static void rtrs_clt_fast_reg_done(struct ib_cq *cq, struct ib_wc *wc)
{
- struct rtrs_clt_con *con = cq->cq_context;
+ struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
if (unlikely(wc->status != IB_WC_SUCCESS)) {
rtrs_err(con->c.sess, "Failed IB_WR_REG_MR: %s\n",
{
struct rtrs_clt_io_req *req =
container_of(wc->wr_cqe, typeof(*req), inv_cqe);
- struct rtrs_clt_con *con = cq->cq_context;
+ struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
if (unlikely(wc->status != IB_WC_SUCCESS)) {
rtrs_err(con->c.sess, "Failed IB_WR_LOCAL_INV: %s\n",
req->in_use = false;
req->con = NULL;
+ if (errno) {
+ rtrs_err_rl(con->c.sess,
+ "IO request failed: error=%d path=%s [%s:%u]\n",
+ errno, kobject_name(&sess->kobj), sess->hca_name,
+ sess->hca_port);
+ }
+
if (notify)
req->conf(req->priv, errno);
}
static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc)
{
- struct rtrs_clt_con *con = cq->cq_context;
+ struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
u32 imm_type, imm_payload;
bool w_inval = false;
} else if (imm_type == RTRS_HB_ACK_IMM) {
WARN_ON(con->c.cid);
sess->s.hb_missed_cnt = 0;
+ sess->s.hb_cur_latency =
+ ktime_sub(ktime_get(), sess->s.hb_last_sent);
if (sess->flags & RTRS_MSG_NEW_RKEY_F)
return rtrs_clt_recv_done(con, wc);
} else {
return min_path;
}
+/**
+ * get_next_path_min_latency() - Returns path with minimal latency.
+ * @it: the path pointer
+ *
+ * Return: a path with the lowest latency or NULL if all paths are tried
+ *
+ * Locks:
+ * rcu_read_lock() must be hold.
+ *
+ * Related to @MP_POLICY_MIN_LATENCY
+ *
+ * This DOES skip an already-tried path.
+ * There is a skip-list to skip a path if the path has tried but failed.
+ * It will try the minimum latency path and then the second minimum latency
+ * path and so on. Finally it will return NULL if all paths are tried.
+ * Therefore the caller MUST check the returned
+ * path is NULL and trigger the IO error.
+ */
+static struct rtrs_clt_sess *get_next_path_min_latency(struct path_it *it)
+{
+ struct rtrs_clt_sess *min_path = NULL;
+ struct rtrs_clt *clt = it->clt;
+ struct rtrs_clt_sess *sess;
+ ktime_t min_latency = INT_MAX;
+ ktime_t latency;
+
+ list_for_each_entry_rcu(sess, &clt->paths_list, s.entry) {
+ if (unlikely(READ_ONCE(sess->state) != RTRS_CLT_CONNECTED))
+ continue;
+
+ if (unlikely(!list_empty(raw_cpu_ptr(sess->mp_skip_entry))))
+ continue;
+
+ latency = sess->s.hb_cur_latency;
+
+ if (latency < min_latency) {
+ min_latency = latency;
+ min_path = sess;
+ }
+ }
+
+ /*
+ * add the path to the skip list, so that next time we can get
+ * a different one
+ */
+ if (min_path)
+ list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list);
+
+ return min_path;
+}
+
static inline void path_it_init(struct path_it *it, struct rtrs_clt *clt)
{
INIT_LIST_HEAD(&it->skip_list);
if (clt->mp_policy == MP_POLICY_RR)
it->next_path = get_next_path_rr;
- else
+ else if (clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
it->next_path = get_next_path_min_inflight;
+ else
+ it->next_path = get_next_path_min_latency;
}
static inline void path_it_deinit(struct path_it *it)
req->usr_len + sizeof(*msg),
imm);
if (unlikely(ret)) {
- rtrs_err(s, "Write request failed: %d\n", ret);
+ rtrs_err_rl(s,
+ "Write request failed: error=%d path=%s [%s:%u]\n",
+ ret, kobject_name(&sess->kobj), sess->hca_name,
+ sess->hca_port);
if (sess->clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
atomic_dec(&sess->stats->inflight);
if (req->sg_cnt)
struct rtrs_sess *s = con->c.sess;
struct rtrs_clt_sess *sess = to_clt_sess(s);
struct rtrs_msg_rdma_read *msg;
- struct rtrs_ib_dev *dev;
+ struct rtrs_ib_dev *dev = sess->s.dev;
struct ib_reg_wr rwr;
struct ib_send_wr *wr = NULL;
const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
- s = &sess->s;
- dev = sess->s.dev;
-
if (unlikely(tsize > sess->chunk_size)) {
rtrs_wrn(s,
"Read request failed, message size is %zu, bigger than CHUNK_SIZE %d\n",
ret = rtrs_post_send_rdma(req->con, req, &sess->rbufs[buf_id],
req->data_len, imm, wr);
if (unlikely(ret)) {
- rtrs_err(s, "Read request failed: %d\n", ret);
+ rtrs_err_rl(s,
+ "Read request failed: error=%d path=%s [%s:%u]\n",
+ ret, kobject_name(&sess->kobj), sess->hca_name,
+ sess->hca_port);
if (sess->clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
atomic_dec(&sess->stats->inflight);
req->need_inv = false;
case RDMA_CM_EVENT_UNREACHABLE:
case RDMA_CM_EVENT_ADDR_CHANGE:
case RDMA_CM_EVENT_TIMEWAIT_EXIT:
- rtrs_wrn(s, "CM error event %d\n", ev->event);
+ rtrs_wrn(s, "CM error (CM event: %s, err: %d)\n",
+ rdma_event_msg(ev->event), ev->status);
cm_err = -ECONNRESET;
break;
case RDMA_CM_EVENT_ADDR_ERROR:
case RDMA_CM_EVENT_ROUTE_ERROR:
- rtrs_wrn(s, "CM error event %d\n", ev->event);
+ rtrs_wrn(s, "CM error (CM event: %s, err: %d)\n",
+ rdma_event_msg(ev->event), ev->status);
cm_err = -EHOSTUNREACH;
break;
case RDMA_CM_EVENT_DEVICE_REMOVAL:
rtrs_clt_close_conns(sess, false);
return 0;
default:
- rtrs_err(s, "Unexpected RDMA CM event (%d)\n", ev->event);
+ rtrs_err(s, "Unexpected RDMA CM error (CM event: %s, err: %d)\n",
+ rdma_event_msg(ev->event), ev->status);
cm_err = -ECONNRESET;
break;
}
static void rtrs_clt_info_req_done(struct ib_cq *cq, struct ib_wc *wc)
{
- struct rtrs_clt_con *con = cq->cq_context;
+ struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
struct rtrs_iu *iu;
static void rtrs_clt_info_rsp_done(struct ib_cq *cq, struct ib_wc *wc)
{
- struct rtrs_clt_con *con = cq->cq_context;
+ struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
struct rtrs_msg_info_rsp *msg;
enum rtrs_clt_state state;
static int init_sess(struct rtrs_clt_sess *sess)
{
int err;
+ char str[NAME_MAX];
+ struct rtrs_addr path = {
+ .src = &sess->s.src_addr,
+ .dst = &sess->s.dst_addr,
+ };
+
+ rtrs_addr_to_str(&path, str, sizeof(str));
mutex_lock(&sess->init_mutex);
err = init_conns(sess);
if (err) {
- rtrs_err(sess->clt, "init_conns(), err: %d\n", err);
+ rtrs_err(sess->clt,
+ "init_conns() failed: err=%d path=%s [%s:%u]\n", err,
+ str, sess->hca_name, sess->hca_port);
goto out;
}
err = rtrs_send_sess_info(sess);
if (err) {
- rtrs_err(sess->clt, "rtrs_send_sess_info(), err: %d\n", err);
+ rtrs_err(
+ sess->clt,
+ "rtrs_send_sess_info() failed: err=%d path=%s [%s:%u]\n",
+ err, str, sess->hca_name, sess->hca_port);
goto out;
}
rtrs_clt_sess_up(sess);
} while (!changed && old_state != RTRS_CLT_DEAD);
if (likely(changed)) {
- rtrs_clt_destroy_sess_files(sess, sysfs_self);
rtrs_clt_remove_path_from_arr(sess);
+ rtrs_clt_destroy_sess_files(sess, sysfs_self);
kobject_put(&sess->kobj);
}
int rtrs_clt_rdma_cq_direct(struct rtrs_clt *clt, unsigned int index)
{
- int cnt;
+ /* If no path, return -1 for block layer not to try again */
+ int cnt = -1;
struct rtrs_con *con;
struct rtrs_clt_sess *sess;
struct path_it it;
return -ECOMM;
attr->queue_depth = clt->queue_depth;
- attr->max_io_size = clt->max_io_size;
- attr->sess_kobj = &clt->dev.kobj;
- strlcpy(attr->sessname, clt->sessname, sizeof(attr->sessname));
+ /* Cap max_io_size to min of remote buffer size and the fr pages */
+ attr->max_io_size = min_t(int, clt->max_io_size,
+ clt->max_segments * SZ_4K);
return 0;
}
projects / J-linux.git / commitdiff