On Sun, Apr 26, 2020 at 08:38:03PM +0200, Magnus Karlsson wrote: > On Sun, Apr 26, 2020 at 6:08 PM Jesse DuMond <suttonshiredev@xxxxxxxxx> wrote: > > > > Hi there, > > > > I am writing some test code that tries repeatedly to create a umem > > and socket and then deletes them both. I'm using the libbpf function > > xsk_umem__create/delete and xsk_socket__create/xsk_socket__delete. I > > am getting an EBUSY from xsk_socket__create when trying to create a > > socket after deleting a socket on the same interface and queue number. > > The attached program reproduces the error. > > > > I've traced the EBUSY error to this call to xdp_umem_assign_dev in the > > kernel: https://github.com/torvalds/linux/blob/v5.3/net/xdp/xsk.c#L500 > > That call is made during the bind call in xsk_socket__create. As far > > as I can tell this error is because the work of xdp_umem_clear_dev is > > actually called asynchronously after xsk_socket__delete. Putting a > > sleep between xsk_socket__delete and xsk_socket__create makes this > > code succeed. > > > > Am I trying to do something that is not supposed to be supported? > > No, you are doing everything correctly. The tear down of a socket is > unfortunately asynchronous due to legacy reasons that are not true > anymore. We do have a item in the backlog to make this synchronous, > but have not gotten around to it yet. So sorry, we have to live with > this for a while more. For now I guess that in case of EBUSY the application can just repeat the call to xsk_socket__create? > > /Magnus
|