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 #include "tracefile.h"
  41
  42 #include "../../include/linux/libcfs/libcfs.h"
  43
  44 /* XXX move things up to the top, comment */
  45 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  46
  47 char cfs_tracefile[TRACEFILE_NAME_SIZE];
  48 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
  49 static struct tracefiled_ctl trace_tctl;
  50 static DEFINE_MUTEX(cfs_trace_thread_mutex);
  51 static int thread_running;
  52
  53 static atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
  54
  55 struct page_collection {
  56         struct list_head        pc_pages;
  57         /*
  58          * if this flag is set, collect_pages() will spill both
  59          * ->tcd_daemon_pages and ->tcd_pages to the ->pc_pages. Otherwise,
  60          * only ->tcd_pages are spilled.
  61          */
  62         int             pc_want_daemon_pages;
  63 };
  64
  65 struct tracefiled_ctl {
  66         struct completion       tctl_start;
  67         struct completion       tctl_stop;
  68         wait_queue_head_t               tctl_waitq;
  69         pid_t                   tctl_pid;
  70         atomic_t                tctl_shutdown;
  71 };
  72
  73 /*
  74  * small data-structure for each page owned by tracefiled.
  75  */
  76 struct cfs_trace_page {
  77         /*
  78          * page itself
  79          */
  80         struct page       *page;
  81         /*
  82          * linkage into one of the lists in trace_data_union or
  83          * page_collection
  84          */
  85         struct list_head           linkage;
  86         /*
  87          * number of bytes used within this page
  88          */
  89         unsigned int     used;
  90         /*
  91          * cpu that owns this page
  92          */
  93         unsigned short       cpu;
  94         /*
  95          * type(context) of this page
  96          */
  97         unsigned short       type;
  98 };
  99
 100 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 101                                          struct cfs_trace_cpu_data *tcd);
 102
 103 static inline struct cfs_trace_page *
 104 cfs_tage_from_list(struct list_head *list)
 105 {
 106         return list_entry(list, struct cfs_trace_page, linkage);
 107 }
 108
 109 static struct cfs_trace_page *cfs_tage_alloc(gfp_t gfp)
 110 {
 111         struct page         *page;
 112         struct cfs_trace_page *tage;
 113
 114         /* My caller is trying to free memory */
 115         if (!in_interrupt() && memory_pressure_get())
 116                 return NULL;
 117
 118         /*
 119          * Don't spam console with allocation failures: they will be reported
 120          * by upper layer anyway.
 121          */
 122         gfp |= __GFP_NOWARN;
 123         page = alloc_page(gfp);
 124         if (!page)
 125                 return NULL;
 126
 127         tage = kmalloc(sizeof(*tage), gfp);
 128         if (!tage) {
 129                 __free_page(page);
 130                 return NULL;
 131         }
 132
 133         tage->page = page;
 134         atomic_inc(&cfs_tage_allocated);
 135         return tage;
 136 }
 137
 138 static void cfs_tage_free(struct cfs_trace_page *tage)
 139 {
 140         __free_page(tage->page);
 141         kfree(tage);
 142         atomic_dec(&cfs_tage_allocated);
 143 }
 144
 145 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
 146                              struct list_head *queue)
 147 {
 148         list_move_tail(&tage->linkage, queue);
 149 }
 150
 151 int cfs_trace_refill_stock(struct cfs_trace_cpu_data *tcd, gfp_t gfp,
 152                            struct list_head *stock)
 153 {
 154         int i;
 155
 156         /*
 157          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 158          * from here: this will lead to infinite recursion.
 159          */
 160
 161         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++i) {
 162                 struct cfs_trace_page *tage;
 163
 164                 tage = cfs_tage_alloc(gfp);
 165                 if (!tage)
 166                         break;
 167                 list_add_tail(&tage->linkage, stock);
 168         }
 169         return i;
 170 }
 171
 172 /* return a page that has 'len' bytes left at the end */
 173 static struct cfs_trace_page *
 174 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
 175 {
 176         struct cfs_trace_page *tage;
 177
 178         if (tcd->tcd_cur_pages > 0) {
 179                 __LASSERT(!list_empty(&tcd->tcd_pages));
 180                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
 181                 if (tage->used + len <= PAGE_SIZE)
 182                         return tage;
 183         }
 184
 185         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 186                 if (tcd->tcd_cur_stock_pages > 0) {
 187                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
 188                         --tcd->tcd_cur_stock_pages;
 189                         list_del_init(&tage->linkage);
 190                 } else {
 191                         tage = cfs_tage_alloc(GFP_ATOMIC);
 192                         if (unlikely(!tage)) {
 193                                 if ((!memory_pressure_get() ||
 194                                      in_interrupt()) && printk_ratelimit())
 195                                         printk(KERN_WARNING
 196                                                "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                 printk(KERN_WARNING "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-- == 0)
 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 == 0)
 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                         printk(KERN_EMERG "negative max_nob: %d\n",
 362                                max_nob);
 363                         mask |= D_ERROR;
 364                         cfs_trace_put_tcd(tcd);
 365                         tcd = NULL;
 366                         goto console;
 367                 }
 368
 369                 needed = 0;
 370                 if (format1) {
 371                         va_copy(ap, args);
 372                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 373                         va_end(ap);
 374                 }
 375
 376                 if (format2) {
 377                         remain = max_nob - needed;
 378                         if (remain < 0)
 379                                 remain = 0;
 380
 381                         va_start(ap, format2);
 382                         needed += vsnprintf(string_buf + needed, remain,
 383                                             format2, ap);
 384                         va_end(ap);
 385                 }
 386
 387                 if (needed < max_nob) /* well. printing ok.. */
 388                         break;
 389         }
 390
 391         if (*(string_buf + needed - 1) != '\n')
 392                 printk(KERN_INFO "format at %s:%d:%s doesn't end in newline\n",
 393                        file, msgdata->msg_line, msgdata->msg_fn);
 394
 395         header.ph_len = known_size + needed;
 396         debug_buf = (char *)page_address(tage->page) + tage->used;
 397
 398         if (libcfs_debug_binary) {
 399                 memcpy(debug_buf, &header, sizeof(header));
 400                 tage->used += sizeof(header);
 401                 debug_buf += sizeof(header);
 402         }
 403
 404         /* indent message according to the nesting level */
 405         while (depth-- > 0) {
 406                 *(debug_buf++) = '.';
 407                 ++tage->used;
 408         }
 409
 410         strcpy(debug_buf, file);
 411         tage->used += strlen(file) + 1;
 412         debug_buf += strlen(file) + 1;
 413
 414         if (msgdata->msg_fn) {
 415                 strcpy(debug_buf, msgdata->msg_fn);
 416                 tage->used += strlen(msgdata->msg_fn) + 1;
 417                 debug_buf += strlen(msgdata->msg_fn) + 1;
 418         }
 419
 420         __LASSERT(debug_buf == string_buf);
 421
 422         tage->used += needed;
 423         __LASSERT(tage->used <= PAGE_SIZE);
 424
 425 console:
 426         if ((mask & libcfs_printk) == 0) {
 427                 /* no console output requested */
 428                 if (tcd)
 429                         cfs_trace_put_tcd(tcd);
 430                 return 1;
 431         }
 432
 433         if (cdls) {
 434                 if (libcfs_console_ratelimit &&
 435                     cdls->cdls_next != 0 &&     /* not first time ever */
 436                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 437                         /* skipping a console message */
 438                         cdls->cdls_count++;
 439                         if (tcd)
 440                                 cfs_trace_put_tcd(tcd);
 441                         return 1;
 442                 }
 443
 444                 if (cfs_time_after(cfs_time_current(),
 445                                    cdls->cdls_next + libcfs_console_max_delay +
 446                                    cfs_time_seconds(10))) {
 447                         /* last timeout was a long time ago */
 448                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 449                 } else {
 450                         cdls->cdls_delay *= libcfs_console_backoff;
 451                 }
 452
 453                 if (cdls->cdls_delay < libcfs_console_min_delay)
 454                         cdls->cdls_delay = libcfs_console_min_delay;
 455                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 456                         cdls->cdls_delay = libcfs_console_max_delay;
 457
 458                 /* ensure cdls_next is never zero after it's been seen */
 459                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 460         }
 461
 462         if (tcd) {
 463                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 464                                      msgdata->msg_fn);
 465                 cfs_trace_put_tcd(tcd);
 466         } else {
 467                 string_buf = cfs_trace_get_console_buffer();
 468
 469                 needed = 0;
 470                 if (format1) {
 471                         va_copy(ap, args);
 472                         needed = vsnprintf(string_buf,
 473                                            CFS_TRACE_CONSOLE_BUFFER_SIZE,
 474                                            format1, ap);
 475                         va_end(ap);
 476                 }
 477                 if (format2) {
 478                         remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
 479                         if (remain > 0) {
 480                                 va_start(ap, format2);
 481                                 needed += vsnprintf(string_buf + needed, remain,
 482                                                     format2, ap);
 483                                 va_end(ap);
 484                         }
 485                 }
 486                 cfs_print_to_console(&header, mask,
 487                                      string_buf, needed, file, msgdata->msg_fn);
 488
 489                 put_cpu();
 490         }
 491
 492         if (cdls && cdls->cdls_count != 0) {
 493                 string_buf = cfs_trace_get_console_buffer();
 494
 495                 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 496                                   "Skipped %d previous similar message%s\n",
 497                                   cdls->cdls_count,
 498                                   (cdls->cdls_count > 1) ? "s" : "");
 499
 500                 cfs_print_to_console(&header, mask,
 501                                      string_buf, needed, file, msgdata->msg_fn);
 502
 503                 put_cpu();
 504                 cdls->cdls_count = 0;
 505         }
 506
 507         return 0;
 508 }
 509 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 510
 511 void
 512 cfs_trace_assertion_failed(const char *str,
 513                            struct libcfs_debug_msg_data *msgdata)
 514 {
 515         struct ptldebug_header hdr;
 516
 517         libcfs_panic_in_progress = 1;
 518         libcfs_catastrophe = 1;
 519         mb();
 520
 521         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 522
 523         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 524                              msgdata->msg_file, msgdata->msg_fn);
 525
 526         panic("Lustre debug assertion failure\n");
 527
 528         /* not reached */
 529 }
 530
 531 static void
 532 panic_collect_pages(struct page_collection *pc)
 533 {
 534         /* Do the collect_pages job on a single CPU: assumes that all other
 535          * CPUs have been stopped during a panic.  If this isn't true for some
 536          * arch, this will have to be implemented separately in each arch.
 537          */
 538         int                     i;
 539         int                     j;
 540         struct cfs_trace_cpu_data *tcd;
 541
 542         INIT_LIST_HEAD(&pc->pc_pages);
 543
 544         cfs_tcd_for_each(tcd, i, j) {
 545                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 546                 tcd->tcd_cur_pages = 0;
 547
 548                 if (pc->pc_want_daemon_pages) {
 549                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 550                         tcd->tcd_cur_daemon_pages = 0;
 551                 }
 552         }
 553 }
 554
 555 static void collect_pages_on_all_cpus(struct page_collection *pc)
 556 {
 557         struct cfs_trace_cpu_data *tcd;
 558         int i, cpu;
 559
 560         for_each_possible_cpu(cpu) {
 561                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 562                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 563                         tcd->tcd_cur_pages = 0;
 564                         if (pc->pc_want_daemon_pages) {
 565                                 list_splice_init(&tcd->tcd_daemon_pages,
 566                                                  &pc->pc_pages);
 567                                 tcd->tcd_cur_daemon_pages = 0;
 568                         }
 569                 }
 570         }
 571 }
 572
 573 static void collect_pages(struct page_collection *pc)
 574 {
 575         INIT_LIST_HEAD(&pc->pc_pages);
 576
 577         if (libcfs_panic_in_progress)
 578                 panic_collect_pages(pc);
 579         else
 580                 collect_pages_on_all_cpus(pc);
 581 }
 582
 583 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 584 {
 585         struct cfs_trace_cpu_data *tcd;
 586         struct list_head *cur_head;
 587         struct cfs_trace_page *tage;
 588         struct cfs_trace_page *tmp;
 589         int i, cpu;
 590
 591         for_each_possible_cpu(cpu) {
 592                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 593                         cur_head = tcd->tcd_pages.next;
 594
 595                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 596                                                  linkage) {
 597                                 __LASSERT_TAGE_INVARIANT(tage);
 598
 599                                 if (tage->cpu != cpu || tage->type != i)
 600                                         continue;
 601
 602                                 cfs_tage_to_tail(tage, cur_head);
 603                                 tcd->tcd_cur_pages++;
 604                         }
 605                 }
 606         }
 607 }
 608
 609 static void put_pages_back(struct page_collection *pc)
 610 {
 611         if (!libcfs_panic_in_progress)
 612                 put_pages_back_on_all_cpus(pc);
 613 }
 614
 615 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 616  * we have a good amount of data at all times for dumping during an LBUG, even
 617  * if we have been steadily writing (and otherwise discarding) pages via the
 618  * debug daemon.
 619  */
 620 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 621                                          struct cfs_trace_cpu_data *tcd)
 622 {
 623         struct cfs_trace_page *tage;
 624         struct cfs_trace_page *tmp;
 625
 626         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 627                 __LASSERT_TAGE_INVARIANT(tage);
 628
 629                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 630                         continue;
 631
 632                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 633                 tcd->tcd_cur_daemon_pages++;
 634
 635                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 636                         struct cfs_trace_page *victim;
 637
 638                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 639                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 640
 641                         __LASSERT_TAGE_INVARIANT(victim);
 642
 643                         list_del(&victim->linkage);
 644                         cfs_tage_free(victim);
 645                         tcd->tcd_cur_daemon_pages--;
 646                 }
 647         }
 648 }
 649
 650 static void put_pages_on_daemon_list(struct page_collection *pc)
 651 {
 652         struct cfs_trace_cpu_data *tcd;
 653         int i, cpu;
 654
 655         for_each_possible_cpu(cpu) {
 656                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 657                         put_pages_on_tcd_daemon_list(pc, tcd);
 658         }
 659 }
 660
 661 void cfs_trace_debug_print(void)
 662 {
 663         struct page_collection pc;
 664         struct cfs_trace_page *tage;
 665         struct cfs_trace_page *tmp;
 666
 667         pc.pc_want_daemon_pages = 1;
 668         collect_pages(&pc);
 669         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 670                 char *p, *file, *fn;
 671                 struct page *page;
 672
 673                 __LASSERT_TAGE_INVARIANT(tage);
 674
 675                 page = tage->page;
 676                 p = page_address(page);
 677                 while (p < ((char *)page_address(page) + tage->used)) {
 678                         struct ptldebug_header *hdr;
 679                         int len;
 680
 681                         hdr = (void *)p;
 682                         p += sizeof(*hdr);
 683                         file = p;
 684                         p += strlen(file) + 1;
 685                         fn = p;
 686                         p += strlen(fn) + 1;
 687                         len = hdr->ph_len - (int)(p - (char *)hdr);
 688
 689                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 690
 691                         p += len;
 692                 }
 693
 694                 list_del(&tage->linkage);
 695                 cfs_tage_free(tage);
 696         }
 697 }
 698
 699 int cfs_tracefile_dump_all_pages(char *filename)
 700 {
 701         struct page_collection  pc;
 702         struct file             *filp;
 703         struct cfs_trace_page   *tage;
 704         struct cfs_trace_page   *tmp;
 705         char                    *buf;
 706         mm_segment_t __oldfs;
 707         int rc;
 708
 709         cfs_tracefile_write_lock();
 710
 711         filp = filp_open(filename, O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE,
 712                          0600);
 713         if (IS_ERR(filp)) {
 714                 rc = PTR_ERR(filp);
 715                 filp = NULL;
 716                 pr_err("LustreError: can't open %s for dump: rc %d\n",
 717                        filename, rc);
 718                 goto out;
 719         }
 720
 721         pc.pc_want_daemon_pages = 1;
 722         collect_pages(&pc);
 723         if (list_empty(&pc.pc_pages)) {
 724                 rc = 0;
 725                 goto close;
 726         }
 727         __oldfs = get_fs();
 728         set_fs(get_ds());
 729
 730         /* ok, for now, just write the pages.  in the future we'll be building
 731          * iobufs with the pages and calling generic_direct_IO
 732          */
 733         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 734                 __LASSERT_TAGE_INVARIANT(tage);
 735
 736                 buf = kmap(tage->page);
 737                 rc = vfs_write(filp, (__force const char __user *)buf,
 738                                tage->used, &filp->f_pos);
 739                 kunmap(tage->page);
 740
 741                 if (rc != (int)tage->used) {
 742                         printk(KERN_WARNING "wanted to write %u but wrote %d\n",
 743                                tage->used, rc);
 744                         put_pages_back(&pc);
 745                         __LASSERT(list_empty(&pc.pc_pages));
 746                         break;
 747                 }
 748                 list_del(&tage->linkage);
 749                 cfs_tage_free(tage);
 750         }
 751         set_fs(__oldfs);
 752         rc = vfs_fsync(filp, 1);
 753         if (rc)
 754                 pr_err("sync returns %d\n", rc);
 755 close:
 756         filp_close(filp, NULL);
 757 out:
 758         cfs_tracefile_write_unlock();
 759         return rc;
 760 }
 761
 762 void cfs_trace_flush_pages(void)
 763 {
 764         struct page_collection pc;
 765         struct cfs_trace_page *tage;
 766         struct cfs_trace_page *tmp;
 767
 768         pc.pc_want_daemon_pages = 1;
 769         collect_pages(&pc);
 770         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 771                 __LASSERT_TAGE_INVARIANT(tage);
 772
 773                 list_del(&tage->linkage);
 774                 cfs_tage_free(tage);
 775         }
 776 }
 777
 778 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 779                             const char __user *usr_buffer, int usr_buffer_nob)
 780 {
 781         int    nob;
 782
 783         if (usr_buffer_nob > knl_buffer_nob)
 784                 return -EOVERFLOW;
 785
 786         if (copy_from_user((void *)knl_buffer,
 787                            usr_buffer, usr_buffer_nob))
 788                 return -EFAULT;
 789
 790         nob = strnlen(knl_buffer, usr_buffer_nob);
 791         while (nob-- >= 0)                    /* strip trailing whitespace */
 792                 if (!isspace(knl_buffer[nob]))
 793                         break;
 794
 795         if (nob < 0)                        /* empty string */
 796                 return -EINVAL;
 797
 798         if (nob == knl_buffer_nob)            /* no space to terminate */
 799                 return -EOVERFLOW;
 800
 801         knl_buffer[nob + 1] = 0;                /* terminate */
 802         return 0;
 803 }
 804 EXPORT_SYMBOL(cfs_trace_copyin_string);
 805
 806 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 807                              const char *knl_buffer, char *append)
 808 {
 809         /*
 810          * NB if 'append' != NULL, it's a single character to append to the
 811          * copied out string - usually "\n" or "" (i.e. a terminating zero byte)
 812          */
 813         int   nob = strlen(knl_buffer);
 814
 815         if (nob > usr_buffer_nob)
 816                 nob = usr_buffer_nob;
 817
 818         if (copy_to_user(usr_buffer, knl_buffer, nob))
 819                 return -EFAULT;
 820
 821         if (append && nob < usr_buffer_nob) {
 822                 if (copy_to_user(usr_buffer + nob, append, 1))
 823                         return -EFAULT;
 824
 825                 nob++;
 826         }
 827
 828         return nob;
 829 }
 830 EXPORT_SYMBOL(cfs_trace_copyout_string);
 831
 832 int cfs_trace_allocate_string_buffer(char **str, int nob)
 833 {
 834         if (nob > 2 * PAGE_SIZE)            /* string must be "sensible" */
 835                 return -EINVAL;
 836
 837         *str = kmalloc(nob, GFP_KERNEL | __GFP_ZERO);
 838         if (!*str)
 839                 return -ENOMEM;
 840
 841         return 0;
 842 }
 843
 844 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
 845 {
 846         char     *str;
 847         int        rc;
 848
 849         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 850         if (rc != 0)
 851                 return rc;
 852
 853         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 854                                      usr_str, usr_str_nob);
 855         if (rc != 0)
 856                 goto out;
 857
 858         if (str[0] != '/') {
 859                 rc = -EINVAL;
 860                 goto out;
 861         }
 862         rc = cfs_tracefile_dump_all_pages(str);
 863 out:
 864         kfree(str);
 865         return rc;
 866 }
 867
 868 int cfs_trace_daemon_command(char *str)
 869 {
 870         int       rc = 0;
 871
 872         cfs_tracefile_write_lock();
 873
 874         if (strcmp(str, "stop") == 0) {
 875                 cfs_tracefile_write_unlock();
 876                 cfs_trace_stop_thread();
 877                 cfs_tracefile_write_lock();
 878                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 879
 880         } else if (strncmp(str, "size=", 5) == 0) {
 881                 unsigned long tmp;
 882
 883                 rc = kstrtoul(str + 5, 10, &tmp);
 884                 if (!rc) {
 885                         if (tmp < 10 || tmp > 20480)
 886                                 cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 887                         else
 888                                 cfs_tracefile_size = tmp << 20;
 889                 }
 890         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 891                 rc = -ENAMETOOLONG;
 892         } else if (str[0] != '/') {
 893                 rc = -EINVAL;
 894         } else {
 895                 strcpy(cfs_tracefile, str);
 896
 897                 printk(KERN_INFO
 898                        "Lustre: debug daemon will attempt to start writing to %s (%lukB max)\n",
 899                        cfs_tracefile,
 900                        (long)(cfs_tracefile_size >> 10));
 901
 902                 cfs_trace_start_thread();
 903         }
 904
 905         cfs_tracefile_write_unlock();
 906         return rc;
 907 }
 908
 909 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
 910 {
 911         char *str;
 912         int   rc;
 913
 914         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 915         if (rc != 0)
 916                 return rc;
 917
 918         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 919                                      usr_str, usr_str_nob);
 920         if (rc == 0)
 921                 rc = cfs_trace_daemon_command(str);
 922
 923         kfree(str);
 924         return rc;
 925 }
 926
 927 int cfs_trace_set_debug_mb(int mb)
 928 {
 929         int i;
 930         int j;
 931         int pages;
 932         int limit = cfs_trace_max_debug_mb();
 933         struct cfs_trace_cpu_data *tcd;
 934
 935         if (mb < num_possible_cpus()) {
 936                 printk(KERN_WARNING
 937                        "Lustre: %d MB is too small for debug buffer size, setting it to %d MB.\n",
 938                        mb, num_possible_cpus());
 939                 mb = num_possible_cpus();
 940         }
 941
 942         if (mb > limit) {
 943                 printk(KERN_WARNING
 944                        "Lustre: %d MB is too large for debug buffer size, setting it to %d MB.\n",
 945                        mb, limit);
 946                 mb = limit;
 947         }
 948
 949         mb /= num_possible_cpus();
 950         pages = mb << (20 - PAGE_SHIFT);
 951
 952         cfs_tracefile_write_lock();
 953
 954         cfs_tcd_for_each(tcd, i, j)
 955                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 956
 957         cfs_tracefile_write_unlock();
 958
 959         return 0;
 960 }
 961
 962 int cfs_trace_get_debug_mb(void)
 963 {
 964         int i;
 965         int j;
 966         struct cfs_trace_cpu_data *tcd;
 967         int total_pages = 0;
 968
 969         cfs_tracefile_read_lock();
 970
 971         cfs_tcd_for_each(tcd, i, j)
 972                 total_pages += tcd->tcd_max_pages;
 973
 974         cfs_tracefile_read_unlock();
 975
 976         return (total_pages >> (20 - PAGE_SHIFT)) + 1;
 977 }
 978
 979 static int tracefiled(void *arg)
 980 {
 981         struct page_collection pc;
 982         struct tracefiled_ctl *tctl = arg;
 983         struct cfs_trace_page *tage;
 984         struct cfs_trace_page *tmp;
 985         mm_segment_t __oldfs;
 986         struct file *filp;
 987         char *buf;
 988         int last_loop = 0;
 989         int rc;
 990
 991         /* we're started late enough that we pick up init's fs context */
 992         /* this is so broken in uml?  what on earth is going on? */
 993
 994         complete(&tctl->tctl_start);
 995
 996         while (1) {
 997                 wait_queue_t __wait;
 998
 999                 pc.pc_want_daemon_pages = 0;
1000                 collect_pages(&pc);
1001                 if (list_empty(&pc.pc_pages))
1002                         goto end_loop;
1003
1004                 filp = NULL;
1005                 cfs_tracefile_read_lock();
1006                 if (cfs_tracefile[0] != 0) {
1007                         filp = filp_open(cfs_tracefile,
1008                                          O_CREAT | O_RDWR | O_LARGEFILE,
1009                                          0600);
1010                         if (IS_ERR(filp)) {
1011                                 rc = PTR_ERR(filp);
1012                                 filp = NULL;
1013                                 printk(KERN_WARNING "couldn't open %s: %d\n",
1014                                        cfs_tracefile, rc);
1015                         }
1016                 }
1017                 cfs_tracefile_read_unlock();
1018                 if (!filp) {
1019                         put_pages_on_daemon_list(&pc);
1020                         __LASSERT(list_empty(&pc.pc_pages));
1021                         goto end_loop;
1022                 }
1023                 __oldfs = get_fs();
1024                 set_fs(get_ds());
1025
1026                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
1027                         static loff_t f_pos;
1028
1029                         __LASSERT_TAGE_INVARIANT(tage);
1030
1031                         if (f_pos >= (off_t)cfs_tracefile_size)
1032                                 f_pos = 0;
1033                         else if (f_pos > i_size_read(file_inode(filp)))
1034                                 f_pos = i_size_read(file_inode(filp));
1035
1036                         buf = kmap(tage->page);
1037                         rc = vfs_write(filp, (__force const char __user *)buf,
1038                                        tage->used, &f_pos);
1039                         kunmap(tage->page);
1040
1041                         if (rc != (int)tage->used) {
1042                                 printk(KERN_WARNING "wanted to write %u but wrote %d\n",
1043                                        tage->used, rc);
1044                                 put_pages_back(&pc);
1045                                 __LASSERT(list_empty(&pc.pc_pages));
1046                                 break;
1047                         }
1048                 }
1049                 set_fs(__oldfs);
1050
1051                 filp_close(filp, NULL);
1052                 put_pages_on_daemon_list(&pc);
1053                 if (!list_empty(&pc.pc_pages)) {
1054                         int i;
1055
1056                         printk(KERN_ALERT "Lustre: trace pages aren't empty\n");
1057                         pr_err("total cpus(%d): ", num_possible_cpus());
1058                         for (i = 0; i < num_possible_cpus(); i++)
1059                                 if (cpu_online(i))
1060                                         pr_cont("%d(on) ", i);
1061                                 else
1062                                         pr_cont("%d(off) ", i);
1063                         pr_cont("\n");
1064
1065                         i = 0;
1066                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1067                                                  linkage)
1068                                 pr_err("page %d belongs to cpu %d\n",
1069                                        ++i, tage->cpu);
1070                         pr_err("There are %d pages unwritten\n", i);
1071                 }
1072                 __LASSERT(list_empty(&pc.pc_pages));
1073 end_loop:
1074                 if (atomic_read(&tctl->tctl_shutdown)) {
1075                         if (last_loop == 0) {
1076                                 last_loop = 1;
1077                                 continue;
1078                         } else {
1079                                 break;
1080                         }
1081                 }
1082                 init_waitqueue_entry(&__wait, current);
1083                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1084                 set_current_state(TASK_INTERRUPTIBLE);
1085                 schedule_timeout(cfs_time_seconds(1));
1086                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1087         }
1088         complete(&tctl->tctl_stop);
1089         return 0;
1090 }
1091
1092 int cfs_trace_start_thread(void)
1093 {
1094         struct tracefiled_ctl *tctl = &trace_tctl;
1095         struct task_struct *task;
1096         int rc = 0;
1097
1098         mutex_lock(&cfs_trace_thread_mutex);
1099         if (thread_running)
1100                 goto out;
1101
1102         init_completion(&tctl->tctl_start);
1103         init_completion(&tctl->tctl_stop);
1104         init_waitqueue_head(&tctl->tctl_waitq);
1105         atomic_set(&tctl->tctl_shutdown, 0);
1106
1107         task = kthread_run(tracefiled, tctl, "ktracefiled");
1108         if (IS_ERR(task)) {
1109                 rc = PTR_ERR(task);
1110                 goto out;
1111         }
1112
1113         wait_for_completion(&tctl->tctl_start);
1114         thread_running = 1;
1115 out:
1116         mutex_unlock(&cfs_trace_thread_mutex);
1117         return rc;
1118 }
1119
1120 void cfs_trace_stop_thread(void)
1121 {
1122         struct tracefiled_ctl *tctl = &trace_tctl;
1123
1124         mutex_lock(&cfs_trace_thread_mutex);
1125         if (thread_running) {
1126                 printk(KERN_INFO
1127                        "Lustre: shutting down debug daemon thread...\n");
1128                 atomic_set(&tctl->tctl_shutdown, 1);
1129                 wait_for_completion(&tctl->tctl_stop);
1130                 thread_running = 0;
1131         }
1132         mutex_unlock(&cfs_trace_thread_mutex);
1133 }
1134
1135 int cfs_tracefile_init(int max_pages)
1136 {
1137         struct cfs_trace_cpu_data *tcd;
1138         int                 i;
1139         int                 j;
1140         int                 rc;
1141         int                 factor;
1142
1143         rc = cfs_tracefile_init_arch();
1144         if (rc != 0)
1145                 return rc;
1146
1147         cfs_tcd_for_each(tcd, i, j) {
1148                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1149                 factor = tcd->tcd_pages_factor;
1150                 INIT_LIST_HEAD(&tcd->tcd_pages);
1151                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1152                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1153                 tcd->tcd_cur_pages = 0;
1154                 tcd->tcd_cur_stock_pages = 0;
1155                 tcd->tcd_cur_daemon_pages = 0;
1156                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1157                 LASSERT(tcd->tcd_max_pages > 0);
1158                 tcd->tcd_shutting_down = 0;
1159         }
1160
1161         return 0;
1162 }
1163
1164 static void trace_cleanup_on_all_cpus(void)
1165 {
1166         struct cfs_trace_cpu_data *tcd;
1167         struct cfs_trace_page *tage;
1168         struct cfs_trace_page *tmp;
1169         int i, cpu;
1170
1171         for_each_possible_cpu(cpu) {
1172                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1173                         tcd->tcd_shutting_down = 1;
1174
1175                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages,
1176                                                  linkage) {
1177                                 __LASSERT_TAGE_INVARIANT(tage);
1178
1179                                 list_del(&tage->linkage);
1180                                 cfs_tage_free(tage);
1181                         }
1182
1183                         tcd->tcd_cur_pages = 0;
1184                 }
1185         }
1186 }
1187
1188 static void cfs_trace_cleanup(void)
1189 {
1190         struct page_collection pc;
1191
1192         INIT_LIST_HEAD(&pc.pc_pages);
1193
1194         trace_cleanup_on_all_cpus();
1195
1196         cfs_tracefile_fini_arch();
1197 }
1198
1199 void cfs_tracefile_exit(void)
1200 {
1201         cfs_trace_stop_thread();
1202         cfs_trace_cleanup();
1203 }