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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * libcfs/libcfs/tracefile.c
  33  *
  34  * Author: Zach Brown <zab@clusterfs.com>
  35  * Author: Phil Schwan <phil@clusterfs.com>
  36  */
  37
  38
  39 #define DEBUG_SUBSYSTEM S_LNET
  40 #define LUSTRE_TRACEFILE_PRIVATE
  41 #include "tracefile.h"
  42
  43 #include <linux/kthread.h>
  44 #include <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 = 0;
  54
  55 static atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
  56
  57 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  58                                         struct cfs_trace_cpu_data *tcd);
  59
  60 static inline struct cfs_trace_page *
  61 cfs_tage_from_list(struct list_head *list)
  62 {
  63         return list_entry(list, struct cfs_trace_page, linkage);
  64 }
  65
  66 static struct cfs_trace_page *cfs_tage_alloc(gfp_t gfp)
  67 {
  68         struct page            *page;
  69         struct cfs_trace_page *tage;
  70
  71         /* My caller is trying to free memory */
  72         if (!in_interrupt() && memory_pressure_get())
  73                 return NULL;
  74
  75         /*
  76          * Don't spam console with allocation failures: they will be reported
  77          * by upper layer anyway.
  78          */
  79         gfp |= __GFP_NOWARN;
  80         page = alloc_page(gfp);
  81         if (page == NULL)
  82                 return NULL;
  83
  84         tage = kmalloc(sizeof(*tage), gfp);
  85         if (tage == NULL) {
  86                 __free_page(page);
  87                 return NULL;
  88         }
  89
  90         tage->page = page;
  91         atomic_inc(&cfs_tage_allocated);
  92         return tage;
  93 }
  94
  95 static void cfs_tage_free(struct cfs_trace_page *tage)
  96 {
  97         __LASSERT(tage != NULL);
  98         __LASSERT(tage->page != NULL);
  99
 100         __free_page(tage->page);
 101         kfree(tage);
 102         atomic_dec(&cfs_tage_allocated);
 103 }
 104
 105 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
 106                              struct list_head *queue)
 107 {
 108         __LASSERT(tage != NULL);
 109         __LASSERT(queue != NULL);
 110
 111         list_move_tail(&tage->linkage, queue);
 112 }
 113
 114 int cfs_trace_refill_stock(struct cfs_trace_cpu_data *tcd, gfp_t gfp,
 115                            struct list_head *stock)
 116 {
 117         int i;
 118
 119         /*
 120          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 121          * from here: this will lead to infinite recursion.
 122          */
 123
 124         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++ i) {
 125                 struct cfs_trace_page *tage;
 126
 127                 tage = cfs_tage_alloc(gfp);
 128                 if (tage == NULL)
 129                         break;
 130                 list_add_tail(&tage->linkage, stock);
 131         }
 132         return i;
 133 }
 134
 135 /* return a page that has 'len' bytes left at the end */
 136 static struct cfs_trace_page *
 137 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
 138 {
 139         struct cfs_trace_page *tage;
 140
 141         if (tcd->tcd_cur_pages > 0) {
 142                 __LASSERT(!list_empty(&tcd->tcd_pages));
 143                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
 144                 if (tage->used + len <= PAGE_SIZE)
 145                         return tage;
 146         }
 147
 148         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 149                 if (tcd->tcd_cur_stock_pages > 0) {
 150                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
 151                         --tcd->tcd_cur_stock_pages;
 152                         list_del_init(&tage->linkage);
 153                 } else {
 154                         tage = cfs_tage_alloc(GFP_ATOMIC);
 155                         if (unlikely(tage == NULL)) {
 156                                 if ((!memory_pressure_get() ||
 157                                      in_interrupt()) && printk_ratelimit())
 158                                         printk(KERN_WARNING
 159                                                "cannot allocate a tage (%ld)\n",
 160                                                tcd->tcd_cur_pages);
 161                                 return NULL;
 162                         }
 163                 }
 164
 165                 tage->used = 0;
 166                 tage->cpu = smp_processor_id();
 167                 tage->type = tcd->tcd_type;
 168                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 169                 tcd->tcd_cur_pages++;
 170
 171                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 172                         struct tracefiled_ctl *tctl = &trace_tctl;
 173                         /*
 174                          * wake up tracefiled to process some pages.
 175                          */
 176                         wake_up(&tctl->tctl_waitq);
 177                 }
 178                 return tage;
 179         }
 180         return NULL;
 181 }
 182
 183 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
 184 {
 185         int pgcount = tcd->tcd_cur_pages / 10;
 186         struct page_collection pc;
 187         struct cfs_trace_page *tage;
 188         struct cfs_trace_page *tmp;
 189
 190         /*
 191          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 192          * from here: this will lead to infinite recursion.
 193          */
 194
 195         if (printk_ratelimit())
 196                 printk(KERN_WARNING "debug daemon buffer overflowed; "
 197                         "discarding 10%% of pages (%d of %ld)\n",
 198                         pgcount + 1, tcd->tcd_cur_pages);
 199
 200         INIT_LIST_HEAD(&pc.pc_pages);
 201
 202         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 203                 if (pgcount-- == 0)
 204                         break;
 205
 206                 list_move_tail(&tage->linkage, &pc.pc_pages);
 207                 tcd->tcd_cur_pages--;
 208         }
 209         put_pages_on_tcd_daemon_list(&pc, tcd);
 210 }
 211
 212 /* return a page that has 'len' bytes left at the end */
 213 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 214                                                  unsigned long len)
 215 {
 216         struct cfs_trace_page *tage;
 217
 218         /*
 219          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 220          * from here: this will lead to infinite recursion.
 221          */
 222
 223         if (len > PAGE_SIZE) {
 224                 printk(KERN_ERR
 225                        "cowardly refusing to write %lu bytes in a page\n", len);
 226                 return NULL;
 227         }
 228
 229         tage = cfs_trace_get_tage_try(tcd, len);
 230         if (tage != NULL)
 231                 return tage;
 232         if (thread_running)
 233                 cfs_tcd_shrink(tcd);
 234         if (tcd->tcd_cur_pages > 0) {
 235                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 236                 tage->used = 0;
 237                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 238         }
 239         return tage;
 240 }
 241
 242 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 243                      const char *format, ...)
 244 {
 245         va_list args;
 246         int     rc;
 247
 248         va_start(args, format);
 249         rc = libcfs_debug_vmsg2(msgdata, format, args, NULL);
 250         va_end(args);
 251
 252         return rc;
 253 }
 254 EXPORT_SYMBOL(libcfs_debug_msg);
 255
 256 int libcfs_debug_vmsg2(struct libcfs_debug_msg_data *msgdata,
 257                        const char *format1, va_list args,
 258                        const char *format2, ...)
 259 {
 260         struct cfs_trace_cpu_data *tcd = NULL;
 261         struct ptldebug_header     header = {0};
 262         struct cfs_trace_page     *tage;
 263         /* string_buf is used only if tcd != NULL, and is always set then */
 264         char                      *string_buf = NULL;
 265         char                      *debug_buf;
 266         int                        known_size;
 267         int                        needed = 85; /* average message length */
 268         int                        max_nob;
 269         va_list                    ap;
 270         int                        i;
 271         int                        remain;
 272         int                        mask = msgdata->msg_mask;
 273         char                      *file = (char *)msgdata->msg_file;
 274         struct cfs_debug_limit_state *cdls = msgdata->msg_cdls;
 275
 276         if (strchr(file, '/'))
 277                 file = strrchr(file, '/') + 1;
 278
 279         tcd = cfs_trace_get_tcd();
 280
 281         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 282          * pins us to a particular CPU.  This avoids an smp_processor_id()
 283          * warning on Linux when debugging is enabled. */
 284         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 285
 286         if (tcd == NULL)                /* arch may not log in IRQ context */
 287                 goto console;
 288
 289         if (tcd->tcd_cur_pages == 0)
 290                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 291
 292         if (tcd->tcd_shutting_down) {
 293                 cfs_trace_put_tcd(tcd);
 294                 tcd = NULL;
 295                 goto console;
 296         }
 297
 298         known_size = strlen(file) + 1;
 299         if (msgdata->msg_fn)
 300                 known_size += strlen(msgdata->msg_fn) + 1;
 301
 302         if (libcfs_debug_binary)
 303                 known_size += sizeof(header);
 304
 305         /*/
 306          * '2' used because vsnprintf return real size required for output
 307          * _without_ terminating NULL.
 308          * if needed is to small for this format.
 309          */
 310         for (i = 0; i < 2; i++) {
 311                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 312                 if (tage == NULL) {
 313                         if (needed + known_size > PAGE_SIZE)
 314                                 mask |= D_ERROR;
 315
 316                         cfs_trace_put_tcd(tcd);
 317                         tcd = NULL;
 318                         goto console;
 319                 }
 320
 321                 string_buf = (char *)page_address(tage->page) +
 322                                         tage->used + known_size;
 323
 324                 max_nob = PAGE_SIZE - tage->used - known_size;
 325                 if (max_nob <= 0) {
 326                         printk(KERN_EMERG "negative max_nob: %d\n",
 327                                max_nob);
 328                         mask |= D_ERROR;
 329                         cfs_trace_put_tcd(tcd);
 330                         tcd = NULL;
 331                         goto console;
 332                 }
 333
 334                 needed = 0;
 335                 if (format1) {
 336                         va_copy(ap, args);
 337                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 338                         va_end(ap);
 339                 }
 340
 341                 if (format2) {
 342                         remain = max_nob - needed;
 343                         if (remain < 0)
 344                                 remain = 0;
 345
 346                         va_start(ap, format2);
 347                         needed += vsnprintf(string_buf + needed, remain,
 348                                             format2, ap);
 349                         va_end(ap);
 350                 }
 351
 352                 if (needed < max_nob) /* well. printing ok.. */
 353                         break;
 354         }
 355
 356         if (*(string_buf+needed-1) != '\n')
 357                 printk(KERN_INFO "format at %s:%d:%s doesn't end in "
 358                        "newline\n", file, msgdata->msg_line, msgdata->msg_fn);
 359
 360         header.ph_len = known_size + needed;
 361         debug_buf = (char *)page_address(tage->page) + tage->used;
 362
 363         if (libcfs_debug_binary) {
 364                 memcpy(debug_buf, &header, sizeof(header));
 365                 tage->used += sizeof(header);
 366                 debug_buf += sizeof(header);
 367         }
 368
 369         strcpy(debug_buf, file);
 370         tage->used += strlen(file) + 1;
 371         debug_buf += strlen(file) + 1;
 372
 373         if (msgdata->msg_fn) {
 374                 strcpy(debug_buf, msgdata->msg_fn);
 375                 tage->used += strlen(msgdata->msg_fn) + 1;
 376                 debug_buf += strlen(msgdata->msg_fn) + 1;
 377         }
 378
 379         __LASSERT(debug_buf == string_buf);
 380
 381         tage->used += needed;
 382         __LASSERT(tage->used <= PAGE_SIZE);
 383
 384 console:
 385         if ((mask & libcfs_printk) == 0) {
 386                 /* no console output requested */
 387                 if (tcd != NULL)
 388                         cfs_trace_put_tcd(tcd);
 389                 return 1;
 390         }
 391
 392         if (cdls != NULL) {
 393                 if (libcfs_console_ratelimit &&
 394                     cdls->cdls_next != 0 &&     /* not first time ever */
 395                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 396                         /* skipping a console message */
 397                         cdls->cdls_count++;
 398                         if (tcd != NULL)
 399                                 cfs_trace_put_tcd(tcd);
 400                         return 1;
 401                 }
 402
 403                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 404                                                        libcfs_console_max_delay
 405                                                        + cfs_time_seconds(10))) {
 406                         /* last timeout was a long time ago */
 407                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 408                 } else {
 409                         cdls->cdls_delay *= libcfs_console_backoff;
 410                 }
 411
 412                 if (cdls->cdls_delay < libcfs_console_min_delay)
 413                         cdls->cdls_delay = libcfs_console_min_delay;
 414                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 415                         cdls->cdls_delay = libcfs_console_max_delay;
 416
 417                 /* ensure cdls_next is never zero after it's been seen */
 418                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 419         }
 420
 421         if (tcd != NULL) {
 422                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 423                                      msgdata->msg_fn);
 424                 cfs_trace_put_tcd(tcd);
 425         } else {
 426                 string_buf = cfs_trace_get_console_buffer();
 427
 428                 needed = 0;
 429                 if (format1 != NULL) {
 430                         va_copy(ap, args);
 431                         needed = vsnprintf(string_buf,
 432                                            CFS_TRACE_CONSOLE_BUFFER_SIZE,
 433                                            format1, ap);
 434                         va_end(ap);
 435                 }
 436                 if (format2 != NULL) {
 437                         remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
 438                         if (remain > 0) {
 439                                 va_start(ap, format2);
 440                                 needed += vsnprintf(string_buf+needed, remain,
 441                                                     format2, ap);
 442                                 va_end(ap);
 443                         }
 444                 }
 445                 cfs_print_to_console(&header, mask,
 446                                      string_buf, needed, file, msgdata->msg_fn);
 447
 448                 put_cpu();
 449         }
 450
 451         if (cdls != NULL && cdls->cdls_count != 0) {
 452                 string_buf = cfs_trace_get_console_buffer();
 453
 454                 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 455                                   "Skipped %d previous similar message%s\n",
 456                                   cdls->cdls_count,
 457                                   (cdls->cdls_count > 1) ? "s" : "");
 458
 459                 cfs_print_to_console(&header, mask,
 460                                      string_buf, needed, file, msgdata->msg_fn);
 461
 462                 put_cpu();
 463                 cdls->cdls_count = 0;
 464         }
 465
 466         return 0;
 467 }
 468 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 469
 470 void
 471 cfs_trace_assertion_failed(const char *str,
 472                            struct libcfs_debug_msg_data *msgdata)
 473 {
 474         struct ptldebug_header hdr;
 475
 476         libcfs_panic_in_progress = 1;
 477         libcfs_catastrophe = 1;
 478         smp_mb();
 479
 480         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 481
 482         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 483                              msgdata->msg_file, msgdata->msg_fn);
 484
 485         panic("Lustre debug assertion failure\n");
 486
 487         /* not reached */
 488 }
 489
 490 static void
 491 panic_collect_pages(struct page_collection *pc)
 492 {
 493         /* Do the collect_pages job on a single CPU: assumes that all other
 494          * CPUs have been stopped during a panic.  If this isn't true for some
 495          * arch, this will have to be implemented separately in each arch.  */
 496         int                        i;
 497         int                        j;
 498         struct cfs_trace_cpu_data *tcd;
 499
 500         INIT_LIST_HEAD(&pc->pc_pages);
 501
 502         cfs_tcd_for_each(tcd, i, j) {
 503                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 504                 tcd->tcd_cur_pages = 0;
 505
 506                 if (pc->pc_want_daemon_pages) {
 507                         list_splice_init(&tcd->tcd_daemon_pages,
 508                                                 &pc->pc_pages);
 509                         tcd->tcd_cur_daemon_pages = 0;
 510                 }
 511         }
 512 }
 513
 514 static void collect_pages_on_all_cpus(struct page_collection *pc)
 515 {
 516         struct cfs_trace_cpu_data *tcd;
 517         int i, cpu;
 518
 519         for_each_possible_cpu(cpu) {
 520                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 521                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 522                         tcd->tcd_cur_pages = 0;
 523                         if (pc->pc_want_daemon_pages) {
 524                                 list_splice_init(&tcd->tcd_daemon_pages,
 525                                                         &pc->pc_pages);
 526                                 tcd->tcd_cur_daemon_pages = 0;
 527                         }
 528                 }
 529         }
 530 }
 531
 532 static void collect_pages(struct page_collection *pc)
 533 {
 534         INIT_LIST_HEAD(&pc->pc_pages);
 535
 536         if (libcfs_panic_in_progress)
 537                 panic_collect_pages(pc);
 538         else
 539                 collect_pages_on_all_cpus(pc);
 540 }
 541
 542 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 543 {
 544         struct cfs_trace_cpu_data *tcd;
 545         struct list_head *cur_head;
 546         struct cfs_trace_page *tage;
 547         struct cfs_trace_page *tmp;
 548         int i, cpu;
 549
 550         for_each_possible_cpu(cpu) {
 551                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 552                         cur_head = tcd->tcd_pages.next;
 553
 554                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 555                                                  linkage) {
 556
 557                                 __LASSERT_TAGE_INVARIANT(tage);
 558
 559                                 if (tage->cpu != cpu || tage->type != i)
 560                                         continue;
 561
 562                                 cfs_tage_to_tail(tage, cur_head);
 563                                 tcd->tcd_cur_pages++;
 564                         }
 565                 }
 566         }
 567 }
 568
 569 static void put_pages_back(struct page_collection *pc)
 570 {
 571         if (!libcfs_panic_in_progress)
 572                 put_pages_back_on_all_cpus(pc);
 573 }
 574
 575 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 576  * we have a good amount of data at all times for dumping during an LBUG, even
 577  * if we have been steadily writing (and otherwise discarding) pages via the
 578  * debug daemon. */
 579 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 580                                          struct cfs_trace_cpu_data *tcd)
 581 {
 582         struct cfs_trace_page *tage;
 583         struct cfs_trace_page *tmp;
 584
 585         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 586                 __LASSERT_TAGE_INVARIANT(tage);
 587
 588                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 589                         continue;
 590
 591                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 592                 tcd->tcd_cur_daemon_pages++;
 593
 594                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 595                         struct cfs_trace_page *victim;
 596
 597                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 598                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 599
 600                         __LASSERT_TAGE_INVARIANT(victim);
 601
 602                         list_del(&victim->linkage);
 603                         cfs_tage_free(victim);
 604                         tcd->tcd_cur_daemon_pages--;
 605                 }
 606         }
 607 }
 608
 609 static void put_pages_on_daemon_list(struct page_collection *pc)
 610 {
 611         struct cfs_trace_cpu_data *tcd;
 612         int i, cpu;
 613
 614         for_each_possible_cpu(cpu) {
 615                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 616                         put_pages_on_tcd_daemon_list(pc, tcd);
 617         }
 618 }
 619
 620 void cfs_trace_debug_print(void)
 621 {
 622         struct page_collection pc;
 623         struct cfs_trace_page *tage;
 624         struct cfs_trace_page *tmp;
 625
 626         pc.pc_want_daemon_pages = 1;
 627         collect_pages(&pc);
 628         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 629                 char *p, *file, *fn;
 630                 struct page *page;
 631
 632                 __LASSERT_TAGE_INVARIANT(tage);
 633
 634                 page = tage->page;
 635                 p = page_address(page);
 636                 while (p < ((char *)page_address(page) + tage->used)) {
 637                         struct ptldebug_header *hdr;
 638                         int len;
 639                         hdr = (void *)p;
 640                         p += sizeof(*hdr);
 641                         file = p;
 642                         p += strlen(file) + 1;
 643                         fn = p;
 644                         p += strlen(fn) + 1;
 645                         len = hdr->ph_len - (int)(p - (char *)hdr);
 646
 647                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 648
 649                         p += len;
 650                 }
 651
 652                 list_del(&tage->linkage);
 653                 cfs_tage_free(tage);
 654         }
 655 }
 656
 657 int cfs_tracefile_dump_all_pages(char *filename)
 658 {
 659         struct page_collection  pc;
 660         struct file             *filp;
 661         struct cfs_trace_page   *tage;
 662         struct cfs_trace_page   *tmp;
 663         mm_segment_t            __oldfs;
 664         char                    *buf;
 665         int rc;
 666
 667         cfs_tracefile_write_lock();
 668
 669         filp = filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600);
 670         if (IS_ERR(filp)) {
 671                 rc = PTR_ERR(filp);
 672                 filp = NULL;
 673                 printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 674                       filename, rc);
 675                 goto out;
 676         }
 677
 678         pc.pc_want_daemon_pages = 1;
 679         collect_pages(&pc);
 680         if (list_empty(&pc.pc_pages)) {
 681                 rc = 0;
 682                 goto close;
 683         }
 684         __oldfs = get_fs();
 685         set_fs(get_ds());
 686
 687         /* ok, for now, just write the pages.  in the future we'll be building
 688          * iobufs with the pages and calling generic_direct_IO */
 689         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 690
 691                 __LASSERT_TAGE_INVARIANT(tage);
 692
 693                 buf = kmap(tage->page);
 694                 rc = vfs_write(filp, (__force const char __user *)buf,
 695                                tage->used, &filp->f_pos);
 696                 kunmap(tage->page);
 697                 if (rc != (int)tage->used) {
 698                         printk(KERN_WARNING "wanted to write %u but wrote "
 699                                "%d\n", tage->used, rc);
 700                         put_pages_back(&pc);
 701                         __LASSERT(list_empty(&pc.pc_pages));
 702                         break;
 703                 }
 704                 list_del(&tage->linkage);
 705                 cfs_tage_free(tage);
 706         }
 707         set_fs(__oldfs);
 708         rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1);
 709         if (rc)
 710                 printk(KERN_ERR "sync returns %d\n", rc);
 711 close:
 712         filp_close(filp, NULL);
 713 out:
 714         cfs_tracefile_write_unlock();
 715         return rc;
 716 }
 717
 718 void cfs_trace_flush_pages(void)
 719 {
 720         struct page_collection pc;
 721         struct cfs_trace_page *tage;
 722         struct cfs_trace_page *tmp;
 723
 724         pc.pc_want_daemon_pages = 1;
 725         collect_pages(&pc);
 726         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 727
 728                 __LASSERT_TAGE_INVARIANT(tage);
 729
 730                 list_del(&tage->linkage);
 731                 cfs_tage_free(tage);
 732         }
 733 }
 734
 735 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 736                             const char __user *usr_buffer, int usr_buffer_nob)
 737 {
 738         int    nob;
 739
 740         if (usr_buffer_nob > knl_buffer_nob)
 741                 return -EOVERFLOW;
 742
 743         if (copy_from_user(knl_buffer, usr_buffer, usr_buffer_nob))
 744                 return -EFAULT;
 745
 746         nob = strnlen(knl_buffer, usr_buffer_nob);
 747         while (nob-- >= 0)                      /* strip trailing whitespace */
 748                 if (!isspace(knl_buffer[nob]))
 749                         break;
 750
 751         if (nob < 0)                            /* empty string */
 752                 return -EINVAL;
 753
 754         if (nob == knl_buffer_nob)              /* no space to terminate */
 755                 return -EOVERFLOW;
 756
 757         knl_buffer[nob + 1] = 0;                /* terminate */
 758         return 0;
 759 }
 760 EXPORT_SYMBOL(cfs_trace_copyin_string);
 761
 762 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 763                              const char *knl_buffer, char *append)
 764 {
 765         /* NB if 'append' != NULL, it's a single character to append to the
 766          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 767          * terminating zero byte) for sysctl entries */
 768         int   nob = strlen(knl_buffer);
 769
 770         if (nob > usr_buffer_nob)
 771                 nob = usr_buffer_nob;
 772
 773         if (copy_to_user(usr_buffer, knl_buffer, nob))
 774                 return -EFAULT;
 775
 776         if (append != NULL && nob < usr_buffer_nob) {
 777                 if (copy_to_user(usr_buffer + nob, append, 1))
 778                         return -EFAULT;
 779
 780                 nob++;
 781         }
 782
 783         return nob;
 784 }
 785 EXPORT_SYMBOL(cfs_trace_copyout_string);
 786
 787 int cfs_trace_allocate_string_buffer(char **str, int nob)
 788 {
 789         if (nob > 2 * PAGE_SIZE)        /* string must be "sensible" */
 790                 return -EINVAL;
 791
 792         *str = kmalloc(nob, GFP_KERNEL | __GFP_ZERO);
 793         if (*str == NULL)
 794                 return -ENOMEM;
 795
 796         return 0;
 797 }
 798
 799 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
 800 {
 801         char         *str;
 802         int           rc;
 803
 804         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 805         if (rc != 0)
 806                 return rc;
 807
 808         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 809                                      usr_str, usr_str_nob);
 810         if (rc != 0)
 811                 goto out;
 812
 813         if (str[0] != '/') {
 814                 rc = -EINVAL;
 815                 goto out;
 816         }
 817         rc = cfs_tracefile_dump_all_pages(str);
 818 out:
 819         kfree(str);
 820         return rc;
 821 }
 822
 823 int cfs_trace_daemon_command(char *str)
 824 {
 825         int       rc = 0;
 826
 827         cfs_tracefile_write_lock();
 828
 829         if (strcmp(str, "stop") == 0) {
 830                 cfs_tracefile_write_unlock();
 831                 cfs_trace_stop_thread();
 832                 cfs_tracefile_write_lock();
 833                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 834
 835         } else if (strncmp(str, "size=", 5) == 0) {
 836                 cfs_tracefile_size = simple_strtoul(str + 5, NULL, 0);
 837                 if (cfs_tracefile_size < 10 || cfs_tracefile_size > 20480)
 838                         cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 839                 else
 840                         cfs_tracefile_size <<= 20;
 841
 842         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 843                 rc = -ENAMETOOLONG;
 844         } else if (str[0] != '/') {
 845                 rc = -EINVAL;
 846         } else {
 847                 strcpy(cfs_tracefile, str);
 848
 849                 printk(KERN_INFO
 850                        "Lustre: debug daemon will attempt to start writing "
 851                        "to %s (%lukB max)\n", cfs_tracefile,
 852                        (long)(cfs_tracefile_size >> 10));
 853
 854                 cfs_trace_start_thread();
 855         }
 856
 857         cfs_tracefile_write_unlock();
 858         return rc;
 859 }
 860
 861 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
 862 {
 863         char *str;
 864         int   rc;
 865
 866         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 867         if (rc != 0)
 868                 return rc;
 869
 870         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 871                                  usr_str, usr_str_nob);
 872         if (rc == 0)
 873                 rc = cfs_trace_daemon_command(str);
 874
 875         kfree(str);
 876         return rc;
 877 }
 878
 879 int cfs_trace_set_debug_mb(int mb)
 880 {
 881         int i;
 882         int j;
 883         int pages;
 884         int limit = cfs_trace_max_debug_mb();
 885         struct cfs_trace_cpu_data *tcd;
 886
 887         if (mb < num_possible_cpus()) {
 888                 printk(KERN_WARNING
 889                        "Lustre: %d MB is too small for debug buffer size, "
 890                        "setting it to %d MB.\n", mb, num_possible_cpus());
 891                 mb = num_possible_cpus();
 892         }
 893
 894         if (mb > limit) {
 895                 printk(KERN_WARNING
 896                        "Lustre: %d MB is too large for debug buffer size, "
 897                        "setting it to %d MB.\n", mb, limit);
 898                 mb = limit;
 899         }
 900
 901         mb /= num_possible_cpus();
 902         pages = mb << (20 - PAGE_SHIFT);
 903
 904         cfs_tracefile_write_lock();
 905
 906         cfs_tcd_for_each(tcd, i, j)
 907                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 908
 909         cfs_tracefile_write_unlock();
 910
 911         return 0;
 912 }
 913
 914 int cfs_trace_get_debug_mb(void)
 915 {
 916         int i;
 917         int j;
 918         struct cfs_trace_cpu_data *tcd;
 919         int total_pages = 0;
 920
 921         cfs_tracefile_read_lock();
 922
 923         cfs_tcd_for_each(tcd, i, j)
 924                 total_pages += tcd->tcd_max_pages;
 925
 926         cfs_tracefile_read_unlock();
 927
 928         return (total_pages >> (20 - PAGE_SHIFT)) + 1;
 929 }
 930
 931 static int tracefiled(void *arg)
 932 {
 933         struct page_collection pc;
 934         struct tracefiled_ctl *tctl = arg;
 935         struct cfs_trace_page *tage;
 936         struct cfs_trace_page *tmp;
 937         mm_segment_t __oldfs;
 938         struct file *filp;
 939         char *buf;
 940         int last_loop = 0;
 941         int rc;
 942
 943         /* we're started late enough that we pick up init's fs context */
 944         /* this is so broken in uml?  what on earth is going on? */
 945
 946         complete(&tctl->tctl_start);
 947
 948         while (1) {
 949                 wait_queue_t __wait;
 950
 951                 pc.pc_want_daemon_pages = 0;
 952                 collect_pages(&pc);
 953                 if (list_empty(&pc.pc_pages))
 954                         goto end_loop;
 955
 956                 filp = NULL;
 957                 cfs_tracefile_read_lock();
 958                 if (cfs_tracefile[0] != 0) {
 959                         filp = filp_open(cfs_tracefile,
 960                                          O_CREAT | O_RDWR | O_LARGEFILE,
 961                                          0600);
 962                         if (IS_ERR(filp)) {
 963                                 rc = PTR_ERR(filp);
 964                                 filp = NULL;
 965                                 printk(KERN_WARNING "couldn't open %s: "
 966                                        "%d\n", cfs_tracefile, rc);
 967                         }
 968                 }
 969                 cfs_tracefile_read_unlock();
 970                 if (filp == NULL) {
 971                         put_pages_on_daemon_list(&pc);
 972                         __LASSERT(list_empty(&pc.pc_pages));
 973                         goto end_loop;
 974                 }
 975                 __oldfs = get_fs();
 976                 set_fs(get_ds());
 977
 978                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 979                         struct dentry *de = file_dentry(filp);
 980                         static loff_t f_pos;
 981
 982                         __LASSERT_TAGE_INVARIANT(tage);
 983
 984                         if (f_pos >= (off_t)cfs_tracefile_size)
 985                                 f_pos = 0;
 986                         else if (f_pos > i_size_read(de->d_inode))
 987                                 f_pos = i_size_read(de->d_inode);
 988
 989                         buf = kmap(tage->page);
 990                         rc = vfs_write(filp, (__force const char __user *)buf,
 991                                        tage->used, &f_pos);
 992                         kunmap(tage->page);
 993                         if (rc != (int)tage->used) {
 994                                 printk(KERN_WARNING "wanted to write %u "
 995                                        "but wrote %d\n", tage->used, rc);
 996                                 put_pages_back(&pc);
 997                                 __LASSERT(list_empty(&pc.pc_pages));
 998                                 break;
 999                         }
1000                 }
1001                 set_fs(__oldfs);
1002
1003                 filp_close(filp, NULL);
1004                 put_pages_on_daemon_list(&pc);
1005                 if (!list_empty(&pc.pc_pages)) {
1006                         int i;
1007
1008                         printk(KERN_ALERT "Lustre: trace pages aren't "
1009                                " empty\n");
1010                         printk(KERN_ERR "total cpus(%d): ",
1011                                num_possible_cpus());
1012                         for (i = 0; i < num_possible_cpus(); i++)
1013                                 if (cpu_online(i))
1014                                         printk(KERN_ERR "%d(on) ", i);
1015                                 else
1016                                         printk(KERN_ERR "%d(off) ", i);
1017                         printk(KERN_ERR "\n");
1018
1019                         i = 0;
1020                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1021                                                      linkage)
1022                                 printk(KERN_ERR "page %d belongs to cpu "
1023                                        "%d\n", ++i, tage->cpu);
1024                         printk(KERN_ERR "There are %d pages unwritten\n",
1025                                i);
1026                 }
1027                 __LASSERT(list_empty(&pc.pc_pages));
1028 end_loop:
1029                 if (atomic_read(&tctl->tctl_shutdown)) {
1030                         if (last_loop == 0) {
1031                                 last_loop = 1;
1032                                 continue;
1033                         } else {
1034                                 break;
1035                         }
1036                 }
1037                 init_waitqueue_entry(&__wait, current);
1038                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1039                 set_current_state(TASK_INTERRUPTIBLE);
1040                 schedule_timeout(cfs_time_seconds(1));
1041                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1042         }
1043         complete(&tctl->tctl_stop);
1044         return 0;
1045 }
1046
1047 int cfs_trace_start_thread(void)
1048 {
1049         struct tracefiled_ctl *tctl = &trace_tctl;
1050         int rc = 0;
1051
1052         mutex_lock(&cfs_trace_thread_mutex);
1053         if (thread_running)
1054                 goto out;
1055
1056         init_completion(&tctl->tctl_start);
1057         init_completion(&tctl->tctl_stop);
1058         init_waitqueue_head(&tctl->tctl_waitq);
1059         atomic_set(&tctl->tctl_shutdown, 0);
1060
1061         if (IS_ERR(kthread_run(tracefiled, tctl, "ktracefiled"))) {
1062                 rc = -ECHILD;
1063                 goto out;
1064         }
1065
1066         wait_for_completion(&tctl->tctl_start);
1067         thread_running = 1;
1068 out:
1069         mutex_unlock(&cfs_trace_thread_mutex);
1070         return rc;
1071 }
1072
1073 void cfs_trace_stop_thread(void)
1074 {
1075         struct tracefiled_ctl *tctl = &trace_tctl;
1076
1077         mutex_lock(&cfs_trace_thread_mutex);
1078         if (thread_running) {
1079                 printk(KERN_INFO
1080                        "Lustre: shutting down debug daemon thread...\n");
1081                 atomic_set(&tctl->tctl_shutdown, 1);
1082                 wait_for_completion(&tctl->tctl_stop);
1083                 thread_running = 0;
1084         }
1085         mutex_unlock(&cfs_trace_thread_mutex);
1086 }
1087
1088 int cfs_tracefile_init(int max_pages)
1089 {
1090         struct cfs_trace_cpu_data *tcd;
1091         int     i;
1092         int     j;
1093         int     rc;
1094         int     factor;
1095
1096         rc = cfs_tracefile_init_arch();
1097         if (rc != 0)
1098                 return rc;
1099
1100         cfs_tcd_for_each(tcd, i, j) {
1101                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1102                 factor = tcd->tcd_pages_factor;
1103                 INIT_LIST_HEAD(&tcd->tcd_pages);
1104                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1105                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1106                 tcd->tcd_cur_pages = 0;
1107                 tcd->tcd_cur_stock_pages = 0;
1108                 tcd->tcd_cur_daemon_pages = 0;
1109                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1110                 LASSERT(tcd->tcd_max_pages > 0);
1111                 tcd->tcd_shutting_down = 0;
1112         }
1113         return 0;
1114 }
1115
1116 static void trace_cleanup_on_all_cpus(void)
1117 {
1118         struct cfs_trace_cpu_data *tcd;
1119         struct cfs_trace_page *tage;
1120         struct cfs_trace_page *tmp;
1121         int i, cpu;
1122
1123         for_each_possible_cpu(cpu) {
1124                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1125                         tcd->tcd_shutting_down = 1;
1126
1127                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1128                                 __LASSERT_TAGE_INVARIANT(tage);
1129
1130                                 list_del(&tage->linkage);
1131                                 cfs_tage_free(tage);
1132                         }
1133                         tcd->tcd_cur_pages = 0;
1134                 }
1135         }
1136 }
1137
1138 static void cfs_trace_cleanup(void)
1139 {
1140         struct page_collection pc;
1141
1142         INIT_LIST_HEAD(&pc.pc_pages);
1143
1144         trace_cleanup_on_all_cpus();
1145
1146         cfs_tracefile_fini_arch();
1147 }
1148
1149 void cfs_tracefile_exit(void)
1150 {
1151         cfs_trace_stop_thread();
1152         cfs_trace_cleanup();
1153 }