drivers/staging/lustre/lnet/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, 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 #define DEBUG_SUBSYSTEM S_LNET
  43 #define LUSTRE_TRACEFILE_PRIVATE
  44 #include "tracefile.h"
  45
  46 #include "../../include/linux/libcfs/libcfs.h"
  47
  48 /* XXX move things up to the top, comment */
  49 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  50
  51 char cfs_tracefile[TRACEFILE_NAME_SIZE];
  52 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
  53 static struct tracefiled_ctl trace_tctl;
  54 static DEFINE_MUTEX(cfs_trace_thread_mutex);
  55 static int thread_running;
  56
  57 static atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
  58
  59 struct page_collection {
  60         struct list_head        pc_pages;
  61         /*
  62          * if this flag is set, collect_pages() will spill both
  63          * ->tcd_daemon_pages and ->tcd_pages to the ->pc_pages. Otherwise,
  64          * only ->tcd_pages are spilled.
  65          */
  66         int             pc_want_daemon_pages;
  67 };
  68
  69 struct tracefiled_ctl {
  70         struct completion       tctl_start;
  71         struct completion       tctl_stop;
  72         wait_queue_head_t               tctl_waitq;
  73         pid_t                   tctl_pid;
  74         atomic_t                tctl_shutdown;
  75 };
  76
  77 /*
  78  * small data-structure for each page owned by tracefiled.
  79  */
  80 struct cfs_trace_page {
  81         /*
  82          * page itself
  83          */
  84         struct page       *page;
  85         /*
  86          * linkage into one of the lists in trace_data_union or
  87          * page_collection
  88          */
  89         struct list_head           linkage;
  90         /*
  91          * number of bytes used within this page
  92          */
  93         unsigned int     used;
  94         /*
  95          * cpu that owns this page
  96          */
  97         unsigned short       cpu;
  98         /*
  99          * type(context) of this page
 100          */
 101         unsigned short       type;
 102 };
 103
 104 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 105                                          struct cfs_trace_cpu_data *tcd);
 106
 107 static inline struct cfs_trace_page *
 108 cfs_tage_from_list(struct list_head *list)
 109 {
 110         return list_entry(list, struct cfs_trace_page, linkage);
 111 }
 112
 113 static struct cfs_trace_page *cfs_tage_alloc(gfp_t gfp)
 114 {
 115         struct page         *page;
 116         struct cfs_trace_page *tage;
 117
 118         /* My caller is trying to free memory */
 119         if (!in_interrupt() && memory_pressure_get())
 120                 return NULL;
 121
 122         /*
 123          * Don't spam console with allocation failures: they will be reported
 124          * by upper layer anyway.
 125          */
 126         gfp |= __GFP_NOWARN;
 127         page = alloc_page(gfp);
 128         if (!page)
 129                 return NULL;
 130
 131         tage = kmalloc(sizeof(*tage), gfp);
 132         if (!tage) {
 133                 __free_page(page);
 134                 return NULL;
 135         }
 136
 137         tage->page = page;
 138         atomic_inc(&cfs_tage_allocated);
 139         return tage;
 140 }
 141
 142 static void cfs_tage_free(struct cfs_trace_page *tage)
 143 {
 144         __free_page(tage->page);
 145         kfree(tage);
 146         atomic_dec(&cfs_tage_allocated);
 147 }
 148
 149 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
 150                              struct list_head *queue)
 151 {
 152         list_move_tail(&tage->linkage, queue);
 153 }
 154
 155 int cfs_trace_refill_stock(struct cfs_trace_cpu_data *tcd, gfp_t gfp,
 156                            struct list_head *stock)
 157 {
 158         int i;
 159
 160         /*
 161          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 162          * from here: this will lead to infinite recursion.
 163          */
 164
 165         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++i) {
 166                 struct cfs_trace_page *tage;
 167
 168                 tage = cfs_tage_alloc(gfp);
 169                 if (!tage)
 170                         break;
 171                 list_add_tail(&tage->linkage, stock);
 172         }
 173         return i;
 174 }
 175
 176 /* return a page that has 'len' bytes left at the end */
 177 static struct cfs_trace_page *
 178 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
 179 {
 180         struct cfs_trace_page *tage;
 181
 182         if (tcd->tcd_cur_pages > 0) {
 183                 __LASSERT(!list_empty(&tcd->tcd_pages));
 184                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
 185                 if (tage->used + len <= PAGE_CACHE_SIZE)
 186                         return tage;
 187         }
 188
 189         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 190                 if (tcd->tcd_cur_stock_pages > 0) {
 191                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
 192                         --tcd->tcd_cur_stock_pages;
 193                         list_del_init(&tage->linkage);
 194                 } else {
 195                         tage = cfs_tage_alloc(GFP_ATOMIC);
 196                         if (unlikely(!tage)) {
 197                                 if ((!memory_pressure_get() ||
 198                                      in_interrupt()) && printk_ratelimit())
 199                                         printk(KERN_WARNING
 200                                                "cannot allocate a tage (%ld)\n",
 201                                                tcd->tcd_cur_pages);
 202                                 return NULL;
 203                         }
 204                 }
 205
 206                 tage->used = 0;
 207                 tage->cpu = smp_processor_id();
 208                 tage->type = tcd->tcd_type;
 209                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 210                 tcd->tcd_cur_pages++;
 211
 212                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 213                         struct tracefiled_ctl *tctl = &trace_tctl;
 214                         /*
 215                          * wake up tracefiled to process some pages.
 216                          */
 217                         wake_up(&tctl->tctl_waitq);
 218                 }
 219                 return tage;
 220         }
 221         return NULL;
 222 }
 223
 224 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
 225 {
 226         int pgcount = tcd->tcd_cur_pages / 10;
 227         struct page_collection pc;
 228         struct cfs_trace_page *tage;
 229         struct cfs_trace_page *tmp;
 230
 231         /*
 232          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 233          * from here: this will lead to infinite recursion.
 234          */
 235
 236         if (printk_ratelimit())
 237                 printk(KERN_WARNING "debug daemon buffer overflowed; discarding 10%% of pages (%d of %ld)\n",
 238                        pgcount + 1, tcd->tcd_cur_pages);
 239
 240         INIT_LIST_HEAD(&pc.pc_pages);
 241
 242         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 243                 if (pgcount-- == 0)
 244                         break;
 245
 246                 list_move_tail(&tage->linkage, &pc.pc_pages);
 247                 tcd->tcd_cur_pages--;
 248         }
 249         put_pages_on_tcd_daemon_list(&pc, tcd);
 250 }
 251
 252 /* return a page that has 'len' bytes left at the end */
 253 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 254                                                  unsigned long len)
 255 {
 256         struct cfs_trace_page *tage;
 257
 258         /*
 259          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 260          * from here: this will lead to infinite recursion.
 261          */
 262
 263         if (len > PAGE_CACHE_SIZE) {
 264                 pr_err("cowardly refusing to write %lu bytes in a page\n", len);
 265                 return NULL;
 266         }
 267
 268         tage = cfs_trace_get_tage_try(tcd, len);
 269         if (tage)
 270                 return tage;
 271         if (thread_running)
 272                 cfs_tcd_shrink(tcd);
 273         if (tcd->tcd_cur_pages > 0) {
 274                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 275                 tage->used = 0;
 276                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 277         }
 278         return tage;
 279 }
 280
 281 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 282                      const char *format, ...)
 283 {
 284         va_list args;
 285         int     rc;
 286
 287         va_start(args, format);
 288         rc = libcfs_debug_vmsg2(msgdata, format, args, NULL);
 289         va_end(args);
 290
 291         return rc;
 292 }
 293 EXPORT_SYMBOL(libcfs_debug_msg);
 294
 295 int libcfs_debug_vmsg2(struct libcfs_debug_msg_data *msgdata,
 296                        const char *format1, va_list args,
 297                        const char *format2, ...)
 298 {
 299         struct cfs_trace_cpu_data *tcd = NULL;
 300         struct ptldebug_header     header = {0};
 301         struct cfs_trace_page     *tage;
 302         /* string_buf is used only if tcd != NULL, and is always set then */
 303         char                  *string_buf = NULL;
 304         char                  *debug_buf;
 305         int                     known_size;
 306         int                     needed = 85; /* average message length */
 307         int                     max_nob;
 308         va_list             ap;
 309         int                     depth;
 310         int                     i;
 311         int                     remain;
 312         int                     mask = msgdata->msg_mask;
 313         const char              *file = kbasename(msgdata->msg_file);
 314         struct cfs_debug_limit_state   *cdls = msgdata->msg_cdls;
 315
 316         tcd = cfs_trace_get_tcd();
 317
 318         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 319          * pins us to a particular CPU.  This avoids an smp_processor_id()
 320          * warning on Linux when debugging is enabled.
 321          */
 322         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 323
 324         if (!tcd)               /* arch may not log in IRQ context */
 325                 goto console;
 326
 327         if (tcd->tcd_cur_pages == 0)
 328                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 329
 330         if (tcd->tcd_shutting_down) {
 331                 cfs_trace_put_tcd(tcd);
 332                 tcd = NULL;
 333                 goto console;
 334         }
 335
 336         depth = __current_nesting_level();
 337         known_size = strlen(file) + 1 + depth;
 338         if (msgdata->msg_fn)
 339                 known_size += strlen(msgdata->msg_fn) + 1;
 340
 341         if (libcfs_debug_binary)
 342                 known_size += sizeof(header);
 343
 344         /*
 345          * '2' used because vsnprintf return real size required for output
 346          * _without_ terminating NULL.
 347          * if needed is to small for this format.
 348          */
 349         for (i = 0; i < 2; i++) {
 350                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 351                 if (!tage) {
 352                         if (needed + known_size > PAGE_CACHE_SIZE)
 353                                 mask |= D_ERROR;
 354
 355                         cfs_trace_put_tcd(tcd);
 356                         tcd = NULL;
 357                         goto console;
 358                 }
 359
 360                 string_buf = (char *)page_address(tage->page) +
 361                                         tage->used + known_size;
 362
 363                 max_nob = PAGE_CACHE_SIZE - tage->used - known_size;
 364                 if (max_nob <= 0) {
 365                         printk(KERN_EMERG "negative max_nob: %d\n",
 366                                max_nob);
 367                         mask |= D_ERROR;
 368                         cfs_trace_put_tcd(tcd);
 369                         tcd = NULL;
 370                         goto console;
 371                 }
 372
 373                 needed = 0;
 374                 if (format1) {
 375                         va_copy(ap, args);
 376                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 377                         va_end(ap);
 378                 }
 379
 380                 if (format2) {
 381                         remain = max_nob - needed;
 382                         if (remain < 0)
 383                                 remain = 0;
 384
 385                         va_start(ap, format2);
 386                         needed += vsnprintf(string_buf + needed, remain,
 387                                             format2, ap);
 388                         va_end(ap);
 389                 }
 390
 391                 if (needed < max_nob) /* well. printing ok.. */
 392                         break;
 393         }
 394
 395         if (*(string_buf + needed - 1) != '\n')
 396                 printk(KERN_INFO "format at %s:%d:%s doesn't end in newline\n",
 397                        file, msgdata->msg_line, msgdata->msg_fn);
 398
 399         header.ph_len = known_size + needed;
 400         debug_buf = (char *)page_address(tage->page) + tage->used;
 401
 402         if (libcfs_debug_binary) {
 403                 memcpy(debug_buf, &header, sizeof(header));
 404                 tage->used += sizeof(header);
 405                 debug_buf += sizeof(header);
 406         }
 407
 408         /* indent message according to the nesting level */
 409         while (depth-- > 0) {
 410                 *(debug_buf++) = '.';
 411                 ++tage->used;
 412         }
 413
 414         strcpy(debug_buf, file);
 415         tage->used += strlen(file) + 1;
 416         debug_buf += strlen(file) + 1;
 417
 418         if (msgdata->msg_fn) {
 419                 strcpy(debug_buf, msgdata->msg_fn);
 420                 tage->used += strlen(msgdata->msg_fn) + 1;
 421                 debug_buf += strlen(msgdata->msg_fn) + 1;
 422         }
 423
 424         __LASSERT(debug_buf == string_buf);
 425
 426         tage->used += needed;
 427         __LASSERT(tage->used <= PAGE_CACHE_SIZE);
 428
 429 console:
 430         if ((mask & libcfs_printk) == 0) {
 431                 /* no console output requested */
 432                 if (tcd)
 433                         cfs_trace_put_tcd(tcd);
 434                 return 1;
 435         }
 436
 437         if (cdls) {
 438                 if (libcfs_console_ratelimit &&
 439                     cdls->cdls_next != 0 &&     /* not first time ever */
 440                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 441                         /* skipping a console message */
 442                         cdls->cdls_count++;
 443                         if (tcd)
 444                                 cfs_trace_put_tcd(tcd);
 445                         return 1;
 446                 }
 447
 448                 if (cfs_time_after(cfs_time_current(),
 449                                    cdls->cdls_next + libcfs_console_max_delay +
 450                                    cfs_time_seconds(10))) {
 451                         /* last timeout was a long time ago */
 452                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 453                 } else {
 454                         cdls->cdls_delay *= libcfs_console_backoff;
 455                 }
 456
 457                 if (cdls->cdls_delay < libcfs_console_min_delay)
 458                         cdls->cdls_delay = libcfs_console_min_delay;
 459                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 460                         cdls->cdls_delay = libcfs_console_max_delay;
 461
 462                 /* ensure cdls_next is never zero after it's been seen */
 463                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 464         }
 465
 466         if (tcd) {
 467                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 468                                      msgdata->msg_fn);
 469                 cfs_trace_put_tcd(tcd);
 470         } else {
 471                 string_buf = cfs_trace_get_console_buffer();
 472
 473                 needed = 0;
 474                 if (format1) {
 475                         va_copy(ap, args);
 476                         needed = vsnprintf(string_buf,
 477                                            CFS_TRACE_CONSOLE_BUFFER_SIZE,
 478                                            format1, ap);
 479                         va_end(ap);
 480                 }
 481                 if (format2) {
 482                         remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
 483                         if (remain > 0) {
 484                                 va_start(ap, format2);
 485                                 needed += vsnprintf(string_buf + needed, remain,
 486                                                     format2, ap);
 487                                 va_end(ap);
 488                         }
 489                 }
 490                 cfs_print_to_console(&header, mask,
 491                                      string_buf, needed, file, msgdata->msg_fn);
 492
 493                 put_cpu();
 494         }
 495
 496         if (cdls && cdls->cdls_count != 0) {
 497                 string_buf = cfs_trace_get_console_buffer();
 498
 499                 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 500                                   "Skipped %d previous similar message%s\n",
 501                                   cdls->cdls_count,
 502                                   (cdls->cdls_count > 1) ? "s" : "");
 503
 504                 cfs_print_to_console(&header, mask,
 505                                      string_buf, needed, file, msgdata->msg_fn);
 506
 507                 put_cpu();
 508                 cdls->cdls_count = 0;
 509         }
 510
 511         return 0;
 512 }
 513 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 514
 515 void
 516 cfs_trace_assertion_failed(const char *str,
 517                            struct libcfs_debug_msg_data *msgdata)
 518 {
 519         struct ptldebug_header hdr;
 520
 521         libcfs_panic_in_progress = 1;
 522         libcfs_catastrophe = 1;
 523         mb();
 524
 525         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 526
 527         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 528                              msgdata->msg_file, msgdata->msg_fn);
 529
 530         panic("Lustre debug assertion failure\n");
 531
 532         /* not reached */
 533 }
 534
 535 static void
 536 panic_collect_pages(struct page_collection *pc)
 537 {
 538         /* Do the collect_pages job on a single CPU: assumes that all other
 539          * CPUs have been stopped during a panic.  If this isn't true for some
 540          * arch, this will have to be implemented separately in each arch.
 541          */
 542         int                     i;
 543         int                     j;
 544         struct cfs_trace_cpu_data *tcd;
 545
 546         INIT_LIST_HEAD(&pc->pc_pages);
 547
 548         cfs_tcd_for_each(tcd, i, j) {
 549                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 550                 tcd->tcd_cur_pages = 0;
 551
 552                 if (pc->pc_want_daemon_pages) {
 553                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 554                         tcd->tcd_cur_daemon_pages = 0;
 555                 }
 556         }
 557 }
 558
 559 static void collect_pages_on_all_cpus(struct page_collection *pc)
 560 {
 561         struct cfs_trace_cpu_data *tcd;
 562         int i, cpu;
 563
 564         for_each_possible_cpu(cpu) {
 565                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 566                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 567                         tcd->tcd_cur_pages = 0;
 568                         if (pc->pc_want_daemon_pages) {
 569                                 list_splice_init(&tcd->tcd_daemon_pages,
 570                                                  &pc->pc_pages);
 571                                 tcd->tcd_cur_daemon_pages = 0;
 572                         }
 573                 }
 574         }
 575 }
 576
 577 static void collect_pages(struct page_collection *pc)
 578 {
 579         INIT_LIST_HEAD(&pc->pc_pages);
 580
 581         if (libcfs_panic_in_progress)
 582                 panic_collect_pages(pc);
 583         else
 584                 collect_pages_on_all_cpus(pc);
 585 }
 586
 587 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 588 {
 589         struct cfs_trace_cpu_data *tcd;
 590         struct list_head *cur_head;
 591         struct cfs_trace_page *tage;
 592         struct cfs_trace_page *tmp;
 593         int i, cpu;
 594
 595         for_each_possible_cpu(cpu) {
 596                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 597                         cur_head = tcd->tcd_pages.next;
 598
 599                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 600                                                  linkage) {
 601                                 __LASSERT_TAGE_INVARIANT(tage);
 602
 603                                 if (tage->cpu != cpu || tage->type != i)
 604                                         continue;
 605
 606                                 cfs_tage_to_tail(tage, cur_head);
 607                                 tcd->tcd_cur_pages++;
 608                         }
 609                 }
 610         }
 611 }
 612
 613 static void put_pages_back(struct page_collection *pc)
 614 {
 615         if (!libcfs_panic_in_progress)
 616                 put_pages_back_on_all_cpus(pc);
 617 }
 618
 619 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 620  * we have a good amount of data at all times for dumping during an LBUG, even
 621  * if we have been steadily writing (and otherwise discarding) pages via the
 622  * debug daemon.
 623  */
 624 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 625                                          struct cfs_trace_cpu_data *tcd)
 626 {
 627         struct cfs_trace_page *tage;
 628         struct cfs_trace_page *tmp;
 629
 630         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 631                 __LASSERT_TAGE_INVARIANT(tage);
 632
 633                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 634                         continue;
 635
 636                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 637                 tcd->tcd_cur_daemon_pages++;
 638
 639                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 640                         struct cfs_trace_page *victim;
 641
 642                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 643                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 644
 645                         __LASSERT_TAGE_INVARIANT(victim);
 646
 647                         list_del(&victim->linkage);
 648                         cfs_tage_free(victim);
 649                         tcd->tcd_cur_daemon_pages--;
 650                 }
 651         }
 652 }
 653
 654 static void put_pages_on_daemon_list(struct page_collection *pc)
 655 {
 656         struct cfs_trace_cpu_data *tcd;
 657         int i, cpu;
 658
 659         for_each_possible_cpu(cpu) {
 660                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 661                         put_pages_on_tcd_daemon_list(pc, tcd);
 662         }
 663 }
 664
 665 void cfs_trace_debug_print(void)
 666 {
 667         struct page_collection pc;
 668         struct cfs_trace_page *tage;
 669         struct cfs_trace_page *tmp;
 670
 671         pc.pc_want_daemon_pages = 1;
 672         collect_pages(&pc);
 673         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 674                 char *p, *file, *fn;
 675                 struct page *page;
 676
 677                 __LASSERT_TAGE_INVARIANT(tage);
 678
 679                 page = tage->page;
 680                 p = page_address(page);
 681                 while (p < ((char *)page_address(page) + tage->used)) {
 682                         struct ptldebug_header *hdr;
 683                         int len;
 684
 685                         hdr = (void *)p;
 686                         p += sizeof(*hdr);
 687                         file = p;
 688                         p += strlen(file) + 1;
 689                         fn = p;
 690                         p += strlen(fn) + 1;
 691                         len = hdr->ph_len - (int)(p - (char *)hdr);
 692
 693                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 694
 695                         p += len;
 696                 }
 697
 698                 list_del(&tage->linkage);
 699                 cfs_tage_free(tage);
 700         }
 701 }
 702
 703 int cfs_tracefile_dump_all_pages(char *filename)
 704 {
 705         struct page_collection  pc;
 706         struct file             *filp;
 707         struct cfs_trace_page   *tage;
 708         struct cfs_trace_page   *tmp;
 709         char                    *buf;
 710         int rc;
 711
 712         DECL_MMSPACE;
 713
 714         cfs_tracefile_write_lock();
 715
 716         filp = filp_open(filename, O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE,
 717                          0600);
 718         if (IS_ERR(filp)) {
 719                 rc = PTR_ERR(filp);
 720                 filp = NULL;
 721                 pr_err("LustreError: can't open %s for dump: rc %d\n",
 722                        filename, rc);
 723                 goto out;
 724         }
 725
 726         pc.pc_want_daemon_pages = 1;
 727         collect_pages(&pc);
 728         if (list_empty(&pc.pc_pages)) {
 729                 rc = 0;
 730                 goto close;
 731         }
 732
 733         /* ok, for now, just write the pages.  in the future we'll be building
 734          * iobufs with the pages and calling generic_direct_IO
 735          */
 736         MMSPACE_OPEN;
 737         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 738                 __LASSERT_TAGE_INVARIANT(tage);
 739
 740                 buf = kmap(tage->page);
 741                 rc = vfs_write(filp, (__force const char __user *)buf,
 742                                tage->used, &filp->f_pos);
 743                 kunmap(tage->page);
 744
 745                 if (rc != (int)tage->used) {
 746                         printk(KERN_WARNING "wanted to write %u but wrote %d\n",
 747                                tage->used, rc);
 748                         put_pages_back(&pc);
 749                         __LASSERT(list_empty(&pc.pc_pages));
 750                         break;
 751                 }
 752                 list_del(&tage->linkage);
 753                 cfs_tage_free(tage);
 754         }
 755         MMSPACE_CLOSE;
 756         rc = vfs_fsync(filp, 1);
 757         if (rc)
 758                 pr_err("sync returns %d\n", rc);
 759 close:
 760         filp_close(filp, NULL);
 761 out:
 762         cfs_tracefile_write_unlock();
 763         return rc;
 764 }
 765
 766 void cfs_trace_flush_pages(void)
 767 {
 768         struct page_collection pc;
 769         struct cfs_trace_page *tage;
 770         struct cfs_trace_page *tmp;
 771
 772         pc.pc_want_daemon_pages = 1;
 773         collect_pages(&pc);
 774         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 775                 __LASSERT_TAGE_INVARIANT(tage);
 776
 777                 list_del(&tage->linkage);
 778                 cfs_tage_free(tage);
 779         }
 780 }
 781
 782 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 783                             const char __user *usr_buffer, int usr_buffer_nob)
 784 {
 785         int    nob;
 786
 787         if (usr_buffer_nob > knl_buffer_nob)
 788                 return -EOVERFLOW;
 789
 790         if (copy_from_user((void *)knl_buffer,
 791                            usr_buffer, usr_buffer_nob))
 792                 return -EFAULT;
 793
 794         nob = strnlen(knl_buffer, usr_buffer_nob);
 795         while (nob-- >= 0)                    /* strip trailing whitespace */
 796                 if (!isspace(knl_buffer[nob]))
 797                         break;
 798
 799         if (nob < 0)                        /* empty string */
 800                 return -EINVAL;
 801
 802         if (nob == knl_buffer_nob)            /* no space to terminate */
 803                 return -EOVERFLOW;
 804
 805         knl_buffer[nob + 1] = 0;                /* terminate */
 806         return 0;
 807 }
 808 EXPORT_SYMBOL(cfs_trace_copyin_string);
 809
 810 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 811                              const char *knl_buffer, char *append)
 812 {
 813         /*
 814          * NB if 'append' != NULL, it's a single character to append to the
 815          * copied out string - usually "\n" or "" (i.e. a terminating zero byte)
 816          */
 817         int   nob = strlen(knl_buffer);
 818
 819         if (nob > usr_buffer_nob)
 820                 nob = usr_buffer_nob;
 821
 822         if (copy_to_user(usr_buffer, knl_buffer, nob))
 823                 return -EFAULT;
 824
 825         if (append && nob < usr_buffer_nob) {
 826                 if (copy_to_user(usr_buffer + nob, append, 1))
 827                         return -EFAULT;
 828
 829                 nob++;
 830         }
 831
 832         return nob;
 833 }
 834 EXPORT_SYMBOL(cfs_trace_copyout_string);
 835
 836 int cfs_trace_allocate_string_buffer(char **str, int nob)
 837 {
 838         if (nob > 2 * PAGE_CACHE_SIZE)      /* string must be "sensible" */
 839                 return -EINVAL;
 840
 841         *str = kmalloc(nob, GFP_KERNEL | __GFP_ZERO);
 842         if (!*str)
 843                 return -ENOMEM;
 844
 845         return 0;
 846 }
 847
 848 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
 849 {
 850         char     *str;
 851         int        rc;
 852
 853         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 854         if (rc != 0)
 855                 return rc;
 856
 857         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 858                                      usr_str, usr_str_nob);
 859         if (rc != 0)
 860                 goto out;
 861
 862         if (str[0] != '/') {
 863                 rc = -EINVAL;
 864                 goto out;
 865         }
 866         rc = cfs_tracefile_dump_all_pages(str);
 867 out:
 868         kfree(str);
 869         return rc;
 870 }
 871
 872 int cfs_trace_daemon_command(char *str)
 873 {
 874         int       rc = 0;
 875
 876         cfs_tracefile_write_lock();
 877
 878         if (strcmp(str, "stop") == 0) {
 879                 cfs_tracefile_write_unlock();
 880                 cfs_trace_stop_thread();
 881                 cfs_tracefile_write_lock();
 882                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 883
 884         } else if (strncmp(str, "size=", 5) == 0) {
 885                 unsigned long tmp;
 886
 887                 rc = kstrtoul(str + 5, 10, &tmp);
 888                 if (!rc) {
 889                         if (tmp < 10 || tmp > 20480)
 890                                 cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 891                         else
 892                                 cfs_tracefile_size = tmp << 20;
 893                 }
 894         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 895                 rc = -ENAMETOOLONG;
 896         } else if (str[0] != '/') {
 897                 rc = -EINVAL;
 898         } else {
 899                 strcpy(cfs_tracefile, str);
 900
 901                 printk(KERN_INFO
 902                        "Lustre: debug daemon will attempt to start writing to %s (%lukB max)\n",
 903                        cfs_tracefile,
 904                        (long)(cfs_tracefile_size >> 10));
 905
 906                 cfs_trace_start_thread();
 907         }
 908
 909         cfs_tracefile_write_unlock();
 910         return rc;
 911 }
 912
 913 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
 914 {
 915         char *str;
 916         int   rc;
 917
 918         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 919         if (rc != 0)
 920                 return rc;
 921
 922         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 923                                      usr_str, usr_str_nob);
 924         if (rc == 0)
 925                 rc = cfs_trace_daemon_command(str);
 926
 927         kfree(str);
 928         return rc;
 929 }
 930
 931 int cfs_trace_set_debug_mb(int mb)
 932 {
 933         int i;
 934         int j;
 935         int pages;
 936         int limit = cfs_trace_max_debug_mb();
 937         struct cfs_trace_cpu_data *tcd;
 938
 939         if (mb < num_possible_cpus()) {
 940                 printk(KERN_WARNING
 941                        "Lustre: %d MB is too small for debug buffer size, setting it to %d MB.\n",
 942                        mb, num_possible_cpus());
 943                 mb = num_possible_cpus();
 944         }
 945
 946         if (mb > limit) {
 947                 printk(KERN_WARNING
 948                        "Lustre: %d MB is too large for debug buffer size, setting it to %d MB.\n",
 949                        mb, limit);
 950                 mb = limit;
 951         }
 952
 953         mb /= num_possible_cpus();
 954         pages = mb << (20 - PAGE_CACHE_SHIFT);
 955
 956         cfs_tracefile_write_lock();
 957
 958         cfs_tcd_for_each(tcd, i, j)
 959                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 960
 961         cfs_tracefile_write_unlock();
 962
 963         return 0;
 964 }
 965
 966 int cfs_trace_get_debug_mb(void)
 967 {
 968         int i;
 969         int j;
 970         struct cfs_trace_cpu_data *tcd;
 971         int total_pages = 0;
 972
 973         cfs_tracefile_read_lock();
 974
 975         cfs_tcd_for_each(tcd, i, j)
 976                 total_pages += tcd->tcd_max_pages;
 977
 978         cfs_tracefile_read_unlock();
 979
 980         return (total_pages >> (20 - PAGE_CACHE_SHIFT)) + 1;
 981 }
 982
 983 static int tracefiled(void *arg)
 984 {
 985         struct page_collection pc;
 986         struct tracefiled_ctl *tctl = arg;
 987         struct cfs_trace_page *tage;
 988         struct cfs_trace_page *tmp;
 989         struct file *filp;
 990         char *buf;
 991         int last_loop = 0;
 992         int rc;
 993
 994         DECL_MMSPACE;
 995
 996         /* we're started late enough that we pick up init's fs context */
 997         /* this is so broken in uml?  what on earth is going on? */
 998
 999         complete(&tctl->tctl_start);
1000
1001         while (1) {
1002                 wait_queue_t __wait;
1003
1004                 pc.pc_want_daemon_pages = 0;
1005                 collect_pages(&pc);
1006                 if (list_empty(&pc.pc_pages))
1007                         goto end_loop;
1008
1009                 filp = NULL;
1010                 cfs_tracefile_read_lock();
1011                 if (cfs_tracefile[0] != 0) {
1012                         filp = filp_open(cfs_tracefile,
1013                                          O_CREAT | O_RDWR | O_LARGEFILE,
1014                                          0600);
1015                         if (IS_ERR(filp)) {
1016                                 rc = PTR_ERR(filp);
1017                                 filp = NULL;
1018                                 printk(KERN_WARNING "couldn't open %s: %d\n",
1019                                        cfs_tracefile, rc);
1020                         }
1021                 }
1022                 cfs_tracefile_read_unlock();
1023                 if (!filp) {
1024                         put_pages_on_daemon_list(&pc);
1025                         __LASSERT(list_empty(&pc.pc_pages));
1026                         goto end_loop;
1027                 }
1028
1029                 MMSPACE_OPEN;
1030
1031                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
1032                         static loff_t f_pos;
1033
1034                         __LASSERT_TAGE_INVARIANT(tage);
1035
1036                         if (f_pos >= (off_t)cfs_tracefile_size)
1037                                 f_pos = 0;
1038                         else if (f_pos > i_size_read(file_inode(filp)))
1039                                 f_pos = i_size_read(file_inode(filp));
1040
1041                         buf = kmap(tage->page);
1042                         rc = vfs_write(filp, (__force const char __user *)buf,
1043                                        tage->used, &f_pos);
1044                         kunmap(tage->page);
1045
1046                         if (rc != (int)tage->used) {
1047                                 printk(KERN_WARNING "wanted to write %u but wrote %d\n",
1048                                        tage->used, rc);
1049                                 put_pages_back(&pc);
1050                                 __LASSERT(list_empty(&pc.pc_pages));
1051                                 break;
1052                         }
1053                 }
1054                 MMSPACE_CLOSE;
1055
1056                 filp_close(filp, NULL);
1057                 put_pages_on_daemon_list(&pc);
1058                 if (!list_empty(&pc.pc_pages)) {
1059                         int i;
1060
1061                         printk(KERN_ALERT "Lustre: trace pages aren't empty\n");
1062                         pr_err("total cpus(%d): ", num_possible_cpus());
1063                         for (i = 0; i < num_possible_cpus(); i++)
1064                                 if (cpu_online(i))
1065                                         pr_cont("%d(on) ", i);
1066                                 else
1067                                         pr_cont("%d(off) ", i);
1068                         pr_cont("\n");
1069
1070                         i = 0;
1071                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1072                                                  linkage)
1073                                 pr_err("page %d belongs to cpu %d\n",
1074                                        ++i, tage->cpu);
1075                         pr_err("There are %d pages unwritten\n", i);
1076                 }
1077                 __LASSERT(list_empty(&pc.pc_pages));
1078 end_loop:
1079                 if (atomic_read(&tctl->tctl_shutdown)) {
1080                         if (last_loop == 0) {
1081                                 last_loop = 1;
1082                                 continue;
1083                         } else {
1084                                 break;
1085                         }
1086                 }
1087                 init_waitqueue_entry(&__wait, current);
1088                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1089                 set_current_state(TASK_INTERRUPTIBLE);
1090                 schedule_timeout(cfs_time_seconds(1));
1091                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1092         }
1093         complete(&tctl->tctl_stop);
1094         return 0;
1095 }
1096
1097 int cfs_trace_start_thread(void)
1098 {
1099         struct tracefiled_ctl *tctl = &trace_tctl;
1100         struct task_struct *task;
1101         int rc = 0;
1102
1103         mutex_lock(&cfs_trace_thread_mutex);
1104         if (thread_running)
1105                 goto out;
1106
1107         init_completion(&tctl->tctl_start);
1108         init_completion(&tctl->tctl_stop);
1109         init_waitqueue_head(&tctl->tctl_waitq);
1110         atomic_set(&tctl->tctl_shutdown, 0);
1111
1112         task = kthread_run(tracefiled, tctl, "ktracefiled");
1113         if (IS_ERR(task)) {
1114                 rc = PTR_ERR(task);
1115                 goto out;
1116         }
1117
1118         wait_for_completion(&tctl->tctl_start);
1119         thread_running = 1;
1120 out:
1121         mutex_unlock(&cfs_trace_thread_mutex);
1122         return rc;
1123 }
1124
1125 void cfs_trace_stop_thread(void)
1126 {
1127         struct tracefiled_ctl *tctl = &trace_tctl;
1128
1129         mutex_lock(&cfs_trace_thread_mutex);
1130         if (thread_running) {
1131                 printk(KERN_INFO
1132                        "Lustre: shutting down debug daemon thread...\n");
1133                 atomic_set(&tctl->tctl_shutdown, 1);
1134                 wait_for_completion(&tctl->tctl_stop);
1135                 thread_running = 0;
1136         }
1137         mutex_unlock(&cfs_trace_thread_mutex);
1138 }
1139
1140 int cfs_tracefile_init(int max_pages)
1141 {
1142         struct cfs_trace_cpu_data *tcd;
1143         int                 i;
1144         int                 j;
1145         int                 rc;
1146         int                 factor;
1147
1148         rc = cfs_tracefile_init_arch();
1149         if (rc != 0)
1150                 return rc;
1151
1152         cfs_tcd_for_each(tcd, i, j) {
1153                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1154                 factor = tcd->tcd_pages_factor;
1155                 INIT_LIST_HEAD(&tcd->tcd_pages);
1156                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1157                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1158                 tcd->tcd_cur_pages = 0;
1159                 tcd->tcd_cur_stock_pages = 0;
1160                 tcd->tcd_cur_daemon_pages = 0;
1161                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1162                 LASSERT(tcd->tcd_max_pages > 0);
1163                 tcd->tcd_shutting_down = 0;
1164         }
1165
1166         return 0;
1167 }
1168
1169 static void trace_cleanup_on_all_cpus(void)
1170 {
1171         struct cfs_trace_cpu_data *tcd;
1172         struct cfs_trace_page *tage;
1173         struct cfs_trace_page *tmp;
1174         int i, cpu;
1175
1176         for_each_possible_cpu(cpu) {
1177                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1178                         tcd->tcd_shutting_down = 1;
1179
1180                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages,
1181                                                  linkage) {
1182                                 __LASSERT_TAGE_INVARIANT(tage);
1183
1184                                 list_del(&tage->linkage);
1185                                 cfs_tage_free(tage);
1186                         }
1187
1188                         tcd->tcd_cur_pages = 0;
1189                 }
1190         }
1191 }
1192
1193 static void cfs_trace_cleanup(void)
1194 {
1195         struct page_collection pc;
1196
1197         INIT_LIST_HEAD(&pc.pc_pages);
1198
1199         trace_cleanup_on_all_cpus();
1200
1201         cfs_tracefile_fini_arch();
1202 }
1203
1204 void cfs_tracefile_exit(void)
1205 {
1206         cfs_trace_stop_thread();
1207         cfs_trace_cleanup();
1208 }