On Thu, 12 Mar 2020 at 10:10, Björn Töpel <bjorn.topel@xxxxxxxxx> wrote:
>On Thu, 12 Mar 2020 at 09:49, Gaul, Maximilian <maximilian.gaul@xxxxxx> wrote:
>>
>> Björn Töpel <bjorn.topel@xxxxxxxxx> wrote:
>>
>> >On Thu, 12 Mar 2020 at 09:20, Gaul, Maximilian <maximilian.gaul@xxxxxx> wrote:
>> >>
>> >> I don't know if this reply works but I will try.
>> >>
>> >
>> >It worked! :-)
>> >
>> >> On Thu, 12 Mar 2020 at 08:55, Björn Töpel <bjorn.topel@xxxxxxxxx> wrote:
>> >> >>
>> >> >> Hello everyone,
>> >> >>
>> >> >
>> >> > Hi! I'm moving this to the XDP newbies list, which is a more proper
>> >> > place for these kind of discussions!
>> >> >
>> >> Sure, no problem. Thank you.
>> >> >>
>> >> >> I am not sure if this is the correct address for my question / problem but I was forwarded to this e-mail from the libbpf github-issue section, so this is my excuse.
>> >> >>
>> >> >>
>> >> >> Just a few information at the start of this e-mail: My program is largely based on: https://github.com/xdp-project/xdp-tutorial/tree/master/advanced03-AF_XDP and I am using libbpf: https://github.com/libbpf/libbpf
>> >> >>
>> >> >>
>> >> >> I am currently trying to build an application that enables me to process multiple udp-multicast streams at once in parallel (each with up to several ten-thousands of packets per second).
>> >> >>
>> >> >>
>> >> >> My first solution was to steer each multicast-stream on a separate RX-Queue on my NIC via `ethtool -N <if> flow-type udp4 ...` and to spawn as much user-space processes (each with a separate AF-XDP socket connected to one of the RX-Queues) as there are streams to process.
>> >> >>
>> >> >>
>> >> >> But because this solution is limited to the amount of RX-Queues the NIC has and I wanted to build something hardware-independent, I looked around a bit and found a feature called `XDP_SHARED_UMEM`.
>> >> >>
>> >>
>> >> > Let's start with defining what shared-umem is: The idea is to share
>> >> > the same umem, fill ring, and completion ring for multiple
>> >> > sockets. The sockets sharing that umem/fr/cr are tied (bound) to one
>> >> > hardware ring. It's a mechanism to load-balance a HW queue over
>> >> > multiple sockets.
>> >> >
>> >> > If I'm reading you correctly, you'd like a solution:
>> >> >
>> >> > hw_q0,
>> >> > xsk_q0_0, xsk_q0_1, xsk_q0_2,
>> >> >
>> >> > instead of:
>> >> >
>> >> > hw_q0, hw_q1, hw_q2,
>> >> > xsk_q0_0, xsk_q1_0, xsk_q2_0,
>> >> >
>> >> > In the first case you'll need to mux the flows in the XDP program
>> >> > using an XSKMAP.
>> >> >
>> >> > Is this what you're trying to do?
>> >> >
>> >> Yes it is. But I had the problem that I couldn't create multiple sockets (no sharing, everyone with its own umem and rx/tx queues) tied to the same RX-Queue. Maybe I did something wrong. But is this possible?
>> >
>> >No; one socket, one umem, one queue. Unless you're using shared umem,
>> >then multiple sockets, one umem, one queue.
>> >
>> >> >>
>> >> >>
>> >> >> As far as I understand (please correct me if I am wrong), at the moment libbpf only supports shared umem between threads of a process but not between processes - right?
>> >> >>
>> >> >
>> >> > Yes, that is correct, and for a reason! :-) Note that if you'd like to
>> >> > do a multi-*process* setup with shared umem, you: need to have a
>> >> > control process that manages the fill/completion rings, and
>> >> > synchronize between the processes, OR re-mmap the fill/completetion
>> >> > ring from the socket owning the umem in multiple processes *and*
>> >> > synchronize the access to them. Neither is pleasant.
>> >> >
>> >> > Honestly, not a setup I'd recommend.
>> >> >
>> >> This indeed sounds very unpleasent. So instead, if I understand correctly, you would go with the version above (the XDP program distributing the packets on the sockets via a XSKMAP). Is there something I have to watch out for? As I said, I wasn't able to create multiple sockets for the same RX-Queue.
>> >
>> >I would probably go for the first option, without shared umem, but
>> >that's really up to you! If you're going for the shared umem, I'd do
>> >it single process.
>> >
>>
>> I am sorry but I am confused, you just said *No; one socket, one umem, one queue.*. How would I be able to follow your rough sketch of
>>
>> hw_q0
>> xsk_q0_0, xsk_q0_1, xsk_q0_2
>>
>> I don't have deep knowledge about XDP and the pipeline, maybe there is something I am missing. I am sorry.
>>
>
>No worries! :-)
>
>Above you wrote "I couldn't create multiple sockets (no sharing,
>everyone with its own umem and rx/tx queues) tied to the same
>RX-Queue. Maybe I did something wrong." You can *only* tie multiple
>sockets to one queue by using shared umem. You said that "everyone
>with its own umem and rx/tx queues) tied to the same RX-Queue".
>
>If you'd like to go for the setup above, you can do this with libbpf
>today (have a look at the sample, where opt_num_xsks > 1). That will
>however be a single process solution.
>
>Clearer?
>
>
>Björn
>
>
Thank you so much Björn!
just to wrap things up:
- if I want to distribute packet processing from a single RX-Queue to multiple sockets I have to use shared umem because it is not possible to bind multiple af-xdp sockets onto the same RX-Queue
- furthermore, you would recommend to go with a single process / multiple threads solution in case of shared umem
is this correct?
Max
>> >> >> I ran unto the problem, that `struct xsk_umem` is hidden in `xsk.c`. This prevents me from copying the content from the original socket / umem into shared memory. I am not sure, what information the sub-process (the one which is using the umem from another process) needs so I figured the simplest solution would be to just copy the whole umem struct.
>> >> >>
>> >> >
>> >> > Just for completeness; To setup shared umem:
>> >> >
>> >> > 1. create socket 0 and register the umem to this.
>> >> > 2. mmap the fr/cr using socket 0
>> >> > 3. create socket 1, 2, n and refer to socket 0 for the umem.
>> >> >
>> >> > So, in a multiprocess solution step 3 would be done in separate
>> >> > processes, and step 2 depending on your application. You'd need to
>> >> > pass socket 0 to the other processes *and* share the umem memory from
>> >> > the process where socket 0 was created. This is pretty much a threaded
>> >> > solution, given all the shared state.
>> >> >
>> >> > I advice not taking this path.
>> >> >
>> >> I am not entirely sure what you mean with *passing socket 0* is this just the fd of the socket? What's about the `struct xsk_umem`? Do I need that? I guess so because `xsk_socket__create()` has a parameter `struct xsk_umem`.
>> >> >>
>> >> >>
>> >> >> So I went with the "quick-fix" to just move the definition of `struct xsk_umem` into `xsk.h` and to copy the umem-information from the original process into a shared memory. This process then calls `fork()` thus spawning a sub-process. This sub-process then reads the previously written umem-information from shared memory and passes it into `xsk_configure_socket` (af_xdp_user.c) which then eventually calls `xsk_socket__create` in `xsk.c`. This function then checks for `umem->refcount` and sets the flags for shared umem accordingly.
>> >> >>
>> >> >>
>> >> >>
>> >> >> After returning from `xsk_socket__create` (we are still in `xsk_configure_socket` in af_xdp_user.c), `bpf_get_link_xdp_id` is called (I don't know if that's necessary). But after that call I exit the function `xsk_socket__create` in the sub-process because I figured it is probably bad to configure the umem a second time by calling `xsk_ring_prod__reserve` after that:
>> >> >>
>> >> >>
>> >> >>
>> >> >>
>> >> >> static struct xsk_socket_info *xsk_configure_socket(struct config *cfg, struct xsk_umem_info *umem) {
>> >> >>
>> >> >> struct xsk_socket_config xsk_cfg;
>> >> >> struct xsk_socket_info *xsk_info;
>> >> >> uint32_t idx;
>> >> >> uint32_t prog_id = 0;
>> >> >> int i;
>> >> >> int ret;
>> >> >>
>> >> >> xsk_info = calloc(1, sizeof(*xsk_info));
>> >> >> if (!xsk_info)
>> >> >> return NULL;
>> >> >>
>> >> >> xsk_info->umem = umem;
>> >> >> xsk_cfg.rx_size = XSK_RING_CONS__DEFAULT_NUM_DESCS;
>> >> >> xsk_cfg.tx_size = XSK_RING_PROD__DEFAULT_NUM_DESCS;
>> >> >> xsk_cfg.libbpf_flags = 0;
>> >> >> xsk_cfg.xdp_flags = cfg->xdp_flags;
>> >> >> xsk_cfg.bind_flags = cfg->xsk_bind_flags;
>> >> >> ret = xsk_socket__create(&xsk_info->xsk, cfg->ifname, cfg->xsk_if_queue, umem->umem, &xsk_info->rx, &xsk_info->tx, &xsk_cfg);
>> >> >>
>> >> >> if (ret) {
>> >> >> fprintf(stderr, "FAIL 1\n");
>> >> >> goto error_exit;
>> >> >> }
>> >> >>
>> >> >> ret = bpf_get_link_xdp_id(cfg->ifindex, &prog_id, cfg->xdp_flags);
>> >> >> if (ret) {
>> >> >> fprintf(stderr, "FAIL 2\n");
>> >> >> goto error_exit;
>> >> >> }
>> >> >>
>> >> >> /* Initialize umem frame allocation */
>> >> >> for (i = 0; i < NUM_FRAMES; i++)
>> >> >> xsk_info->umem_frame_addr[i] = i * FRAME_SIZE;
>> >> >>
>> >> >> xsk_info->umem_frame_free = NUM_FRAMES;
>> >> >>
>> >> >> if(cfg->use_shrd_umem) {
>> >> >> return xsk_info;
>> >> >> }
>> >> >> ...
>> >> >> }
>> >> >>
>> >> >> Somehow what I am doing doesn't work because my sub-process dies in `xsk_configure_socket`. I am not able to debug it properly with GDB though. Another point I don't understand is the statement:
>> >> >>
>> >> >> However, note that you need to supply the XSK_LIBBPF_FLAGS__INHIBIT_PROG_LOAD libbpf_flag with the xsk_socket__create calls and load your own XDP program as there is no built in one in libbpf that will route the traffic for you.
>> >> >>
>> >> >> from https://www.kernel.org/doc/html/latest/networking/af_xdp.html#xdp-shared-umem-bind-flag
>> >> >>
>> >> >> I didn't know that libbpf loads a XDP-program? Why would it do that? I am using my own af-xdp program which filters for udp-packets. If I set `xsk_cfg.libbpf_flags = XSK_LIBBPF_FLAGS__INHIBIT_PROG_LOAD;` in `xsk_configure_socket`, the af-xdp-socket fd is not put into the kernel `xsks-map` which basically means that I don't receive any packets.
>> >> >>
>> >> >> As you probably already noticed, I am overstrained with the concept of Shared Umem and I have to say, there is no documentation about it besides the two sentences in https://www.kernel.org/doc/html/latest/networking/af_xdp.html#xdp-shared-umem-bind-flag and a mail in a linux mailbox from Nov. 2019 stating that this feature is now implemented.
>> >> >>
>> >> >> Can you please help?
>> >> >>
>> >> >
>> >> > XDP sockets always use an XDP program, it just that a default one is
>> >> > provided if the use doesn't explicitly add one. Have a look at
>> >> > tools/lib/bpf/xsk.c:xsk_load_xdp_prog. So, for shared umem you need to
>> >> > explicitly have a program that muxes over the sockets. A naïve variant
>> >> > can be found in samples/bpf/xdpsock_kern.c
>> >> >
>> >> >
>> >> > Cheers,
>> >> > Björn
>> >> >
>> >> >> Best regards
>> >> >>
>> >> >> Max
>> >>
|