From: Kaike Wan <kaike.wan@xxxxxxxxx>
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@xxxxxxxxx>
Signed-off-by: Ashutosh Dixit <ashutosh.dixit@xxxxxxxxx>
Signed-off-by: Mitko Haralanov <mitko.haralanov@xxxxxxxxx>
Signed-off-by: Kaike Wan <kaike.wan@xxxxxxxxx>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@xxxxxxxxx>
---
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
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 */
