Before opening new IPv4/IPv6 server sockets, chronyd will check for
matching reusable sockets passed from the service manager (for example,
passed via systemd socket activation:
https://www.freedesktop.org/software/systemd/man/latest/sd_listen_fds.html)
and use those instead.
Aside from IPV6_V6ONLY (which cannot be set on already-bound sockets),
the daemon sets the same socket options on reusable sockets as it would
on sockets it opens itself.
Unit tests test the correct parsing of the LISTEN_FDS environment
variable.
Add 011-systemd system test to test socket activation for DGRAM and
STREAM sockets (both IPv4 and IPv6). The tests use the
systemd-socket-activate test tool, which has some limitations requiring
workarounds discussed in inline comments.
Add two new fields to the NTP_Local_Timestamp structure:
- receive duration as the time it takes to receive the ethernet frame,
currently known only with HW timestamping
- network correction as a generalized PTP correction
The PTP correction is provided by transparent clocks in the correction
field of PTP messages to remove the receive, processing and queueing
delays of network switches and routers. Only one-step end-to-end unicast
transparent clocks are useful for NTP-over-PTP. Two-step transparent
clocks use follow-up messages and peer-to-peer transparent clocks don't
handle delay requests.
The RX duration will be included in the network correction to compensate
for asymmetric link speeds of the server and client as the NTP RX
timestamp corresponds to the end of the reception (in order to
compensate for the asymmetry in the normal case when no corrections
are applied).
Clients sockets are closed immediately after receiving valid response.
Don't wait for the first early HW TX timestamp to enable waiting for
late timestamps. It may take a long time or never come if the HW/driver
is consistently slow. It's a chicken and egg problem.
Instead, simply check if HW timestamping is enabled on at least one
interface. Responses from NTP sources on other interfaces will always be
saved (for 1 millisecond by default).
Initialize the unused part of shorter server NTS keys (AES-128-GCM-SIV)
loaded from ntsdumpdir to avoid sending uninitialized data in requests
to the NTS-KE helper process.
Do that also for newly generated keys in case the memory will be
allocated dynamically.
Fixes: b1230efac3 ("nts: add support for encrypting cookies with AES-128-GCM-SIV")
Rework handling of late HW TX timestamps. Instead of suspending reading
from client-only sockets that have HW TX timestamping enabled, save the
whole response if it is valid and a HW TX timestamp was received for the
source before. When the timestamp is received, or the configurable
timeout is reached, process the saved response again, but skip the
authentication test as the NTS code allows only one response per
request. Only one valid response per source can be saved. If a second
valid response is received while waiting for the timestamp, process both
responses immediately in the order they were received.
The main advantage of this approach is that it works on all sockets, i.e.
even in the symmetric mode and with NTP-over-PTP, and the kernel does
not need to buffer invalid responses.
Add a structure for 64-bit integers without requiring 64-bit alignment
to be usable in CMD_Reply without struct packing.
Add utility functions for conversion to/from network order. Avoid using
be64toh() and htobe64() as they don't seem to be available on all
supported systems.
If the authenticator SIV encryption fails (e.g. due to wrong nonce
length), decrement the number of extension fields to keep the packet
info consistent.
Keep a server SIV instance for each available algorithm.
Select AES-128-GCM-SIV if requested by NTS-KE client as the first
supported algorithm.
Instead of encoding the AEAD ID in the cookie, select the algorithm
according to the length of decrypted keys. (This can work as a long as
all supported algorithms use keys with different lengths.)
If AES-128-GCM-SIV is available on the client, add it to the requested
algorithms in NTS-KE as the first (preferred) entry.
If supported on the server, it will make the cookies shorter, which
will get the length of NTP messages containing only one cookie below
200 octets. This should make NTS more reliable in networks where longer
NTP packets are filtered as a mitigation against amplification attacks
exploiting the ntpd mode 6/7 protocol.
While AES-SIV-CMAC allows nonces of any length, AES-GCM-SIV requires
exactly 12 bytes, which is less than the unpadded minimum length of 16
used in the NTS authenticator field. These functions will be needed to
support both ciphers in the NTS code.
This is a newer nonce misuse-resistant cipher specified in RFC 8452,
which is now supported in the development code of the Nettle library.
The advantages over AES-SIV-CMAC-256 are shorter keys and better
performance.
If the randomly generated timestamps are close to the current time, the
source can be selected for synchronization, which causes a crash when
logging the source name due to uninitialized ntp_sources.
Specify the source with the noselect option to prevent selection.
Call the function with current time instead of latest sample of the
first source to avoid undefined conversion of negative double to long
int.
Fixes: 07600cbd71 ("test: extend sources unit test")
With the first interleaved response coming after a basic response the
client is forced to select the four timestamps covering most of the last
polling interval, which makes measured delay very sensitive to the
frequency offset between server and client. To avoid corrupting the
minimum delay held in sourcestats (which can cause testC failures),
reject the first interleaved response in the client/server mode as
failing the test A.
This does not change anything for the symmetric mode, where both sets of
the four timestamps generally cover a significant part of the polling
interval.
Estimate the 1st and 2nd 10-quantile of the reading delay and accept
only readings between them unless the error of the offset predicted from
previous samples is larger than the minimum reading error. With the 25
PHC readings per ioctl it should combine about 2-3 readings.
This should improve hwclock tracking and synchronization stability when
a PHC reading delay occasionally falls below the normal expected
minimum, or all readings in the batch are delayed significantly (e.g.
due to high PCIe load).
Add estimation of quantiles using the Frugal-2U streaming algorithm
(https://arxiv.org/pdf/1407.1121v1.pdf). It does not need to save
previous samples and adapts to changes in the distribution.
Allow multiple estimates of the same quantile and select the median for
better stability.
Move processing of PHC readings from sys_linux to hwclock, where
statistics can be collected and filtering improved.
In the PHC refclock driver accumulate the samples even if not in the
external timestamping mode to update the context which will be needed
for improved filtering.
Allow sources to accumulate samples with the leap status set to not
synchronized. Define a new state for them to be ignored in the
selection. This is intended for sources that are never synchronized and
will be used only for stabilization.
Avoid searching the hash table of sources when a packet in the client
mode is received. It cannot be a response from our source. Analogously,
avoid source lookups for transmitted packets in the server mode. This
doesn't change anything for packets in symmetric modes, which can be
requests and responses at the same time.
This slightly improves the maximum packet rate handled as a server.
Close /dev/urandom and drop cached getrandom() data after forking helper
processes to avoid them getting the same sequence of random numbers
(e.g. two NTS-KE helpers generating cookies with identical nonces).
arc4random() is assumed to be able to detect forks and reseed
automatically.
This is not strictly necessary with the current code, which does not use
the GetRandom functions before the NTS-KE helper processes are forked,
but that could change in future.
Also, call the reset function before exit to close /dev/urandom in order
to avoid valgrind reporting the file object as "still reachable".
The maximum value of the new 32-bit fields is slightly less than 16,
which can cause the NTP test #7 to pass for a server which has a zero
root delay but maximum root dispersion.
Interpret the maximum value as the maximum value of the original 32-bit
fields (~65536.0 seconds) for better compatibility with NTPv4.
Since commit fdfcabd79b ("ntp: drop support for long NTPv4 MACs"), the
parser doesn't need to check validify of MACs in NTPv4 packets to
distinguish them from extension fields. Move the parser to ntp_core to
avoid having a separate iteration looking for non-authentication
extension fields.
When the server clock was updated between saving of the RX timestamp and
updating the TX timestamp, a client using interleaved mode with the four
timestamps which minimize error in measured delay (e.g. chrony) had the
server clock adjustment included in the measured delay, which could
disrupt the sample filtering and weighting.
Add a handler to track the slew epoch and remember the last offset. Undo
the adjustment in TX timestamps which have their RX timestamp in the
previous epoch to fix the delay observed by the clients.
If an unknown clock step is detected, drop all timestamps.
Instead of keeping one pair of RX and TX timestamp for each address, add
a separate RX->TX map using an ordered circular buffer. Save the RX
timestamps as 64-bit integers and search them with a combined linear
interpolation and binary algorithm.
This enables the server to support multiple interleaved clients sharing
the same IP address (e.g. NAT) and it will allow other improvements to
be implemented later. A drawback is that a single broken client sending
interleaved requests at a high rate (without spoofing the source
address) can now prevent clients on other addresses from getting
interleaved responses.
The total number of saved timestamps does not change. It's still
determined by the clientloglimit directive. A new option may be added
later if needed. The whole buffer is allocated at once, but only on
first use to not waste memory on client-only configurations.
On some systems (e.g. Solaris/OpenIndiana) rand() and random() have
different ranges. RAND_MAX is the maximum value returned by rand(),
but random() should always have a range of 0 through 2^31-1.
This fixes multiple failures in different tests.
gnutls running in the FIPS140-2 mode does not allow MD5 to be
initialized, which breaks chronyd using MD5 to calculate reference ID
of IPv6 addresses. Specify a new hash algorithm for non-security MD5 use
and temporarily switch to the lax mode when initializing the hash
function.
Don't require timespec/timeval-double conversion tests to produce
correctly rounded results to handle x86 and other archs with wider
intermediate results.
