perf evsel: Free evsel->counts in perf_evsel__exit()

[linux/fpc-iii.git] / Documentation / networking / timestamping.txt

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,
                   sizeof(val));

val is an integer with any of the following bits set. Setting other
bit returns EINVAL and does not change the current state.

The socket option configures timestamp generation for individual
sk_buffs (1.3.1), timestamp reporting to the socket's error
queue (1.3.2) and options (1.3.3). Timestamp generation can also
be enabled for individual sendmsg calls using cmsg (1.3.4).


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. This flag
  can be enabled via both socket options and control messages.

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.
  This flag can be enabled via both socket options and control messages.


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. This flag can be enabled
  via both socket options and control messages.

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.
  This flag can be enabled via both socket options and control messages.


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.


1.3.4. Enabling timestamps via control messages

In addition to socket options, timestamp generation can be requested
per write via cmsg, only for SOF_TIMESTAMPING_TX_* (see Section 1.3.1).
Using this feature, applications can sample timestamps per sendmsg()
without paying the overhead of enabling and disabling timestamps via
setsockopt:

  struct msghdr *msg;
  ...
  cmsg                         = CMSG_FIRSTHDR(msg);
  cmsg->cmsg_level             = SOL_SOCKET;
  cmsg->cmsg_type              = SO_TIMESTAMPING;
  cmsg->cmsg_len               = CMSG_LEN(sizeof(__u32));
  *((__u32 *) CMSG_DATA(cmsg)) = SOF_TIMESTAMPING_TX_SCHED |
                                 SOF_TIMESTAMPING_TX_SOFTWARE |
                                 SOF_TIMESTAMPING_TX_ACK;
  err = sendmsg(fd, msg, 0);

The SOF_TIMESTAMPING_TX_* flags set via cmsg will override
the SOF_TIMESTAMPING_TX_* flags set via setsockopt.

Moreover, applications must still enable timestamp reporting via
setsockopt to receive timestamps:

  __u32 val = SOF_TIMESTAMPING_SOFTWARE |
              SOF_TIMESTAMPING_OPT_ID /* or any other flag */;
  err = setsockopt(fd, SOL_SOCKET, SO_TIMESTAMPING, (void *) val,
                   sizeof(val));


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.

Drivers are free to use a more permissive configuration than the requested
configuration. It is expected that drivers should only implement directly the
most generic mode that can be supported. For example if the hardware can
support HWTSTAMP_FILTER_V2_EVENT, then it should generally always upscale
HWTSTAMP_FILTER_V2_L2_SYNC_MESSAGE, and so forth, as HWTSTAMP_FILTER_V2_EVENT
is more generic (and more useful to applications).

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.