Re: [PATCH 3/3] virarptable: End parsing earlier in case of NLMSG_DONE

On 8/16/24 8:45 AM, Martin Kletzander wrote: Also: Resolves: https://bugzilla.redhat.com/2302245 Fixes: a176d67cdfaf5b8237a7e3a80d8be0e6bdf2d8fd (Hmm, and BTW I had suggested that patch 2/3 should have that Fixes: line, but I guess it can't be *totally* fixed until this patch, since prior to that we're checking the values of a struct ndmsg before checking the message type (i.e. it might not be an RTM_NEWNEIGH)) I'm pretty sure there are other issues with our netlink code, but the effect of those issues is unknown, and unrelated to the current bug which your code fixes. So, IMO, "this" doesn't have some issues :-) Well, virArpTableGet() calls virNetlinkGetNeighbor(), which allows responses where the *first* message in the response is NLMSG_NEWNEIGH or NLMSG_ERROR, and that calls virNetlinkTalk() which fails on any NLMSG_ERROR that isn't 0 (an NLMSG_ERRO with error==0 is just an ack packet). virNetlinkTalk() calls virNetLinkCommand(), which just calls nl_recv(), and as far as I can tell from the libnl API docs, there's no special handling there for any type of message. Have you ever tried to assimilate the libnl documentation? (https://www.infradead.org/~tgr/libnl/doc/api/) I'll refer you to my comment further down marked with "[*]". This somehow reminds me of code that we had in netcf that I think needed to get a list of interfaces, and we used the nl_cache_* functions for this, which in the end (I thought) made the code even more complicated than if we'd just used lower level functions. There is some very useful stuff in libnl, but I think some times it complicates in the name of "simplifying". I guess I'm trying to say "Maybe, but don't get your hopes up". Everything about netlink from the very beginning has been "hard to go through", especially when mixed in with libnl. I think originally there were a couple different usages of netlink/libnl in libvirt authored by a couple different people, and then another usage or two came along with each new user trying to refactor existing code to work for them too, but without ever any underlying plan. Or something like that. The result is that (IMO) trying to understand some of our netlink code might do you more harm than good :-/. Not that I'm cynical or anything. (P.S. Ditching the current abstraction(s) and making something new that works better for everyone would be wonderful to see. As long as someone else did it. The dreadful side of that is that I may soon need to add yet another usage of netlink to the code, and that could mean that either I will a) go through the pain and time consuming frustration of doing said ditching/rewriting so that my new code can be nicer, or b) live with the embarrasment of adding yet another hacky house of netlink cards to the code :-/) One other problem I noticed during my short dip into our netlink code last week is that in all cases except getting VFINFO (for SRIOV devices), we only read a single response packet after we make a request, when any response could have multiple packets. [*]Good luck with that. Be sure the beer cooler next to your desk is well stocked. I would prefer "break;" here rather than "return table;" just in case someone in the future adds something that requires some sort of cleanup code at the end of the function. This is fine too though. Reviewed-by: Laine Stump <laine(a)redhat.com&gt;