I was able to reproduce #3579 in Linux by running:
`sudo sysctl net.core.wmem_default=10000`
If a subscription message was too big to be sent at once, it was
possible to break a client by sending a reply to an other message sent
by the client. Eg:
- Write 8192 out of 11612 bytes of a workspace event.
- Blockingly write the reply to a workspace change message.
- Write the rest 3420 bytes of the workspace event.
This commit fixes this by utilizing the ipc queue for all types of
writes.
ipc_receive_message can only be called from a callback started in
ipc_new_client. This callback uses the same file descriptor with the
client also created in ipc_new_client. When the client is deleted, the
read callback is now also stopped. Thus, we can assume that whenever
ipc_receive_message is called, the corresponding client should still
exist.
- ipc_client now contains pointers to both write and read watchers. When
freed, a client will stop both of them.
- IPC_HANDLERs now work with ipc_clients instead of fds.
Fixes#3579.
5aa0459be0