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 <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() || in_interrupt())
 195                                         pr_warn_ratelimited("cannot allocate a tage (%ld)\n",
 196                                                             tcd->tcd_cur_pages);
 197                                 return NULL;
 198                         }
 199                 }
 200
 201                 tage->used = 0;
 202                 tage->cpu = smp_processor_id();
 203                 tage->type = tcd->tcd_type;
 204                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 205                 tcd->tcd_cur_pages++;
 206
 207                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 208                         struct tracefiled_ctl *tctl = &trace_tctl;
 209                         /*
 210                          * wake up tracefiled to process some pages.
 211                          */
 212                         wake_up(&tctl->tctl_waitq);
 213                 }
 214                 return tage;
 215         }
 216         return NULL;
 217 }
 218
 219 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
 220 {
 221         int pgcount = tcd->tcd_cur_pages / 10;
 222         struct page_collection pc;
 223         struct cfs_trace_page *tage;
 224         struct cfs_trace_page *tmp;
 225
 226         /*
 227          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 228          * from here: this will lead to infinite recursion.
 229          */
 230
 231         pr_warn_ratelimited("debug daemon buffer overflowed; discarding 10%% of pages (%d of %ld)\n",
 232                             pgcount + 1, tcd->tcd_cur_pages);
 233
 234         INIT_LIST_HEAD(&pc.pc_pages);
 235
 236         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 237                 if (!pgcount--)
 238                         break;
 239
 240                 list_move_tail(&tage->linkage, &pc.pc_pages);
 241                 tcd->tcd_cur_pages--;
 242         }
 243         put_pages_on_tcd_daemon_list(&pc, tcd);
 244 }
 245
 246 /* return a page that has 'len' bytes left at the end */
 247 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 248                                                  unsigned long len)
 249 {
 250         struct cfs_trace_page *tage;
 251
 252         /*
 253          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 254          * from here: this will lead to infinite recursion.
 255          */
 256
 257         if (len > PAGE_SIZE) {
 258                 pr_err("cowardly refusing to write %lu bytes in a page\n", len);
 259                 return NULL;
 260         }
 261
 262         tage = cfs_trace_get_tage_try(tcd, len);
 263         if (tage)
 264                 return tage;
 265         if (thread_running)
 266                 cfs_tcd_shrink(tcd);
 267         if (tcd->tcd_cur_pages > 0) {
 268                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 269                 tage->used = 0;
 270                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 271         }
 272         return tage;
 273 }
 274
 275 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 276                      const char *format, ...)
 277 {
 278         va_list args;
 279         int rc;
 280
 281         va_start(args, format);
 282         rc = libcfs_debug_vmsg2(msgdata, format, args, NULL);
 283         va_end(args);
 284
 285         return rc;
 286 }
 287 EXPORT_SYMBOL(libcfs_debug_msg);
 288
 289 int libcfs_debug_vmsg2(struct libcfs_debug_msg_data *msgdata,
 290                        const char *format1, va_list args,
 291                        const char *format2, ...)
 292 {
 293         struct cfs_trace_cpu_data *tcd = NULL;
 294         struct ptldebug_header header = { 0 };
 295         struct cfs_trace_page *tage;
 296         /* string_buf is used only if tcd != NULL, and is always set then */
 297         char *string_buf = NULL;
 298         char *debug_buf;
 299         int known_size;
 300         int needed = 85; /* average message length */
 301         int max_nob;
 302         va_list ap;
 303         int depth;
 304         int i;
 305         int remain;
 306         int mask = msgdata->msg_mask;
 307         const char *file = kbasename(msgdata->msg_file);
 308         struct cfs_debug_limit_state *cdls = msgdata->msg_cdls;
 309
 310         tcd = cfs_trace_get_tcd();
 311
 312         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 313          * pins us to a particular CPU.  This avoids an smp_processor_id()
 314          * warning on Linux when debugging is enabled.
 315          */
 316         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 317
 318         if (!tcd)               /* arch may not log in IRQ context */
 319                 goto console;
 320
 321         if (!tcd->tcd_cur_pages)
 322                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 323
 324         if (tcd->tcd_shutting_down) {
 325                 cfs_trace_put_tcd(tcd);
 326                 tcd = NULL;
 327                 goto console;
 328         }
 329
 330         depth = __current_nesting_level();
 331         known_size = strlen(file) + 1 + depth;
 332         if (msgdata->msg_fn)
 333                 known_size += strlen(msgdata->msg_fn) + 1;
 334
 335         if (libcfs_debug_binary)
 336                 known_size += sizeof(header);
 337
 338         /*
 339          * '2' used because vsnprintf return real size required for output
 340          * _without_ terminating NULL.
 341          * if needed is to small for this format.
 342          */
 343         for (i = 0; i < 2; i++) {
 344                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 345                 if (!tage) {
 346                         if (needed + known_size > PAGE_SIZE)
 347                                 mask |= D_ERROR;
 348
 349                         cfs_trace_put_tcd(tcd);
 350                         tcd = NULL;
 351                         goto console;
 352                 }
 353
 354                 string_buf = (char *)page_address(tage->page) +
 355                                         tage->used + known_size;
 356
 357                 max_nob = PAGE_SIZE - tage->used - known_size;
 358                 if (max_nob <= 0) {
 359                         pr_emerg("negative max_nob: %d\n", max_nob);
 360                         mask |= D_ERROR;
 361                         cfs_trace_put_tcd(tcd);
 362                         tcd = NULL;
 363                         goto console;
 364                 }
 365
 366                 needed = 0;
 367                 if (format1) {
 368                         va_copy(ap, args);
 369                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 370                         va_end(ap);
 371                 }
 372
 373                 if (format2) {
 374                         remain = max_nob - needed;
 375                         if (remain < 0)
 376                                 remain = 0;
 377
 378                         va_start(ap, format2);
 379                         needed += vsnprintf(string_buf + needed, remain,
 380                                             format2, ap);
 381                         va_end(ap);
 382                 }
 383
 384                 if (needed < max_nob) /* well. printing ok.. */
 385                         break;
 386         }
 387
 388         if (*(string_buf + needed - 1) != '\n')
 389                 pr_info("format at %s:%d:%s doesn't end in newline\n", file,
 390                         msgdata->msg_line, msgdata->msg_fn);
 391
 392         header.ph_len = known_size + needed;
 393         debug_buf = (char *)page_address(tage->page) + tage->used;
 394
 395         if (libcfs_debug_binary) {
 396                 memcpy(debug_buf, &header, sizeof(header));
 397                 tage->used += sizeof(header);
 398                 debug_buf += sizeof(header);
 399         }
 400
 401         /* indent message according to the nesting level */
 402         while (depth-- > 0) {
 403                 *(debug_buf++) = '.';
 404                 ++tage->used;
 405         }
 406
 407         strcpy(debug_buf, file);
 408         tage->used += strlen(file) + 1;
 409         debug_buf += strlen(file) + 1;
 410
 411         if (msgdata->msg_fn) {
 412                 strcpy(debug_buf, msgdata->msg_fn);
 413                 tage->used += strlen(msgdata->msg_fn) + 1;
 414                 debug_buf += strlen(msgdata->msg_fn) + 1;
 415         }
 416
 417         __LASSERT(debug_buf == string_buf);
 418
 419         tage->used += needed;
 420         __LASSERT(tage->used <= PAGE_SIZE);
 421
 422 console:
 423         if (!(mask & libcfs_printk)) {
 424                 /* no console output requested */
 425                 if (tcd)
 426                         cfs_trace_put_tcd(tcd);
 427                 return 1;
 428         }
 429
 430         if (cdls) {
 431                 if (libcfs_console_ratelimit &&
 432                     cdls->cdls_next &&          /* not first time ever */
 433                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 434                         /* skipping a console message */
 435                         cdls->cdls_count++;
 436                         if (tcd)
 437                                 cfs_trace_put_tcd(tcd);
 438                         return 1;
 439                 }
 440
 441                 if (cfs_time_after(cfs_time_current(),
 442                                    cdls->cdls_next + libcfs_console_max_delay +
 443                                    cfs_time_seconds(10))) {
 444                         /* last timeout was a long time ago */
 445                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 446                 } else {
 447                         cdls->cdls_delay *= libcfs_console_backoff;
 448                 }
 449
 450                 if (cdls->cdls_delay < libcfs_console_min_delay)
 451                         cdls->cdls_delay = libcfs_console_min_delay;
 452                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 453                         cdls->cdls_delay = libcfs_console_max_delay;
 454
 455                 /* ensure cdls_next is never zero after it's been seen */
 456                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 457         }
 458
 459         if (tcd) {
 460                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 461                                      msgdata->msg_fn);
 462                 cfs_trace_put_tcd(tcd);
 463         } else {
 464                 string_buf = cfs_trace_get_console_buffer();
 465
 466                 needed = 0;
 467                 if (format1) {
 468                         va_copy(ap, args);
 469                         needed = vsnprintf(string_buf,
 470                                            CFS_TRACE_CONSOLE_BUFFER_SIZE,
 471                                            format1, ap);
 472                         va_end(ap);
 473                 }
 474                 if (format2) {
 475                         remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
 476                         if (remain > 0) {
 477                                 va_start(ap, format2);
 478                                 needed += vsnprintf(string_buf + needed, remain,
 479                                                     format2, ap);
 480                                 va_end(ap);
 481                         }
 482                 }
 483                 cfs_print_to_console(&header, mask,
 484                                      string_buf, needed, file, msgdata->msg_fn);
 485
 486                 put_cpu();
 487         }
 488
 489         if (cdls && cdls->cdls_count) {
 490                 string_buf = cfs_trace_get_console_buffer();
 491
 492                 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 493                                   "Skipped %d previous similar message%s\n",
 494                                   cdls->cdls_count,
 495                                   (cdls->cdls_count > 1) ? "s" : "");
 496
 497                 cfs_print_to_console(&header, mask,
 498                                      string_buf, needed, file, msgdata->msg_fn);
 499
 500                 put_cpu();
 501                 cdls->cdls_count = 0;
 502         }
 503
 504         return 0;
 505 }
 506 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 507
 508 void
 509 cfs_trace_assertion_failed(const char *str,
 510                            struct libcfs_debug_msg_data *msgdata)
 511 {
 512         struct ptldebug_header hdr;
 513
 514         libcfs_panic_in_progress = 1;
 515         libcfs_catastrophe = 1;
 516         mb();
 517
 518         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 519
 520         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 521                              msgdata->msg_file, msgdata->msg_fn);
 522
 523         panic("Lustre debug assertion failure\n");
 524
 525         /* not reached */
 526 }
 527
 528 static void
 529 panic_collect_pages(struct page_collection *pc)
 530 {
 531         /* Do the collect_pages job on a single CPU: assumes that all other
 532          * CPUs have been stopped during a panic.  If this isn't true for some
 533          * arch, this will have to be implemented separately in each arch.
 534          */
 535         struct cfs_trace_cpu_data *tcd;
 536         int i;
 537         int j;
 538
 539         INIT_LIST_HEAD(&pc->pc_pages);
 540
 541         cfs_tcd_for_each(tcd, i, j) {
 542                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 543                 tcd->tcd_cur_pages = 0;
 544
 545                 if (pc->pc_want_daemon_pages) {
 546                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 547                         tcd->tcd_cur_daemon_pages = 0;
 548                 }
 549         }
 550 }
 551
 552 static void collect_pages_on_all_cpus(struct page_collection *pc)
 553 {
 554         struct cfs_trace_cpu_data *tcd;
 555         int i, cpu;
 556
 557         for_each_possible_cpu(cpu) {
 558                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 559                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 560                         tcd->tcd_cur_pages = 0;
 561                         if (pc->pc_want_daemon_pages) {
 562                                 list_splice_init(&tcd->tcd_daemon_pages,
 563                                                  &pc->pc_pages);
 564                                 tcd->tcd_cur_daemon_pages = 0;
 565                         }
 566                 }
 567         }
 568 }
 569
 570 static void collect_pages(struct page_collection *pc)
 571 {
 572         INIT_LIST_HEAD(&pc->pc_pages);
 573
 574         if (libcfs_panic_in_progress)
 575                 panic_collect_pages(pc);
 576         else
 577                 collect_pages_on_all_cpus(pc);
 578 }
 579
 580 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 581 {
 582         struct cfs_trace_cpu_data *tcd;
 583         struct list_head *cur_head;
 584         struct cfs_trace_page *tage;
 585         struct cfs_trace_page *tmp;
 586         int i, cpu;
 587
 588         for_each_possible_cpu(cpu) {
 589                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 590                         cur_head = tcd->tcd_pages.next;
 591
 592                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 593                                                  linkage) {
 594                                 __LASSERT_TAGE_INVARIANT(tage);
 595
 596                                 if (tage->cpu != cpu || tage->type != i)
 597                                         continue;
 598
 599                                 cfs_tage_to_tail(tage, cur_head);
 600                                 tcd->tcd_cur_pages++;
 601                         }
 602                 }
 603         }
 604 }
 605
 606 static void put_pages_back(struct page_collection *pc)
 607 {
 608         if (!libcfs_panic_in_progress)
 609                 put_pages_back_on_all_cpus(pc);
 610 }
 611
 612 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 613  * we have a good amount of data at all times for dumping during an LBUG, even
 614  * if we have been steadily writing (and otherwise discarding) pages via the
 615  * debug daemon.
 616  */
 617 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 618                                          struct cfs_trace_cpu_data *tcd)
 619 {
 620         struct cfs_trace_page *tage;
 621         struct cfs_trace_page *tmp;
 622
 623         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 624                 __LASSERT_TAGE_INVARIANT(tage);
 625
 626                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 627                         continue;
 628
 629                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 630                 tcd->tcd_cur_daemon_pages++;
 631
 632                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 633                         struct cfs_trace_page *victim;
 634
 635                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 636                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 637
 638                         __LASSERT_TAGE_INVARIANT(victim);
 639
 640                         list_del(&victim->linkage);
 641                         cfs_tage_free(victim);
 642                         tcd->tcd_cur_daemon_pages--;
 643                 }
 644         }
 645 }
 646
 647 static void put_pages_on_daemon_list(struct page_collection *pc)
 648 {
 649         struct cfs_trace_cpu_data *tcd;
 650         int i, cpu;
 651
 652         for_each_possible_cpu(cpu) {
 653                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 654                         put_pages_on_tcd_daemon_list(pc, tcd);
 655         }
 656 }
 657
 658 void cfs_trace_debug_print(void)
 659 {
 660         struct page_collection pc;
 661         struct cfs_trace_page *tage;
 662         struct cfs_trace_page *tmp;
 663
 664         pc.pc_want_daemon_pages = 1;
 665         collect_pages(&pc);
 666         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 667                 char *p, *file, *fn;
 668                 struct page *page;
 669
 670                 __LASSERT_TAGE_INVARIANT(tage);
 671
 672                 page = tage->page;
 673                 p = page_address(page);
 674                 while (p < ((char *)page_address(page) + tage->used)) {
 675                         struct ptldebug_header *hdr;
 676                         int len;
 677
 678                         hdr = (void *)p;
 679                         p += sizeof(*hdr);
 680                         file = p;
 681                         p += strlen(file) + 1;
 682                         fn = p;
 683                         p += strlen(fn) + 1;
 684                         len = hdr->ph_len - (int)(p - (char *)hdr);
 685
 686                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 687
 688                         p += len;
 689                 }
 690
 691                 list_del(&tage->linkage);
 692                 cfs_tage_free(tage);
 693         }
 694 }
 695
 696 int cfs_tracefile_dump_all_pages(char *filename)
 697 {
 698         struct page_collection pc;
 699         struct file *filp;
 700         struct cfs_trace_page *tage;
 701         struct cfs_trace_page *tmp;
 702         char *buf;
 703         mm_segment_t __oldfs;
 704         int rc;
 705
 706         cfs_tracefile_write_lock();
 707
 708         filp = filp_open(filename, O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE,
 709                          0600);
 710         if (IS_ERR(filp)) {
 711                 rc = PTR_ERR(filp);
 712                 filp = NULL;
 713                 pr_err("LustreError: can't open %s for dump: rc %d\n",
 714                        filename, rc);
 715                 goto out;
 716         }
 717
 718         pc.pc_want_daemon_pages = 1;
 719         collect_pages(&pc);
 720         if (list_empty(&pc.pc_pages)) {
 721                 rc = 0;
 722                 goto close;
 723         }
 724         __oldfs = get_fs();
 725         set_fs(get_ds());
 726
 727         /* ok, for now, just write the pages.  in the future we'll be building
 728          * iobufs with the pages and calling generic_direct_IO
 729          */
 730         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 731                 __LASSERT_TAGE_INVARIANT(tage);
 732
 733                 buf = kmap(tage->page);
 734                 rc = kernel_write(filp, buf, tage->used, &filp->f_pos);
 735                 kunmap(tage->page);
 736
 737                 if (rc != (int)tage->used) {
 738                         pr_warn("wanted to write %u but wrote %d\n", tage->used,
 739                                 rc);
 740                         put_pages_back(&pc);
 741                         __LASSERT(list_empty(&pc.pc_pages));
 742                         break;
 743                 }
 744                 list_del(&tage->linkage);
 745                 cfs_tage_free(tage);
 746         }
 747         set_fs(__oldfs);
 748         rc = vfs_fsync(filp, 1);
 749         if (rc)
 750                 pr_err("sync returns %d\n", rc);
 751 close:
 752         filp_close(filp, NULL);
 753 out:
 754         cfs_tracefile_write_unlock();
 755         return rc;
 756 }
 757
 758 void cfs_trace_flush_pages(void)
 759 {
 760         struct page_collection pc;
 761         struct cfs_trace_page *tage;
 762         struct cfs_trace_page *tmp;
 763
 764         pc.pc_want_daemon_pages = 1;
 765         collect_pages(&pc);
 766         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 767                 __LASSERT_TAGE_INVARIANT(tage);
 768
 769                 list_del(&tage->linkage);
 770                 cfs_tage_free(tage);
 771         }
 772 }
 773
 774 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 775                             const char __user *usr_buffer, int usr_buffer_nob)
 776 {
 777         int nob;
 778
 779         if (usr_buffer_nob > knl_buffer_nob)
 780                 return -EOVERFLOW;
 781
 782         if (copy_from_user((void *)knl_buffer,
 783                            usr_buffer, usr_buffer_nob))
 784                 return -EFAULT;
 785
 786         nob = strnlen(knl_buffer, usr_buffer_nob);
 787         while (nob-- >= 0)                    /* strip trailing whitespace */
 788                 if (!isspace(knl_buffer[nob]))
 789                         break;
 790
 791         if (nob < 0)                        /* empty string */
 792                 return -EINVAL;
 793
 794         if (nob == knl_buffer_nob)            /* no space to terminate */
 795                 return -EOVERFLOW;
 796
 797         knl_buffer[nob + 1] = 0;                /* terminate */
 798         return 0;
 799 }
 800 EXPORT_SYMBOL(cfs_trace_copyin_string);
 801
 802 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 803                              const char *knl_buffer, char *append)
 804 {
 805         /*
 806          * NB if 'append' != NULL, it's a single character to append to the
 807          * copied out string - usually "\n" or "" (i.e. a terminating zero byte)
 808          */
 809         int nob = strlen(knl_buffer);
 810
 811         if (nob > usr_buffer_nob)
 812                 nob = usr_buffer_nob;
 813
 814         if (copy_to_user(usr_buffer, knl_buffer, nob))
 815                 return -EFAULT;
 816
 817         if (append && nob < usr_buffer_nob) {
 818                 if (copy_to_user(usr_buffer + nob, append, 1))
 819                         return -EFAULT;
 820
 821                 nob++;
 822         }
 823
 824         return nob;
 825 }
 826 EXPORT_SYMBOL(cfs_trace_copyout_string);
 827
 828 int cfs_trace_allocate_string_buffer(char **str, int nob)
 829 {
 830         if (nob > 2 * PAGE_SIZE)            /* string must be "sensible" */
 831                 return -EINVAL;
 832
 833         *str = kmalloc(nob, GFP_KERNEL | __GFP_ZERO);
 834         if (!*str)
 835                 return -ENOMEM;
 836
 837         return 0;
 838 }
 839
 840 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
 841 {
 842         char *str;
 843         int rc;
 844
 845         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 846         if (rc)
 847                 return rc;
 848
 849         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 850                                      usr_str, usr_str_nob);
 851         if (rc)
 852                 goto out;
 853
 854         if (str[0] != '/') {
 855                 rc = -EINVAL;
 856                 goto out;
 857         }
 858         rc = cfs_tracefile_dump_all_pages(str);
 859 out:
 860         kfree(str);
 861         return rc;
 862 }
 863
 864 int cfs_trace_daemon_command(char *str)
 865 {
 866         int rc = 0;
 867
 868         cfs_tracefile_write_lock();
 869
 870         if (!strcmp(str, "stop")) {
 871                 cfs_tracefile_write_unlock();
 872                 cfs_trace_stop_thread();
 873                 cfs_tracefile_write_lock();
 874                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 875
 876         } else if (!strncmp(str, "size=", 5)) {
 877                 unsigned long tmp;
 878
 879                 rc = kstrtoul(str + 5, 10, &tmp);
 880                 if (!rc) {
 881                         if (tmp < 10 || tmp > 20480)
 882                                 cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 883                         else
 884                                 cfs_tracefile_size = tmp << 20;
 885                 }
 886         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 887                 rc = -ENAMETOOLONG;
 888         } else if (str[0] != '/') {
 889                 rc = -EINVAL;
 890         } else {
 891                 strcpy(cfs_tracefile, str);
 892
 893                 pr_info("debug daemon will attempt to start writing to %s (%lukB max)\n",
 894                         cfs_tracefile,
 895                         (long)(cfs_tracefile_size >> 10));
 896
 897                 cfs_trace_start_thread();
 898         }
 899
 900         cfs_tracefile_write_unlock();
 901         return rc;
 902 }
 903
 904 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
 905 {
 906         char *str;
 907         int rc;
 908
 909         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 910         if (rc)
 911                 return rc;
 912
 913         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 914                                      usr_str, usr_str_nob);
 915         if (!rc)
 916                 rc = cfs_trace_daemon_command(str);
 917
 918         kfree(str);
 919         return rc;
 920 }
 921
 922 int cfs_trace_set_debug_mb(int mb)
 923 {
 924         int i;
 925         int j;
 926         int pages;
 927         int limit = cfs_trace_max_debug_mb();
 928         struct cfs_trace_cpu_data *tcd;
 929
 930         if (mb < num_possible_cpus()) {
 931                 pr_warn("%d MB is too small for debug buffer size, setting it to %d MB.\n",
 932                         mb, num_possible_cpus());
 933                 mb = num_possible_cpus();
 934         }
 935
 936         if (mb > limit) {
 937                 pr_warn("%d MB is too large for debug buffer size, setting it to %d MB.\n",
 938                         mb, limit);
 939                 mb = limit;
 940         }
 941
 942         mb /= num_possible_cpus();
 943         pages = mb << (20 - PAGE_SHIFT);
 944
 945         cfs_tracefile_write_lock();
 946
 947         cfs_tcd_for_each(tcd, i, j)
 948                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 949
 950         cfs_tracefile_write_unlock();
 951
 952         return 0;
 953 }
 954
 955 int cfs_trace_get_debug_mb(void)
 956 {
 957         int i;
 958         int j;
 959         struct cfs_trace_cpu_data *tcd;
 960         int total_pages = 0;
 961
 962         cfs_tracefile_read_lock();
 963
 964         cfs_tcd_for_each(tcd, i, j)
 965                 total_pages += tcd->tcd_max_pages;
 966
 967         cfs_tracefile_read_unlock();
 968
 969         return (total_pages >> (20 - PAGE_SHIFT)) + 1;
 970 }
 971
 972 static int tracefiled(void *arg)
 973 {
 974         struct page_collection pc;
 975         struct tracefiled_ctl *tctl = arg;
 976         struct cfs_trace_page *tage;
 977         struct cfs_trace_page *tmp;
 978         struct file *filp;
 979         char *buf;
 980         int last_loop = 0;
 981         int rc;
 982
 983         /* we're started late enough that we pick up init's fs context */
 984         /* this is so broken in uml?  what on earth is going on? */
 985
 986         complete(&tctl->tctl_start);
 987
 988         while (1) {
 989                 wait_queue_entry_t __wait;
 990
 991                 pc.pc_want_daemon_pages = 0;
 992                 collect_pages(&pc);
 993                 if (list_empty(&pc.pc_pages))
 994                         goto end_loop;
 995
 996                 filp = NULL;
 997                 cfs_tracefile_read_lock();
 998                 if (cfs_tracefile[0]) {
 999                         filp = filp_open(cfs_tracefile,
1000                                          O_CREAT | O_RDWR | O_LARGEFILE,
1001                                          0600);
1002                         if (IS_ERR(filp)) {
1003                                 rc = PTR_ERR(filp);
1004                                 filp = NULL;
1005                                 pr_warn("couldn't open %s: %d\n", cfs_tracefile,
1006                                         rc);
1007                         }
1008                 }
1009                 cfs_tracefile_read_unlock();
1010                 if (!filp) {
1011                         put_pages_on_daemon_list(&pc);
1012                         __LASSERT(list_empty(&pc.pc_pages));
1013                         goto end_loop;
1014                 }
1015
1016                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
1017                         static loff_t f_pos;
1018
1019                         __LASSERT_TAGE_INVARIANT(tage);
1020
1021                         if (f_pos >= (off_t)cfs_tracefile_size)
1022                                 f_pos = 0;
1023                         else if (f_pos > i_size_read(file_inode(filp)))
1024                                 f_pos = i_size_read(file_inode(filp));
1025
1026                         buf = kmap(tage->page);
1027                         rc = kernel_write(filp, buf, tage->used, &f_pos);
1028                         kunmap(tage->page);
1029
1030                         if (rc != (int)tage->used) {
1031                                 pr_warn("wanted to write %u but wrote %d\n",
1032                                         tage->used, rc);
1033                                 put_pages_back(&pc);
1034                                 __LASSERT(list_empty(&pc.pc_pages));
1035                                 break;
1036                         }
1037                 }
1038
1039                 filp_close(filp, NULL);
1040                 put_pages_on_daemon_list(&pc);
1041                 if (!list_empty(&pc.pc_pages)) {
1042                         int i;
1043
1044                         pr_alert("trace pages aren't empty\n");
1045                         pr_err("total cpus(%d): ", num_possible_cpus());
1046                         for (i = 0; i < num_possible_cpus(); i++)
1047                                 if (cpu_online(i))
1048                                         pr_cont("%d(on) ", i);
1049                                 else
1050                                         pr_cont("%d(off) ", i);
1051                         pr_cont("\n");
1052
1053                         i = 0;
1054                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1055                                                  linkage)
1056                                 pr_err("page %d belongs to cpu %d\n",
1057                                        ++i, tage->cpu);
1058                         pr_err("There are %d pages unwritten\n", i);
1059                 }
1060                 __LASSERT(list_empty(&pc.pc_pages));
1061 end_loop:
1062                 if (atomic_read(&tctl->tctl_shutdown)) {
1063                         if (!last_loop) {
1064                                 last_loop = 1;
1065                                 continue;
1066                         } else {
1067                                 break;
1068                         }
1069                 }
1070                 init_waitqueue_entry(&__wait, current);
1071                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1072                 set_current_state(TASK_INTERRUPTIBLE);
1073                 schedule_timeout(cfs_time_seconds(1));
1074                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1075         }
1076         complete(&tctl->tctl_stop);
1077         return 0;
1078 }
1079
1080 int cfs_trace_start_thread(void)
1081 {
1082         struct tracefiled_ctl *tctl = &trace_tctl;
1083         struct task_struct *task;
1084         int rc = 0;
1085
1086         mutex_lock(&cfs_trace_thread_mutex);
1087         if (thread_running)
1088                 goto out;
1089
1090         init_completion(&tctl->tctl_start);
1091         init_completion(&tctl->tctl_stop);
1092         init_waitqueue_head(&tctl->tctl_waitq);
1093         atomic_set(&tctl->tctl_shutdown, 0);
1094
1095         task = kthread_run(tracefiled, tctl, "ktracefiled");
1096         if (IS_ERR(task)) {
1097                 rc = PTR_ERR(task);
1098                 goto out;
1099         }
1100
1101         wait_for_completion(&tctl->tctl_start);
1102         thread_running = 1;
1103 out:
1104         mutex_unlock(&cfs_trace_thread_mutex);
1105         return rc;
1106 }
1107
1108 void cfs_trace_stop_thread(void)
1109 {
1110         struct tracefiled_ctl *tctl = &trace_tctl;
1111
1112         mutex_lock(&cfs_trace_thread_mutex);
1113         if (thread_running) {
1114                 pr_info("shutting down debug daemon thread...\n");
1115                 atomic_set(&tctl->tctl_shutdown, 1);
1116                 wait_for_completion(&tctl->tctl_stop);
1117                 thread_running = 0;
1118         }
1119         mutex_unlock(&cfs_trace_thread_mutex);
1120 }
1121
1122 int cfs_tracefile_init(int max_pages)
1123 {
1124         struct cfs_trace_cpu_data *tcd;
1125         int i;
1126         int j;
1127         int rc;
1128         int factor;
1129
1130         rc = cfs_tracefile_init_arch();
1131         if (rc)
1132                 return rc;
1133
1134         cfs_tcd_for_each(tcd, i, j) {
1135                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1136                 factor = tcd->tcd_pages_factor;
1137                 INIT_LIST_HEAD(&tcd->tcd_pages);
1138                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1139                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1140                 tcd->tcd_cur_pages = 0;
1141                 tcd->tcd_cur_stock_pages = 0;
1142                 tcd->tcd_cur_daemon_pages = 0;
1143                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1144                 LASSERT(tcd->tcd_max_pages > 0);
1145                 tcd->tcd_shutting_down = 0;
1146         }
1147
1148         return 0;
1149 }
1150
1151 static void trace_cleanup_on_all_cpus(void)
1152 {
1153         struct cfs_trace_cpu_data *tcd;
1154         struct cfs_trace_page *tage;
1155         struct cfs_trace_page *tmp;
1156         int i, cpu;
1157
1158         for_each_possible_cpu(cpu) {
1159                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1160                         tcd->tcd_shutting_down = 1;
1161
1162                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages,
1163                                                  linkage) {
1164                                 __LASSERT_TAGE_INVARIANT(tage);
1165
1166                                 list_del(&tage->linkage);
1167                                 cfs_tage_free(tage);
1168                         }
1169
1170                         tcd->tcd_cur_pages = 0;
1171                 }
1172         }
1173 }
1174
1175 static void cfs_trace_cleanup(void)
1176 {
1177         struct page_collection pc;
1178
1179         INIT_LIST_HEAD(&pc.pc_pages);
1180
1181         trace_cleanup_on_all_cpus();
1182
1183         cfs_tracefile_fini_arch();
1184 }
1185
1186 void cfs_tracefile_exit(void)
1187 {
1188         cfs_trace_stop_thread();
1189         cfs_trace_cleanup();
1190 }