Merge branch 'tid-write' into hfi1-tid

Here is the final set of patches for TID RDMA. Again this is code which
was previously submitted but re-organized so as to be easier to review.

Similar to how the READ series was organized the patches to build,
receive, allocate resources etc are broken out. For details on TID RDMA
as a whole again refer to the original cover letter.

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

* tid-write: (23 commits)
  IB/hfi1: Prioritize the sending of ACK packets
  IB/hfi1: Add static trace for TID RDMA WRITE protocol
  IB/hfi1: Enable TID RDMA WRITE protocol
  IB/hfi1: Add interlock between TID RDMA WRITE and other requests
  IB/hfi1: Add TID RDMA WRITE functionality into RDMA verbs
  IB/hfi1: Add the dual leg code
  IB/hfi1: Add the TID second leg ACK packet builder
  IB/hfi1: Add the TID second leg send packet builder
  IB/hfi1: Resend the TID RDMA WRITE DATA packets
  IB/hfi1: Add a function to receive TID RDMA RESYNC packet
  IB/hfi1: Add a function to build TID RDMA RESYNC packet
  IB/hfi1: Add TID RDMA retry timer
  IB/hfi1: Add a function to receive TID RDMA ACK packet
  IB/hfi1: Add a function to build TID RDMA ACK packet
  IB/hfi1: Add a function to receive TID RDMA WRITE DATA packet
  IB/hfi1: Add a function to build TID RDMA WRITE DATA packet
  IB/hfi1: Add a function to receive TID RDMA WRITE response
  IB/hfi1: Add TID resource timer
  IB/hfi1: Add a function to build TID RDMA WRITE response
  IB/hfi1: Add functions to receive TID RDMA WRITE request
  ...
Signed-off-by: Doug Ledford <dledford@redhat.com>

drivers/infiniband/hw/hfi1/init.c

@@ -1512,6 +1512,7 @@ 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/iowait.c

@@ -6,6 +6,9 @@
 #include "iowait.h"
 #include "trace_iowait.h"
 
+/* 1 priority == 16 starve_cnt */
+#define IOWAIT_PRIORITY_STARVE_SHIFT 4
+
 void iowait_set_flag(struct iowait *wait, u32 flag)
 {
 	trace_hfi1_iowait_set(wait, flag);
@@ -44,7 +47,8 @@ void iowait_init(struct iowait *wait, u32 tx_limit,
 			      uint seq,
 			      bool pkts_sent),
 		 void (*wakeup)(struct iowait *wait, int reason),
-		 void (*sdma_drained)(struct iowait *wait))
+		 void (*sdma_drained)(struct iowait *wait),
+		 void (*init_priority)(struct iowait *wait))
 {
 	int i;
 
@@ -58,6 +62,7 @@ void iowait_init(struct iowait *wait, u32 tx_limit,
 	wait->sleep = sleep;
 	wait->wakeup = wakeup;
 	wait->sdma_drained = sdma_drained;
+	wait->init_priority = init_priority;
 	wait->flags = 0;
 	for (i = 0; i < IOWAIT_SES; i++) {
 		wait->wait[i].iow = wait;
@@ -92,3 +97,30 @@ int iowait_set_work_flag(struct iowait_work *w)
 	iowait_set_flag(w->iow, IOWAIT_PENDING_TID);
 	return IOWAIT_TID_SE;
 }
+
+/**
+ * iowait_priority_update_top - update the top priority entry
+ * @w: the iowait struct
+ * @top: a pointer to the top priority entry
+ * @idx: the index of the current iowait in an array
+ * @top_idx: the array index for the iowait entry that has the top priority
+ *
+ * This function is called to compare the priority of a given
+ * iowait with the given top priority entry. The top index will
+ * be returned.
+ */
+uint iowait_priority_update_top(struct iowait *w,
+				struct iowait *top,
+				uint idx, uint top_idx)
+{
+	u8 cnt, tcnt;
+
+	/* Convert priority into starve_cnt and compare the total.*/
+	cnt = (w->priority << IOWAIT_PRIORITY_STARVE_SHIFT) + w->starved_cnt;
+	tcnt = (top->priority << IOWAIT_PRIORITY_STARVE_SHIFT) +
+		top->starved_cnt;
+	if (cnt > tcnt)
+		return idx;
+	else
+		return top_idx;
+}

drivers/infiniband/hw/hfi1/iowait.h

@@ -100,6 +100,7 @@ struct iowait_work {
  * @sleep: no space callback
  * @wakeup: space callback wakeup
  * @sdma_drained: sdma count drained
+ * @init_priority: callback to manipulate priority
  * @lock: lock protected head of wait queue
  * @iowork: workqueue overhead
  * @wait_dma: wait for sdma_busy == 0
@@ -109,7 +110,7 @@ struct iowait_work {
  * @tx_limit: limit for overflow queuing
  * @tx_count: number of tx entry's in tx_head'ed list
  * @flags: wait flags (one per QP)
- * @wait: SE array
+ * @wait: SE array for multiple legs
  *
  * This is to be embedded in user's state structure
  * (QP or PQ).
@@ -120,10 +121,13 @@ struct iowait_work {
  * are callbacks for the ULP to implement
  * what ever queuing/dequeuing of
  * the embedded iowait and its containing struct
- * when a resource shortage like SDMA ring space is seen.
+ * when a resource shortage like SDMA ring space
+ * or PIO credit space is seen.
  *
  * Both potentially have locks help
- * so sleeping is not allowed.
+ * so sleeping is not allowed and it is not
+ * supported to submit txreqs from the wakeup
+ * call directly because of lock conflicts.
  *
  * The wait_dma member along with the iow
  *
@@ -143,6 +147,7 @@ struct iowait {
 		);
 	void (*wakeup)(struct iowait *wait, int reason);
 	void (*sdma_drained)(struct iowait *wait);
+	void (*init_priority)(struct iowait *wait);
 	seqlock_t *lock;
 	wait_queue_head_t wait_dma;
 	wait_queue_head_t wait_pio;
@@ -152,6 +157,7 @@ struct iowait {
 	u32 tx_limit;
 	u32 tx_count;
 	u8 starved_cnt;
+	u8 priority;
 	unsigned long flags;
 	struct iowait_work wait[IOWAIT_SES];
 };
@@ -171,7 +177,8 @@ void iowait_init(struct iowait *wait, u32 tx_limit,
 			      uint seq,
 			      bool pkts_sent),
 		 void (*wakeup)(struct iowait *wait, int reason),
-		 void (*sdma_drained)(struct iowait *wait));
+		 void (*sdma_drained)(struct iowait *wait),
+		 void (*init_priority)(struct iowait *wait));
 
 /**
  * iowait_schedule() - schedule the default send engine work
@@ -185,6 +192,18 @@ static inline bool iowait_schedule(struct iowait *wait,
 	return !!queue_work_on(cpu, wq, &wait->wait[IOWAIT_IB_SE].iowork);
 }
 
+/**
+ * iowait_tid_schedule - schedule the tid SE
+ * @wait: the iowait structure
+ * @wq: the work queue
+ * @cpu: the cpu
+ */
+static inline bool iowait_tid_schedule(struct iowait *wait,
+				       struct workqueue_struct *wq, int cpu)
+{
+	return !!queue_work_on(cpu, wq, &wait->wait[IOWAIT_TID_SE].iowork);
+}
+
 /**
  * iowait_sdma_drain() - wait for DMAs to drain
  *
@@ -327,6 +346,8 @@ static inline u16 iowait_get_desc(struct iowait_work *w)
 		tx = list_first_entry(&w->tx_head, struct sdma_txreq,
 				      list);
 		num_desc = tx->num_desc;
+		if (tx->flags & SDMA_TXREQ_F_VIP)
+			w->iow->priority++;
 	}
 	return num_desc;
 }
@@ -340,6 +361,37 @@ static inline u32 iowait_get_all_desc(struct iowait *w)
 	return num_desc;
 }
 
+static inline void iowait_update_priority(struct iowait_work *w)
+{
+	struct sdma_txreq *tx = NULL;
+
+	if (!list_empty(&w->tx_head)) {
+		tx = list_first_entry(&w->tx_head, struct sdma_txreq,
+				      list);
+		if (tx->flags & SDMA_TXREQ_F_VIP)
+			w->iow->priority++;
+	}
+}
+
+static inline void iowait_update_all_priority(struct iowait *w)
+{
+	iowait_update_priority(&w->wait[IOWAIT_IB_SE]);
+	iowait_update_priority(&w->wait[IOWAIT_TID_SE]);
+}
+
+static inline void iowait_init_priority(struct iowait *w)
+{
+	w->priority = 0;
+	if (w->init_priority)
+		w->init_priority(w);
+}
+
+static inline void iowait_get_priority(struct iowait *w)
+{
+	iowait_init_priority(w);
+	iowait_update_all_priority(w);
+}
+
 /**
  * iowait_queue - Put the iowait on a wait queue
  * @pkts_sent: have some packets been sent before queuing?
@@ -356,14 +408,18 @@ static inline void iowait_queue(bool pkts_sent, struct iowait *w,
 	/*
 	 * To play fair, insert the iowait at the tail of the wait queue if it
 	 * has already sent some packets; Otherwise, put it at the head.
+	 * However, if it has priority packets to send, also put it at the
+	 * head.
 	 */
-	if (pkts_sent) {
-		list_add_tail(&w->list, wait_head);
+	if (pkts_sent)
 		w->starved_cnt = 0;
-	} else {
-		list_add(&w->list, wait_head);
+	else
 		w->starved_cnt++;
-	}
+
+	if (w->priority > 0 || !pkts_sent)
+		list_add(&w->list, wait_head);
+	else
+		list_add_tail(&w->list, wait_head);
 }
 
 /**
@@ -380,27 +436,10 @@ static inline void iowait_starve_clear(bool pkts_sent, struct iowait *w)
 		w->starved_cnt = 0;
 }
 
-/**
- * iowait_starve_find_max - Find the maximum of the starve count
- * @w: the iowait struct
- * @max: a variable containing the max starve count
- * @idx: the index of the current iowait in an array
- * @max_idx: a variable containing the array index for the
- *         iowait entry that has the max starve count
- *
- * This function is called to compare the starve count of a
- * given iowait with the given max starve count. The max starve
- * count and the index will be updated if the iowait's start
- * count is larger.
- */
-static inline void iowait_starve_find_max(struct iowait *w, u8 *max,
-					  uint idx, uint *max_idx)
-{
-	if (w->starved_cnt > *max) {
-		*max = w->starved_cnt;
-		*max_idx = idx;
-	}
-}
+/* Update the top priority index */
+uint iowait_priority_update_top(struct iowait *w,
+				struct iowait *top,
+				uint idx, uint top_idx);
 
 /**
  * iowait_packet_queued() - determine if a packet is queued

drivers/infiniband/hw/hfi1/opfn.c

@@ -245,10 +245,15 @@ void opfn_qp_init(struct rvt_qp *qp, struct ib_qp_attr *attr, int attr_mask)
 	struct hfi1_qp_priv *priv = qp->priv;
 	unsigned long flags;
 
+	if (attr_mask & IB_QP_RETRY_CNT)
+		priv->s_retry = attr->retry_cnt;
+
 	spin_lock_irqsave(&priv->opfn.lock, flags);
 	if (ibqp->qp_type == IB_QPT_RC && HFI1_CAP_IS_KSET(TID_RDMA)) {
 		struct tid_rdma_params *local = &priv->tid_rdma.local;
 
+		if (attr_mask & IB_QP_TIMEOUT)
+			priv->tid_retry_timeout_jiffies = qp->timeout_jiffies;
 		if (qp->pmtu == enum_to_mtu(OPA_MTU_4096) ||
 		    qp->pmtu == enum_to_mtu(OPA_MTU_8192)) {
 			tid_rdma_opfn_init(qp, local);

drivers/infiniband/hw/hfi1/pio.c

@@ -1599,8 +1599,7 @@ static void sc_piobufavail(struct send_context *sc)
 	struct rvt_qp *qp;
 	struct hfi1_qp_priv *priv;
 	unsigned long flags;
-	uint i, n = 0, max_idx = 0;
-	u8 max_starved_cnt = 0;
+	uint i, n = 0, top_idx = 0;
 
 	if (dd->send_contexts[sc->sw_index].type != SC_KERNEL &&
 	    dd->send_contexts[sc->sw_index].type != SC_VL15)
@@ -1619,11 +1618,18 @@ static void sc_piobufavail(struct send_context *sc)
 		if (n == ARRAY_SIZE(qps))
 			break;
 		wait = list_first_entry(list, struct iowait, list);
+		iowait_get_priority(wait);
 		qp = iowait_to_qp(wait);
 		priv = qp->priv;
 		list_del_init(&priv->s_iowait.list);
 		priv->s_iowait.lock = NULL;
-		iowait_starve_find_max(wait, &max_starved_cnt, n, &max_idx);
+		if (n) {
+			priv = qps[top_idx]->priv;
+			top_idx = iowait_priority_update_top(wait,
+							     &priv->s_iowait,
+							     n, top_idx);
+		}
+
 		/* refcount held until actual wake up */
 		qps[n++] = qp;
 	}
@@ -1638,12 +1644,12 @@ static void sc_piobufavail(struct send_context *sc)
 	}
 	write_sequnlock_irqrestore(&sc->waitlock, flags);
 
-	/* Wake up the most starved one first */
+	/* Wake up the top-priority one first */
 	if (n)
-		hfi1_qp_wakeup(qps[max_idx],
+		hfi1_qp_wakeup(qps[top_idx],
 			       RVT_S_WAIT_PIO | HFI1_S_WAIT_PIO_DRAIN);
 	for (i = 0; i < n; i++)
-		if (i != max_idx)
+		if (i != top_idx)
 			hfi1_qp_wakeup(qps[i],
 				       RVT_S_WAIT_PIO | HFI1_S_WAIT_PIO_DRAIN);
 }

drivers/infiniband/hw/hfi1/qp.c

@@ -138,6 +138,12 @@ const struct rvt_operation_params hfi1_post_parms[RVT_OPERATION_MAX] = {
 	.flags = RVT_OPERATION_USE_RESERVE,
 },
 
+[IB_WR_TID_RDMA_WRITE] = {
+	.length = sizeof(struct ib_rdma_wr),
+	.qpt_support = BIT(IB_QPT_RC),
+	.flags = RVT_OPERATION_IGN_RNR_CNT,
+},
+
 };
 
 static void flush_list_head(struct list_head *l)
@@ -431,6 +437,11 @@ static void hfi1_qp_schedule(struct rvt_qp *qp)
 		if (ret)
 			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
 	}
+	if (iowait_flag_set(&priv->s_iowait, IOWAIT_PENDING_TID)) {
+		ret = hfi1_schedule_tid_send(qp);
+		if (ret)
+			iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
+	}
 }
 
 void hfi1_qp_wakeup(struct rvt_qp *qp, u32 flag)
@@ -450,8 +461,27 @@ void hfi1_qp_wakeup(struct rvt_qp *qp, u32 flag)
 
 void hfi1_qp_unbusy(struct rvt_qp *qp, struct iowait_work *wait)
 {
-	if (iowait_set_work_flag(wait) == IOWAIT_IB_SE)
+	struct hfi1_qp_priv *priv = qp->priv;
+
+	if (iowait_set_work_flag(wait) == IOWAIT_IB_SE) {
 		qp->s_flags &= ~RVT_S_BUSY;
+		/*
+		 * If we are sending a first-leg packet from the second leg,
+		 * we need to clear the busy flag from priv->s_flags to
+		 * avoid a race condition when the qp wakes up before
+		 * the call to hfi1_verbs_send() returns to the second
+		 * leg. In that case, the second leg will terminate without
+		 * being re-scheduled, resulting in failure to send TID RDMA
+		 * WRITE DATA and TID RDMA ACK packets.
+		 */
+		if (priv->s_flags & HFI1_S_TID_BUSY_SET) {
+			priv->s_flags &= ~(HFI1_S_TID_BUSY_SET |
+					   RVT_S_BUSY);
+			iowait_set_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
+		}
+	} else {
+		priv->s_flags &= ~RVT_S_BUSY;
+	}
 }
 
 static int iowait_sleep(
@@ -488,6 +518,7 @@ static int iowait_sleep(
 
 			ibp->rvp.n_dmawait++;
 			qp->s_flags |= RVT_S_WAIT_DMA_DESC;
+			iowait_get_priority(&priv->s_iowait);
 			iowait_queue(pkts_sent, &priv->s_iowait,
 				     &sde->dmawait);
 			priv->s_iowait.lock = &sde->waitlock;
@@ -537,6 +568,17 @@ static void iowait_sdma_drained(struct iowait *wait)
 	spin_unlock_irqrestore(&qp->s_lock, flags);
 }
 
+static void hfi1_init_priority(struct iowait *w)
+{
+	struct rvt_qp *qp = iowait_to_qp(w);
+	struct hfi1_qp_priv *priv = qp->priv;
+
+	if (qp->s_flags & RVT_S_ACK_PENDING)
+		w->priority++;
+	if (priv->s_flags & RVT_S_ACK_PENDING)
+		w->priority++;
+}
+
 /**
  * qp_to_sdma_engine - map a qp to a send engine
  * @qp: the QP
@@ -694,10 +736,11 @@ void *qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp)
 		&priv->s_iowait,
 		1,
 		_hfi1_do_send,
-		NULL,
+		_hfi1_do_tid_send,
 		iowait_sleep,
 		iowait_wakeup,
-		iowait_sdma_drained);
+		iowait_sdma_drained,
+		hfi1_init_priority);
 	return priv;
 }
 
@@ -755,6 +798,8 @@ void quiesce_qp(struct rvt_qp *qp)
 {
 	struct hfi1_qp_priv *priv = qp->priv;
 
+	hfi1_del_tid_reap_timer(qp);
+	hfi1_del_tid_retry_timer(qp);
 	iowait_sdma_drain(&priv->s_iowait);
 	qp_pio_drain(qp);
 	flush_tx_list(qp);
@@ -850,7 +895,8 @@ void notify_error_qp(struct rvt_qp *qp)
 	if (lock) {
 		write_seqlock(lock);
 		if (!list_empty(&priv->s_iowait.list) &&
-		    !(qp->s_flags & RVT_S_BUSY)) {
+		    !(qp->s_flags & RVT_S_BUSY) &&
+		    !(priv->s_flags & RVT_S_BUSY)) {
 			qp->s_flags &= ~RVT_S_ANY_WAIT_IO;
 			list_del_init(&priv->s_iowait.list);
 			priv->s_iowait.lock = NULL;
@@ -859,7 +905,8 @@ void notify_error_qp(struct rvt_qp *qp)
 		write_sequnlock(lock);
 	}
 
-	if (!(qp->s_flags & RVT_S_BUSY)) {
+	if (!(qp->s_flags & RVT_S_BUSY) && !(priv->s_flags & RVT_S_BUSY)) {
+		qp->s_hdrwords = 0;
 		if (qp->s_rdma_mr) {
 			rvt_put_mr(qp->s_rdma_mr);
 			qp->s_rdma_mr = NULL;

drivers/infiniband/hw/hfi1/qp.h

@@ -64,12 +64,16 @@ extern const struct rvt_operation_params hfi1_post_parms[];
  * 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_WAIT_TID_RESP - waiting for a TID RDMA WRITE response
+ * HFI1_S_WAIT_HALT - halt the first leg send engine
  * 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_WAIT_TID_RESP     0x08000000
+#define HFI1_S_WAIT_HALT         0x04000000
 #define HFI1_S_MIN_BIT_MASK      0x01000000
 
 /*
@@ -78,6 +82,7 @@ extern const struct rvt_operation_params hfi1_post_parms[];
 
 #define HFI1_S_ANY_WAIT_IO (RVT_S_ANY_WAIT_IO | HFI1_S_WAIT_PIO_DRAIN)
 #define HFI1_S_ANY_WAIT (HFI1_S_ANY_WAIT_IO | RVT_S_ANY_WAIT_SEND)
+#define HFI1_S_ANY_TID_WAIT_SEND (RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_DMA)
 
 /*
  * Send if not busy or waiting for I/O and either

drivers/infiniband/hw/hfi1/rc.h

@@ -18,6 +18,7 @@ static inline void update_ack_queue(struct rvt_qp *qp, unsigned int n)
 	if (next > rvt_size_atomic(ib_to_rvt(qp->ibqp.device)))
 		next = 0;
 	qp->s_tail_ack_queue = next;
+	qp->s_acked_ack_queue = next;
 	qp->s_ack_state = OP(ACKNOWLEDGE);
 }
 

drivers/infiniband/hw/hfi1/ruc.c

@@ -453,11 +453,13 @@ void hfi1_make_ruc_header(struct rvt_qp *qp, struct ib_other_headers *ohdr,
 #define SEND_RESCHED_TIMEOUT (5 * HZ)  /* 5s in jiffies */
 
 /**
- * schedule_send_yield - test for a yield required for QP send engine
+ * hfi1_schedule_send_yield - test for a yield required for QP
+ * send engine
  * @timeout: Final time for timeout slice for jiffies
  * @qp: a pointer to QP
  * @ps: a pointer to a structure with commonly lookup values for
  *      the the send engine progress
+ * @tid - true if it is the tid leg
  *
  * This routine checks if the time slice for the QP has expired
  * for RC QPs, if so an additional work entry is queued. At this
@@ -465,8 +467,8 @@ void hfi1_make_ruc_header(struct rvt_qp *qp, struct ib_other_headers *ohdr,
  * returns true if a yield is required, otherwise, false
  * is returned.
  */
-static bool schedule_send_yield(struct rvt_qp *qp,
-				struct hfi1_pkt_state *ps)
+bool hfi1_schedule_send_yield(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
+			      bool tid)
 {
 	ps->pkts_sent = true;
 
@@ -474,8 +476,24 @@ static bool schedule_send_yield(struct rvt_qp *qp,
 		if (!ps->in_thread ||
 		    workqueue_congested(ps->cpu, ps->ppd->hfi1_wq)) {
 			spin_lock_irqsave(&qp->s_lock, ps->flags);
-			qp->s_flags &= ~RVT_S_BUSY;
-			hfi1_schedule_send(qp);
+			if (!tid) {
+				qp->s_flags &= ~RVT_S_BUSY;
+				hfi1_schedule_send(qp);
+			} else {
+				struct hfi1_qp_priv *priv = qp->priv;
+
+				if (priv->s_flags &
+				    HFI1_S_TID_BUSY_SET) {
+					qp->s_flags &= ~RVT_S_BUSY;
+					priv->s_flags &=
+						~(HFI1_S_TID_BUSY_SET |
+						  RVT_S_BUSY);
+				} else {
+					priv->s_flags &= ~RVT_S_BUSY;
+				}
+				hfi1_schedule_tid_send(qp);
+			}
+
 			spin_unlock_irqrestore(&qp->s_lock, ps->flags);
 			this_cpu_inc(*ps->ppd->dd->send_schedule);
 			trace_hfi1_rc_expired_time_slice(qp, true);
@@ -576,6 +594,8 @@ void hfi1_do_send(struct rvt_qp *qp, bool in_thread)
 	do {
 		/* Check for a constructed packet to be sent. */
 		if (ps.s_txreq) {
+			if (priv->s_flags & HFI1_S_TID_BUSY_SET)
+				qp->s_flags |= RVT_S_BUSY;
 			spin_unlock_irqrestore(&qp->s_lock, ps.flags);
 			/*
 			 * If the packet cannot be sent now, return and
@@ -585,7 +605,7 @@ void hfi1_do_send(struct rvt_qp *qp, bool in_thread)
 				return;
 
 			/* allow other tasks to run */
-			if (schedule_send_yield(qp, &ps))
+			if (hfi1_schedule_send_yield(qp, &ps, false))
 				return;
 
 			spin_lock_irqsave(&qp->s_lock, ps.flags);

drivers/infiniband/hw/hfi1/sdma.c

@@ -1747,10 +1747,9 @@ static inline u16 sdma_gethead(struct sdma_engine *sde)
  */
 static void sdma_desc_avail(struct sdma_engine *sde, uint avail)
 {
-	struct iowait *wait, *nw;
+	struct iowait *wait, *nw, *twait;
 	struct iowait *waits[SDMA_WAIT_BATCH_SIZE];
-	uint i, n = 0, seq, max_idx = 0;
-	u8 max_starved_cnt = 0;
+	uint i, n = 0, seq, tidx = 0;
 
 #ifdef CONFIG_SDMA_VERBOSITY
 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", sde->this_idx,
@@ -1775,13 +1774,20 @@ static void sdma_desc_avail(struct sdma_engine *sde, uint avail)
 					continue;
 				if (n == ARRAY_SIZE(waits))
 					break;
+				iowait_init_priority(wait);
 				num_desc = iowait_get_all_desc(wait);
 				if (num_desc > avail)
 					break;
 				avail -= num_desc;
-				/* Find the most starved wait memeber */
-				iowait_starve_find_max(wait, &max_starved_cnt,
-						       n, &max_idx);
+				/* Find the top-priority wait memeber */
+				if (n) {
+					twait = waits[tidx];
+					tidx =
+					    iowait_priority_update_top(wait,
+								       twait,
+								       n,
+								       tidx);
+				}
 				list_del_init(&wait->list);
 				waits[n++] = wait;
 			}
@@ -1790,12 +1796,12 @@ static void sdma_desc_avail(struct sdma_engine *sde, uint avail)
 		}
 	} while (read_seqretry(&sde->waitlock, seq));
 
-	/* Schedule the most starved one first */
+	/* Schedule the top-priority entry first */
 	if (n)
-		waits[max_idx]->wakeup(waits[max_idx], SDMA_AVAIL_REASON);
+		waits[tidx]->wakeup(waits[tidx], SDMA_AVAIL_REASON);
 
 	for (i = 0; i < n; i++)
-		if (i != max_idx)
+		if (i != tidx)
 			waits[i]->wakeup(waits[i], SDMA_AVAIL_REASON);
 }
 

drivers/infiniband/hw/hfi1/sdma_txreq.h

@@ -91,6 +91,7 @@ struct sdma_desc {
 #define SDMA_TXREQ_F_URGENT       0x0001
 #define SDMA_TXREQ_F_AHG_COPY     0x0002
 #define SDMA_TXREQ_F_USE_AHG      0x0004
+#define SDMA_TXREQ_F_VIP          0x0010
 
 struct sdma_txreq;
 typedef void (*callback_t)(struct sdma_txreq *, int);

drivers/infiniband/hw/hfi1/tid_rdma.h

@@ -25,8 +25,34 @@
  * s_flags, there are no collisions.
  *
  * HFI1_S_TID_WAIT_INTERLCK - QP is waiting for requester interlock
+ * HFI1_R_TID_WAIT_INTERLCK - QP is waiting for responder interlock
  */
+#define HFI1_S_TID_BUSY_SET       BIT(0)
+/* BIT(1) reserved for RVT_S_BUSY. */
+#define HFI1_R_TID_RSC_TIMER      BIT(2)
+/* BIT(3) reserved for RVT_S_RESP_PENDING. */
+/* BIT(4) reserved for RVT_S_ACK_PENDING. */
 #define HFI1_S_TID_WAIT_INTERLCK  BIT(5)
+#define HFI1_R_TID_WAIT_INTERLCK  BIT(6)
+/* BIT(7) - BIT(15) reserved for RVT_S_WAIT_*. */
+/* BIT(16) reserved for RVT_S_SEND_ONE */
+#define HFI1_S_TID_RETRY_TIMER    BIT(17)
+/* BIT(18) reserved for RVT_S_ECN. */
+#define HFI1_R_TID_SW_PSN         BIT(19)
+/* BIT(26) reserved for HFI1_S_WAIT_HALT */
+/* BIT(27) reserved for HFI1_S_WAIT_TID_RESP */
+/* BIT(28) reserved for HFI1_S_WAIT_TID_SPACE */
+
+/*
+ * Unlike regular IB RDMA VERBS, which do not require an entry
+ * in the s_ack_queue, TID RDMA WRITE requests do because they
+ * generate responses.
+ * Therefore, the s_ack_queue needs to be extended by a certain
+ * amount. The key point is that the queue needs to be extended
+ * without letting the "user" know so they user doesn't end up
+ * using these extra entries.
+ */
+#define HFI1_TID_RDMA_WRITE_CNT 8
 
 struct tid_rdma_params {
 	struct rcu_head rcu_head;
@@ -78,20 +104,25 @@ struct tid_rdma_request {
 	} e;
 
 	struct tid_rdma_flow *flows;	/* array of tid flows */
+	struct rvt_sge_state ss; /* SGE state for TID RDMA requests */
 	u16 n_flows;		/* size of the flow buffer window */
 	u16 setup_head;		/* flow index we are setting up */
 	u16 clear_tail;		/* flow index we are clearing */
 	u16 flow_idx;		/* flow index most recently set up */
+	u16 acked_tail;
 
 	u32 seg_len;
 	u32 total_len;
+	u32 r_ack_psn;          /* next expected ack PSN */
 	u32 r_flow_psn;         /* IB PSN of next segment start */
+	u32 r_last_acked;       /* IB PSN of last ACK'ed packet */
 	u32 s_next_psn;		/* IB PSN of next segment start for read */
 
 	u32 total_segs;		/* segments required to complete a request */
 	u32 cur_seg;		/* index of current segment */
 	u32 comp_seg;           /* index of last completed segment */
 	u32 ack_seg;            /* index of last ack'ed segment */
+	u32 alloc_seg;          /* index of next segment to be allocated */
 	u32 isge;		/* index of "current" sge */
 	u32 ack_pending;        /* num acks pending for this request */
 
@@ -158,11 +189,18 @@ struct tid_rdma_flow {
 	u8 npagesets;
 	u8 npkts;
 	u8 pkt;
+	u8 resync_npkts;
 	struct kern_tid_node tnode[TID_RDMA_MAX_PAGES];
 	struct tid_rdma_pageset pagesets[TID_RDMA_MAX_PAGES];
 	u32 tid_entry[TID_RDMA_MAX_PAGES];
 };
 
+enum tid_rnr_nak_state {
+	TID_RNR_NAK_INIT = 0,
+	TID_RNR_NAK_SEND,
+	TID_RNR_NAK_SENT,
+};
+
 bool tid_rdma_conn_req(struct rvt_qp *qp, u64 *data);
 bool tid_rdma_conn_reply(struct rvt_qp *qp, u64 data);
 bool tid_rdma_conn_resp(struct rvt_qp *qp, u64 *data);
@@ -228,9 +266,57 @@ static inline void hfi1_setup_tid_rdma_wqe(struct rvt_qp *qp,
 					   struct rvt_swqe *wqe)
 {
 	if (wqe->priv &&
-	    wqe->wr.opcode == IB_WR_RDMA_READ &&
+	    (wqe->wr.opcode == IB_WR_RDMA_READ ||
+	     wqe->wr.opcode == IB_WR_RDMA_WRITE) &&
 	    wqe->length >= TID_RDMA_MIN_SEGMENT_SIZE)
 		setup_tid_rdma_wqe(qp, wqe);
 }
 
+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,
+				   struct ib_other_headers *ohdr, u32 *bth1,
+				   u32 bth2, u32 *len,
+				   struct rvt_sge_state **ss);
+
+void hfi1_del_tid_reap_timer(struct rvt_qp *qp);
+
+void hfi1_rc_rcv_tid_rdma_write_resp(struct hfi1_packet *packet);
+
+bool hfi1_build_tid_rdma_packet(struct rvt_swqe *wqe,
+				struct ib_other_headers *ohdr,
+				u32 *bth1, u32 *bth2, u32 *len);
+
+void hfi1_rc_rcv_tid_rdma_write_data(struct hfi1_packet *packet);
+
+u32 hfi1_build_tid_rdma_write_ack(struct rvt_qp *qp, struct rvt_ack_entry *e,
+				  struct ib_other_headers *ohdr, u16 iflow,
+				  u32 *bth1, u32 *bth2);
+
+void hfi1_rc_rcv_tid_rdma_ack(struct hfi1_packet *packet);
+
+void hfi1_add_tid_retry_timer(struct rvt_qp *qp);
+void hfi1_del_tid_retry_timer(struct rvt_qp *qp);
+
+u32 hfi1_build_tid_rdma_resync(struct rvt_qp *qp, struct rvt_swqe *wqe,
+			       struct ib_other_headers *ohdr, u32 *bth1,
+			       u32 *bth2, u16 fidx);
+
+void hfi1_rc_rcv_tid_rdma_resync(struct hfi1_packet *packet);
+
+struct hfi1_pkt_state;
+int hfi1_make_tid_rdma_pkt(struct rvt_qp *qp, struct hfi1_pkt_state *ps);
+
+void _hfi1_do_tid_send(struct work_struct *work);
+
+bool hfi1_schedule_tid_send(struct rvt_qp *qp);
+
+bool hfi1_tid_rdma_ack_interlock(struct rvt_qp *qp, struct rvt_ack_entry *e);
+
 #endif /* HFI1_TID_RDMA_H */

drivers/infiniband/hw/hfi1/trace.c

@@ -133,6 +133,11 @@ const char *hfi1_trace_get_packet_l2_str(u8 l2)
 #define TID_RDMA_KDETH_DATA "kdeth0 0x%x: kver %u sh %u intr %u tidctrl %u tid %x offset %x kdeth1 0x%x: jkey %x"
 #define TID_READ_REQ_PRN "tid_flow_psn 0x%x tid_flow_qp 0x%x verbs_qp 0x%x"
 #define TID_READ_RSP_PRN "verbs_qp 0x%x"
+#define TID_WRITE_REQ_PRN "original_qp 0x%x"
+#define TID_WRITE_RSP_PRN "tid_flow_psn 0x%x tid_flow_qp 0x%x verbs_qp 0x%x"
+#define TID_WRITE_DATA_PRN "verbs_qp 0x%x"
+#define TID_ACK_PRN "tid_flow_psn 0x%x verbs_psn 0x%x tid_flow_qp 0x%x verbs_qp 0x%x"
+#define TID_RESYNC_PRN "verbs_qp 0x%x"
 
 #define OP(transport, op) IB_OPCODE_## transport ## _ ## op
 
@@ -327,6 +332,45 @@ const char *parse_everbs_hdrs(
 				 parse_syndrome(be32_to_cpu(eh->aeth) >> 24),
 				 be32_to_cpu(eh->aeth) & IB_MSN_MASK);
 		break;
+	case OP(TID_RDMA, WRITE_REQ):
+		trace_seq_printf(p, TID_RDMA_KDETH " " RETH_PRN " "
+				 TID_WRITE_REQ_PRN,
+				 le32_to_cpu(eh->tid_rdma.w_req.kdeth0),
+				 le32_to_cpu(eh->tid_rdma.w_req.kdeth1),
+				 ib_u64_get(&eh->tid_rdma.w_req.reth.vaddr),
+				 be32_to_cpu(eh->tid_rdma.w_req.reth.rkey),
+				 be32_to_cpu(eh->tid_rdma.w_req.reth.length),
+				 be32_to_cpu(eh->tid_rdma.w_req.verbs_qp));
+		break;
+	case OP(TID_RDMA, WRITE_RESP):
+		trace_seq_printf(p, TID_RDMA_KDETH " " AETH_PRN " "
+				 TID_WRITE_RSP_PRN,
+				 le32_to_cpu(eh->tid_rdma.w_rsp.kdeth0),
+				 le32_to_cpu(eh->tid_rdma.w_rsp.kdeth1),
+				 be32_to_cpu(eh->tid_rdma.w_rsp.aeth) >> 24,
+				 parse_syndrome(/* aeth */
+					 be32_to_cpu(eh->tid_rdma.w_rsp.aeth)
+					 >> 24),
+				 (be32_to_cpu(eh->tid_rdma.w_rsp.aeth) &
+				  IB_MSN_MASK),
+				 be32_to_cpu(eh->tid_rdma.w_rsp.tid_flow_psn),
+				 be32_to_cpu(eh->tid_rdma.w_rsp.tid_flow_qp),
+				 be32_to_cpu(eh->tid_rdma.w_rsp.verbs_qp));
+		break;
+	case OP(TID_RDMA, WRITE_DATA_LAST):
+	case OP(TID_RDMA, WRITE_DATA):
+		trace_seq_printf(p, TID_RDMA_KDETH_DATA " " TID_WRITE_DATA_PRN,
+				 le32_to_cpu(eh->tid_rdma.w_data.kdeth0),
+				 KDETH_GET(eh->tid_rdma.w_data.kdeth0, KVER),
+				 KDETH_GET(eh->tid_rdma.w_data.kdeth0, SH),
+				 KDETH_GET(eh->tid_rdma.w_data.kdeth0, INTR),
+				 KDETH_GET(eh->tid_rdma.w_data.kdeth0, TIDCTRL),
+				 KDETH_GET(eh->tid_rdma.w_data.kdeth0, TID),
+				 KDETH_GET(eh->tid_rdma.w_data.kdeth0, OFFSET),
+				 le32_to_cpu(eh->tid_rdma.w_data.kdeth1),
+				 KDETH_GET(eh->tid_rdma.w_data.kdeth1, JKEY),
+				 be32_to_cpu(eh->tid_rdma.w_data.verbs_qp));
+		break;
 	case OP(TID_RDMA, READ_REQ):
 		trace_seq_printf(p, TID_RDMA_KDETH " " RETH_PRN " "
 				 TID_READ_REQ_PRN,
@@ -359,6 +403,28 @@ const char *parse_everbs_hdrs(
 				  IB_MSN_MASK),
 				 be32_to_cpu(eh->tid_rdma.r_rsp.verbs_qp));
 		break;
+	case OP(TID_RDMA, ACK):
+		trace_seq_printf(p, TID_RDMA_KDETH " " AETH_PRN " "
+				 TID_ACK_PRN,
+				 le32_to_cpu(eh->tid_rdma.ack.kdeth0),
+				 le32_to_cpu(eh->tid_rdma.ack.kdeth1),
+				 be32_to_cpu(eh->tid_rdma.ack.aeth) >> 24,
+				 parse_syndrome(/* aeth */
+					 be32_to_cpu(eh->tid_rdma.ack.aeth)
+					 >> 24),
+				 (be32_to_cpu(eh->tid_rdma.ack.aeth) &
+				  IB_MSN_MASK),
+				 be32_to_cpu(eh->tid_rdma.ack.tid_flow_psn),
+				 be32_to_cpu(eh->tid_rdma.ack.verbs_psn),
+				 be32_to_cpu(eh->tid_rdma.ack.tid_flow_qp),
+				 be32_to_cpu(eh->tid_rdma.ack.verbs_qp));
+		break;
+	case OP(TID_RDMA, RESYNC):
+		trace_seq_printf(p, TID_RDMA_KDETH " " TID_RESYNC_PRN,
+				 le32_to_cpu(eh->tid_rdma.resync.kdeth0),
+				 le32_to_cpu(eh->tid_rdma.resync.kdeth1),
+				 be32_to_cpu(eh->tid_rdma.resync.verbs_qp));
+		break;
 	/* aeth + atomicacketh */
 	case OP(RC, ATOMIC_ACKNOWLEDGE):
 		trace_seq_printf(p, AETH_PRN " " ATOMICACKETH_PRN,

drivers/infiniband/hw/hfi1/trace_ibhdrs.h

@@ -79,8 +79,14 @@ __print_symbolic(opcode,                                   \
 	ib_opcode_name(RC_ATOMIC_ACKNOWLEDGE),             \
 	ib_opcode_name(RC_COMPARE_SWAP),                   \
 	ib_opcode_name(RC_FETCH_ADD),                      \
+	ib_opcode_name(TID_RDMA_WRITE_REQ),	           \
+	ib_opcode_name(TID_RDMA_WRITE_RESP),	           \
+	ib_opcode_name(TID_RDMA_WRITE_DATA),	           \
+	ib_opcode_name(TID_RDMA_WRITE_DATA_LAST),          \
 	ib_opcode_name(TID_RDMA_READ_REQ),	           \
 	ib_opcode_name(TID_RDMA_READ_RESP),	           \
+	ib_opcode_name(TID_RDMA_RESYNC),	           \
+	ib_opcode_name(TID_RDMA_ACK),                      \
 	ib_opcode_name(UC_SEND_FIRST),                     \
 	ib_opcode_name(UC_SEND_MIDDLE),                    \
 	ib_opcode_name(UC_SEND_LAST),                      \

drivers/infiniband/hw/hfi1/trace_tx.h

@@ -846,6 +846,12 @@ DEFINE_EVENT(
 	TP_ARGS(qp, flag)
 );
 
+DEFINE_EVENT(/* event */
+	hfi1_do_send_template, hfi1_rc_do_tid_send,
+	TP_PROTO(struct rvt_qp *qp, bool flag),
+	TP_ARGS(qp, flag)
+);
+
 DEFINE_EVENT(
 	hfi1_do_send_template, hfi1_rc_expired_time_slice,
 	TP_PROTO(struct rvt_qp *qp, bool flag),