int sheep_exec_req(const struct node_id *nid, struct sd_req *hdr, void *buf)
{
struct sd_rsp *rsp = (struct sd_rsp *)hdr;
struct sockfd *sfd;
int ret;
assert(is_worker_thread());
sfd = sockfd_cache_get(nid);
if (!sfd)
return SD_RES_NETWORK_ERROR;
ret = exec_req(sfd->fd, hdr, buf, sheep_need_retry, hdr->epoch,
MAX_RETRY_COUNT);
if (ret) {
sd_dprintf("remote node might have gone away");
sockfd_cache_del(nid, sfd);
return SD_RES_NETWORK_ERROR;
}
ret = rsp->result;
if (ret != SD_RES_SUCCESS)
sd_eprintf("failed %s", sd_strerror(ret));
sockfd_cache_put(nid, sfd);
return ret;
}
/*
* Wait for all forward requests completion.
*
* Even if something goes wrong, we have to wait forward requests completion to
* avoid interleaved requests.
*
* Return error code if any one request fails.
*/
static int wait_forward_request(struct write_info *wi, struct request *req)
{
int nr_sent, err_ret = SD_RES_SUCCESS, ret, pollret, i,
repeat = MAX_RETRY_COUNT;
struct pfd_info pi;
struct sd_rsp *rsp = &req->rp;
again:
pfd_info_init(wi, &pi);
pollret = poll(pi.pfds, pi.nr, 1000 * POLL_TIMEOUT);
if (pollret < 0) {
if (errno == EINTR)
goto again;
panic("%m");
} else if (pollret == 0) {
/*
* If IO NIC is down, epoch isn't incremented, so we can't retry
* for ever.
*/
if (sheep_need_retry(req->rq.epoch) && repeat) {
repeat--;
sd_warn("poll timeout %d, disks of some nodes or "
"network is busy. Going to poll-wait again",
wi->nr_sent);
goto again;
}
nr_sent = wi->nr_sent;
/* XXX Blinedly close all the connections */
for (i = 0; i < nr_sent; i++)
sockfd_cache_del(wi->ent[i].nid, wi->ent[i].sfd);
return SD_RES_NETWORK_ERROR;
}
nr_sent = wi->nr_sent;
for (i = 0; i < nr_sent; i++)
if (pi.pfds[i].revents & POLLIN)
break;
if (i < nr_sent) {
int re = pi.pfds[i].revents;
sd_debug("%d, revents %x", i, re);
if (re & (POLLERR | POLLHUP | POLLNVAL)) {
err_ret = SD_RES_NETWORK_ERROR;
finish_one_write_err(wi, i);
goto finish_write;
}
if (do_read(pi.pfds[i].fd, rsp, sizeof(*rsp), sheep_need_retry,
req->rq.epoch, MAX_RETRY_COUNT)) {
sd_err("remote node might have gone away");
err_ret = SD_RES_NETWORK_ERROR;
finish_one_write_err(wi, i);
goto finish_write;
}
ret = rsp->result;
if (ret != SD_RES_SUCCESS) {
sd_err("fail %"PRIx64", %s", req->rq.obj.oid,
sd_strerror(ret));
err_ret = ret;
}
finish_one_write(wi, i);
}
finish_write:
if (wi->nr_sent > 0)
goto again;
return err_ret;
}
static inline void finish_one_write_err(struct write_info *wi, int i)
{
sockfd_cache_del(wi->ent[i].nid, wi->ent[i].sfd);
write_info_update(wi, i);
}
