RDMA/rxe: Move atomic responder res to atomic_reply
authorBob Pearson <rpearsonhpe@gmail.com>
Mon, 6 Jun 2022 14:38:35 +0000 (09:38 -0500)
committerJason Gunthorpe <jgg@nvidia.com>
Thu, 30 Jun 2022 17:00:21 +0000 (14:00 -0300)
Move the allocation of the atomic responder resource up into
rxe_atomic_reply() from send_atomic_ack(). In preparation for merging the
normal and retry atomic responder flows.

Link: https://lore.kernel.org/r/20220606143836.3323-4-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
drivers/infiniband/sw/rxe/rxe_resp.c

index 66d4d02..7489256 100644 (file)
@@ -554,12 +554,36 @@ out:
 /* Guarantee atomicity of atomic operations at the machine level. */
 static DEFINE_SPINLOCK(atomic_ops_lock);
 
+static struct resp_res *rxe_prepare_atomic_res(struct rxe_qp *qp,
+                                              struct rxe_pkt_info *pkt)
+{
+       struct resp_res *res;
+
+       res = &qp->resp.resources[qp->resp.res_head];
+       rxe_advance_resp_resource(qp);
+       free_rd_atomic_resource(qp, res);
+
+       res->type = RXE_ATOMIC_MASK;
+       res->first_psn = pkt->psn;
+       res->last_psn = pkt->psn;
+       res->cur_psn = pkt->psn;
+       res->replay = 0;
+
+       return res;
+}
+
 static enum resp_states rxe_atomic_reply(struct rxe_qp *qp,
                                         struct rxe_pkt_info *pkt)
 {
        u64 *vaddr;
        enum resp_states ret;
        struct rxe_mr *mr = qp->resp.mr;
+       struct resp_res *res = qp->resp.res;
+
+       if (!res) {
+               res = rxe_prepare_atomic_res(qp, pkt);
+               qp->resp.res = res;
+       }
 
        if (mr->state != RXE_MR_STATE_VALID) {
                ret = RESPST_ERR_RKEY_VIOLATION;
@@ -1026,32 +1050,15 @@ err1:
        return err;
 }
 
-static struct resp_res *rxe_prepare_atomic_res(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
-{
-       struct resp_res *res;
-
-       res = &qp->resp.resources[qp->resp.res_head];
-       rxe_advance_resp_resource(qp);
-       free_rd_atomic_resource(qp, res);
-
-       res->type = RXE_ATOMIC_MASK;
-       res->first_psn = pkt->psn;
-       res->last_psn = pkt->psn;
-       res->cur_psn = pkt->psn;
-
-       return res;
-}
-
-static int send_atomic_ack(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
-                          u8 syndrome)
+static int send_atomic_ack(struct rxe_qp *qp, u8 syndrome, u32 psn)
 {
        int err = 0;
        struct rxe_pkt_info ack_pkt;
        struct sk_buff *skb;
-       struct resp_res *res;
+       struct resp_res *res = qp->resp.res;
 
        skb = prepare_ack_packet(qp, &ack_pkt, IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE,
-                                0, pkt->psn, syndrome);
+                                0, psn, syndrome);
        if (!skb) {
                err = -ENOMEM;
                goto out;
@@ -1059,7 +1066,6 @@ static int send_atomic_ack(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
 
        skb_get(skb);
 
-       res = rxe_prepare_atomic_res(qp, pkt);
        res->atomic.skb = skb;
 
        err = rxe_xmit_packet(qp, &ack_pkt, skb);
@@ -1067,6 +1073,11 @@ static int send_atomic_ack(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
                pr_err_ratelimited("Failed sending ack\n");
                rxe_put(qp);
        }
+
+       /* have to clear this since it is used to trigger
+        * long read replies
+        */
+       qp->resp.res = NULL;
 out:
        return err;
 }
@@ -1080,7 +1091,7 @@ static enum resp_states acknowledge(struct rxe_qp *qp,
        if (qp->resp.aeth_syndrome != AETH_ACK_UNLIMITED)
                send_ack(qp, qp->resp.aeth_syndrome, pkt->psn);
        else if (pkt->mask & RXE_ATOMIC_MASK)
-               send_atomic_ack(qp, pkt, AETH_ACK_UNLIMITED);
+               send_atomic_ack(qp, AETH_ACK_UNLIMITED, pkt->psn);
        else if (bth_ack(pkt))
                send_ack(qp, AETH_ACK_UNLIMITED, pkt->psn);