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