aboutsummaryrefslogtreecommitdiff
path: root/drivers
diff options
context:
space:
mode:
Diffstat (limited to 'drivers')
-rw-r--r--drivers/infiniband/hw/mlx4/cq.c4
-rw-r--r--drivers/infiniband/hw/mlx4/mlx4_ib.h5
-rw-r--r--drivers/infiniband/hw/mlx4/qp.c101
-rw-r--r--drivers/infiniband/hw/mlx4/user.h9
-rw-r--r--drivers/net/mlx4/fw.c2
5 files changed, 83 insertions, 38 deletions
diff --git a/drivers/infiniband/hw/mlx4/cq.c b/drivers/infiniband/hw/mlx4/cq.c
index 048c5274ab1..e940521e9c8 100644
--- a/drivers/infiniband/hw/mlx4/cq.c
+++ b/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;
}
diff --git a/drivers/infiniband/hw/mlx4/mlx4_ib.h b/drivers/infiniband/hw/mlx4/mlx4_ib.h
index 93dac71f323..24ccadd6e4f 100644
--- a/drivers/infiniband/hw/mlx4/mlx4_ib.h
+++ b/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;
};
diff --git a/drivers/infiniband/hw/mlx4/qp.c b/drivers/infiniband/hw/mlx4/qp.c
index 4c15fa3426b..8fabe0da323 100644
--- a/drivers/infiniband/hw/mlx4/qp.c
+++ b/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);
}
}
@@ -1124,7 +1153,7 @@ 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);
@@ -1132,7 +1161,7 @@ static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq
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 @@ out:
* 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:
diff --git a/drivers/infiniband/hw/mlx4/user.h b/drivers/infiniband/hw/mlx4/user.h
index 88c72d56368..e2d11be4525 100644
--- a/drivers/infiniband/hw/mlx4/user.h
+++ b/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 */
diff --git a/drivers/net/mlx4/fw.c b/drivers/net/mlx4/fw.c
index e7ca118c8df..81fc546a1c4 100644
--- a/drivers/net/mlx4/fw.c
+++ b/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);
generated by cgit v1.2.3 (git 2.25.1) at 2024-11-22 09:53:54 +0000