When checking of a 'foot --server' instance is already running, we try
to connect to the UNIX socket we're planning on listening on.
In most cases, this will fail hard and fast. But under certain
circumstances, we can get stuck in connect() waiting for a connection
timeout.
Since it should be possible to establish a connection immediately *if*
there's someone actually listening on the socket, rely on the fact
that connect() will fail with a EINPROGRESS if a connection can *not*
be established immediately (and thus no one is listening on it).
Since fonts are cached, this adds no additional memory. However, it
makes the first terminal window in --server mode start much faster,
since the (primary) fonts have already been loaded.
Fallback fonts are still loaded on-demand.
While unusual, it *is* possible for a client *not* to terminate when
we close ptmx.
We need to handle this *somehow*. Since it is so unusual, we'll go
with a fairly easy, but synchronous method:
* Register a signal handler for SIGALRM, and setup a 2 second alarm
* Wait for slave to die
* If it didn't die, sent SIGTERM, then re-set the alarm for another 2
seconds.
* If it still hasn't died, send SIGKILL (this time without an alarm).
When there hasn't been a timeout (or in our case, there was a timeout,
but we reset the timer before we got to the read()), read() returns,
not 0, but -1 with errno == EAGAIN.
Since we cancel the timers every now and then, there's a (small)
chance that one handler cancels a timer that has triggered in the same
epoll() iteration.
When this happens, read() blocks.
Fix by making the timer FDs non-blocking, and simply returning when we
read 0 bytes.
When we need to create a new buffer (because the cache doesn't have
any buffers of correct size, or because they're all busy), purge
buffers with a size mismatch.
Each font instance has a ref-counter. Whenever we want to instantiate
a font that has already been loaded, we instead return the
already-loaded instance, and bump the ref counter.
When the last font instance is destroyed, it is also removed from the
cache.
In this mode, foot listens on a UNIX socket and creates terminal
windows when clients connect.
A connecting client sends argc/argv to the server, and the server
instantiates a new terminal window.
When the terminal window is closed, the exit code is sent back to the
client.
This will trigger e.d. keyboard_leave() and wl_pointer_leave() events,
which ensures there aren't any references to the destroyed window from
the global wayland struct.
Call wl_display_roundtrip() to trigger those events *before* we
destroy the window.
It is perfectly possible, and legal, for a FDM handler to delete
another handler. The problem however is when the epoll returned array
of FD events contain the removed FD handler *after* the handler that
removed it.
That is, if the epoll returned array is:
[FD=13, FD=37]
and the handler for FD=13 removes FD=37, then given the current
implementation, the FD user data (our handler callback etc) will point
to a free:d address.
Add support for this situation by deferring FD handler removal *when
called from another handler*.
This is done by "locking" the FDM struct before looping the handlers
for FDs with events (and unlocking afterwards).
In fdm_del(), we check if the FDM has been locked, in which case the
FD is marked as deleted, and put on a deferred list. But
not *actually* deleted.
Meanwhile, in the FDM poll function, we skip calling handlers for
marked-for-deletion FDs.
Then, when all FDs have been processed, we loop the deferred list and
finally deletes the FDs for real.
