On Fri, 2018-08-03 at 13:43 -0600, Jason Gunthorpe wrote: +AD4- On Fri, Aug 03, 2018 at 10:24:31PM +-0300, Dan Carpenter wrote: +AD4- +AD4- A couple of the callers assume that ib+AF8-post+AF8-send() initializes +AD4- +AD4- +ACo-bad+AF8-send+AF8-wr. It doesn't look like we can rely on the -+AD4-post+AF8-send() +AD4- +AD4- function to initialize it. For example in the i40iw+AF8-post+AF8-recv() +AD4- +AD4- function and there are some error paths there which don't set it. +AD4- +AD4- I think those are bugs in the drivers, if bad+AF8-wr is provided, and +AD4- post+AF8-send fails then it must be set to the wr that has a problem, left +AD4- unset/uninit is incorrect. +AD4- +AD4- This is a high performance call path, so I'd prefer not to see +AD4- unnecessary babying of drivers.. +AD4- +AD4- Bart? Hello Dan and Jason, The documentation of the bad+AF8-wr parameter for ib+AF8-post+AF8-send(), ib+AF8-post+AF8-recv() and ib+AF8-post+AF8-srq+AF8-recv() is as follows: +ACo- +AEA-bad+AF8-...wr: On an immediate failure, this parameter will reference +ACo- the work request that failed to be posted on the QP. In other words, nothing is guaranteed about the +ACo-bad+AF8-wr value if these functions return 0. My proposal is to proceed as follows: +ACo- In the ULPs that read +ACo-bad+AF8-wr if one of these functions returns 0, initialize +ACo-bad+AF8-wr before calling one of these functions (I think all ULPs already do this). +ACo- Make sure that all HW drivers set +ACo-bad+AF8-wr if the return value is not 0. Thanks, Bart. -- 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
