 rc = -EINVAL;
 871 #endif
 872         } else {
 873                 strcpy(cfs_tracefile, str);
 874
 875                 printk(KERN_INFO
 876                        "Lustre: debug daemon will attempt to start writing "
 877                        "to %s (%lukB max)\n", cfs_tracefile,
 878                        (long)(cfs_tracefile_size >> 10));
 879
 880                 cfs_trace_start_thread();
 881         }
 882
 883         cfs_tracefile_write_unlock();
 884         return rc;
 885 }
 886
 887 int cfs_trace_daemon_command_usrstr(void *usr_str, int usr_str_nob)
 888 {
 889         char *str;
 890         int   rc;
 891
 892         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 893         if (rc != 0)
 894                 return rc;
 895
 896         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 897                                  usr_str, usr_str_nob);
 898         if (rc == 0)
 899                 rc = cfs_trace_daemon_command(str);
 900
 901         cfs_trace_free_string_buffer(str, usr_str_nob + 1);
 902         return rc;
 903 }
 904
 905 int cfs_trace_set_debug_mb(int mb)
 906 {
 907         int i;
 908         int j;
 909         int pages;
 910         int limit = cfs_trace_max_debug_mb();
 911         struct cfs_trace_cpu_data *tcd;
 912
 913         if (mb < num_possible_cpus()) {
 914                 printk(KERN_WARNING
 915                        "Lustre: %d MB is too small for debug buffer size, "
 916                        "setting it to %d MB.\n", mb, num_possible_cpus());
 917                 mb = num_possible_cpus();
 918         }
 919
 920         if (mb > limit) {
 921                 printk(KERN_WARNING
 922                        "Lustre: %d MB is too large for debug buffer size, "
 923                        "setting it to %d MB.\n", mb, limit);
 924                 mb = limit;
 925         }
 926
 927         mb /= num_possible_cpus();
 928         pages = mb << (20 - PAGE_CACHE_SHIFT);
 929
 930         cfs_tracefile_write_lock();
 931
 932         cfs_tcd_for_each(tcd, i, j)
 933                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 934
 935         cfs_tracefile_write_unlock();
 936
 937         return 0;
 938 }
 939
 940 int cfs_trace_set_debug_mb_usrstr(void *usr_str, int usr_str_nob)
 941 {
 942         char     str[32];
 943         int      rc;
 944
 945         rc = cfs_trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 946         if (rc < 0)
 947                 return rc;
 948
 949         return cfs_trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 950 }
 951
 952 int cfs_trace_get_debug_mb(void)
 953 {
 954         int i;
 955         int j;
 956         struct cfs_trace_cpu_data *tcd;
 957         int total_pages = 0;
 958
 959         cfs_tracefile_read_lock();
 960
 961         cfs_tcd_for_each(tcd, i, j)
 962                 total_pages += tcd->tcd_max_pages;
 963
 964         cfs_tracefile_read_unlock();
 965
 966         return (total_pages >> (20 - PAGE_CACHE_SHIFT)) + 1;
 967 }
 968
 969 static int tracefiled(void *arg)
 970 {
 971         struct page_collection pc;
 972         struct tracefiled_ctl *tctl = arg;
 973         struct cfs_trace_page *tage;
 974         struct cfs_trace_page *tmp;
 975         struct file *filp;
 976         int last_loop = 0;
 977         int rc;
 978
 979         DECL_MMSPACE;
 980
 981         /* we're started late enough that we pick up init's fs context */
 982         /* this is so broken in uml?  what on earth is going on? */
 983
 984         complete(&tctl->tctl_start);
 985
 986         while (1) {
 987                 wait_queue_t __wait;
 988
 989                 pc.pc_want_daemon_pages = 0;
 990                 collect_pages(&pc);
 991                 if (cfs_list_empty(&pc.pc_pages))
 992                         goto end_loop;
 993
 994                 filp = NULL;
 995                 cfs_tracefile_read_lock();
 996                 if (cfs_tracefile[0] != 0) {
 997                         filp = filp_open(cfs_tracefile,
 998                                          O_CREAT | O_RDWR | O_LARGEFILE,
 999                                          0600);
1000                         if (IS_ERR(filp)) {
1001                                 rc = PTR_ERR(filp);
1002                                 filp = NULL;
1003                                 printk(KERN_WARNING "couldn't open %s: "
1004                                        "%d\n", cfs_tracefile, rc);
1005                         }
1006                 }
1007                 cfs_tracefile_read_unlock();
1008                 if (filp == NULL) {
1009                         put_pages_on_daemon_list(&pc);
1010                         __LASSERT(cfs_list_empty(&pc.pc_pages));
1011                         goto end_loop;
1012                 }
1013
1014                 MMSPACE_OPEN;
1015
1016                 cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
1017                                                    struct cfs_trace_page,
1018                                                    linkage) {
1019                         static loff_t f_pos;
1020
1021                         __LASSERT_TAGE_INVARIANT(tage);
1022
1023                         if (f_pos >= (off_t)cfs_tracefile_size)
1024                                 f_pos = 0;
1025                         else if (f_pos > (off_t)filp_size(filp))
1026                                 f_pos = filp_size(filp);
1027
1028                         rc = filp_write(filp, page_address(tage->page),
1029                                         tage->used, &f_pos);
1030                         if (rc != (int)tage->used) {
1031                                 printk(KERN_WARNING "wanted to write %u "
1032                                        "but wrote %d\n", tage->used, rc);
1033                                 put_pages_back(&pc);
1034                                 __LASSERT(cfs_list_empty(&pc.pc_pages));
1035                         }
1036                 }
1037                 MMSPACE_CLOSE;
1038
1039                 filp_close(filp, NULL);
1040                 put_pages_on_daemon_list(&pc);
1041                 if (!cfs_list_empty(&pc.pc_pages)) {
1042                         int i;
1043
1044                         printk(KERN_ALERT "Lustre: trace pages aren't "
1045                                " empty\n");
1046                         printk(KERN_ERR "total cpus(%d): ",
1047                                num_possible_cpus());
1048                         for (i = 0; i < num_possible_cpus(); i++)
1049                                 if (cpu_online(i))
1050                                         printk(KERN_ERR "%d(on) ", i);
1051                                 else
1052                                         printk(KERN_ERR "%d(off) ", i);
1053                         printk(KERN_ERR "\n");
1054
1055                         i = 0;
1056                         cfs_list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1057                                                      linkage)
1058                                 printk(KERN_ERR "page %d belongs to cpu "
1059                                        "%d\n", ++i, tage->cpu);
1060                         printk(KERN_ERR "There are %d pages unwritten\n",
1061                                i);
1062                 }
1063                 __LASSERT(cfs_list_empty(&pc.pc_pages));
1064 end_loop:
1065                 if (atomic_read(&tctl->tctl_shutdown)) {
1066                         if (last_loop == 0) {
1067                                 last_loop = 1;
1068                                 continue;
1069                         } else {
1070                                 break;
1071                         }
1072                 }
1073                 init_waitqueue_entry_current(&__wait);
1074                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1075                 set_current_state(TASK_INTERRUPTIBLE);
1076                 waitq_timedwait(&__wait, TASK_INTERRUPTIBLE,
1077                                 cfs_time_seconds(1));
1078                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1079         }
1080         complete(&tctl->tctl_stop);
1081         return 0;
1082 }
1083
1084 int cfs_trace_start_thread(void)
1085 {
1086         struct tracefiled_ctl *tctl = &trace_tctl;
1087         int rc = 0;
1088
1089         mutex_lock(&cfs_trace_thread_mutex);
1090         if (thread_running)
1091                 goto out;
1092
1093         init_completion(&tctl->tctl_start);
1094         init_completion(&tctl->tctl_stop);
1095         init_waitqueue_head(&tctl->tctl_waitq);
1096         atomic_set(&tctl->tctl_shutdown, 0);
1097
1098         if (IS_ERR(kthread_run(tracefiled, tctl, "ktracefiled"))) {
1099                 rc = -ECHILD;
1100                 goto out;
1101         }
1102
1103         wait_for_completion(&tctl->tctl_start);
1104         thread_running = 1;
1105 out:
1106         mutex_unlock(&cfs_trace_thread_mutex);
1107         return rc;
1108 }
1109
1110 void cfs_trace_stop_thread(void)
1111 {
1112         struct tracefiled_ctl *tctl = &trace_tctl;
1113
1114         mutex_lock(&cfs_trace_thread_mutex);
1115         if (thread_running) {
1116                 printk(KERN_INFO
1117                        "Lustre: shutting down debug daemon thread...\n");
1118                 atomic_set(&tctl->tctl_shutdown, 1);
1119                 wait_for_completion(&tctl->tctl_stop);
1120                 thread_running = 0;
1121         }
1122         mutex_unlock(&cfs_trace_thread_mutex);
1123 }
1124
1125 int cfs_tracefile_init(int max_pages)
1126 {
1127         struct cfs_trace_cpu_data *tcd;
1128         int                    i;
1129         int                    j;
1130         int                    rc;
1131         int                    factor;
1132
1133         rc = cfs_tracefile_init_arch();
1134         if (rc != 0)
1135                 return rc;
1136
1137         cfs_tcd_for_each(tcd, i, j) {
1138                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1139                 factor = tcd->tcd_pages_factor;
1140                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
1141                 CFS_INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1142                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1143                 tcd->tcd_cur_pages = 0;
1144                 tcd->tcd_cur_stock_pages = 0;
1145                 tcd->tcd_cur_daemon_pages = 0;
1146                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1147                 LASSERT(tcd->tcd_max_pages > 0);
1148                 tcd->tcd_shutting_down = 0;
1149         }
1150
1151         return 0;
1152 }
1153
1154 static void trace_cleanup_on_all_cpus(void)
1155 {
1156         struct cfs_trace_cpu_data *tcd;
1157         struct cfs_trace_page *tage;
1158         struct cfs_trace_page *tmp;
1159         int i, cpu;
1160
1161         cfs_for_each_possible_cpu(cpu) {
1162                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1163                         tcd->tcd_shutting_down = 1;
1164
1165                         cfs_list_for_each_entry_safe_typed(tage, tmp,
1166                                                            &tcd->tcd_pages,
1167                                                            struct cfs_trace_page,
1168                                                            linkage) {
1169                                 __LASSERT_TAGE_INVARIANT(tage);
1170
1171                                 cfs_list_del(&tage->linkage);
1172                                 cfs_tage_free(tage);
1173                         }
1174
1175                         tcd->tcd_cur_pages = 0;
1176                 }
1177         }
1178 }
1179
1180 static void cfs_trace_cleanup(void)
1181 {
1182         struct page_collection pc;
1183
1184         CFS_INIT_LIST_HEAD(&pc.pc_pages);
1185
1186         trace_cleanup_on_all_cpus();
1187
1188         cfs_tracefile_fini_arch();
1189 }
1190
1191 void cfs_tracefile_exit(void)
1192 {
1193         cfs_trace_stop_thread();
1194         cfs_trace_cleanup();
1195 }