Files
mscp/src/main.c
Ryo Nakamura 3d98451bba set default nr_threads to floor(log(cores) * 2) + 1)
This change prevents mscp from establishing too many ssh connections
on many-core machines in default.
2022-12-11 14:01:52 +09:00

713 lines
16 KiB
C

#include <stdio.h>
#include <stdlib.h>
#include <stdbool.h>
#include <unistd.h>
#include <signal.h>
#include <sys/time.h>
#include <sys/ioctl.h>
#include <math.h>
#include <pthread.h>
#include <list.h>
#include <util.h>
#include <ssh.h>
#include <file.h>
#include <pprint.h>
#include <atomic.h>
#include <platform.h>
#ifndef _VERSION /* passed through cmake */
#define VERSION "(unknown)"
#else
#define VERSION _VERSION
#endif
#define DEFAULT_MIN_CHUNK_SZ (64 << 20) /* 64MB */
#define DEFAULT_NR_AHEAD 32
#define DEFAULT_BUF_SZ 16384
/* XXX: we use 16384 byte buffer pointed by
* https://api.libssh.org/stable/libssh_tutor_sftp.html. The larget
* read length from sftp_async_read is 65536 byte. Read sizes larger
* than 65536 cause a situation where data remainds but
* sftp_async_read returns 0.
*/
struct mscp_thread {
sftp_session sftp;
pthread_t tid;
int cpu;
size_t done; /* copied bytes */
bool finished;
int ret;
};
struct mscp {
char *host; /* remote host (and username) */
struct ssh_opts *opts; /* ssh parameters */
struct list_head file_list;
struct list_head chunk_list; /* stack of chunks */
lock chunk_lock; /* lock for chunk list */
char *target;
int nr_threads; /* number of threads */
int buf_sz; /* i/o buf size */
int nr_ahead; /* # of ahead read command for remote to local copy */
struct mscp_thread *threads;
} m;
void *mscp_copy_thread(void *arg);
int mscp_stat_init();
void mscp_stat_final();
void stop_copy_threads(int sig)
{
int n;
pr("stopping...\n");
for (n = 0; n < m.nr_threads; n++) {
if (m.threads[n].tid && !m.threads[n].finished)
pthread_cancel(m.threads[n].tid);
}
}
void usage(bool print_help) {
printf("mscp v" VERSION ": copy files over multiple ssh connections\n"
"\n"
"Usage: mscp [vqDCHdNh] [-n nr_conns] [-m coremask]\n"
" [-s min_chunk_sz] [-S max_chunk_sz] [-a nr_ahead] [-b buf_sz]\n"
" [-l login_name] [-p port] [-i identity_file]\n"
" [-c cipher_spec] [-M hmac_spec] source ... target\n"
"\n");
if (!print_help)
return;
printf(" -n NR_CONNECTIONS number of connections "
"(default: floor(log(cores)*2)+1)\n"
" -m COREMASK hex value to specify cores where threads pinned\n"
" -s MIN_CHUNK_SIZE min chunk size (default: 64MB)\n"
" -S MAX_CHUNK_SIZE max chunk size (default: filesize/nr_conn)\n"
"\n"
" -a NR_AHEAD number of inflight SFTP commands (default: 32)\n"
" -b BUF_SZ buffer size for i/o and transfer\n"
"\n"
" -v increment verbose output level\n"
" -q disable output\n"
" -D dry run\n"
"\n"
" -l LOGIN_NAME login name\n"
" -p PORT port number\n"
" -i IDENTITY identity file for public key authentication\n"
" -c CIPHER cipher spec\n"
" -M HMAC hmac spec\n"
" -C enable compression on libssh\n"
" -H disable hostkey check\n"
" -d increment ssh debug output level\n"
" -N disable tcp nodelay (default on)\n"
" -h print this help\n"
"\n");
}
char *find_hostname(int ind, int argc, char **argv)
{
char *h, *hostnames[argc];
int n, cnt = 0;
for (n = ind; n < argc; n++) {
h = file_find_hostname(argv[n]);
if (h)
hostnames[cnt++] = h;
}
if (cnt == 0)
return NULL;
/* check all hostnames are identical */
for (n = 1; n < cnt; n++) {
int s1 = strlen(hostnames[n - 1]);
int s2 = strlen(hostnames[n]);
if (s1 != s2) {
pr_err("different hostnames: %s and %s\n",
hostnames[n - 1], hostnames[n]);
goto err_out;
}
if (strncmp(hostnames[n - 1], hostnames[n], s1) != 0) {
pr_err("different hostnames: %s and %s\n",
hostnames[n - 1], hostnames[n]);
goto err_out;
}
}
for (n = 1; n < cnt; n++) {
free(hostnames[n]);
}
return hostnames[0];
err_out:
for (n = 0; n < cnt; n++) {
free(hostnames[n]);
}
return NULL;
}
int expand_coremask(const char *coremask, int **cores, int *nr_cores)
{
int n, *core_list, core_list_len = 0, nr_usable, nr_all;
char c[2] = { 'x', '\0' };
const char *_coremask;
long v, needle;
int ncores = nr_cpus();
/*
* This function returns array of usable cores in `cores` and
* returns the number of usable cores (array length) through
* nr_cores.
*/
if (strncmp(coremask, "0x", 2) == 0)
_coremask = coremask + 2;
else
_coremask = coremask;
core_list = realloc(NULL, sizeof(int) * 64);
if (!core_list) {
pr_err("failed to realloc: %s\n", strerrno());
return -1;
}
nr_usable = 0;
nr_all = 0;
for (n = strlen(_coremask) - 1; n >=0; n--) {
c[0] = _coremask[n];
v = strtol(c, NULL, 16);
if (v == LONG_MIN || v == LONG_MAX) {
pr_err("invalid coremask: %s\n", coremask);
return -1;
}
for (needle = 0x01; needle < 0x10; needle <<= 1) {
nr_all++;
if (nr_all > ncores)
break; /* too long coremask */
if (v & needle) {
nr_usable++;
core_list = realloc(core_list, sizeof(int) * nr_usable);
if (!core_list) {
pr_err("failed to realloc: %s\n", strerrno());
return -1;
}
core_list[nr_usable - 1] = nr_all - 1;
}
}
}
if (nr_usable < 1) {
pr_err("invalid core mask: %s\n", coremask);
return -1;
}
*cores = core_list;
*nr_cores = nr_usable;
return 0;
}
int default_nr_threads()
{
return (int)(floor(log(nr_cpus()) * 2) + 1);
}
int main(int argc, char **argv)
{
struct ssh_opts opts;
sftp_session ctrl;
int min_chunk_sz = DEFAULT_MIN_CHUNK_SZ;
int max_chunk_sz = 0;
char *coremask = NULL;;
int verbose = 1;
bool dryrun = false;
int ret = 0, n;
int *cores, nr_cores;
char ch;
memset(&opts, 0, sizeof(opts));
opts.nodelay = 1;
memset(&m, 0, sizeof(m));
INIT_LIST_HEAD(&m.file_list);
INIT_LIST_HEAD(&m.chunk_list);
lock_init(&m.chunk_lock);
m.nr_ahead = DEFAULT_NR_AHEAD;
m.buf_sz = DEFAULT_BUF_SZ;
m.nr_threads = default_nr_threads();
while ((ch = getopt(argc, argv, "n:m:s:S:a:b:vqDl:p:i:c:M:CHdNh")) != -1) {
switch (ch) {
case 'n':
m.nr_threads = atoi(optarg);
if (m.nr_threads < 1) {
pr_err("invalid number of connections: %s\n", optarg);
return 1;
}
break;
case 'm':
coremask = optarg;
break;
case 's':
min_chunk_sz = atoi(optarg);
if (min_chunk_sz < getpagesize()) {
pr_err("min chunk size must be "
"larger than or equal to %d: %s\n",
getpagesize(), optarg);
return 1;
}
if (min_chunk_sz % getpagesize() != 0) {
pr_err("min chunk size must be "
"multiple of page size %d: %s\n",
getpagesize(), optarg);
return -1;
}
break;
case 'S':
max_chunk_sz = atoi(optarg);
if (max_chunk_sz < getpagesize()) {
pr_err("max chunk size must be "
"larger than or equal to %d: %s\n",
getpagesize(), optarg);
return 1;
}
if (max_chunk_sz % getpagesize() != 0) {
pr_err("max chunk size must be "
"multiple of page size %d: %s\n",
getpagesize(), optarg);
return -1;
}
break;
case 'a':
m.nr_ahead = atoi(optarg);
if (m.nr_ahead < 1) {
pr_err("invalid number of ahead: %s\n", optarg);
return -1;
}
break;
case 'b':
m.buf_sz = atoi(optarg);
if (m.buf_sz < 1) {
pr_err("invalid buffer size: %s\n", optarg);
return -1;
}
break;
case 'v':
verbose++;
break;
case 'q':
verbose = -1;
break;
case 'D':
dryrun = true;
break;
case 'l':
opts.login_name = optarg;
break;
case 'p':
opts.port = optarg;
break;
case 'i':
opts.identity = optarg;
break;
case 'c':
opts.cipher = optarg;
break;
case 'M':
opts.hmac = optarg;
break;
case 'C':
opts.compress++;
break;
case 'H':
opts.no_hostkey_check = true;
break;
case 'd':
opts.debuglevel++;
break;
case 'N':
opts.nodelay = 0;
break;
case 'h':
usage(true);
return 0;
default:
usage(false);
return 1;
}
}
pprint_set_level(verbose);
if (argc - optind < 2) {
/* mscp needs at lease 2 (src and target) argument */
usage(false);
return 1;
}
m.target = argv[argc - 1];
if (max_chunk_sz > 0 && min_chunk_sz > max_chunk_sz) {
pr_err("smaller max chunk size than min chunk size: %d < %d\n",
max_chunk_sz, min_chunk_sz);
return 1;
}
/* expand usable cores from coremask */
if (coremask) {
if (expand_coremask(coremask, &cores, &nr_cores) < 0)
return -1;
pprint(2, "cpu cores:");
for (n = 0; n < nr_cores; n++)
pprint(2, " %d", cores[n]);
pprint(2, "\n");
}
pprint2("number of connections: %d\n", m.nr_threads);
/* create control session */
m.host = find_hostname(optind, argc, argv);
if (!m.host) {
pr_err("no remote host given\n");
return 1;
}
pprint3("connecting to %s for checking destinations...\n", m.host);
ctrl = ssh_init_sftp_session(m.host, &opts);
if (!ctrl)
return 1;
m.opts = &opts; /* save ssh-able ssh_opts */
/* fill file list */
ret = file_fill(ctrl, &m.file_list, &argv[optind], argc - optind - 1, m.target);
if (ret < 0)
goto out;
#ifdef DEBUG
file_dump(&m.file_list);
#endif
/* fill chunk list */
ret = chunk_fill(&m.file_list, &m.chunk_list,
m.nr_threads, min_chunk_sz, max_chunk_sz);
if (ret < 0)
goto out;
#ifdef DEBUG
chunk_dump(&m.chunk_list);
#endif
if (dryrun) {
ssh_sftp_close(ctrl);
return 0;
}
/* prepare thread instances */
if ((n = list_count(&m.chunk_list)) < m.nr_threads) {
pprint2("we have only %d chunk(s). "
"set number of connections to %d\n", n, n);
m.nr_threads = n;
}
m.threads = calloc(m.nr_threads, sizeof(struct mscp_thread));
memset(m.threads, 0, m.nr_threads * sizeof(struct mscp_thread));
for (n = 0; n < m.nr_threads; n++) {
struct mscp_thread *t = &m.threads[n];
t->finished = false;
if (!coremask)
t->cpu = -1;
else
t->cpu = cores[n % nr_cores];
if (n == 0) {
t->sftp = ctrl; /* reuse ctrl sftp session */
ctrl = NULL;
} else {
pprint3("connecting to %s for a copy thread...\n", m.host);
t->sftp = ssh_init_sftp_session(m.host, m.opts);
}
if (!t->sftp) {
ret = 1;
goto out;
}
}
/* init mscp stat for printing progress bar */
if (mscp_stat_init() < 0) {
ret = 1;
goto out;
}
/* spawn copy threads */
for (n = 0; n < m.nr_threads; n++) {
struct mscp_thread *t = &m.threads[n];
ret = pthread_create(&t->tid, NULL, mscp_copy_thread, t);
if (ret < 0) {
pr_err("pthread_create error: %d\n", ret);
stop_copy_threads(0);
ret = 1;
goto join_out;
}
}
/* register SIGINT to stop threads */
if (signal(SIGINT, stop_copy_threads) == SIG_ERR) {
pr_err("cannot set signal: %s\n", strerrno());
ret = 1;
goto out;
}
join_out:
/* waiting for threads join... */
for (n = 0; n < m.nr_threads; n++) {
if (m.threads[n].tid) {
pthread_join(m.threads[n].tid, NULL);
if (m.threads[n].ret < 0)
ret = m.threads[n].ret;
}
}
/* print final result */
mscp_stat_final();
out:
if (ctrl)
ssh_sftp_close(ctrl);
if (m.threads) {
for (n = 0; n < m.nr_threads; n++) {
struct mscp_thread *t = &m.threads[n];
if (t->sftp)
ssh_sftp_close(t->sftp);
}
}
return ret;
}
void mscp_copy_thread_cleanup(void *arg)
{
struct mscp_thread *t = arg;
t->finished = true;
}
void *mscp_copy_thread(void *arg)
{
struct mscp_thread *t = arg;
sftp_session sftp = t->sftp;
struct chunk *c;
if (t->cpu > -1) {
if (set_thread_affinity(pthread_self(), t->cpu) < 0)
return NULL;
}
pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, NULL);
pthread_setcanceltype(PTHREAD_CANCEL_DEFERRED, NULL);
pthread_cleanup_push(mscp_copy_thread_cleanup, t);
while (1) {
LOCK_ACQUIRE_THREAD(&m.chunk_lock);
c = chunk_acquire(&m.chunk_list);
LOCK_RELEASE_THREAD();
if (!c)
break; /* no more chunks */
if ((t->ret = chunk_prepare(c, sftp)) < 0)
break;
if ((t->ret = chunk_copy(c, sftp, m.nr_ahead, m.buf_sz, &t->done)) < 0)
break;
}
pthread_cleanup_pop(1);
if (t->ret < 0)
pr_err("copy failed: chunk %s 0x%010lx-0x%010lx\n",
c->f->src_path, c->off, c->off + c->len);
return NULL;
}
/* progress bar-related functions */
double calculate_timedelta(struct timeval *b, struct timeval *a)
{
double sec, usec;
if (a->tv_usec < b->tv_usec) {
a->tv_usec += 1000000;
a->tv_sec--;
}
sec = a->tv_sec - b->tv_sec;
usec = a->tv_usec - b->tv_usec;
sec += usec / 1000000;
return sec;
}
double calculate_bps(size_t diff, struct timeval *b, struct timeval *a)
{
return (double)diff / calculate_timedelta(b, a);
}
char *calculate_eta(size_t remain, size_t diff, struct timeval *b, struct timeval *a)
{
static char buf[16];
double elapsed = calculate_timedelta(b, a);
double eta;
if (diff == 0)
snprintf(buf, sizeof(buf), "--:-- ETA");
else {
eta = remain / (diff / elapsed);
snprintf(buf, sizeof(buf), "%02d:%02d ETA",
(int)floor(eta / 60), (int)round(eta) % 60);
}
return buf;
}
void print_progress_bar(double percent, char *suffix)
{
int n, thresh, bar_width;
struct winsize ws;
char buf[128];
/*
* [=======> ] XX% SUFFIX
*/
buf[0] = '\0';
if (ioctl(STDOUT_FILENO, TIOCGWINSZ, &ws) < 0)
return; /* XXX */
bar_width = min(sizeof(buf), ws.ws_col) - strlen(suffix) - 7;
memset(buf, 0, sizeof(buf));
if (bar_width > 8) {
thresh = floor(bar_width * (percent / 100)) - 1;
for (n = 1; n < bar_width - 1; n++) {
if (n <= thresh)
buf[n] = '=';
else
buf[n] = ' ';
}
buf[thresh] = '>';
buf[0] = '[';
buf[bar_width - 1] = ']';
snprintf(buf + bar_width, sizeof(buf) - bar_width,
" %3d%% ", (int)floor(percent));
}
pprint1("%s%s", buf, suffix);
}
void print_progress(struct timeval *b, struct timeval *a,
size_t total, size_t last, size_t done)
{
char *bps_units[] = { "B/s ", "KB/s", "MB/s", "GB/s" };
char *byte_units[] = { "B ", "KB", "MB", "GB", "TB", "PB" };
char suffix[128];
int bps_u, byte_tu, byte_du;
size_t total_round, done_round;
int percent;
double bps;
#define array_size(a) (sizeof(a) / sizeof(a[0]))
if (total <= 0) {
pprint1("total 0 byte transferred");
return; /* copy 0-byte file(s) */
}
total_round = total;
for (byte_tu = 0; total_round > 1000 && byte_tu < array_size(byte_units) - 1;
byte_tu++)
total_round /= 1024;
bps = calculate_bps(done - last, b, a);
for (bps_u = 0; bps > 1000 && bps_u < array_size(bps_units); bps_u++)
bps /= 1000;
percent = floor(((double)(done) / (double)total) * 100);
done_round = done;
for (byte_du = 0; done_round > 1000 && byte_du < array_size(byte_units) - 1;
byte_du++)
done_round /= 1024;
snprintf(suffix, sizeof(suffix), "%4lu%s/%lu%s %6.1f%s %s",
done_round, byte_units[byte_du], total_round, byte_units[byte_tu],
bps, bps_units[bps_u], calculate_eta(total - done, done - last, b, a));
print_progress_bar(percent, suffix);
}
struct mscp_stat {
struct timeval start, before, after;
size_t total;
size_t last;
size_t done;
} s;
void mscp_stat_handler(int signum)
{
int n;
for (s.done = 0, n = 0; n < m.nr_threads; n++)
s.done += m.threads[n].done;
gettimeofday(&s.after, NULL);
if (signum == SIGALRM) {
alarm(1);
print_progress(&s.before, &s.after, s.total, s.last, s.done);
s.before = s.after;
s.last = s.done;
} else {
/* called from mscp_stat_final. calculate progress from the beginning */
print_progress(&s.start, &s.after, s.total, 0, s.done);
}
}
int mscp_stat_init()
{
struct file *f;
memset(&s, 0, sizeof(s));
list_for_each_entry(f, &m.file_list, list) {
s.total += f->size;
}
if (signal(SIGALRM, mscp_stat_handler) == SIG_ERR) {
pr_err("signal: %s\n", strerrno());
return -1;
}
gettimeofday(&s.start, NULL);
s.before = s.start;
alarm(1);
return 0;
}
void mscp_stat_final()
{
alarm(0);
mscp_stat_handler(0);
}