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, 2014, 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 <linux/kthread.h>
  48 #include <libcfs/libcfs.h>
  49
  50 /* XXX move things up to the top, comment */
  51 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  52
  53 char cfs_tracefile[TRACEFILE_NAME_SIZE];
  54 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
  55 static struct tracefiled_ctl trace_tctl;
  56 static DEFINE_MUTEX(cfs_trace_thread_mutex);
  57 static int thread_running = 0;
  58
  59 static atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
  60
  61 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  62                                         struct cfs_trace_cpu_data *tcd);
  63
  64 static inline struct cfs_trace_page *
  65 cfs_tage_from_list(struct list_head *list)
  66 {
  67         return list_entry(list, struct cfs_trace_page, linkage);
  68 }
  69
  70 static struct cfs_trace_page *cfs_tage_alloc(gfp_t gfp)
  71 {
  72         struct page            *page;
  73         struct cfs_trace_page *tage;
  74
  75         /* My caller is trying to free memory */
  76         if (!in_interrupt() && memory_pressure_get())
  77                 return NULL;
  78
  79         /*
  80          * Don't spam console with allocation failures: they will be reported
  81          * by upper layer anyway.
  82          */
  83         gfp |= __GFP_NOWARN;
  84         page = alloc_page(gfp);
  85         if (page == NULL)
  86                 return NULL;
  87
  88         tage = kmalloc(sizeof(*tage), gfp);
  89         if (tage == NULL) {
  90                 __free_page(page);
  91                 return NULL;
  92         }
  93
  94         tage->page = page;
  95         atomic_inc(&cfs_tage_allocated);
  96         return tage;
  97 }
  98
  99 static void cfs_tage_free(struct cfs_trace_page *tage)
 100 {
 101         __LASSERT(tage != NULL);
 102         __LASSERT(tage->page != NULL);
 103
 104         __free_page(tage->page);
 105         kfree(tage);
 106         atomic_dec(&cfs_tage_allocated);
 107 }
 108
 109 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
 110                              struct list_head *queue)
 111 {
 112         __LASSERT(tage != NULL);
 113         __LASSERT(queue != NULL);
 114
 115         list_move_tail(&tage->linkage, queue);
 116 }
 117
 118 int cfs_trace_refill_stock(struct cfs_trace_cpu_data *tcd, gfp_t gfp,
 119                            struct list_head *stock)
 120 {
 121         int i;
 122
 123         /*
 124          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 125          * from here: this will lead to infinite recursion.
 126          */
 127
 128         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++ i) {
 129                 struct cfs_trace_page *tage;
 130
 131                 tage = cfs_tage_alloc(gfp);
 132                 if (tage == NULL)
 133                         break;
 134                 list_add_tail(&tage->linkage, stock);
 135         }
 136         return i;
 137 }
 138
 139 /* return a page that has 'len' bytes left at the end */
 140 static struct cfs_trace_page *
 141 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
 142 {
 143         struct cfs_trace_page *tage;
 144
 145         if (tcd->tcd_cur_pages > 0) {
 146                 __LASSERT(!list_empty(&tcd->tcd_pages));
 147                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
 148                 if (tage->used + len <= PAGE_SIZE)
 149                         return tage;
 150         }
 151
 152         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 153                 if (tcd->tcd_cur_stock_pages > 0) {
 154                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
 155                         --tcd->tcd_cur_stock_pages;
 156                         list_del_init(&tage->linkage);
 157                 } else {
 158                         tage = cfs_tage_alloc(GFP_ATOMIC);
 159                         if (unlikely(tage == NULL)) {
 160                                 if ((!memory_pressure_get() ||
 161                                      in_interrupt()) && printk_ratelimit())
 162                                         printk(KERN_WARNING
 163                                                "cannot allocate a tage (%ld)\n",
 164                                                tcd->tcd_cur_pages);
 165                                 return NULL;
 166                         }
 167                 }
 168
 169                 tage->used = 0;
 170                 tage->cpu = smp_processor_id();
 171                 tage->type = tcd->tcd_type;
 172                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 173                 tcd->tcd_cur_pages++;
 174
 175                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 176                         struct tracefiled_ctl *tctl = &trace_tctl;
 177                         /*
 178                          * wake up tracefiled to process some pages.
 179                          */
 180                         wake_up(&tctl->tctl_waitq);
 181                 }
 182                 return tage;
 183         }
 184         return NULL;
 185 }
 186
 187 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
 188 {
 189         int pgcount = tcd->tcd_cur_pages / 10;
 190         struct page_collection pc;
 191         struct cfs_trace_page *tage;
 192         struct cfs_trace_page *tmp;
 193
 194         /*
 195          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 196          * from here: this will lead to infinite recursion.
 197          */
 198
 199         if (printk_ratelimit())
 200                 printk(KERN_WARNING "debug daemon buffer overflowed; "
 201                         "discarding 10%% of pages (%d of %ld)\n",
 202                         pgcount + 1, tcd->tcd_cur_pages);
 203
 204         INIT_LIST_HEAD(&pc.pc_pages);
 205
 206         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 207                 if (pgcount-- == 0)
 208                         break;
 209
 210                 list_move_tail(&tage->linkage, &pc.pc_pages);
 211                 tcd->tcd_cur_pages--;
 212         }
 213         put_pages_on_tcd_daemon_list(&pc, tcd);
 214 }
 215
 216 /* return a page that has 'len' bytes left at the end */
 217 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 218                                                  unsigned long len)
 219 {
 220         struct cfs_trace_page *tage;
 221
 222         /*
 223          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 224          * from here: this will lead to infinite recursion.
 225          */
 226
 227         if (len > PAGE_SIZE) {
 228                 printk(KERN_ERR
 229                        "cowardly refusing to write %lu bytes in a page\n", len);
 230                 return NULL;
 231         }
 232
 233         tage = cfs_trace_get_tage_try(tcd, len);
 234         if (tage != NULL)
 235                 return tage;
 236         if (thread_running)
 237                 cfs_tcd_shrink(tcd);
 238         if (tcd->tcd_cur_pages > 0) {
 239                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 240                 tage->used = 0;
 241                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 242         }
 243         return tage;
 244 }
 245
 246 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 247                      const char *format, ...)
 248 {
 249         va_list args;
 250         int     rc;
 251
 252         va_start(args, format);
 253         rc = libcfs_debug_vmsg2(msgdata, format, args, NULL);
 254         va_end(args);
 255
 256         return rc;
 257 }
 258 EXPORT_SYMBOL(libcfs_debug_msg);
 259
 260 int libcfs_debug_vmsg2(struct libcfs_debug_msg_data *msgdata,
 261                        const char *format1, va_list args,
 262                        const char *format2, ...)
 263 {
 264         struct cfs_trace_cpu_data *tcd = NULL;
 265         struct ptldebug_header     header = {0};
 266         struct cfs_trace_page     *tage;
 267         /* string_buf is used only if tcd != NULL, and is always set then */
 268         char                      *string_buf = NULL;
 269         char                      *debug_buf;
 270         int                        known_size;
 271         int                        needed = 85; /* average message length */
 272         int                        max_nob;
 273         va_list                    ap;
 274         int                        i;
 275         int                        remain;
 276         int                        mask = msgdata->msg_mask;
 277         char                      *file = (char *)msgdata->msg_file;
 278         struct cfs_debug_limit_state *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         known_size = strlen(file) + 1;
 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 > 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 *)page_address(tage->page) +
 326                                         tage->used + known_size;
 327
 328                 max_nob = PAGE_SIZE - tage->used - known_size;
 329                 if (max_nob <= 0) {
 330                         printk(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(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 *)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         strcpy(debug_buf, file);
 374         tage->used += strlen(file) + 1;
 375         debug_buf += strlen(file) + 1;
 376
 377         if (msgdata->msg_fn) {
 378                 strcpy(debug_buf, msgdata->msg_fn);
 379                 tage->used += strlen(msgdata->msg_fn) + 1;
 380                 debug_buf += strlen(msgdata->msg_fn) + 1;
 381         }
 382
 383         __LASSERT(debug_buf == string_buf);
 384
 385         tage->used += needed;
 386         __LASSERT(tage->used <= PAGE_SIZE);
 387
 388 console:
 389         if ((mask & libcfs_printk) == 0) {
 390                 /* no console output requested */
 391                 if (tcd != NULL)
 392                         cfs_trace_put_tcd(tcd);
 393                 return 1;
 394         }
 395
 396         if (cdls != NULL) {
 397                 if (libcfs_console_ratelimit &&
 398                     cdls->cdls_next != 0 &&     /* not first time ever */
 399                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 400                         /* skipping a console message */
 401                         cdls->cdls_count++;
 402                         if (tcd != NULL)
 403                                 cfs_trace_put_tcd(tcd);
 404                         return 1;
 405                 }
 406
 407                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 408                                                        libcfs_console_max_delay
 409                                                        + cfs_time_seconds(10))) {
 410                         /* last timeout was a long time ago */
 411                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 412                 } else {
 413                         cdls->cdls_delay *= libcfs_console_backoff;
 414                 }
 415
 416                 if (cdls->cdls_delay < libcfs_console_min_delay)
 417                         cdls->cdls_delay = libcfs_console_min_delay;
 418                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 419                         cdls->cdls_delay = libcfs_console_max_delay;
 420
 421                 /* ensure cdls_next is never zero after it's been seen */
 422                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 423         }
 424
 425         if (tcd != NULL) {
 426                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 427                                      msgdata->msg_fn);
 428                 cfs_trace_put_tcd(tcd);
 429         } else {
 430                 string_buf = cfs_trace_get_console_buffer();
 431
 432                 needed = 0;
 433                 if (format1 != NULL) {
 434                         va_copy(ap, args);
 435                         needed = vsnprintf(string_buf,
 436                                            CFS_TRACE_CONSOLE_BUFFER_SIZE,
 437                                            format1, ap);
 438                         va_end(ap);
 439                 }
 440                 if (format2 != NULL) {
 441                         remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
 442                         if (remain > 0) {
 443                                 va_start(ap, format2);
 444                                 needed += vsnprintf(string_buf+needed, remain,
 445                                                     format2, ap);
 446                                 va_end(ap);
 447                         }
 448                 }
 449                 cfs_print_to_console(&header, mask,
 450                                      string_buf, needed, file, msgdata->msg_fn);
 451
 452                 put_cpu();
 453         }
 454
 455         if (cdls != NULL && cdls->cdls_count != 0) {
 456                 string_buf = cfs_trace_get_console_buffer();
 457
 458                 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 459                                   "Skipped %d previous similar message%s\n",
 460                                   cdls->cdls_count,
 461                                   (cdls->cdls_count > 1) ? "s" : "");
 462
 463                 cfs_print_to_console(&header, mask,
 464                                      string_buf, needed, file, msgdata->msg_fn);
 465
 466                 put_cpu();
 467                 cdls->cdls_count = 0;
 468         }
 469
 470         return 0;
 471 }
 472 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 473
 474 void
 475 cfs_trace_assertion_failed(const char *str,
 476                            struct libcfs_debug_msg_data *msgdata)
 477 {
 478         struct ptldebug_header hdr;
 479
 480         libcfs_panic_in_progress = 1;
 481         libcfs_catastrophe = 1;
 482         smp_mb();
 483
 484         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 485
 486         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 487                              msgdata->msg_file, msgdata->msg_fn);
 488
 489         panic("Lustre debug assertion failure\n");
 490
 491         /* not reached */
 492 }
 493
 494 static void
 495 panic_collect_pages(struct page_collection *pc)
 496 {
 497         /* Do the collect_pages job on a single CPU: assumes that all other
 498          * CPUs have been stopped during a panic.  If this isn't true for some
 499          * arch, this will have to be implemented separately in each arch.  */
 500         int                        i;
 501         int                        j;
 502         struct cfs_trace_cpu_data *tcd;
 503
 504         INIT_LIST_HEAD(&pc->pc_pages);
 505
 506         cfs_tcd_for_each(tcd, i, j) {
 507                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 508                 tcd->tcd_cur_pages = 0;
 509
 510                 if (pc->pc_want_daemon_pages) {
 511                         list_splice_init(&tcd->tcd_daemon_pages,
 512                                                 &pc->pc_pages);
 513                         tcd->tcd_cur_daemon_pages = 0;
 514                 }
 515         }
 516 }
 517
 518 static void collect_pages_on_all_cpus(struct page_collection *pc)
 519 {
 520         struct cfs_trace_cpu_data *tcd;
 521         int i, cpu;
 522
 523         for_each_possible_cpu(cpu) {
 524                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 525                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 526                         tcd->tcd_cur_pages = 0;
 527                         if (pc->pc_want_daemon_pages) {
 528                                 list_splice_init(&tcd->tcd_daemon_pages,
 529                                                         &pc->pc_pages);
 530                                 tcd->tcd_cur_daemon_pages = 0;
 531                         }
 532                 }
 533         }
 534 }
 535
 536 static void collect_pages(struct page_collection *pc)
 537 {
 538         INIT_LIST_HEAD(&pc->pc_pages);
 539
 540         if (libcfs_panic_in_progress)
 541                 panic_collect_pages(pc);
 542         else
 543                 collect_pages_on_all_cpus(pc);
 544 }
 545
 546 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 547 {
 548         struct cfs_trace_cpu_data *tcd;
 549         struct list_head *cur_head;
 550         struct cfs_trace_page *tage;
 551         struct cfs_trace_page *tmp;
 552         int i, cpu;
 553
 554         for_each_possible_cpu(cpu) {
 555                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 556                         cur_head = tcd->tcd_pages.next;
 557
 558                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 559                                                  linkage) {
 560
 561                                 __LASSERT_TAGE_INVARIANT(tage);
 562
 563                                 if (tage->cpu != cpu || tage->type != i)
 564                                         continue;
 565
 566                                 cfs_tage_to_tail(tage, cur_head);
 567                                 tcd->tcd_cur_pages++;
 568                         }
 569                 }
 570         }
 571 }
 572
 573 static void put_pages_back(struct page_collection *pc)
 574 {
 575         if (!libcfs_panic_in_progress)
 576                 put_pages_back_on_all_cpus(pc);
 577 }
 578
 579 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 580  * we have a good amount of data at all times for dumping during an LBUG, even
 581  * if we have been steadily writing (and otherwise discarding) pages via the
 582  * debug daemon. */
 583 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 584                                          struct cfs_trace_cpu_data *tcd)
 585 {
 586         struct cfs_trace_page *tage;
 587         struct cfs_trace_page *tmp;
 588
 589         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 590                 __LASSERT_TAGE_INVARIANT(tage);
 591
 592                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 593                         continue;
 594
 595                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 596                 tcd->tcd_cur_daemon_pages++;
 597
 598                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 599                         struct cfs_trace_page *victim;
 600
 601                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 602                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 603
 604                         __LASSERT_TAGE_INVARIANT(victim);
 605
 606                         list_del(&victim->linkage);
 607                         cfs_tage_free(victim);
 608                         tcd->tcd_cur_daemon_pages--;
 609                 }
 610         }
 611 }
 612
 613 static void put_pages_on_daemon_list(struct page_collection *pc)
 614 {
 615         struct cfs_trace_cpu_data *tcd;
 616         int i, cpu;
 617
 618         for_each_possible_cpu(cpu) {
 619                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 620                         put_pages_on_tcd_daemon_list(pc, tcd);
 621         }
 622 }
 623
 624 void cfs_trace_debug_print(void)
 625 {
 626         struct page_collection pc;
 627         struct cfs_trace_page *tage;
 628         struct cfs_trace_page *tmp;
 629
 630         pc.pc_want_daemon_pages = 1;
 631         collect_pages(&pc);
 632         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 633                 char *p, *file, *fn;
 634                 struct page *page;
 635
 636                 __LASSERT_TAGE_INVARIANT(tage);
 637
 638                 page = tage->page;
 639                 p = page_address(page);
 640                 while (p < ((char *)page_address(page) + tage->used)) {
 641                         struct ptldebug_header *hdr;
 642                         int len;
 643                         hdr = (void *)p;
 644                         p += sizeof(*hdr);
 645                         file = p;
 646                         p += strlen(file) + 1;
 647                         fn = p;
 648                         p += strlen(fn) + 1;
 649                         len = hdr->ph_len - (int)(p - (char *)hdr);
 650
 651                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 652
 653                         p += len;
 654                 }
 655
 656                 list_del(&tage->linkage);
 657                 cfs_tage_free(tage);
 658         }
 659 }
 660
 661 int cfs_tracefile_dump_all_pages(char *filename)
 662 {
 663         struct page_collection  pc;
 664         struct file             *filp;
 665         struct cfs_trace_page   *tage;
 666         struct cfs_trace_page   *tmp;
 667         mm_segment_t            __oldfs;
 668         char                    *buf;
 669         int rc;
 670
 671         cfs_tracefile_write_lock();
 672
 673         filp = filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600);
 674         if (IS_ERR(filp)) {
 675                 rc = PTR_ERR(filp);
 676                 filp = NULL;
 677                 printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 678                       filename, rc);
 679                 goto out;
 680         }
 681
 682         pc.pc_want_daemon_pages = 1;
 683         collect_pages(&pc);
 684         if (list_empty(&pc.pc_pages)) {
 685                 rc = 0;
 686                 goto close;
 687         }
 688         __oldfs = get_fs();
 689         set_fs(get_ds());
 690
 691         /* ok, for now, just write the pages.  in the future we'll be building
 692          * iobufs with the pages and calling generic_direct_IO */
 693         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 694
 695                 __LASSERT_TAGE_INVARIANT(tage);
 696
 697                 buf = kmap(tage->page);
 698                 rc = vfs_write(filp, (__force const char __user *)buf,
 699                                tage->used, &filp->f_pos);
 700                 kunmap(tage->page);
 701                 if (rc != (int)tage->used) {
 702                         printk(KERN_WARNING "wanted to write %u but wrote "
 703                                "%d\n", tage->used, rc);
 704                         put_pages_back(&pc);
 705                         __LASSERT(list_empty(&pc.pc_pages));
 706                         break;
 707                 }
 708                 list_del(&tage->linkage);
 709                 cfs_tage_free(tage);
 710         }
 711         set_fs(__oldfs);
 712         rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1);
 713         if (rc)
 714                 printk(KERN_ERR "sync returns %d\n", rc);
 715 close:
 716         filp_close(filp, NULL);
 717 out:
 718         cfs_tracefile_write_unlock();
 719         return rc;
 720 }
 721
 722 void cfs_trace_flush_pages(void)
 723 {
 724         struct page_collection pc;
 725         struct cfs_trace_page *tage;
 726         struct cfs_trace_page *tmp;
 727
 728         pc.pc_want_daemon_pages = 1;
 729         collect_pages(&pc);
 730         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 731
 732                 __LASSERT_TAGE_INVARIANT(tage);
 733
 734                 list_del(&tage->linkage);
 735                 cfs_tage_free(tage);
 736         }
 737 }
 738
 739 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 740                             const char __user *usr_buffer, int usr_buffer_nob)
 741 {
 742         int    nob;
 743
 744         if (usr_buffer_nob > knl_buffer_nob)
 745                 return -EOVERFLOW;
 746
 747         if (copy_from_user(knl_buffer, usr_buffer, usr_buffer_nob))
 748                 return -EFAULT;
 749
 750         nob = strnlen(knl_buffer, usr_buffer_nob);
 751         while (nob-- >= 0)                      /* strip trailing whitespace */
 752                 if (!isspace(knl_buffer[nob]))
 753                         break;
 754
 755         if (nob < 0)                            /* empty string */
 756                 return -EINVAL;
 757
 758         if (nob == knl_buffer_nob)              /* no space to terminate */
 759                 return -EOVERFLOW;
 760
 761         knl_buffer[nob + 1] = 0;                /* terminate */
 762         return 0;
 763 }
 764 EXPORT_SYMBOL(cfs_trace_copyin_string);
 765
 766 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 767                              const char *knl_buffer, char *append)
 768 {
 769         /* NB if 'append' != NULL, it's a single character to append to the
 770          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 771          * terminating zero byte) for sysctl entries */
 772         int   nob = strlen(knl_buffer);
 773
 774         if (nob > usr_buffer_nob)
 775                 nob = usr_buffer_nob;
 776
 777         if (copy_to_user(usr_buffer, knl_buffer, nob))
 778                 return -EFAULT;
 779
 780         if (append != NULL && nob < usr_buffer_nob) {
 781                 if (copy_to_user(usr_buffer + nob, append, 1))
 782                         return -EFAULT;
 783
 784                 nob++;
 785         }
 786
 787         return nob;
 788 }
 789 EXPORT_SYMBOL(cfs_trace_copyout_string);
 790
 791 int cfs_trace_allocate_string_buffer(char **str, int nob)
 792 {
 793         if (nob > 2 * PAGE_SIZE)        /* string must be "sensible" */
 794                 return -EINVAL;
 795
 796         *str = kmalloc(nob, GFP_KERNEL | __GFP_ZERO);
 797         if (*str == NULL)
 798                 return -ENOMEM;
 799
 800         return 0;
 801 }
 802
 803 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
 804 {
 805         char         *str;
 806         int           rc;
 807
 808         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 809         if (rc != 0)
 810                 return rc;
 811
 812         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 813                                      usr_str, usr_str_nob);
 814         if (rc != 0)
 815                 goto out;
 816
 817         if (str[0] != '/') {
 818                 rc = -EINVAL;
 819                 goto out;
 820         }
 821         rc = cfs_tracefile_dump_all_pages(str);
 822 out:
 823         kfree(str);
 824         return rc;
 825 }
 826
 827 int cfs_trace_daemon_command(char *str)
 828 {
 829         int       rc = 0;
 830
 831         cfs_tracefile_write_lock();
 832
 833         if (strcmp(str, "stop") == 0) {
 834                 cfs_tracefile_write_unlock();
 835                 cfs_trace_stop_thread();
 836                 cfs_tracefile_write_lock();
 837                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 838
 839         } else if (strncmp(str, "size=", 5) == 0) {
 840                 cfs_tracefile_size = simple_strtoul(str + 5, NULL, 0);
 841                 if (cfs_tracefile_size < 10 || cfs_tracefile_size > 20480)
 842                         cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 843                 else
 844                         cfs_tracefile_size <<= 20;
 845
 846         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 847                 rc = -ENAMETOOLONG;
 848         } else if (str[0] != '/') {
 849                 rc = -EINVAL;
 850         } else {
 851                 strcpy(cfs_tracefile, str);
 852
 853                 printk(KERN_INFO
 854                        "Lustre: debug daemon will attempt to start writing "
 855                        "to %s (%lukB max)\n", cfs_tracefile,
 856                        (long)(cfs_tracefile_size >> 10));
 857
 858                 cfs_trace_start_thread();
 859         }
 860
 861         cfs_tracefile_write_unlock();
 862         return rc;
 863 }
 864
 865 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
 866 {
 867         char *str;
 868         int   rc;
 869
 870         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 871         if (rc != 0)
 872                 return rc;
 873
 874         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 875                                  usr_str, usr_str_nob);
 876         if (rc == 0)
 877                 rc = cfs_trace_daemon_command(str);
 878
 879         kfree(str);
 880         return rc;
 881 }
 882
 883 int cfs_trace_set_debug_mb(int mb)
 884 {
 885         int i;
 886         int j;
 887         int pages;
 888         int limit = cfs_trace_max_debug_mb();
 889         struct cfs_trace_cpu_data *tcd;
 890
 891         if (mb < num_possible_cpus()) {
 892                 printk(KERN_WARNING
 893                        "Lustre: %d MB is too small for debug buffer size, "
 894                        "setting it to %d MB.\n", mb, num_possible_cpus());
 895                 mb = num_possible_cpus();
 896         }
 897
 898         if (mb > limit) {
 899                 printk(KERN_WARNING
 900                        "Lustre: %d MB is too large for debug buffer size, "
 901                        "setting it to %d MB.\n", mb, limit);
 902                 mb = limit;
 903         }
 904
 905         mb /= num_possible_cpus();
 906         pages = mb << (20 - PAGE_SHIFT);
 907
 908         cfs_tracefile_write_lock();
 909
 910         cfs_tcd_for_each(tcd, i, j)
 911                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 912
 913         cfs_tracefile_write_unlock();
 914
 915         return 0;
 916 }
 917
 918 int cfs_trace_set_debug_mb_usrstr(void __user *usr_str, int usr_str_nob)
 919 {
 920         char     str[32];
 921         int      rc;
 922
 923         rc = cfs_trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 924         if (rc < 0)
 925                 return rc;
 926
 927         return cfs_trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 928 }
 929
 930 int cfs_trace_get_debug_mb(void)
 931 {
 932         int i;
 933         int j;
 934         struct cfs_trace_cpu_data *tcd;
 935         int total_pages = 0;
 936
 937         cfs_tracefile_read_lock();
 938
 939         cfs_tcd_for_each(tcd, i, j)
 940                 total_pages += tcd->tcd_max_pages;
 941
 942         cfs_tracefile_read_unlock();
 943
 944         return (total_pages >> (20 - PAGE_SHIFT)) + 1;
 945 }
 946
 947 static int tracefiled(void *arg)
 948 {
 949         struct page_collection pc;
 950         struct tracefiled_ctl *tctl = arg;
 951         struct cfs_trace_page *tage;
 952         struct cfs_trace_page *tmp;
 953         mm_segment_t __oldfs;
 954         struct file *filp;
 955         char *buf;
 956         int last_loop = 0;
 957         int rc;
 958
 959         /* we're started late enough that we pick up init's fs context */
 960         /* this is so broken in uml?  what on earth is going on? */
 961
 962         complete(&tctl->tctl_start);
 963
 964         while (1) {
 965                 wait_queue_t __wait;
 966
 967                 pc.pc_want_daemon_pages = 0;
 968                 collect_pages(&pc);
 969                 if (list_empty(&pc.pc_pages))
 970                         goto end_loop;
 971
 972                 filp = NULL;
 973                 cfs_tracefile_read_lock();
 974                 if (cfs_tracefile[0] != 0) {
 975                         filp = filp_open(cfs_tracefile,
 976                                          O_CREAT | O_RDWR | O_LARGEFILE,
 977                                          0600);
 978                         if (IS_ERR(filp)) {
 979                                 rc = PTR_ERR(filp);
 980                                 filp = NULL;
 981                                 printk(KERN_WARNING "couldn't open %s: "
 982                                        "%d\n", cfs_tracefile, rc);
 983                         }
 984                 }
 985                 cfs_tracefile_read_unlock();
 986                 if (filp == NULL) {
 987                         put_pages_on_daemon_list(&pc);
 988                         __LASSERT(list_empty(&pc.pc_pages));
 989                         goto end_loop;
 990                 }
 991                 __oldfs = get_fs();
 992                 set_fs(get_ds());
 993
 994                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 995                         struct dentry *de = file_dentry(filp);
 996                         static loff_t f_pos;
 997
 998                         __LASSERT_TAGE_INVARIANT(tage);
 999
1000                         if (f_pos >= (off_t)cfs_tracefile_size)
1001                                 f_pos = 0;
1002                         else if (f_pos > i_size_read(de->d_inode))
1003                                 f_pos = i_size_read(de->d_inode);
1004
1005                         buf = kmap(tage->page);
1006                         rc = vfs_write(filp, (__force const char __user *)buf,
1007                                        tage->used, &f_pos);
1008                         kunmap(tage->page);
1009                         if (rc != (int)tage->used) {
1010                                 printk(KERN_WARNING "wanted to write %u "
1011                                        "but wrote %d\n", tage->used, rc);
1012                                 put_pages_back(&pc);
1013                                 __LASSERT(list_empty(&pc.pc_pages));
1014                                 break;
1015                         }
1016                 }
1017                 set_fs(__oldfs);
1018
1019                 filp_close(filp, NULL);
1020                 put_pages_on_daemon_list(&pc);
1021                 if (!list_empty(&pc.pc_pages)) {
1022                         int i;
1023
1024                         printk(KERN_ALERT "Lustre: trace pages aren't "
1025                                " empty\n");
1026                         printk(KERN_ERR "total cpus(%d): ",
1027                                num_possible_cpus());
1028                         for (i = 0; i < num_possible_cpus(); i++)
1029                                 if (cpu_online(i))
1030                                         printk(KERN_ERR "%d(on) ", i);
1031                                 else
1032                                         printk(KERN_ERR "%d(off) ", i);
1033                         printk(KERN_ERR "\n");
1034
1035                         i = 0;
1036                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1037                                                      linkage)
1038                                 printk(KERN_ERR "page %d belongs to cpu "
1039                                        "%d\n", ++i, tage->cpu);
1040                         printk(KERN_ERR "There are %d pages unwritten\n",
1041                                i);
1042                 }
1043                 __LASSERT(list_empty(&pc.pc_pages));
1044 end_loop:
1045                 if (atomic_read(&tctl->tctl_shutdown)) {
1046                         if (last_loop == 0) {
1047                                 last_loop = 1;
1048                                 continue;
1049                         } else {
1050                                 break;
1051                         }
1052                 }
1053                 init_waitqueue_entry(&__wait, current);
1054                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1055                 set_current_state(TASK_INTERRUPTIBLE);
1056                 schedule_timeout(cfs_time_seconds(1));
1057                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1058         }
1059         complete(&tctl->tctl_stop);
1060         return 0;
1061 }
1062
1063 int cfs_trace_start_thread(void)
1064 {
1065         struct tracefiled_ctl *tctl = &trace_tctl;
1066         int rc = 0;
1067
1068         mutex_lock(&cfs_trace_thread_mutex);
1069         if (thread_running)
1070                 goto out;
1071
1072         init_completion(&tctl->tctl_start);
1073         init_completion(&tctl->tctl_stop);
1074         init_waitqueue_head(&tctl->tctl_waitq);
1075         atomic_set(&tctl->tctl_shutdown, 0);
1076
1077         if (IS_ERR(kthread_run(tracefiled, tctl, "ktracefiled"))) {
1078                 rc = -ECHILD;
1079                 goto out;
1080         }
1081
1082         wait_for_completion(&tctl->tctl_start);
1083         thread_running = 1;
1084 out:
1085         mutex_unlock(&cfs_trace_thread_mutex);
1086         return rc;
1087 }
1088
1089 void cfs_trace_stop_thread(void)
1090 {
1091         struct tracefiled_ctl *tctl = &trace_tctl;
1092
1093         mutex_lock(&cfs_trace_thread_mutex);
1094         if (thread_running) {
1095                 printk(KERN_INFO
1096                        "Lustre: shutting down debug daemon thread...\n");
1097                 atomic_set(&tctl->tctl_shutdown, 1);
1098                 wait_for_completion(&tctl->tctl_stop);
1099                 thread_running = 0;
1100         }
1101         mutex_unlock(&cfs_trace_thread_mutex);
1102 }
1103
1104 int cfs_tracefile_init(int max_pages)
1105 {
1106         struct cfs_trace_cpu_data *tcd;
1107         int     i;
1108         int     j;
1109         int     rc;
1110         int     factor;
1111
1112         rc = cfs_tracefile_init_arch();
1113         if (rc != 0)
1114                 return rc;
1115
1116         cfs_tcd_for_each(tcd, i, j) {
1117                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1118                 factor = tcd->tcd_pages_factor;
1119                 INIT_LIST_HEAD(&tcd->tcd_pages);
1120                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1121                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1122                 tcd->tcd_cur_pages = 0;
1123                 tcd->tcd_cur_stock_pages = 0;
1124                 tcd->tcd_cur_daemon_pages = 0;
1125                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1126                 LASSERT(tcd->tcd_max_pages > 0);
1127                 tcd->tcd_shutting_down = 0;
1128         }
1129         return 0;
1130 }
1131
1132 static void trace_cleanup_on_all_cpus(void)
1133 {
1134         struct cfs_trace_cpu_data *tcd;
1135         struct cfs_trace_page *tage;
1136         struct cfs_trace_page *tmp;
1137         int i, cpu;
1138
1139         for_each_possible_cpu(cpu) {
1140                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1141                         tcd->tcd_shutting_down = 1;
1142
1143                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1144                                 __LASSERT_TAGE_INVARIANT(tage);
1145
1146                                 list_del(&tage->linkage);
1147                                 cfs_tage_free(tage);
1148                         }
1149                         tcd->tcd_cur_pages = 0;
1150                 }
1151         }
1152 }
1153
1154 static void cfs_trace_cleanup(void)
1155 {
1156         struct page_collection pc;
1157
1158         INIT_LIST_HEAD(&pc.pc_pages);
1159
1160         trace_cleanup_on_all_cpus();
1161
1162         cfs_tracefile_fini_arch();
1163 }
1164
1165 void cfs_tracefile_exit(void)
1166 {
1167         cfs_trace_stop_thread();
1168         cfs_trace_cleanup();
1169 }