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, 2017, 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/ctype.h>
  44 #include <linux/fs.h>
  45 #include <linux/kthread.h>
  46 #include <linux/pagemap.h>
  47 #include <linux/uaccess.h>
  48 #include <libcfs/linux/linux-fs.h>
  49 #include <libcfs/libcfs.h>
  50
  51 /* XXX move things up to the top, comment */
  52 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  53
  54 char cfs_tracefile[TRACEFILE_NAME_SIZE];
  55 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
  56 static struct tracefiled_ctl trace_tctl;
  57 static DEFINE_MUTEX(cfs_trace_thread_mutex);
  58 static int thread_running = 0;
  59
  60 static atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
  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_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_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         struct cfs_debug_limit_state *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_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_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_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                 put_cpu();
 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                 put_cpu();
 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         char                    *buf;
 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                 buf = kmap(tage->page);
 699                 rc = vfs_write(filp, (__force const char __user *)buf,
 700                                tage->used, &filp->f_pos);
 701                 kunmap(tage->page);
 702                 if (rc != (int)tage->used) {
 703                         printk(KERN_WARNING "wanted to write %u but wrote "
 704                                "%d\n", tage->used, rc);
 705                         put_pages_back(&pc);
 706                         __LASSERT(list_empty(&pc.pc_pages));
 707                         break;
 708                 }
 709                 list_del(&tage->linkage);
 710                 cfs_tage_free(tage);
 711         }
 712         set_fs(__oldfs);
 713         rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1);
 714         if (rc)
 715                 printk(KERN_ERR "sync returns %d\n", rc);
 716 close:
 717         filp_close(filp, NULL);
 718 out:
 719         cfs_tracefile_write_unlock();
 720         return rc;
 721 }
 722
 723 void cfs_trace_flush_pages(void)
 724 {
 725         struct page_collection pc;
 726         struct cfs_trace_page *tage;
 727         struct cfs_trace_page *tmp;
 728
 729         pc.pc_want_daemon_pages = 1;
 730         collect_pages(&pc);
 731         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 732
 733                 __LASSERT_TAGE_INVARIANT(tage);
 734
 735                 list_del(&tage->linkage);
 736                 cfs_tage_free(tage);
 737         }
 738 }
 739
 740 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 741                             const char __user *usr_buffer, int usr_buffer_nob)
 742 {
 743         int    nob;
 744
 745         if (usr_buffer_nob > knl_buffer_nob)
 746                 return -EOVERFLOW;
 747
 748         if (copy_from_user(knl_buffer, usr_buffer, usr_buffer_nob))
 749                 return -EFAULT;
 750
 751         nob = strnlen(knl_buffer, usr_buffer_nob);
 752         while (nob-- >= 0)                      /* strip trailing whitespace */
 753                 if (!isspace(knl_buffer[nob]))
 754                         break;
 755
 756         if (nob < 0)                            /* empty string */
 757                 return -EINVAL;
 758
 759         if (nob == knl_buffer_nob)              /* no space to terminate */
 760                 return -EOVERFLOW;
 761
 762         knl_buffer[nob + 1] = 0;                /* terminate */
 763         return 0;
 764 }
 765 EXPORT_SYMBOL(cfs_trace_copyin_string);
 766
 767 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 768                              const char *knl_buffer, char *append)
 769 {
 770         /* NB if 'append' != NULL, it's a single character to append to the
 771          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 772          * terminating zero byte) for sysctl entries */
 773         int   nob = strlen(knl_buffer);
 774
 775         if (nob > usr_buffer_nob)
 776                 nob = usr_buffer_nob;
 777
 778         if (copy_to_user(usr_buffer, knl_buffer, nob))
 779                 return -EFAULT;
 780
 781         if (append != NULL && nob < usr_buffer_nob) {
 782                 if (copy_to_user(usr_buffer + nob, append, 1))
 783                         return -EFAULT;
 784
 785                 nob++;
 786         }
 787
 788         return nob;
 789 }
 790 EXPORT_SYMBOL(cfs_trace_copyout_string);
 791
 792 int cfs_trace_allocate_string_buffer(char **str, int nob)
 793 {
 794         if (nob > 2 * PAGE_SIZE)        /* string must be "sensible" */
 795                 return -EINVAL;
 796
 797         *str = kmalloc(nob, GFP_KERNEL | __GFP_ZERO);
 798         if (*str == NULL)
 799                 return -ENOMEM;
 800
 801         return 0;
 802 }
 803
 804 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
 805 {
 806         char         *str;
 807         int           rc;
 808
 809         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 810         if (rc != 0)
 811                 return rc;
 812
 813         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 814                                      usr_str, usr_str_nob);
 815         if (rc != 0)
 816                 goto out;
 817
 818         if (str[0] != '/') {
 819                 rc = -EINVAL;
 820                 goto out;
 821         }
 822         rc = cfs_tracefile_dump_all_pages(str);
 823 out:
 824         kfree(str);
 825         return rc;
 826 }
 827
 828 int cfs_trace_daemon_command(char *str)
 829 {
 830         int       rc = 0;
 831
 832         cfs_tracefile_write_lock();
 833
 834         if (strcmp(str, "stop") == 0) {
 835                 cfs_tracefile_write_unlock();
 836                 cfs_trace_stop_thread();
 837                 cfs_tracefile_write_lock();
 838                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 839
 840         } else if (strncmp(str, "size=", 5) == 0) {
 841                 cfs_tracefile_size = simple_strtoul(str + 5, NULL, 0);
 842                 if (cfs_tracefile_size < 10 || cfs_tracefile_size > 20480)
 843                         cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 844                 else
 845                         cfs_tracefile_size <<= 20;
 846
 847         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 848                 rc = -ENAMETOOLONG;
 849         } else if (str[0] != '/') {
 850                 rc = -EINVAL;
 851         } else {
 852                 strcpy(cfs_tracefile, str);
 853
 854                 printk(KERN_INFO
 855                        "Lustre: debug daemon will attempt to start writing "
 856                        "to %s (%lukB max)\n", cfs_tracefile,
 857                        (long)(cfs_tracefile_size >> 10));
 858
 859                 cfs_trace_start_thread();
 860         }
 861
 862         cfs_tracefile_write_unlock();
 863         return rc;
 864 }
 865
 866 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
 867 {
 868         char *str;
 869         int   rc;
 870
 871         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 872         if (rc != 0)
 873                 return rc;
 874
 875         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 876                                  usr_str, usr_str_nob);
 877         if (rc == 0)
 878                 rc = cfs_trace_daemon_command(str);
 879
 880         kfree(str);
 881         return rc;
 882 }
 883
 884 int cfs_trace_set_debug_mb(int mb)
 885 {
 886         int i;
 887         int j;
 888         int pages;
 889         int limit = cfs_trace_max_debug_mb();
 890         struct cfs_trace_cpu_data *tcd;
 891
 892         if (mb < num_possible_cpus()) {
 893                 printk(KERN_WARNING
 894                        "Lustre: %d MB is too small for debug buffer size, "
 895                        "setting it to %d MB.\n", mb, num_possible_cpus());
 896                 mb = num_possible_cpus();
 897         }
 898
 899         if (mb > limit) {
 900                 printk(KERN_WARNING
 901                        "Lustre: %d MB is too large for debug buffer size, "
 902                        "setting it to %d MB.\n", mb, limit);
 903                 mb = limit;
 904         }
 905
 906         mb /= num_possible_cpus();
 907         pages = mb << (20 - PAGE_SHIFT);
 908
 909         cfs_tracefile_write_lock();
 910
 911         cfs_tcd_for_each(tcd, i, j)
 912                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 913
 914         cfs_tracefile_write_unlock();
 915
 916         return 0;
 917 }
 918
 919 int cfs_trace_get_debug_mb(void)
 920 {
 921         int i;
 922         int j;
 923         struct cfs_trace_cpu_data *tcd;
 924         int total_pages = 0;
 925
 926         cfs_tracefile_read_lock();
 927
 928         cfs_tcd_for_each(tcd, i, j)
 929                 total_pages += tcd->tcd_max_pages;
 930
 931         cfs_tracefile_read_unlock();
 932
 933         return (total_pages >> (20 - PAGE_SHIFT)) + 1;
 934 }
 935
 936 static int tracefiled(void *arg)
 937 {
 938         struct page_collection pc;
 939         struct tracefiled_ctl *tctl = arg;
 940         struct cfs_trace_page *tage;
 941         struct cfs_trace_page *tmp;
 942         mm_segment_t __oldfs;
 943         struct file *filp;
 944         char *buf;
 945         int last_loop = 0;
 946         int rc;
 947
 948         /* we're started late enough that we pick up init's fs context */
 949         /* this is so broken in uml?  what on earth is going on? */
 950
 951         complete(&tctl->tctl_start);
 952
 953         while (1) {
 954                 wait_queue_t __wait;
 955
 956                 pc.pc_want_daemon_pages = 0;
 957                 collect_pages(&pc);
 958                 if (list_empty(&pc.pc_pages))
 959                         goto end_loop;
 960
 961                 filp = NULL;
 962                 cfs_tracefile_read_lock();
 963                 if (cfs_tracefile[0] != 0) {
 964                         filp = filp_open(cfs_tracefile,
 965                                          O_CREAT | O_RDWR | O_LARGEFILE,
 966                                          0600);
 967                         if (IS_ERR(filp)) {
 968                                 rc = PTR_ERR(filp);
 969                                 filp = NULL;
 970                                 printk(KERN_WARNING "couldn't open %s: "
 971                                        "%d\n", cfs_tracefile, rc);
 972                         }
 973                 }
 974                 cfs_tracefile_read_unlock();
 975                 if (filp == NULL) {
 976                         put_pages_on_daemon_list(&pc);
 977                         __LASSERT(list_empty(&pc.pc_pages));
 978                         goto end_loop;
 979                 }
 980                 __oldfs = get_fs();
 981                 set_fs(get_ds());
 982
 983                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 984                         struct dentry *de = file_dentry(filp);
 985                         static loff_t f_pos;
 986
 987                         __LASSERT_TAGE_INVARIANT(tage);
 988
 989                         if (f_pos >= (off_t)cfs_tracefile_size)
 990                                 f_pos = 0;
 991                         else if (f_pos > i_size_read(de->d_inode))
 992                                 f_pos = i_size_read(de->d_inode);
 993
 994                         buf = kmap(tage->page);
 995                         rc = vfs_write(filp, (__force const char __user *)buf,
 996                                        tage->used, &f_pos);
 997                         kunmap(tage->page);
 998                         if (rc != (int)tage->used) {
 999                                 printk(KERN_WARNING "wanted to write %u "
1000                                        "but wrote %d\n", tage->used, rc);
1001                                 put_pages_back(&pc);
1002                                 __LASSERT(list_empty(&pc.pc_pages));
1003                                 break;
1004                         }
1005                 }
1006                 set_fs(__oldfs);
1007
1008                 filp_close(filp, NULL);
1009                 put_pages_on_daemon_list(&pc);
1010                 if (!list_empty(&pc.pc_pages)) {
1011                         int i;
1012
1013                         printk(KERN_ALERT "Lustre: trace pages aren't "
1014                                " empty\n");
1015                         printk(KERN_ERR "total cpus(%d): ",
1016                                num_possible_cpus());
1017                         for (i = 0; i < num_possible_cpus(); i++)
1018                                 if (cpu_online(i))
1019                                         printk(KERN_ERR "%d(on) ", i);
1020                                 else
1021                                         printk(KERN_ERR "%d(off) ", i);
1022                         printk(KERN_ERR "\n");
1023
1024                         i = 0;
1025                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1026                                                      linkage)
1027                                 printk(KERN_ERR "page %d belongs to cpu "
1028                                        "%d\n", ++i, tage->cpu);
1029                         printk(KERN_ERR "There are %d pages unwritten\n",
1030                                i);
1031                 }
1032                 __LASSERT(list_empty(&pc.pc_pages));
1033 end_loop:
1034                 if (atomic_read(&tctl->tctl_shutdown)) {
1035                         if (last_loop == 0) {
1036                                 last_loop = 1;
1037                                 continue;
1038                         } else {
1039                                 break;
1040                         }
1041                 }
1042                 init_waitqueue_entry(&__wait, current);
1043                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1044                 set_current_state(TASK_INTERRUPTIBLE);
1045                 schedule_timeout(cfs_time_seconds(1));
1046                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1047         }
1048         complete(&tctl->tctl_stop);
1049         return 0;
1050 }
1051
1052 int cfs_trace_start_thread(void)
1053 {
1054         struct tracefiled_ctl *tctl = &trace_tctl;
1055         int rc = 0;
1056
1057         mutex_lock(&cfs_trace_thread_mutex);
1058         if (thread_running)
1059                 goto out;
1060
1061         init_completion(&tctl->tctl_start);
1062         init_completion(&tctl->tctl_stop);
1063         init_waitqueue_head(&tctl->tctl_waitq);
1064         atomic_set(&tctl->tctl_shutdown, 0);
1065
1066         if (IS_ERR(kthread_run(tracefiled, tctl, "ktracefiled"))) {
1067                 rc = -ECHILD;
1068                 goto out;
1069         }
1070
1071         wait_for_completion(&tctl->tctl_start);
1072         thread_running = 1;
1073 out:
1074         mutex_unlock(&cfs_trace_thread_mutex);
1075         return rc;
1076 }
1077
1078 void cfs_trace_stop_thread(void)
1079 {
1080         struct tracefiled_ctl *tctl = &trace_tctl;
1081
1082         mutex_lock(&cfs_trace_thread_mutex);
1083         if (thread_running) {
1084                 printk(KERN_INFO
1085                        "Lustre: shutting down debug daemon thread...\n");
1086                 atomic_set(&tctl->tctl_shutdown, 1);
1087                 wait_for_completion(&tctl->tctl_stop);
1088                 thread_running = 0;
1089         }
1090         mutex_unlock(&cfs_trace_thread_mutex);
1091 }
1092
1093 int cfs_tracefile_init(int max_pages)
1094 {
1095         struct cfs_trace_cpu_data *tcd;
1096         int     i;
1097         int     j;
1098         int     rc;
1099         int     factor;
1100
1101         rc = cfs_tracefile_init_arch();
1102         if (rc != 0)
1103                 return rc;
1104
1105         cfs_tcd_for_each(tcd, i, j) {
1106                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1107                 factor = tcd->tcd_pages_factor;
1108                 INIT_LIST_HEAD(&tcd->tcd_pages);
1109                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1110                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1111                 tcd->tcd_cur_pages = 0;
1112                 tcd->tcd_cur_stock_pages = 0;
1113                 tcd->tcd_cur_daemon_pages = 0;
1114                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1115                 LASSERT(tcd->tcd_max_pages > 0);
1116                 tcd->tcd_shutting_down = 0;
1117         }
1118         return 0;
1119 }
1120
1121 static void trace_cleanup_on_all_cpus(void)
1122 {
1123         struct cfs_trace_cpu_data *tcd;
1124         struct cfs_trace_page *tage;
1125         struct cfs_trace_page *tmp;
1126         int i, cpu;
1127
1128         for_each_possible_cpu(cpu) {
1129                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1130                         tcd->tcd_shutting_down = 1;
1131
1132                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1133                                 __LASSERT_TAGE_INVARIANT(tage);
1134
1135                                 list_del(&tage->linkage);
1136                                 cfs_tage_free(tage);
1137                         }
1138                         tcd->tcd_cur_pages = 0;
1139                 }
1140         }
1141 }
1142
1143 static void cfs_trace_cleanup(void)
1144 {
1145         struct page_collection pc;
1146
1147         INIT_LIST_HEAD(&pc.pc_pages);
1148
1149         trace_cleanup_on_all_cpus();
1150
1151         cfs_tracefile_fini_arch();
1152 }
1153
1154 void cfs_tracefile_exit(void)
1155 {
1156         cfs_trace_stop_thread();
1157         cfs_trace_cleanup();
1158 }