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