Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma

Pull RDMA fixes from Jason Gunthorpe:
 "Bug fixes for old bugs in the hns and hfi1 drivers:

   - Calculate various values in hns properly to avoid over/underflows
     in some cases

   - Fix an oops, PCI negotiation on Gen4 systems, and bugs related to
     retries"

* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma:
  RDMA/hns: Correct the value of srq_desc_size
  RDMA/hns: Correct the value of HNS_ROCE_HEM_CHUNK_LEN
  IB/hfi1: TID RDMA WRITE should not return IB_WC_RNR_RETRY_EXC_ERR
  IB/hfi1: Calculate flow weight based on QP MTU for TID RDMA
  IB/hfi1: Ensure r_tid_ack is valid before building TID RDMA ACK packet
  IB/hfi1: Ensure full Gen3 speed in a Gen4 system

drivers/infiniband/hw/hfi1/init.c

@@ -1489,7 +1489,6 @@ static int __init hfi1_mod_init(void)
 		goto bail_dev;
 	}
 
-	hfi1_compute_tid_rdma_flow_wt();
 	/*
 	 * These must be called before the driver is registered with
 	 * the PCI subsystem.

drivers/infiniband/hw/hfi1/pcie.c

@@ -319,7 +319,9 @@ int pcie_speeds(struct hfi1_devdata *dd)
 	/*
 	 * bus->max_bus_speed is set from the bridge's linkcap Max Link Speed
 	 */
-	if (parent && dd->pcidev->bus->max_bus_speed != PCIE_SPEED_8_0GT) {
+	if (parent &&
+	    (dd->pcidev->bus->max_bus_speed == PCIE_SPEED_2_5GT ||
+	     dd->pcidev->bus->max_bus_speed == PCIE_SPEED_5_0GT)) {
 		dd_dev_info(dd, "Parent PCIe bridge does not support Gen3\n");
 		dd->link_gen3_capable = 0;
 	}

drivers/infiniband/hw/hfi1/rc.c

@@ -2209,15 +2209,15 @@ int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode,
 		if (qp->s_flags & RVT_S_WAIT_RNR)
 			goto bail_stop;
 		rdi = ib_to_rvt(qp->ibqp.device);
-		if (qp->s_rnr_retry == 0 &&
-		    !((rdi->post_parms[wqe->wr.opcode].flags &
-		      RVT_OPERATION_IGN_RNR_CNT) &&
-		      qp->s_rnr_retry_cnt == 0)) {
+		if (!(rdi->post_parms[wqe->wr.opcode].flags &
+		       RVT_OPERATION_IGN_RNR_CNT)) {
+			if (qp->s_rnr_retry == 0) {
 				status = IB_WC_RNR_RETRY_EXC_ERR;
 				goto class_b;
 			}
 			if (qp->s_rnr_retry_cnt < 7 && qp->s_rnr_retry_cnt > 0)
 				qp->s_rnr_retry--;
+		}
 
 		/*
 		 * The last valid PSN is the previous PSN. For TID RDMA WRITE

drivers/infiniband/hw/hfi1/tid_rdma.c

@@ -107,8 +107,6 @@ static u32 mask_generation(u32 a)
  * C - Capcode
  */
 
-static u32 tid_rdma_flow_wt;
-
 static void tid_rdma_trigger_resume(struct work_struct *work);
 static void hfi1_kern_exp_rcv_free_flows(struct tid_rdma_request *req);
 static int hfi1_kern_exp_rcv_alloc_flows(struct tid_rdma_request *req,
@@ -136,6 +134,26 @@ static void update_r_next_psn_fecn(struct hfi1_packet *packet,
 				   struct tid_rdma_flow *flow,
 				   bool fecn);
 
+static void validate_r_tid_ack(struct hfi1_qp_priv *priv)
+{
+	if (priv->r_tid_ack == HFI1_QP_WQE_INVALID)
+		priv->r_tid_ack = priv->r_tid_tail;
+}
+
+static void tid_rdma_schedule_ack(struct rvt_qp *qp)
+{
+	struct hfi1_qp_priv *priv = qp->priv;
+
+	priv->s_flags |= RVT_S_ACK_PENDING;
+	hfi1_schedule_tid_send(qp);
+}
+
+static void tid_rdma_trigger_ack(struct rvt_qp *qp)
+{
+	validate_r_tid_ack(qp->priv);
+	tid_rdma_schedule_ack(qp);
+}
+
 static u64 tid_rdma_opfn_encode(struct tid_rdma_params *p)
 {
 	return
@@ -3005,10 +3023,7 @@ bool hfi1_handle_kdeth_eflags(struct hfi1_ctxtdata *rcd,
 		qpriv->s_nak_state = IB_NAK_PSN_ERROR;
 		/* We are NAK'ing the next expected PSN */
 		qpriv->s_nak_psn = mask_psn(flow->flow_state.r_next_psn);
-		qpriv->s_flags |= RVT_S_ACK_PENDING;
-		if (qpriv->r_tid_ack == HFI1_QP_WQE_INVALID)
-			qpriv->r_tid_ack = qpriv->r_tid_tail;
-		hfi1_schedule_tid_send(qp);
+		tid_rdma_trigger_ack(qp);
 	}
 	goto unlock;
 }
@@ -3371,18 +3386,17 @@ u32 hfi1_build_tid_rdma_write_req(struct rvt_qp *qp, struct rvt_swqe *wqe,
 	return sizeof(ohdr->u.tid_rdma.w_req) / sizeof(u32);
 }
 
-void hfi1_compute_tid_rdma_flow_wt(void)
+static u32 hfi1_compute_tid_rdma_flow_wt(struct rvt_qp *qp)
 {
 	/*
 	 * Heuristic for computing the RNR timeout when waiting on the flow
 	 * queue. Rather than a computationaly expensive exact estimate of when
 	 * a flow will be available, we assume that if a QP is at position N in
 	 * the flow queue it has to wait approximately (N + 1) * (number of
-	 * segments between two sync points), assuming PMTU of 4K. The rationale
-	 * for this is that flows are released and recycled at each sync point.
+	 * segments between two sync points). The rationale for this is that
+	 * flows are released and recycled at each sync point.
 	 */
-	tid_rdma_flow_wt = MAX_TID_FLOW_PSN * enum_to_mtu(OPA_MTU_4096) /
-		TID_RDMA_MAX_SEGMENT_SIZE;
+	return (MAX_TID_FLOW_PSN * qp->pmtu) >> TID_RDMA_SEGMENT_SHIFT;
 }
 
 static u32 position_in_queue(struct hfi1_qp_priv *qpriv,
@@ -3505,7 +3519,7 @@ static void hfi1_tid_write_alloc_resources(struct rvt_qp *qp, bool intr_ctx)
 		if (qpriv->flow_state.index >= RXE_NUM_TID_FLOWS) {
 			ret = hfi1_kern_setup_hw_flow(qpriv->rcd, qp);
 			if (ret) {
-				to_seg = tid_rdma_flow_wt *
+				to_seg = hfi1_compute_tid_rdma_flow_wt(qp) *
 					position_in_queue(qpriv,
 							  &rcd->flow_queue);
 				break;
@@ -3526,7 +3540,7 @@ static void hfi1_tid_write_alloc_resources(struct rvt_qp *qp, bool intr_ctx)
 		/*
 		 * If overtaking req->acked_tail, send an RNR NAK. Because the
 		 * QP is not queued in this case, and the issue can only be
-		 * caused due a delay in scheduling the second leg which we
+		 * caused by a delay in scheduling the second leg which we
 		 * cannot estimate, we use a rather arbitrary RNR timeout of
 		 * (MAX_FLOWS / 2) segments
 		 */
@@ -3534,8 +3548,7 @@ static void hfi1_tid_write_alloc_resources(struct rvt_qp *qp, bool intr_ctx)
 				MAX_FLOWS)) {
 			ret = -EAGAIN;
 			to_seg = MAX_FLOWS >> 1;
-			qpriv->s_flags |= RVT_S_ACK_PENDING;
-			hfi1_schedule_tid_send(qp);
+			tid_rdma_trigger_ack(qp);
 			break;
 		}
 
@@ -4335,8 +4348,7 @@ void hfi1_rc_rcv_tid_rdma_write_data(struct hfi1_packet *packet)
 	trace_hfi1_tid_req_rcv_write_data(qp, 0, e->opcode, e->psn, e->lpsn,
 					  req);
 	trace_hfi1_tid_write_rsp_rcv_data(qp);
-	if (priv->r_tid_ack == HFI1_QP_WQE_INVALID)
-		priv->r_tid_ack = priv->r_tid_tail;
+	validate_r_tid_ack(priv);
 
 	if (opcode == TID_OP(WRITE_DATA_LAST)) {
 		release_rdma_sge_mr(e);
@@ -4375,8 +4387,7 @@ void hfi1_rc_rcv_tid_rdma_write_data(struct hfi1_packet *packet)
 	}
 
 done:
-	priv->s_flags |= RVT_S_ACK_PENDING;
-	hfi1_schedule_tid_send(qp);
+	tid_rdma_schedule_ack(qp);
 exit:
 	priv->r_next_psn_kdeth = flow->flow_state.r_next_psn;
 	if (fecn)
@@ -4388,10 +4399,7 @@ void hfi1_rc_rcv_tid_rdma_write_data(struct hfi1_packet *packet)
 	if (!priv->s_nak_state) {
 		priv->s_nak_state = IB_NAK_PSN_ERROR;
 		priv->s_nak_psn = flow->flow_state.r_next_psn;
-		priv->s_flags |= RVT_S_ACK_PENDING;
-		if (priv->r_tid_ack == HFI1_QP_WQE_INVALID)
-			priv->r_tid_ack = priv->r_tid_tail;
-		hfi1_schedule_tid_send(qp);
+		tid_rdma_trigger_ack(qp);
 	}
 	goto done;
 }
@@ -4939,8 +4947,7 @@ void hfi1_rc_rcv_tid_rdma_resync(struct hfi1_packet *packet)
 	qpriv->resync = true;
 	/* RESYNC request always gets a TID RDMA ACK. */
 	qpriv->s_nak_state = 0;
-	qpriv->s_flags |= RVT_S_ACK_PENDING;
-	hfi1_schedule_tid_send(qp);
+	tid_rdma_trigger_ack(qp);
 bail:
 	if (fecn)
 		qp->s_flags |= RVT_S_ECN;

drivers/infiniband/hw/hfi1/tid_rdma.h

@@ -17,6 +17,7 @@
 #define TID_RDMA_MIN_SEGMENT_SIZE       BIT(18)   /* 256 KiB (for now) */
 #define TID_RDMA_MAX_SEGMENT_SIZE       BIT(18)   /* 256 KiB (for now) */
 #define TID_RDMA_MAX_PAGES              (BIT(18) >> PAGE_SHIFT)
+#define TID_RDMA_SEGMENT_SHIFT		18
 
 /*
  * Bit definitions for priv->s_flags.
@@ -274,8 +275,6 @@ u32 hfi1_build_tid_rdma_write_req(struct rvt_qp *qp, struct rvt_swqe *wqe,
 				  struct ib_other_headers *ohdr,
 				  u32 *bth1, u32 *bth2, u32 *len);
 
-void hfi1_compute_tid_rdma_flow_wt(void);
-
 void hfi1_rc_rcv_tid_rdma_write_req(struct hfi1_packet *packet);
 
 u32 hfi1_build_tid_rdma_write_resp(struct rvt_qp *qp, struct rvt_ack_entry *e,

drivers/infiniband/hw/hns/hns_roce_hem.h

@@ -59,7 +59,7 @@ enum {
 
 #define HNS_ROCE_HEM_CHUNK_LEN	\
 	 ((256 - sizeof(struct list_head) - 2 * sizeof(int)) /	 \
-	 (sizeof(struct scatterlist)))
+	 (sizeof(struct scatterlist) + sizeof(void *)))
 
 #define check_whether_bt_num_3(type, hop_num) \
 	(type < HEM_TYPE_MTT && hop_num == 2)

drivers/infiniband/hw/hns/hns_roce_srq.c

@@ -376,7 +376,7 @@ int hns_roce_create_srq(struct ib_srq *ib_srq,
 	srq->max = roundup_pow_of_two(srq_init_attr->attr.max_wr + 1);
 	srq->max_gs = srq_init_attr->attr.max_sge;
 
-	srq_desc_size = max(16, 16 * srq->max_gs);
+	srq_desc_size = roundup_pow_of_two(max(16, 16 * srq->max_gs));
 
 	srq->wqe_shift = ilog2(srq_desc_size);