Re: [PATCH for-next 3/5] RDMA/hns: Optimize the wr opcode conversion from ib to hns

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 2020/3/5 14:18, Leon Romanovsky wrote:
> On Mon, Mar 02, 2020 at 08:11:31PM +0800, Weihang Li wrote:
>> From: Xi Wang <wangxi11@xxxxxxxxxx>
>>
>> Simplify the wr opcode conversion from ib to hns by using a map table
>> instead of the switch-case statement.
>>
>> Signed-off-by: Xi Wang <wangxi11@xxxxxxxxxx>
>> Signed-off-by: Weihang Li <liweihang@xxxxxxxxxx>
>> ---
>>  drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 70 ++++++++++++++++++------------
>>  1 file changed, 43 insertions(+), 27 deletions(-)
>>
>> diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
>> index c8c345f..ea61ccb 100644
>> --- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
>> +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
>> @@ -56,6 +56,47 @@ static void set_data_seg_v2(struct hns_roce_v2_wqe_data_seg *dseg,
>>  	dseg->len  = cpu_to_le32(sg->length);
>>  }
>>
>> +/*
>> + * mapped-value = 1 + real-value
>> + * The hns wr opcode real value is start from 0, In order to distinguish between
>> + * initialized and uninitialized map values, we plus 1 to the actual value when
>> + * defining the mapping, so that the validity can be identified by checking the
>> + * mapped value is greater than 0.
>> + */
>> +#define HR_OPC_MAP(ib_key, hr_key) \
>> +		[IB_WR_ ## ib_key] = 1 + HNS_ROCE_V2_WQE_OP_ ## hr_key
>> +
>> +static const u32 hns_roce_op_code[] = {
>> +	HR_OPC_MAP(RDMA_WRITE,			RDMA_WRITE),
>> +	HR_OPC_MAP(RDMA_WRITE_WITH_IMM,		RDMA_WRITE_WITH_IMM),
>> +	HR_OPC_MAP(SEND,			SEND),
>> +	HR_OPC_MAP(SEND_WITH_IMM,		SEND_WITH_IMM),
>> +	HR_OPC_MAP(RDMA_READ,			RDMA_READ),
>> +	HR_OPC_MAP(ATOMIC_CMP_AND_SWP,		ATOM_CMP_AND_SWAP),
>> +	HR_OPC_MAP(ATOMIC_FETCH_AND_ADD,	ATOM_FETCH_AND_ADD),
>> +	HR_OPC_MAP(SEND_WITH_INV,		SEND_WITH_INV),
>> +	HR_OPC_MAP(LOCAL_INV,			LOCAL_INV),
>> +	HR_OPC_MAP(MASKED_ATOMIC_CMP_AND_SWP,	ATOM_MSK_CMP_AND_SWAP),
>> +	HR_OPC_MAP(MASKED_ATOMIC_FETCH_AND_ADD,	ATOM_MSK_FETCH_AND_ADD),
>> +	HR_OPC_MAP(REG_MR,			FAST_REG_PMR),
>> +	[IB_WR_RESERVED1] = 0,
> 
> hns_roce_op_code[] is declared as static, everything is initialized to
> 0, there is no need to set 0 again.

OK, thank you.

> 
>> +};
>> +
>> +static inline u32 to_hr_opcode(u32 ib_opcode)
> 
> No inline functions in *.c, please.

Hi Leon,

Thanks for your comments.

But I'm confused about when we should use static inline and when we should
use macros if a function is only used in a *.c. A few days ago, Jason
suggested me to use static inline functions, you can check the link below:

https://patchwork.kernel.org/patch/11372851/

Are there any rules about that in kernel or in our rdma subsystem? Should
I use a macro, just remove the keyword "inline" from this definition or
move this definition to .h?

> 
>> +{
>> +	u32 hr_opcode = 0;
>> +
>> +	if (ib_opcode < IB_WR_RESERVED1)
> 
> if (ib_opcode > ARRAY_SIZE(hns_roce_op_code) - 1)
> 	return HNS_ROCE_V2_WQE_OP_MASK;
> 
> return hns_roce_op_code[ib_opcode];
> 

The index of ib_key in hns_roce_op_code[] is not continuous, so there
are some invalid ib_wr_opcode for hns between the valid index.

For hardware of HIP08, HNS_ROCE_V2_WQE_OP_MASK means invalid opcode but
not zero. So we have to check if the ib_wr_opcode has a mapping value in
hns_roce_op_code[], and if the mapping result is zero, we have to return
HNS_ROCE_V2_WQE_OP_MASK. Is it ok like this?

{
	u32 hr_opcode = 0;

	if (ib_opcode > ARRAY_SIZE(hns_roce_op_code) - 1)
 		return HNS_ROCE_V2_WQE_OP_MASK;

	hr_opcode = hns_roce_op_code[ib_opcode];

	/* exist a valid mapping definition for ib code */
	if (hr_opcode > 0)
		return hr_opcode - 1;
	else
		return HNS_ROCE_V2_WQE_OP_MASK;
}

Thanks,
Weihang

> 
>> +		hr_opcode = hns_roce_op_code[ib_opcode];
>> +
>> +	/* exist a valid mapping definition for ib code */
>> +	if (hr_opcode > 0)
>> +		return hr_opcode - 1;
>> +
>> +	/* default hns roce wr opcode */
>> +	return HNS_ROCE_V2_WQE_OP_MASK;
>> +}
>> +
>>  static void set_frmr_seg(struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
>>  			 void *wqe, const struct ib_reg_wr *wr)
>>  {
>> @@ -303,7 +344,6 @@ static int hns_roce_v2_post_send(struct ib_qp *ibqp,
>>  	void *wqe = NULL;
>>  	bool loopback;
>>  	u32 tmp_len;
>> -	u32 hr_op;
>>  	u8 *smac;
>>  	int nreq;
>>  	int ret;
>> @@ -517,76 +557,52 @@ static int hns_roce_v2_post_send(struct ib_qp *ibqp,
>>  			wqe += sizeof(struct hns_roce_v2_rc_send_wqe);
>>  			switch (wr->opcode) {
>>  			case IB_WR_RDMA_READ:
>> -				hr_op = HNS_ROCE_V2_WQE_OP_RDMA_READ;
>>  				rc_sq_wqe->rkey =
>>  					cpu_to_le32(rdma_wr(wr)->rkey);
>>  				rc_sq_wqe->va =
>>  					cpu_to_le64(rdma_wr(wr)->remote_addr);
>>  				break;
>>  			case IB_WR_RDMA_WRITE:
>> -				hr_op = HNS_ROCE_V2_WQE_OP_RDMA_WRITE;
>>  				rc_sq_wqe->rkey =
>>  					cpu_to_le32(rdma_wr(wr)->rkey);
>>  				rc_sq_wqe->va =
>>  					cpu_to_le64(rdma_wr(wr)->remote_addr);
>>  				break;
>>  			case IB_WR_RDMA_WRITE_WITH_IMM:
>> -				hr_op = HNS_ROCE_V2_WQE_OP_RDMA_WRITE_WITH_IMM;
>>  				rc_sq_wqe->rkey =
>>  					cpu_to_le32(rdma_wr(wr)->rkey);
>>  				rc_sq_wqe->va =
>>  					cpu_to_le64(rdma_wr(wr)->remote_addr);
>>  				break;
>> -			case IB_WR_SEND:
>> -				hr_op = HNS_ROCE_V2_WQE_OP_SEND;
>> -				break;
>> -			case IB_WR_SEND_WITH_INV:
>> -				hr_op = HNS_ROCE_V2_WQE_OP_SEND_WITH_INV;
>> -				break;
>> -			case IB_WR_SEND_WITH_IMM:
>> -				hr_op = HNS_ROCE_V2_WQE_OP_SEND_WITH_IMM;
>> -				break;
>>  			case IB_WR_LOCAL_INV:
>> -				hr_op = HNS_ROCE_V2_WQE_OP_LOCAL_INV;
>>  				roce_set_bit(rc_sq_wqe->byte_4,
>>  					       V2_RC_SEND_WQE_BYTE_4_SO_S, 1);
>>  				rc_sq_wqe->inv_key =
>>  					    cpu_to_le32(wr->ex.invalidate_rkey);
>>  				break;
>>  			case IB_WR_REG_MR:
>> -				hr_op = HNS_ROCE_V2_WQE_OP_FAST_REG_PMR;
>>  				set_frmr_seg(rc_sq_wqe, wqe, reg_wr(wr));
>>  				break;
>>  			case IB_WR_ATOMIC_CMP_AND_SWP:
>> -				hr_op = HNS_ROCE_V2_WQE_OP_ATOM_CMP_AND_SWAP;
>>  				rc_sq_wqe->rkey =
>>  					cpu_to_le32(atomic_wr(wr)->rkey);
>>  				rc_sq_wqe->va =
>>  					cpu_to_le64(atomic_wr(wr)->remote_addr);
>>  				break;
>>  			case IB_WR_ATOMIC_FETCH_AND_ADD:
>> -				hr_op = HNS_ROCE_V2_WQE_OP_ATOM_FETCH_AND_ADD;
>>  				rc_sq_wqe->rkey =
>>  					cpu_to_le32(atomic_wr(wr)->rkey);
>>  				rc_sq_wqe->va =
>>  					cpu_to_le64(atomic_wr(wr)->remote_addr);
>>  				break;
>> -			case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
>> -				hr_op =
>> -				       HNS_ROCE_V2_WQE_OP_ATOM_MSK_CMP_AND_SWAP;
>> -				break;
>> -			case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
>> -				hr_op =
>> -				      HNS_ROCE_V2_WQE_OP_ATOM_MSK_FETCH_AND_ADD;
>> -				break;
>>  			default:
>> -				hr_op = HNS_ROCE_V2_WQE_OP_MASK;
>>  				break;
>>  			}
>>
>>  			roce_set_field(rc_sq_wqe->byte_4,
>>  				       V2_RC_SEND_WQE_BYTE_4_OPCODE_M,
>> -				       V2_RC_SEND_WQE_BYTE_4_OPCODE_S, hr_op);
>> +				       V2_RC_SEND_WQE_BYTE_4_OPCODE_S,
>> +				       to_hr_opcode(wr->opcode));
>>
>>  			if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
>>  			    wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
>> --
>> 2.8.1
>>
> 




[Index of Archives]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Photo]     [Yosemite News]     [Yosemite Photos]     [Linux Kernel]     [Linux SCSI]     [XFree86]

  Powered by Linux