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         struct cfs_trace_cpu_data *tcd = NULL;
 251         struct ptldebug_header     header = {0};
 252         struct cfs_trace_page     *tage;
 253         /* string_buf is used only if tcd != NULL, and is always set then */
 254         char                      *string_buf = NULL;
 255         char                      *debug_buf;
 256         int                        known_size;
 257         int                        needed = 85; /* average message length */
 258         int                        max_nob;
 259         va_list                    ap;
 260         int                        i;
 261         int                        remain;
 262         int                        mask = msgdata->msg_mask;
 263         char                      *file = (char *)msgdata->msg_file;
 264         struct cfs_debug_limit_state *cdls = msgdata->msg_cdls;
 265
 266         if (strchr(file, '/'))
 267                 file = strrchr(file, '/') + 1;
 268
 269         tcd = cfs_trace_get_tcd();
 270
 271         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 272          * pins us to a particular CPU.  This avoids an smp_processor_id()
 273          * warning on Linux when debugging is enabled. */
 274         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 275
 276         if (tcd == NULL)                /* arch may not log in IRQ context */
 277                 goto console;
 278
 279         if (tcd->tcd_cur_pages == 0)
 280                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 281
 282         if (tcd->tcd_shutting_down) {
 283                 cfs_trace_put_tcd(tcd);
 284                 tcd = NULL;
 285                 goto console;
 286         }
 287
 288         known_size = strlen(file) + 1;
 289         if (msgdata->msg_fn)
 290                 known_size += strlen(msgdata->msg_fn) + 1;
 291
 292         if (libcfs_debug_binary)
 293                 known_size += sizeof(header);
 294
 295         /*/
 296          * '2' used because vsnprintf return real size required for output
 297          * _without_ terminating NULL.
 298          * if needed is to small for this format.
 299          */
 300         for (i = 0; i < 2; i++) {
 301                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 302                 if (tage == NULL) {
 303                         if (needed + known_size > PAGE_SIZE)
 304                                 mask |= D_ERROR;
 305
 306                         cfs_trace_put_tcd(tcd);
 307                         tcd = NULL;
 308                         goto console;
 309                 }
 310
 311                 string_buf = (char *)page_address(tage->page) +
 312                              tage->used + known_size;
 313
 314                 max_nob = PAGE_SIZE - tage->used - known_size;
 315                 if (max_nob <= 0) {
 316                         printk(KERN_EMERG "negative max_nob: %d\n",
 317                                max_nob);
 318                         mask |= D_ERROR;
 319                         cfs_trace_put_tcd(tcd);
 320                         tcd = NULL;
 321                         goto console;
 322                 }
 323
 324                 needed = 0;
 325                 remain = max_nob - needed;
 326                 if (remain < 0)
 327                         remain = 0;
 328
 329                 va_start(ap, format);
 330                 needed += vsnprintf(string_buf + needed, remain,
 331                                     format, ap);
 332                 va_end(ap);
 333
 334                 if (needed < max_nob) /* well. printing ok.. */
 335                         break;
 336         }
 337
 338         if (*(string_buf + needed - 1) != '\n') {
 339                 printk(KERN_INFO "format at %s:%d:%s doesn't end in "
 340                        "newline\n", file, msgdata->msg_line, msgdata->msg_fn);
 341         } else if (mask & D_TTY) {
 342                 /* TTY needs '\r\n' to move carriage to leftmost position */
 343                 if (needed < 2 || *(string_buf + needed - 2) != '\r')
 344                         printk(KERN_INFO "format at %s:%d:%s doesn't end in "
 345                                "'\\r\\n'\n", file, msgdata->msg_line,
 346                                msgdata->msg_fn);
 347         }
 348
 349         header.ph_len = known_size + needed;
 350         debug_buf = (char *)page_address(tage->page) + tage->used;
 351
 352         if (libcfs_debug_binary) {
 353                 memcpy(debug_buf, &header, sizeof(header));
 354                 tage->used += sizeof(header);
 355                 debug_buf += sizeof(header);
 356         }
 357
 358         strcpy(debug_buf, file);
 359         tage->used += strlen(file) + 1;
 360         debug_buf += strlen(file) + 1;
 361
 362         if (msgdata->msg_fn) {
 363                 strcpy(debug_buf, msgdata->msg_fn);
 364                 tage->used += strlen(msgdata->msg_fn) + 1;
 365                 debug_buf += strlen(msgdata->msg_fn) + 1;
 366         }
 367
 368         __LASSERT(debug_buf == string_buf);
 369
 370         tage->used += needed;
 371         __LASSERT(tage->used <= PAGE_SIZE);
 372
 373 console:
 374         if ((mask & libcfs_printk) == 0) {
 375                 /* no console output requested */
 376                 if (tcd != NULL)
 377                         cfs_trace_put_tcd(tcd);
 378                 return 1;
 379         }
 380
 381         if (cdls != NULL) {
 382                 if (libcfs_console_ratelimit &&
 383                     cdls->cdls_next != 0 &&     /* not first time ever */
 384                     time_before(jiffies, cdls->cdls_next)) {
 385                         /* skipping a console message */
 386                         cdls->cdls_count++;
 387                         if (tcd != NULL)
 388                                 cfs_trace_put_tcd(tcd);
 389                         return 1;
 390                 }
 391
 392                 if (time_after(jiffies, cdls->cdls_next +
 393                                         libcfs_console_max_delay +
 394                                         cfs_time_seconds(10))) {
 395                         /* last timeout was a long time ago */
 396                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 397                 } else {
 398                         cdls->cdls_delay *= libcfs_console_backoff;
 399                 }
 400
 401                 if (cdls->cdls_delay < libcfs_console_min_delay)
 402                         cdls->cdls_delay = libcfs_console_min_delay;
 403                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 404                         cdls->cdls_delay = libcfs_console_max_delay;
 405
 406                 /* ensure cdls_next is never zero after it's been seen */
 407                 cdls->cdls_next = (jiffies + cdls->cdls_delay) | 1;
 408         }
 409
 410         if (tcd != NULL) {
 411                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 412                                      msgdata->msg_fn);
 413                 cfs_trace_put_tcd(tcd);
 414         } else {
 415                 string_buf = cfs_trace_get_console_buffer();
 416
 417                 needed = 0;
 418                 remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
 419                 if (remain > 0) {
 420                         va_start(ap, format);
 421                         needed += vsnprintf(string_buf+needed, remain,
 422                                             format, ap);
 423                         va_end(ap);
 424                 }
 425
 426                 cfs_print_to_console(&header, mask,
 427                                      string_buf, needed, file, msgdata->msg_fn);
 428
 429                 put_cpu();
 430         }
 431
 432         if (cdls != NULL && cdls->cdls_count != 0) {
 433                 string_buf = cfs_trace_get_console_buffer();
 434
 435                 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 436                                   "Skipped %d previous similar message%s\n",
 437                                   cdls->cdls_count,
 438                                   (cdls->cdls_count > 1) ? "s" : "");
 439
 440                 /* Do not allow print this to TTY */
 441                 cfs_print_to_console(&header, mask & ~D_TTY, string_buf,
 442                                      needed, file, msgdata->msg_fn);
 443
 444                 put_cpu();
 445                 cdls->cdls_count = 0;
 446         }
 447
 448         return 0;
 449 }
 450 EXPORT_SYMBOL(libcfs_debug_msg);
 451
 452 void
 453 cfs_trace_assertion_failed(const char *str,
 454                            struct libcfs_debug_msg_data *msgdata)
 455 {
 456         struct ptldebug_header hdr;
 457
 458         libcfs_panic_in_progress = 1;
 459         libcfs_catastrophe = 1;
 460         smp_mb();
 461
 462         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 463
 464         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 465                              msgdata->msg_file, msgdata->msg_fn);
 466
 467         panic("Lustre debug assertion failure\n");
 468
 469         /* not reached */
 470 }
 471
 472 static void
 473 panic_collect_pages(struct page_collection *pc)
 474 {
 475         /* Do the collect_pages job on a single CPU: assumes that all other
 476          * CPUs have been stopped during a panic.  If this isn't true for some
 477          * arch, this will have to be implemented separately in each arch.  */
 478         int                        i;
 479         int                        j;
 480         struct cfs_trace_cpu_data *tcd;
 481
 482         INIT_LIST_HEAD(&pc->pc_pages);
 483
 484         cfs_tcd_for_each(tcd, i, j) {
 485                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 486                 tcd->tcd_cur_pages = 0;
 487
 488                 if (pc->pc_want_daemon_pages) {
 489                         list_splice_init(&tcd->tcd_daemon_pages,
 490                                                 &pc->pc_pages);
 491                         tcd->tcd_cur_daemon_pages = 0;
 492                 }
 493         }
 494 }
 495
 496 static void collect_pages_on_all_cpus(struct page_collection *pc)
 497 {
 498         struct cfs_trace_cpu_data *tcd;
 499         int i, cpu;
 500
 501         for_each_possible_cpu(cpu) {
 502                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 503                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 504                         tcd->tcd_cur_pages = 0;
 505                         if (pc->pc_want_daemon_pages) {
 506                                 list_splice_init(&tcd->tcd_daemon_pages,
 507                                                         &pc->pc_pages);
 508                                 tcd->tcd_cur_daemon_pages = 0;
 509                         }
 510                 }
 511         }
 512 }
 513
 514 static void collect_pages(struct page_collection *pc)
 515 {
 516         INIT_LIST_HEAD(&pc->pc_pages);
 517
 518         if (libcfs_panic_in_progress)
 519                 panic_collect_pages(pc);
 520         else
 521                 collect_pages_on_all_cpus(pc);
 522 }
 523
 524 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 525 {
 526         struct cfs_trace_cpu_data *tcd;
 527         struct list_head *cur_head;
 528         struct cfs_trace_page *tage;
 529         struct cfs_trace_page *tmp;
 530         int i, cpu;
 531
 532         for_each_possible_cpu(cpu) {
 533                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 534                         cur_head = tcd->tcd_pages.next;
 535
 536                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 537                                                  linkage) {
 538
 539                                 __LASSERT_TAGE_INVARIANT(tage);
 540
 541                                 if (tage->cpu != cpu || tage->type != i)
 542                                         continue;
 543
 544                                 cfs_tage_to_tail(tage, cur_head);
 545                                 tcd->tcd_cur_pages++;
 546                         }
 547                 }
 548         }
 549 }
 550
 551 static void put_pages_back(struct page_collection *pc)
 552 {
 553         if (!libcfs_panic_in_progress)
 554                 put_pages_back_on_all_cpus(pc);
 555 }
 556
 557 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 558  * we have a good amount of data at all times for dumping during an LBUG, even
 559  * if we have been steadily writing (and otherwise discarding) pages via the
 560  * debug daemon. */
 561 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 562                                          struct cfs_trace_cpu_data *tcd)
 563 {
 564         struct cfs_trace_page *tage;
 565         struct cfs_trace_page *tmp;
 566
 567         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 568                 __LASSERT_TAGE_INVARIANT(tage);
 569
 570                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 571                         continue;
 572
 573                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 574                 tcd->tcd_cur_daemon_pages++;
 575
 576                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 577                         struct cfs_trace_page *victim;
 578
 579                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 580                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 581
 582                         __LASSERT_TAGE_INVARIANT(victim);
 583
 584                         list_del(&victim->linkage);
 585                         cfs_tage_free(victim);
 586                         tcd->tcd_cur_daemon_pages--;
 587                 }
 588         }
 589 }
 590
 591 static void put_pages_on_daemon_list(struct page_collection *pc)
 592 {
 593         struct cfs_trace_cpu_data *tcd;
 594         int i, cpu;
 595
 596         for_each_possible_cpu(cpu) {
 597                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 598                         put_pages_on_tcd_daemon_list(pc, tcd);
 599         }
 600 }
 601
 602 void cfs_trace_debug_print(void)
 603 {
 604         struct page_collection pc;
 605         struct cfs_trace_page *tage;
 606         struct cfs_trace_page *tmp;
 607
 608         pc.pc_want_daemon_pages = 1;
 609         collect_pages(&pc);
 610         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 611                 char *p, *file, *fn;
 612                 struct page *page;
 613
 614                 __LASSERT_TAGE_INVARIANT(tage);
 615
 616                 page = tage->page;
 617                 p = page_address(page);
 618                 while (p < ((char *)page_address(page) + tage->used)) {
 619                         struct ptldebug_header *hdr;
 620                         int len;
 621                         hdr = (void *)p;
 622                         p += sizeof(*hdr);
 623                         file = p;
 624                         p += strlen(file) + 1;
 625                         fn = p;
 626                         p += strlen(fn) + 1;
 627                         len = hdr->ph_len - (int)(p - (char *)hdr);
 628
 629                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 630
 631                         p += len;
 632                 }
 633
 634                 list_del(&tage->linkage);
 635                 cfs_tage_free(tage);
 636         }
 637 }
 638
 639 int cfs_tracefile_dump_all_pages(char *filename)
 640 {
 641         struct page_collection  pc;
 642         struct file             *filp;
 643         struct cfs_trace_page   *tage;
 644         struct cfs_trace_page   *tmp;
 645         char                    *buf;
 646         int rc;
 647
 648         cfs_tracefile_write_lock();
 649
 650         filp = filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600);
 651         if (IS_ERR(filp)) {
 652                 rc = PTR_ERR(filp);
 653                 filp = NULL;
 654                 printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 655                       filename, rc);
 656                 goto out;
 657         }
 658
 659         pc.pc_want_daemon_pages = 1;
 660         collect_pages(&pc);
 661         if (list_empty(&pc.pc_pages)) {
 662                 rc = 0;
 663                 goto close;
 664         }
 665
 666         /* ok, for now, just write the pages.  in the future we'll be building
 667          * iobufs with the pages and calling generic_direct_IO */
 668         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 669
 670                 __LASSERT_TAGE_INVARIANT(tage);
 671
 672                 buf = kmap(tage->page);
 673                 rc = cfs_kernel_write(filp, buf, tage->used, &filp->f_pos);
 674                 kunmap(tage->page);
 675                 if (rc != (int)tage->used) {
 676                         printk(KERN_WARNING "wanted to write %u but wrote "
 677                                "%d\n", tage->used, rc);
 678                         put_pages_back(&pc);
 679                         __LASSERT(list_empty(&pc.pc_pages));
 680                         break;
 681                 }
 682                 list_del(&tage->linkage);
 683                 cfs_tage_free(tage);
 684         }
 685
 686         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
 687         if (rc)
 688                 printk(KERN_ERR "sync returns %d\n", rc);
 689 close:
 690         filp_close(filp, NULL);
 691 out:
 692         cfs_tracefile_write_unlock();
 693         return rc;
 694 }
 695
 696 void cfs_trace_flush_pages(void)
 697 {
 698         struct page_collection pc;
 699         struct cfs_trace_page *tage;
 700         struct cfs_trace_page *tmp;
 701
 702         pc.pc_want_daemon_pages = 1;
 703         collect_pages(&pc);
 704         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 705
 706                 __LASSERT_TAGE_INVARIANT(tage);
 707
 708                 list_del(&tage->linkage);
 709                 cfs_tage_free(tage);
 710         }
 711 }
 712
 713 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 714                             const char __user *usr_buffer, int usr_buffer_nob)
 715 {
 716         int    nob;
 717
 718         if (usr_buffer_nob > knl_buffer_nob)
 719                 return -EOVERFLOW;
 720
 721         if (copy_from_user(knl_buffer, usr_buffer, usr_buffer_nob))
 722                 return -EFAULT;
 723
 724         nob = strnlen(knl_buffer, usr_buffer_nob);
 725         while (--nob >= 0)                      /* strip trailing whitespace */
 726                 if (!isspace(knl_buffer[nob]))
 727                         break;
 728
 729         if (nob < 0)                            /* empty string */
 730                 return -EINVAL;
 731
 732         if (nob == knl_buffer_nob)              /* no space to terminate */
 733                 return -EOVERFLOW;
 734
 735         knl_buffer[nob + 1] = 0;                /* terminate */
 736         return 0;
 737 }
 738 EXPORT_SYMBOL(cfs_trace_copyin_string);
 739
 740 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 741                              const char *knl_buffer, char *append)
 742 {
 743         /* NB if 'append' != NULL, it's a single character to append to the
 744          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 745          * terminating zero byte) for sysctl entries */
 746         int   nob = strlen(knl_buffer);
 747
 748         if (nob > usr_buffer_nob)
 749                 nob = usr_buffer_nob;
 750
 751         if (copy_to_user(usr_buffer, knl_buffer, nob))
 752                 return -EFAULT;
 753
 754         if (append != NULL && nob < usr_buffer_nob) {
 755                 if (copy_to_user(usr_buffer + nob, append, 1))
 756                         return -EFAULT;
 757
 758                 nob++;
 759         }
 760
 761         return nob;
 762 }
 763 EXPORT_SYMBOL(cfs_trace_copyout_string);
 764
 765 int cfs_trace_allocate_string_buffer(char **str, int nob)
 766 {
 767         if (nob > 2 * PAGE_SIZE)        /* string must be "sensible" */
 768                 return -EINVAL;
 769
 770         *str = kmalloc(nob, GFP_KERNEL | __GFP_ZERO);
 771         if (*str == NULL)
 772                 return -ENOMEM;
 773
 774         return 0;
 775 }
 776
 777 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
 778 {
 779         char         *str;
 780         int           rc;
 781
 782         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 783         if (rc != 0)
 784                 return rc;
 785
 786         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 787                                      usr_str, usr_str_nob);
 788         if (rc != 0)
 789                 goto out;
 790
 791         if (str[0] != '/') {
 792                 rc = -EINVAL;
 793                 goto out;
 794         }
 795         rc = cfs_tracefile_dump_all_pages(str);
 796 out:
 797         kfree(str);
 798         return rc;
 799 }
 800
 801 int cfs_trace_daemon_command(char *str)
 802 {
 803         int       rc = 0;
 804
 805         cfs_tracefile_write_lock();
 806
 807         if (strcmp(str, "stop") == 0) {
 808                 cfs_tracefile_write_unlock();
 809                 cfs_trace_stop_thread();
 810                 cfs_tracefile_write_lock();
 811                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 812
 813         } else if (strncmp(str, "size=", 5) == 0) {
 814                 unsigned long tmp;
 815
 816                 rc = kstrtoul(str + 5, 10, &tmp);
 817                 if (!rc) {
 818                         if (tmp < 10 || tmp > 20480)
 819                                 cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 820                         else
 821                                 cfs_tracefile_size = tmp << 20;
 822                 }
 823         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 824                 rc = -ENAMETOOLONG;
 825         } else if (str[0] != '/') {
 826                 rc = -EINVAL;
 827         } else {
 828                 strcpy(cfs_tracefile, str);
 829
 830                 printk(KERN_INFO
 831                        "Lustre: debug daemon will attempt to start writing "
 832                        "to %s (%lukB max)\n", cfs_tracefile,
 833                        (long)(cfs_tracefile_size >> 10));
 834
 835                 cfs_trace_start_thread();
 836         }
 837
 838         cfs_tracefile_write_unlock();
 839         return rc;
 840 }
 841
 842 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
 843 {
 844         char *str;
 845         int   rc;
 846
 847         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 848         if (rc != 0)
 849                 return rc;
 850
 851         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 852                                  usr_str, usr_str_nob);
 853         if (rc == 0)
 854                 rc = cfs_trace_daemon_command(str);
 855
 856         kfree(str);
 857         return rc;
 858 }
 859
 860 int cfs_trace_set_debug_mb(int mb)
 861 {
 862         int i;
 863         int j;
 864         int pages;
 865         int limit = cfs_trace_max_debug_mb();
 866         struct cfs_trace_cpu_data *tcd;
 867
 868         if (mb < num_possible_cpus()) {
 869                 printk(KERN_WARNING
 870                        "Lustre: %d MB is too small for debug buffer size, "
 871                        "setting it to %d MB.\n", mb, num_possible_cpus());
 872                 mb = num_possible_cpus();
 873         }
 874
 875         if (mb > limit) {
 876                 printk(KERN_WARNING
 877                        "Lustre: %d MB is too large for debug buffer size, "
 878                        "setting it to %d MB.\n", mb, limit);
 879                 mb = limit;
 880         }
 881
 882         mb /= num_possible_cpus();
 883         pages = mb << (20 - PAGE_SHIFT);
 884
 885         cfs_tracefile_write_lock();
 886
 887         cfs_tcd_for_each(tcd, i, j)
 888                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 889
 890         cfs_tracefile_write_unlock();
 891
 892         return 0;
 893 }
 894
 895 int cfs_trace_get_debug_mb(void)
 896 {
 897         int i;
 898         int j;
 899         struct cfs_trace_cpu_data *tcd;
 900         int total_pages = 0;
 901
 902         cfs_tracefile_read_lock();
 903
 904         cfs_tcd_for_each(tcd, i, j)
 905                 total_pages += tcd->tcd_max_pages;
 906
 907         cfs_tracefile_read_unlock();
 908
 909         return (total_pages >> (20 - PAGE_SHIFT)) + 1;
 910 }
 911
 912 static int tracefiled(void *arg)
 913 {
 914         struct page_collection pc;
 915         struct tracefiled_ctl *tctl = arg;
 916         struct cfs_trace_page *tage;
 917         struct cfs_trace_page *tmp;
 918         struct file *filp;
 919         char *buf;
 920         int last_loop = 0;
 921         int rc;
 922
 923         /* we're started late enough that we pick up init's fs context */
 924         /* this is so broken in uml?  what on earth is going on? */
 925
 926         complete(&tctl->tctl_start);
 927
 928         while (1) {
 929                 wait_queue_entry_t __wait;
 930
 931                 pc.pc_want_daemon_pages = 0;
 932                 collect_pages(&pc);
 933                 if (list_empty(&pc.pc_pages))
 934                         goto end_loop;
 935
 936                 filp = NULL;
 937                 cfs_tracefile_read_lock();
 938                 if (cfs_tracefile[0] != 0) {
 939                         filp = filp_open(cfs_tracefile,
 940                                          O_CREAT | O_RDWR | O_LARGEFILE,
 941                                          0600);
 942                         if (IS_ERR(filp)) {
 943                                 rc = PTR_ERR(filp);
 944                                 filp = NULL;
 945                                 printk(KERN_WARNING "couldn't open %s: "
 946                                        "%d\n", cfs_tracefile, rc);
 947                         }
 948                 }
 949                 cfs_tracefile_read_unlock();
 950                 if (filp == NULL) {
 951                         put_pages_on_daemon_list(&pc);
 952                         __LASSERT(list_empty(&pc.pc_pages));
 953                         goto end_loop;
 954                 }
 955
 956                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 957                         struct dentry *de = file_dentry(filp);
 958                         static loff_t f_pos;
 959
 960                         __LASSERT_TAGE_INVARIANT(tage);
 961
 962                         if (f_pos >= (off_t)cfs_tracefile_size)
 963                                 f_pos = 0;
 964                         else if (f_pos > i_size_read(de->d_inode))
 965                                 f_pos = i_size_read(de->d_inode);
 966
 967                         buf = kmap(tage->page);
 968                         rc = cfs_kernel_write(filp, buf, tage->used, &f_pos);
 969                         kunmap(tage->page);
 970                         if (rc != (int)tage->used) {
 971                                 printk(KERN_WARNING "wanted to write %u "
 972                                        "but wrote %d\n", tage->used, rc);
 973                                 put_pages_back(&pc);
 974                                 __LASSERT(list_empty(&pc.pc_pages));
 975                                 break;
 976                         }
 977                 }
 978
 979                 filp_close(filp, NULL);
 980                 put_pages_on_daemon_list(&pc);
 981                 if (!list_empty(&pc.pc_pages)) {
 982                         int i;
 983
 984                         printk(KERN_ALERT "Lustre: trace pages aren't "
 985                                " empty\n");
 986                         printk(KERN_ERR "total cpus(%d): ",
 987                                num_possible_cpus());
 988                         for (i = 0; i < num_possible_cpus(); i++)
 989                                 if (cpu_online(i))
 990                                         printk(KERN_ERR "%d(on) ", i);
 991                                 else
 992                                         printk(KERN_ERR "%d(off) ", i);
 993                         printk(KERN_ERR "\n");
 994
 995                         i = 0;
 996                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
 997                                                      linkage)
 998                                 printk(KERN_ERR "page %d belongs to cpu "
 999                                        "%d\n", ++i, tage->cpu);
1000                         printk(KERN_ERR "There are %d pages unwritten\n",
1001                                i);
1002                 }
1003                 __LASSERT(list_empty(&pc.pc_pages));
1004 end_loop:
1005                 if (atomic_read(&tctl->tctl_shutdown)) {
1006                         if (last_loop == 0) {
1007                                 last_loop = 1;
1008                                 continue;
1009                         } else {
1010                                 break;
1011                         }
1012                 }
1013                 init_waitqueue_entry(&__wait, current);
1014                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1015                 set_current_state(TASK_INTERRUPTIBLE);
1016                 schedule_timeout(cfs_time_seconds(1));
1017                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1018         }
1019         complete(&tctl->tctl_stop);
1020         return 0;
1021 }
1022
1023 int cfs_trace_start_thread(void)
1024 {
1025         struct tracefiled_ctl *tctl = &trace_tctl;
1026         int rc = 0;
1027
1028         mutex_lock(&cfs_trace_thread_mutex);
1029         if (thread_running)
1030                 goto out;
1031
1032         init_completion(&tctl->tctl_start);
1033         init_completion(&tctl->tctl_stop);
1034         init_waitqueue_head(&tctl->tctl_waitq);
1035         atomic_set(&tctl->tctl_shutdown, 0);
1036
1037         if (IS_ERR(kthread_run(tracefiled, tctl, "ktracefiled"))) {
1038                 rc = -ECHILD;
1039                 goto out;
1040         }
1041
1042         wait_for_completion(&tctl->tctl_start);
1043         thread_running = 1;
1044 out:
1045         mutex_unlock(&cfs_trace_thread_mutex);
1046         return rc;
1047 }
1048
1049 void cfs_trace_stop_thread(void)
1050 {
1051         struct tracefiled_ctl *tctl = &trace_tctl;
1052
1053         mutex_lock(&cfs_trace_thread_mutex);
1054         if (thread_running) {
1055                 printk(KERN_INFO
1056                        "Lustre: shutting down debug daemon thread...\n");
1057                 atomic_set(&tctl->tctl_shutdown, 1);
1058                 wait_for_completion(&tctl->tctl_stop);
1059                 thread_running = 0;
1060         }
1061         mutex_unlock(&cfs_trace_thread_mutex);
1062 }
1063
1064 int cfs_tracefile_init(int max_pages)
1065 {
1066         struct cfs_trace_cpu_data *tcd;
1067         int     i;
1068         int     j;
1069         int     rc;
1070         int     factor;
1071
1072         rc = cfs_tracefile_init_arch();
1073         if (rc != 0)
1074                 return rc;
1075
1076         cfs_tcd_for_each(tcd, i, j) {
1077                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1078                 factor = tcd->tcd_pages_factor;
1079                 INIT_LIST_HEAD(&tcd->tcd_pages);
1080                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1081                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1082                 tcd->tcd_cur_pages = 0;
1083                 tcd->tcd_cur_stock_pages = 0;
1084                 tcd->tcd_cur_daemon_pages = 0;
1085                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1086                 LASSERT(tcd->tcd_max_pages > 0);
1087                 tcd->tcd_shutting_down = 0;
1088         }
1089         return 0;
1090 }
1091
1092 static void trace_cleanup_on_all_cpus(void)
1093 {
1094         struct cfs_trace_cpu_data *tcd;
1095         struct cfs_trace_page *tage;
1096         struct cfs_trace_page *tmp;
1097         int i, cpu;
1098
1099         for_each_possible_cpu(cpu) {
1100                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1101                         tcd->tcd_shutting_down = 1;
1102
1103                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1104                                 __LASSERT_TAGE_INVARIANT(tage);
1105
1106                                 list_del(&tage->linkage);
1107                                 cfs_tage_free(tage);
1108                         }
1109                         tcd->tcd_cur_pages = 0;
1110                 }
1111         }
1112 }
1113
1114 static void cfs_trace_cleanup(void)
1115 {
1116         struct page_collection pc;
1117
1118         INIT_LIST_HEAD(&pc.pc_pages);
1119
1120         trace_cleanup_on_all_cpus();
1121
1122         cfs_tracefile_fini_arch();
1123 }
1124
1125 void cfs_tracefile_exit(void)
1126 {
1127         cfs_trace_stop_thread();
1128         cfs_trace_cleanup();
1129 }