| Age | Commit message (Collapse) | Author |
|---|
|
gettimeofday(). It was also bugging me for some time to have all
these checks of gettimeofday()'s return value: it should not fail. So
this diff introduces a void getmonotime(struct timeval *tv) that calls
clock_gettime(CLOCK_MONOTONIC, &ts) and converts the output to a
struct timeval that can be used with the existing code and the
timeval-specific timer functions (timerclear, timersub, ...). It does
not return a status but calls fatal() on error-that-should-not-happen.
ok sthen@ chris@
|
|
closes pr system/6627
diff submitted by Martin Matuska, thanks
ok benno@
|
|
the config. Solves a race at startup time where processes can send status
messages about hosts that other processes don't know about yet.
(and have relayd abort with "desynchronized" or "invalid host id")
ok henning pyr deraadt
solves the problem ok from benno todd
|
|
ok mikeb
|
|
daemon infrastructure. The previous design made it fairly hard to
reload the complex data structures, especially relays and protocols.
One of the reasons was that the privsep'd relayd processes had two
ways of getting their configuration: 1) from memory after forking from
the parent process and 2) and (partially) via imsgs after reload. The
new implementation first forks the privsep'd children before the
parents loads the configuration and sends it via imsgs to them; so it
is only like 2) before. It is based on an approach that I first
implemented for iked(8) and I also fixed many bugs in the code.
Thanks to many testers including dlg@ sthen@ phessler@
ok pyr@ dlg@ sthen@
|
|
functions that are based on work for iked and smtpd. This simplifies
the setup of privsep processes and moves some redundant and repeated
code to a single place - which is always good from a quality and
security point of view. The relayd version of proc.c is different to
the current version in iked because it uses 1:N communications between
processes, eg. a single parent process is talking to many forked relay
children while iked only needs 1:1 communications.
ok sthen@ pyr@
|
|
instead of static function names. __func__ is C99 and perfectly fine
to use. It also avoids printing errors; for example if a statement
log_debug("foo:"..) was moved or copied from function foo() to bar()
and the log message was not updated...
|
|
ok pyr@
|
|
resource limits to the maximum of the daemon class but the host check
process (hce/health checks) didn't and was limited to a fairly low
default of 128 open sockets (openfiles-cur=128 in login.conf). This
was reached fairly quickly with "check tcp" of many hosts. This diff
increases the maximum number of monitored hosts and concurrent health
checks in relayd in a significant way and may fix issues for people
that have around 100 or more hosts (or fewer hosts with multiple checked
ports).
tested by phessler@
ok jsg@
|
|
and follows a suggestion in event.h. also don't mix signal() and
signal_set()/signal_add().
ok jsg@ gilles@
|
|
like several other things in the tree.
ok reyk@ looks fine claudio@
|
|
Stop pushing event handling in the imsg framework.
Instead, provide a small glue layer on top of both imsg and libevent.
This finally clearly separates event handling and imsg construction.
Sidetrack bonus: remove the mega-ugly hack of having a dummy imsg_event_add
stub in relayctl.
This will make bgpd (and thus henning) happy. Next up are smtpd and ospfd.
ok eric@
|
|
|
|
imsg_event_add function to be provided (which ended up being a named
callback).
Instead provide a wrapper in the daemon and call that everywhere.
Previsously discussed with the usual suspects,
ok eric@ though not too happy about the function name
(imsg_compose_event).
|
|
caching can kick in on subsequent checks, making them faster and
lighter on the server.
From camield, closes PR 6137 (modified diff)
|
|
required because it is called later and there is no return before.
ok gilles@
|
|
So change the code accordingly to allow that.
Found by claudio@ in ospfd
|
|
From Camiel Dobbelaar, closes PR 6066
|
|
arbitrary string in debugging mode, it will store an error code
(HCE_*) for each host. the error code can be translated to a string
(in log.c) for debugging but it will also be passed to relayctl via
the control socket.
from a user point of view, this will print a human-readable error
message in the "relayctl show hosts" output if a host is down because
the check failed. the relayctl(8) manpage includes detailed
explanations of the error messages including mitigations for the
most-common problems.
ok jmc@ (manpages)
ok phessler@
|
|
|
|
OK reyk@
|
|
singly-linked SLIST instead. the only noticeable change is the
reversed order to notify the children but it does not really matter
here. also only walk through the children host list if the host
itself is a potential parent.
|
|
inherit the state from another host with the specified Id; no
additional check will be for the inheriting host. This helps in
scenarios with lots of IP aliases that all point to the same service
on the same host (like web hosting with many SSL domains).
discussed with pyr, tested in different setups
|
|
|
|
a SNMP trap when the state of a host changes. relayd(8) will try to
(re-)connect to snmpd(8) and request it to send a trap to the
registered trap receivers, see snmpd.conf(5) for more information
about the configuration.
ok pyr@ thib@
|
|
for code, start with struct relayd. finally.
ok thib@
|
|
(for instance: rename struct service to struct rdr), refer to redirects
otherwise (hoststatectl output).
ok reyk@
|
|
what the daemon does - it is a relayer that pays attention to the status
of pools of hosts; not a status checkers that happens to do some relaying
|
|
|
|
that check interval can differ from table to table.
ok reyk@
|
|
Table specific intervals must be multiples of the global interval.
help and ok reyk@
|
|
|
|
Especially useful when interval is rather long.
I was supposed to commit this before 4.2.
|
|
ok pyr@
|
|
|
|
libevent just to ignore it, use SIG_IGN instead.
this syncs hoststated with bgpd and (soon) ospfd.
|
|
of our socket pairs. Instead disable listening on the pipe, terminate the
event loop, and let the parent process's SIGCHLD handler do a clean
shutdown.
from an ospfd diff by claudio, ok claudio@
|
|
needed for layer 7 reload support.
ok pyr@
|
|
Hoststated can be reloaded either by sending SIGHUP to the parent process
or by using ``hoststatectl reload''
discussed and ok reyk@
|
|
at reconfiguration time.
|
|
|
|
split the code to start the event loop in two functions.
introduce merge_config which will be used later on.
|
|
for custom evaluations.
pyr agrees to put it in now but to do some improvements of the timeout
handling later.
|
|
allow purging of parts of the hoststated environment structure.
start using this function now to only keep vital information in
hoststated children processes.
ok reyk@
|
|
First split out hosts, tables and services into to structs, one that
contains the runtime fields and one (inside the runtime) that contains
mostly static fields that will be sent over the socket during reload.
Also move the demoted field of tables inside the flags field as its
just a boolean.
ok reyk@
|
|
* make parse_config allocate the hoststated function by itself
* make as many sockets as necessary to talk to the relay children
* add send_all for talking to all children
with advise and ok reyk@
|
|
layer 7 loadbalancing.
- allow to run relays with tables without depending on services
- show hosts and tables assigned to relays in hoststatectl show commands
ok pyr@ deraadt@ with some input from mcbride@
|
|
|
|
loadbalancing, SSL acceleration, general-purpose TCP relaying, and
transparent proxying.
see hoststated.conf(5) and my upcoming article on undeadly.org for
details.
ok to commit deraadt@ pyr@
|
|
notifications after completed host checks. either only log the
"updates" to new states or log "all" state notifications, even if the
state didn't change. the log messages will be reported to syslog or to
stderr if the daemon is running in foreground mode.
ok claudio@ pyr@
|
