void *writer(void *unused) {
jlog_ctx *ctx;
int i;
char foo[72];
ctx = jlog_new(LOGNAME);
memset(foo, 'X', sizeof(foo)-1);
foo[sizeof(foo)-1] = '\0';
if(jlog_ctx_open_writer(ctx) != 0) {
fprintf(stderr, "jlog_ctx_open_writer failed: %d %s\n", jlog_ctx_err(ctx), jlog_ctx_err_string(ctx));
croak();
}
#ifdef TEST_UNIT_LIMIT
newsize = 1024*1024 + (((intptr_t)unused) % (1024 * 1024));
fprintf(stderr, "writer setting new unit_limit to %d\n", (int)newsize);
jlog_ctx_alter_journal_size(ctx, newsize);
#endif
for(i=0;i<10000;i++) {
int rv;
rv = jlog_ctx_write(ctx, foo, strlen(foo));
if(rv != 0) {
fprintf(stderr, "jlog_ctx_write_message failed: %d %s\n", jlog_ctx_err(ctx), jlog_ctx_err_string(ctx));
/* abort(); */
}
usleep(100);
}
fprintf(stderr, "writer thinks unit_limit is %d\n", ctx->meta->unit_limit);
jlog_ctx_close(ctx);
writer_done = 1;
return 0;
}
void jcreate(jlog_safety s) {
jlog_ctx *ctx;
const char *label = NULL;
switch (s) {
case JLOG_ALMOST_SAFE: label = "almost safe"; break;
case JLOG_UNSAFE: label = "unsafe"; break;
case JLOG_SAFE: label = "safe"; break;
}
fprintf(stderr, "jcreate %s in %s mode\n", LOGNAME, label);
ctx = jlog_new(LOGNAME);
jlog_ctx_alter_journal_size(ctx, 102400);
jlog_ctx_alter_safety(ctx, s);
if(jlog_ctx_init(ctx) != 0) {
fprintf(stderr, "jlog_ctx_init failed: %d %s\n", jlog_ctx_err(ctx), jlog_ctx_err_string(ctx));
if(jlog_ctx_err(ctx) != JLOG_ERR_CREATE_EXISTS) exit(0);
} else {
jlog_ctx_add_subscriber(ctx, MASTER, JLOG_BEGIN);
}
jlog_ctx_close(ctx);
}
static fq_msg *queue_jlog_dequeue(fqd_queue_impl_data f) {
struct queue_jlog *d = (struct queue_jlog *)f;
jlog_message msg;
fq_msg *m;
if(d->count == 0 && d->last_seen_nenqueued == d->nenqueued) return NULL;
retry:
if(d->count <= 0) {
d->count = jlog_ctx_read_interval(d->reader, &d->start, &d->finish);
fq_debug(FQ_DEBUG_IO, "jlog read batch count -> %d\n", d->count);
if(d->count < 0) {
char idxfile[PATH_MAX];
fq_debug(FQ_DEBUG_IO, "jlog_ctx_read_interval: %s\n",
jlog_ctx_err_string(d->reader));
switch (jlog_ctx_err(d->reader)) {
case JLOG_ERR_FILE_CORRUPT:
case JLOG_ERR_IDX_CORRUPT:
jlog_repair_datafile(d->reader, d->start.log);
jlog_repair_datafile(d->reader, d->start.log + 1);
fq_debug(FQ_DEBUG_IO,
"jlog reconstructed, deleting corresponding index.\n");
STRSETDATAFILE(d->reader, idxfile, d->start.log);
strncpy(idxfile + strlen(idxfile), INDEX_EXT, sizeof(idxfile) - strlen(idxfile));
unlink(idxfile);
STRSETDATAFILE(d->reader, idxfile, d->start.log + 1);
strncpy(idxfile + strlen(idxfile), INDEX_EXT, sizeof(idxfile) - strlen(idxfile));
unlink(idxfile);
break;
default:
break;
}
}
if(d->count <= 0) return NULL;
}
if(jlog_ctx_read_message(d->reader, &d->start, &msg) == -1) {
d->count = 0;
return NULL;
}
if(d->last_dequeued.log > d->start.log ||
(d->last_dequeued.log == d->start.log &&
d->last_dequeued.marker > d->start.marker)) {
d->count--;
JLOG_ID_ADVANCE(&d->start);
goto retry;
}
if(msg.mess_len < sizeof(fq_msg)-1)
m = NULL;
else {
off_t expected_len;
uint32_t payload_len;
m = (fq_msg *)msg.mess;
memcpy(&payload_len, &m->payload_len, sizeof(m->payload_len));
expected_len = offsetof(fq_msg, payload) + payload_len;
if(expected_len != msg.mess_len) m = NULL;
else {
m = malloc(expected_len);
memcpy(m, msg.mess, expected_len);
m->sender_msgid.id.u32.p3 = d->start.log;
m->sender_msgid.id.u32.p4 = d->start.marker;
}
}
d->count--;
fq_debug(FQ_DEBUG_IO, "jlog batch count -> %d\n", d->count);
if(d->count == 0) {
if(d->auto_chkpt) {
jlog_ctx_read_checkpoint(d->reader, &d->start);
}
}
d->last_dequeued = d->start;
JLOG_ID_ADVANCE(&d->start);
ck_pr_inc_uint(&d->last_seen_nenqueued);
return m;
}
void *
noit_jlog_thread_main(void *e_vptr) {
int mask, bytes_read;
eventer_t e = e_vptr;
acceptor_closure_t *ac = e->closure;
noit_jlog_closure_t *jcl = ac->service_ctx;
char inbuff[sizeof(jlog_id)];
eventer_set_fd_blocking(e->fd);
while(1) {
jlog_id client_chkpt;
int sleeptime = (ac->cmd == NOIT_JLOG_DATA_TEMP_FEED) ?
1 : DEFAULT_SECONDS_BETWEEN_BATCHES;
jlog_get_checkpoint(jcl->jlog, ac->remote_cn, &jcl->chkpt);
jcl->count = jlog_ctx_read_interval(jcl->jlog, &jcl->start, &jcl->finish);
if(jcl->count < 0) {
char idxfile[PATH_MAX];
noitL(noit_error, "jlog_ctx_read_interval: %s\n",
jlog_ctx_err_string(jcl->jlog));
switch (jlog_ctx_err(jcl->jlog)) {
case JLOG_ERR_FILE_CORRUPT:
case JLOG_ERR_IDX_CORRUPT:
jlog_repair_datafile(jcl->jlog, jcl->start.log);
jlog_repair_datafile(jcl->jlog, jcl->start.log + 1);
noitL(noit_error,
"jlog reconstructed, deleting corresponding index.\n");
STRSETDATAFILE(jcl->jlog, idxfile, jcl->start.log);
strlcat(idxfile, INDEX_EXT, sizeof(idxfile));
unlink(idxfile);
STRSETDATAFILE(jcl->jlog, idxfile, jcl->start.log + 1);
strlcat(idxfile, INDEX_EXT, sizeof(idxfile));
unlink(idxfile);
goto alldone;
break;
default:
goto alldone;
}
}
if(jcl->count > MAX_ROWS_AT_ONCE) {
/* Artificially set down the range to make the batches a bit easier
* to handle on the stratcond/postgres end.
* However, we must have more data, so drop the sleeptime to 0
*/
jcl->count = MAX_ROWS_AT_ONCE;
jcl->finish.marker = jcl->start.marker + jcl->count;
sleeptime = 0;
}
if(jcl->count > 0) {
if(noit_jlog_push(e, jcl)) {
goto alldone;
}
/* Read our jlog_id accounting for possibly short reads */
bytes_read = 0;
while(bytes_read < sizeof(jlog_id)) {
int len;
if((len = e->opset->read(e->fd, inbuff + bytes_read,
sizeof(jlog_id) - bytes_read,
&mask, e)) <= 0)
goto alldone;
bytes_read += len;
}
memcpy(&client_chkpt, inbuff, sizeof(jlog_id));
/* Fix the endian */
client_chkpt.log = ntohl(client_chkpt.log);
client_chkpt.marker = ntohl(client_chkpt.marker);
if(memcmp(&jcl->chkpt, &client_chkpt, sizeof(jlog_id))) {
noitL(noit_error,
"client %s submitted invalid checkpoint %u:%u expected %u:%u\n",
ac->remote_cn, client_chkpt.log, client_chkpt.marker,
jcl->chkpt.log, jcl->chkpt.marker);
goto alldone;
}
gettimeofday(&jcl->feed_stats->last_checkpoint, NULL);
jlog_ctx_read_checkpoint(jcl->jlog, &jcl->chkpt);
}
else {
/* we have nothing to write -- maybe we have no checks configured...
* If this is the case "forever", the remote might disconnect and
* we would never know. Do the painful work of detecting a
* disconnected client.
*/
struct pollfd pfd;
pfd.fd = e->fd;
pfd.events = POLLIN | POLLHUP | POLLRDNORM;
pfd.revents = 0;
if(poll(&pfd, 1, 0) != 0) {
/* normally, we'd recv PEEK|DONTWAIT. However, the client should
* not be writing to us. So, we know we can't have any legitimate
* data on this socket (true even though this is SSL). So, if we're
* here then "shit went wrong"
*/
noitL(noit_error, "jlog client %s disconnected while idle\n",
ac->remote_cn);
goto alldone;
}
}
if(sleeptime) sleep(sleeptime);
}
alldone:
e->opset->close(e->fd, &mask, e);
noit_atomic_dec32(&jcl->feed_stats->connections);
noit_jlog_closure_free(jcl);
acceptor_closure_free(ac);
return NULL;
}
int main(int argc, char **argv) {
int i;
char *toremove = NULL;
jlog_safety safety = JLOG_ALMOST_SAFE;
pthread_t tid[THRCNT], wtid[WTHRCNT];
void *foo;
#if _WIN32
mem_init();
#endif
if(argc == 3) {
if(!strcmp(argv[1], "safety")) {
if(!strcmp(argv[2], "unsafe"))
safety = JLOG_UNSAFE;
else if(!strcmp(argv[2], "almost_safe"))
safety = JLOG_ALMOST_SAFE;
else if(!strcmp(argv[2], "safe"))
safety = JLOG_SAFE;
else {
fprintf(stderr, "invalid safety option\n");
usage();
}
} else if(!strcmp(argv[1], "only")) {
if(!strcmp(argv[2], "read")) only_read = 1;
else if(!strcmp(argv[2], "write")) only_write = 1;
else usage();
} else if(!strcmp(argv[1], "remove")) {
toremove = argv[2];
} else {
usage();
}
} else if(argc < 3 || argc > 3) {
usage();
}
jcreate(safety);
if(toremove) {
jlog_ctx *ctx;
ctx = jlog_new(LOGNAME);
if(jlog_ctx_open_writer(ctx) != 0) {
fprintf(stderr, "jlog_ctx_open_writer failed: %d %s\n", jlog_ctx_err(ctx), jlog_ctx_err_string(ctx));
croak();
}
jlog_ctx_remove_subscriber(ctx, argv[2]);
jlog_ctx_close(ctx);
exit(0);
}
if(!only_write) {
for(i=0; i<THRCNT; i++) {
pthread_create(&tid[i], NULL, reader, (void *)(uintptr_t)i);
fprintf(stderr, "[%d] started reader\n", (int)(uintptr_t)tid[i]);
}
}
if(!only_read) {
fprintf(stderr, "starting writers..\n");
for(i=0; i<WTHRCNT; i++) {
pthread_create(&wtid[i], NULL, writer, (void *)(uintptr_t)i);
fprintf(stderr, "[%d] started writer\n", (int)(uintptr_t)wtid[i]);
}
} else {
sleep(5);
writer_done = 1;
}
if(!only_write) {
for(i=0; i<THRCNT; i++) {
pthread_join(tid[i], &foo);
fprintf(stderr, "[%d] joined, read %d\n", i, (int)(uintptr_t)foo);
}
}
if(!only_read) {
for(i=0; i<WTHRCNT; i++) {
pthread_join(wtid[i], &foo);
fprintf(stderr, "[%d] joined, write %d\n", i, (int)(uintptr_t)foo);
}
}
if(error) fprintf(stderr, "errors occurred\n");
return 0;
}
void *reader(void *unused) {
jlog_ctx *ctx;
char subname[32];
int tcount = 0, fcount = 0;
int prev_err = 0;
int subno = (int)(uintptr_t)unused;
snprintf(subname, sizeof(subname), "sub-%02d", subno);
reader_retry:
ctx = jlog_new(LOGNAME);
if(jlog_ctx_open_reader(ctx, subname) != 0) {
if(prev_err == 0) {
prev_err = jlog_ctx_err(ctx);
jlog_ctx_close(ctx);
ctx = jlog_new(LOGNAME);
if(prev_err == JLOG_ERR_INVALID_SUBSCRIBER) {
fprintf(stderr, "[%02d] invalid subscriber, init...\n", subno);
if(jlog_ctx_open_writer(ctx) != 0) {
fprintf(stderr, "[%02d] jlog_ctx_open_writer failed: %d %s\n", subno, jlog_ctx_err(ctx), jlog_ctx_err_string(ctx));
} else {
if(jlog_ctx_add_subscriber(ctx, subname, JLOG_BEGIN) != 0) {
fprintf(stderr, "[%02d] jlog_ctx_add_subscriber failed: %d %s\n", subno, jlog_ctx_err(ctx), jlog_ctx_err_string(ctx));
} else {
jlog_ctx_close(ctx);
goto reader_retry;
}
}
}
}
fprintf(stderr, "[%02d] jlog_ctx_open_reader failed: %d %s\n", subno, jlog_ctx_err(ctx), jlog_ctx_err_string(ctx));
croak();
}
fprintf(stderr, "[%02d] reader started\n", subno);
while(1) {
char begins[20], ends[20];
jlog_id begin, end;
int count;
jlog_message message;
if((count = jlog_ctx_read_interval(ctx, &begin, &end)) == -1) {
fprintf(stderr, "jlog_ctx_read_interval failed: %d %s\n", jlog_ctx_err(ctx), jlog_ctx_err_string(ctx));
croak();
}
jlog_snprint_logid(begins, sizeof(begins), &begin);
jlog_snprint_logid(ends, sizeof(ends), &end);
if(count > 0) {
int i;
//fprintf(stderr, "[%02d] reader (%s, %s] count: %d\n", subno, begins, ends, count);
for(i=0; i<count; i++, JLOG_ID_ADVANCE(&begin)) {
end = begin;
if(jlog_ctx_read_message(ctx, &begin, &message) != 0) {
if(jlog_ctx_err(ctx) == JLOG_ERR_CLOSE_LOGID) {
/* fine */
} else {
fcount++;
jlog_snprint_logid(begins, sizeof(begins), &begin);
fprintf(stderr, "[%02d] read failed @ %s: %d %s\n", subno, begins, jlog_ctx_err(ctx), jlog_ctx_err_string(ctx));
}
} else {
tcount++;
jlog_snprint_logid(begins, sizeof(begins), &begin);
/* fprintf(stderr, "[%02d] read: [%s]\n\t'%.*s'\n", subno, begins,
message.mess_len, (char *)message.mess); */
}
}
if(jlog_ctx_read_checkpoint(ctx, &end) != 0) {
fprintf(stderr, "[%02d] checkpoint failed: %d %s\n", subno, jlog_ctx_err(ctx), jlog_ctx_err_string(ctx));
} else {
fprintf(stderr, "[%02d] \tcheckpointed...\n", subno);
}
} else {
if(writer_done == 1) break;
}
}
fprintf(stderr, "[%02d] reader read %d, failed %d\n", subno, tcount, fcount);
jlog_ctx_close(ctx);
return (void *)(uintptr_t)tcount;
}
