libcfs/libcfs/tracefile.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * libcfs/libcfs/tracefile.c
  37  *
  38  * Author: Zach Brown <zab@clusterfs.com>
  39  * Author: Phil Schwan <phil@clusterfs.com>
  40  */
  41
  42
  43 #define DEBUG_SUBSYSTEM S_LNET
  44 #define LUSTRE_TRACEFILE_PRIVATE
  45 #include "tracefile.h"
  46
  47 #include <libcfs/libcfs.h>
  48
  49 /* XXX move things up to the top, comment */
  50 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  51
  52 char cfs_tracefile[TRACEFILE_NAME_SIZE];
  53 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
  54 static struct tracefiled_ctl trace_tctl;
  55 struct mutex cfs_trace_thread_mutex;
  56 static int thread_running = 0;
  57
  58 atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
  59
  60 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  61                                         struct cfs_trace_cpu_data *tcd);
  62
  63 static inline struct cfs_trace_page *
  64 cfs_tage_from_list(struct list_head *list)
  65 {
  66         return list_entry(list, struct cfs_trace_page, linkage);
  67 }
  68
  69 static struct cfs_trace_page *cfs_tage_alloc(int gfp)
  70 {
  71         struct page            *page;
  72         struct cfs_trace_page *tage;
  73
  74         /* My caller is trying to free memory */
  75         if (!in_interrupt() && memory_pressure_get())
  76                 return NULL;
  77
  78         /*
  79          * Don't spam console with allocation failures: they will be reported
  80          * by upper layer anyway.
  81          */
  82         gfp |= __GFP_NOWARN;
  83         page = alloc_page(gfp);
  84         if (page == NULL)
  85                 return NULL;
  86
  87         tage = kmalloc(sizeof(*tage), gfp);
  88         if (tage == NULL) {
  89                 __free_page(page);
  90                 return NULL;
  91         }
  92
  93         tage->page = page;
  94         atomic_inc(&cfs_tage_allocated);
  95         return tage;
  96 }
  97
  98 static void cfs_tage_free(struct cfs_trace_page *tage)
  99 {
 100         __LASSERT(tage != NULL);
 101         __LASSERT(tage->page != NULL);
 102
 103         __free_page(tage->page);
 104         kfree(tage);
 105         atomic_dec(&cfs_tage_allocated);
 106 }
 107
 108 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
 109                              struct list_head *queue)
 110 {
 111         __LASSERT(tage != NULL);
 112         __LASSERT(queue != NULL);
 113
 114         list_move_tail(&tage->linkage, queue);
 115 }
 116
 117 int cfs_trace_refill_stock(struct cfs_trace_cpu_data *tcd, int gfp,
 118                            struct list_head *stock)
 119 {
 120         int i;
 121
 122         /*
 123          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 124          * from here: this will lead to infinite recursion.
 125          */
 126
 127         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++ i) {
 128                 struct cfs_trace_page *tage;
 129
 130                 tage = cfs_tage_alloc(gfp);
 131                 if (tage == NULL)
 132                         break;
 133                 list_add_tail(&tage->linkage, stock);
 134         }
 135         return i;
 136 }
 137
 138 /* return a page that has 'len' bytes left at the end */
 139 static struct cfs_trace_page *
 140 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
 141 {
 142         struct cfs_trace_page *tage;
 143
 144         if (tcd->tcd_cur_pages > 0) {
 145                 __LASSERT(!list_empty(&tcd->tcd_pages));
 146                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
 147                 if (tage->used + len <= PAGE_CACHE_SIZE)
 148                         return tage;
 149         }
 150
 151         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 152                 if (tcd->tcd_cur_stock_pages > 0) {
 153                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
 154                         --tcd->tcd_cur_stock_pages;
 155                         list_del_init(&tage->linkage);
 156                 } else {
 157                         tage = cfs_tage_alloc(GFP_ATOMIC);
 158                         if (unlikely(tage == NULL)) {
 159                                 if ((!memory_pressure_get() ||
 160                                      in_interrupt()) && printk_ratelimit())
 161                                         printk(KERN_WARNING
 162                                                "cannot allocate a tage (%ld)\n",
 163                                                tcd->tcd_cur_pages);
 164                                 return NULL;
 165                         }
 166                 }
 167
 168                 tage->used = 0;
 169                 tage->cpu = smp_processor_id();
 170                 tage->type = tcd->tcd_type;
 171                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 172                 tcd->tcd_cur_pages++;
 173
 174                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 175                         struct tracefiled_ctl *tctl = &trace_tctl;
 176                         /*
 177                          * wake up tracefiled to process some pages.
 178                          */
 179                         wake_up(&tctl->tctl_waitq);
 180                 }
 181                 return tage;
 182         }
 183         return NULL;
 184 }
 185
 186 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
 187 {
 188         int pgcount = tcd->tcd_cur_pages / 10;
 189         struct page_collection pc;
 190         struct cfs_trace_page *tage;
 191         struct cfs_trace_page *tmp;
 192
 193         /*
 194          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 195          * from here: this will lead to infinite recursion.
 196          */
 197
 198         if (printk_ratelimit())
 199                 printk(KERN_WARNING "debug daemon buffer overflowed; "
 200                         "discarding 10%% of pages (%d of %ld)\n",
 201                         pgcount + 1, tcd->tcd_cur_pages);
 202
 203         INIT_LIST_HEAD(&pc.pc_pages);
 204
 205         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 206                 if (pgcount-- == 0)
 207                         break;
 208
 209                 list_move_tail(&tage->linkage, &pc.pc_pages);
 210                 tcd->tcd_cur_pages--;
 211         }
 212         put_pages_on_tcd_daemon_list(&pc, tcd);
 213 }
 214
 215 /* return a page that has 'len' bytes left at the end */
 216 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 217                                                  unsigned long len)
 218 {
 219         struct cfs_trace_page *tage;
 220
 221         /*
 222          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 223          * from here: this will lead to infinite recursion.
 224          */
 225
 226         if (len > PAGE_CACHE_SIZE) {
 227                 printk(KERN_ERR
 228                        "cowardly refusing to write %lu bytes in a page\n", len);
 229                 return NULL;
 230         }
 231
 232         tage = cfs_trace_get_tage_try(tcd, len);
 233         if (tage != NULL)
 234                 return tage;
 235         if (thread_running)
 236                 cfs_tcd_shrink(tcd);
 237         if (tcd->tcd_cur_pages > 0) {
 238                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 239                 tage->used = 0;
 240                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 241         }
 242         return tage;
 243 }
 244
 245 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 246                      const char *format, ...)
 247 {
 248         va_list args;
 249         int     rc;
 250
 251         va_start(args, format);
 252         rc = libcfs_debug_vmsg2(msgdata, format, args, NULL);
 253         va_end(args);
 254
 255         return rc;
 256 }
 257 EXPORT_SYMBOL(libcfs_debug_msg);
 258
 259 int libcfs_debug_vmsg2(struct libcfs_debug_msg_data *msgdata,
 260                        const char *format1, va_list args,
 261                        const char *format2, ...)
 262 {
 263         struct cfs_trace_cpu_data *tcd = NULL;
 264         struct ptldebug_header     header = {0};
 265         struct cfs_trace_page     *tage;
 266         /* string_buf is used only if tcd != NULL, and is always set then */
 267         char                      *string_buf = NULL;
 268         char                      *debug_buf;
 269         int                        known_size;
 270         int                        needed = 85; /* average message length */
 271         int                        max_nob;
 272         va_list                    ap;
 273         int                        depth;
 274         int                        i;
 275         int                        remain;
 276         int                        mask = msgdata->msg_mask;
 277         char                      *file = (char *)msgdata->msg_file;
 278         cfs_debug_limit_state_t   *cdls = msgdata->msg_cdls;
 279
 280         if (strchr(file, '/'))
 281                 file = strrchr(file, '/') + 1;
 282
 283         tcd = cfs_trace_get_tcd();
 284
 285         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 286          * pins us to a particular CPU.  This avoids an smp_processor_id()
 287          * warning on Linux when debugging is enabled. */
 288         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 289
 290         if (tcd == NULL)                /* arch may not log in IRQ context */
 291                 goto console;
 292
 293         if (tcd->tcd_cur_pages == 0)
 294                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 295
 296         if (tcd->tcd_shutting_down) {
 297                 cfs_trace_put_tcd(tcd);
 298                 tcd = NULL;
 299                 goto console;
 300         }
 301
 302         depth = __current_nesting_level();
 303         known_size = strlen(file) + 1 + depth;
 304         if (msgdata->msg_fn)
 305                 known_size += strlen(msgdata->msg_fn) + 1;
 306
 307         if (libcfs_debug_binary)
 308                 known_size += sizeof(header);
 309
 310         /*/
 311          * '2' used because vsnprintf return real size required for output
 312          * _without_ terminating NULL.
 313          * if needed is to small for this format.
 314          */
 315         for (i = 0; i < 2; i++) {
 316                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 317                 if (tage == NULL) {
 318                         if (needed + known_size > PAGE_CACHE_SIZE)
 319                                 mask |= D_ERROR;
 320
 321                         cfs_trace_put_tcd(tcd);
 322                         tcd = NULL;
 323                         goto console;
 324                 }
 325
 326                 string_buf = (char *)page_address(tage->page) +
 327                                         tage->used + known_size;
 328
 329                 max_nob = PAGE_CACHE_SIZE - tage->used - known_size;
 330                 if (max_nob <= 0) {
 331                         printk(KERN_EMERG "negative max_nob: %d\n",
 332                                max_nob);
 333                         mask |= D_ERROR;
 334                         cfs_trace_put_tcd(tcd);
 335                         tcd = NULL;
 336                         goto console;
 337                 }
 338
 339                 needed = 0;
 340                 if (format1) {
 341                         va_copy(ap, args);
 342                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 343                         va_end(ap);
 344                 }
 345
 346                 if (format2) {
 347                         remain = max_nob - needed;
 348                         if (remain < 0)
 349                                 remain = 0;
 350
 351                         va_start(ap, format2);
 352                         needed += vsnprintf(string_buf + needed, remain,
 353                                             format2, ap);
 354                         va_end(ap);
 355                 }
 356
 357                 if (needed < max_nob) /* well. printing ok.. */
 358                         break;
 359         }
 360
 361         if (*(string_buf+needed-1) != '\n')
 362                 printk(KERN_INFO "format at %s:%d:%s doesn't end in "
 363                        "newline\n", file, msgdata->msg_line, msgdata->msg_fn);
 364
 365         header.ph_len = known_size + needed;
 366         debug_buf = (char *)page_address(tage->page) + tage->used;
 367
 368         if (libcfs_debug_binary) {
 369                 memcpy(debug_buf, &header, sizeof(header));
 370                 tage->used += sizeof(header);
 371                 debug_buf += sizeof(header);
 372         }
 373
 374         /* indent message according to the nesting level */
 375         while (depth-- > 0) {
 376                 *(debug_buf++) = '.';
 377                 ++ tage->used;
 378         }
 379
 380         strcpy(debug_buf, file);
 381         tage->used += strlen(file) + 1;
 382         debug_buf += strlen(file) + 1;
 383
 384         if (msgdata->msg_fn) {
 385                 strcpy(debug_buf, msgdata->msg_fn);
 386                 tage->used += strlen(msgdata->msg_fn) + 1;
 387                 debug_buf += strlen(msgdata->msg_fn) + 1;
 388         }
 389
 390         __LASSERT(debug_buf == string_buf);
 391
 392         tage->used += needed;
 393         __LASSERT(tage->used <= PAGE_CACHE_SIZE);
 394
 395 console:
 396         if ((mask & libcfs_printk) == 0) {
 397                 /* no console output requested */
 398                 if (tcd != NULL)
 399                         cfs_trace_put_tcd(tcd);
 400                 return 1;
 401         }
 402
 403         if (cdls != NULL) {
 404                 if (libcfs_console_ratelimit &&
 405                     cdls->cdls_next != 0 &&     /* not first time ever */
 406                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 407                         /* skipping a console message */
 408                         cdls->cdls_count++;
 409                         if (tcd != NULL)
 410                                 cfs_trace_put_tcd(tcd);
 411                         return 1;
 412                 }
 413
 414                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 415                                                        libcfs_console_max_delay
 416                                                        + cfs_time_seconds(10))) {
 417                         /* last timeout was a long time ago */
 418                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 419                 } else {
 420                         cdls->cdls_delay *= libcfs_console_backoff;
 421                 }
 422
 423                 if (cdls->cdls_delay < libcfs_console_min_delay)
 424                         cdls->cdls_delay = libcfs_console_min_delay;
 425                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 426                         cdls->cdls_delay = libcfs_console_max_delay;
 427
 428                 /* ensure cdls_next is never zero after it's been seen */
 429                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 430         }
 431
 432         if (tcd != NULL) {
 433                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 434                                      msgdata->msg_fn);
 435                 cfs_trace_put_tcd(tcd);
 436         } else {
 437                 string_buf = cfs_trace_get_console_buffer();
 438
 439                 needed = 0;
 440                 if (format1 != NULL) {
 441                         va_copy(ap, args);
 442                         needed = vsnprintf(string_buf,
 443                                            CFS_TRACE_CONSOLE_BUFFER_SIZE,
 444                                            format1, ap);
 445                         va_end(ap);
 446                 }
 447                 if (format2 != NULL) {
 448                         remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
 449                         if (remain > 0) {
 450                                 va_start(ap, format2);
 451                                 needed += vsnprintf(string_buf+needed, remain,
 452                                                     format2, ap);
 453                                 va_end(ap);
 454                         }
 455                 }
 456                 cfs_print_to_console(&header, mask,
 457                                      string_buf, needed, file, msgdata->msg_fn);
 458
 459                 cfs_trace_put_console_buffer(string_buf);
 460         }
 461
 462         if (cdls != NULL && cdls->cdls_count != 0) {
 463                 string_buf = cfs_trace_get_console_buffer();
 464
 465                 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 466                                   "Skipped %d previous similar message%s\n",
 467                                   cdls->cdls_count,
 468                                   (cdls->cdls_count > 1) ? "s" : "");
 469
 470                 cfs_print_to_console(&header, mask,
 471                                      string_buf, needed, file, msgdata->msg_fn);
 472
 473                 cfs_trace_put_console_buffer(string_buf);
 474                 cdls->cdls_count = 0;
 475         }
 476
 477         return 0;
 478 }
 479 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 480
 481 void
 482 cfs_trace_assertion_failed(const char *str,
 483                            struct libcfs_debug_msg_data *msgdata)
 484 {
 485         struct ptldebug_header hdr;
 486
 487         libcfs_panic_in_progress = 1;
 488         libcfs_catastrophe = 1;
 489         smp_mb();
 490
 491         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 492
 493         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 494                              msgdata->msg_file, msgdata->msg_fn);
 495
 496         panic("Lustre debug assertion failure\n");
 497
 498         /* not reached */
 499 }
 500
 501 static void
 502 panic_collect_pages(struct page_collection *pc)
 503 {
 504         /* Do the collect_pages job on a single CPU: assumes that all other
 505          * CPUs have been stopped during a panic.  If this isn't true for some
 506          * arch, this will have to be implemented separately in each arch.  */
 507         int                        i;
 508         int                        j;
 509         struct cfs_trace_cpu_data *tcd;
 510
 511         INIT_LIST_HEAD(&pc->pc_pages);
 512
 513         cfs_tcd_for_each(tcd, i, j) {
 514                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 515                 tcd->tcd_cur_pages = 0;
 516
 517                 if (pc->pc_want_daemon_pages) {
 518                         list_splice_init(&tcd->tcd_daemon_pages,
 519                                                 &pc->pc_pages);
 520                         tcd->tcd_cur_daemon_pages = 0;
 521                 }
 522         }
 523 }
 524
 525 static void collect_pages_on_all_cpus(struct page_collection *pc)
 526 {
 527         struct cfs_trace_cpu_data *tcd;
 528         int i, cpu;
 529
 530         cfs_for_each_possible_cpu(cpu) {
 531                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 532                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 533                         tcd->tcd_cur_pages = 0;
 534                         if (pc->pc_want_daemon_pages) {
 535                                 list_splice_init(&tcd->tcd_daemon_pages,
 536                                                         &pc->pc_pages);
 537                                 tcd->tcd_cur_daemon_pages = 0;
 538                         }
 539                 }
 540         }
 541 }
 542
 543 static void collect_pages(struct page_collection *pc)
 544 {
 545         INIT_LIST_HEAD(&pc->pc_pages);
 546
 547         if (libcfs_panic_in_progress)
 548                 panic_collect_pages(pc);
 549         else
 550                 collect_pages_on_all_cpus(pc);
 551 }
 552
 553 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 554 {
 555         struct cfs_trace_cpu_data *tcd;
 556         struct list_head *cur_head;
 557         struct cfs_trace_page *tage;
 558         struct cfs_trace_page *tmp;
 559         int i, cpu;
 560
 561         cfs_for_each_possible_cpu(cpu) {
 562                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 563                         cur_head = tcd->tcd_pages.next;
 564
 565                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 566                                                  linkage) {
 567
 568                                 __LASSERT_TAGE_INVARIANT(tage);
 569
 570                                 if (tage->cpu != cpu || tage->type != i)
 571                                         continue;
 572
 573                                 cfs_tage_to_tail(tage, cur_head);
 574                                 tcd->tcd_cur_pages++;
 575                         }
 576                 }
 577         }
 578 }
 579
 580 static void put_pages_back(struct page_collection *pc)
 581 {
 582         if (!libcfs_panic_in_progress)
 583                 put_pages_back_on_all_cpus(pc);
 584 }
 585
 586 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 587  * we have a good amount of data at all times for dumping during an LBUG, even
 588  * if we have been steadily writing (and otherwise discarding) pages via the
 589  * debug daemon. */
 590 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 591                                          struct cfs_trace_cpu_data *tcd)
 592 {
 593         struct cfs_trace_page *tage;
 594         struct cfs_trace_page *tmp;
 595
 596         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 597                 __LASSERT_TAGE_INVARIANT(tage);
 598
 599                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 600                         continue;
 601
 602                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 603                 tcd->tcd_cur_daemon_pages++;
 604
 605                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 606                         struct cfs_trace_page *victim;
 607
 608                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 609                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 610
 611                         __LASSERT_TAGE_INVARIANT(victim);
 612
 613                         list_del(&victim->linkage);
 614                         cfs_tage_free(victim);
 615                         tcd->tcd_cur_daemon_pages--;
 616                 }
 617         }
 618 }
 619
 620 static void put_pages_on_daemon_list(struct page_collection *pc)
 621 {
 622         struct cfs_trace_cpu_data *tcd;
 623         int i, cpu;
 624
 625         cfs_for_each_possible_cpu(cpu) {
 626                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 627                         put_pages_on_tcd_daemon_list(pc, tcd);
 628         }
 629 }
 630
 631 void cfs_trace_debug_print(void)
 632 {
 633         struct page_collection pc;
 634         struct cfs_trace_page *tage;
 635         struct cfs_trace_page *tmp;
 636
 637         pc.pc_want_daemon_pages = 1;
 638         collect_pages(&pc);
 639         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 640                 char *p, *file, *fn;
 641                 struct page *page;
 642
 643                 __LASSERT_TAGE_INVARIANT(tage);
 644
 645                 page = tage->page;
 646                 p = page_address(page);
 647                 while (p < ((char *)page_address(page) + tage->used)) {
 648                         struct ptldebug_header *hdr;
 649                         int len;
 650                         hdr = (void *)p;
 651                         p += sizeof(*hdr);
 652                         file = p;
 653                         p += strlen(file) + 1;
 654                         fn = p;
 655                         p += strlen(fn) + 1;
 656                         len = hdr->ph_len - (int)(p - (char *)hdr);
 657
 658                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 659
 660                         p += len;
 661                 }
 662
 663                 list_del(&tage->linkage);
 664                 cfs_tage_free(tage);
 665         }
 666 }
 667
 668 int cfs_tracefile_dump_all_pages(char *filename)
 669 {
 670         struct page_collection  pc;
 671         struct file             *filp;
 672         struct cfs_trace_page   *tage;
 673         struct cfs_trace_page   *tmp;
 674         int rc;
 675
 676         DECL_MMSPACE;
 677
 678         cfs_tracefile_write_lock();
 679
 680         filp = filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600);
 681         if (IS_ERR(filp)) {
 682                 rc = PTR_ERR(filp);
 683                 filp = NULL;
 684                 printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 685                       filename, rc);
 686                 goto out;
 687         }
 688
 689         pc.pc_want_daemon_pages = 1;
 690         collect_pages(&pc);
 691         if (list_empty(&pc.pc_pages)) {
 692                 rc = 0;
 693                 goto close;
 694         }
 695
 696         /* ok, for now, just write the pages.  in the future we'll be building
 697          * iobufs with the pages and calling generic_direct_IO */
 698         MMSPACE_OPEN;
 699         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 700
 701                 __LASSERT_TAGE_INVARIANT(tage);
 702
 703                 rc = filp_write(filp, page_address(tage->page),
 704                                 tage->used, filp_poff(filp));
 705                 if (rc != (int)tage->used) {
 706                         printk(KERN_WARNING "wanted to write %u but wrote "
 707                                "%d\n", tage->used, rc);
 708                         put_pages_back(&pc);
 709                         __LASSERT(list_empty(&pc.pc_pages));
 710                         break;
 711                 }
 712                 list_del(&tage->linkage);
 713                 cfs_tage_free(tage);
 714         }
 715         MMSPACE_CLOSE;
 716         rc = filp_fsync(filp, 0, LLONG_MAX);
 717         if (rc)
 718                 printk(KERN_ERR "sync returns %d\n", rc);
 719 close:
 720         filp_close(filp, NULL);
 721 out:
 722         cfs_tracefile_write_unlock();
 723         return rc;
 724 }
 725
 726 void cfs_trace_flush_pages(void)
 727 {
 728         struct page_collection pc;
 729         struct cfs_trace_page *tage;
 730         struct cfs_trace_page *tmp;
 731
 732         pc.pc_want_daemon_pages = 1;
 733         collect_pages(&pc);
 734         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 735
 736                 __LASSERT_TAGE_INVARIANT(tage);
 737
 738                 list_del(&tage->linkage);
 739                 cfs_tage_free(tage);
 740         }
 741 }
 742
 743 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 744                             const char *usr_buffer, int usr_buffer_nob)
 745 {
 746         int    nob;
 747
 748         if (usr_buffer_nob > knl_buffer_nob)
 749                 return -EOVERFLOW;
 750
 751         if (copy_from_user((void *)knl_buffer,
 752                            (void *)usr_buffer, usr_buffer_nob))
 753                 return -EFAULT;
 754
 755         nob = strnlen(knl_buffer, usr_buffer_nob);
 756         while (nob-- >= 0)                      /* strip trailing whitespace */
 757                 if (!isspace(knl_buffer[nob]))
 758                         break;
 759
 760         if (nob < 0)                            /* empty string */
 761                 return -EINVAL;
 762
 763         if (nob == knl_buffer_nob)              /* no space to terminate */
 764                 return -EOVERFLOW;
 765
 766         knl_buffer[nob + 1] = 0;                /* terminate */
 767         return 0;
 768 }
 769 EXPORT_SYMBOL(cfs_trace_copyin_string);
 770
 771 int cfs_trace_copyout_string(char *usr_buffer, int usr_buffer_nob,
 772                              const char *knl_buffer, char *append)
 773 {
 774         /* NB if 'append' != NULL, it's a single character to append to the
 775          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 776          * terminating zero byte) for sysctl entries */
 777         int   nob = strlen(knl_buffer);
 778
 779         if (nob > usr_buffer_nob)
 780                 nob = usr_buffer_nob;
 781
 782         if (copy_to_user(usr_buffer, knl_buffer, nob))
 783                 return -EFAULT;
 784
 785         if (append != NULL && nob < usr_buffer_nob) {
 786                 if (copy_to_user(usr_buffer + nob, append, 1))
 787                         return -EFAULT;
 788
 789                 nob++;
 790         }
 791
 792         return nob;
 793 }
 794 EXPORT_SYMBOL(cfs_trace_copyout_string);
 795
 796 int cfs_trace_allocate_string_buffer(char **str, int nob)
 797 {
 798         if (nob > 2 * PAGE_CACHE_SIZE)  /* string must be "sensible" */
 799                 return -EINVAL;
 800
 801         *str = kmalloc(nob, GFP_IOFS | __GFP_ZERO);
 802         if (*str == NULL)
 803                 return -ENOMEM;
 804
 805         return 0;
 806 }
 807
 808 void cfs_trace_free_string_buffer(char *str, int nob)
 809 {
 810         kfree(str);
 811 }
 812
 813 int cfs_trace_dump_debug_buffer_usrstr(void *usr_str, int usr_str_nob)
 814 {
 815         char         *str;
 816         int           rc;
 817
 818         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 819         if (rc != 0)
 820                 return rc;
 821
 822         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 823                                      usr_str, usr_str_nob);
 824         if (rc != 0)
 825                 goto out;
 826
 827 #if !defined(__WINNT__)
 828         if (str[0] != '/') {
 829                 rc = -EINVAL;
 830                 goto out;
 831         }
 832 #endif
 833         rc = cfs_tracefile_dump_all_pages(str);
 834 out:
 835         cfs_trace_free_string_buffer(str, usr_str_nob + 1);
 836         return rc;
 837 }
 838
 839 int cfs_trace_daemon_command(char *str)
 840 {
 841         int       rc = 0;
 842
 843         cfs_tracefile_write_lock();
 844
 845         if (strcmp(str, "stop") == 0) {
 846                 cfs_tracefile_write_unlock();
 847                 cfs_trace_stop_thread();
 848                 cfs_tracefile_write_lock();
 849                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 850
 851         } else if (strncmp(str, "size=", 5) == 0) {
 852                 cfs_tracefile_size = simple_strtoul(str + 5, NULL, 0);
 853                 if (cfs_tracefile_size < 10 || cfs_tracefile_size > 20480)
 854                         cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 855                 else
 856                         cfs_tracefile_size <<= 20;
 857
 858         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 859                 rc = -ENAMETOOLONG;
 860 #ifndef __WINNT__
 861         } else if (str[0] != '/') {
 862                 rc = -EINVAL;
 863 #endif
 864         } else {
 865                 strcpy(cfs_tracefile, str);
 866
 867                 printk(KERN_INFO
 868                        "Lustre: debug daemon will attempt to start writing "
 869                        "to %s (%lukB max)\n", cfs_tracefile,
 870                        (long)(cfs_tracefile_size >> 10));
 871
 872                 cfs_trace_start_thread();
 873         }
 874
 875         cfs_tracefile_write_unlock();
 876         return rc;
 877 }
 878
 879 int cfs_trace_daemon_command_usrstr(void *usr_str, int usr_str_nob)
 880 {
 881         char *str;
 882         int   rc;
 883
 884         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 885         if (rc != 0)
 886                 return rc;
 887
 888         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 889                                  usr_str, usr_str_nob);
 890         if (rc == 0)
 891                 rc = cfs_trace_daemon_command(str);
 892
 893         cfs_trace_free_string_buffer(str, usr_str_nob + 1);
 894         return rc;
 895 }
 896
 897 int cfs_trace_set_debug_mb(int mb)
 898 {
 899         int i;
 900         int j;
 901         int pages;
 902         int limit = cfs_trace_max_debug_mb();
 903         struct cfs_trace_cpu_data *tcd;
 904
 905         if (mb < num_possible_cpus()) {
 906                 printk(KERN_WARNING
 907                        "Lustre: %d MB is too small for debug buffer size, "
 908                        "setting it to %d MB.\n", mb, num_possible_cpus());
 909                 mb = num_possible_cpus();
 910         }
 911
 912         if (mb > limit) {
 913                 printk(KERN_WARNING
 914                        "Lustre: %d MB is too large for debug buffer size, "
 915                        "setting it to %d MB.\n", mb, limit);
 916                 mb = limit;
 917         }
 918
 919         mb /= num_possible_cpus();
 920         pages = mb << (20 - PAGE_CACHE_SHIFT);
 921
 922         cfs_tracefile_write_lock();
 923
 924         cfs_tcd_for_each(tcd, i, j)
 925                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 926
 927         cfs_tracefile_write_unlock();
 928
 929         return 0;
 930 }
 931
 932 int cfs_trace_set_debug_mb_usrstr(void *usr_str, int usr_str_nob)
 933 {
 934         char     str[32];
 935         int      rc;
 936
 937         rc = cfs_trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 938         if (rc < 0)
 939                 return rc;
 940
 941         return cfs_trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 942 }
 943
 944 int cfs_trace_get_debug_mb(void)
 945 {
 946         int i;
 947         int j;
 948         struct cfs_trace_cpu_data *tcd;
 949         int total_pages = 0;
 950
 951         cfs_tracefile_read_lock();
 952
 953         cfs_tcd_for_each(tcd, i, j)
 954                 total_pages += tcd->tcd_max_pages;
 955
 956         cfs_tracefile_read_unlock();
 957
 958         return (total_pages >> (20 - PAGE_CACHE_SHIFT)) + 1;
 959 }
 960
 961 static int tracefiled(void *arg)
 962 {
 963         struct page_collection pc;
 964         struct tracefiled_ctl *tctl = arg;
 965         struct cfs_trace_page *tage;
 966         struct cfs_trace_page *tmp;
 967         struct file *filp;
 968         int last_loop = 0;
 969         int rc;
 970
 971         DECL_MMSPACE;
 972
 973         /* we're started late enough that we pick up init's fs context */
 974         /* this is so broken in uml?  what on earth is going on? */
 975
 976         complete(&tctl->tctl_start);
 977
 978         while (1) {
 979                 wait_queue_t __wait;
 980
 981                 pc.pc_want_daemon_pages = 0;
 982                 collect_pages(&pc);
 983                 if (list_empty(&pc.pc_pages))
 984                         goto end_loop;
 985
 986                 filp = NULL;
 987                 cfs_tracefile_read_lock();
 988                 if (cfs_tracefile[0] != 0) {
 989                         filp = filp_open(cfs_tracefile,
 990                                          O_CREAT | O_RDWR | O_LARGEFILE,
 991                                          0600);
 992                         if (IS_ERR(filp)) {
 993                                 rc = PTR_ERR(filp);
 994                                 filp = NULL;
 995                                 printk(KERN_WARNING "couldn't open %s: "
 996                                        "%d\n", cfs_tracefile, rc);
 997                         }
 998                 }
 999                 cfs_tracefile_read_unlock();
1000                 if (filp == NULL) {
1001                         put_pages_on_daemon_list(&pc);
1002                         __LASSERT(list_empty(&pc.pc_pages));
1003                         goto end_loop;
1004                 }
1005
1006                 MMSPACE_OPEN;
1007
1008                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
1009                         static loff_t f_pos;
1010
1011                         __LASSERT_TAGE_INVARIANT(tage);
1012
1013                         if (f_pos >= (off_t)cfs_tracefile_size)
1014                                 f_pos = 0;
1015                         else if (f_pos > (off_t)filp_size(filp))
1016                                 f_pos = filp_size(filp);
1017
1018                         rc = filp_write(filp, page_address(tage->page),
1019                                         tage->used, &f_pos);
1020                         if (rc != (int)tage->used) {
1021                                 printk(KERN_WARNING "wanted to write %u "
1022                                        "but wrote %d\n", tage->used, rc);
1023                                 put_pages_back(&pc);
1024                                 __LASSERT(list_empty(&pc.pc_pages));
1025                                 break;
1026                         }
1027                 }
1028                 MMSPACE_CLOSE;
1029
1030                 filp_close(filp, NULL);
1031                 put_pages_on_daemon_list(&pc);
1032                 if (!list_empty(&pc.pc_pages)) {
1033                         int i;
1034
1035                         printk(KERN_ALERT "Lustre: trace pages aren't "
1036                                " empty\n");
1037                         printk(KERN_ERR "total cpus(%d): ",
1038                                num_possible_cpus());
1039                         for (i = 0; i < num_possible_cpus(); i++)
1040                                 if (cpu_online(i))
1041                                         printk(KERN_ERR "%d(on) ", i);
1042                                 else
1043                                         printk(KERN_ERR "%d(off) ", i);
1044                         printk(KERN_ERR "\n");
1045
1046                         i = 0;
1047                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1048                                                      linkage)
1049                                 printk(KERN_ERR "page %d belongs to cpu "
1050                                        "%d\n", ++i, tage->cpu);
1051                         printk(KERN_ERR "There are %d pages unwritten\n",
1052                                i);
1053                 }
1054                 __LASSERT(list_empty(&pc.pc_pages));
1055 end_loop:
1056                 if (atomic_read(&tctl->tctl_shutdown)) {
1057                         if (last_loop == 0) {
1058                                 last_loop = 1;
1059                                 continue;
1060                         } else {
1061                                 break;
1062                         }
1063                 }
1064                 init_waitqueue_entry_current(&__wait);
1065                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1066                 set_current_state(TASK_INTERRUPTIBLE);
1067                 waitq_timedwait(&__wait, TASK_INTERRUPTIBLE,
1068                                 cfs_time_seconds(1));
1069                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1070         }
1071         complete(&tctl->tctl_stop);
1072         return 0;
1073 }
1074
1075 int cfs_trace_start_thread(void)
1076 {
1077         struct tracefiled_ctl *tctl = &trace_tctl;
1078         int rc = 0;
1079
1080         mutex_lock(&cfs_trace_thread_mutex);
1081         if (thread_running)
1082                 goto out;
1083
1084         init_completion(&tctl->tctl_start);
1085         init_completion(&tctl->tctl_stop);
1086         init_waitqueue_head(&tctl->tctl_waitq);
1087         atomic_set(&tctl->tctl_shutdown, 0);
1088
1089         if (IS_ERR(kthread_run(tracefiled, tctl, "ktracefiled"))) {
1090                 rc = -ECHILD;
1091                 goto out;
1092         }
1093
1094         wait_for_completion(&tctl->tctl_start);
1095         thread_running = 1;
1096 out:
1097         mutex_unlock(&cfs_trace_thread_mutex);
1098         return rc;
1099 }
1100
1101 void cfs_trace_stop_thread(void)
1102 {
1103         struct tracefiled_ctl *tctl = &trace_tctl;
1104
1105         mutex_lock(&cfs_trace_thread_mutex);
1106         if (thread_running) {
1107                 printk(KERN_INFO
1108                        "Lustre: shutting down debug daemon thread...\n");
1109                 atomic_set(&tctl->tctl_shutdown, 1);
1110                 wait_for_completion(&tctl->tctl_stop);
1111                 thread_running = 0;
1112         }
1113         mutex_unlock(&cfs_trace_thread_mutex);
1114 }
1115
1116 int cfs_tracefile_init(int max_pages)
1117 {
1118         struct cfs_trace_cpu_data *tcd;
1119         int     i;
1120         int     j;
1121         int     rc;
1122         int     factor;
1123
1124         rc = cfs_tracefile_init_arch();
1125         if (rc != 0)
1126                 return rc;
1127
1128         cfs_tcd_for_each(tcd, i, j) {
1129                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1130                 factor = tcd->tcd_pages_factor;
1131                 INIT_LIST_HEAD(&tcd->tcd_pages);
1132                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1133                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1134                 tcd->tcd_cur_pages = 0;
1135                 tcd->tcd_cur_stock_pages = 0;
1136                 tcd->tcd_cur_daemon_pages = 0;
1137                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1138                 LASSERT(tcd->tcd_max_pages > 0);
1139                 tcd->tcd_shutting_down = 0;
1140         }
1141         return 0;
1142 }
1143
1144 static void trace_cleanup_on_all_cpus(void)
1145 {
1146         struct cfs_trace_cpu_data *tcd;
1147         struct cfs_trace_page *tage;
1148         struct cfs_trace_page *tmp;
1149         int i, cpu;
1150
1151         cfs_for_each_possible_cpu(cpu) {
1152                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1153                         tcd->tcd_shutting_down = 1;
1154
1155                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1156                                 __LASSERT_TAGE_INVARIANT(tage);
1157
1158                                 list_del(&tage->linkage);
1159                                 cfs_tage_free(tage);
1160                         }
1161                         tcd->tcd_cur_pages = 0;
1162                 }
1163         }
1164 }
1165
1166 static void cfs_trace_cleanup(void)
1167 {
1168         struct page_collection pc;
1169
1170         INIT_LIST_HEAD(&pc.pc_pages);
1171
1172         trace_cleanup_on_all_cpus();
1173
1174         cfs_tracefile_fini_arch();
1175 }
1176
1177 void cfs_tracefile_exit(void)
1178 {
1179         cfs_trace_stop_thread();
1180         cfs_trace_cleanup();
1181 }