On 01/05/2021 18:32 Ryan Beethe ryan@splintermail.com wrote:
I'm interested in writing a custom imap command that behaves a bit like IDLE but synchronizes some state that is specific to my mail client / mail server.
I found that stateless commands were trivial to understand, and I really like the plugin pattern for registering custom commands.
But I have a few questions on how to write a long-running command that I was not able to answer by reading through the code. This mailing list seems like the best place to ask them.
For reference, the source code for my custom command is here:
https://github.com/Splintermail/splintermail-client/blob/dev/server/xkeysync.cThanks,
Ryan
Why does cmd-idle.c sometimes call client_command_free()? But sometimes it doesn't?
For example, cmd_idle_continue() frees it in some branches but not others. That makes no sense to me; it seems like it should be based on your entrypoint (mailbox notify callback vs input callback vs timeout callback), not based on which branch of logic within that entrypoint.
Why does cmd-idle.c ever call client_destroy()? That seems like something that should be invoked only by the imap process, not by any command.
It calls it in cmd-idle.c:idle_callback (which is a mailbox notify callback). It invokes it after idle_sync_now() when it detects that client->disconnected is set. Maybe that happens in imap_sync_init() or something?
Why does cmd-idle.c ever call client_disconnect()? That also seems like the responsibility of the imap process, and not any command.
idle_client_input_more() detects when i_stream_read returns -1, meaning that the client has *already disconnected*. Then it calls client_disconnect().
I think this is the crazy part... the istream is effectively unique to the imap process, so it seems unreasonable that any command is responsible for cleaning it up; it should just always happen at the imap process level before exiting, right?
IDLE cmd can be sometimes delegated to a separate worker called imap-hibernate, in which case the connection is moved to another process. This explains about all your questions.
What does client_continue_pending_input() actually do, and under what conditions does it need to be called?
There is one place that you *can't* call it; there is a section in imap-client.c:client_handle_input() that calls imap-client.c:client_handle_next_command(), which calls the cmd->func(). That makes sense; that's the input trigger for the command plugin, so maybe you only have to trigger it when you are receiving input that doesn't fit into the normal command args behavior.
It has a comment that says "this function is called at the end of I/O callbacks (and only there)". It _is_ called by client_input() and by client_output(), but also by: that even exist, so that explains them. I can't explain the
- cmd-idle.c:idle_client_input (io_add_istream callback)
- cmd-append.c:client_input_append (io_add_istream callback)
- imap-search.c:cmd_search_more_callback (timeout_add callback) The first two cases seem to be the only io_add_istream() commands
imap-search.c case at all.
Reading through it I have really no idea what client_continue_pending_input is really doing.
My command has a DONE mechanic just like IDLE so I'm pretty sure I need to invoke this function, I'm just concerned I'm going to do it wrong if I don't understand the mechanics of it.
It means that you did not consume all the input there was.
You probably shold look some much more simple commands as insipiration. Try looking e.g. how cmd_id is implemented instead.
Aki