IB/mlx4: Handle new FW requirement for send request prefetching

New ConnectX firmware introduces FW command interface revision 2,
which requires that for each QP, a chunk of send queue entries (the
"headroom") is kept marked as invalid, so that the HCA doesn't get
confused if it prefetches entries that haven't been posted yet.  Add
code to the driver to do this, and also update the user ABI so that
userspace can request that the prefetcher be turned off for userspace
QPs (we just leave the prefetcher on for all kernel QPs).

Unfortunately, marking send queue entries this way is confuses older
firmware, so we change the driver to allow only FW command interface
revisions 2.  This means that users will have to update their firmware
to work with the new driver, but the firmware is changing quickly and
the old firmware has lots of other bugs anyway, so this shouldn't be too
big a deal.

Based on a patch from Jack Morgenstein <jackm@dev.mellanox.co.il>.

Signed-off-by: Roland Dreier <rolandd@cisco.com>

drivers/infiniband/hw/mlx4/cq.c

@@ -355,7 +355,7 @@ static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
 		wq = &(*cur_qp)->sq;
 		wqe_ctr = be16_to_cpu(cqe->wqe_index);
 		wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
-		wc->wr_id = wq->wrid[wq->tail & (wq->max - 1)];
+		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
 		++wq->tail;
 	} else if ((*cur_qp)->ibqp.srq) {
 		srq = to_msrq((*cur_qp)->ibqp.srq);
@@ -364,7 +364,7 @@ static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
 		mlx4_ib_free_srq_wqe(srq, wqe_ctr);
 	} else {
 		wq	  = &(*cur_qp)->rq;
-		wc->wr_id = wq->wrid[wq->tail & (wq->max - 1)];
+		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
 		++wq->tail;
 	}
 

drivers/infiniband/hw/mlx4/mlx4_ib.h

@@ -95,7 +95,8 @@ struct mlx4_ib_mr {
 struct mlx4_ib_wq {
 	u64		       *wrid;
 	spinlock_t		lock;
-	int			max;
+	int			wqe_cnt;
+	int			max_post;
 	int			max_gs;
 	int			offset;
 	int			wqe_shift;
@@ -113,6 +114,7 @@ struct mlx4_ib_qp {
 
 	u32			doorbell_qpn;
 	__be32			sq_signal_bits;
+	int			sq_spare_wqes;
 	struct mlx4_ib_wq	sq;
 
 	struct ib_umem	       *umem;
@@ -123,6 +125,7 @@ struct mlx4_ib_qp {
 	u8			alt_port;
 	u8			atomic_rd_en;
 	u8			resp_depth;
+	u8			sq_no_prefetch;
 	u8			state;
 };
 

drivers/infiniband/hw/mlx4/qp.c

@@ -109,6 +109,20 @@ static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
 	return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
 }
 
+/*
+ * Stamp a SQ WQE so that it is invalid if prefetched by marking the
+ * first four bytes of every 64 byte chunk with 0xffffffff, except for
+ * the very first chunk of the WQE.
+ */
+static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n)
+{
+	u32 *wqe = get_send_wqe(qp, n);
+	int i;
+
+	for (i = 16; i < 1 << (qp->sq.wqe_shift - 2); i += 16)
+		wqe[i] = 0xffffffff;
+}
+
 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
 {
 	struct ib_event event;
@@ -201,18 +215,18 @@ static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
 		if (cap->max_recv_wr)
 			return -EINVAL;
 
-		qp->rq.max = qp->rq.max_gs = 0;
+		qp->rq.wqe_cnt = qp->rq.max_gs = 0;
 	} else {
 		/* HW requires >= 1 RQ entry with >= 1 gather entry */
 		if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
 			return -EINVAL;
 
-		qp->rq.max	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
+		qp->rq.wqe_cnt	 = roundup_pow_of_two(max(1U, cap->max_recv_wr));
 		qp->rq.max_gs	 = roundup_pow_of_two(max(1U, cap->max_recv_sge));
 		qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
 	}
 
-	cap->max_recv_wr  = qp->rq.max;
+	cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
 	cap->max_recv_sge = qp->rq.max_gs;
 
 	return 0;
@@ -236,8 +250,6 @@ static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
 	    cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
 		return -EINVAL;
 
-	qp->sq.max = cap->max_send_wr ? roundup_pow_of_two(cap->max_send_wr) : 1;
-
 	qp->sq.wqe_shift = ilog2(roundup_pow_of_two(max(cap->max_send_sge *
 							sizeof (struct mlx4_wqe_data_seg),
 							cap->max_inline_data +
@@ -246,18 +258,25 @@ static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
 	qp->sq.max_gs    = ((1 << qp->sq.wqe_shift) - send_wqe_overhead(type)) /
 		sizeof (struct mlx4_wqe_data_seg);
 
-	qp->buf_size = (qp->rq.max << qp->rq.wqe_shift) +
-		(qp->sq.max << qp->sq.wqe_shift);
+	/*
+	 * We need to leave 2 KB + 1 WQE of headroom in the SQ to
+	 * allow HW to prefetch.
+	 */
+	qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + 1;
+	qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr + qp->sq_spare_wqes);
+
+	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
+		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
 	if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
 		qp->rq.offset = 0;
-		qp->sq.offset = qp->rq.max << qp->rq.wqe_shift;
+		qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
 	} else {
-		qp->rq.offset = qp->sq.max << qp->sq.wqe_shift;
+		qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
 		qp->sq.offset = 0;
 	}
 
-	cap->max_send_wr     = qp->sq.max;
-	cap->max_send_sge    = qp->sq.max_gs;
+	cap->max_send_wr  = qp->sq.max_post = qp->sq.wqe_cnt - qp->sq_spare_wqes;
+	cap->max_send_sge = qp->sq.max_gs;
 	cap->max_inline_data = (1 << qp->sq.wqe_shift) - send_wqe_overhead(type) -
 		sizeof (struct mlx4_wqe_inline_seg);
 
@@ -267,11 +286,11 @@ static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
 static int set_user_sq_size(struct mlx4_ib_qp *qp,
 			    struct mlx4_ib_create_qp *ucmd)
 {
-	qp->sq.max       = 1 << ucmd->log_sq_bb_count;
+	qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
 	qp->sq.wqe_shift = ucmd->log_sq_stride;
 
-	qp->buf_size = (qp->rq.max << qp->rq.wqe_shift) +
-		(qp->sq.max << qp->sq.wqe_shift);
+	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
+		(qp->sq.wqe_cnt << qp->sq.wqe_shift);
 
 	return 0;
 }
@@ -307,6 +326,8 @@ static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
 			goto err;
 		}
 
+		qp->sq_no_prefetch = ucmd.sq_no_prefetch;
+
 		err = set_user_sq_size(qp, &ucmd);
 		if (err)
 			goto err;
@@ -334,6 +355,8 @@ static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
 				goto err_mtt;
 		}
 	} else {
+		qp->sq_no_prefetch = 0;
+
 		err = set_kernel_sq_size(dev, &init_attr->cap, init_attr->qp_type, qp);
 		if (err)
 			goto err;
@@ -360,8 +383,8 @@ static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
 		if (err)
 			goto err_mtt;
 
-		qp->sq.wrid  = kmalloc(qp->sq.max * sizeof (u64), GFP_KERNEL);
-		qp->rq.wrid  = kmalloc(qp->rq.max * sizeof (u64), GFP_KERNEL);
+		qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), GFP_KERNEL);
+		qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), GFP_KERNEL);
 
 		if (!qp->sq.wrid || !qp->rq.wrid) {
 			err = -ENOMEM;
@@ -743,14 +766,17 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
 		context->mtu_msgmax = (attr->path_mtu << 5) | 31;
 	}
 
-	if (qp->rq.max)
-		context->rq_size_stride = ilog2(qp->rq.max) << 3;
+	if (qp->rq.wqe_cnt)
+		context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
 	context->rq_size_stride |= qp->rq.wqe_shift - 4;
 
-	if (qp->sq.max)
-		context->sq_size_stride = ilog2(qp->sq.max) << 3;
+	if (qp->sq.wqe_cnt)
+		context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
 	context->sq_size_stride |= qp->sq.wqe_shift - 4;
 
+	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
+		context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
+
 	if (qp->ibqp.uobject)
 		context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
 	else
@@ -884,16 +910,19 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
 
 	/*
 	 * Before passing a kernel QP to the HW, make sure that the
-	 * ownership bits of the send queue are set so that the
-	 * hardware doesn't start processing stale work requests.
+	 * ownership bits of the send queue are set and the SQ
+	 * headroom is stamped so that the hardware doesn't start
+	 * processing stale work requests.
 	 */
 	if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
 		struct mlx4_wqe_ctrl_seg *ctrl;
 		int i;
 
-		for (i = 0; i < qp->sq.max; ++i) {
+		for (i = 0; i < qp->sq.wqe_cnt; ++i) {
 			ctrl = get_send_wqe(qp, i);
 			ctrl->owner_opcode = cpu_to_be32(1 << 31);
+
+			stamp_send_wqe(qp, i);
 		}
 	}
 
@@ -1124,15 +1153,15 @@ static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq
 	struct mlx4_ib_cq *cq;
 
 	cur = wq->head - wq->tail;
-	if (likely(cur + nreq < wq->max))
+	if (likely(cur + nreq < wq->max_post))
 		return 0;
 
 	cq = to_mcq(ib_cq);
 	spin_lock(&cq->lock);
 	cur = wq->head - wq->tail;
 	spin_unlock(&cq->lock);
 
-	return cur + nreq >= wq->max;
+	return cur + nreq >= wq->max_post;
 }
 
 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
@@ -1165,8 +1194,8 @@ int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 			goto out;
 		}
 
-		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.max - 1));
-		qp->sq.wrid[ind & (qp->sq.max - 1)] = wr->wr_id;
+		ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
+		qp->sq.wrid[ind & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
 
 		ctrl->srcrb_flags =
 			(wr->send_flags & IB_SEND_SIGNALED ?
@@ -1301,7 +1330,16 @@ int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 		}
 
 		ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
-			(ind & qp->sq.max ? cpu_to_be32(1 << 31) : 0);
+			(ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
+
+		/*
+		 * We can improve latency by not stamping the last
+		 * send queue WQE until after ringing the doorbell, so
+		 * only stamp here if there are still more WQEs to post.
+		 */
+		if (wr->next)
+			stamp_send_wqe(qp, (ind + qp->sq_spare_wqes) &
+				       (qp->sq.wqe_cnt - 1));
 
 		++ind;
 	}
@@ -1324,6 +1362,9 @@ int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 		 * and reach the HCA out of order.
 		 */
 		mmiowb();
+
+		stamp_send_wqe(qp, (ind + qp->sq_spare_wqes - 1) &
+			       (qp->sq.wqe_cnt - 1));
 	}
 
 	spin_unlock_irqrestore(&qp->rq.lock, flags);
@@ -1344,7 +1385,7 @@ int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 
 	spin_lock_irqsave(&qp->rq.lock, flags);
 
-	ind = qp->rq.head & (qp->rq.max - 1);
+	ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
 
 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
 		if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.send_cq)) {
@@ -1375,7 +1416,7 @@ int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 
 		qp->rq.wrid[ind] = wr->wr_id;
 
-		ind = (ind + 1) & (qp->rq.max - 1);
+		ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
 	}
 
 out:

drivers/infiniband/hw/mlx4/user.h

@@ -39,7 +39,7 @@
  * Increment this value if any changes that break userspace ABI
  * compatibility are made.
  */
-#define MLX4_IB_UVERBS_ABI_VERSION	2
+#define MLX4_IB_UVERBS_ABI_VERSION	3
 
 /*
  * Make sure that all structs defined in this file remain laid out so
@@ -87,9 +87,10 @@ struct mlx4_ib_create_srq_resp {
 struct mlx4_ib_create_qp {
 	__u64	buf_addr;
 	__u64	db_addr;
-        __u8	log_sq_bb_count;
-        __u8	log_sq_stride;
-        __u8	reserved[6];
+	__u8	log_sq_bb_count;
+	__u8	log_sq_stride;
+	__u8	sq_no_prefetch;
+	__u8	reserved[5];
 };
 
 #endif /* MLX4_IB_USER_H */

drivers/net/mlx4/fw.c

@@ -38,7 +38,7 @@
 #include "icm.h"
 
 enum {
-	MLX4_COMMAND_INTERFACE_REV	= 1
+	MLX4_COMMAND_INTERFACE_REV	= 2,
 };
 
 extern void __buggy_use_of_MLX4_GET(void);