We now keep multiple commands for a json_connection, and an array of
json_streams.
When a command wants to write something, we allocate a new json_stream
at the end of the array.
We always output from the first available json_stream; once that
command has finished, we free that and move to the next. Once all are
done, we wake the reader.
This means we won't read a new command if output is still pending, but
as most commands don't start writing until they're ready to write
everything, we still get command parallelism.
In particular, you can now 'waitinvoice' and 'delinvoice' and it will
work even though the 'waitinvoice' blocks.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
json_stream_success / json_stream_fail belong in jsonrpc.c, and the
json_tok helpers for special types belong in json.x
json_add_object() isn't used, remove it rather than moving it.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We promote 'struct json_stream' to contain the membuf; we only attach
the json_stream to the command when we actually call
json_stream_success / json_stream_fail.
This means we are closer to 'struct json_stream' being an independent
layer; the tests are already modified to use it directly to create
JSON.
This is also the first step toward re-enabling non-serial command
execution.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This doesn't make a performance difference, but even better, it
simplifies the code.
We hacked test_multirpc to send 200x as many commands, and timed the
pytest over 20 runs:
Before:
=================== 1 passed, 136 deselected in 8.550000-9.400000(9.0045+/-0.2) seconds ===================
After:
=================== 1 passed, 136 deselected in 8.540000-9.370000(8.97286+/-0.16) seconds ===================
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We need to keep the remaining buffer, and we need to try to parse it
before we read the next. I first tried keeping it in the object, but
its lifetime is that of the *socket*, which we actually reopen for
every command.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This was hanging sometimes in travis, but actually checking the result
of the commands makes it *always* hang. We remove the waitinvoice
which will not return.
ZmnSCPxj points out that this behavior, introduced in
ce0bd7abd3, is a regression: it would be
nice to be able to cancel a waitinvoice. But that fix is more complex,
and will have to be another PR.
This test will now hang, but it's OK: we're about to fix it!
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This is the final step to get the plugins working. After parsing the
early options (including `--plugin`), then starting and asking the
plugins for options, and finally reading in the options we just
registered, we just need to assemble the options and send them over.
Signed-off-by: Christian Decker <@cdecker>
We also make `--help` a non-early arg so it allows for the plugins to
register their options before printing the help message. The options
themselves are stored in a separate struct inbetween them being
registered and them being forwarded to the plugin. Currently only
supports string options.
Signed-off-by: Christian Decker <@cdecker>
Also includes some sanity checks for the results returned by the
plugin, such as ensuring that the ID is as expected and that we have
either an error or a real result.
The idea is that `plugin` is an early arg that is parsed (from command
line or the config file). We can then start the plugins and have them
tell us about the options they'd like to add to the mix, before we
actually parse them.
Signed-off-by: Christian Decker <@cdecker>
This is kind of a hack, but let's make it a complete hack. GCC with
-flto noticed we use different definitions of 'struct io_conn' here
and gave the warning:
ccan/ccan/io/io.h:620:17: warning: type of ‘io_close’ does not match original declaration [-Wlto-type-mismatch]
struct io_plan *io_close(struct io_conn *conn);
^
ccan/ccan/io/io.c:449:17: note: ‘io_close’ was previously declared here
struct io_plan *io_close(struct io_conn *conn)
^
ccan/ccan/io/io.c:449:17: note: code may be misoptimized unless -fno-strict-aliasing is used
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The spec says so, and it's right: with the right pattern of packet loss
(thanks Travis!) the other end can still be in channeld, waiting for our
`shutdown` message.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
If you steal something onto its own child, you create a loop. These are
expensive to check for at runtime, but they can hide from memleak and are
usually a bad idea. So we add a tal_steal() notify which does this work
in DEVELOPER mdoe.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This way there's no need for a context pointer, and freeing a msg_queue
frees its contents, as expected.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
It means an extra allocation at startup, but it means we can hide the definition,
and use standard patterns (new_daemon_conn and typesafe callbacks).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Rusty Russell
Conor Scott
arowser
Christian Decker
William Casarin