Re: [PATCH 1/2] RDMA/rxe: copy send WQE to kernel buffer before processing

在 2026/8/18 6:33, Tristan Madani 写道:
> Hi Yanjun,
> 
> Thank you for the review.
> 
> Regarding concern 2: unless I am mistaken, rxe_sender() calls
> rxe_requester() then rxe_completer() sequentially in the same
> work item (rxe_req.c:836), and do_task() prevents re-entry, so
> the requester and completer should not run concurrently on the
> same QP. But please correct me if there is a scheduling path
> I missed.
> 
> Either way, using WRITE_ONCE() for the state and status fields
> instead of a bulk memcpy() is cleaner and eliminates any concern
> about tearing from concurrent userspace writes (concern 1). I
> will send v2 with targeted WRITE_ONCE() writebacks for just the
> fields that need to be visible, dropping the bulk memcpy.

Sounds good. Please go ahead and send v2 so we can look at the actual 
implementation and discuss based on the code.

Best regards,
Yanjun Zhu

> 
> Best regards,
> Tristan Madani