This commit introduces intruduces a small helper class that wraps any writer
and provides a flush operation that performs the corresponding action if the
writer is an AsyncJsonWriter and does nothing otherwise.
Replacing invalid UTF-8 characters beforehand by our selves doesn't make
any sense, the serializer can literally perform the same replacement ops
with the exact same Unicode replacement character (U+FFFD) on its own.
So, why not just use it directly? Instead of wasting memory on a temporary
`String` object to always UTF-8 validate every and each value, we just
use the serializer to directly to dump the replaced char (if any) into
the output writer. No memory waste, no fuss!
The Array, Dictionary, and Namespace types provide a Freeze() method that makes
them read-only. So far, there was the possibility to call some methods with
`overrideFrozen=true` which would then bypass the corresponding check and allow
modification of the data structures nonetheless.
With 24b57f0d3a222835178e88489eabd595755ed883, this possibility was already
removed from the Namespace type. However, for interface compatibility, it kept
the parameter and just ignores it, throwing an exception on any modification on
a frozen instance.
The only place using `overrideFrozen` was processing of the `-D`/`--define`
command line flag that allows setting additional variables in the DSL. At the
time it is evaluated, there are no user-created data structures yet that could
be frozen, so the only frozen objects that could be encountered are Namespaces
(Icinga doesn't freeze other types by itself) and for these, `overrideFrozen`
already has no effect.
Hence, there is no harm in removing `overrideFrozen` altogether. This
simplifies the code and also means that frozen objects are now indeed read-only
without exceptions, allowing further optimizations regarding locking in the
future.
The wait group gets passed to HttpServerConnection, then down to the
HttpHandlers. For those handlers that modify the program state, the
wait group is locked so ApiListener will wait on Stop() for the
request to complete. If the request iterates over config objects,
a further check on the state of the wait group is added to abort early
and not delay program shutdown. In that case, 503 responses will be
sent to the client.
Additionally, in HttpServerConnection, no further requests than the
one already started will be allowed once the wait group is joining.
CalcEventID's internal logic uses the TimestampToMilliseconds function
to convert the given eventTime to milliseconds. Within this function,
the timestamp is capped to prevent an overflow.
On three occasions, the input timestamp given to CalcEventID had already
been converted using TimestampToMilliseconds. The second
TimestampToMilliseconds function then checked the value and always
returned the capped maximum value. Consequently, CalcEventID returned
the same hash value for different timestamps.
This affected SendFlappingChange, SendAcknowledgementSet, and
SendAcknowledgementCleared. For example, when two acknowledgments were
created for the same service, the calculated event_id representing the
history table row would be identical.
Fixes#10465
Since the types and states attributes are user configurable and allowed to change at
runtime, we need to update the actual filter bitsets whenever these attributes change.
Otherwise, the filter bitsets would be stale and not reflect their current state.
No external user needs to manipulate the actual object dependency
graphs. This was maybe introduced for debugging purposes at that time
but if someone messes with this in prod - good luck with that. Oh, apart
from that it's broken :( and doesn't track parents as its implies but
children.
Not sure why it's introduced in the first place, maybe for debugging
purposes at the early stage of Icinga 2 dev but I failed to see an
actual useful use case for it that's worth its maintenance burden. So,
this commit dropped it entirely from the DSL language.
Boost `1.88.0` introduced a feature [^1] that makes use of the Windows API, but it
uses API functions that are only available with `PSAPI_VERSION >= 2` and
Windows VISTA only supports `PSAPI_VERSION == 1`. Actually, that new feature
can also be disabled by setting the `BOOST_STACKTRACE_DISABLE_OFFSET_ADDR_BASE`
macro, but since it seems to be a useful feature and isn't even disabled by default,
we can just drop it that ancient Windows version instead of disabling it.
[^1]: https://github.com/boostorg/stacktrace/pull/200
With this mutex, we can have deadlock in the following case:
1/ Thread A processes a /v1/actions/acknowledge-problem request and locks the checkable
2/ Thread B processes a /v1/actions/add-comment and enters first the ConfigItem::ActivateItems() method and locks the static mutex there and starts the just created comment object, which triggers the OnCommentAdded() event.
3/ Thread A wants to activate the just created ack comment as well but since the mutex is already locked by TB, it blocks.
4/ Thread B's OnCommentAdded() even dispatch causes the IcingaDB::CommentAddedHandler() to be called and implicitly triggers full state update for the checkable. Now, the state serialization of that checkable (remember that's the same checkable currently locked by TA) also includes computing its severity, thus it calls either service->GetSeverity() or host->GetSeverity(). However, since computing the checkable severity (as of now) requires acquiring the object lock, and boom - they deadlock each other.
