On Thu, Mar 16, 2017 at 12:45 AM, Leon Romanovsky <leon@xxxxxxxxxx> wrote:
> On Wed, Mar 15, 2017 at 06:37:32AM -0400, Devesh Sharma wrote:
>> This patch adds support for compare-and-swap and fetch-and-add
>> atomic operations in user library.
>>
>> v1->v2
>> -- Fixed the missing "break"
>> -- Changed macros to inline function
>>
>> Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@xxxxxxxxxxxx>
>> Signed-off-by: Somnath Kotur <somnath.kotur@xxxxxxxxxxxx>
>> Signed-off-by: Selvin Xavier <selvin.xavier@xxxxxxxxxxxx>
>> Signed-off-by: Devesh Sharma <devesh.sharma@xxxxxxxxxxxx>
>> ---
>> providers/bnxt_re/bnxt_re-abi.h | 3 ++-
>> providers/bnxt_re/main.h | 8 +++++-
>> providers/bnxt_re/memory.h | 10 +++++++
>> providers/bnxt_re/verbs.c | 58 +++++++++++++++++++++++++++++++++++------
>> 4 files changed, 69 insertions(+), 10 deletions(-)
>>
>> diff --git a/providers/bnxt_re/bnxt_re-abi.h b/providers/bnxt_re/bnxt_re-abi.h
>> index 581e1b7..557221b 100644
>> --- a/providers/bnxt_re/bnxt_re-abi.h
>> +++ b/providers/bnxt_re/bnxt_re-abi.h
>> @@ -54,7 +54,8 @@ enum bnxt_re_wr_opcode {
>> BNXT_RE_WR_OPCD_ATOMIC_FA = 0x0B,
>> BNXT_RE_WR_OPCD_LOC_INVAL = 0x0C,
>> BNXT_RE_WR_OPCD_BIND = 0x0E,
>> - BNXT_RE_WR_OPCD_RECV = 0x80
>> + BNXT_RE_WR_OPCD_RECV = 0x80,
>> + BNXT_RE_WR_OPCD_INVAL = 0xFF
>> };
>>
>> enum bnxt_re_wr_flags {
>> diff --git a/providers/bnxt_re/main.h b/providers/bnxt_re/main.h
>> index 9d99c64..a417328 100644
>> --- a/providers/bnxt_re/main.h
>> +++ b/providers/bnxt_re/main.h
>> @@ -236,9 +236,15 @@ static inline uint8_t bnxt_re_ibv_to_bnxt_wr_opcd(uint8_t ibv_opcd)
>> case IBV_WR_RDMA_READ:
>> bnxt_opcd = BNXT_RE_WR_OPCD_RDMA_READ;
>> break;
>> + case IBV_WR_ATOMIC_CMP_AND_SWP:
>> + bnxt_opcd = BNXT_RE_WR_OPCD_ATOMIC_CS;
>> + break;
>> + case IBV_WR_ATOMIC_FETCH_AND_ADD:
>> + bnxt_opcd = BNXT_RE_WR_OPCD_ATOMIC_FA;
>> + break;
>> /* TODO: Add other opcodes */
>> default:
>> - bnxt_opcd = 0xFF;
>> + bnxt_opcd = BNXT_RE_WR_OPCD_INVAL;
>> break;
>> };
>>
>> diff --git a/providers/bnxt_re/memory.h b/providers/bnxt_re/memory.h
>> index debb31a..d7e6a92 100644
>> --- a/providers/bnxt_re/memory.h
>> +++ b/providers/bnxt_re/memory.h
>> @@ -83,6 +83,16 @@ static inline void iowrite32(__u32 *dst, uint32_t *src)
>> *(volatile __u32 *)dst = *src;
>> }
>>
>> +static inline __u32 upper_32_bits(uint64_t n)
>> +{
>> + return (__u32)((n >> 16) >> 16);
>> +}
>> +
>> +static inline __u32 lower_32_bits(uint64_t n)
>> +{
>> + return (__u32)(n & 0xFFFFFFFFUL);
>> +}
>> +
>> /* Basic queue operation */
>> static inline uint32_t bnxt_re_is_que_full(struct bnxt_re_queue *que)
>> {
>> diff --git a/providers/bnxt_re/verbs.c b/providers/bnxt_re/verbs.c
>> index e60010d..85d77cd 100644
>> --- a/providers/bnxt_re/verbs.c
>> +++ b/providers/bnxt_re/verbs.c
>> @@ -1068,6 +1068,9 @@ static int bnxt_re_build_send_sqe(struct bnxt_re_qp *qp, void *wqe,
>>
>> /* Fill Header */
>> opcode = bnxt_re_ibv_to_bnxt_wr_opcd(wr->opcode);
>> + if (opcode == BNXT_RE_WR_OPCD_INVAL)
>> + return -EINVAL;
>> +
>> hdr->rsv_ws_fl_wt |= (opcode & BNXT_RE_HDR_WT_MASK);
>>
>> if (is_inline) {
>> @@ -1116,6 +1119,44 @@ static int bnxt_re_build_rdma_sqe(struct bnxt_re_qp *qp, void *wqe,
>> return len;
>> }
>>
>> +static int bnxt_re_build_cns_sqe(struct bnxt_re_qp *qp, void *wqe,
>> + struct ibv_send_wr *wr)
>> +{
>> + struct bnxt_re_bsqe *hdr = wqe;
>> + struct bnxt_re_atomic *sqe = ((void *)wqe +
>> + sizeof(struct bnxt_re_bsqe));
>> + int len;
>> +
>> + len = bnxt_re_build_send_sqe(qp, wqe, wr, false);
>> + hdr->key_immd = wr->wr.atomic.rkey;
>> + sqe->rva_lo = lower_32_bits(wr->wr.atomic.remote_addr);
>> + sqe->rva_hi = upper_32_bits(wr->wr.atomic.remote_addr);
>> + sqe->cmp_dt_lo = lower_32_bits(wr->wr.atomic.compare_add);
>> + sqe->cmp_dt_hi = upper_32_bits(wr->wr.atomic.compare_add);
>> + sqe->swp_dt_lo = lower_32_bits(wr->wr.atomic.swap);
>> + sqe->swp_dt_hi = upper_32_bits(wr->wr.atomic.swap);
>> +
>
> Sorry for my question, but why do you need to separate uint64_t variable
> to two __u32? Why can't you declare one __u64?
It may help 32 bit CPU during byte order conversion.
>
>> + return len;
>> +}
>> +
>> +static int bnxt_re_build_fna_sqe(struct bnxt_re_qp *qp, void *wqe,
>> + struct ibv_send_wr *wr)
>> +{
>> + struct bnxt_re_bsqe *hdr = wqe;
>> + struct bnxt_re_atomic *sqe = ((void *)wqe +
>> + sizeof(struct bnxt_re_bsqe));
>> + int len;
>> +
>> + len = bnxt_re_build_send_sqe(qp, wqe, wr, false);
>> + hdr->key_immd = wr->wr.atomic.rkey;
>> + sqe->rva_lo = lower_32_bits(wr->wr.atomic.remote_addr);
>> + sqe->rva_hi = upper_32_bits(wr->wr.atomic.remote_addr);
>> + sqe->cmp_dt_lo = lower_32_bits(wr->wr.atomic.compare_add);
>> + sqe->cmp_dt_hi = upper_32_bits(wr->wr.atomic.compare_add);
>> +
>> + return len;
>> +}
>> +
>> int bnxt_re_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
>> struct ibv_send_wr **bad)
>> {
>> @@ -1169,27 +1210,28 @@ int bnxt_re_post_send(struct ibv_qp *ibvqp, struct ibv_send_wr *wr,
>> else
>> bytes = bnxt_re_build_send_sqe(qp, sqe, wr,
>> is_inline);
>> - if (bytes < 0)
>> - ret = (bytes == -EINVAL) ? EINVAL : ENOMEM;
>> break;
>> case IBV_WR_RDMA_WRITE_WITH_IMM:
>> hdr->key_immd = wr->imm_data;
>> case IBV_WR_RDMA_WRITE:
>> bytes = bnxt_re_build_rdma_sqe(qp, sqe, wr, is_inline);
>> - if (bytes < 0)
>> - ret = ENOMEM;
>> break;
>> case IBV_WR_RDMA_READ:
>> bytes = bnxt_re_build_rdma_sqe(qp, sqe, wr, false);
>> - if (bytes < 0)
>> - ret = ENOMEM;
>> + break;
>> + case IBV_WR_ATOMIC_CMP_AND_SWP:
>> + bytes = bnxt_re_build_cns_sqe(qp, sqe, wr);
>> + break;
>> + case IBV_WR_ATOMIC_FETCH_AND_ADD:
>> + bytes = bnxt_re_build_fna_sqe(qp, sqe, wr);
>> break;
>> default:
>> - ret = EINVAL;
>> + bytes = -EINVAL;
>> break;
>> }
>>
>> - if (ret) {
>> + if (bytes < 0) {
>> + ret = (bytes == -EINVAL) ? EINVAL : ENOMEM;
>> *bad = wr;
>> break;
>> }
>> --
>> 1.8.3.1
>>
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
>> the body of a message to majordomo@xxxxxxxxxxxxxxx
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
