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 #define DEBUG_SUBSYSTEM S_LNET
  39 #include "tracefile.h"
  40
  41 #include <linux/ctype.h>
  42 #include <linux/fs.h>
  43 #include <linux/kthread.h>
  44 #include <linux/pagemap.h>
  45 #include <linux/poll.h>
  46 #include <linux/tty.h>
  47 #include <linux/uaccess.h>
  48 #include <libcfs/linux/linux-fs.h>
  49 #include <libcfs/libcfs.h>
  50
  51 #define CFS_TRACE_CONSOLE_BUFFER_SIZE   1024
  52
  53 enum cfs_trace_buf_type {
  54         CFS_TCD_TYPE_PROC = 0,
  55         CFS_TCD_TYPE_SOFTIRQ,
  56         CFS_TCD_TYPE_IRQ,
  57         CFS_TCD_TYPE_CNT
  58 };
  59
  60 union cfs_trace_data_union (*cfs_trace_data[CFS_TCD_TYPE_CNT])[NR_CPUS] __cacheline_aligned;
  61
  62 char *cfs_trace_console_buffers[NR_CPUS][CFS_TCD_TYPE_CNT];
  63 char cfs_tracefile[TRACEFILE_NAME_SIZE];
  64 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
  65 static struct tracefiled_ctl trace_tctl;
  66 static DEFINE_MUTEX(cfs_trace_thread_mutex);
  67 static int thread_running = 0;
  68
  69 static atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
  70 static DECLARE_RWSEM(cfs_tracefile_sem);
  71
  72 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  73                                         struct cfs_trace_cpu_data *tcd);
  74
  75 /* trace file lock routines */
  76 /* The walking argument indicates the locking comes from all tcd types
  77  * iterator and we must lock it and dissable local irqs to avoid deadlocks
  78  * with other interrupt locks that might be happening. See LU-1311
  79  * for details.
  80  */
  81 int cfs_trace_lock_tcd(struct cfs_trace_cpu_data *tcd, int walking)
  82         __acquires(&tcd->tcd_lock)
  83 {
  84         __LASSERT(tcd->tcd_type < CFS_TCD_TYPE_CNT);
  85         if (tcd->tcd_type == CFS_TCD_TYPE_IRQ)
  86                 spin_lock_irqsave(&tcd->tcd_lock, tcd->tcd_lock_flags);
  87         else if (tcd->tcd_type == CFS_TCD_TYPE_SOFTIRQ)
  88                 spin_lock_bh(&tcd->tcd_lock);
  89         else if (unlikely(walking))
  90                 spin_lock_irq(&tcd->tcd_lock);
  91         else
  92                 spin_lock(&tcd->tcd_lock);
  93         return 1;
  94 }
  95
  96 void cfs_trace_unlock_tcd(struct cfs_trace_cpu_data *tcd, int walking)
  97         __releases(&tcd->tcd_lock)
  98 {
  99         __LASSERT(tcd->tcd_type < CFS_TCD_TYPE_CNT);
 100         if (tcd->tcd_type == CFS_TCD_TYPE_IRQ)
 101                 spin_unlock_irqrestore(&tcd->tcd_lock, tcd->tcd_lock_flags);
 102         else if (tcd->tcd_type == CFS_TCD_TYPE_SOFTIRQ)
 103                 spin_unlock_bh(&tcd->tcd_lock);
 104         else if (unlikely(walking))
 105                 spin_unlock_irq(&tcd->tcd_lock);
 106         else
 107                 spin_unlock(&tcd->tcd_lock);
 108 }
 109
 110 #define cfs_tcd_for_each(tcd, i, j)                                     \
 111         for (i = 0; i < CFS_TCD_TYPE_CNT && cfs_trace_data[i]; i++)     \
 112                 for (j = 0, ((tcd) = &(*cfs_trace_data[i])[j].tcd);     \
 113                      j < num_possible_cpus();                           \
 114                      j++, (tcd) = &(*cfs_trace_data[i])[j].tcd)
 115
 116 #define cfs_tcd_for_each_type_lock(tcd, i, cpu)                         \
 117         for (i = 0; i < CFS_TCD_TYPE_CNT && cfs_trace_data[i] &&        \
 118              (tcd = &(*cfs_trace_data[i])[cpu].tcd) &&                  \
 119              cfs_trace_lock_tcd(tcd, 1); cfs_trace_unlock_tcd(tcd, 1), i++)
 120
 121 enum cfs_trace_buf_type cfs_trace_buf_idx_get(void)
 122 {
 123         if (in_irq())
 124                 return CFS_TCD_TYPE_IRQ;
 125         if (in_softirq())
 126                 return CFS_TCD_TYPE_SOFTIRQ;
 127         return CFS_TCD_TYPE_PROC;
 128 }
 129
 130 static inline char *cfs_trace_get_console_buffer(void)
 131 {
 132         unsigned int i = get_cpu();
 133         unsigned int j = cfs_trace_buf_idx_get();
 134
 135         return cfs_trace_console_buffers[i][j];
 136 }
 137
 138 static inline struct cfs_trace_cpu_data *
 139 cfs_trace_get_tcd(void)
 140 {
 141         struct cfs_trace_cpu_data *tcd =
 142                 &(*cfs_trace_data[cfs_trace_buf_idx_get()])[get_cpu()].tcd;
 143
 144         cfs_trace_lock_tcd(tcd, 0);
 145
 146         return tcd;
 147 }
 148
 149 static inline void cfs_trace_put_tcd(struct cfs_trace_cpu_data *tcd)
 150 {
 151         cfs_trace_unlock_tcd(tcd, 0);
 152
 153         put_cpu();
 154 }
 155
 156 static inline struct cfs_trace_page *
 157 cfs_tage_from_list(struct list_head *list)
 158 {
 159         return list_entry(list, struct cfs_trace_page, linkage);
 160 }
 161
 162 static struct cfs_trace_page *cfs_tage_alloc(gfp_t gfp)
 163 {
 164         struct page            *page;
 165         struct cfs_trace_page *tage;
 166
 167         /* My caller is trying to free memory */
 168         if (!in_interrupt() && (current->flags & PF_MEMALLOC))
 169                 return NULL;
 170
 171         /*
 172          * Don't spam console with allocation failures: they will be reported
 173          * by upper layer anyway.
 174          */
 175         gfp |= __GFP_NOWARN;
 176         page = alloc_page(gfp);
 177         if (page == NULL)
 178                 return NULL;
 179
 180         tage = kmalloc(sizeof(*tage), gfp);
 181         if (tage == NULL) {
 182                 __free_page(page);
 183                 return NULL;
 184         }
 185
 186         tage->page = page;
 187         atomic_inc(&cfs_tage_allocated);
 188         return tage;
 189 }
 190
 191 static void cfs_tage_free(struct cfs_trace_page *tage)
 192 {
 193         __LASSERT(tage != NULL);
 194         __LASSERT(tage->page != NULL);
 195
 196         __free_page(tage->page);
 197         kfree(tage);
 198         atomic_dec(&cfs_tage_allocated);
 199 }
 200
 201 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
 202                              struct list_head *queue)
 203 {
 204         __LASSERT(tage != NULL);
 205         __LASSERT(queue != NULL);
 206
 207         list_move_tail(&tage->linkage, queue);
 208 }
 209
 210 /* return a page that has 'len' bytes left at the end */
 211 static struct cfs_trace_page *
 212 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
 213 {
 214         struct cfs_trace_page *tage;
 215
 216         if (tcd->tcd_cur_pages > 0) {
 217                 __LASSERT(!list_empty(&tcd->tcd_pages));
 218                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
 219                 if (tage->used + len <= PAGE_SIZE)
 220                         return tage;
 221         }
 222
 223         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 224                 if (tcd->tcd_cur_stock_pages > 0) {
 225                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
 226                         --tcd->tcd_cur_stock_pages;
 227                         list_del_init(&tage->linkage);
 228                 } else {
 229                         tage = cfs_tage_alloc(GFP_ATOMIC);
 230                         if (unlikely(tage == NULL)) {
 231                                 if ((!(current->flags & PF_MEMALLOC) ||
 232                                      in_interrupt()) && printk_ratelimit())
 233                                         pr_warn("Lustre: cannot allocate a tage (%ld)\n",
 234                                                 tcd->tcd_cur_pages);
 235                                 return NULL;
 236                         }
 237                 }
 238
 239                 tage->used = 0;
 240                 tage->cpu = smp_processor_id();
 241                 tage->type = tcd->tcd_type;
 242                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 243                 tcd->tcd_cur_pages++;
 244
 245                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 246                         struct tracefiled_ctl *tctl = &trace_tctl;
 247                         /*
 248                          * wake up tracefiled to process some pages.
 249                          */
 250                         wake_up(&tctl->tctl_waitq);
 251                 }
 252                 return tage;
 253         }
 254         return NULL;
 255 }
 256
 257 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
 258 {
 259         int pgcount = tcd->tcd_cur_pages / 10;
 260         struct page_collection pc;
 261         struct cfs_trace_page *tage;
 262         struct cfs_trace_page *tmp;
 263
 264         /*
 265          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 266          * from here: this will lead to infinite recursion.
 267          */
 268
 269         if (printk_ratelimit())
 270                 pr_warn("Lustre: debug daemon buffer overflowed; discarding 10%% of pages (%d of %ld)\n",
 271                         pgcount + 1, tcd->tcd_cur_pages);
 272
 273         INIT_LIST_HEAD(&pc.pc_pages);
 274
 275         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 276                 if (pgcount-- == 0)
 277                         break;
 278
 279                 list_move_tail(&tage->linkage, &pc.pc_pages);
 280                 tcd->tcd_cur_pages--;
 281         }
 282         put_pages_on_tcd_daemon_list(&pc, tcd);
 283 }
 284
 285 /* return a page that has 'len' bytes left at the end */
 286 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 287                                                  unsigned long len)
 288 {
 289         struct cfs_trace_page *tage;
 290
 291         /*
 292          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 293          * from here: this will lead to infinite recursion.
 294          */
 295
 296         if (len > PAGE_SIZE) {
 297                 pr_err("LustreError: cowardly refusing to write %lu bytes in a page\n",
 298                        len);
 299                 return NULL;
 300         }
 301
 302         tage = cfs_trace_get_tage_try(tcd, len);
 303         if (tage != NULL)
 304                 return tage;
 305         if (thread_running)
 306                 cfs_tcd_shrink(tcd);
 307         if (tcd->tcd_cur_pages > 0) {
 308                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 309                 tage->used = 0;
 310                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 311         }
 312         return tage;
 313 }
 314
 315 static void cfs_set_ptldebug_header(struct ptldebug_header *header,
 316                                     struct libcfs_debug_msg_data *msgdata,
 317                                     unsigned long stack)
 318 {
 319         struct timespec64 ts;
 320
 321         ktime_get_real_ts64(&ts);
 322
 323         header->ph_subsys = msgdata->msg_subsys;
 324         header->ph_mask = msgdata->msg_mask;
 325         header->ph_cpu_id = smp_processor_id();
 326         header->ph_type = cfs_trace_buf_idx_get();
 327         /* y2038 safe since all user space treats this as unsigned, but
 328          * will overflow in 2106
 329          */
 330         header->ph_sec = (u32)ts.tv_sec;
 331         header->ph_usec = ts.tv_nsec / NSEC_PER_USEC;
 332         header->ph_stack = stack;
 333         header->ph_pid = current->pid;
 334         header->ph_line_num = msgdata->msg_line;
 335         header->ph_extern_pid = 0;
 336 }
 337
 338 /**
 339  * tty_write_msg - write a message to a certain tty, not just the console.
 340  * @tty: the destination tty_struct
 341  * @msg: the message to write
 342  *
 343  * tty_write_message is not exported, so write a same function for it
 344  *
 345  */
 346 static void tty_write_msg(struct tty_struct *tty, const char *msg)
 347 {
 348         mutex_lock(&tty->atomic_write_lock);
 349         tty_lock(tty);
 350         if (tty->ops->write && tty->count > 0)
 351                 tty->ops->write(tty, msg, strlen(msg));
 352         tty_unlock(tty);
 353         mutex_unlock(&tty->atomic_write_lock);
 354         wake_up_interruptible_poll(&tty->write_wait, POLLOUT);
 355 }
 356
 357 static void cfs_tty_write_message(const char *prefix, int mask, const char *msg)
 358 {
 359         struct tty_struct *tty;
 360
 361         tty = get_current_tty();
 362         if (!tty)
 363                 return;
 364
 365         tty_write_msg(tty, prefix);
 366         if ((mask & D_EMERG) || (mask & D_ERROR))
 367                 tty_write_msg(tty, "Error");
 368         tty_write_msg(tty, ": ");
 369         tty_write_msg(tty, msg);
 370         tty_kref_put(tty);
 371 }
 372
 373 static void cfs_print_to_console(struct ptldebug_header *hdr, int mask,
 374                                  const char *buf, int len, const char *file,
 375                                  const char *fn)
 376 {
 377         char *prefix = "Lustre";
 378
 379         if (hdr->ph_subsys == S_LND || hdr->ph_subsys == S_LNET)
 380                 prefix = "LNet";
 381
 382         if (mask & D_CONSOLE) {
 383                 if (mask & D_EMERG)
 384                         pr_emerg("%sError: %.*s", prefix, len, buf);
 385                 else if (mask & D_ERROR)
 386                         pr_err("%sError: %.*s", prefix, len, buf);
 387                 else if (mask & D_WARNING)
 388                         pr_warn("%s: %.*s", prefix, len, buf);
 389                 else if (mask & libcfs_printk)
 390                         pr_info("%s: %.*s", prefix, len, buf);
 391         } else {
 392                 if (mask & D_EMERG)
 393                         pr_emerg("%sError: %d:%d:(%s:%d:%s()) %.*s", prefix,
 394                                  hdr->ph_pid, hdr->ph_extern_pid, file,
 395                                  hdr->ph_line_num, fn, len, buf);
 396                 else if (mask & D_ERROR)
 397                         pr_err("%sError: %d:%d:(%s:%d:%s()) %.*s", prefix,
 398                                hdr->ph_pid, hdr->ph_extern_pid, file,
 399                                hdr->ph_line_num, fn, len, buf);
 400                 else if (mask & D_WARNING)
 401                         pr_warn("%s: %d:%d:(%s:%d:%s()) %.*s", prefix,
 402                                 hdr->ph_pid, hdr->ph_extern_pid, file,
 403                                 hdr->ph_line_num, fn, len, buf);
 404                 else if (mask & (D_CONSOLE | libcfs_printk))
 405                         pr_info("%s: %.*s", prefix, len, buf);
 406         }
 407
 408         if (mask & D_TTY)
 409                 cfs_tty_write_message(prefix, mask, buf);
 410 }
 411
 412 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 413                      const char *format, ...)
 414 {
 415         struct cfs_trace_cpu_data *tcd = NULL;
 416         struct ptldebug_header header = {0};
 417         struct cfs_trace_page *tage;
 418         /* string_buf is used only if tcd != NULL, and is always set then */
 419         char *string_buf = NULL;
 420         char *debug_buf;
 421         int known_size;
 422         int needed = 85; /* seeded with average message length */
 423         int max_nob;
 424         va_list ap;
 425         int retry;
 426         int mask = msgdata->msg_mask;
 427         char *file = (char *)msgdata->msg_file;
 428         struct cfs_debug_limit_state *cdls = msgdata->msg_cdls;
 429
 430         if (strchr(file, '/'))
 431                 file = strrchr(file, '/') + 1;
 432
 433         tcd = cfs_trace_get_tcd();
 434
 435         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 436          * pins us to a particular CPU.  This avoids an smp_processor_id()
 437          * warning on Linux when debugging is enabled.
 438          */
 439         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 440
 441         if (!tcd)                /* arch may not log in IRQ context */
 442                 goto console;
 443
 444         if (tcd->tcd_cur_pages == 0)
 445                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 446
 447         if (tcd->tcd_shutting_down) {
 448                 cfs_trace_put_tcd(tcd);
 449                 tcd = NULL;
 450                 goto console;
 451         }
 452
 453         known_size = strlen(file) + 1;
 454         if (msgdata->msg_fn)
 455                 known_size += strlen(msgdata->msg_fn) + 1;
 456
 457         if (libcfs_debug_binary)
 458                 known_size += sizeof(header);
 459
 460         /*
 461          * May perform an additional pass to update 'needed' and increase
 462          * tage buffer size to match vsnprintf reported size required
 463          * On the second pass (retry=1) use vscnprintf [which returns
 464          * number of bytes written not including the terminating nul]
 465          * to clarify `needed` is used as number of bytes written
 466          * for the remainder of this function
 467          */
 468         for (retry = 0; retry < 2; retry++) {
 469                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 470                 if (!tage) {
 471                         if (needed + known_size > PAGE_SIZE)
 472                                 mask |= D_ERROR;
 473
 474                         cfs_trace_put_tcd(tcd);
 475                         tcd = NULL;
 476                         goto console;
 477                 }
 478
 479                 string_buf = (char *)page_address(tage->page) +
 480                              tage->used + known_size;
 481
 482                 max_nob = PAGE_SIZE - tage->used - known_size;
 483                 if (max_nob <= 0) {
 484                         pr_emerg("LustreError: negative max_nob: %d\n",
 485                                  max_nob);
 486                         mask |= D_ERROR;
 487                         cfs_trace_put_tcd(tcd);
 488                         tcd = NULL;
 489                         goto console;
 490                 }
 491
 492                 va_start(ap, format);
 493                 if (retry)
 494                         needed = vscnprintf(string_buf, max_nob, format, ap);
 495                 else
 496                         needed = vsnprintf(string_buf, max_nob, format, ap);
 497                 va_end(ap);
 498
 499                 if (needed < max_nob) /* well. printing ok.. */
 500                         break;
 501         }
 502
 503         /* `needed` is actual bytes written to string_buf */
 504         if (*(string_buf + needed - 1) != '\n') {
 505                 pr_info("Lustre: format at %s:%d:%s doesn't end in newline\n",
 506                         file, msgdata->msg_line, msgdata->msg_fn);
 507         } else if (mask & D_TTY) {
 508                 /* TTY needs '\r\n' to move carriage to leftmost position */
 509                 if (needed < 2 || *(string_buf + needed - 2) != '\r')
 510                         pr_info("Lustre: format at %s:%d:%s doesn't end in '\\r\\n'\n",
 511                                 file, msgdata->msg_line, msgdata->msg_fn);
 512         }
 513
 514         header.ph_len = known_size + needed;
 515         debug_buf = (char *)page_address(tage->page) + tage->used;
 516
 517         if (libcfs_debug_binary) {
 518                 memcpy(debug_buf, &header, sizeof(header));
 519                 tage->used += sizeof(header);
 520                 debug_buf += sizeof(header);
 521         }
 522
 523         strlcpy(debug_buf, file, PAGE_SIZE - tage->used);
 524         tage->used += strlen(file) + 1;
 525         debug_buf += strlen(file) + 1;
 526
 527         if (msgdata->msg_fn) {
 528                 strlcpy(debug_buf, msgdata->msg_fn, PAGE_SIZE - tage->used);
 529                 tage->used += strlen(msgdata->msg_fn) + 1;
 530                 debug_buf += strlen(msgdata->msg_fn) + 1;
 531         }
 532
 533         __LASSERT(debug_buf == string_buf);
 534
 535         tage->used += needed;
 536         __LASSERT(tage->used <= PAGE_SIZE);
 537
 538 console:
 539         if ((mask & libcfs_printk) == 0) {
 540                 /* no console output requested */
 541                 if (tcd != NULL)
 542                         cfs_trace_put_tcd(tcd);
 543                 return 1;
 544         }
 545
 546         if (cdls != NULL) {
 547                 if (libcfs_console_ratelimit &&
 548                     cdls->cdls_next != 0 &&     /* not first time ever */
 549                     time_before(jiffies, cdls->cdls_next)) {
 550                         /* skipping a console message */
 551                         cdls->cdls_count++;
 552                         if (tcd != NULL)
 553                                 cfs_trace_put_tcd(tcd);
 554                         return 1;
 555                 }
 556
 557                 if (time_after(jiffies, cdls->cdls_next +
 558                                         libcfs_console_max_delay +
 559                                         cfs_time_seconds(10))) {
 560                         /* last timeout was a long time ago */
 561                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 562                 } else {
 563                         cdls->cdls_delay *= libcfs_console_backoff;
 564                 }
 565
 566                 if (cdls->cdls_delay < libcfs_console_min_delay)
 567                         cdls->cdls_delay = libcfs_console_min_delay;
 568                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 569                         cdls->cdls_delay = libcfs_console_max_delay;
 570
 571                 /* ensure cdls_next is never zero after it's been seen */
 572                 cdls->cdls_next = (jiffies + cdls->cdls_delay) | 1;
 573         }
 574
 575         if (tcd) {
 576                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 577                                      msgdata->msg_fn);
 578                 cfs_trace_put_tcd(tcd);
 579         } else {
 580                 string_buf = cfs_trace_get_console_buffer();
 581
 582                 va_start(ap, format);
 583                 needed = vscnprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 584                                     format, ap);
 585                 va_end(ap);
 586
 587                 cfs_print_to_console(&header, mask,
 588                                      string_buf, needed, file, msgdata->msg_fn);
 589
 590                 put_cpu();
 591         }
 592
 593         if (cdls != NULL && cdls->cdls_count != 0) {
 594                 string_buf = cfs_trace_get_console_buffer();
 595
 596                 needed = scnprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 597                                    "Skipped %d previous similar message%s\n",
 598                                    cdls->cdls_count,
 599                                    (cdls->cdls_count > 1) ? "s" : "");
 600
 601                 /* Do not allow print this to TTY */
 602                 cfs_print_to_console(&header, mask & ~D_TTY, string_buf,
 603                                      needed, file, msgdata->msg_fn);
 604
 605                 put_cpu();
 606                 cdls->cdls_count = 0;
 607         }
 608
 609         return 0;
 610 }
 611 EXPORT_SYMBOL(libcfs_debug_msg);
 612
 613 void
 614 cfs_trace_assertion_failed(const char *str,
 615                            struct libcfs_debug_msg_data *msgdata)
 616 {
 617         struct ptldebug_header hdr;
 618
 619         libcfs_panic_in_progress = 1;
 620         libcfs_catastrophe = 1;
 621         smp_mb();
 622
 623         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 624
 625         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 626                              msgdata->msg_file, msgdata->msg_fn);
 627
 628         panic("Lustre debug assertion failure\n");
 629
 630         /* not reached */
 631 }
 632
 633 static void
 634 panic_collect_pages(struct page_collection *pc)
 635 {
 636         /* Do the collect_pages job on a single CPU: assumes that all other
 637          * CPUs have been stopped during a panic.  If this isn't true for some
 638          * arch, this will have to be implemented separately in each arch.  */
 639         int                        i;
 640         int                        j;
 641         struct cfs_trace_cpu_data *tcd;
 642
 643         INIT_LIST_HEAD(&pc->pc_pages);
 644
 645         cfs_tcd_for_each(tcd, i, j) {
 646                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 647                 tcd->tcd_cur_pages = 0;
 648
 649                 if (pc->pc_want_daemon_pages) {
 650                         list_splice_init(&tcd->tcd_daemon_pages,
 651                                                 &pc->pc_pages);
 652                         tcd->tcd_cur_daemon_pages = 0;
 653                 }
 654         }
 655 }
 656
 657 static void collect_pages_on_all_cpus(struct page_collection *pc)
 658 {
 659         struct cfs_trace_cpu_data *tcd;
 660         int i, cpu;
 661
 662         for_each_possible_cpu(cpu) {
 663                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 664                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 665                         tcd->tcd_cur_pages = 0;
 666                         if (pc->pc_want_daemon_pages) {
 667                                 list_splice_init(&tcd->tcd_daemon_pages,
 668                                                         &pc->pc_pages);
 669                                 tcd->tcd_cur_daemon_pages = 0;
 670                         }
 671                 }
 672         }
 673 }
 674
 675 static void collect_pages(struct page_collection *pc)
 676 {
 677         INIT_LIST_HEAD(&pc->pc_pages);
 678
 679         if (libcfs_panic_in_progress)
 680                 panic_collect_pages(pc);
 681         else
 682                 collect_pages_on_all_cpus(pc);
 683 }
 684
 685 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 686 {
 687         struct cfs_trace_cpu_data *tcd;
 688         struct list_head *cur_head;
 689         struct cfs_trace_page *tage;
 690         struct cfs_trace_page *tmp;
 691         int i, cpu;
 692
 693         for_each_possible_cpu(cpu) {
 694                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 695                         cur_head = tcd->tcd_pages.next;
 696
 697                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 698                                                  linkage) {
 699
 700                                 __LASSERT_TAGE_INVARIANT(tage);
 701
 702                                 if (tage->cpu != cpu || tage->type != i)
 703                                         continue;
 704
 705                                 cfs_tage_to_tail(tage, cur_head);
 706                                 tcd->tcd_cur_pages++;
 707                         }
 708                 }
 709         }
 710 }
 711
 712 static void put_pages_back(struct page_collection *pc)
 713 {
 714         if (!libcfs_panic_in_progress)
 715                 put_pages_back_on_all_cpus(pc);
 716 }
 717
 718 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 719  * we have a good amount of data at all times for dumping during an LBUG, even
 720  * if we have been steadily writing (and otherwise discarding) pages via the
 721  * debug daemon. */
 722 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 723                                          struct cfs_trace_cpu_data *tcd)
 724 {
 725         struct cfs_trace_page *tage;
 726         struct cfs_trace_page *tmp;
 727
 728         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 729                 __LASSERT_TAGE_INVARIANT(tage);
 730
 731                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 732                         continue;
 733
 734                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 735                 tcd->tcd_cur_daemon_pages++;
 736
 737                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 738                         struct cfs_trace_page *victim;
 739
 740                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 741                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 742
 743                         __LASSERT_TAGE_INVARIANT(victim);
 744
 745                         list_del(&victim->linkage);
 746                         cfs_tage_free(victim);
 747                         tcd->tcd_cur_daemon_pages--;
 748                 }
 749         }
 750 }
 751
 752 static void put_pages_on_daemon_list(struct page_collection *pc)
 753 {
 754         struct cfs_trace_cpu_data *tcd;
 755         int i, cpu;
 756
 757         for_each_possible_cpu(cpu) {
 758                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 759                         put_pages_on_tcd_daemon_list(pc, tcd);
 760         }
 761 }
 762
 763 void cfs_trace_debug_print(void)
 764 {
 765         struct page_collection pc;
 766         struct cfs_trace_page *tage;
 767         struct cfs_trace_page *tmp;
 768
 769         pc.pc_want_daemon_pages = 1;
 770         collect_pages(&pc);
 771         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 772                 char *p, *file, *fn;
 773                 struct page *page;
 774
 775                 __LASSERT_TAGE_INVARIANT(tage);
 776
 777                 page = tage->page;
 778                 p = page_address(page);
 779                 while (p < ((char *)page_address(page) + tage->used)) {
 780                         struct ptldebug_header *hdr;
 781                         int len;
 782                         hdr = (void *)p;
 783                         p += sizeof(*hdr);
 784                         file = p;
 785                         p += strlen(file) + 1;
 786                         fn = p;
 787                         p += strlen(fn) + 1;
 788                         len = hdr->ph_len - (int)(p - (char *)hdr);
 789
 790                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 791
 792                         p += len;
 793                 }
 794
 795                 list_del(&tage->linkage);
 796                 cfs_tage_free(tage);
 797         }
 798 }
 799
 800 int cfs_tracefile_dump_all_pages(char *filename)
 801 {
 802         struct page_collection  pc;
 803         struct file             *filp;
 804         struct cfs_trace_page   *tage;
 805         struct cfs_trace_page   *tmp;
 806         char                    *buf;
 807         int rc;
 808
 809         down_write(&cfs_tracefile_sem);
 810
 811         filp = filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600);
 812         if (IS_ERR(filp)) {
 813                 rc = PTR_ERR(filp);
 814                 filp = NULL;
 815                 pr_err("LustreError: can't open %s for dump: rc = %d\n",
 816                       filename, rc);
 817                 goto out;
 818         }
 819
 820         pc.pc_want_daemon_pages = 1;
 821         collect_pages(&pc);
 822         if (list_empty(&pc.pc_pages)) {
 823                 rc = 0;
 824                 goto close;
 825         }
 826
 827         /* ok, for now, just write the pages.  in the future we'll be building
 828          * iobufs with the pages and calling generic_direct_IO */
 829         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 830
 831                 __LASSERT_TAGE_INVARIANT(tage);
 832
 833                 buf = kmap(tage->page);
 834                 rc = cfs_kernel_write(filp, buf, tage->used, &filp->f_pos);
 835                 kunmap(tage->page);
 836                 if (rc != (int)tage->used) {
 837                         pr_warn("Lustre: wanted to write %u but wrote %d\n",
 838                                 tage->used, rc);
 839                         put_pages_back(&pc);
 840                         __LASSERT(list_empty(&pc.pc_pages));
 841                         break;
 842                 }
 843                 list_del(&tage->linkage);
 844                 cfs_tage_free(tage);
 845         }
 846
 847         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
 848         if (rc)
 849                 pr_err("LustreError: sync returns: rc = %d\n", rc);
 850 close:
 851         filp_close(filp, NULL);
 852 out:
 853         up_write(&cfs_tracefile_sem);
 854         return rc;
 855 }
 856
 857 void cfs_trace_flush_pages(void)
 858 {
 859         struct page_collection pc;
 860         struct cfs_trace_page *tage;
 861         struct cfs_trace_page *tmp;
 862
 863         pc.pc_want_daemon_pages = 1;
 864         collect_pages(&pc);
 865         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 866
 867                 __LASSERT_TAGE_INVARIANT(tage);
 868
 869                 list_del(&tage->linkage);
 870                 cfs_tage_free(tage);
 871         }
 872 }
 873
 874 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 875                             const char __user *usr_buffer, int usr_buffer_nob)
 876 {
 877         int    nob;
 878
 879         if (usr_buffer_nob > knl_buffer_nob)
 880                 return -EOVERFLOW;
 881
 882         if (copy_from_user(knl_buffer, usr_buffer, usr_buffer_nob))
 883                 return -EFAULT;
 884
 885         nob = strnlen(knl_buffer, usr_buffer_nob);
 886         while (--nob >= 0)                      /* strip trailing whitespace */
 887                 if (!isspace(knl_buffer[nob]))
 888                         break;
 889
 890         if (nob < 0)                            /* empty string */
 891                 return -EINVAL;
 892
 893         if (nob == knl_buffer_nob)              /* no space to terminate */
 894                 return -EOVERFLOW;
 895
 896         knl_buffer[nob + 1] = 0;                /* terminate */
 897         return 0;
 898 }
 899 EXPORT_SYMBOL(cfs_trace_copyin_string);
 900
 901 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 902                              const char *knl_buffer, char *append)
 903 {
 904         /* NB if 'append' != NULL, it's a single character to append to the
 905          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 906          * terminating zero byte) for sysctl entries */
 907         int   nob = strlen(knl_buffer);
 908
 909         if (nob > usr_buffer_nob)
 910                 nob = usr_buffer_nob;
 911
 912         if (copy_to_user(usr_buffer, knl_buffer, nob))
 913                 return -EFAULT;
 914
 915         if (append != NULL && nob < usr_buffer_nob) {
 916                 if (copy_to_user(usr_buffer + nob, append, 1))
 917                         return -EFAULT;
 918
 919                 nob++;
 920         }
 921
 922         return nob;
 923 }
 924 EXPORT_SYMBOL(cfs_trace_copyout_string);
 925
 926 int cfs_trace_allocate_string_buffer(char **str, int nob)
 927 {
 928         if (nob > 2 * PAGE_SIZE)        /* string must be "sensible" */
 929                 return -EINVAL;
 930
 931         *str = kmalloc(nob, GFP_KERNEL | __GFP_ZERO);
 932         if (*str == NULL)
 933                 return -ENOMEM;
 934
 935         return 0;
 936 }
 937
 938 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
 939 {
 940         char         *str;
 941         int           rc;
 942
 943         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 944         if (rc != 0)
 945                 return rc;
 946
 947         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 948                                      usr_str, usr_str_nob);
 949         if (rc != 0)
 950                 goto out;
 951
 952         if (str[0] != '/') {
 953                 rc = -EINVAL;
 954                 goto out;
 955         }
 956         rc = cfs_tracefile_dump_all_pages(str);
 957 out:
 958         kfree(str);
 959         return rc;
 960 }
 961
 962 int cfs_trace_daemon_command(char *str)
 963 {
 964         int       rc = 0;
 965
 966         down_write(&cfs_tracefile_sem);
 967
 968         if (strcmp(str, "stop") == 0) {
 969                 up_write(&cfs_tracefile_sem);
 970                 cfs_trace_stop_thread();
 971                 down_write(&cfs_tracefile_sem);
 972                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 973
 974         } else if (strncmp(str, "size=", 5) == 0) {
 975                 unsigned long tmp;
 976
 977                 rc = kstrtoul(str + 5, 10, &tmp);
 978                 if (!rc) {
 979                         if (tmp < 10 || tmp > 20480)
 980                                 cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 981                         else
 982                                 cfs_tracefile_size = tmp << 20;
 983                 }
 984         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 985                 rc = -ENAMETOOLONG;
 986         } else if (str[0] != '/') {
 987                 rc = -EINVAL;
 988         } else {
 989                 strcpy(cfs_tracefile, str);
 990
 991                 pr_info("Lustre: debug daemon will attempt to start writing to %s (%lukB max)\n",
 992                         cfs_tracefile, (long)(cfs_tracefile_size >> 10));
 993
 994                 cfs_trace_start_thread();
 995         }
 996
 997         up_write(&cfs_tracefile_sem);
 998         return rc;
 999 }
1000
1001 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
1002 {
1003         char *str;
1004         int   rc;
1005
1006         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
1007         if (rc != 0)
1008                 return rc;
1009
1010         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
1011                                  usr_str, usr_str_nob);
1012         if (rc == 0)
1013                 rc = cfs_trace_daemon_command(str);
1014
1015         kfree(str);
1016         return rc;
1017 }
1018
1019 int cfs_trace_set_debug_mb(int mb)
1020 {
1021         int i;
1022         int j;
1023         unsigned long pages;
1024         unsigned long total_mb = (cfs_totalram_pages() >> (20 - PAGE_SHIFT));
1025         unsigned long limit = max_t(unsigned long, 512, (total_mb * 4) / 5);
1026         struct cfs_trace_cpu_data *tcd;
1027
1028         if (mb < num_possible_cpus()) {
1029                 pr_warn("Lustre: %d MB is too small for debug buffer size, setting it to %d MB.\n",
1030                         mb, num_possible_cpus());
1031                 mb = num_possible_cpus();
1032         }
1033
1034         if (mb > limit) {
1035                 pr_warn("Lustre: %d MB is too large for debug buffer size, setting it to %lu MB.\n",
1036                         mb, limit);
1037                 mb = limit;
1038         }
1039
1040         mb /= num_possible_cpus();
1041         pages = mb << (20 - PAGE_SHIFT);
1042
1043         down_write(&cfs_tracefile_sem);
1044
1045         cfs_tcd_for_each(tcd, i, j)
1046                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
1047
1048         up_write(&cfs_tracefile_sem);
1049
1050         return mb;
1051 }
1052
1053 int cfs_trace_get_debug_mb(void)
1054 {
1055         int i;
1056         int j;
1057         struct cfs_trace_cpu_data *tcd;
1058         int total_pages = 0;
1059
1060         down_read(&cfs_tracefile_sem);
1061
1062         cfs_tcd_for_each(tcd, i, j)
1063                 total_pages += tcd->tcd_max_pages;
1064
1065         up_read(&cfs_tracefile_sem);
1066
1067         return (total_pages >> (20 - PAGE_SHIFT)) + 1;
1068 }
1069
1070 static int tracefiled(void *arg)
1071 {
1072         struct page_collection pc;
1073         struct tracefiled_ctl *tctl = arg;
1074         struct cfs_trace_page *tage;
1075         struct cfs_trace_page *tmp;
1076         struct file *filp;
1077         char *buf;
1078         int last_loop = 0;
1079         int rc;
1080
1081         /* we're started late enough that we pick up init's fs context */
1082         /* this is so broken in uml?  what on earth is going on? */
1083
1084         complete(&tctl->tctl_start);
1085
1086         pc.pc_want_daemon_pages = 0;
1087
1088         while (!last_loop) {
1089                 wait_event_timeout(tctl->tctl_waitq,
1090                                    ({ collect_pages(&pc);
1091                                      !list_empty(&pc.pc_pages); }) ||
1092                                    atomic_read(&tctl->tctl_shutdown),
1093                                    cfs_time_seconds(1));
1094                 if (atomic_read(&tctl->tctl_shutdown))
1095                         last_loop = 1;
1096                 if (list_empty(&pc.pc_pages))
1097                         continue;
1098
1099                 filp = NULL;
1100                 down_read(&cfs_tracefile_sem);
1101                 if (cfs_tracefile[0] != 0) {
1102                         filp = filp_open(cfs_tracefile,
1103                                          O_CREAT | O_RDWR | O_LARGEFILE,
1104                                          0600);
1105                         if (IS_ERR(filp)) {
1106                                 rc = PTR_ERR(filp);
1107                                 filp = NULL;
1108                                 pr_warn("Lustre: couldn't open %s: rc = %d\n",
1109                                         cfs_tracefile, rc);
1110                         }
1111                 }
1112                 up_read(&cfs_tracefile_sem);
1113                 if (filp == NULL) {
1114                         put_pages_on_daemon_list(&pc);
1115                         __LASSERT(list_empty(&pc.pc_pages));
1116                         continue;
1117                 }
1118
1119                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
1120                         struct dentry *de = file_dentry(filp);
1121                         static loff_t f_pos;
1122
1123                         __LASSERT_TAGE_INVARIANT(tage);
1124
1125                         if (f_pos >= (off_t)cfs_tracefile_size)
1126                                 f_pos = 0;
1127                         else if (f_pos > i_size_read(de->d_inode))
1128                                 f_pos = i_size_read(de->d_inode);
1129
1130                         buf = kmap(tage->page);
1131                         rc = cfs_kernel_write(filp, buf, tage->used, &f_pos);
1132                         kunmap(tage->page);
1133                         if (rc != (int)tage->used) {
1134                                 pr_warn("Lustre: wanted to write %u but wrote %d\n",
1135                                         tage->used, rc);
1136                                 put_pages_back(&pc);
1137                                 __LASSERT(list_empty(&pc.pc_pages));
1138                                 break;
1139                         }
1140                 }
1141
1142                 filp_close(filp, NULL);
1143                 put_pages_on_daemon_list(&pc);
1144                 if (!list_empty(&pc.pc_pages)) {
1145                         int i;
1146
1147                         pr_alert("Lustre: trace pages aren't empty\n");
1148                         pr_err("Lustre: total cpus(%d): ", num_possible_cpus());
1149                         for (i = 0; i < num_possible_cpus(); i++)
1150                                 if (cpu_online(i))
1151                                         pr_cont("%d(on) ", i);
1152                                 else
1153                                         pr_cont("%d(off) ", i);
1154                         pr_cont("\n");
1155
1156                         i = 0;
1157                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1158                                                  linkage)
1159                                 pr_err("Lustre: page %d belongs to cpu %d\n",
1160                                        ++i, tage->cpu);
1161                         pr_err("Lustre: There are %d pages unwritten\n", i);
1162                 }
1163                 __LASSERT(list_empty(&pc.pc_pages));
1164         }
1165         complete(&tctl->tctl_stop);
1166         return 0;
1167 }
1168
1169 int cfs_trace_start_thread(void)
1170 {
1171         struct tracefiled_ctl *tctl = &trace_tctl;
1172         int rc = 0;
1173
1174         mutex_lock(&cfs_trace_thread_mutex);
1175         if (thread_running)
1176                 goto out;
1177
1178         init_completion(&tctl->tctl_start);
1179         init_completion(&tctl->tctl_stop);
1180         init_waitqueue_head(&tctl->tctl_waitq);
1181         atomic_set(&tctl->tctl_shutdown, 0);
1182
1183         if (IS_ERR(kthread_run(tracefiled, tctl, "ktracefiled"))) {
1184                 rc = -ECHILD;
1185                 goto out;
1186         }
1187
1188         wait_for_completion(&tctl->tctl_start);
1189         thread_running = 1;
1190 out:
1191         mutex_unlock(&cfs_trace_thread_mutex);
1192         return rc;
1193 }
1194
1195 void cfs_trace_stop_thread(void)
1196 {
1197         struct tracefiled_ctl *tctl = &trace_tctl;
1198
1199         mutex_lock(&cfs_trace_thread_mutex);
1200         if (thread_running) {
1201                 pr_info("Lustre: shutting down debug daemon thread...\n");
1202                 atomic_set(&tctl->tctl_shutdown, 1);
1203                 wait_for_completion(&tctl->tctl_stop);
1204                 thread_running = 0;
1205         }
1206         mutex_unlock(&cfs_trace_thread_mutex);
1207 }
1208
1209 /* percents to share the total debug memory for each type */
1210 static unsigned int pages_factor[CFS_TCD_TYPE_CNT] = {
1211         80, /* 80% pages for CFS_TCD_TYPE_PROC */
1212         10, /* 10% pages for CFS_TCD_TYPE_SOFTIRQ */
1213         10  /* 10% pages for CFS_TCD_TYPE_IRQ */
1214 };
1215
1216 int cfs_tracefile_init(int max_pages)
1217 {
1218         struct cfs_trace_cpu_data *tcd;
1219         int i;
1220         int j;
1221
1222         /* initialize trace_data */
1223         memset(cfs_trace_data, 0, sizeof(cfs_trace_data));
1224         for (i = 0; i < CFS_TCD_TYPE_CNT; i++) {
1225                 cfs_trace_data[i] =
1226                         kmalloc_array(num_possible_cpus(),
1227                                       sizeof(union cfs_trace_data_union),
1228                                       GFP_KERNEL);
1229                 if (!cfs_trace_data[i])
1230                         goto out_trace_data;
1231         }
1232
1233         /* arch related info initialized */
1234         cfs_tcd_for_each(tcd, i, j) {
1235                 int factor = pages_factor[i];
1236
1237                 spin_lock_init(&tcd->tcd_lock);
1238                 tcd->tcd_pages_factor = factor;
1239                 tcd->tcd_type = i;
1240                 tcd->tcd_cpu = j;
1241
1242                 INIT_LIST_HEAD(&tcd->tcd_pages);
1243                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1244                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1245                 tcd->tcd_cur_pages = 0;
1246                 tcd->tcd_cur_stock_pages = 0;
1247                 tcd->tcd_cur_daemon_pages = 0;
1248                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1249                 LASSERT(tcd->tcd_max_pages > 0);
1250                 tcd->tcd_shutting_down = 0;
1251         }
1252
1253         for (i = 0; i < num_possible_cpus(); i++)
1254                 for (j = 0; j < 3; j++) {
1255                         cfs_trace_console_buffers[i][j] =
1256                                 kmalloc(CFS_TRACE_CONSOLE_BUFFER_SIZE,
1257                                         GFP_KERNEL);
1258                         if (!cfs_trace_console_buffers[i][j])
1259                                 goto out_buffers;
1260                 }
1261
1262         return 0;
1263
1264 out_buffers:
1265         for (i = 0; i < num_possible_cpus(); i++)
1266                 for (j = 0; j < 3; j++) {
1267                         kfree(cfs_trace_console_buffers[i][j]);
1268                         cfs_trace_console_buffers[i][j] = NULL;
1269                 }
1270 out_trace_data:
1271         for (i = 0; cfs_trace_data[i]; i++) {
1272                 kfree(cfs_trace_data[i]);
1273                 cfs_trace_data[i] = NULL;
1274         }
1275         pr_err("lnet: Not enough memory\n");
1276         return -ENOMEM;
1277 }
1278
1279 static void trace_cleanup_on_all_cpus(void)
1280 {
1281         struct cfs_trace_cpu_data *tcd;
1282         struct cfs_trace_page *tage;
1283         struct cfs_trace_page *tmp;
1284         int i, cpu;
1285
1286         for_each_possible_cpu(cpu) {
1287                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1288                         if (!tcd->tcd_pages_factor)
1289                                 /* Not initialised */
1290                                 continue;
1291                         tcd->tcd_shutting_down = 1;
1292
1293                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1294                                 __LASSERT_TAGE_INVARIANT(tage);
1295
1296                                 list_del(&tage->linkage);
1297                                 cfs_tage_free(tage);
1298                         }
1299                         tcd->tcd_cur_pages = 0;
1300                 }
1301         }
1302 }
1303
1304 static void cfs_trace_cleanup(void)
1305 {
1306         struct page_collection pc;
1307         int i;
1308         int j;
1309
1310         INIT_LIST_HEAD(&pc.pc_pages);
1311
1312         trace_cleanup_on_all_cpus();
1313
1314         for (i = 0; i < num_possible_cpus(); i++)
1315                 for (j = 0; j < 3; j++) {
1316                         kfree(cfs_trace_console_buffers[i][j]);
1317                         cfs_trace_console_buffers[i][j] = NULL;
1318                 }
1319
1320         for (i = 0; i < CFS_TCD_TYPE_CNT && cfs_trace_data[i]; i++) {
1321                 kfree(cfs_trace_data[i]);
1322                 cfs_trace_data[i] = NULL;
1323         }
1324 }
1325
1326 void cfs_tracefile_exit(void)
1327 {
1328         cfs_trace_stop_thread();
1329         cfs_trace_cleanup();
1330 }