libcfs/libcfs/tracefile.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * libcfs/libcfs/tracefile.c
  37  *
  38  * Author: Zach Brown <zab@clusterfs.com>
  39  * Author: Phil Schwan <phil@clusterfs.com>
  40  */
  41
  42
  43 #define DEBUG_SUBSYSTEM S_LNET
  44 #define LUSTRE_TRACEFILE_PRIVATE
  45 #include "tracefile.h"
  46
  47 #include <libcfs/libcfs.h>
  48
  49 /* XXX move things up to the top, comment */
  50 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  51
  52 char cfs_tracefile[TRACEFILE_NAME_SIZE];
  53 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
  54 static struct tracefiled_ctl trace_tctl;
  55 static DEFINE_MUTEX(cfs_trace_thread_mutex);
  56 static int thread_running = 0;
  57
  58 static atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
  59
  60 #define filp_size(f)    (i_size_read((f)->f_dentry->d_inode))
  61
  62 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  63                                         struct cfs_trace_cpu_data *tcd);
  64
  65 static inline struct cfs_trace_page *
  66 cfs_tage_from_list(struct list_head *list)
  67 {
  68         return list_entry(list, struct cfs_trace_page, linkage);
  69 }
  70
  71 static struct cfs_trace_page *cfs_tage_alloc(gfp_t gfp)
  72 {
  73         struct page            *page;
  74         struct cfs_trace_page *tage;
  75
  76         /* My caller is trying to free memory */
  77         if (!in_interrupt() && memory_pressure_get())
  78                 return NULL;
  79
  80         /*
  81          * Don't spam console with allocation failures: they will be reported
  82          * by upper layer anyway.
  83          */
  84         gfp |= __GFP_NOWARN;
  85         page = alloc_page(gfp);
  86         if (page == NULL)
  87                 return NULL;
  88
  89         tage = kmalloc(sizeof(*tage), gfp);
  90         if (tage == NULL) {
  91                 __free_page(page);
  92                 return NULL;
  93         }
  94
  95         tage->page = page;
  96         atomic_inc(&cfs_tage_allocated);
  97         return tage;
  98 }
  99
 100 static void cfs_tage_free(struct cfs_trace_page *tage)
 101 {
 102         __LASSERT(tage != NULL);
 103         __LASSERT(tage->page != NULL);
 104
 105         __free_page(tage->page);
 106         kfree(tage);
 107         atomic_dec(&cfs_tage_allocated);
 108 }
 109
 110 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
 111                              struct list_head *queue)
 112 {
 113         __LASSERT(tage != NULL);
 114         __LASSERT(queue != NULL);
 115
 116         list_move_tail(&tage->linkage, queue);
 117 }
 118
 119 int cfs_trace_refill_stock(struct cfs_trace_cpu_data *tcd, gfp_t gfp,
 120                            struct list_head *stock)
 121 {
 122         int i;
 123
 124         /*
 125          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 126          * from here: this will lead to infinite recursion.
 127          */
 128
 129         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++ i) {
 130                 struct cfs_trace_page *tage;
 131
 132                 tage = cfs_tage_alloc(gfp);
 133                 if (tage == NULL)
 134                         break;
 135                 list_add_tail(&tage->linkage, stock);
 136         }
 137         return i;
 138 }
 139
 140 /* return a page that has 'len' bytes left at the end */
 141 static struct cfs_trace_page *
 142 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
 143 {
 144         struct cfs_trace_page *tage;
 145
 146         if (tcd->tcd_cur_pages > 0) {
 147                 __LASSERT(!list_empty(&tcd->tcd_pages));
 148                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
 149                 if (tage->used + len <= PAGE_CACHE_SIZE)
 150                         return tage;
 151         }
 152
 153         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 154                 if (tcd->tcd_cur_stock_pages > 0) {
 155                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
 156                         --tcd->tcd_cur_stock_pages;
 157                         list_del_init(&tage->linkage);
 158                 } else {
 159                         tage = cfs_tage_alloc(GFP_ATOMIC);
 160                         if (unlikely(tage == NULL)) {
 161                                 if ((!memory_pressure_get() ||
 162                                      in_interrupt()) && printk_ratelimit())
 163                                         printk(KERN_WARNING
 164                                                "cannot allocate a tage (%ld)\n",
 165                                                tcd->tcd_cur_pages);
 166                                 return NULL;
 167                         }
 168                 }
 169
 170                 tage->used = 0;
 171                 tage->cpu = smp_processor_id();
 172                 tage->type = tcd->tcd_type;
 173                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 174                 tcd->tcd_cur_pages++;
 175
 176                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 177                         struct tracefiled_ctl *tctl = &trace_tctl;
 178                         /*
 179                          * wake up tracefiled to process some pages.
 180                          */
 181                         wake_up(&tctl->tctl_waitq);
 182                 }
 183                 return tage;
 184         }
 185         return NULL;
 186 }
 187
 188 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
 189 {
 190         int pgcount = tcd->tcd_cur_pages / 10;
 191         struct page_collection pc;
 192         struct cfs_trace_page *tage;
 193         struct cfs_trace_page *tmp;
 194
 195         /*
 196          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 197          * from here: this will lead to infinite recursion.
 198          */
 199
 200         if (printk_ratelimit())
 201                 printk(KERN_WARNING "debug daemon buffer overflowed; "
 202                         "discarding 10%% of pages (%d of %ld)\n",
 203                         pgcount + 1, tcd->tcd_cur_pages);
 204
 205         INIT_LIST_HEAD(&pc.pc_pages);
 206
 207         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 208                 if (pgcount-- == 0)
 209                         break;
 210
 211                 list_move_tail(&tage->linkage, &pc.pc_pages);
 212                 tcd->tcd_cur_pages--;
 213         }
 214         put_pages_on_tcd_daemon_list(&pc, tcd);
 215 }
 216
 217 /* return a page that has 'len' bytes left at the end */
 218 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 219                                                  unsigned long len)
 220 {
 221         struct cfs_trace_page *tage;
 222
 223         /*
 224          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 225          * from here: this will lead to infinite recursion.
 226          */
 227
 228         if (len > PAGE_CACHE_SIZE) {
 229                 printk(KERN_ERR
 230                        "cowardly refusing to write %lu bytes in a page\n", len);
 231                 return NULL;
 232         }
 233
 234         tage = cfs_trace_get_tage_try(tcd, len);
 235         if (tage != NULL)
 236                 return tage;
 237         if (thread_running)
 238                 cfs_tcd_shrink(tcd);
 239         if (tcd->tcd_cur_pages > 0) {
 240                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 241                 tage->used = 0;
 242                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 243         }
 244         return tage;
 245 }
 246
 247 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 248                      const char *format, ...)
 249 {
 250         va_list args;
 251         int     rc;
 252
 253         va_start(args, format);
 254         rc = libcfs_debug_vmsg2(msgdata, format, args, NULL);
 255         va_end(args);
 256
 257         return rc;
 258 }
 259 EXPORT_SYMBOL(libcfs_debug_msg);
 260
 261 int libcfs_debug_vmsg2(struct libcfs_debug_msg_data *msgdata,
 262                        const char *format1, va_list args,
 263                        const char *format2, ...)
 264 {
 265         struct cfs_trace_cpu_data *tcd = NULL;
 266         struct ptldebug_header     header = {0};
 267         struct cfs_trace_page     *tage;
 268         /* string_buf is used only if tcd != NULL, and is always set then */
 269         char                      *string_buf = NULL;
 270         char                      *debug_buf;
 271         int                        known_size;
 272         int                        needed = 85; /* average message length */
 273         int                        max_nob;
 274         va_list                    ap;
 275         int                        i;
 276         int                        remain;
 277         int                        mask = msgdata->msg_mask;
 278         char                      *file = (char *)msgdata->msg_file;
 279         cfs_debug_limit_state_t   *cdls = msgdata->msg_cdls;
 280
 281         if (strchr(file, '/'))
 282                 file = strrchr(file, '/') + 1;
 283
 284         tcd = cfs_trace_get_tcd();
 285
 286         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 287          * pins us to a particular CPU.  This avoids an smp_processor_id()
 288          * warning on Linux when debugging is enabled. */
 289         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 290
 291         if (tcd == NULL)                /* arch may not log in IRQ context */
 292                 goto console;
 293
 294         if (tcd->tcd_cur_pages == 0)
 295                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 296
 297         if (tcd->tcd_shutting_down) {
 298                 cfs_trace_put_tcd(tcd);
 299                 tcd = NULL;
 300                 goto console;
 301         }
 302
 303         known_size = strlen(file) + 1;
 304         if (msgdata->msg_fn)
 305                 known_size += strlen(msgdata->msg_fn) + 1;
 306
 307         if (libcfs_debug_binary)
 308                 known_size += sizeof(header);
 309
 310         /*/
 311          * '2' used because vsnprintf return real size required for output
 312          * _without_ terminating NULL.
 313          * if needed is to small for this format.
 314          */
 315         for (i = 0; i < 2; i++) {
 316                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 317                 if (tage == NULL) {
 318                         if (needed + known_size > PAGE_CACHE_SIZE)
 319                                 mask |= D_ERROR;
 320
 321                         cfs_trace_put_tcd(tcd);
 322                         tcd = NULL;
 323                         goto console;
 324                 }
 325
 326                 string_buf = (char *)page_address(tage->page) +
 327                                         tage->used + known_size;
 328
 329                 max_nob = PAGE_CACHE_SIZE - tage->used - known_size;
 330                 if (max_nob <= 0) {
 331                         printk(KERN_EMERG "negative max_nob: %d\n",
 332                                max_nob);
 333                         mask |= D_ERROR;
 334                         cfs_trace_put_tcd(tcd);
 335                         tcd = NULL;
 336                         goto console;
 337                 }
 338
 339                 needed = 0;
 340                 if (format1) {
 341                         va_copy(ap, args);
 342                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 343                         va_end(ap);
 344                 }
 345
 346                 if (format2) {
 347                         remain = max_nob - needed;
 348                         if (remain < 0)
 349                                 remain = 0;
 350
 351                         va_start(ap, format2);
 352                         needed += vsnprintf(string_buf + needed, remain,
 353                                             format2, ap);
 354                         va_end(ap);
 355                 }
 356
 357                 if (needed < max_nob) /* well. printing ok.. */
 358                         break;
 359         }
 360
 361         if (*(string_buf+needed-1) != '\n')
 362                 printk(KERN_INFO "format at %s:%d:%s doesn't end in "
 363                        "newline\n", file, msgdata->msg_line, msgdata->msg_fn);
 364
 365         header.ph_len = known_size + needed;
 366         debug_buf = (char *)page_address(tage->page) + tage->used;
 367
 368         if (libcfs_debug_binary) {
 369                 memcpy(debug_buf, &header, sizeof(header));
 370                 tage->used += sizeof(header);
 371                 debug_buf += sizeof(header);
 372         }
 373
 374         strcpy(debug_buf, file);
 375         tage->used += strlen(file) + 1;
 376         debug_buf += strlen(file) + 1;
 377
 378         if (msgdata->msg_fn) {
 379                 strcpy(debug_buf, msgdata->msg_fn);
 380                 tage->used += strlen(msgdata->msg_fn) + 1;
 381                 debug_buf += strlen(msgdata->msg_fn) + 1;
 382         }
 383
 384         __LASSERT(debug_buf == string_buf);
 385
 386         tage->used += needed;
 387         __LASSERT(tage->used <= PAGE_CACHE_SIZE);
 388
 389 console:
 390         if ((mask & libcfs_printk) == 0) {
 391                 /* no console output requested */
 392                 if (tcd != NULL)
 393                         cfs_trace_put_tcd(tcd);
 394                 return 1;
 395         }
 396
 397         if (cdls != NULL) {
 398                 if (libcfs_console_ratelimit &&
 399                     cdls->cdls_next != 0 &&     /* not first time ever */
 400                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 401                         /* skipping a console message */
 402                         cdls->cdls_count++;
 403                         if (tcd != NULL)
 404                                 cfs_trace_put_tcd(tcd);
 405                         return 1;
 406                 }
 407
 408                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 409                                                        libcfs_console_max_delay
 410                                                        + cfs_time_seconds(10))) {
 411                         /* last timeout was a long time ago */
 412                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 413                 } else {
 414                         cdls->cdls_delay *= libcfs_console_backoff;
 415                 }
 416
 417                 if (cdls->cdls_delay < libcfs_console_min_delay)
 418                         cdls->cdls_delay = libcfs_console_min_delay;
 419                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 420                         cdls->cdls_delay = libcfs_console_max_delay;
 421
 422                 /* ensure cdls_next is never zero after it's been seen */
 423                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 424         }
 425
 426         if (tcd != NULL) {
 427                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 428                                      msgdata->msg_fn);
 429                 cfs_trace_put_tcd(tcd);
 430         } else {
 431                 string_buf = cfs_trace_get_console_buffer();
 432
 433                 needed = 0;
 434                 if (format1 != NULL) {
 435                         va_copy(ap, args);
 436                         needed = vsnprintf(string_buf,
 437                                            CFS_TRACE_CONSOLE_BUFFER_SIZE,
 438                                            format1, ap);
 439                         va_end(ap);
 440                 }
 441                 if (format2 != NULL) {
 442                         remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
 443                         if (remain > 0) {
 444                                 va_start(ap, format2);
 445                                 needed += vsnprintf(string_buf+needed, remain,
 446                                                     format2, ap);
 447                                 va_end(ap);
 448                         }
 449                 }
 450                 cfs_print_to_console(&header, mask,
 451                                      string_buf, needed, file, msgdata->msg_fn);
 452
 453                 cfs_trace_put_console_buffer(string_buf);
 454         }
 455
 456         if (cdls != NULL && cdls->cdls_count != 0) {
 457                 string_buf = cfs_trace_get_console_buffer();
 458
 459                 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 460                                   "Skipped %d previous similar message%s\n",
 461                                   cdls->cdls_count,
 462                                   (cdls->cdls_count > 1) ? "s" : "");
 463
 464                 cfs_print_to_console(&header, mask,
 465                                      string_buf, needed, file, msgdata->msg_fn);
 466
 467                 cfs_trace_put_console_buffer(string_buf);
 468                 cdls->cdls_count = 0;
 469         }
 470
 471         return 0;
 472 }
 473 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 474
 475 void
 476 cfs_trace_assertion_failed(const char *str,
 477                            struct libcfs_debug_msg_data *msgdata)
 478 {
 479         struct ptldebug_header hdr;
 480
 481         libcfs_panic_in_progress = 1;
 482         libcfs_catastrophe = 1;
 483         smp_mb();
 484
 485         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 486
 487         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 488                              msgdata->msg_file, msgdata->msg_fn);
 489
 490         panic("Lustre debug assertion failure\n");
 491
 492         /* not reached */
 493 }
 494
 495 static void
 496 panic_collect_pages(struct page_collection *pc)
 497 {
 498         /* Do the collect_pages job on a single CPU: assumes that all other
 499          * CPUs have been stopped during a panic.  If this isn't true for some
 500          * arch, this will have to be implemented separately in each arch.  */
 501         int                        i;
 502         int                        j;
 503         struct cfs_trace_cpu_data *tcd;
 504
 505         INIT_LIST_HEAD(&pc->pc_pages);
 506
 507         cfs_tcd_for_each(tcd, i, j) {
 508                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 509                 tcd->tcd_cur_pages = 0;
 510
 511                 if (pc->pc_want_daemon_pages) {
 512                         list_splice_init(&tcd->tcd_daemon_pages,
 513                                                 &pc->pc_pages);
 514                         tcd->tcd_cur_daemon_pages = 0;
 515                 }
 516         }
 517 }
 518
 519 static void collect_pages_on_all_cpus(struct page_collection *pc)
 520 {
 521         struct cfs_trace_cpu_data *tcd;
 522         int i, cpu;
 523
 524         for_each_possible_cpu(cpu) {
 525                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 526                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 527                         tcd->tcd_cur_pages = 0;
 528                         if (pc->pc_want_daemon_pages) {
 529                                 list_splice_init(&tcd->tcd_daemon_pages,
 530                                                         &pc->pc_pages);
 531                                 tcd->tcd_cur_daemon_pages = 0;
 532                         }
 533                 }
 534         }
 535 }
 536
 537 static void collect_pages(struct page_collection *pc)
 538 {
 539         INIT_LIST_HEAD(&pc->pc_pages);
 540
 541         if (libcfs_panic_in_progress)
 542                 panic_collect_pages(pc);
 543         else
 544                 collect_pages_on_all_cpus(pc);
 545 }
 546
 547 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 548 {
 549         struct cfs_trace_cpu_data *tcd;
 550         struct list_head *cur_head;
 551         struct cfs_trace_page *tage;
 552         struct cfs_trace_page *tmp;
 553         int i, cpu;
 554
 555         for_each_possible_cpu(cpu) {
 556                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 557                         cur_head = tcd->tcd_pages.next;
 558
 559                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 560                                                  linkage) {
 561
 562                                 __LASSERT_TAGE_INVARIANT(tage);
 563
 564                                 if (tage->cpu != cpu || tage->type != i)
 565                                         continue;
 566
 567                                 cfs_tage_to_tail(tage, cur_head);
 568                                 tcd->tcd_cur_pages++;
 569                         }
 570                 }
 571         }
 572 }
 573
 574 static void put_pages_back(struct page_collection *pc)
 575 {
 576         if (!libcfs_panic_in_progress)
 577                 put_pages_back_on_all_cpus(pc);
 578 }
 579
 580 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 581  * we have a good amount of data at all times for dumping during an LBUG, even
 582  * if we have been steadily writing (and otherwise discarding) pages via the
 583  * debug daemon. */
 584 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 585                                          struct cfs_trace_cpu_data *tcd)
 586 {
 587         struct cfs_trace_page *tage;
 588         struct cfs_trace_page *tmp;
 589
 590         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 591                 __LASSERT_TAGE_INVARIANT(tage);
 592
 593                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 594                         continue;
 595
 596                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 597                 tcd->tcd_cur_daemon_pages++;
 598
 599                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 600                         struct cfs_trace_page *victim;
 601
 602                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 603                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 604
 605                         __LASSERT_TAGE_INVARIANT(victim);
 606
 607                         list_del(&victim->linkage);
 608                         cfs_tage_free(victim);
 609                         tcd->tcd_cur_daemon_pages--;
 610                 }
 611         }
 612 }
 613
 614 static void put_pages_on_daemon_list(struct page_collection *pc)
 615 {
 616         struct cfs_trace_cpu_data *tcd;
 617         int i, cpu;
 618
 619         for_each_possible_cpu(cpu) {
 620                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 621                         put_pages_on_tcd_daemon_list(pc, tcd);
 622         }
 623 }
 624
 625 void cfs_trace_debug_print(void)
 626 {
 627         struct page_collection pc;
 628         struct cfs_trace_page *tage;
 629         struct cfs_trace_page *tmp;
 630
 631         pc.pc_want_daemon_pages = 1;
 632         collect_pages(&pc);
 633         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 634                 char *p, *file, *fn;
 635                 struct page *page;
 636
 637                 __LASSERT_TAGE_INVARIANT(tage);
 638
 639                 page = tage->page;
 640                 p = page_address(page);
 641                 while (p < ((char *)page_address(page) + tage->used)) {
 642                         struct ptldebug_header *hdr;
 643                         int len;
 644                         hdr = (void *)p;
 645                         p += sizeof(*hdr);
 646                         file = p;
 647                         p += strlen(file) + 1;
 648                         fn = p;
 649                         p += strlen(fn) + 1;
 650                         len = hdr->ph_len - (int)(p - (char *)hdr);
 651
 652                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 653
 654                         p += len;
 655                 }
 656
 657                 list_del(&tage->linkage);
 658                 cfs_tage_free(tage);
 659         }
 660 }
 661
 662 int cfs_tracefile_dump_all_pages(char *filename)
 663 {
 664         struct page_collection  pc;
 665         struct file             *filp;
 666         struct cfs_trace_page   *tage;
 667         struct cfs_trace_page   *tmp;
 668         mm_segment_t            __oldfs;
 669         int rc;
 670
 671         cfs_tracefile_write_lock();
 672
 673         filp = filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600);
 674         if (IS_ERR(filp)) {
 675                 rc = PTR_ERR(filp);
 676                 filp = NULL;
 677                 printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 678                       filename, rc);
 679                 goto out;
 680         }
 681
 682         pc.pc_want_daemon_pages = 1;
 683         collect_pages(&pc);
 684         if (list_empty(&pc.pc_pages)) {
 685                 rc = 0;
 686                 goto close;
 687         }
 688         __oldfs = get_fs();
 689         set_fs(get_ds());
 690
 691         /* ok, for now, just write the pages.  in the future we'll be building
 692          * iobufs with the pages and calling generic_direct_IO */
 693         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 694
 695                 __LASSERT_TAGE_INVARIANT(tage);
 696
 697                 rc = vfs_write(filp, page_address(tage->page), tage->used,
 698                                &filp->f_pos);
 699                 if (rc != (int)tage->used) {
 700                         printk(KERN_WARNING "wanted to write %u but wrote "
 701                                "%d\n", tage->used, rc);
 702                         put_pages_back(&pc);
 703                         __LASSERT(list_empty(&pc.pc_pages));
 704                         break;
 705                 }
 706                 list_del(&tage->linkage);
 707                 cfs_tage_free(tage);
 708         }
 709         set_fs(__oldfs);
 710         rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1);
 711         if (rc)
 712                 printk(KERN_ERR "sync returns %d\n", rc);
 713 close:
 714         filp_close(filp, NULL);
 715 out:
 716         cfs_tracefile_write_unlock();
 717         return rc;
 718 }
 719
 720 void cfs_trace_flush_pages(void)
 721 {
 722         struct page_collection pc;
 723         struct cfs_trace_page *tage;
 724         struct cfs_trace_page *tmp;
 725
 726         pc.pc_want_daemon_pages = 1;
 727         collect_pages(&pc);
 728         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 729
 730                 __LASSERT_TAGE_INVARIANT(tage);
 731
 732                 list_del(&tage->linkage);
 733                 cfs_tage_free(tage);
 734         }
 735 }
 736
 737 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 738                             const char __user *usr_buffer, int usr_buffer_nob)
 739 {
 740         int    nob;
 741
 742         if (usr_buffer_nob > knl_buffer_nob)
 743                 return -EOVERFLOW;
 744
 745         if (copy_from_user(knl_buffer, usr_buffer, usr_buffer_nob))
 746                 return -EFAULT;
 747
 748         nob = strnlen(knl_buffer, usr_buffer_nob);
 749         while (nob-- >= 0)                      /* strip trailing whitespace */
 750                 if (!isspace(knl_buffer[nob]))
 751                         break;
 752
 753         if (nob < 0)                            /* empty string */
 754                 return -EINVAL;
 755
 756         if (nob == knl_buffer_nob)              /* no space to terminate */
 757                 return -EOVERFLOW;
 758
 759         knl_buffer[nob + 1] = 0;                /* terminate */
 760         return 0;
 761 }
 762 EXPORT_SYMBOL(cfs_trace_copyin_string);
 763
 764 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 765                              const char *knl_buffer, char *append)
 766 {
 767         /* NB if 'append' != NULL, it's a single character to append to the
 768          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 769          * terminating zero byte) for sysctl entries */
 770         int   nob = strlen(knl_buffer);
 771
 772         if (nob > usr_buffer_nob)
 773                 nob = usr_buffer_nob;
 774
 775         if (copy_to_user(usr_buffer, knl_buffer, nob))
 776                 return -EFAULT;
 777
 778         if (append != NULL && nob < usr_buffer_nob) {
 779                 if (copy_to_user(usr_buffer + nob, append, 1))
 780                         return -EFAULT;
 781
 782                 nob++;
 783         }
 784
 785         return nob;
 786 }
 787 EXPORT_SYMBOL(cfs_trace_copyout_string);
 788
 789 int cfs_trace_allocate_string_buffer(char **str, int nob)
 790 {
 791         if (nob > 2 * PAGE_CACHE_SIZE)  /* string must be "sensible" */
 792                 return -EINVAL;
 793
 794         *str = kmalloc(nob, GFP_IOFS | __GFP_ZERO);
 795         if (*str == NULL)
 796                 return -ENOMEM;
 797
 798         return 0;
 799 }
 800
 801 void cfs_trace_free_string_buffer(char *str, int nob)
 802 {
 803         kfree(str);
 804 }
 805
 806 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
 807 {
 808         char         *str;
 809         int           rc;
 810
 811         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 812         if (rc != 0)
 813                 return rc;
 814
 815         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 816                                      usr_str, usr_str_nob);
 817         if (rc != 0)
 818                 goto out;
 819
 820         if (str[0] != '/') {
 821                 rc = -EINVAL;
 822                 goto out;
 823         }
 824         rc = cfs_tracefile_dump_all_pages(str);
 825 out:
 826         cfs_trace_free_string_buffer(str, usr_str_nob + 1);
 827         return rc;
 828 }
 829
 830 int cfs_trace_daemon_command(char *str)
 831 {
 832         int       rc = 0;
 833
 834         cfs_tracefile_write_lock();
 835
 836         if (strcmp(str, "stop") == 0) {
 837                 cfs_tracefile_write_unlock();
 838                 cfs_trace_stop_thread();
 839                 cfs_tracefile_write_lock();
 840                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 841
 842         } else if (strncmp(str, "size=", 5) == 0) {
 843                 cfs_tracefile_size = simple_strtoul(str + 5, NULL, 0);
 844                 if (cfs_tracefile_size < 10 || cfs_tracefile_size > 20480)
 845                         cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 846                 else
 847                         cfs_tracefile_size <<= 20;
 848
 849         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 850                 rc = -ENAMETOOLONG;
 851         } else if (str[0] != '/') {
 852                 rc = -EINVAL;
 853         } else {
 854                 strcpy(cfs_tracefile, str);
 855
 856                 printk(KERN_INFO
 857                        "Lustre: debug daemon will attempt to start writing "
 858                        "to %s (%lukB max)\n", cfs_tracefile,
 859                        (long)(cfs_tracefile_size >> 10));
 860
 861                 cfs_trace_start_thread();
 862         }
 863
 864         cfs_tracefile_write_unlock();
 865         return rc;
 866 }
 867
 868 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
 869 {
 870         char *str;
 871         int   rc;
 872
 873         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 874         if (rc != 0)
 875                 return rc;
 876
 877         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 878                                  usr_str, usr_str_nob);
 879         if (rc == 0)
 880                 rc = cfs_trace_daemon_command(str);
 881
 882         cfs_trace_free_string_buffer(str, usr_str_nob + 1);
 883         return rc;
 884 }
 885
 886 int cfs_trace_set_debug_mb(int mb)
 887 {
 888         int i;
 889         int j;
 890         int pages;
 891         int limit = cfs_trace_max_debug_mb();
 892         struct cfs_trace_cpu_data *tcd;
 893
 894         if (mb < num_possible_cpus()) {
 895                 printk(KERN_WARNING
 896                        "Lustre: %d MB is too small for debug buffer size, "
 897                        "setting it to %d MB.\n", mb, num_possible_cpus());
 898                 mb = num_possible_cpus();
 899         }
 900
 901         if (mb > limit) {
 902                 printk(KERN_WARNING
 903                        "Lustre: %d MB is too large for debug buffer size, "
 904                        "setting it to %d MB.\n", mb, limit);
 905                 mb = limit;
 906         }
 907
 908         mb /= num_possible_cpus();
 909         pages = mb << (20 - PAGE_CACHE_SHIFT);
 910
 911         cfs_tracefile_write_lock();
 912
 913         cfs_tcd_for_each(tcd, i, j)
 914                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 915
 916         cfs_tracefile_write_unlock();
 917
 918         return 0;
 919 }
 920
 921 int cfs_trace_set_debug_mb_usrstr(void __user *usr_str, int usr_str_nob)
 922 {
 923         char     str[32];
 924         int      rc;
 925
 926         rc = cfs_trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 927         if (rc < 0)
 928                 return rc;
 929
 930         return cfs_trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 931 }
 932
 933 int cfs_trace_get_debug_mb(void)
 934 {
 935         int i;
 936         int j;
 937         struct cfs_trace_cpu_data *tcd;
 938         int total_pages = 0;
 939
 940         cfs_tracefile_read_lock();
 941
 942         cfs_tcd_for_each(tcd, i, j)
 943                 total_pages += tcd->tcd_max_pages;
 944
 945         cfs_tracefile_read_unlock();
 946
 947         return (total_pages >> (20 - PAGE_CACHE_SHIFT)) + 1;
 948 }
 949
 950 static int tracefiled(void *arg)
 951 {
 952         struct page_collection pc;
 953         struct tracefiled_ctl *tctl = arg;
 954         struct cfs_trace_page *tage;
 955         struct cfs_trace_page *tmp;
 956         mm_segment_t __oldfs;
 957         struct file *filp;
 958         int last_loop = 0;
 959         int rc;
 960
 961         /* we're started late enough that we pick up init's fs context */
 962         /* this is so broken in uml?  what on earth is going on? */
 963
 964         complete(&tctl->tctl_start);
 965
 966         while (1) {
 967                 wait_queue_t __wait;
 968
 969                 pc.pc_want_daemon_pages = 0;
 970                 collect_pages(&pc);
 971                 if (list_empty(&pc.pc_pages))
 972                         goto end_loop;
 973
 974                 filp = NULL;
 975                 cfs_tracefile_read_lock();
 976                 if (cfs_tracefile[0] != 0) {
 977                         filp = filp_open(cfs_tracefile,
 978                                          O_CREAT | O_RDWR | O_LARGEFILE,
 979                                          0600);
 980                         if (IS_ERR(filp)) {
 981                                 rc = PTR_ERR(filp);
 982                                 filp = NULL;
 983                                 printk(KERN_WARNING "couldn't open %s: "
 984                                        "%d\n", cfs_tracefile, rc);
 985                         }
 986                 }
 987                 cfs_tracefile_read_unlock();
 988                 if (filp == NULL) {
 989                         put_pages_on_daemon_list(&pc);
 990                         __LASSERT(list_empty(&pc.pc_pages));
 991                         goto end_loop;
 992                 }
 993                 __oldfs = get_fs();
 994                 set_fs(get_ds());
 995
 996                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 997                         static loff_t f_pos;
 998
 999                         __LASSERT_TAGE_INVARIANT(tage);
1000
1001                         if (f_pos >= (off_t)cfs_tracefile_size)
1002                                 f_pos = 0;
1003                         else if (f_pos > (off_t)filp_size(filp))
1004                                 f_pos = filp_size(filp);
1005
1006                         rc = vfs_write(filp, page_address(tage->page),
1007                                        tage->used, &f_pos);
1008                         if (rc != (int)tage->used) {
1009                                 printk(KERN_WARNING "wanted to write %u "
1010                                        "but wrote %d\n", tage->used, rc);
1011                                 put_pages_back(&pc);
1012                                 __LASSERT(list_empty(&pc.pc_pages));
1013                                 break;
1014                         }
1015                 }
1016                 set_fs(__oldfs);
1017
1018                 filp_close(filp, NULL);
1019                 put_pages_on_daemon_list(&pc);
1020                 if (!list_empty(&pc.pc_pages)) {
1021                         int i;
1022
1023                         printk(KERN_ALERT "Lustre: trace pages aren't "
1024                                " empty\n");
1025                         printk(KERN_ERR "total cpus(%d): ",
1026                                num_possible_cpus());
1027                         for (i = 0; i < num_possible_cpus(); i++)
1028                                 if (cpu_online(i))
1029                                         printk(KERN_ERR "%d(on) ", i);
1030                                 else
1031                                         printk(KERN_ERR "%d(off) ", i);
1032                         printk(KERN_ERR "\n");
1033
1034                         i = 0;
1035                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1036                                                      linkage)
1037                                 printk(KERN_ERR "page %d belongs to cpu "
1038                                        "%d\n", ++i, tage->cpu);
1039                         printk(KERN_ERR "There are %d pages unwritten\n",
1040                                i);
1041                 }
1042                 __LASSERT(list_empty(&pc.pc_pages));
1043 end_loop:
1044                 if (atomic_read(&tctl->tctl_shutdown)) {
1045                         if (last_loop == 0) {
1046                                 last_loop = 1;
1047                                 continue;
1048                         } else {
1049                                 break;
1050                         }
1051                 }
1052                 init_waitqueue_entry_current(&__wait);
1053                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1054                 set_current_state(TASK_INTERRUPTIBLE);
1055                 waitq_timedwait(&__wait, TASK_INTERRUPTIBLE,
1056                                 cfs_time_seconds(1));
1057                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1058         }
1059         complete(&tctl->tctl_stop);
1060         return 0;
1061 }
1062
1063 int cfs_trace_start_thread(void)
1064 {
1065         struct tracefiled_ctl *tctl = &trace_tctl;
1066         int rc = 0;
1067
1068         mutex_lock(&cfs_trace_thread_mutex);
1069         if (thread_running)
1070                 goto out;
1071
1072         init_completion(&tctl->tctl_start);
1073         init_completion(&tctl->tctl_stop);
1074         init_waitqueue_head(&tctl->tctl_waitq);
1075         atomic_set(&tctl->tctl_shutdown, 0);
1076
1077         if (IS_ERR(kthread_run(tracefiled, tctl, "ktracefiled"))) {
1078                 rc = -ECHILD;
1079                 goto out;
1080         }
1081
1082         wait_for_completion(&tctl->tctl_start);
1083         thread_running = 1;
1084 out:
1085         mutex_unlock(&cfs_trace_thread_mutex);
1086         return rc;
1087 }
1088
1089 void cfs_trace_stop_thread(void)
1090 {
1091         struct tracefiled_ctl *tctl = &trace_tctl;
1092
1093         mutex_lock(&cfs_trace_thread_mutex);
1094         if (thread_running) {
1095                 printk(KERN_INFO
1096                        "Lustre: shutting down debug daemon thread...\n");
1097                 atomic_set(&tctl->tctl_shutdown, 1);
1098                 wait_for_completion(&tctl->tctl_stop);
1099                 thread_running = 0;
1100         }
1101         mutex_unlock(&cfs_trace_thread_mutex);
1102 }
1103
1104 int cfs_tracefile_init(int max_pages)
1105 {
1106         struct cfs_trace_cpu_data *tcd;
1107         int     i;
1108         int     j;
1109         int     rc;
1110         int     factor;
1111
1112         rc = cfs_tracefile_init_arch();
1113         if (rc != 0)
1114                 return rc;
1115
1116         cfs_tcd_for_each(tcd, i, j) {
1117                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1118                 factor = tcd->tcd_pages_factor;
1119                 INIT_LIST_HEAD(&tcd->tcd_pages);
1120                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1121                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1122                 tcd->tcd_cur_pages = 0;
1123                 tcd->tcd_cur_stock_pages = 0;
1124                 tcd->tcd_cur_daemon_pages = 0;
1125                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1126                 LASSERT(tcd->tcd_max_pages > 0);
1127                 tcd->tcd_shutting_down = 0;
1128         }
1129         return 0;
1130 }
1131
1132 static void trace_cleanup_on_all_cpus(void)
1133 {
1134         struct cfs_trace_cpu_data *tcd;
1135         struct cfs_trace_page *tage;
1136         struct cfs_trace_page *tmp;
1137         int i, cpu;
1138
1139         for_each_possible_cpu(cpu) {
1140                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1141                         tcd->tcd_shutting_down = 1;
1142
1143                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1144                                 __LASSERT_TAGE_INVARIANT(tage);
1145
1146                                 list_del(&tage->linkage);
1147                                 cfs_tage_free(tage);
1148                         }
1149                         tcd->tcd_cur_pages = 0;
1150                 }
1151         }
1152 }
1153
1154 static void cfs_trace_cleanup(void)
1155 {
1156         struct page_collection pc;
1157
1158         INIT_LIST_HEAD(&pc.pc_pages);
1159
1160         trace_cleanup_on_all_cpus();
1161
1162         cfs_tracefile_fini_arch();
1163 }
1164
1165 void cfs_tracefile_exit(void)
1166 {
1167         cfs_trace_stop_thread();
1168         cfs_trace_cleanup();
1169 }