Please drop this patch. It is mlx4 specific driver issue.
On Wed, Apr 03, 2019 at 10:54:18AM +0800, Honggang Li wrote:
> Signed-off-by: Honggang Li <honli@xxxxxxxxxx>
> ---
> libibverbs/man/ibv_poll_cq.3 | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/libibverbs/man/ibv_poll_cq.3 b/libibverbs/man/ibv_poll_cq.3
> index 957fd151495a..db94436f848b 100644
> --- a/libibverbs/man/ibv_poll_cq.3
> +++ b/libibverbs/man/ibv_poll_cq.3
> @@ -30,7 +30,7 @@ uint64_t wr_id; /* ID of the completed Work Request (WR)
> enum ibv_wc_status status; /* Status of the operation */
> enum ibv_wc_opcode opcode; /* Operation type specified in the completed WR */
> uint32_t vendor_err; /* Vendor error syndrome */
> -uint32_t byte_len; /* Number of bytes transferred */
> +uint32_t byte_len; /* Number of bytes received */
> union {
> .in +8
> __be32 imm_data; /* Immediate data (in network byte order) */
> @@ -78,6 +78,8 @@ found. On failure, a negative value is returned.
> .PP
> Each polled completion is removed from the CQ and cannot be returned to it.
> .PP
> +ibv_poll_cq will not update the "byte_len" field when the "opcode" of completed WR is IBV_WC_SEND.
> +.PP
> The user should consume work completions at a rate that prevents CQ
> overrun from occurrence. In case of a CQ overrun, the async event
> .B IBV_EVENT_CQ_ERR
> --
> 2.15.0-rc1
>
