[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [PATCH 4/5] ebpf_rss_helper: Added helper for eBPF RSS.
From: |
Jason Wang |
Subject: |
Re: [PATCH 4/5] ebpf_rss_helper: Added helper for eBPF RSS. |
Date: |
Fri, 10 Sep 2021 09:37:29 +0800 |
On Fri, Sep 10, 2021 at 7:44 AM Yuri Benditovich
<yuri.benditovich@daynix.com> wrote:
>
> On Thu, Sep 9, 2021 at 4:16 AM Jason Wang <jasowang@redhat.com> wrote:
> >
> > On Thu, Sep 9, 2021 at 8:00 AM Yuri Benditovich
> > <yuri.benditovich@daynix.com> wrote:
> > >
> > > On Wed, Sep 8, 2021 at 6:45 AM Jason Wang <jasowang@redhat.com> wrote:
> > > >
> > > > On Tue, Sep 7, 2021 at 6:40 PM Yuri Benditovich
> > > > <yuri.benditovich@daynix.com> wrote:
> > > > >
> > > > > On Wed, Sep 1, 2021 at 9:42 AM Jason Wang <jasowang@redhat.com> wrote:
> > > > > >
> > > > > >
> > > > > > 在 2021/8/31 上午1:07, Yuri Benditovich 写道:
> > > > > > > On Fri, Aug 20, 2021 at 6:41 AM Jason Wang <jasowang@redhat.com>
> > > > > > > wrote:
> > > > > > >>
> > > > > > >> 在 2021/7/13 下午11:37, Andrew Melnychenko 写道:
> > > > > > >>> Helper program. Loads eBPF RSS program and maps and passes them
> > > > > > >>> through unix socket.
> > > > > > >>> Libvirt may launch this helper and pass eBPF fds to qemu
> > > > > > >>> virtio-net.
> > > > > > >>
> > > > > > >> I wonder if this can be done as helper for TAP/bridge.
> > > > > > >>
> > > > > > >> E.g it's the qemu to launch those helper with set-uid.
> > > > > > >>
> > > > > > >> Then libvirt won't even need to care about that?
> > > > > > >>
> > > > > > > There are pros and cons for such a solution with set-uid.
> > > > > > > From my point of view one of the cons is that set-uid is
> > > > > > > efficient
> > > > > > > only at install time so the coexistence of different qemu builds
> > > > > > > (and
> > > > > > > different helpers for each one) is kind of problematic.
> > > > > > > With the current solution this does not present any problem: the
> > > > > > > developer can have several different builds, each one
> > > > > > > automatically
> > > > > > > has its own helper and there is no conflict between these builds
> > > > > > > and
> > > > > > > between these builds and installed qemu package. Changing the
> > > > > > > 'emulator' in the libvirt profile automatically brings the proper
> > > > > > > helper to work.
> > > > > >
> > > > > >
> > > > > > I'm not sure I get you here. We can still have default/sample
> > > > > > helper to
> > > > > > make sure it works for different builds.
> > > > > >
> > > > > > If we can avoid the involvement of libvirt, that would be better.
> > > > >
> > > > > Hi Jason,
> > > > >
> > > > > Indeed I did not get the idea, can you please explain it in more
> > > > > details (as detailed as possible to avoid future misunderstanding),
> > > > > especially how exactly we can use the set-uid and what is the
> > > > > 'default' helper.
> > > > > We also would prefer to do everything from qemu but we do not see how
> > > > > we can do that.
> > > >
> > > >
> > > Some more questions to understand the idea better:
> > > > Something like:
> > > >
> > > > 1) -netdev tap,rss_helper=/path/to/name
> > >
> > > So, on each editing of 'emulator' in the xml the helper path should
> > > be set manually or be default?
> >
> > It could done manually, or we can have a default path.
> >
> > >
> > > > 2) having a sample/default helper implemented in Qemu
> > >
> > > Does it mean the default helper is the code in the qemu (without
> > > running additional executable, like it does today)
> >
> > Yes.
> If the "default helper" is just a keyword and it is like what we have
> today (i.e. part of qemu)
Yes, it's something like "-netdev tap,script=no"
> it can't work under libvirt and should never
> be used by libvirt.
Any reason for this?
> >
> > or this is qemu
> > > itself with dedicated command line?
> > > As far as I remember Daniel had strong objections of ever running qemu
> > > with capabilities
> >
> > Qemu won't run with capabilities but the helper.
> So under libvirt the helper is always separate executable and not
> "default helper"
That's fine, we can ship the helper as an independent package I think.
>
> >
> > >
> > > > 3) we can introduce something special path like "default", then if
> > > > -netdev tap,rss_helper="default" is specified, qemu will use the
> > > > sample helper
> > >
> > > Probably this is not so important but the rss helper and rss in
> > > general has no relation to netdev, much more they are related to
> > > virtio-net
> >
> > So I think the reason for this is that we currently only support
> > eBPF/RSS for tap.
>
> This is just because only tap supports respective ioctls.
>
> >
> > >
> > > >
> > > > So we have:
> > > > 1) set set-uid for the helper
> > > Who and when does set-uid to the helper binary? Only installer or
> > > libvirt can do that, correct?
> >
> > Yes, it could be done the installer, or other system provision tools.
>
> So this changes the rule of the game: currently libvirt runs helpers
> that require privileges and qemu runs helpers that do not require any
> privileges (like TPM).
I don't understand here, if the helper doesn't need any privilege, why
not do that in qemu itself?
> If we follow your suggestion - only for RSS we will create the helper
> that must be used with set-uid.
> Who are stakeholders we need to have a consensus with?
I think we need inputs from libvirt guys. Daniel, any idea on this?
Thanks
>
> >
> > >
> > > > 2) libvirt may just choose to launch the default helper
> > > All this discussion is to avoid launching the helper from libvirt,
> > > correct?
> >
> > Sorry, it's a typo. I meant, libvirt launch qemu, and then qemu will
> > launch the helper.
> >
> > Thanks
> >
> > >
> > > >
> > > > >
> > > > > Our main points (what should be addressed):
> > > > > - qemu should be able to load ebpf and use the maps when it runs from
> > > > > libvirt (without special caps) and standalone (with caps)
> > > >
> > > > This is solved by leaving the privileged operations to the helper with
> > > > set-uid.
> > > >
> > > > > - it is possible that there are different qemu builds on the machine,
> > > > > one of them might be installed, their ebpf's might be different and
> > > > > the interface between qemu and ebpf (exact content of maps and number
> > > > > of maps)
> > > >
> > > > We can use different helpers in this way.
> > > >
> > > > > - qemu configures the RSS dynamically according to the commands
> > > > > provided by the guest
> > > >
> > > > Consider we decided to use mmap() based maps, this is not an issue.
> > > >
> > > > Or am I missing something?
> > > >
> > > > Thanks
> > > >
> > > > >
> > > > > Thanks in advance
> > > > > Yuri
> > > > >
> > > > > >
> > > > > > Thanks
> > > > > >
> > > > > >
> > > > > > >
> > > > > > >>> Also, libbpf dependency now exclusively for Linux.
> > > > > > >>> Libbpf is used for eBPF RSS steering, which is supported only
> > > > > > >>> by Linux TAP.
> > > > > > >>> There is no reason yet to build eBPF loader and helper for non
> > > > > > >>> Linux systems,
> > > > > > >>> even if libbpf is present.
> > > > > > >>>
> > > > > > >>> Signed-off-by: Andrew Melnychenko <andrew@daynix.com>
> > > > > > >>> ---
> > > > > > >>> ebpf/qemu-ebpf-rss-helper.c | 130
> > > > > > >>> ++++++++++++++++++++++++++++++++++++
> > > > > > >>> meson.build | 37 ++++++----
> > > > > > >>> 2 files changed, 154 insertions(+), 13 deletions(-)
> > > > > > >>> create mode 100644 ebpf/qemu-ebpf-rss-helper.c
> > > > > > >>>
> > > > > > >>> diff --git a/ebpf/qemu-ebpf-rss-helper.c
> > > > > > >>> b/ebpf/qemu-ebpf-rss-helper.c
> > > > > > >>> new file mode 100644
> > > > > > >>> index 0000000000..fe68758f57
> > > > > > >>> --- /dev/null
> > > > > > >>> +++ b/ebpf/qemu-ebpf-rss-helper.c
> > > > > > >>> @@ -0,0 +1,130 @@
> > > > > > >>> +/*
> > > > > > >>> + * eBPF RSS Helper
> > > > > > >>> + *
> > > > > > >>> + * Developed by Daynix Computing LTD (http://www.daynix.com)
> > > > > > >>> + *
> > > > > > >>> + * Authors:
> > > > > > >>> + * Andrew Melnychenko <andrew@daynix.com>
> > > > > > >>> + *
> > > > > > >>> + * This work is licensed under the terms of the GNU GPL,
> > > > > > >>> version 2. See
> > > > > > >>> + * the COPYING file in the top-level directory.
> > > > > > >>> + *
> > > > > > >>> + * Description: This is helper program for libvirtd.
> > > > > > >>> + * It loads eBPF RSS program and passes fds
> > > > > > >>> through unix socket.
> > > > > > >>> + * Built by meson, target -
> > > > > > >>> 'qemu-ebpf-rss-helper'.
> > > > > > >>> + */
> > > > > > >>> +
> > > > > > >>> +#include <stdio.h>
> > > > > > >>> +#include <stdint.h>
> > > > > > >>> +#include <stdlib.h>
> > > > > > >>> +#include <stdbool.h>
> > > > > > >>> +#include <getopt.h>
> > > > > > >>> +#include <memory.h>
> > > > > > >>> +#include <errno.h>
> > > > > > >>> +#include <sys/socket.h>
> > > > > > >>> +
> > > > > > >>> +#include "ebpf_rss.h"
> > > > > > >>> +
> > > > > > >>> +#include "qemu-helper-stamp.h"
> > > > > > >>> +
> > > > > > >>> +void QEMU_HELPER_STAMP(void) {}
> > > > > > >>> +
> > > > > > >>> +static int send_fds(int socket, int *fds, int n)
> > > > > > >>> +{
> > > > > > >>> + struct msghdr msg = {};
> > > > > > >>> + struct cmsghdr *cmsg = NULL;
> > > > > > >>> + char buf[CMSG_SPACE(n * sizeof(int))];
> > > > > > >>> + char dummy_buffer = 0;
> > > > > > >>> + struct iovec io = { .iov_base = &dummy_buffer,
> > > > > > >>> + .iov_len = sizeof(dummy_buffer) };
> > > > > > >>> +
> > > > > > >>> + memset(buf, 0, sizeof(buf));
> > > > > > >>> +
> > > > > > >>> + msg.msg_iov = &io;
> > > > > > >>> + msg.msg_iovlen = 1;
> > > > > > >>> + msg.msg_control = buf;
> > > > > > >>> + msg.msg_controllen = sizeof(buf);
> > > > > > >>> +
> > > > > > >>> + cmsg = CMSG_FIRSTHDR(&msg);
> > > > > > >>> + cmsg->cmsg_level = SOL_SOCKET;
> > > > > > >>> + cmsg->cmsg_type = SCM_RIGHTS;
> > > > > > >>> + cmsg->cmsg_len = CMSG_LEN(n * sizeof(int));
> > > > > > >>> +
> > > > > > >>> + memcpy(CMSG_DATA(cmsg), fds, n * sizeof(int));
> > > > > > >>> +
> > > > > > >>> + return sendmsg(socket, &msg, 0);
> > > > > > >>> +}
> > > > > > >>> +
> > > > > > >>> +static void print_help_and_exit(const char *prog, int exitcode)
> > > > > > >>> +{
> > > > > > >>> + fprintf(stderr, "%s - load eBPF RSS program for qemu and
> > > > > > >>> pass eBPF fds"
> > > > > > >>> + " through unix socket.\n", prog);
> > > > > > >>> + fprintf(stderr, "\t--fd <num>, -f <num> - unix socket file
> > > > > > >>> descriptor"
> > > > > > >>> + " used to pass eBPF fds.\n");
> > > > > > >>> + fprintf(stderr, "\t--help, -h - this help.\n");
> > > > > > >>> + exit(exitcode);
> > > > > > >>> +}
> > > > > > >>> +
> > > > > > >>> +int main(int argc, char **argv)
> > > > > > >>> +{
> > > > > > >>> + char *fd_string = NULL;
> > > > > > >>> + int unix_fd = 0;
> > > > > > >>> + struct EBPFRSSContext ctx = {};
> > > > > > >>> + int fds[EBPF_RSS_MAX_FDS] = {};
> > > > > > >>> + int ret = -1;
> > > > > > >>> +
> > > > > > >>> + for (;;) {
> > > > > > >>> + int c;
> > > > > > >>> + static struct option long_options[] = {
> > > > > > >>> + {"help", no_argument, 0, 'h'},
> > > > > > >>> + {"fd", required_argument, 0, 'f'},
> > > > > > >>> + {0, 0, 0, 0}
> > > > > > >>> + };
> > > > > > >>> + c = getopt_long(argc, argv, "hf:",
> > > > > > >>> + long_options, NULL);
> > > > > > >>> +
> > > > > > >>> + if (c == -1) {
> > > > > > >>> + break;
> > > > > > >>> + }
> > > > > > >>> +
> > > > > > >>> + switch (c) {
> > > > > > >>> + case 'f':
> > > > > > >>> + fd_string = optarg;
> > > > > > >>> + break;
> > > > > > >>> + case 'h':
> > > > > > >>> + default:
> > > > > > >>> + print_help_and_exit(argv[0],
> > > > > > >>> + c == 'h' ? EXIT_SUCCESS : EXIT_FAILURE);
> > > > > > >>> + }
> > > > > > >>> + }
> > > > > > >>> +
> > > > > > >>> + if (!fd_string) {
> > > > > > >>> + fprintf(stderr, "Unix file descriptor not present.\n");
> > > > > > >>> + print_help_and_exit(argv[0], EXIT_FAILURE);
> > > > > > >>> + }
> > > > > > >>> +
> > > > > > >>> + unix_fd = atoi(fd_string);
> > > > > > >>> +
> > > > > > >>> + if (!unix_fd) {
> > > > > > >>> + fprintf(stderr, "Unix file descriptor is invalid.\n");
> > > > > > >>> + return EXIT_FAILURE;
> > > > > > >>> + }
> > > > > > >>> +
> > > > > > >>> + ebpf_rss_init(&ctx);
> > > > > > >>> + if (!ebpf_rss_load(&ctx)) {
> > > > > > >>> + fprintf(stderr, "Can't load ebpf.\n");
> > > > > > >>> + return EXIT_FAILURE;
> > > > > > >>> + }
> > > > > > >>> + fds[0] = ctx.program_fd;
> > > > > > >>> + fds[1] = ctx.map_configuration;
> > > > > > >>> +
> > > > > > >>> + ret = send_fds(unix_fd, fds, EBPF_RSS_MAX_FDS);
> > > > > > >>> + if (ret < 0) {
> > > > > > >>> + fprintf(stderr, "Issue while sending fds: %s.\n",
> > > > > > >>> strerror(errno));
> > > > > > >>> + }
> > > > > > >>> +
> > > > > > >>> + ebpf_rss_unload(&ctx);
> > > > > > >>> +
> > > > > > >>> + return ret < 0 ? EXIT_FAILURE : EXIT_SUCCESS;
> > > > > > >>> +}
> > > > > > >>> +
> > > > > > >>> diff --git a/meson.build b/meson.build
> > > > > > >>> index 257e51d91b..913aa1fee5 100644
> > > > > > >>> --- a/meson.build
> > > > > > >>> +++ b/meson.build
> > > > > > >>> @@ -1033,19 +1033,22 @@ if not
> > > > > > >>> get_option('fuse_lseek').disabled()
> > > > > > >>> endif
> > > > > > >>>
> > > > > > >>> # libbpf
> > > > > > >>> -libbpf = dependency('libbpf', required: get_option('bpf'),
> > > > > > >>> method: 'pkg-config')
> > > > > > >>> -if libbpf.found() and not cc.links('''
> > > > > > >>> - #include <bpf/libbpf.h>
> > > > > > >>> - int main(void)
> > > > > > >>> - {
> > > > > > >>> - bpf_object__destroy_skeleton(NULL);
> > > > > > >>> - return 0;
> > > > > > >>> - }''', dependencies: libbpf)
> > > > > > >>> - libbpf = not_found
> > > > > > >>> - if get_option('bpf').enabled()
> > > > > > >>> - error('libbpf skeleton test failed')
> > > > > > >>> - else
> > > > > > >>> - warning('libbpf skeleton test failed, disabling')
> > > > > > >>> +libbpf = not_found
> > > > > > >>> +if targetos == 'linux'
> > > > > > >>> + libbpf = dependency('libbpf', required: get_option('bpf'),
> > > > > > >>> method: 'pkg-config')
> > > > > > >>> + if libbpf.found() and not cc.links('''
> > > > > > >>> + #include <bpf/libbpf.h>
> > > > > > >>> + int main(void)
> > > > > > >>> + {
> > > > > > >>> + bpf_object__destroy_skeleton(NULL);
> > > > > > >>
> > > > > > >> Do we need to test whether the bpf can do mmap() here?
> > > > > > >>
> > > > > > >> Thanks
> > > > > > >>
> > > > > > >>
> > > > > > >>> + return 0;
> > > > > > >>> + }''', dependencies: libbpf)
> > > > > > >>> + libbpf = not_found
> > > > > > >>> + if get_option('bpf').enabled()
> > > > > > >>> + error('libbpf skeleton test failed')
> > > > > > >>> + else
> > > > > > >>> + warning('libbpf skeleton test failed, disabling')
> > > > > > >>> + endif
> > > > > > >>> endif
> > > > > > >>> endif
> > > > > > >>>
> > > > > > >>> @@ -2423,6 +2426,14 @@ if have_tools
> > > > > > >>> dependencies: [authz, crypto, io, qom,
> > > > > > >>> qemuutil,
> > > > > > >>> libcap_ng, mpathpersist],
> > > > > > >>> install: true)
> > > > > > >>> +
> > > > > > >>> + if libbpf.found()
> > > > > > >>> + executable('qemu-ebpf-rss-helper', files(
> > > > > > >>> + 'ebpf/qemu-ebpf-rss-helper.c',
> > > > > > >>> 'ebpf/ebpf_rss.c'),
> > > > > > >>> + dependencies: [qemuutil, libbpf, glib],
> > > > > > >>> + install: true,
> > > > > > >>> + install_dir: get_option('libexecdir'))
> > > > > > >>> + endif
> > > > > > >>> endif
> > > > > > >>>
> > > > > > >>> if 'CONFIG_IVSHMEM' in config_host
> > > > > >
> > > > >
> > > >
> > >
> >
>