Commit ec8607a6 authored by Junxian Huang's avatar Junxian Huang
Browse files

RDMA/hns: Fix mbox timing out by adding retry mechanism 

maillist inclusion
category: bugfix
bugzilla: https://gitee.com/openeuler/kernel/issues/IBSLDS
CVE: NA

Reference: https://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma.git/commit/?h=wip/leon-for-rc&id=9747c0c7791d4a5a62018a0c9c563dd2e6f6c1c0



----------------------------------------------------------------------

If a QP is modified to error state and a flush CQE process is triggered,
the subsequent QP destruction mbox can still be successfully posted but
will be blocked in HW until the flush CQE process finishes. This causes
further mbox posting timeouts in driver. The blocking time is related
to QP depth. Considering an extreme case where SQ depth and RQ depth
are both 32K, the blocking time can reach about 135ms.

This patch adds a retry mechanism for mbox posting. For each try, FW
waits 15ms for HW to complete the previous mbox, otherwise return a
timeout error code to driver. Counting other time consumption in FW,
set 8 tries for mbox posting and a 5ms time gap before each retry to
increase to a sufficient timeout limit.

Fixes: 0425e3e6 ("RDMA/hns: Support flush cqe for hip08 in kernel space")
Signed-off-by: default avatarJunxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20250208105930.522796-1-huangjunxian6@hisilicon.com


Signed-off-by: default avatarLeon Romanovsky <leon@kernel.org>
Signed-off-by: default avatarXinghai Cen <cenxinghai@h-partners.com>
parent 5d94761f

drivers/infiniband/hw/hns/hns_roce_hw_v2.c

+48 −16
Original line number Diff line number Diff line
@@ -1395,10 +1395,8 @@ static u32 hns_roce_cmdq_tx_timeout(u16 opcode, u32 tx_timeout) 
	return tx_timeout;

}



static void hns_roce_wait_csq_done(struct hns_roce_dev *hr_dev, u16 opcode)

static void hns_roce_wait_csq_done(struct hns_roce_dev *hr_dev, u32 tx_timeout)

{

	struct hns_roce_v2_priv *priv = hr_dev->priv;

	u32 tx_timeout = hns_roce_cmdq_tx_timeout(opcode, priv->cmq.tx_timeout);

	u32 timeout = 0;



	do {
@@ -1408,8 +1406,9 @@ static void hns_roce_wait_csq_done(struct hns_roce_dev *hr_dev, u16 opcode) 
	} while (++timeout < tx_timeout);

}



static int __hns_roce_cmq_send(struct hns_roce_dev *hr_dev,

			       struct hns_roce_cmq_desc *desc, int num)

static int __hns_roce_cmq_send_one(struct hns_roce_dev *hr_dev,

				   struct hns_roce_cmq_desc *desc,

				   int num, u32 tx_timeout)

{

	struct hns_roce_v2_priv *priv = hr_dev->priv;

	struct hns_roce_v2_cmq_ring *csq = &priv->cmq.csq;
@@ -1418,8 +1417,6 @@ static int __hns_roce_cmq_send(struct hns_roce_dev *hr_dev, 
	int ret;

	int i;



	spin_lock_bh(&csq->lock);



	tail = csq->head;



	for (i = 0; i < num; i++) {
@@ -1433,22 +1430,17 @@ static int __hns_roce_cmq_send(struct hns_roce_dev *hr_dev, 


	atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CMDS_CNT]);



	hns_roce_wait_csq_done(hr_dev, le16_to_cpu(desc->opcode));

	hns_roce_wait_csq_done(hr_dev, tx_timeout);

	if (hns_roce_cmq_csq_done(hr_dev)) {

		ret = 0;

		for (i = 0; i < num; i++) {

			/* check the result of hardware write back */

			desc[i] = csq->desc[tail++];

			desc_ret = le16_to_cpu(csq->desc[tail++].retval);

			if (tail == csq->desc_num)

				tail = 0;



			desc_ret = le16_to_cpu(desc[i].retval);

			if (likely(desc_ret == CMD_EXEC_SUCCESS))

				continue;



			dev_err_ratelimited(hr_dev->dev,

					    "Cmdq IO error, opcode = 0x%x, return = 0x%x.\n",

					    desc->opcode, desc_ret);

			ret = hns_roce_cmd_err_convert_errno(desc_ret);

		}

	} else {
@@ -1463,10 +1455,50 @@ static int __hns_roce_cmq_send(struct hns_roce_dev *hr_dev, 
		ret = -EAGAIN;

	}



	if (ret)

		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CMDS_ERR_CNT]);



	return ret;

}



static int __hns_roce_cmq_send(struct hns_roce_dev *hr_dev,

			       struct hns_roce_cmq_desc *desc, int num)

{

	struct hns_roce_v2_priv *priv = hr_dev->priv;

	struct hns_roce_v2_cmq_ring *csq = &priv->cmq.csq;

	u16 opcode = le16_to_cpu(desc->opcode);

	u32 tx_timeout = hns_roce_cmdq_tx_timeout(opcode, priv->cmq.tx_timeout);

	u8 try_cnt = HNS_ROCE_OPC_POST_MB_TRY_CNT;

	u32 rsv_tail;

	int ret;

	int i;



	while (try_cnt) {

		try_cnt--;



		spin_lock_bh(&csq->lock);

		rsv_tail = csq->head;

		ret = __hns_roce_cmq_send_one(hr_dev, desc, num, tx_timeout);

		if (opcode == HNS_ROCE_OPC_POST_MB && ret == -ETIME &&

		    try_cnt) {

			spin_unlock_bh(&csq->lock);

			mdelay(HNS_ROCE_OPC_POST_MB_RETRY_GAP_MSEC);

			continue;

		}



		for (i = 0; i < num; i++) {

			desc[i] = csq->desc[rsv_tail++];

			if (rsv_tail == csq->desc_num)

				rsv_tail = 0;

		}

		spin_unlock_bh(&csq->lock);

		break;

	}



	if (ret)

		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CMDS_ERR_CNT]);

		dev_err_ratelimited(hr_dev->dev,

				    "Cmdq IO error, opcode = 0x%x, return = %d.\n",

				    opcode, ret);



	return ret;

}

drivers/infiniband/hw/hns/hns_roce_hw_v2.h

+2 −0
Original line number Diff line number Diff line
@@ -243,6 +243,8 @@ enum hns_roce_opcode_type { 
};



#define HNS_ROCE_OPC_POST_MB_TIMEOUT 35000

#define HNS_ROCE_OPC_POST_MB_TRY_CNT 8

#define HNS_ROCE_OPC_POST_MB_RETRY_GAP_MSEC 5

struct hns_roce_cmdq_tx_timeout_map {

	u16 opcode;

	u32 tx_timeout;