* stm32wl: check HAL_FLASH_Unlock() return in _internal_flash_erase
_internal_flash_prog already checks HAL_FLASH_Unlock() and returns
LFS_ERR_IO on failure. _internal_flash_erase discarded the return
value, proceeding to erase even if the flash was not unlocked.
Apply the same check for consistency and safety.
Signed-off-by: Andrew Yong <me@ndoo.sg>
Assisted-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* stm32wl: fix _internal_flash_prog to abort on first write error
Previously the programming loop continued to the next doubleword after
HAL_FLASH_Program() failed, potentially writing to invalid addresses
and returning a misleading error code only at the end (last iteration).
HAL_FLASH_Lock() was also skipped on the mid-loop early return path.
- Move bounds check before the loop (validate full range at once)
- Break on first HAL error so subsequent doublewords are not written
- Move HAL_FLASH_Lock() after the loop so it always runs
Signed-off-by: Andrew Yong <me@ndoo.sg>
Assisted-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* stm32wl: clear stale flash SR error flags before erase and program
Stale error flags in FLASH->SR from a previous failed operation can
cause HAL_FLASH_Program() or HAL_FLASHEx_Erase() to return HAL_ERROR
immediately without attempting the operation.
Add __HAL_FLASH_CLEAR_FLAG(FLASH_FLAG_ALL_ERRORS) after each
HAL_FLASH_Unlock() in both _internal_flash_prog and
_internal_flash_erase to ensure a clean state before each operation.
Signed-off-by: Andrew Yong <me@ndoo.sg>
Assisted-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* stm32wl: reject flash prog writes not aligned to 8-byte doubleword
The STM32WL HAL minimum write unit is one 64-bit doubleword (8 bytes).
_internal_flash_prog silently truncated any trailing bytes when size % 8
!= 0 because dw_count = size / 8 drops the remainder. Return LFS_ERR_INVAL
early so LittleFS sees the error rather than a silent short write.
Signed-off-by: Andrew Yong <me@ndoo.sg>
Assisted-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(nrf52,fs): use atomic SafeFile rename instead of direct write
NRF52 was bypassing the .tmp/readback/rename path entirely — openFile()
deleted the target file and wrote directly to it, and close() returned
true without verifying the write or renaming anything.
Adafruit_LittleFS::rename() calls lfs_rename() directly (confirmed at
Adafruit_LittleFS.cpp:205). Remove both ARCH_NRF52 guards so NRF52
follows the same write-to-.tmp → readback-hash → rename path used by
all other platforms.
Signed-off-by: Andrew Yong <me@ndoo.sg>
Assisted-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(admin): skip uiconfig.proto save on devices without a screen
handleStoreDeviceUIConfig() was writing /prefs/uiconfig.proto
unconditionally. MenuHandler.cpp is already gated behind #if HAS_SCREEN,
so there is no path that populates UI config on screen-less platforms.
Guard the save with #if HAS_SCREEN to avoid wasting a flash block on
devices that will never use it.
The read path (handleGetDeviceUIConfig) does not touch the filesystem
and needs no change.
Signed-off-by: Andrew Yong <me@ndoo.sg>
Assisted-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* fs: enable format-on-retry for all platforms in saveToDisk
The FSCom.format() call on save failure was guarded to ARCH_NRF52 with
a comment that other platforms were not ready (bug #4184). STM32WL was
added to the guard in a prior commit. All platforms now expose format
semantics and the retry logic is identical — remove the guard.
To keep NodeDB.cpp platform-agnostic and fix a CI failure on native-tft
(portduino's fs::FS has no format() method), introduce fsFormat() in
FSCommon as the single call-site for all callers:
- Embedded (ESP32, NRF52, STM32WL, RP2040): delegates to FSCom.format()
- Portduino: rmDir("/prefs") + FSBegin() (a no-op on portduino).
rmDir("/prefs") is already called unconditionally by factoryReset()
(NodeDB.cpp:504), so both primitives are proven on portduino.
Replace both direct FSCom.format() calls in NodeDB.cpp with fsFormat().
Note: we do not run portduino locally — portduino/native build testers
please verify the format-on-retry path.
Signed-off-by: Andrew Yong <me@ndoo.sg>
Assisted-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* DO NOT MERGE: nrf52(fs): add File() default constructor bound to InternalFS
Adds File() to the Adafruit LittleFS File class (in the Meshtastic
Adafruit_nRF52_Arduino fork), delegating to File(InternalFS). This
matches the default-constructible File API on all other platforms.
The constructor is implemented in Adafruit_LittleFS_File.cpp rather
than inline in the header to avoid a circular include between
Adafruit_LittleFS_File.h and InternalFileSystem.h.
FOLLOW-UP REQUIRED: nrf52.ini points to a commit SHA on the
mesh-malaysia/Adafruit_nRF52_Arduino fork instead of the upstream
meshtastic framework. Once meshtastic/Adafruit_nRF52_Arduino#5 is
merged, revert nrf52.ini to point back to the upstream meshtastic
framework URL.
Signed-off-by: Andrew Yong <me@ndoo.sg>
Assisted-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* stm32wl(fs): add File() default constructor and document LFS tunables
Adds File() to STM32_LittleFS_Namespace::File, delegating to
File(InternalFS). Implemented in the .cpp to avoid a circular include
between STM32_LittleFS_File.h (which cannot include LittleFS.h) and
the InternalFS extern declaration.
This matches the File API on ESP32/RP2040/Portduino and is a
prerequisite for removing the ARCH_STM32WL guard in xmodem.h.
No behavior change — the constructor leaves the file in the same
closed/unattached state as File(InternalFS) would.
Signed-off-by: Andrew Yong <me@ndoo.sg>
Assisted-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* fs: remove arch-specific ifdefs from FSCommon, SafeFile, xmodem
Now that NRF52 and STM32WL have File() default constructors and NRF52
has working atomic SafeFile rename, the capability gaps are closed.
Remove all per-arch guards across the shared FS layer:
FSCommon.cpp — renameFile():
Use FSCom.rename() on all platforms. Adafruit_LittleFS::rename()
calls lfs_rename() directly (Adafruit_LittleFS.cpp:205). The
copy+delete fallback on NRF52/RP2040 was never necessary.
FSCommon.cpp — getFiles():
Replace four ARCH_ESP32 guards with a single filepath pointer at
the top of the loop (file.path() on ESP32, file.name() elsewhere).
Fix strcpy(fileInfo.file_name, filepath): bounded to
sizeof(fileInfo.file_name)-1 with explicit NUL termination to prevent
overflow of the 228-byte meshtastic_FileInfo::file_name array.
FSCommon.cpp — listDir():
Same filepath pointer approach. NRF52/STM32WL were in an else-branch
that only logged but never deleted — now all platforms follow the
unified del path. 12 guards → 2.
Fix three strncpy(buffer, ..., sizeof(buffer)) calls that did not
NUL-terminate when source length >= sizeof(buffer) (255 bytes).
Add explicit buffer[sizeof(buffer)-1] = '\0' after each.
FSCommon.cpp — rmDir():
Use listDir(del=true) everywhere. The ARCH_NRF52 rmdir_r() path and
the ARCH_ESP32|RP2040|PORTDUINO listDir() path collapse to one line.
SafeFile.cpp:
ARCH_NRF52 bypass removed (handled in preceding commit).
xmodem.h:
File file; now works on all platforms via default constructors
added in the two preceding commits.
Remaining #ifdef ARCH_ESP32 in FSCommon.cpp: exactly 4, all for the
file.path() vs file.name() API difference (ESP32 Arduino LittleFS
returns the full path; all others return only the name). That
difference lives in the framework and cannot be closed here.
Signed-off-by: Andrew Yong <me@ndoo.sg>
Assisted-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* stm32wl(fs): add write-behind page cache, reduce virtual block size and FS reservation (FORMAT BREAK)
Adds a write-behind (RMW) page cache to the STM32WL LittleFS driver,
modelled after the NRF52 Adafruit approach (flash_cache.c). This allows
LFS to use 256-byte virtual blocks backed by 2048-byte physical pages:
the erase/prog callbacks accumulate changes in a 2 KB RAM buffer; the
sync callback (and page eviction on page-change) flushes with a single
HAL physical-erase + doubleword-program pass.
LFS tunables changed (FORMAT BREAK — superblock parameters):
block_size: 2048 B → 256 B (8 virtual blocks per physical page)
read_size: 2048 B → 256 B (= block_size)
prog_size: 2048 B → 256 B (= block_size; hardware min is 8 B)
block_count: 112 → 80 (14 phys pages → 10 phys pages = 20 KiB)
Benefits:
- Internal fragmentation: max 2047 B/file → max 255 B/file
- Heap per open LFS file: ~4 KB → 512 B (prog + read buffers)
- Code flash headroom: 6.7 KB → ~14.1 KB (+7.4 KB)
- Block budget: 80 virtual blocks, worst-case peak ~20, ~60 free
Updates board_upload.maximum_size in wio-e5/platformio.ini from 233472
(256 KB − 28 KB) to 241664 (256 KB − 20 KB) to match the reduced FS
reservation.
Justification for the format break: the prior STM32WL firmware had
several flash write bugs fixed earlier in this series (missing error
flag clearing, no abort on first write failure, unaligned write
acceptance). These bugs very likely caused silent config corruption on
deployed devices. The format break should be treated as an enhancement:
it provides a clean, reliably-written starting point. Users will need
to reconfigure their device once after this update.
Correctness fixes applied to the cache implementation:
- alignas(8) on _page_cache: the buffer was uint8_t[] (alignment 1)
but _flash_cache_flush casts it to const uint64_t* — undefined
behaviour per C++ standard, potential Cortex-M hardfault. alignas(8)
guarantees the required alignment for the doubleword cast.
- HAL_FLASH_Lock() return value: was discarded. Now assigned to
lock_rc and propagated into rc if prior writes succeeded, so LFS
sees the error rather than a false success.
Signed-off-by: Andrew Yong <me@ndoo.sg>
Assisted-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* stm32wl(fs): reduce FS reservation from 10 pages to 7 pages (FORMAT BREAK)
Reduces LFS_FLASH_TOTAL_SIZE from 10 × 2 KiB pages (20 KiB) to
7 × 2 KiB pages (14 KiB), freeing 6 KiB for firmware.
board_upload.maximum_size updated accordingly across all STM32WL variants:
241664 (256 KiB - 20 KiB) → 247808 (256 KiB - 14 KiB)
This is a FORMAT BREAK: existing filesystems must be erased before use.
Assisted-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: Andrew Yong <me@ndoo.sg>
* fix(fs): return false in renameFile() when FSCom is not defined
Avoids undefined behavior and -Wreturn-type warnings in configurations
that compile FSCommon.cpp without a filesystem backend.
Signed-off-by: Andrew Yong <me@ndoo.sg>
Assisted-by: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Signed-off-by: Andrew Yong <me@ndoo.sg>
Co-authored-by: Ben Meadors <benmmeadors@gmail.com>
Use distro provided Python at build time (instead of the `python` images from dockerhub) and install `grpcio-tools` using the distro provided packages.
This should speed up build times, ESPECIALLY on riscv64 (where prebuilt `grpcio-tools` wheels are not provided on pip).
Co-authored-by: Copilot <copilot@github.com>
* Fix MAC_from_string to use input parameter instead of global config for MAC address parsing
* Enhance MAC_from_string validation and error handling
* Add missing include for <cctype> in PortduinoGlue.cpp
* Add TCP support for Meshtastic MCP interface / tests and update docs
* Address TCP endpoint validation and error handling in connection
* TCP connection handling and device listing logic
* Fix docstring formatting in normalize_tcp_endpoint function
isConnectedToNetwork() always returned false on ARCH_PORTDUINO
because none of HAS_WIFI, HAS_ETHERNET, or USE_WS5500 are defined
for Linux native builds. This caused wantsLink() to always return
false, preventing the MQTT thread from ever connecting at boot.
Fix: return true for ARCH_PORTDUINO since Linux always has network
access available.
Co-authored-by: Jonathan Bennett <jbennett@incomsystems.biz>
Upstream support has been added in Debian and Alpine.
Only build as part of `docker_manifest` (Beta/Alpha/Daily) releases, because these will take a **while** thanks to qemu.
Co-authored-by: Copilot <copilot@github.com>
* Add clamping logic for milliseconds conversion and unit tests
* Simplify comments in secondsToMsClamped function
Removed detailed comments about seconds to milliseconds conversion.
* macOS: enable CH341 LoRa-hardware path — fix serial truncation, document setup
Verified on Apple Silicon with a CH341A USB-SPI bridge (VID 0x1A86,
PID 0x5512) wired to an SX1262 (Meshstick variant) that the existing
`pine64/libch341-spi-userspace` lib_dep works on macOS as-is — Apple's
bundled CH34x driver only matches the CH340 *UART* variant
(PID 0x7523), so the CH341A's interface 0 is left unclaimed and
libusb opens / configures / claims it directly via IOUSBHostInterface.
End-to-end test: meshtasticd boots, libusb claim succeeds, SX1262 init
returns 0, TCP API serves the meshtastic CLI's --info / --sendtext flow.
Two changes:
1. **`PortduinoGlue.cpp:497`**: pass `sizeof(serial)` (= 9) instead of
the literal `8` to `Ch341Hal::getSerialString()`. The function in
`USBHal.h:61-68` treats `len` as buffer size and reserves one slot
for the null terminator (`bytesCopied = (len - 1) < 8 ? (len - 1) : 8`),
so passing 8 produced a 7-char serial — which then broke the
`strlen(serial) == 8` check at line 502, skipping the auto-MAC
derivation from serial + product string. On Linux this was masked
by the BlueZ HCI MAC fallback in `getMacAddr()` at lines 139-157,
but on macOS that fallback is `__linux__`-guarded so the serial path
is mandatory and the truncation left `mac_address` empty, causing
the daemon to exit with `*** Blank MAC Address not allowed!`.
2. **`variants/native/portduino/platformio.ini`**: expand the
`[env:native-macos]` comment block with a "Real LoRa hardware on
macOS" section. Documents:
- Why no upstream library change is needed (Apple kext targets
CH340/UART, not CH341A/SPI; libusb's `#ifdef __linux__` skip is
correct for macOS in this case).
- How to point `meshtasticd` at an existing platform-agnostic
`bin/config.d/lora-*.yaml` for CH341 hardware.
- The auto-MAC-derivation contract (now working with this fix).
- `ioreg` and `LIBUSB_DEBUG=4` diagnostic recipes for the failure
mode where a third-party WCH `CH34xVCPDriver` *would* claim
interface 0 (`kmutil unload -b <bundleID>` workaround).
No upstream library forks, no PR chain, no additional lib_deps —
the existing `pine64/libch341-spi-userspace` + libusb-1.0 stack does
the right thing on macOS already.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Apply suggestion from @Copilot
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* macOS: enable CH341 LoRa-hardware path — fix serial truncation, document setup
Verified on Apple Silicon with a CH341A USB-SPI bridge (VID 0x1A86,
PID 0x5512) wired to an SX1262 (Meshstick variant) that the existing
`pine64/libch341-spi-userspace` lib_dep works on macOS as-is — Apple's
bundled CH34x driver only matches the CH340 *UART* variant
(PID 0x7523), so the CH341A's interface 0 is left unclaimed and
libusb opens / configures / claims it directly via IOUSBHostInterface.
End-to-end test: meshtasticd boots, libusb claim succeeds, SX1262 init
returns 0, TCP API serves the meshtastic CLI's --info / --sendtext flow.
Two changes:
1. **`PortduinoGlue.cpp:497`**: pass `sizeof(serial)` (= 9) instead of
the literal `8` to `Ch341Hal::getSerialString()`. The function in
`USBHal.h:61-68` treats `len` as buffer size and reserves one slot
for the null terminator (`bytesCopied = (len - 1) < 8 ? (len - 1) : 8`),
so passing 8 produced a 7-char serial — which then broke the
`strlen(serial) == 8` check at line 502, skipping the auto-MAC
derivation from serial + product string. On Linux this was masked
by the BlueZ HCI MAC fallback in `getMacAddr()` at lines 139-157,
but on macOS that fallback is `__linux__`-guarded so the serial path
is mandatory and the truncation left `mac_address` empty, causing
the daemon to exit with `*** Blank MAC Address not allowed!`.
2. **`variants/native/portduino/platformio.ini`**: expand the
`[env:native-macos]` comment block with a "Real LoRa hardware on
macOS" section. Documents:
- Why no upstream library change is needed (Apple kext targets
CH340/UART, not CH341A/SPI; libusb's `#ifdef __linux__` skip is
correct for macOS in this case).
- How to point `meshtasticd` at an existing platform-agnostic
`bin/config.d/lora-*.yaml` for CH341 hardware.
- The auto-MAC-derivation contract (now working with this fix).
- `ioreg` and `LIBUSB_DEBUG=4` diagnostic recipes for the failure
mode where a third-party WCH `CH34xVCPDriver` *would* claim
interface 0 (`kmutil unload -b <bundleID>` workaround).
No upstream library forks, no PR chain, no additional lib_deps —
the existing `pine64/libch341-spi-userspace` + libusb-1.0 stack does
the right thing on macOS already.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Apply suggestion from @Copilot
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Router::handleReceived stores its allocCopy of the encrypted packet in the
class member p_encrypted. callModules() invokes module replies that re-enter
the router via MeshService::sendToMesh -> Router::sendLocal, which on a
broadcast reply recursively calls handleReceived. The inner call overwrites
the outer's p_encrypted without releasing it; on the way out it nulls the
member, the outer release(p_encrypted) now releases nullptr, and the original
allocation is permanently leaked. ~381 B per recursion.
Promote p_encrypted to a function-local so each invocation owns its own copy
for its full lifetime. The MQTT-publish null check at the call site (added by
PR #9136 as a workaround for this bug) stays in place because allocCopy can
still legitimately return nullptr on packetPool exhaustion.
Copilot's review of PR #8999 (the original introduction) flagged this exact
pattern at merge time:
"Storing p_encrypted as a class member can cause issues with recursive or
concurrent calls to handleReceived() since each call would overwrite the
previous packet pointer."
The historical reason for the member (S&F needing to retain the encrypted
copy across calls) was satisfied differently by PR #9799 (ServerAPI converted
to std::unique_ptr + cleanup on connection close), so the member is no longer
load-bearing.
Reproduces issues #9632 / #10101 / #8729 (heap leak when MeshMonitor
connected; TCP drops on Station G2 / LILYGO ServerAPI dump abort).
Hardware A/B on Station G2 under sustained TCP-API retry storm (open :4403,
request config, disconnect mid-stream, repeat at ~0.6/s) - 9 min run:
| Build | heapFree drift | rebootCount delta |
| this patch | -1.5 KB (noise)| 0 |
| stock 2.7.13 | -73 KB (8.1KB/min) | +1 (OOM crash) |
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
Co-authored-by: Ben Meadors <benmmeadors@gmail.com>
Router::handleReceived stores its allocCopy of the encrypted packet in the
class member p_encrypted. callModules() invokes module replies that re-enter
the router via MeshService::sendToMesh -> Router::sendLocal, which on a
broadcast reply recursively calls handleReceived. The inner call overwrites
the outer's p_encrypted without releasing it; on the way out it nulls the
member, the outer release(p_encrypted) now releases nullptr, and the original
allocation is permanently leaked. ~381 B per recursion.
Promote p_encrypted to a function-local so each invocation owns its own copy
for its full lifetime. The MQTT-publish null check at the call site (added by
PR #9136 as a workaround for this bug) stays in place because allocCopy can
still legitimately return nullptr on packetPool exhaustion.
Copilot's review of PR #8999 (the original introduction) flagged this exact
pattern at merge time:
"Storing p_encrypted as a class member can cause issues with recursive or
concurrent calls to handleReceived() since each call would overwrite the
previous packet pointer."
The historical reason for the member (S&F needing to retain the encrypted
copy across calls) was satisfied differently by PR #9799 (ServerAPI converted
to std::unique_ptr + cleanup on connection close), so the member is no longer
load-bearing.
Reproduces issues #9632 / #10101 / #8729 (heap leak when MeshMonitor
connected; TCP drops on Station G2 / LILYGO ServerAPI dump abort).
Hardware A/B on Station G2 under sustained TCP-API retry storm (open :4403,
request config, disconnect mid-stream, repeat at ~0.6/s) - 9 min run:
| Build | heapFree drift | rebootCount delta |
| this patch | -1.5 KB (noise)| 0 |
| stock 2.7.13 | -73 KB (8.1KB/min) | +1 (OOM crash) |
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
Co-authored-by: Ben Meadors <benmmeadors@gmail.com>
* True Colors on TFT (Heltec Mesh Node T114, Heltec Vision Master T190, CardPuter Adv, T-Deck, T-Lora Pager)
* Theme support - New and some Classic Themes!
* Colored Compass
---------
Co-authored-by: Jason P <applewiz@mac.com>
Co-authored-by: Jonathan Bennett <jbennett@incomsystems.biz>
Co-authored-by: Ben Meadors <benmmeadors@gmail.com>
Fixes issues with #includes inherited from `configuration.h` when building for pioarduino.
Aligns t5s3_epaper with other variants like t_deck_pro.
Co-authored-by: Copilot <copilot@github.com>
Co-authored-by: Ben Meadors <benmmeadors@gmail.com>
