8.9 KiB
Extension Developer FAQ
# # # Work in Progress # # #
Changes in IDF compared to non-OS/RTOS SDK
This is a non-exhaustive list, obviously, but some key points are:
-
No more
c_types.h
. Standard C types are finally the norm.stdint.h
for all your[u]intX_t
needsstdbool.h
for bool (notetrue
/false
vs oldTRUE
/FALSE
)stddef.h
forsize_t
-
A real C library. All the
os_
,ets_
andc_
prefixes for standard library functions are gone (except for the special case c_getenv, for now). stdout/stdin are wired up to the UART0 console. The NodeMCU vfs layer is not currently hooked up to the C library, but that would be a nice thing to do. -
Everything builds on at least C99 level, with plenty of warnings enabled. Fix the code so it doesn't produce warnings - don't turn off the warnings! Yes, there may be exceptions, but they're rare.
-
user_config.h
is no more. All configuration is now handled via Kconfig (make menuconfig
to configure). From the developer perspective, simply includesdkconfig.h
and test the corresponding CONFIG_YOUR_FEATURE macro. Theplatform.h
header is guaranteed to includesdkconfig.h
, btw. -
user_modules.h
is also gone. Module selection is now done via Kconfig. Rather than adding a #define touser_modules.h
, add an option incomponents/modules/Kconfig
of the formNODEMCU_CMODULE_XYZ
, and the existingNODEMCU_MODULE()
macros will take care of the rest. Example Kconfig entry:config NODEMCU_CMODULE_XYZ bool "Xyz module" default "y" help Includes the XYZ module. Provides features X, Y and Z.
-
Preemptive multithreading. The network stack and other drivers now run in their own threads with private stacks. This makes for a more robust architecture, but does mean proper synchronization MUST be employed between the threads. This includes between API callbacks and main Lua thread. Note that many API callbacks have turned into events in the IDF, and NodeMCU handle those in the context of the main Lua thread.
-
Logical flash partitions. Rather than hardcoding assumptions of flash area usage, there is now an actual logical partition table kept on the flash.
NodeMCU task paradigm
NodeMCU uses a task based approach for scheduling work to run within
the Lua VM. This interface is defined in task/task.h
, and comprises
three aspects:
- Task registration (via
task_getid()
) - Task posting (via
task_post()
and associated macros) - Task processing (via
task_pump_messages()
)
This NodeMCU task API is designed to complement the Lua library model, so that a library can declare one or more task handlers and that both ISRs and library functions can then post a message for delivery to a task handler.
Note that NodeMCU tasks must not be confused with the FreeRTOS tasks. A FreeRTOS task is fully preemptible thread of execution managed by the OS, while a NodeMCU task is a non-preemptive* a callback invoked by the Lua FreeRTOS task. It effectively implements cooperative multitasking. To reduce confusion, from here on FreeRTOS tasks will be referred to as threads, and NodeMCU tasks simply as tasks. Most NodeMCU developers will not need to concern themselves with threads unless they're doing low-level driver development.
The Lua runtime is NOT reentrant, and hence any code which calls any Lua API must run within the Lua task context. ISRs, other threads and API callbacks must not access the Lua API or Lua resources. This typically means that messages need to be posted using the NodeMCU task API for handling within the correct thread context. Depending on the scenario, data may need to be put on a FreeRTOS queue to transfer ownership safely between the threads, or posted directly across with the NodeMCU task API.
The application has no control over the relative time ordering of tasks and API callbacks, and no assumptions can be made about whether a task and any posted successors will run consecutively.
*) Non-preemptive in the sense that other NodeMCU tasks will not preempt it. The RTOS may still preempt the task to run ISRs or other threads.
Task registration
Each module which wishes to receive events and process those within the
context of the Lua thread need to register their callback. This is done
by calling task_get_id(module_callback_fn)
. A non-zero return value
indicates successful registration, and the returned handle can be used
to post events to this task. The task registration is typically done in
the module_init function.
Task posting
To signal a task, a message is posted to it using
task_post(prio, handle, param)
or the helper macros task_post_low()/task_post_medium()/task_post_high()
. Each message carries a single parameter value
of type intptr_t
and may be used to carry either pointers or raw values.
Each task defines its own schema depending on its needs.
Note that the message queues can theoretically fill up, and this is reported
by a false
return value from the task_post()
call. In this case no
message was posted. In most cases nothing much can be done, and the best
approach may be to simply ignore the failure and drop the data. Be careful
not to accidentally leak memory in this circumstance however! A good
habit would be to do something like this:
char *buf = malloc (len);
// do something with buf...
if (!task_post_medium(handle, buf))
free (buf);
The task_post*()
function and macros can be safely called from any ISR or
thread.
Task processing and priorities
The Lua runtime executes in a single thread, and at its root level runs the task message processing loop. Task messages arrive on three queues, one for each priority level. The queues are services in order of their priority, so while a higher-priority queue contains messages, no lower-priority messages will be delivered. It is thus quite possible to jam up the Lua runtime by continually posting high-priority messages. Don't do that.
Unless there are particular reasons not to, messages should be posted with medium priority. This is the friendly, cooperative level. If something is time sensitive (e.g. audio buffer running low), high priority may be warranted. The low priority level is intended for background processing during otherwise idle time.
Processing system events
The IDF is quite flexible in how system events may be handled, and NodeMCU takes advantage of this to get all system events handled by the main Lua thread. Event listening registration is very similar to module registration, and happens at link-time. The following snippet shows how to use this:
#include "nodemcu_esp_event.h"
static void on_got_ip (const system_event_t *evt)
{
// Do stuff, maybe invoke a Lua callback
}
// Register for the event SYSTEM_EVENT_STA_GOT_IP (see esp_event.h for list).
NODEMCU_ESP_EVENT(SYSTEM_EVENT_STA_GOT_IP, on_got_ip);
Memory allocation in modules
There are three main ways of allocating memory when writing modules for NodeMCU - stack (for temporaries), heap, and Lua heap. Using the stack is the same as for other embedded development, i.e. feel free to put small(ish) temporary objects there, but don't expect to get away with hundreds of bytes on the stack. Under good circumstances RTOS will detect a stack overflow and throw an error, under less good circumstances anything can happen.
Heap allocation is through the standard C malloc and friends. Their use is best limited to third-party libraries which do not allow for custom allocators. Also, if dynamic memory allocation is done in a thread other than the main Lua thread, this is the only available option.
The best way to allocate memory in a module however, is to use the luaM
memory allocation routines. What makes this the best option is that an
allocation made this way may trigger a garbage collection. Where a regular
C malloc might have failed due to out of memory, the Lua allocation can
succeed by virtue of freeing up memory first. The downside of course is
that this is only possible to do while running in the main Lua thread.
Note that memory allocated via e.g. luaM_malloc()
is not subject to
garbage collection, it behaves just like memory from the C malloc()
,
and must be explicitly free'd via luaM_free()
.
A quick example:
static int some_func (lua_State *L)
{
char *buf = luaM_malloc (L, 512);
int result = calc_using_large_buf (buf, 1, 2, 3);
lua_pushinteger (L, result);
luaM_free (buf);
return 1;
}
Caution
Do note that luaM_malloc()
raises a Lua error on allocation failure, and
will exit the calling function right then and there. This can lead to
resource leaks if care is not taken, though in most cases a failure
will result in a Lua panic and require a reboot.
On the upside, there is never a need to test the return value from
luaM_malloc()
for NULL, as on failure the function does not return.