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 #define TCD_MAX_TYPES                   8
  53
  54 enum cfs_trace_buf_type {
  55         CFS_TCD_TYPE_PROC = 0,
  56         CFS_TCD_TYPE_SOFTIRQ,
  57         CFS_TCD_TYPE_IRQ,
  58         CFS_TCD_TYPE_MAX
  59 };
  60
  61 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  62
  63 char *cfs_trace_console_buffers[NR_CPUS][CFS_TCD_TYPE_MAX];
  64 char cfs_tracefile[TRACEFILE_NAME_SIZE];
  65 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
  66 static struct tracefiled_ctl trace_tctl;
  67 static DEFINE_MUTEX(cfs_trace_thread_mutex);
  68 static int thread_running = 0;
  69
  70 static atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
  71 static DECLARE_RWSEM(cfs_tracefile_sem);
  72
  73 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  74                                         struct cfs_trace_cpu_data *tcd);
  75
  76 /* trace file lock routines */
  77 /* The walking argument indicates the locking comes from all tcd types
  78  * iterator and we must lock it and dissable local irqs to avoid deadlocks
  79  * with other interrupt locks that might be happening. See LU-1311
  80  * for details.
  81  */
  82 int cfs_trace_lock_tcd(struct cfs_trace_cpu_data *tcd, int walking)
  83         __acquires(&tcd->tcd_lock)
  84 {
  85         __LASSERT(tcd->tcd_type < CFS_TCD_TYPE_MAX);
  86         if (tcd->tcd_type == CFS_TCD_TYPE_IRQ)
  87                 spin_lock_irqsave(&tcd->tcd_lock, tcd->tcd_lock_flags);
  88         else if (tcd->tcd_type == CFS_TCD_TYPE_SOFTIRQ)
  89                 spin_lock_bh(&tcd->tcd_lock);
  90         else if (unlikely(walking))
  91                 spin_lock_irq(&tcd->tcd_lock);
  92         else
  93                 spin_lock(&tcd->tcd_lock);
  94         return 1;
  95 }
  96
  97 void cfs_trace_unlock_tcd(struct cfs_trace_cpu_data *tcd, int walking)
  98         __releases(&tcd->tcd_lock)
  99 {
 100         __LASSERT(tcd->tcd_type < CFS_TCD_TYPE_MAX);
 101         if (tcd->tcd_type == CFS_TCD_TYPE_IRQ)
 102                 spin_unlock_irqrestore(&tcd->tcd_lock, tcd->tcd_lock_flags);
 103         else if (tcd->tcd_type == CFS_TCD_TYPE_SOFTIRQ)
 104                 spin_unlock_bh(&tcd->tcd_lock);
 105         else if (unlikely(walking))
 106                 spin_unlock_irq(&tcd->tcd_lock);
 107         else
 108                 spin_unlock(&tcd->tcd_lock);
 109 }
 110
 111 #define cfs_tcd_for_each(tcd, i, j)                                     \
 112         for (i = 0; cfs_trace_data[i]; i++)                             \
 113                 for (j = 0, ((tcd) = &(*cfs_trace_data[i])[j].tcd);     \
 114                      j < num_possible_cpus();                           \
 115                      j++, (tcd) = &(*cfs_trace_data[i])[j].tcd)
 116
 117 #define cfs_tcd_for_each_type_lock(tcd, i, cpu)                         \
 118         for (i = 0; cfs_trace_data[i] &&                                \
 119              (tcd = &(*cfs_trace_data[i])[cpu].tcd) &&                  \
 120              cfs_trace_lock_tcd(tcd, 1); cfs_trace_unlock_tcd(tcd, 1), i++)
 121
 122 enum cfs_trace_buf_type cfs_trace_buf_idx_get(void)
 123 {
 124         if (in_irq())
 125                 return CFS_TCD_TYPE_IRQ;
 126         if (in_softirq())
 127                 return CFS_TCD_TYPE_SOFTIRQ;
 128         return CFS_TCD_TYPE_PROC;
 129 }
 130
 131 static inline char *cfs_trace_get_console_buffer(void)
 132 {
 133         unsigned int i = get_cpu();
 134         unsigned int j = cfs_trace_buf_idx_get();
 135
 136         return cfs_trace_console_buffers[i][j];
 137 }
 138
 139 static inline struct cfs_trace_cpu_data *
 140 cfs_trace_get_tcd(void)
 141 {
 142         struct cfs_trace_cpu_data *tcd =
 143                 &(*cfs_trace_data[cfs_trace_buf_idx_get()])[get_cpu()].tcd;
 144
 145         cfs_trace_lock_tcd(tcd, 0);
 146
 147         return tcd;
 148 }
 149
 150 static inline void cfs_trace_put_tcd(struct cfs_trace_cpu_data *tcd)
 151 {
 152         cfs_trace_unlock_tcd(tcd, 0);
 153
 154         put_cpu();
 155 }
 156
 157 static inline struct cfs_trace_page *
 158 cfs_tage_from_list(struct list_head *list)
 159 {
 160         return list_entry(list, struct cfs_trace_page, linkage);
 161 }
 162
 163 static struct cfs_trace_page *cfs_tage_alloc(gfp_t gfp)
 164 {
 165         struct page            *page;
 166         struct cfs_trace_page *tage;
 167
 168         /* My caller is trying to free memory */
 169         if (!in_interrupt() && (current->flags & PF_MEMALLOC))
 170                 return NULL;
 171
 172         /*
 173          * Don't spam console with allocation failures: they will be reported
 174          * by upper layer anyway.
 175          */
 176         gfp |= __GFP_NOWARN;
 177         page = alloc_page(gfp);
 178         if (page == NULL)
 179                 return NULL;
 180
 181         tage = kmalloc(sizeof(*tage), gfp);
 182         if (tage == NULL) {
 183                 __free_page(page);
 184                 return NULL;
 185         }
 186
 187         tage->page = page;
 188         atomic_inc(&cfs_tage_allocated);
 189         return tage;
 190 }
 191
 192 static void cfs_tage_free(struct cfs_trace_page *tage)
 193 {
 194         __LASSERT(tage != NULL);
 195         __LASSERT(tage->page != NULL);
 196
 197         __free_page(tage->page);
 198         kfree(tage);
 199         atomic_dec(&cfs_tage_allocated);
 200 }
 201
 202 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
 203                              struct list_head *queue)
 204 {
 205         __LASSERT(tage != NULL);
 206         __LASSERT(queue != NULL);
 207
 208         list_move_tail(&tage->linkage, queue);
 209 }
 210
 211 /* return a page that has 'len' bytes left at the end */
 212 static struct cfs_trace_page *
 213 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
 214 {
 215         struct cfs_trace_page *tage;
 216
 217         if (tcd->tcd_cur_pages > 0) {
 218                 __LASSERT(!list_empty(&tcd->tcd_pages));
 219                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
 220                 if (tage->used + len <= PAGE_SIZE)
 221                         return tage;
 222         }
 223
 224         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 225                 if (tcd->tcd_cur_stock_pages > 0) {
 226                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
 227                         --tcd->tcd_cur_stock_pages;
 228                         list_del_init(&tage->linkage);
 229                 } else {
 230                         tage = cfs_tage_alloc(GFP_ATOMIC);
 231                         if (unlikely(tage == NULL)) {
 232                                 if ((!(current->flags & PF_MEMALLOC) ||
 233                                      in_interrupt()) && printk_ratelimit())
 234                                         pr_warn("Lustre: cannot allocate a tage (%ld)\n",
 235                                                 tcd->tcd_cur_pages);
 236                                 return NULL;
 237                         }
 238                 }
 239
 240                 tage->used = 0;
 241                 tage->cpu = smp_processor_id();
 242                 tage->type = tcd->tcd_type;
 243                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 244                 tcd->tcd_cur_pages++;
 245
 246                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 247                         struct tracefiled_ctl *tctl = &trace_tctl;
 248                         /*
 249                          * wake up tracefiled to process some pages.
 250                          */
 251                         wake_up(&tctl->tctl_waitq);
 252                 }
 253                 return tage;
 254         }
 255         return NULL;
 256 }
 257
 258 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
 259 {
 260         int pgcount = tcd->tcd_cur_pages / 10;
 261         struct page_collection pc;
 262         struct cfs_trace_page *tage;
 263         struct cfs_trace_page *tmp;
 264
 265         /*
 266          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 267          * from here: this will lead to infinite recursion.
 268          */
 269
 270         if (printk_ratelimit())
 271                 pr_warn("Lustre: debug daemon buffer overflowed; discarding 10%% of pages (%d of %ld)\n",
 272                         pgcount + 1, tcd->tcd_cur_pages);
 273
 274         INIT_LIST_HEAD(&pc.pc_pages);
 275
 276         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 277                 if (pgcount-- == 0)
 278                         break;
 279
 280                 list_move_tail(&tage->linkage, &pc.pc_pages);
 281                 tcd->tcd_cur_pages--;
 282         }
 283         put_pages_on_tcd_daemon_list(&pc, tcd);
 284 }
 285
 286 /* return a page that has 'len' bytes left at the end */
 287 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 288                                                  unsigned long len)
 289 {
 290         struct cfs_trace_page *tage;
 291
 292         /*
 293          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 294          * from here: this will lead to infinite recursion.
 295          */
 296
 297         if (len > PAGE_SIZE) {
 298                 pr_err("LustreError: cowardly refusing to write %lu bytes in a page\n",
 299                        len);
 300                 return NULL;
 301         }
 302
 303         tage = cfs_trace_get_tage_try(tcd, len);
 304         if (tage != NULL)
 305                 return tage;
 306         if (thread_running)
 307                 cfs_tcd_shrink(tcd);
 308         if (tcd->tcd_cur_pages > 0) {
 309                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 310                 tage->used = 0;
 311                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 312         }
 313         return tage;
 314 }
 315
 316 static void cfs_set_ptldebug_header(struct ptldebug_header *header,
 317                                     struct libcfs_debug_msg_data *msgdata,
 318                                     unsigned long stack)
 319 {
 320         struct timespec64 ts;
 321
 322         ktime_get_real_ts64(&ts);
 323
 324         header->ph_subsys = msgdata->msg_subsys;
 325         header->ph_mask = msgdata->msg_mask;
 326         header->ph_cpu_id = smp_processor_id();
 327         header->ph_type = cfs_trace_buf_idx_get();
 328         /* y2038 safe since all user space treats this as unsigned, but
 329          * will overflow in 2106
 330          */
 331         header->ph_sec = (u32)ts.tv_sec;
 332         header->ph_usec = ts.tv_nsec / NSEC_PER_USEC;
 333         header->ph_stack = stack;
 334         header->ph_pid = current->pid;
 335         header->ph_line_num = msgdata->msg_line;
 336         header->ph_extern_pid = 0;
 337 }
 338
 339 /**
 340  * tty_write_msg - write a message to a certain tty, not just the console.
 341  * @tty: the destination tty_struct
 342  * @msg: the message to write
 343  *
 344  * tty_write_message is not exported, so write a same function for it
 345  *
 346  */
 347 static void tty_write_msg(struct tty_struct *tty, const char *msg)
 348 {
 349         mutex_lock(&tty->atomic_write_lock);
 350         tty_lock(tty);
 351         if (tty->ops->write && tty->count > 0)
 352                 tty->ops->write(tty, msg, strlen(msg));
 353         tty_unlock(tty);
 354         mutex_unlock(&tty->atomic_write_lock);
 355         wake_up_interruptible_poll(&tty->write_wait, POLLOUT);
 356 }
 357
 358 static void cfs_tty_write_message(const char *prefix, int mask, const char *msg)
 359 {
 360         struct tty_struct *tty;
 361
 362         tty = get_current_tty();
 363         if (!tty)
 364                 return;
 365
 366         tty_write_msg(tty, prefix);
 367         if ((mask & D_EMERG) || (mask & D_ERROR))
 368                 tty_write_msg(tty, "Error");
 369         tty_write_msg(tty, ": ");
 370         tty_write_msg(tty, msg);
 371         tty_kref_put(tty);
 372 }
 373
 374 static void cfs_print_to_console(struct ptldebug_header *hdr, int mask,
 375                                  const char *buf, int len, const char *file,
 376                                  const char *fn)
 377 {
 378         char *prefix = "Lustre";
 379
 380         if (hdr->ph_subsys == S_LND || hdr->ph_subsys == S_LNET)
 381                 prefix = "LNet";
 382
 383         if (mask & D_CONSOLE) {
 384                 if (mask & D_EMERG)
 385                         pr_emerg("%sError: %.*s", prefix, len, buf);
 386                 else if (mask & D_ERROR)
 387                         pr_err("%sError: %.*s", prefix, len, buf);
 388                 else if (mask & D_WARNING)
 389                         pr_warn("%s: %.*s", prefix, len, buf);
 390                 else if (mask & libcfs_printk)
 391                         pr_info("%s: %.*s", prefix, len, buf);
 392         } else {
 393                 if (mask & D_EMERG)
 394                         pr_emerg("%sError: %d:%d:(%s:%d:%s()) %.*s", prefix,
 395                                  hdr->ph_pid, hdr->ph_extern_pid, file,
 396                                  hdr->ph_line_num, fn, len, buf);
 397                 else if (mask & D_ERROR)
 398                         pr_err("%sError: %d:%d:(%s:%d:%s()) %.*s", prefix,
 399                                hdr->ph_pid, hdr->ph_extern_pid, file,
 400                                hdr->ph_line_num, fn, len, buf);
 401                 else if (mask & D_WARNING)
 402                         pr_warn("%s: %d:%d:(%s:%d:%s()) %.*s", prefix,
 403                                 hdr->ph_pid, hdr->ph_extern_pid, file,
 404                                 hdr->ph_line_num, fn, len, buf);
 405                 else if (mask & (D_CONSOLE | libcfs_printk))
 406                         pr_info("%s: %.*s", prefix, len, buf);
 407         }
 408
 409         if (mask & D_TTY)
 410                 cfs_tty_write_message(prefix, mask, buf);
 411 }
 412
 413 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 414                      const char *format, ...)
 415 {
 416         struct cfs_trace_cpu_data *tcd = NULL;
 417         struct ptldebug_header header = {0};
 418         struct cfs_trace_page *tage;
 419         /* string_buf is used only if tcd != NULL, and is always set then */
 420         char *string_buf = NULL;
 421         char *debug_buf;
 422         int known_size;
 423         int needed = 85; /* seeded with average message length */
 424         int max_nob;
 425         va_list ap;
 426         int retry;
 427         int mask = msgdata->msg_mask;
 428         char *file = (char *)msgdata->msg_file;
 429         struct cfs_debug_limit_state *cdls = msgdata->msg_cdls;
 430
 431         if (strchr(file, '/'))
 432                 file = strrchr(file, '/') + 1;
 433
 434         tcd = cfs_trace_get_tcd();
 435
 436         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 437          * pins us to a particular CPU.  This avoids an smp_processor_id()
 438          * warning on Linux when debugging is enabled.
 439          */
 440         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 441
 442         if (!tcd)                /* arch may not log in IRQ context */
 443                 goto console;
 444
 445         if (tcd->tcd_cur_pages == 0)
 446                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 447
 448         if (tcd->tcd_shutting_down) {
 449                 cfs_trace_put_tcd(tcd);
 450                 tcd = NULL;
 451                 goto console;
 452         }
 453
 454         known_size = strlen(file) + 1;
 455         if (msgdata->msg_fn)
 456                 known_size += strlen(msgdata->msg_fn) + 1;
 457
 458         if (libcfs_debug_binary)
 459                 known_size += sizeof(header);
 460
 461         /*
 462          * May perform an additional pass to update 'needed' and increase
 463          * tage buffer size to match vsnprintf reported size required
 464          * On the second pass (retry=1) use vscnprintf [which returns
 465          * number of bytes written not including the terminating nul]
 466          * to clarify `needed` is used as number of bytes written
 467          * for the remainder of this function
 468          */
 469         for (retry = 0; retry < 2; retry++) {
 470                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 471                 if (!tage) {
 472                         if (needed + known_size > PAGE_SIZE)
 473                                 mask |= D_ERROR;
 474
 475                         cfs_trace_put_tcd(tcd);
 476                         tcd = NULL;
 477                         goto console;
 478                 }
 479
 480                 string_buf = (char *)page_address(tage->page) +
 481                              tage->used + known_size;
 482
 483                 max_nob = PAGE_SIZE - tage->used - known_size;
 484                 if (max_nob <= 0) {
 485                         pr_emerg("LustreError: negative max_nob: %d\n",
 486                                  max_nob);
 487                         mask |= D_ERROR;
 488                         cfs_trace_put_tcd(tcd);
 489                         tcd = NULL;
 490                         goto console;
 491                 }
 492
 493                 va_start(ap, format);
 494                 if (retry)
 495                         needed = vscnprintf(string_buf, max_nob, format, ap);
 496                 else
 497                         needed = vsnprintf(string_buf, max_nob, format, ap);
 498                 va_end(ap);
 499
 500                 if (needed < max_nob) /* well. printing ok.. */
 501                         break;
 502         }
 503
 504         /* `needed` is actual bytes written to string_buf */
 505         if (*(string_buf + needed - 1) != '\n') {
 506                 pr_info("Lustre: format at %s:%d:%s doesn't end in newline\n",
 507                         file, msgdata->msg_line, msgdata->msg_fn);
 508         } else if (mask & D_TTY) {
 509                 /* TTY needs '\r\n' to move carriage to leftmost position */
 510                 if (needed < 2 || *(string_buf + needed - 2) != '\r')
 511                         pr_info("Lustre: format at %s:%d:%s doesn't end in '\\r\\n'\n",
 512                                 file, msgdata->msg_line, msgdata->msg_fn);
 513         }
 514
 515         header.ph_len = known_size + needed;
 516         debug_buf = (char *)page_address(tage->page) + tage->used;
 517
 518         if (libcfs_debug_binary) {
 519                 memcpy(debug_buf, &header, sizeof(header));
 520                 tage->used += sizeof(header);
 521                 debug_buf += sizeof(header);
 522         }
 523
 524         strlcpy(debug_buf, file, PAGE_SIZE - tage->used);
 525         tage->used += strlen(file) + 1;
 526         debug_buf += strlen(file) + 1;
 527
 528         if (msgdata->msg_fn) {
 529                 strlcpy(debug_buf, msgdata->msg_fn, PAGE_SIZE - tage->used);
 530                 tage->used += strlen(msgdata->msg_fn) + 1;
 531                 debug_buf += strlen(msgdata->msg_fn) + 1;
 532         }
 533
 534         __LASSERT(debug_buf == string_buf);
 535
 536         tage->used += needed;
 537         __LASSERT(tage->used <= PAGE_SIZE);
 538
 539 console:
 540         if ((mask & libcfs_printk) == 0) {
 541                 /* no console output requested */
 542                 if (tcd != NULL)
 543                         cfs_trace_put_tcd(tcd);
 544                 return 1;
 545         }
 546
 547         if (cdls != NULL) {
 548                 if (libcfs_console_ratelimit &&
 549                     cdls->cdls_next != 0 &&     /* not first time ever */
 550                     time_before(jiffies, cdls->cdls_next)) {
 551                         /* skipping a console message */
 552                         cdls->cdls_count++;
 553                         if (tcd != NULL)
 554                                 cfs_trace_put_tcd(tcd);
 555                         return 1;
 556                 }
 557
 558                 if (time_after(jiffies, cdls->cdls_next +
 559                                         libcfs_console_max_delay +
 560                                         cfs_time_seconds(10))) {
 561                         /* last timeout was a long time ago */
 562                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 563                 } else {
 564                         cdls->cdls_delay *= libcfs_console_backoff;
 565                 }
 566
 567                 if (cdls->cdls_delay < libcfs_console_min_delay)
 568                         cdls->cdls_delay = libcfs_console_min_delay;
 569                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 570                         cdls->cdls_delay = libcfs_console_max_delay;
 571
 572                 /* ensure cdls_next is never zero after it's been seen */
 573                 cdls->cdls_next = (jiffies + cdls->cdls_delay) | 1;
 574         }
 575
 576         if (tcd) {
 577                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 578                                      msgdata->msg_fn);
 579                 cfs_trace_put_tcd(tcd);
 580         } else {
 581                 string_buf = cfs_trace_get_console_buffer();
 582
 583                 va_start(ap, format);
 584                 needed = vscnprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 585                                     format, ap);
 586                 va_end(ap);
 587
 588                 cfs_print_to_console(&header, mask,
 589                                      string_buf, needed, file, msgdata->msg_fn);
 590
 591                 put_cpu();
 592         }
 593
 594         if (cdls != NULL && cdls->cdls_count != 0) {
 595                 string_buf = cfs_trace_get_console_buffer();
 596
 597                 needed = scnprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 598                                    "Skipped %d previous similar message%s\n",
 599                                    cdls->cdls_count,
 600                                    (cdls->cdls_count > 1) ? "s" : "");
 601
 602                 /* Do not allow print this to TTY */
 603                 cfs_print_to_console(&header, mask & ~D_TTY, string_buf,
 604                                      needed, file, msgdata->msg_fn);
 605
 606                 put_cpu();
 607                 cdls->cdls_count = 0;
 608         }
 609
 610         return 0;
 611 }
 612 EXPORT_SYMBOL(libcfs_debug_msg);
 613
 614 void
 615 cfs_trace_assertion_failed(const char *str,
 616                            struct libcfs_debug_msg_data *msgdata)
 617 {
 618         struct ptldebug_header hdr;
 619
 620         libcfs_panic_in_progress = 1;
 621         libcfs_catastrophe = 1;
 622         smp_mb();
 623
 624         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 625
 626         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 627                              msgdata->msg_file, msgdata->msg_fn);
 628
 629         panic("Lustre debug assertion failure\n");
 630
 631         /* not reached */
 632 }
 633
 634 static void
 635 panic_collect_pages(struct page_collection *pc)
 636 {
 637         /* Do the collect_pages job on a single CPU: assumes that all other
 638          * CPUs have been stopped during a panic.  If this isn't true for some
 639          * arch, this will have to be implemented separately in each arch.  */
 640         int                        i;
 641         int                        j;
 642         struct cfs_trace_cpu_data *tcd;
 643
 644         INIT_LIST_HEAD(&pc->pc_pages);
 645
 646         cfs_tcd_for_each(tcd, i, j) {
 647                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 648                 tcd->tcd_cur_pages = 0;
 649
 650                 if (pc->pc_want_daemon_pages) {
 651                         list_splice_init(&tcd->tcd_daemon_pages,
 652                                                 &pc->pc_pages);
 653                         tcd->tcd_cur_daemon_pages = 0;
 654                 }
 655         }
 656 }
 657
 658 static void collect_pages_on_all_cpus(struct page_collection *pc)
 659 {
 660         struct cfs_trace_cpu_data *tcd;
 661         int i, cpu;
 662
 663         for_each_possible_cpu(cpu) {
 664                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 665                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 666                         tcd->tcd_cur_pages = 0;
 667                         if (pc->pc_want_daemon_pages) {
 668                                 list_splice_init(&tcd->tcd_daemon_pages,
 669                                                         &pc->pc_pages);
 670                                 tcd->tcd_cur_daemon_pages = 0;
 671                         }
 672                 }
 673         }
 674 }
 675
 676 static void collect_pages(struct page_collection *pc)
 677 {
 678         INIT_LIST_HEAD(&pc->pc_pages);
 679
 680         if (libcfs_panic_in_progress)
 681                 panic_collect_pages(pc);
 682         else
 683                 collect_pages_on_all_cpus(pc);
 684 }
 685
 686 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 687 {
 688         struct cfs_trace_cpu_data *tcd;
 689         struct list_head *cur_head;
 690         struct cfs_trace_page *tage;
 691         struct cfs_trace_page *tmp;
 692         int i, cpu;
 693
 694         for_each_possible_cpu(cpu) {
 695                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 696                         cur_head = tcd->tcd_pages.next;
 697
 698                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 699                                                  linkage) {
 700
 701                                 __LASSERT_TAGE_INVARIANT(tage);
 702
 703                                 if (tage->cpu != cpu || tage->type != i)
 704                                         continue;
 705
 706                                 cfs_tage_to_tail(tage, cur_head);
 707                                 tcd->tcd_cur_pages++;
 708                         }
 709                 }
 710         }
 711 }
 712
 713 static void put_pages_back(struct page_collection *pc)
 714 {
 715         if (!libcfs_panic_in_progress)
 716                 put_pages_back_on_all_cpus(pc);
 717 }
 718
 719 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 720  * we have a good amount of data at all times for dumping during an LBUG, even
 721  * if we have been steadily writing (and otherwise discarding) pages via the
 722  * debug daemon. */
 723 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 724                                          struct cfs_trace_cpu_data *tcd)
 725 {
 726         struct cfs_trace_page *tage;
 727         struct cfs_trace_page *tmp;
 728
 729         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 730                 __LASSERT_TAGE_INVARIANT(tage);
 731
 732                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 733                         continue;
 734
 735                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 736                 tcd->tcd_cur_daemon_pages++;
 737
 738                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 739                         struct cfs_trace_page *victim;
 740
 741                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 742                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 743
 744                         __LASSERT_TAGE_INVARIANT(victim);
 745
 746                         list_del(&victim->linkage);
 747                         cfs_tage_free(victim);
 748                         tcd->tcd_cur_daemon_pages--;
 749                 }
 750         }
 751 }
 752
 753 static void put_pages_on_daemon_list(struct page_collection *pc)
 754 {
 755         struct cfs_trace_cpu_data *tcd;
 756         int i, cpu;
 757
 758         for_each_possible_cpu(cpu) {
 759                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 760                         put_pages_on_tcd_daemon_list(pc, tcd);
 761         }
 762 }
 763
 764 void cfs_trace_debug_print(void)
 765 {
 766         struct page_collection pc;
 767         struct cfs_trace_page *tage;
 768         struct cfs_trace_page *tmp;
 769
 770         pc.pc_want_daemon_pages = 1;
 771         collect_pages(&pc);
 772         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 773                 char *p, *file, *fn;
 774                 struct page *page;
 775
 776                 __LASSERT_TAGE_INVARIANT(tage);
 777
 778                 page = tage->page;
 779                 p = page_address(page);
 780                 while (p < ((char *)page_address(page) + tage->used)) {
 781                         struct ptldebug_header *hdr;
 782                         int len;
 783                         hdr = (void *)p;
 784                         p += sizeof(*hdr);
 785                         file = p;
 786                         p += strlen(file) + 1;
 787                         fn = p;
 788                         p += strlen(fn) + 1;
 789                         len = hdr->ph_len - (int)(p - (char *)hdr);
 790
 791                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 792
 793                         p += len;
 794                 }
 795
 796                 list_del(&tage->linkage);
 797                 cfs_tage_free(tage);
 798         }
 799 }
 800
 801 int cfs_tracefile_dump_all_pages(char *filename)
 802 {
 803         struct page_collection  pc;
 804         struct file             *filp;
 805         struct cfs_trace_page   *tage;
 806         struct cfs_trace_page   *tmp;
 807         char                    *buf;
 808         int rc;
 809
 810         down_write(&cfs_tracefile_sem);
 811
 812         filp = filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600);
 813         if (IS_ERR(filp)) {
 814                 rc = PTR_ERR(filp);
 815                 filp = NULL;
 816                 pr_err("LustreError: can't open %s for dump: rc = %d\n",
 817                       filename, rc);
 818                 goto out;
 819         }
 820
 821         pc.pc_want_daemon_pages = 1;
 822         collect_pages(&pc);
 823         if (list_empty(&pc.pc_pages)) {
 824                 rc = 0;
 825                 goto close;
 826         }
 827
 828         /* ok, for now, just write the pages.  in the future we'll be building
 829          * iobufs with the pages and calling generic_direct_IO */
 830         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 831
 832                 __LASSERT_TAGE_INVARIANT(tage);
 833
 834                 buf = kmap(tage->page);
 835                 rc = cfs_kernel_write(filp, buf, tage->used, &filp->f_pos);
 836                 kunmap(tage->page);
 837                 if (rc != (int)tage->used) {
 838                         pr_warn("Lustre: wanted to write %u but wrote %d\n",
 839                                 tage->used, rc);
 840                         put_pages_back(&pc);
 841                         __LASSERT(list_empty(&pc.pc_pages));
 842                         break;
 843                 }
 844                 list_del(&tage->linkage);
 845                 cfs_tage_free(tage);
 846         }
 847
 848         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
 849         if (rc)
 850                 pr_err("LustreError: sync returns: rc = %d\n", rc);
 851 close:
 852         filp_close(filp, NULL);
 853 out:
 854         up_write(&cfs_tracefile_sem);
 855         return rc;
 856 }
 857
 858 void cfs_trace_flush_pages(void)
 859 {
 860         struct page_collection pc;
 861         struct cfs_trace_page *tage;
 862         struct cfs_trace_page *tmp;
 863
 864         pc.pc_want_daemon_pages = 1;
 865         collect_pages(&pc);
 866         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 867
 868                 __LASSERT_TAGE_INVARIANT(tage);
 869
 870                 list_del(&tage->linkage);
 871                 cfs_tage_free(tage);
 872         }
 873 }
 874
 875 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 876                             const char __user *usr_buffer, int usr_buffer_nob)
 877 {
 878         int    nob;
 879
 880         if (usr_buffer_nob > knl_buffer_nob)
 881                 return -EOVERFLOW;
 882
 883         if (copy_from_user(knl_buffer, usr_buffer, usr_buffer_nob))
 884                 return -EFAULT;
 885
 886         nob = strnlen(knl_buffer, usr_buffer_nob);
 887         while (--nob >= 0)                      /* strip trailing whitespace */
 888                 if (!isspace(knl_buffer[nob]))
 889                         break;
 890
 891         if (nob < 0)                            /* empty string */
 892                 return -EINVAL;
 893
 894         if (nob == knl_buffer_nob)              /* no space to terminate */
 895                 return -EOVERFLOW;
 896
 897         knl_buffer[nob + 1] = 0;                /* terminate */
 898         return 0;
 899 }
 900 EXPORT_SYMBOL(cfs_trace_copyin_string);
 901
 902 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 903                              const char *knl_buffer, char *append)
 904 {
 905         /* NB if 'append' != NULL, it's a single character to append to the
 906          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 907          * terminating zero byte) for sysctl entries */
 908         int   nob = strlen(knl_buffer);
 909
 910         if (nob > usr_buffer_nob)
 911                 nob = usr_buffer_nob;
 912
 913         if (copy_to_user(usr_buffer, knl_buffer, nob))
 914                 return -EFAULT;
 915
 916         if (append != NULL && nob < usr_buffer_nob) {
 917                 if (copy_to_user(usr_buffer + nob, append, 1))
 918                         return -EFAULT;
 919
 920                 nob++;
 921         }
 922
 923         return nob;
 924 }
 925 EXPORT_SYMBOL(cfs_trace_copyout_string);
 926
 927 int cfs_trace_allocate_string_buffer(char **str, int nob)
 928 {
 929         if (nob > 2 * PAGE_SIZE)        /* string must be "sensible" */
 930                 return -EINVAL;
 931
 932         *str = kmalloc(nob, GFP_KERNEL | __GFP_ZERO);
 933         if (*str == NULL)
 934                 return -ENOMEM;
 935
 936         return 0;
 937 }
 938
 939 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
 940 {
 941         char         *str;
 942         int           rc;
 943
 944         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 945         if (rc != 0)
 946                 return rc;
 947
 948         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 949                                      usr_str, usr_str_nob);
 950         if (rc != 0)
 951                 goto out;
 952
 953         if (str[0] != '/') {
 954                 rc = -EINVAL;
 955                 goto out;
 956         }
 957         rc = cfs_tracefile_dump_all_pages(str);
 958 out:
 959         kfree(str);
 960         return rc;
 961 }
 962
 963 int cfs_trace_daemon_command(char *str)
 964 {
 965         int       rc = 0;
 966
 967         down_write(&cfs_tracefile_sem);
 968
 969         if (strcmp(str, "stop") == 0) {
 970                 up_write(&cfs_tracefile_sem);
 971                 cfs_trace_stop_thread();
 972                 down_write(&cfs_tracefile_sem);
 973                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 974
 975         } else if (strncmp(str, "size=", 5) == 0) {
 976                 unsigned long tmp;
 977
 978                 rc = kstrtoul(str + 5, 10, &tmp);
 979                 if (!rc) {
 980                         if (tmp < 10 || tmp > 20480)
 981                                 cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 982                         else
 983                                 cfs_tracefile_size = tmp << 20;
 984                 }
 985         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 986                 rc = -ENAMETOOLONG;
 987         } else if (str[0] != '/') {
 988                 rc = -EINVAL;
 989         } else {
 990                 strcpy(cfs_tracefile, str);
 991
 992                 pr_info("Lustre: debug daemon will attempt to start writing to %s (%lukB max)\n",
 993                         cfs_tracefile, (long)(cfs_tracefile_size >> 10));
 994
 995                 cfs_trace_start_thread();
 996         }
 997
 998         up_write(&cfs_tracefile_sem);
 999         return rc;
1000 }
1001
1002 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
1003 {
1004         char *str;
1005         int   rc;
1006
1007         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
1008         if (rc != 0)
1009                 return rc;
1010
1011         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
1012                                  usr_str, usr_str_nob);
1013         if (rc == 0)
1014                 rc = cfs_trace_daemon_command(str);
1015
1016         kfree(str);
1017         return rc;
1018 }
1019
1020 int cfs_trace_set_debug_mb(int mb)
1021 {
1022         int i;
1023         int j;
1024         unsigned long pages;
1025         unsigned long total_mb = (cfs_totalram_pages() >> (20 - PAGE_SHIFT));
1026         unsigned long limit = max_t(unsigned long, 512, (total_mb * 4) / 5);
1027         struct cfs_trace_cpu_data *tcd;
1028
1029         if (mb < num_possible_cpus()) {
1030                 pr_warn("Lustre: %d MB is too small for debug buffer size, setting it to %d MB.\n",
1031                         mb, num_possible_cpus());
1032                 mb = num_possible_cpus();
1033         }
1034
1035         if (mb > limit) {
1036                 pr_warn("Lustre: %d MB is too large for debug buffer size, setting it to %lu MB.\n",
1037                         mb, limit);
1038                 mb = limit;
1039         }
1040
1041         mb /= num_possible_cpus();
1042         pages = mb << (20 - PAGE_SHIFT);
1043
1044         down_write(&cfs_tracefile_sem);
1045
1046         cfs_tcd_for_each(tcd, i, j)
1047                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
1048
1049         up_write(&cfs_tracefile_sem);
1050
1051         return mb;
1052 }
1053
1054 int cfs_trace_get_debug_mb(void)
1055 {
1056         int i;
1057         int j;
1058         struct cfs_trace_cpu_data *tcd;
1059         int total_pages = 0;
1060
1061         down_read(&cfs_tracefile_sem);
1062
1063         cfs_tcd_for_each(tcd, i, j)
1064                 total_pages += tcd->tcd_max_pages;
1065
1066         up_read(&cfs_tracefile_sem);
1067
1068         return (total_pages >> (20 - PAGE_SHIFT)) + 1;
1069 }
1070
1071 static int tracefiled(void *arg)
1072 {
1073         struct page_collection pc;
1074         struct tracefiled_ctl *tctl = arg;
1075         struct cfs_trace_page *tage;
1076         struct cfs_trace_page *tmp;
1077         struct file *filp;
1078         char *buf;
1079         int last_loop = 0;
1080         int rc;
1081
1082         /* we're started late enough that we pick up init's fs context */
1083         /* this is so broken in uml?  what on earth is going on? */
1084
1085         complete(&tctl->tctl_start);
1086
1087         while (1) {
1088                 wait_queue_entry_t __wait;
1089
1090                 pc.pc_want_daemon_pages = 0;
1091                 collect_pages(&pc);
1092                 if (list_empty(&pc.pc_pages))
1093                         goto end_loop;
1094
1095                 filp = NULL;
1096                 down_read(&cfs_tracefile_sem);
1097                 if (cfs_tracefile[0] != 0) {
1098                         filp = filp_open(cfs_tracefile,
1099                                          O_CREAT | O_RDWR | O_LARGEFILE,
1100                                          0600);
1101                         if (IS_ERR(filp)) {
1102                                 rc = PTR_ERR(filp);
1103                                 filp = NULL;
1104                                 pr_warn("Lustre: couldn't open %s: rc = %d\n",
1105                                         cfs_tracefile, rc);
1106                         }
1107                 }
1108                 up_read(&cfs_tracefile_sem);
1109                 if (filp == NULL) {
1110                         put_pages_on_daemon_list(&pc);
1111                         __LASSERT(list_empty(&pc.pc_pages));
1112                         goto end_loop;
1113                 }
1114
1115                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
1116                         struct dentry *de = file_dentry(filp);
1117                         static loff_t f_pos;
1118
1119                         __LASSERT_TAGE_INVARIANT(tage);
1120
1121                         if (f_pos >= (off_t)cfs_tracefile_size)
1122                                 f_pos = 0;
1123                         else if (f_pos > i_size_read(de->d_inode))
1124                                 f_pos = i_size_read(de->d_inode);
1125
1126                         buf = kmap(tage->page);
1127                         rc = cfs_kernel_write(filp, buf, tage->used, &f_pos);
1128                         kunmap(tage->page);
1129                         if (rc != (int)tage->used) {
1130                                 pr_warn("Lustre: wanted to write %u but wrote %d\n",
1131                                         tage->used, rc);
1132                                 put_pages_back(&pc);
1133                                 __LASSERT(list_empty(&pc.pc_pages));
1134                                 break;
1135                         }
1136                 }
1137
1138                 filp_close(filp, NULL);
1139                 put_pages_on_daemon_list(&pc);
1140                 if (!list_empty(&pc.pc_pages)) {
1141                         int i;
1142
1143                         pr_alert("Lustre: trace pages aren't empty\n");
1144                         pr_err("Lustre: total cpus(%d): ", num_possible_cpus());
1145                         for (i = 0; i < num_possible_cpus(); i++)
1146                                 if (cpu_online(i))
1147                                         pr_cont("%d(on) ", i);
1148                                 else
1149                                         pr_cont("%d(off) ", i);
1150                         pr_cont("\n");
1151
1152                         i = 0;
1153                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1154                                                  linkage)
1155                                 pr_err("Lustre: page %d belongs to cpu %d\n",
1156                                        ++i, tage->cpu);
1157                         pr_err("Lustre: There are %d pages unwritten\n", i);
1158                 }
1159                 __LASSERT(list_empty(&pc.pc_pages));
1160 end_loop:
1161                 if (atomic_read(&tctl->tctl_shutdown)) {
1162                         if (last_loop == 0) {
1163                                 last_loop = 1;
1164                                 continue;
1165                         } else {
1166                                 break;
1167                         }
1168                 }
1169                 init_wait(&__wait);
1170                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1171                 schedule_timeout_interruptible(cfs_time_seconds(1));
1172                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1173         }
1174         complete(&tctl->tctl_stop);
1175         return 0;
1176 }
1177
1178 int cfs_trace_start_thread(void)
1179 {
1180         struct tracefiled_ctl *tctl = &trace_tctl;
1181         int rc = 0;
1182
1183         mutex_lock(&cfs_trace_thread_mutex);
1184         if (thread_running)
1185                 goto out;
1186
1187         init_completion(&tctl->tctl_start);
1188         init_completion(&tctl->tctl_stop);
1189         init_waitqueue_head(&tctl->tctl_waitq);
1190         atomic_set(&tctl->tctl_shutdown, 0);
1191
1192         if (IS_ERR(kthread_run(tracefiled, tctl, "ktracefiled"))) {
1193                 rc = -ECHILD;
1194                 goto out;
1195         }
1196
1197         wait_for_completion(&tctl->tctl_start);
1198         thread_running = 1;
1199 out:
1200         mutex_unlock(&cfs_trace_thread_mutex);
1201         return rc;
1202 }
1203
1204 void cfs_trace_stop_thread(void)
1205 {
1206         struct tracefiled_ctl *tctl = &trace_tctl;
1207
1208         mutex_lock(&cfs_trace_thread_mutex);
1209         if (thread_running) {
1210                 pr_info("Lustre: shutting down debug daemon thread...\n");
1211                 atomic_set(&tctl->tctl_shutdown, 1);
1212                 wait_for_completion(&tctl->tctl_stop);
1213                 thread_running = 0;
1214         }
1215         mutex_unlock(&cfs_trace_thread_mutex);
1216 }
1217
1218 /* percents to share the total debug memory for each type */
1219 static unsigned int pages_factor[CFS_TCD_TYPE_MAX] = {
1220         80, /* 80% pages for CFS_TCD_TYPE_PROC */
1221         10, /* 10% pages for CFS_TCD_TYPE_SOFTIRQ */
1222         10  /* 10% pages for CFS_TCD_TYPE_IRQ */
1223 };
1224
1225 int cfs_tracefile_init(int max_pages)
1226 {
1227         struct cfs_trace_cpu_data *tcd;
1228         int i;
1229         int j;
1230
1231         /* initialize trace_data */
1232         memset(cfs_trace_data, 0, sizeof(cfs_trace_data));
1233         for (i = 0; i < CFS_TCD_TYPE_MAX; i++) {
1234                 cfs_trace_data[i] =
1235                         kmalloc_array(num_possible_cpus(),
1236                                       sizeof(union cfs_trace_data_union),
1237                                       GFP_KERNEL);
1238                 if (!cfs_trace_data[i])
1239                         goto out_trace_data;
1240         }
1241
1242         /* arch related info initialized */
1243         cfs_tcd_for_each(tcd, i, j) {
1244                 int factor = pages_factor[i];
1245
1246                 spin_lock_init(&tcd->tcd_lock);
1247                 tcd->tcd_pages_factor = factor;
1248                 tcd->tcd_type = i;
1249                 tcd->tcd_cpu = j;
1250
1251                 INIT_LIST_HEAD(&tcd->tcd_pages);
1252                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1253                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1254                 tcd->tcd_cur_pages = 0;
1255                 tcd->tcd_cur_stock_pages = 0;
1256                 tcd->tcd_cur_daemon_pages = 0;
1257                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1258                 LASSERT(tcd->tcd_max_pages > 0);
1259                 tcd->tcd_shutting_down = 0;
1260         }
1261
1262         for (i = 0; i < num_possible_cpus(); i++)
1263                 for (j = 0; j < 3; j++) {
1264                         cfs_trace_console_buffers[i][j] =
1265                                 kmalloc(CFS_TRACE_CONSOLE_BUFFER_SIZE,
1266                                         GFP_KERNEL);
1267                         if (!cfs_trace_console_buffers[i][j])
1268                                 goto out_buffers;
1269                 }
1270
1271         return 0;
1272
1273 out_buffers:
1274         for (i = 0; i < num_possible_cpus(); i++)
1275                 for (j = 0; j < 3; j++) {
1276                         kfree(cfs_trace_console_buffers[i][j]);
1277                         cfs_trace_console_buffers[i][j] = NULL;
1278                 }
1279 out_trace_data:
1280         for (i = 0; cfs_trace_data[i]; i++) {
1281                 kfree(cfs_trace_data[i]);
1282                 cfs_trace_data[i] = NULL;
1283         }
1284         pr_err("lnet: Not enough memory\n");
1285         return -ENOMEM;
1286 }
1287
1288 static void trace_cleanup_on_all_cpus(void)
1289 {
1290         struct cfs_trace_cpu_data *tcd;
1291         struct cfs_trace_page *tage;
1292         struct cfs_trace_page *tmp;
1293         int i, cpu;
1294
1295         for_each_possible_cpu(cpu) {
1296                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1297                         if (!tcd->tcd_pages_factor)
1298                                 /* Not initialised */
1299                                 continue;
1300                         tcd->tcd_shutting_down = 1;
1301
1302                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1303                                 __LASSERT_TAGE_INVARIANT(tage);
1304
1305                                 list_del(&tage->linkage);
1306                                 cfs_tage_free(tage);
1307                         }
1308                         tcd->tcd_cur_pages = 0;
1309                 }
1310         }
1311 }
1312
1313 static void cfs_trace_cleanup(void)
1314 {
1315         struct page_collection pc;
1316         int i;
1317         int j;
1318
1319         INIT_LIST_HEAD(&pc.pc_pages);
1320
1321         trace_cleanup_on_all_cpus();
1322
1323         for (i = 0; i < num_possible_cpus(); i++)
1324                 for (j = 0; j < 3; j++) {
1325                         kfree(cfs_trace_console_buffers[i][j]);
1326                         cfs_trace_console_buffers[i][j] = NULL;
1327                 }
1328
1329         for (i = 0; cfs_trace_data[i]; i++) {
1330                 kfree(cfs_trace_data[i]);
1331                 cfs_trace_data[i] = NULL;
1332         }
1333 }
1334
1335 void cfs_tracefile_exit(void)
1336 {
1337         cfs_trace_stop_thread();
1338         cfs_trace_cleanup();
1339 }