static void analyze_datafile(jlog_ctx *ctx, u_int32_t logid) {
char idxfile[MAXPATHLEN];
if (jlog_inspect_datafile(ctx, logid) > 0) {
fprintf(stderr, "One or more errors were found.\n");
if(repair_datafiles) {
jlog_repair_datafile(ctx, logid);
fprintf(stderr,
"Log file reconstructed, deleting the corresponding idx file.\n");
STRSETDATAFILE(ctx, idxfile, logid);
strcat(idxfile, INDEX_EXT);
unlink(idxfile);
}
}
}
static jlog_file *__jlog_open_reader(jlog_ctx *ctx, u_int32_t log) {
char file[MAXPATHLEN];
if(ctx->current_log != log) {
__jlog_close_reader(ctx);
__jlog_close_indexer(ctx);
}
if(ctx->data) {
return ctx->data;
}
STRSETDATAFILE(ctx, file, log);
#ifdef DEBUG
fprintf(stderr, "opening log file[ro]: '%s'\n", file);
#endif
ctx->data = jlog_file_open(file, 0, ctx->file_mode);
ctx->current_log = log;
return ctx->data;
}
static jlog_file *__jlog_open_writer(jlog_ctx *ctx) {
char file[MAXPATHLEN];
if(ctx->data) {
/* Still open */
return ctx->data;
}
if(!jlog_file_lock(ctx->metastore))
SYS_FAIL(JLOG_ERR_LOCK);
if(__jlog_restore_metastore(ctx, 1))
SYS_FAIL(JLOG_ERR_META_OPEN);
STRSETDATAFILE(ctx, file, ctx->storage.log);
#ifdef DEBUG
fprintf(stderr, "opening log file[rw]: '%s'\n", file);
#endif
ctx->data = jlog_file_open(file, O_CREAT, ctx->file_mode);
finish:
jlog_file_unlock(ctx->metastore);
return ctx->data;
}
static jlog_file *__jlog_open_indexer(jlog_ctx *ctx, u_int32_t log) {
char file[MAXPATHLEN];
int len;
if(ctx->current_log != log) {
__jlog_close_reader(ctx);
__jlog_close_indexer(ctx);
}
if(ctx->index) {
return ctx->index;
}
STRSETDATAFILE(ctx, file, log);
len = strlen(file);
if((len + sizeof(INDEX_EXT)) > sizeof(file)) return NULL;
memcpy(file + len, INDEX_EXT, sizeof(INDEX_EXT));
#ifdef DEBUG
fprintf(stderr, "opening index file: '%s'\n", idx);
#endif
ctx->index = jlog_file_open(file, O_CREAT, ctx->file_mode);
ctx->current_log = log;
return ctx->index;
}
static int __jlog_unlink_datafile(jlog_ctx *ctx, u_int32_t log) {
char file[MAXPATHLEN];
int len;
if(ctx->current_log == log) {
__jlog_close_reader(ctx);
__jlog_close_indexer(ctx);
}
STRSETDATAFILE(ctx, file, log);
#ifdef DEBUG
fprintf(stderr, "unlinking %s\n", file);
#endif
unlink(file);
len = strlen(file);
if((len + sizeof(INDEX_EXT)) > sizeof(file)) return -1;
memcpy(file + len, INDEX_EXT, sizeof(INDEX_EXT));
#ifdef DEBUG
fprintf(stderr, "unlinking %s\n", file);
#endif
unlink(file);
return 0;
}
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;
}
static int __jlog_find_first_log_after(jlog_ctx *ctx, jlog_id *chkpt,
jlog_id *start, jlog_id *finish) {
jlog_id last;
int closed;
memcpy(start, chkpt, sizeof(*chkpt));
attempt:
if(__jlog_resync_index(ctx, start->log, &last, &closed) != 0) {
if(ctx->last_error == JLOG_ERR_FILE_OPEN &&
ctx->last_errno == ENOENT) {
char file[MAXPATHLEN];
int ferr, len;
struct stat sb = {0};
STRSETDATAFILE(ctx, file, start->log + 1);
while((ferr = stat(file, &sb)) == -1 && errno == EINTR);
/* That file doesn't exist... bad, but we can fake a recovery by
advancing the next file that does exist */
ctx->last_error = JLOG_ERR_SUCCESS;
if(start->log >= ctx->storage.log || ferr != 0 || sb.st_size == 0) {
/* We don't advance past where people are writing */
memcpy(finish, start, sizeof(*start));
return 0;
}
if(__jlog_resync_index(ctx, start->log + 1, &last, &closed) != 0) {
/* We don't advance past where people are writing */
memcpy(finish, start, sizeof(*start));
return 0;
}
len = strlen(file);
if((len + sizeof(INDEX_EXT)) > sizeof(file)) return -1;
memcpy(file + len, INDEX_EXT, sizeof(INDEX_EXT));
while((ferr = stat(file, &sb)) == -1 && errno == EINTR);
if(ferr != 0 || sb.st_size == 0) {
/* We don't advance past where people are writing */
memcpy(finish, start, sizeof(*start));
return 0;
}
start->marker = 0;
start->log++; /* BE SMARTER! */
goto attempt;
}
return -1; /* Just persist resync's error state */
}
/* If someone checkpoints off the end, be nice */
if(last.log == start->log && last.marker < start->marker)
memcpy(start, &last, sizeof(*start));
if(!memcmp(start, &last, sizeof(last)) && closed) {
char file[MAXPATHLEN];
int ferr, len;
struct stat sb = {0};
STRSETDATAFILE(ctx, file, start->log + 1);
while((ferr = stat(file, &sb)) == -1 && errno == EINTR);
if(start->log >= ctx->storage.log || ferr != 0 || sb.st_size == 0) {
/* We don't advance past where people are writing */
memcpy(finish, start, sizeof(*start));
return 0;
}
if(__jlog_resync_index(ctx, start->log + 1, &last, &closed) != 0) {
/* We don't advance past where people are writing */
memcpy(finish, start, sizeof(*start));
return 0;
}
len = strlen(file);
if((len + sizeof(INDEX_EXT)) > sizeof(file)) return -1;
memcpy(file + len, INDEX_EXT, sizeof(INDEX_EXT));
while((ferr = stat(file, &sb)) == -1 && errno == EINTR);
if(ferr != 0 || sb.st_size == 0) {
/* We don't advance past where people are writing */
memcpy(finish, start, sizeof(*start));
return 0;
}
start->marker = 0;
start->log++;
goto attempt;
}
memcpy(finish, &last, sizeof(last));
return 0;
}
