> -----Original Message----- > From: David Howells <dhowells@xxxxxxxxxx> > Sent: Monday, 20 March 2023 12:09 > To: Bernard Metzler <BMT@xxxxxxxxxxxxxx> > Cc: David Howells <dhowells@xxxxxxxxxx>; Tom Talpey <tom@xxxxxxxxxx>; > linux-rdma@xxxxxxxxxxxxxxx > Subject: [EXTERNAL] Re: [RFC PATCH 08/28] siw: Inline do_tcp_sendpages() > > Bernard Metzler <BMT@xxxxxxxxxxxxxx> wrote: > > > > /* > > > - * 0copy TCP transmit interface: Use do_tcp_sendpages. > > > + * 0copy TCP transmit interface: Use MSG_SPLICE_PAGES. > > > * > > > * Using sendpage to push page by page appears to be less efficient > > > * than using sendmsg, even if data are copied. > > > > That is an interesting observation. Is efficiency to be read as > > CPU load, or throughput on the wire, or both? > > Um. The observation in the comment is one you made, not me according to > git Haha, yes. So sorry for that. I am getting older ;) I need to put on some more sanity checks before posting here! > blame. I merely changed "do_tcp_sendpages" to "MSG_SPLICE_PAGES" in the > first > line of the comment. > > > Back in the days, I introduced that zcopy path for efficiency > > reasons - getting both better throughput and less CPU load. > > I looked at both WRITE and READ performance. Using > > do_tcp_sendpages() is currently limited to processing work > > which is not registered with local completion generation. > > Replying to a remote READ request is a typical case. Did > > you check with READ? > > Ah - you're talking about ksmbd there? I haven't tested the patch with > that. Did you test with both kernel ULPs and user level applications? > > > > - rv = do_tcp_sendpages(sk, page[i], offset, bytes, flags); > > > + rv = tcp_sendmsg_locked(sk, &msg, size); > > > > Would that tcp_sendmsg_locked() with a msg flagged > > MSG_SPLICE_PAGES still have zero copy semantics? > > Yes - though I am considering making it conditional on whether the pages in > the iterator belong to the slab allocator (in which case they get copied) > or > not. Sounds good to me! > > David
