2 * block queue tracing application
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
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
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.
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.
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
39 #include <sys/ioctl.h>
40 #include <sys/types.h>
44 #include <sys/param.h>
46 #include <sys/resource.h>
47 #include <sys/socket.h>
48 #include <netinet/in.h>
49 #include <arpa/inet.h>
51 #include <sys/sendfile.h>
57 * You may want to increase this even more, if you are logging at a high
58 * rate and see skipped/missed events
60 #define BUF_SIZE (512 * 1024)
63 #define FILE_VBUF_SIZE (128 * 1024)
65 #define DEBUGFS_TYPE (0x64626720)
66 #define TRACE_NET_PORT (8462)
81 * Generic stats collected: nevents can be _roughly_ estimated by data_read
82 * (discounting pdu...)
84 * These fields are updated w/ pdc_dr_update & pdc_nev_update below.
87 unsigned long long data_read;
88 unsigned long long nevents;
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;
97 unsigned long long drops;
100 * For piped output only:
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.
107 struct tracer_devpath_head *heads;
110 * For network server mode only:
114 time_t cl_connect_time;
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.
122 * The main thread will then collect trace buffers from each of lists in turn.
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
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).
135 struct list_head head;
141 struct tracer_devpath_head {
142 pthread_mutex_t mutex;
143 struct list_head head;
144 struct trace_buf *prev;
148 * Used to handle the mmap() interfaces for output file (containing traces)
152 unsigned long long fs_size, fs_max_size, fs_off, fs_buf_len;
153 unsigned long buf_size, buf_nr;
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
164 struct list_head head;
169 volatile int status, is_done;
173 * networking stuff follows. we include a magic number so we know whether
174 * to endianness convert or not.
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
180 * The cpu field is overloaded on close: it will contain the number of drops.
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 */
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 */
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.
201 struct list_head head;
202 struct list_head conn_list;
203 struct list_head devpaths;
204 struct net_server_s *ns;
206 struct in_addr cl_in_addr;
207 int connects, ndevs, cl_opens;
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.
216 struct list_head ch_head, ns_head;
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.
228 struct net_server_s {
229 struct list_head conn_list;
230 struct list_head ch_list;
232 int listen_fd, connects, nchs;
233 struct sockaddr_in addr;
237 * This structure is (generically) used to providide information
238 * for a read-to-write set of values.
240 * ifn & ifd represent input information
242 * ofn, ofd, ofp, obuf & mmap_info are used for output file (optionally).
248 struct cl_conn *nc; /* Server network connection */
251 * mmap controlled output files
253 struct mmap_info mmap_info;
256 * Client network fields
259 unsigned long long data_queued;
262 * Input/output file descriptors & names
265 char ifn[MAXPATHLEN + 64];
266 char ofn[MAXPATHLEN + 64];
269 static char blktrace_version[] = "2.0.0";
272 * Linkage to blktrace helper routines (trace conversions)
274 int data_is_native = -1;
279 static int act_mask = ~0U;
280 static int kill_running_trace;
281 static int stop_watch;
282 static int piped_output;
284 static char *debugfs_path = "/sys/kernel/debug";
285 static char *output_name;
286 static char *output_dir;
288 static unsigned long buf_size = BUF_SIZE;
289 static unsigned long buf_nr = BUF_NR;
293 static LIST_HEAD(devpaths);
294 static LIST_HEAD(tracers);
296 static volatile int done;
299 * tracer threads add entries, the main thread takes them off and processes
300 * them. These protect the dp_entries variable.
302 static pthread_cond_t dp_cond = PTHREAD_COND_INITIALIZER;
303 static pthread_mutex_t dp_mutex = PTHREAD_MUTEX_INITIALIZER;
304 static volatile int dp_entries;
307 * These synchronize master / thread interactions.
309 static pthread_cond_t mt_cond = PTHREAD_COND_INITIALIZER;
310 static pthread_mutex_t mt_mutex = PTHREAD_MUTEX_INITIALIZER;
311 static volatile int nthreads_running;
312 static volatile int nthreads_leaving;
313 static volatile int nthreads_error;
314 static volatile int tracers_run;
317 * network cmd line params
319 static struct sockaddr_in hostname_addr;
320 static char hostname[MAXHOSTNAMELEN];
321 static int net_port = TRACE_NET_PORT;
322 static int net_use_sendfile = 1;
326 static int (*handle_pfds)(struct tracer *, int, int);
327 static int (*handle_list)(struct tracer_devpath_head *, struct list_head *);
329 #define S_OPTS "d:a:A:r:o:kw:vVb:n:D:lh:p:sI:"
330 static struct option l_opts[] = {
333 .has_arg = required_argument,
338 .name = "input-devs",
339 .has_arg = required_argument,
345 .has_arg = required_argument,
351 .has_arg = required_argument,
357 .has_arg = required_argument,
363 .has_arg = required_argument,
369 .has_arg = no_argument,
375 .has_arg = required_argument,
381 .has_arg = no_argument,
387 .has_arg = no_argument,
392 .name = "buffer-size",
393 .has_arg = required_argument,
398 .name = "num-sub-buffers",
399 .has_arg = required_argument,
404 .name = "output-dir",
405 .has_arg = required_argument,
411 .has_arg = no_argument,
417 .has_arg = required_argument,
423 .has_arg = required_argument,
428 .name = "no-sendfile",
429 .has_arg = no_argument,
438 static char usage_str[] = "\n\n" \
439 "-d <dev> | --dev=<dev>\n" \
440 "[ -r <debugfs path> | --relay=<debugfs path> ]\n" \
441 "[ -o <file> | --output=<file>]\n" \
442 "[ -D <dir> | --output-dir=<dir>\n" \
443 "[ -w <time> | --stopwatch=<time>]\n" \
444 "[ -a <action field> | --act-mask=<action field>]\n" \
445 "[ -A <action mask> | --set-mask=<action mask>]\n" \
446 "[ -b <size> | --buffer-size]\n" \
447 "[ -n <number> | --num-sub-buffers=<number>]\n" \
448 "[ -l | --listen]\n" \
449 "[ -h <hostname> | --host=<hostname>]\n" \
450 "[ -p <port number> | --port=<port number>]\n" \
451 "[ -s | --no-sendfile]\n" \
452 "[ -I <devs file> | --input-devs=<devs file>]\n" \
453 "[ -v <version> | --version]\n" \
454 "[ -V <version> | --version]\n" \
456 "\t-d Use specified device. May also be given last after options\n" \
457 "\t-r Path to mounted debugfs, defaults to /sys/kernel/debug\n" \
458 "\t-o File(s) to send output to\n" \
459 "\t-D Directory to prepend to output file names\n" \
460 "\t-w Stop after defined time, in seconds\n" \
461 "\t-a Only trace specified actions. See documentation\n" \
462 "\t-A Give trace mask as a single value. See documentation\n" \
463 "\t-b Sub buffer size in KiB (default 512)\n" \
464 "\t-n Number of sub buffers (default 4)\n" \
465 "\t-l Run in network listen mode (blktrace server)\n" \
466 "\t-h Run in network client mode, connecting to the given host\n" \
467 "\t-p Network port to use (default 8462)\n" \
468 "\t-s Make the network client NOT use sendfile() to transfer data\n" \
469 "\t-I Add devices found in <devs file>\n" \
470 "\t-v Print program version info\n" \
471 "\t-V Print program version info\n\n";
473 static void clear_events(struct pollfd *pfd)
479 static inline int net_client_use_sendfile(void)
481 return net_mode == Net_client && net_use_sendfile;
484 static inline int net_client_use_send(void)
486 return net_mode == Net_client && !net_use_sendfile;
489 static inline int use_tracer_devpaths(void)
491 return piped_output || net_client_use_send();
494 static inline int in_addr_eq(struct in_addr a, struct in_addr b)
496 return a.s_addr == b.s_addr;
499 static inline void pdc_dr_update(struct devpath *dpp, int cpu, int data_read)
501 dpp->stats[cpu].data_read += data_read;
504 static inline void pdc_nev_update(struct devpath *dpp, int cpu, int nevents)
506 dpp->stats[cpu].nevents += nevents;
509 static void show_usage(char *prog)
511 fprintf(stderr, "Usage: %s %s", prog, usage_str);
515 * Create a timespec 'msec' milliseconds into the future
517 static inline void make_timespec(struct timespec *tsp, long delta_msec)
521 gettimeofday(&now, NULL);
522 tsp->tv_sec = now.tv_sec;
523 tsp->tv_nsec = 1000L * now.tv_usec;
525 tsp->tv_nsec += (delta_msec * 1000000L);
526 if (tsp->tv_nsec > 1000000000L) {
527 long secs = tsp->tv_nsec / 1000000000L;
530 tsp->tv_nsec -= (secs * 1000000000L);
535 * Add a timer to ensure wait ends
537 static void t_pthread_cond_wait(pthread_cond_t *cond, pthread_mutex_t *mutex)
541 make_timespec(&ts, 50);
542 pthread_cond_timedwait(cond, mutex, &ts);
545 static void unblock_tracers(void)
547 pthread_mutex_lock(&mt_mutex);
549 pthread_cond_broadcast(&mt_cond);
550 pthread_mutex_unlock(&mt_mutex);
553 static void tracer_wait_unblock(struct tracer *tp)
555 pthread_mutex_lock(&mt_mutex);
556 while (!tp->is_done && !tracers_run)
557 pthread_cond_wait(&mt_cond, &mt_mutex);
558 pthread_mutex_unlock(&mt_mutex);
561 static void tracer_signal_ready(struct tracer *tp,
562 enum thread_status th_status,
565 pthread_mutex_lock(&mt_mutex);
568 if (th_status == Th_running)
570 else if (th_status == Th_error)
575 pthread_cond_signal(&mt_cond);
576 pthread_mutex_unlock(&mt_mutex);
579 static void wait_tracers_ready(int ncpus_started)
581 pthread_mutex_lock(&mt_mutex);
582 while ((nthreads_running + nthreads_error) < ncpus_started)
583 t_pthread_cond_wait(&mt_cond, &mt_mutex);
584 pthread_mutex_unlock(&mt_mutex);
587 static void wait_tracers_leaving(void)
589 pthread_mutex_lock(&mt_mutex);
590 while (nthreads_leaving < nthreads_running)
591 t_pthread_cond_wait(&mt_cond, &mt_mutex);
592 pthread_mutex_unlock(&mt_mutex);
595 static void init_mmap_info(struct mmap_info *mip)
597 mip->buf_size = buf_size;
598 mip->buf_nr = buf_nr;
599 mip->pagesize = pagesize;
602 static void net_close_connection(int *fd)
604 shutdown(*fd, SHUT_RDWR);
609 static void dpp_free(struct devpath *dpp)
618 free(dpp->buts_name);
622 static int lock_on_cpu(int cpu)
627 CPU_SET(cpu, &cpu_mask);
628 if (sched_setaffinity(0, sizeof(cpu_mask), &cpu_mask) < 0)
634 static int increase_limit(int resource, rlim_t increase)
637 int save_errno = errno;
639 if (!getrlimit(resource, &rlim)) {
640 rlim.rlim_cur += increase;
641 if (rlim.rlim_cur >= rlim.rlim_max)
642 rlim.rlim_max = rlim.rlim_cur + increase;
644 if (!setrlimit(resource, &rlim))
652 static int handle_open_failure(void)
654 if (errno == ENFILE || errno == EMFILE)
655 return increase_limit(RLIMIT_NOFILE, 16);
659 static int handle_mem_failure(size_t length)
662 return handle_open_failure();
663 else if (errno == ENOMEM)
664 return increase_limit(RLIMIT_MEMLOCK, 2 * length);
668 static FILE *my_fopen(const char *path, const char *mode)
673 fp = fopen(path, mode);
674 } while (fp == NULL && handle_open_failure());
679 static int my_open(const char *path, int flags)
684 fd = open(path, flags);
685 } while (fd < 0 && handle_open_failure());
690 static int my_socket(int domain, int type, int protocol)
695 fd = socket(domain, type, protocol);
696 } while (fd < 0 && handle_open_failure());
701 static int my_accept(int sockfd, struct sockaddr *addr, socklen_t *addrlen)
706 fd = accept(sockfd, addr, addrlen);
707 } while (fd < 0 && handle_open_failure());
712 static void *my_mmap(void *addr, size_t length, int prot, int flags, int fd,
718 new = mmap(addr, length, prot, flags, fd, offset);
719 } while (new == MAP_FAILED && handle_mem_failure(length));
724 static int my_mlock(struct tracer *tp,
725 const void *addr, size_t len)
730 ret = mlock(addr, len);
731 if ((retry >= 10) && tp && tp->is_done)
734 } while (ret < 0 && handle_mem_failure(len));
739 static int setup_mmap(int fd, unsigned int maxlen,
740 struct mmap_info *mip,
743 if (mip->fs_off + maxlen > mip->fs_buf_len) {
744 unsigned long nr = max(16, mip->buf_nr);
747 munlock(mip->fs_buf, mip->fs_buf_len);
748 munmap(mip->fs_buf, mip->fs_buf_len);
752 mip->fs_off = mip->fs_size & (mip->pagesize - 1);
753 mip->fs_buf_len = (nr * mip->buf_size) - mip->fs_off;
754 mip->fs_max_size += mip->fs_buf_len;
756 if (ftruncate(fd, mip->fs_max_size) < 0) {
757 perror("setup_mmap: ftruncate");
761 mip->fs_buf = my_mmap(NULL, mip->fs_buf_len, PROT_WRITE,
763 mip->fs_size - mip->fs_off);
764 if (mip->fs_buf == MAP_FAILED) {
765 perror("setup_mmap: mmap");
768 if (my_mlock(tp, mip->fs_buf, mip->fs_buf_len) < 0) {
769 perror("setup_mlock: mlock");
777 static int __stop_trace(int fd)
780 * Should be stopped, don't complain if it isn't
782 ioctl(fd, BLKTRACESTOP);
783 return ioctl(fd, BLKTRACETEARDOWN);
786 static int write_data(char *buf, int len)
791 ret = fwrite(buf, len, 1, pfp);
792 if (ferror(pfp) || ret != 1) {
793 if (errno == EINTR) {
798 if (!piped_output || (errno != EPIPE && errno != EBADF)) {
799 fprintf(stderr, "write(%d) failed: %d/%s\n",
800 len, errno, strerror(errno));
814 * Returns the number of bytes read (successfully)
816 static int __net_recv_data(int fd, void *buf, unsigned int len)
818 unsigned int bytes_left = len;
820 while (bytes_left && !done) {
821 int ret = recv(fd, buf, bytes_left, MSG_WAITALL);
826 if (errno == EAGAIN) {
830 perror("server: net_recv_data: recv failed");
838 return len - bytes_left;
841 static int net_recv_data(int fd, void *buf, unsigned int len)
843 return __net_recv_data(fd, buf, len);
847 * Returns number of bytes written
849 static int net_send_data(int fd, void *buf, unsigned int buf_len)
852 unsigned int bytes_left = buf_len;
855 ret = send(fd, buf, bytes_left, 0);
865 return buf_len - bytes_left;
868 static int net_send_header(int fd, int cpu, char *buts_name, int len)
870 struct blktrace_net_hdr hdr;
872 memset(&hdr, 0, sizeof(hdr));
874 hdr.magic = BLK_IO_TRACE_MAGIC;
875 strncpy(hdr.buts_name, buts_name, sizeof(hdr.buts_name));
876 hdr.buts_name[sizeof(hdr.buts_name)-1] = '\0';
878 hdr.max_cpus = ncpus;
880 hdr.cl_id = getpid();
881 hdr.buf_size = buf_size;
883 hdr.page_size = pagesize;
885 return net_send_data(fd, &hdr, sizeof(hdr)) != sizeof(hdr);
888 static void net_send_open_close(int fd, int cpu, char *buts_name, int len)
890 struct blktrace_net_hdr ret_hdr;
892 net_send_header(fd, cpu, buts_name, len);
893 net_recv_data(fd, &ret_hdr, sizeof(ret_hdr));
896 static void net_send_open(int fd, int cpu, char *buts_name)
898 net_send_open_close(fd, cpu, buts_name, 0);
901 static void net_send_close(int fd, char *buts_name, int drops)
904 * Overload CPU w/ number of drops
906 * XXX: Need to clear/set done around call - done=1 (which
907 * is true here) stops reads from happening... :-(
910 net_send_open_close(fd, drops, buts_name, 1);
914 static void ack_open_close(int fd, char *buts_name)
916 net_send_header(fd, 0, buts_name, 2);
919 static void net_send_drops(int fd)
923 __list_for_each(p, &devpaths) {
924 struct devpath *dpp = list_entry(p, struct devpath, head);
926 net_send_close(fd, dpp->buts_name, dpp->drops);
936 static int net_get_header(struct cl_conn *nc, struct blktrace_net_hdr *bnh)
939 int fl = fcntl(nc->fd, F_GETFL);
941 fcntl(nc->fd, F_SETFL, fl | O_NONBLOCK);
942 bytes_read = __net_recv_data(nc->fd, bnh, sizeof(*bnh));
943 fcntl(nc->fd, F_SETFL, fl & ~O_NONBLOCK);
945 if (bytes_read == sizeof(*bnh))
947 else if (bytes_read == 0)
953 static int net_setup_addr(void)
955 struct sockaddr_in *addr = &hostname_addr;
957 memset(addr, 0, sizeof(*addr));
958 addr->sin_family = AF_INET;
959 addr->sin_port = htons(net_port);
961 if (inet_aton(hostname, &addr->sin_addr) != 1) {
962 struct hostent *hent;
964 hent = gethostbyname(hostname);
966 if (h_errno == TRY_AGAIN) {
969 } else if (h_errno == NO_RECOVERY) {
970 fprintf(stderr, "gethostbyname(%s)"
971 "non-recoverable error encountered\n",
975 * HOST_NOT_FOUND, NO_ADDRESS or NO_DATA
977 fprintf(stderr, "Host %s not found\n",
983 memcpy(&addr->sin_addr, hent->h_addr, 4);
984 strcpy(hostname, hent->h_name);
990 static int net_setup_client(void)
993 struct sockaddr_in *addr = &hostname_addr;
995 fd = my_socket(AF_INET, SOCK_STREAM, 0);
997 perror("client: socket");
1001 if (connect(fd, (struct sockaddr *)addr, sizeof(*addr)) < 0) {
1002 if (errno == ECONNREFUSED)
1004 "\nclient: Connection to %s refused, "
1005 "perhaps the server is not started?\n\n",
1008 perror("client: connect");
1017 static int open_client_connections(void)
1021 cl_fds = calloc(ncpus, sizeof(*cl_fds));
1022 for (cpu = 0; cpu < ncpus; cpu++) {
1023 cl_fds[cpu] = net_setup_client();
1024 if (cl_fds[cpu] < 0)
1031 close(cl_fds[cpu--]);
1036 static void close_client_connections(void)
1041 for (cpu = 0, fdp = cl_fds; cpu < ncpus; cpu++, fdp++) {
1043 net_send_drops(*fdp);
1044 net_close_connection(fdp);
1051 static void setup_buts(void)
1053 struct list_head *p;
1055 __list_for_each(p, &devpaths) {
1056 struct blk_user_trace_setup buts;
1057 struct devpath *dpp = list_entry(p, struct devpath, head);
1059 memset(&buts, 0, sizeof(buts));
1060 buts.buf_size = buf_size;
1061 buts.buf_nr = buf_nr;
1062 buts.act_mask = act_mask;
1064 if (ioctl(dpp->fd, BLKTRACESETUP, &buts) >= 0) {
1066 dpp->buts_name = strdup(buts.name);
1069 dpp->stats = calloc(dpp->ncpus, sizeof(*dpp->stats));
1070 memset(dpp->stats, 0, dpp->ncpus * sizeof(*dpp->stats));
1072 fprintf(stderr, "BLKTRACESETUP(2) %s failed: %d/%s\n",
1073 dpp->path, errno, strerror(errno));
1077 static void start_buts(void)
1079 struct list_head *p;
1081 __list_for_each(p, &devpaths) {
1082 struct devpath *dpp = list_entry(p, struct devpath, head);
1084 if (ioctl(dpp->fd, BLKTRACESTART) < 0) {
1085 fprintf(stderr, "BLKTRACESTART %s failed: %d/%s\n",
1086 dpp->path, errno, strerror(errno));
1091 static int get_drops(struct devpath *dpp)
1094 char fn[MAXPATHLEN + 64], tmp[256];
1096 snprintf(fn, sizeof(fn), "%s/block/%s/dropped", debugfs_path,
1099 fd = my_open(fn, O_RDONLY);
1102 * This may be ok: the kernel may not support
1105 if (errno != ENOENT)
1106 fprintf(stderr, "Could not open %s: %d/%s\n",
1107 fn, errno, strerror(errno));
1109 } else if (read(fd, tmp, sizeof(tmp)) < 0) {
1110 fprintf(stderr, "Could not read %s: %d/%s\n",
1111 fn, errno, strerror(errno));
1119 static void get_all_drops(void)
1121 struct list_head *p;
1123 __list_for_each(p, &devpaths) {
1124 struct devpath *dpp = list_entry(p, struct devpath, head);
1126 dpp->drops = get_drops(dpp);
1130 static inline struct trace_buf *alloc_trace_buf(int cpu, int bufsize)
1132 struct trace_buf *tbp;
1134 tbp = malloc(sizeof(*tbp) + bufsize);
1135 INIT_LIST_HEAD(&tbp->head);
1137 tbp->buf = (void *)(tbp + 1);
1139 tbp->dpp = NULL; /* Will be set when tbp is added */
1144 static void free_tracer_heads(struct devpath *dpp)
1147 struct tracer_devpath_head *hd;
1149 for (cpu = 0, hd = dpp->heads; cpu < ncpus; cpu++, hd++) {
1153 pthread_mutex_destroy(&hd->mutex);
1158 static int setup_tracer_devpaths(void)
1160 struct list_head *p;
1162 if (net_client_use_send())
1163 if (open_client_connections())
1166 __list_for_each(p, &devpaths) {
1168 struct tracer_devpath_head *hd;
1169 struct devpath *dpp = list_entry(p, struct devpath, head);
1171 dpp->heads = calloc(ncpus, sizeof(struct tracer_devpath_head));
1172 for (cpu = 0, hd = dpp->heads; cpu < ncpus; cpu++, hd++) {
1173 INIT_LIST_HEAD(&hd->head);
1174 pthread_mutex_init(&hd->mutex, NULL);
1182 static inline void add_trace_buf(struct devpath *dpp, int cpu,
1183 struct trace_buf **tbpp)
1185 struct trace_buf *tbp = *tbpp;
1186 struct tracer_devpath_head *hd = &dpp->heads[cpu];
1190 pthread_mutex_lock(&hd->mutex);
1191 list_add_tail(&tbp->head, &hd->head);
1192 pthread_mutex_unlock(&hd->mutex);
1194 *tbpp = alloc_trace_buf(cpu, buf_size);
1197 static inline void incr_entries(int entries_handled)
1199 pthread_mutex_lock(&dp_mutex);
1200 if (dp_entries == 0)
1201 pthread_cond_signal(&dp_cond);
1202 dp_entries += entries_handled;
1203 pthread_mutex_unlock(&dp_mutex);
1206 static void decr_entries(int handled)
1208 pthread_mutex_lock(&dp_mutex);
1209 dp_entries -= handled;
1210 pthread_mutex_unlock(&dp_mutex);
1213 static int wait_empty_entries(void)
1215 pthread_mutex_lock(&dp_mutex);
1216 while (!done && dp_entries == 0)
1217 t_pthread_cond_wait(&dp_cond, &dp_mutex);
1218 pthread_mutex_unlock(&dp_mutex);
1223 static int add_devpath(char *path)
1226 struct devpath *dpp;
1227 struct list_head *p;
1230 * Verify device is not duplicated
1232 __list_for_each(p, &devpaths) {
1233 struct devpath *tmp = list_entry(p, struct devpath, head);
1234 if (!strcmp(tmp->path, path))
1238 * Verify device is valid before going too far
1240 fd = my_open(path, O_RDONLY | O_NONBLOCK);
1242 fprintf(stderr, "Invalid path %s specified: %d/%s\n",
1243 path, errno, strerror(errno));
1247 dpp = malloc(sizeof(*dpp));
1248 memset(dpp, 0, sizeof(*dpp));
1249 dpp->path = strdup(path);
1252 list_add_tail(&dpp->head, &devpaths);
1257 static void rel_devpaths(void)
1259 struct list_head *p, *q;
1261 list_for_each_safe(p, q, &devpaths) {
1262 struct devpath *dpp = list_entry(p, struct devpath, head);
1264 list_del(&dpp->head);
1265 __stop_trace(dpp->fd);
1269 free_tracer_heads(dpp);
1276 static int flush_subbuf_net(struct trace_buf *tbp)
1278 int fd = cl_fds[tbp->cpu];
1279 struct devpath *dpp = tbp->dpp;
1281 if (net_send_header(fd, tbp->cpu, dpp->buts_name, tbp->len))
1283 else if (net_send_data(fd, tbp->buf, tbp->len) != tbp->len)
1290 handle_list_net(__attribute__((__unused__))struct tracer_devpath_head *hd,
1291 struct list_head *list)
1293 struct trace_buf *tbp;
1294 struct list_head *p, *q;
1295 int entries_handled = 0;
1297 list_for_each_safe(p, q, list) {
1298 tbp = list_entry(p, struct trace_buf, head);
1300 list_del(&tbp->head);
1303 if (cl_fds[tbp->cpu] >= 0) {
1304 if (flush_subbuf_net(tbp)) {
1305 close(cl_fds[tbp->cpu]);
1306 cl_fds[tbp->cpu] = -1;
1313 return entries_handled;
1317 * Tack 'tbp's buf onto the tail of 'prev's buf
1319 static struct trace_buf *tb_combine(struct trace_buf *prev,
1320 struct trace_buf *tbp)
1322 unsigned long tot_len;
1324 tot_len = prev->len + tbp->len;
1325 if (tot_len > buf_size) {
1327 * tbp->head isn't connected (it was 'prev'
1328 * so it had been taken off of the list
1329 * before). Therefore, we can realloc
1330 * the whole structures, as the other fields
1333 prev = realloc(prev->buf, sizeof(*prev) + tot_len);
1334 prev->buf = (void *)(prev + 1);
1337 memcpy(prev->buf + prev->len, tbp->buf, tbp->len);
1338 prev->len = tot_len;
1344 static int handle_list_file(struct tracer_devpath_head *hd,
1345 struct list_head *list)
1347 int off, t_len, nevents;
1348 struct blk_io_trace *t;
1349 struct list_head *p, *q;
1350 int entries_handled = 0;
1351 struct trace_buf *tbp, *prev;
1354 list_for_each_safe(p, q, list) {
1355 tbp = list_entry(p, struct trace_buf, head);
1356 list_del(&tbp->head);
1360 * If there was some leftover before, tack this new
1361 * entry onto the tail of the previous one.
1364 tbp = tb_combine(prev, tbp);
1367 * See how many whole traces there are - send them
1368 * all out in one go.
1372 while (off + (int)sizeof(*t) <= tbp->len) {
1373 t = (struct blk_io_trace *)(tbp->buf + off);
1374 t_len = sizeof(*t) + t->pdu_len;
1375 if (off + t_len > tbp->len)
1382 pdc_nev_update(tbp->dpp, tbp->cpu, nevents);
1385 * Write any full set of traces, any remaining data is kept
1386 * for the next pass.
1389 if (write_data(tbp->buf, off) || off == tbp->len) {
1395 * Move valid data to beginning of buffer
1398 memmove(tbp->buf, tbp->buf + off, tbp->len);
1406 return entries_handled;
1409 static void __process_trace_bufs(void)
1412 struct list_head *p;
1413 struct list_head list;
1416 __list_for_each(p, &devpaths) {
1417 struct devpath *dpp = list_entry(p, struct devpath, head);
1418 struct tracer_devpath_head *hd = dpp->heads;
1420 for (cpu = 0; cpu < ncpus; cpu++, hd++) {
1421 pthread_mutex_lock(&hd->mutex);
1422 if (list_empty(&hd->head)) {
1423 pthread_mutex_unlock(&hd->mutex);
1427 list_replace_init(&hd->head, &list);
1428 pthread_mutex_unlock(&hd->mutex);
1430 handled += handle_list(hd, &list);
1435 decr_entries(handled);
1438 static void process_trace_bufs(void)
1440 while (wait_empty_entries())
1441 __process_trace_bufs();
1444 static void clean_trace_bufs(void)
1447 * No mutex needed here: we're only reading from the lists,
1451 __process_trace_bufs();
1454 static inline void read_err(int cpu, char *ifn)
1456 if (errno != EAGAIN)
1457 fprintf(stderr, "Thread %d failed read of %s: %d/%s\n",
1458 cpu, ifn, errno, strerror(errno));
1461 static int net_sendfile(struct io_info *iop)
1465 ret = sendfile(iop->ofd, iop->ifd, NULL, iop->ready);
1469 } else if (ret < (int)iop->ready) {
1470 fprintf(stderr, "short sendfile send (%d of %d)\n",
1478 static inline int net_sendfile_data(struct tracer *tp, struct io_info *iop)
1480 struct devpath *dpp = iop->dpp;
1482 if (net_send_header(iop->ofd, tp->cpu, dpp->buts_name, iop->ready))
1484 return net_sendfile(iop);
1487 static int fill_ofname(struct io_info *iop, int cpu)
1491 char *dst = iop->ofn;
1494 len = snprintf(iop->ofn, sizeof(iop->ofn), "%s/", output_dir);
1496 len = snprintf(iop->ofn, sizeof(iop->ofn), "./");
1498 if (net_mode == Net_server) {
1499 struct cl_conn *nc = iop->nc;
1501 len += sprintf(dst + len, "%s-", nc->ch->hostname);
1502 len += strftime(dst + len, 64, "%F-%T/",
1503 gmtime(&iop->dpp->cl_connect_time));
1506 if (stat(iop->ofn, &sb) < 0) {
1507 if (errno != ENOENT) {
1509 "Destination dir %s stat failed: %d/%s\n",
1510 iop->ofn, errno, strerror(errno));
1514 * There is no synchronization between multiple threads
1515 * trying to create the directory at once. It's harmless
1516 * to let them try, so just detect the problem and move on.
1518 if (mkdir(iop->ofn, 0755) < 0 && errno != EEXIST) {
1520 "Destination dir %s can't be made: %d/%s\n",
1521 iop->ofn, errno, strerror(errno));
1527 snprintf(iop->ofn + len, sizeof(iop->ofn), "%s.blktrace.%d",
1530 snprintf(iop->ofn + len, sizeof(iop->ofn), "%s.blktrace.%d",
1531 iop->dpp->buts_name, cpu);
1536 static int set_vbuf(struct io_info *iop, int mode, size_t size)
1538 iop->obuf = malloc(size);
1539 if (setvbuf(iop->ofp, iop->obuf, mode, size) < 0) {
1540 fprintf(stderr, "setvbuf(%s, %d) failed: %d/%s\n",
1541 iop->dpp->path, (int)size, errno,
1550 static int iop_open(struct io_info *iop, int cpu)
1553 if (fill_ofname(iop, cpu))
1556 iop->ofp = my_fopen(iop->ofn, "w+");
1557 if (iop->ofp == NULL) {
1558 fprintf(stderr, "Open output file %s failed: %d/%s\n",
1559 iop->ofn, errno, strerror(errno));
1563 if (set_vbuf(iop, _IOLBF, FILE_VBUF_SIZE)) {
1564 fprintf(stderr, "set_vbuf for file %s failed: %d/%s\n",
1565 iop->ofn, errno, strerror(errno));
1570 iop->ofd = fileno(iop->ofp);
1574 static void close_iop(struct io_info *iop)
1576 struct mmap_info *mip = &iop->mmap_info;
1579 munmap(mip->fs_buf, mip->fs_buf_len);
1581 if (!piped_output) {
1582 if (ftruncate(fileno(iop->ofp), mip->fs_size) < 0) {
1584 "Ignoring err: ftruncate(%s): %d/%s\n",
1585 iop->ofn, errno, strerror(errno));
1595 static void close_ios(struct tracer *tp)
1597 while (tp->nios > 0) {
1598 struct io_info *iop = &tp->ios[--tp->nios];
1600 iop->dpp->drops = get_drops(iop->dpp);
1606 else if (iop->ofd >= 0) {
1607 struct devpath *dpp = iop->dpp;
1609 net_send_close(iop->ofd, dpp->buts_name, dpp->drops);
1610 net_close_connection(&iop->ofd);
1618 static int open_ios(struct tracer *tp)
1621 struct io_info *iop;
1622 struct list_head *p;
1624 tp->ios = calloc(ndevs, sizeof(struct io_info));
1625 memset(tp->ios, 0, ndevs * sizeof(struct io_info));
1627 tp->pfds = calloc(ndevs, sizeof(struct pollfd));
1628 memset(tp->pfds, 0, ndevs * sizeof(struct pollfd));
1633 __list_for_each(p, &devpaths) {
1634 struct devpath *dpp = list_entry(p, struct devpath, head);
1638 snprintf(iop->ifn, sizeof(iop->ifn), "%s/block/%s/trace%d",
1639 debugfs_path, dpp->buts_name, tp->cpu);
1641 iop->ifd = my_open(iop->ifn, O_RDONLY | O_NONBLOCK);
1643 fprintf(stderr, "Thread %d failed open %s: %d/%s\n",
1644 tp->cpu, iop->ifn, errno, strerror(errno));
1648 init_mmap_info(&iop->mmap_info);
1651 pfd->events = POLLIN;
1655 else if (net_client_use_sendfile()) {
1656 iop->ofd = net_setup_client();
1659 net_send_open(iop->ofd, tp->cpu, dpp->buts_name);
1660 } else if (net_mode == Net_none) {
1661 if (iop_open(iop, tp->cpu))
1665 * This ensures that the server knows about all
1666 * connections & devices before _any_ closes
1668 net_send_open(cl_fds[tp->cpu], tp->cpu, dpp->buts_name);
1679 close(iop->ifd); /* tp->nios _not_ bumped */
1684 static int handle_pfds_file(struct tracer *tp, int nevs, int force_read)
1686 struct mmap_info *mip;
1687 int i, ret, nentries = 0;
1688 struct pollfd *pfd = tp->pfds;
1689 struct io_info *iop = tp->ios;
1691 for (i = 0; nevs > 0 && i < ndevs; i++, pfd++, iop++) {
1692 if (pfd->revents & POLLIN || force_read) {
1693 mip = &iop->mmap_info;
1695 ret = setup_mmap(iop->ofd, buf_size, mip, tp);
1701 ret = read(iop->ifd, mip->fs_buf + mip->fs_off,
1704 pdc_dr_update(iop->dpp, tp->cpu, ret);
1705 mip->fs_size += ret;
1708 } else if (ret == 0) {
1710 * Short reads after we're done stop us
1711 * from trying reads.
1716 read_err(tp->cpu, iop->ifn);
1717 if (errno != EAGAIN || tp->is_done)
1727 static int handle_pfds_netclient(struct tracer *tp, int nevs, int force_read)
1730 int i, nentries = 0;
1731 struct pdc_stats *sp;
1732 struct pollfd *pfd = tp->pfds;
1733 struct io_info *iop = tp->ios;
1735 for (i = 0; i < ndevs; i++, pfd++, iop++, sp++) {
1736 if (pfd->revents & POLLIN || force_read) {
1737 if (fstat(iop->ifd, &sb) < 0) {
1740 } else if (sb.st_size > (off_t)iop->data_queued) {
1741 iop->ready = sb.st_size - iop->data_queued;
1742 iop->data_queued = sb.st_size;
1744 if (!net_sendfile_data(tp, iop)) {
1745 pdc_dr_update(iop->dpp, tp->cpu,
1757 incr_entries(nentries);
1762 static int handle_pfds_entries(struct tracer *tp, int nevs, int force_read)
1764 int i, nentries = 0;
1765 struct trace_buf *tbp;
1766 struct pollfd *pfd = tp->pfds;
1767 struct io_info *iop = tp->ios;
1769 tbp = alloc_trace_buf(tp->cpu, buf_size);
1770 for (i = 0; i < ndevs; i++, pfd++, iop++) {
1771 if (pfd->revents & POLLIN || force_read) {
1772 tbp->len = read(iop->ifd, tbp->buf, buf_size);
1774 pdc_dr_update(iop->dpp, tp->cpu, tbp->len);
1775 add_trace_buf(iop->dpp, tp->cpu, &tbp);
1777 } else if (tbp->len == 0) {
1779 * Short reads after we're done stop us
1780 * from trying reads.
1785 read_err(tp->cpu, iop->ifn);
1786 if (errno != EAGAIN || tp->is_done)
1789 if (!piped_output && --nevs == 0)
1796 incr_entries(nentries);
1801 static void *thread_main(void *arg)
1803 int ret, ndone, to_val;
1804 struct tracer *tp = arg;
1806 ret = lock_on_cpu(tp->cpu);
1815 to_val = 50; /* Frequent partial handles */
1817 to_val = 500; /* 1/2 second intervals */
1820 tracer_signal_ready(tp, Th_running, 0);
1821 tracer_wait_unblock(tp);
1823 while (!tp->is_done) {
1824 ndone = poll(tp->pfds, ndevs, to_val);
1825 if (ndone || piped_output)
1826 (void)handle_pfds(tp, ndone, piped_output);
1827 else if (ndone < 0 && errno != EINTR)
1828 fprintf(stderr, "Thread %d poll failed: %d/%s\n",
1829 tp->cpu, errno, strerror(errno));
1833 * Trace is stopped, pull data until we get a short read
1835 while (handle_pfds(tp, ndevs, 1) > 0)
1839 tracer_signal_ready(tp, Th_leaving, 0);
1843 tracer_signal_ready(tp, Th_error, ret);
1847 static int start_tracer(int cpu)
1851 tp = malloc(sizeof(*tp));
1852 memset(tp, 0, sizeof(*tp));
1854 INIT_LIST_HEAD(&tp->head);
1858 if (pthread_create(&tp->thread, NULL, thread_main, tp)) {
1859 fprintf(stderr, "FAILED to start thread on CPU %d: %d/%s\n",
1860 cpu, errno, strerror(errno));
1865 list_add_tail(&tp->head, &tracers);
1869 static void start_tracers(void)
1872 struct list_head *p;
1874 for (cpu = 0; cpu < ncpus; cpu++)
1875 if (start_tracer(cpu))
1878 wait_tracers_ready(cpu);
1880 __list_for_each(p, &tracers) {
1881 struct tracer *tp = list_entry(p, struct tracer, head);
1884 "FAILED to start thread on CPU %d: %d/%s\n",
1885 tp->cpu, tp->status, strerror(tp->status));
1889 static void stop_tracers(void)
1891 struct list_head *p;
1894 * Stop the tracing - makes the tracer threads clean up quicker.
1896 __list_for_each(p, &devpaths) {
1897 struct devpath *dpp = list_entry(p, struct devpath, head);
1898 (void)ioctl(dpp->fd, BLKTRACESTOP);
1902 * Tell each tracer to quit
1904 __list_for_each(p, &tracers) {
1905 struct tracer *tp = list_entry(p, struct tracer, head);
1910 static void del_tracers(void)
1912 struct list_head *p, *q;
1914 list_for_each_safe(p, q, &tracers) {
1915 struct tracer *tp = list_entry(p, struct tracer, head);
1917 list_del(&tp->head);
1922 static void wait_tracers(void)
1924 struct list_head *p;
1926 if (use_tracer_devpaths())
1927 process_trace_bufs();
1929 wait_tracers_leaving();
1931 __list_for_each(p, &tracers) {
1933 struct tracer *tp = list_entry(p, struct tracer, head);
1935 ret = pthread_join(tp->thread, NULL);
1937 fprintf(stderr, "Thread join %d failed %d\n",
1941 if (use_tracer_devpaths())
1947 static void exit_tracing(void)
1949 signal(SIGINT, SIG_IGN);
1950 signal(SIGHUP, SIG_IGN);
1951 signal(SIGTERM, SIG_IGN);
1952 signal(SIGALRM, SIG_IGN);
1960 static void handle_sigint(__attribute__((__unused__)) int sig)
1966 static void show_stats(struct list_head *devpaths)
1969 struct list_head *p;
1970 unsigned long long nevents, data_read;
1971 unsigned long long total_drops = 0;
1972 unsigned long long total_events = 0;
1975 ofp = my_fopen("/dev/null", "w");
1979 __list_for_each(p, devpaths) {
1981 struct pdc_stats *sp;
1982 struct devpath *dpp = list_entry(p, struct devpath, head);
1984 if (net_mode == Net_server)
1985 printf("server: end of run for %s:%s\n",
1986 dpp->ch->hostname, dpp->buts_name);
1991 fprintf(ofp, "=== %s ===\n", dpp->buts_name);
1992 for (cpu = 0, sp = dpp->stats; cpu < dpp->ncpus; cpu++, sp++) {
1994 * Estimate events if not known...
1996 if (sp->nevents == 0) {
1997 sp->nevents = sp->data_read /
1998 sizeof(struct blk_io_trace);
2002 " CPU%3d: %20llu events, %8llu KiB data\n",
2003 cpu, sp->nevents, (sp->data_read + 1023) >> 10);
2005 data_read += sp->data_read;
2006 nevents += sp->nevents;
2009 fprintf(ofp, " Total: %20llu events (dropped %llu),"
2010 " %8llu KiB data\n", nevents,
2011 dpp->drops, (data_read + 1024) >> 10);
2013 total_drops += dpp->drops;
2014 total_events += (nevents + dpp->drops);
2022 double drops_ratio = 1.0;
2025 drops_ratio = (double)total_drops/(double)total_events;
2027 fprintf(stderr, "\nYou have %llu (%5.1lf%%) dropped events\n"
2028 "Consider using a larger buffer size (-b) "
2029 "and/or more buffers (-n)\n",
2030 total_drops, 100.0 * drops_ratio);
2034 static int handle_args(int argc, char *argv[])
2038 int act_mask_tmp = 0;
2040 while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) >= 0) {
2043 i = find_mask_map(optarg);
2045 fprintf(stderr, "Invalid action mask %s\n",
2053 if ((sscanf(optarg, "%x", &i) != 1) ||
2054 !valid_act_opt(i)) {
2056 "Invalid set action mask %s/0x%x\n",
2064 if (add_devpath(optarg) != 0)
2070 FILE *ifp = my_fopen(optarg, "r");
2074 "Invalid file for devices %s\n",
2079 while (fscanf(ifp, "%s\n", dev_line) == 1)
2080 if (add_devpath(dev_line) != 0)
2086 debugfs_path = optarg;
2090 output_name = optarg;
2093 kill_running_trace = 1;
2096 stop_watch = atoi(optarg);
2097 if (stop_watch <= 0) {
2099 "Invalid stopwatch value (%d secs)\n",
2106 printf("%s version %s\n", argv[0], blktrace_version);
2110 buf_size = strtoul(optarg, NULL, 10);
2111 if (buf_size <= 0 || buf_size > 16*1024) {
2112 fprintf(stderr, "Invalid buffer size (%lu)\n",
2119 buf_nr = strtoul(optarg, NULL, 10);
2122 "Invalid buffer nr (%lu)\n", buf_nr);
2127 output_dir = optarg;
2130 net_mode = Net_client;
2131 strcpy(hostname, optarg);
2134 net_mode = Net_server;
2137 net_port = atoi(optarg);
2140 net_use_sendfile = 0;
2143 show_usage(argv[0]);
2149 while (optind < argc)
2150 if (add_devpath(argv[optind++]) != 0)
2153 if (net_mode != Net_server && ndevs == 0) {
2154 show_usage(argv[0]);
2158 if (statfs(debugfs_path, &st) < 0 || st.f_type != (long)DEBUGFS_TYPE) {
2159 fprintf(stderr, "Invalid debug path %s: %d/%s\n",
2160 debugfs_path, errno, strerror(errno));
2164 if (act_mask_tmp != 0)
2165 act_mask = act_mask_tmp;
2167 if (net_mode == Net_client && net_setup_addr())
2171 * Set up for appropriate PFD handler based upon output name.
2173 if (net_client_use_sendfile())
2174 handle_pfds = handle_pfds_netclient;
2175 else if (net_client_use_send())
2176 handle_pfds = handle_pfds_entries;
2177 else if (output_name && (strcmp(output_name, "-") == 0)) {
2179 handle_pfds = handle_pfds_entries;
2181 setvbuf(pfp, NULL, _IONBF, 0);
2183 handle_pfds = handle_pfds_file;
2187 static void ch_add_connection(struct net_server_s *ns, struct cl_host *ch,
2192 nc = malloc(sizeof(*nc));
2193 memset(nc, 0, sizeof(*nc));
2195 time(&nc->connect_time);
2200 list_add_tail(&nc->ch_head, &ch->conn_list);
2203 list_add_tail(&nc->ns_head, &ns->conn_list);
2205 ns->pfds = realloc(ns->pfds, (ns->connects+1) * sizeof(struct pollfd));
2208 static void ch_rem_connection(struct net_server_s *ns, struct cl_host *ch,
2211 net_close_connection(&nc->fd);
2213 list_del(&nc->ch_head);
2216 list_del(&nc->ns_head);
2218 ns->pfds = realloc(ns->pfds, (ns->connects+1) * sizeof(struct pollfd));
2223 static struct cl_host *net_find_client_host(struct net_server_s *ns,
2224 struct in_addr cl_in_addr)
2226 struct list_head *p;
2228 __list_for_each(p, &ns->ch_list) {
2229 struct cl_host *ch = list_entry(p, struct cl_host, head);
2231 if (in_addr_eq(ch->cl_in_addr, cl_in_addr))
2238 static struct cl_host *net_add_client_host(struct net_server_s *ns,
2239 struct sockaddr_in *addr)
2243 ch = malloc(sizeof(*ch));
2244 memset(ch, 0, sizeof(*ch));
2247 ch->cl_in_addr = addr->sin_addr;
2248 list_add_tail(&ch->head, &ns->ch_list);
2251 ch->hostname = strdup(inet_ntoa(addr->sin_addr));
2252 printf("server: connection from %s\n", ch->hostname);
2254 INIT_LIST_HEAD(&ch->conn_list);
2255 INIT_LIST_HEAD(&ch->devpaths);
2260 static void device_done(struct devpath *dpp, int ncpus)
2263 struct io_info *iop;
2265 for (cpu = 0, iop = dpp->ios; cpu < ncpus; cpu++, iop++)
2268 list_del(&dpp->head);
2272 static void net_ch_remove(struct cl_host *ch, int ncpus)
2274 struct list_head *p, *q;
2275 struct net_server_s *ns = ch->ns;
2277 list_for_each_safe(p, q, &ch->devpaths) {
2278 struct devpath *dpp = list_entry(p, struct devpath, head);
2279 device_done(dpp, ncpus);
2282 list_for_each_safe(p, q, &ch->conn_list) {
2283 struct cl_conn *nc = list_entry(p, struct cl_conn, ch_head);
2285 ch_rem_connection(ns, ch, nc);
2288 list_del(&ch->head);
2296 static void net_add_connection(struct net_server_s *ns)
2300 socklen_t socklen = sizeof(ns->addr);
2302 fd = my_accept(ns->listen_fd, (struct sockaddr *)&ns->addr, &socklen);
2305 * This is OK: we just won't accept this connection,
2310 ch = net_find_client_host(ns, ns->addr.sin_addr);
2312 ch = net_add_client_host(ns, &ns->addr);
2314 ch_add_connection(ns, ch, fd);
2318 static struct devpath *nc_add_dpp(struct cl_conn *nc,
2319 struct blktrace_net_hdr *bnh,
2320 time_t connect_time)
2323 struct io_info *iop;
2324 struct devpath *dpp;
2326 dpp = malloc(sizeof(*dpp));
2327 memset(dpp, 0, sizeof(*dpp));
2329 dpp->buts_name = strdup(bnh->buts_name);
2330 dpp->path = strdup(bnh->buts_name);
2333 dpp->cl_id = bnh->cl_id;
2334 dpp->cl_connect_time = connect_time;
2335 dpp->ncpus = nc->ncpus;
2336 dpp->stats = calloc(dpp->ncpus, sizeof(*dpp->stats));
2337 memset(dpp->stats, 0, dpp->ncpus * sizeof(*dpp->stats));
2339 list_add_tail(&dpp->head, &nc->ch->devpaths);
2342 dpp->ios = calloc(nc->ncpus, sizeof(*iop));
2343 memset(dpp->ios, 0, ndevs * sizeof(*iop));
2345 for (cpu = 0, iop = dpp->ios; cpu < nc->ncpus; cpu++, iop++) {
2348 init_mmap_info(&iop->mmap_info);
2350 if (iop_open(iop, cpu))
2358 * Need to unravel what's been done...
2361 close_iop(&dpp->ios[cpu--]);
2367 static struct devpath *nc_find_dpp(struct cl_conn *nc,
2368 struct blktrace_net_hdr *bnh)
2370 struct list_head *p;
2371 time_t connect_time = nc->connect_time;
2373 __list_for_each(p, &nc->ch->devpaths) {
2374 struct devpath *dpp = list_entry(p, struct devpath, head);
2376 if (!strcmp(dpp->buts_name, bnh->buts_name))
2379 if (dpp->cl_id == bnh->cl_id)
2380 connect_time = dpp->cl_connect_time;
2383 return nc_add_dpp(nc, bnh, connect_time);
2386 static void net_client_read_data(struct cl_conn *nc, struct devpath *dpp,
2387 struct blktrace_net_hdr *bnh)
2390 struct io_info *iop = &dpp->ios[bnh->cpu];
2391 struct mmap_info *mip = &iop->mmap_info;
2393 if (setup_mmap(iop->ofd, bnh->len, &iop->mmap_info, NULL)) {
2394 fprintf(stderr, "ncd(%s:%d): mmap failed\n",
2395 nc->ch->hostname, nc->fd);
2399 ret = net_recv_data(nc->fd, mip->fs_buf + mip->fs_off, bnh->len);
2401 pdc_dr_update(dpp, bnh->cpu, ret);
2402 mip->fs_size += ret;
2409 * Returns 1 if we closed a host - invalidates other polling information
2410 * that may be present.
2412 static int net_client_data(struct cl_conn *nc)
2415 struct devpath *dpp;
2416 struct blktrace_net_hdr bnh;
2418 ret = net_get_header(nc, &bnh);
2423 fprintf(stderr, "ncd(%d): header read failed\n", nc->fd);
2427 if (data_is_native == -1 && check_data_endianness(bnh.magic)) {
2428 fprintf(stderr, "ncd(%d): received data is bad\n", nc->fd);
2432 if (!data_is_native) {
2433 bnh.magic = be32_to_cpu(bnh.magic);
2434 bnh.cpu = be32_to_cpu(bnh.cpu);
2435 bnh.max_cpus = be32_to_cpu(bnh.max_cpus);
2436 bnh.len = be32_to_cpu(bnh.len);
2437 bnh.cl_id = be32_to_cpu(bnh.cl_id);
2438 bnh.buf_size = be32_to_cpu(bnh.buf_size);
2439 bnh.buf_nr = be32_to_cpu(bnh.buf_nr);
2440 bnh.page_size = be32_to_cpu(bnh.page_size);
2443 if ((bnh.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
2444 fprintf(stderr, "ncd(%s:%d): bad data magic\n",
2445 nc->ch->hostname, nc->fd);
2449 if (nc->ncpus == -1)
2450 nc->ncpus = bnh.max_cpus;
2453 * len == 0 means the other end is sending us a new connection/dpp
2454 * len == 1 means that the other end signalled end-of-run
2456 dpp = nc_find_dpp(nc, &bnh);
2459 * Just adding in the dpp above is enough
2461 ack_open_close(nc->fd, dpp->buts_name);
2463 } else if (bnh.len == 1) {
2465 * overload cpu count with dropped events
2467 dpp->drops = bnh.cpu;
2469 ack_open_close(nc->fd, dpp->buts_name);
2470 if (--nc->ch->cl_opens == 0) {
2471 show_stats(&nc->ch->devpaths);
2472 net_ch_remove(nc->ch, nc->ncpus);
2476 net_client_read_data(nc, dpp, &bnh);
2481 static void handle_client_data(struct net_server_s *ns, int events)
2485 struct list_head *p, *q;
2488 list_for_each_safe(p, q, &ns->conn_list) {
2489 if (pfd->revents & POLLIN) {
2490 nc = list_entry(p, struct cl_conn, ns_head);
2492 if (net_client_data(nc) || --events == 0)
2499 static void net_setup_pfds(struct net_server_s *ns)
2502 struct list_head *p;
2504 ns->pfds[0].fd = ns->listen_fd;
2505 ns->pfds[0].events = POLLIN;
2508 __list_for_each(p, &ns->conn_list) {
2509 struct cl_conn *nc = list_entry(p, struct cl_conn, ns_head);
2512 pfd->events = POLLIN;
2517 static int net_server_handle_connections(struct net_server_s *ns)
2521 printf("server: waiting for connections...\n");
2525 events = poll(ns->pfds, ns->connects + 1, -1);
2527 if (errno != EINTR) {
2528 perror("FATAL: poll error");
2531 } else if (events > 0) {
2532 if (ns->pfds[0].revents & POLLIN) {
2533 net_add_connection(ns);
2538 handle_client_data(ns, events);
2545 static int net_server(void)
2549 struct net_server_s net_server;
2550 struct net_server_s *ns = &net_server;
2552 memset(ns, 0, sizeof(*ns));
2553 INIT_LIST_HEAD(&ns->ch_list);
2554 INIT_LIST_HEAD(&ns->conn_list);
2555 ns->pfds = malloc(sizeof(struct pollfd));
2557 fd = my_socket(AF_INET, SOCK_STREAM, 0);
2559 perror("server: socket");
2564 if (setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)) < 0) {
2565 perror("setsockopt");
2569 memset(&ns->addr, 0, sizeof(ns->addr));
2570 ns->addr.sin_family = AF_INET;
2571 ns->addr.sin_addr.s_addr = htonl(INADDR_ANY);
2572 ns->addr.sin_port = htons(net_port);
2574 if (bind(fd, (struct sockaddr *) &ns->addr, sizeof(ns->addr)) < 0) {
2579 if (listen(fd, 1) < 0) {
2585 * The actual server looping is done here:
2588 ret = net_server_handle_connections(ns);
2591 * Clean up and return...
2598 static int run_tracers(void)
2600 atexit(exit_tracing);
2601 if (net_mode == Net_client)
2602 printf("blktrace: connecting to %s\n", hostname);
2606 if (use_tracer_devpaths()) {
2607 if (setup_tracer_devpaths())
2611 handle_list = handle_list_file;
2613 handle_list = handle_list_net;
2617 if (nthreads_running == ncpus) {
2620 if (net_mode == Net_client)
2621 printf("blktrace: connected!\n");
2628 if (nthreads_running == ncpus)
2629 show_stats(&devpaths);
2630 if (net_client_use_send())
2631 close_client_connections();
2637 int main(int argc, char *argv[])
2641 setlocale(LC_NUMERIC, "en_US");
2642 pagesize = getpagesize();
2643 ncpus = sysconf(_SC_NPROCESSORS_ONLN);
2645 fprintf(stderr, "sysconf(_SC_NPROCESSORS_ONLN) failed %d/%s\n",
2646 errno, strerror(errno));
2649 } else if (handle_args(argc, argv)) {
2654 if (ndevs > 1 && output_name && strcmp(output_name, "-") != 0) {
2655 fprintf(stderr, "-o not supported with multiple devices\n");
2660 signal(SIGINT, handle_sigint);
2661 signal(SIGHUP, handle_sigint);
2662 signal(SIGTERM, handle_sigint);
2663 signal(SIGALRM, handle_sigint);
2664 signal(SIGPIPE, SIG_IGN);
2666 if (kill_running_trace) {
2667 struct devpath *dpp;
2668 struct list_head *p;
2670 __list_for_each(p, &devpaths) {
2671 dpp = list_entry(p, struct devpath, head);
2672 if (__stop_trace(dpp->fd)) {
2674 "BLKTRACETEARDOWN %s failed: %d/%s\n",
2675 dpp->path, errno, strerror(errno));
2678 } else if (net_mode == Net_server) {
2680 fprintf(stderr, "-o ignored in server mode\n");
2685 ret = run_tracers();