When loading the state of a Route fails (here
"unknown enumerator SurroundMaster in PBD::EnumWriter"), the
routes which have already been loaded are not added to the
Session's routelist.
Already existing routes that have an InternalSend or have
a circular reference:
The Send's `_send_from` holds a shared pointer
`<Route>(shared_from_this())` to the Route, and the
Route's ProcessorList contains the InternalSend.
This leads to various
"SessionHandleRef exists across session deletion" of
IO, Ports, GraphNode, Graph, etc
which causes issues when loading another session.
Session::destroy() cleans calls drop_references for
each route in the RouteList, which breaks the circular
dependency (InternalSend drops reference to Route).
But here the RouteList is empty.
Crash fixed:
* Load a session that fails to load a Route
(here a session created on with the vapor branch, on master)
* Then load another session without restarting Ardour.
When copying a mono LV2 plugin to a stereo track, the
state of the copied mono plugin is copied to the replicated
instance.
However at this point in time PBD::Stateful::ForceIDRegeneration
is still enabled, and the state ID has not been set.
This used to not be an issue. Older versions of liblilv
handle non-existent paths just fine.
However in lilv v0.24.20-10-gdd5e851 `lilv_path_absolute`
was replaced with `zix_canonical_path` which returns NULL
if a state file does not [yet] exist. This lead to a segfault
due to strlen(NULL) in `serd_node_new_file_uri`:
#0 strlen -- SEGV on unknown address 0x000000000000
#1 serd_node_new_file_uri () at /lib/x86_64-linux-gnu/libserd-0.so.0
#2 lilv_state_new_from_file () at /lib/x86_64-linux-gnu/liblilv-0.so.0
#3 ARDOUR::LV2Plugin::set_state(XMLNode const&, int) at ../libs/ardour/lv2_plugin.cc:2320
* yabridge runs the plugin's process function in a dedicated
bridged thread. Ardour's process thread is not (it just waits)
* When a plugin calls `restartComponent` from the realtime
thread. yabridge uses a host notification thread to perform
the callback.
Unlike other VST3 implementations that use a notification thread
(eg. JUCE), yabridge blocks and waits for the notification to
complete before the realtime thread can continue.
This leads to a deadlock.
However, we know that yabridge always synchronizes the
callback and concurrent calls are prevented by yabridge's design.
https://github.com/robbert-vdh/yabridge/issues/266
Locations::ripple can never add/remove markers, hence
Locations::changed is not applicable.
That signal is to indicate when more than one location is
added or removed from the location list.
Undo sets the state of ALL Locations, which resulted in
at least two signals for each Location (name changed,
start+end changed), even if there was no change.
This is in preparation to reduce signals during
Location Drag motion (which operates on a copied
Location, that causes [static] signal emissions,
but the location itself is not in any list).
