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, 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])[CFS_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 cfs_atomic_t cfs_tage_allocated = CFS_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         cfs_page_t            *page;
  72         struct cfs_trace_page *tage;
  73
  74         /* My caller is trying to free memory */
  75         if (!cfs_in_interrupt() && cfs_memory_pressure_get())
  76                 return NULL;
  77
  78         /*
  79          * Don't spam console with allocation failures: they will be reported
  80          * by upper layer anyway.
  81          */
  82         gfp |= CFS_ALLOC_NOWARN;
  83         page = cfs_alloc_page(gfp);
  84         if (page == NULL)
  85                 return NULL;
  86
  87         tage = cfs_alloc(sizeof(*tage), gfp);
  88         if (tage == NULL) {
  89                 cfs_free_page(page);
  90                 return NULL;
  91         }
  92
  93         tage->page = page;
  94         cfs_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         cfs_free_page(tage->page);
 104         cfs_free(tage);
 105         cfs_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 <= CFS_PAGE_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(CFS_ALLOC_ATOMIC);
 158                         if (unlikely(tage == NULL)) {
 159                                 if ((!cfs_memory_pressure_get() ||
 160                                      cfs_in_interrupt()) && printk_ratelimit())
 161                                         printk(CFS_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 = cfs_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                         cfs_waitq_signal(&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(CFS_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         spin_lock_init(&pc.pc_lock);
 205
 206         cfs_list_for_each_entry_safe_typed(tage, tmp, &tcd->tcd_pages,
 207                                            struct cfs_trace_page, linkage) {
 208                 if (pgcount-- == 0)
 209                         break;
 210
 211                 cfs_list_move_tail(&tage->linkage, &pc.pc_pages);
 212                 tcd->tcd_cur_pages--;
 213         }
 214         put_pages_on_tcd_daemon_list(&pc, tcd);
 215 }
 216
 217 /* return a page that has 'len' bytes left at the end */
 218 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 219                                                  unsigned long len)
 220 {
 221         struct cfs_trace_page *tage;
 222
 223         /*
 224          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 225          * from here: this will lead to infinite recursion.
 226          */
 227
 228         if (len > CFS_PAGE_SIZE) {
 229                 printk(CFS_KERN_ERR
 230                        "cowardly refusing to write %lu bytes in a page\n", len);
 231                 return NULL;
 232         }
 233
 234         tage = cfs_trace_get_tage_try(tcd, len);
 235         if (tage != NULL)
 236                 return tage;
 237         if (thread_running)
 238                 cfs_tcd_shrink(tcd);
 239         if (tcd->tcd_cur_pages > 0) {
 240                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 241                 tage->used = 0;
 242                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 243         }
 244         return tage;
 245 }
 246
 247 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 248                      const char *format, ...)
 249 {
 250         va_list args;
 251         int     rc;
 252
 253         va_start(args, format);
 254         rc = libcfs_debug_vmsg2(msgdata, format, args, NULL);
 255         va_end(args);
 256
 257         return rc;
 258 }
 259 EXPORT_SYMBOL(libcfs_debug_msg);
 260
 261 int libcfs_debug_vmsg2(struct libcfs_debug_msg_data *msgdata,
 262                        const char *format1, va_list args,
 263                        const char *format2, ...)
 264 {
 265         struct cfs_trace_cpu_data *tcd = NULL;
 266         struct ptldebug_header     header = {0};
 267         struct cfs_trace_page     *tage;
 268         /* string_buf is used only if tcd != NULL, and is always set then */
 269         char                      *string_buf = NULL;
 270         char                      *debug_buf;
 271         int                        known_size;
 272         int                        needed = 85; /* average message length */
 273         int                        max_nob;
 274         va_list                    ap;
 275         int                        depth;
 276         int                        i;
 277         int                        remain;
 278         int                        mask = msgdata->msg_mask;
 279         char                      *file = (char *)msgdata->msg_file;
 280         cfs_debug_limit_state_t   *cdls = msgdata->msg_cdls;
 281
 282         if (strchr(file, '/'))
 283                 file = strrchr(file, '/') + 1;
 284
 285         tcd = cfs_trace_get_tcd();
 286
 287         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 288          * pins us to a particular CPU.  This avoids an smp_processor_id()
 289          * warning on Linux when debugging is enabled. */
 290         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 291
 292         if (tcd == NULL)                /* arch may not log in IRQ context */
 293                 goto console;
 294
 295         if (tcd->tcd_cur_pages == 0)
 296                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 297
 298         if (tcd->tcd_shutting_down) {
 299                 cfs_trace_put_tcd(tcd);
 300                 tcd = NULL;
 301                 goto console;
 302         }
 303
 304         depth = __current_nesting_level();
 305         known_size = strlen(file) + 1 + depth;
 306         if (msgdata->msg_fn)
 307                 known_size += strlen(msgdata->msg_fn) + 1;
 308
 309         if (libcfs_debug_binary)
 310                 known_size += sizeof(header);
 311
 312         /*/
 313          * '2' used because vsnprintf return real size required for output
 314          * _without_ terminating NULL.
 315          * if needed is to small for this format.
 316          */
 317         for (i = 0; i < 2; i++) {
 318                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 319                 if (tage == NULL) {
 320                         if (needed + known_size > CFS_PAGE_SIZE)
 321                                 mask |= D_ERROR;
 322
 323                         cfs_trace_put_tcd(tcd);
 324                         tcd = NULL;
 325                         goto console;
 326                 }
 327
 328                 string_buf = (char *)cfs_page_address(tage->page) +
 329                                         tage->used + known_size;
 330
 331                 max_nob = CFS_PAGE_SIZE - tage->used - known_size;
 332                 if (max_nob <= 0) {
 333                         printk(CFS_KERN_EMERG "negative max_nob: %d\n",
 334                                max_nob);
 335                         mask |= D_ERROR;
 336                         cfs_trace_put_tcd(tcd);
 337                         tcd = NULL;
 338                         goto console;
 339                 }
 340
 341                 needed = 0;
 342                 if (format1) {
 343                         va_copy(ap, args);
 344                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 345                         va_end(ap);
 346                 }
 347
 348                 if (format2) {
 349                         remain = max_nob - needed;
 350                         if (remain < 0)
 351                                 remain = 0;
 352
 353                         va_start(ap, format2);
 354                         needed += vsnprintf(string_buf + needed, remain,
 355                                             format2, ap);
 356                         va_end(ap);
 357                 }
 358
 359                 if (needed < max_nob) /* well. printing ok.. */
 360                         break;
 361         }
 362
 363         if (*(string_buf+needed-1) != '\n')
 364                 printk(CFS_KERN_INFO "format at %s:%d:%s doesn't end in "
 365                        "newline\n", file, msgdata->msg_line, msgdata->msg_fn);
 366
 367         header.ph_len = known_size + needed;
 368         debug_buf = (char *)cfs_page_address(tage->page) + tage->used;
 369
 370         if (libcfs_debug_binary) {
 371                 memcpy(debug_buf, &header, sizeof(header));
 372                 tage->used += sizeof(header);
 373                 debug_buf += sizeof(header);
 374         }
 375
 376         /* indent message according to the nesting level */
 377         while (depth-- > 0) {
 378                 *(debug_buf++) = '.';
 379                 ++ tage->used;
 380         }
 381
 382         strcpy(debug_buf, file);
 383         tage->used += strlen(file) + 1;
 384         debug_buf += strlen(file) + 1;
 385
 386         if (msgdata->msg_fn) {
 387                 strcpy(debug_buf, msgdata->msg_fn);
 388                 tage->used += strlen(msgdata->msg_fn) + 1;
 389                 debug_buf += strlen(msgdata->msg_fn) + 1;
 390         }
 391
 392         __LASSERT(debug_buf == string_buf);
 393
 394         tage->used += needed;
 395         __LASSERT (tage->used <= CFS_PAGE_SIZE);
 396
 397 console:
 398         if ((mask & libcfs_printk) == 0) {
 399                 /* no console output requested */
 400                 if (tcd != NULL)
 401                         cfs_trace_put_tcd(tcd);
 402                 return 1;
 403         }
 404
 405         if (cdls != NULL) {
 406                 if (libcfs_console_ratelimit &&
 407                     cdls->cdls_next != 0 &&     /* not first time ever */
 408                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 409                         /* skipping a console message */
 410                         cdls->cdls_count++;
 411                         if (tcd != NULL)
 412                                 cfs_trace_put_tcd(tcd);
 413                         return 1;
 414                 }
 415
 416                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 417                                                        libcfs_console_max_delay
 418                                                        + cfs_time_seconds(10))) {
 419                         /* last timeout was a long time ago */
 420                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 421                 } else {
 422                         cdls->cdls_delay *= libcfs_console_backoff;
 423
 424                         if (cdls->cdls_delay < libcfs_console_min_delay)
 425                                 cdls->cdls_delay = libcfs_console_min_delay;
 426                         else if (cdls->cdls_delay > libcfs_console_max_delay)
 427                                 cdls->cdls_delay = libcfs_console_max_delay;
 428                 }
 429
 430                 /* ensure cdls_next is never zero after it's been seen */
 431                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 432         }
 433
 434         if (tcd != NULL) {
 435                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 436                                      msgdata->msg_fn);
 437                 cfs_trace_put_tcd(tcd);
 438         } else {
 439                 string_buf = cfs_trace_get_console_buffer();
 440
 441                 needed = 0;
 442                 if (format1 != NULL) {
 443                         va_copy(ap, args);
 444                         needed = vsnprintf(string_buf,
 445                                            CFS_TRACE_CONSOLE_BUFFER_SIZE,
 446                                            format1, ap);
 447                         va_end(ap);
 448                 }
 449                 if (format2 != NULL) {
 450                         remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
 451                         if (remain > 0) {
 452                                 va_start(ap, format2);
 453                                 needed += vsnprintf(string_buf+needed, remain,
 454                                                     format2, ap);
 455                                 va_end(ap);
 456                         }
 457                 }
 458                 cfs_print_to_console(&header, mask,
 459                                      string_buf, needed, file, msgdata->msg_fn);
 460
 461                 cfs_trace_put_console_buffer(string_buf);
 462         }
 463
 464         if (cdls != NULL && cdls->cdls_count != 0) {
 465                 string_buf = cfs_trace_get_console_buffer();
 466
 467                 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 468                                   "Skipped %d previous similar message%s\n",
 469                                   cdls->cdls_count,
 470                                   (cdls->cdls_count > 1) ? "s" : "");
 471
 472                 cfs_print_to_console(&header, mask,
 473                                      string_buf, needed, file, msgdata->msg_fn);
 474
 475                 cfs_trace_put_console_buffer(string_buf);
 476                 cdls->cdls_count = 0;
 477         }
 478
 479         return 0;
 480 }
 481 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 482
 483 void
 484 cfs_trace_assertion_failed(const char *str,
 485                            struct libcfs_debug_msg_data *msgdata)
 486 {
 487         struct ptldebug_header hdr;
 488
 489         libcfs_panic_in_progress = 1;
 490         libcfs_catastrophe = 1;
 491         cfs_mb();
 492
 493         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 494
 495         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 496                              msgdata->msg_file, msgdata->msg_fn);
 497
 498         LIBCFS_PANIC("Lustre debug assertion failure\n");
 499
 500         /* not reached */
 501 }
 502
 503 static void
 504 panic_collect_pages(struct page_collection *pc)
 505 {
 506         /* Do the collect_pages job on a single CPU: assumes that all other
 507          * CPUs have been stopped during a panic.  If this isn't true for some
 508          * arch, this will have to be implemented separately in each arch.  */
 509         int                        i;
 510         int                        j;
 511         struct cfs_trace_cpu_data *tcd;
 512
 513         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 514
 515         cfs_tcd_for_each(tcd, i, j) {
 516                 cfs_list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 517                 tcd->tcd_cur_pages = 0;
 518
 519                 if (pc->pc_want_daemon_pages) {
 520                         cfs_list_splice_init(&tcd->tcd_daemon_pages,
 521                                              &pc->pc_pages);
 522                         tcd->tcd_cur_daemon_pages = 0;
 523                 }
 524         }
 525 }
 526
 527 static void collect_pages_on_all_cpus(struct page_collection *pc)
 528 {
 529         struct cfs_trace_cpu_data *tcd;
 530         int i, cpu;
 531
 532         spin_lock(&pc->pc_lock);
 533         cfs_for_each_possible_cpu(cpu) {
 534                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 535                         cfs_list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 536                         tcd->tcd_cur_pages = 0;
 537                         if (pc->pc_want_daemon_pages) {
 538                                 cfs_list_splice_init(&tcd->tcd_daemon_pages,
 539                                                      &pc->pc_pages);
 540                                 tcd->tcd_cur_daemon_pages = 0;
 541                         }
 542                 }
 543         }
 544         spin_unlock(&pc->pc_lock);
 545 }
 546
 547 static void collect_pages(struct page_collection *pc)
 548 {
 549         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 550
 551         if (libcfs_panic_in_progress)
 552                 panic_collect_pages(pc);
 553         else
 554                 collect_pages_on_all_cpus(pc);
 555 }
 556
 557 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 558 {
 559         struct cfs_trace_cpu_data *tcd;
 560         cfs_list_t *cur_head;
 561         struct cfs_trace_page *tage;
 562         struct cfs_trace_page *tmp;
 563         int i, cpu;
 564
 565         spin_lock(&pc->pc_lock);
 566         cfs_for_each_possible_cpu(cpu) {
 567                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 568                         cur_head = tcd->tcd_pages.next;
 569
 570                         cfs_list_for_each_entry_safe_typed(tage, tmp,
 571                                                            &pc->pc_pages,
 572                                                            struct cfs_trace_page,
 573                                                            linkage) {
 574
 575                                 __LASSERT_TAGE_INVARIANT(tage);
 576
 577                                 if (tage->cpu != cpu || tage->type != i)
 578                                         continue;
 579
 580                                 cfs_tage_to_tail(tage, cur_head);
 581                                 tcd->tcd_cur_pages++;
 582                         }
 583                 }
 584         }
 585         spin_unlock(&pc->pc_lock);
 586 }
 587
 588 static void put_pages_back(struct page_collection *pc)
 589 {
 590         if (!libcfs_panic_in_progress)
 591                 put_pages_back_on_all_cpus(pc);
 592 }
 593
 594 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 595  * we have a good amount of data at all times for dumping during an LBUG, even
 596  * if we have been steadily writing (and otherwise discarding) pages via the
 597  * debug daemon. */
 598 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 599                                          struct cfs_trace_cpu_data *tcd)
 600 {
 601         struct cfs_trace_page *tage;
 602         struct cfs_trace_page *tmp;
 603
 604         spin_lock(&pc->pc_lock);
 605         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc->pc_pages,
 606                                            struct cfs_trace_page, linkage) {
 607
 608                 __LASSERT_TAGE_INVARIANT(tage);
 609
 610                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 611                         continue;
 612
 613                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 614                 tcd->tcd_cur_daemon_pages++;
 615
 616                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 617                         struct cfs_trace_page *victim;
 618
 619                         __LASSERT(!cfs_list_empty(&tcd->tcd_daemon_pages));
 620                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 621
 622                         __LASSERT_TAGE_INVARIANT(victim);
 623
 624                         cfs_list_del(&victim->linkage);
 625                         cfs_tage_free(victim);
 626                         tcd->tcd_cur_daemon_pages--;
 627                 }
 628         }
 629         spin_unlock(&pc->pc_lock);
 630 }
 631
 632 static void put_pages_on_daemon_list(struct page_collection *pc)
 633 {
 634         struct cfs_trace_cpu_data *tcd;
 635         int i, cpu;
 636
 637         cfs_for_each_possible_cpu(cpu) {
 638                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 639                         put_pages_on_tcd_daemon_list(pc, tcd);
 640         }
 641 }
 642
 643 void cfs_trace_debug_print(void)
 644 {
 645         struct page_collection pc;
 646         struct cfs_trace_page *tage;
 647         struct cfs_trace_page *tmp;
 648
 649         spin_lock_init(&pc.pc_lock);
 650
 651         pc.pc_want_daemon_pages = 1;
 652         collect_pages(&pc);
 653         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 654                                            struct cfs_trace_page, linkage) {
 655                 char *p, *file, *fn;
 656                 cfs_page_t *page;
 657
 658                 __LASSERT_TAGE_INVARIANT(tage);
 659
 660                 page = tage->page;
 661                 p = cfs_page_address(page);
 662                 while (p < ((char *)cfs_page_address(page) + tage->used)) {
 663                         struct ptldebug_header *hdr;
 664                         int len;
 665                         hdr = (void *)p;
 666                         p += sizeof(*hdr);
 667                         file = p;
 668                         p += strlen(file) + 1;
 669                         fn = p;
 670                         p += strlen(fn) + 1;
 671                         len = hdr->ph_len - (int)(p - (char *)hdr);
 672
 673                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 674
 675                         p += len;
 676                 }
 677
 678                 cfs_list_del(&tage->linkage);
 679                 cfs_tage_free(tage);
 680         }
 681 }
 682
 683 int cfs_tracefile_dump_all_pages(char *filename)
 684 {
 685         struct page_collection  pc;
 686         struct file             *filp;
 687         struct cfs_trace_page   *tage;
 688         struct cfs_trace_page   *tmp;
 689         int rc;
 690
 691         CFS_DECL_MMSPACE;
 692
 693         cfs_tracefile_write_lock();
 694
 695         filp = filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600);
 696         if (IS_ERR(filp)) {
 697                 rc = PTR_ERR(filp);
 698                 filp = NULL;
 699                 printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 700                       filename, rc);
 701                 goto out;
 702         }
 703
 704         spin_lock_init(&pc.pc_lock);
 705         pc.pc_want_daemon_pages = 1;
 706         collect_pages(&pc);
 707         if (cfs_list_empty(&pc.pc_pages)) {
 708                 rc = 0;
 709                 goto close;
 710         }
 711
 712         /* ok, for now, just write the pages.  in the future we'll be building
 713          * iobufs with the pages and calling generic_direct_IO */
 714         CFS_MMSPACE_OPEN;
 715         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 716                                            struct cfs_trace_page, linkage) {
 717
 718                 __LASSERT_TAGE_INVARIANT(tage);
 719
 720                 rc = filp_write(filp, cfs_page_address(tage->page),
 721                                 tage->used, filp_poff(filp));
 722                 if (rc != (int)tage->used) {
 723                         printk(CFS_KERN_WARNING "wanted to write %u but wrote "
 724                                "%d\n", tage->used, rc);
 725                         put_pages_back(&pc);
 726                         __LASSERT(cfs_list_empty(&pc.pc_pages));
 727                         break;
 728                 }
 729                 cfs_list_del(&tage->linkage);
 730                 cfs_tage_free(tage);
 731         }
 732         CFS_MMSPACE_CLOSE;
 733         rc = filp_fsync(filp);
 734         if (rc)
 735                 printk(CFS_KERN_ERR "sync returns %d\n", rc);
 736 close:
 737         filp_close(filp, NULL);
 738 out:
 739         cfs_tracefile_write_unlock();
 740         return rc;
 741 }
 742
 743 void cfs_trace_flush_pages(void)
 744 {
 745         struct page_collection pc;
 746         struct cfs_trace_page *tage;
 747         struct cfs_trace_page *tmp;
 748
 749         spin_lock_init(&pc.pc_lock);
 750
 751         pc.pc_want_daemon_pages = 1;
 752         collect_pages(&pc);
 753         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 754                                            struct cfs_trace_page, linkage) {
 755
 756                 __LASSERT_TAGE_INVARIANT(tage);
 757
 758                 cfs_list_del(&tage->linkage);
 759                 cfs_tage_free(tage);
 760         }
 761 }
 762
 763 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 764                             const char *usr_buffer, int usr_buffer_nob)
 765 {
 766         int    nob;
 767
 768         if (usr_buffer_nob > knl_buffer_nob)
 769                 return -EOVERFLOW;
 770
 771         if (cfs_copy_from_user((void *)knl_buffer,
 772                            (void *)usr_buffer, usr_buffer_nob))
 773                 return -EFAULT;
 774
 775         nob = strnlen(knl_buffer, usr_buffer_nob);
 776         while (nob-- >= 0)                      /* strip trailing whitespace */
 777                 if (!isspace(knl_buffer[nob]))
 778                         break;
 779
 780         if (nob < 0)                            /* empty string */
 781                 return -EINVAL;
 782
 783         if (nob == knl_buffer_nob)              /* no space to terminate */
 784                 return -EOVERFLOW;
 785
 786         knl_buffer[nob + 1] = 0;                /* terminate */
 787         return 0;
 788 }
 789 EXPORT_SYMBOL(cfs_trace_copyin_string);
 790
 791 int cfs_trace_copyout_string(char *usr_buffer, int usr_buffer_nob,
 792                              const char *knl_buffer, char *append)
 793 {
 794         /* NB if 'append' != NULL, it's a single character to append to the
 795          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 796          * terminating zero byte) for sysctl entries */
 797         int   nob = strlen(knl_buffer);
 798
 799         if (nob > usr_buffer_nob)
 800                 nob = usr_buffer_nob;
 801
 802         if (cfs_copy_to_user(usr_buffer, knl_buffer, nob))
 803                 return -EFAULT;
 804
 805         if (append != NULL && nob < usr_buffer_nob) {
 806                 if (cfs_copy_to_user(usr_buffer + nob, append, 1))
 807                         return -EFAULT;
 808
 809                 nob++;
 810         }
 811
 812         return nob;
 813 }
 814 EXPORT_SYMBOL(cfs_trace_copyout_string);
 815
 816 int cfs_trace_allocate_string_buffer(char **str, int nob)
 817 {
 818         if (nob > 2 * CFS_PAGE_SIZE)            /* string must be "sensible" */
 819                 return -EINVAL;
 820
 821         *str = cfs_alloc(nob, CFS_ALLOC_STD | CFS_ALLOC_ZERO);
 822         if (*str == NULL)
 823                 return -ENOMEM;
 824
 825         return 0;
 826 }
 827
 828 void cfs_trace_free_string_buffer(char *str, int nob)
 829 {
 830         cfs_free(str);
 831 }
 832
 833 int cfs_trace_dump_debug_buffer_usrstr(void *usr_str, int usr_str_nob)
 834 {
 835         char         *str;
 836         int           rc;
 837
 838         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 839         if (rc != 0)
 840                 return rc;
 841
 842         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 843                                      usr_str, usr_str_nob);
 844         if (rc != 0)
 845                 goto out;
 846
 847 #if !defined(__WINNT__)
 848         if (str[0] != '/') {
 849                 rc = -EINVAL;
 850                 goto out;
 851         }
 852 #endif
 853         rc = cfs_tracefile_dump_all_pages(str);
 854 out:
 855         cfs_trace_free_string_buffer(str, usr_str_nob + 1);
 856         return rc;
 857 }
 858
 859 int cfs_trace_daemon_command(char *str)
 860 {
 861         int       rc = 0;
 862
 863         cfs_tracefile_write_lock();
 864
 865         if (strcmp(str, "stop") == 0) {
 866                 cfs_tracefile_write_unlock();
 867                 cfs_trace_stop_thread();
 868                 cfs_tracefile_write_lock();
 869                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 870
 871         } else if (strncmp(str, "size=", 5) == 0) {
 872                 cfs_tracefile_size = simple_strtoul(str + 5, NULL, 0);
 873                 if (cfs_tracefile_size < 10 || cfs_tracefile_size > 20480)
 874                         cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 875                 else
 876                         cfs_tracefile_size <<= 20;
 877
 878         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 879                 rc = -ENAMETOOLONG;
 880 #ifndef __WINNT__
 881         } else if (str[0] != '/') {
 882                 rc = -EINVAL;
 883 #endif
 884         } else {
 885                 strcpy(cfs_tracefile, str);
 886
 887                 printk(CFS_KERN_INFO
 888                        "Lustre: debug daemon will attempt to start writing "
 889                        "to %s (%lukB max)\n", cfs_tracefile,
 890                        (long)(cfs_tracefile_size >> 10));
 891
 892                 cfs_trace_start_thread();
 893         }
 894
 895         cfs_tracefile_write_unlock();
 896         return rc;
 897 }
 898
 899 int cfs_trace_daemon_command_usrstr(void *usr_str, int usr_str_nob)
 900 {
 901         char *str;
 902         int   rc;
 903
 904         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 905         if (rc != 0)
 906                 return rc;
 907
 908         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 909                                  usr_str, usr_str_nob);
 910         if (rc == 0)
 911                 rc = cfs_trace_daemon_command(str);
 912
 913         cfs_trace_free_string_buffer(str, usr_str_nob + 1);
 914         return rc;
 915 }
 916
 917 int cfs_trace_set_debug_mb(int mb)
 918 {
 919         int i;
 920         int j;
 921         int pages;
 922         int limit = cfs_trace_max_debug_mb();
 923         struct cfs_trace_cpu_data *tcd;
 924
 925         if (mb < cfs_num_possible_cpus()) {
 926                 printk(CFS_KERN_WARNING
 927                        "Lustre: %d MB is too small for debug buffer size, "
 928                        "setting it to %d MB.\n", mb, cfs_num_possible_cpus());
 929                 mb = cfs_num_possible_cpus();
 930         }
 931
 932         if (mb > limit) {
 933                 printk(CFS_KERN_WARNING
 934                        "Lustre: %d MB is too large for debug buffer size, "
 935                        "setting it to %d MB.\n", mb, limit);
 936                 mb = limit;
 937         }
 938
 939         mb /= cfs_num_possible_cpus();
 940         pages = mb << (20 - CFS_PAGE_SHIFT);
 941
 942         cfs_tracefile_write_lock();
 943
 944         cfs_tcd_for_each(tcd, i, j)
 945                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 946
 947         cfs_tracefile_write_unlock();
 948
 949         return 0;
 950 }
 951
 952 int cfs_trace_set_debug_mb_usrstr(void *usr_str, int usr_str_nob)
 953 {
 954         char     str[32];
 955         int      rc;
 956
 957         rc = cfs_trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 958         if (rc < 0)
 959                 return rc;
 960
 961         return cfs_trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 962 }
 963
 964 int cfs_trace_get_debug_mb(void)
 965 {
 966         int i;
 967         int j;
 968         struct cfs_trace_cpu_data *tcd;
 969         int total_pages = 0;
 970
 971         cfs_tracefile_read_lock();
 972
 973         cfs_tcd_for_each(tcd, i, j)
 974                 total_pages += tcd->tcd_max_pages;
 975
 976         cfs_tracefile_read_unlock();
 977
 978         return (total_pages >> (20 - CFS_PAGE_SHIFT)) + 1;
 979 }
 980
 981 static int tracefiled(void *arg)
 982 {
 983         struct page_collection pc;
 984         struct tracefiled_ctl *tctl = arg;
 985         struct cfs_trace_page *tage;
 986         struct cfs_trace_page *tmp;
 987         struct file *filp;
 988         int last_loop = 0;
 989         int rc;
 990
 991         CFS_DECL_MMSPACE;
 992
 993         /* we're started late enough that we pick up init's fs context */
 994         /* this is so broken in uml?  what on earth is going on? */
 995         cfs_daemonize("ktracefiled");
 996
 997         spin_lock_init(&pc.pc_lock);
 998         complete(&tctl->tctl_start);
 999
1000         while (1) {
1001                 cfs_waitlink_t __wait;
1002
1003                 pc.pc_want_daemon_pages = 0;
1004                 collect_pages(&pc);
1005                 if (cfs_list_empty(&pc.pc_pages))
1006                         goto end_loop;
1007
1008                 filp = NULL;
1009                 cfs_tracefile_read_lock();
1010                 if (cfs_tracefile[0] != 0) {
1011                         filp = filp_open(cfs_tracefile,
1012                                          O_CREAT | O_RDWR | O_LARGEFILE,
1013                                          0600);
1014                         if (IS_ERR(filp)) {
1015                                 rc = PTR_ERR(filp);
1016                                 filp = NULL;
1017                                 printk(CFS_KERN_WARNING "couldn't open %s: "
1018                                        "%d\n", cfs_tracefile, rc);
1019                         }
1020                 }
1021                 cfs_tracefile_read_unlock();
1022                 if (filp == NULL) {
1023                         put_pages_on_daemon_list(&pc);
1024                         __LASSERT(cfs_list_empty(&pc.pc_pages));
1025                         goto end_loop;
1026                 }
1027
1028                 CFS_MMSPACE_OPEN;
1029
1030                 cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
1031                                                    struct cfs_trace_page,
1032                                                    linkage) {
1033                         static loff_t f_pos;
1034
1035                         __LASSERT_TAGE_INVARIANT(tage);
1036
1037                         if (f_pos >= (off_t)cfs_tracefile_size)
1038                                 f_pos = 0;
1039                         else if (f_pos > (off_t)filp_size(filp))
1040                                 f_pos = filp_size(filp);
1041
1042                         rc = filp_write(filp, cfs_page_address(tage->page),
1043                                         tage->used, &f_pos);
1044                         if (rc != (int)tage->used) {
1045                                 printk(CFS_KERN_WARNING "wanted to write %u "
1046                                        "but wrote %d\n", tage->used, rc);
1047                                 put_pages_back(&pc);
1048                                 __LASSERT(cfs_list_empty(&pc.pc_pages));
1049                         }
1050                 }
1051                 CFS_MMSPACE_CLOSE;
1052
1053                 filp_close(filp, NULL);
1054                 put_pages_on_daemon_list(&pc);
1055                 if (!cfs_list_empty(&pc.pc_pages)) {
1056                         int i;
1057
1058                         printk(CFS_KERN_ALERT "Lustre: trace pages aren't "
1059                                " empty\n");
1060                         printk(CFS_KERN_ERR "total cpus(%d): ",
1061                                cfs_num_possible_cpus());
1062                         for (i = 0; i < cfs_num_possible_cpus(); i++)
1063                                 if (cpu_online(i))
1064                                         printk(CFS_KERN_ERR "%d(on) ", i);
1065                                 else
1066                                         printk(CFS_KERN_ERR "%d(off) ", i);
1067                         printk(CFS_KERN_ERR "\n");
1068
1069                         i = 0;
1070                         cfs_list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1071                                                      linkage)
1072                                 printk(CFS_KERN_ERR "page %d belongs to cpu "
1073                                        "%d\n", ++i, tage->cpu);
1074                         printk(CFS_KERN_ERR "There are %d pages unwritten\n",
1075                                i);
1076                 }
1077                 __LASSERT(cfs_list_empty(&pc.pc_pages));
1078 end_loop:
1079                 if (cfs_atomic_read(&tctl->tctl_shutdown)) {
1080                         if (last_loop == 0) {
1081                                 last_loop = 1;
1082                                 continue;
1083                         } else {
1084                                 break;
1085                         }
1086                 }
1087                 cfs_waitlink_init(&__wait);
1088                 cfs_waitq_add(&tctl->tctl_waitq, &__wait);
1089                 cfs_set_current_state(CFS_TASK_INTERRUPTIBLE);
1090                 cfs_waitq_timedwait(&__wait, CFS_TASK_INTERRUPTIBLE,
1091                                     cfs_time_seconds(1));
1092                 cfs_waitq_del(&tctl->tctl_waitq, &__wait);
1093         }
1094         complete(&tctl->tctl_stop);
1095         return 0;
1096 }
1097
1098 int cfs_trace_start_thread(void)
1099 {
1100         struct tracefiled_ctl *tctl = &trace_tctl;
1101         int rc = 0;
1102
1103         mutex_lock(&cfs_trace_thread_mutex);
1104         if (thread_running)
1105                 goto out;
1106
1107         init_completion(&tctl->tctl_start);
1108         init_completion(&tctl->tctl_stop);
1109         cfs_waitq_init(&tctl->tctl_waitq);
1110         cfs_atomic_set(&tctl->tctl_shutdown, 0);
1111
1112         if (cfs_create_thread(tracefiled, tctl, 0) < 0) {
1113                 rc = -ECHILD;
1114                 goto out;
1115         }
1116
1117         wait_for_completion(&tctl->tctl_start);
1118         thread_running = 1;
1119 out:
1120         mutex_unlock(&cfs_trace_thread_mutex);
1121         return rc;
1122 }
1123
1124 void cfs_trace_stop_thread(void)
1125 {
1126         struct tracefiled_ctl *tctl = &trace_tctl;
1127
1128         mutex_lock(&cfs_trace_thread_mutex);
1129         if (thread_running) {
1130                 printk(CFS_KERN_INFO
1131                        "Lustre: shutting down debug daemon thread...\n");
1132                 cfs_atomic_set(&tctl->tctl_shutdown, 1);
1133                 wait_for_completion(&tctl->tctl_stop);
1134                 thread_running = 0;
1135         }
1136         mutex_unlock(&cfs_trace_thread_mutex);
1137 }
1138
1139 int cfs_tracefile_init(int max_pages)
1140 {
1141         struct cfs_trace_cpu_data *tcd;
1142         int                    i;
1143         int                    j;
1144         int                    rc;
1145         int                    factor;
1146
1147         rc = cfs_tracefile_init_arch();
1148         if (rc != 0)
1149                 return rc;
1150
1151         cfs_tcd_for_each(tcd, i, j) {
1152                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1153                 factor = tcd->tcd_pages_factor;
1154                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
1155                 CFS_INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1156                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1157                 tcd->tcd_cur_pages = 0;
1158                 tcd->tcd_cur_stock_pages = 0;
1159                 tcd->tcd_cur_daemon_pages = 0;
1160                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1161                 LASSERT(tcd->tcd_max_pages > 0);
1162                 tcd->tcd_shutting_down = 0;
1163         }
1164
1165         return 0;
1166 }
1167
1168 static void trace_cleanup_on_all_cpus(void)
1169 {
1170         struct cfs_trace_cpu_data *tcd;
1171         struct cfs_trace_page *tage;
1172         struct cfs_trace_page *tmp;
1173         int i, cpu;
1174
1175         cfs_for_each_possible_cpu(cpu) {
1176                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1177                         tcd->tcd_shutting_down = 1;
1178
1179                         cfs_list_for_each_entry_safe_typed(tage, tmp,
1180                                                            &tcd->tcd_pages,
1181                                                            struct cfs_trace_page,
1182                                                            linkage) {
1183                                 __LASSERT_TAGE_INVARIANT(tage);
1184
1185                                 cfs_list_del(&tage->linkage);
1186                                 cfs_tage_free(tage);
1187                         }
1188
1189                         tcd->tcd_cur_pages = 0;
1190                 }
1191         }
1192 }
1193
1194 static void cfs_trace_cleanup(void)
1195 {
1196         struct page_collection pc;
1197
1198         CFS_INIT_LIST_HEAD(&pc.pc_pages);
1199         spin_lock_init(&pc.pc_lock);
1200
1201         trace_cleanup_on_all_cpus();
1202
1203         cfs_tracefile_fini_arch();
1204 }
1205
1206 void cfs_tracefile_exit(void)
1207 {
1208         cfs_trace_stop_thread();
1209         cfs_trace_cleanup();
1210 }