diff options
Diffstat (limited to 'kernel/Documentation/networking/timestamping.txt')
-rw-r--r-- | kernel/Documentation/networking/timestamping.txt | 457 |
1 files changed, 457 insertions, 0 deletions
diff --git a/kernel/Documentation/networking/timestamping.txt b/kernel/Documentation/networking/timestamping.txt new file mode 100644 index 000000000..5f0922613 --- /dev/null +++ b/kernel/Documentation/networking/timestamping.txt @@ -0,0 +1,457 @@ + +1. Control Interfaces + +The interfaces for receiving network packages timestamps are: + +* SO_TIMESTAMP + Generates a timestamp for each incoming packet in (not necessarily + monotonic) system time. Reports the timestamp via recvmsg() in a + control message as struct timeval (usec resolution). + +* SO_TIMESTAMPNS + Same timestamping mechanism as SO_TIMESTAMP, but reports the + timestamp as struct timespec (nsec resolution). + +* IP_MULTICAST_LOOP + SO_TIMESTAMP[NS] + Only for multicast:approximate transmit timestamp obtained by + reading the looped packet receive timestamp. + +* SO_TIMESTAMPING + Generates timestamps on reception, transmission or both. Supports + multiple timestamp sources, including hardware. Supports generating + timestamps for stream sockets. + + +1.1 SO_TIMESTAMP: + +This socket option enables timestamping of datagrams on the reception +path. Because the destination socket, if any, is not known early in +the network stack, the feature has to be enabled for all packets. The +same is true for all early receive timestamp options. + +For interface details, see `man 7 socket`. + + +1.2 SO_TIMESTAMPNS: + +This option is identical to SO_TIMESTAMP except for the returned data type. +Its struct timespec allows for higher resolution (ns) timestamps than the +timeval of SO_TIMESTAMP (ms). + + +1.3 SO_TIMESTAMPING: + +Supports multiple types of timestamp requests. As a result, this +socket option takes a bitmap of flags, not a boolean. In + + err = setsockopt(fd, SOL_SOCKET, SO_TIMESTAMPING, (void *) val, &val); + +val is an integer with any of the following bits set. Setting other +bit returns EINVAL and does not change the current state. + + +1.3.1 Timestamp Generation + +Some bits are requests to the stack to try to generate timestamps. Any +combination of them is valid. Changes to these bits apply to newly +created packets, not to packets already in the stack. As a result, it +is possible to selectively request timestamps for a subset of packets +(e.g., for sampling) by embedding an send() call within two setsockopt +calls, one to enable timestamp generation and one to disable it. +Timestamps may also be generated for reasons other than being +requested by a particular socket, such as when receive timestamping is +enabled system wide, as explained earlier. + +SOF_TIMESTAMPING_RX_HARDWARE: + Request rx timestamps generated by the network adapter. + +SOF_TIMESTAMPING_RX_SOFTWARE: + Request rx timestamps when data enters the kernel. These timestamps + are generated just after a device driver hands a packet to the + kernel receive stack. + +SOF_TIMESTAMPING_TX_HARDWARE: + Request tx timestamps generated by the network adapter. + +SOF_TIMESTAMPING_TX_SOFTWARE: + Request tx timestamps when data leaves the kernel. These timestamps + are generated in the device driver as close as possible, but always + prior to, passing the packet to the network interface. Hence, they + require driver support and may not be available for all devices. + +SOF_TIMESTAMPING_TX_SCHED: + Request tx timestamps prior to entering the packet scheduler. Kernel + transmit latency is, if long, often dominated by queuing delay. The + difference between this timestamp and one taken at + SOF_TIMESTAMPING_TX_SOFTWARE will expose this latency independent + of protocol processing. The latency incurred in protocol + processing, if any, can be computed by subtracting a userspace + timestamp taken immediately before send() from this timestamp. On + machines with virtual devices where a transmitted packet travels + through multiple devices and, hence, multiple packet schedulers, + a timestamp is generated at each layer. This allows for fine + grained measurement of queuing delay. + +SOF_TIMESTAMPING_TX_ACK: + Request tx timestamps when all data in the send buffer has been + acknowledged. This only makes sense for reliable protocols. It is + currently only implemented for TCP. For that protocol, it may + over-report measurement, because the timestamp is generated when all + data up to and including the buffer at send() was acknowledged: the + cumulative acknowledgment. The mechanism ignores SACK and FACK. + + +1.3.2 Timestamp Reporting + +The other three bits control which timestamps will be reported in a +generated control message. Changes to the bits take immediate +effect at the timestamp reporting locations in the stack. Timestamps +are only reported for packets that also have the relevant timestamp +generation request set. + +SOF_TIMESTAMPING_SOFTWARE: + Report any software timestamps when available. + +SOF_TIMESTAMPING_SYS_HARDWARE: + This option is deprecated and ignored. + +SOF_TIMESTAMPING_RAW_HARDWARE: + Report hardware timestamps as generated by + SOF_TIMESTAMPING_TX_HARDWARE when available. + + +1.3.3 Timestamp Options + +The interface supports the options + +SOF_TIMESTAMPING_OPT_ID: + + Generate a unique identifier along with each packet. A process can + have multiple concurrent timestamping requests outstanding. Packets + can be reordered in the transmit path, for instance in the packet + scheduler. In that case timestamps will be queued onto the error + queue out of order from the original send() calls. It is not always + possible to uniquely match timestamps to the original send() calls + based on timestamp order or payload inspection alone, then. + + This option associates each packet at send() with a unique + identifier and returns that along with the timestamp. The identifier + is derived from a per-socket u32 counter (that wraps). For datagram + sockets, the counter increments with each sent packet. For stream + sockets, it increments with every byte. + + The counter starts at zero. It is initialized the first time that + the socket option is enabled. It is reset each time the option is + enabled after having been disabled. Resetting the counter does not + change the identifiers of existing packets in the system. + + This option is implemented only for transmit timestamps. There, the + timestamp is always looped along with a struct sock_extended_err. + The option modifies field ee_data to pass an id that is unique + among all possibly concurrently outstanding timestamp requests for + that socket. + + +SOF_TIMESTAMPING_OPT_CMSG: + + Support recv() cmsg for all timestamped packets. Control messages + are already supported unconditionally on all packets with receive + timestamps and on IPv6 packets with transmit timestamp. This option + extends them to IPv4 packets with transmit timestamp. One use case + is to correlate packets with their egress device, by enabling socket + option IP_PKTINFO simultaneously. + + +SOF_TIMESTAMPING_OPT_TSONLY: + + Applies to transmit timestamps only. Makes the kernel return the + timestamp as a cmsg alongside an empty packet, as opposed to + alongside the original packet. This reduces the amount of memory + charged to the socket's receive budget (SO_RCVBUF) and delivers + the timestamp even if sysctl net.core.tstamp_allow_data is 0. + This option disables SOF_TIMESTAMPING_OPT_CMSG. + + +New applications are encouraged to pass SOF_TIMESTAMPING_OPT_ID to +disambiguate timestamps and SOF_TIMESTAMPING_OPT_TSONLY to operate +regardless of the setting of sysctl net.core.tstamp_allow_data. + +An exception is when a process needs additional cmsg data, for +instance SOL_IP/IP_PKTINFO to detect the egress network interface. +Then pass option SOF_TIMESTAMPING_OPT_CMSG. This option depends on +having access to the contents of the original packet, so cannot be +combined with SOF_TIMESTAMPING_OPT_TSONLY. + + +1.4 Bytestream Timestamps + +The SO_TIMESTAMPING interface supports timestamping of bytes in a +bytestream. Each request is interpreted as a request for when the +entire contents of the buffer has passed a timestamping point. That +is, for streams option SOF_TIMESTAMPING_TX_SOFTWARE will record +when all bytes have reached the device driver, regardless of how +many packets the data has been converted into. + +In general, bytestreams have no natural delimiters and therefore +correlating a timestamp with data is non-trivial. A range of bytes +may be split across segments, any segments may be merged (possibly +coalescing sections of previously segmented buffers associated with +independent send() calls). Segments can be reordered and the same +byte range can coexist in multiple segments for protocols that +implement retransmissions. + +It is essential that all timestamps implement the same semantics, +regardless of these possible transformations, as otherwise they are +incomparable. Handling "rare" corner cases differently from the +simple case (a 1:1 mapping from buffer to skb) is insufficient +because performance debugging often needs to focus on such outliers. + +In practice, timestamps can be correlated with segments of a +bytestream consistently, if both semantics of the timestamp and the +timing of measurement are chosen correctly. This challenge is no +different from deciding on a strategy for IP fragmentation. There, the +definition is that only the first fragment is timestamped. For +bytestreams, we chose that a timestamp is generated only when all +bytes have passed a point. SOF_TIMESTAMPING_TX_ACK as defined is easy to +implement and reason about. An implementation that has to take into +account SACK would be more complex due to possible transmission holes +and out of order arrival. + +On the host, TCP can also break the simple 1:1 mapping from buffer to +skbuff as a result of Nagle, cork, autocork, segmentation and GSO. The +implementation ensures correctness in all cases by tracking the +individual last byte passed to send(), even if it is no longer the +last byte after an skbuff extend or merge operation. It stores the +relevant sequence number in skb_shinfo(skb)->tskey. Because an skbuff +has only one such field, only one timestamp can be generated. + +In rare cases, a timestamp request can be missed if two requests are +collapsed onto the same skb. A process can detect this situation by +enabling SOF_TIMESTAMPING_OPT_ID and comparing the byte offset at +send time with the value returned for each timestamp. It can prevent +the situation by always flushing the TCP stack in between requests, +for instance by enabling TCP_NODELAY and disabling TCP_CORK and +autocork. + +These precautions ensure that the timestamp is generated only when all +bytes have passed a timestamp point, assuming that the network stack +itself does not reorder the segments. The stack indeed tries to avoid +reordering. The one exception is under administrator control: it is +possible to construct a packet scheduler configuration that delays +segments from the same stream differently. Such a setup would be +unusual. + + +2 Data Interfaces + +Timestamps are read using the ancillary data feature of recvmsg(). +See `man 3 cmsg` for details of this interface. The socket manual +page (`man 7 socket`) describes how timestamps generated with +SO_TIMESTAMP and SO_TIMESTAMPNS records can be retrieved. + + +2.1 SCM_TIMESTAMPING records + +These timestamps are returned in a control message with cmsg_level +SOL_SOCKET, cmsg_type SCM_TIMESTAMPING, and payload of type + +struct scm_timestamping { + struct timespec ts[3]; +}; + +The structure can return up to three timestamps. This is a legacy +feature. Only one field is non-zero at any time. Most timestamps +are passed in ts[0]. Hardware timestamps are passed in ts[2]. + +ts[1] used to hold hardware timestamps converted to system time. +Instead, expose the hardware clock device on the NIC directly as +a HW PTP clock source, to allow time conversion in userspace and +optionally synchronize system time with a userspace PTP stack such +as linuxptp. For the PTP clock API, see Documentation/ptp/ptp.txt. + +2.1.1 Transmit timestamps with MSG_ERRQUEUE + +For transmit timestamps the outgoing packet is looped back to the +socket's error queue with the send timestamp(s) attached. A process +receives the timestamps by calling recvmsg() with flag MSG_ERRQUEUE +set and with a msg_control buffer sufficiently large to receive the +relevant metadata structures. The recvmsg call returns the original +outgoing data packet with two ancillary messages attached. + +A message of cm_level SOL_IP(V6) and cm_type IP(V6)_RECVERR +embeds a struct sock_extended_err. This defines the error type. For +timestamps, the ee_errno field is ENOMSG. The other ancillary message +will have cm_level SOL_SOCKET and cm_type SCM_TIMESTAMPING. This +embeds the struct scm_timestamping. + + +2.1.1.2 Timestamp types + +The semantics of the three struct timespec are defined by field +ee_info in the extended error structure. It contains a value of +type SCM_TSTAMP_* to define the actual timestamp passed in +scm_timestamping. + +The SCM_TSTAMP_* types are 1:1 matches to the SOF_TIMESTAMPING_* +control fields discussed previously, with one exception. For legacy +reasons, SCM_TSTAMP_SND is equal to zero and can be set for both +SOF_TIMESTAMPING_TX_HARDWARE and SOF_TIMESTAMPING_TX_SOFTWARE. It +is the first if ts[2] is non-zero, the second otherwise, in which +case the timestamp is stored in ts[0]. + + +2.1.1.3 Fragmentation + +Fragmentation of outgoing datagrams is rare, but is possible, e.g., by +explicitly disabling PMTU discovery. If an outgoing packet is fragmented, +then only the first fragment is timestamped and returned to the sending +socket. + + +2.1.1.4 Packet Payload + +The calling application is often not interested in receiving the whole +packet payload that it passed to the stack originally: the socket +error queue mechanism is just a method to piggyback the timestamp on. +In this case, the application can choose to read datagrams with a +smaller buffer, possibly even of length 0. The payload is truncated +accordingly. Until the process calls recvmsg() on the error queue, +however, the full packet is queued, taking up budget from SO_RCVBUF. + + +2.1.1.5 Blocking Read + +Reading from the error queue is always a non-blocking operation. To +block waiting on a timestamp, use poll or select. poll() will return +POLLERR in pollfd.revents if any data is ready on the error queue. +There is no need to pass this flag in pollfd.events. This flag is +ignored on request. See also `man 2 poll`. + + +2.1.2 Receive timestamps + +On reception, there is no reason to read from the socket error queue. +The SCM_TIMESTAMPING ancillary data is sent along with the packet data +on a normal recvmsg(). Since this is not a socket error, it is not +accompanied by a message SOL_IP(V6)/IP(V6)_RECVERROR. In this case, +the meaning of the three fields in struct scm_timestamping is +implicitly defined. ts[0] holds a software timestamp if set, ts[1] +is again deprecated and ts[2] holds a hardware timestamp if set. + + +3. Hardware Timestamping configuration: SIOCSHWTSTAMP and SIOCGHWTSTAMP + +Hardware time stamping must also be initialized for each device driver +that is expected to do hardware time stamping. The parameter is defined in +/include/linux/net_tstamp.h as: + +struct hwtstamp_config { + int flags; /* no flags defined right now, must be zero */ + int tx_type; /* HWTSTAMP_TX_* */ + int rx_filter; /* HWTSTAMP_FILTER_* */ +}; + +Desired behavior is passed into the kernel and to a specific device by +calling ioctl(SIOCSHWTSTAMP) with a pointer to a struct ifreq whose +ifr_data points to a struct hwtstamp_config. The tx_type and +rx_filter are hints to the driver what it is expected to do. If +the requested fine-grained filtering for incoming packets is not +supported, the driver may time stamp more than just the requested types +of packets. + +A driver which supports hardware time stamping shall update the struct +with the actual, possibly more permissive configuration. If the +requested packets cannot be time stamped, then nothing should be +changed and ERANGE shall be returned (in contrast to EINVAL, which +indicates that SIOCSHWTSTAMP is not supported at all). + +Only a processes with admin rights may change the configuration. User +space is responsible to ensure that multiple processes don't interfere +with each other and that the settings are reset. + +Any process can read the actual configuration by passing this +structure to ioctl(SIOCGHWTSTAMP) in the same way. However, this has +not been implemented in all drivers. + +/* possible values for hwtstamp_config->tx_type */ +enum { + /* + * no outgoing packet will need hardware time stamping; + * should a packet arrive which asks for it, no hardware + * time stamping will be done + */ + HWTSTAMP_TX_OFF, + + /* + * enables hardware time stamping for outgoing packets; + * the sender of the packet decides which are to be + * time stamped by setting SOF_TIMESTAMPING_TX_SOFTWARE + * before sending the packet + */ + HWTSTAMP_TX_ON, +}; + +/* possible values for hwtstamp_config->rx_filter */ +enum { + /* time stamp no incoming packet at all */ + HWTSTAMP_FILTER_NONE, + + /* time stamp any incoming packet */ + HWTSTAMP_FILTER_ALL, + + /* return value: time stamp all packets requested plus some others */ + HWTSTAMP_FILTER_SOME, + + /* PTP v1, UDP, any kind of event packet */ + HWTSTAMP_FILTER_PTP_V1_L4_EVENT, + + /* for the complete list of values, please check + * the include file /include/linux/net_tstamp.h + */ +}; + +3.1 Hardware Timestamping Implementation: Device Drivers + +A driver which supports hardware time stamping must support the +SIOCSHWTSTAMP ioctl and update the supplied struct hwtstamp_config with +the actual values as described in the section on SIOCSHWTSTAMP. It +should also support SIOCGHWTSTAMP. + +Time stamps for received packets must be stored in the skb. To get a pointer +to the shared time stamp structure of the skb call skb_hwtstamps(). Then +set the time stamps in the structure: + +struct skb_shared_hwtstamps { + /* hardware time stamp transformed into duration + * since arbitrary point in time + */ + ktime_t hwtstamp; +}; + +Time stamps for outgoing packets are to be generated as follows: +- In hard_start_xmit(), check if (skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) + is set no-zero. If yes, then the driver is expected to do hardware time + stamping. +- If this is possible for the skb and requested, then declare + that the driver is doing the time stamping by setting the flag + SKBTX_IN_PROGRESS in skb_shinfo(skb)->tx_flags , e.g. with + + skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS; + + You might want to keep a pointer to the associated skb for the next step + and not free the skb. A driver not supporting hardware time stamping doesn't + do that. A driver must never touch sk_buff::tstamp! It is used to store + software generated time stamps by the network subsystem. +- Driver should call skb_tx_timestamp() as close to passing sk_buff to hardware + as possible. skb_tx_timestamp() provides a software time stamp if requested + and hardware timestamping is not possible (SKBTX_IN_PROGRESS not set). +- As soon as the driver has sent the packet and/or obtained a + hardware time stamp for it, it passes the time stamp back by + calling skb_hwtstamp_tx() with the original skb, the raw + hardware time stamp. skb_hwtstamp_tx() clones the original skb and + adds the timestamps, therefore the original skb has to be freed now. + If obtaining the hardware time stamp somehow fails, then the driver + should not fall back to software time stamping. The rationale is that + this would occur at a later time in the processing pipeline than other + software time stamping and therefore could lead to unexpected deltas + between time stamps. |