blktrace.c

   1 /*
   2  * block queue tracing application
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * Rewrite to have a single thread per CPU (managing all devices on that CPU)
   8  *      Alan D. Brunelle <alan.brunelle@hp.com> - January 2009
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License as published by
  12  *  the Free Software Foundation; either version 2 of the License, or
  13  *  (at your option) any later version.
  14  *
  15  *  This program is distributed in the hope that it will be useful,
  16  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  *  GNU General Public License for more details.
  19  *
  20  *  You should have received a copy of the GNU General Public License
  21  *  along with this program; if not, write to the Free Software
  22  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  23  *
  24  */
  25
  26 #include <errno.h>
  27 #include <stdarg.h>
  28 #include <stdio.h>
  29 #include <stdlib.h>
  30 #include <string.h>
  31 #include <fcntl.h>
  32 #include <getopt.h>
  33 #include <sched.h>
  34 #include <unistd.h>
  35 #include <poll.h>
  36 #include <signal.h>
  37 #include <pthread.h>
  38 #include <locale.h>
  39 #include <sys/ioctl.h>
  40 #include <sys/types.h>
  41 #include <sys/stat.h>
  42 #include <sys/vfs.h>
  43 #include <sys/mman.h>
  44 #include <sys/param.h>
  45 #include <sys/time.h>
  46 #include <sys/resource.h>
  47 #include <sys/socket.h>
  48 #include <netinet/in.h>
  49 #include <arpa/inet.h>
  50 #include <netdb.h>
  51 #include <sys/sendfile.h>
  52
  53 #include "btt/list.h"
  54 #include "blktrace.h"
  55
  56 /*
  57  * You may want to increase this even more, if you are logging at a high
  58  * rate and see skipped/missed events
  59  */
  60 #define BUF_SIZE                (512 * 1024)
  61 #define BUF_NR                  (4)
  62
  63 #define FILE_VBUF_SIZE          (128 * 1024)
  64
  65 #define DEBUGFS_TYPE            (0x64626720)
  66 #define TRACE_NET_PORT          (8462)
  67
  68 enum {
  69         Net_none = 0,
  70         Net_server,
  71         Net_client,
  72 };
  73
  74 enum thread_status {
  75         Th_running,
  76         Th_leaving,
  77         Th_error
  78 };
  79
  80 /*
  81  * Generic stats collected: nevents can be _roughly_ estimated by data_read
  82  * (discounting pdu...)
  83  *
  84  * These fields are updated w/ pdc_dr_update & pdc_nev_update below.
  85  */
  86 struct pdc_stats {
  87         unsigned long long data_read;
  88         unsigned long long nevents;
  89 };
  90
  91 struct devpath {
  92         struct list_head head;
  93         char *path;                     /* path to device special file */
  94         char *buts_name;                /* name returned from bt kernel code */
  95         struct pdc_stats *stats;
  96         int fd, ncpus;
  97         unsigned long long drops;
  98
  99         /*
 100          * For piped output only:
 101          *
 102          * Each tracer will have a tracer_devpath_head that it will add new
 103          * data onto. It's list is protected above (tracer_devpath_head.mutex)
 104          * and it will signal the processing thread using the dp_cond,
 105          * dp_mutex & dp_entries variables above.
 106          */
 107         struct tracer_devpath_head *heads;
 108
 109         /*
 110          * For network server mode only:
 111          */
 112         struct cl_host *ch;
 113         u32 cl_id;
 114         time_t cl_connect_time;
 115         struct io_info *ios;
 116 };
 117
 118 /*
 119  * For piped output to stdout we will have each tracer thread (one per dev)
 120  * tack buffers read from the relay queues on a per-device list.
 121  *
 122  * The main thread will then collect trace buffers from each of lists in turn.
 123  *
 124  * We will use a mutex to guard each of the trace_buf list. The tracers
 125  * can then signal the main thread using <dp_cond,dp_mutex> and
 126  * dp_entries. (When dp_entries is 0, and a tracer adds an entry it will
 127  * signal. When dp_entries is 0, the main thread will wait for that condition
 128  * to be signalled.)
 129  *
 130  * adb: It may be better just to have a large buffer per tracer per dev,
 131  * and then use it as a ring-buffer. This would certainly cut down a lot
 132  * of malloc/free thrashing, at the cost of more memory movements (potentially).
 133  */
 134 struct trace_buf {
 135         struct list_head head;
 136         struct devpath *dpp;
 137         void *buf;
 138         int cpu, len;
 139 };
 140
 141 struct tracer_devpath_head {
 142         pthread_mutex_t mutex;
 143         struct list_head head;
 144         struct trace_buf *prev;
 145 };
 146
 147 /*
 148  * Used to handle the mmap() interfaces for output file (containing traces)
 149  */
 150 struct mmap_info {
 151         void *fs_buf;
 152         unsigned long long fs_size, fs_max_size, fs_off, fs_buf_len;
 153         unsigned long buf_size, buf_nr;
 154         int pagesize;
 155 };
 156
 157 /*
 158  * Each thread doing work on a (client) side of blktrace will have one
 159  * of these. The ios array contains input/output information, pfds holds
 160  * poll() data. The volatile's provide flags to/from the main executing
 161  * thread.
 162  */
 163 struct tracer {
 164         struct list_head head;
 165         struct io_info *ios;
 166         struct pollfd *pfds;
 167         pthread_t thread;
 168         int cpu, nios;
 169         volatile int status, is_done;
 170 };
 171
 172 /*
 173  * networking stuff follows. we include a magic number so we know whether
 174  * to endianness convert or not.
 175  *
 176  * The len field is overloaded:
 177  *      0 - Indicates an "open" - allowing the server to set up for a dev/cpu
 178  *      1 - Indicates a "close" - Shut down connection orderly
 179  *
 180  * The cpu field is overloaded on close: it will contain the number of drops.
 181  */
 182 struct blktrace_net_hdr {
 183         u32 magic;              /* same as trace magic */
 184         char buts_name[32];     /* trace name */
 185         u32 cpu;                /* for which cpu */
 186         u32 max_cpus;
 187         u32 len;                /* length of following trace data */
 188         u32 cl_id;              /* id for set of client per-cpu connections */
 189         u32 buf_size;           /* client buf_size for this trace  */
 190         u32 buf_nr;             /* client buf_nr for this trace  */
 191         u32 page_size;          /* client page_size for this trace  */
 192 };
 193
 194 /*
 195  * Each host encountered has one of these. The head is used to link this
 196  * on to the network server's ch_list. Connections associated with this
 197  * host are linked on conn_list, and any devices traced on that host
 198  * are connected on the devpaths list.
 199  */
 200 struct cl_host {
 201         struct list_head head;
 202         struct list_head conn_list;
 203         struct list_head devpaths;
 204         struct net_server_s *ns;
 205         char *hostname;
 206         struct in_addr cl_in_addr;
 207         int connects, ndevs, cl_opens;
 208 };
 209
 210 /*
 211  * Each connection (client to server socket ('fd')) has one of these. A
 212  * back reference to the host ('ch'), and lists headers (for the host
 213  * list, and the network server conn_list) are also included.
 214  */
 215 struct cl_conn {
 216         struct list_head ch_head, ns_head;
 217         struct cl_host *ch;
 218         int fd, ncpus;
 219         time_t connect_time;
 220 };
 221
 222 /*
 223  * The network server requires some poll structures to be maintained -
 224  * one per conection currently on conn_list. The nchs/ch_list values
 225  * are for each host connected to this server. The addr field is used
 226  * for scratch as new connections are established.
 227  */
 228 struct net_server_s {
 229         struct list_head conn_list;
 230         struct list_head ch_list;
 231         struct pollfd *pfds;
 232         int listen_fd, connects, nchs;
 233         struct sockaddr_in addr;
 234 };
 235
 236 /*
 237  * This structure is (generically) used to providide information
 238  * for a read-to-write set of values.
 239  *
 240  * ifn & ifd represent input information
 241  *
 242  * ofn, ofd, ofp, obuf & mmap_info are used for output file (optionally).
 243  */
 244 struct io_info {
 245         struct devpath *dpp;
 246         FILE *ofp;
 247         char *obuf;
 248         struct cl_conn *nc;     /* Server network connection */
 249
 250         /*
 251          * mmap controlled output files
 252          */
 253         struct mmap_info mmap_info;
 254
 255         /*
 256          * Client network fields
 257          */
 258         unsigned int ready;
 259         unsigned long long data_queued;
 260
 261         /*
 262          * Input/output file descriptors & names
 263          */
 264         int ifd, ofd;
 265         char ifn[MAXPATHLEN + 64];
 266         char ofn[MAXPATHLEN + 64];
 267 };
 268
 269 static char blktrace_version[] = "2.0.0";
 270
 271 /*
 272  * Linkage to blktrace helper routines (trace conversions)
 273  */
 274 int data_is_native = -1;
 275
 276 static int ndevs;
 277 static int max_cpus;
 278 static int ncpus;
 279 static cpu_set_t *online_cpus;
 280 static int pagesize;
 281 static int act_mask = ~0U;
 282 static int kill_running_trace;
 283 static int stop_watch;
 284 static int piped_output;
 285
 286 static char *debugfs_path = "/sys/kernel/debug";
 287 static char *output_name;
 288 static char *output_dir;
 289
 290 static unsigned long buf_size = BUF_SIZE;
 291 static unsigned long buf_nr = BUF_NR;
 292
 293 static FILE *pfp;
 294
 295 static LIST_HEAD(devpaths);
 296 static LIST_HEAD(tracers);
 297
 298 static volatile int done;
 299
 300 /*
 301  * tracer threads add entries, the main thread takes them off and processes
 302  * them. These protect the dp_entries variable.
 303  */
 304 static pthread_cond_t dp_cond = PTHREAD_COND_INITIALIZER;
 305 static pthread_mutex_t dp_mutex = PTHREAD_MUTEX_INITIALIZER;
 306 static volatile int dp_entries;
 307
 308 /*
 309  * These synchronize master / thread interactions.
 310  */
 311 static pthread_cond_t mt_cond = PTHREAD_COND_INITIALIZER;
 312 static pthread_mutex_t mt_mutex = PTHREAD_MUTEX_INITIALIZER;
 313 static volatile int nthreads_running;
 314 static volatile int nthreads_leaving;
 315 static volatile int nthreads_error;
 316 static volatile int tracers_run;
 317
 318 /*
 319  * network cmd line params
 320  */
 321 static struct sockaddr_in hostname_addr;
 322 static char hostname[MAXHOSTNAMELEN];
 323 static int net_port = TRACE_NET_PORT;
 324 static int net_use_sendfile = 1;
 325 static int net_mode;
 326 static int *cl_fds;
 327
 328 static int (*handle_pfds)(struct tracer *, int, int);
 329 static int (*handle_list)(struct tracer_devpath_head *, struct list_head *);
 330
 331 #define S_OPTS  "d:a:A:r:o:kw:vVb:n:D:lh:p:sI:"
 332 static struct option l_opts[] = {
 333         {
 334                 .name = "dev",
 335                 .has_arg = required_argument,
 336                 .flag = NULL,
 337                 .val = 'd'
 338         },
 339         {
 340                 .name = "input-devs",
 341                 .has_arg = required_argument,
 342                 .flag = NULL,
 343                 .val = 'I'
 344         },
 345         {
 346                 .name = "act-mask",
 347                 .has_arg = required_argument,
 348                 .flag = NULL,
 349                 .val = 'a'
 350         },
 351         {
 352                 .name = "set-mask",
 353                 .has_arg = required_argument,
 354                 .flag = NULL,
 355                 .val = 'A'
 356         },
 357         {
 358                 .name = "relay",
 359                 .has_arg = required_argument,
 360                 .flag = NULL,
 361                 .val = 'r'
 362         },
 363         {
 364                 .name = "output",
 365                 .has_arg = required_argument,
 366                 .flag = NULL,
 367                 .val = 'o'
 368         },
 369         {
 370                 .name = "kill",
 371                 .has_arg = no_argument,
 372                 .flag = NULL,
 373                 .val = 'k'
 374         },
 375         {
 376                 .name = "stopwatch",
 377                 .has_arg = required_argument,
 378                 .flag = NULL,
 379                 .val = 'w'
 380         },
 381         {
 382                 .name = "version",
 383                 .has_arg = no_argument,
 384                 .flag = NULL,
 385                 .val = 'v'
 386         },
 387         {
 388                 .name = "version",
 389                 .has_arg = no_argument,
 390                 .flag = NULL,
 391                 .val = 'V'
 392         },
 393         {
 394                 .name = "buffer-size",
 395                 .has_arg = required_argument,
 396                 .flag = NULL,
 397                 .val = 'b'
 398         },
 399         {
 400                 .name = "num-sub-buffers",
 401                 .has_arg = required_argument,
 402                 .flag = NULL,
 403                 .val = 'n'
 404         },
 405         {
 406                 .name = "output-dir",
 407                 .has_arg = required_argument,
 408                 .flag = NULL,
 409                 .val = 'D'
 410         },
 411         {
 412                 .name = "listen",
 413                 .has_arg = no_argument,
 414                 .flag = NULL,
 415                 .val = 'l'
 416         },
 417         {
 418                 .name = "host",
 419                 .has_arg = required_argument,
 420                 .flag = NULL,
 421                 .val = 'h'
 422         },
 423         {
 424                 .name = "port",
 425                 .has_arg = required_argument,
 426                 .flag = NULL,
 427                 .val = 'p'
 428         },
 429         {
 430                 .name = "no-sendfile",
 431                 .has_arg = no_argument,
 432                 .flag = NULL,
 433                 .val = 's'
 434         },
 435         {
 436                 .name = NULL,
 437         }
 438 };
 439
 440 static char usage_str[] = "\n\n" \
 441         "-d <dev>             | --dev=<dev>\n" \
 442         "[ -r <debugfs path>  | --relay=<debugfs path> ]\n" \
 443         "[ -o <file>          | --output=<file>]\n" \
 444         "[ -D <dir>           | --output-dir=<dir>\n" \
 445         "[ -w <time>          | --stopwatch=<time>]\n" \
 446         "[ -a <action field>  | --act-mask=<action field>]\n" \
 447         "[ -A <action mask>   | --set-mask=<action mask>]\n" \
 448         "[ -b <size>          | --buffer-size]\n" \
 449         "[ -n <number>        | --num-sub-buffers=<number>]\n" \
 450         "[ -l                 | --listen]\n" \
 451         "[ -h <hostname>      | --host=<hostname>]\n" \
 452         "[ -p <port number>   | --port=<port number>]\n" \
 453         "[ -s                 | --no-sendfile]\n" \
 454         "[ -I <devs file>     | --input-devs=<devs file>]\n" \
 455         "[ -v <version>       | --version]\n" \
 456         "[ -V <version>       | --version]\n" \
 457
 458         "\t-d Use specified device. May also be given last after options\n" \
 459         "\t-r Path to mounted debugfs, defaults to /sys/kernel/debug\n" \
 460         "\t-o File(s) to send output to\n" \
 461         "\t-D Directory to prepend to output file names\n" \
 462         "\t-w Stop after defined time, in seconds\n" \
 463         "\t-a Only trace specified actions. See documentation\n" \
 464         "\t-A Give trace mask as a single value. See documentation\n" \
 465         "\t-b Sub buffer size in KiB (default 512)\n" \
 466         "\t-n Number of sub buffers (default 4)\n" \
 467         "\t-l Run in network listen mode (blktrace server)\n" \
 468         "\t-h Run in network client mode, connecting to the given host\n" \
 469         "\t-p Network port to use (default 8462)\n" \
 470         "\t-s Make the network client NOT use sendfile() to transfer data\n" \
 471         "\t-I Add devices found in <devs file>\n" \
 472         "\t-v Print program version info\n" \
 473         "\t-V Print program version info\n\n";
 474
 475 static void clear_events(struct pollfd *pfd)
 476 {
 477         pfd->events = 0;
 478         pfd->revents = 0;
 479 }
 480
 481 static inline int net_client_use_sendfile(void)
 482 {
 483         return net_mode == Net_client && net_use_sendfile;
 484 }
 485
 486 static inline int net_client_use_send(void)
 487 {
 488         return net_mode == Net_client && !net_use_sendfile;
 489 }
 490
 491 static inline int use_tracer_devpaths(void)
 492 {
 493         return piped_output || net_client_use_send();
 494 }
 495
 496 static inline int in_addr_eq(struct in_addr a, struct in_addr b)
 497 {
 498         return a.s_addr == b.s_addr;
 499 }
 500
 501 static inline void pdc_dr_update(struct devpath *dpp, int cpu, int data_read)
 502 {
 503         dpp->stats[cpu].data_read += data_read;
 504 }
 505
 506 static inline void pdc_nev_update(struct devpath *dpp, int cpu, int nevents)
 507 {
 508         dpp->stats[cpu].nevents += nevents;
 509 }
 510
 511 static void show_usage(char *prog)
 512 {
 513         fprintf(stderr, "Usage: %s %s", prog, usage_str);
 514 }
 515
 516 /*
 517  * Create a timespec 'msec' milliseconds into the future
 518  */
 519 static inline void make_timespec(struct timespec *tsp, long delta_msec)
 520 {
 521         struct timeval now;
 522
 523         gettimeofday(&now, NULL);
 524         tsp->tv_sec = now.tv_sec;
 525         tsp->tv_nsec = 1000L * now.tv_usec;
 526
 527         tsp->tv_nsec += (delta_msec * 1000000L);
 528         if (tsp->tv_nsec > 1000000000L) {
 529                 long secs = tsp->tv_nsec / 1000000000L;
 530
 531                 tsp->tv_sec += secs;
 532                 tsp->tv_nsec -= (secs * 1000000000L);
 533         }
 534 }
 535
 536 /*
 537  * Add a timer to ensure wait ends
 538  */
 539 static void t_pthread_cond_wait(pthread_cond_t *cond, pthread_mutex_t *mutex)
 540 {
 541         struct timespec ts;
 542
 543         make_timespec(&ts, 50);
 544         pthread_cond_timedwait(cond, mutex, &ts);
 545 }
 546
 547 static void unblock_tracers(void)
 548 {
 549         pthread_mutex_lock(&mt_mutex);
 550         tracers_run = 1;
 551         pthread_cond_broadcast(&mt_cond);
 552         pthread_mutex_unlock(&mt_mutex);
 553 }
 554
 555 static void tracer_wait_unblock(struct tracer *tp)
 556 {
 557         pthread_mutex_lock(&mt_mutex);
 558         while (!tp->is_done && !tracers_run)
 559                 pthread_cond_wait(&mt_cond, &mt_mutex);
 560         pthread_mutex_unlock(&mt_mutex);
 561 }
 562
 563 static void tracer_signal_ready(struct tracer *tp,
 564                                 enum thread_status th_status,
 565                                 int status)
 566 {
 567         pthread_mutex_lock(&mt_mutex);
 568         tp->status = status;
 569
 570         if (th_status == Th_running)
 571                 nthreads_running++;
 572         else if (th_status == Th_error)
 573                 nthreads_error++;
 574         else
 575                 nthreads_leaving++;
 576
 577         pthread_cond_signal(&mt_cond);
 578         pthread_mutex_unlock(&mt_mutex);
 579 }
 580
 581 static void wait_tracers_ready(int ncpus_started)
 582 {
 583         pthread_mutex_lock(&mt_mutex);
 584         while ((nthreads_running + nthreads_error) < ncpus_started)
 585                 t_pthread_cond_wait(&mt_cond, &mt_mutex);
 586         pthread_mutex_unlock(&mt_mutex);
 587 }
 588
 589 static void wait_tracers_leaving(void)
 590 {
 591         pthread_mutex_lock(&mt_mutex);
 592         while (nthreads_leaving < nthreads_running)
 593                 t_pthread_cond_wait(&mt_cond, &mt_mutex);
 594         pthread_mutex_unlock(&mt_mutex);
 595 }
 596
 597 static void init_mmap_info(struct mmap_info *mip)
 598 {
 599         mip->buf_size = buf_size;
 600         mip->buf_nr = buf_nr;
 601         mip->pagesize = pagesize;
 602 }
 603
 604 static void net_close_connection(int *fd)
 605 {
 606         shutdown(*fd, SHUT_RDWR);
 607         close(*fd);
 608         *fd = -1;
 609 }
 610
 611 static void dpp_free(struct devpath *dpp)
 612 {
 613         if (dpp->stats)
 614                 free(dpp->stats);
 615         if (dpp->ios)
 616                 free(dpp->ios);
 617         if (dpp->path)
 618                 free(dpp->path);
 619         if (dpp->buts_name)
 620                 free(dpp->buts_name);
 621         free(dpp);
 622 }
 623
 624 static int lock_on_cpu(int cpu)
 625 {
 626         cpu_set_t * cpu_mask;
 627         size_t size;
 628
 629         cpu_mask = CPU_ALLOC(max_cpus);
 630         size = CPU_ALLOC_SIZE(max_cpus);
 631
 632         CPU_ZERO_S(size, cpu_mask);
 633         CPU_SET_S(cpu, size, cpu_mask);
 634         if (sched_setaffinity(0, size, cpu_mask) < 0) {
 635                 CPU_FREE(cpu_mask);
 636                 return errno;
 637         }
 638
 639         CPU_FREE(cpu_mask);
 640         return 0;
 641 }
 642
 643 static int increase_limit(int resource, rlim_t increase)
 644 {
 645         struct rlimit rlim;
 646         int save_errno = errno;
 647
 648         if (!getrlimit(resource, &rlim)) {
 649                 rlim.rlim_cur += increase;
 650                 if (rlim.rlim_cur >= rlim.rlim_max)
 651                         rlim.rlim_max = rlim.rlim_cur + increase;
 652
 653                 if (!setrlimit(resource, &rlim))
 654                         return 1;
 655         }
 656
 657         errno = save_errno;
 658         return 0;
 659 }
 660
 661 static int handle_open_failure(void)
 662 {
 663         if (errno == ENFILE || errno == EMFILE)
 664                 return increase_limit(RLIMIT_NOFILE, 16);
 665         return 0;
 666 }
 667
 668 static int handle_mem_failure(size_t length)
 669 {
 670         if (errno == ENFILE)
 671                 return handle_open_failure();
 672         else if (errno == ENOMEM)
 673                 return increase_limit(RLIMIT_MEMLOCK, 2 * length);
 674         return 0;
 675 }
 676
 677 static FILE *my_fopen(const char *path, const char *mode)
 678 {
 679         FILE *fp;
 680
 681         do {
 682                 fp = fopen(path, mode);
 683         } while (fp == NULL && handle_open_failure());
 684
 685         return fp;
 686 }
 687
 688 static int my_open(const char *path, int flags)
 689 {
 690         int fd;
 691
 692         do {
 693                 fd = open(path, flags);
 694         } while (fd < 0 && handle_open_failure());
 695
 696         return fd;
 697 }
 698
 699 static int my_socket(int domain, int type, int protocol)
 700 {
 701         int fd;
 702
 703         do {
 704                 fd = socket(domain, type, protocol);
 705         } while (fd < 0 && handle_open_failure());
 706
 707         return fd;
 708 }
 709
 710 static int my_accept(int sockfd, struct sockaddr *addr, socklen_t *addrlen)
 711 {
 712         int fd;
 713
 714         do {
 715                 fd = accept(sockfd, addr, addrlen);
 716         } while (fd < 0 && handle_open_failure());
 717
 718         return fd;
 719 }
 720
 721 static void *my_mmap(void *addr, size_t length, int prot, int flags, int fd,
 722                      off_t offset)
 723 {
 724         void *new;
 725
 726         do {
 727                 new = mmap(addr, length, prot, flags, fd, offset);
 728         } while (new == MAP_FAILED && handle_mem_failure(length));
 729
 730         return new;
 731 }
 732
 733 static int my_mlock(struct tracer *tp,
 734                     const void *addr, size_t len)
 735 {
 736         int ret, retry = 0;
 737
 738         do {
 739                 ret = mlock(addr, len);
 740                 if ((retry >= 10) && tp && tp->is_done)
 741                         break;
 742                 retry++;
 743         } while (ret < 0 && handle_mem_failure(len));
 744
 745         return ret;
 746 }
 747
 748 static int setup_mmap(int fd, unsigned int maxlen,
 749                       struct mmap_info *mip,
 750                       struct tracer *tp)
 751 {
 752         if (mip->fs_off + maxlen > mip->fs_buf_len) {
 753                 unsigned long nr = max(16, mip->buf_nr);
 754
 755                 if (mip->fs_buf) {
 756                         munlock(mip->fs_buf, mip->fs_buf_len);
 757                         munmap(mip->fs_buf, mip->fs_buf_len);
 758                         mip->fs_buf = NULL;
 759                 }
 760
 761                 mip->fs_off = mip->fs_size & (mip->pagesize - 1);
 762                 mip->fs_buf_len = (nr * mip->buf_size) - mip->fs_off;
 763                 mip->fs_max_size += mip->fs_buf_len;
 764
 765                 if (ftruncate(fd, mip->fs_max_size) < 0) {
 766                         perror("setup_mmap: ftruncate");
 767                         return 1;
 768                 }
 769
 770                 mip->fs_buf = my_mmap(NULL, mip->fs_buf_len, PROT_WRITE,
 771                                       MAP_SHARED, fd,
 772                                       mip->fs_size - mip->fs_off);
 773                 if (mip->fs_buf == MAP_FAILED) {
 774                         perror("setup_mmap: mmap");
 775                         return 1;
 776                 }
 777                 if (my_mlock(tp, mip->fs_buf, mip->fs_buf_len) < 0) {
 778                         perror("setup_mlock: mlock");
 779                         return 1;
 780                 }
 781         }
 782
 783         return 0;
 784 }
 785
 786 static int __stop_trace(int fd)
 787 {
 788         /*
 789          * Should be stopped, don't complain if it isn't
 790          */
 791         ioctl(fd, BLKTRACESTOP);
 792         return ioctl(fd, BLKTRACETEARDOWN);
 793 }
 794
 795 static int write_data(char *buf, int len)
 796 {
 797         int ret;
 798
 799 rewrite:
 800         ret = fwrite(buf, len, 1, pfp);
 801         if (ferror(pfp) || ret != 1) {
 802                 if (errno == EINTR) {
 803                         clearerr(pfp);
 804                         goto rewrite;
 805                 }
 806
 807                 if (!piped_output || (errno != EPIPE && errno != EBADF)) {
 808                         fprintf(stderr, "write(%d) failed: %d/%s\n",
 809                                 len, errno, strerror(errno));
 810                 }
 811                 goto err;
 812         }
 813
 814         fflush(pfp);
 815         return 0;
 816
 817 err:
 818         clearerr(pfp);
 819         return 1;
 820 }
 821
 822 /*
 823  * Returns the number of bytes read (successfully)
 824  */
 825 static int __net_recv_data(int fd, void *buf, unsigned int len)
 826 {
 827         unsigned int bytes_left = len;
 828
 829         while (bytes_left && !done) {
 830                 int ret = recv(fd, buf, bytes_left, MSG_WAITALL);
 831
 832                 if (ret == 0)
 833                         break;
 834                 else if (ret < 0) {
 835                         if (errno == EAGAIN) {
 836                                 usleep(50);
 837                                 continue;
 838                         }
 839                         perror("server: net_recv_data: recv failed");
 840                         break;
 841                 } else {
 842                         buf += ret;
 843                         bytes_left -= ret;
 844                 }
 845         }
 846
 847         return len - bytes_left;
 848 }
 849
 850 static int net_recv_data(int fd, void *buf, unsigned int len)
 851 {
 852         return __net_recv_data(fd, buf, len);
 853 }
 854
 855 /*
 856  * Returns number of bytes written
 857  */
 858 static int net_send_data(int fd, void *buf, unsigned int buf_len)
 859 {
 860         int ret;
 861         unsigned int bytes_left = buf_len;
 862
 863         while (bytes_left) {
 864                 ret = send(fd, buf, bytes_left, 0);
 865                 if (ret < 0) {
 866                         perror("send");
 867                         break;
 868                 }
 869
 870                 buf += ret;
 871                 bytes_left -= ret;
 872         }
 873
 874         return buf_len - bytes_left;
 875 }
 876
 877 static int net_send_header(int fd, int cpu, char *buts_name, int len)
 878 {
 879         struct blktrace_net_hdr hdr;
 880
 881         memset(&hdr, 0, sizeof(hdr));
 882
 883         hdr.magic = BLK_IO_TRACE_MAGIC;
 884         memset(hdr.buts_name, 0, sizeof(hdr.buts_name));
 885         strncpy(hdr.buts_name, buts_name, sizeof(hdr.buts_name));
 886         hdr.buts_name[sizeof(hdr.buts_name) - 1] = '\0';
 887         hdr.cpu = cpu;
 888         hdr.max_cpus = max_cpus;
 889         hdr.len = len;
 890         hdr.cl_id = getpid();
 891         hdr.buf_size = buf_size;
 892         hdr.buf_nr = buf_nr;
 893         hdr.page_size = pagesize;
 894
 895         return net_send_data(fd, &hdr, sizeof(hdr)) != sizeof(hdr);
 896 }
 897
 898 static void net_send_open_close(int fd, int cpu, char *buts_name, int len)
 899 {
 900         struct blktrace_net_hdr ret_hdr;
 901
 902         net_send_header(fd, cpu, buts_name, len);
 903         net_recv_data(fd, &ret_hdr, sizeof(ret_hdr));
 904 }
 905
 906 static void net_send_open(int fd, int cpu, char *buts_name)
 907 {
 908         net_send_open_close(fd, cpu, buts_name, 0);
 909 }
 910
 911 static void net_send_close(int fd, char *buts_name, int drops)
 912 {
 913         /*
 914          * Overload CPU w/ number of drops
 915          *
 916          * XXX: Need to clear/set done around call - done=1 (which
 917          * is true here) stops reads from happening... :-(
 918          */
 919         done = 0;
 920         net_send_open_close(fd, drops, buts_name, 1);
 921         done = 1;
 922 }
 923
 924 static void ack_open_close(int fd, char *buts_name)
 925 {
 926         net_send_header(fd, 0, buts_name, 2);
 927 }
 928
 929 static void net_send_drops(int fd)
 930 {
 931         struct list_head *p;
 932
 933         __list_for_each(p, &devpaths) {
 934                 struct devpath *dpp = list_entry(p, struct devpath, head);
 935
 936                 net_send_close(fd, dpp->buts_name, dpp->drops);
 937         }
 938 }
 939
 940 /*
 941  * Returns:
 942  *       0: "EOF"
 943  *       1: OK
 944  *      -1: Error
 945  */
 946 static int net_get_header(struct cl_conn *nc, struct blktrace_net_hdr *bnh)
 947 {
 948         int bytes_read;
 949         int fl = fcntl(nc->fd, F_GETFL);
 950
 951         fcntl(nc->fd, F_SETFL, fl | O_NONBLOCK);
 952         bytes_read = __net_recv_data(nc->fd, bnh, sizeof(*bnh));
 953         fcntl(nc->fd, F_SETFL, fl & ~O_NONBLOCK);
 954
 955         if (bytes_read == sizeof(*bnh))
 956                 return 1;
 957         else if (bytes_read == 0)
 958                 return 0;
 959         else
 960                 return -1;
 961 }
 962
 963 static int net_setup_addr(void)
 964 {
 965         struct sockaddr_in *addr = &hostname_addr;
 966
 967         memset(addr, 0, sizeof(*addr));
 968         addr->sin_family = AF_INET;
 969         addr->sin_port = htons(net_port);
 970
 971         if (inet_aton(hostname, &addr->sin_addr) != 1) {
 972                 struct hostent *hent;
 973 retry:
 974                 hent = gethostbyname(hostname);
 975                 if (!hent) {
 976                         if (h_errno == TRY_AGAIN) {
 977                                 usleep(100);
 978                                 goto retry;
 979                         } else if (h_errno == NO_RECOVERY) {
 980                                 fprintf(stderr, "gethostbyname(%s)"
 981                                         "non-recoverable error encountered\n",
 982                                         hostname);
 983                         } else {
 984                                 /*
 985                                  * HOST_NOT_FOUND, NO_ADDRESS or NO_DATA
 986                                  */
 987                                 fprintf(stderr, "Host %s not found\n",
 988                                         hostname);
 989                         }
 990                         return 1;
 991                 }
 992
 993                 memcpy(&addr->sin_addr, hent->h_addr, 4);
 994                 memset(hostname, 0, sizeof(hostname));
 995                 strncpy(hostname, hent->h_name, sizeof(hostname));
 996                 hostname[sizeof(hostname) - 1] = '\0';
 997         }
 998
 999         return 0;
1000 }
1001
1002 static int net_setup_client(void)
1003 {
1004         int fd;
1005         struct sockaddr_in *addr = &hostname_addr;
1006
1007         fd = my_socket(AF_INET, SOCK_STREAM, 0);
1008         if (fd < 0) {
1009                 perror("client: socket");
1010                 return -1;
1011         }
1012
1013         if (connect(fd, (struct sockaddr *)addr, sizeof(*addr)) < 0) {
1014                 if (errno == ECONNREFUSED)
1015                         fprintf(stderr,
1016                                 "\nclient: Connection to %s refused, "
1017                                 "perhaps the server is not started?\n\n",
1018                                 hostname);
1019                 else
1020                         perror("client: connect");
1021
1022                 close(fd);
1023                 return -1;
1024         }
1025
1026         return fd;
1027 }
1028
1029 static int open_client_connections(void)
1030 {
1031         int cpu;
1032         size_t alloc_size = CPU_ALLOC_SIZE(max_cpus);
1033
1034         cl_fds = calloc(ncpus, sizeof(*cl_fds));
1035         for (cpu = 0; cpu < max_cpus; cpu++) {
1036                 if (!CPU_ISSET_S(cpu, alloc_size, online_cpus))
1037                         continue;
1038                 cl_fds[cpu] = net_setup_client();
1039                 if (cl_fds[cpu] < 0)
1040                         goto err;
1041         }
1042         return 0;
1043
1044 err:
1045         while (cpu > 0)
1046                 close(cl_fds[cpu--]);
1047         free(cl_fds);
1048         return 1;
1049 }
1050
1051 static void close_client_connections(void)
1052 {
1053         if (cl_fds) {
1054                 int cpu, *fdp;
1055                 size_t alloc_size = CPU_ALLOC_SIZE(max_cpus);
1056
1057                 for (cpu = 0, fdp = cl_fds; cpu < max_cpus; cpu++, fdp++) {
1058                         if (!CPU_ISSET_S(cpu, alloc_size, online_cpus))
1059                                 continue;
1060                         if (*fdp >= 0) {
1061                                 net_send_drops(*fdp);
1062                                 net_close_connection(fdp);
1063                         }
1064                 }
1065                 free(cl_fds);
1066         }
1067 }
1068
1069 static void setup_buts(void)
1070 {
1071         struct list_head *p;
1072
1073         __list_for_each(p, &devpaths) {
1074                 struct blk_user_trace_setup buts;
1075                 struct devpath *dpp = list_entry(p, struct devpath, head);
1076
1077                 memset(&buts, 0, sizeof(buts));
1078                 buts.buf_size = buf_size;
1079                 buts.buf_nr = buf_nr;
1080                 buts.act_mask = act_mask;
1081
1082                 if (ioctl(dpp->fd, BLKTRACESETUP, &buts) >= 0) {
1083                         dpp->ncpus = max_cpus;
1084                         dpp->buts_name = strdup(buts.name);
1085                         if (dpp->stats)
1086                                 free(dpp->stats);
1087                         dpp->stats = calloc(dpp->ncpus, sizeof(*dpp->stats));
1088                         memset(dpp->stats, 0, dpp->ncpus * sizeof(*dpp->stats));
1089                 } else
1090                         fprintf(stderr, "BLKTRACESETUP(2) %s failed: %d/%s\n",
1091                                 dpp->path, errno, strerror(errno));
1092         }
1093 }
1094
1095 static void start_buts(void)
1096 {
1097         struct list_head *p;
1098
1099         __list_for_each(p, &devpaths) {
1100                 struct devpath *dpp = list_entry(p, struct devpath, head);
1101
1102                 if (ioctl(dpp->fd, BLKTRACESTART) < 0) {
1103                         fprintf(stderr, "BLKTRACESTART %s failed: %d/%s\n",
1104                                 dpp->path, errno, strerror(errno));
1105                 }
1106         }
1107 }
1108
1109 static int get_drops(struct devpath *dpp)
1110 {
1111         int fd, drops = 0;
1112         char fn[MAXPATHLEN + 64], tmp[256];
1113
1114         snprintf(fn, sizeof(fn), "%s/block/%s/dropped", debugfs_path,
1115                  dpp->buts_name);
1116
1117         fd = my_open(fn, O_RDONLY);
1118         if (fd < 0) {
1119                 /*
1120                  * This may be ok: the kernel may not support
1121                  * dropped counts.
1122                  */
1123                 if (errno != ENOENT)
1124                         fprintf(stderr, "Could not open %s: %d/%s\n",
1125                                 fn, errno, strerror(errno));
1126                 return 0;
1127         } else if (read(fd, tmp, sizeof(tmp)) < 0) {
1128                 fprintf(stderr, "Could not read %s: %d/%s\n",
1129                         fn, errno, strerror(errno));
1130         } else
1131                 drops = atoi(tmp);
1132         close(fd);
1133
1134         return drops;
1135 }
1136
1137 static void get_all_drops(void)
1138 {
1139         struct list_head *p;
1140
1141         __list_for_each(p, &devpaths) {
1142                 struct devpath *dpp = list_entry(p, struct devpath, head);
1143
1144                 dpp->drops = get_drops(dpp);
1145         }
1146 }
1147
1148 static inline struct trace_buf *alloc_trace_buf(int cpu, int bufsize)
1149 {
1150         struct trace_buf *tbp;
1151
1152         tbp = malloc(sizeof(*tbp) + bufsize);
1153         INIT_LIST_HEAD(&tbp->head);
1154         tbp->len = 0;
1155         tbp->buf = (void *)(tbp + 1);
1156         tbp->cpu = cpu;
1157         tbp->dpp = NULL;        /* Will be set when tbp is added */
1158
1159         return tbp;
1160 }
1161
1162 static void free_tracer_heads(struct devpath *dpp)
1163 {
1164         int cpu;
1165         struct tracer_devpath_head *hd;
1166
1167         for (cpu = 0, hd = dpp->heads; cpu < max_cpus; cpu++, hd++) {
1168                 if (hd->prev)
1169                         free(hd->prev);
1170
1171                 pthread_mutex_destroy(&hd->mutex);
1172         }
1173         free(dpp->heads);
1174 }
1175
1176 static int setup_tracer_devpaths(void)
1177 {
1178         struct list_head *p;
1179
1180         if (net_client_use_send())
1181                 if (open_client_connections())
1182                         return 1;
1183
1184         __list_for_each(p, &devpaths) {
1185                 int cpu;
1186                 struct tracer_devpath_head *hd;
1187                 struct devpath *dpp = list_entry(p, struct devpath, head);
1188
1189                 dpp->heads = calloc(max_cpus, sizeof(struct tracer_devpath_head));
1190                 for (cpu = 0, hd = dpp->heads; cpu < max_cpus; cpu++, hd++) {
1191                         INIT_LIST_HEAD(&hd->head);
1192                         pthread_mutex_init(&hd->mutex, NULL);
1193                         hd->prev = NULL;
1194                 }
1195         }
1196
1197         return 0;
1198 }
1199
1200 static inline void add_trace_buf(struct devpath *dpp, int cpu,
1201                                                 struct trace_buf **tbpp)
1202 {
1203         struct trace_buf *tbp = *tbpp;
1204         struct tracer_devpath_head *hd = &dpp->heads[cpu];
1205
1206         tbp->dpp = dpp;
1207
1208         pthread_mutex_lock(&hd->mutex);
1209         list_add_tail(&tbp->head, &hd->head);
1210         pthread_mutex_unlock(&hd->mutex);
1211
1212         *tbpp = alloc_trace_buf(cpu, buf_size);
1213 }
1214
1215 static inline void incr_entries(int entries_handled)
1216 {
1217         pthread_mutex_lock(&dp_mutex);
1218         if (dp_entries == 0)
1219                 pthread_cond_signal(&dp_cond);
1220         dp_entries += entries_handled;
1221         pthread_mutex_unlock(&dp_mutex);
1222 }
1223
1224 static void decr_entries(int handled)
1225 {
1226         pthread_mutex_lock(&dp_mutex);
1227         dp_entries -= handled;
1228         pthread_mutex_unlock(&dp_mutex);
1229 }
1230
1231 static int wait_empty_entries(void)
1232 {
1233         pthread_mutex_lock(&dp_mutex);
1234         while (!done && dp_entries == 0)
1235                 t_pthread_cond_wait(&dp_cond, &dp_mutex);
1236         pthread_mutex_unlock(&dp_mutex);
1237
1238         return !done;
1239 }
1240
1241 static int add_devpath(char *path)
1242 {
1243         int fd;
1244         struct devpath *dpp;
1245         struct list_head *p;
1246
1247         /*
1248          * Verify device is not duplicated
1249          */
1250         __list_for_each(p, &devpaths) {
1251                struct devpath *tmp = list_entry(p, struct devpath, head);
1252                if (!strcmp(tmp->path, path))
1253                         return 0;
1254         }
1255         /*
1256          * Verify device is valid before going too far
1257          */
1258         fd = my_open(path, O_RDONLY | O_NONBLOCK);
1259         if (fd < 0) {
1260                 fprintf(stderr, "Invalid path %s specified: %d/%s\n",
1261                         path, errno, strerror(errno));
1262                 return 1;
1263         }
1264
1265         dpp = malloc(sizeof(*dpp));
1266         memset(dpp, 0, sizeof(*dpp));
1267         dpp->path = strdup(path);
1268         dpp->fd = fd;
1269         ndevs++;
1270         list_add_tail(&dpp->head, &devpaths);
1271
1272         return 0;
1273 }
1274
1275 static void rel_devpaths(void)
1276 {
1277         struct list_head *p, *q;
1278
1279         list_for_each_safe(p, q, &devpaths) {
1280                 struct devpath *dpp = list_entry(p, struct devpath, head);
1281
1282                 list_del(&dpp->head);
1283                 __stop_trace(dpp->fd);
1284                 close(dpp->fd);
1285
1286                 if (dpp->heads)
1287                         free_tracer_heads(dpp);
1288
1289                 dpp_free(dpp);
1290                 ndevs--;
1291         }
1292 }
1293
1294 static int flush_subbuf_net(struct trace_buf *tbp)
1295 {
1296         int fd = cl_fds[tbp->cpu];
1297         struct devpath *dpp = tbp->dpp;
1298
1299         if (net_send_header(fd, tbp->cpu, dpp->buts_name, tbp->len))
1300                 return 1;
1301         else if (net_send_data(fd, tbp->buf, tbp->len) != tbp->len)
1302                 return 1;
1303
1304         return 0;
1305 }
1306
1307 static int
1308 handle_list_net(__attribute__((__unused__))struct tracer_devpath_head *hd,
1309                 struct list_head *list)
1310 {
1311         struct trace_buf *tbp;
1312         struct list_head *p, *q;
1313         int entries_handled = 0;
1314
1315         list_for_each_safe(p, q, list) {
1316                 tbp = list_entry(p, struct trace_buf, head);
1317
1318                 list_del(&tbp->head);
1319                 entries_handled++;
1320
1321                 if (cl_fds[tbp->cpu] >= 0) {
1322                         if (flush_subbuf_net(tbp)) {
1323                                 close(cl_fds[tbp->cpu]);
1324                                 cl_fds[tbp->cpu] = -1;
1325                         }
1326                 }
1327
1328                 free(tbp);
1329         }
1330
1331         return entries_handled;
1332 }
1333
1334 /*
1335  * Tack 'tbp's buf onto the tail of 'prev's buf
1336  */
1337 static struct trace_buf *tb_combine(struct trace_buf *prev,
1338                                     struct trace_buf *tbp)
1339 {
1340         unsigned long tot_len;
1341
1342         tot_len = prev->len + tbp->len;
1343         if (tot_len > buf_size) {
1344                 /*
1345                  * tbp->head isn't connected (it was 'prev'
1346                  * so it had been taken off of the list
1347                  * before). Therefore, we can realloc
1348                  * the whole structures, as the other fields
1349                  * are "static".
1350                  */
1351                 prev = realloc(prev, sizeof(*prev) + tot_len);
1352                 prev->buf = (void *)(prev + 1);
1353         }
1354
1355         memcpy(prev->buf + prev->len, tbp->buf, tbp->len);
1356         prev->len = tot_len;
1357
1358         free(tbp);
1359         return prev;
1360 }
1361
1362 static int handle_list_file(struct tracer_devpath_head *hd,
1363                             struct list_head *list)
1364 {
1365         int off, t_len, nevents;
1366         struct blk_io_trace *t;
1367         struct list_head *p, *q;
1368         int entries_handled = 0;
1369         struct trace_buf *tbp, *prev;
1370
1371         prev = hd->prev;
1372         list_for_each_safe(p, q, list) {
1373                 tbp = list_entry(p, struct trace_buf, head);
1374                 list_del(&tbp->head);
1375                 entries_handled++;
1376
1377                 /*
1378                  * If there was some leftover before, tack this new
1379                  * entry onto the tail of the previous one.
1380                  */
1381                 if (prev)
1382                         tbp = tb_combine(prev, tbp);
1383
1384                 /*
1385                  * See how many whole traces there are - send them
1386                  * all out in one go.
1387                  */
1388                 off = 0;
1389                 nevents = 0;
1390                 while (off + (int)sizeof(*t) <= tbp->len) {
1391                         t = (struct blk_io_trace *)(tbp->buf + off);
1392                         t_len = sizeof(*t) + t->pdu_len;
1393                         if (off + t_len > tbp->len)
1394                                 break;
1395
1396                         off += t_len;
1397                         nevents++;
1398                 }
1399                 if (nevents)
1400                         pdc_nev_update(tbp->dpp, tbp->cpu, nevents);
1401
1402                 /*
1403                  * Write any full set of traces, any remaining data is kept
1404                  * for the next pass.
1405                  */
1406                 if (off) {
1407                         if (write_data(tbp->buf, off) || off == tbp->len) {
1408                                 free(tbp);
1409                                 prev = NULL;
1410                         }
1411                         else {
1412                                 /*
1413                                  * Move valid data to beginning of buffer
1414                                  */
1415                                 tbp->len -= off;
1416                                 memmove(tbp->buf, tbp->buf + off, tbp->len);
1417                                 prev = tbp;
1418                         }
1419                 } else
1420                         prev = tbp;
1421         }
1422         hd->prev = prev;
1423
1424         return entries_handled;
1425 }
1426
1427 static void __process_trace_bufs(void)
1428 {
1429         int cpu;
1430         struct list_head *p;
1431         struct list_head list;
1432         int handled = 0;
1433
1434         __list_for_each(p, &devpaths) {
1435                 struct devpath *dpp = list_entry(p, struct devpath, head);
1436                 struct tracer_devpath_head *hd = dpp->heads;
1437
1438                 for (cpu = 0; cpu < max_cpus; cpu++, hd++) {
1439                         pthread_mutex_lock(&hd->mutex);
1440                         if (list_empty(&hd->head)) {
1441                                 pthread_mutex_unlock(&hd->mutex);
1442                                 continue;
1443                         }
1444
1445                         list_replace_init(&hd->head, &list);
1446                         pthread_mutex_unlock(&hd->mutex);
1447
1448                         handled += handle_list(hd, &list);
1449                 }
1450         }
1451
1452         if (handled)
1453                 decr_entries(handled);
1454 }
1455
1456 static void process_trace_bufs(void)
1457 {
1458         while (wait_empty_entries())
1459                 __process_trace_bufs();
1460 }
1461
1462 static void clean_trace_bufs(void)
1463 {
1464         /*
1465          * No mutex needed here: we're only reading from the lists,
1466          * tracers are done
1467          */
1468         while (dp_entries)
1469                 __process_trace_bufs();
1470 }
1471
1472 static inline void read_err(int cpu, char *ifn)
1473 {
1474         if (errno != EAGAIN)
1475                 fprintf(stderr, "Thread %d failed read of %s: %d/%s\n",
1476                         cpu, ifn, errno, strerror(errno));
1477 }
1478
1479 static int net_sendfile(struct io_info *iop)
1480 {
1481         int ret;
1482
1483         ret = sendfile(iop->ofd, iop->ifd, NULL, iop->ready);
1484         if (ret < 0) {
1485                 perror("sendfile");
1486                 return 1;
1487         } else if (ret < (int)iop->ready) {
1488                 fprintf(stderr, "short sendfile send (%d of %d)\n",
1489                         ret, iop->ready);
1490                 return 1;
1491         }
1492
1493         return 0;
1494 }
1495
1496 static inline int net_sendfile_data(struct tracer *tp, struct io_info *iop)
1497 {
1498         struct devpath *dpp = iop->dpp;
1499
1500         if (net_send_header(iop->ofd, tp->cpu, dpp->buts_name, iop->ready))
1501                 return 1;
1502         return net_sendfile(iop);
1503 }
1504
1505 static int fill_ofname(struct io_info *iop, int cpu)
1506 {
1507         int len;
1508         struct stat sb;
1509         char *dst = iop->ofn;
1510
1511         if (output_dir)
1512                 len = snprintf(iop->ofn, sizeof(iop->ofn), "%s/", output_dir);
1513         else
1514                 len = snprintf(iop->ofn, sizeof(iop->ofn), "./");
1515
1516         if (net_mode == Net_server) {
1517                 struct cl_conn *nc = iop->nc;
1518
1519                 len += sprintf(dst + len, "%s-", nc->ch->hostname);
1520                 len += strftime(dst + len, 64, "%F-%T/",
1521                                 gmtime(&iop->dpp->cl_connect_time));
1522         }
1523
1524         if (stat(iop->ofn, &sb) < 0) {
1525                 if (errno != ENOENT) {
1526                         fprintf(stderr,
1527                                 "Destination dir %s stat failed: %d/%s\n",
1528                                 iop->ofn, errno, strerror(errno));
1529                         return 1;
1530                 }
1531                 /*
1532                  * There is no synchronization between multiple threads
1533                  * trying to create the directory at once.  It's harmless
1534                  * to let them try, so just detect the problem and move on.
1535                  */
1536                 if (mkdir(iop->ofn, 0755) < 0 && errno != EEXIST) {
1537                         fprintf(stderr,
1538                                 "Destination dir %s can't be made: %d/%s\n",
1539                                 iop->ofn, errno, strerror(errno));
1540                         return 1;
1541                 }
1542         }
1543
1544         if (output_name)
1545                 snprintf(iop->ofn + len, sizeof(iop->ofn), "%s.blktrace.%d",
1546                          output_name, cpu);
1547         else
1548                 snprintf(iop->ofn + len, sizeof(iop->ofn), "%s.blktrace.%d",
1549                          iop->dpp->buts_name, cpu);
1550
1551         return 0;
1552 }
1553
1554 static int set_vbuf(struct io_info *iop, int mode, size_t size)
1555 {
1556         iop->obuf = malloc(size);
1557         if (setvbuf(iop->ofp, iop->obuf, mode, size) < 0) {
1558                 fprintf(stderr, "setvbuf(%s, %d) failed: %d/%s\n",
1559                         iop->dpp->path, (int)size, errno,
1560                         strerror(errno));
1561                 free(iop->obuf);
1562                 return 1;
1563         }
1564
1565         return 0;
1566 }
1567
1568 static int iop_open(struct io_info *iop, int cpu)
1569 {
1570         iop->ofd = -1;
1571         if (fill_ofname(iop, cpu))
1572                 return 1;
1573
1574         iop->ofp = my_fopen(iop->ofn, "w+");
1575         if (iop->ofp == NULL) {
1576                 fprintf(stderr, "Open output file %s failed: %d/%s\n",
1577                         iop->ofn, errno, strerror(errno));
1578                 return 1;
1579         }
1580
1581         if (set_vbuf(iop, _IOLBF, FILE_VBUF_SIZE)) {
1582                 fprintf(stderr, "set_vbuf for file %s failed: %d/%s\n",
1583                         iop->ofn, errno, strerror(errno));
1584                 fclose(iop->ofp);
1585                 return 1;
1586         }
1587
1588         iop->ofd = fileno(iop->ofp);
1589         return 0;
1590 }
1591
1592 static void close_iop(struct io_info *iop)
1593 {
1594         struct mmap_info *mip = &iop->mmap_info;
1595
1596         if (mip->fs_buf)
1597                 munmap(mip->fs_buf, mip->fs_buf_len);
1598
1599         if (!piped_output) {
1600                 if (ftruncate(fileno(iop->ofp), mip->fs_size) < 0) {
1601                         fprintf(stderr,
1602                                 "Ignoring err: ftruncate(%s): %d/%s\n",
1603                                 iop->ofn, errno, strerror(errno));
1604                 }
1605         }
1606
1607         if (iop->ofp)
1608                 fclose(iop->ofp);
1609         if (iop->obuf)
1610                 free(iop->obuf);
1611 }
1612
1613 static void close_ios(struct tracer *tp)
1614 {
1615         while (tp->nios > 0) {
1616                 struct io_info *iop = &tp->ios[--tp->nios];
1617
1618                 iop->dpp->drops = get_drops(iop->dpp);
1619                 if (iop->ifd >= 0)
1620                         close(iop->ifd);
1621
1622                 if (iop->ofp)
1623                         close_iop(iop);
1624                 else if (iop->ofd >= 0) {
1625                         struct devpath *dpp = iop->dpp;
1626
1627                         net_send_close(iop->ofd, dpp->buts_name, dpp->drops);
1628                         net_close_connection(&iop->ofd);
1629                 }
1630         }
1631
1632         free(tp->ios);
1633         free(tp->pfds);
1634 }
1635
1636 static int open_ios(struct tracer *tp)
1637 {
1638         struct pollfd *pfd;
1639         struct io_info *iop;
1640         struct list_head *p;
1641
1642         tp->ios = calloc(ndevs, sizeof(struct io_info));
1643         memset(tp->ios, 0, ndevs * sizeof(struct io_info));
1644
1645         tp->pfds = calloc(ndevs, sizeof(struct pollfd));
1646         memset(tp->pfds, 0, ndevs * sizeof(struct pollfd));
1647
1648         tp->nios = 0;
1649         iop = tp->ios;
1650         pfd = tp->pfds;
1651         __list_for_each(p, &devpaths) {
1652                 struct devpath *dpp = list_entry(p, struct devpath, head);
1653
1654                 iop->dpp = dpp;
1655                 iop->ofd = -1;
1656                 snprintf(iop->ifn, sizeof(iop->ifn), "%s/block/%s/trace%d",
1657                         debugfs_path, dpp->buts_name, tp->cpu);
1658
1659                 iop->ifd = my_open(iop->ifn, O_RDONLY | O_NONBLOCK);
1660                 if (iop->ifd < 0) {
1661                         fprintf(stderr, "Thread %d failed open %s: %d/%s\n",
1662                                 tp->cpu, iop->ifn, errno, strerror(errno));
1663                         return 1;
1664                 }
1665
1666                 init_mmap_info(&iop->mmap_info);
1667
1668                 pfd->fd = iop->ifd;
1669                 pfd->events = POLLIN;
1670
1671                 if (piped_output)
1672                         ;
1673                 else if (net_client_use_sendfile()) {
1674                         iop->ofd = net_setup_client();
1675                         if (iop->ofd < 0)
1676                                 goto err;
1677                         net_send_open(iop->ofd, tp->cpu, dpp->buts_name);
1678                 } else if (net_mode == Net_none) {
1679                         if (iop_open(iop, tp->cpu))
1680                                 goto err;
1681                 } else {
1682                         /*
1683                          * This ensures that the server knows about all
1684                          * connections & devices before _any_ closes
1685                          */
1686                         net_send_open(cl_fds[tp->cpu], tp->cpu, dpp->buts_name);
1687                 }
1688
1689                 pfd++;
1690                 iop++;
1691                 tp->nios++;
1692         }
1693
1694         return 0;
1695
1696 err:
1697         close(iop->ifd);        /* tp->nios _not_ bumped */
1698         close_ios(tp);
1699         return 1;
1700 }
1701
1702 static int handle_pfds_file(struct tracer *tp, int nevs, int force_read)
1703 {
1704         struct mmap_info *mip;
1705         int i, ret, nentries = 0;
1706         struct pollfd *pfd = tp->pfds;
1707         struct io_info *iop = tp->ios;
1708
1709         for (i = 0; nevs > 0 && i < ndevs; i++, pfd++, iop++) {
1710                 if (pfd->revents & POLLIN || force_read) {
1711                         mip = &iop->mmap_info;
1712
1713                         ret = setup_mmap(iop->ofd, buf_size, mip, tp);
1714                         if (ret < 0) {
1715                                 pfd->events = 0;
1716                                 break;
1717                         }
1718
1719                         ret = read(iop->ifd, mip->fs_buf + mip->fs_off,
1720                                    buf_size);
1721                         if (ret > 0) {
1722                                 pdc_dr_update(iop->dpp, tp->cpu, ret);
1723                                 mip->fs_size += ret;
1724                                 mip->fs_off += ret;
1725                                 nentries++;
1726                         } else if (ret == 0) {
1727                                 /*
1728                                  * Short reads after we're done stop us
1729                                  * from trying reads.
1730                                  */
1731                                 if (tp->is_done)
1732                                         clear_events(pfd);
1733                         } else {
1734                                 read_err(tp->cpu, iop->ifn);
1735                                 if (errno != EAGAIN || tp->is_done)
1736                                         clear_events(pfd);
1737                         }
1738                         nevs--;
1739                 }
1740         }
1741
1742         return nentries;
1743 }
1744
1745 static int handle_pfds_netclient(struct tracer *tp, int nevs, int force_read)
1746 {
1747         struct stat sb;
1748         int i, nentries = 0;
1749         struct pollfd *pfd = tp->pfds;
1750         struct io_info *iop = tp->ios;
1751
1752         for (i = 0; i < ndevs; i++, pfd++, iop++) {
1753                 if (pfd->revents & POLLIN || force_read) {
1754                         if (fstat(iop->ifd, &sb) < 0) {
1755                                 perror(iop->ifn);
1756                                 pfd->events = 0;
1757                         } else if (sb.st_size > (off_t)iop->data_queued) {
1758                                 iop->ready = sb.st_size - iop->data_queued;
1759                                 iop->data_queued = sb.st_size;
1760
1761                                 if (!net_sendfile_data(tp, iop)) {
1762                                         pdc_dr_update(iop->dpp, tp->cpu,
1763                                                       iop->ready);
1764                                         nentries++;
1765                                 } else
1766                                         clear_events(pfd);
1767                         }
1768                         if (--nevs == 0)
1769                                 break;
1770                 }
1771         }
1772
1773         if (nentries)
1774                 incr_entries(nentries);
1775
1776         return nentries;
1777 }
1778
1779 static int handle_pfds_entries(struct tracer *tp, int nevs, int force_read)
1780 {
1781         int i, nentries = 0;
1782         struct trace_buf *tbp;
1783         struct pollfd *pfd = tp->pfds;
1784         struct io_info *iop = tp->ios;
1785
1786         tbp = alloc_trace_buf(tp->cpu, buf_size);
1787         for (i = 0; i < ndevs; i++, pfd++, iop++) {
1788                 if (pfd->revents & POLLIN || force_read) {
1789                         tbp->len = read(iop->ifd, tbp->buf, buf_size);
1790                         if (tbp->len > 0) {
1791                                 pdc_dr_update(iop->dpp, tp->cpu, tbp->len);
1792                                 add_trace_buf(iop->dpp, tp->cpu, &tbp);
1793                                 nentries++;
1794                         } else if (tbp->len == 0) {
1795                                 /*
1796                                  * Short reads after we're done stop us
1797                                  * from trying reads.
1798                                  */
1799                                 if (tp->is_done)
1800                                         clear_events(pfd);
1801                         } else {
1802                                 read_err(tp->cpu, iop->ifn);
1803                                 if (errno != EAGAIN || tp->is_done)
1804                                         clear_events(pfd);
1805                         }
1806                         if (!piped_output && --nevs == 0)
1807                                 break;
1808                 }
1809         }
1810         free(tbp);
1811
1812         if (nentries)
1813                 incr_entries(nentries);
1814
1815         return nentries;
1816 }
1817
1818 static void *thread_main(void *arg)
1819 {
1820         int ret, ndone, to_val;
1821         struct tracer *tp = arg;
1822
1823         ret = lock_on_cpu(tp->cpu);
1824         if (ret)
1825                 goto err;
1826
1827         ret = open_ios(tp);
1828         if (ret)
1829                 goto err;
1830
1831         if (piped_output)
1832                 to_val = 50;            /* Frequent partial handles */
1833         else
1834                 to_val = 500;           /* 1/2 second intervals */
1835
1836
1837         tracer_signal_ready(tp, Th_running, 0);
1838         tracer_wait_unblock(tp);
1839
1840         while (!tp->is_done) {
1841                 ndone = poll(tp->pfds, ndevs, to_val);
1842                 if (ndone || piped_output)
1843                         (void)handle_pfds(tp, ndone, piped_output);
1844                 else if (ndone < 0 && errno != EINTR)
1845                         fprintf(stderr, "Thread %d poll failed: %d/%s\n",
1846                                 tp->cpu, errno, strerror(errno));
1847         }
1848
1849         /*
1850          * Trace is stopped, pull data until we get a short read
1851          */
1852         while (handle_pfds(tp, ndevs, 1) > 0)
1853                 ;
1854
1855         close_ios(tp);
1856         tracer_signal_ready(tp, Th_leaving, 0);
1857         return NULL;
1858
1859 err:
1860         tracer_signal_ready(tp, Th_error, ret);
1861         return NULL;
1862 }
1863
1864 static int start_tracer(int cpu)
1865 {
1866         struct tracer *tp;
1867
1868         tp = malloc(sizeof(*tp));
1869         memset(tp, 0, sizeof(*tp));
1870
1871         INIT_LIST_HEAD(&tp->head);
1872         tp->status = 0;
1873         tp->cpu = cpu;
1874
1875         if (pthread_create(&tp->thread, NULL, thread_main, tp)) {
1876                 fprintf(stderr, "FAILED to start thread on CPU %d: %d/%s\n",
1877                         cpu, errno, strerror(errno));
1878                 free(tp);
1879                 return 1;
1880         }
1881
1882         list_add_tail(&tp->head, &tracers);
1883         return 0;
1884 }
1885
1886 static void start_tracers(void)
1887 {
1888         int cpu, started = 0;
1889         struct list_head *p;
1890         size_t alloc_size = CPU_ALLOC_SIZE(max_cpus);
1891
1892         for (cpu = 0; cpu < max_cpus; cpu++) {
1893                 if (!CPU_ISSET_S(cpu, alloc_size, online_cpus))
1894                         continue;
1895                 if (start_tracer(cpu))
1896                         break;
1897                 started++;
1898         }
1899
1900         wait_tracers_ready(started);
1901
1902         __list_for_each(p, &tracers) {
1903                 struct tracer *tp = list_entry(p, struct tracer, head);
1904                 if (tp->status)
1905                         fprintf(stderr,
1906                                 "FAILED to start thread on CPU %d: %d/%s\n",
1907                                 tp->cpu, tp->status, strerror(tp->status));
1908         }
1909 }
1910
1911 static void stop_tracers(void)
1912 {
1913         struct list_head *p;
1914
1915         /*
1916          * Stop the tracing - makes the tracer threads clean up quicker.
1917          */
1918         __list_for_each(p, &devpaths) {
1919                 struct devpath *dpp = list_entry(p, struct devpath, head);
1920                 (void)ioctl(dpp->fd, BLKTRACESTOP);
1921         }
1922
1923         /*
1924          * Tell each tracer to quit
1925          */
1926         __list_for_each(p, &tracers) {
1927                 struct tracer *tp = list_entry(p, struct tracer, head);
1928                 tp->is_done = 1;
1929         }
1930         pthread_cond_broadcast(&mt_cond);
1931 }
1932
1933 static void del_tracers(void)
1934 {
1935         struct list_head *p, *q;
1936
1937         list_for_each_safe(p, q, &tracers) {
1938                 struct tracer *tp = list_entry(p, struct tracer, head);
1939
1940                 list_del(&tp->head);
1941                 free(tp);
1942         }
1943 }
1944
1945 static void wait_tracers(void)
1946 {
1947         struct list_head *p;
1948
1949         if (use_tracer_devpaths())
1950                 process_trace_bufs();
1951
1952         wait_tracers_leaving();
1953
1954         __list_for_each(p, &tracers) {
1955                 int ret;
1956                 struct tracer *tp = list_entry(p, struct tracer, head);
1957
1958                 ret = pthread_join(tp->thread, NULL);
1959                 if (ret)
1960                         fprintf(stderr, "Thread join %d failed %d\n",
1961                                 tp->cpu, ret);
1962         }
1963
1964         if (use_tracer_devpaths())
1965                 clean_trace_bufs();
1966
1967         get_all_drops();
1968 }
1969
1970 static void exit_tracing(void)
1971 {
1972         signal(SIGINT, SIG_IGN);
1973         signal(SIGHUP, SIG_IGN);
1974         signal(SIGTERM, SIG_IGN);
1975         signal(SIGALRM, SIG_IGN);
1976
1977         stop_tracers();
1978         wait_tracers();
1979         del_tracers();
1980         rel_devpaths();
1981 }
1982
1983 static void handle_sigint(__attribute__((__unused__)) int sig)
1984 {
1985         done = 1;
1986         stop_tracers();
1987 }
1988
1989 static void show_stats(struct list_head *devpaths)
1990 {
1991         FILE *ofp;
1992         struct list_head *p;
1993         unsigned long long nevents, data_read;
1994         unsigned long long total_drops = 0;
1995         unsigned long long total_events = 0;
1996
1997         if (piped_output)
1998                 ofp = my_fopen("/dev/null", "w");
1999         else
2000                 ofp = stdout;
2001
2002         __list_for_each(p, devpaths) {
2003                 int cpu;
2004                 struct pdc_stats *sp;
2005                 struct devpath *dpp = list_entry(p, struct devpath, head);
2006
2007                 if (net_mode == Net_server)
2008                         printf("server: end of run for %s:%s\n",
2009                                 dpp->ch->hostname, dpp->buts_name);
2010
2011                 data_read = 0;
2012                 nevents = 0;
2013
2014                 fprintf(ofp, "=== %s ===\n", dpp->buts_name);
2015                 for (cpu = 0, sp = dpp->stats; cpu < dpp->ncpus; cpu++, sp++) {
2016                         /*
2017                          * Estimate events if not known...
2018                          */
2019                         if (sp->nevents == 0) {
2020                                 sp->nevents = sp->data_read /
2021                                                 sizeof(struct blk_io_trace);
2022                         }
2023
2024                         fprintf(ofp,
2025                                 "  CPU%3d: %20llu events, %8llu KiB data\n",
2026                                 cpu, sp->nevents, (sp->data_read + 1023) >> 10);
2027
2028                         data_read += sp->data_read;
2029                         nevents += sp->nevents;
2030                 }
2031
2032                 fprintf(ofp, "  Total:  %20llu events (dropped %llu),"
2033                              " %8llu KiB data\n", nevents,
2034                              dpp->drops, (data_read + 1024) >> 10);
2035
2036                 total_drops += dpp->drops;
2037                 total_events += (nevents + dpp->drops);
2038         }
2039
2040         fflush(ofp);
2041         if (piped_output)
2042                 fclose(ofp);
2043
2044         if (total_drops) {
2045                 double drops_ratio = 1.0;
2046
2047                 if (total_events)
2048                         drops_ratio = (double)total_drops/(double)total_events;
2049
2050                 fprintf(stderr, "\nYou have %llu (%5.1lf%%) dropped events\n"
2051                                 "Consider using a larger buffer size (-b) "
2052                                 "and/or more buffers (-n)\n",
2053                         total_drops, 100.0 * drops_ratio);
2054         }
2055 }
2056
2057 static int handle_args(int argc, char *argv[])
2058 {
2059         int c, i;
2060         struct statfs st;
2061         int act_mask_tmp = 0;
2062
2063         while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) >= 0) {
2064                 switch (c) {
2065                 case 'a':
2066                         i = find_mask_map(optarg);
2067                         if (i < 0) {
2068                                 fprintf(stderr, "Invalid action mask %s\n",
2069                                         optarg);
2070                                 return 1;
2071                         }
2072                         act_mask_tmp |= i;
2073                         break;
2074
2075                 case 'A':
2076                         if ((sscanf(optarg, "%x", &i) != 1) ||
2077                                                         !valid_act_opt(i)) {
2078                                 fprintf(stderr,
2079                                         "Invalid set action mask %s/0x%x\n",
2080                                         optarg, i);
2081                                 return 1;
2082                         }
2083                         act_mask_tmp = i;
2084                         break;
2085
2086                 case 'd':
2087                         if (add_devpath(optarg) != 0)
2088                                 return 1;
2089                         break;
2090
2091                 case 'I': {
2092                         char dev_line[256];
2093                         FILE *ifp = my_fopen(optarg, "r");
2094
2095                         if (!ifp) {
2096                                 fprintf(stderr,
2097                                         "Invalid file for devices %s\n",
2098                                         optarg);
2099                                 return 1;
2100                         }
2101
2102                         while (fscanf(ifp, "%s\n", dev_line) == 1) {
2103                                 if (add_devpath(dev_line) != 0) {
2104                                         fclose(ifp);
2105                                         return 1;
2106                                 }
2107                         }
2108                         fclose(ifp);
2109                         break;
2110                 }
2111
2112                 case 'r':
2113                         debugfs_path = optarg;
2114                         break;
2115
2116                 case 'o':
2117                         output_name = optarg;
2118                         break;
2119                 case 'k':
2120                         kill_running_trace = 1;
2121                         break;
2122                 case 'w':
2123                         stop_watch = atoi(optarg);
2124                         if (stop_watch <= 0) {
2125                                 fprintf(stderr,
2126                                         "Invalid stopwatch value (%d secs)\n",
2127                                         stop_watch);
2128                                 return 1;
2129                         }
2130                         break;
2131                 case 'V':
2132                 case 'v':
2133                         printf("%s version %s\n", argv[0], blktrace_version);
2134                         exit(0);
2135                         /*NOTREACHED*/
2136                 case 'b':
2137                         buf_size = strtoul(optarg, NULL, 10);
2138                         if (buf_size <= 0 || buf_size > 16*1024) {
2139                                 fprintf(stderr, "Invalid buffer size (%lu)\n",
2140                                         buf_size);
2141                                 return 1;
2142                         }
2143                         buf_size <<= 10;
2144                         break;
2145                 case 'n':
2146                         buf_nr = strtoul(optarg, NULL, 10);
2147                         if (buf_nr <= 0) {
2148                                 fprintf(stderr,
2149                                         "Invalid buffer nr (%lu)\n", buf_nr);
2150                                 return 1;
2151                         }
2152                         break;
2153                 case 'D':
2154                         output_dir = optarg;
2155                         break;
2156                 case 'h':
2157                         net_mode = Net_client;
2158                         memset(hostname, 0, sizeof(hostname));
2159                         strncpy(hostname, optarg, sizeof(hostname));
2160                         hostname[sizeof(hostname) - 1] = '\0';
2161                         break;
2162                 case 'l':
2163                         net_mode = Net_server;
2164                         break;
2165                 case 'p':
2166                         net_port = atoi(optarg);
2167                         break;
2168                 case 's':
2169                         net_use_sendfile = 0;
2170                         break;
2171                 default:
2172                         show_usage(argv[0]);
2173                         exit(1);
2174                         /*NOTREACHED*/
2175                 }
2176         }
2177
2178         while (optind < argc)
2179                 if (add_devpath(argv[optind++]) != 0)
2180                         return 1;
2181
2182         if (net_mode != Net_server && ndevs == 0) {
2183                 show_usage(argv[0]);
2184                 return 1;
2185         }
2186
2187         if (statfs(debugfs_path, &st) < 0) {
2188                 fprintf(stderr, "Invalid debug path %s: %d/%s\n",
2189                         debugfs_path, errno, strerror(errno));
2190                 return 1;
2191         }
2192
2193         if (st.f_type != (long)DEBUGFS_TYPE) {
2194                 fprintf(stderr, "Debugfs is not mounted at %s\n", debugfs_path);
2195                 return 1;
2196         }
2197
2198         if (act_mask_tmp != 0)
2199                 act_mask = act_mask_tmp;
2200
2201         if (net_mode == Net_client && net_setup_addr())
2202                 return 1;
2203
2204         /*
2205          * Set up for appropriate PFD handler based upon output name.
2206          */
2207         if (net_client_use_sendfile())
2208                 handle_pfds = handle_pfds_netclient;
2209         else if (net_client_use_send())
2210                 handle_pfds = handle_pfds_entries;
2211         else if (output_name && (strcmp(output_name, "-") == 0)) {
2212                 piped_output = 1;
2213                 handle_pfds = handle_pfds_entries;
2214                 pfp = stdout;
2215                 if (setvbuf(pfp, NULL, _IONBF, 0)) {
2216                         perror("setvbuf stdout");
2217                         return 1;
2218                 }
2219         } else
2220                 handle_pfds = handle_pfds_file;
2221         return 0;
2222 }
2223
2224 static void ch_add_connection(struct net_server_s *ns, struct cl_host *ch,
2225                               int fd)
2226 {
2227         struct cl_conn *nc;
2228
2229         nc = malloc(sizeof(*nc));
2230         memset(nc, 0, sizeof(*nc));
2231
2232         time(&nc->connect_time);
2233         nc->ch = ch;
2234         nc->fd = fd;
2235         nc->ncpus = -1;
2236
2237         list_add_tail(&nc->ch_head, &ch->conn_list);
2238         ch->connects++;
2239
2240         list_add_tail(&nc->ns_head, &ns->conn_list);
2241         ns->connects++;
2242         ns->pfds = realloc(ns->pfds, (ns->connects+1) * sizeof(struct pollfd));
2243 }
2244
2245 static void ch_rem_connection(struct net_server_s *ns, struct cl_host *ch,
2246                               struct cl_conn *nc)
2247 {
2248         net_close_connection(&nc->fd);
2249
2250         list_del(&nc->ch_head);
2251         ch->connects--;
2252
2253         list_del(&nc->ns_head);
2254         ns->connects--;
2255         ns->pfds = realloc(ns->pfds, (ns->connects+1) * sizeof(struct pollfd));
2256
2257         free(nc);
2258 }
2259
2260 static struct cl_host *net_find_client_host(struct net_server_s *ns,
2261                                             struct in_addr cl_in_addr)
2262 {
2263         struct list_head *p;
2264
2265         __list_for_each(p, &ns->ch_list) {
2266                 struct cl_host *ch = list_entry(p, struct cl_host, head);
2267
2268                 if (in_addr_eq(ch->cl_in_addr, cl_in_addr))
2269                         return ch;
2270         }
2271
2272         return NULL;
2273 }
2274
2275 static struct cl_host *net_add_client_host(struct net_server_s *ns,
2276                                            struct sockaddr_in *addr)
2277 {
2278         struct cl_host *ch;
2279
2280         ch = malloc(sizeof(*ch));
2281         memset(ch, 0, sizeof(*ch));
2282
2283         ch->ns = ns;
2284         ch->cl_in_addr = addr->sin_addr;
2285         list_add_tail(&ch->head, &ns->ch_list);
2286         ns->nchs++;
2287
2288         ch->hostname = strdup(inet_ntoa(addr->sin_addr));
2289         printf("server: connection from %s\n", ch->hostname);
2290
2291         INIT_LIST_HEAD(&ch->conn_list);
2292         INIT_LIST_HEAD(&ch->devpaths);
2293
2294         return ch;
2295 }
2296
2297 static void device_done(struct devpath *dpp, int ncpus)
2298 {
2299         int cpu;
2300         struct io_info *iop;
2301
2302         for (cpu = 0, iop = dpp->ios; cpu < ncpus; cpu++, iop++)
2303                 close_iop(iop);
2304
2305         list_del(&dpp->head);
2306         dpp_free(dpp);
2307 }
2308
2309 static void net_ch_remove(struct cl_host *ch, int ncpus)
2310 {
2311         struct list_head *p, *q;
2312         struct net_server_s *ns = ch->ns;
2313
2314         list_for_each_safe(p, q, &ch->devpaths) {
2315                 struct devpath *dpp = list_entry(p, struct devpath, head);
2316                 device_done(dpp, ncpus);
2317         }
2318
2319         list_for_each_safe(p, q, &ch->conn_list) {
2320                 struct cl_conn *nc = list_entry(p, struct cl_conn, ch_head);
2321
2322                 ch_rem_connection(ns, ch, nc);
2323         }
2324
2325         list_del(&ch->head);
2326         ns->nchs--;
2327
2328         if (ch->hostname)
2329                 free(ch->hostname);
2330         free(ch);
2331 }
2332
2333 static void net_add_connection(struct net_server_s *ns)
2334 {
2335         int fd;
2336         struct cl_host *ch;
2337         socklen_t socklen = sizeof(ns->addr);
2338
2339         fd = my_accept(ns->listen_fd, (struct sockaddr *)&ns->addr, &socklen);
2340         if (fd < 0) {
2341                 /*
2342                  * This is OK: we just won't accept this connection,
2343                  * nothing fatal.
2344                  */
2345                 perror("accept");
2346         } else {
2347                 ch = net_find_client_host(ns, ns->addr.sin_addr);
2348                 if (!ch)
2349                         ch = net_add_client_host(ns, &ns->addr);
2350
2351                 ch_add_connection(ns, ch, fd);
2352         }
2353 }
2354
2355 static struct devpath *nc_add_dpp(struct cl_conn *nc,
2356                                   struct blktrace_net_hdr *bnh,
2357                                   time_t connect_time)
2358 {
2359         int cpu;
2360         struct io_info *iop;
2361         struct devpath *dpp;
2362
2363         dpp = malloc(sizeof(*dpp));
2364         memset(dpp, 0, sizeof(*dpp));
2365
2366         dpp->buts_name = strdup(bnh->buts_name);
2367         dpp->path = strdup(bnh->buts_name);
2368         dpp->fd = -1;
2369         dpp->ch = nc->ch;
2370         dpp->cl_id = bnh->cl_id;
2371         dpp->cl_connect_time = connect_time;
2372         dpp->ncpus = nc->ncpus;
2373         dpp->stats = calloc(dpp->ncpus, sizeof(*dpp->stats));
2374         memset(dpp->stats, 0, dpp->ncpus * sizeof(*dpp->stats));
2375
2376         list_add_tail(&dpp->head, &nc->ch->devpaths);
2377         nc->ch->ndevs++;
2378
2379         dpp->ios = calloc(nc->ncpus, sizeof(*iop));
2380         memset(dpp->ios, 0, ndevs * sizeof(*iop));
2381
2382         for (cpu = 0, iop = dpp->ios; cpu < nc->ncpus; cpu++, iop++) {
2383                 iop->dpp = dpp;
2384                 iop->nc = nc;
2385                 init_mmap_info(&iop->mmap_info);
2386
2387                 if (iop_open(iop, cpu))
2388                         goto err;
2389         }
2390
2391         return dpp;
2392
2393 err:
2394         /*
2395          * Need to unravel what's been done...
2396          */
2397         while (cpu >= 0)
2398                 close_iop(&dpp->ios[cpu--]);
2399         dpp_free(dpp);
2400
2401         return NULL;
2402 }
2403
2404 static struct devpath *nc_find_dpp(struct cl_conn *nc,
2405                                    struct blktrace_net_hdr *bnh)
2406 {
2407         struct list_head *p;
2408         time_t connect_time = nc->connect_time;
2409
2410         __list_for_each(p, &nc->ch->devpaths) {
2411                 struct devpath *dpp = list_entry(p, struct devpath, head);
2412
2413                 if (!strcmp(dpp->buts_name, bnh->buts_name))
2414                         return dpp;
2415
2416                 if (dpp->cl_id == bnh->cl_id)
2417                         connect_time = dpp->cl_connect_time;
2418         }
2419
2420         return nc_add_dpp(nc, bnh, connect_time);
2421 }
2422
2423 static void net_client_read_data(struct cl_conn *nc, struct devpath *dpp,
2424                                  struct blktrace_net_hdr *bnh)
2425 {
2426         int ret;
2427         struct io_info *iop = &dpp->ios[bnh->cpu];
2428         struct mmap_info *mip = &iop->mmap_info;
2429
2430         if (setup_mmap(iop->ofd, bnh->len, &iop->mmap_info, NULL)) {
2431                 fprintf(stderr, "ncd(%s:%d): mmap failed\n",
2432                         nc->ch->hostname, nc->fd);
2433                 exit(1);
2434         }
2435
2436         ret = net_recv_data(nc->fd, mip->fs_buf + mip->fs_off, bnh->len);
2437         if (ret > 0) {
2438                 pdc_dr_update(dpp, bnh->cpu, ret);
2439                 mip->fs_size += ret;
2440                 mip->fs_off += ret;
2441         } else if (ret < 0)
2442                 exit(1);
2443 }
2444
2445 /*
2446  * Returns 1 if we closed a host - invalidates other polling information
2447  * that may be present.
2448  */
2449 static int net_client_data(struct cl_conn *nc)
2450 {
2451         int ret;
2452         struct devpath *dpp;
2453         struct blktrace_net_hdr bnh;
2454
2455         ret = net_get_header(nc, &bnh);
2456         if (ret == 0)
2457                 return 0;
2458
2459         if (ret < 0) {
2460                 fprintf(stderr, "ncd(%d): header read failed\n", nc->fd);
2461                 exit(1);
2462         }
2463
2464         if (data_is_native == -1 && check_data_endianness(bnh.magic)) {
2465                 fprintf(stderr, "ncd(%d): received data is bad\n", nc->fd);
2466                 exit(1);
2467         }
2468
2469         if (!data_is_native) {
2470                 bnh.magic = be32_to_cpu(bnh.magic);
2471                 bnh.cpu = be32_to_cpu(bnh.cpu);
2472                 bnh.max_cpus = be32_to_cpu(bnh.max_cpus);
2473                 bnh.len = be32_to_cpu(bnh.len);
2474                 bnh.cl_id = be32_to_cpu(bnh.cl_id);
2475                 bnh.buf_size = be32_to_cpu(bnh.buf_size);
2476                 bnh.buf_nr = be32_to_cpu(bnh.buf_nr);
2477                 bnh.page_size = be32_to_cpu(bnh.page_size);
2478         }
2479
2480         if ((bnh.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
2481                 fprintf(stderr, "ncd(%s:%d): bad data magic\n",
2482                         nc->ch->hostname, nc->fd);
2483                 exit(1);
2484         }
2485
2486         if (nc->ncpus == -1)
2487                 nc->ncpus = bnh.max_cpus;
2488
2489         /*
2490          * len == 0 means the other end is sending us a new connection/dpp
2491          * len == 1 means that the other end signalled end-of-run
2492          */
2493         dpp = nc_find_dpp(nc, &bnh);
2494         if (bnh.len == 0) {
2495                 /*
2496                  * Just adding in the dpp above is enough
2497                  */
2498                 ack_open_close(nc->fd, dpp->buts_name);
2499                 nc->ch->cl_opens++;
2500         } else if (bnh.len == 1) {
2501                 /*
2502                  * overload cpu count with dropped events
2503                  */
2504                 dpp->drops = bnh.cpu;
2505
2506                 ack_open_close(nc->fd, dpp->buts_name);
2507                 if (--nc->ch->cl_opens == 0) {
2508                         show_stats(&nc->ch->devpaths);
2509                         net_ch_remove(nc->ch, nc->ncpus);
2510                         return 1;
2511                 }
2512         } else
2513                 net_client_read_data(nc, dpp, &bnh);
2514
2515         return 0;
2516 }
2517
2518 static void handle_client_data(struct net_server_s *ns, int events)
2519 {
2520         struct cl_conn *nc;
2521         struct pollfd *pfd;
2522         struct list_head *p, *q;
2523
2524         pfd = &ns->pfds[1];
2525         list_for_each_safe(p, q, &ns->conn_list) {
2526                 if (pfd->revents & POLLIN) {
2527                         nc = list_entry(p, struct cl_conn, ns_head);
2528
2529                         if (net_client_data(nc) || --events == 0)
2530                                 break;
2531                 }
2532                 pfd++;
2533         }
2534 }
2535
2536 static void net_setup_pfds(struct net_server_s *ns)
2537 {
2538         struct pollfd *pfd;
2539         struct list_head *p;
2540
2541         ns->pfds[0].fd = ns->listen_fd;
2542         ns->pfds[0].events = POLLIN;
2543
2544         pfd = &ns->pfds[1];
2545         __list_for_each(p, &ns->conn_list) {
2546                 struct cl_conn *nc = list_entry(p, struct cl_conn, ns_head);
2547
2548                 pfd->fd = nc->fd;
2549                 pfd->events = POLLIN;
2550                 pfd++;
2551         }
2552 }
2553
2554 static int net_server_handle_connections(struct net_server_s *ns)
2555 {
2556         int events;
2557
2558         printf("server: waiting for connections...\n");
2559
2560         while (!done) {
2561                 net_setup_pfds(ns);
2562                 events = poll(ns->pfds, ns->connects + 1, -1);
2563                 if (events < 0) {
2564                         if (errno != EINTR) {
2565                                 perror("FATAL: poll error");
2566                                 return 1;
2567                         }
2568                 } else if (events > 0) {
2569                         if (ns->pfds[0].revents & POLLIN) {
2570                                 net_add_connection(ns);
2571                                 events--;
2572                         }
2573
2574                         if (events)
2575                                 handle_client_data(ns, events);
2576                 }
2577         }
2578
2579         return 0;
2580 }
2581
2582 static int net_server(void)
2583 {
2584         int fd, opt;
2585         int ret = 1;
2586         struct net_server_s net_server;
2587         struct net_server_s *ns = &net_server;
2588
2589         memset(ns, 0, sizeof(*ns));
2590         INIT_LIST_HEAD(&ns->ch_list);
2591         INIT_LIST_HEAD(&ns->conn_list);
2592         ns->pfds = malloc(sizeof(struct pollfd));
2593
2594         fd = my_socket(AF_INET, SOCK_STREAM, 0);
2595         if (fd < 0) {
2596                 perror("server: socket");
2597                 goto out;
2598         }
2599
2600         opt = 1;
2601         if (setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)) < 0) {
2602                 perror("setsockopt");
2603                 goto out;
2604         }
2605
2606         memset(&ns->addr, 0, sizeof(ns->addr));
2607         ns->addr.sin_family = AF_INET;
2608         ns->addr.sin_addr.s_addr = htonl(INADDR_ANY);
2609         ns->addr.sin_port = htons(net_port);
2610
2611         if (bind(fd, (struct sockaddr *) &ns->addr, sizeof(ns->addr)) < 0) {
2612                 perror("bind");
2613                 goto out;
2614         }
2615
2616         if (listen(fd, 1) < 0) {
2617                 perror("listen");
2618                 goto out;
2619         }
2620
2621         /*
2622          * The actual server looping is done here:
2623          */
2624         ns->listen_fd = fd;
2625         ret = net_server_handle_connections(ns);
2626
2627         /*
2628          * Clean up and return...
2629          */
2630 out:
2631         free(ns->pfds);
2632         return ret;
2633 }
2634
2635 static int run_tracers(void)
2636 {
2637         atexit(exit_tracing);
2638         if (net_mode == Net_client)
2639                 printf("blktrace: connecting to %s\n", hostname);
2640
2641         setup_buts();
2642
2643         if (use_tracer_devpaths()) {
2644                 if (setup_tracer_devpaths())
2645                         return 1;
2646
2647                 if (piped_output)
2648                         handle_list = handle_list_file;
2649                 else
2650                         handle_list = handle_list_net;
2651         }
2652
2653         start_tracers();
2654         if (nthreads_running == ncpus) {
2655                 unblock_tracers();
2656                 start_buts();
2657                 if (net_mode == Net_client)
2658                         printf("blktrace: connected!\n");
2659                 if (stop_watch)
2660                         alarm(stop_watch);
2661         } else
2662                 stop_tracers();
2663
2664         wait_tracers();
2665         if (nthreads_running == ncpus)
2666                 show_stats(&devpaths);
2667         if (net_client_use_send())
2668                 close_client_connections();
2669         del_tracers();
2670
2671         return 0;
2672 }
2673
2674 static cpu_set_t *get_online_cpus(void)
2675 {
2676         FILE *cpus;
2677         cpu_set_t *set;
2678         size_t alloc_size;
2679         int cpuid, prevcpuid = -1;
2680         char nextch;
2681         int n, ncpu, curcpu = 0;
2682         int *cpu_nums;
2683
2684         ncpu = sysconf(_SC_NPROCESSORS_CONF);
2685         if (ncpu < 0)
2686                 return NULL;
2687
2688         cpu_nums = malloc(sizeof(int)*ncpu);
2689         if (!cpu_nums) {
2690                 errno = ENOMEM;
2691                 return NULL;
2692         }
2693
2694         /*
2695          * There is no way to easily get maximum CPU number. So we have to
2696          * parse the file first to find it out and then create appropriate
2697          * cpuset
2698          */
2699         cpus = my_fopen("/sys/devices/system/cpu/online", "r");
2700         for (;;) {
2701                 n = fscanf(cpus, "%d%c", &cpuid, &nextch);
2702                 if (n <= 0)
2703                         break;
2704                 if (n == 2 && nextch == '-') {
2705                         prevcpuid = cpuid;
2706                         continue;
2707                 }
2708                 if (prevcpuid == -1)
2709                         prevcpuid = cpuid;
2710                 while (prevcpuid <= cpuid) {
2711                         /* More CPUs listed than configured? */
2712                         if (curcpu >= ncpu) {
2713                                 errno = EINVAL;
2714                                 return NULL;
2715                         }
2716                         cpu_nums[curcpu++] = prevcpuid++;
2717                 }
2718                 prevcpuid = -1;
2719         }
2720         fclose(cpus);
2721
2722         ncpu = curcpu;
2723         max_cpus = cpu_nums[ncpu - 1] + 1;
2724
2725         /* Now that we have maximum cpu number, create a cpuset */
2726         set = CPU_ALLOC(max_cpus);
2727         if (!set) {
2728                 errno = ENOMEM;
2729                 return NULL;
2730         }
2731         alloc_size = CPU_ALLOC_SIZE(max_cpus);
2732         CPU_ZERO_S(alloc_size, set);
2733
2734         for (curcpu = 0; curcpu < ncpu; curcpu++)
2735                 CPU_SET_S(cpu_nums[curcpu], alloc_size, set);
2736
2737         free(cpu_nums);
2738
2739         return set;
2740 }
2741
2742 int main(int argc, char *argv[])
2743 {
2744         int ret = 0;
2745
2746         setlocale(LC_NUMERIC, "en_US");
2747         pagesize = getpagesize();
2748         online_cpus = get_online_cpus();
2749         if (!online_cpus) {
2750                 fprintf(stderr, "cannot get online cpus %d/%s\n",
2751                         errno, strerror(errno));
2752                 ret = 1;
2753                 goto out;
2754         } else if (handle_args(argc, argv)) {
2755                 ret = 1;
2756                 goto out;
2757         }
2758
2759         ncpus = CPU_COUNT_S(CPU_ALLOC_SIZE(max_cpus), online_cpus);
2760         if (ndevs > 1 && output_name && strcmp(output_name, "-") != 0) {
2761                 fprintf(stderr, "-o not supported with multiple devices\n");
2762                 ret = 1;
2763                 goto out;
2764         }
2765
2766         signal(SIGINT, handle_sigint);
2767         signal(SIGHUP, handle_sigint);
2768         signal(SIGTERM, handle_sigint);
2769         signal(SIGALRM, handle_sigint);
2770         signal(SIGPIPE, SIG_IGN);
2771
2772         if (kill_running_trace) {
2773                 struct devpath *dpp;
2774                 struct list_head *p;
2775
2776                 __list_for_each(p, &devpaths) {
2777                         dpp = list_entry(p, struct devpath, head);
2778                         if (__stop_trace(dpp->fd)) {
2779                                 fprintf(stderr,
2780                                         "BLKTRACETEARDOWN %s failed: %d/%s\n",
2781                                         dpp->path, errno, strerror(errno));
2782                         }
2783                 }
2784         } else if (net_mode == Net_server) {
2785                 if (output_name) {
2786                         fprintf(stderr, "-o ignored in server mode\n");
2787                         output_name = NULL;
2788                 }
2789                 ret = net_server();
2790         } else
2791                 ret = run_tracers();
2792
2793 out:
2794         if (pfp)
2795                 fclose(pfp);
2796         rel_devpaths();
2797         return ret;
2798 }