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/poll.h>
  48 #include <linux/tty.h>
  49 #include <linux/uaccess.h>
  50 #include <libcfs/linux/linux-fs.h>
  51 #include <libcfs/libcfs.h>
  52
  53 /* XXX move things up to the top, comment */
  54 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  55
  56 char *cfs_trace_console_buffers[NR_CPUS][CFS_TCD_TYPE_MAX];
  57 char cfs_tracefile[TRACEFILE_NAME_SIZE];
  58 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
  59 static struct tracefiled_ctl trace_tctl;
  60 static DEFINE_MUTEX(cfs_trace_thread_mutex);
  61 static int thread_running = 0;
  62
  63 static atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
  64 static DECLARE_RWSEM(cfs_tracefile_sem);
  65
  66 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  67                                         struct cfs_trace_cpu_data *tcd);
  68
  69 /* trace file lock routines */
  70 /* The walking argument indicates the locking comes from all tcd types
  71  * iterator and we must lock it and dissable local irqs to avoid deadlocks
  72  * with other interrupt locks that might be happening. See LU-1311
  73  * for details.
  74  */
  75 int cfs_trace_lock_tcd(struct cfs_trace_cpu_data *tcd, int walking)
  76         __acquires(&tcd->tcd_lock)
  77 {
  78         __LASSERT(tcd->tcd_type < CFS_TCD_TYPE_MAX);
  79         if (tcd->tcd_type == CFS_TCD_TYPE_IRQ)
  80                 spin_lock_irqsave(&tcd->tcd_lock, tcd->tcd_lock_flags);
  81         else if (tcd->tcd_type == CFS_TCD_TYPE_SOFTIRQ)
  82                 spin_lock_bh(&tcd->tcd_lock);
  83         else if (unlikely(walking))
  84                 spin_lock_irq(&tcd->tcd_lock);
  85         else
  86                 spin_lock(&tcd->tcd_lock);
  87         return 1;
  88 }
  89
  90 void cfs_trace_unlock_tcd(struct cfs_trace_cpu_data *tcd, int walking)
  91         __releases(&tcd->tcd_lock)
  92 {
  93         __LASSERT(tcd->tcd_type < CFS_TCD_TYPE_MAX);
  94         if (tcd->tcd_type == CFS_TCD_TYPE_IRQ)
  95                 spin_unlock_irqrestore(&tcd->tcd_lock, tcd->tcd_lock_flags);
  96         else if (tcd->tcd_type == CFS_TCD_TYPE_SOFTIRQ)
  97                 spin_unlock_bh(&tcd->tcd_lock);
  98         else if (unlikely(walking))
  99                 spin_unlock_irq(&tcd->tcd_lock);
 100         else
 101                 spin_unlock(&tcd->tcd_lock);
 102 }
 103
 104 #define cfs_tcd_for_each_type_lock(tcd, i, cpu)                         \
 105         for (i = 0; cfs_trace_data[i] &&                                \
 106              (tcd = &(*cfs_trace_data[i])[cpu].tcd) &&                  \
 107              cfs_trace_lock_tcd(tcd, 1); cfs_trace_unlock_tcd(tcd, 1), i++)
 108
 109 enum cfs_trace_buf_type cfs_trace_buf_idx_get(void)
 110 {
 111         if (in_irq())
 112                 return CFS_TCD_TYPE_IRQ;
 113         if (in_softirq())
 114                 return CFS_TCD_TYPE_SOFTIRQ;
 115         return CFS_TCD_TYPE_PROC;
 116 }
 117
 118 static inline struct cfs_trace_cpu_data *
 119 cfs_trace_get_tcd(void)
 120 {
 121         struct cfs_trace_cpu_data *tcd =
 122                 &(*cfs_trace_data[cfs_trace_buf_idx_get()])[get_cpu()].tcd;
 123
 124         cfs_trace_lock_tcd(tcd, 0);
 125
 126         return tcd;
 127 }
 128
 129 static inline void cfs_trace_put_tcd(struct cfs_trace_cpu_data *tcd)
 130 {
 131         cfs_trace_unlock_tcd(tcd, 0);
 132
 133         put_cpu();
 134 }
 135
 136 /* percents to share the total debug memory for each type */
 137 static unsigned int pages_factor[CFS_TCD_TYPE_MAX] = {
 138         80,     /* 80% pages for CFS_TCD_TYPE_PROC */
 139         10,     /* 10% pages for CFS_TCD_TYPE_SOFTIRQ */
 140         10      /* 10% pages for CFS_TCD_TYPE_IRQ */
 141 };
 142
 143 int cfs_tracefile_init_arch(void)
 144 {
 145         struct cfs_trace_cpu_data *tcd;
 146         int i;
 147         int j;
 148
 149         /* initialize trace_data */
 150         memset(cfs_trace_data, 0, sizeof(cfs_trace_data));
 151         for (i = 0; i < CFS_TCD_TYPE_MAX; i++) {
 152                 cfs_trace_data[i] =
 153                         kmalloc_array(num_possible_cpus(),
 154                                       sizeof(union cfs_trace_data_union),
 155                                       GFP_KERNEL);
 156                 if (!cfs_trace_data[i])
 157                         goto out_trace_data;
 158         }
 159
 160         /* arch related info initialized */
 161         cfs_tcd_for_each(tcd, i, j) {
 162                 spin_lock_init(&tcd->tcd_lock);
 163                 tcd->tcd_pages_factor = pages_factor[i];
 164                 tcd->tcd_type = i;
 165                 tcd->tcd_cpu = j;
 166         }
 167
 168         for (i = 0; i < num_possible_cpus(); i++)
 169                 for (j = 0; j < 3; j++) {
 170                         cfs_trace_console_buffers[i][j] =
 171                                 kmalloc(CFS_TRACE_CONSOLE_BUFFER_SIZE,
 172                                         GFP_KERNEL);
 173
 174                         if (!cfs_trace_console_buffers[i][j])
 175                                 goto out_buffers;
 176                 }
 177
 178         return 0;
 179
 180 out_buffers:
 181         for (i = 0; i < num_possible_cpus(); i++)
 182                 for (j = 0; j < 3; j++) {
 183                         kfree(cfs_trace_console_buffers[i][j]);
 184                         cfs_trace_console_buffers[i][j] = NULL;
 185                 }
 186 out_trace_data:
 187         for (i = 0; cfs_trace_data[i]; i++) {
 188                 kfree(cfs_trace_data[i]);
 189                 cfs_trace_data[i] = NULL;
 190         }
 191         pr_err("lnet: Not enough memory\n");
 192         return -ENOMEM;
 193 }
 194
 195 void cfs_tracefile_fini_arch(void)
 196 {
 197         int i;
 198         int j;
 199
 200         for (i = 0; i < num_possible_cpus(); i++)
 201                 for (j = 0; j < 3; j++) {
 202                         kfree(cfs_trace_console_buffers[i][j]);
 203                         cfs_trace_console_buffers[i][j] = NULL;
 204                 }
 205
 206         for (i = 0; cfs_trace_data[i]; i++) {
 207                 kfree(cfs_trace_data[i]);
 208                 cfs_trace_data[i] = NULL;
 209         }
 210 }
 211
 212 static inline struct cfs_trace_page *
 213 cfs_tage_from_list(struct list_head *list)
 214 {
 215         return list_entry(list, struct cfs_trace_page, linkage);
 216 }
 217
 218 static struct cfs_trace_page *cfs_tage_alloc(gfp_t gfp)
 219 {
 220         struct page            *page;
 221         struct cfs_trace_page *tage;
 222
 223         /* My caller is trying to free memory */
 224         if (!in_interrupt() && (current->flags & PF_MEMALLOC))
 225                 return NULL;
 226
 227         /*
 228          * Don't spam console with allocation failures: they will be reported
 229          * by upper layer anyway.
 230          */
 231         gfp |= __GFP_NOWARN;
 232         page = alloc_page(gfp);
 233         if (page == NULL)
 234                 return NULL;
 235
 236         tage = kmalloc(sizeof(*tage), gfp);
 237         if (tage == NULL) {
 238                 __free_page(page);
 239                 return NULL;
 240         }
 241
 242         tage->page = page;
 243         atomic_inc(&cfs_tage_allocated);
 244         return tage;
 245 }
 246
 247 static void cfs_tage_free(struct cfs_trace_page *tage)
 248 {
 249         __LASSERT(tage != NULL);
 250         __LASSERT(tage->page != NULL);
 251
 252         __free_page(tage->page);
 253         kfree(tage);
 254         atomic_dec(&cfs_tage_allocated);
 255 }
 256
 257 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
 258                              struct list_head *queue)
 259 {
 260         __LASSERT(tage != NULL);
 261         __LASSERT(queue != NULL);
 262
 263         list_move_tail(&tage->linkage, queue);
 264 }
 265
 266 /* return a page that has 'len' bytes left at the end */
 267 static struct cfs_trace_page *
 268 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
 269 {
 270         struct cfs_trace_page *tage;
 271
 272         if (tcd->tcd_cur_pages > 0) {
 273                 __LASSERT(!list_empty(&tcd->tcd_pages));
 274                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
 275                 if (tage->used + len <= PAGE_SIZE)
 276                         return tage;
 277         }
 278
 279         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 280                 if (tcd->tcd_cur_stock_pages > 0) {
 281                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
 282                         --tcd->tcd_cur_stock_pages;
 283                         list_del_init(&tage->linkage);
 284                 } else {
 285                         tage = cfs_tage_alloc(GFP_ATOMIC);
 286                         if (unlikely(tage == NULL)) {
 287                                 if ((!(current->flags & PF_MEMALLOC) ||
 288                                      in_interrupt()) && printk_ratelimit())
 289                                         pr_warn("Lustre: cannot allocate a tage (%ld)\n",
 290                                                 tcd->tcd_cur_pages);
 291                                 return NULL;
 292                         }
 293                 }
 294
 295                 tage->used = 0;
 296                 tage->cpu = smp_processor_id();
 297                 tage->type = tcd->tcd_type;
 298                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 299                 tcd->tcd_cur_pages++;
 300
 301                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 302                         struct tracefiled_ctl *tctl = &trace_tctl;
 303                         /*
 304                          * wake up tracefiled to process some pages.
 305                          */
 306                         wake_up(&tctl->tctl_waitq);
 307                 }
 308                 return tage;
 309         }
 310         return NULL;
 311 }
 312
 313 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
 314 {
 315         int pgcount = tcd->tcd_cur_pages / 10;
 316         struct page_collection pc;
 317         struct cfs_trace_page *tage;
 318         struct cfs_trace_page *tmp;
 319
 320         /*
 321          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 322          * from here: this will lead to infinite recursion.
 323          */
 324
 325         if (printk_ratelimit())
 326                 pr_warn("Lustre: debug daemon buffer overflowed; discarding 10%% of pages (%d of %ld)\n",
 327                         pgcount + 1, tcd->tcd_cur_pages);
 328
 329         INIT_LIST_HEAD(&pc.pc_pages);
 330
 331         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 332                 if (pgcount-- == 0)
 333                         break;
 334
 335                 list_move_tail(&tage->linkage, &pc.pc_pages);
 336                 tcd->tcd_cur_pages--;
 337         }
 338         put_pages_on_tcd_daemon_list(&pc, tcd);
 339 }
 340
 341 /* return a page that has 'len' bytes left at the end */
 342 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 343                                                  unsigned long len)
 344 {
 345         struct cfs_trace_page *tage;
 346
 347         /*
 348          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 349          * from here: this will lead to infinite recursion.
 350          */
 351
 352         if (len > PAGE_SIZE) {
 353                 pr_err("LustreError: cowardly refusing to write %lu bytes in a page\n",
 354                        len);
 355                 return NULL;
 356         }
 357
 358         tage = cfs_trace_get_tage_try(tcd, len);
 359         if (tage != NULL)
 360                 return tage;
 361         if (thread_running)
 362                 cfs_tcd_shrink(tcd);
 363         if (tcd->tcd_cur_pages > 0) {
 364                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 365                 tage->used = 0;
 366                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 367         }
 368         return tage;
 369 }
 370
 371 static void cfs_set_ptldebug_header(struct ptldebug_header *header,
 372                                     struct libcfs_debug_msg_data *msgdata,
 373                                     unsigned long stack)
 374 {
 375         struct timespec64 ts;
 376
 377         ktime_get_real_ts64(&ts);
 378
 379         header->ph_subsys = msgdata->msg_subsys;
 380         header->ph_mask = msgdata->msg_mask;
 381         header->ph_cpu_id = smp_processor_id();
 382         header->ph_type = cfs_trace_buf_idx_get();
 383         /* y2038 safe since all user space treats this as unsigned, but
 384          * will overflow in 2106
 385          */
 386         header->ph_sec = (u32)ts.tv_sec;
 387         header->ph_usec = ts.tv_nsec / NSEC_PER_USEC;
 388         header->ph_stack = stack;
 389         header->ph_pid = current->pid;
 390         header->ph_line_num = msgdata->msg_line;
 391         header->ph_extern_pid = 0;
 392 }
 393
 394 /**
 395  * tty_write_msg - write a message to a certain tty, not just the console.
 396  * @tty: the destination tty_struct
 397  * @msg: the message to write
 398  *
 399  * tty_write_message is not exported, so write a same function for it
 400  *
 401  */
 402 static void tty_write_msg(struct tty_struct *tty, const char *msg)
 403 {
 404         mutex_lock(&tty->atomic_write_lock);
 405         tty_lock(tty);
 406         if (tty->ops->write && tty->count > 0)
 407                 tty->ops->write(tty, msg, strlen(msg));
 408         tty_unlock(tty);
 409         mutex_unlock(&tty->atomic_write_lock);
 410         wake_up_interruptible_poll(&tty->write_wait, POLLOUT);
 411 }
 412
 413 static void cfs_tty_write_message(const char *prefix, int mask, const char *msg)
 414 {
 415         struct tty_struct *tty;
 416
 417         tty = get_current_tty();
 418         if (!tty)
 419                 return;
 420
 421         tty_write_msg(tty, prefix);
 422         if ((mask & D_EMERG) || (mask & D_ERROR))
 423                 tty_write_msg(tty, "Error");
 424         tty_write_msg(tty, ": ");
 425         tty_write_msg(tty, msg);
 426         tty_kref_put(tty);
 427 }
 428
 429 static void cfs_print_to_console(struct ptldebug_header *hdr, int mask,
 430                                  const char *buf, int len, const char *file,
 431                                  const char *fn)
 432 {
 433         char *prefix = "Lustre";
 434
 435         if (hdr->ph_subsys == S_LND || hdr->ph_subsys == S_LNET)
 436                 prefix = "LNet";
 437
 438         if (mask & D_CONSOLE) {
 439                 if (mask & D_EMERG)
 440                         pr_emerg("%sError: %.*s", prefix, len, buf);
 441                 else if (mask & D_ERROR)
 442                         pr_err("%sError: %.*s", prefix, len, buf);
 443                 else if (mask & D_WARNING)
 444                         pr_warn("%s: %.*s", prefix, len, buf);
 445                 else if (mask & libcfs_printk)
 446                         pr_info("%s: %.*s", prefix, len, buf);
 447         } else {
 448                 if (mask & D_EMERG)
 449                         pr_emerg("%sError: %d:%d:(%s:%d:%s()) %.*s", prefix,
 450                                  hdr->ph_pid, hdr->ph_extern_pid, file,
 451                                  hdr->ph_line_num, fn, len, buf);
 452                 else if (mask & D_ERROR)
 453                         pr_err("%sError: %d:%d:(%s:%d:%s()) %.*s", prefix,
 454                                hdr->ph_pid, hdr->ph_extern_pid, file,
 455                                hdr->ph_line_num, fn, len, buf);
 456                 else if (mask & D_WARNING)
 457                         pr_warn("%s: %d:%d:(%s:%d:%s()) %.*s", prefix,
 458                                 hdr->ph_pid, hdr->ph_extern_pid, file,
 459                                 hdr->ph_line_num, fn, len, buf);
 460                 else if (mask & (D_CONSOLE | libcfs_printk))
 461                         pr_info("%s: %.*s", prefix, len, buf);
 462         }
 463
 464         if (mask & D_TTY)
 465                 cfs_tty_write_message(prefix, mask, buf);
 466 }
 467
 468 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 469                      const char *format, ...)
 470 {
 471         struct cfs_trace_cpu_data *tcd = NULL;
 472         struct ptldebug_header header = {0};
 473         struct cfs_trace_page *tage;
 474         /* string_buf is used only if tcd != NULL, and is always set then */
 475         char *string_buf = NULL;
 476         char *debug_buf;
 477         int known_size;
 478         int needed = 85; /* seeded with average message length */
 479         int max_nob;
 480         va_list ap;
 481         int retry;
 482         int mask = msgdata->msg_mask;
 483         char *file = (char *)msgdata->msg_file;
 484         struct cfs_debug_limit_state *cdls = msgdata->msg_cdls;
 485
 486         if (strchr(file, '/'))
 487                 file = strrchr(file, '/') + 1;
 488
 489         tcd = cfs_trace_get_tcd();
 490
 491         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 492          * pins us to a particular CPU.  This avoids an smp_processor_id()
 493          * warning on Linux when debugging is enabled.
 494          */
 495         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 496
 497         if (!tcd)                /* arch may not log in IRQ context */
 498                 goto console;
 499
 500         if (tcd->tcd_cur_pages == 0)
 501                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 502
 503         if (tcd->tcd_shutting_down) {
 504                 cfs_trace_put_tcd(tcd);
 505                 tcd = NULL;
 506                 goto console;
 507         }
 508
 509         known_size = strlen(file) + 1;
 510         if (msgdata->msg_fn)
 511                 known_size += strlen(msgdata->msg_fn) + 1;
 512
 513         if (libcfs_debug_binary)
 514                 known_size += sizeof(header);
 515
 516         /*
 517          * May perform an additional pass to update 'needed' and increase
 518          * tage buffer size to match vsnprintf reported size required
 519          * On the second pass (retry=1) use vscnprintf [which returns
 520          * number of bytes written not including the terminating nul]
 521          * to clarify `needed` is used as number of bytes written
 522          * for the remainder of this function
 523          */
 524         for (retry = 0; retry < 2; retry++) {
 525                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 526                 if (!tage) {
 527                         if (needed + known_size > PAGE_SIZE)
 528                                 mask |= D_ERROR;
 529
 530                         cfs_trace_put_tcd(tcd);
 531                         tcd = NULL;
 532                         goto console;
 533                 }
 534
 535                 string_buf = (char *)page_address(tage->page) +
 536                              tage->used + known_size;
 537
 538                 max_nob = PAGE_SIZE - tage->used - known_size;
 539                 if (max_nob <= 0) {
 540                         pr_emerg("LustreError: negative max_nob: %d\n",
 541                                  max_nob);
 542                         mask |= D_ERROR;
 543                         cfs_trace_put_tcd(tcd);
 544                         tcd = NULL;
 545                         goto console;
 546                 }
 547
 548                 va_start(ap, format);
 549                 if (retry)
 550                         needed = vscnprintf(string_buf, max_nob, format, ap);
 551                 else
 552                         needed = vsnprintf(string_buf, max_nob, format, ap);
 553                 va_end(ap);
 554
 555                 if (needed < max_nob) /* well. printing ok.. */
 556                         break;
 557         }
 558
 559         /* `needed` is actual bytes written to string_buf */
 560         if (*(string_buf + needed - 1) != '\n') {
 561                 pr_info("Lustre: format at %s:%d:%s doesn't end in newline\n",
 562                         file, msgdata->msg_line, msgdata->msg_fn);
 563         } else if (mask & D_TTY) {
 564                 /* TTY needs '\r\n' to move carriage to leftmost position */
 565                 if (needed < 2 || *(string_buf + needed - 2) != '\r')
 566                         pr_info("Lustre: format at %s:%d:%s doesn't end in '\\r\\n'\n",
 567                                 file, msgdata->msg_line, msgdata->msg_fn);
 568         }
 569
 570         header.ph_len = known_size + needed;
 571         debug_buf = (char *)page_address(tage->page) + tage->used;
 572
 573         if (libcfs_debug_binary) {
 574                 memcpy(debug_buf, &header, sizeof(header));
 575                 tage->used += sizeof(header);
 576                 debug_buf += sizeof(header);
 577         }
 578
 579         strlcpy(debug_buf, file, PAGE_SIZE - tage->used);
 580         tage->used += strlen(file) + 1;
 581         debug_buf += strlen(file) + 1;
 582
 583         if (msgdata->msg_fn) {
 584                 strlcpy(debug_buf, msgdata->msg_fn, PAGE_SIZE - tage->used);
 585                 tage->used += strlen(msgdata->msg_fn) + 1;
 586                 debug_buf += strlen(msgdata->msg_fn) + 1;
 587         }
 588
 589         __LASSERT(debug_buf == string_buf);
 590
 591         tage->used += needed;
 592         __LASSERT(tage->used <= PAGE_SIZE);
 593
 594 console:
 595         if ((mask & libcfs_printk) == 0) {
 596                 /* no console output requested */
 597                 if (tcd != NULL)
 598                         cfs_trace_put_tcd(tcd);
 599                 return 1;
 600         }
 601
 602         if (cdls != NULL) {
 603                 if (libcfs_console_ratelimit &&
 604                     cdls->cdls_next != 0 &&     /* not first time ever */
 605                     time_before(jiffies, cdls->cdls_next)) {
 606                         /* skipping a console message */
 607                         cdls->cdls_count++;
 608                         if (tcd != NULL)
 609                                 cfs_trace_put_tcd(tcd);
 610                         return 1;
 611                 }
 612
 613                 if (time_after(jiffies, cdls->cdls_next +
 614                                         libcfs_console_max_delay +
 615                                         cfs_time_seconds(10))) {
 616                         /* last timeout was a long time ago */
 617                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 618                 } else {
 619                         cdls->cdls_delay *= libcfs_console_backoff;
 620                 }
 621
 622                 if (cdls->cdls_delay < libcfs_console_min_delay)
 623                         cdls->cdls_delay = libcfs_console_min_delay;
 624                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 625                         cdls->cdls_delay = libcfs_console_max_delay;
 626
 627                 /* ensure cdls_next is never zero after it's been seen */
 628                 cdls->cdls_next = (jiffies + cdls->cdls_delay) | 1;
 629         }
 630
 631         if (tcd) {
 632                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 633                                      msgdata->msg_fn);
 634                 cfs_trace_put_tcd(tcd);
 635         } else {
 636                 string_buf = cfs_trace_get_console_buffer();
 637
 638                 va_start(ap, format);
 639                 needed = vscnprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 640                                     format, ap);
 641                 va_end(ap);
 642
 643                 cfs_print_to_console(&header, mask,
 644                                      string_buf, needed, file, msgdata->msg_fn);
 645
 646                 put_cpu();
 647         }
 648
 649         if (cdls != NULL && cdls->cdls_count != 0) {
 650                 string_buf = cfs_trace_get_console_buffer();
 651
 652                 needed = scnprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 653                                    "Skipped %d previous similar message%s\n",
 654                                    cdls->cdls_count,
 655                                    (cdls->cdls_count > 1) ? "s" : "");
 656
 657                 /* Do not allow print this to TTY */
 658                 cfs_print_to_console(&header, mask & ~D_TTY, string_buf,
 659                                      needed, file, msgdata->msg_fn);
 660
 661                 put_cpu();
 662                 cdls->cdls_count = 0;
 663         }
 664
 665         return 0;
 666 }
 667 EXPORT_SYMBOL(libcfs_debug_msg);
 668
 669 void
 670 cfs_trace_assertion_failed(const char *str,
 671                            struct libcfs_debug_msg_data *msgdata)
 672 {
 673         struct ptldebug_header hdr;
 674
 675         libcfs_panic_in_progress = 1;
 676         libcfs_catastrophe = 1;
 677         smp_mb();
 678
 679         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 680
 681         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 682                              msgdata->msg_file, msgdata->msg_fn);
 683
 684         panic("Lustre debug assertion failure\n");
 685
 686         /* not reached */
 687 }
 688
 689 static void
 690 panic_collect_pages(struct page_collection *pc)
 691 {
 692         /* Do the collect_pages job on a single CPU: assumes that all other
 693          * CPUs have been stopped during a panic.  If this isn't true for some
 694          * arch, this will have to be implemented separately in each arch.  */
 695         int                        i;
 696         int                        j;
 697         struct cfs_trace_cpu_data *tcd;
 698
 699         INIT_LIST_HEAD(&pc->pc_pages);
 700
 701         cfs_tcd_for_each(tcd, i, j) {
 702                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 703                 tcd->tcd_cur_pages = 0;
 704
 705                 if (pc->pc_want_daemon_pages) {
 706                         list_splice_init(&tcd->tcd_daemon_pages,
 707                                                 &pc->pc_pages);
 708                         tcd->tcd_cur_daemon_pages = 0;
 709                 }
 710         }
 711 }
 712
 713 static void collect_pages_on_all_cpus(struct page_collection *pc)
 714 {
 715         struct cfs_trace_cpu_data *tcd;
 716         int i, cpu;
 717
 718         for_each_possible_cpu(cpu) {
 719                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 720                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 721                         tcd->tcd_cur_pages = 0;
 722                         if (pc->pc_want_daemon_pages) {
 723                                 list_splice_init(&tcd->tcd_daemon_pages,
 724                                                         &pc->pc_pages);
 725                                 tcd->tcd_cur_daemon_pages = 0;
 726                         }
 727                 }
 728         }
 729 }
 730
 731 static void collect_pages(struct page_collection *pc)
 732 {
 733         INIT_LIST_HEAD(&pc->pc_pages);
 734
 735         if (libcfs_panic_in_progress)
 736                 panic_collect_pages(pc);
 737         else
 738                 collect_pages_on_all_cpus(pc);
 739 }
 740
 741 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 742 {
 743         struct cfs_trace_cpu_data *tcd;
 744         struct list_head *cur_head;
 745         struct cfs_trace_page *tage;
 746         struct cfs_trace_page *tmp;
 747         int i, cpu;
 748
 749         for_each_possible_cpu(cpu) {
 750                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 751                         cur_head = tcd->tcd_pages.next;
 752
 753                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 754                                                  linkage) {
 755
 756                                 __LASSERT_TAGE_INVARIANT(tage);
 757
 758                                 if (tage->cpu != cpu || tage->type != i)
 759                                         continue;
 760
 761                                 cfs_tage_to_tail(tage, cur_head);
 762                                 tcd->tcd_cur_pages++;
 763                         }
 764                 }
 765         }
 766 }
 767
 768 static void put_pages_back(struct page_collection *pc)
 769 {
 770         if (!libcfs_panic_in_progress)
 771                 put_pages_back_on_all_cpus(pc);
 772 }
 773
 774 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 775  * we have a good amount of data at all times for dumping during an LBUG, even
 776  * if we have been steadily writing (and otherwise discarding) pages via the
 777  * debug daemon. */
 778 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 779                                          struct cfs_trace_cpu_data *tcd)
 780 {
 781         struct cfs_trace_page *tage;
 782         struct cfs_trace_page *tmp;
 783
 784         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 785                 __LASSERT_TAGE_INVARIANT(tage);
 786
 787                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 788                         continue;
 789
 790                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 791                 tcd->tcd_cur_daemon_pages++;
 792
 793                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 794                         struct cfs_trace_page *victim;
 795
 796                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 797                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 798
 799                         __LASSERT_TAGE_INVARIANT(victim);
 800
 801                         list_del(&victim->linkage);
 802                         cfs_tage_free(victim);
 803                         tcd->tcd_cur_daemon_pages--;
 804                 }
 805         }
 806 }
 807
 808 static void put_pages_on_daemon_list(struct page_collection *pc)
 809 {
 810         struct cfs_trace_cpu_data *tcd;
 811         int i, cpu;
 812
 813         for_each_possible_cpu(cpu) {
 814                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 815                         put_pages_on_tcd_daemon_list(pc, tcd);
 816         }
 817 }
 818
 819 void cfs_trace_debug_print(void)
 820 {
 821         struct page_collection pc;
 822         struct cfs_trace_page *tage;
 823         struct cfs_trace_page *tmp;
 824
 825         pc.pc_want_daemon_pages = 1;
 826         collect_pages(&pc);
 827         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 828                 char *p, *file, *fn;
 829                 struct page *page;
 830
 831                 __LASSERT_TAGE_INVARIANT(tage);
 832
 833                 page = tage->page;
 834                 p = page_address(page);
 835                 while (p < ((char *)page_address(page) + tage->used)) {
 836                         struct ptldebug_header *hdr;
 837                         int len;
 838                         hdr = (void *)p;
 839                         p += sizeof(*hdr);
 840                         file = p;
 841                         p += strlen(file) + 1;
 842                         fn = p;
 843                         p += strlen(fn) + 1;
 844                         len = hdr->ph_len - (int)(p - (char *)hdr);
 845
 846                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 847
 848                         p += len;
 849                 }
 850
 851                 list_del(&tage->linkage);
 852                 cfs_tage_free(tage);
 853         }
 854 }
 855
 856 int cfs_tracefile_dump_all_pages(char *filename)
 857 {
 858         struct page_collection  pc;
 859         struct file             *filp;
 860         struct cfs_trace_page   *tage;
 861         struct cfs_trace_page   *tmp;
 862         char                    *buf;
 863         int rc;
 864
 865         down_write(&cfs_tracefile_sem);
 866
 867         filp = filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600);
 868         if (IS_ERR(filp)) {
 869                 rc = PTR_ERR(filp);
 870                 filp = NULL;
 871                 pr_err("LustreError: can't open %s for dump: rc = %d\n",
 872                       filename, rc);
 873                 goto out;
 874         }
 875
 876         pc.pc_want_daemon_pages = 1;
 877         collect_pages(&pc);
 878         if (list_empty(&pc.pc_pages)) {
 879                 rc = 0;
 880                 goto close;
 881         }
 882
 883         /* ok, for now, just write the pages.  in the future we'll be building
 884          * iobufs with the pages and calling generic_direct_IO */
 885         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 886
 887                 __LASSERT_TAGE_INVARIANT(tage);
 888
 889                 buf = kmap(tage->page);
 890                 rc = cfs_kernel_write(filp, buf, tage->used, &filp->f_pos);
 891                 kunmap(tage->page);
 892                 if (rc != (int)tage->used) {
 893                         pr_warn("Lustre: wanted to write %u but wrote %d\n",
 894                                 tage->used, rc);
 895                         put_pages_back(&pc);
 896                         __LASSERT(list_empty(&pc.pc_pages));
 897                         break;
 898                 }
 899                 list_del(&tage->linkage);
 900                 cfs_tage_free(tage);
 901         }
 902
 903         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
 904         if (rc)
 905                 pr_err("LustreError: sync returns: rc = %d\n", rc);
 906 close:
 907         filp_close(filp, NULL);
 908 out:
 909         up_write(&cfs_tracefile_sem);
 910         return rc;
 911 }
 912
 913 void cfs_trace_flush_pages(void)
 914 {
 915         struct page_collection pc;
 916         struct cfs_trace_page *tage;
 917         struct cfs_trace_page *tmp;
 918
 919         pc.pc_want_daemon_pages = 1;
 920         collect_pages(&pc);
 921         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 922
 923                 __LASSERT_TAGE_INVARIANT(tage);
 924
 925                 list_del(&tage->linkage);
 926                 cfs_tage_free(tage);
 927         }
 928 }
 929
 930 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 931                             const char __user *usr_buffer, int usr_buffer_nob)
 932 {
 933         int    nob;
 934
 935         if (usr_buffer_nob > knl_buffer_nob)
 936                 return -EOVERFLOW;
 937
 938         if (copy_from_user(knl_buffer, usr_buffer, usr_buffer_nob))
 939                 return -EFAULT;
 940
 941         nob = strnlen(knl_buffer, usr_buffer_nob);
 942         while (--nob >= 0)                      /* strip trailing whitespace */
 943                 if (!isspace(knl_buffer[nob]))
 944                         break;
 945
 946         if (nob < 0)                            /* empty string */
 947                 return -EINVAL;
 948
 949         if (nob == knl_buffer_nob)              /* no space to terminate */
 950                 return -EOVERFLOW;
 951
 952         knl_buffer[nob + 1] = 0;                /* terminate */
 953         return 0;
 954 }
 955 EXPORT_SYMBOL(cfs_trace_copyin_string);
 956
 957 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 958                              const char *knl_buffer, char *append)
 959 {
 960         /* NB if 'append' != NULL, it's a single character to append to the
 961          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 962          * terminating zero byte) for sysctl entries */
 963         int   nob = strlen(knl_buffer);
 964
 965         if (nob > usr_buffer_nob)
 966                 nob = usr_buffer_nob;
 967
 968         if (copy_to_user(usr_buffer, knl_buffer, nob))
 969                 return -EFAULT;
 970
 971         if (append != NULL && nob < usr_buffer_nob) {
 972                 if (copy_to_user(usr_buffer + nob, append, 1))
 973                         return -EFAULT;
 974
 975                 nob++;
 976         }
 977
 978         return nob;
 979 }
 980 EXPORT_SYMBOL(cfs_trace_copyout_string);
 981
 982 int cfs_trace_allocate_string_buffer(char **str, int nob)
 983 {
 984         if (nob > 2 * PAGE_SIZE)        /* string must be "sensible" */
 985                 return -EINVAL;
 986
 987         *str = kmalloc(nob, GFP_KERNEL | __GFP_ZERO);
 988         if (*str == NULL)
 989                 return -ENOMEM;
 990
 991         return 0;
 992 }
 993
 994 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
 995 {
 996         char         *str;
 997         int           rc;
 998
 999         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
1000         if (rc != 0)
1001                 return rc;
1002
1003         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
1004                                      usr_str, usr_str_nob);
1005         if (rc != 0)
1006                 goto out;
1007
1008         if (str[0] != '/') {
1009                 rc = -EINVAL;
1010                 goto out;
1011         }
1012         rc = cfs_tracefile_dump_all_pages(str);
1013 out:
1014         kfree(str);
1015         return rc;
1016 }
1017
1018 int cfs_trace_daemon_command(char *str)
1019 {
1020         int       rc = 0;
1021
1022         down_write(&cfs_tracefile_sem);
1023
1024         if (strcmp(str, "stop") == 0) {
1025                 up_write(&cfs_tracefile_sem);
1026                 cfs_trace_stop_thread();
1027                 down_write(&cfs_tracefile_sem);
1028                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
1029
1030         } else if (strncmp(str, "size=", 5) == 0) {
1031                 unsigned long tmp;
1032
1033                 rc = kstrtoul(str + 5, 10, &tmp);
1034                 if (!rc) {
1035                         if (tmp < 10 || tmp > 20480)
1036                                 cfs_tracefile_size = CFS_TRACEFILE_SIZE;
1037                         else
1038                                 cfs_tracefile_size = tmp << 20;
1039                 }
1040         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
1041                 rc = -ENAMETOOLONG;
1042         } else if (str[0] != '/') {
1043                 rc = -EINVAL;
1044         } else {
1045                 strcpy(cfs_tracefile, str);
1046
1047                 pr_info("Lustre: debug daemon will attempt to start writing to %s (%lukB max)\n",
1048                         cfs_tracefile, (long)(cfs_tracefile_size >> 10));
1049
1050                 cfs_trace_start_thread();
1051         }
1052
1053         up_write(&cfs_tracefile_sem);
1054         return rc;
1055 }
1056
1057 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
1058 {
1059         char *str;
1060         int   rc;
1061
1062         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
1063         if (rc != 0)
1064                 return rc;
1065
1066         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
1067                                  usr_str, usr_str_nob);
1068         if (rc == 0)
1069                 rc = cfs_trace_daemon_command(str);
1070
1071         kfree(str);
1072         return rc;
1073 }
1074
1075 int cfs_trace_set_debug_mb(int mb)
1076 {
1077         int i;
1078         int j;
1079         unsigned long pages;
1080         unsigned long total_mb = (cfs_totalram_pages() >> (20 - PAGE_SHIFT));
1081         unsigned long limit = max_t(unsigned long, 512, (total_mb * 4) / 5);
1082         struct cfs_trace_cpu_data *tcd;
1083
1084         if (mb < num_possible_cpus()) {
1085                 pr_warn("Lustre: %d MB is too small for debug buffer size, setting it to %d MB.\n",
1086                         mb, num_possible_cpus());
1087                 mb = num_possible_cpus();
1088         }
1089
1090         if (mb > limit) {
1091                 pr_warn("Lustre: %d MB is too large for debug buffer size, setting it to %lu MB.\n",
1092                         mb, limit);
1093                 mb = limit;
1094         }
1095
1096         mb /= num_possible_cpus();
1097         pages = mb << (20 - PAGE_SHIFT);
1098
1099         down_write(&cfs_tracefile_sem);
1100
1101         cfs_tcd_for_each(tcd, i, j)
1102                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
1103
1104         up_write(&cfs_tracefile_sem);
1105
1106         return mb;
1107 }
1108
1109 int cfs_trace_get_debug_mb(void)
1110 {
1111         int i;
1112         int j;
1113         struct cfs_trace_cpu_data *tcd;
1114         int total_pages = 0;
1115
1116         down_read(&cfs_tracefile_sem);
1117
1118         cfs_tcd_for_each(tcd, i, j)
1119                 total_pages += tcd->tcd_max_pages;
1120
1121         up_read(&cfs_tracefile_sem);
1122
1123         return (total_pages >> (20 - PAGE_SHIFT)) + 1;
1124 }
1125
1126 static int tracefiled(void *arg)
1127 {
1128         struct page_collection pc;
1129         struct tracefiled_ctl *tctl = arg;
1130         struct cfs_trace_page *tage;
1131         struct cfs_trace_page *tmp;
1132         struct file *filp;
1133         char *buf;
1134         int last_loop = 0;
1135         int rc;
1136
1137         /* we're started late enough that we pick up init's fs context */
1138         /* this is so broken in uml?  what on earth is going on? */
1139
1140         complete(&tctl->tctl_start);
1141
1142         while (1) {
1143                 wait_queue_entry_t __wait;
1144
1145                 pc.pc_want_daemon_pages = 0;
1146                 collect_pages(&pc);
1147                 if (list_empty(&pc.pc_pages))
1148                         goto end_loop;
1149
1150                 filp = NULL;
1151                 down_read(&cfs_tracefile_sem);
1152                 if (cfs_tracefile[0] != 0) {
1153                         filp = filp_open(cfs_tracefile,
1154                                          O_CREAT | O_RDWR | O_LARGEFILE,
1155                                          0600);
1156                         if (IS_ERR(filp)) {
1157                                 rc = PTR_ERR(filp);
1158                                 filp = NULL;
1159                                 pr_warn("Lustre: couldn't open %s: rc = %d\n",
1160                                         cfs_tracefile, rc);
1161                         }
1162                 }
1163                 up_read(&cfs_tracefile_sem);
1164                 if (filp == NULL) {
1165                         put_pages_on_daemon_list(&pc);
1166                         __LASSERT(list_empty(&pc.pc_pages));
1167                         goto end_loop;
1168                 }
1169
1170                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
1171                         struct dentry *de = file_dentry(filp);
1172                         static loff_t f_pos;
1173
1174                         __LASSERT_TAGE_INVARIANT(tage);
1175
1176                         if (f_pos >= (off_t)cfs_tracefile_size)
1177                                 f_pos = 0;
1178                         else if (f_pos > i_size_read(de->d_inode))
1179                                 f_pos = i_size_read(de->d_inode);
1180
1181                         buf = kmap(tage->page);
1182                         rc = cfs_kernel_write(filp, buf, tage->used, &f_pos);
1183                         kunmap(tage->page);
1184                         if (rc != (int)tage->used) {
1185                                 pr_warn("Lustre: wanted to write %u but wrote %d\n",
1186                                         tage->used, rc);
1187                                 put_pages_back(&pc);
1188                                 __LASSERT(list_empty(&pc.pc_pages));
1189                                 break;
1190                         }
1191                 }
1192
1193                 filp_close(filp, NULL);
1194                 put_pages_on_daemon_list(&pc);
1195                 if (!list_empty(&pc.pc_pages)) {
1196                         int i;
1197
1198                         pr_alert("Lustre: trace pages aren't empty\n");
1199                         pr_err("Lustre: total cpus(%d): ", num_possible_cpus());
1200                         for (i = 0; i < num_possible_cpus(); i++)
1201                                 if (cpu_online(i))
1202                                         pr_cont("%d(on) ", i);
1203                                 else
1204                                         pr_cont("%d(off) ", i);
1205                         pr_cont("\n");
1206
1207                         i = 0;
1208                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1209                                                  linkage)
1210                                 pr_err("Lustre: page %d belongs to cpu %d\n",
1211                                        ++i, tage->cpu);
1212                         pr_err("Lustre: There are %d pages unwritten\n", i);
1213                 }
1214                 __LASSERT(list_empty(&pc.pc_pages));
1215 end_loop:
1216                 if (atomic_read(&tctl->tctl_shutdown)) {
1217                         if (last_loop == 0) {
1218                                 last_loop = 1;
1219                                 continue;
1220                         } else {
1221                                 break;
1222                         }
1223                 }
1224                 init_waitqueue_entry(&__wait, current);
1225                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1226                 schedule_timeout_interruptible(cfs_time_seconds(1));
1227                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1228         }
1229         complete(&tctl->tctl_stop);
1230         return 0;
1231 }
1232
1233 int cfs_trace_start_thread(void)
1234 {
1235         struct tracefiled_ctl *tctl = &trace_tctl;
1236         int rc = 0;
1237
1238         mutex_lock(&cfs_trace_thread_mutex);
1239         if (thread_running)
1240                 goto out;
1241
1242         init_completion(&tctl->tctl_start);
1243         init_completion(&tctl->tctl_stop);
1244         init_waitqueue_head(&tctl->tctl_waitq);
1245         atomic_set(&tctl->tctl_shutdown, 0);
1246
1247         if (IS_ERR(kthread_run(tracefiled, tctl, "ktracefiled"))) {
1248                 rc = -ECHILD;
1249                 goto out;
1250         }
1251
1252         wait_for_completion(&tctl->tctl_start);
1253         thread_running = 1;
1254 out:
1255         mutex_unlock(&cfs_trace_thread_mutex);
1256         return rc;
1257 }
1258
1259 void cfs_trace_stop_thread(void)
1260 {
1261         struct tracefiled_ctl *tctl = &trace_tctl;
1262
1263         mutex_lock(&cfs_trace_thread_mutex);
1264         if (thread_running) {
1265                 pr_info("Lustre: shutting down debug daemon thread...\n");
1266                 atomic_set(&tctl->tctl_shutdown, 1);
1267                 wait_for_completion(&tctl->tctl_stop);
1268                 thread_running = 0;
1269         }
1270         mutex_unlock(&cfs_trace_thread_mutex);
1271 }
1272
1273 int cfs_tracefile_init(int max_pages)
1274 {
1275         struct cfs_trace_cpu_data *tcd;
1276         int     i;
1277         int     j;
1278         int     rc;
1279         int     factor;
1280
1281         rc = cfs_tracefile_init_arch();
1282         if (rc != 0)
1283                 return rc;
1284
1285         cfs_tcd_for_each(tcd, i, j) {
1286                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1287                 factor = tcd->tcd_pages_factor;
1288                 INIT_LIST_HEAD(&tcd->tcd_pages);
1289                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1290                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1291                 tcd->tcd_cur_pages = 0;
1292                 tcd->tcd_cur_stock_pages = 0;
1293                 tcd->tcd_cur_daemon_pages = 0;
1294                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1295                 LASSERT(tcd->tcd_max_pages > 0);
1296                 tcd->tcd_shutting_down = 0;
1297         }
1298         return 0;
1299 }
1300
1301 static void trace_cleanup_on_all_cpus(void)
1302 {
1303         struct cfs_trace_cpu_data *tcd;
1304         struct cfs_trace_page *tage;
1305         struct cfs_trace_page *tmp;
1306         int i, cpu;
1307
1308         for_each_possible_cpu(cpu) {
1309                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1310                         tcd->tcd_shutting_down = 1;
1311
1312                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1313                                 __LASSERT_TAGE_INVARIANT(tage);
1314
1315                                 list_del(&tage->linkage);
1316                                 cfs_tage_free(tage);
1317                         }
1318                         tcd->tcd_cur_pages = 0;
1319                 }
1320         }
1321 }
1322
1323 static void cfs_trace_cleanup(void)
1324 {
1325         struct page_collection pc;
1326
1327         INIT_LIST_HEAD(&pc.pc_pages);
1328
1329         trace_cleanup_on_all_cpus();
1330
1331         cfs_tracefile_fini_arch();
1332 }
1333
1334 void cfs_tracefile_exit(void)
1335 {
1336         cfs_trace_stop_thread();
1337         cfs_trace_cleanup();
1338 }