This is done by performing one synchronous query for carrier state at
the start of the program; after that, we just monitor the nlsocket for
carrier state changes and update the cached state accordingly.
The benefit is that ifchd needs to do a lot less work and this should
reduce the CPU cycle consumption; prior to this commit, the CPU time
ends up being a few CPU-minutes per month.
There's really no advantage to using signalfd in ndhc, particularly
since the normal POSIX signal API is now used for handling SIGCHLD in
ndhc-master. So just use the tried and true volatile sig_atomic_t set
and check approach.
The only intended behavior change is in the dhcp RELEASE state --
before there would be a spurious attempt at renewing a nonexistent
lease when the RENEW signal was received.
Suppose a system call such as bind() fails in the sockd subprocess in
request_sockd_fd(). sockd will suicide(). This will send a SIGCHLD to
the master process, which the master process should respond to by
calling suicide(), forcing a process supervisor to respawn the entire
ndhc program.
But, this doesn't reliably happen prior to this commit because of the
interaction between request_sock_fd() and signalfd() [or equivalently
self-pipe-trick] signal handling.
request_sock_fd() makes ndhc-master synchronously wait for a response
from sockd via safe_recvmsg(). The normal goto-like signal handling
path is suppressed when using signalfd() , so when SIGCHLD is received,
it will not be handled until io is dispatched for the signalfd or pipe.
But such code will never be reached because ndhc-master is waiting in
safe_recvmsg() and thus never polls signal fd status.
So, revert to using traditional POSIX sigaction() for SIGCHLD, which
provides exactly the required behavior for proper functioning.
Apparently I had forgotten the counter-intuitive semantics of
signalfd(): it's necessary to BLOCK the signals that will be
handled exclusively by signalfd() so that the default POSIX
signal handling mechanism won't intercept the signals first.
The lack of response to ctrl+c is a legitimate bug that is
now properly fixed; ba046c02c7 fixed that issue, but
regressed the handling of other signals.
'(c)' may not be a valid substitute for 'Copyright' in some legal
domains/interpretations. So be safe, since I obviously am asserting
copyright on my legal work.
It breaks with the existing whitelists on the latest glibc and is
just too much maintenance burden. It also causes the most questions
for new users.
Something like openbsd's pledge() would be fine, but I have no
intention of maintaining such a thing.
Most of the value-gain would come from disallowing high-risk
syscalls like ptrace() and the perf syscalls, anyway.
ndhc already uses extensive defense-in-depth and wasn't using
seccomp on non-(x86|x86-64) platforms, so it's not a huge loss.
ARP packets aren't split across multiple receive events, so
reply_offset is pointless, and we implicitly assume that the
previous ARP packet data is still available after a forced sleep.
We instead check carrier status as needed. This approach is more
robust. For a simple example, imagine link state changes that happen
while the machine is suspended.
There was no way to disable writing pidfiles before.
pidfiles are an unreliable method of tracking processes, anyway; process
supervisors are strongly recommended. If a pidfile is really needed, it
can be explicitly specified.
Before the exit code would be success and no error message would print,
and it required a bit of control flow tracing to determine what would
actually happen.
No direct functional change (unless the supervising process cares about
the return code of ndhc on exit).
Lease times and arp timeouts are all calculated using long long,
but epoll takes its timeout argument as an int. Guard against
timeouts > INT_MAX but < UINT_MAX wrapping and causing spuriously
short timeouts when converted to a signed int.
This problem has been observed in the wild. Thanks to thypon
for a detailed strace that pointed me towards this issue.
This corrects a bug where stale dhcp packets would get reprocessed,
causing very bad behavior; an issue that was introduced in the
coroutine conversion.
If ndhc were a high-performance program that handled lots of events,
this change would harm performance. But it is not, and it implicitly
believes that events come in one at a time. Processing batches
would make it harder to assure correctness while also never allocating
memory at runtime.
The previous structure was fine when everything was handled
immediately by callbacks, but it isn't now.
This change makes it much easier to reason about ndhc's behavior
and properly handle errors.
It is a very large changeset, but there is no way to make this
sort of change incrementally. Lease acquisition is tested to
work.
It is highly likely that some bugs were both introduced and
squashed here. Some obvious code cleanups will quickly follow.
This change paves the way for allowing ifch to notify the core ndhc
about failures. It would be far too difficult to reason about the
state machine if the requests to ifch were asynchronous.
Currently ndhc assumes that ifch requests never fail, but this
is not always true because of eg, rfkill.
