int ret = -ENOMEM;
if (recv->r_iwinc == NULL) {
- if (atomic_read(&rds_iw_allocation) >= rds_iw_sysctl_max_recv_allocation) {
+ if (!atomic_add_unless(&rds_iw_allocation, 1, rds_iw_sysctl_max_recv_allocation)) {
rds_iw_stats_inc(s_iw_rx_alloc_limit);
goto out;
}
recv->r_iwinc = kmem_cache_alloc(rds_iw_incoming_slab,
kptr_gfp);
- if (recv->r_iwinc == NULL)
+ if (recv->r_iwinc == NULL) {
+ atomic_dec(&rds_iw_allocation);
goto out;
- atomic_inc(&rds_iw_allocation);
+ }
INIT_LIST_HEAD(&recv->r_iwinc->ii_frags);
rds_inc_init(&recv->r_iwinc->ii_inc, conn, conn->c_faddr);
}
int ret = 0;
u32 pos;
- while ((prefill || rds_conn_up(conn))
- && rds_iw_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
+ while ((prefill || rds_conn_up(conn)) &&
+ rds_iw_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
if (pos >= ic->i_recv_ring.w_nr) {
printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
pos);
* room for it beyond the ring size. Send completion notices its special
* wr_id and avoids working with the ring in that case.
*/
+#ifndef KERNEL_HAS_ATOMIC64
+static void rds_iw_set_ack(struct rds_iw_connection *ic, u64 seq,
+ int ack_required)
+{
+ unsigned long flags;
+
+ spin_lock_irqsave(&ic->i_ack_lock, flags);
+ ic->i_ack_next = seq;
+ if (ack_required)
+ set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
+ spin_unlock_irqrestore(&ic->i_ack_lock, flags);
+}
+
+static u64 rds_iw_get_ack(struct rds_iw_connection *ic)
+{
+ unsigned long flags;
+ u64 seq;
+
+ clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
+
+ spin_lock_irqsave(&ic->i_ack_lock, flags);
+ seq = ic->i_ack_next;
+ spin_unlock_irqrestore(&ic->i_ack_lock, flags);
+
+ return seq;
+}
+#else
static void rds_iw_set_ack(struct rds_iw_connection *ic, u64 seq,
int ack_required)
{
- rds_iw_set_64bit(&ic->i_ack_next, seq);
+ atomic64_set(&ic->i_ack_next, seq);
if (ack_required) {
smp_mb__before_clear_bit();
set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
smp_mb__after_clear_bit();
- return ic->i_ack_next;
+ return atomic64_read(&ic->i_ack_next);
}
+#endif
+
static void rds_iw_send_ack(struct rds_iw_connection *ic, unsigned int adv_credits)
{
* - i_ack_next, which is the last sequence number we received
*
* Potentially, send queue and receive queue handlers can run concurrently.
+ * It would be nice to not have to use a spinlock to synchronize things,
+ * but the one problem that rules this out is that 64bit updates are
+ * not atomic on all platforms. Things would be a lot simpler if
+ * we had atomic64 or maybe cmpxchg64 everywhere.
*
* Reconnecting complicates this picture just slightly. When we
* reconnect, we may be seeing duplicate packets. The peer
}
/* Can we get a send credit? */
- if (!rds_iw_send_grab_credits(ic, 1, &adv_credits, 0)) {
+ if (!rds_iw_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) {
rds_iw_stats_inc(s_iw_tx_throttle);
clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
return;
hdr = &iwinc->ii_inc.i_hdr;
/* We can't just use memcmp here; fragments of a
* single message may carry different ACKs */
- if (hdr->h_sequence != ihdr->h_sequence
- || hdr->h_len != ihdr->h_len
- || hdr->h_sport != ihdr->h_sport
- || hdr->h_dport != ihdr->h_dport) {
+ if (hdr->h_sequence != ihdr->h_sequence ||
+ hdr->h_len != ihdr->h_len ||
+ hdr->h_sport != ihdr->h_sport ||
+ hdr->h_dport != ihdr->h_dport) {
rds_iw_conn_error(conn,
"fragment header mismatch; forcing reconnect\n");
return;
{
struct rds_connection *conn = context;
struct rds_iw_connection *ic = conn->c_transport_data;
- struct ib_wc wc;
- struct rds_iw_ack_state state = { 0, };
- struct rds_iw_recv_work *recv;
rdsdebug("conn %p cq %p\n", conn, cq);
rds_iw_stats_inc(s_iw_rx_cq_call);
- ib_req_notify_cq(cq, IB_CQ_SOLICITED);
+ tasklet_schedule(&ic->i_recv_tasklet);
+}
- while (ib_poll_cq(cq, 1, &wc) > 0) {
+static inline void rds_poll_cq(struct rds_iw_connection *ic,
+ struct rds_iw_ack_state *state)
+{
+ struct rds_connection *conn = ic->conn;
+ struct ib_wc wc;
+ struct rds_iw_recv_work *recv;
+
+ while (ib_poll_cq(ic->i_recv_cq, 1, &wc) > 0) {
rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
(unsigned long long)wc.wr_id, wc.status, wc.byte_len,
be32_to_cpu(wc.ex.imm_data));
if (rds_conn_up(conn) || rds_conn_connecting(conn)) {
/* We expect errors as the qp is drained during shutdown */
if (wc.status == IB_WC_SUCCESS) {
- rds_iw_process_recv(conn, recv, wc.byte_len, &state);
+ rds_iw_process_recv(conn, recv, wc.byte_len, state);
} else {
rds_iw_conn_error(conn, "recv completion on "
"%pI4 had status %u, disconnecting and "
rds_iw_ring_free(&ic->i_recv_ring, 1);
}
+}
+
+void rds_iw_recv_tasklet_fn(unsigned long data)
+{
+ struct rds_iw_connection *ic = (struct rds_iw_connection *) data;
+ struct rds_connection *conn = ic->conn;
+ struct rds_iw_ack_state state = { 0, };
+
+ rds_poll_cq(ic, &state);
+ ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
+ rds_poll_cq(ic, &state);
if (state.ack_next_valid)
rds_iw_set_ack(ic, state.ack_next, state.ack_required);