On Mon, Nov 26, 2018 at 1:04 PM Paolo Abeni [off-list ref] wrote:

On Mon, 2018-11-26 at 12:59 -0500, Willem de Bruijn wrote:

quoted

The callers of this function do flush the queue of the other skbs on
error, but only after the call to sock_zerocopy_put_abort.

sock_zerocopy_put_abort depends on total rollback to revert the
sk_zckey increment and suppress the completion notification (which
must not happen on return with error).

I don't immediately have a fix. Need to think about this some more..

[still out of sheer ignorance] How about tacking a refcnt for the whole
ip_append_data() scope, like in the tcp case? that will add an atomic
op per loop (likely, hitting the cache) but will remove some code hunk
in sock_zerocopy_put_abort() and sock_zerocopy_alloc().

The atomic op pair is indeed what I was trying to avoid. But I also need
to solve the problem that the final decrement will happen from the freeing
of the other skbs in __ip_flush_pending_frames, and will not suppress
the notification.

Freeing the entire queue inside __ip_append_data, effectively making it
a true noop on error is one approach. But that is invasive, also to non
zerocopy codepaths, so I would rather avoid that.

Perhaps I need to handle the abort logic in udp_sendmsg directly,
after both __ip_append_data and __ip_flush_pending_frames.