libcfs/libcfs/tracefile.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * libcfs/libcfs/tracefile.c
  33  *
  34  * Author: Zach Brown <zab@clusterfs.com>
  35  * Author: Phil Schwan <phil@clusterfs.com>
  36  */
  37
  38
  39 #define DEBUG_SUBSYSTEM S_LNET
  40 #define LUSTRE_TRACEFILE_PRIVATE
  41 #include "tracefile.h"
  42
  43 #include <linux/ctype.h>
  44 #include <linux/fs.h>
  45 #include <linux/kthread.h>
  46 #include <linux/pagemap.h>
  47 #include <linux/uaccess.h>
  48 #include <libcfs/linux/linux-fs.h>
  49 #include <libcfs/libcfs.h>
  50
  51 /* XXX move things up to the top, comment */
  52 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  53
  54 char cfs_tracefile[TRACEFILE_NAME_SIZE];
  55 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
  56 static struct tracefiled_ctl trace_tctl;
  57 static DEFINE_MUTEX(cfs_trace_thread_mutex);
  58 static int thread_running = 0;
  59
  60 static atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
  61
  62 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  63                                         struct cfs_trace_cpu_data *tcd);
  64
  65 static inline struct cfs_trace_page *
  66 cfs_tage_from_list(struct list_head *list)
  67 {
  68         return list_entry(list, struct cfs_trace_page, linkage);
  69 }
  70
  71 static struct cfs_trace_page *cfs_tage_alloc(gfp_t gfp)
  72 {
  73         struct page            *page;
  74         struct cfs_trace_page *tage;
  75
  76         /* My caller is trying to free memory */
  77         if (!in_interrupt() && memory_pressure_get())
  78                 return NULL;
  79
  80         /*
  81          * Don't spam console with allocation failures: they will be reported
  82          * by upper layer anyway.
  83          */
  84         gfp |= __GFP_NOWARN;
  85         page = alloc_page(gfp);
  86         if (page == NULL)
  87                 return NULL;
  88
  89         tage = kmalloc(sizeof(*tage), gfp);
  90         if (tage == NULL) {
  91                 __free_page(page);
  92                 return NULL;
  93         }
  94
  95         tage->page = page;
  96         atomic_inc(&cfs_tage_allocated);
  97         return tage;
  98 }
  99
 100 static void cfs_tage_free(struct cfs_trace_page *tage)
 101 {
 102         __LASSERT(tage != NULL);
 103         __LASSERT(tage->page != NULL);
 104
 105         __free_page(tage->page);
 106         kfree(tage);
 107         atomic_dec(&cfs_tage_allocated);
 108 }
 109
 110 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
 111                              struct list_head *queue)
 112 {
 113         __LASSERT(tage != NULL);
 114         __LASSERT(queue != NULL);
 115
 116         list_move_tail(&tage->linkage, queue);
 117 }
 118
 119 int cfs_trace_refill_stock(struct cfs_trace_cpu_data *tcd, gfp_t gfp,
 120                            struct list_head *stock)
 121 {
 122         int i;
 123
 124         /*
 125          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 126          * from here: this will lead to infinite recursion.
 127          */
 128
 129         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++ i) {
 130                 struct cfs_trace_page *tage;
 131
 132                 tage = cfs_tage_alloc(gfp);
 133                 if (tage == NULL)
 134                         break;
 135                 list_add_tail(&tage->linkage, stock);
 136         }
 137         return i;
 138 }
 139
 140 /* return a page that has 'len' bytes left at the end */
 141 static struct cfs_trace_page *
 142 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
 143 {
 144         struct cfs_trace_page *tage;
 145
 146         if (tcd->tcd_cur_pages > 0) {
 147                 __LASSERT(!list_empty(&tcd->tcd_pages));
 148                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
 149                 if (tage->used + len <= PAGE_SIZE)
 150                         return tage;
 151         }
 152
 153         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 154                 if (tcd->tcd_cur_stock_pages > 0) {
 155                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
 156                         --tcd->tcd_cur_stock_pages;
 157                         list_del_init(&tage->linkage);
 158                 } else {
 159                         tage = cfs_tage_alloc(GFP_ATOMIC);
 160                         if (unlikely(tage == NULL)) {
 161                                 if ((!memory_pressure_get() ||
 162                                      in_interrupt()) && printk_ratelimit())
 163                                         printk(KERN_WARNING
 164                                                "cannot allocate a tage (%ld)\n",
 165                                                tcd->tcd_cur_pages);
 166                                 return NULL;
 167                         }
 168                 }
 169
 170                 tage->used = 0;
 171                 tage->cpu = smp_processor_id();
 172                 tage->type = tcd->tcd_type;
 173                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 174                 tcd->tcd_cur_pages++;
 175
 176                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 177                         struct tracefiled_ctl *tctl = &trace_tctl;
 178                         /*
 179                          * wake up tracefiled to process some pages.
 180                          */
 181                         wake_up(&tctl->tctl_waitq);
 182                 }
 183                 return tage;
 184         }
 185         return NULL;
 186 }
 187
 188 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
 189 {
 190         int pgcount = tcd->tcd_cur_pages / 10;
 191         struct page_collection pc;
 192         struct cfs_trace_page *tage;
 193         struct cfs_trace_page *tmp;
 194
 195         /*
 196          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 197          * from here: this will lead to infinite recursion.
 198          */
 199
 200         if (printk_ratelimit())
 201                 printk(KERN_WARNING "debug daemon buffer overflowed; "
 202                         "discarding 10%% of pages (%d of %ld)\n",
 203                         pgcount + 1, tcd->tcd_cur_pages);
 204
 205         INIT_LIST_HEAD(&pc.pc_pages);
 206
 207         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 208                 if (pgcount-- == 0)
 209                         break;
 210
 211                 list_move_tail(&tage->linkage, &pc.pc_pages);
 212                 tcd->tcd_cur_pages--;
 213         }
 214         put_pages_on_tcd_daemon_list(&pc, tcd);
 215 }
 216
 217 /* return a page that has 'len' bytes left at the end */
 218 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 219                                                  unsigned long len)
 220 {
 221         struct cfs_trace_page *tage;
 222
 223         /*
 224          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 225          * from here: this will lead to infinite recursion.
 226          */
 227
 228         if (len > PAGE_SIZE) {
 229                 printk(KERN_ERR
 230                        "cowardly refusing to write %lu bytes in a page\n", len);
 231                 return NULL;
 232         }
 233
 234         tage = cfs_trace_get_tage_try(tcd, len);
 235         if (tage != NULL)
 236                 return tage;
 237         if (thread_running)
 238                 cfs_tcd_shrink(tcd);
 239         if (tcd->tcd_cur_pages > 0) {
 240                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 241                 tage->used = 0;
 242                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 243         }
 244         return tage;
 245 }
 246
 247 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 248                      const char *format, ...)
 249 {
 250         va_list args;
 251         int     rc;
 252
 253         va_start(args, format);
 254         rc = libcfs_debug_vmsg2(msgdata, format, args, NULL);
 255         va_end(args);
 256
 257         return rc;
 258 }
 259 EXPORT_SYMBOL(libcfs_debug_msg);
 260
 261 int libcfs_debug_vmsg2(struct libcfs_debug_msg_data *msgdata,
 262                        const char *format1, va_list args,
 263                        const char *format2, ...)
 264 {
 265         struct cfs_trace_cpu_data *tcd = NULL;
 266         struct ptldebug_header     header = {0};
 267         struct cfs_trace_page     *tage;
 268         /* string_buf is used only if tcd != NULL, and is always set then */
 269         char                      *string_buf = NULL;
 270         char                      *debug_buf;
 271         int                        known_size;
 272         int                        needed = 85; /* average message length */
 273         int                        max_nob;
 274         va_list                    ap;
 275         int                        i;
 276         int                        remain;
 277         int                        mask = msgdata->msg_mask;
 278         char                      *file = (char *)msgdata->msg_file;
 279         struct cfs_debug_limit_state *cdls = msgdata->msg_cdls;
 280
 281         if (strchr(file, '/'))
 282                 file = strrchr(file, '/') + 1;
 283
 284         tcd = cfs_trace_get_tcd();
 285
 286         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 287          * pins us to a particular CPU.  This avoids an smp_processor_id()
 288          * warning on Linux when debugging is enabled. */
 289         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 290
 291         if (tcd == NULL)                /* arch may not log in IRQ context */
 292                 goto console;
 293
 294         if (tcd->tcd_cur_pages == 0)
 295                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 296
 297         if (tcd->tcd_shutting_down) {
 298                 cfs_trace_put_tcd(tcd);
 299                 tcd = NULL;
 300                 goto console;
 301         }
 302
 303         known_size = strlen(file) + 1;
 304         if (msgdata->msg_fn)
 305                 known_size += strlen(msgdata->msg_fn) + 1;
 306
 307         if (libcfs_debug_binary)
 308                 known_size += sizeof(header);
 309
 310         /*/
 311          * '2' used because vsnprintf return real size required for output
 312          * _without_ terminating NULL.
 313          * if needed is to small for this format.
 314          */
 315         for (i = 0; i < 2; i++) {
 316                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 317                 if (tage == NULL) {
 318                         if (needed + known_size > PAGE_SIZE)
 319                                 mask |= D_ERROR;
 320
 321                         cfs_trace_put_tcd(tcd);
 322                         tcd = NULL;
 323                         goto console;
 324                 }
 325
 326                 string_buf = (char *)page_address(tage->page) +
 327                                         tage->used + known_size;
 328
 329                 max_nob = PAGE_SIZE - tage->used - known_size;
 330                 if (max_nob <= 0) {
 331                         printk(KERN_EMERG "negative max_nob: %d\n",
 332                                max_nob);
 333                         mask |= D_ERROR;
 334                         cfs_trace_put_tcd(tcd);
 335                         tcd = NULL;
 336                         goto console;
 337                 }
 338
 339                 needed = 0;
 340                 if (format1) {
 341                         va_copy(ap, args);
 342                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 343                         va_end(ap);
 344                 }
 345
 346                 if (format2) {
 347                         remain = max_nob - needed;
 348                         if (remain < 0)
 349                                 remain = 0;
 350
 351                         va_start(ap, format2);
 352                         needed += vsnprintf(string_buf + needed, remain,
 353                                             format2, ap);
 354                         va_end(ap);
 355                 }
 356
 357                 if (needed < max_nob) /* well. printing ok.. */
 358                         break;
 359         }
 360
 361         if (*(string_buf + needed - 1) != '\n') {
 362                 printk(KERN_INFO "format at %s:%d:%s doesn't end in "
 363                        "newline\n", file, msgdata->msg_line, msgdata->msg_fn);
 364         } else if (mask & D_TTY) {
 365                 /* TTY needs '\r\n' to move carriage to leftmost position */
 366                 if (needed < 2 || *(string_buf + needed - 2) != '\r')
 367                         printk(KERN_INFO "format at %s:%d:%s doesn't end in "
 368                                "'\\r\\n'\n", file, msgdata->msg_line,
 369                                msgdata->msg_fn);
 370         }
 371
 372         header.ph_len = known_size + needed;
 373         debug_buf = (char *)page_address(tage->page) + tage->used;
 374
 375         if (libcfs_debug_binary) {
 376                 memcpy(debug_buf, &header, sizeof(header));
 377                 tage->used += sizeof(header);
 378                 debug_buf += sizeof(header);
 379         }
 380
 381         strcpy(debug_buf, file);
 382         tage->used += strlen(file) + 1;
 383         debug_buf += strlen(file) + 1;
 384
 385         if (msgdata->msg_fn) {
 386                 strcpy(debug_buf, msgdata->msg_fn);
 387                 tage->used += strlen(msgdata->msg_fn) + 1;
 388                 debug_buf += strlen(msgdata->msg_fn) + 1;
 389         }
 390
 391         __LASSERT(debug_buf == string_buf);
 392
 393         tage->used += needed;
 394         __LASSERT(tage->used <= PAGE_SIZE);
 395
 396 console:
 397         if ((mask & libcfs_printk) == 0) {
 398                 /* no console output requested */
 399                 if (tcd != NULL)
 400                         cfs_trace_put_tcd(tcd);
 401                 return 1;
 402         }
 403
 404         if (cdls != NULL) {
 405                 if (libcfs_console_ratelimit &&
 406                     cdls->cdls_next != 0 &&     /* not first time ever */
 407                     time_before(jiffies, cdls->cdls_next)) {
 408                         /* skipping a console message */
 409                         cdls->cdls_count++;
 410                         if (tcd != NULL)
 411                                 cfs_trace_put_tcd(tcd);
 412                         return 1;
 413                 }
 414
 415                 if (time_after(jiffies, cdls->cdls_next +
 416                                         libcfs_console_max_delay +
 417                                         cfs_time_seconds(10))) {
 418                         /* last timeout was a long time ago */
 419                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 420                 } else {
 421                         cdls->cdls_delay *= libcfs_console_backoff;
 422                 }
 423
 424                 if (cdls->cdls_delay < libcfs_console_min_delay)
 425                         cdls->cdls_delay = libcfs_console_min_delay;
 426                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 427                         cdls->cdls_delay = libcfs_console_max_delay;
 428
 429                 /* ensure cdls_next is never zero after it's been seen */
 430                 cdls->cdls_next = (jiffies + cdls->cdls_delay) | 1;
 431         }
 432
 433         if (tcd != NULL) {
 434                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 435                                      msgdata->msg_fn);
 436                 cfs_trace_put_tcd(tcd);
 437         } else {
 438                 string_buf = cfs_trace_get_console_buffer();
 439
 440                 needed = 0;
 441                 if (format1 != NULL) {
 442                         va_copy(ap, args);
 443                         needed = vsnprintf(string_buf,
 444                                            CFS_TRACE_CONSOLE_BUFFER_SIZE,
 445                                            format1, ap);
 446                         va_end(ap);
 447                 }
 448                 if (format2 != NULL) {
 449                         remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
 450                         if (remain > 0) {
 451                                 va_start(ap, format2);
 452                                 needed += vsnprintf(string_buf+needed, remain,
 453                                                     format2, ap);
 454                                 va_end(ap);
 455                         }
 456                 }
 457                 cfs_print_to_console(&header, mask,
 458                                      string_buf, needed, file, msgdata->msg_fn);
 459
 460                 put_cpu();
 461         }
 462
 463         if (cdls != NULL && cdls->cdls_count != 0) {
 464                 string_buf = cfs_trace_get_console_buffer();
 465
 466                 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 467                                   "Skipped %d previous similar message%s\n",
 468                                   cdls->cdls_count,
 469                                   (cdls->cdls_count > 1) ? "s" : "");
 470
 471                 /* Do not allow print this to TTY */
 472                 cfs_print_to_console(&header, mask & ~D_TTY, string_buf,
 473                                      needed, file, msgdata->msg_fn);
 474
 475                 put_cpu();
 476                 cdls->cdls_count = 0;
 477         }
 478
 479         return 0;
 480 }
 481 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 482
 483 void
 484 cfs_trace_assertion_failed(const char *str,
 485                            struct libcfs_debug_msg_data *msgdata)
 486 {
 487         struct ptldebug_header hdr;
 488
 489         libcfs_panic_in_progress = 1;
 490         libcfs_catastrophe = 1;
 491         smp_mb();
 492
 493         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 494
 495         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 496                              msgdata->msg_file, msgdata->msg_fn);
 497
 498         panic("Lustre debug assertion failure\n");
 499
 500         /* not reached */
 501 }
 502
 503 static void
 504 panic_collect_pages(struct page_collection *pc)
 505 {
 506         /* Do the collect_pages job on a single CPU: assumes that all other
 507          * CPUs have been stopped during a panic.  If this isn't true for some
 508          * arch, this will have to be implemented separately in each arch.  */
 509         int                        i;
 510         int                        j;
 511         struct cfs_trace_cpu_data *tcd;
 512
 513         INIT_LIST_HEAD(&pc->pc_pages);
 514
 515         cfs_tcd_for_each(tcd, i, j) {
 516                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 517                 tcd->tcd_cur_pages = 0;
 518
 519                 if (pc->pc_want_daemon_pages) {
 520                         list_splice_init(&tcd->tcd_daemon_pages,
 521                                                 &pc->pc_pages);
 522                         tcd->tcd_cur_daemon_pages = 0;
 523                 }
 524         }
 525 }
 526
 527 static void collect_pages_on_all_cpus(struct page_collection *pc)
 528 {
 529         struct cfs_trace_cpu_data *tcd;
 530         int i, cpu;
 531
 532         for_each_possible_cpu(cpu) {
 533                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 534                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 535                         tcd->tcd_cur_pages = 0;
 536                         if (pc->pc_want_daemon_pages) {
 537                                 list_splice_init(&tcd->tcd_daemon_pages,
 538                                                         &pc->pc_pages);
 539                                 tcd->tcd_cur_daemon_pages = 0;
 540                         }
 541                 }
 542         }
 543 }
 544
 545 static void collect_pages(struct page_collection *pc)
 546 {
 547         INIT_LIST_HEAD(&pc->pc_pages);
 548
 549         if (libcfs_panic_in_progress)
 550                 panic_collect_pages(pc);
 551         else
 552                 collect_pages_on_all_cpus(pc);
 553 }
 554
 555 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 556 {
 557         struct cfs_trace_cpu_data *tcd;
 558         struct list_head *cur_head;
 559         struct cfs_trace_page *tage;
 560         struct cfs_trace_page *tmp;
 561         int i, cpu;
 562
 563         for_each_possible_cpu(cpu) {
 564                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 565                         cur_head = tcd->tcd_pages.next;
 566
 567                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 568                                                  linkage) {
 569
 570                                 __LASSERT_TAGE_INVARIANT(tage);
 571
 572                                 if (tage->cpu != cpu || tage->type != i)
 573                                         continue;
 574
 575                                 cfs_tage_to_tail(tage, cur_head);
 576                                 tcd->tcd_cur_pages++;
 577                         }
 578                 }
 579         }
 580 }
 581
 582 static void put_pages_back(struct page_collection *pc)
 583 {
 584         if (!libcfs_panic_in_progress)
 585                 put_pages_back_on_all_cpus(pc);
 586 }
 587
 588 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 589  * we have a good amount of data at all times for dumping during an LBUG, even
 590  * if we have been steadily writing (and otherwise discarding) pages via the
 591  * debug daemon. */
 592 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 593                                          struct cfs_trace_cpu_data *tcd)
 594 {
 595         struct cfs_trace_page *tage;
 596         struct cfs_trace_page *tmp;
 597
 598         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 599                 __LASSERT_TAGE_INVARIANT(tage);
 600
 601                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 602                         continue;
 603
 604                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 605                 tcd->tcd_cur_daemon_pages++;
 606
 607                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 608                         struct cfs_trace_page *victim;
 609
 610                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 611                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 612
 613                         __LASSERT_TAGE_INVARIANT(victim);
 614
 615                         list_del(&victim->linkage);
 616                         cfs_tage_free(victim);
 617                         tcd->tcd_cur_daemon_pages--;
 618                 }
 619         }
 620 }
 621
 622 static void put_pages_on_daemon_list(struct page_collection *pc)
 623 {
 624         struct cfs_trace_cpu_data *tcd;
 625         int i, cpu;
 626
 627         for_each_possible_cpu(cpu) {
 628                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 629                         put_pages_on_tcd_daemon_list(pc, tcd);
 630         }
 631 }
 632
 633 void cfs_trace_debug_print(void)
 634 {
 635         struct page_collection pc;
 636         struct cfs_trace_page *tage;
 637         struct cfs_trace_page *tmp;
 638
 639         pc.pc_want_daemon_pages = 1;
 640         collect_pages(&pc);
 641         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 642                 char *p, *file, *fn;
 643                 struct page *page;
 644
 645                 __LASSERT_TAGE_INVARIANT(tage);
 646
 647                 page = tage->page;
 648                 p = page_address(page);
 649                 while (p < ((char *)page_address(page) + tage->used)) {
 650                         struct ptldebug_header *hdr;
 651                         int len;
 652                         hdr = (void *)p;
 653                         p += sizeof(*hdr);
 654                         file = p;
 655                         p += strlen(file) + 1;
 656                         fn = p;
 657                         p += strlen(fn) + 1;
 658                         len = hdr->ph_len - (int)(p - (char *)hdr);
 659
 660                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 661
 662                         p += len;
 663                 }
 664
 665                 list_del(&tage->linkage);
 666                 cfs_tage_free(tage);
 667         }
 668 }
 669
 670 int cfs_tracefile_dump_all_pages(char *filename)
 671 {
 672         struct page_collection  pc;
 673         struct file             *filp;
 674         struct cfs_trace_page   *tage;
 675         struct cfs_trace_page   *tmp;
 676         char                    *buf;
 677         int rc;
 678
 679         cfs_tracefile_write_lock();
 680
 681         filp = filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600);
 682         if (IS_ERR(filp)) {
 683                 rc = PTR_ERR(filp);
 684                 filp = NULL;
 685                 printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 686                       filename, rc);
 687                 goto out;
 688         }
 689
 690         pc.pc_want_daemon_pages = 1;
 691         collect_pages(&pc);
 692         if (list_empty(&pc.pc_pages)) {
 693                 rc = 0;
 694                 goto close;
 695         }
 696
 697         /* ok, for now, just write the pages.  in the future we'll be building
 698          * iobufs with the pages and calling generic_direct_IO */
 699         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 700
 701                 __LASSERT_TAGE_INVARIANT(tage);
 702
 703                 buf = kmap(tage->page);
 704                 rc = cfs_kernel_write(filp, buf, tage->used, &filp->f_pos);
 705                 kunmap(tage->page);
 706                 if (rc != (int)tage->used) {
 707                         printk(KERN_WARNING "wanted to write %u but wrote "
 708                                "%d\n", tage->used, rc);
 709                         put_pages_back(&pc);
 710                         __LASSERT(list_empty(&pc.pc_pages));
 711                         break;
 712                 }
 713                 list_del(&tage->linkage);
 714                 cfs_tage_free(tage);
 715         }
 716
 717         rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1);
 718         if (rc)
 719                 printk(KERN_ERR "sync returns %d\n", rc);
 720 close:
 721         filp_close(filp, NULL);
 722 out:
 723         cfs_tracefile_write_unlock();
 724         return rc;
 725 }
 726
 727 void cfs_trace_flush_pages(void)
 728 {
 729         struct page_collection pc;
 730         struct cfs_trace_page *tage;
 731         struct cfs_trace_page *tmp;
 732
 733         pc.pc_want_daemon_pages = 1;
 734         collect_pages(&pc);
 735         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 736
 737                 __LASSERT_TAGE_INVARIANT(tage);
 738
 739                 list_del(&tage->linkage);
 740                 cfs_tage_free(tage);
 741         }
 742 }
 743
 744 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 745                             const char __user *usr_buffer, int usr_buffer_nob)
 746 {
 747         int    nob;
 748
 749         if (usr_buffer_nob > knl_buffer_nob)
 750                 return -EOVERFLOW;
 751
 752         if (copy_from_user(knl_buffer, usr_buffer, usr_buffer_nob))
 753                 return -EFAULT;
 754
 755         nob = strnlen(knl_buffer, usr_buffer_nob);
 756         while (nob-- >= 0)                      /* strip trailing whitespace */
 757                 if (!isspace(knl_buffer[nob]))
 758                         break;
 759
 760         if (nob < 0)                            /* empty string */
 761                 return -EINVAL;
 762
 763         if (nob == knl_buffer_nob)              /* no space to terminate */
 764                 return -EOVERFLOW;
 765
 766         knl_buffer[nob + 1] = 0;                /* terminate */
 767         return 0;
 768 }
 769 EXPORT_SYMBOL(cfs_trace_copyin_string);
 770
 771 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 772                              const char *knl_buffer, char *append)
 773 {
 774         /* NB if 'append' != NULL, it's a single character to append to the
 775          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 776          * terminating zero byte) for sysctl entries */
 777         int   nob = strlen(knl_buffer);
 778
 779         if (nob > usr_buffer_nob)
 780                 nob = usr_buffer_nob;
 781
 782         if (copy_to_user(usr_buffer, knl_buffer, nob))
 783                 return -EFAULT;
 784
 785         if (append != NULL && nob < usr_buffer_nob) {
 786                 if (copy_to_user(usr_buffer + nob, append, 1))
 787                         return -EFAULT;
 788
 789                 nob++;
 790         }
 791
 792         return nob;
 793 }
 794 EXPORT_SYMBOL(cfs_trace_copyout_string);
 795
 796 int cfs_trace_allocate_string_buffer(char **str, int nob)
 797 {
 798         if (nob > 2 * PAGE_SIZE)        /* string must be "sensible" */
 799                 return -EINVAL;
 800
 801         *str = kmalloc(nob, GFP_KERNEL | __GFP_ZERO);
 802         if (*str == NULL)
 803                 return -ENOMEM;
 804
 805         return 0;
 806 }
 807
 808 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
 809 {
 810         char         *str;
 811         int           rc;
 812
 813         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 814         if (rc != 0)
 815                 return rc;
 816
 817         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 818                                      usr_str, usr_str_nob);
 819         if (rc != 0)
 820                 goto out;
 821
 822         if (str[0] != '/') {
 823                 rc = -EINVAL;
 824                 goto out;
 825         }
 826         rc = cfs_tracefile_dump_all_pages(str);
 827 out:
 828         kfree(str);
 829         return rc;
 830 }
 831
 832 int cfs_trace_daemon_command(char *str)
 833 {
 834         int       rc = 0;
 835
 836         cfs_tracefile_write_lock();
 837
 838         if (strcmp(str, "stop") == 0) {
 839                 cfs_tracefile_write_unlock();
 840                 cfs_trace_stop_thread();
 841                 cfs_tracefile_write_lock();
 842                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 843
 844         } else if (strncmp(str, "size=", 5) == 0) {
 845                 unsigned long tmp;
 846
 847                 rc = kstrtoul(str + 5, 10, &tmp);
 848                 if (!rc) {
 849                         if (tmp < 10 || tmp > 20480)
 850                                 cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 851                         else
 852                                 cfs_tracefile_size = tmp << 20;
 853                 }
 854         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 855                 rc = -ENAMETOOLONG;
 856         } else if (str[0] != '/') {
 857                 rc = -EINVAL;
 858         } else {
 859                 strcpy(cfs_tracefile, str);
 860
 861                 printk(KERN_INFO
 862                        "Lustre: debug daemon will attempt to start writing "
 863                        "to %s (%lukB max)\n", cfs_tracefile,
 864                        (long)(cfs_tracefile_size >> 10));
 865
 866                 cfs_trace_start_thread();
 867         }
 868
 869         cfs_tracefile_write_unlock();
 870         return rc;
 871 }
 872
 873 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
 874 {
 875         char *str;
 876         int   rc;
 877
 878         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 879         if (rc != 0)
 880                 return rc;
 881
 882         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 883                                  usr_str, usr_str_nob);
 884         if (rc == 0)
 885                 rc = cfs_trace_daemon_command(str);
 886
 887         kfree(str);
 888         return rc;
 889 }
 890
 891 int cfs_trace_set_debug_mb(int mb)
 892 {
 893         int i;
 894         int j;
 895         int pages;
 896         int limit = cfs_trace_max_debug_mb();
 897         struct cfs_trace_cpu_data *tcd;
 898
 899         if (mb < num_possible_cpus()) {
 900                 printk(KERN_WARNING
 901                        "Lustre: %d MB is too small for debug buffer size, "
 902                        "setting it to %d MB.\n", mb, num_possible_cpus());
 903                 mb = num_possible_cpus();
 904         }
 905
 906         if (mb > limit) {
 907                 printk(KERN_WARNING
 908                        "Lustre: %d MB is too large for debug buffer size, "
 909                        "setting it to %d MB.\n", mb, limit);
 910                 mb = limit;
 911         }
 912
 913         mb /= num_possible_cpus();
 914         pages = mb << (20 - PAGE_SHIFT);
 915
 916         cfs_tracefile_write_lock();
 917
 918         cfs_tcd_for_each(tcd, i, j)
 919                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 920
 921         cfs_tracefile_write_unlock();
 922
 923         return 0;
 924 }
 925
 926 int cfs_trace_get_debug_mb(void)
 927 {
 928         int i;
 929         int j;
 930         struct cfs_trace_cpu_data *tcd;
 931         int total_pages = 0;
 932
 933         cfs_tracefile_read_lock();
 934
 935         cfs_tcd_for_each(tcd, i, j)
 936                 total_pages += tcd->tcd_max_pages;
 937
 938         cfs_tracefile_read_unlock();
 939
 940         return (total_pages >> (20 - PAGE_SHIFT)) + 1;
 941 }
 942
 943 static int tracefiled(void *arg)
 944 {
 945         struct page_collection pc;
 946         struct tracefiled_ctl *tctl = arg;
 947         struct cfs_trace_page *tage;
 948         struct cfs_trace_page *tmp;
 949         struct file *filp;
 950         char *buf;
 951         int last_loop = 0;
 952         int rc;
 953
 954         /* we're started late enough that we pick up init's fs context */
 955         /* this is so broken in uml?  what on earth is going on? */
 956
 957         complete(&tctl->tctl_start);
 958
 959         while (1) {
 960                 wait_queue_entry_t __wait;
 961
 962                 pc.pc_want_daemon_pages = 0;
 963                 collect_pages(&pc);
 964                 if (list_empty(&pc.pc_pages))
 965                         goto end_loop;
 966
 967                 filp = NULL;
 968                 cfs_tracefile_read_lock();
 969                 if (cfs_tracefile[0] != 0) {
 970                         filp = filp_open(cfs_tracefile,
 971                                          O_CREAT | O_RDWR | O_LARGEFILE,
 972                                          0600);
 973                         if (IS_ERR(filp)) {
 974                                 rc = PTR_ERR(filp);
 975                                 filp = NULL;
 976                                 printk(KERN_WARNING "couldn't open %s: "
 977                                        "%d\n", cfs_tracefile, rc);
 978                         }
 979                 }
 980                 cfs_tracefile_read_unlock();
 981                 if (filp == NULL) {
 982                         put_pages_on_daemon_list(&pc);
 983                         __LASSERT(list_empty(&pc.pc_pages));
 984                         goto end_loop;
 985                 }
 986
 987                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 988                         struct dentry *de = file_dentry(filp);
 989                         static loff_t f_pos;
 990
 991                         __LASSERT_TAGE_INVARIANT(tage);
 992
 993                         if (f_pos >= (off_t)cfs_tracefile_size)
 994                                 f_pos = 0;
 995                         else if (f_pos > i_size_read(de->d_inode))
 996                                 f_pos = i_size_read(de->d_inode);
 997
 998                         buf = kmap(tage->page);
 999                         rc = cfs_kernel_write(filp, buf, tage->used, &f_pos);
1000                         kunmap(tage->page);
1001                         if (rc != (int)tage->used) {
1002                                 printk(KERN_WARNING "wanted to write %u "
1003                                        "but wrote %d\n", tage->used, rc);
1004                                 put_pages_back(&pc);
1005                                 __LASSERT(list_empty(&pc.pc_pages));
1006                                 break;
1007                         }
1008                 }
1009
1010                 filp_close(filp, NULL);
1011                 put_pages_on_daemon_list(&pc);
1012                 if (!list_empty(&pc.pc_pages)) {
1013                         int i;
1014
1015                         printk(KERN_ALERT "Lustre: trace pages aren't "
1016                                " empty\n");
1017                         printk(KERN_ERR "total cpus(%d): ",
1018                                num_possible_cpus());
1019                         for (i = 0; i < num_possible_cpus(); i++)
1020                                 if (cpu_online(i))
1021                                         printk(KERN_ERR "%d(on) ", i);
1022                                 else
1023                                         printk(KERN_ERR "%d(off) ", i);
1024                         printk(KERN_ERR "\n");
1025
1026                         i = 0;
1027                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1028                                                      linkage)
1029                                 printk(KERN_ERR "page %d belongs to cpu "
1030                                        "%d\n", ++i, tage->cpu);
1031                         printk(KERN_ERR "There are %d pages unwritten\n",
1032                                i);
1033                 }
1034                 __LASSERT(list_empty(&pc.pc_pages));
1035 end_loop:
1036                 if (atomic_read(&tctl->tctl_shutdown)) {
1037                         if (last_loop == 0) {
1038                                 last_loop = 1;
1039                                 continue;
1040                         } else {
1041                                 break;
1042                         }
1043                 }
1044                 init_waitqueue_entry(&__wait, current);
1045                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1046                 set_current_state(TASK_INTERRUPTIBLE);
1047                 schedule_timeout(cfs_time_seconds(1));
1048                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1049         }
1050         complete(&tctl->tctl_stop);
1051         return 0;
1052 }
1053
1054 int cfs_trace_start_thread(void)
1055 {
1056         struct tracefiled_ctl *tctl = &trace_tctl;
1057         int rc = 0;
1058
1059         mutex_lock(&cfs_trace_thread_mutex);
1060         if (thread_running)
1061                 goto out;
1062
1063         init_completion(&tctl->tctl_start);
1064         init_completion(&tctl->tctl_stop);
1065         init_waitqueue_head(&tctl->tctl_waitq);
1066         atomic_set(&tctl->tctl_shutdown, 0);
1067
1068         if (IS_ERR(kthread_run(tracefiled, tctl, "ktracefiled"))) {
1069                 rc = -ECHILD;
1070                 goto out;
1071         }
1072
1073         wait_for_completion(&tctl->tctl_start);
1074         thread_running = 1;
1075 out:
1076         mutex_unlock(&cfs_trace_thread_mutex);
1077         return rc;
1078 }
1079
1080 void cfs_trace_stop_thread(void)
1081 {
1082         struct tracefiled_ctl *tctl = &trace_tctl;
1083
1084         mutex_lock(&cfs_trace_thread_mutex);
1085         if (thread_running) {
1086                 printk(KERN_INFO
1087                        "Lustre: shutting down debug daemon thread...\n");
1088                 atomic_set(&tctl->tctl_shutdown, 1);
1089                 wait_for_completion(&tctl->tctl_stop);
1090                 thread_running = 0;
1091         }
1092         mutex_unlock(&cfs_trace_thread_mutex);
1093 }
1094
1095 int cfs_tracefile_init(int max_pages)
1096 {
1097         struct cfs_trace_cpu_data *tcd;
1098         int     i;
1099         int     j;
1100         int     rc;
1101         int     factor;
1102
1103         rc = cfs_tracefile_init_arch();
1104         if (rc != 0)
1105                 return rc;
1106
1107         cfs_tcd_for_each(tcd, i, j) {
1108                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1109                 factor = tcd->tcd_pages_factor;
1110                 INIT_LIST_HEAD(&tcd->tcd_pages);
1111                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1112                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1113                 tcd->tcd_cur_pages = 0;
1114                 tcd->tcd_cur_stock_pages = 0;
1115                 tcd->tcd_cur_daemon_pages = 0;
1116                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1117                 LASSERT(tcd->tcd_max_pages > 0);
1118                 tcd->tcd_shutting_down = 0;
1119         }
1120         return 0;
1121 }
1122
1123 static void trace_cleanup_on_all_cpus(void)
1124 {
1125         struct cfs_trace_cpu_data *tcd;
1126         struct cfs_trace_page *tage;
1127         struct cfs_trace_page *tmp;
1128         int i, cpu;
1129
1130         for_each_possible_cpu(cpu) {
1131                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1132                         tcd->tcd_shutting_down = 1;
1133
1134                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1135                                 __LASSERT_TAGE_INVARIANT(tage);
1136
1137                                 list_del(&tage->linkage);
1138                                 cfs_tage_free(tage);
1139                         }
1140                         tcd->tcd_cur_pages = 0;
1141                 }
1142         }
1143 }
1144
1145 static void cfs_trace_cleanup(void)
1146 {
1147         struct page_collection pc;
1148
1149         INIT_LIST_HEAD(&pc.pc_pages);
1150
1151         trace_cleanup_on_all_cpus();
1152
1153         cfs_tracefile_fini_arch();
1154 }
1155
1156 void cfs_tracefile_exit(void)
1157 {
1158         cfs_trace_stop_thread();
1159         cfs_trace_cleanup();
1160 }