diff mbox series

[for-next,3/6] IB/hfi1: OPFN interface

Message ID 20190124032047.29585.5623.stgit@scvm10.sc.intel.com (mailing list archive)
State Accepted
Headers show
Series IB/hfi1: Add OPFN | expand

Commit Message

Dennis Dalessandro Jan. 24, 2019, 3:20 a.m. UTC
From: Kaike Wan <kaike.wan@intel.com>

OPFN allows a pair of connected RC QPs to exchange a set of parameters
in succession. The parameter exchange itself is done using the IB compare
and swap request with a special virtual address. The request is triggered
using a reserved IB work request opcode. This patch implements the OPFN
interface to initialize, start, process, and reset the OPFN request.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
---
 drivers/infiniband/hw/hfi1/Makefile |    1 
 drivers/infiniband/hw/hfi1/hfi.h    |    2 
 drivers/infiniband/hw/hfi1/init.c   |    1 
 drivers/infiniband/hw/hfi1/opfn.c   |  304 +++++++++++++++++++++++++++++++++++
 drivers/infiniband/hw/hfi1/opfn.h   |   27 +++
 5 files changed, 334 insertions(+), 1 deletions(-)
 create mode 100644 drivers/infiniband/hw/hfi1/opfn.c

Comments

Doug Ledford Jan. 30, 2019, 7 p.m. UTC | #1
On Wed, 2019-01-23 at 19:20 -0800, Dennis Dalessandro wrote:
> +       /* 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;
> +       return;
> +err:
> +       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
> +        */

This seems clunky.  Why not if (!ret) return; and skip the whole goto?
diff mbox series

Patch

diff --git a/drivers/infiniband/hw/hfi1/Makefile b/drivers/infiniband/hw/hfi1/Makefile
index 3ce9dc8..4044a8c 100644
--- a/drivers/infiniband/hw/hfi1/Makefile
+++ b/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 \
diff --git a/drivers/infiniband/hw/hfi1/hfi.h b/drivers/infiniband/hw/hfi1/hfi.h
index ddfcf2f..9aa0357 100644
--- a/drivers/infiniband/hw/hfi1/hfi.h
+++ b/drivers/infiniband/hw/hfi1/hfi.h
@@ -99,6 +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) \
diff --git a/drivers/infiniband/hw/hfi1/init.c b/drivers/infiniband/hw/hfi1/init.c
index f1d0c95..d772a01 100644
--- a/drivers/infiniband/hw/hfi1/init.c
+++ b/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
  */
diff --git a/drivers/infiniband/hw/hfi1/opfn.c b/drivers/infiniband/hw/hfi1/opfn.c
new file mode 100644
index 0000000..2d46c91
--- /dev/null
+++ b/drivers/infiniband/hw/hfi1/opfn.c
@@ -0,0 +1,304 @@ 
+// 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;
+
+	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;
+	}
+
+	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;
+	return;
+err:
+	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;
+
+	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;
+
+	capcode = data & 0xf;
+	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;
+	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;
+
+	capcode = data & 0xf;
+	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;
+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;
+
+	/*
+	 * 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;
+
+	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 (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;
+	}
+}
diff --git a/drivers/infiniband/hw/hfi1/opfn.h b/drivers/infiniband/hw/hfi1/opfn.h
index 1927c98..5f2011c 100644
--- a/drivers/infiniband/hw/hfi1/opfn.h
+++ b/drivers/infiniband/hw/hfi1/opfn.h
@@ -49,10 +49,37 @@ 
 
 /* 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 */