1.. SPDX-License-Identifier: GPL-2.0
2
3============
4Timestamping
5============
6
7
81. Control Interfaces
9=====================
10
11The interfaces for receiving network packages timestamps are:
12
13SO_TIMESTAMP
14  Generates a timestamp for each incoming packet in (not necessarily
15  monotonic) system time. Reports the timestamp via recvmsg() in a
16  control message in usec resolution.
17  SO_TIMESTAMP is defined as SO_TIMESTAMP_NEW or SO_TIMESTAMP_OLD
18  based on the architecture type and time_t representation of libc.
19  Control message format is in struct __kernel_old_timeval for
20  SO_TIMESTAMP_OLD and in struct __kernel_sock_timeval for
21  SO_TIMESTAMP_NEW options respectively.
22
23SO_TIMESTAMPNS
24  Same timestamping mechanism as SO_TIMESTAMP, but reports the
25  timestamp as struct timespec in nsec resolution.
26  SO_TIMESTAMPNS is defined as SO_TIMESTAMPNS_NEW or SO_TIMESTAMPNS_OLD
27  based on the architecture type and time_t representation of libc.
28  Control message format is in struct timespec for SO_TIMESTAMPNS_OLD
29  and in struct __kernel_timespec for SO_TIMESTAMPNS_NEW options
30  respectively.
31
32IP_MULTICAST_LOOP + SO_TIMESTAMP[NS]
33  Only for multicast:approximate transmit timestamp obtained by
34  reading the looped packet receive timestamp.
35
36SO_TIMESTAMPING
37  Generates timestamps on reception, transmission or both. Supports
38  multiple timestamp sources, including hardware. Supports generating
39  timestamps for stream sockets.
40
41
421.1 SO_TIMESTAMP (also SO_TIMESTAMP_OLD and SO_TIMESTAMP_NEW)
43-------------------------------------------------------------
44
45This socket option enables timestamping of datagrams on the reception
46path. Because the destination socket, if any, is not known early in
47the network stack, the feature has to be enabled for all packets. The
48same is true for all early receive timestamp options.
49
50For interface details, see `man 7 socket`.
51
52Always use SO_TIMESTAMP_NEW timestamp to always get timestamp in
53struct __kernel_sock_timeval format.
54
55SO_TIMESTAMP_OLD returns incorrect timestamps after the year 2038
56on 32 bit machines.
57
581.2 SO_TIMESTAMPNS (also SO_TIMESTAMPNS_OLD and SO_TIMESTAMPNS_NEW):
59
60This option is identical to SO_TIMESTAMP except for the returned data type.
61Its struct timespec allows for higher resolution (ns) timestamps than the
62timeval of SO_TIMESTAMP (ms).
63
64Always use SO_TIMESTAMPNS_NEW timestamp to always get timestamp in
65struct __kernel_timespec format.
66
67SO_TIMESTAMPNS_OLD returns incorrect timestamps after the year 2038
68on 32 bit machines.
69
701.3 SO_TIMESTAMPING (also SO_TIMESTAMPING_OLD and SO_TIMESTAMPING_NEW)
71----------------------------------------------------------------------
72
73Supports multiple types of timestamp requests. As a result, this
74socket option takes a bitmap of flags, not a boolean. In::
75
76  err = setsockopt(fd, SOL_SOCKET, SO_TIMESTAMPING, &val, sizeof(val));
77
78val is an integer with any of the following bits set. Setting other
79bit returns EINVAL and does not change the current state.
80
81The socket option configures timestamp generation for individual
82sk_buffs (1.3.1), timestamp reporting to the socket's error
83queue (1.3.2) and options (1.3.3). Timestamp generation can also
84be enabled for individual sendmsg calls using cmsg (1.3.4).
85
86
871.3.1 Timestamp Generation
88^^^^^^^^^^^^^^^^^^^^^^^^^^
89
90Some bits are requests to the stack to try to generate timestamps. Any
91combination of them is valid. Changes to these bits apply to newly
92created packets, not to packets already in the stack. As a result, it
93is possible to selectively request timestamps for a subset of packets
94(e.g., for sampling) by embedding an send() call within two setsockopt
95calls, one to enable timestamp generation and one to disable it.
96Timestamps may also be generated for reasons other than being
97requested by a particular socket, such as when receive timestamping is
98enabled system wide, as explained earlier.
99
100SOF_TIMESTAMPING_RX_HARDWARE:
101  Request rx timestamps generated by the network adapter.
102
103SOF_TIMESTAMPING_RX_SOFTWARE:
104  Request rx timestamps when data enters the kernel. These timestamps
105  are generated just after a device driver hands a packet to the
106  kernel receive stack.
107
108SOF_TIMESTAMPING_TX_HARDWARE:
109  Request tx timestamps generated by the network adapter. This flag
110  can be enabled via both socket options and control messages.
111
112SOF_TIMESTAMPING_TX_SOFTWARE:
113  Request tx timestamps when data leaves the kernel. These timestamps
114  are generated in the device driver as close as possible, but always
115  prior to, passing the packet to the network interface. Hence, they
116  require driver support and may not be available for all devices.
117  This flag can be enabled via both socket options and control messages.
118
119SOF_TIMESTAMPING_TX_SCHED:
120  Request tx timestamps prior to entering the packet scheduler. Kernel
121  transmit latency is, if long, often dominated by queuing delay. The
122  difference between this timestamp and one taken at
123  SOF_TIMESTAMPING_TX_SOFTWARE will expose this latency independent
124  of protocol processing. The latency incurred in protocol
125  processing, if any, can be computed by subtracting a userspace
126  timestamp taken immediately before send() from this timestamp. On
127  machines with virtual devices where a transmitted packet travels
128  through multiple devices and, hence, multiple packet schedulers,
129  a timestamp is generated at each layer. This allows for fine
130  grained measurement of queuing delay. This flag can be enabled
131  via both socket options and control messages.
132
133SOF_TIMESTAMPING_TX_ACK:
134  Request tx timestamps when all data in the send buffer has been
135  acknowledged. This only makes sense for reliable protocols. It is
136  currently only implemented for TCP. For that protocol, it may
137  over-report measurement, because the timestamp is generated when all
138  data up to and including the buffer at send() was acknowledged: the
139  cumulative acknowledgment. The mechanism ignores SACK and FACK.
140  This flag can be enabled via both socket options and control messages.
141
142
1431.3.2 Timestamp Reporting
144^^^^^^^^^^^^^^^^^^^^^^^^^
145
146The other three bits control which timestamps will be reported in a
147generated control message. Changes to the bits take immediate
148effect at the timestamp reporting locations in the stack. Timestamps
149are only reported for packets that also have the relevant timestamp
150generation request set.
151
152SOF_TIMESTAMPING_SOFTWARE:
153  Report any software timestamps when available.
154
155SOF_TIMESTAMPING_SYS_HARDWARE:
156  This option is deprecated and ignored.
157
158SOF_TIMESTAMPING_RAW_HARDWARE:
159  Report hardware timestamps as generated by
160  SOF_TIMESTAMPING_TX_HARDWARE when available.
161
162
1631.3.3 Timestamp Options
164^^^^^^^^^^^^^^^^^^^^^^^
165
166The interface supports the options
167
168SOF_TIMESTAMPING_OPT_ID:
169  Generate a unique identifier along with each packet. A process can
170  have multiple concurrent timestamping requests outstanding. Packets
171  can be reordered in the transmit path, for instance in the packet
172  scheduler. In that case timestamps will be queued onto the error
173  queue out of order from the original send() calls. It is not always
174  possible to uniquely match timestamps to the original send() calls
175  based on timestamp order or payload inspection alone, then.
176
177  This option associates each packet at send() with a unique
178  identifier and returns that along with the timestamp. The identifier
179  is derived from a per-socket u32 counter (that wraps). For datagram
180  sockets, the counter increments with each sent packet. For stream
181  sockets, it increments with every byte.
182
183  The counter starts at zero. It is initialized the first time that
184  the socket option is enabled. It is reset each time the option is
185  enabled after having been disabled. Resetting the counter does not
186  change the identifiers of existing packets in the system.
187
188  This option is implemented only for transmit timestamps. There, the
189  timestamp is always looped along with a struct sock_extended_err.
190  The option modifies field ee_data to pass an id that is unique
191  among all possibly concurrently outstanding timestamp requests for
192  that socket.
193
194
195SOF_TIMESTAMPING_OPT_CMSG:
196  Support recv() cmsg for all timestamped packets. Control messages
197  are already supported unconditionally on all packets with receive
198  timestamps and on IPv6 packets with transmit timestamp. This option
199  extends them to IPv4 packets with transmit timestamp. One use case
200  is to correlate packets with their egress device, by enabling socket
201  option IP_PKTINFO simultaneously.
202
203
204SOF_TIMESTAMPING_OPT_TSONLY:
205  Applies to transmit timestamps only. Makes the kernel return the
206  timestamp as a cmsg alongside an empty packet, as opposed to
207  alongside the original packet. This reduces the amount of memory
208  charged to the socket's receive budget (SO_RCVBUF) and delivers
209  the timestamp even if sysctl net.core.tstamp_allow_data is 0.
210  This option disables SOF_TIMESTAMPING_OPT_CMSG.
211
212SOF_TIMESTAMPING_OPT_STATS:
213  Optional stats that are obtained along with the transmit timestamps.
214  It must be used together with SOF_TIMESTAMPING_OPT_TSONLY. When the
215  transmit timestamp is available, the stats are available in a
216  separate control message of type SCM_TIMESTAMPING_OPT_STATS, as a
217  list of TLVs (struct nlattr) of types. These stats allow the
218  application to associate various transport layer stats with
219  the transmit timestamps, such as how long a certain block of
220  data was limited by peer's receiver window.
221
222SOF_TIMESTAMPING_OPT_PKTINFO:
223  Enable the SCM_TIMESTAMPING_PKTINFO control message for incoming
224  packets with hardware timestamps. The message contains struct
225  scm_ts_pktinfo, which supplies the index of the real interface which
226  received the packet and its length at layer 2. A valid (non-zero)
227  interface index will be returned only if CONFIG_NET_RX_BUSY_POLL is
228  enabled and the driver is using NAPI. The struct contains also two
229  other fields, but they are reserved and undefined.
230
231SOF_TIMESTAMPING_OPT_TX_SWHW:
232  Request both hardware and software timestamps for outgoing packets
233  when SOF_TIMESTAMPING_TX_HARDWARE and SOF_TIMESTAMPING_TX_SOFTWARE
234  are enabled at the same time. If both timestamps are generated,
235  two separate messages will be looped to the socket's error queue,
236  each containing just one timestamp.
237
238New applications are encouraged to pass SOF_TIMESTAMPING_OPT_ID to
239disambiguate timestamps and SOF_TIMESTAMPING_OPT_TSONLY to operate
240regardless of the setting of sysctl net.core.tstamp_allow_data.
241
242An exception is when a process needs additional cmsg data, for
243instance SOL_IP/IP_PKTINFO to detect the egress network interface.
244Then pass option SOF_TIMESTAMPING_OPT_CMSG. This option depends on
245having access to the contents of the original packet, so cannot be
246combined with SOF_TIMESTAMPING_OPT_TSONLY.
247
248
2491.3.4. Enabling timestamps via control messages
250^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
251
252In addition to socket options, timestamp generation can be requested
253per write via cmsg, only for SOF_TIMESTAMPING_TX_* (see Section 1.3.1).
254Using this feature, applications can sample timestamps per sendmsg()
255without paying the overhead of enabling and disabling timestamps via
256setsockopt::
257
258  struct msghdr *msg;
259  ...
260  cmsg			       = CMSG_FIRSTHDR(msg);
261  cmsg->cmsg_level	       = SOL_SOCKET;
262  cmsg->cmsg_type	       = SO_TIMESTAMPING;
263  cmsg->cmsg_len	       = CMSG_LEN(sizeof(__u32));
264  *((__u32 *) CMSG_DATA(cmsg)) = SOF_TIMESTAMPING_TX_SCHED |
265				 SOF_TIMESTAMPING_TX_SOFTWARE |
266				 SOF_TIMESTAMPING_TX_ACK;
267  err = sendmsg(fd, msg, 0);
268
269The SOF_TIMESTAMPING_TX_* flags set via cmsg will override
270the SOF_TIMESTAMPING_TX_* flags set via setsockopt.
271
272Moreover, applications must still enable timestamp reporting via
273setsockopt to receive timestamps::
274
275  __u32 val = SOF_TIMESTAMPING_SOFTWARE |
276	      SOF_TIMESTAMPING_OPT_ID /* or any other flag */;
277  err = setsockopt(fd, SOL_SOCKET, SO_TIMESTAMPING, &val, sizeof(val));
278
279
2801.4 Bytestream Timestamps
281-------------------------
282
283The SO_TIMESTAMPING interface supports timestamping of bytes in a
284bytestream. Each request is interpreted as a request for when the
285entire contents of the buffer has passed a timestamping point. That
286is, for streams option SOF_TIMESTAMPING_TX_SOFTWARE will record
287when all bytes have reached the device driver, regardless of how
288many packets the data has been converted into.
289
290In general, bytestreams have no natural delimiters and therefore
291correlating a timestamp with data is non-trivial. A range of bytes
292may be split across segments, any segments may be merged (possibly
293coalescing sections of previously segmented buffers associated with
294independent send() calls). Segments can be reordered and the same
295byte range can coexist in multiple segments for protocols that
296implement retransmissions.
297
298It is essential that all timestamps implement the same semantics,
299regardless of these possible transformations, as otherwise they are
300incomparable. Handling "rare" corner cases differently from the
301simple case (a 1:1 mapping from buffer to skb) is insufficient
302because performance debugging often needs to focus on such outliers.
303
304In practice, timestamps can be correlated with segments of a
305bytestream consistently, if both semantics of the timestamp and the
306timing of measurement are chosen correctly. This challenge is no
307different from deciding on a strategy for IP fragmentation. There, the
308definition is that only the first fragment is timestamped. For
309bytestreams, we chose that a timestamp is generated only when all
310bytes have passed a point. SOF_TIMESTAMPING_TX_ACK as defined is easy to
311implement and reason about. An implementation that has to take into
312account SACK would be more complex due to possible transmission holes
313and out of order arrival.
314
315On the host, TCP can also break the simple 1:1 mapping from buffer to
316skbuff as a result of Nagle, cork, autocork, segmentation and GSO. The
317implementation ensures correctness in all cases by tracking the
318individual last byte passed to send(), even if it is no longer the
319last byte after an skbuff extend or merge operation. It stores the
320relevant sequence number in skb_shinfo(skb)->tskey. Because an skbuff
321has only one such field, only one timestamp can be generated.
322
323In rare cases, a timestamp request can be missed if two requests are
324collapsed onto the same skb. A process can detect this situation by
325enabling SOF_TIMESTAMPING_OPT_ID and comparing the byte offset at
326send time with the value returned for each timestamp. It can prevent
327the situation by always flushing the TCP stack in between requests,
328for instance by enabling TCP_NODELAY and disabling TCP_CORK and
329autocork.
330
331These precautions ensure that the timestamp is generated only when all
332bytes have passed a timestamp point, assuming that the network stack
333itself does not reorder the segments. The stack indeed tries to avoid
334reordering. The one exception is under administrator control: it is
335possible to construct a packet scheduler configuration that delays
336segments from the same stream differently. Such a setup would be
337unusual.
338
339
3402 Data Interfaces
341==================
342
343Timestamps are read using the ancillary data feature of recvmsg().
344See `man 3 cmsg` for details of this interface. The socket manual
345page (`man 7 socket`) describes how timestamps generated with
346SO_TIMESTAMP and SO_TIMESTAMPNS records can be retrieved.
347
348
3492.1 SCM_TIMESTAMPING records
350----------------------------
351
352These timestamps are returned in a control message with cmsg_level
353SOL_SOCKET, cmsg_type SCM_TIMESTAMPING, and payload of type
354
355For SO_TIMESTAMPING_OLD::
356
357	struct scm_timestamping {
358		struct timespec ts[3];
359	};
360
361For SO_TIMESTAMPING_NEW::
362
363	struct scm_timestamping64 {
364		struct __kernel_timespec ts[3];
365
366Always use SO_TIMESTAMPING_NEW timestamp to always get timestamp in
367struct scm_timestamping64 format.
368
369SO_TIMESTAMPING_OLD returns incorrect timestamps after the year 2038
370on 32 bit machines.
371
372The structure can return up to three timestamps. This is a legacy
373feature. At least one field is non-zero at any time. Most timestamps
374are passed in ts[0]. Hardware timestamps are passed in ts[2].
375
376ts[1] used to hold hardware timestamps converted to system time.
377Instead, expose the hardware clock device on the NIC directly as
378a HW PTP clock source, to allow time conversion in userspace and
379optionally synchronize system time with a userspace PTP stack such
380as linuxptp. For the PTP clock API, see Documentation/driver-api/ptp.rst.
381
382Note that if the SO_TIMESTAMP or SO_TIMESTAMPNS option is enabled
383together with SO_TIMESTAMPING using SOF_TIMESTAMPING_SOFTWARE, a false
384software timestamp will be generated in the recvmsg() call and passed
385in ts[0] when a real software timestamp is missing. This happens also
386on hardware transmit timestamps.
387
3882.1.1 Transmit timestamps with MSG_ERRQUEUE
389^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
390
391For transmit timestamps the outgoing packet is looped back to the
392socket's error queue with the send timestamp(s) attached. A process
393receives the timestamps by calling recvmsg() with flag MSG_ERRQUEUE
394set and with a msg_control buffer sufficiently large to receive the
395relevant metadata structures. The recvmsg call returns the original
396outgoing data packet with two ancillary messages attached.
397
398A message of cm_level SOL_IP(V6) and cm_type IP(V6)_RECVERR
399embeds a struct sock_extended_err. This defines the error type. For
400timestamps, the ee_errno field is ENOMSG. The other ancillary message
401will have cm_level SOL_SOCKET and cm_type SCM_TIMESTAMPING. This
402embeds the struct scm_timestamping.
403
404
4052.1.1.2 Timestamp types
406~~~~~~~~~~~~~~~~~~~~~~~
407
408The semantics of the three struct timespec are defined by field
409ee_info in the extended error structure. It contains a value of
410type SCM_TSTAMP_* to define the actual timestamp passed in
411scm_timestamping.
412
413The SCM_TSTAMP_* types are 1:1 matches to the SOF_TIMESTAMPING_*
414control fields discussed previously, with one exception. For legacy
415reasons, SCM_TSTAMP_SND is equal to zero and can be set for both
416SOF_TIMESTAMPING_TX_HARDWARE and SOF_TIMESTAMPING_TX_SOFTWARE. It
417is the first if ts[2] is non-zero, the second otherwise, in which
418case the timestamp is stored in ts[0].
419
420
4212.1.1.3 Fragmentation
422~~~~~~~~~~~~~~~~~~~~~
423
424Fragmentation of outgoing datagrams is rare, but is possible, e.g., by
425explicitly disabling PMTU discovery. If an outgoing packet is fragmented,
426then only the first fragment is timestamped and returned to the sending
427socket.
428
429
4302.1.1.4 Packet Payload
431~~~~~~~~~~~~~~~~~~~~~~
432
433The calling application is often not interested in receiving the whole
434packet payload that it passed to the stack originally: the socket
435error queue mechanism is just a method to piggyback the timestamp on.
436In this case, the application can choose to read datagrams with a
437smaller buffer, possibly even of length 0. The payload is truncated
438accordingly. Until the process calls recvmsg() on the error queue,
439however, the full packet is queued, taking up budget from SO_RCVBUF.
440
441
4422.1.1.5 Blocking Read
443~~~~~~~~~~~~~~~~~~~~~
444
445Reading from the error queue is always a non-blocking operation. To
446block waiting on a timestamp, use poll or select. poll() will return
447POLLERR in pollfd.revents if any data is ready on the error queue.
448There is no need to pass this flag in pollfd.events. This flag is
449ignored on request. See also `man 2 poll`.
450
451
4522.1.2 Receive timestamps
453^^^^^^^^^^^^^^^^^^^^^^^^
454
455On reception, there is no reason to read from the socket error queue.
456The SCM_TIMESTAMPING ancillary data is sent along with the packet data
457on a normal recvmsg(). Since this is not a socket error, it is not
458accompanied by a message SOL_IP(V6)/IP(V6)_RECVERROR. In this case,
459the meaning of the three fields in struct scm_timestamping is
460implicitly defined. ts[0] holds a software timestamp if set, ts[1]
461is again deprecated and ts[2] holds a hardware timestamp if set.
462
463
4643. Hardware Timestamping configuration: SIOCSHWTSTAMP and SIOCGHWTSTAMP
465=======================================================================
466
467Hardware time stamping must also be initialized for each device driver
468that is expected to do hardware time stamping. The parameter is defined in
469include/uapi/linux/net_tstamp.h as::
470
471	struct hwtstamp_config {
472		int flags;	/* no flags defined right now, must be zero */
473		int tx_type;	/* HWTSTAMP_TX_* */
474		int rx_filter;	/* HWTSTAMP_FILTER_* */
475	};
476
477Desired behavior is passed into the kernel and to a specific device by
478calling ioctl(SIOCSHWTSTAMP) with a pointer to a struct ifreq whose
479ifr_data points to a struct hwtstamp_config. The tx_type and
480rx_filter are hints to the driver what it is expected to do. If
481the requested fine-grained filtering for incoming packets is not
482supported, the driver may time stamp more than just the requested types
483of packets.
484
485Drivers are free to use a more permissive configuration than the requested
486configuration. It is expected that drivers should only implement directly the
487most generic mode that can be supported. For example if the hardware can
488support HWTSTAMP_FILTER_V2_EVENT, then it should generally always upscale
489HWTSTAMP_FILTER_V2_L2_SYNC_MESSAGE, and so forth, as HWTSTAMP_FILTER_V2_EVENT
490is more generic (and more useful to applications).
491
492A driver which supports hardware time stamping shall update the struct
493with the actual, possibly more permissive configuration. If the
494requested packets cannot be time stamped, then nothing should be
495changed and ERANGE shall be returned (in contrast to EINVAL, which
496indicates that SIOCSHWTSTAMP is not supported at all).
497
498Only a processes with admin rights may change the configuration. User
499space is responsible to ensure that multiple processes don't interfere
500with each other and that the settings are reset.
501
502Any process can read the actual configuration by passing this
503structure to ioctl(SIOCGHWTSTAMP) in the same way.  However, this has
504not been implemented in all drivers.
505
506::
507
508    /* possible values for hwtstamp_config->tx_type */
509    enum {
510	    /*
511	    * no outgoing packet will need hardware time stamping;
512	    * should a packet arrive which asks for it, no hardware
513	    * time stamping will be done
514	    */
515	    HWTSTAMP_TX_OFF,
516
517	    /*
518	    * enables hardware time stamping for outgoing packets;
519	    * the sender of the packet decides which are to be
520	    * time stamped by setting SOF_TIMESTAMPING_TX_SOFTWARE
521	    * before sending the packet
522	    */
523	    HWTSTAMP_TX_ON,
524    };
525
526    /* possible values for hwtstamp_config->rx_filter */
527    enum {
528	    /* time stamp no incoming packet at all */
529	    HWTSTAMP_FILTER_NONE,
530
531	    /* time stamp any incoming packet */
532	    HWTSTAMP_FILTER_ALL,
533
534	    /* return value: time stamp all packets requested plus some others */
535	    HWTSTAMP_FILTER_SOME,
536
537	    /* PTP v1, UDP, any kind of event packet */
538	    HWTSTAMP_FILTER_PTP_V1_L4_EVENT,
539
540	    /* for the complete list of values, please check
541	    * the include file include/uapi/linux/net_tstamp.h
542	    */
543    };
544
5453.1 Hardware Timestamping Implementation: Device Drivers
546--------------------------------------------------------
547
548A driver which supports hardware time stamping must support the
549SIOCSHWTSTAMP ioctl and update the supplied struct hwtstamp_config with
550the actual values as described in the section on SIOCSHWTSTAMP.  It
551should also support SIOCGHWTSTAMP.
552
553Time stamps for received packets must be stored in the skb. To get a pointer
554to the shared time stamp structure of the skb call skb_hwtstamps(). Then
555set the time stamps in the structure::
556
557    struct skb_shared_hwtstamps {
558	    /* hardware time stamp transformed into duration
559	    * since arbitrary point in time
560	    */
561	    ktime_t	hwtstamp;
562    };
563
564Time stamps for outgoing packets are to be generated as follows:
565
566- In hard_start_xmit(), check if (skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP)
567  is set no-zero. If yes, then the driver is expected to do hardware time
568  stamping.
569- If this is possible for the skb and requested, then declare
570  that the driver is doing the time stamping by setting the flag
571  SKBTX_IN_PROGRESS in skb_shinfo(skb)->tx_flags , e.g. with::
572
573      skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
574
575  You might want to keep a pointer to the associated skb for the next step
576  and not free the skb. A driver not supporting hardware time stamping doesn't
577  do that. A driver must never touch sk_buff::tstamp! It is used to store
578  software generated time stamps by the network subsystem.
579- Driver should call skb_tx_timestamp() as close to passing sk_buff to hardware
580  as possible. skb_tx_timestamp() provides a software time stamp if requested
581  and hardware timestamping is not possible (SKBTX_IN_PROGRESS not set).
582- As soon as the driver has sent the packet and/or obtained a
583  hardware time stamp for it, it passes the time stamp back by
584  calling skb_hwtstamp_tx() with the original skb, the raw
585  hardware time stamp. skb_hwtstamp_tx() clones the original skb and
586  adds the timestamps, therefore the original skb has to be freed now.
587  If obtaining the hardware time stamp somehow fails, then the driver
588  should not fall back to software time stamping. The rationale is that
589  this would occur at a later time in the processing pipeline than other
590  software time stamping and therefore could lead to unexpected deltas
591  between time stamps.
592
5933.2 Special considerations for stacked PTP Hardware Clocks
594----------------------------------------------------------
595
596There are situations when there may be more than one PHC (PTP Hardware Clock)
597in the data path of a packet. The kernel has no explicit mechanism to allow the
598user to select which PHC to use for timestamping Ethernet frames. Instead, the
599assumption is that the outermost PHC is always the most preferable, and that
600kernel drivers collaborate towards achieving that goal. Currently there are 3
601cases of stacked PHCs, detailed below:
602
6033.2.1 DSA (Distributed Switch Architecture) switches
604^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
605
606These are Ethernet switches which have one of their ports connected to an
607(otherwise completely unaware) host Ethernet interface, and perform the role of
608a port multiplier with optional forwarding acceleration features.  Each DSA
609switch port is visible to the user as a standalone (virtual) network interface,
610and its network I/O is performed, under the hood, indirectly through the host
611interface (redirecting to the host port on TX, and intercepting frames on RX).
612
613When a DSA switch is attached to a host port, PTP synchronization has to
614suffer, since the switch's variable queuing delay introduces a path delay
615jitter between the host port and its PTP partner. For this reason, some DSA
616switches include a timestamping clock of their own, and have the ability to
617perform network timestamping on their own MAC, such that path delays only
618measure wire and PHY propagation latencies. Timestamping DSA switches are
619supported in Linux and expose the same ABI as any other network interface (save
620for the fact that the DSA interfaces are in fact virtual in terms of network
621I/O, they do have their own PHC).  It is typical, but not mandatory, for all
622interfaces of a DSA switch to share the same PHC.
623
624By design, PTP timestamping with a DSA switch does not need any special
625handling in the driver for the host port it is attached to.  However, when the
626host port also supports PTP timestamping, DSA will take care of intercepting
627the ``.ndo_do_ioctl`` calls towards the host port, and block attempts to enable
628hardware timestamping on it. This is because the SO_TIMESTAMPING API does not
629allow the delivery of multiple hardware timestamps for the same packet, so
630anybody else except for the DSA switch port must be prevented from doing so.
631
632In code, DSA provides for most of the infrastructure for timestamping already,
633in generic code: a BPF classifier (``ptp_classify_raw``) is used to identify
634PTP event messages (any other packets, including PTP general messages, are not
635timestamped), and provides two hooks to drivers:
636
637- ``.port_txtstamp()``: The driver is passed a clone of the timestampable skb
638  to be transmitted, before actually transmitting it. Typically, a switch will
639  have a PTP TX timestamp register (or sometimes a FIFO) where the timestamp
640  becomes available. There may be an IRQ that is raised upon this timestamp's
641  availability, or the driver might have to poll after invoking
642  ``dev_queue_xmit()`` towards the host interface. Either way, in the
643  ``.port_txtstamp()`` method, the driver only needs to save the clone for
644  later use (when the timestamp becomes available). Each skb is annotated with
645  a pointer to its clone, in ``DSA_SKB_CB(skb)->clone``, to ease the driver's
646  job of keeping track of which clone belongs to which skb.
647
648- ``.port_rxtstamp()``: The original (and only) timestampable skb is provided
649  to the driver, for it to annotate it with a timestamp, if that is immediately
650  available, or defer to later. On reception, timestamps might either be
651  available in-band (through metadata in the DSA header, or attached in other
652  ways to the packet), or out-of-band (through another RX timestamping FIFO).
653  Deferral on RX is typically necessary when retrieving the timestamp needs a
654  sleepable context. In that case, it is the responsibility of the DSA driver
655  to call ``netif_rx_ni()`` on the freshly timestamped skb.
656
6573.2.2 Ethernet PHYs
658^^^^^^^^^^^^^^^^^^^
659
660These are devices that typically fulfill a Layer 1 role in the network stack,
661hence they do not have a representation in terms of a network interface as DSA
662switches do. However, PHYs may be able to detect and timestamp PTP packets, for
663performance reasons: timestamps taken as close as possible to the wire have the
664potential to yield a more stable and precise synchronization.
665
666A PHY driver that supports PTP timestamping must create a ``struct
667mii_timestamper`` and add a pointer to it in ``phydev->mii_ts``. The presence
668of this pointer will be checked by the networking stack.
669
670Since PHYs do not have network interface representations, the timestamping and
671ethtool ioctl operations for them need to be mediated by their respective MAC
672driver.  Therefore, as opposed to DSA switches, modifications need to be done
673to each individual MAC driver for PHY timestamping support. This entails:
674
675- Checking, in ``.ndo_do_ioctl``, whether ``phy_has_hwtstamp(netdev->phydev)``
676  is true or not. If it is, then the MAC driver should not process this request
677  but instead pass it on to the PHY using ``phy_mii_ioctl()``.
678
679- On RX, special intervention may or may not be needed, depending on the
680  function used to deliver skb's up the network stack. In the case of plain
681  ``netif_rx()`` and similar, MAC drivers must check whether
682  ``skb_defer_rx_timestamp(skb)`` is necessary or not - and if it is, don't
683  call ``netif_rx()`` at all.  If ``CONFIG_NETWORK_PHY_TIMESTAMPING`` is
684  enabled, and ``skb->dev->phydev->mii_ts`` exists, its ``.rxtstamp()`` hook
685  will be called now, to determine, using logic very similar to DSA, whether
686  deferral for RX timestamping is necessary.  Again like DSA, it becomes the
687  responsibility of the PHY driver to send the packet up the stack when the
688  timestamp is available.
689
690  For other skb receive functions, such as ``napi_gro_receive`` and
691  ``netif_receive_skb``, the stack automatically checks whether
692  ``skb_defer_rx_timestamp()`` is necessary, so this check is not needed inside
693  the driver.
694
695- On TX, again, special intervention might or might not be needed.  The
696  function that calls the ``mii_ts->txtstamp()`` hook is named
697  ``skb_clone_tx_timestamp()``. This function can either be called directly
698  (case in which explicit MAC driver support is indeed needed), but the
699  function also piggybacks from the ``skb_tx_timestamp()`` call, which many MAC
700  drivers already perform for software timestamping purposes. Therefore, if a
701  MAC supports software timestamping, it does not need to do anything further
702  at this stage.
703
7043.2.3 MII bus snooping devices
705^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
706
707These perform the same role as timestamping Ethernet PHYs, save for the fact
708that they are discrete devices and can therefore be used in conjunction with
709any PHY even if it doesn't support timestamping. In Linux, they are
710discoverable and attachable to a ``struct phy_device`` through Device Tree, and
711for the rest, they use the same mii_ts infrastructure as those. See
712Documentation/devicetree/bindings/ptp/timestamper.txt for more details.
713
7143.2.4 Other caveats for MAC drivers
715^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
716
717Stacked PHCs, especially DSA (but not only) - since that doesn't require any
718modification to MAC drivers, so it is more difficult to ensure correctness of
719all possible code paths - is that they uncover bugs which were impossible to
720trigger before the existence of stacked PTP clocks.  One example has to do with
721this line of code, already presented earlier::
722
723      skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
724
725Any TX timestamping logic, be it a plain MAC driver, a DSA switch driver, a PHY
726driver or a MII bus snooping device driver, should set this flag.
727But a MAC driver that is unaware of PHC stacking might get tripped up by
728somebody other than itself setting this flag, and deliver a duplicate
729timestamp.
730For example, a typical driver design for TX timestamping might be to split the
731transmission part into 2 portions:
732
7331. "TX": checks whether PTP timestamping has been previously enabled through
734   the ``.ndo_do_ioctl`` ("``priv->hwtstamp_tx_enabled == true``") and the
735   current skb requires a TX timestamp ("``skb_shinfo(skb)->tx_flags &
736   SKBTX_HW_TSTAMP``"). If this is true, it sets the
737   "``skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS``" flag. Note: as
738   described above, in the case of a stacked PHC system, this condition should
739   never trigger, as this MAC is certainly not the outermost PHC. But this is
740   not where the typical issue is.  Transmission proceeds with this packet.
741
7422. "TX confirmation": Transmission has finished. The driver checks whether it
743   is necessary to collect any TX timestamp for it. Here is where the typical
744   issues are: the MAC driver takes a shortcut and only checks whether
745   "``skb_shinfo(skb)->tx_flags & SKBTX_IN_PROGRESS``" was set. With a stacked
746   PHC system, this is incorrect because this MAC driver is not the only entity
747   in the TX data path who could have enabled SKBTX_IN_PROGRESS in the first
748   place.
749
750The correct solution for this problem is for MAC drivers to have a compound
751check in their "TX confirmation" portion, not only for
752"``skb_shinfo(skb)->tx_flags & SKBTX_IN_PROGRESS``", but also for
753"``priv->hwtstamp_tx_enabled == true``". Because the rest of the system ensures
754that PTP timestamping is not enabled for anything other than the outermost PHC,
755this enhanced check will avoid delivering a duplicated TX timestamp to user
756space.
757