Merge branch 'tid-read' into hfi1-tid

This is the series for adding TID RDMA read. Kaike put in a lot of
effort into making this more consumable for review so special thanks to
him.

Allocating resources and tracing are separated out followed by patches
which build up the read request. Then we have the patches to receive
incoming TID RDMA read requests and handle integration with the RC
protocol.

See the cover letter of the original posting for more of a detailed
overview of TID.

https://www.spinics.net/lists/linux-rdma/msg66611.html

* tid-read:
  IB/hfi1: Add static trace for TID RDMA READ protocol
  IB/hfi1: Enable TID RDMA READ protocol
  IB/hfi1: Add interlock between a TID RDMA request and other requests
  IB/hfi1: Integrate TID RDMA READ protocol into RC protocol
  IB/hfi1: Increment the retry timeout value for TID RDMA READ request
  IB/hfi1: Add functions for restarting TID RDMA READ request
  IB/hfi1: Add TID RDMA handlers
  IB/hfi1: Add functions to receive TID RDMA READ response
  IB/hfi1: Add a function to build TID RDMA READ response
  IB/hfi1: Add functions to receive TID RDMA READ request
  IB/hfi1: Set PbcInsertHcrc for TID RDMA packets
  IB/hfi1: Add functions to build TID RDMA READ request
  IB/hfi1: Add static trace for flow and TID management functions
  IB/hfi1: Add the counter n_tidwait
  IB/hfi1: TID RDMA RcvArray programming and TID allocation
  IB/hfi1: TID RDMA flow allocation
  IB/hfi: Move RC functions into a header file

Signed-off-by: Doug Ledford <dledford@redhat.com>

27 files changed, 4669 insertions, 173 deletions

diff --git a/drivers/infiniband/hw/hfi1/chip.c b/drivers/infiniband/hw/hfi1/chip.c
index 4d40311f082e..612f04190ed8 100644
--- a/drivers/infiniband/hw/hfi1/chip.c
+++ b/drivers/infiniband/hw/hfi1/chip.c
@@ -4253,6 +4253,8 @@ static struct cntr_entry dev_cntrs[DEV_CNTR_LAST] = {
 			    access_sw_pio_drain),
 [C_SW_KMEM_WAIT] = CNTR_ELEM("KmemWait", 0, 0, CNTR_NORMAL,
 			    access_sw_kmem_wait),
+[C_SW_TID_WAIT] = CNTR_ELEM("TidWait", 0, 0, CNTR_NORMAL,
+			    hfi1_access_sw_tid_wait),
 [C_SW_SEND_SCHED] = CNTR_ELEM("SendSched", 0, 0, CNTR_NORMAL,
 			    access_sw_send_schedule),
 [C_SDMA_DESC_FETCHED_CNT] = CNTR_ELEM("SDEDscFdCn",
diff --git a/drivers/infiniband/hw/hfi1/chip.h b/drivers/infiniband/hw/hfi1/chip.h
index ba3d99e6e33b..6c27c1c6a868 100644
--- a/drivers/infiniband/hw/hfi1/chip.h
+++ b/drivers/infiniband/hw/hfi1/chip.h
@@ -927,6 +927,7 @@ enum {
 	C_SW_PIO_WAIT,
 	C_SW_PIO_DRAIN,
 	C_SW_KMEM_WAIT,
+	C_SW_TID_WAIT,
 	C_SW_SEND_SCHED,
 	C_SDMA_DESC_FETCHED_CNT,
 	C_SDMA_INT_CNT,
diff --git a/drivers/infiniband/hw/hfi1/common.h b/drivers/infiniband/hw/hfi1/common.h
index 40d3cfb58bd1..7310a5dba420 100644
--- a/drivers/infiniband/hw/hfi1/common.h
+++ b/drivers/infiniband/hw/hfi1/common.h
@@ -340,6 +340,10 @@ struct diag_pkt {
 
 #define HFI1_PSM_IOC_BASE_SEQ 0x0
 
+/* Number of BTH.PSN bits used for sequence number in expected rcvs */
+#define HFI1_KDETH_BTH_SEQ_SHIFT 11
+#define HFI1_KDETH_BTH_SEQ_MASK (BIT(HFI1_KDETH_BTH_SEQ_SHIFT) - 1)
+
 static inline __u64 rhf_to_cpu(const __le32 *rbuf)
 {
 	return __le64_to_cpu(*((__le64 *)rbuf));
diff --git a/drivers/infiniband/hw/hfi1/driver.c b/drivers/infiniband/hw/hfi1/driver.c
index a8ad70730203..2a9d2912f5db 100644
--- a/drivers/infiniband/hw/hfi1/driver.c
+++ b/drivers/infiniband/hw/hfi1/driver.c
@@ -1575,25 +1575,32 @@ drop:
 	return -EINVAL;
 }
 
-void handle_eflags(struct hfi1_packet *packet)
+static void show_eflags_errs(struct hfi1_packet *packet)
 {
 	struct hfi1_ctxtdata *rcd = packet->rcd;
 	u32 rte = rhf_rcv_type_err(packet->rhf);
 
+	dd_dev_err(rcd->dd,
+		   "receive context %d: rhf 0x%016llx, errs [ %s%s%s%s%s%s%s%s] rte 0x%x\n",
+		   rcd->ctxt, packet->rhf,
+		   packet->rhf & RHF_K_HDR_LEN_ERR ? "k_hdr_len " : "",
+		   packet->rhf & RHF_DC_UNC_ERR ? "dc_unc " : "",
+		   packet->rhf & RHF_DC_ERR ? "dc " : "",
+		   packet->rhf & RHF_TID_ERR ? "tid " : "",
+		   packet->rhf & RHF_LEN_ERR ? "len " : "",
+		   packet->rhf & RHF_ECC_ERR ? "ecc " : "",
+		   packet->rhf & RHF_VCRC_ERR ? "vcrc " : "",
+		   packet->rhf & RHF_ICRC_ERR ? "icrc " : "",
+		   rte);
+}
+
+void handle_eflags(struct hfi1_packet *packet)
+{
+	struct hfi1_ctxtdata *rcd = packet->rcd;
+
 	rcv_hdrerr(rcd, rcd->ppd, packet);
 	if (rhf_err_flags(packet->rhf))
-		dd_dev_err(rcd->dd,
-			   "receive context %d: rhf 0x%016llx, errs [ %s%s%s%s%s%s%s%s] rte 0x%x\n",
-			   rcd->ctxt, packet->rhf,
-			   packet->rhf & RHF_K_HDR_LEN_ERR ? "k_hdr_len " : "",
-			   packet->rhf & RHF_DC_UNC_ERR ? "dc_unc " : "",
-			   packet->rhf & RHF_DC_ERR ? "dc " : "",
-			   packet->rhf & RHF_TID_ERR ? "tid " : "",
-			   packet->rhf & RHF_LEN_ERR ? "len " : "",
-			   packet->rhf & RHF_ECC_ERR ? "ecc " : "",
-			   packet->rhf & RHF_VCRC_ERR ? "vcrc " : "",
-			   packet->rhf & RHF_ICRC_ERR ? "icrc " : "",
-			   rte);
+		show_eflags_errs(packet);
 }
 
 /*
@@ -1699,11 +1706,14 @@ static int kdeth_process_expected(struct hfi1_packet *packet)
 	if (unlikely(hfi1_dbg_should_fault_rx(packet)))
 		return RHF_RCV_CONTINUE;
 
-	if (unlikely(rhf_err_flags(packet->rhf)))
-		handle_eflags(packet);
+	if (unlikely(rhf_err_flags(packet->rhf))) {
+		struct hfi1_ctxtdata *rcd = packet->rcd;
 
-	dd_dev_err(packet->rcd->dd,
-		   "Unhandled expected packet received. Dropping.\n");
+		if (hfi1_handle_kdeth_eflags(rcd, rcd->ppd, packet))
+			return RHF_RCV_CONTINUE;
+	}
+
+	hfi1_kdeth_expected_rcv(packet);
 	return RHF_RCV_CONTINUE;
 }
 
@@ -1712,11 +1722,17 @@ static int kdeth_process_eager(struct hfi1_packet *packet)
 	hfi1_setup_9B_packet(packet);
 	if (unlikely(hfi1_dbg_should_fault_rx(packet)))
 		return RHF_RCV_CONTINUE;
-	if (unlikely(rhf_err_flags(packet->rhf)))
-		handle_eflags(packet);
 
-	dd_dev_err(packet->rcd->dd,
-		   "Unhandled eager packet received. Dropping.\n");
+	trace_hfi1_rcvhdr(packet);
+	if (unlikely(rhf_err_flags(packet->rhf))) {
+		struct hfi1_ctxtdata *rcd = packet->rcd;
+
+		show_eflags_errs(packet);
+		if (hfi1_handle_kdeth_eflags(rcd, rcd->ppd, packet))
+			return RHF_RCV_CONTINUE;
+	}
+
+	hfi1_kdeth_eager_rcv(packet);
 	return RHF_RCV_CONTINUE;
 }
 
diff --git a/drivers/infiniband/hw/hfi1/hfi.h b/drivers/infiniband/hw/hfi1/hfi.h
index 9aa0357e17b7..6582184cc985 100644
--- a/drivers/infiniband/hw/hfi1/hfi.h
+++ b/drivers/infiniband/hw/hfi1/hfi.h
@@ -198,6 +198,14 @@ struct exp_tid_set {
 };
 
 typedef int (*rhf_rcv_function_ptr)(struct hfi1_packet *packet);
+
+struct tid_queue {
+	struct list_head queue_head;
+			/* queue head for QP TID resource waiters */
+	u32 enqueue;	/* count of tid enqueues */
+	u32 dequeue;	/* count of tid dequeues */
+};
+
 struct hfi1_ctxtdata {
 	/* rcvhdrq base, needs mmap before useful */
 	void *rcvhdrq;
@@ -291,6 +299,12 @@ struct hfi1_ctxtdata {
 	/* PSM Specific fields */
 	/* lock protecting all Expected TID data */
 	struct mutex exp_mutex;
+	/* lock protecting all Expected TID data of kernel contexts */
+	spinlock_t exp_lock;
+	/* Queue for QP's waiting for HW TID flows */
+	struct tid_queue flow_queue;
+	/* Queue for QP's waiting for HW receive array entries */
+	struct tid_queue rarr_queue;
 	/* when waiting for rcv or pioavail */
 	wait_queue_head_t wait;
 	/* uuid from PSM */
@@ -323,6 +337,9 @@ struct hfi1_ctxtdata {
 	 */
 	u8 subctxt_cnt;
 
+	/* Bit mask to track free TID RDMA HW flows */
+	unsigned long flow_mask;
+	struct tid_flow_state flows[RXE_NUM_TID_FLOWS];
 };
 
 /**
@@ -2103,7 +2120,7 @@ static inline u64 hfi1_pkt_default_send_ctxt_mask(struct hfi1_devdata *dd,
 			SEND_CTXT_CHECK_ENABLE_DISALLOW_PBC_TEST_SMASK |
 #endif
 			HFI1_PKT_USER_SC_INTEGRITY;
-	else
+	else if (ctxt_type != SC_KERNEL)
 		base_sc_integrity |= HFI1_PKT_KERNEL_SC_INTEGRITY;
 
 	/* turn on send-side job key checks if !A0 */
diff --git a/drivers/infiniband/hw/hfi1/init.c b/drivers/infiniband/hw/hfi1/init.c
index a8dbd0f191f5..d13304f7340d 100644
--- a/drivers/infiniband/hw/hfi1/init.c
+++ b/drivers/infiniband/hw/hfi1/init.c
@@ -370,6 +370,9 @@ int hfi1_create_ctxtdata(struct hfi1_pportdata *ppd, int numa,
 		rcd->rhf_rcv_function_map = normal_rhf_rcv_functions;
 
 		mutex_init(&rcd->exp_mutex);
+		spin_lock_init(&rcd->exp_lock);
+		INIT_LIST_HEAD(&rcd->flow_queue.queue_head);
+		INIT_LIST_HEAD(&rcd->rarr_queue.queue_head);
 
 		hfi1_cdbg(PROC, "setting up context %u\n", rcd->ctxt);
 
@@ -472,6 +475,9 @@ int hfi1_create_ctxtdata(struct hfi1_pportdata *ppd, int numa,
 						    GFP_KERNEL, numa);
 			if (!rcd->opstats)
 				goto bail;
+
+			/* Initialize TID flow generations for the context */
+			hfi1_kern_init_ctxt_generations(rcd);
 		}
 
 		*context = rcd;
@@ -771,6 +777,8 @@ static void enable_chip(struct hfi1_devdata *dd)
 			rcvmask |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
 		if (HFI1_CAP_KGET_MASK(rcd->flags, NODROP_EGR_FULL))
 			rcvmask |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
+		if (HFI1_CAP_IS_KSET(TID_RDMA))
+			rcvmask |= HFI1_RCVCTRL_TIDFLOW_ENB;
 		hfi1_rcvctrl(dd, rcvmask, rcd);
 		sc_enable(rcd->sc);
 		hfi1_rcd_put(rcd);
@@ -1589,7 +1597,7 @@ static void cleanup_device_data(struct hfi1_devdata *dd)
 		struct hfi1_ctxtdata *rcd = dd->rcd[ctxt];
 
 		if (rcd) {
-			hfi1_clear_tids(rcd);
+			hfi1_free_ctxt_rcv_groups(rcd);
 			hfi1_free_ctxt(rcd);
 		}
 	}
diff --git a/drivers/infiniband/hw/hfi1/qp.c b/drivers/infiniband/hw/hfi1/qp.c
index f822f92b415f..acdd9eba189b 100644
--- a/drivers/infiniband/hw/hfi1/qp.c
+++ b/drivers/infiniband/hw/hfi1/qp.c
@@ -319,6 +319,7 @@ int hfi1_setup_wqe(struct rvt_qp *qp, struct rvt_swqe *wqe, bool *call_send)
 
 	switch (qp->ibqp.qp_type) {
 	case IB_QPT_RC:
+		hfi1_setup_tid_rdma_wqe(qp, wqe);
 	case IB_QPT_UC:
 		if (wqe->length > 0x80000000U)
 			return -EINVAL;
@@ -738,6 +739,7 @@ void flush_qp_waiters(struct rvt_qp *qp)
 {
 	lockdep_assert_held(&qp->s_lock);
 	flush_iowait(qp);
+	hfi1_tid_rdma_flush_wait(qp);
 }
 
 void stop_send_queue(struct rvt_qp *qp)
@@ -745,6 +747,8 @@ void stop_send_queue(struct rvt_qp *qp)
 	struct hfi1_qp_priv *priv = qp->priv;
 
 	iowait_cancel_work(&priv->s_iowait);
+	if (cancel_work_sync(&priv->tid_rdma.trigger_work))
+		rvt_put_qp(qp);
 }
 
 void quiesce_qp(struct rvt_qp *qp)
@@ -758,6 +762,7 @@ void quiesce_qp(struct rvt_qp *qp)
 
 void notify_qp_reset(struct rvt_qp *qp)
 {
+	hfi1_qp_kern_exp_rcv_clear_all(qp);
 	qp->r_adefered = 0;
 	clear_ahg(qp);
 
diff --git a/drivers/infiniband/hw/hfi1/qp.h b/drivers/infiniband/hw/hfi1/qp.h
index 7adb6dff6813..ce25a27aa4a1 100644
--- a/drivers/infiniband/hw/hfi1/qp.h
+++ b/drivers/infiniband/hw/hfi1/qp.h
@@ -63,11 +63,13 @@ extern const struct rvt_operation_params hfi1_post_parms[];
  * HFI1_S_AHG_VALID - ahg header valid on chip
  * HFI1_S_AHG_CLEAR - have send engine clear ahg state
  * HFI1_S_WAIT_PIO_DRAIN - qp waiting for PIOs to drain
+ * HFI1_S_WAIT_TID_SPACE - a QP is waiting for TID resource
  * HFI1_S_MIN_BIT_MASK - the lowest bit that can be used by hfi1
  */
 #define HFI1_S_AHG_VALID         0x80000000
 #define HFI1_S_AHG_CLEAR         0x40000000
 #define HFI1_S_WAIT_PIO_DRAIN    0x20000000
+#define HFI1_S_WAIT_TID_SPACE    0x10000000
 #define HFI1_S_MIN_BIT_MASK      0x01000000
 
 /*
diff --git a/drivers/infiniband/hw/hfi1/rc.c b/drivers/infiniband/hw/hfi1/rc.c
index 092d5eba980f..6c9ef572fc69 100644
--- a/drivers/infiniband/hw/hfi1/rc.c
+++ b/drivers/infiniband/hw/hfi1/rc.c
@@ -51,28 +51,48 @@
 
 #include "hfi.h"
 #include "qp.h"
+#include "rc.h"
 #include "verbs_txreq.h"
 #include "trace.h"
 
-/* cut down ridiculously long IB macro names */
-#define OP(x) RC_OP(x)
-
-static struct rvt_swqe *do_rc_completion(struct rvt_qp *qp,
-					 struct rvt_swqe *wqe,
-					 struct hfi1_ibport *ibp);
-
-static u32 restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe,
-		       u32 psn, u32 pmtu)
+struct rvt_ack_entry *find_prev_entry(struct rvt_qp *qp, u32 psn, u8 *prev,
+				      u8 *prev_ack, bool *scheduled)
+	__must_hold(&qp->s_lock)
 {
-	u32 len;
-
-	len = delta_psn(psn, wqe->psn) * pmtu;
-	ss->sge = wqe->sg_list[0];
-	ss->sg_list = wqe->sg_list + 1;
-	ss->num_sge = wqe->wr.num_sge;
-	ss->total_len = wqe->length;
-	rvt_skip_sge(ss, len, false);
-	return wqe->length - len;
+	struct rvt_ack_entry *e = NULL;
+	u8 i, p;
+	bool s = true;
+
+	for (i = qp->r_head_ack_queue; ; i = p) {
+		if (i == qp->s_tail_ack_queue)
+			s = false;
+		if (i)
+			p = i - 1;
+		else
+			p = rvt_size_atomic(ib_to_rvt(qp->ibqp.device));
+		if (p == qp->r_head_ack_queue) {
+			e = NULL;
+			break;
+		}
+		e = &qp->s_ack_queue[p];
+		if (!e->opcode) {
+			e = NULL;
+			break;
+		}
+		if (cmp_psn(psn, e->psn) >= 0) {
+			if (p == qp->s_tail_ack_queue &&
+			    cmp_psn(psn, e->lpsn) <= 0)
+				s = false;
+			break;
+		}
+	}
+	if (prev)
+		*prev = p;
+	if (prev_ack)
+		*prev_ack = i;
+	if (scheduled)
+		*scheduled = s;
+	return e;
 }
 
 /**
@@ -92,13 +112,16 @@ static int make_rc_ack(struct hfi1_ibdev *dev, struct rvt_qp *qp,
 {
 	struct rvt_ack_entry *e;
 	u32 hwords;
-	u32 len;
-	u32 bth0, bth2;
+	u32 len = 0;
+	u32 bth0 = 0, bth2 = 0;
 	u32 bth1 = qp->remote_qpn | (HFI1_CAP_IS_KSET(OPFN) << IB_BTHE_E_SHIFT);
 	int middle = 0;
 	u32 pmtu = qp->pmtu;
 	struct hfi1_qp_priv *priv = qp->priv;
+	bool last_pkt;
+	u32 delta;
 
+	trace_hfi1_rsp_make_rc_ack(qp, 0);
 	lockdep_assert_held(&qp->s_lock);
 	/* Don't send an ACK if we aren't supposed to. */
 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
@@ -170,6 +193,26 @@ static int make_rc_ack(struct hfi1_ibdev *dev, struct rvt_qp *qp,
 			hwords++;
 			qp->s_ack_rdma_psn = e->psn;
 			bth2 = mask_psn(qp->s_ack_rdma_psn++);
+		} else if (e->opcode == TID_OP(READ_REQ)) {
+			/*
+			 * If a TID RDMA read response is being resent and
+			 * we haven't seen the duplicate request yet,
+			 * then stop sending the remaining responses the
+			 * responder has seen until the requester re-sends it.
+			 */
+			len = e->rdma_sge.sge_length;
+			if (len && !e->rdma_sge.mr) {
+				qp->s_tail_ack_queue = qp->r_head_ack_queue;
+				goto bail;
+			}
+			/* Copy SGE state in case we need to resend */
+			ps->s_txreq->mr = e->rdma_sge.mr;
+			if (ps->s_txreq->mr)
+				rvt_get_mr(ps->s_txreq->mr);
+			qp->s_ack_rdma_sge.sge = e->rdma_sge;
+			qp->s_ack_rdma_sge.num_sge = 1;
+			qp->s_ack_state = TID_OP(READ_RESP);
+			goto read_resp;
 		} else {
 			/* COMPARE_SWAP or FETCH_ADD */
 			ps->s_txreq->ss = NULL;
@@ -207,6 +250,28 @@ static int make_rc_ack(struct hfi1_ibdev *dev, struct rvt_qp *qp,
 		bth2 = mask_psn(qp->s_ack_rdma_psn++);
 		break;
 
+	case TID_OP(READ_RESP):
+read_resp:
+		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
+		ps->s_txreq->ss = &qp->s_ack_rdma_sge;
+		delta = hfi1_build_tid_rdma_read_resp(qp, e, ohdr, &bth0,
+						      &bth1, &bth2, &len,
+						      &last_pkt);
+		if (delta == 0)
+			goto error_qp;
+		hwords += delta;
+		if (last_pkt) {
+			e->sent = 1;
+			/*
+			 * Increment qp->s_tail_ack_queue through s_ack_state
+			 * transition.
+			 */
+			qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
+		}
+		break;
+	case TID_OP(READ_REQ):
+		goto bail;
+
 	default:
 normal:
 		/*
@@ -236,7 +301,14 @@ normal:
 	ps->s_txreq->hdr_dwords = hwords;
 	hfi1_make_ruc_header(qp, ohdr, bth0, bth1, bth2, middle, ps);
 	return 1;
-
+error_qp:
+	spin_unlock_irqrestore(&qp->s_lock, ps->flags);
+	spin_lock_irqsave(&qp->r_lock, ps->flags);
+	spin_lock(&qp->s_lock);
+	rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
+	spin_unlock(&qp->s_lock);
+	spin_unlock_irqrestore(&qp->r_lock, ps->flags);
+	spin_lock_irqsave(&qp->s_lock, ps->flags);
 bail:
 	qp->s_ack_state = OP(ACKNOWLEDGE);
 	/*
@@ -263,17 +335,22 @@ int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 	struct hfi1_qp_priv *priv = qp->priv;
 	struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
 	struct ib_other_headers *ohdr;
-	struct rvt_sge_state *ss;
+	struct rvt_sge_state *ss = NULL;
 	struct rvt_swqe *wqe;
-	u32 hwords;
-	u32 len;
-	u32 bth0 = 0, bth2;
+	struct hfi1_swqe_priv *wpriv;
+	struct tid_rdma_request *req = NULL;
+	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
+	u32 hwords = 5;
+	u32 len = 0;
+	u32 bth0 = 0, bth2 = 0;
 	u32 bth1 = qp->remote_qpn | (HFI1_CAP_IS_KSET(OPFN) << IB_BTHE_E_SHIFT);
 	u32 pmtu = qp->pmtu;
 	char newreq;
 	int middle = 0;
 	int delta;
+	struct tid_rdma_flow *flow = NULL;
 
+	trace_hfi1_sender_make_rc_req(qp);
 	lockdep_assert_held(&qp->s_lock);
 	ps->s_txreq = get_txreq(ps->dev, qp);
 	if (!ps->s_txreq)
@@ -314,8 +391,8 @@ int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 		}
 		clear_ahg(qp);
 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
-		rvt_send_complete(qp, wqe, qp->s_last != qp->s_acked ?
-			IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR);
+		hfi1_trdma_send_complete(qp, wqe, qp->s_last != qp->s_acked ?
+					 IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR);
 		/* will get called again */
 		goto done_free_tx;
 	}
@@ -334,6 +411,7 @@ int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 
 	/* Send a request. */
 	wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
+check_s_state:
 	switch (qp->s_state) {
 	default:
 		if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK))
@@ -355,9 +433,13 @@ int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 			/*
 			 * If a fence is requested, wait for previous
 			 * RDMA read and atomic operations to finish.
+			 * However, there is no need to guard against
+			 * TID RDMA READ after TID RDMA READ.
 			 */
 			if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
-			    qp->s_num_rd_atomic) {
+			    qp->s_num_rd_atomic &&
+			    (wqe->wr.opcode != IB_WR_TID_RDMA_READ ||
+			     priv->pending_tid_r_segs < qp->s_num_rd_atomic)) {
 				qp->s_flags |= RVT_S_WAIT_FENCE;
 				goto bail;
 			}
@@ -402,6 +484,15 @@ int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 		len = wqe->length;
 		ss = &qp->s_sge;
 		bth2 = mask_psn(qp->s_psn);
+
+		/*
+		 * Interlock between various IB requests and TID RDMA
+		 * if necessary.
+		 */
+		if ((priv->s_flags & HFI1_S_TID_WAIT_INTERLCK) ||
+		    hfi1_tid_rdma_wqe_interlock(qp, wqe))
+			goto bail;
+
 		switch (wqe->wr.opcode) {
 		case IB_WR_SEND:
 		case IB_WR_SEND_WITH_IMM:
@@ -483,16 +574,14 @@ no_flow_control:
 			 * Don't allow more operations to be started
 			 * than the QP limits allow.
 			 */
-			if (newreq) {
-				if (qp->s_num_rd_atomic >=
-				    qp->s_max_rd_atomic) {
-					qp->s_flags |= RVT_S_WAIT_RDMAR;
-					goto bail;
-				}
-				qp->s_num_rd_atomic++;
-				if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
-					qp->s_lsn++;
+			if (qp->s_num_rd_atomic >=
+			    qp->s_max_rd_atomic) {
+				qp->s_flags |= RVT_S_WAIT_RDMAR;
+				goto bail;
 			}
+			qp->s_num_rd_atomic++;
+			if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
+				qp->s_lsn++;
 			put_ib_reth_vaddr(
 				wqe->rdma_wr.remote_addr,
 				&ohdr->u.rc.reth);
@@ -508,20 +597,92 @@ no_flow_control:
 				qp->s_cur = 0;
 			break;
 
+		case IB_WR_TID_RDMA_READ:
+			trace_hfi1_tid_read_sender_make_req(qp, newreq);
+			wpriv = wqe->priv;
+			req = wqe_to_tid_req(wqe);
+			trace_hfi1_tid_req_make_req_read(qp, newreq,
+							 wqe->wr.opcode,
+							 wqe->psn, wqe->lpsn,
+							 req);
+			delta = cmp_psn(qp->s_psn, wqe->psn);
+
+			/*
+			 * Don't allow more operations to be started
+			 * than the QP limits allow. We could get here under
+			 * three conditions; (1) It's a new request; (2) We are
+			 * sending the second or later segment of a request,
+			 * but the qp->s_state is set to OP(RDMA_READ_REQUEST)
+			 * when the last segment of a previous request is
+			 * received just before this; (3) We are re-sending a
+			 * request.
+			 */
+			if (qp->s_num_rd_atomic >= qp->s_max_rd_atomic) {
+				qp->s_flags |= RVT_S_WAIT_RDMAR;
+				goto bail;
+			}
+			if (newreq) {
+				struct tid_rdma_flow *flow =
+					&req->flows[req->setup_head];
+
+				/*
+				 * Set up s_sge as it is needed for TID
+				 * allocation. However, if the pages have been
+				 * walked and mapped, skip it. An earlier try
+				 * has failed to allocate the TID entries.
+				 */
+				if (!flow->npagesets) {
+					qp->s_sge.sge = wqe->sg_list[0];
+					qp->s_sge.sg_list = wqe->sg_list + 1;
+					qp->s_sge.num_sge = wqe->wr.num_sge;
+					qp->s_sge.total_len = wqe->length;
+					qp->s_len = wqe->length;
+					req->isge = 0;
+					req->clear_tail = req->setup_head;
+					req->flow_idx = req->setup_head;
+					req->state = TID_REQUEST_ACTIVE;
+				}
+			} else if (delta == 0) {
+				/* Re-send a request */
+				req->cur_seg = 0;
+				req->comp_seg = 0;
+				req->ack_pending = 0;
+				req->flow_idx = req->clear_tail;
+				req->state = TID_REQUEST_RESEND;
+			}
+			req->s_next_psn = qp->s_psn;
+			/* Read one segment at a time */
+			len = min_t(u32, req->seg_len,
+				    wqe->length - req->seg_len * req->cur_seg);
+			delta = hfi1_build_tid_rdma_read_req(qp, wqe, ohdr,
+							     &bth1, &bth2,
+							     &len);
+			if (delta <= 0) {
+				/* Wait for TID space */
+				goto bail;
+			}
+			if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
+				qp->s_lsn++;
+			hwords += delta;
+			ss = &wpriv->ss;
+			/* Check if this is the last segment */
+			if (req->cur_seg >= req->total_segs &&
+			    ++qp->s_cur == qp->s_size)
+				qp->s_cur = 0;
+			break;
+
 		case IB_WR_ATOMIC_CMP_AND_SWP:
 		case IB_WR_ATOMIC_FETCH_AND_ADD:
 			/*
 			 * Don't allow more operations to be started
 			 * than the QP limits allow.
 			 */
-			if (newreq) {
-				if (qp->s_num_rd_atomic >=
-				    qp->s_max_rd_atomic) {
-					qp->s_flags |= RVT_S_WAIT_RDMAR;
-					goto bail;
-				}
-				qp->s_num_rd_atomic++;
+			if (qp->s_num_rd_atomic >=
+			    qp->s_max_rd_atomic) {
+				qp->s_flags |= RVT_S_WAIT_RDMAR;
+				goto bail;
 			}
+			qp->s_num_rd_atomic++;
 
 			/* FALLTHROUGH */
 		case IB_WR_OPFN:
@@ -555,11 +716,13 @@ no_flow_control:
 		default:
 			goto bail;
 		}
-		qp->s_sge.sge = wqe->sg_list[0];
-		qp->s_sge.sg_list = wqe->sg_list + 1;
-		qp->s_sge.num_sge = wqe->wr.num_sge;
-		qp->s_sge.total_len = wqe->length;
-		qp->s_len = wqe->length;
+		if (wqe->wr.opcode != IB_WR_TID_RDMA_READ) {
+			qp->s_sge.sge = wqe->sg_list[0];
+			qp->s_sge.sg_list = wqe->sg_list + 1;
+			qp->s_sge.num_sge = wqe->wr.num_sge;
+			qp->s_sge.total_len = wqe->length;
+			qp->s_len = wqe->length;
+		}
 		if (newreq) {
 			qp->s_tail++;
 			if (qp->s_tail >= qp->s_size)
@@ -567,6 +730,8 @@ no_flow_control:
 		}
 		if (wqe->wr.opcode == IB_WR_RDMA_READ)
 			qp->s_psn = wqe->lpsn + 1;
+		else if (wqe->wr.opcode == IB_WR_TID_RDMA_READ)
+			qp->s_psn = req->s_next_psn;
 		else
 			qp->s_psn++;
 		break;
@@ -683,6 +848,103 @@ no_flow_control:
 		if (qp->s_cur == qp->s_size)
 			qp->s_cur = 0;
 		break;
+	case TID_OP(READ_RESP):
+		if (wqe->wr.opcode != IB_WR_TID_RDMA_READ)
+			goto bail;
+		/* This is used to restart a TID read request */
+		req = wqe_to_tid_req(wqe);
+		wpriv = wqe->priv;
+		/*
+		 * Back down. The field qp->s_psn has been set to the psn with
+		 * which the request should be restart. It's OK to use division
+		 * as this is on the retry path.
+		 */
+		req->cur_seg = delta_psn(qp->s_psn, wqe->psn) / priv->pkts_ps;
+
+		/*
+		 * The following function need to be redefined to return the
+		 * status to make sure that we find the flow. At the same
+		 * time, we can use the req->state change to check if the
+		 * call succeeds or not.
+		 */
+		req->state = TID_REQUEST_RESEND;
+		hfi1_tid_rdma_restart_req(qp, wqe, &bth2);
+		if (req->state != TID_REQUEST_ACTIVE) {
+			/*
+			 * Failed to find the flow. Release all allocated tid
+			 * resources.
+			 */
+			hfi1_kern_exp_rcv_clear_all(req);
+			hfi1_kern_clear_hw_flow(priv->rcd, qp);
+
+			hfi1_trdma_send_complete(qp, wqe, IB_WC_LOC_QP_OP_ERR);
+			goto bail;
+		}
+		req->state = TID_REQUEST_RESEND;
+		len = min_t(u32, req->seg_len,
+			    wqe->length - req->seg_len * req->cur_seg);
+		flow = &req->flows[req->flow_idx];
+		len -= flow->sent;
+		req->s_next_psn = flow->flow_state.ib_lpsn + 1;
+		delta = hfi1_build_tid_rdma_read_packet(wqe, ohdr, &bth1,
+							&bth2, &len);
+		if (delta <= 0) {
+			/* Wait for TID space */
+			goto bail;
+		}
+		hwords += delta;
+		ss = &wpriv->ss;
+		/* Check if this is the last segment */
+		if (req->cur_seg >= req->total_segs &&
+		    ++qp->s_cur == qp->s_size)
+			qp->s_cur = 0;
+		qp->s_psn = req->s_next_psn;
+		trace_hfi1_tid_req_make_req_read(qp, 0, wqe->wr.opcode,
+						 wqe->psn, wqe->lpsn, req);
+		break;
+	case TID_OP(READ_REQ):
+		req = wqe_to_tid_req(wqe);
+		delta = cmp_psn(qp->s_psn, wqe->psn);
+		/*
+		 * If the current WR is not TID RDMA READ, or this is the start
+		 * of a new request, we need to change the qp->s_state so that
+		 * the request can be set up properly.
+		 */
+		if (wqe->wr.opcode != IB_WR_TID_RDMA_READ || delta == 0 ||
+		    qp->s_cur == qp->s_tail) {
+			qp->s_state = OP(RDMA_READ_REQUEST);
+			if (delta == 0 || qp->s_cur == qp->s_tail)
+				goto check_s_state;
+			else
+				goto bail;
+		}
+
+		/* Rate limiting */
+		if (qp->s_num_rd_atomic >= qp->s_max_rd_atomic) {
+			qp->s_flags |= RVT_S_WAIT_RDMAR;
+			goto bail;
+		}
+
+		wpriv = wqe->priv;
+		/* Read one segment at a time */
+		len = min_t(u32, req->seg_len,
+			    wqe->length - req->seg_len * req->cur_seg);
+		delta = hfi1_build_tid_rdma_read_req(qp, wqe, ohdr, &bth1,
+						     &bth2, &len);
+		if (delta <= 0) {
+			/* Wait for TID space */
+			goto bail;
+		}
+		hwords += delta;
+		ss = &wpriv->ss;
+		/* Check if this is the last segment */
+		if (req->cur_seg >= req->total_segs &&
+		    ++qp->s_cur == qp->s_size)
+			qp->s_cur = 0;
+		qp->s_psn = req->s_next_psn;
+		trace_hfi1_tid_req_make_req_read(qp, 0, wqe->wr.opcode,
+						 wqe->psn, wqe->lpsn, req);
+		break;
 	}
 	qp->s_sending_hpsn = bth2;
 	delta = delta_psn(bth2, wqe->psn);
@@ -951,6 +1213,43 @@ void hfi1_send_rc_ack(struct hfi1_packet *packet, bool is_fecn)
 }
 
 /**
+ * update_num_rd_atomic - update the qp->s_num_rd_atomic
+ * @qp: the QP
+ * @psn: the packet sequence number to restart at
+ * @wqe: the wqe
+ *
+ * This is called from reset_psn() to update qp->s_num_rd_atomic
+ * for the current wqe.
+ * Called at interrupt level with the QP s_lock held.
+ */
+static void update_num_rd_atomic(struct rvt_qp *qp, u32 psn,
+				 struct rvt_swqe *wqe)
+{
+	u32 opcode = wqe->wr.opcode;
+
+	if (opcode == IB_WR_RDMA_READ ||
+	    opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
+	    opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
+		qp->s_num_rd_atomic++;
+	} else if (opcode == IB_WR_TID_RDMA_READ) {
+		struct tid_rdma_request *req = wqe_to_tid_req(wqe);
+		stru