drivers/staging/lustre/lnet/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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * libcfs/libcfs/tracefile.c
  33  *
  34  * Author: Zach Brown <zab@clusterfs.com>
  35  * Author: Phil Schwan <phil@clusterfs.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_LNET
  39 #define LUSTRE_TRACEFILE_PRIVATE
  40 #define pr_fmt(fmt) "Lustre: " fmt
  41 #include "tracefile.h"
  42
  43 #include "../../include/linux/libcfs/libcfs.h"
  44
  45 /* XXX move things up to the top, comment */
  46 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  47
  48 char cfs_tracefile[TRACEFILE_NAME_SIZE];
  49 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
  50 static struct tracefiled_ctl trace_tctl;
  51 static DEFINE_MUTEX(cfs_trace_thread_mutex);
  52 static int thread_running;
  53
  54 static atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
  55
  56 struct page_collection {
  57         struct list_head        pc_pages;
  58         /*
  59          * if this flag is set, collect_pages() will spill both
  60          * ->tcd_daemon_pages and ->tcd_pages to the ->pc_pages. Otherwise,
  61          * only ->tcd_pages are spilled.
  62          */
  63         int                     pc_want_daemon_pages;
  64 };
  65
  66 struct tracefiled_ctl {
  67         struct completion       tctl_start;
  68         struct completion       tctl_stop;
  69         wait_queue_head_t       tctl_waitq;
  70         pid_t                   tctl_pid;
  71         atomic_t                tctl_shutdown;
  72 };
  73
  74 /*
  75  * small data-structure for each page owned by tracefiled.
  76  */
  77 struct cfs_trace_page {
  78         /*
  79          * page itself
  80          */
  81         struct page             *page;
  82         /*
  83          * linkage into one of the lists in trace_data_union or
  84          * page_collection
  85          */
  86         struct list_head        linkage;
  87         /*
  88          * number of bytes used within this page
  89          */
  90         unsigned int            used;
  91         /*
  92          * cpu that owns this page
  93          */
  94         unsigned short          cpu;
  95         /*
  96          * type(context) of this page
  97          */
  98         unsigned short          type;
  99 };
 100
 101 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 102                                          struct cfs_trace_cpu_data *tcd);
 103
 104 static inline struct cfs_trace_page *
 105 cfs_tage_from_list(struct list_head *list)
 106 {
 107         return list_entry(list, struct cfs_trace_page, linkage);
 108 }
 109
 110 static struct cfs_trace_page *cfs_tage_alloc(gfp_t gfp)
 111 {
 112         struct page *page;
 113         struct cfs_trace_page *tage;
 114
 115         /* My caller is trying to free memory */
 116         if (!in_interrupt() && memory_pressure_get())
 117                 return NULL;
 118
 119         /*
 120          * Don't spam console with allocation failures: they will be reported
 121          * by upper layer anyway.
 122          */
 123         gfp |= __GFP_NOWARN;
 124         page = alloc_page(gfp);
 125         if (!page)
 126                 return NULL;
 127
 128         tage = kmalloc(sizeof(*tage), gfp);
 129         if (!tage) {
 130                 __free_page(page);
 131                 return NULL;
 132         }
 133
 134         tage->page = page;
 135         atomic_inc(&cfs_tage_allocated);
 136         return tage;
 137 }
 138
 139 static void cfs_tage_free(struct cfs_trace_page *tage)
 140 {
 141         __free_page(tage->page);
 142         kfree(tage);
 143         atomic_dec(&cfs_tage_allocated);
 144 }
 145
 146 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
 147                              struct list_head *queue)
 148 {
 149         list_move_tail(&tage->linkage, queue);
 150 }
 151
 152 int cfs_trace_refill_stock(struct cfs_trace_cpu_data *tcd, gfp_t gfp,
 153                            struct list_head *stock)
 154 {
 155         int i;
 156
 157         /*
 158          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 159          * from here: this will lead to infinite recursion.
 160          */
 161
 162         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++i) {
 163                 struct cfs_trace_page *tage;
 164
 165                 tage = cfs_tage_alloc(gfp);
 166                 if (!tage)
 167                         break;
 168                 list_add_tail(&tage->linkage, stock);
 169         }
 170         return i;
 171 }
 172
 173 /* return a page that has 'len' bytes left at the end */
 174 static struct cfs_trace_page *
 175 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
 176 {
 177         struct cfs_trace_page *tage;
 178
 179         if (tcd->tcd_cur_pages > 0) {
 180                 __LASSERT(!list_empty(&tcd->tcd_pages));
 181                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
 182                 if (tage->used + len <= PAGE_SIZE)
 183                         return tage;
 184         }
 185
 186         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 187                 if (tcd->tcd_cur_stock_pages > 0) {
 188                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
 189                         --tcd->tcd_cur_stock_pages;
 190                         list_del_init(&tage->linkage);
 191                 } else {
 192                         tage = cfs_tage_alloc(GFP_ATOMIC);
 193                         if (unlikely(!tage)) {
 194                                 if ((!memory_pressure_get() ||
 195                                      in_interrupt()) && printk_ratelimit())
 196                                         pr_warn("cannot allocate a tage (%ld)\n",
 197                                                 tcd->tcd_cur_pages);
 198                                 return NULL;
 199                         }
 200                 }
 201
 202                 tage->used = 0;
 203                 tage->cpu = smp_processor_id();
 204                 tage->type = tcd->tcd_type;
 205                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 206                 tcd->tcd_cur_pages++;
 207
 208                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 209                         struct tracefiled_ctl *tctl = &trace_tctl;
 210                         /*
 211                          * wake up tracefiled to process some pages.
 212                          */
 213                         wake_up(&tctl->tctl_waitq);
 214                 }
 215                 return tage;
 216         }
 217         return NULL;
 218 }
 219
 220 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
 221 {
 222         int pgcount = tcd->tcd_cur_pages / 10;
 223         struct page_collection pc;
 224         struct cfs_trace_page *tage;
 225         struct cfs_trace_page *tmp;
 226
 227         /*
 228          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 229          * from here: this will lead to infinite recursion.
 230          */
 231
 232         if (printk_ratelimit())
 233                 pr_warn("debug daemon buffer overflowed; discarding 10%% of pages (%d of %ld)\n",
 234                         pgcount + 1, tcd->tcd_cur_pages);
 235
 236         INIT_LIST_HEAD(&pc.pc_pages);
 237
 238         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 239                 if (!pgcount--)
 240                         break;
 241
 242                 list_move_tail(&tage->linkage, &pc.pc_pages);
 243                 tcd->tcd_cur_pages--;
 244         }
 245         put_pages_on_tcd_daemon_list(&pc, tcd);
 246 }
 247
 248 /* return a page that has 'len' bytes left at the end */
 249 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 250                                                  unsigned long len)
 251 {
 252         struct cfs_trace_page *tage;
 253
 254         /*
 255          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 256          * from here: this will lead to infinite recursion.
 257          */
 258
 259         if (len > PAGE_SIZE) {
 260                 pr_err("cowardly refusing to write %lu bytes in a page\n", len);
 261                 return NULL;
 262         }
 263
 264         tage = cfs_trace_get_tage_try(tcd, len);
 265         if (tage)
 266                 return tage;
 267         if (thread_running)
 268                 cfs_tcd_shrink(tcd);
 269         if (tcd->tcd_cur_pages > 0) {
 270                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 271                 tage->used = 0;
 272                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 273         }
 274         return tage;
 275 }
 276
 277 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 278                      const char *format, ...)
 279 {
 280         va_list args;
 281         int rc;
 282
 283         va_start(args, format);
 284         rc = libcfs_debug_vmsg2(msgdata, format, args, NULL);
 285         va_end(args);
 286
 287         return rc;
 288 }
 289 EXPORT_SYMBOL(libcfs_debug_msg);
 290
 291 int libcfs_debug_vmsg2(struct libcfs_debug_msg_data *msgdata,
 292                        const char *format1, va_list args,
 293                        const char *format2, ...)
 294 {
 295         struct cfs_trace_cpu_data *tcd = NULL;
 296         struct ptldebug_header header = { 0 };
 297         struct cfs_trace_page *tage;
 298         /* string_buf is used only if tcd != NULL, and is always set then */
 299         char *string_buf = NULL;
 300         char *debug_buf;
 301         int known_size;
 302         int needed = 85; /* average message length */
 303         int max_nob;
 304         va_list ap;
 305         int depth;
 306         int i;
 307         int remain;
 308         int mask = msgdata->msg_mask;
 309         const char *file = kbasename(msgdata->msg_file);
 310         struct cfs_debug_limit_state *cdls = msgdata->msg_cdls;
 311
 312         tcd = cfs_trace_get_tcd();
 313
 314         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 315          * pins us to a particular CPU.  This avoids an smp_processor_id()
 316          * warning on Linux when debugging is enabled.
 317          */
 318         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 319
 320         if (!tcd)               /* arch may not log in IRQ context */
 321                 goto console;
 322
 323         if (!tcd->tcd_cur_pages)
 324                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 325
 326         if (tcd->tcd_shutting_down) {
 327                 cfs_trace_put_tcd(tcd);
 328                 tcd = NULL;
 329                 goto console;
 330         }
 331
 332         depth = __current_nesting_level();
 333         known_size = strlen(file) + 1 + depth;
 334         if (msgdata->msg_fn)
 335                 known_size += strlen(msgdata->msg_fn) + 1;
 336
 337         if (libcfs_debug_binary)
 338                 known_size += sizeof(header);
 339
 340         /*
 341          * '2' used because vsnprintf return real size required for output
 342          * _without_ terminating NULL.
 343          * if needed is to small for this format.
 344          */
 345         for (i = 0; i < 2; i++) {
 346                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 347                 if (!tage) {
 348                         if (needed + known_size > PAGE_SIZE)
 349                                 mask |= D_ERROR;
 350
 351                         cfs_trace_put_tcd(tcd);
 352                         tcd = NULL;
 353                         goto console;
 354                 }
 355
 356                 string_buf = (char *)page_address(tage->page) +
 357                                         tage->used + known_size;
 358
 359                 max_nob = PAGE_SIZE - tage->used - known_size;
 360                 if (max_nob <= 0) {
 361                         pr_emerg("negative max_nob: %d\n", max_nob);
 362                         mask |= D_ERROR;
 363                         cfs_trace_put_tcd(tcd);
 364                         tcd = NULL;
 365                         goto console;
 366                 }
 367
 368                 needed = 0;
 369                 if (format1) {
 370                         va_copy(ap, args);
 371                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 372                         va_end(ap);
 373                 }
 374
 375                 if (format2) {
 376                         remain = max_nob - needed;
 377                         if (remain < 0)
 378                                 remain = 0;
 379
 380                         va_start(ap, format2);
 381                         needed += vsnprintf(string_buf + needed, remain,
 382                                             format2, ap);
 383                         va_end(ap);
 384                 }
 385
 386                 if (needed < max_nob) /* well. printing ok.. */
 387                         break;
 388         }
 389
 390         if (*(string_buf + needed - 1) != '\n')
 391                 pr_info("format at %s:%d:%s doesn't end in newline\n", file,
 392                         msgdata->msg_line, msgdata->msg_fn);
 393
 394         header.ph_len = known_size + needed;
 395         debug_buf = (char *)page_address(tage->page) + tage->used;
 396
 397         if (libcfs_debug_binary) {
 398                 memcpy(debug_buf, &header, sizeof(header));
 399                 tage->used += sizeof(header);
 400                 debug_buf += sizeof(header);
 401         }
 402
 403         /* indent message according to the nesting level */
 404         while (depth-- > 0) {
 405                 *(debug_buf++) = '.';
 406                 ++tage->used;
 407         }
 408
 409         strcpy(debug_buf, file);
 410         tage->used += strlen(file) + 1;
 411         debug_buf += strlen(file) + 1;
 412
 413         if (msgdata->msg_fn) {
 414                 strcpy(debug_buf, msgdata->msg_fn);
 415                 tage->used += strlen(msgdata->msg_fn) + 1;
 416                 debug_buf += strlen(msgdata->msg_fn) + 1;
 417         }
 418
 419         __LASSERT(debug_buf == string_buf);
 420
 421         tage->used += needed;
 422         __LASSERT(tage->used <= PAGE_SIZE);
 423
 424 console:
 425         if (!(mask & libcfs_printk)) {
 426                 /* no console output requested */
 427                 if (tcd)
 428                         cfs_trace_put_tcd(tcd);
 429                 return 1;
 430         }
 431
 432         if (cdls) {
 433                 if (libcfs_console_ratelimit &&
 434                     cdls->cdls_next &&          /* not first time ever */
 435                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 436                         /* skipping a console message */
 437                         cdls->cdls_count++;
 438                         if (tcd)
 439                                 cfs_trace_put_tcd(tcd);
 440                         return 1;
 441                 }
 442
 443                 if (cfs_time_after(cfs_time_current(),
 444                                    cdls->cdls_next + libcfs_console_max_delay +
 445                                    cfs_time_seconds(10))) {
 446                         /* last timeout was a long time ago */
 447                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 448                 } else {
 449                         cdls->cdls_delay *= libcfs_console_backoff;
 450                 }
 451
 452                 if (cdls->cdls_delay < libcfs_console_min_delay)
 453                         cdls->cdls_delay = libcfs_console_min_delay;
 454                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 455                         cdls->cdls_delay = libcfs_console_max_delay;
 456
 457                 /* ensure cdls_next is never zero after it's been seen */
 458                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 459         }
 460
 461         if (tcd) {
 462                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 463                                      msgdata->msg_fn);
 464                 cfs_trace_put_tcd(tcd);
 465         } else {
 466                 string_buf = cfs_trace_get_console_buffer();
 467
 468                 needed = 0;
 469                 if (format1) {
 470                         va_copy(ap, args);
 471                         needed = vsnprintf(string_buf,
 472                                            CFS_TRACE_CONSOLE_BUFFER_SIZE,
 473                                            format1, ap);
 474                         va_end(ap);
 475                 }
 476                 if (format2) {
 477                         remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
 478                         if (remain > 0) {
 479                                 va_start(ap, format2);
 480                                 needed += vsnprintf(string_buf + needed, remain,
 481                                                     format2, ap);
 482                                 va_end(ap);
 483                         }
 484                 }
 485                 cfs_print_to_console(&header, mask,
 486                                      string_buf, needed, file, msgdata->msg_fn);
 487
 488                 put_cpu();
 489         }
 490
 491         if (cdls && cdls->cdls_count) {
 492                 string_buf = cfs_trace_get_console_buffer();
 493
 494                 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 495                                   "Skipped %d previous similar message%s\n",
 496                                   cdls->cdls_count,
 497                                   (cdls->cdls_count > 1) ? "s" : "");
 498
 499                 cfs_print_to_console(&header, mask,
 500                                      string_buf, needed, file, msgdata->msg_fn);
 501
 502                 put_cpu();
 503                 cdls->cdls_count = 0;
 504         }
 505
 506         return 0;
 507 }
 508 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 509
 510 void
 511 cfs_trace_assertion_failed(const char *str,
 512                            struct libcfs_debug_msg_data *msgdata)
 513 {
 514         struct ptldebug_header hdr;
 515
 516         libcfs_panic_in_progress = 1;
 517         libcfs_catastrophe = 1;
 518         mb();
 519
 520         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 521
 522         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 523                              msgdata->msg_file, msgdata->msg_fn);
 524
 525         panic("Lustre debug assertion failure\n");
 526
 527         /* not reached */
 528 }
 529
 530 static void
 531 panic_collect_pages(struct page_collection *pc)
 532 {
 533         /* Do the collect_pages job on a single CPU: assumes that all other
 534          * CPUs have been stopped during a panic.  If this isn't true for some
 535          * arch, this will have to be implemented separately in each arch.
 536          */
 537         struct cfs_trace_cpu_data *tcd;
 538         int i;
 539         int j;
 540
 541         INIT_LIST_HEAD(&pc->pc_pages);
 542
 543         cfs_tcd_for_each(tcd, i, j) {
 544                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 545                 tcd->tcd_cur_pages = 0;
 546
 547                 if (pc->pc_want_daemon_pages) {
 548                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 549                         tcd->tcd_cur_daemon_pages = 0;
 550                 }
 551         }
 552 }
 553
 554 static void collect_pages_on_all_cpus(struct page_collection *pc)
 555 {
 556         struct cfs_trace_cpu_data *tcd;
 557         int i, cpu;
 558
 559         for_each_possible_cpu(cpu) {
 560                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 561                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 562                         tcd->tcd_cur_pages = 0;
 563                         if (pc->pc_want_daemon_pages) {
 564                                 list_splice_init(&tcd->tcd_daemon_pages,
 565                                                  &pc->pc_pages);
 566                                 tcd->tcd_cur_daemon_pages = 0;
 567                         }
 568                 }
 569         }
 570 }
 571
 572 static void collect_pages(struct page_collection *pc)
 573 {
 574         INIT_LIST_HEAD(&pc->pc_pages);
 575
 576         if (libcfs_panic_in_progress)
 577                 panic_collect_pages(pc);
 578         else
 579                 collect_pages_on_all_cpus(pc);
 580 }
 581
 582 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 583 {
 584         struct cfs_trace_cpu_data *tcd;
 585         struct list_head *cur_head;
 586         struct cfs_trace_page *tage;
 587         struct cfs_trace_page *tmp;
 588         int i, cpu;
 589
 590         for_each_possible_cpu(cpu) {
 591                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 592                         cur_head = tcd->tcd_pages.next;
 593
 594                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 595                                                  linkage) {
 596                                 __LASSERT_TAGE_INVARIANT(tage);
 597
 598                                 if (tage->cpu != cpu || tage->type != i)
 599                                         continue;
 600
 601                                 cfs_tage_to_tail(tage, cur_head);
 602                                 tcd->tcd_cur_pages++;
 603                         }
 604                 }
 605         }
 606 }
 607
 608 static void put_pages_back(struct page_collection *pc)
 609 {
 610         if (!libcfs_panic_in_progress)
 611                 put_pages_back_on_all_cpus(pc);
 612 }
 613
 614 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 615  * we have a good amount of data at all times for dumping during an LBUG, even
 616  * if we have been steadily writing (and otherwise discarding) pages via the
 617  * debug daemon.
 618  */
 619 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 620                                          struct cfs_trace_cpu_data *tcd)
 621 {
 622         struct cfs_trace_page *tage;
 623         struct cfs_trace_page *tmp;
 624
 625         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 626                 __LASSERT_TAGE_INVARIANT(tage);
 627
 628                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 629                         continue;
 630
 631                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 632                 tcd->tcd_cur_daemon_pages++;
 633
 634                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 635                         struct cfs_trace_page *victim;
 636
 637                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 638                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 639
 640                         __LASSERT_TAGE_INVARIANT(victim);
 641
 642                         list_del(&victim->linkage);
 643                         cfs_tage_free(victim);
 644                         tcd->tcd_cur_daemon_pages--;
 645                 }
 646         }
 647 }
 648
 649 static void put_pages_on_daemon_list(struct page_collection *pc)
 650 {
 651         struct cfs_trace_cpu_data *tcd;
 652         int i, cpu;
 653
 654         for_each_possible_cpu(cpu) {
 655                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 656                         put_pages_on_tcd_daemon_list(pc, tcd);
 657         }
 658 }
 659
 660 void cfs_trace_debug_print(void)
 661 {
 662         struct page_collection pc;
 663         struct cfs_trace_page *tage;
 664         struct cfs_trace_page *tmp;
 665
 666         pc.pc_want_daemon_pages = 1;
 667         collect_pages(&pc);
 668         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 669                 char *p, *file, *fn;
 670                 struct page *page;
 671
 672                 __LASSERT_TAGE_INVARIANT(tage);
 673
 674                 page = tage->page;
 675                 p = page_address(page);
 676                 while (p < ((char *)page_address(page) + tage->used)) {
 677                         struct ptldebug_header *hdr;
 678                         int len;
 679
 680                         hdr = (void *)p;
 681                         p += sizeof(*hdr);
 682                         file = p;
 683                         p += strlen(file) + 1;
 684                         fn = p;
 685                         p += strlen(fn) + 1;
 686                         len = hdr->ph_len - (int)(p - (char *)hdr);
 687
 688                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 689
 690                         p += len;
 691                 }
 692
 693                 list_del(&tage->linkage);
 694                 cfs_tage_free(tage);
 695         }
 696 }
 697
 698 int cfs_tracefile_dump_all_pages(char *filename)
 699 {
 700         struct page_collection pc;
 701         struct file *filp;
 702         struct cfs_trace_page *tage;
 703         struct cfs_trace_page *tmp;
 704         char *buf;
 705         mm_segment_t __oldfs;
 706         int rc;
 707
 708         cfs_tracefile_write_lock();
 709
 710         filp = filp_open(filename, O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE,
 711                          0600);
 712         if (IS_ERR(filp)) {
 713                 rc = PTR_ERR(filp);
 714                 filp = NULL;
 715                 pr_err("LustreError: can't open %s for dump: rc %d\n",
 716                        filename, rc);
 717                 goto out;
 718         }
 719
 720         pc.pc_want_daemon_pages = 1;
 721         collect_pages(&pc);
 722         if (list_empty(&pc.pc_pages)) {
 723                 rc = 0;
 724                 goto close;
 725         }
 726         __oldfs = get_fs();
 727         set_fs(get_ds());
 728
 729         /* ok, for now, just write the pages.  in the future we'll be building
 730          * iobufs with the pages and calling generic_direct_IO
 731          */
 732         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 733                 __LASSERT_TAGE_INVARIANT(tage);
 734
 735                 buf = kmap(tage->page);
 736                 rc = vfs_write(filp, (__force const char __user *)buf,
 737                                tage->used, &filp->f_pos);
 738                 kunmap(tage->page);
 739
 740                 if (rc != (int)tage->used) {
 741                         pr_warn("wanted to write %u but wrote %d\n", tage->used,
 742                                 rc);
 743                         put_pages_back(&pc);
 744                         __LASSERT(list_empty(&pc.pc_pages));
 745                         break;
 746                 }
 747                 list_del(&tage->linkage);
 748                 cfs_tage_free(tage);
 749         }
 750         set_fs(__oldfs);
 751         rc = vfs_fsync(filp, 1);
 752         if (rc)
 753                 pr_err("sync returns %d\n", rc);
 754 close:
 755         filp_close(filp, NULL);
 756 out:
 757         cfs_tracefile_write_unlock();
 758         return rc;
 759 }
 760
 761 void cfs_trace_flush_pages(void)
 762 {
 763         struct page_collection pc;
 764         struct cfs_trace_page *tage;
 765         struct cfs_trace_page *tmp;
 766
 767         pc.pc_want_daemon_pages = 1;
 768         collect_pages(&pc);
 769         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 770                 __LASSERT_TAGE_INVARIANT(tage);
 771
 772                 list_del(&tage->linkage);
 773                 cfs_tage_free(tage);
 774         }
 775 }
 776
 777 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 778                             const char __user *usr_buffer, int usr_buffer_nob)
 779 {
 780         int nob;
 781
 782         if (usr_buffer_nob > knl_buffer_nob)
 783                 return -EOVERFLOW;
 784
 785         if (copy_from_user((void *)knl_buffer,
 786                            usr_buffer, usr_buffer_nob))
 787                 return -EFAULT;
 788
 789         nob = strnlen(knl_buffer, usr_buffer_nob);
 790         while (nob-- >= 0)                    /* strip trailing whitespace */
 791                 if (!isspace(knl_buffer[nob]))
 792                         break;
 793
 794         if (nob < 0)                        /* empty string */
 795                 return -EINVAL;
 796
 797         if (nob == knl_buffer_nob)            /* no space to terminate */
 798                 return -EOVERFLOW;
 799
 800         knl_buffer[nob + 1] = 0;                /* terminate */
 801         return 0;
 802 }
 803 EXPORT_SYMBOL(cfs_trace_copyin_string);
 804
 805 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 806                              const char *knl_buffer, char *append)
 807 {
 808         /*
 809          * NB if 'append' != NULL, it's a single character to append to the
 810          * copied out string - usually "\n" or "" (i.e. a terminating zero byte)
 811          */
 812         int nob = strlen(knl_buffer);
 813
 814         if (nob > usr_buffer_nob)
 815                 nob = usr_buffer_nob;
 816
 817         if (copy_to_user(usr_buffer, knl_buffer, nob))
 818                 return -EFAULT;
 819
 820         if (append && nob < usr_buffer_nob) {
 821                 if (copy_to_user(usr_buffer + nob, append, 1))
 822                         return -EFAULT;
 823
 824                 nob++;
 825         }
 826
 827         return nob;
 828 }
 829 EXPORT_SYMBOL(cfs_trace_copyout_string);
 830
 831 int cfs_trace_allocate_string_buffer(char **str, int nob)
 832 {
 833         if (nob > 2 * PAGE_SIZE)            /* string must be "sensible" */
 834                 return -EINVAL;
 835
 836         *str = kmalloc(nob, GFP_KERNEL | __GFP_ZERO);
 837         if (!*str)
 838                 return -ENOMEM;
 839
 840         return 0;
 841 }
 842
 843 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
 844 {
 845         char *str;
 846         int rc;
 847
 848         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 849         if (rc)
 850                 return rc;
 851
 852         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 853                                      usr_str, usr_str_nob);
 854         if (rc)
 855                 goto out;
 856
 857         if (str[0] != '/') {
 858                 rc = -EINVAL;
 859                 goto out;
 860         }
 861         rc = cfs_tracefile_dump_all_pages(str);
 862 out:
 863         kfree(str);
 864         return rc;
 865 }
 866
 867 int cfs_trace_daemon_command(char *str)
 868 {
 869         int rc = 0;
 870
 871         cfs_tracefile_write_lock();
 872
 873         if (!strcmp(str, "stop")) {
 874                 cfs_tracefile_write_unlock();
 875                 cfs_trace_stop_thread();
 876                 cfs_tracefile_write_lock();
 877                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 878
 879         } else if (!strncmp(str, "size=", 5)) {
 880                 unsigned long tmp;
 881
 882                 rc = kstrtoul(str + 5, 10, &tmp);
 883                 if (!rc) {
 884                         if (tmp < 10 || tmp > 20480)
 885                                 cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 886                         else
 887                                 cfs_tracefile_size = tmp << 20;
 888                 }
 889         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 890                 rc = -ENAMETOOLONG;
 891         } else if (str[0] != '/') {
 892                 rc = -EINVAL;
 893         } else {
 894                 strcpy(cfs_tracefile, str);
 895
 896                 pr_info("debug daemon will attempt to start writing to %s (%lukB max)\n",
 897                         cfs_tracefile,
 898                         (long)(cfs_tracefile_size >> 10));
 899
 900                 cfs_trace_start_thread();
 901         }
 902
 903         cfs_tracefile_write_unlock();
 904         return rc;
 905 }
 906
 907 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
 908 {
 909         char *str;
 910         int rc;
 911
 912         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 913         if (rc)
 914                 return rc;
 915
 916         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 917                                      usr_str, usr_str_nob);
 918         if (!rc)
 919                 rc = cfs_trace_daemon_command(str);
 920
 921         kfree(str);
 922         return rc;
 923 }
 924
 925 int cfs_trace_set_debug_mb(int mb)
 926 {
 927         int i;
 928         int j;
 929         int pages;
 930         int limit = cfs_trace_max_debug_mb();
 931         struct cfs_trace_cpu_data *tcd;
 932
 933         if (mb < num_possible_cpus()) {
 934                 pr_warn("%d MB is too small for debug buffer size, setting it to %d MB.\n",
 935                         mb, num_possible_cpus());
 936                 mb = num_possible_cpus();
 937         }
 938
 939         if (mb > limit) {
 940                 pr_warn("%d MB is too large for debug buffer size, setting it to %d MB.\n",
 941                         mb, limit);
 942                 mb = limit;
 943         }
 944
 945         mb /= num_possible_cpus();
 946         pages = mb << (20 - 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_get_debug_mb(void)
 959 {
 960         int i;
 961         int j;
 962         struct cfs_trace_cpu_data *tcd;
 963         int total_pages = 0;
 964
 965         cfs_tracefile_read_lock();
 966
 967         cfs_tcd_for_each(tcd, i, j)
 968                 total_pages += tcd->tcd_max_pages;
 969
 970         cfs_tracefile_read_unlock();
 971
 972         return (total_pages >> (20 - PAGE_SHIFT)) + 1;
 973 }
 974
 975 static int tracefiled(void *arg)
 976 {
 977         struct page_collection pc;
 978         struct tracefiled_ctl *tctl = arg;
 979         struct cfs_trace_page *tage;
 980         struct cfs_trace_page *tmp;
 981         mm_segment_t __oldfs;
 982         struct file *filp;
 983         char *buf;
 984         int last_loop = 0;
 985         int rc;
 986
 987         /* we're started late enough that we pick up init's fs context */
 988         /* this is so broken in uml?  what on earth is going on? */
 989
 990         complete(&tctl->tctl_start);
 991
 992         while (1) {
 993                 wait_queue_entry_t __wait;
 994
 995                 pc.pc_want_daemon_pages = 0;
 996                 collect_pages(&pc);
 997                 if (list_empty(&pc.pc_pages))
 998                         goto end_loop;
 999
1000                 filp = NULL;
1001                 cfs_tracefile_read_lock();
1002                 if (cfs_tracefile[0]) {
1003                         filp = filp_open(cfs_tracefile,
1004                                          O_CREAT | O_RDWR | O_LARGEFILE,
1005                                          0600);
1006                         if (IS_ERR(filp)) {
1007                                 rc = PTR_ERR(filp);
1008                                 filp = NULL;
1009                                 pr_warn("couldn't open %s: %d\n", cfs_tracefile,
1010                                         rc);
1011                         }
1012                 }
1013                 cfs_tracefile_read_unlock();
1014                 if (!filp) {
1015                         put_pages_on_daemon_list(&pc);
1016                         __LASSERT(list_empty(&pc.pc_pages));
1017                         goto end_loop;
1018                 }
1019                 __oldfs = get_fs();
1020                 set_fs(get_ds());
1021
1022                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
1023                         static loff_t f_pos;
1024
1025                         __LASSERT_TAGE_INVARIANT(tage);
1026
1027                         if (f_pos >= (off_t)cfs_tracefile_size)
1028                                 f_pos = 0;
1029                         else if (f_pos > i_size_read(file_inode(filp)))
1030                                 f_pos = i_size_read(file_inode(filp));
1031
1032                         buf = kmap(tage->page);
1033                         rc = vfs_write(filp, (__force const char __user *)buf,
1034                                        tage->used, &f_pos);
1035                         kunmap(tage->page);
1036
1037                         if (rc != (int)tage->used) {
1038                                 pr_warn("wanted to write %u but wrote %d\n",
1039                                         tage->used, rc);
1040                                 put_pages_back(&pc);
1041                                 __LASSERT(list_empty(&pc.pc_pages));
1042                                 break;
1043                         }
1044                 }
1045                 set_fs(__oldfs);
1046
1047                 filp_close(filp, NULL);
1048                 put_pages_on_daemon_list(&pc);
1049                 if (!list_empty(&pc.pc_pages)) {
1050                         int i;
1051
1052                         pr_alert("trace pages aren't empty\n");
1053                         pr_err("total cpus(%d): ", num_possible_cpus());
1054                         for (i = 0; i < num_possible_cpus(); i++)
1055                                 if (cpu_online(i))
1056                                         pr_cont("%d(on) ", i);
1057                                 else
1058                                         pr_cont("%d(off) ", i);
1059                         pr_cont("\n");
1060
1061                         i = 0;
1062                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1063                                                  linkage)
1064                                 pr_err("page %d belongs to cpu %d\n",
1065                                        ++i, tage->cpu);
1066                         pr_err("There are %d pages unwritten\n", i);
1067                 }
1068                 __LASSERT(list_empty(&pc.pc_pages));
1069 end_loop:
1070                 if (atomic_read(&tctl->tctl_shutdown)) {
1071                         if (!last_loop) {
1072                                 last_loop = 1;
1073                                 continue;
1074                         } else {
1075                                 break;
1076                         }
1077                 }
1078                 init_waitqueue_entry(&__wait, current);
1079                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1080                 set_current_state(TASK_INTERRUPTIBLE);
1081                 schedule_timeout(cfs_time_seconds(1));
1082                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1083         }
1084         complete(&tctl->tctl_stop);
1085         return 0;
1086 }
1087
1088 int cfs_trace_start_thread(void)
1089 {
1090         struct tracefiled_ctl *tctl = &trace_tctl;
1091         struct task_struct *task;
1092         int rc = 0;
1093
1094         mutex_lock(&cfs_trace_thread_mutex);
1095         if (thread_running)
1096                 goto out;
1097
1098         init_completion(&tctl->tctl_start);
1099         init_completion(&tctl->tctl_stop);
1100         init_waitqueue_head(&tctl->tctl_waitq);
1101         atomic_set(&tctl->tctl_shutdown, 0);
1102
1103         task = kthread_run(tracefiled, tctl, "ktracefiled");
1104         if (IS_ERR(task)) {
1105                 rc = PTR_ERR(task);
1106                 goto out;
1107         }
1108
1109         wait_for_completion(&tctl->tctl_start);
1110         thread_running = 1;
1111 out:
1112         mutex_unlock(&cfs_trace_thread_mutex);
1113         return rc;
1114 }
1115
1116 void cfs_trace_stop_thread(void)
1117 {
1118         struct tracefiled_ctl *tctl = &trace_tctl;
1119
1120         mutex_lock(&cfs_trace_thread_mutex);
1121         if (thread_running) {
1122                 pr_info("shutting down debug daemon thread...\n");
1123                 atomic_set(&tctl->tctl_shutdown, 1);
1124                 wait_for_completion(&tctl->tctl_stop);
1125                 thread_running = 0;
1126         }
1127         mutex_unlock(&cfs_trace_thread_mutex);
1128 }
1129
1130 int cfs_tracefile_init(int max_pages)
1131 {
1132         struct cfs_trace_cpu_data *tcd;
1133         int i;
1134         int j;
1135         int rc;
1136         int factor;
1137
1138         rc = cfs_tracefile_init_arch();
1139         if (rc)
1140                 return rc;
1141
1142         cfs_tcd_for_each(tcd, i, j) {
1143                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1144                 factor = tcd->tcd_pages_factor;
1145                 INIT_LIST_HEAD(&tcd->tcd_pages);
1146                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1147                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1148                 tcd->tcd_cur_pages = 0;
1149                 tcd->tcd_cur_stock_pages = 0;
1150                 tcd->tcd_cur_daemon_pages = 0;
1151                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1152                 LASSERT(tcd->tcd_max_pages > 0);
1153                 tcd->tcd_shutting_down = 0;
1154         }
1155
1156         return 0;
1157 }
1158
1159 static void trace_cleanup_on_all_cpus(void)
1160 {
1161         struct cfs_trace_cpu_data *tcd;
1162         struct cfs_trace_page *tage;
1163         struct cfs_trace_page *tmp;
1164         int i, cpu;
1165
1166         for_each_possible_cpu(cpu) {
1167                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1168                         tcd->tcd_shutting_down = 1;
1169
1170                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages,
1171                                                  linkage) {
1172                                 __LASSERT_TAGE_INVARIANT(tage);
1173
1174                                 list_del(&tage->linkage);
1175                                 cfs_tage_free(tage);
1176                         }
1177
1178                         tcd->tcd_cur_pages = 0;
1179                 }
1180         }
1181 }
1182
1183 static void cfs_trace_cleanup(void)
1184 {
1185         struct page_collection pc;
1186
1187         INIT_LIST_HEAD(&pc.pc_pages);
1188
1189         trace_cleanup_on_all_cpus();
1190
1191         cfs_tracefile_fini_arch();
1192 }
1193
1194 void cfs_tracefile_exit(void)
1195 {
1196         cfs_trace_stop_thread();
1197         cfs_trace_cleanup();
1198 }