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