I thought the nature of trans_fd would have prevented any sort of true zero copy, but I suppose one less is always welcome :)
-eric On Sun, Sep 6, 2015 at 1:55 AM, Dominique Martinet <dominique.marti...@cea.fr> wrote: > Eric Van Hensbergen wrote on Sat, Sep 05, 2015: >> On Thu, Sep 3, 2015 at 4:38 AM, Dominique Martinet >> <dominique.marti...@cea.fr> wrote: >> > To be honest, I think it might be better to just bail out if we get in >> > this switch (m->req->rc == NULL after p9_tag_lookup) and not try to >> > allocate more, because if we get there it's likely a race condition and >> > silently re-allocating will end up in more troubles than trying to >> > recover is worth. >> > Thoughts ? >> > >> >> Hmmm...trying to rattle my brain and remember why I put it in there >> back in 2008. >> It might have just been over-defensive programming -- or more likely it just >> pre-dated all the zero copy infrastructure which pretty much guaranteed we >> had >> an rc allocated and what is there is vestigial. I'm happy to accept a >> patch which >> makes this an assert, or perhaps just resets the connection because something >> has gone horribly wrong (similar to the ENOMEM path that is there now). > > Yeah, it looks like the safety comes from the zero-copy stuff that came > much later. > Let's go with resetting the connection then. Hmm. EIO is a bit too > generic so would be good to avoid that if possible, but can't think of > anything better... > > > Speaking of zero-copy, I believe it should be fairly straight-forward to > implement for trans_fd now I've actually looked at it, since we do the > payload read after a p9_tag_lookup, would just need m->req to point to a > zc buffer. Write is similar, if there's a zc buffer just send it after > the header. > The cost is a couple more pointers in req and an extra if in both > workers, that seems pretty reasonable. > > Well, I'm not using trans_fd much here (and unfortunately zero-copy > isn't possible at all given the transport protocol for RDMA, at least > for recv), but if anyone cares it probably could be done without too > much hassle for the fd workers. > > -- > Dominique -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/