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