drivers/staging/lustre/lnet/libcfs/tracefile.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * GPL HEADER START
   4  *
   5  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License version 2 only,
   9  * as published by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope that it will be useful, but
  12  * WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * General Public License version 2 for more details (a copy is included
  15  * in the LICENSE file that accompanied this code).
  16  *
  17  * You should have received a copy of the GNU General Public License
  18  * version 2 along with this program; If not, see
  19  * http://www.gnu.org/licenses/gpl-2.0.html
  20  *
  21  * GPL HEADER END
  22  */
  23 /*
  24  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  25  * Use is subject to license terms.
  26  *
  27  * Copyright (c) 2012, Intel Corporation.
  28  */
  29 /*
  30  * This file is part of Lustre, http://www.lustre.org/
  31  * Lustre is a trademark of Sun Microsystems, Inc.
  32  *
  33  * libcfs/libcfs/tracefile.c
  34  *
  35  * Author: Zach Brown <zab@clusterfs.com>
  36  * Author: Phil Schwan <phil@clusterfs.com>
  37  */
  38
  39 #define DEBUG_SUBSYSTEM S_LNET
  40 #define LUSTRE_TRACEFILE_PRIVATE
  41 #define pr_fmt(fmt) "Lustre: " fmt
  42 #include "tracefile.h"
  43
  44 #include <linux/libcfs/libcfs.h>
  45
  46 /* XXX move things up to the top, comment */
  47 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  48
  49 char cfs_tracefile[TRACEFILE_NAME_SIZE];
  50 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
  51 static struct tracefiled_ctl trace_tctl;
  52 static DEFINE_MUTEX(cfs_trace_thread_mutex);
  53 static int thread_running;
  54
  55 static atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
  56
  57 struct page_collection {
  58         struct list_head        pc_pages;
  59         /*
  60          * if this flag is set, collect_pages() will spill both
  61          * ->tcd_daemon_pages and ->tcd_pages to the ->pc_pages. Otherwise,
  62          * only ->tcd_pages are spilled.
  63          */
  64         int                     pc_want_daemon_pages;
  65 };
  66
  67 struct tracefiled_ctl {
  68         struct completion       tctl_start;
  69         struct completion       tctl_stop;
  70         wait_queue_head_t       tctl_waitq;
  71         pid_t                   tctl_pid;
  72         atomic_t                tctl_shutdown;
  73 };
  74
  75 /*
  76  * small data-structure for each page owned by tracefiled.
  77  */
  78 struct cfs_trace_page {
  79         /*
  80          * page itself
  81          */
  82         struct page             *page;
  83         /*
  84          * linkage into one of the lists in trace_data_union or
  85          * page_collection
  86          */
  87         struct list_head        linkage;
  88         /*
  89          * number of bytes used within this page
  90          */
  91         unsigned int            used;
  92         /*
  93          * cpu that owns this page
  94          */
  95         unsigned short          cpu;
  96         /*
  97          * type(context) of this page
  98          */
  99         unsigned short          type;
 100 };
 101
 102 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 103                                          struct cfs_trace_cpu_data *tcd);
 104
 105 static inline struct cfs_trace_page *
 106 cfs_tage_from_list(struct list_head *list)
 107 {
 108         return list_entry(list, struct cfs_trace_page, linkage);
 109 }
 110
 111 static struct cfs_trace_page *cfs_tage_alloc(gfp_t gfp)
 112 {
 113         struct page *page;
 114         struct cfs_trace_page *tage;
 115
 116         /* My caller is trying to free memory */
 117         if (!in_interrupt() && memory_pressure_get())
 118                 return NULL;
 119
 120         /*
 121          * Don't spam console with allocation failures: they will be reported
 122          * by upper layer anyway.
 123          */
 124         gfp |= __GFP_NOWARN;
 125         page = alloc_page(gfp);
 126         if (!page)
 127                 return NULL;
 128
 129         tage = kmalloc(sizeof(*tage), gfp);
 130         if (!tage) {
 131                 __free_page(page);
 132                 return NULL;
 133         }
 134
 135         tage->page = page;
 136         atomic_inc(&cfs_tage_allocated);
 137         return tage;
 138 }
 139
 140 static void cfs_tage_free(struct cfs_trace_page *tage)
 141 {
 142         __free_page(tage->page);
 143         kfree(tage);
 144         atomic_dec(&cfs_tage_allocated);
 145 }
 146
 147 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
 148                              struct list_head *queue)
 149 {
 150         list_move_tail(&tage->linkage, queue);
 151 }
 152
 153 int cfs_trace_refill_stock(struct cfs_trace_cpu_data *tcd, gfp_t gfp,
 154                            struct list_head *stock)
 155 {
 156         int i;
 157
 158         /*
 159          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 160          * from here: this will lead to infinite recursion.
 161          */
 162
 163         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++i) {
 164                 struct cfs_trace_page *tage;
 165
 166                 tage = cfs_tage_alloc(gfp);
 167                 if (!tage)
 168                         break;
 169                 list_add_tail(&tage->linkage, stock);
 170         }
 171         return i;
 172 }
 173
 174 /* return a page that has 'len' bytes left at the end */
 175 static struct cfs_trace_page *
 176 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
 177 {
 178         struct cfs_trace_page *tage;
 179
 180         if (tcd->tcd_cur_pages > 0) {
 181                 __LASSERT(!list_empty(&tcd->tcd_pages));
 182                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
 183                 if (tage->used + len <= PAGE_SIZE)
 184                         return tage;
 185         }
 186
 187         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 188                 if (tcd->tcd_cur_stock_pages > 0) {
 189                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
 190                         --tcd->tcd_cur_stock_pages;
 191                         list_del_init(&tage->linkage);
 192                 } else {
 193                         tage = cfs_tage_alloc(GFP_ATOMIC);
 194                         if (unlikely(!tage)) {
 195                                 if (!memory_pressure_get() || in_interrupt())
 196                                         pr_warn_ratelimited("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         pr_warn_ratelimited("debug daemon buffer overflowed; discarding 10%% of pages (%d of %ld)\n",
 233                             pgcount + 1, tcd->tcd_cur_pages);
 234
 235         INIT_LIST_HEAD(&pc.pc_pages);
 236
 237         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 238                 if (!pgcount--)
 239                         break;
 240
 241                 list_move_tail(&tage->linkage, &pc.pc_pages);
 242                 tcd->tcd_cur_pages--;
 243         }
 244         put_pages_on_tcd_daemon_list(&pc, tcd);
 245 }
 246
 247 /* return a page that has 'len' bytes left at the end */
 248 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 249                                                  unsigned long len)
 250 {
 251         struct cfs_trace_page *tage;
 252
 253         /*
 254          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 255          * from here: this will lead to infinite recursion.
 256          */
 257
 258         if (len > PAGE_SIZE) {
 259                 pr_err("cowardly refusing to write %lu bytes in a page\n", len);
 260                 return NULL;
 261         }
 262
 263         tage = cfs_trace_get_tage_try(tcd, len);
 264         if (tage)
 265                 return tage;
 266         if (thread_running)
 267                 cfs_tcd_shrink(tcd);
 268         if (tcd->tcd_cur_pages > 0) {
 269                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 270                 tage->used = 0;
 271                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 272         }
 273         return tage;
 274 }
 275
 276 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 277                      const char *format, ...)
 278 {
 279         va_list args;
 280         int rc;
 281
 282         va_start(args, format);
 283         rc = libcfs_debug_vmsg2(msgdata, format, args, NULL);
 284         va_end(args);
 285
 286         return rc;
 287 }
 288 EXPORT_SYMBOL(libcfs_debug_msg);
 289
 290 int libcfs_debug_vmsg2(struct libcfs_debug_msg_data *msgdata,
 291                        const char *format1, va_list args,
 292                        const char *format2, ...)
 293 {
 294         struct cfs_trace_cpu_data *tcd = NULL;
 295         struct ptldebug_header header = { 0 };
 296         struct cfs_trace_page *tage;
 297         /* string_buf is used only if tcd != NULL, and is always set then */
 298         char *string_buf = NULL;
 299         char *debug_buf;
 300         int known_size;
 301         int needed = 85; /* average message length */
 302         int max_nob;
 303         va_list ap;
 304         int depth;
 305         int i;
 306         int remain;
 307         int mask = msgdata->msg_mask;
 308         const char *file = kbasename(msgdata->msg_file);
 309         struct cfs_debug_limit_state *cdls = msgdata->msg_cdls;
 310
 311         tcd = cfs_trace_get_tcd();
 312
 313         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 314          * pins us to a particular CPU.  This avoids an smp_processor_id()
 315          * warning on Linux when debugging is enabled.
 316          */
 317         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 318
 319         if (!tcd)               /* arch may not log in IRQ context */
 320                 goto console;
 321
 322         if (!tcd->tcd_cur_pages)
 323                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 324
 325         if (tcd->tcd_shutting_down) {
 326                 cfs_trace_put_tcd(tcd);
 327                 tcd = NULL;
 328                 goto console;
 329         }
 330
 331         depth = __current_nesting_level();
 332         known_size = strlen(file) + 1 + depth;
 333         if (msgdata->msg_fn)
 334                 known_size += strlen(msgdata->msg_fn) + 1;
 335
 336         if (libcfs_debug_binary)
 337                 known_size += sizeof(header);
 338
 339         /*
 340          * '2' used because vsnprintf return real size required for output
 341          * _without_ terminating NULL.
 342          * if needed is to small for this format.
 343          */
 344         for (i = 0; i < 2; i++) {
 345                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 346                 if (!tage) {
 347                         if (needed + known_size > PAGE_SIZE)
 348                                 mask |= D_ERROR;
 349
 350                         cfs_trace_put_tcd(tcd);
 351                         tcd = NULL;
 352                         goto console;
 353                 }
 354
 355                 string_buf = (char *)page_address(tage->page) +
 356                                         tage->used + known_size;
 357
 358                 max_nob = PAGE_SIZE - tage->used - known_size;
 359                 if (max_nob <= 0) {
 360                         pr_emerg("negative max_nob: %d\n", max_nob);
 361                         mask |= D_ERROR;
 362                         cfs_trace_put_tcd(tcd);
 363                         tcd = NULL;
 364                         goto console;
 365                 }
 366
 367                 needed = 0;
 368                 if (format1) {
 369                         va_copy(ap, args);
 370                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 371                         va_end(ap);
 372                 }
 373
 374                 if (format2) {
 375                         remain = max_nob - needed;
 376                         if (remain < 0)
 377                                 remain = 0;
 378
 379                         va_start(ap, format2);
 380                         needed += vsnprintf(string_buf + needed, remain,
 381                                             format2, ap);
 382                         va_end(ap);
 383                 }
 384
 385                 if (needed < max_nob) /* well. printing ok.. */
 386                         break;
 387         }
 388
 389         if (*(string_buf + needed - 1) != '\n')
 390                 pr_info("format at %s:%d:%s doesn't end in newline\n", file,
 391                         msgdata->msg_line, msgdata->msg_fn);
 392
 393         header.ph_len = known_size + needed;
 394         debug_buf = (char *)page_address(tage->page) + tage->used;
 395
 396         if (libcfs_debug_binary) {
 397                 memcpy(debug_buf, &header, sizeof(header));
 398                 tage->used += sizeof(header);
 399                 debug_buf += sizeof(header);
 400         }
 401
 402         /* indent message according to the nesting level */
 403         while (depth-- > 0) {
 404                 *(debug_buf++) = '.';
 405                 ++tage->used;
 406         }
 407
 408         strcpy(debug_buf, file);
 409         tage->used += strlen(file) + 1;
 410         debug_buf += strlen(file) + 1;
 411
 412         if (msgdata->msg_fn) {
 413                 strcpy(debug_buf, msgdata->msg_fn);
 414                 tage->used += strlen(msgdata->msg_fn) + 1;
 415                 debug_buf += strlen(msgdata->msg_fn) + 1;
 416         }
 417
 418         __LASSERT(debug_buf == string_buf);
 419
 420         tage->used += needed;
 421         __LASSERT(tage->used <= PAGE_SIZE);
 422
 423 console:
 424         if (!(mask & libcfs_printk)) {
 425                 /* no console output requested */
 426                 if (tcd)
 427                         cfs_trace_put_tcd(tcd);
 428                 return 1;
 429         }
 430
 431         if (cdls) {
 432                 if (libcfs_console_ratelimit &&
 433                     cdls->cdls_next &&          /* not first time ever */
 434                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 435                         /* skipping a console message */
 436                         cdls->cdls_count++;
 437                         if (tcd)
 438                                 cfs_trace_put_tcd(tcd);
 439                         return 1;
 440                 }
 441
 442                 if (cfs_time_after(cfs_time_current(),
 443                                    cdls->cdls_next + libcfs_console_max_delay +
 444                                    cfs_time_seconds(10))) {
 445                         /* last timeout was a long time ago */
 446                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 447                 } else {
 448                         cdls->cdls_delay *= libcfs_console_backoff;
 449                 }
 450
 451                 if (cdls->cdls_delay < libcfs_console_min_delay)
 452                         cdls->cdls_delay = libcfs_console_min_delay;
 453                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 454                         cdls->cdls_delay = libcfs_console_max_delay;
 455
 456                 /* ensure cdls_next is never zero after it's been seen */
 457                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 458         }
 459
 460         if (tcd) {
 461                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 462                                      msgdata->msg_fn);
 463                 cfs_trace_put_tcd(tcd);
 464         } else {
 465                 string_buf = cfs_trace_get_console_buffer();
 466
 467                 needed = 0;
 468                 if (format1) {
 469                         va_copy(ap, args);
 470                         needed = vsnprintf(string_buf,
 471                                            CFS_TRACE_CONSOLE_BUFFER_SIZE,
 472                                            format1, ap);
 473                         va_end(ap);
 474                 }
 475                 if (format2) {
 476                         remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
 477                         if (remain > 0) {
 478                                 va_start(ap, format2);
 479                                 needed += vsnprintf(string_buf + needed, remain,
 480                                                     format2, ap);
 481                                 va_end(ap);
 482                         }
 483                 }
 484                 cfs_print_to_console(&header, mask,
 485                                      string_buf, needed, file, msgdata->msg_fn);
 486
 487                 put_cpu();
 488         }
 489
 490         if (cdls && cdls->cdls_count) {
 491                 string_buf = cfs_trace_get_console_buffer();
 492
 493                 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 494                                   "Skipped %d previous similar message%s\n",
 495                                   cdls->cdls_count,
 496                                   (cdls->cdls_count > 1) ? "s" : "");
 497
 498                 cfs_print_to_console(&header, mask,
 499                                      string_buf, needed, file, msgdata->msg_fn);
 500
 501                 put_cpu();
 502                 cdls->cdls_count = 0;
 503         }
 504
 505         return 0;
 506 }
 507 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 508
 509 void
 510 cfs_trace_assertion_failed(const char *str,
 511                            struct libcfs_debug_msg_data *msgdata)
 512 {
 513         struct ptldebug_header hdr;
 514
 515         libcfs_panic_in_progress = 1;
 516         libcfs_catastrophe = 1;
 517         mb();
 518
 519         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 520
 521         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 522                              msgdata->msg_file, msgdata->msg_fn);
 523
 524         panic("Lustre debug assertion failure\n");
 525
 526         /* not reached */
 527 }
 528
 529 static void
 530 panic_collect_pages(struct page_collection *pc)
 531 {
 532         /* Do the collect_pages job on a single CPU: assumes that all other
 533          * CPUs have been stopped during a panic.  If this isn't true for some
 534          * arch, this will have to be implemented separately in each arch.
 535          */
 536         struct cfs_trace_cpu_data *tcd;
 537         int i;
 538         int j;
 539
 540         INIT_LIST_HEAD(&pc->pc_pages);
 541
 542         cfs_tcd_for_each(tcd, i, j) {
 543                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 544                 tcd->tcd_cur_pages = 0;
 545
 546                 if (pc->pc_want_daemon_pages) {
 547                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 548                         tcd->tcd_cur_daemon_pages = 0;
 549                 }
 550         }
 551 }
 552
 553 static void collect_pages_on_all_cpus(struct page_collection *pc)
 554 {
 555         struct cfs_trace_cpu_data *tcd;
 556         int i, cpu;
 557
 558         for_each_possible_cpu(cpu) {
 559                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 560                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 561                         tcd->tcd_cur_pages = 0;
 562                         if (pc->pc_want_daemon_pages) {
 563                                 list_splice_init(&tcd->tcd_daemon_pages,
 564                                                  &pc->pc_pages);
 565                                 tcd->tcd_cur_daemon_pages = 0;
 566                         }
 567                 }
 568         }
 569 }
 570
 571 static void collect_pages(struct page_collection *pc)
 572 {
 573         INIT_LIST_HEAD(&pc->pc_pages);
 574
 575         if (libcfs_panic_in_progress)
 576                 panic_collect_pages(pc);
 577         else
 578                 collect_pages_on_all_cpus(pc);
 579 }
 580
 581 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 582 {
 583         struct cfs_trace_cpu_data *tcd;
 584         struct list_head *cur_head;
 585         struct cfs_trace_page *tage;
 586         struct cfs_trace_page *tmp;
 587         int i, cpu;
 588
 589         for_each_possible_cpu(cpu) {
 590                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 591                         cur_head = tcd->tcd_pages.next;
 592
 593                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 594                                                  linkage) {
 595                                 __LASSERT_TAGE_INVARIANT(tage);
 596
 597                                 if (tage->cpu != cpu || tage->type != i)
 598                                         continue;
 599
 600                                 cfs_tage_to_tail(tage, cur_head);
 601                                 tcd->tcd_cur_pages++;
 602                         }
 603                 }
 604         }
 605 }
 606
 607 static void put_pages_back(struct page_collection *pc)
 608 {
 609         if (!libcfs_panic_in_progress)
 610                 put_pages_back_on_all_cpus(pc);
 611 }
 612
 613 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 614  * we have a good amount of data at all times for dumping during an LBUG, even
 615  * if we have been steadily writing (and otherwise discarding) pages via the
 616  * debug daemon.
 617  */
 618 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 619                                          struct cfs_trace_cpu_data *tcd)
 620 {
 621         struct cfs_trace_page *tage;
 622         struct cfs_trace_page *tmp;
 623
 624         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 625                 __LASSERT_TAGE_INVARIANT(tage);
 626
 627                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 628                         continue;
 629
 630                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 631                 tcd->tcd_cur_daemon_pages++;
 632
 633                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 634                         struct cfs_trace_page *victim;
 635
 636                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 637                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 638
 639                         __LASSERT_TAGE_INVARIANT(victim);
 640
 641                         list_del(&victim->linkage);
 642                         cfs_tage_free(victim);
 643                         tcd->tcd_cur_daemon_pages--;
 644                 }
 645         }
 646 }
 647
 648 static void put_pages_on_daemon_list(struct page_collection *pc)
 649 {
 650         struct cfs_trace_cpu_data *tcd;
 651         int i, cpu;
 652
 653         for_each_possible_cpu(cpu) {
 654                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 655                         put_pages_on_tcd_daemon_list(pc, tcd);
 656         }
 657 }
 658
 659 void cfs_trace_debug_print(void)
 660 {
 661         struct page_collection pc;
 662         struct cfs_trace_page *tage;
 663         struct cfs_trace_page *tmp;
 664
 665         pc.pc_want_daemon_pages = 1;
 666         collect_pages(&pc);
 667         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 668                 char *p, *file, *fn;
 669                 struct page *page;
 670
 671                 __LASSERT_TAGE_INVARIANT(tage);
 672
 673                 page = tage->page;
 674                 p = page_address(page);
 675                 while (p < ((char *)page_address(page) + tage->used)) {
 676                         struct ptldebug_header *hdr;
 677                         int len;
 678
 679                         hdr = (void *)p;
 680                         p += sizeof(*hdr);
 681                         file = p;
 682                         p += strlen(file) + 1;
 683                         fn = p;
 684                         p += strlen(fn) + 1;
 685                         len = hdr->ph_len - (int)(p - (char *)hdr);
 686
 687                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 688
 689                         p += len;
 690                 }
 691
 692                 list_del(&tage->linkage);
 693                 cfs_tage_free(tage);
 694         }
 695 }
 696
 697 int cfs_tracefile_dump_all_pages(char *filename)
 698 {
 699         struct page_collection pc;
 700         struct file *filp;
 701         struct cfs_trace_page *tage;
 702         struct cfs_trace_page *tmp;
 703         char *buf;
 704         mm_segment_t __oldfs;
 705         int rc;
 706
 707         cfs_tracefile_write_lock();
 708
 709         filp = filp_open(filename, O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE,
 710                          0600);
 711         if (IS_ERR(filp)) {
 712                 rc = PTR_ERR(filp);
 713                 filp = NULL;
 714                 pr_err("LustreError: can't open %s for dump: rc %d\n",
 715                        filename, rc);
 716                 goto out;
 717         }
 718
 719         pc.pc_want_daemon_pages = 1;
 720         collect_pages(&pc);
 721         if (list_empty(&pc.pc_pages)) {
 722                 rc = 0;
 723                 goto close;
 724         }
 725         __oldfs = get_fs();
 726         set_fs(get_ds());
 727
 728         /* ok, for now, just write the pages.  in the future we'll be building
 729          * iobufs with the pages and calling generic_direct_IO
 730          */
 731         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 732                 __LASSERT_TAGE_INVARIANT(tage);
 733
 734                 buf = kmap(tage->page);
 735                 rc = kernel_write(filp, buf, tage->used, &filp->f_pos);
 736                 kunmap(tage->page);
 737
 738                 if (rc != (int)tage->used) {
 739                         pr_warn("wanted to write %u but wrote %d\n", tage->used,
 740                                 rc);
 741                         put_pages_back(&pc);
 742                         __LASSERT(list_empty(&pc.pc_pages));
 743                         break;
 744                 }
 745                 list_del(&tage->linkage);
 746                 cfs_tage_free(tage);
 747         }
 748         set_fs(__oldfs);
 749         rc = vfs_fsync(filp, 1);
 750         if (rc)
 751                 pr_err("sync returns %d\n", rc);
 752 close:
 753         filp_close(filp, NULL);
 754 out:
 755         cfs_tracefile_write_unlock();
 756         return rc;
 757 }
 758
 759 void cfs_trace_flush_pages(void)
 760 {
 761         struct page_collection pc;
 762         struct cfs_trace_page *tage;
 763         struct cfs_trace_page *tmp;
 764
 765         pc.pc_want_daemon_pages = 1;
 766         collect_pages(&pc);
 767         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 768                 __LASSERT_TAGE_INVARIANT(tage);
 769
 770                 list_del(&tage->linkage);
 771                 cfs_tage_free(tage);
 772         }
 773 }
 774
 775 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 776                             const char __user *usr_buffer, int usr_buffer_nob)
 777 {
 778         int nob;
 779
 780         if (usr_buffer_nob > knl_buffer_nob)
 781                 return -EOVERFLOW;
 782
 783         if (copy_from_user((void *)knl_buffer,
 784                            usr_buffer, usr_buffer_nob))
 785                 return -EFAULT;
 786
 787         nob = strnlen(knl_buffer, usr_buffer_nob);
 788         while (--nob >= 0)                    /* strip trailing whitespace */
 789                 if (!isspace(knl_buffer[nob]))
 790                         break;
 791
 792         if (nob < 0)                        /* empty string */
 793                 return -EINVAL;
 794
 795         if (nob == knl_buffer_nob)            /* no space to terminate */
 796                 return -EOVERFLOW;
 797
 798         knl_buffer[nob + 1] = 0;                /* terminate */
 799         return 0;
 800 }
 801 EXPORT_SYMBOL(cfs_trace_copyin_string);
 802
 803 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 804                              const char *knl_buffer, char *append)
 805 {
 806         /*
 807          * NB if 'append' != NULL, it's a single character to append to the
 808          * copied out string - usually "\n" or "" (i.e. a terminating zero byte)
 809          */
 810         int nob = strlen(knl_buffer);
 811
 812         if (nob > usr_buffer_nob)
 813                 nob = usr_buffer_nob;
 814
 815         if (copy_to_user(usr_buffer, knl_buffer, nob))
 816                 return -EFAULT;
 817
 818         if (append && nob < usr_buffer_nob) {
 819                 if (copy_to_user(usr_buffer + nob, append, 1))
 820                         return -EFAULT;
 821
 822                 nob++;
 823         }
 824
 825         return nob;
 826 }
 827 EXPORT_SYMBOL(cfs_trace_copyout_string);
 828
 829 int cfs_trace_allocate_string_buffer(char **str, int nob)
 830 {
 831         if (nob > 2 * PAGE_SIZE)            /* string must be "sensible" */
 832                 return -EINVAL;
 833
 834         *str = kmalloc(nob, GFP_KERNEL | __GFP_ZERO);
 835         if (!*str)
 836                 return -ENOMEM;
 837
 838         return 0;
 839 }
 840
 841 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
 842 {
 843         char *str;
 844         int rc;
 845
 846         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 847         if (rc)
 848                 return rc;
 849
 850         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 851                                      usr_str, usr_str_nob);
 852         if (rc)
 853                 goto out;
 854
 855         if (str[0] != '/') {
 856                 rc = -EINVAL;
 857                 goto out;
 858         }
 859         rc = cfs_tracefile_dump_all_pages(str);
 860 out:
 861         kfree(str);
 862         return rc;
 863 }
 864
 865 int cfs_trace_daemon_command(char *str)
 866 {
 867         int rc = 0;
 868
 869         cfs_tracefile_write_lock();
 870
 871         if (!strcmp(str, "stop")) {
 872                 cfs_tracefile_write_unlock();
 873                 cfs_trace_stop_thread();
 874                 cfs_tracefile_write_lock();
 875                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 876
 877         } else if (!strncmp(str, "size=", 5)) {
 878                 unsigned long tmp;
 879
 880                 rc = kstrtoul(str + 5, 10, &tmp);
 881                 if (!rc) {
 882                         if (tmp < 10 || tmp > 20480)
 883                                 cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 884                         else
 885                                 cfs_tracefile_size = tmp << 20;
 886                 }
 887         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 888                 rc = -ENAMETOOLONG;
 889         } else if (str[0] != '/') {
 890                 rc = -EINVAL;
 891         } else {
 892                 strcpy(cfs_tracefile, str);
 893
 894                 pr_info("debug daemon will attempt to start writing to %s (%lukB max)\n",
 895                         cfs_tracefile,
 896                         (long)(cfs_tracefile_size >> 10));
 897
 898                 cfs_trace_start_thread();
 899         }
 900
 901         cfs_tracefile_write_unlock();
 902         return rc;
 903 }
 904
 905 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
 906 {
 907         char *str;
 908         int rc;
 909
 910         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 911         if (rc)
 912                 return rc;
 913
 914         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 915                                      usr_str, usr_str_nob);
 916         if (!rc)
 917                 rc = cfs_trace_daemon_command(str);
 918
 919         kfree(str);
 920         return rc;
 921 }
 922
 923 int cfs_trace_set_debug_mb(int mb)
 924 {
 925         int i;
 926         int j;
 927         int pages;
 928         int limit = cfs_trace_max_debug_mb();
 929         struct cfs_trace_cpu_data *tcd;
 930
 931         if (mb < num_possible_cpus()) {
 932                 pr_warn("%d MB is too small for debug buffer size, setting it to %d MB.\n",
 933                         mb, num_possible_cpus());
 934                 mb = num_possible_cpus();
 935         }
 936
 937         if (mb > limit) {
 938                 pr_warn("%d MB is too large for debug buffer size, setting it to %d MB.\n",
 939                         mb, limit);
 940                 mb = limit;
 941         }
 942
 943         mb /= num_possible_cpus();
 944         pages = mb << (20 - PAGE_SHIFT);
 945
 946         cfs_tracefile_write_lock();
 947
 948         cfs_tcd_for_each(tcd, i, j)
 949                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 950
 951         cfs_tracefile_write_unlock();
 952
 953         return 0;
 954 }
 955
 956 int cfs_trace_get_debug_mb(void)
 957 {
 958         int i;
 959         int j;
 960         struct cfs_trace_cpu_data *tcd;
 961         int total_pages = 0;
 962
 963         cfs_tracefile_read_lock();
 964
 965         cfs_tcd_for_each(tcd, i, j)
 966                 total_pages += tcd->tcd_max_pages;
 967
 968         cfs_tracefile_read_unlock();
 969
 970         return (total_pages >> (20 - PAGE_SHIFT)) + 1;
 971 }
 972
 973 static int tracefiled(void *arg)
 974 {
 975         struct page_collection pc;
 976         struct tracefiled_ctl *tctl = arg;
 977         struct cfs_trace_page *tage;
 978         struct cfs_trace_page *tmp;
 979         struct file *filp;
 980         char *buf;
 981         int last_loop = 0;
 982         int rc;
 983
 984         /* we're started late enough that we pick up init's fs context */
 985         /* this is so broken in uml?  what on earth is going on? */
 986
 987         complete(&tctl->tctl_start);
 988
 989         while (1) {
 990                 wait_queue_entry_t __wait;
 991
 992                 pc.pc_want_daemon_pages = 0;
 993                 collect_pages(&pc);
 994                 if (list_empty(&pc.pc_pages))
 995                         goto end_loop;
 996
 997                 filp = NULL;
 998                 cfs_tracefile_read_lock();
 999                 if (cfs_tracefile[0]) {
1000                         filp = filp_open(cfs_tracefile,
1001                                          O_CREAT | O_RDWR | O_LARGEFILE,
1002                                          0600);
1003                         if (IS_ERR(filp)) {
1004                                 rc = PTR_ERR(filp);
1005                                 filp = NULL;
1006                                 pr_warn("couldn't open %s: %d\n", cfs_tracefile,
1007                                         rc);
1008                         }
1009                 }
1010                 cfs_tracefile_read_unlock();
1011                 if (!filp) {
1012                         put_pages_on_daemon_list(&pc);
1013                         __LASSERT(list_empty(&pc.pc_pages));
1014                         goto end_loop;
1015                 }
1016
1017                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
1018                         static loff_t f_pos;
1019
1020                         __LASSERT_TAGE_INVARIANT(tage);
1021
1022                         if (f_pos >= (off_t)cfs_tracefile_size)
1023                                 f_pos = 0;
1024                         else if (f_pos > i_size_read(file_inode(filp)))
1025                                 f_pos = i_size_read(file_inode(filp));
1026
1027                         buf = kmap(tage->page);
1028                         rc = kernel_write(filp, buf, tage->used, &f_pos);
1029                         kunmap(tage->page);
1030
1031                         if (rc != (int)tage->used) {
1032                                 pr_warn("wanted to write %u but wrote %d\n",
1033                                         tage->used, rc);
1034                                 put_pages_back(&pc);
1035                                 __LASSERT(list_empty(&pc.pc_pages));
1036                                 break;
1037                         }
1038                 }
1039
1040                 filp_close(filp, NULL);
1041                 put_pages_on_daemon_list(&pc);
1042                 if (!list_empty(&pc.pc_pages)) {
1043                         int i;
1044
1045                         pr_alert("trace pages aren't empty\n");
1046                         pr_err("total cpus(%d): ", num_possible_cpus());
1047                         for (i = 0; i < num_possible_cpus(); i++)
1048                                 if (cpu_online(i))
1049                                         pr_cont("%d(on) ", i);
1050                                 else
1051                                         pr_cont("%d(off) ", i);
1052                         pr_cont("\n");
1053
1054                         i = 0;
1055                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1056                                                  linkage)
1057                                 pr_err("page %d belongs to cpu %d\n",
1058                                        ++i, tage->cpu);
1059                         pr_err("There are %d pages unwritten\n", i);
1060                 }
1061                 __LASSERT(list_empty(&pc.pc_pages));
1062 end_loop:
1063                 if (atomic_read(&tctl->tctl_shutdown)) {
1064                         if (!last_loop) {
1065                                 last_loop = 1;
1066                                 continue;
1067                         } else {
1068                                 break;
1069                         }
1070                 }
1071                 init_waitqueue_entry(&__wait, current);
1072                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1073                 set_current_state(TASK_INTERRUPTIBLE);
1074                 schedule_timeout(cfs_time_seconds(1));
1075                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1076         }
1077         complete(&tctl->tctl_stop);
1078         return 0;
1079 }
1080
1081 int cfs_trace_start_thread(void)
1082 {
1083         struct tracefiled_ctl *tctl = &trace_tctl;
1084         struct task_struct *task;
1085         int rc = 0;
1086
1087         mutex_lock(&cfs_trace_thread_mutex);
1088         if (thread_running)
1089                 goto out;
1090
1091         init_completion(&tctl->tctl_start);
1092         init_completion(&tctl->tctl_stop);
1093         init_waitqueue_head(&tctl->tctl_waitq);
1094         atomic_set(&tctl->tctl_shutdown, 0);
1095
1096         task = kthread_run(tracefiled, tctl, "ktracefiled");
1097         if (IS_ERR(task)) {
1098                 rc = PTR_ERR(task);
1099                 goto out;
1100         }
1101
1102         wait_for_completion(&tctl->tctl_start);
1103         thread_running = 1;
1104 out:
1105         mutex_unlock(&cfs_trace_thread_mutex);
1106         return rc;
1107 }
1108
1109 void cfs_trace_stop_thread(void)
1110 {
1111         struct tracefiled_ctl *tctl = &trace_tctl;
1112
1113         mutex_lock(&cfs_trace_thread_mutex);
1114         if (thread_running) {
1115                 pr_info("shutting down debug daemon thread...\n");
1116                 atomic_set(&tctl->tctl_shutdown, 1);
1117                 wait_for_completion(&tctl->tctl_stop);
1118                 thread_running = 0;
1119         }
1120         mutex_unlock(&cfs_trace_thread_mutex);
1121 }
1122
1123 int cfs_tracefile_init(int max_pages)
1124 {
1125         struct cfs_trace_cpu_data *tcd;
1126         int i;
1127         int j;
1128         int rc;
1129         int factor;
1130
1131         rc = cfs_tracefile_init_arch();
1132         if (rc)
1133                 return rc;
1134
1135         cfs_tcd_for_each(tcd, i, j) {
1136                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1137                 factor = tcd->tcd_pages_factor;
1138                 INIT_LIST_HEAD(&tcd->tcd_pages);
1139                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1140                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1141                 tcd->tcd_cur_pages = 0;
1142                 tcd->tcd_cur_stock_pages = 0;
1143                 tcd->tcd_cur_daemon_pages = 0;
1144                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1145                 LASSERT(tcd->tcd_max_pages > 0);
1146                 tcd->tcd_shutting_down = 0;
1147         }
1148
1149         return 0;
1150 }
1151
1152 static void trace_cleanup_on_all_cpus(void)
1153 {
1154         struct cfs_trace_cpu_data *tcd;
1155         struct cfs_trace_page *tage;
1156         struct cfs_trace_page *tmp;
1157         int i, cpu;
1158
1159         for_each_possible_cpu(cpu) {
1160                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1161                         tcd->tcd_shutting_down = 1;
1162
1163                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages,
1164                                                  linkage) {
1165                                 __LASSERT_TAGE_INVARIANT(tage);
1166
1167                                 list_del(&tage->linkage);
1168                                 cfs_tage_free(tage);
1169                         }
1170
1171                         tcd->tcd_cur_pages = 0;
1172                 }
1173         }
1174 }
1175
1176 static void cfs_trace_cleanup(void)
1177 {
1178         struct page_collection pc;
1179
1180         INIT_LIST_HEAD(&pc.pc_pages);
1181
1182         trace_cleanup_on_all_cpus();
1183
1184         cfs_tracefile_fini_arch();
1185 }
1186
1187 void cfs_tracefile_exit(void)
1188 {
1189         cfs_trace_stop_thread();
1190         cfs_trace_cleanup();
1191 }