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 /*
  31  * This file is part of Lustre, http://www.lustre.org/
  32  * Lustre is a trademark of Sun Microsystems, Inc.
  33  *
  34  * libcfs/libcfs/tracefile.c
  35  *
  36  * Author: Zach Brown <zab@clusterfs.com>
  37  * Author: Phil Schwan <phil@clusterfs.com>
  38  */
  39
  40
  41 #define DEBUG_SUBSYSTEM S_LNET
  42 #define LUSTRE_TRACEFILE_PRIVATE
  43 #include "tracefile.h"
  44
  45 #include <libcfs/libcfs.h>
  46
  47 /* XXX move things up to the top, comment */
  48 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[CFS_NR_CPUS] __cacheline_aligned;
  49
  50 char cfs_tracefile[TRACEFILE_NAME_SIZE];
  51 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
  52 static struct tracefiled_ctl trace_tctl;
  53 cfs_mutex_t cfs_trace_thread_mutex;
  54 static int thread_running = 0;
  55
  56 cfs_atomic_t cfs_tage_allocated = CFS_ATOMIC_INIT(0);
  57
  58 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  59                                          struct cfs_trace_cpu_data *tcd);
  60
  61 static inline struct cfs_trace_page *
  62 cfs_tage_from_list(cfs_list_t *list)
  63 {
  64         return cfs_list_entry(list, struct cfs_trace_page, linkage);
  65 }
  66
  67 static struct cfs_trace_page *cfs_tage_alloc(int gfp)
  68 {
  69         cfs_page_t            *page;
  70         struct cfs_trace_page *tage;
  71
  72         /* My caller is trying to free memory */
  73         if (!cfs_in_interrupt() && cfs_memory_pressure_get())
  74                 return NULL;
  75
  76         /*
  77          * Don't spam console with allocation failures: they will be reported
  78          * by upper layer anyway.
  79          */
  80         gfp |= CFS_ALLOC_NOWARN;
  81         page = cfs_alloc_page(gfp);
  82         if (page == NULL)
  83                 return NULL;
  84
  85         tage = cfs_alloc(sizeof(*tage), gfp);
  86         if (tage == NULL) {
  87                 cfs_free_page(page);
  88                 return NULL;
  89         }
  90
  91         tage->page = page;
  92         cfs_atomic_inc(&cfs_tage_allocated);
  93         return tage;
  94 }
  95
  96 static void cfs_tage_free(struct cfs_trace_page *tage)
  97 {
  98         __LASSERT(tage != NULL);
  99         __LASSERT(tage->page != NULL);
 100
 101         cfs_free_page(tage->page);
 102         cfs_free(tage);
 103         cfs_atomic_dec(&cfs_tage_allocated);
 104 }
 105
 106 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
 107                              cfs_list_t *queue)
 108 {
 109         __LASSERT(tage != NULL);
 110         __LASSERT(queue != NULL);
 111
 112         cfs_list_move_tail(&tage->linkage, queue);
 113 }
 114
 115 int cfs_trace_refill_stock(struct cfs_trace_cpu_data *tcd, int gfp,
 116                            cfs_list_t *stock)
 117 {
 118         int i;
 119
 120         /*
 121          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 122          * from here: this will lead to infinite recursion.
 123          */
 124
 125         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++ i) {
 126                 struct cfs_trace_page *tage;
 127
 128                 tage = cfs_tage_alloc(gfp);
 129                 if (tage == NULL)
 130                         break;
 131                 cfs_list_add_tail(&tage->linkage, stock);
 132         }
 133         return i;
 134 }
 135
 136 /* return a page that has 'len' bytes left at the end */
 137 static struct cfs_trace_page *
 138 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
 139 {
 140         struct cfs_trace_page *tage;
 141
 142         if (tcd->tcd_cur_pages > 0) {
 143                 __LASSERT(!cfs_list_empty(&tcd->tcd_pages));
 144                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
 145                 if (tage->used + len <= CFS_PAGE_SIZE)
 146                         return tage;
 147         }
 148
 149         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 150                 if (tcd->tcd_cur_stock_pages > 0) {
 151                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
 152                         -- tcd->tcd_cur_stock_pages;
 153                         cfs_list_del_init(&tage->linkage);
 154                 } else {
 155                         tage = cfs_tage_alloc(CFS_ALLOC_ATOMIC);
 156                         if (tage == NULL) {
 157                                 if (printk_ratelimit())
 158                                         printk(CFS_KERN_WARNING
 159                                                "cannot allocate a tage (%ld)\n",
 160                                        tcd->tcd_cur_pages);
 161                                 return NULL;
 162                         }
 163                 }
 164
 165                 tage->used = 0;
 166                 tage->cpu = cfs_smp_processor_id();
 167                 tage->type = tcd->tcd_type;
 168                 cfs_list_add_tail(&tage->linkage, &tcd->tcd_pages);
 169                 tcd->tcd_cur_pages++;
 170
 171                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 172                         struct tracefiled_ctl *tctl = &trace_tctl;
 173                         /*
 174                          * wake up tracefiled to process some pages.
 175                          */
 176                         cfs_waitq_signal(&tctl->tctl_waitq);
 177                 }
 178                 return tage;
 179         }
 180         return NULL;
 181 }
 182
 183 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
 184 {
 185         int pgcount = tcd->tcd_cur_pages / 10;
 186         struct page_collection pc;
 187         struct cfs_trace_page *tage;
 188         struct cfs_trace_page *tmp;
 189
 190         /*
 191          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 192          * from here: this will lead to infinite recursion.
 193          */
 194
 195         if (printk_ratelimit())
 196                 printk(CFS_KERN_WARNING "debug daemon buffer overflowed; "
 197                        "discarding 10%% of pages (%d of %ld)\n",
 198                        pgcount + 1, tcd->tcd_cur_pages);
 199
 200         CFS_INIT_LIST_HEAD(&pc.pc_pages);
 201         cfs_spin_lock_init(&pc.pc_lock);
 202
 203         cfs_list_for_each_entry_safe_typed(tage, tmp, &tcd->tcd_pages,
 204                                            struct cfs_trace_page, linkage) {
 205                 if (pgcount-- == 0)
 206                         break;
 207
 208                 cfs_list_move_tail(&tage->linkage, &pc.pc_pages);
 209                 tcd->tcd_cur_pages--;
 210         }
 211         put_pages_on_tcd_daemon_list(&pc, tcd);
 212 }
 213
 214 /* return a page that has 'len' bytes left at the end */
 215 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 216                                                  unsigned long len)
 217 {
 218         struct cfs_trace_page *tage;
 219
 220         /*
 221          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 222          * from here: this will lead to infinite recursion.
 223          */
 224
 225         if (len > CFS_PAGE_SIZE) {
 226                 printk(CFS_KERN_ERR
 227                        "cowardly refusing to write %lu bytes in a page\n", len);
 228                 return NULL;
 229         }
 230
 231         tage = cfs_trace_get_tage_try(tcd, len);
 232         if (tage != NULL)
 233                 return tage;
 234         if (thread_running)
 235                 cfs_tcd_shrink(tcd);
 236         if (tcd->tcd_cur_pages > 0) {
 237                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 238                 tage->used = 0;
 239                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 240         }
 241         return tage;
 242 }
 243
 244 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 245                      const char *format, ...)
 246 {
 247         va_list args;
 248         int     rc;
 249
 250         va_start(args, format);
 251         rc = libcfs_debug_vmsg2(msgdata, format, args, NULL);
 252         va_end(args);
 253
 254         return rc;
 255 }
 256 EXPORT_SYMBOL(libcfs_debug_msg);
 257
 258 int libcfs_debug_vmsg2(struct libcfs_debug_msg_data *msgdata,
 259                        const char *format1, va_list args,
 260                        const char *format2, ...)
 261 {
 262         struct cfs_trace_cpu_data *tcd = NULL;
 263         struct ptldebug_header     header = {0};
 264         struct cfs_trace_page     *tage;
 265         /* string_buf is used only if tcd != NULL, and is always set then */
 266         char                      *string_buf = NULL;
 267         char                      *debug_buf;
 268         int                        known_size;
 269         int                        needed = 85; /* average message length */
 270         int                        max_nob;
 271         va_list                    ap;
 272         int                        depth;
 273         int                        i;
 274         int                        remain;
 275         int                        mask = msgdata->msg_mask;
 276         char                      *file = (char *)msgdata->msg_file;
 277         cfs_debug_limit_state_t   *cdls = msgdata->msg_cdls;
 278
 279         if (strchr(file, '/'))
 280                 file = strrchr(file, '/') + 1;
 281
 282         tcd = cfs_trace_get_tcd();
 283
 284         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 285          * pins us to a particular CPU.  This avoids an smp_processor_id()
 286          * warning on Linux when debugging is enabled. */
 287         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 288
 289         if (tcd == NULL)                /* arch may not log in IRQ context */
 290                 goto console;
 291
 292         if (tcd->tcd_cur_pages == 0)
 293                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 294
 295         if (tcd->tcd_shutting_down) {
 296                 cfs_trace_put_tcd(tcd);
 297                 tcd = NULL;
 298                 goto console;
 299         }
 300
 301         depth = __current_nesting_level();
 302         known_size = strlen(file) + 1 + depth;
 303         if (msgdata->msg_fn)
 304                 known_size += strlen(msgdata->msg_fn) + 1;
 305
 306         if (libcfs_debug_binary)
 307                 known_size += sizeof(header);
 308
 309         /*/
 310          * '2' used because vsnprintf return real size required for output
 311          * _without_ terminating NULL.
 312          * if needed is to small for this format.
 313          */
 314         for (i = 0; i < 2; i++) {
 315                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 316                 if (tage == NULL) {
 317                         if (needed + known_size > CFS_PAGE_SIZE)
 318                                 mask |= D_ERROR;
 319
 320                         cfs_trace_put_tcd(tcd);
 321                         tcd = NULL;
 322                         goto console;
 323                 }
 324
 325                 string_buf = (char *)cfs_page_address(tage->page) +
 326                                         tage->used + known_size;
 327
 328                 max_nob = CFS_PAGE_SIZE - tage->used - known_size;
 329                 if (max_nob <= 0) {
 330                         printk(CFS_KERN_EMERG "negative max_nob: %d\n",
 331                                max_nob);
 332                         mask |= D_ERROR;
 333                         cfs_trace_put_tcd(tcd);
 334                         tcd = NULL;
 335                         goto console;
 336                 }
 337
 338                 needed = 0;
 339                 if (format1) {
 340                         va_copy(ap, args);
 341                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 342                         va_end(ap);
 343                 }
 344
 345                 if (format2) {
 346                         remain = max_nob - needed;
 347                         if (remain < 0)
 348                                 remain = 0;
 349
 350                         va_start(ap, format2);
 351                         needed += vsnprintf(string_buf + needed, remain,
 352                                             format2, ap);
 353                         va_end(ap);
 354                 }
 355
 356                 if (needed < max_nob) /* well. printing ok.. */
 357                         break;
 358         }
 359
 360         if (*(string_buf+needed-1) != '\n')
 361                 printk(CFS_KERN_INFO "format at %s:%d:%s doesn't end in "
 362                        "newline\n", file, msgdata->msg_line, msgdata->msg_fn);
 363
 364         header.ph_len = known_size + needed;
 365         debug_buf = (char *)cfs_page_address(tage->page) + tage->used;
 366
 367         if (libcfs_debug_binary) {
 368                 memcpy(debug_buf, &header, sizeof(header));
 369                 tage->used += sizeof(header);
 370                 debug_buf += sizeof(header);
 371         }
 372
 373         /* indent message according to the nesting level */
 374         while (depth-- > 0) {
 375                 *(debug_buf++) = '.';
 376                 ++ tage->used;
 377         }
 378
 379         strcpy(debug_buf, file);
 380         tage->used += strlen(file) + 1;
 381         debug_buf += strlen(file) + 1;
 382
 383         if (msgdata->msg_fn) {
 384                 strcpy(debug_buf, msgdata->msg_fn);
 385                 tage->used += strlen(msgdata->msg_fn) + 1;
 386                 debug_buf += strlen(msgdata->msg_fn) + 1;
 387         }
 388
 389         __LASSERT(debug_buf == string_buf);
 390
 391         tage->used += needed;
 392         __LASSERT (tage->used <= CFS_PAGE_SIZE);
 393
 394 console:
 395         if ((mask & libcfs_printk) == 0) {
 396                 /* no console output requested */
 397                 if (tcd != NULL)
 398                         cfs_trace_put_tcd(tcd);
 399                 return 1;
 400         }
 401
 402         if (cdls != NULL) {
 403                 if (libcfs_console_ratelimit &&
 404                     cdls->cdls_next != 0 &&     /* not first time ever */
 405                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 406                         /* skipping a console message */
 407                         cdls->cdls_count++;
 408                         if (tcd != NULL)
 409                                 cfs_trace_put_tcd(tcd);
 410                         return 1;
 411                 }
 412
 413                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 414                                                        libcfs_console_max_delay
 415                                                        + cfs_time_seconds(10))) {
 416                         /* last timeout was a long time ago */
 417                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 418                 } else {
 419                         cdls->cdls_delay *= libcfs_console_backoff;
 420
 421                         if (cdls->cdls_delay < libcfs_console_min_delay)
 422                                 cdls->cdls_delay = libcfs_console_min_delay;
 423                         else if (cdls->cdls_delay > libcfs_console_max_delay)
 424                                 cdls->cdls_delay = libcfs_console_max_delay;
 425                 }
 426
 427                 /* ensure cdls_next is never zero after it's been seen */
 428                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 429         }
 430
 431         if (tcd != NULL) {
 432                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 433                                      msgdata->msg_fn);
 434                 cfs_trace_put_tcd(tcd);
 435         } else {
 436                 string_buf = cfs_trace_get_console_buffer();
 437
 438                 needed = 0;
 439                 if (format1 != NULL) {
 440                         va_copy(ap, args);
 441                         needed = vsnprintf(string_buf,
 442                                            CFS_TRACE_CONSOLE_BUFFER_SIZE,
 443                                            format1, ap);
 444                         va_end(ap);
 445                 }
 446                 if (format2 != NULL) {
 447                         remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
 448                         if (remain > 0) {
 449                                 va_start(ap, format2);
 450                                 needed += vsnprintf(string_buf+needed, remain,
 451                                                     format2, ap);
 452                                 va_end(ap);
 453                         }
 454                 }
 455                 cfs_print_to_console(&header, mask,
 456                                      string_buf, needed, file, msgdata->msg_fn);
 457
 458                 cfs_trace_put_console_buffer(string_buf);
 459         }
 460
 461         if (cdls != NULL && cdls->cdls_count != 0) {
 462                 string_buf = cfs_trace_get_console_buffer();
 463
 464                 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 465                                   "Skipped %d previous similar message%s\n",
 466                                   cdls->cdls_count,
 467                                   (cdls->cdls_count > 1) ? "s" : "");
 468
 469                 cfs_print_to_console(&header, mask,
 470                                      string_buf, needed, file, msgdata->msg_fn);
 471
 472                 cfs_trace_put_console_buffer(string_buf);
 473                 cdls->cdls_count = 0;
 474         }
 475
 476         return 0;
 477 }
 478 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 479
 480 void
 481 libcfs_assertion_failed(const char *expr, struct libcfs_debug_msg_data *msgdata)
 482 {
 483         libcfs_debug_msg(msgdata, "ASSERTION(%s) failed\n", expr);
 484         /* cfs_enter_debugger(); */
 485         lbug_with_loc(msgdata);
 486 }
 487 EXPORT_SYMBOL(libcfs_assertion_failed);
 488
 489 void
 490 cfs_trace_assertion_failed(const char *str,
 491                            struct libcfs_debug_msg_data *msgdata)
 492 {
 493         struct ptldebug_header hdr;
 494
 495         libcfs_panic_in_progress = 1;
 496         libcfs_catastrophe = 1;
 497         cfs_mb();
 498
 499         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 500
 501         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 502                              msgdata->msg_file, msgdata->msg_fn);
 503
 504         LIBCFS_PANIC("Lustre debug assertion failure\n");
 505
 506         /* not reached */
 507 }
 508
 509 static void
 510 panic_collect_pages(struct page_collection *pc)
 511 {
 512         /* Do the collect_pages job on a single CPU: assumes that all other
 513          * CPUs have been stopped during a panic.  If this isn't true for some
 514          * arch, this will have to be implemented separately in each arch.  */
 515         int                        i;
 516         int                        j;
 517         struct cfs_trace_cpu_data *tcd;
 518
 519         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 520
 521         cfs_tcd_for_each(tcd, i, j) {
 522                 cfs_list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 523                 tcd->tcd_cur_pages = 0;
 524
 525                 if (pc->pc_want_daemon_pages) {
 526                         cfs_list_splice_init(&tcd->tcd_daemon_pages,
 527                                              &pc->pc_pages);
 528                         tcd->tcd_cur_daemon_pages = 0;
 529                 }
 530         }
 531 }
 532
 533 static void collect_pages_on_all_cpus(struct page_collection *pc)
 534 {
 535         struct cfs_trace_cpu_data *tcd;
 536         int i, cpu;
 537
 538         cfs_spin_lock(&pc->pc_lock);
 539         cfs_for_each_possible_cpu(cpu) {
 540                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 541                         cfs_list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 542                         tcd->tcd_cur_pages = 0;
 543                         if (pc->pc_want_daemon_pages) {
 544                                 cfs_list_splice_init(&tcd->tcd_daemon_pages,
 545                                                      &pc->pc_pages);
 546                                 tcd->tcd_cur_daemon_pages = 0;
 547                         }
 548                 }
 549         }
 550         cfs_spin_unlock(&pc->pc_lock);
 551 }
 552
 553 static void collect_pages(struct page_collection *pc)
 554 {
 555         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 556
 557         if (libcfs_panic_in_progress)
 558                 panic_collect_pages(pc);
 559         else
 560                 collect_pages_on_all_cpus(pc);
 561 }
 562
 563 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 564 {
 565         struct cfs_trace_cpu_data *tcd;
 566         cfs_list_t *cur_head;
 567         struct cfs_trace_page *tage;
 568         struct cfs_trace_page *tmp;
 569         int i, cpu;
 570
 571         cfs_spin_lock(&pc->pc_lock);
 572         cfs_for_each_possible_cpu(cpu) {
 573                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 574                         cur_head = tcd->tcd_pages.next;
 575
 576                         cfs_list_for_each_entry_safe_typed(tage, tmp,
 577                                                            &pc->pc_pages,
 578                                                            struct cfs_trace_page,
 579                                                            linkage) {
 580
 581                                 __LASSERT_TAGE_INVARIANT(tage);
 582
 583                                 if (tage->cpu != cpu || tage->type != i)
 584                                         continue;
 585
 586                                 cfs_tage_to_tail(tage, cur_head);
 587                                 tcd->tcd_cur_pages++;
 588                         }
 589                 }
 590         }
 591         cfs_spin_unlock(&pc->pc_lock);
 592 }
 593
 594 static void put_pages_back(struct page_collection *pc)
 595 {
 596         if (!libcfs_panic_in_progress)
 597                 put_pages_back_on_all_cpus(pc);
 598 }
 599
 600 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 601  * we have a good amount of data at all times for dumping during an LBUG, even
 602  * if we have been steadily writing (and otherwise discarding) pages via the
 603  * debug daemon. */
 604 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 605                                          struct cfs_trace_cpu_data *tcd)
 606 {
 607         struct cfs_trace_page *tage;
 608         struct cfs_trace_page *tmp;
 609
 610         cfs_spin_lock(&pc->pc_lock);
 611         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc->pc_pages,
 612                                            struct cfs_trace_page, linkage) {
 613
 614                 __LASSERT_TAGE_INVARIANT(tage);
 615
 616                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 617                         continue;
 618
 619                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 620                 tcd->tcd_cur_daemon_pages++;
 621
 622                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 623                         struct cfs_trace_page *victim;
 624
 625                         __LASSERT(!cfs_list_empty(&tcd->tcd_daemon_pages));
 626                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 627
 628                         __LASSERT_TAGE_INVARIANT(victim);
 629
 630                         cfs_list_del(&victim->linkage);
 631                         cfs_tage_free(victim);
 632                         tcd->tcd_cur_daemon_pages--;
 633                 }
 634         }
 635         cfs_spin_unlock(&pc->pc_lock);
 636 }
 637
 638 static void put_pages_on_daemon_list(struct page_collection *pc)
 639 {
 640         struct cfs_trace_cpu_data *tcd;
 641         int i, cpu;
 642
 643         cfs_for_each_possible_cpu(cpu) {
 644                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 645                         put_pages_on_tcd_daemon_list(pc, tcd);
 646         }
 647 }
 648
 649 void cfs_trace_debug_print(void)
 650 {
 651         struct page_collection pc;
 652         struct cfs_trace_page *tage;
 653         struct cfs_trace_page *tmp;
 654
 655         cfs_spin_lock_init(&pc.pc_lock);
 656
 657         pc.pc_want_daemon_pages = 1;
 658         collect_pages(&pc);
 659         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 660                                            struct cfs_trace_page, linkage) {
 661                 char *p, *file, *fn;
 662                 cfs_page_t *page;
 663
 664                 __LASSERT_TAGE_INVARIANT(tage);
 665
 666                 page = tage->page;
 667                 p = cfs_page_address(page);
 668                 while (p < ((char *)cfs_page_address(page) + tage->used)) {
 669                         struct ptldebug_header *hdr;
 670                         int len;
 671                         hdr = (void *)p;
 672                         p += sizeof(*hdr);
 673                         file = p;
 674                         p += strlen(file) + 1;
 675                         fn = p;
 676                         p += strlen(fn) + 1;
 677                         len = hdr->ph_len - (int)(p - (char *)hdr);
 678
 679                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 680
 681                         p += len;
 682                 }
 683
 684                 cfs_list_del(&tage->linkage);
 685                 cfs_tage_free(tage);
 686         }
 687 }
 688
 689 int cfs_tracefile_dump_all_pages(char *filename)
 690 {
 691         struct page_collection pc;
 692         cfs_file_t *filp;
 693         struct cfs_trace_page *tage;
 694         struct cfs_trace_page *tmp;
 695         int rc;
 696
 697         CFS_DECL_MMSPACE;
 698
 699         cfs_tracefile_write_lock();
 700
 701         filp = cfs_filp_open(filename,
 702                              O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600, &rc);
 703         if (!filp) {
 704                 if (rc != -EEXIST)
 705                         printk(CFS_KERN_ERR
 706                                "LustreError: can't open %s for dump: rc %d\n",
 707                                filename, rc);
 708                 goto out;
 709         }
 710
 711         cfs_spin_lock_init(&pc.pc_lock);
 712         pc.pc_want_daemon_pages = 1;
 713         collect_pages(&pc);
 714         if (cfs_list_empty(&pc.pc_pages)) {
 715                 rc = 0;
 716                 goto close;
 717         }
 718
 719         /* ok, for now, just write the pages.  in the future we'll be building
 720          * iobufs with the pages and calling generic_direct_IO */
 721         CFS_MMSPACE_OPEN;
 722         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 723                                            struct cfs_trace_page, linkage) {
 724
 725                 __LASSERT_TAGE_INVARIANT(tage);
 726
 727                 rc = cfs_filp_write(filp, cfs_page_address(tage->page),
 728                                     tage->used, cfs_filp_poff(filp));
 729                 if (rc != (int)tage->used) {
 730                         printk(CFS_KERN_WARNING "wanted to write %u but wrote "
 731                                "%d\n", tage->used, rc);
 732                         put_pages_back(&pc);
 733                         __LASSERT(cfs_list_empty(&pc.pc_pages));
 734                         break;
 735                 }
 736                 cfs_list_del(&tage->linkage);
 737                 cfs_tage_free(tage);
 738         }
 739         CFS_MMSPACE_CLOSE;
 740         rc = cfs_filp_fsync(filp);
 741         if (rc)
 742                 printk(CFS_KERN_ERR "sync returns %d\n", rc);
 743  close:
 744         cfs_filp_close(filp);
 745  out:
 746         cfs_tracefile_write_unlock();
 747         return rc;
 748 }
 749
 750 void cfs_trace_flush_pages(void)
 751 {
 752         struct page_collection pc;
 753         struct cfs_trace_page *tage;
 754         struct cfs_trace_page *tmp;
 755
 756         cfs_spin_lock_init(&pc.pc_lock);
 757
 758         pc.pc_want_daemon_pages = 1;
 759         collect_pages(&pc);
 760         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 761                                            struct cfs_trace_page, linkage) {
 762
 763                 __LASSERT_TAGE_INVARIANT(tage);
 764
 765                 cfs_list_del(&tage->linkage);
 766                 cfs_tage_free(tage);
 767         }
 768 }
 769
 770 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 771                             const char *usr_buffer, int usr_buffer_nob)
 772 {
 773         int    nob;
 774
 775         if (usr_buffer_nob > knl_buffer_nob)
 776                 return -EOVERFLOW;
 777
 778         if (cfs_copy_from_user((void *)knl_buffer,
 779                            (void *)usr_buffer, usr_buffer_nob))
 780                 return -EFAULT;
 781
 782         nob = strnlen(knl_buffer, usr_buffer_nob);
 783         while (nob-- >= 0)                      /* strip trailing whitespace */
 784                 if (!isspace(knl_buffer[nob]))
 785                         break;
 786
 787         if (nob < 0)                            /* empty string */
 788                 return -EINVAL;
 789
 790         if (nob == knl_buffer_nob)              /* no space to terminate */
 791                 return -EOVERFLOW;
 792
 793         knl_buffer[nob + 1] = 0;                /* terminate */
 794         return 0;
 795 }
 796
 797 int cfs_trace_copyout_string(char *usr_buffer, int usr_buffer_nob,
 798                              const char *knl_buffer, char *append)
 799 {
 800         /* NB if 'append' != NULL, it's a single character to append to the
 801          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 802          * terminating zero byte) for sysctl entries */
 803         int   nob = strlen(knl_buffer);
 804
 805         if (nob > usr_buffer_nob)
 806                 nob = usr_buffer_nob;
 807
 808         if (cfs_copy_to_user(usr_buffer, knl_buffer, nob))
 809                 return -EFAULT;
 810
 811         if (append != NULL && nob < usr_buffer_nob) {
 812                 if (cfs_copy_to_user(usr_buffer + nob, append, 1))
 813                         return -EFAULT;
 814
 815                 nob++;
 816         }
 817
 818         return nob;
 819 }
 820 EXPORT_SYMBOL(cfs_trace_copyout_string);
 821
 822 int cfs_trace_allocate_string_buffer(char **str, int nob)
 823 {
 824         if (nob > 2 * CFS_PAGE_SIZE)            /* string must be "sensible" */
 825                 return -EINVAL;
 826
 827         *str = cfs_alloc(nob, CFS_ALLOC_STD | CFS_ALLOC_ZERO);
 828         if (*str == NULL)
 829                 return -ENOMEM;
 830
 831         return 0;
 832 }
 833
 834 void cfs_trace_free_string_buffer(char *str, int nob)
 835 {
 836         cfs_free(str);
 837 }
 838
 839 int cfs_trace_dump_debug_buffer_usrstr(void *usr_str, int usr_str_nob)
 840 {
 841         char         *str;
 842         int           rc;
 843
 844         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 845         if (rc != 0)
 846                 return rc;
 847
 848         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 849                                      usr_str, usr_str_nob);
 850         if (rc != 0)
 851                 goto out;
 852
 853 #if !defined(__WINNT__)
 854         if (str[0] != '/') {
 855                 rc = -EINVAL;
 856                 goto out;
 857         }
 858 #endif
 859         rc = cfs_tracefile_dump_all_pages(str);
 860 out:
 861         cfs_trace_free_string_buffer(str, usr_str_nob + 1);
 862         return rc;
 863 }
 864
 865 int cfs_trace_daemon_command(char *str)
 866 {
 867         int       rc = 0;
 868
 869         cfs_tracefile_write_lock();
 870
 871         if (strcmp(str, "stop") == 0) {
 872                 cfs_tracefile_write_unlock();
 873                 cfs_trace_stop_thread();
 874                 cfs_tracefile_write_lock();
 875                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 876
 877         } else if (strncmp(str, "size=", 5) == 0) {
 878                 cfs_tracefile_size = simple_strtoul(str + 5, NULL, 0);
 879                 if (cfs_tracefile_size < 10 || cfs_tracefile_size > 20480)
 880                         cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 881                 else
 882                         cfs_tracefile_size <<= 20;
 883
 884         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 885                 rc = -ENAMETOOLONG;
 886 #ifndef __WINNT__
 887         } else if (str[0] != '/') {
 888                 rc = -EINVAL;
 889 #endif
 890         } else {
 891                 strcpy(cfs_tracefile, str);
 892
 893                 printk(CFS_KERN_INFO
 894                        "Lustre: debug daemon will attempt to start writing "
 895                        "to %s (%lukB max)\n", cfs_tracefile,
 896                        (long)(cfs_tracefile_size >> 10));
 897
 898                 cfs_trace_start_thread();
 899         }
 900
 901         cfs_tracefile_write_unlock();
 902         return rc;
 903 }
 904
 905 int cfs_trace_daemon_command_usrstr(void *usr_str, int usr_str_nob)
 906 {
 907         char *str;
 908         int   rc;
 909
 910         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 911         if (rc != 0)
 912                 return rc;
 913
 914         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 915                                  usr_str, usr_str_nob);
 916         if (rc == 0)
 917                 rc = cfs_trace_daemon_command(str);
 918
 919         cfs_trace_free_string_buffer(str, usr_str_nob + 1);
 920         return rc;
 921 }
 922
 923 int cfs_trace_set_debug_mb(int mb)
 924 {
 925         int i;
 926         int j;
 927         int pages;
 928         int limit = cfs_trace_max_debug_mb();
 929         struct cfs_trace_cpu_data *tcd;
 930
 931         if (mb < cfs_num_possible_cpus()) {
 932                 printk(CFS_KERN_WARNING
 933                        "Lustre: %d MB is too small for debug buffer size, "
 934                        "setting it to %d MB.\n", mb, cfs_num_possible_cpus());
 935                 mb = cfs_num_possible_cpus();
 936         }
 937
 938         if (mb > limit) {
 939                 printk(CFS_KERN_WARNING
 940                        "Lustre: %d MB is too large for debug buffer size, "
 941                        "setting it to %d MB.\n", mb, limit);
 942                 mb = limit;
 943         }
 944
 945         mb /= cfs_num_possible_cpus();
 946         pages = mb << (20 - CFS_PAGE_SHIFT);
 947
 948         cfs_tracefile_write_lock();
 949
 950         cfs_tcd_for_each(tcd, i, j)
 951                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 952
 953         cfs_tracefile_write_unlock();
 954
 955         return 0;
 956 }
 957
 958 int cfs_trace_set_debug_mb_usrstr(void *usr_str, int usr_str_nob)
 959 {
 960         char     str[32];
 961         int      rc;
 962
 963         rc = cfs_trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 964         if (rc < 0)
 965                 return rc;
 966
 967         return cfs_trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 968 }
 969
 970 int cfs_trace_get_debug_mb(void)
 971 {
 972         int i;
 973         int j;
 974         struct cfs_trace_cpu_data *tcd;
 975         int total_pages = 0;
 976
 977         cfs_tracefile_read_lock();
 978
 979         cfs_tcd_for_each(tcd, i, j)
 980                 total_pages += tcd->tcd_max_pages;
 981
 982         cfs_tracefile_read_unlock();
 983
 984         return (total_pages >> (20 - CFS_PAGE_SHIFT)) + 1;
 985 }
 986
 987 static int tracefiled(void *arg)
 988 {
 989         struct page_collection pc;
 990         struct tracefiled_ctl *tctl = arg;
 991         struct cfs_trace_page *tage;
 992         struct cfs_trace_page *tmp;
 993         cfs_file_t *filp;
 994         int last_loop = 0;
 995         int rc;
 996
 997         CFS_DECL_MMSPACE;
 998
 999         /* we're started late enough that we pick up init's fs context */
1000         /* this is so broken in uml?  what on earth is going on? */
1001         cfs_daemonize("ktracefiled");
1002
1003         cfs_spin_lock_init(&pc.pc_lock);
1004         cfs_complete(&tctl->tctl_start);
1005
1006         while (1) {
1007                 cfs_waitlink_t __wait;
1008
1009                 pc.pc_want_daemon_pages = 0;
1010                 collect_pages(&pc);
1011                 if (cfs_list_empty(&pc.pc_pages))
1012                         goto end_loop;
1013
1014                 filp = NULL;
1015                 cfs_tracefile_read_lock();
1016                 if (cfs_tracefile[0] != 0) {
1017                         filp = cfs_filp_open(cfs_tracefile,
1018                                              O_CREAT | O_RDWR | O_LARGEFILE,
1019                                              0600, &rc);
1020                         if (!(filp))
1021                                 printk(CFS_KERN_WARNING "couldn't open %s: "
1022                                        "%d\n", cfs_tracefile, rc);
1023                 }
1024                 cfs_tracefile_read_unlock();
1025                 if (filp == NULL) {
1026                         put_pages_on_daemon_list(&pc);
1027                         __LASSERT(cfs_list_empty(&pc.pc_pages));
1028                         goto end_loop;
1029                 }
1030
1031                 CFS_MMSPACE_OPEN;
1032
1033                 cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
1034                                                    struct cfs_trace_page,
1035                                                    linkage) {
1036                         static loff_t f_pos;
1037
1038                         __LASSERT_TAGE_INVARIANT(tage);
1039
1040                         if (f_pos >= (off_t)cfs_tracefile_size)
1041                                 f_pos = 0;
1042                         else if (f_pos > (off_t)cfs_filp_size(filp))
1043                                 f_pos = cfs_filp_size(filp);
1044
1045                         rc = cfs_filp_write(filp, cfs_page_address(tage->page),
1046                                             tage->used, &f_pos);
1047                         if (rc != (int)tage->used) {
1048                                 printk(CFS_KERN_WARNING "wanted to write %u "
1049                                        "but wrote %d\n", tage->used, rc);
1050                                 put_pages_back(&pc);
1051                                 __LASSERT(cfs_list_empty(&pc.pc_pages));
1052                         }
1053                 }
1054                 CFS_MMSPACE_CLOSE;
1055
1056                 cfs_filp_close(filp);
1057                 put_pages_on_daemon_list(&pc);
1058                 if (!cfs_list_empty(&pc.pc_pages)) {
1059                         int i;
1060
1061                         printk(CFS_KERN_ALERT "Lustre: trace pages aren't "
1062                                " empty\n");
1063                         printk(CFS_KERN_ERR "total cpus(%d): ",
1064                                cfs_num_possible_cpus());
1065                         for (i = 0; i < cfs_num_possible_cpus(); i++)
1066                                 if (cpu_online(i))
1067                                         printk(CFS_KERN_ERR "%d(on) ", i);
1068                                 else
1069                                         printk(CFS_KERN_ERR "%d(off) ", i);
1070                         printk(CFS_KERN_ERR "\n");
1071
1072                         i = 0;
1073                         cfs_list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1074                                                      linkage)
1075                                 printk(CFS_KERN_ERR "page %d belongs to cpu "
1076                                        "%d\n", ++i, tage->cpu);
1077                         printk(CFS_KERN_ERR "There are %d pages unwritten\n",
1078                                i);
1079                 }
1080                 __LASSERT(cfs_list_empty(&pc.pc_pages));
1081 end_loop:
1082                 if (cfs_atomic_read(&tctl->tctl_shutdown)) {
1083                         if (last_loop == 0) {
1084                                 last_loop = 1;
1085                                 continue;
1086                         } else {
1087                                 break;
1088                         }
1089                 }
1090                 cfs_waitlink_init(&__wait);
1091                 cfs_waitq_add(&tctl->tctl_waitq, &__wait);
1092                 cfs_set_current_state(CFS_TASK_INTERRUPTIBLE);
1093                 cfs_waitq_timedwait(&__wait, CFS_TASK_INTERRUPTIBLE,
1094                                     cfs_time_seconds(1));
1095                 cfs_waitq_del(&tctl->tctl_waitq, &__wait);
1096         }
1097         cfs_complete(&tctl->tctl_stop);
1098         return 0;
1099 }
1100
1101 int cfs_trace_start_thread(void)
1102 {
1103         struct tracefiled_ctl *tctl = &trace_tctl;
1104         int rc = 0;
1105
1106         cfs_mutex_lock(&cfs_trace_thread_mutex);
1107         if (thread_running)
1108                 goto out;
1109
1110         cfs_init_completion(&tctl->tctl_start);
1111         cfs_init_completion(&tctl->tctl_stop);
1112         cfs_waitq_init(&tctl->tctl_waitq);
1113         cfs_atomic_set(&tctl->tctl_shutdown, 0);
1114
1115         if (cfs_create_thread(tracefiled, tctl, 0) < 0) {
1116                 rc = -ECHILD;
1117                 goto out;
1118         }
1119
1120         cfs_wait_for_completion(&tctl->tctl_start);
1121         thread_running = 1;
1122 out:
1123         cfs_mutex_unlock(&cfs_trace_thread_mutex);
1124         return rc;
1125 }
1126
1127 void cfs_trace_stop_thread(void)
1128 {
1129         struct tracefiled_ctl *tctl = &trace_tctl;
1130
1131         cfs_mutex_lock(&cfs_trace_thread_mutex);
1132         if (thread_running) {
1133                 printk(CFS_KERN_INFO
1134                        "Lustre: shutting down debug daemon thread...\n");
1135                 cfs_atomic_set(&tctl->tctl_shutdown, 1);
1136                 cfs_wait_for_completion(&tctl->tctl_stop);
1137                 thread_running = 0;
1138         }
1139         cfs_mutex_unlock(&cfs_trace_thread_mutex);
1140 }
1141
1142 int cfs_tracefile_init(int max_pages)
1143 {
1144         struct cfs_trace_cpu_data *tcd;
1145         int                    i;
1146         int                    j;
1147         int                    rc;
1148         int                    factor;
1149
1150         rc = cfs_tracefile_init_arch();
1151         if (rc != 0)
1152                 return rc;
1153
1154         cfs_tcd_for_each(tcd, i, j) {
1155                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1156                 factor = tcd->tcd_pages_factor;
1157                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
1158                 CFS_INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1159                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1160                 tcd->tcd_cur_pages = 0;
1161                 tcd->tcd_cur_stock_pages = 0;
1162                 tcd->tcd_cur_daemon_pages = 0;
1163                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1164                 LASSERT(tcd->tcd_max_pages > 0);
1165                 tcd->tcd_shutting_down = 0;
1166         }
1167
1168         return 0;
1169 }
1170
1171 static void trace_cleanup_on_all_cpus(void)
1172 {
1173         struct cfs_trace_cpu_data *tcd;
1174         struct cfs_trace_page *tage;
1175         struct cfs_trace_page *tmp;
1176         int i, cpu;
1177
1178         cfs_for_each_possible_cpu(cpu) {
1179                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1180                         tcd->tcd_shutting_down = 1;
1181
1182                         cfs_list_for_each_entry_safe_typed(tage, tmp,
1183                                                            &tcd->tcd_pages,
1184                                                            struct cfs_trace_page,
1185                                                            linkage) {
1186                                 __LASSERT_TAGE_INVARIANT(tage);
1187
1188                                 cfs_list_del(&tage->linkage);
1189                                 cfs_tage_free(tage);
1190                         }
1191
1192                         tcd->tcd_cur_pages = 0;
1193                 }
1194         }
1195 }
1196
1197 static void cfs_trace_cleanup(void)
1198 {
1199         struct page_collection pc;
1200
1201         CFS_INIT_LIST_HEAD(&pc.pc_pages);
1202         cfs_spin_lock_init(&pc.pc_lock);
1203
1204         trace_cleanup_on_all_cpus();
1205
1206         cfs_tracefile_fini_arch();
1207 }
1208
1209 void cfs_tracefile_exit(void)
1210 {
1211         cfs_trace_stop_thread();
1212         cfs_trace_cleanup();
1213 }