libcfs/libcfs/tracefile.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * libcfs/libcfs/tracefile.c
  33  *
  34  * Author: Zach Brown <zab@clusterfs.com>
  35  * Author: Phil Schwan <phil@clusterfs.com>
  36  */
  37
  38
  39 #define DEBUG_SUBSYSTEM S_LNET
  40 #define LUSTRE_TRACEFILE_PRIVATE
  41 #include "tracefile.h"
  42
  43 #include <linux/ctype.h>
  44 #include <linux/fs.h>
  45 #include <linux/kthread.h>
  46 #include <linux/pagemap.h>
  47 #include <linux/poll.h>
  48 #include <linux/tty.h>
  49 #include <linux/uaccess.h>
  50 #include <libcfs/linux/linux-fs.h>
  51 #include <libcfs/libcfs.h>
  52
  53 /* XXX move things up to the top, comment */
  54 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  55
  56 char *cfs_trace_console_buffers[NR_CPUS][CFS_TCD_TYPE_MAX];
  57 char cfs_tracefile[TRACEFILE_NAME_SIZE];
  58 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
  59 static struct tracefiled_ctl trace_tctl;
  60 static DEFINE_MUTEX(cfs_trace_thread_mutex);
  61 static int thread_running = 0;
  62
  63 static atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
  64 static DECLARE_RWSEM(cfs_tracefile_sem);
  65
  66 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  67                                         struct cfs_trace_cpu_data *tcd);
  68
  69 /* trace file lock routines */
  70 /* The walking argument indicates the locking comes from all tcd types
  71  * iterator and we must lock it and dissable local irqs to avoid deadlocks
  72  * with other interrupt locks that might be happening. See LU-1311
  73  * for details.
  74  */
  75 int cfs_trace_lock_tcd(struct cfs_trace_cpu_data *tcd, int walking)
  76         __acquires(&tcd->tcd_lock)
  77 {
  78         __LASSERT(tcd->tcd_type < CFS_TCD_TYPE_MAX);
  79         if (tcd->tcd_type == CFS_TCD_TYPE_IRQ)
  80                 spin_lock_irqsave(&tcd->tcd_lock, tcd->tcd_lock_flags);
  81         else if (tcd->tcd_type == CFS_TCD_TYPE_SOFTIRQ)
  82                 spin_lock_bh(&tcd->tcd_lock);
  83         else if (unlikely(walking))
  84                 spin_lock_irq(&tcd->tcd_lock);
  85         else
  86                 spin_lock(&tcd->tcd_lock);
  87         return 1;
  88 }
  89
  90 void cfs_trace_unlock_tcd(struct cfs_trace_cpu_data *tcd, int walking)
  91         __releases(&tcd->tcd_lock)
  92 {
  93         __LASSERT(tcd->tcd_type < CFS_TCD_TYPE_MAX);
  94         if (tcd->tcd_type == CFS_TCD_TYPE_IRQ)
  95                 spin_unlock_irqrestore(&tcd->tcd_lock, tcd->tcd_lock_flags);
  96         else if (tcd->tcd_type == CFS_TCD_TYPE_SOFTIRQ)
  97                 spin_unlock_bh(&tcd->tcd_lock);
  98         else if (unlikely(walking))
  99                 spin_unlock_irq(&tcd->tcd_lock);
 100         else
 101                 spin_unlock(&tcd->tcd_lock);
 102 }
 103
 104 #define cfs_tcd_for_each_type_lock(tcd, i, cpu)                         \
 105         for (i = 0; cfs_trace_data[i] &&                                \
 106              (tcd = &(*cfs_trace_data[i])[cpu].tcd) &&                  \
 107              cfs_trace_lock_tcd(tcd, 1); cfs_trace_unlock_tcd(tcd, 1), i++)
 108
 109 enum cfs_trace_buf_type cfs_trace_buf_idx_get(void)
 110 {
 111         if (in_irq())
 112                 return CFS_TCD_TYPE_IRQ;
 113         if (in_softirq())
 114                 return CFS_TCD_TYPE_SOFTIRQ;
 115         return CFS_TCD_TYPE_PROC;
 116 }
 117
 118 static inline struct cfs_trace_cpu_data *
 119 cfs_trace_get_tcd(void)
 120 {
 121         struct cfs_trace_cpu_data *tcd =
 122                 &(*cfs_trace_data[cfs_trace_buf_idx_get()])[get_cpu()].tcd;
 123
 124         cfs_trace_lock_tcd(tcd, 0);
 125
 126         return tcd;
 127 }
 128
 129 static inline void cfs_trace_put_tcd(struct cfs_trace_cpu_data *tcd)
 130 {
 131         cfs_trace_unlock_tcd(tcd, 0);
 132
 133         put_cpu();
 134 }
 135
 136 /* percents to share the total debug memory for each type */
 137 static unsigned int pages_factor[CFS_TCD_TYPE_MAX] = {
 138         80,     /* 80% pages for CFS_TCD_TYPE_PROC */
 139         10,     /* 10% pages for CFS_TCD_TYPE_SOFTIRQ */
 140         10      /* 10% pages for CFS_TCD_TYPE_IRQ */
 141 };
 142
 143 int cfs_tracefile_init_arch(void)
 144 {
 145         struct cfs_trace_cpu_data *tcd;
 146         int i;
 147         int j;
 148
 149         /* initialize trace_data */
 150         memset(cfs_trace_data, 0, sizeof(cfs_trace_data));
 151         for (i = 0; i < CFS_TCD_TYPE_MAX; i++) {
 152                 cfs_trace_data[i] =
 153                         kmalloc_array(num_possible_cpus(),
 154                                       sizeof(union cfs_trace_data_union),
 155                                       GFP_KERNEL);
 156                 if (!cfs_trace_data[i])
 157                         goto out_trace_data;
 158         }
 159
 160         /* arch related info initialized */
 161         cfs_tcd_for_each(tcd, i, j) {
 162                 spin_lock_init(&tcd->tcd_lock);
 163                 tcd->tcd_pages_factor = pages_factor[i];
 164                 tcd->tcd_type = i;
 165                 tcd->tcd_cpu = j;
 166         }
 167
 168         for (i = 0; i < num_possible_cpus(); i++)
 169                 for (j = 0; j < 3; j++) {
 170                         cfs_trace_console_buffers[i][j] =
 171                                 kmalloc(CFS_TRACE_CONSOLE_BUFFER_SIZE,
 172                                         GFP_KERNEL);
 173
 174                         if (!cfs_trace_console_buffers[i][j])
 175                                 goto out_buffers;
 176                 }
 177
 178         return 0;
 179
 180 out_buffers:
 181         for (i = 0; i < num_possible_cpus(); i++)
 182                 for (j = 0; j < 3; j++) {
 183                         kfree(cfs_trace_console_buffers[i][j]);
 184                         cfs_trace_console_buffers[i][j] = NULL;
 185                 }
 186 out_trace_data:
 187         for (i = 0; cfs_trace_data[i]; i++) {
 188                 kfree(cfs_trace_data[i]);
 189                 cfs_trace_data[i] = NULL;
 190         }
 191         pr_err("lnet: Not enough memory\n");
 192         return -ENOMEM;
 193 }
 194
 195 void cfs_tracefile_fini_arch(void)
 196 {
 197         int i;
 198         int j;
 199
 200         for (i = 0; i < num_possible_cpus(); i++)
 201                 for (j = 0; j < 3; j++) {
 202                         kfree(cfs_trace_console_buffers[i][j]);
 203                         cfs_trace_console_buffers[i][j] = NULL;
 204                 }
 205
 206         for (i = 0; cfs_trace_data[i]; i++) {
 207                 kfree(cfs_trace_data[i]);
 208                 cfs_trace_data[i] = NULL;
 209         }
 210 }
 211
 212 static inline struct cfs_trace_page *
 213 cfs_tage_from_list(struct list_head *list)
 214 {
 215         return list_entry(list, struct cfs_trace_page, linkage);
 216 }
 217
 218 static struct cfs_trace_page *cfs_tage_alloc(gfp_t gfp)
 219 {
 220         struct page            *page;
 221         struct cfs_trace_page *tage;
 222
 223         /* My caller is trying to free memory */
 224         if (!in_interrupt() && (current->flags & PF_MEMALLOC))
 225                 return NULL;
 226
 227         /*
 228          * Don't spam console with allocation failures: they will be reported
 229          * by upper layer anyway.
 230          */
 231         gfp |= __GFP_NOWARN;
 232         page = alloc_page(gfp);
 233         if (page == NULL)
 234                 return NULL;
 235
 236         tage = kmalloc(sizeof(*tage), gfp);
 237         if (tage == NULL) {
 238                 __free_page(page);
 239                 return NULL;
 240         }
 241
 242         tage->page = page;
 243         atomic_inc(&cfs_tage_allocated);
 244         return tage;
 245 }
 246
 247 static void cfs_tage_free(struct cfs_trace_page *tage)
 248 {
 249         __LASSERT(tage != NULL);
 250         __LASSERT(tage->page != NULL);
 251
 252         __free_page(tage->page);
 253         kfree(tage);
 254         atomic_dec(&cfs_tage_allocated);
 255 }
 256
 257 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
 258                              struct list_head *queue)
 259 {
 260         __LASSERT(tage != NULL);
 261         __LASSERT(queue != NULL);
 262
 263         list_move_tail(&tage->linkage, queue);
 264 }
 265
 266 int cfs_trace_refill_stock(struct cfs_trace_cpu_data *tcd, gfp_t gfp,
 267                            struct list_head *stock)
 268 {
 269         int i;
 270
 271         /*
 272          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 273          * from here: this will lead to infinite recursion.
 274          */
 275
 276         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++ i) {
 277                 struct cfs_trace_page *tage;
 278
 279                 tage = cfs_tage_alloc(gfp);
 280                 if (tage == NULL)
 281                         break;
 282                 list_add_tail(&tage->linkage, stock);
 283         }
 284         return i;
 285 }
 286
 287 /* return a page that has 'len' bytes left at the end */
 288 static struct cfs_trace_page *
 289 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
 290 {
 291         struct cfs_trace_page *tage;
 292
 293         if (tcd->tcd_cur_pages > 0) {
 294                 __LASSERT(!list_empty(&tcd->tcd_pages));
 295                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
 296                 if (tage->used + len <= PAGE_SIZE)
 297                         return tage;
 298         }
 299
 300         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 301                 if (tcd->tcd_cur_stock_pages > 0) {
 302                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
 303                         --tcd->tcd_cur_stock_pages;
 304                         list_del_init(&tage->linkage);
 305                 } else {
 306                         tage = cfs_tage_alloc(GFP_ATOMIC);
 307                         if (unlikely(tage == NULL)) {
 308                                 if ((!(current->flags & PF_MEMALLOC) ||
 309                                      in_interrupt()) && printk_ratelimit())
 310                                         pr_warn("Lustre: cannot allocate a tage (%ld)\n",
 311                                                 tcd->tcd_cur_pages);
 312                                 return NULL;
 313                         }
 314                 }
 315
 316                 tage->used = 0;
 317                 tage->cpu = smp_processor_id();
 318                 tage->type = tcd->tcd_type;
 319                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 320                 tcd->tcd_cur_pages++;
 321
 322                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 323                         struct tracefiled_ctl *tctl = &trace_tctl;
 324                         /*
 325                          * wake up tracefiled to process some pages.
 326                          */
 327                         wake_up(&tctl->tctl_waitq);
 328                 }
 329                 return tage;
 330         }
 331         return NULL;
 332 }
 333
 334 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
 335 {
 336         int pgcount = tcd->tcd_cur_pages / 10;
 337         struct page_collection pc;
 338         struct cfs_trace_page *tage;
 339         struct cfs_trace_page *tmp;
 340
 341         /*
 342          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 343          * from here: this will lead to infinite recursion.
 344          */
 345
 346         if (printk_ratelimit())
 347                 pr_warn("Lustre: debug daemon buffer overflowed; discarding 10%% of pages (%d of %ld)\n",
 348                         pgcount + 1, tcd->tcd_cur_pages);
 349
 350         INIT_LIST_HEAD(&pc.pc_pages);
 351
 352         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
 353                 if (pgcount-- == 0)
 354                         break;
 355
 356                 list_move_tail(&tage->linkage, &pc.pc_pages);
 357                 tcd->tcd_cur_pages--;
 358         }
 359         put_pages_on_tcd_daemon_list(&pc, tcd);
 360 }
 361
 362 /* return a page that has 'len' bytes left at the end */
 363 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
 364                                                  unsigned long len)
 365 {
 366         struct cfs_trace_page *tage;
 367
 368         /*
 369          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 370          * from here: this will lead to infinite recursion.
 371          */
 372
 373         if (len > PAGE_SIZE) {
 374                 pr_err("LustreError: cowardly refusing to write %lu bytes in a page\n",
 375                        len);
 376                 return NULL;
 377         }
 378
 379         tage = cfs_trace_get_tage_try(tcd, len);
 380         if (tage != NULL)
 381                 return tage;
 382         if (thread_running)
 383                 cfs_tcd_shrink(tcd);
 384         if (tcd->tcd_cur_pages > 0) {
 385                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
 386                 tage->used = 0;
 387                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
 388         }
 389         return tage;
 390 }
 391
 392 static void cfs_set_ptldebug_header(struct ptldebug_header *header,
 393                                     struct libcfs_debug_msg_data *msgdata,
 394                                     unsigned long stack)
 395 {
 396         struct timespec64 ts;
 397
 398         ktime_get_real_ts64(&ts);
 399
 400         header->ph_subsys = msgdata->msg_subsys;
 401         header->ph_mask = msgdata->msg_mask;
 402         header->ph_cpu_id = smp_processor_id();
 403         header->ph_type = cfs_trace_buf_idx_get();
 404         /* y2038 safe since all user space treats this as unsigned, but
 405          * will overflow in 2106
 406          */
 407         header->ph_sec = (u32)ts.tv_sec;
 408         header->ph_usec = ts.tv_nsec / NSEC_PER_USEC;
 409         header->ph_stack = stack;
 410         header->ph_pid = current->pid;
 411         header->ph_line_num = msgdata->msg_line;
 412         header->ph_extern_pid = 0;
 413 }
 414
 415 /**
 416  * tty_write_msg - write a message to a certain tty, not just the console.
 417  * @tty: the destination tty_struct
 418  * @msg: the message to write
 419  *
 420  * tty_write_message is not exported, so write a same function for it
 421  *
 422  */
 423 static void tty_write_msg(struct tty_struct *tty, const char *msg)
 424 {
 425         mutex_lock(&tty->atomic_write_lock);
 426         tty_lock(tty);
 427         if (tty->ops->write && tty->count > 0)
 428                 tty->ops->write(tty, msg, strlen(msg));
 429         tty_unlock(tty);
 430         mutex_unlock(&tty->atomic_write_lock);
 431         wake_up_interruptible_poll(&tty->write_wait, POLLOUT);
 432 }
 433
 434 static void cfs_tty_write_message(const char *prefix, int mask, const char *msg)
 435 {
 436         struct tty_struct *tty;
 437
 438         tty = get_current_tty();
 439         if (!tty)
 440                 return;
 441
 442         tty_write_msg(tty, prefix);
 443         if ((mask & D_EMERG) || (mask & D_ERROR))
 444                 tty_write_msg(tty, "Error");
 445         tty_write_msg(tty, ": ");
 446         tty_write_msg(tty, msg);
 447         tty_kref_put(tty);
 448 }
 449
 450 static void cfs_print_to_console(struct ptldebug_header *hdr, int mask,
 451                                  const char *buf, int len, const char *file,
 452                                  const char *fn)
 453 {
 454         char *prefix = "Lustre";
 455
 456         if (hdr->ph_subsys == S_LND || hdr->ph_subsys == S_LNET)
 457                 prefix = "LNet";
 458
 459         if (mask & D_CONSOLE) {
 460                 if (mask & D_EMERG)
 461                         pr_emerg("%sError: %.*s", prefix, len, buf);
 462                 else if (mask & D_ERROR)
 463                         pr_err("%sError: %.*s", prefix, len, buf);
 464                 else if (mask & D_WARNING)
 465                         pr_warn("%s: %.*s", prefix, len, buf);
 466                 else if (mask & libcfs_printk)
 467                         pr_info("%s: %.*s", prefix, len, buf);
 468         } else {
 469                 if (mask & D_EMERG)
 470                         pr_emerg("%sError: %d:%d:(%s:%d:%s()) %.*s", prefix,
 471                                  hdr->ph_pid, hdr->ph_extern_pid, file,
 472                                  hdr->ph_line_num, fn, len, buf);
 473                 else if (mask & D_ERROR)
 474                         pr_err("%sError: %d:%d:(%s:%d:%s()) %.*s", prefix,
 475                                hdr->ph_pid, hdr->ph_extern_pid, file,
 476                                hdr->ph_line_num, fn, len, buf);
 477                 else if (mask & D_WARNING)
 478                         pr_warn("%s: %d:%d:(%s:%d:%s()) %.*s", prefix,
 479                                 hdr->ph_pid, hdr->ph_extern_pid, file,
 480                                 hdr->ph_line_num, fn, len, buf);
 481                 else if (mask & (D_CONSOLE | libcfs_printk))
 482                         pr_info("%s: %.*s", prefix, len, buf);
 483         }
 484
 485         if (mask & D_TTY)
 486                 cfs_tty_write_message(prefix, mask, buf);
 487 }
 488
 489 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
 490                      const char *format, ...)
 491 {
 492         struct cfs_trace_cpu_data *tcd = NULL;
 493         struct ptldebug_header header = {0};
 494         struct cfs_trace_page *tage;
 495         /* string_buf is used only if tcd != NULL, and is always set then */
 496         char *string_buf = NULL;
 497         char *debug_buf;
 498         int known_size;
 499         int needed = 85; /* seeded with average message length */
 500         int max_nob;
 501         va_list ap;
 502         int retry;
 503         int mask = msgdata->msg_mask;
 504         char *file = (char *)msgdata->msg_file;
 505         struct cfs_debug_limit_state *cdls = msgdata->msg_cdls;
 506
 507         if (strchr(file, '/'))
 508                 file = strrchr(file, '/') + 1;
 509
 510         tcd = cfs_trace_get_tcd();
 511
 512         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
 513          * pins us to a particular CPU.  This avoids an smp_processor_id()
 514          * warning on Linux when debugging is enabled.
 515          */
 516         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
 517
 518         if (!tcd)                /* arch may not log in IRQ context */
 519                 goto console;
 520
 521         if (tcd->tcd_cur_pages == 0)
 522                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
 523
 524         if (tcd->tcd_shutting_down) {
 525                 cfs_trace_put_tcd(tcd);
 526                 tcd = NULL;
 527                 goto console;
 528         }
 529
 530         known_size = strlen(file) + 1;
 531         if (msgdata->msg_fn)
 532                 known_size += strlen(msgdata->msg_fn) + 1;
 533
 534         if (libcfs_debug_binary)
 535                 known_size += sizeof(header);
 536
 537         /*
 538          * May perform an additional pass to update 'needed' and increase
 539          * tage buffer size to match vsnprintf reported size required
 540          * On the second pass (retry=1) use vscnprintf [which returns
 541          * number of bytes written not including the terminating nul]
 542          * to clarify `needed` is used as number of bytes written
 543          * for the remainder of this function
 544          */
 545         for (retry = 0; retry < 2; retry++) {
 546                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
 547                 if (!tage) {
 548                         if (needed + known_size > PAGE_SIZE)
 549                                 mask |= D_ERROR;
 550
 551                         cfs_trace_put_tcd(tcd);
 552                         tcd = NULL;
 553                         goto console;
 554                 }
 555
 556                 string_buf = (char *)page_address(tage->page) +
 557                              tage->used + known_size;
 558
 559                 max_nob = PAGE_SIZE - tage->used - known_size;
 560                 if (max_nob <= 0) {
 561                         pr_emerg("LustreError: negative max_nob: %d\n",
 562                                  max_nob);
 563                         mask |= D_ERROR;
 564                         cfs_trace_put_tcd(tcd);
 565                         tcd = NULL;
 566                         goto console;
 567                 }
 568
 569                 va_start(ap, format);
 570                 if (retry)
 571                         needed = vscnprintf(string_buf, max_nob, format, ap);
 572                 else
 573                         needed = vsnprintf(string_buf, max_nob, format, ap);
 574                 va_end(ap);
 575
 576                 if (needed < max_nob) /* well. printing ok.. */
 577                         break;
 578         }
 579
 580         /* `needed` is actual bytes written to string_buf */
 581         if (*(string_buf + needed - 1) != '\n') {
 582                 pr_info("Lustre: format at %s:%d:%s doesn't end in newline\n",
 583                         file, msgdata->msg_line, msgdata->msg_fn);
 584         } else if (mask & D_TTY) {
 585                 /* TTY needs '\r\n' to move carriage to leftmost position */
 586                 if (needed < 2 || *(string_buf + needed - 2) != '\r')
 587                         pr_info("Lustre: format at %s:%d:%s doesn't end in '\\r\\n'\n",
 588                                 file, msgdata->msg_line, msgdata->msg_fn);
 589         }
 590
 591         header.ph_len = known_size + needed;
 592         debug_buf = (char *)page_address(tage->page) + tage->used;
 593
 594         if (libcfs_debug_binary) {
 595                 memcpy(debug_buf, &header, sizeof(header));
 596                 tage->used += sizeof(header);
 597                 debug_buf += sizeof(header);
 598         }
 599
 600         strlcpy(debug_buf, file, PAGE_SIZE - tage->used);
 601         tage->used += strlen(file) + 1;
 602         debug_buf += strlen(file) + 1;
 603
 604         if (msgdata->msg_fn) {
 605                 strlcpy(debug_buf, msgdata->msg_fn, PAGE_SIZE - tage->used);
 606                 tage->used += strlen(msgdata->msg_fn) + 1;
 607                 debug_buf += strlen(msgdata->msg_fn) + 1;
 608         }
 609
 610         __LASSERT(debug_buf == string_buf);
 611
 612         tage->used += needed;
 613         __LASSERT(tage->used <= PAGE_SIZE);
 614
 615 console:
 616         if ((mask & libcfs_printk) == 0) {
 617                 /* no console output requested */
 618                 if (tcd != NULL)
 619                         cfs_trace_put_tcd(tcd);
 620                 return 1;
 621         }
 622
 623         if (cdls != NULL) {
 624                 if (libcfs_console_ratelimit &&
 625                     cdls->cdls_next != 0 &&     /* not first time ever */
 626                     time_before(jiffies, cdls->cdls_next)) {
 627                         /* skipping a console message */
 628                         cdls->cdls_count++;
 629                         if (tcd != NULL)
 630                                 cfs_trace_put_tcd(tcd);
 631                         return 1;
 632                 }
 633
 634                 if (time_after(jiffies, cdls->cdls_next +
 635                                         libcfs_console_max_delay +
 636                                         cfs_time_seconds(10))) {
 637                         /* last timeout was a long time ago */
 638                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 639                 } else {
 640                         cdls->cdls_delay *= libcfs_console_backoff;
 641                 }
 642
 643                 if (cdls->cdls_delay < libcfs_console_min_delay)
 644                         cdls->cdls_delay = libcfs_console_min_delay;
 645                 else if (cdls->cdls_delay > libcfs_console_max_delay)
 646                         cdls->cdls_delay = libcfs_console_max_delay;
 647
 648                 /* ensure cdls_next is never zero after it's been seen */
 649                 cdls->cdls_next = (jiffies + cdls->cdls_delay) | 1;
 650         }
 651
 652         if (tcd) {
 653                 cfs_print_to_console(&header, mask, string_buf, needed, file,
 654                                      msgdata->msg_fn);
 655                 cfs_trace_put_tcd(tcd);
 656         } else {
 657                 string_buf = cfs_trace_get_console_buffer();
 658
 659                 va_start(ap, format);
 660                 needed = vscnprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 661                                     format, ap);
 662                 va_end(ap);
 663
 664                 cfs_print_to_console(&header, mask,
 665                                      string_buf, needed, file, msgdata->msg_fn);
 666
 667                 put_cpu();
 668         }
 669
 670         if (cdls != NULL && cdls->cdls_count != 0) {
 671                 string_buf = cfs_trace_get_console_buffer();
 672
 673                 needed = scnprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
 674                                    "Skipped %d previous similar message%s\n",
 675                                    cdls->cdls_count,
 676                                    (cdls->cdls_count > 1) ? "s" : "");
 677
 678                 /* Do not allow print this to TTY */
 679                 cfs_print_to_console(&header, mask & ~D_TTY, string_buf,
 680                                      needed, file, msgdata->msg_fn);
 681
 682                 put_cpu();
 683                 cdls->cdls_count = 0;
 684         }
 685
 686         return 0;
 687 }
 688 EXPORT_SYMBOL(libcfs_debug_msg);
 689
 690 void
 691 cfs_trace_assertion_failed(const char *str,
 692                            struct libcfs_debug_msg_data *msgdata)
 693 {
 694         struct ptldebug_header hdr;
 695
 696         libcfs_panic_in_progress = 1;
 697         libcfs_catastrophe = 1;
 698         smp_mb();
 699
 700         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
 701
 702         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
 703                              msgdata->msg_file, msgdata->msg_fn);
 704
 705         panic("Lustre debug assertion failure\n");
 706
 707         /* not reached */
 708 }
 709
 710 static void
 711 panic_collect_pages(struct page_collection *pc)
 712 {
 713         /* Do the collect_pages job on a single CPU: assumes that all other
 714          * CPUs have been stopped during a panic.  If this isn't true for some
 715          * arch, this will have to be implemented separately in each arch.  */
 716         int                        i;
 717         int                        j;
 718         struct cfs_trace_cpu_data *tcd;
 719
 720         INIT_LIST_HEAD(&pc->pc_pages);
 721
 722         cfs_tcd_for_each(tcd, i, j) {
 723                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 724                 tcd->tcd_cur_pages = 0;
 725
 726                 if (pc->pc_want_daemon_pages) {
 727                         list_splice_init(&tcd->tcd_daemon_pages,
 728                                                 &pc->pc_pages);
 729                         tcd->tcd_cur_daemon_pages = 0;
 730                 }
 731         }
 732 }
 733
 734 static void collect_pages_on_all_cpus(struct page_collection *pc)
 735 {
 736         struct cfs_trace_cpu_data *tcd;
 737         int i, cpu;
 738
 739         for_each_possible_cpu(cpu) {
 740                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 741                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 742                         tcd->tcd_cur_pages = 0;
 743                         if (pc->pc_want_daemon_pages) {
 744                                 list_splice_init(&tcd->tcd_daemon_pages,
 745                                                         &pc->pc_pages);
 746                                 tcd->tcd_cur_daemon_pages = 0;
 747                         }
 748                 }
 749         }
 750 }
 751
 752 static void collect_pages(struct page_collection *pc)
 753 {
 754         INIT_LIST_HEAD(&pc->pc_pages);
 755
 756         if (libcfs_panic_in_progress)
 757                 panic_collect_pages(pc);
 758         else
 759                 collect_pages_on_all_cpus(pc);
 760 }
 761
 762 static void put_pages_back_on_all_cpus(struct page_collection *pc)
 763 {
 764         struct cfs_trace_cpu_data *tcd;
 765         struct list_head *cur_head;
 766         struct cfs_trace_page *tage;
 767         struct cfs_trace_page *tmp;
 768         int i, cpu;
 769
 770         for_each_possible_cpu(cpu) {
 771                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
 772                         cur_head = tcd->tcd_pages.next;
 773
 774                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
 775                                                  linkage) {
 776
 777                                 __LASSERT_TAGE_INVARIANT(tage);
 778
 779                                 if (tage->cpu != cpu || tage->type != i)
 780                                         continue;
 781
 782                                 cfs_tage_to_tail(tage, cur_head);
 783                                 tcd->tcd_cur_pages++;
 784                         }
 785                 }
 786         }
 787 }
 788
 789 static void put_pages_back(struct page_collection *pc)
 790 {
 791         if (!libcfs_panic_in_progress)
 792                 put_pages_back_on_all_cpus(pc);
 793 }
 794
 795 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 796  * we have a good amount of data at all times for dumping during an LBUG, even
 797  * if we have been steadily writing (and otherwise discarding) pages via the
 798  * debug daemon. */
 799 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 800                                          struct cfs_trace_cpu_data *tcd)
 801 {
 802         struct cfs_trace_page *tage;
 803         struct cfs_trace_page *tmp;
 804
 805         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
 806                 __LASSERT_TAGE_INVARIANT(tage);
 807
 808                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
 809                         continue;
 810
 811                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
 812                 tcd->tcd_cur_daemon_pages++;
 813
 814                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 815                         struct cfs_trace_page *victim;
 816
 817                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 818                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
 819
 820                         __LASSERT_TAGE_INVARIANT(victim);
 821
 822                         list_del(&victim->linkage);
 823                         cfs_tage_free(victim);
 824                         tcd->tcd_cur_daemon_pages--;
 825                 }
 826         }
 827 }
 828
 829 static void put_pages_on_daemon_list(struct page_collection *pc)
 830 {
 831         struct cfs_trace_cpu_data *tcd;
 832         int i, cpu;
 833
 834         for_each_possible_cpu(cpu) {
 835                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
 836                         put_pages_on_tcd_daemon_list(pc, tcd);
 837         }
 838 }
 839
 840 void cfs_trace_debug_print(void)
 841 {
 842         struct page_collection pc;
 843         struct cfs_trace_page *tage;
 844         struct cfs_trace_page *tmp;
 845
 846         pc.pc_want_daemon_pages = 1;
 847         collect_pages(&pc);
 848         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 849                 char *p, *file, *fn;
 850                 struct page *page;
 851
 852                 __LASSERT_TAGE_INVARIANT(tage);
 853
 854                 page = tage->page;
 855                 p = page_address(page);
 856                 while (p < ((char *)page_address(page) + tage->used)) {
 857                         struct ptldebug_header *hdr;
 858                         int len;
 859                         hdr = (void *)p;
 860                         p += sizeof(*hdr);
 861                         file = p;
 862                         p += strlen(file) + 1;
 863                         fn = p;
 864                         p += strlen(fn) + 1;
 865                         len = hdr->ph_len - (int)(p - (char *)hdr);
 866
 867                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
 868
 869                         p += len;
 870                 }
 871
 872                 list_del(&tage->linkage);
 873                 cfs_tage_free(tage);
 874         }
 875 }
 876
 877 int cfs_tracefile_dump_all_pages(char *filename)
 878 {
 879         struct page_collection  pc;
 880         struct file             *filp;
 881         struct cfs_trace_page   *tage;
 882         struct cfs_trace_page   *tmp;
 883         char                    *buf;
 884         int rc;
 885
 886         down_write(&cfs_tracefile_sem);
 887
 888         filp = filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600);
 889         if (IS_ERR(filp)) {
 890                 rc = PTR_ERR(filp);
 891                 filp = NULL;
 892                 pr_err("LustreError: can't open %s for dump: rc = %d\n",
 893                       filename, rc);
 894                 goto out;
 895         }
 896
 897         pc.pc_want_daemon_pages = 1;
 898         collect_pages(&pc);
 899         if (list_empty(&pc.pc_pages)) {
 900                 rc = 0;
 901                 goto close;
 902         }
 903
 904         /* ok, for now, just write the pages.  in the future we'll be building
 905          * iobufs with the pages and calling generic_direct_IO */
 906         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 907
 908                 __LASSERT_TAGE_INVARIANT(tage);
 909
 910                 buf = kmap(tage->page);
 911                 rc = cfs_kernel_write(filp, buf, tage->used, &filp->f_pos);
 912                 kunmap(tage->page);
 913                 if (rc != (int)tage->used) {
 914                         pr_warn("Lustre: wanted to write %u but wrote %d\n",
 915                                 tage->used, rc);
 916                         put_pages_back(&pc);
 917                         __LASSERT(list_empty(&pc.pc_pages));
 918                         break;
 919                 }
 920                 list_del(&tage->linkage);
 921                 cfs_tage_free(tage);
 922         }
 923
 924         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
 925         if (rc)
 926                 pr_err("LustreError: sync returns: rc = %d\n", rc);
 927 close:
 928         filp_close(filp, NULL);
 929 out:
 930         up_write(&cfs_tracefile_sem);
 931         return rc;
 932 }
 933
 934 void cfs_trace_flush_pages(void)
 935 {
 936         struct page_collection pc;
 937         struct cfs_trace_page *tage;
 938         struct cfs_trace_page *tmp;
 939
 940         pc.pc_want_daemon_pages = 1;
 941         collect_pages(&pc);
 942         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
 943
 944                 __LASSERT_TAGE_INVARIANT(tage);
 945
 946                 list_del(&tage->linkage);
 947                 cfs_tage_free(tage);
 948         }
 949 }
 950
 951 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 952                             const char __user *usr_buffer, int usr_buffer_nob)
 953 {
 954         int    nob;
 955
 956         if (usr_buffer_nob > knl_buffer_nob)
 957                 return -EOVERFLOW;
 958
 959         if (copy_from_user(knl_buffer, usr_buffer, usr_buffer_nob))
 960                 return -EFAULT;
 961
 962         nob = strnlen(knl_buffer, usr_buffer_nob);
 963         while (--nob >= 0)                      /* strip trailing whitespace */
 964                 if (!isspace(knl_buffer[nob]))
 965                         break;
 966
 967         if (nob < 0)                            /* empty string */
 968                 return -EINVAL;
 969
 970         if (nob == knl_buffer_nob)              /* no space to terminate */
 971                 return -EOVERFLOW;
 972
 973         knl_buffer[nob + 1] = 0;                /* terminate */
 974         return 0;
 975 }
 976 EXPORT_SYMBOL(cfs_trace_copyin_string);
 977
 978 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
 979                              const char *knl_buffer, char *append)
 980 {
 981         /* NB if 'append' != NULL, it's a single character to append to the
 982          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 983          * terminating zero byte) for sysctl entries */
 984         int   nob = strlen(knl_buffer);
 985
 986         if (nob > usr_buffer_nob)
 987                 nob = usr_buffer_nob;
 988
 989         if (copy_to_user(usr_buffer, knl_buffer, nob))
 990                 return -EFAULT;
 991
 992         if (append != NULL && nob < usr_buffer_nob) {
 993                 if (copy_to_user(usr_buffer + nob, append, 1))
 994                         return -EFAULT;
 995
 996                 nob++;
 997         }
 998
 999         return nob;
1000 }
1001 EXPORT_SYMBOL(cfs_trace_copyout_string);
1002
1003 int cfs_trace_allocate_string_buffer(char **str, int nob)
1004 {
1005         if (nob > 2 * PAGE_SIZE)        /* string must be "sensible" */
1006                 return -EINVAL;
1007
1008         *str = kmalloc(nob, GFP_KERNEL | __GFP_ZERO);
1009         if (*str == NULL)
1010                 return -ENOMEM;
1011
1012         return 0;
1013 }
1014
1015 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
1016 {
1017         char         *str;
1018         int           rc;
1019
1020         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
1021         if (rc != 0)
1022                 return rc;
1023
1024         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
1025                                      usr_str, usr_str_nob);
1026         if (rc != 0)
1027                 goto out;
1028
1029         if (str[0] != '/') {
1030                 rc = -EINVAL;
1031                 goto out;
1032         }
1033         rc = cfs_tracefile_dump_all_pages(str);
1034 out:
1035         kfree(str);
1036         return rc;
1037 }
1038
1039 int cfs_trace_daemon_command(char *str)
1040 {
1041         int       rc = 0;
1042
1043         down_write(&cfs_tracefile_sem);
1044
1045         if (strcmp(str, "stop") == 0) {
1046                 up_write(&cfs_tracefile_sem);
1047                 cfs_trace_stop_thread();
1048                 down_write(&cfs_tracefile_sem);
1049                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
1050
1051         } else if (strncmp(str, "size=", 5) == 0) {
1052                 unsigned long tmp;
1053
1054                 rc = kstrtoul(str + 5, 10, &tmp);
1055                 if (!rc) {
1056                         if (tmp < 10 || tmp > 20480)
1057                                 cfs_tracefile_size = CFS_TRACEFILE_SIZE;
1058                         else
1059                                 cfs_tracefile_size = tmp << 20;
1060                 }
1061         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
1062                 rc = -ENAMETOOLONG;
1063         } else if (str[0] != '/') {
1064                 rc = -EINVAL;
1065         } else {
1066                 strcpy(cfs_tracefile, str);
1067
1068                 pr_info("Lustre: debug daemon will attempt to start writing to %s (%lukB max)\n",
1069                         cfs_tracefile, (long)(cfs_tracefile_size >> 10));
1070
1071                 cfs_trace_start_thread();
1072         }
1073
1074         up_write(&cfs_tracefile_sem);
1075         return rc;
1076 }
1077
1078 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
1079 {
1080         char *str;
1081         int   rc;
1082
1083         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
1084         if (rc != 0)
1085                 return rc;
1086
1087         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
1088                                  usr_str, usr_str_nob);
1089         if (rc == 0)
1090                 rc = cfs_trace_daemon_command(str);
1091
1092         kfree(str);
1093         return rc;
1094 }
1095
1096 int cfs_trace_set_debug_mb(int mb)
1097 {
1098         int i;
1099         int j;
1100         unsigned long pages;
1101         unsigned long total_mb = (cfs_totalram_pages() >> (20 - PAGE_SHIFT));
1102         unsigned long limit = max_t(unsigned long, 512, (total_mb * 4) / 5);
1103         struct cfs_trace_cpu_data *tcd;
1104
1105         if (mb < num_possible_cpus()) {
1106                 pr_warn("Lustre: %d MB is too small for debug buffer size, setting it to %d MB.\n",
1107                         mb, num_possible_cpus());
1108                 mb = num_possible_cpus();
1109         }
1110
1111         if (mb > limit) {
1112                 pr_warn("Lustre: %d MB is too large for debug buffer size, setting it to %lu MB.\n",
1113                         mb, limit);
1114                 mb = limit;
1115         }
1116
1117         mb /= num_possible_cpus();
1118         pages = mb << (20 - PAGE_SHIFT);
1119
1120         down_write(&cfs_tracefile_sem);
1121
1122         cfs_tcd_for_each(tcd, i, j)
1123                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
1124
1125         up_write(&cfs_tracefile_sem);
1126
1127         return mb;
1128 }
1129
1130 int cfs_trace_get_debug_mb(void)
1131 {
1132         int i;
1133         int j;
1134         struct cfs_trace_cpu_data *tcd;
1135         int total_pages = 0;
1136
1137         down_read(&cfs_tracefile_sem);
1138
1139         cfs_tcd_for_each(tcd, i, j)
1140                 total_pages += tcd->tcd_max_pages;
1141
1142         up_read(&cfs_tracefile_sem);
1143
1144         return (total_pages >> (20 - PAGE_SHIFT)) + 1;
1145 }
1146
1147 static int tracefiled(void *arg)
1148 {
1149         struct page_collection pc;
1150         struct tracefiled_ctl *tctl = arg;
1151         struct cfs_trace_page *tage;
1152         struct cfs_trace_page *tmp;
1153         struct file *filp;
1154         char *buf;
1155         int last_loop = 0;
1156         int rc;
1157
1158         /* we're started late enough that we pick up init's fs context */
1159         /* this is so broken in uml?  what on earth is going on? */
1160
1161         complete(&tctl->tctl_start);
1162
1163         while (1) {
1164                 wait_queue_entry_t __wait;
1165
1166                 pc.pc_want_daemon_pages = 0;
1167                 collect_pages(&pc);
1168                 if (list_empty(&pc.pc_pages))
1169                         goto end_loop;
1170
1171                 filp = NULL;
1172                 down_read(&cfs_tracefile_sem);
1173                 if (cfs_tracefile[0] != 0) {
1174                         filp = filp_open(cfs_tracefile,
1175                                          O_CREAT | O_RDWR | O_LARGEFILE,
1176                                          0600);
1177                         if (IS_ERR(filp)) {
1178                                 rc = PTR_ERR(filp);
1179                                 filp = NULL;
1180                                 pr_warn("Lustre: couldn't open %s: rc = %d\n",
1181                                         cfs_tracefile, rc);
1182                         }
1183                 }
1184                 up_read(&cfs_tracefile_sem);
1185                 if (filp == NULL) {
1186                         put_pages_on_daemon_list(&pc);
1187                         __LASSERT(list_empty(&pc.pc_pages));
1188                         goto end_loop;
1189                 }
1190
1191                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
1192                         struct dentry *de = file_dentry(filp);
1193                         static loff_t f_pos;
1194
1195                         __LASSERT_TAGE_INVARIANT(tage);
1196
1197                         if (f_pos >= (off_t)cfs_tracefile_size)
1198                                 f_pos = 0;
1199                         else if (f_pos > i_size_read(de->d_inode))
1200                                 f_pos = i_size_read(de->d_inode);
1201
1202                         buf = kmap(tage->page);
1203                         rc = cfs_kernel_write(filp, buf, tage->used, &f_pos);
1204                         kunmap(tage->page);
1205                         if (rc != (int)tage->used) {
1206                                 pr_warn("Lustre: wanted to write %u but wrote %d\n",
1207                                         tage->used, rc);
1208                                 put_pages_back(&pc);
1209                                 __LASSERT(list_empty(&pc.pc_pages));
1210                                 break;
1211                         }
1212                 }
1213
1214                 filp_close(filp, NULL);
1215                 put_pages_on_daemon_list(&pc);
1216                 if (!list_empty(&pc.pc_pages)) {
1217                         int i;
1218
1219                         pr_alert("Lustre: trace pages aren't empty\n");
1220                         pr_err("Lustre: total cpus(%d): ", num_possible_cpus());
1221                         for (i = 0; i < num_possible_cpus(); i++)
1222                                 if (cpu_online(i))
1223                                         pr_cont("%d(on) ", i);
1224                                 else
1225                                         pr_cont("%d(off) ", i);
1226                         pr_cont("\n");
1227
1228                         i = 0;
1229                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1230                                                  linkage)
1231                                 pr_err("Lustre: page %d belongs to cpu %d\n",
1232                                        ++i, tage->cpu);
1233                         pr_err("Lustre: There are %d pages unwritten\n", i);
1234                 }
1235                 __LASSERT(list_empty(&pc.pc_pages));
1236 end_loop:
1237                 if (atomic_read(&tctl->tctl_shutdown)) {
1238                         if (last_loop == 0) {
1239                                 last_loop = 1;
1240                                 continue;
1241                         } else {
1242                                 break;
1243                         }
1244                 }
1245                 init_waitqueue_entry(&__wait, current);
1246                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1247                 schedule_timeout_interruptible(cfs_time_seconds(1));
1248                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1249         }
1250         complete(&tctl->tctl_stop);
1251         return 0;
1252 }
1253
1254 int cfs_trace_start_thread(void)
1255 {
1256         struct tracefiled_ctl *tctl = &trace_tctl;
1257         int rc = 0;
1258
1259         mutex_lock(&cfs_trace_thread_mutex);
1260         if (thread_running)
1261                 goto out;
1262
1263         init_completion(&tctl->tctl_start);
1264         init_completion(&tctl->tctl_stop);
1265         init_waitqueue_head(&tctl->tctl_waitq);
1266         atomic_set(&tctl->tctl_shutdown, 0);
1267
1268         if (IS_ERR(kthread_run(tracefiled, tctl, "ktracefiled"))) {
1269                 rc = -ECHILD;
1270                 goto out;
1271         }
1272
1273         wait_for_completion(&tctl->tctl_start);
1274         thread_running = 1;
1275 out:
1276         mutex_unlock(&cfs_trace_thread_mutex);
1277         return rc;
1278 }
1279
1280 void cfs_trace_stop_thread(void)
1281 {
1282         struct tracefiled_ctl *tctl = &trace_tctl;
1283
1284         mutex_lock(&cfs_trace_thread_mutex);
1285         if (thread_running) {
1286                 pr_info("Lustre: shutting down debug daemon thread...\n");
1287                 atomic_set(&tctl->tctl_shutdown, 1);
1288                 wait_for_completion(&tctl->tctl_stop);
1289                 thread_running = 0;
1290         }
1291         mutex_unlock(&cfs_trace_thread_mutex);
1292 }
1293
1294 int cfs_tracefile_init(int max_pages)
1295 {
1296         struct cfs_trace_cpu_data *tcd;
1297         int     i;
1298         int     j;
1299         int     rc;
1300         int     factor;
1301
1302         rc = cfs_tracefile_init_arch();
1303         if (rc != 0)
1304                 return rc;
1305
1306         cfs_tcd_for_each(tcd, i, j) {
1307                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1308                 factor = tcd->tcd_pages_factor;
1309                 INIT_LIST_HEAD(&tcd->tcd_pages);
1310                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1311                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1312                 tcd->tcd_cur_pages = 0;
1313                 tcd->tcd_cur_stock_pages = 0;
1314                 tcd->tcd_cur_daemon_pages = 0;
1315                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1316                 LASSERT(tcd->tcd_max_pages > 0);
1317                 tcd->tcd_shutting_down = 0;
1318         }
1319         return 0;
1320 }
1321
1322 static void trace_cleanup_on_all_cpus(void)
1323 {
1324         struct cfs_trace_cpu_data *tcd;
1325         struct cfs_trace_page *tage;
1326         struct cfs_trace_page *tmp;
1327         int i, cpu;
1328
1329         for_each_possible_cpu(cpu) {
1330                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1331                         tcd->tcd_shutting_down = 1;
1332
1333                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1334                                 __LASSERT_TAGE_INVARIANT(tage);
1335
1336                                 list_del(&tage->linkage);
1337                                 cfs_tage_free(tage);
1338                         }
1339                         tcd->tcd_cur_pages = 0;
1340                 }
1341         }
1342 }
1343
1344 static void cfs_trace_cleanup(void)
1345 {
1346         struct page_collection pc;
1347
1348         INIT_LIST_HEAD(&pc.pc_pages);
1349
1350         trace_cleanup_on_all_cpus();
1351
1352         cfs_tracefile_fini_arch();
1353 }
1354
1355 void cfs_tracefile_exit(void)
1356 {
1357         cfs_trace_stop_thread();
1358         cfs_trace_cleanup();
1359 }