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