Merge branch 'hfi1-tid' into wip/dl-for-next

Omni-Path TID RDMA Feature

Intel Omni-Path (OPA) TID RDMA support is a feature that accelerates
data movement between two OPA nodes through the IB Verbs interface. It
improves RDMA READ/WRITE performance by delivering the data payload to a
user buffer directly without any software copying.

Architecture
=============
The TID RDMA protocol is implemented on the hfi1 driver level and is
therefore transparent to the ULPs. It is designed to facilitate the data
transactions for two specific RDMA requests:
  - RDMA READ;
  - RDMA WRITE.
Previously, when a verbs data packet is received at the destination
(requester side for RDMA READ and responder side for RDMA WRITE), the
data payload is copied to the user buffer by software, which slows down
the performance significantly for large requests.

Internally, hfi1 converts qualified RDMA READ/WRITE requests into TID
RDMA READ/WRITE requests when the requests are post sent to the hfi1
driver. Non-qualified RDMA requests are handled by normal RDMA protocol.

For TID RDMA requests, hardware resources (hardware flow and TID entries)
are allocated on the destination side (the requester side for TID RDMA
READ and the responder side for TID RDMA WRITE). The information for
these resources is conveyed to the data source side (the responder side
for TID RDMA READ and the requester side for TID RDMA WRITE) and embedded
in data packets. When data packets are received by the destination,
hardware will deliver the data payload to the destination buffer without
involving software and therefore improve the performance.

Details
=======
RDMA READ/WRITE requests are qualified by the following:
  - Total data length >= 256k;
  - Totoal data length is a multiple of 4K pages.

Additional qualifications are enforced for the destination buffers:
  For RDMA RAED:
    - Each destination sge buffer is 4K aligned;
    - Each destination sge buffer is a multiple of 4K pages.
  For RDMA WRITE:
    - The destination number is 4K aligned.

In addition, in an OPA fabric, some nodes may support TID RDMA while
others may not. As such, it is important for two transaction nodes to
exchange the information about the features they support. This discovery
mechanism is called OPA Feature Negotion (OPFN) and is described in
details in the patch series. Through OPFN, two nodes can find whether
they both support TID RDMA and subsequently convert RDMA requests into
TID RDMA requests.

* hfi1-tid: (46 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/Makefile

@@ -24,6 +24,7 @@ hfi1-y := \
 	mad.o \
 	mmu_rb.o \
 	msix.o \
+	opfn.o \
 	pcie.o \
 	pio.o \
 	pio_copy.o \

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",
@@ -5222,6 +5224,17 @@ int is_bx(struct hfi1_devdata *dd)
 	return (chip_rev_minor & 0xF0) == 0x10;
 }
 
+/* return true is kernel urg disabled for rcd */
+bool is_urg_masked(struct hfi1_ctxtdata *rcd)
+{
+	u64 mask;
+	u32 is = IS_RCVURGENT_START + rcd->ctxt;
+	u8 bit = is % 64;
+
+	mask = read_csr(rcd->dd, CCE_INT_MASK + (8 * (is / 64)));
+	return !(mask & BIT_ULL(bit));
+}
+
 /*
  * Append string s to buffer buf.  Arguments curp and len are the current
  * position and remaining length, respectively.

drivers/infiniband/hw/hfi1/chip.h

 #ifndef _CHIP_H
 #define _CHIP_H
 /*
- * Copyright(c) 2015 - 2017 Intel Corporation.
+ * Copyright(c) 2015 - 2018 Intel Corporation.
  *
  * This file is provided under a dual BSD/GPLv2 license.  When using or
  * redistributing this file, you may do so under either license.
@@ -804,6 +804,7 @@ void clear_linkup_counters(struct hfi1_devdata *dd);
 u32 hdrqempty(struct hfi1_ctxtdata *rcd);
 int is_ax(struct hfi1_devdata *dd);
 int is_bx(struct hfi1_devdata *dd);
+bool is_urg_masked(struct hfi1_ctxtdata *rcd);
 u32 read_physical_state(struct hfi1_devdata *dd);
 u32 chip_to_opa_pstate(struct hfi1_devdata *dd, u32 chip_pstate);
 const char *opa_lstate_name(u32 lstate);
@@ -926,6 +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,

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));

drivers/infiniband/hw/hfi1/driver.c

@@ -1575,25 +1575,32 @@ static int hfi1_setup_bypass_packet(struct hfi1_packet *packet)
 	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;
 }
 

drivers/infiniband/hw/hfi1/hfi.h

@@ -73,6 +73,7 @@
 
 #include "chip_registers.h"
 #include "common.h"
+#include "opfn.h"
 #include "verbs.h"
 #include "pio.h"
 #include "chip.h"
@@ -98,6 +99,8 @@
 #define NEIGHBOR_TYPE_HFI		0
 #define NEIGHBOR_TYPE_SWITCH	1
 
+#define HFI1_MAX_ACTIVE_WORKQUEUE_ENTRIES 5
+
 extern unsigned long hfi1_cap_mask;
 #define HFI1_CAP_KGET_MASK(mask, cap) ((mask) & HFI1_CAP_##cap)
 #define HFI1_CAP_UGET_MASK(mask, cap) \
@@ -195,6 +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;
@@ -288,6 +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 */
@@ -320,6 +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];
 };
 
 /**
@@ -2100,7 +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 */

drivers/infiniband/hw/hfi1/init.c

@@ -72,7 +72,6 @@
 #undef pr_fmt
 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
 
-#define HFI1_MAX_ACTIVE_WORKQUEUE_ENTRIES 5
 /*
  * min buffers we want to have per context, after driver
  */
@@ -371,6 +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);
 
@@ -473,6 +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;
@@ -772,6 +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);
@@ -927,6 +934,8 @@ int hfi1_init(struct hfi1_devdata *dd, int reinit)
 		lastfail = hfi1_create_rcvhdrq(dd, rcd);
 		if (!lastfail)
 			lastfail = hfi1_setup_eagerbufs(rcd);
+		if (!lastfail)
+			lastfail = hfi1_kern_exp_rcv_init(rcd, reinit);
 		if (lastfail) {
 			dd_dev_err(dd,
 				   "failed to allocate kernel ctxt's rcvhdrq and/or egr bufs\n");
@@ -1497,6 +1506,13 @@ static int __init hfi1_mod_init(void)
 	/* sanitize link CRC options */
 	link_crc_mask &= SUPPORTED_CRCS;
 
+	ret = opfn_init();
+	if (ret < 0) {
+		pr_err("Failed to allocate opfn_wq");
+		goto bail_dev;
+	}
+
+	hfi1_compute_tid_rdma_flow_wt();
 	/*
 	 * These must be called before the driver is registered with
 	 * the PCI subsystem.
@@ -1527,6 +1543,7 @@ module_init(hfi1_mod_init);
 static void __exit hfi1_mod_cleanup(void)
 {
 	pci_unregister_driver(&hfi1_pci_driver);
+	opfn_exit();
 	node_affinity_destroy_all();
 	hfi1_dbg_exit();
 
@@ -1581,7 +1598,7 @@ 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);
 		}
 	}

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

+// SPDX-License-Identifier: (GPL-2.0 OR BSD-3-Clause)
+/*
+ * Copyright(c) 2018 Intel Corporation.
+ *
+ */
+#include "hfi.h"
+#include "trace.h"
+#include "qp.h"
+#include "opfn.h"
+
+#define IB_BTHE_E                 BIT(IB_BTHE_E_SHIFT)
+
+#define OPFN_CODE(code) BIT((code) - 1)
+#define OPFN_MASK(code) OPFN_CODE(STL_VERBS_EXTD_##code)
+
+struct hfi1_opfn_type {
+	bool (*request)(struct rvt_qp *qp, u64 *data);
+	bool (*response)(struct rvt_qp *qp, u64 *data);
+	bool (*reply)(struct rvt_qp *qp, u64 data);
+	void (*error)(struct rvt_qp *qp);
+};
+
+static struct hfi1_opfn_type hfi1_opfn_handlers[STL_VERBS_EXTD_MAX] = {
+	[STL_VERBS_EXTD_TID_RDMA] = {
+		.request = tid_rdma_conn_req,
+		.response = tid_rdma_conn_resp,
+		.reply = tid_rdma_conn_reply,
+		.error = tid_rdma_conn_error,
+	},
+};
+
+static struct workqueue_struct *opfn_wq;
+
+static void opfn_schedule_conn_request(struct rvt_qp *qp);
+
+static bool hfi1_opfn_extended(u32 bth1)
+{
+	return !!(bth1 & IB_BTHE_E);
+}
+
+static void opfn_conn_request(struct rvt_qp *qp)
+{
+	struct hfi1_qp_priv *priv = qp->priv;
+	struct ib_atomic_wr wr;
+	u16 mask, capcode;
+	struct hfi1_opfn_type *extd;
+	u64 data;
+	unsigned long flags;
+	int ret = 0;
+
+	trace_hfi1_opfn_state_conn_request(qp);
+	spin_lock_irqsave(&priv->opfn.lock, flags);
+	/*
+	 * Exit if the extended bit is not set, or if nothing is requested, or
+	 * if we have completed all requests, or if a previous request is in
+	 * progress
+	 */
+	if (!priv->opfn.extended || !priv->opfn.requested ||
+	    priv->opfn.requested == priv->opfn.completed || priv->opfn.curr)
+		goto done;
+
+	mask = priv->opfn.requested & ~priv->opfn.completed;
+	capcode = ilog2(mask & ~(mask - 1)) + 1;
+	if (capcode >= STL_VERBS_EXTD_MAX) {
+		priv->opfn.completed |= OPFN_CODE(capcode);
+		goto done;
+	}
+
+	extd = &hfi1_opfn_handlers[capcode];
+	if (!extd || !extd->request || !extd->request(qp, &data)) {
+		/*
+		 * Either there is no handler for this capability or the request
+		 * packet could not be generated. Either way, mark it as done so
+		 * we don't keep attempting to complete it.
+		 */
+		priv->opfn.completed |= OPFN_CODE(capcode);
+		goto done;
+	}
+
+	trace_hfi1_opfn_data_conn_request(qp, capcode, data);
+	data = (data & ~0xf) | capcode;
+
+	memset(&wr, 0, sizeof(wr));
+	wr.wr.opcode = IB_WR_OPFN;
+	wr.remote_addr = HFI1_VERBS_E_ATOMIC_VADDR;
+	wr.compare_add = data;
+
+	priv->opfn.curr = capcode;	/* A new request is now in progress */
+	/* Drop opfn.lock before calling ib_post_send() */
+	spin_unlock_irqrestore(&priv->opfn.lock, flags);
+
+	ret = ib_post_send(&qp->ibqp, &wr.wr, NULL);
+	if (ret)
+		goto err;
+	trace_hfi1_opfn_state_conn_request(qp);
+	return;
+err:
+	trace_hfi1_msg_opfn_conn_request(qp, "ib_ost_send failed: ret = ",
+					 (u64)ret);
+	spin_lock_irqsave(&priv->opfn.lock, flags);
+	/*
+	 * In case of an unexpected error return from ib_post_send
+	 * clear opfn.curr and reschedule to try again
+	 */
+	priv->opfn.curr = STL_VERBS_EXTD_NONE;
+	opfn_schedule_conn_request(qp);
+done:
+	spin_unlock_irqrestore(&priv->opfn.lock, flags);
+}
+
+void opfn_send_conn_request(struct work_struct *work)
+{
+	struct hfi1_opfn_data *od;
+	struct hfi1_qp_priv *qpriv;
+
+	od = container_of(work, struct hfi1_opfn_data, opfn_work);
+	qpriv = container_of(od, struct hfi1_qp_priv, opfn);
+
+	opfn_conn_request(qpriv->owner);
+}
+
+/*
+ * When QP s_lock is held in the caller, the OPFN request must be scheduled
+ * to a different workqueue to avoid double locking QP s_lock in call to
+ * ib_post_send in opfn_conn_request
+ */
+static void opfn_schedule_conn_request(struct rvt_qp *qp)
+{
+	struct hfi1_qp_priv *priv = qp->priv;
+
+	trace_hfi1_opfn_state_sched_conn_request(qp);
+	queue_work(opfn_wq, &priv->opfn.opfn_work);
+}
+
+void opfn_conn_response(struct rvt_qp *qp, struct rvt_ack_entry *e,
+			struct ib_atomic_eth *ateth)
+{
+	struct hfi1_qp_priv *priv = qp->priv;
+	u64 data = be64_to_cpu(ateth->compare_data);
+	struct hfi1_opfn_type *extd;
+	u8 capcode;
+	unsigned long flags;
+
+	trace_hfi1_opfn_state_conn_response(qp);
+	capcode = data & 0xf;
+	trace_hfi1_opfn_data_conn_response(qp, capcode, data);
+	if (!capcode || capcode >= STL_VERBS_EXTD_MAX)
+		return;
+
+	extd = &hfi1_opfn_handlers[capcode];
+
+	if (!extd || !extd->response) {
+		e->atomic_data = capcode;
+		return;
+	}
+
+	spin_lock_irqsave(&priv->opfn.lock, flags);
+	if (priv->opfn.completed & OPFN_CODE(capcode)) {
+		/*
+		 * We are receiving a request for a feature that has already
+		 * been negotiated. This may mean that the other side has reset
+		 */
+		priv->opfn.completed &= ~OPFN_CODE(capcode);
+		if (extd->error)
+			extd->error(qp);
+	}
+
+	if (extd->response(qp, &data))
+		priv->opfn.completed |= OPFN_CODE(capcode);
+	e->atomic_data = (data & ~0xf) | capcode;
+	trace_hfi1_opfn_state_conn_response(qp);
+	spin_unlock_irqrestore(&priv->opfn.lock, flags);
+}
+
+void opfn_conn_reply(struct rvt_qp *qp, u64 data)
+{
+	struct hfi1_qp_priv *priv = qp->priv;
+	struct hfi1_opfn_type *extd;
+	u8 capcode;
+	unsigned long flags;
+
+	trace_hfi1_opfn_state_conn_reply(qp);
+	capcode = data & 0xf;
+	trace_hfi1_opfn_data_conn_reply(qp, capcode, data);
+	if (!capcode || capcode >= STL_VERBS_EXTD_MAX)
+		return;
+
+	spin_lock_irqsave(&priv->opfn.lock, flags);
+	/*
+	 * Either there is no previous request or the reply is not for the
+	 * current request
+	 */
+	if (!priv->opfn.curr || capcode != priv->opfn.curr)
+		goto done;
+
+	extd = &hfi1_opfn_handlers[capcode];
+
+	if (!extd || !extd->reply)
+		goto clear;
+
+	if (extd->reply(qp, data))
+		priv->opfn.completed |= OPFN_CODE(capcode);
+clear:
+	/*
+	 * Clear opfn.curr to indicate that the previous request is no longer in
+	 * progress
+	 */
+	priv->opfn.curr = STL_VERBS_EXTD_NONE;
+	trace_hfi1_opfn_state_conn_reply(qp);
+done:
+	spin_unlock_irqrestore(&priv->opfn.lock, flags);
+}
+
+void opfn_conn_error(struct rvt_qp *qp)
+{
+	struct hfi1_qp_priv *priv = qp->priv;
+	struct hfi1_opfn_type *extd = NULL;
+	unsigned long flags;
+	u16 capcode;
+
+	trace_hfi1_opfn_state_conn_error(qp);
+	trace_hfi1_msg_opfn_conn_error(qp, "error. qp state ", (u64)qp->state);
+	/*
+	 * The QP has gone into the Error state. We have to invalidate all
+	 * negotiated feature, including the one in progress (if any). The RC
+	 * QP handling will clean the WQE for the connection request.
+	 */
+	spin_lock_irqsave(&priv->opfn.lock, flags);
+	while (priv->opfn.completed) {
+		capcode = priv->opfn.completed & ~(priv->opfn.completed - 1);
+		extd = &hfi1_opfn_handlers[ilog2(capcode) + 1];
+		if (extd->error)
+			extd->error(qp);
+		priv->opfn.completed &= ~OPFN_CODE(capcode);
+	}
+	priv->opfn.extended = 0;
+	priv->opfn.requested = 0;
+	priv->opfn.curr = STL_VERBS_EXTD_NONE;
+	spin_unlock_irqrestore(&priv->opfn.lock, flags);
+}
+
+void opfn_qp_init(struct rvt_qp *qp, struct ib_qp_attr *attr, int attr_mask)
+{
+	struct ib_qp *ibqp = &qp->ibqp;
+	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);
+			/*
+			 * We only want to set the OPFN requested bit when the
+			 * QP transitions to RTS.
+			 */
+			if (attr_mask & IB_QP_STATE &&
+			    attr->qp_state == IB_QPS_RTS) {
+				priv->opfn.requested |= OPFN_MASK(TID_RDMA);
+				/*
+				 * If the QP is transitioning to RTS and the
+				 * opfn.completed for TID RDMA has already been
+				 * set, the QP is being moved *back* into RTS.
+				 * We can now renegotiate the TID RDMA
+				 * parameters.
+				 */
+				if (priv->opfn.completed &
+				    OPFN_MASK(TID_RDMA)) {
+					priv->opfn.completed &=
+						~OPFN_MASK(TID_RDMA);
+					/*
+					 * Since the opfn.completed bit was
+					 * already set, it is safe to assume
+					 * that the opfn.extended is also set.
+					 */
+					opfn_schedule_conn_request(qp);
+				}
+			}
+		} else {
+			memset(local, 0, sizeof(*local));
+		}
+	}
+	spin_unlock_irqrestore(&priv->opfn.lock, flags);
+}
+
+void opfn_trigger_conn_request(struct rvt_qp *qp, u32 bth1)
+{
+	struct hfi1_qp_priv *priv = qp->priv;
+
+	if (!priv->opfn.extended && hfi1_opfn_extended(bth1) &&
+	    HFI1_CAP_IS_KSET(OPFN)) {
+		priv->opfn.extended = 1;
+		if (qp->state == IB_QPS_RTS)
+			opfn_conn_request(qp);
+	}
+}
+
+int opfn_init(void)
+{
+	opfn_wq = alloc_workqueue("hfi_opfn",
+				  WQ_SYSFS | WQ_HIGHPRI | WQ_CPU_INTENSIVE |
+				  WQ_MEM_RECLAIM,
+				  HFI1_MAX_ACTIVE_WORKQUEUE_ENTRIES);
+	if (!opfn_wq)
+		return -ENOMEM;
+
+	return 0;
+}
+
+void opfn_exit(void)
+{
+	if (opfn_wq) {
+		destroy_workqueue(opfn_wq);
+		opfn_wq = NULL;
+	}
+}

drivers/infiniband/hw/hfi1/opfn.h

+/* SPDX-License-Identifier: (GPL-2.0 OR BSD-3-Clause) */
+/*
+ * Copyright(c) 2018 Intel Corporation.
+ *
+ */
+#ifndef _HFI1_OPFN_H
+#define _HFI1_OPFN_H
+
+/**
+ * DOC: Omni Path Feature Negotion (OPFN)
+ *
+ * OPFN is a discovery protocol for Intel Omni-Path fabric that
+ * allows two RC QPs to negotiate a common feature that both QPs
+ * can support. Currently, the only OPA feature that OPFN
+ * supports is TID RDMA.
+ *
+ * Architecture
+ *
+ * OPFN involves the communication between two QPs on the HFI
+ * level on an Omni-Path fabric, and ULPs have no knowledge of
+ * OPFN at all.
+ *
+ * Implementation
+ *
+ * OPFN extends the existing IB RC protocol with the following
+ * changes:
+ * -- Uses Bit 24 (reserved) of DWORD 1 of Base Transport
+ *    Header (BTH1) to indicate that the RC QP supports OPFN;
+ * -- Uses a combination of RC COMPARE_SWAP opcode (0x13) and
+ *    the address U64_MAX (0xFFFFFFFFFFFFFFFF) as an OPFN
+ *    request; The 64-bit data carried with the request/response
+ *    contains the parameters for negotiation and will be
+ *    defined in tid_rdma.c file;
+ * -- Defines IB_WR_RESERVED3 as IB_WR_OPFN.
+ *
+ * The OPFN communication will be triggered when an RC QP
+ * receives a request with Bit 24 of BTH1 set. The responder QP
+ * will then post send an OPFN request with its local
+ * parameters, which will be sent to the requester QP once all
+ * existing requests on the responder QP side have been sent.
+ * Once the requester QP receives the OPFN request, it will
+ * keep a copy of the responder QP's parameters, and return a
+ * response packet with its own local parameters. The responder
+ * QP receives the response packet and keeps a copy of the requester
+ * QP's parameters. After this exchange, each side has the parameters
+ * for both sides and therefore can select the right parameters
+ * for future transactions
+ */
+
+/* STL Verbs Extended */
+#define IB_BTHE_E_SHIFT           24
+#define HFI1_VERBS_E_ATOMIC_VADDR U64_MAX
+
+struct ib_atomic_eth;
+
+enum hfi1_opfn_codes {
+	STL_VERBS_EXTD_NONE = 0,
+	STL_VERBS_EXTD_TID_RDMA,
+	STL_VERBS_EXTD_MAX
+};
+
+struct hfi1_opfn_data {
+	u8 extended;
+	u16 requested;
+	u16 completed;
+	enum hfi1_opfn_codes curr;
+	/* serialize opfn function calls */
+	spinlock_t lock;
+	struct work_struct opfn_work;
+};
+
+/* WR opcode for OPFN */
+#define IB_WR_OPFN IB_WR_RESERVED3
+
+void opfn_send_conn_request(struct work_struct *work);
+void opfn_conn_response(struct rvt_qp *qp, struct rvt_ack_entry *e,
+			struct ib_atomic_eth *ateth);
+void opfn_conn_reply(struct rvt_qp *qp, u64 data);
+void opfn_conn_error(struct rvt_qp *qp);
+void opfn_qp_init(struct rvt_qp *qp, struct ib_qp_attr *attr, int attr_mask);
+void opfn_trigger_conn_request(struct rvt_qp *qp, u32 bth1);
+int opfn_init(void);
+void opfn_exit(void);
+
+#endif /* _HFI1_OPFN_H */

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

@@ -132,6 +132,18 @@ const struct rvt_operation_params hfi1_post_parms[RVT_OPERATION_MAX] = {
 	.qpt_support = BIT(IB_QPT_RC),
 },
 
+[IB_WR_OPFN] = {
+	.length = sizeof(struct ib_atomic_wr),
+	.qpt_support = BIT(IB_QPT_RC),
+	.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)
@@ -285,6 +297,8 @@ void hfi1_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
 		priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
 		qp_set_16b(qp);
 	}
+
+	opfn_qp_init(qp, attr, attr_mask);
 }
 
 /**
@@ -311,6 +325,7 @@ 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;
@@ -422,6 +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)
@@ -441,8 +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(
@@ -479,6 +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;
@@ -528,6 +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
@@ -685,10 +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;
 }
 
@@ -696,6 +748,7 @@ void qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
 {
 	struct hfi1_qp_priv *priv = qp->priv;
 
+	hfi1_qp_priv_tid_free(rdi, qp);
 	kfree(priv->s_ahg);
 	kfree(priv);
 }
@@ -729,6 +782,7 @@ 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)
@@ -736,12 +790,16 @@ 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)
 {
 	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);
@@ -749,8 +807,13 @@ 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);
+
+	/* Clear any OPFN state */
+	if (qp->ibqp.qp_type == IB_QPT_RC)
+		opfn_conn_error(qp);
 }
 
 /*
@@ -832,7 +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;
@@ -841,7 +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

@@ -63,11 +63,17 @@ 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_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
 
 /*
@@ -76,6 +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

+/* SPDX-License-Identifier: (GPL-2.0 OR BSD-3-Clause) */
+/*
+ * Copyright(c) 2018 Intel Corporation.
+ *
+ */
+
+#ifndef HFI1_RC_H
+#define HFI1_RC_H
+
+/* cut down ridiculously long IB macro names */
+#define OP(x) IB_OPCODE_RC_##x
+
+static inline void update_ack_queue(struct rvt_qp *qp, unsigned int n)
+{
+	unsigned int next;
+
+	next = n + 1;
+	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);
+}
+
+static inline void rc_defered_ack(struct hfi1_ctxtdata *rcd,
+				  struct rvt_qp *qp)
+{
+	if (list_empty(&qp->rspwait)) {
+		qp->r_flags |= RVT_R_RSP_NAK;
+		rvt_get_qp(qp);
+		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
+	}
+}
+
+static inline u32 restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe,
+			      u32 psn, u32 pmtu)
+{
+	u32 len;
+
+	len = delta_psn(psn, wqe->psn) * pmtu;
+	return rvt_restart_sge(ss, wqe, len);
+}
+
+struct rvt_ack_entry *find_prev_entry(struct rvt_qp *qp, u32 psn, u8 *prev,
+				      u8 *prev_ack, bool *scheduled);
+int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode, u64 val,
+	      struct hfi1_ctxtdata *rcd);
+struct rvt_swqe *do_rc_completion(struct rvt_qp *qp, struct rvt_swqe *wqe,
+				  struct hfi1_ibport *ibp);
+
+#endif /* HFI1_RC_H */

drivers/infiniband/hw/hfi1/ruc.c

@@ -250,7 +250,6 @@ static inline void hfi1_make_ruc_bth(struct rvt_qp *qp,
 				     struct ib_other_headers *ohdr,
 				     u32 bth0, u32 bth1, u32 bth2)
 {
-	bth1 |= qp->remote_qpn;
 	ohdr->bth[0] = cpu_to_be32(bth0);
 	ohdr->bth[1] = cpu_to_be32(bth1);
 	ohdr->bth[2] = cpu_to_be32(bth2);
@@ -272,13 +271,13 @@ static inline void hfi1_make_ruc_bth(struct rvt_qp *qp,
  */
 static inline void hfi1_make_ruc_header_16B(struct rvt_qp *qp,
 					    struct ib_other_headers *ohdr,
-					    u32 bth0, u32 bth2, int middle,
+					    u32 bth0, u32 bth1, u32 bth2,
+					    int middle,
 					    struct hfi1_pkt_state *ps)
 {
 	struct hfi1_qp_priv *priv = qp->priv;
 	struct hfi1_ibport *ibp = ps->ibp;
 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
-	u32 bth1 = 0;
 	u32 slid;
 	u16 pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
 	u8 l4 = OPA_16B_L4_IB_LOCAL;
@@ -360,12 +359,12 @@ static inline void hfi1_make_ruc_header_16B(struct rvt_qp *qp,
  */
 static inline void hfi1_make_ruc_header_9B(struct rvt_qp *qp,
 					   struct ib_other_headers *ohdr,
-					   u32 bth0, u32 bth2, int middle,
+					   u32 bth0, u32 bth1, u32 bth2,
+					   int middle,
 					   struct hfi1_pkt_state *ps)
 {
 	struct hfi1_qp_priv *priv = qp->priv;
 	struct hfi1_ibport *ibp = ps->ibp;
-	u32 bth1 = 0;
 	u16 pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
 	u16 lrh0 = HFI1_LRH_BTH;
 	u8 extra_bytes = -ps->s_txreq->s_cur_size & 3;
@@ -415,7 +414,7 @@ static inline void hfi1_make_ruc_header_9B(struct rvt_qp *qp,
 
 typedef void (*hfi1_make_ruc_hdr)(struct rvt_qp *qp,
 				  struct ib_other_headers *ohdr,
-				  u32 bth0, u32 bth2, int middle,
+				  u32 bth0, u32 bth1, u32 bth2, int middle,
 				  struct hfi1_pkt_state *ps);
 
 /* We support only two types - 9B and 16B for now */
@@ -425,7 +424,7 @@ static const hfi1_make_ruc_hdr hfi1_ruc_header_tbl[2] = {
 };
 
 void hfi1_make_ruc_header(struct rvt_qp *qp, struct ib_other_headers *ohdr,
-			  u32 bth0, u32 bth2, int middle,
+			  u32 bth0, u32 bth1, u32 bth2, int middle,
 			  struct hfi1_pkt_state *ps)
 {
 	struct hfi1_qp_priv *priv = qp->priv;
@@ -446,18 +445,21 @@ void hfi1_make_ruc_header(struct rvt_qp *qp, struct ib_other_headers *ohdr,
 	priv->s_ahg->ahgidx = 0;
 
 	/* Make the appropriate header */
-	hfi1_ruc_header_tbl[priv->hdr_type](qp, ohdr, bth0, bth2, middle, ps);
+	hfi1_ruc_header_tbl[priv->hdr_type](qp, ohdr, bth0, bth1, bth2, middle,
+					    ps);
 }
 
 /* when sending, force a reschedule every one of these periods */
 #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/trace.c

@@ -46,6 +46,7 @@
  */
 #define CREATE_TRACE_POINTS
 #include "trace.h"
+#include "exp_rcv.h"
 
 static u8 __get_ib_hdr_len(struct ib_header *hdr)
 {
@@ -128,6 +129,15 @@ const char *hfi1_trace_get_packet_l2_str(u8 l2)
 #define IETH_PRN "ieth rkey:0x%.8x"
 #define ATOMICACKETH_PRN "origdata:%llx"
 #define ATOMICETH_PRN "vaddr:0x%llx rkey:0x%.8x sdata:%llx cdata:%llx"
+#define TID_RDMA_KDETH "kdeth0 0x%x kdeth1 0x%x"
+#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
 
@@ -322,6 +332,99 @@ 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,
+				 le32_to_cpu(eh->tid_rdma.r_req.kdeth0),
+				 le32_to_cpu(eh->tid_rdma.r_req.kdeth1),
+				 ib_u64_get(&eh->tid_rdma.r_req.reth.vaddr),
+				 be32_to_cpu(eh->tid_rdma.r_req.reth.rkey),
+				 be32_to_cpu(eh->tid_rdma.r_req.reth.length),
+				 be32_to_cpu(eh->tid_rdma.r_req.tid_flow_psn),
+				 be32_to_cpu(eh->tid_rdma.r_req.tid_flow_qp),
+				 be32_to_cpu(eh->tid_rdma.r_req.verbs_qp));
+		break;
+	case OP(TID_RDMA, READ_RESP):
+		trace_seq_printf(p, TID_RDMA_KDETH_DATA " " AETH_PRN " "
+				 TID_READ_RSP_PRN,
+				 le32_to_cpu(eh->tid_rdma.r_rsp.kdeth0),
+				 KDETH_GET(eh->tid_rdma.r_rsp.kdeth0, KVER),
+				 KDETH_GET(eh->tid_rdma.r_rsp.kdeth0, SH),
+				 KDETH_GET(eh->tid_rdma.r_rsp.kdeth0, INTR),
+				 KDETH_GET(eh->tid_rdma.r_rsp.kdeth0, TIDCTRL),
+				 KDETH_GET(eh->tid_rdma.r_rsp.kdeth0, TID),
+				 KDETH_GET(eh->tid_rdma.r_rsp.kdeth0, OFFSET),
+				 le32_to_cpu(eh->tid_rdma.r_rsp.kdeth1),
+				 KDETH_GET(eh->tid_rdma.r_rsp.kdeth1, JKEY),
+				 be32_to_cpu(eh->tid_rdma.r_rsp.aeth) >> 24,
+				 parse_syndrome(/* aeth */
+					 be32_to_cpu(eh->tid_rdma.r_rsp.aeth)
+					 >> 24),
+				 (be32_to_cpu(eh->tid_rdma.r_rsp.aeth) &
+				  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,
@@ -394,6 +497,21 @@ const char *print_u32_array(
 	return ret;
 }
 
+u8 hfi1_trace_get_tid_ctrl(u32 ent)
+{
+	return EXP_TID_GET(ent, CTRL);
+}
+
+u16 hfi1_trace_get_tid_len(u32 ent)
+{
+	return EXP_TID_GET(ent, LEN);
+}
+
+u16 hfi1_trace_get_tid_idx(u32 ent)
+{
+	return EXP_TID_GET(ent, IDX);
+}
+
 __hfi1_trace_fn(AFFINITY);
 __hfi1_trace_fn(PKT);
 __hfi1_trace_fn(PROC);

drivers/infiniband/hw/hfi1/trace.h

@@ -63,3 +63,4 @@ __print_symbolic(etype,                         \
 #include "trace_tx.h"
 #include "trace_mmu.h"
 #include "trace_iowait.h"
+#include "trace_tid.h"

drivers/infiniband/hw/hfi1/trace_ibhdrs.h

@@ -79,6 +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_rc.h

@@ -109,6 +109,54 @@ DEFINE_EVENT(hfi1_rc_template, hfi1_rcv_error,
 	     TP_ARGS(qp, psn)
 );
 
+DEFINE_EVENT(/* event */
+	hfi1_rc_template, hfi1_rc_completion,
+	TP_PROTO(struct rvt_qp *qp, u32 psn),
+	TP_ARGS(qp, psn)
+);
+
+DECLARE_EVENT_CLASS(/* rc_ack */
+	hfi1_rc_ack_template,
+	TP_PROTO(struct rvt_qp *qp, u32 aeth, u32 psn,
+		 struct rvt_swqe *wqe),
+	TP_ARGS(qp, aeth, psn, wqe),
+	TP_STRUCT__entry(/* entry */
+		DD_DEV_ENTRY(dd_from_ibdev(qp->ibqp.device))
+		__field(u32, qpn)
+		__field(u32, aeth)
+		__field(u32, psn)
+		__field(u8, opcode)
+		__field(u32, spsn)
+		__field(u32, lpsn)
+	),
+	TP_fast_assign(/* assign */
+		DD_DEV_ASSIGN(dd_from_ibdev(qp->ibqp.device))
+		__entry->qpn = qp->ibqp.qp_num;
+		__entry->aeth = aeth;
+		__entry->psn = psn;
+		__entry->opcode = wqe->wr.opcode;
+		__entry->spsn = wqe->psn;
+		__entry->lpsn = wqe->lpsn;
+	),
+	TP_printk(/* print */
+		"[%s] qpn 0x%x aeth 0x%x psn 0x%x opcode 0x%x spsn 0x%x lpsn 0x%x",
+		__get_str(dev),
+		__entry->qpn,
+		__entry->aeth,
+		__entry->psn,
+		__entry->opcode,
+		__entry->spsn,
+		__entry->lpsn
+	)
+);
+
+DEFINE_EVENT(/* do_rc_ack */
+	hfi1_rc_ack_template, hfi1_rc_ack_do,
+	TP_PROTO(struct rvt_qp *qp, u32 aeth, u32 psn,
+		 struct rvt_swqe *wqe),
+	TP_ARGS(qp, aeth, psn, wqe)
+);
+
 #endif /* __HFI1_TRACE_RC_H */
 
 #undef TRACE_INCLUDE_PATH

drivers/infiniband/hw/hfi1/trace_rx.h

 /*
- * Copyright(c) 2015 - 2017 Intel Corporation.
+ * Copyright(c) 2015 - 2018 Intel Corporation.
  *
  * This file is provided under a dual BSD/GPLv2 license.  When using or
  * redistributing this file, you may do so under either license.
@@ -128,111 +128,6 @@ TRACE_EVENT(hfi1_receive_interrupt,
 		      )
 );
 
-DECLARE_EVENT_CLASS(
-	    hfi1_exp_tid_reg_unreg,
-	    TP_PROTO(unsigned int ctxt, u16 subctxt, u32 rarr,
-		     u32 npages, unsigned long va, unsigned long pa,
-		     dma_addr_t dma),
-	    TP_ARGS(ctxt, subctxt, rarr, npages, va, pa, dma),
-	    TP_STRUCT__entry(
-			     __field(unsigned int, ctxt)
-			     __field(u16, subctxt)
-			     __field(u32, rarr)
-			     __field(u32, npages)
-			     __field(unsigned long, va)
-			     __field(unsigned long, pa)
-			     __field(dma_addr_t, dma)
-			     ),
-	    TP_fast_assign(
-			   __entry->ctxt = ctxt;
-			   __entry->subctxt = subctxt;
-			   __entry->rarr = rarr;
-			   __entry->npages = npages;
-			   __entry->va = va;
-			   __entry->pa = pa;
-			   __entry->dma = dma;
-			   ),
-	    TP_printk("[%u:%u] entry:%u, %u pages @ 0x%lx, va:0x%lx dma:0x%llx",
-		      __entry->ctxt,
-		      __entry->subctxt,
-		      __entry->rarr,
-		      __entry->npages,
-		      __entry->pa,
-		      __entry->va,
-		      __entry->dma
-		      )
-	);
-
-DEFINE_EVENT(
-	hfi1_exp_tid_reg_unreg, hfi1_exp_tid_unreg,
-	TP_PROTO(unsigned int ctxt, u16 subctxt, u32 rarr, u32 npages,
-		 unsigned long va, unsigned long pa, dma_addr_t dma),
-	TP_ARGS(ctxt, subctxt, rarr, npages, va, pa, dma));
-
-DEFINE_EVENT(
-	hfi1_exp_tid_reg_unreg, hfi1_exp_tid_reg,
-	TP_PROTO(unsigned int ctxt, u16 subctxt, u32 rarr, u32 npages,
-		 unsigned long va, unsigned long pa, dma_addr_t dma),
-	TP_ARGS(ctxt, subctxt, rarr, npages, va, pa, dma));
-
-TRACE_EVENT(
-	hfi1_put_tid,
-	TP_PROTO(struct hfi1_devdata *dd,
-		 u32 index, u32 type, unsigned long pa, u16 order),
-	TP_ARGS(dd, index, type, pa, order),
-	TP_STRUCT__entry(
-		DD_DEV_ENTRY(dd)
-		__field(unsigned long, pa);
-		__field(u32, index);
-		__field(u32, type);
-		__field(u16, order);
-	),
-	TP_fast_assign(
-		DD_DEV_ASSIGN(dd);
-		__entry->pa = pa;
-		__entry->index = index;
-		__entry->type = type;
-		__entry->order = order;
-	),
-	TP_printk("[%s] type %s pa %lx index %u order %u",
-		  __get_str(dev),
-		  show_tidtype(__entry->type),
-		  __entry->pa,
-		  __entry->index,
-		  __entry->order
-	)
-);
-
-TRACE_EVENT(hfi1_exp_tid_inval,
-	    TP_PROTO(unsigned int ctxt, u16 subctxt, unsigned long va, u32 rarr,
-		     u32 npages, dma_addr_t dma),
-	    TP_ARGS(ctxt, subctxt, va, rarr, npages, dma),
-	    TP_STRUCT__entry(
-			     __field(unsigned int, ctxt)
-			     __field(u16, subctxt)
-			     __field(unsigned long, va)
-			     __field(u32, rarr)
-			     __field(u32, npages)
-			     __field(dma_addr_t, dma)
-			     ),
-	    TP_fast_assign(
-			   __entry->ctxt = ctxt;
-			   __entry->subctxt = subctxt;
-			   __entry->va = va;
-			   __entry->rarr = rarr;
-			   __entry->npages = npages;
-			   __entry->dma = dma;
-			  ),
-	    TP_printk("[%u:%u] entry:%u, %u pages @ 0x%lx dma: 0x%llx",
-		      __entry->ctxt,
-		      __entry->subctxt,
-		      __entry->rarr,
-		      __entry->npages,
-		      __entry->va,
-		      __entry->dma
-		      )
-	    );
-
 TRACE_EVENT(hfi1_mmu_invalidate,
 	    TP_PROTO(unsigned int ctxt, u16 subctxt, const char *type,
 		     unsigned long start, unsigned long end),

drivers/infiniband/hw/hfi1/trace_tx.h

@@ -114,19 +114,27 @@ DECLARE_EVENT_CLASS(hfi1_qpsleepwakeup_template,
 		    __field(u32, qpn)
 		    __field(u32, flags)
 		    __field(u32, s_flags)
+		    __field(u32, ps_flags)
+		    __field(unsigned long, iow_flags)
 		    ),
 		    TP_fast_assign(
 		    DD_DEV_ASSIGN(dd_from_ibdev(qp->ibqp.device))
 		    __entry->flags = flags;
 		    __entry->qpn = qp->ibqp.qp_num;
 		    __entry->s_flags = qp->s_flags;
+		    __entry->ps_flags =
+			((struct hfi1_qp_priv *)qp->priv)->s_flags;
+		    __entry->iow_flags =
+			((struct hfi1_qp_priv *)qp->priv)->s_iowait.flags;
 		    ),
 		    TP_printk(
-		    "[%s] qpn 0x%x flags 0x%x s_flags 0x%x",
+		    "[%s] qpn 0x%x flags 0x%x s_flags 0x%x ps_flags 0x%x iow_flags 0x%lx",
 		    __get_str(dev),
 		    __entry->qpn,
 		    __entry->flags,
-		    __entry->s_flags
+		    __entry->s_flags,
+		    __entry->ps_flags,
+		    __entry->iow_flags
 		    )
 );
 
@@ -838,6 +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),

drivers/infiniband/hw/hfi1/uc.c

@@ -271,7 +271,8 @@ int hfi1_make_uc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 	ps->s_txreq->ss = &qp->s_sge;
 	ps->s_txreq->s_cur_size = len;
 	hfi1_make_ruc_header(qp, ohdr, bth0 | (qp->s_state << 24),
-			     mask_psn(qp->s_psn++), middle, ps);
+			     qp->remote_qpn, mask_psn(qp->s_psn++),
+			     middle, ps);
 	return 1;
 
 done_free_tx:

drivers/infiniband/hw/hfi1/user_exp_rcv.h

@@ -48,7 +48,6 @@
  */
 
 #include "hfi.h"
-
 #include "exp_rcv.h"
 
 struct tid_pageset {

drivers/infiniband/hw/hfi1/user_sdma.c

@@ -144,8 +144,10 @@ static int defer_packet_queue(
 	 */
 	xchg(&pq->state, SDMA_PKT_Q_DEFERRED);
 	write_seqlock(&sde->waitlock);
-	if (list_empty(&pq->busy.list))
+	if (list_empty(&pq->busy.list)) {
+		iowait_get_priority(&pq->busy);
 		iowait_queue(pkts_sent, &pq->busy, &sde->dmawait);
+	}
 	write_sequnlock(&sde->waitlock);
 	return -EBUSY;
 eagain:
@@ -191,7 +193,7 @@ int hfi1_user_sdma_alloc_queues(struct hfi1_ctxtdata *uctxt,
 	pq->mm = fd->mm;
 
 	iowait_init(&pq->busy, 0, NULL, NULL, defer_packet_queue,
-		    activate_packet_queue, NULL);
+		    activate_packet_queue, NULL, NULL);
 	pq->reqidx = 0;
 
 	pq->reqs = kcalloc(hfi1_sdma_comp_ring_size,
@@ -1126,7 +1128,8 @@ static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags)
 			0xffffffull),
 		psn = val & mask;
 	if (expct)
-		psn = (psn & ~BTH_SEQ_MASK) | ((psn + frags) & BTH_SEQ_MASK);
+		psn = (psn & ~HFI1_KDETH_BTH_SEQ_MASK) |
+			((psn + frags) & HFI1_KDETH_BTH_SEQ_MASK);
 	else
 		psn = psn + frags;
 	return psn & mask;

drivers/infiniband/hw/hfi1/verbs.h

@@ -72,6 +72,7 @@ struct hfi1_packet;
 
 #include "iowait.h"
 #include "tid_rdma.h"
+#include "opfn.h"
 
 #define HFI1_MAX_RDMA_ATOMIC     16
 
@@ -158,10 +159,68 @@ struct hfi1_qp_priv {
 	struct sdma_engine *s_sde;                /* current sde */
 	struct send_context *s_sendcontext;       /* current sendcontext */
 	struct hfi1_ctxtdata *rcd;                /* QP's receive context */
+	struct page **pages;                      /* for TID page scan */
+	u32 tid_enqueue;                          /* saved when tid waited */
 	u8 s_sc;		                  /* SC[0..4] for next packet */
 	struct iowait s_iowait;
+	struct timer_list s_tid_timer;            /* for timing tid wait */
+	struct timer_list s_tid_retry_timer;      /* for timing tid ack */
+	struct list_head tid_wait;                /* for queueing tid space */
+	struct hfi1_opfn_data opfn;
+	struct tid_flow_state flow_state;
+	struct tid_rdma_qp_params tid_rdma;
 	struct rvt_qp *owner;
 	u8 hdr_type; /* 9B or 16B */
+	struct rvt_sge_state tid_ss;       /* SGE state pointer for 2nd leg */
+	atomic_t n_requests;               /* # of TID RDMA requests in the */
+					   /* queue */
+	atomic_t n_tid_requests;            /* # of sent TID RDMA requests */
+	unsigned long tid_timer_timeout_jiffies;
+	unsigned long tid_retry_timeout_jiffies;
+
+	/* variables for the TID RDMA SE state machine */
+	u8 s_state;
+	u8 s_retry;
+	u8 rnr_nak_state;       /* RNR NAK state */
+	u8 s_nak_state;
+	u32 s_nak_psn;
+	u32 s_flags;
+	u32 s_tid_cur;
+	u32 s_tid_head;
+	u32 s_tid_tail;
+	u32 r_tid_head;     /* Most recently added TID RDMA request */
+	u32 r_tid_tail;     /* the last completed TID RDMA request */
+	u32 r_tid_ack;      /* the TID RDMA request to be ACK'ed */
+	u32 r_tid_alloc;    /* Request for which we are allocating resources */
+	u32 pending_tid_w_segs; /* Num of pending tid write segments */
+	u32 pending_tid_w_resp; /* Num of pending tid write responses */
+	u32 alloc_w_segs;       /* Number of segments for which write */
+			       /* resources have been allocated for this QP */
+
+	/* For TID RDMA READ */
+	u32 tid_r_reqs;         /* Num of tid reads requested */
+	u32 tid_r_comp;         /* Num of tid reads completed */
+	u32 pending_tid_r_segs; /* Num of pending tid read segments */
+	u16 pkts_ps;            /* packets per segment */
+	u8 timeout_shift;       /* account for number of packets per segment */
+
+	u32 r_next_psn_kdeth;
+	u32 r_next_psn_kdeth_save;
+	u32 s_resync_psn;
+	u8 sync_pt;           /* Set when QP reaches sync point */
+	u8 resync;
+};
+
+#define HFI1_QP_WQE_INVALID   ((u32)-1)
+
+struct hfi1_swqe_priv {
+	struct tid_rdma_request tid_req;
+	struct rvt_sge_state ss;  /* Used for TID RDMA READ Request */
+};
+
+struct hfi1_ack_priv {
+	struct rvt_sge_state ss;               /* used for TID WRITE RESP */
+	struct tid_rdma_request tid_req;
 };
 
 /*
@@ -225,6 +284,7 @@ struct hfi1_ibdev {
 	struct kmem_cache *verbs_txreq_cache;
 	u64 n_txwait;
 	u64 n_kmem_wait;
+	u64 n_tidwait;
 
 	/* protect iowait lists */
 	seqlock_t iowait_lock ____cacheline_aligned_in_smp;
@@ -312,6 +372,31 @@ static inline u32 delta_psn(u32 a, u32 b)
 	return (((int)a - (int)b) << PSN_SHIFT) >> PSN_SHIFT;
 }
 
+static inline struct tid_rdma_request *wqe_to_tid_req(struct rvt_swqe *wqe)
+{
+	return &((struct hfi1_swqe_priv *)wqe->priv)->tid_req;
+}
+
+static inline struct tid_rdma_request *ack_to_tid_req(struct rvt_ack_entry *e)
+{
+	return &((struct hfi1_ack_priv *)e->priv)->tid_req;
+}
+
+/*
+ * Look through all the active flows for a TID RDMA request and find
+ * the one (if it exists) that contains the specified PSN.
+ */
+static inline u32 __full_flow_psn(struct flow_state *state, u32 psn)
+{
+	return mask_psn((state->generation << HFI1_KDETH_BTH_SEQ_SHIFT) |
+			(psn & HFI1_KDETH_BTH_SEQ_MASK));
+}
+
+static inline u32 full_flow_psn(struct tid_rdma_flow *flow, u32 psn)
+{
+	return __full_flow_psn(&flow->flow_state, psn);
+}
+
 struct verbs_txreq;
 void hfi1_put_txreq(struct verbs_txreq *tx);
 
@@ -356,9 +441,12 @@ u32 hfi1_make_grh(struct hfi1_ibport *ibp, struct ib_grh *hdr,
 		  const struct ib_global_route *grh, u32 hwords, u32 nwords);
 
 void hfi1_make_ruc_header(struct rvt_qp *qp, struct ib_other_headers *ohdr,
-			  u32 bth0, u32 bth2, int middle,
+			  u32 bth0, u32 bth1, u32 bth2, int middle,
 			  struct hfi1_pkt_state *ps);
 
+bool hfi1_schedule_send_yield(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
+			      bool tid);
+
 void _hfi1_do_send(struct work_struct *work);
 
 void hfi1_do_send_from_rvt(struct rvt_qp *qp);
@@ -377,6 +465,10 @@ int hfi1_register_ib_device(struct hfi1_devdata *);
 
 void hfi1_unregister_ib_device(struct hfi1_devdata *);
 
+void hfi1_kdeth_eager_rcv(struct hfi1_packet *packet);
+
+void hfi1_kdeth_expected_rcv(struct hfi1_packet *packet);
+
 void hfi1_ib_rcv(struct hfi1_packet *packet);
 
 void hfi1_16B_rcv(struct hfi1_packet *packet);
@@ -394,6 +486,16 @@ static inline bool opa_bth_is_migration(struct ib_other_headers *ohdr)
 	return ohdr->bth[1] & cpu_to_be32(OPA_BTH_MIG_REQ);
 }
 
+void hfi1_wait_kmem(struct rvt_qp *qp);
+
+static inline void hfi1_trdma_send_complete(struct rvt_qp *qp,
+					    struct rvt_swqe *wqe,
+					    enum ib_wc_status status)
+{
+	trdma_clean_swqe(qp, wqe);
+	rvt_send_complete(qp, wqe, status);
+}
+
 extern const enum ib_wc_opcode ib_hfi1_wc_opcode[];
 
 extern const u8 hdr_len_by_opcode[];

drivers/infiniband/hw/hfi1/verbs_txreq.h

@@ -94,6 +94,7 @@ static inline struct verbs_txreq *get_txreq(struct hfi1_ibdev *dev,
 	tx->txreq.num_desc = 0;
 	/* Set the header type */
 	tx->phdr.hdr.hdr_type = priv->hdr_type;
+	tx->txreq.flags = 0;
 	return tx;
 }
 

drivers/infiniband/hw/hfi1/vnic_sdma.c

@@ -240,8 +240,10 @@ static int hfi1_vnic_sdma_sleep(struct sdma_engine *sde,
 	}
 
 	vnic_sdma->state = HFI1_VNIC_SDMA_Q_DEFERRED;
-	if (list_empty(&vnic_sdma->wait.list))
+	if (list_empty(&vnic_sdma->wait.list)) {
+		iowait_get_priority(wait->iow);
 		iowait_queue(pkts_sent, wait->iow, &sde->dmawait);
+	}
 	write_sequnlock(&sde->waitlock);
 	return -EBUSY;
 }
@@ -281,7 +283,7 @@ void hfi1_vnic_sdma_init(struct hfi1_vnic_vport_info *vinfo)
 
 		iowait_init(&vnic_sdma->wait, 0, NULL, NULL,
 			    hfi1_vnic_sdma_sleep,
-			    hfi1_vnic_sdma_wakeup, NULL);
+			    hfi1_vnic_sdma_wakeup, NULL, NULL);
 		vnic_sdma->sde = &vinfo->dd->per_sdma[i];
 		vnic_sdma->dd = vinfo->dd;
 		vnic_sdma->vinfo = vinfo;

drivers/infiniband/hw/qib/qib_rc.c

@@ -45,12 +45,7 @@ static u32 restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe,
 	u32 len;
 
 	len = ((psn - wqe->psn) & QIB_PSN_MASK) * 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;
+	return rvt_restart_sge(ss, wqe, len);
 }
 
 /**

drivers/infiniband/sw/rdmavt/qp.c

@@ -854,6 +854,7 @@ static void rvt_init_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp,
 	qp->s_mig_state = IB_MIG_MIGRATED;
 	qp->r_head_ack_queue = 0;
 	qp->s_tail_ack_queue = 0;
+	qp->s_acked_ack_queue = 0;
 	qp->s_num_rd_atomic = 0;
 	if (qp->r_rq.wq) {
 		qp->r_rq.wq->head = 0;
@@ -1642,11 +1643,11 @@ int rvt_destroy_qp(struct ib_qp *ibqp)
 		kref_put(&qp->ip->ref, rvt_release_mmap_info);
 	else
 		vfree(qp->r_rq.wq);
-	vfree(qp->s_wq);
 	rdi->driver_f.qp_priv_free(rdi, qp);
 	kfree(qp->s_ack_queue);
 	rdma_destroy_ah_attr(&qp->remote_ah_attr);
 	rdma_destroy_ah_attr(&qp->alt_ah_attr);
+	vfree(qp->s_wq);
 	kfree(qp);
 	return 0;
 }
@@ -2393,11 +2394,12 @@ static inline unsigned long rvt_aeth_to_usec(u32 aeth)
 }
 
 /*
- *  rvt_add_retry_timer - add/start a retry timer
+ *  rvt_add_retry_timer_ext - add/start a retry timer
  *  @qp - the QP
+ *  @shift - timeout shift to wait for multiple packets
  *  add a retry timer on the QP
  */
-void rvt_add_retry_timer(struct rvt_qp *qp)
+void rvt_add_retry_timer_ext(struct rvt_qp *qp, u8 shift)
 {
 	struct ib_qp *ibqp = &qp->ibqp;
 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
@@ -2405,11 +2407,11 @@ void rvt_add_retry_timer(struct rvt_qp *qp)
 	lockdep_assert_held(&qp->s_lock);
 	qp->s_flags |= RVT_S_TIMER;
        /* 4.096 usec. * (1 << qp->timeout) */
-	qp->s_timer.expires = jiffies + qp->timeout_jiffies +
-			     rdi->busy_jiffies;
+	qp->s_timer.expires = jiffies + rdi->busy_jiffies +
+			      (qp->timeout_jiffies << shift);
 	add_timer(&qp->s_timer);
 }
-EXPORT_SYMBOL(rvt_add_retry_timer);
+EXPORT_SYMBOL(rvt_add_retry_timer_ext);
 
 /**
  * rvt_add_rnr_timer - add/start an rnr timer

drivers/infiniband/sw/rdmavt/rc.c

@@ -187,3 +187,16 @@ void rvt_get_credit(struct rvt_qp *qp, u32 aeth)
 	}
 }
 EXPORT_SYMBOL(rvt_get_credit);
+
+/* rvt_restart_sge - rewind the sge state for a wqe */
+u32 rvt_restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe, u32 len)
+{
+	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;
+}
+EXPORT_SYMBOL(rvt_restart_sge);
+

include/rdma/ib_hdrs.h

 /*
- * Copyright(c) 2016 Intel Corporation.
+ * Copyright(c) 2016 - 2018 Intel Corporation.
  *
  * This file is provided under a dual BSD/GPLv2 license.  When using or
  * redistributing this file, you may do so under either license.
@@ -100,6 +100,8 @@ struct ib_atomic_eth {
 	__be64 compare_data; /* potentially unaligned */
 } __packed;
 
+#include <rdma/tid_rdma_defs.h>
+
 union ib_ehdrs {
 	struct {
 		__be32 deth[2];
@@ -117,6 +119,16 @@ union ib_ehdrs {
 	__be32 aeth;
 	__be32 ieth;
 	struct ib_atomic_eth atomic_eth;
+	/* TID RDMA headers */
+	union {
+		struct tid_rdma_read_req r_req;
+		struct tid_rdma_read_resp r_rsp;
+		struct tid_rdma_write_req w_req;
+		struct tid_rdma_write_resp w_rsp;
+		struct tid_rdma_write_data w_data;
+		struct tid_rdma_resync resync;
+		struct tid_rdma_ack ack;
+	} tid_rdma;
 }  __packed;
 
 struct ib_other_headers {

include/rdma/rdma_vt.h

@@ -182,6 +182,7 @@ struct rvt_driver_params {
 	u32 max_mad_size;
 	u8 qos_shift;
 	u8 max_rdma_atomic;
+	u8 extra_rdma_atomic;
 	u8 reserved_operations;
 };
 
@@ -519,7 +520,14 @@ static inline unsigned rvt_get_npkeys(struct rvt_dev_info *rdi)
  */
 static inline unsigned int rvt_max_atomic(struct rvt_dev_info *rdi)
 {
-	return rdi->dparms.max_rdma_atomic + 1;
+	return rdi->dparms.max_rdma_atomic +
+		rdi->dparms.extra_rdma_atomic + 1;
+}
+
+static inline unsigned int rvt_size_atomic(struct rvt_dev_info *rdi)
+{
+	return rdi->dparms.max_rdma_atomic +
+		rdi->dparms.extra_rdma_atomic;
 }
 
 /*
@@ -566,9 +574,10 @@ static inline struct rvt_qp *rvt_lookup_qpn(struct rvt_dev_info *rdi,
 /**
  * rvt_mod_retry_timer - mod a retry timer
  * @qp - the QP
+ * @shift - timeout shift to wait for multiple packets
  * Modify a potentially already running retry timer
  */
-static inline void rvt_mod_retry_timer(struct rvt_qp *qp)
+static inline void rvt_mod_retry_timer_ext(struct rvt_qp *qp, u8 shift)
 {
 	struct ib_qp *ibqp = &qp->ibqp;
 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
@@ -576,8 +585,13 @@ static inline void rvt_mod_retry_timer(struct rvt_qp *qp)
 	lockdep_assert_held(&qp->s_lock);
 	qp->s_flags |= RVT_S_TIMER;
 	/* 4.096 usec. * (1 << qp->timeout) */
-	mod_timer(&qp->s_timer, jiffies + qp->timeout_jiffies +
-		  rdi->busy_jiffies);
+	mod_timer(&qp->s_timer, jiffies + rdi->busy_jiffies +
+		  (qp->timeout_jiffies << shift));
+}
+
+static inline void rvt_mod_retry_timer(struct rvt_qp *qp)
+{
+	return rvt_mod_retry_timer_ext(qp, 0);
 }
 
 struct rvt_dev_info *rvt_alloc_device(size_t size, int nports);

include/rdma/rdmavt_qp.h

@@ -174,6 +174,7 @@ struct rvt_swqe {
 	u32 lpsn;               /* last packet sequence number */
 	u32 ssn;                /* send sequence number */
 	u32 length;             /* total length of data in sg_list */
+	void *priv;             /* driver dependent field */
 	struct rvt_sge sg_list[0];
 };
 
@@ -235,6 +236,7 @@ struct rvt_ack_entry {
 	u32 lpsn;
 	u8 opcode;
 	u8 sent;
+	void *priv;
 };
 
 #define	RC_QP_SCALING_INTERVAL	5
@@ -244,6 +246,7 @@ struct rvt_ack_entry {
 #define RVT_OPERATION_ATOMIC_SGE  0x00000004
 #define RVT_OPERATION_LOCAL       0x00000008
 #define RVT_OPERATION_USE_RESERVE 0x00000010
+#define RVT_OPERATION_IGN_RNR_CNT 0x00000020
 
 #define RVT_OPERATION_MAX (IB_WR_RESERVED10 + 1)
 
@@ -373,6 +376,7 @@ struct rvt_qp {
 	u8 s_rnr_retry;         /* requester RNR retry counter */
 	u8 s_num_rd_atomic;     /* number of RDMA read/atomic pending */
 	u8 s_tail_ack_queue;    /* index into s_ack_queue[] */
+	u8 s_acked_ack_queue;   /* index into s_ack_queue[] */
 
 	struct rvt_sge_state s_ack_rdma_sge;
 	struct timer_list s_timer;
@@ -628,6 +632,16 @@ __be32 rvt_compute_aeth(struct rvt_qp *qp);
  */
 void rvt_get_credit(struct rvt_qp *qp, u32 aeth);
 
+/**
+ * rvt_restart_sge - rewind the sge state for a wqe
+ * @ss: the sge state pointer
+ * @wqe: the wqe to rewind
+ * @len: the data length from the start of the wqe in bytes
+ *
+ * Returns the remaining data length.
+ */
+u32 rvt_restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe, u32 len);
+
 /**
  * @qp - the qp pair
  * @len - the length
@@ -676,7 +690,11 @@ enum hrtimer_restart rvt_rc_rnr_retry(struct hrtimer *t);
 void rvt_add_rnr_timer(struct rvt_qp *qp, u32 aeth);
 void rvt_del_timers_sync(struct rvt_qp *qp);
 void rvt_stop_rc_timers(struct rvt_qp *qp);
-void rvt_add_retry_timer(struct rvt_qp *qp);
+void rvt_add_retry_timer_ext(struct rvt_qp *qp, u8 shift);
+static inline void rvt_add_retry_timer(struct rvt_qp *qp)
+{
+	rvt_add_retry_timer_ext(qp, 0);
+}
 
 void rvt_copy_sge(struct rvt_qp *qp, struct rvt_sge_state *ss,
 		  void *data, u32 length,

include/rdma/tid_rdma_defs.h

+/* SPDX-License-Identifier: (GPL-2.0 OR BSD-3-Clause) */
+/*
+ * Copyright(c) 2018 Intel Corporation.
+ *
+ */
+
+#ifndef TID_RDMA_DEFS_H
+#define TID_RDMA_DEFS_H
+
+#include <rdma/ib_pack.h>
+
+struct tid_rdma_read_req {
+	__le32 kdeth0;
+	__le32 kdeth1;
+	struct ib_reth reth;
+	__be32 tid_flow_psn;
+	__be32 tid_flow_qp;
+	__be32 verbs_qp;
+};
+
+struct tid_rdma_read_resp {
+	__le32 kdeth0;
+	__le32 kdeth1;
+	__be32 aeth;
+	__be32 reserved[4];
+	__be32 verbs_psn;
+	__be32 verbs_qp;
+};
+
+struct tid_rdma_write_req {
+	__le32 kdeth0;
+	__le32 kdeth1;
+	struct ib_reth reth;
+	__be32 reserved[2];
+	__be32 verbs_qp;
+};
+
+struct tid_rdma_write_resp {
+	__le32 kdeth0;
+	__le32 kdeth1;
+	__be32 aeth;
+	__be32 reserved[3];
+	__be32 tid_flow_psn;
+	__be32 tid_flow_qp;
+	__be32 verbs_qp;
+};
+
+struct tid_rdma_write_data {
+	__le32 kdeth0;
+	__le32 kdeth1;
+	__be32 reserved[6];
+	__be32 verbs_qp;
+};
+
+struct tid_rdma_resync {
+	__le32 kdeth0;
+	__le32 kdeth1;
+	__be32 reserved[6];
+	__be32 verbs_qp;
+};
+
+struct tid_rdma_ack {
+	__le32 kdeth0;
+	__le32 kdeth1;
+	__be32 aeth;
+	__be32 reserved[2];
+	__be32 tid_flow_psn;
+	__be32 verbs_psn;
+	__be32 tid_flow_qp;
+	__be32 verbs_qp;
+};
+
+/*
+ * TID RDMA Opcodes
+ */
+#define IB_OPCODE_TID_RDMA 0xe0
+enum {
+	IB_OPCODE_WRITE_REQ       = 0x0,
+	IB_OPCODE_WRITE_RESP      = 0x1,
+	IB_OPCODE_WRITE_DATA      = 0x2,
+	IB_OPCODE_WRITE_DATA_LAST = 0x3,
+	IB_OPCODE_READ_REQ        = 0x4,
+	IB_OPCODE_READ_RESP       = 0x5,
+	IB_OPCODE_RESYNC          = 0x6,
+	IB_OPCODE_ACK             = 0x7,
+
+	IB_OPCODE(TID_RDMA, WRITE_REQ),
+	IB_OPCODE(TID_RDMA, WRITE_RESP),
+	IB_OPCODE(TID_RDMA, WRITE_DATA),
+	IB_OPCODE(TID_RDMA, WRITE_DATA_LAST),
+	IB_OPCODE(TID_RDMA, READ_REQ),
+	IB_OPCODE(TID_RDMA, READ_RESP),
+	IB_OPCODE(TID_RDMA, RESYNC),
+	IB_OPCODE(TID_RDMA, ACK),
+};
+
+#define TID_OP(x) IB_OPCODE_TID_RDMA_##x
+
+/*
+ * Define TID RDMA specific WR opcodes. The ib_wr_opcode
+ * enum already provides some reserved values for use by
+ * low level drivers. Two of those are used but renamed
+ * to be more descriptive.
+ */
+#define IB_WR_TID_RDMA_WRITE IB_WR_RESERVED1
+#define IB_WR_TID_RDMA_READ  IB_WR_RESERVED2
+
+#endif /* TID_RDMA_DEFS_H */