Date: Wed, 29 Nov 2017 21:55:02 -0600 From: Xiaoye Sun <Xiaoye.Sun@rice.edu> To: Vincenzo Maffione <v.maffione@gmail.com> Cc: Luigi Rizzo <rizzo@iet.unipi.it>, "freebsd-net@freebsd.org" <freebsd-net@freebsd.org>, Victor Detoni <victordetoni@gmail.com>, Pavel Odintsov <pavel.odintsov@gmail.com>, Giuseppe Lettieri <g.lettieri@iet.unipi.it> Subject: Re: swaping ring slots between NIC ring and Host ring does not always success Message-ID: <CAJnByziKswKV5kYukGZgUUzy_ia%2B6En-d_E3kH_NEfPNPAYK-g@mail.gmail.com> In-Reply-To: <CA%2B_eA9jHT_R5id2kZBMe_j-Lc0-%2BvXuX4e8bqvy1cXymvaQN-A@mail.gmail.com> References: <CAJnByzj6Dj3vouZ2NbxqvCV-2-7TVtTR4FaWKuCFaaRN2X%2ByAA@mail.gmail.com> <CALgsdbd3XuE3wMYp4ey%2B1aer%2BHSVNojLYoVqwqTBPAXXdf9i%2BQ@mail.gmail.com> <CAJnByzirLXdCe-kwHV2s_E6ytGJG0Dth=0Ms12RrEk7FK_%2B8Og@mail.gmail.com> <CA%2BhQ2%2BgMWY0eabjHGw0=PJCAkS-wO=RBrN5brSbaqWc3_AOYoQ@mail.gmail.com> <CAJnByziBS8o6LtmpUrUu5xtRUd008Z2hnCsp=WVFv35r2J0rHw@mail.gmail.com> <CA%2BhQ2%2Bim9nFfYnqDS2HgRbAzdf5D0iaLCmCYhfXQVVRMouUFuw@mail.gmail.com> <CAJnByzht-qfDcm8oEg1aSRyVBZ1ygPvc2eMuoyJcq4geueTZ0Q@mail.gmail.com> <CA%2BhQ2%2BiERgWJ=cdFB-cByfT3r11T1kKr-5HiuCYZY-rxbjf=XA@mail.gmail.com> <CAJnByziDzdR2C6DcSRNPtrWACLq0XFpe4X1Ek9yXtFP9ivqWQw@mail.gmail.com> <CA%2BhQ2%2BhjnuGo1xKgc8CQ7gP35tiaZG7%2BroZBmX8aBgb8qWnLgg@mail.gmail.com> <CAJnByzh-VrRZeYdpkRFtCUGEN_arFBkemcN7byb51XV6UPswyg@mail.gmail.com> <CA%2BhQ2%2BiMw3kxjpcZy77vgOEsfk2UY0-farh9C8RKXZHMU7D8kw@mail.gmail.com> <CAJnByzgsuNBhdfPJsGrrHcU79xjK%2Bdq2RENgUkbZcehFm8MUxg@mail.gmail.com> <CAJnByzgNZ9YsYd7tBgYxiQPvuS_VZbhZNGvsPS-0apCDga7XFA@mail.gmail.com> <CANpwN=uHk-VwOoFz7NaPE9A-0B=MAapqxJ-uyCBtn=oMdacYnw@mail.gmail.com> <CAJnByzgjEEAzmWZu7BsSWHXmpjUtZcqXFGN8umCqmvgME1Jv%2BA@mail.gmail.com> <CANpwN=tfqitQW0BTXA7bU%2BTfmP8=wr7gE8wAP=hjAamjD7ny9Q@mail.gmail.com> <CAJnByzi5WHHvwcrmEOkJOHf5SJekbTtQoUgLmPbMtwTotc8mzA@mail.gmail.com> <CAJnByzhJnrmiwiLEEQV0meg7%2BDnLJ6Jq_J=6L=35Z9Lgw1GcyA@mail.gmail.com> <CA%2BhQ2%2BjTXyYdN4N4aWOkDdkBr%2BD3quocHn%2Bc8MA%2Bxc9yLs9V4w@mail.gmail.com> <CAJnByzg0PCXCyjnypS_2%2BRhKB0mf_4s8X1niFiyfedaCLUku7Q@mail.gmail.com> <CAJnByzipicNHhmw0kGq0cXnfgXqq1xP0aJ4dqr12SryBV1Ny3w@mail.gmail.com> <CA%2B_eA9h3NL6Ga4=HRjEiDtNaANkR8m7oZ9%2BCCpoWVB%2BaYW3%2Btw@mail.gmail.com> <CAJnByziVfVJ6gX_aL%2BwFvmW9ZsfwPV1RRCkxgEbZMoe1kSo35A@mail.gmail.com> <CA%2B_eA9jHT_R5id2kZBMe_j-Lc0-%2BvXuX4e8bqvy1cXymvaQN-A@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, Nov 27, 2017 at 2:33 AM, Vincenzo Maffione <v.maffione@gmail.com> wrote: > Hi, > > If you think it's a bug can you please open an issue on the github ( > https://github.com/luigirizzo/netmap/issues)? > > 2017-11-24 22:11 GMT+01:00 Xiaoye Sun <Xiaoye.Sun@rice.edu>: > >> Hi Vincenzo, >> >> Let me clarify my problem. (please ignore the previous incompleted email) >> >> I have a program, which is an extension of bridge.c >> https://github.com/luigirizzo/netmap/blob/master/apps/bridge/bridge.c >> The only difference is that my program also generates customized packets >> sent to the NIC directly. >> These customized packets have increasing sequence numbers. >> So, this program not only sends these customized packets but also >> forwards packets between NIC and host stack using zerocopy. >> The program only takes one NIC queue and there is only one thread. >> >> I think the problem is that there is a chance where netmap does not >> update the pointer to the buffer even when NS_BUF_CHANGED is set >> (buf_idx is changed). >> > > Can you disable zerocopy in bridge.c to see if the problem goes away? This > would be an useful information. > > disable zerocopy make this problem disappeared > >> Let's say the NIC tx ring has 4096 slots. The customized packet sequence >> 16 is filled in the buffer of slot 2057. >> The customized packets keep filling the slots until the next available >> slot is 2056. >> > > Do you mean that your program fills the TX ring slots > 2057,2058...2054,2055 with custom packets? This would mean you filled all > the available slots, since one slot is left empty. > > >> Now the customised packet sequence 4111 is filled to 2056. >> > > You cannot fill the slot 2056 if 2055 has not been NIOCTXYSINC'd. Aren't > you using nm_ring_empty() and nm_ring_space() functions to check > for available space in TX ring (assuming you update rinig->head/ring->cur > before calling those functions)? > > the slots are not filled-in once. NIOCTXYSINC are called at most every 512 slots are filled. I always use nm_ring_space() to check the number of remaining slots in the ring. > Cheers, > Vincenzo > > >> Then the netmap program is notified that there is a packet from the host >> stack sent to the NIC. >> The netmap program swaps the buf_idx between slot 2057 and the >> corresponding slot in the host rx ring and set the NS_BUF_CHANGED flag >> of both slots. >> Then the netmap program fills sequence 4112 to slot 2058. >> However, the buffer swap seems not succeed so that the original content >> of slot 2057 (sequence 16) is sent out. >> So that at the receiver side, the receiver sees two sequence >> 16s.(16,17...4110,4111,16,4112,4113). >> >> So think the root of the problem is that the buffer pointer is not always >> successfully/timely updated even after the NS_BUF_CHANGED flag is set >> and the buf_idx is updated. >> >> Best, >> Xiaoye >> >> >> >> On Wed, Nov 22, 2017 at 7:39 AM, Vincenzo Maffione <v.maffione@gmail.com> >> wrote: >> >>> Hi, >>> >>> 2017-11-21 7:51 GMT+01:00 Xiaoye Sun <Xiaoye.Sun@rice.edu>: >>> >>>> Hi, >>>> >>>> Recently I found another problem with netmap. I think this new problem >>>> could be related to the problems in this threads so I just post the new >>>> problem here. >>>> >>>> In my setup, I have a sender program having a netmap ring (a pair of >>>> RX/TX ring) for the NIC and a ring for the host stack. The sender >>>> program >>>> puts customized packets (each packet has a unique sequence number and >>>> the >>>> sender sends the packet in a sequence number increasing order) to the >>>> NIC >>>> TX ring directly and also forwards the packets from the host RX ring to >>>> the >>>> NIC TX ring using "zerocopy" by swapping the buffer indices. >>>> However, the receiver sees duplicated customized packets. For example, >>>> in >>>> the case where the ring size is 32 (32 slots in a ring) the order of the >>>> sequence numbers the receiver see is 1,2,3,4,5,...,68,69,*70* >>>> ,71,72,73,...,99,100,*70*,101,102,103,... . An interesting thing I >>>> found is >>>> that the "gaps" between these two duplicated packets (70 in the example) >>>> are always a number very close to the ring size, 32 in this example. In >>>> my >>>> experiment, I use a ring with 4096 slots and the gap is always more than >>>> 4090 and close to 4096. I verified that this duplication happens due to >>>> the >>>> sender, not the receiver. Assuming my sender's implementation is >>>> correct, >>>> then this duplication may happen in netmap and the NIC driver (ixgbe). >>>> >>> >>> Netmap itself doesn't do any duplication nor takes a look at the >>> packets. It just passes >>> down ring->cur/ring->head to the ixgbe driver (after validation). >>> The ixgbe driver datapath is bypassed and replaced with a netmap-enabled >>> datapath (see https://github.com/luigirizzo/ >>> netmap/blob/master/LINUX/ixgbe_netmap_linux.h#L294-L461); >>> no duplication should happen there as each netmap slot (1 TX packet) is >>> used >>> only once. >>> >>>> >>>> >>>> Thinking back to the original problem in this post, I think these >>>> problems >>>> may be related. It seems to me that there could be multiple threads >>>> pulling >>>> the packets from the NIC TX ring (or the thread moved to other CPUs when >>>> the problem occurs) and these threads may run on different cores so that >>>> the outdated content in the buffer may be sent out when new content is >>>> written to the buffer. >>>> >>>> >>> There are no such threads pulling from the NIC TX ring. Your application >>> directly >>> puts new packets to be transmitted in the netmap buffers referenced in >>> the netmap TX >>> ring. When then you call NIOCTXSYNC or poll(), all the new TX buffers >>> (e.g. all >>> the ones from the previous value ring->head (included) to the new value >>> of ring->head (excluded)) >>> are moved to the NIC TX ring. This happens in the context of your >>> application thread, >>> no worker threads are used. Then the NIC hardware starts the >>> transmission. >>> >>> >>>> I am wondering if there is a way to pin the NIC driver of the netmap >>>> module >>>> to a specific core. or is there a way to know the root of such problem? >>>> >>> >>> The only threads are the ones of your application. >>> Maybe your problem comes from concurrent accesses to the netmap TX ring >>> from different threads? Only one thread at a given time should update a >>> netmap >>> TX/RX ring. Otherwise the behaviour is unspecified. >>> >>> Cheers, >>> Vincenzo >>> >>> >>>> >>>> Best, >>>> Xiaoye >>>> >>>> >>> -- >>> Vincenzo Maffione >>> >> >> > > > -- > Vincenzo Maffione >
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CAJnByziKswKV5kYukGZgUUzy_ia%2B6En-d_E3kH_NEfPNPAYK-g>