Re: [PATCH v2] avoid race condition between start_xmit and cm_rep_handler

On 8/20/18 12:28 PM, Jason Gunthorpe wrote:

On Mon, Aug 20, 2018 at 09:36:53AM +0300, Erez Shitrit wrote:

...

Hi,

Did you check the option to hold the netif_tx_lock_xxx() in ipoib_cm_rep_handler function (over the line set_bit(IPOIB_FLAG_OPER_UP)) instead of in the send data path flow?

That does seem better, then the test_bit in the datapath could become non-atomic too :)

Jason

Thanks for the feedback! I've not tried that but I certainly can.

-Aaron