This is intended to be a more portable abstraction: rather than
hard-coding offsets in the macro invocations, it relies on the existence
of __builtin_offset and its availability even within inline asm blocks.
While here, rename some of the functions involved to hopefully improve
clarity
Discovered over on the dev-esp32-idf4 branch.
- Off by one error in loadLFS, leading to slight memory leak and
potential corruption.
- Insufficient return value check in loadLFS, where uzlib may return
one of two success conditions, one of which would result in an
out-of-bounds access and related pain.
- One case of a side effect within a lua_assert(), leading to
silently broken LFS image handling when compiling without asserts
enabled, the issue showing up as module names being shuffled around.
- Incorrect encoding of TValues in LFS when 64bit numbers in use.
LWIP_RAND() return type is int, value returned is sometimes negative.
This causes timer to sometimes (often) go outside of max_time, which in turn causes IGMP snoopers or IGMP routers to drop the subscription
* apa102: remove dead code
We can't store strings of hundreds of thousands of characters in RAM, so
this can't possibly have fired, historically. Pixbufs are still RAM
objects, so that's still out. With LFS, it would take a pathological
example to hit the required 400KB TSTRING.
* Add IRQ management functions
* ws2812: fill UART FIFOs with IRQs off
Refactor code to make the use of two fill loops less gross.
`ws2812` buffer extracted to new `pixbuf` module.
* The new pixbuf module has more functionality than the `ws2812`-specific buffer it replaces.
* This is work in progress towards https://github.com/nodemcu/nodemcu-firmware/issues/2916
* The LED driver modules `ws2812`, `ws2801`, `apa102`, and `tm1829` have sprouted `pixbuf` support.
* `NTest` tests for `pixbuf` now exist.
While here, document the ws2812 UART-based overlapping with mainline
execution. Fixes https://github.com/nodemcu/nodemcu-firmware/issues/3140
Co-authored-by: Gregor Hartmann <HHHartmann@users.noreply.github.com>
* mqtt: remove concept of connection timeout
Just rely on the network stack to tell us when things have gone south.
* mqtt: remove write-only mqtt_state.port field
* mqtt: drop useless conditional
* mqtt: decouple message sent flag from timer
* mqtt: reconnect callback does not need to hang up
The network stack has certainly done that for us at this point.
Similarly, since we're about to call mqtt_socket_disconnected, don't
bother unregistering the timer here, either.
* mqtt: don't tick once per second
Set the timer for the duration of the wait and cancel it on the other side.
* mqtt: defer message queue destruction to _disconnect
We're going to want to publish a disconnect message for real, so doing
this in _close does no one any favors
* mqtt: miscellaneous cleanups
No functional change intended
* mqtt: close() should send disconnect message for real
This means waiting for _sent() to fire again before telling the network
stack to disconnect.
* mqtt: tidy connect and dns
- Push the self-ref to after all allocations and error returns
- Don't try to extract IPv4 from the domain string ourselves, let the
resolver, since it can
- Don't try to connect to localhost. That can't possibly work.
* mqtt: common up some callback invocations
* mqtt: don't retransmit messages on timeout
There's no point in retransmitting messages on timeout; the network
stack will be trying to do it for us anyway.
* mqtt: remove unnecessary NULL udata checks
* mqtt: hold strings in Lua, not C
Eliminates a host of C-side allocations.
While here, move the rest of the mqtt_connect_info structure out to its
own thing, and pack some flags using a bitfield.
* mqtt: mqtt_socket_on use lua_checkoption
* mqtt: slightly augment debug messages
These changes have made some debugging ever so slightly easier.
These have been handy within the test harness for streaming files and
LFS images across to the DUTs.
Add nrec method to pipe
Co-authored-by: Gregor Hartmann <HHHartmann@users.noreply.github.com>
unread contained two subtle bugs:
- it created a buffer chunk at pipe[1], when that position is reserved
for the pipe reader function. Because it shifted pipe[i] to
pipe[i+1], including pipe[1], this was likely to manifest as the pipe
later attempting to dequeue a function rather than a buffer chunk
user datum. Solve this by adjusting the loop bounds and creation
index.
- when it created a new buffer chunk, it left that chunk's ->start and
->end fields at 0, but would then exit the loop to do the tail fill,
which was in turn assuming the exit condition of the loop when an
existing chunk had room, namely that data had been moved to the right
and so ->start > l. Solve this by making new chunks empty but with
->start = ->end = LUAL_BUFFERSIZE. It may be slightly better to
instead try to leave room at both ends if the buffer was empty prior
to this unread.
FIXES: https://github.com/nodemcu/nodemcu-firmware/issues/3155
* It appears that the rf_pre_init is not called any more. Also cleaned up the code in
wifi_common.
* Log a message (at the right baud rate) if the hostname is invalid
* Updated the comment in the user_config.h file
* Don't adjust the clock until after we deal with rtctime...
* Switched to using START_OPTION_CPU_FREQ_MAX instead.
* Use setfield and add caching of the startup option
* Put the startupcounts warning into a warning box
Nathaniel Wesley Filardo
Andreas Deininger
Johan Ström
Johny Mattsson
Lukáš Voborský
Gregor Hartmann
M.K
Philip Gladstone
Cody Cutrer
philip