Re: [libvirt] [PATCH v1] libvirtd: Increase NL buffer size for lots of interface

On 12/16/2015 10:24 AM, Michal Privoznik wrote: Apparently "LTC-Bugzilla" refers to a bugzilla instance for the Linux Technology Center at IBM, but I don't see how to get to it. References to bug reports are always nice in a patch, because it makes it much easier to spelunk the discussion leading up to the fix (now and later), but they can't be included unless they point to a publicly accessible report. Is LTC-Bugzilla (and in particular, these two reports) publicly accessible? If so, a clickable link would be better. The above function doesn't exist in libnl 1.1 (still used in RHEL6/CentOS6, for example), so that would cause a build failure on some systems. In libnl 1.1 the function is called nl_set_buffer_size(). Also, how did you arrive at 128k for the default buffer size? What kind of sizes are you seeing? According to a link I followed from another message on this topic last week, libnl's message peeking can't be guaranteed to always work, because netlink doesn't always return the proper buffer size (depending on version). I had the same memory. So I just looked back through the history of bug reports about this issue, and found the following: * libnl-1.1 and libnl-3 both originally set the default message buffersize to 4096 bytes, with MSG_PEEK turned off. * when this problem came up in RHEL6, it was unfortunately reported as a private BZ (a pet peeve of mine), and the result of the discussion about it was that libnl-1.1 (the version used in RHEL6) was patched *upstream* to set the default message buffersize to 16384 bytes (getpagesize() * 4), which would solve the problem for even very large numbers of VFs. That was in 2013 and there have been no further reports against RHEL6. * Although I had assumed the problem was solved, it again came up in RHEL7 (which uses libnl-3 - a slightly different API, and maintained in a separate git repo), this time in a public BZ: https://bugzilla.redhat.com/show_bug.cgi?id=1040626 I asked if perhaps the change that had been made upstream in libnl-1.1 hadn't also been made to libnl-3 (this is what I assumed during the previous incident). It hadn't. So the same change was made for libnl-3, both upstream and as a backport to RHEL7, and everyone was happy. I have very little detailed memory of that time (the above was all recalled by looking at archives of discussions) but what had stuck in my mind was "This problem has been fixed in libnl, so libvirt should NOT put in "workarounds" for broken versions of libnl." But if you are using a version of libnl3 with this patch (which was in libnl-3.2.22 upstream, and is in the libnl-3.2.21 that's in RHEL7.0+), : https://github.com/tgraf/libnl/commit/807fddc4cd9ecb12ba64e1b7fa26d86b6c2... then the change to quadruple the buffer size in libnl was insufficient, (and also, when I looked back at the discussion now, I see that the libnl maintainer had said "The permanent fix would be for libvirt to enable message peeking", so I suppose it's time to "bite the bullet" and enable netlink message peeking in libvirt (but, since there are apparently versions of netlink that don't properly inform libnl when a re-read is necessary, we also need to increase the default buffer size). However, your patch is only fixing the problem in one place. There are several places that we allocate netlink sockets, and they should all get the same fix, implying that there should be a common function called by all three. Fortunately, we already have a macro called "virNetlinkAlloc" that is #defined differently depending on the libnl version - this macro can simply be made into a static function that is defined differently depending on libnl version. It can call nl_handle_alloc or nl_socket_alloc depending on libnl version, then call nl_socket_set_buffer_size/nl_set_buffer_size depending on version, and finally call nl_socket_enable_msg_peek. A bit of due diligence about the default buffer size is in order though - just to make sure that we don't open a ton of netlink sockets at the same time, each with an unnecessarily huge buffer.