load_more_messages() (background fill, main loop) ran each batch's
metadata reads while holding the LVGL lock. A cold batch is 2+ LittleFS
reads; on the degraded device a fill batch measured ~4s in earlier
captures, within a hair of the 5s deadlock guard. The fill starts
immediately after any conversation open (first page > INITIAL_RENDER),
so an open triggered a second under-lock stall right after the first
was fixed.
Same restructure: guard lock copies the batch's hashes, metadata I/O
runs off-lock, the bubble prepend commits under a brief lock. A
_fill_generation counter (bumped on every list rebuild: open, prepare
commit, refresh) discards a batch whose conversation changed mid-I/O.
Display order is preserved (newest-first read order + push_front gives
oldest-to-newest, matching the old loop).
Opening a conversation crashed the same way sending did. load_conversation()
(LVGL task, under the LVGL lock held by replace_route) ran the full open
pipeline synchronously: identity recall (ustore), display-name read, the
message-index read, and the per-message metadata reads. On this device's
degraded LittleFS each op is 0.4-2s, so a cold open of a dozen-message
conversation held the LVGL mutex past the 5s deadlock guard and asserted at
LVGLLock.h:45. The send path already got the mailbox fix; the open path never
did.
Restructure with the same pattern:
- load_conversation() (LVGL task) now only navigates + resets the list and
shows the truncated hash in the header. Same-peer re-opens return early
with zero store I/O (rows are still built).
- prepare_conversation() (main loop, called from update()) does the store
I/O between a short guard lock and a short commit lock, then commits the
header name + initial bubbles + background-fill arming under a brief
LVGL_LOCK. A generation counter discards a stale in-flight prepare when
the conversation changes mid-I/O.
- refresh() re-arms the prepare instead of re-reading under the lock.
The 1Hz store 'not found in index' fetch is pre-existing (present on
2527c6d) and is being tracked separately as a flash-wear follow-up.
Build tdeck SUCCESS, 170/170 contract tests pass.
Every message send on the device was deterministically rebooting it:
send_message() ran the full pipeline (identity recall, message
construction, RouterLock-scoped router admission, and LittleFS
persistence) synchronously on LVGL's 8 KiB task while holding the LVGL
mutex. On this device's degraded filesystem a single save takes ~7s of
400ms-2s per-op gaps, tripping the 5s LVGL deadlock guard and asserting
at LVGLLock.h:45 (assert failed: LVGL mutex timeout (5s)). The receive
path already carries the fix pattern for exactly this failure class
(see on_message_received); the send path never got it.
Restructure the send path as a mailbox handoff, following the existing
CallStartMailbox / LocationShareCommandMailbox precedent:
- send_message() (LVGL task) now only validates and publishes
(destination, content, source) into a mutex-guarded single-slot
OutgoingSendMailbox. No router lock, no I/O, no message construction.
- update() services the mailbox in service_pending_sends() on the main
loop, before the big LVGL_LOCK() — the only place in the send path
that may take the router lock, block on admission, or wait on
LittleFS.
- On acceptance, a brief LVGL_LOCK in apply_outbound_result() commits
the UI (add_message / clear_composer / compose->chat navigation,
route-guarded). The admitted packed form is unpacked for display
with incoming/state flags restored.
- On rejection (storage error, router busy, queue full) the user's
input is retained for retry, matching the old behavior.
The 500-char UI cap bounds the mailbox payload.
Build tdeck SUCCESS, 170/170 contract tests pass.
Address the three P1 findings from the Greptile review of PR #96:
1. Network LoRa toggle reverted on save. on_interface_switch_changed
saved without propagating the Network switch state, so the save path
read the stale Radio-page twin and wrote it back over the mirror.
The handler now mirrors the flipped switch to its twin (and toggles
the LoRa params visibility) before saving, mirroring
on_lora_enabled_changed in the other direction.
2. Immediate controls committed unsaved drafts from other sub-views.
Every immediate handler ran the full-screen update_settings_from_ui,
so typing a TCP host and then adjusting brightness silently persisted
the host. update_settings_from_ui is now view-scoped: save_settings
reads only the sub-view that just changed (LVGL events only fire for
visible widgets, so _view is always the touched view). The lora
handle merges as Radio-page switch wins, Network mirror when the
Radio page is unbuilt.
3. Hidden controls retained focus. hide() removed only the header
buttons and hub cards, and the focus_group_for cleanup list was
missing the three Radio dropdowns — so hidden sub-view controls
stayed in the default group (and Radio dropdowns lingered after
leaving the Radio view). Extracted the full removal list into
remove_all_from_focus_group (now including _dropdown_lora_bandwidth/
_sf/_cr) and use it from both focus_group_for and hide.
Verification: tdeck build SUCCESS (RAM 23.1%, Flash 92.7%),
tests/build_scripts 170/170. All save_settings() call sites confirmed
event-driven so _view always matches the touched control.
Signed-off-by: Torlando <tyler@torlando.tech>
- Move the LoRa TX Power slider from x=75 to x=86: the 'TX Power:' label
in Montserrat 14 is ~78px wide, so the slider track started on top of
the colon.
- Network sub-view: the Reticulum interface toggles (TCP/AUTO/LoRa/BLE)
are now row 1, relabeled 'Reticulum Interfaces:', with WiFi SSID,
WiFi password, TCP host and port following in their original order.
- Rename the ambiguous 'Reconnect' button to 'WiFi Reconnect' — it
re-applies the WiFi SSID/password fields to the running stack and has
nothing to do with the TCP host. Sized 110px to fit the label.
Signed-off-by: Torlando <tyler@torlando.tech>
'TX Power:' in Montserrat 14 is ~78px wide but the slider track started
at x=75, tucking the colon under the track. Start the track at x=86
(8px clear; still 51px from the right-aligned value label).
Signed-off-by: Torlando <tyler@torlando.tech>
Pre-building all seven settings sub-views at construction created
~100 LVGL objects that (being <256B) allocated into internal DRAM via
the hybrid allocator. On the T-Deck that fragmented internal heap to
an 8180B largest block at LVGL task-creation time, and the 8KB task
(needs ~8.5KB with TCB) failed: xTaskCreatePinnedToCore returned
errQUEUE_FULL, the device parked on the boot splash, and
confirm_running_firmware never ran.
Root cause proven with an on-device heap probe (largest internal
block 8180B, 8192-stack task FAIL, 4096-stack OK), then fixed by
building each sub-view on first navigation. Boot-time internal DRAM
rises to a 47KB largest block and the task starts cleanly.
Signed-off-by: Torlando <tyler@torlando.tech>
Boot-loop regression from the hub-and-spoke rewrite: lv_label_set_text()
on a plain lv_obj object wrote label internals into a non-label and
freed a bogus pointer (heap_caps_free assert 'outside heap areas') during
UIManager::init -> SettingsScreen::create_advanced_view, crashing every
boot at 'Initializing UIManager'. Verified by decoding the on-device
backtrace (ELF 62cf1d0e...) against the exact flashed firmware.
Replace the flat Settings screen with a card hub + dedicated sub-views.
- Hub: 8 navigation cards (Status, Network, Identity, Radio, Delivery,
Appearance, Advanced, Transport) reusing the Network-screen card widget;
Transport stays last (danger invariant preserved).
- Tap a card -> dedicated sub-view holding that area's controls; no accordion.
- Network sub-view gains a LoRa interface toggle alongside TCP/Auto/BLE;
it two-way-mirrors the Radio page's canonical lora_enabled switch.
- Save model: simple controls apply immediately; a Save button appears only
on the form sub-views (Network, Radio, Identity).
- Identity sub-view adds a View Identity row routing to the existing lxma://
QR screen (the Status Share button already reached it).
- Focus group rebuilt per view (only-visible objects) so the auto-scroll-to-
bottom class of bug cannot recur; each sub-view entry scrolls to top.
Contracts updated for the new structure; 170/170 build-script tests pass.
tdeck build green: RAM 23.1% (75,848 B, unchanged), Flash 92.7%
(2,914,713 B, +2,640 B). Not flashed, not PR'd.
The explicit scroll-to-top in show() was confirmed working on the device
(open-at-top on every entry, including repeat entries after scrolling to
the bottom). Remove the temporary Serial logging and the 30ms deviation
watcher.
Temporary [SCROLLDIAG] build: show() logs _content scroll_y before and
after the explicit reset, and a 30ms LVGL timer logs the first deviation
from 0 after entry. Captures the actual mechanism behind the bottom-open
instead of guessing. Remove before PR.
The one-shot LVGL timer's callback deleted the timer but left
_entry_scroll_reset_timer dangling. On the 2nd+ screen entry show()
took the 'else' path and called lv_timer_reset() on freed memory, so
the view was never reset to the top — it stayed wherever the group
re-focus had auto-scrolled it (the bottom). Null the pointer before
lv_timer_del so each entry recreates a fresh timer.
LVGL re-focuses the default input group when the screen is shown; the
focused member is the last focusable object (the transport switch), so
the view landed at the bottom of the list. A one-shot 50ms LVGL timer
after show() scrolls the content back to the top without touching the
focus state, and leaves subsequent user scrolling alone.
Settings held live status (GPS fix, storage/RAM/identity) that belongs
on the Status screen, plus a per-second tick() doing SPI flash stat
reads and label churn mid-scroll — the main cause of laggy scrolling.
- GPS section (sats/location/altitude/HDOP/time) -> StatusScreen
- System Info (firmware build, storage, RAM) -> StatusScreen, with
storage/RAM stat reads throttled to ~5s and stack-buffer snprintfs
instead of Arduino String concatenation
- Settings gains a Status link row (trackball-reachable) that opens
Route::STATUS; the per-second SettingsScreen tick/refresh is deleted
- Reordered sections by frequency of use: General (name/brightness/
timeout/kb-light), Notifications, Network (now includes the
TCP/Auto/BLE interface switches), Radio (LoRa + params), Delivery,
Advanced, DANGER: Transport Mode (still final)
- Identity/LXMF hashes shown in Settings were truncated duplicates of
the Status screen's full display; removed
- main.cpp publishes firmware build + GPS to the Status screen
- Contract test for the storage readout follows the code to StatusScreen
Greptile follow-up on 67c1083: if xSemaphoreCreateMutex() failed,
QueueGuard silently locked nothing and the deferred queues still ran
unsynchronized. The constructor now logs the allocation failure and
every guarded site refuses mutation when _queue_mutex is null:
requesters stop enqueuing, flushers stop draining (nothing queued can
exist), and the one-shot preview commit is skipped. The screen keeps
rendering and chatting; only the deferred list mutations degrade.
Tracks microLXMF PR #11 head d7e05fd: archived-branch cache sync,
bulk-delete cache invalidation, and the uniform capped-content
contract for load_message_metadata(). No Pyxis code change — the
chat/list paths already rely on the capped preview and the full-view
accessor.
Greptile P1 on the list-perf commit: the deferred-queue pair
(requester task -> main-loop flush) was unsynchronized. Producers
push from the LVGL task (click handlers, refresh() during navigation)
while flush_*() swaps the queues on the main loop without the LVGL
lock, so concurrent push/swap is a data race.
A single FreeRTOS mutex (QueueGuard, fail-closed) now guards all
four shared queues: _pending_mark_reads, _pending_drops,
_pending_name_writes, and the _index_commit_pending flag. Critical
sections are bounded vector operations only — no store I/O, no LVGL
lock, so acquisition cannot deadlock and the LVGL task is never
stalled on LittleFS. Lock order is LVGL-lock -> queue-mutex only;
queue sections never take the LVGL lock.
The mutex is created before the screen's LVGL_LOCK section and
deleted after it in the destructor.
Uses the microLXMF in-process message-metadata cache (bumped pin
6bea23c -> 58a6eb0, branch feat/conversation-preview-cache):
- Reopening a conversation (and background page fill / paging) is now
O(1) in-memory once warmed instead of re-reading each message file
from SPI LittleFS (~230ms/read, ~2.3s per open measured on the T-Deck).
- The cache table is PSRAM-allocated on ESP32; a static .bss placement
starved internal DRAM and the LVGL task's 8 KiB stack allocation
failed at boot ("Failed to create LVGL task", hang at startup logo).
- Long-press full-message view now defers to the main loop: the LVGL
event handler only records the hash; tick_pending_full_message() does
load_message_content() (uncapped, no msgpack unpack) off the LVGL
task, then builds the modal. Fixes the capped (600-char) text that
the in-memory rows carry.
- A peer change cancels an in-flight background fill from the previous
conversation (it would otherwise prepend the old conversation's
rows into the new one).
Validated on the T-Deck (T-Deck Plus, 8MB PSRAM): warm opens
reads=3 sync=0-2ms total=15-35ms (was ~2.3s); cold first open per
conversation still pays ~0.6-1.2s disk while warming the cache; no
panics/watchdogs over a ~6-minute interaction session.
Host gates: microLXMF conformance 6/6 (incl. 32-assertion
test_message_metadata_cache), native reference test against pinned
58a6eb0, tdeck build SUCCESS.
On-device capture confirmed the fix (cold-boot tap: fallbacks=0,
gather=2ms, total=48ms vs 2086ms pre-fix; repeat taps no longer
perpetually fall back). All [PERF] timing markers and stage
variables are removed from refresh()/show()/render_route; the
branch is now clean of diagnostic code.
On-device [PERF] capture decomposed the remaining ~1s tap latency:
each uncached conversation costs one SPI-LittleFS metadata read
(~230ms). Repeat taps still paid 2 perpetual fallbacks (empty-
content tails never got cached), and the first tap after every
boot paid all 9 (the in-memory repop was never committed to the
index).
- bump microLXMF pin c8d3156 -> 6bea23c (feat/conversation-
preview-cache): preview_valid index flag so 'cached empty'
differs from 'unpopulated'; bounded preview copy (the old
strncpy read past the non-terminated content Bytes); public
commit_index().
- refresh() re-pops empty-content tails as a valid cached
preview and arms a one-shot deferred index commit;
UIManager::update() drains it out-of-lock (flush_pending_
index_commit, between drops and mark-read) so the warmed
previews persist and the next cold boot reads them from the
index.
- fix the [PERF] skip-log total= wrap (printed p_t_diff - p_t0,
a uint32 underflow; total was already p_t_diff).
[PERF] instrumentation stays in this commit (temporary,
marked); it is removed before merge once the fix is validated
on-device.
Temporary, removable instrumentation (marked [PERF] throughout):
- refresh(): gather / diff / rebuild stage ms + metadata-fallback count,
one [PERF] convlist line per call (skip vs build)
- show(): unhide + focus-group ms
- render_route(MESSAGES): whole route window incl. hide_all_screens
All INFO-level so they survive DEBUG-off and land in the serial capture.
refresh() unconditionally ran lv_obj_clean(_list) and recreated 5-7 LVGL
objects per row on every call - every navigation back to Messages, every
750ms coalesced inbound batch, and the periodic name-resolution sweep -
even when nothing had changed. With the store read now O(1) (index
preview cache), that widget churn was the remaining gap vs NomadNet /
Network / Maps, which build once and just unhide.
refresh() now gathers row data first (index preview + hash + unread, no
message-file I/O) and diffs it against the rendered rows; the rebuild
only happens when data actually changes. Rows keep focus-group
membership and the screen keeps its scroll position on the no-change
path. The per-refresh 'Found N conversations' log drops from INFO to
DEBUG (serial output under the LVGL lock stalls the render task).
get_conversations() ordering is deterministic (last_activity desc with
peer-hash tie-break), so the index-aligned comparison is stable.
refresh() previously opened + parsed each conversation's newest message
file (load_message_metadata) on every list refresh, which dominated
list-load time on SPI LittleFS. It now reads the per-conversation
last-message preview + timestamp from the store's in-memory index
(O(1), zero I/O) and falls back to load_message_metadata only when the
index has no cached preview — the first refresh after a firmware
upgrade or after a corrupt-tail drop — then writes the preview back
through so the fallback happens at most once per conversation per
firmware generation. The write-through is skipped when drops were
queued (the drained deletes move the tail, so a preview written for the
old tail would be stale for one refresh).
Host benchmark (x86 + POSIX fs, 24 msgs/conv): the per-conversation
store-load work drops from ~0.149ms (9 convs) / ~0.342ms (20 convs) to
~0.002ms / ~0.005ms, and the cold-boot path (store reconstructed from
disk) matches the warm path because the preview now persists in the
index.
Bumps the microLXMF pin to c8d3156 (feat/conversation-preview-cache)
in platformio.ini, the release audit, and the native reference test.
Adds tests/microlxmf/bench_conversation_list_load.cpp (baseline vs
index warm/cold) and the bench target in tests/microlxmf/CMakeLists.txt.
When a conversation's newest message is unreadable, walk the index
newest-to-oldest for the newest readable preview and queue the
unreadable messages for deletion (deferred out of the LVGL lock,
drained by UIManager::update()). The store's delete_message() commits
the index and updates last_message_hash, so the next refresh converges
to the same preview and the row never disappears over one bad message.
PR #10 merged to main; swap the temporary branch-head pin (59ca70a) for
the merge commit so the consumer tracks a stable main revision. Update
the release-audit allowlist in the same commit. Firmware source is
byte-identical to 59ca70a (the merge added only the CI workflow commit).
refresh() did heavy unnecessary work per conversation, on the LVGL
render task under the render lock:
- get_messages_for_conversation() copied the full 256-slot hash array
(8KB) just to read the newest hash (messages.back());
- load_message() on that hash read the payload file ~3x, JSON-parsed
it twice (including the large hex 'packed' blob), hex-decoded and
msgpack-unpacked the whole message — to extract a 30-char preview
and a timestamp;
- the unread badge was rendered from a hardwired 0 even though the
store maintains and persists unread_count.
Now:
- newest hash via the new O(1) MessageStore::get_last_message_hash
(index tail — always hot-tier, no I/O, no array copy);
- preview/timestamp via load_message_metadata (single open + filtered
parse, the fast path ChatScreen already uses for the same fields);
- unread badge from MessageStore::get_conversation_unread_count;
- badge cleared + mark-read on open (click) and when a message lands
in the currently-viewed chat, with the LittleFS index commit
deferred out of the LVGL lock (UIManager::update), matching the
existing deferred display-name write-through pattern.
Pins microLXMF 59ca70a (PR #10, temporary branch head) for the two new
accessors.
Greptile note: the derive-before-apply position check could pass even if
the color call regressed to the wrong branch. Tighten the contract test
to assert exactly one label-color call, located after the marker loop's
hidden-marker continue (i.e. in the main render branch, not the hide
path), and that the once-per-frame derivation precedes it.
Marker (pin) labels previously inherited the app's default text color,
which reads white and disappears on light basemaps. Set each label's
text color in applyFrame() against the active style: black by default
(light basemaps osm-bright/positron/toner) and white only on the one
dark basemap (dark-matter). Re-evaluated every frame so a style switch
re-colors visible labels on the next applied frame.
Add a contract test pinning the dark-basemap detection and the
black-by-default / white-on-dark ternary.
The microreticulum-native job installs libbz2-dev without a preceding
apt-get update. When the ubuntu-latest runner's apt index is stale, the
install 404s against the security mirror (pinned libbz2 1.0.8-5.1build0.1
no longer at the indexed URL) and the job fails before any test runs.
This flake is pre-existing — main's own merge commit for PR #92 (520554b)
failed the same way. The sibling pyxis-pytest job already updates before
its install; apply the same here.
A message whose source identity is KNOWN but whose signature fails to
validate is spoofed or malicious and must not be rendered. The
opportunistic (on_packet) and direct (on_resource_concluded) router
paths already reject these, but the propagated (store-and-forward) path
in process_propagated_lxmf queues them without a signature check, so
UIManager::on_message_received is the single choke point that covers
all three inbound routes.
Drop the message at the top of on_message_received — before the key
request, location ingest, persistence, chat render, and notification
beep — when !signature_validated() && reason == SIGNATURE_INVALID.
SOURCE_UNKNOWN (first contact) is untouched: those still render and
trigger the bounded key request from PR #92. Validated messages are
unaffected.
Add a source-level contract test locking in the drop gate's ordering
relative to every side effect and its enum specificity.
The tdeck toolchain rejects aggregate initialization of Entry (NSDMI
makes the class non-aggregate under C++11). Use a default constructor
and field assignment in record_request.
Greptile P2 on PR #92: the per-inbound-location-frame INFOF logged peer
bytes, precise coordinates, and source-clock skew unconditionally in the
release build. The missing-map-pin investigation is closed (announce
timing, verified physically), so remove the temporary instrumentation
and its ingest_diag_result plumbing, and update the feature comment to
describe the final rate-limit semantics instead of 'remove with the
fix'.
Greptile P2 on PR #92: with 64 tracked sources, evicting the oldest
entry discarded its open cooldown, so a flood of distinct bogus
identities reset other sources' windows and forced unbounded path
requests (each answered by every peer holding the announce).
- Per-identity cooldown 5 min -> 30 min.
- The table no longer evicts an open window: while all 64 slots hold
unexpired windows, never-before-seen identities are deferred until a
slot frees (at most one 30-minute window) instead of dropping
someone else's cooldown. Open windows are only ever pruned after
they expire.
- Aggregate bound: at most kMaxTrackedSources automatic path requests
per rolling 30-minute window, capping the worst-case network cost of
a rotating-identity flood.
Host tests extended: saturated-table deferral, 512-identity flood bound
(one window and across consecutive windows), lazy expiry/prune,
boundary checks at the 30-minute cooldown.
Move the rate-limit/cooldown/cap decision out of the UIManager.cpp
anonymous namespace into a pure, header-only policy
(UI/LXMF/UnknownSourceKeyRequest.h) so it is host-testable without the
ESP/microReticulum stack. UIManager keeps only the side effect
(Transport::request_path).
Adds tests/native/test_unknown_source_key_request.{cpp,py} (24 checks,
ASan+UBSan): new-source request, 5-min cooldown boundary, re-record
resets the window, per-source independence, 64-entry cap with oldest
eviction, and steady-state cost.
on_message_received now fires a rate-limited RNS path request when an
LXMF message arrives from a SOURCE_UNKNOWN sender, mirroring Sideband's
'Query Network For Keys' button (RNS.Transport.request_path). Any peer
that already knows the source's announce (hub, phone, other node) can
answer with the cached identity, letting the NEXT message from that
peer validate and reach location ingest.
5-minute per-source cooldown, 64-entry cap. App-only diagnostic;
remove with the other ingest diagnostics once the root cause closes.
Torlando