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