On (02/21/18 17:50), Willem de Bruijn wrote: > > In the common case no more than one notification will be outstanding, > but with a fixed number of notifications per packet, in edge cases this > list may be long. : > Socket functions block if sk_err is non-zero. See for instance > tcp_sendmsg_locked. It is set by most code that also calls > sock_queue_err_skb and also on dequeue from err skb. > > This is the main reason that I would consider dropping error > queue completely if you expect all users of RDS to use the > cmsg on regular read to get these notifications.
I see. That's a good point, and maybe it makes sense to just have a struct sk_buff_head rs_zcookie_quese on the rds_sock, and have rds_rm_zerocopy_callback chain cookies ot this rs_zcookie_queue. [discussion regarding rds_recvmsg return values elided] > Okay. If callers must already handle 0 as a valid return code, then > it is fine to add another case that does this. > > The extra branch in the hot path is still rather unfortunately. Could > this be integrated in the existing if (nonblock) branch below? that's where I first started. it got even hairier because the callers expect a retval of 0 (-EAGAIN threw rds-stress into an infinite loop of continulally trying to recv) and the end result was just confusing code with the same number of branches.. let me revisit this when I spin out V2 without the sk_error_queue.. --Sowmini
