@@ -5222,6 +5222,17 @@ 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.
@@ -1,7 +1,7 @@
#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 update_usrhead(struct hfi1_ctxtdata *rcd, u32 hd, u32 updegr, u32 egrhd,
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);
@@ -927,6 +927,8 @@ 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");
@@ -50,4 +50,9 @@
/* STL Verbs Extended */
#define IB_BTHE_E_SHIFT 24
+struct hfi1_opfn_data {
+ /* serialize opfn function calls */
+ spinlock_t lock;
+};
+
#endif /* _HFI1_OPFN_H */
@@ -8,6 +8,208 @@
#include "verbs.h"
#include "tid_rdma.h"
+/*
+ * J_KEY for kernel contexts when TID RDMA is used.
+ * See generate_jkey() in hfi.h for more information.
+ */
+#define TID_RDMA_JKEY 32
+#define HFI1_KERNEL_MIN_JKEY HFI1_ADMIN_JKEY_RANGE
+#define HFI1_KERNEL_MAX_JKEY (2 * HFI1_ADMIN_JKEY_RANGE - 1)
+
+#define TID_RDMA_MAX_READ_SEGS_PER_REQ 6
+#define TID_RDMA_MAX_WRITE_SEGS_PER_REQ 4
+
+#define TID_OPFN_QP_CTXT_MASK 0xff
+#define TID_OPFN_QP_CTXT_SHIFT 56
+#define TID_OPFN_QP_KDETH_MASK 0xff
+#define TID_OPFN_QP_KDETH_SHIFT 48
+#define TID_OPFN_MAX_LEN_MASK 0x7ff
+#define TID_OPFN_MAX_LEN_SHIFT 37
+#define TID_OPFN_TIMEOUT_MASK 0x1f
+#define TID_OPFN_TIMEOUT_SHIFT 32
+#define TID_OPFN_RESERVED_MASK 0x3f
+#define TID_OPFN_RESERVED_SHIFT 26
+#define TID_OPFN_URG_MASK 0x1
+#define TID_OPFN_URG_SHIFT 25
+#define TID_OPFN_VER_MASK 0x7
+#define TID_OPFN_VER_SHIFT 22
+#define TID_OPFN_JKEY_MASK 0x3f
+#define TID_OPFN_JKEY_SHIFT 16
+#define TID_OPFN_MAX_READ_MASK 0x3f
+#define TID_OPFN_MAX_READ_SHIFT 10
+#define TID_OPFN_MAX_WRITE_MASK 0x3f
+#define TID_OPFN_MAX_WRITE_SHIFT 4
+
+/*
+ * OPFN TID layout
+ *
+ * 63 47 31 15
+ * NNNNNNNNKKKKKKKK MMMMMMMMMMMTTTTT DDDDDDUVVVJJJJJJ RRRRRRWWWWWWCCCC
+ * 3210987654321098 7654321098765432 1098765432109876 5432109876543210
+ * N - the context Number
+ * K - the Kdeth_qp
+ * M - Max_len
+ * T - Timeout
+ * D - reserveD
+ * V - version
+ * U - Urg capable
+ * J - Jkey
+ * R - max_Read
+ * W - max_Write
+ * C - Capcode
+ */
+
+static u64 tid_rdma_opfn_encode(struct tid_rdma_params *p)
+{
+ return
+ (((u64)p->qp & TID_OPFN_QP_CTXT_MASK) <<
+ TID_OPFN_QP_CTXT_SHIFT) |
+ ((((u64)p->qp >> 16) & TID_OPFN_QP_KDETH_MASK) <<
+ TID_OPFN_QP_KDETH_SHIFT) |
+ (((u64)((p->max_len >> PAGE_SHIFT) - 1) &
+ TID_OPFN_MAX_LEN_MASK) << TID_OPFN_MAX_LEN_SHIFT) |
+ (((u64)p->timeout & TID_OPFN_TIMEOUT_MASK) <<
+ TID_OPFN_TIMEOUT_SHIFT) |
+ (((u64)p->urg & TID_OPFN_URG_MASK) << TID_OPFN_URG_SHIFT) |
+ (((u64)p->jkey & TID_OPFN_JKEY_MASK) << TID_OPFN_JKEY_SHIFT) |
+ (((u64)p->max_read & TID_OPFN_MAX_READ_MASK) <<
+ TID_OPFN_MAX_READ_SHIFT) |
+ (((u64)p->max_write & TID_OPFN_MAX_WRITE_MASK) <<
+ TID_OPFN_MAX_WRITE_SHIFT);
+}
+
+static void tid_rdma_opfn_decode(struct tid_rdma_params *p, u64 data)
+{
+ p->max_len = (((data >> TID_OPFN_MAX_LEN_SHIFT) &
+ TID_OPFN_MAX_LEN_MASK) + 1) << PAGE_SHIFT;
+ p->jkey = (data >> TID_OPFN_JKEY_SHIFT) & TID_OPFN_JKEY_MASK;
+ p->max_write = (data >> TID_OPFN_MAX_WRITE_SHIFT) &
+ TID_OPFN_MAX_WRITE_MASK;
+ p->max_read = (data >> TID_OPFN_MAX_READ_SHIFT) &
+ TID_OPFN_MAX_READ_MASK;
+ p->qp =
+ ((((data >> TID_OPFN_QP_KDETH_SHIFT) & TID_OPFN_QP_KDETH_MASK)
+ << 16) |
+ ((data >> TID_OPFN_QP_CTXT_SHIFT) & TID_OPFN_QP_CTXT_MASK));
+ p->urg = (data >> TID_OPFN_URG_SHIFT) & TID_OPFN_URG_MASK;
+ p->timeout = (data >> TID_OPFN_TIMEOUT_SHIFT) & TID_OPFN_TIMEOUT_MASK;
+}
+
+void tid_rdma_opfn_init(struct rvt_qp *qp, struct tid_rdma_params *p)
+{
+ struct hfi1_qp_priv *priv = qp->priv;
+
+ p->qp = (kdeth_qp << 16) | priv->rcd->ctxt;
+ p->max_len = TID_RDMA_MAX_SEGMENT_SIZE;
+ p->jkey = priv->rcd->jkey;
+ p->max_read = TID_RDMA_MAX_READ_SEGS_PER_REQ;
+ p->max_write = TID_RDMA_MAX_WRITE_SEGS_PER_REQ;
+ p->timeout = qp->timeout;
+ p->urg = is_urg_masked(priv->rcd);
+}
+
+bool tid_rdma_conn_req(struct rvt_qp *qp, u64 *data)
+{
+ struct hfi1_qp_priv *priv = qp->priv;
+
+ *data = tid_rdma_opfn_encode(&priv->tid_rdma.local);
+ return true;
+}
+
+bool tid_rdma_conn_reply(struct rvt_qp *qp, u64 data)
+{
+ struct hfi1_qp_priv *priv = qp->priv;
+ struct tid_rdma_params *remote, *old;
+ bool ret = true;
+
+ old = rcu_dereference_protected(priv->tid_rdma.remote,
+ lockdep_is_held(&priv->opfn.lock));
+ data &= ~0xfULL;
+ /*
+ * If data passed in is zero, return true so as not to continue the
+ * negotiation process
+ */
+ if (!data || !HFI1_CAP_IS_KSET(TID_RDMA))
+ goto null;
+ /*
+ * If kzalloc fails, return false. This will result in:
+ * * at the requester a new OPFN request being generated to retry
+ * the negotiation
+ * * at the responder, 0 being returned to the requester so as to
+ * disable TID RDMA at both the requester and the responder
+ */
+ remote = kzalloc(sizeof(*remote), GFP_ATOMIC);
+ if (!remote) {
+ ret = false;
+ goto null;
+ }
+
+ tid_rdma_opfn_decode(remote, data);
+ priv->tid_timer_timeout_jiffies =
+ usecs_to_jiffies((((4096UL * (1UL << remote->timeout)) /
+ 1000UL) << 3) * 7);
+ rcu_assign_pointer(priv->tid_rdma.remote, remote);
+ /*
+ * A TID RDMA READ request's segment size is not equal to
+ * remote->max_len only when the request's data length is smaller
+ * than remote->max_len. In that case, there will be only one segment.
+ * Therefore, when priv->pkts_ps is used to calculate req->cur_seg
+ * during retry, it will lead to req->cur_seg = 0, which is exactly
+ * what is expected.
+ */
+ priv->pkts_ps = (u16)rvt_div_mtu(qp, remote->max_len);
+ priv->timeout_shift = ilog2(priv->pkts_ps - 1) + 1;
+ goto free;
+null:
+ RCU_INIT_POINTER(priv->tid_rdma.remote, NULL);
+ priv->timeout_shift = 0;
+free:
+ if (old)
+ kfree_rcu(old, rcu_head);
+ return ret;
+}
+
+bool tid_rdma_conn_resp(struct rvt_qp *qp, u64 *data)
+{
+ bool ret;
+
+ ret = tid_rdma_conn_reply(qp, *data);
+ *data = 0;
+ /*
+ * If tid_rdma_conn_reply() returns error, set *data as 0 to indicate
+ * TID RDMA could not be enabled. This will result in TID RDMA being
+ * disabled at the requester too.
+ */
+ if (ret)
+ (void)tid_rdma_conn_req(qp, data);
+ return ret;
+}
+
+void tid_rdma_conn_error(struct rvt_qp *qp)
+{
+ struct hfi1_qp_priv *priv = qp->priv;
+ struct tid_rdma_params *old;
+
+ old = rcu_dereference_protected(priv->tid_rdma.remote,
+ lockdep_is_held(&priv->opfn.lock));
+ RCU_INIT_POINTER(priv->tid_rdma.remote, NULL);
+ if (old)
+ kfree_rcu(old, rcu_head);
+}
+
+/* This is called at context initialization time */
+int hfi1_kern_exp_rcv_init(struct hfi1_ctxtdata *rcd, int reinit)
+{
+ if (reinit)
+ return 0;
+
+ BUILD_BUG_ON(TID_RDMA_JKEY < HFI1_KERNEL_MIN_JKEY);
+ BUILD_BUG_ON(TID_RDMA_JKEY > HFI1_KERNEL_MAX_JKEY);
+ rcd->jkey = TID_RDMA_JKEY;
+ hfi1_set_ctxt_jkey(rcd->dd, rcd, rcd->jkey);
+ return 0;
+}
+
/**
* qp_to_rcd - determine the receive context used by a qp
* @qp - the qp
@@ -6,8 +6,34 @@
#ifndef HFI1_TID_RDMA_H
#define HFI1_TID_RDMA_H
+#define TID_RDMA_MAX_SEGMENT_SIZE BIT(18) /* 256 KiB (for now) */
+
+struct tid_rdma_params {
+ struct rcu_head rcu_head;
+ u32 qp;
+ u32 max_len;
+ u16 jkey;
+ u8 max_read;
+ u8 max_write;
+ u8 timeout;
+ u8 urg;
+ u8 version;
+};
+
+struct tid_rdma_qp_params {
+ struct tid_rdma_params local;
+ struct tid_rdma_params __rcu *remote;
+};
+
+bool tid_rdma_conn_req(struct rvt_qp *qp, u64 *data);
+bool tid_rdma_conn_reply(struct rvt_qp *qp, u64 data);
+bool tid_rdma_conn_resp(struct rvt_qp *qp, u64 *data);
+void tid_rdma_conn_error(struct rvt_qp *qp);
+void tid_rdma_opfn_init(struct rvt_qp *qp, struct tid_rdma_params *p);
+
+int hfi1_kern_exp_rcv_init(struct hfi1_ctxtdata *rcd, int reinit);
+
int hfi1_qp_priv_init(struct rvt_dev_info *rdi, struct rvt_qp *qp,
struct ib_qp_init_attr *init_attr);
#endif /* HFI1_TID_RDMA_H */
-
@@ -161,8 +161,13 @@ struct hfi1_qp_priv {
struct hfi1_ctxtdata *rcd; /* QP's receive context */
u8 s_sc; /* SC[0..4] for next packet */
struct iowait s_iowait;
+ struct hfi1_opfn_data opfn;
+ struct tid_rdma_qp_params tid_rdma;
struct rvt_qp *owner;
u8 hdr_type; /* 9B or 16B */
+ unsigned long tid_timer_timeout_jiffies;
+ u16 pkts_ps; /* packets per segment */
+ u8 timeout_shift; /* account for number of packets per segment */
};
/*