On Wed, 23 Jul 2025 06:32:45 +0400 (+04) Ivan Malov <ivan.ma...@arknetworks.am> wrote:
> Hi Stephen, > > On Tue, 22 Jul 2025, Stephen Hemminger wrote: > > > In new model, the packet was already copied, only need > > Copied? But what if it was "indirect attached" instead, as the model > envisages? > > Perhaps this is a silly question of mine, but it may not be clear what happens > in case of 'RTE_ETH_MIRROR_INDIRECT_FLAG' - whether it is safe to modify the > mbuf and whether the 'indirect' clone has to be "freed" in the sense of being > detached/refcnt updated after successful pcapng write, to avoid memory leaks? The indirect attached is for the direct (no pcap case). Code should have checked for flag conflict. I dropped the flag in next version, too may possible bugs and not useful for the case of packet capture. The problem is that indirect method assumes that transmit can handle reference counts (won't work with fast free), and on receive that upper layers do not overwrite packet with refcount > 1. The Linux and BSD kernel stack have these semantics, but DPDK does not make the requirement explicit. I.e too many risks for something that is experimental without current users.
