libcfs/libcfs/tracefile.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * libcfs/libcfs/tracefile.c
  33  *
  34  * Author: Zach Brown <zab@clusterfs.com>
  35  * Author: Phil Schwan <phil@clusterfs.com>
  36  */
  37
  38
  39 #define DEBUG_SUBSYSTEM S_LNET
  40 #define LUSTRE_TRACEFILE_PRIVATE
  41 #include "tracefile.h"
  42
  43 #include <linux/ctype.h>
  44 #include <linux/fs.h>
  45 #include <linux/kthread.h>
  46 #include <linux/pagemap.h>
  47 #include <linux/poll.h>
  48 #include <linux/tty.h>
  49 #include <linux/uaccess.h>
  50 #include <libcfs/linux/linux-fs.h>
  51 #include <libcfs/libcfs.h>
  52
  53 #define TCD_MAX_TYPES                   8
  54
  55 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  56
  57 char *cfs_trace_console_buffers[NR_CPUS][CFS_TCD_TYPE_MAX];
  58 char cfs_tracefile[TRACEFILE_NAME_SIZE];
  59 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
  60 static struct tracefiled_ctl trace_tctl;
  61 static DEFINE_MUTEX(cfs_trace_thread_mutex);
  62 static int thread_running = 0;
  63
  64 static atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
  65 static DECLARE_RWSEM(cfs_tracefile_sem);
  66
  67 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  68                                         struct cfs_trace_cpu_data *tcd);
  69
  70 /* trace file lock routines */
  71 /* The walking argument indicates the locking comes from all tcd types
  72  * iterator and we must lock it and dissable local irqs to avoid deadlocks
  73  * with other interrupt locks that might be happening. See LU-1311
  74  * for details.
  75  */
  76 int cfs_trace_lock_tcd(struct cfs_trace_cpu_data *tcd, int walking)
  77         __acquires(&tcd->tcd_lock)
  78 {
  79         __LASSERT(tcd->tcd_type < CFS_TCD_TYPE_MAX);
  80         if (tcd->tcd_type == CFS_TCD_TYPE_IRQ)
  81                 spin_lock_irqsave(&tcd->tcd_lock, tcd->tcd_lock_flags);
  82         else if (tcd->tcd_type == CFS_TCD_TYPE_SOFTIRQ)
  83                 spin_lock_bh(&tcd->tcd_lock);
  84         else if (unlikely(walking))
  85                 spin_lock_irq(&tcd->tcd_lock);
  86         else
  87                 spin_lock(&tcd->tcd_lock);
  88         return 1;
  89 }
  90
  91 void cfs_trace_unlock_tcd(struct cfs_trace_cpu_data *tcd, int walking)
  92         __releases(&tcd->tcd_lock)
  93 {
  94         __LASSERT(tcd->tcd_type < CFS_TCD_TYPE_MAX);
  95         if (tcd->tcd_type == CFS_TCD_TYPE_IRQ)
  96                 spin_unlock_irqrestore(&tcd->tcd_lock, tcd->tcd_lock_flags);
  97         else if (tcd->tcd_type == CFS_TCD_TYPE_SOFTIRQ)
  98                 spin_unlock_bh(&tcd->tcd_lock);
  99         else if (unlikely(walking))
 100                 spin_unlock_irq(&tcd->tcd_lock);
 101         else
 102                 spin_unlock(&tcd->tcd_lock);
 103 }
 104
 105 #define cfs_tcd_for_each(tcd, i, j)                                     \
 106         for (i = 0; cfs_trace_data[i]; i++)                             \
 107                 for (j = 0, ((tcd) = &(*cfs_trace_data[i])[j].tcd);     \
 108                      j < num_possible_cpus();                           \
 109                      j++, (tcd) = &(*cfs_trace_data[i])[j].tcd)
 110
 111 #define cfs_tcd_for_each_type_lock(tcd, i, cpu)                         \
 112         for (i = 0; cfs_trace_data[i] &&                                \
 113              (tcd = &(*cfs_trace_data[i])[cpu].tcd) &&                  \
 114              cfs_trace_lock_tcd(tcd, 1); cfs_trace_unlock_tcd(tcd, 1), i++)
 115
 116 enum cfs_trace_buf_type cfs_trace_buf_idx_get(void)
 117 {
 118         if (in_irq())
 119                 return CFS_TCD_TYPE_IRQ;
 120         if (in_softirq())
 121                 return CFS_TCD_TYPE_SOFTIRQ;
 122         return CFS_TCD_TYPE_PROC;
 123 }
 124
 125 static inline struct cfs_trace_cpu_data *
 126 cfs_trace_get_tcd(void)
 127 {
 128         struct cfs_trace_cpu_data *tcd =
 129                 &(*cfs_trace_data[cfs_trace_buf_idx_get()])[get_cpu()].tcd;
 130
 131         cfs_trace_lock_tcd(tcd, 0);
 132
 133         return tcd;
 134 }
 135
 136 static inline void cfs_trace_put_tcd(struct cfs_trace_cpu_data *tcd)
 137 {
 138         cfs_trace_unlock_tcd(tcd, 0);
 139
 140         put_cpu();
 141 }
 142
 143 static inline struct cfs_trace_page *
 144 cfs_tage_from_list(struct list_head *list)
 145 {
 146         return list_entry(list, struct cfs_trace_page, linkage);
 147 }
 148
 149 static struct cfs_trace_page *cfs_tage_alloc(gfp_t gfp)
 150 {
 151         struct page            *page;
 152         struct cfs_trace_page *tage;
 153
 154         /* My caller is trying to free memory */
 155         if (!in_interrupt() && (current->flags & PF_MEMALLOC))
 156                 return NULL;
 157
 158         /*
 159          * Don't spam console with allocation failures: they will be reported
 160          * by upper layer anyway.
 161          */
 162         gfp |= __GFP_NOWARN;
 163         page = alloc_page(gfp);
 164         if (page == NULL)
 165                 return NULL;
 166
 167         tage = kmalloc(sizeof(*tage), gfp);
 168         if (tage == NULL) {
 169                 __free_page(page);
 170                 return NULL;
 171         }
 172
 173         tage->page = page;
 174         atomic_inc(&cfs_tage_allocated);
 175         return tage;
 176 }
 177
 178 static void cfs_tage_free(struct cfs_trace_page *tage)
 179 {
 180         __LASSERT(tage != NULL);
 181         __LASSERT(tage->page != NULL);
 182
 183         __free_page(tage->page);
 184         kfree(tage);
 185         atomic_dec(&cfs_tage_allocated);
 186 }
 187
 188 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
 189                              struct list_head *queue)
 190 {
 191         __LASSERT(tage != NULL);
 192         __LASSERT(queue != NULL);
 193
 194         list_move_tail(&tage->linkage, queue);
 195 }
 196
 197 /* return a page that has 'len' bytes left at the end */
 198 static struct cfs_trace_page *
 199 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
 200 {
 201         struct cfs_trace_page *tage;
 202
 203         if (tcd->tcd_cur_pages > 0) {
 204                 __LASSERT(!list_empty(&tcd->tcd_pages));
 205                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
 206                 if (tage->used + len <= PAGE_SIZE)
 207                         return tage;
 208         }
 209
 210         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 211                 if (tcd->tcd_cur_stock_pages > 0) {
 212                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
 213                         --tcd->tcd_cur_stock_pages;
 214                         list_del_init(&tage->linkage);
 215                 } else {
 216                         tage = cfs_tage_alloc(GFP_ATOMIC);
 217                         if (unlikely(tage == NULL)) {
 218                                 if ((!(current->flags & PF_MEMALLOC) ||
 219                                      in_interrupt()) && printk_ratelimit())
 220                                         pr_warn("Lustre: cannot allocate a tage (%ld)\n",
 221                                                 tcd->tcd_cur_pages);
 222                                 return NULL;
 223                         }
 224                 }
 225
 226                 tage->used = 0;
 227                 tage->cpu = smp_processor_id();
 228                 tage->type = tcd->tcd_type;
 229                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 230                 tcd->tcd_cur_pages++;
 231
 232                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 233                         struct tracefiled_ctl *tctl = &trace_tctl;
 234                         /*
 235                          * wake up tracefiled to process some pages.
 236                          */
 237                         wake_up(&tctl->tctl_waitq);
 238                 }
 239                 return tage;
 240         }
 241         return NULL;
 242 }
 243
 244 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
 245 {
 246         int pgcount = tcd->tcd_cur_pages / 10;
 247         struct page_collection pc;
 248         struct cfs_trace_page *tage;
 249         struct cfs_trace_page *tmp;
 250
 251         /*
 252          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 253          * from here: this will lead to infinite recursion.
 254          */
 255
 256         if (printk_ratelimit())
 257                 pr_warn("Lustre: debug daemon buffer overflowed; discarding 10%% of pages (%d of %ld)\n",
 258                         pgcount + 1, tcd->tcd_cur_pages);
 259
 260         INIT_LIST_HEAD(&pc.pc_pages);
 261
 262         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 263                 if (pgcount-- == 0)
 264                         break;
 265
 266                 list_move_tail(&tage->linkage, &pc.pc_pages);
 267                 tcd->tcd_cur_pages--;
 268         }
 269         put_pages_on_tcd_daemon_list(&pc, tcd);
 270 }
 271
 272 /* return a page that has 'len' bytes left at the end */
 273 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 274                                                  unsigned long len)
 275 {
 276         struct cfs_trace_page *tage;
 277
 278         /*
 279          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 280          * from here: this will lead to infinite recursion.
 281          */
 282
 283         if (len > PAGE_SIZE) {
 284                 pr_err("LustreError: cowardly refusing to write %lu bytes in a page\n",
 285                        len);
 286                 return NULL;
 287         }
 288
 289         tage = cfs_trace_get_tage_try(tcd, len);
 290         if (tage != NULL)
 291                 return tage;
 292         if (thread_running)
 293                 cfs_tcd_shrink(tcd);
 294         if (tcd->tcd_cur_pages > 0) {
 295                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 296                 tage->used = 0;
 297                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 298         }
 299         return tage;
 300 }
 301
 302 static void cfs_set_ptldebug_header(struct ptldebug_header *header,
 303                                     struct libcfs_debug_msg_data *msgdata,
 304                                     unsigned long stack)
 305 {
 306         struct timespec64 ts;
 307
 308         ktime_get_real_ts64(&ts);
 309
 310         header->ph_subsys = msgdata->msg_subsys;
 311         header->ph_mask = msgdata->msg_mask;
 312         header->ph_cpu_id = smp_processor_id();
 313         header->ph_type = cfs_trace_buf_idx_get();
 314         /* y2038 safe since all user space treats this as unsigned, but
 315          * will overflow in 2106
 316          */
 317         header->ph_sec = (u32)ts.tv_sec;
 318         header->ph_usec = ts.tv_nsec / NSEC_PER_USEC;
 319         header->ph_stack = stack;
 320         header->ph_pid = current->pid;
 321         header->ph_line_num = msgdata->msg_line;
 322         header->ph_extern_pid = 0;
 323 }
 324
 325 /**
 326  * tty_write_msg - write a message to a certain tty, not just the console.
 327  * @tty: the destination tty_struct
 328  * @msg: the message to write
 329  *
 330  * tty_write_message is not exported, so write a same function for it
 331  *
 332  */
 333 static void tty_write_msg(struct tty_struct *tty, const char *msg)
 334 {
 335         mutex_lock(&tty->atomic_write_lock);
 336         tty_lock(tty);
 337         if (tty->ops->write && tty->count > 0)
 338                 tty->ops->write(tty, msg, strlen(msg));
 339         tty_unlock(tty);
 340         mutex_unlock(&tty->atomic_write_lock);
 341         wake_up_interruptible_poll(&tty->write_wait, POLLOUT);
 342 }
 343
 344 static void cfs_tty_write_message(const char *prefix, int mask, const char *msg)
 345 {
 346         struct tty_struct *tty;
 347
 348         tty = get_current_tty();
 349         if (!tty)
 350                 return;
 351
 352         tty_write_msg(tty, prefix);
 353         if ((mask & D_EMERG) || (mask & D_ERROR))
 354                 tty_write_msg(tty, "Error");
 355         tty_write_msg(tty, ": ");
 356         tty_write_msg(tty, msg);
 357         tty_kref_put(tty);
 358 }
 359
 360 static void cfs_print_to_console(struct ptldebug_header *hdr, int mask,
 361                                  const char *buf, int len, const char *file,
 362                                  const char *fn)
 363 {
 364         char *prefix = "Lustre";
 365
 366         if (hdr->ph_subsys == S_LND || hdr->ph_subsys == S_LNET)
 367                 prefix = "LNet";
 368
 369         if (mask & D_CONSOLE) {
 370                 if (mask & D_EMERG)
 371                         pr_emerg("%sError: %.*s", prefix, len, buf);
 372                 else if (mask & D_ERROR)
 373                         pr_err("%sError: %.*s", prefix, len, buf);
 374                 else if (mask & D_WARNING)
 375                         pr_warn("%s: %.*s", prefix, len, buf);
 376                 else if (mask & libcfs_printk)
 377                         pr_info("%s: %.*s", prefix, len, buf);
 378         } else {
 379                 if (mask & D_EMERG)
 380                         pr_emerg("%sError: %d:%d:(%s:%d:%s()) %.*s", prefix,
 381                                  hdr->ph_pid, hdr->ph_extern_pid, file,
 382                                  hdr->ph_line_num, fn, len, buf);
 383                 else if (mask & D_ERROR)
 384                         pr_err("%sError: %d:%d:(%s:%d:%s()) %.*s", prefix,
 385                                hdr->ph_pid, hdr->ph_extern_pid, file,
 386                                hdr->ph_line_num, fn, len, buf);
 387                 else if (mask & D_WARNING)
 388                         pr_warn("%s: %d:%d:(%s:%d:%s()) %.*s", prefix,
 389                                 hdr->ph_pid, hdr->ph_extern_pid, file,
 390                                 hdr->ph_line_num, fn, len, buf);
 391                 else if (mask & (D_CONSOLE | libcfs_printk))
 392                         pr_info("%s: %.*s", prefix, len, buf);
 393         }
 394
 395         if (mask & D_TTY)
 396                 cfs_tty_write_message(prefix, mask, buf);
 397 }
 398
 399 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 400                      const char *format, ...)
 401 {
 402         struct cfs_trace_cpu_data *tcd = NULL;
 403         struct ptldebug_header header = {0};
 404         struct cfs_trace_page *tage;
 405         /* string_buf is used only if tcd != NULL, and is always set then */
 406         char *string_buf = NULL;
 407         char *debug_buf;
 408         int known_size;
 409         int needed = 85; /* seeded with average message length */
 410         int max_nob;
 411         va_list ap;
 412         int retry;
 413         int mask = msgdata->msg_mask;
 414         char *file = (char *)msgdata->msg_file;
 415         struct cfs_debug_limit_state *cdls = msgdata->msg_cdls;
 416
 417         if (strchr(file, '/'))
 418                 file = strrchr(file, '/') + 1;
 419
 420         tcd = cfs_trace_get_tcd();
 421
 422         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 423          * pins us to a particular CPU.  This avoids an smp_processor_id()
 424          * warning on Linux when debugging is enabled.
 425          */
 426         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 427
 428         if (!tcd)                /* arch may not log in IRQ context */
 429                 goto console;
 430
 431         if (tcd->tcd_cur_pages == 0)
 432                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 433
 434         if (tcd->tcd_shutting_down) {
 435                 cfs_trace_put_tcd(tcd);
 436                 tcd = NULL;
 437                 goto console;
 438         }
 439
 440         known_size = strlen(file) + 1;
 441         if (msgdata->msg_fn)
 442                 known_size += strlen(msgdata->msg_fn) + 1;
 443
 444         if (libcfs_debug_binary)
 445                 known_size += sizeof(header);
 446
 447         /*
 448          * May perform an additional pass to update 'needed' and increase
 449          * tage buffer size to match vsnprintf reported size required
 450          * On the second pass (retry=1) use vscnprintf [which returns
 451          * number of bytes written not including the terminating nul]
 452          * to clarify `needed` is used as number of bytes written
 453          * for the remainder of this function
 454          */
 455         for (retry = 0; retry < 2; retry++) {
 456                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 457                 if (!tage) {
 458                         if (needed + known_size > PAGE_SIZE)
 459                                 mask |= D_ERROR;
 460
 461                         cfs_trace_put_tcd(tcd);
 462                         tcd = NULL;
 463                         goto console;
 464                 }
 465
 466                 string_buf = (char *)page_address(tage->page) +
 467                              tage->used + known_size;
 468
 469                 max_nob = PAGE_SIZE - tage->used - known_size;
 470                 if (max_nob <= 0) {
 471                         pr_emerg("LustreError: negative max_nob: %d\n",
 472                                  max_nob);
 473                         mask |= D_ERROR;
 474                         cfs_trace_put_tcd(tcd);
 475                         tcd = NULL;
 476                         goto console;
 477                 }
 478
 479                 va_start(ap, format);
 480                 if (retry)
 481                         needed = vscnprintf(string_buf, max_nob, format, ap);
 482                 else
 483                         needed = vsnprintf(string_buf, max_nob, format, ap);
 484                 va_end(ap);
 485
 486                 if (needed < max_nob) /* well. printing ok.. */
 487                         break;
 488         }
 489
 490         /* `needed` is actual bytes written to string_buf */
 491         if (*(string_buf + needed - 1) != '\n') {
 492                 pr_info("Lustre: format at %s:%d:%s doesn't end in newline\n",
 493                         file, msgdata->msg_line, msgdata->msg_fn);
 494         } else if (mask & D_TTY) {
 495                 /* TTY needs '\r\n' to move carriage to leftmost position */
 496                 if (needed < 2 || *(string_buf + needed - 2) != '\r')
 497                         pr_info("Lustre: format at %s:%d:%s doesn't end in '\\r\\n'\n",
 498                                 file, msgdata->msg_line, msgdata->msg_fn);
 499         }
 500
 501         header.ph_len = known_size + needed;
 502         debug_buf = (char *)page_address(tage->page) + tage->used;
 503
 504         if (libcfs_debug_binary) {
 505                 memcpy(debug_buf, &header, sizeof(header));
 506                 tage->used += sizeof(header);
 507                 debug_buf += sizeof(header);
 508         }
 509
 510         strlcpy(debug_buf, file, PAGE_SIZE - tage->used);
 511         tage->used += strlen(file) + 1;
 512         debug_buf += strlen(file) + 1;
 513
 514         if (msgdata->msg_fn) {
 515                 strlcpy(debug_buf, msgdata->msg_fn, PAGE_SIZE - tage->used);
 516                 tage->used += strlen(msgdata->msg_fn) + 1;
 517                 debug_buf += strlen(msgdata->msg_fn) + 1;
 518         }
 519
 520         __LASSERT(debug_buf == string_buf);
 521
 522         tage->used += needed;
 523         __LASSERT(tage->used <= PAGE_SIZE);
 524
 525 console:
 526         if ((mask & libcfs_printk) == 0) {
 527                 /* no console output requested */
 528                 if (tcd != NULL)
 529                         cfs_trace_put_tcd(tcd);
 530                 return 1;
 531         }
 532
 533         if (cdls != NULL) {
 534                 if (libcfs_console_ratelimit &&
 535                     cdls->cdls_next != 0 &&     /* not first time ever */
 536                     time_before(jiffies, cdls->cdls_next)) {
 537                         /* skipping a console message */
 538                         cdls->cdls_count++;
 539                         if (tcd != NULL)
 540                                 cfs_trace_put_tcd(tcd);
 541                         return 1;
 542                 }
 543
 544                 if (time_after(jiffies, cdls->cdls_next +
 545                                         libcfs_console_max_delay +
 546                                         cfs_time_seconds(10))) {
 547                         /* last timeout was a long time ago */
 548                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 549                 } else {
 550                         cdls->cdls_delay *= libcfs_console_backoff;
 551                 }
 552
 553                 if (cdls->cdls_delay < libcfs_console_min_delay)
 554                         cdls->cdls_delay = libcfs_console_min_delay;
 555                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 556                         cdls->cdls_delay = libcfs_console_max_delay;
 557
 558                 /* ensure cdls_next is never zero after it's been seen */
 559                 cdls->cdls_next = (jiffies + cdls->cdls_delay) | 1;
 560         }
 561
 562         if (tcd) {
 563                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 564                                      msgdata->msg_fn);
 565                 cfs_trace_put_tcd(tcd);
 566         } else {
 567                 string_buf = cfs_trace_get_console_buffer();
 568
 569                 va_start(ap, format);
 570                 needed = vscnprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 571                                     format, ap);
 572                 va_end(ap);
 573
 574                 cfs_print_to_console(&header, mask,
 575                                      string_buf, needed, file, msgdata->msg_fn);
 576
 577                 put_cpu();
 578         }
 579
 580         if (cdls != NULL && cdls->cdls_count != 0) {
 581                 string_buf = cfs_trace_get_console_buffer();
 582
 583                 needed = scnprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 584                                    "Skipped %d previous similar message%s\n",
 585                                    cdls->cdls_count,
 586                                    (cdls->cdls_count > 1) ? "s" : "");
 587
 588                 /* Do not allow print this to TTY */
 589                 cfs_print_to_console(&header, mask & ~D_TTY, string_buf,
 590                                      needed, file, msgdata->msg_fn);
 591
 592                 put_cpu();
 593                 cdls->cdls_count = 0;
 594         }
 595
 596         return 0;
 597 }
 598 EXPORT_SYMBOL(libcfs_debug_msg);
 599
 600 void
 601 cfs_trace_assertion_failed(const char *str,
 602                            struct libcfs_debug_msg_data *msgdata)
 603 {
 604         struct ptldebug_header hdr;
 605
 606         libcfs_panic_in_progress = 1;
 607         libcfs_catastrophe = 1;
 608         smp_mb();
 609
 610         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 611
 612         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 613                              msgdata->msg_file, msgdata->msg_fn);
 614
 615         panic("Lustre debug assertion failure\n");
 616
 617         /* not reached */
 618 }
 619
 620 static void
 621 panic_collect_pages(struct page_collection *pc)
 622 {
 623         /* Do the collect_pages job on a single CPU: assumes that all other
 624          * CPUs have been stopped during a panic.  If this isn't true for some
 625          * arch, this will have to be implemented separately in each arch.  */
 626         int                        i;
 627         int                        j;
 628         struct cfs_trace_cpu_data *tcd;
 629
 630         INIT_LIST_HEAD(&pc->pc_pages);
 631
 632         cfs_tcd_for_each(tcd, i, j) {
 633                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 634                 tcd->tcd_cur_pages = 0;
 635
 636                 if (pc->pc_want_daemon_pages) {
 637                         list_splice_init(&tcd->tcd_daemon_pages,
 638                                                 &pc->pc_pages);
 639                         tcd->tcd_cur_daemon_pages = 0;
 640                 }
 641         }
 642 }
 643
 644 static void collect_pages_on_all_cpus(struct page_collection *pc)
 645 {
 646         struct cfs_trace_cpu_data *tcd;
 647         int i, cpu;
 648
 649         for_each_possible_cpu(cpu) {
 650                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 651                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 652                         tcd->tcd_cur_pages = 0;
 653                         if (pc->pc_want_daemon_pages) {
 654                                 list_splice_init(&tcd->tcd_daemon_pages,
 655                                                         &pc->pc_pages);
 656                                 tcd->tcd_cur_daemon_pages = 0;
 657                         }
 658                 }
 659         }
 660 }
 661
 662 static void collect_pages(struct page_collection *pc)
 663 {
 664         INIT_LIST_HEAD(&pc->pc_pages);
 665
 666         if (libcfs_panic_in_progress)
 667                 panic_collect_pages(pc);
 668         else
 669                 collect_pages_on_all_cpus(pc);
 670 }
 671
 672 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 673 {
 674         struct cfs_trace_cpu_data *tcd;
 675         struct list_head *cur_head;
 676         struct cfs_trace_page *tage;
 677         struct cfs_trace_page *tmp;
 678         int i, cpu;
 679
 680         for_each_possible_cpu(cpu) {
 681                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 682                         cur_head = tcd->tcd_pages.next;
 683
 684                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 685                                                  linkage) {
 686
 687                                 __LASSERT_TAGE_INVARIANT(tage);
 688
 689                                 if (tage->cpu != cpu || tage->type != i)
 690                                         continue;
 691
 692                                 cfs_tage_to_tail(tage, cur_head);
 693                                 tcd->tcd_cur_pages++;
 694                         }
 695                 }
 696         }
 697 }
 698
 699 static void put_pages_back(struct page_collection *pc)
 700 {
 701         if (!libcfs_panic_in_progress)
 702                 put_pages_back_on_all_cpus(pc);
 703 }
 704
 705 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 706  * we have a good amount of data at all times for dumping during an LBUG, even
 707  * if we have been steadily writing (and otherwise discarding) pages via the
 708  * debug daemon. */
 709 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 710                                          struct cfs_trace_cpu_data *tcd)
 711 {
 712         struct cfs_trace_page *tage;
 713         struct cfs_trace_page *tmp;
 714
 715         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 716                 __LASSERT_TAGE_INVARIANT(tage);
 717
 718                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 719                         continue;
 720
 721                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 722                 tcd->tcd_cur_daemon_pages++;
 723
 724                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 725                         struct cfs_trace_page *victim;
 726
 727                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 728                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 729
 730                         __LASSERT_TAGE_INVARIANT(victim);
 731
 732                         list_del(&victim->linkage);
 733                         cfs_tage_free(victim);
 734                         tcd->tcd_cur_daemon_pages--;
 735                 }
 736         }
 737 }
 738
 739 static void put_pages_on_daemon_list(struct page_collection *pc)
 740 {
 741         struct cfs_trace_cpu_data *tcd;
 742         int i, cpu;
 743
 744         for_each_possible_cpu(cpu) {
 745                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 746                         put_pages_on_tcd_daemon_list(pc, tcd);
 747         }
 748 }
 749
 750 void cfs_trace_debug_print(void)
 751 {
 752         struct page_collection pc;
 753         struct cfs_trace_page *tage;
 754         struct cfs_trace_page *tmp;
 755
 756         pc.pc_want_daemon_pages = 1;
 757         collect_pages(&pc);
 758         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 759                 char *p, *file, *fn;
 760                 struct page *page;
 761
 762                 __LASSERT_TAGE_INVARIANT(tage);
 763
 764                 page = tage->page;
 765                 p = page_address(page);
 766                 while (p < ((char *)page_address(page) + tage->used)) {
 767                         struct ptldebug_header *hdr;
 768                         int len;
 769                         hdr = (void *)p;
 770                         p += sizeof(*hdr);
 771                         file = p;
 772                         p += strlen(file) + 1;
 773                         fn = p;
 774                         p += strlen(fn) + 1;
 775                         len = hdr->ph_len - (int)(p - (char *)hdr);
 776
 777                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 778
 779                         p += len;
 780                 }
 781
 782                 list_del(&tage->linkage);
 783                 cfs_tage_free(tage);
 784         }
 785 }
 786
 787 int cfs_tracefile_dump_all_pages(char *filename)
 788 {
 789         struct page_collection  pc;
 790         struct file             *filp;
 791         struct cfs_trace_page   *tage;
 792         struct cfs_trace_page   *tmp;
 793         char                    *buf;
 794         int rc;
 795
 796         down_write(&cfs_tracefile_sem);
 797
 798         filp = filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600);
 799         if (IS_ERR(filp)) {
 800                 rc = PTR_ERR(filp);
 801                 filp = NULL;
 802                 pr_err("LustreError: can't open %s for dump: rc = %d\n",
 803                       filename, rc);
 804                 goto out;
 805         }
 806
 807         pc.pc_want_daemon_pages = 1;
 808         collect_pages(&pc);
 809         if (list_empty(&pc.pc_pages)) {
 810                 rc = 0;
 811                 goto close;
 812         }
 813
 814         /* ok, for now, just write the pages.  in the future we'll be building
 815          * iobufs with the pages and calling generic_direct_IO */
 816         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 817
 818                 __LASSERT_TAGE_INVARIANT(tage);
 819
 820                 buf = kmap(tage->page);
 821                 rc = cfs_kernel_write(filp, buf, tage->used, &filp->f_pos);
 822                 kunmap(tage->page);
 823                 if (rc != (int)tage->used) {
 824                         pr_warn("Lustre: wanted to write %u but wrote %d\n",
 825                                 tage->used, rc);
 826                         put_pages_back(&pc);
 827                         __LASSERT(list_empty(&pc.pc_pages));
 828                         break;
 829                 }
 830                 list_del(&tage->linkage);
 831                 cfs_tage_free(tage);
 832         }
 833
 834         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
 835         if (rc)
 836                 pr_err("LustreError: sync returns: rc = %d\n", rc);
 837 close:
 838         filp_close(filp, NULL);
 839 out:
 840         up_write(&cfs_tracefile_sem);
 841         return rc;
 842 }
 843
 844 void cfs_trace_flush_pages(void)
 845 {
 846         struct page_collection pc;
 847         struct cfs_trace_page *tage;
 848         struct cfs_trace_page *tmp;
 849
 850         pc.pc_want_daemon_pages = 1;
 851         collect_pages(&pc);
 852         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 853
 854                 __LASSERT_TAGE_INVARIANT(tage);
 855
 856                 list_del(&tage->linkage);
 857                 cfs_tage_free(tage);
 858         }
 859 }
 860
 861 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 862                             const char __user *usr_buffer, int usr_buffer_nob)
 863 {
 864         int    nob;
 865
 866         if (usr_buffer_nob > knl_buffer_nob)
 867                 return -EOVERFLOW;
 868
 869         if (copy_from_user(knl_buffer, usr_buffer, usr_buffer_nob))
 870                 return -EFAULT;
 871
 872         nob = strnlen(knl_buffer, usr_buffer_nob);
 873         while (--nob >= 0)                      /* strip trailing whitespace */
 874                 if (!isspace(knl_buffer[nob]))
 875                         break;
 876
 877         if (nob < 0)                            /* empty string */
 878                 return -EINVAL;
 879
 880         if (nob == knl_buffer_nob)              /* no space to terminate */
 881                 return -EOVERFLOW;
 882
 883         knl_buffer[nob + 1] = 0;                /* terminate */
 884         return 0;
 885 }
 886 EXPORT_SYMBOL(cfs_trace_copyin_string);
 887
 888 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 889                              const char *knl_buffer, char *append)
 890 {
 891         /* NB if 'append' != NULL, it's a single character to append to the
 892          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 893          * terminating zero byte) for sysctl entries */
 894         int   nob = strlen(knl_buffer);
 895
 896         if (nob > usr_buffer_nob)
 897                 nob = usr_buffer_nob;
 898
 899         if (copy_to_user(usr_buffer, knl_buffer, nob))
 900                 return -EFAULT;
 901
 902         if (append != NULL && nob < usr_buffer_nob) {
 903                 if (copy_to_user(usr_buffer + nob, append, 1))
 904                         return -EFAULT;
 905
 906                 nob++;
 907         }
 908
 909         return nob;
 910 }
 911 EXPORT_SYMBOL(cfs_trace_copyout_string);
 912
 913 int cfs_trace_allocate_string_buffer(char **str, int nob)
 914 {
 915         if (nob > 2 * PAGE_SIZE)        /* string must be "sensible" */
 916                 return -EINVAL;
 917
 918         *str = kmalloc(nob, GFP_KERNEL | __GFP_ZERO);
 919         if (*str == NULL)
 920                 return -ENOMEM;
 921
 922         return 0;
 923 }
 924
 925 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
 926 {
 927         char         *str;
 928         int           rc;
 929
 930         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 931         if (rc != 0)
 932                 return rc;
 933
 934         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 935                                      usr_str, usr_str_nob);
 936         if (rc != 0)
 937                 goto out;
 938
 939         if (str[0] != '/') {
 940                 rc = -EINVAL;
 941                 goto out;
 942         }
 943         rc = cfs_tracefile_dump_all_pages(str);
 944 out:
 945         kfree(str);
 946         return rc;
 947 }
 948
 949 int cfs_trace_daemon_command(char *str)
 950 {
 951         int       rc = 0;
 952
 953         down_write(&cfs_tracefile_sem);
 954
 955         if (strcmp(str, "stop") == 0) {
 956                 up_write(&cfs_tracefile_sem);
 957                 cfs_trace_stop_thread();
 958                 down_write(&cfs_tracefile_sem);
 959                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
 960
 961         } else if (strncmp(str, "size=", 5) == 0) {
 962                 unsigned long tmp;
 963
 964                 rc = kstrtoul(str + 5, 10, &tmp);
 965                 if (!rc) {
 966                         if (tmp < 10 || tmp > 20480)
 967                                 cfs_tracefile_size = CFS_TRACEFILE_SIZE;
 968                         else
 969                                 cfs_tracefile_size = tmp << 20;
 970                 }
 971         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
 972                 rc = -ENAMETOOLONG;
 973         } else if (str[0] != '/') {
 974                 rc = -EINVAL;
 975         } else {
 976                 strcpy(cfs_tracefile, str);
 977
 978                 pr_info("Lustre: debug daemon will attempt to start writing to %s (%lukB max)\n",
 979                         cfs_tracefile, (long)(cfs_tracefile_size >> 10));
 980
 981                 cfs_trace_start_thread();
 982         }
 983
 984         up_write(&cfs_tracefile_sem);
 985         return rc;
 986 }
 987
 988 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
 989 {
 990         char *str;
 991         int   rc;
 992
 993         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
 994         if (rc != 0)
 995                 return rc;
 996
 997         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
 998                                  usr_str, usr_str_nob);
 999         if (rc == 0)
1000                 rc = cfs_trace_daemon_command(str);
1001
1002         kfree(str);
1003         return rc;
1004 }
1005
1006 int cfs_trace_set_debug_mb(int mb)
1007 {
1008         int i;
1009         int j;
1010         unsigned long pages;
1011         unsigned long total_mb = (cfs_totalram_pages() >> (20 - PAGE_SHIFT));
1012         unsigned long limit = max_t(unsigned long, 512, (total_mb * 4) / 5);
1013         struct cfs_trace_cpu_data *tcd;
1014
1015         if (mb < num_possible_cpus()) {
1016                 pr_warn("Lustre: %d MB is too small for debug buffer size, setting it to %d MB.\n",
1017                         mb, num_possible_cpus());
1018                 mb = num_possible_cpus();
1019         }
1020
1021         if (mb > limit) {
1022                 pr_warn("Lustre: %d MB is too large for debug buffer size, setting it to %lu MB.\n",
1023                         mb, limit);
1024                 mb = limit;
1025         }
1026
1027         mb /= num_possible_cpus();
1028         pages = mb << (20 - PAGE_SHIFT);
1029
1030         down_write(&cfs_tracefile_sem);
1031
1032         cfs_tcd_for_each(tcd, i, j)
1033                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
1034
1035         up_write(&cfs_tracefile_sem);
1036
1037         return mb;
1038 }
1039
1040 int cfs_trace_get_debug_mb(void)
1041 {
1042         int i;
1043         int j;
1044         struct cfs_trace_cpu_data *tcd;
1045         int total_pages = 0;
1046
1047         down_read(&cfs_tracefile_sem);
1048
1049         cfs_tcd_for_each(tcd, i, j)
1050                 total_pages += tcd->tcd_max_pages;
1051
1052         up_read(&cfs_tracefile_sem);
1053
1054         return (total_pages >> (20 - PAGE_SHIFT)) + 1;
1055 }
1056
1057 static int tracefiled(void *arg)
1058 {
1059         struct page_collection pc;
1060         struct tracefiled_ctl *tctl = arg;
1061         struct cfs_trace_page *tage;
1062         struct cfs_trace_page *tmp;
1063         struct file *filp;
1064         char *buf;
1065         int last_loop = 0;
1066         int rc;
1067
1068         /* we're started late enough that we pick up init's fs context */
1069         /* this is so broken in uml?  what on earth is going on? */
1070
1071         complete(&tctl->tctl_start);
1072
1073         while (1) {
1074                 wait_queue_entry_t __wait;
1075
1076                 pc.pc_want_daemon_pages = 0;
1077                 collect_pages(&pc);
1078                 if (list_empty(&pc.pc_pages))
1079                         goto end_loop;
1080
1081                 filp = NULL;
1082                 down_read(&cfs_tracefile_sem);
1083                 if (cfs_tracefile[0] != 0) {
1084                         filp = filp_open(cfs_tracefile,
1085                                          O_CREAT | O_RDWR | O_LARGEFILE,
1086                                          0600);
1087                         if (IS_ERR(filp)) {
1088                                 rc = PTR_ERR(filp);
1089                                 filp = NULL;
1090                                 pr_warn("Lustre: couldn't open %s: rc = %d\n",
1091                                         cfs_tracefile, rc);
1092                         }
1093                 }
1094                 up_read(&cfs_tracefile_sem);
1095                 if (filp == NULL) {
1096                         put_pages_on_daemon_list(&pc);
1097                         __LASSERT(list_empty(&pc.pc_pages));
1098                         goto end_loop;
1099                 }
1100
1101                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
1102                         struct dentry *de = file_dentry(filp);
1103                         static loff_t f_pos;
1104
1105                         __LASSERT_TAGE_INVARIANT(tage);
1106
1107                         if (f_pos >= (off_t)cfs_tracefile_size)
1108                                 f_pos = 0;
1109                         else if (f_pos > i_size_read(de->d_inode))
1110                                 f_pos = i_size_read(de->d_inode);
1111
1112                         buf = kmap(tage->page);
1113                         rc = cfs_kernel_write(filp, buf, tage->used, &f_pos);
1114                         kunmap(tage->page);
1115                         if (rc != (int)tage->used) {
1116                                 pr_warn("Lustre: wanted to write %u but wrote %d\n",
1117                                         tage->used, rc);
1118                                 put_pages_back(&pc);
1119                                 __LASSERT(list_empty(&pc.pc_pages));
1120                                 break;
1121                         }
1122                 }
1123
1124                 filp_close(filp, NULL);
1125                 put_pages_on_daemon_list(&pc);
1126                 if (!list_empty(&pc.pc_pages)) {
1127                         int i;
1128
1129                         pr_alert("Lustre: trace pages aren't empty\n");
1130                         pr_err("Lustre: total cpus(%d): ", num_possible_cpus());
1131                         for (i = 0; i < num_possible_cpus(); i++)
1132                                 if (cpu_online(i))
1133                                         pr_cont("%d(on) ", i);
1134                                 else
1135                                         pr_cont("%d(off) ", i);
1136                         pr_cont("\n");
1137
1138                         i = 0;
1139                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1140                                                  linkage)
1141                                 pr_err("Lustre: page %d belongs to cpu %d\n",
1142                                        ++i, tage->cpu);
1143                         pr_err("Lustre: There are %d pages unwritten\n", i);
1144                 }
1145                 __LASSERT(list_empty(&pc.pc_pages));
1146 end_loop:
1147                 if (atomic_read(&tctl->tctl_shutdown)) {
1148                         if (last_loop == 0) {
1149                                 last_loop = 1;
1150                                 continue;
1151                         } else {
1152                                 break;
1153                         }
1154                 }
1155                 init_waitqueue_entry(&__wait, current);
1156                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1157                 schedule_timeout_interruptible(cfs_time_seconds(1));
1158                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1159         }
1160         complete(&tctl->tctl_stop);
1161         return 0;
1162 }
1163
1164 int cfs_trace_start_thread(void)
1165 {
1166         struct tracefiled_ctl *tctl = &trace_tctl;
1167         int rc = 0;
1168
1169         mutex_lock(&cfs_trace_thread_mutex);
1170         if (thread_running)
1171                 goto out;
1172
1173         init_completion(&tctl->tctl_start);
1174         init_completion(&tctl->tctl_stop);
1175         init_waitqueue_head(&tctl->tctl_waitq);
1176         atomic_set(&tctl->tctl_shutdown, 0);
1177
1178         if (IS_ERR(kthread_run(tracefiled, tctl, "ktracefiled"))) {
1179                 rc = -ECHILD;
1180                 goto out;
1181         }
1182
1183         wait_for_completion(&tctl->tctl_start);
1184         thread_running = 1;
1185 out:
1186         mutex_unlock(&cfs_trace_thread_mutex);
1187         return rc;
1188 }
1189
1190 void cfs_trace_stop_thread(void)
1191 {
1192         struct tracefiled_ctl *tctl = &trace_tctl;
1193
1194         mutex_lock(&cfs_trace_thread_mutex);
1195         if (thread_running) {
1196                 pr_info("Lustre: shutting down debug daemon thread...\n");
1197                 atomic_set(&tctl->tctl_shutdown, 1);
1198                 wait_for_completion(&tctl->tctl_stop);
1199                 thread_running = 0;
1200         }
1201         mutex_unlock(&cfs_trace_thread_mutex);
1202 }
1203
1204 /* percents to share the total debug memory for each type */
1205 static unsigned int pages_factor[CFS_TCD_TYPE_MAX] = {
1206         80, /* 80% pages for CFS_TCD_TYPE_PROC */
1207         10, /* 10% pages for CFS_TCD_TYPE_SOFTIRQ */
1208         10  /* 10% pages for CFS_TCD_TYPE_IRQ */
1209 };
1210
1211 int cfs_tracefile_init(int max_pages)
1212 {
1213         struct cfs_trace_cpu_data *tcd;
1214         int i;
1215         int j;
1216
1217         /* initialize trace_data */
1218         memset(cfs_trace_data, 0, sizeof(cfs_trace_data));
1219         for (i = 0; i < CFS_TCD_TYPE_MAX; i++) {
1220                 cfs_trace_data[i] =
1221                         kmalloc_array(num_possible_cpus(),
1222                                       sizeof(union cfs_trace_data_union),
1223                                       GFP_KERNEL);
1224                 if (!cfs_trace_data[i])
1225                         goto out_trace_data;
1226         }
1227
1228         /* arch related info initialized */
1229         cfs_tcd_for_each(tcd, i, j) {
1230                 int factor = pages_factor[i];
1231
1232                 spin_lock_init(&tcd->tcd_lock);
1233                 tcd->tcd_pages_factor = factor;
1234                 tcd->tcd_type = i;
1235                 tcd->tcd_cpu = j;
1236
1237                 INIT_LIST_HEAD(&tcd->tcd_pages);
1238                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1239                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1240                 tcd->tcd_cur_pages = 0;
1241                 tcd->tcd_cur_stock_pages = 0;
1242                 tcd->tcd_cur_daemon_pages = 0;
1243                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1244                 LASSERT(tcd->tcd_max_pages > 0);
1245                 tcd->tcd_shutting_down = 0;
1246         }
1247
1248         for (i = 0; i < num_possible_cpus(); i++)
1249                 for (j = 0; j < 3; j++) {
1250                         cfs_trace_console_buffers[i][j] =
1251                                 kmalloc(CFS_TRACE_CONSOLE_BUFFER_SIZE,
1252                                         GFP_KERNEL);
1253                         if (!cfs_trace_console_buffers[i][j])
1254                                 goto out_buffers;
1255                 }
1256
1257         return 0;
1258
1259 out_buffers:
1260         for (i = 0; i < num_possible_cpus(); i++)
1261                 for (j = 0; j < 3; j++) {
1262                         kfree(cfs_trace_console_buffers[i][j]);
1263                         cfs_trace_console_buffers[i][j] = NULL;
1264                 }
1265 out_trace_data:
1266         for (i = 0; cfs_trace_data[i]; i++) {
1267                 kfree(cfs_trace_data[i]);
1268                 cfs_trace_data[i] = NULL;
1269         }
1270         pr_err("lnet: Not enough memory\n");
1271         return -ENOMEM;
1272 }
1273
1274 static void trace_cleanup_on_all_cpus(void)
1275 {
1276         struct cfs_trace_cpu_data *tcd;
1277         struct cfs_trace_page *tage;
1278         struct cfs_trace_page *tmp;
1279         int i, cpu;
1280
1281         for_each_possible_cpu(cpu) {
1282                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1283                         if (!tcd->tcd_pages_factor)
1284                                 /* Not initialised */
1285                                 continue;
1286                         tcd->tcd_shutting_down = 1;
1287
1288                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1289                                 __LASSERT_TAGE_INVARIANT(tage);
1290
1291                                 list_del(&tage->linkage);
1292                                 cfs_tage_free(tage);
1293                         }
1294                         tcd->tcd_cur_pages = 0;
1295                 }
1296         }
1297 }
1298
1299 static void cfs_trace_cleanup(void)
1300 {
1301         struct page_collection pc;
1302         int i;
1303         int j;
1304
1305         INIT_LIST_HEAD(&pc.pc_pages);
1306
1307         trace_cleanup_on_all_cpus();
1308
1309         for (i = 0; i < num_possible_cpus(); i++)
1310                 for (j = 0; j < 3; j++) {
1311                         kfree(cfs_trace_console_buffers[i][j]);
1312                         cfs_trace_console_buffers[i][j] = NULL;
1313                 }
1314
1315         for (i = 0; cfs_trace_data[i]; i++) {
1316                 kfree(cfs_trace_data[i]);
1317                 cfs_trace_data[i] = NULL;
1318         }
1319 }
1320
1321 void cfs_tracefile_exit(void)
1322 {
1323         cfs_trace_stop_thread();
1324         cfs_trace_cleanup();
1325 }