RDMA/cxgb4: Support variable sized work requests
T4 EQ entries are in multiples of 64 bytes. Currently the RDMA SQ and
RQ use fixed sized entries composed of 4 EQ entries for the SQ and 2
EQ entries for the RQ. For optimial latency with small IO, we need to
change this so the HW only needs to DMA the EQ entries actually used
by a given work request.
Implementation:
- add wq_pidx counter to track where we are in the EQ. cidx/pidx are
used for the sw sq/rq tracking and flow control.
- the variable part of work requests is the SGL. Add new functions to
build the SGL and/or immediate data directly in the EQ memory
wrapping when needed.
- adjust the min burst size for the EQ contexts to 64B.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
diff --git a/drivers/infiniband/hw/cxgb4/t4.h b/drivers/infiniband/hw/cxgb4/t4.h
index 9cf8d85..aef55f4 100644
--- a/drivers/infiniband/hw/cxgb4/t4.h
+++ b/drivers/infiniband/hw/cxgb4/t4.h
@@ -65,10 +65,10 @@
u8 db_off;
};
-#define T4_EQ_SIZE 64
+#define T4_EQ_ENTRY_SIZE 64
#define T4_SQ_NUM_SLOTS 4
-#define T4_SQ_NUM_BYTES (T4_EQ_SIZE * T4_SQ_NUM_SLOTS)
+#define T4_SQ_NUM_BYTES (T4_EQ_ENTRY_SIZE * T4_SQ_NUM_SLOTS)
#define T4_MAX_SEND_SGE ((T4_SQ_NUM_BYTES - sizeof(struct fw_ri_send_wr) - \
sizeof(struct fw_ri_isgl)) / sizeof(struct fw_ri_sge))
#define T4_MAX_SEND_INLINE ((T4_SQ_NUM_BYTES - sizeof(struct fw_ri_send_wr) - \
@@ -84,7 +84,7 @@
#define T4_MAX_FR_DEPTH (T4_MAX_FR_IMMD / sizeof(u64))
#define T4_RQ_NUM_SLOTS 2
-#define T4_RQ_NUM_BYTES (T4_EQ_SIZE * T4_RQ_NUM_SLOTS)
+#define T4_RQ_NUM_BYTES (T4_EQ_ENTRY_SIZE * T4_RQ_NUM_SLOTS)
#define T4_MAX_RECV_SGE 4
union t4_wr {
@@ -97,20 +97,18 @@
struct fw_ri_fr_nsmr_wr fr;
struct fw_ri_inv_lstag_wr inv;
struct t4_status_page status;
- __be64 flits[T4_EQ_SIZE / sizeof(__be64) * T4_SQ_NUM_SLOTS];
+ __be64 flits[T4_EQ_ENTRY_SIZE / sizeof(__be64) * T4_SQ_NUM_SLOTS];
};
union t4_recv_wr {
struct fw_ri_recv_wr recv;
struct t4_status_page status;
- __be64 flits[T4_EQ_SIZE / sizeof(__be64) * T4_RQ_NUM_SLOTS];
+ __be64 flits[T4_EQ_ENTRY_SIZE / sizeof(__be64) * T4_RQ_NUM_SLOTS];
};
static inline void init_wr_hdr(union t4_wr *wqe, u16 wrid,
enum fw_wr_opcodes opcode, u8 flags, u8 len16)
{
- int slots_used;
-
wqe->send.opcode = (u8)opcode;
wqe->send.flags = flags;
wqe->send.wrid = wrid;
@@ -118,12 +116,6 @@
wqe->send.r1[1] = 0;
wqe->send.r1[2] = 0;
wqe->send.len16 = len16;
-
- slots_used = DIV_ROUND_UP(len16*16, T4_EQ_SIZE);
- while (slots_used < T4_SQ_NUM_SLOTS) {
- wqe->flits[slots_used * T4_EQ_SIZE / sizeof(__be64)] = 0;
- slots_used++;
- }
}
/* CQE/AE status codes */
@@ -289,6 +281,7 @@
u16 size;
u16 cidx;
u16 pidx;
+ u16 wq_pidx;
};
struct t4_swrqe {
@@ -310,6 +303,7 @@
u16 size;
u16 cidx;
u16 pidx;
+ u16 wq_pidx;
};
struct t4_wq {
@@ -340,11 +334,14 @@
return wq->rq.size - 1 - wq->rq.in_use;
}
-static inline void t4_rq_produce(struct t4_wq *wq)
+static inline void t4_rq_produce(struct t4_wq *wq, u8 len16)
{
wq->rq.in_use++;
if (++wq->rq.pidx == wq->rq.size)
wq->rq.pidx = 0;
+ wq->rq.wq_pidx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
+ if (wq->rq.wq_pidx >= wq->rq.size * T4_RQ_NUM_SLOTS)
+ wq->rq.wq_pidx %= wq->rq.size * T4_RQ_NUM_SLOTS;
}
static inline void t4_rq_consume(struct t4_wq *wq)
@@ -370,11 +367,14 @@
return wq->sq.size - 1 - wq->sq.in_use;
}
-static inline void t4_sq_produce(struct t4_wq *wq)
+static inline void t4_sq_produce(struct t4_wq *wq, u8 len16)
{
wq->sq.in_use++;
if (++wq->sq.pidx == wq->sq.size)
wq->sq.pidx = 0;
+ wq->sq.wq_pidx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
+ if (wq->sq.wq_pidx >= wq->sq.size * T4_SQ_NUM_SLOTS)
+ wq->sq.wq_pidx %= wq->sq.size * T4_SQ_NUM_SLOTS;
}
static inline void t4_sq_consume(struct t4_wq *wq)
@@ -386,14 +386,12 @@
static inline void t4_ring_sq_db(struct t4_wq *wq, u16 inc)
{
- inc *= T4_SQ_NUM_SLOTS;
wmb();
writel(QID(wq->sq.qid) | PIDX(inc), wq->db);
}
static inline void t4_ring_rq_db(struct t4_wq *wq, u16 inc)
{
- inc *= T4_RQ_NUM_SLOTS;
wmb();
writel(QID(wq->rq.qid) | PIDX(inc), wq->db);
}