Combine tcell events with WorkerMessages to better synchronize state
with IO and UI. Remove Tick loop for rendering. Use events to trigger
renders.
Signed-off-by: Tim Culverhouse <tim@timculverhouse.com>
Acked-by: Robin Jarry <robin@jarry.cc>
The worker uses a buffered channel to queue tasks. Buffered channels
are effective at FIFO, but are prone to blocking. The design of aerc is
such that the UI must always accept a response from the backends, and
the backends must always accept a request from the UI. By using buffered
channels for both of these communication channels, a deadlock will
occur.
Break the chain by using a doubly linked list (container/list from the
standard library) to queue tasks for the worker. Essentially, this is an
infinitely buffered channel - but more memory efficient as it can change
size dynamically.
Signed-off-by: Tim Culverhouse <tim@timculverhouse.com>
Signed-off-by: Robin Jarry <robin@jarry.cc>
Commit 716ade8968 ("worker: lock access to callback maps") introduced
locks to the worker callback maps. The locks also locked the processing
of the callback, which had the unintended side effect of deadlocking the
worker if any callbacks attempted to post a new action or message.
Refactor the locks to only lock the worker while accessing the maps.
Fixes: 716ade8968 ("worker: lock access to callback maps")
Signed-off-by: Tim Culverhouse <tim@timculverhouse.com>
Acked-by: Robin Jarry <robin@jarry.cc>
Worker callbacks are inherently set and called from different
goroutines. Protect access to all callback maps with a mutex.
Signed-off-by: Moritz Poldrack <moritz@poldrack.dev>
Signed-off-by: Tim Culverhouse <tim@timculverhouse.com>
Acked-by: Robin Jarry <robin@jarry.cc>
When an error is encountered fetching a header, the backends respond
with a type.Error worker message. On receipt of this message, the UI
deletes all pending headers. The headers are all requested again as they
remain on the screen, resulting in an infinite request loop - and an
infinite logging loop. The user only ever sees the spinner unless they
check the logs.
A previous commit intended to fix this, however it introduced a
regression where any message that was part of the fetch request would
also be marked as erroneous. This commit is reverted with commit
2aad2fea7d36 ("msgstore: revert 9fdc7acf5b48").
Send an erroneous message info message from the backend when an error is
encountered for a specific UID.
Fixes: 01f80721e2 ("msgstore: post MessageInfo on erroneous fetch")
Signed-off-by: Tim Culverhouse <tim@timculverhouse.com>
Acked-by: Robin Jarry <robin@jarry.cc>
This reverts commit 7473571159.
The commit has introduced a regression that lead to the pager not being
filled with content, thereby making reading mails impossible.
Signed-off-by: Moritz Poldrack <git@moritz.sh>
Acked-by: Robin Jarry <robin@jarry.cc>
Send to worker.Messages in goroutine to prevent deadlocks: the UI can
fill the worker.Actions channel. The worker can generate more than one
Message per action, and if it generates enough to fill the
worker.Messages channel from a single message while the worker.Actions
channel is full, a deadlock occurs.
Signed-off-by: Tim Culverhouse <tim@timculverhouse.com>
Acked-by: Robin Jarry <robin@jarry.cc>
The IMAP worker has a few methods that post a new Action to itself. This
can create a deadlock when the worker.Actions channel is full: The
worker can't accept a new Action because it's trying to post an action.
This is most noticeable when cached headers are enabled and the message
list is scrolled fast.
Use a goroutine to post actions to the worker when posting from within
the worker.
Fixes: https://todo.sr.ht/~rjarry/aerc/45
Fixes: 7aa71d334b ("imap: add option to cache headers")
Signed-off-by: Tim Culverhouse <tim@timculverhouse.com>
Acked-by: Robin Jarry <robin@jarry.cc>
Do not pass logger objects around anymore. Shuffle some messages to make
them consistent with the new logging API. Avoid using %v when a more
specific verb exists for the argument types.
The loggers are completely disabled (i.e. Sprintf is not even called)
by default. They are only enabled when redirecting stdout to a file.
Signed-off-by: Robin Jarry <robin@jarry.cc>
Acked-by: Moritz Poldrack <moritz@poldrack.dev>
Worker.callbacks contains two types of callbacks: some are action callbacks,
some are message callbacks. Each of those is access from one side of the
communication channel (UI goroutine vs. worker goroutine).
Instead of using a channel, we can use two different maps for each kind. This
simplifies the code and also ensures we don't call an action callback instead
of a message callback (or the other way around).
Message IDs are allocated for both messages from UI to workers and the other
way around. Hence, the global nextId variable is accessed from multiple
goroutines.
Instead, use atomic to access the global counter.
Worker.Process* functions were called in different goroutines than
Worker.Post*. Protect the map with a mutex. Also make the map unexported to
prevent external unprotected accesses.
Worker.Process* functions used to delete items from the map. However they
didn't delete the element they retrieved: callbacks[msg.InResponseTo()] was
read while callbacks[msg] was deleted. I'm not sure I understand why. I tried
to delete the element that was accessed - but this broke everything (UI froze
at "Connecting..."). I don't believe any elements were actually removed from
the map, so the new code just doesn't remove anything.
Tim Culverhouse
Moritz Poldrack
Robin Jarry
Drew DeVault
Simon Ser