BLE lifecycle memory: true BLE-off on nRF52, NimBLE host pools to PSRAM on S3 by thebentern · Pull Request #10902 · meshtastic/firmware

thebentern · 2026-07-06T14:15:19Z

Draft: needs on-hardware validation before review (checklist below). Part of the heap-recovery series (#10898, #10901, #10900, #10899).

nRF52: being "off" should be free

When config.bluetooth.enabled == false, the firmware previously fully initialized SoftDevice + Bluefruit and then muted advertising — the #4055-era workaround ("device hangs several days after boot", root cause never found). Being off cost ~3.5 KB of Bluefruit heap plus ~6 KB of BLE/SOC FreeRTOS task stacks, out of a ~115 KB arena.

This PR leaves the SoftDevice entirely off in that case:

setBluetoothEnable() short-circuits and clears useSoftDevice, so checkSDEvents() stops polling.
The deep-sleep GPREGRET writes (DFU_MAGIC_SKIP) now mirror lfs_assert()'s existing pattern: try sd_power_gpregret_*, fall back to a direct NRF_POWER->GPREGRET write when the SD isn't enabled.
NRF52Bluetooth::startDisabled() is removed. clearBonds() keeps its lazy bring-up for explicit bond maintenance, and re-enabling bluetooth applies via reboot, so no runtime re-init state machine is needed.

Why this is safe: with the SD disabled, the Adafruit BSP flash driver checks sd_softdevice_is_enabled() per operation and takes its synchronous path — the exact path every boot already uses for all filesystem work before Bluefruit.begin() runs. HardwareRNG uses the CC310, not the SD. sd_* calls with the SD disabled return an error code rather than faulting (the lfs_assert fallback has relied on this for years). The failure mode this deliberately revisits is the un-root-caused #4055 hang — hence draft + soak test below, and the change is a one-branch revert if the ghost reappears.

(Note: the nrf54l15 platform carries its own copy of the startDisabled() workaround — left untouched here, flagged as follow-up.)

ESP32-S3 + PSRAM: NimBLE host pools out of internal DRAM

New esp32s3_psram_base (variants/esp32s3/esp32s3.ini) adds CONFIG_BT_NIMBLE_MEM_ALLOC_MODE_EXTERNAL=y; all PSRAM S3 envs re-parent to it. NimBLE host allocations (MSYS mbufs, GATT/CCCD tables, bond storage) move to PSRAM, freeing ~5–8 KB of internal DRAM where WiFi+BLE coexistence hurts most. Controller memory (CONFIG_BT_CTRL_*) intentionally stays internal. Bond record formats are unchanged, so purgeIncompatibleBleBonds() (#10741) does not trigger.

Non-PSRAM envs are untouched — verified by statically resolving the merged custom_sdkconfig per env (t-deck/station-g2/thinknode-g3 gain the flag; heltec-v3 does not).

Verification so far

Native suite: 535/535 passed; rak4631 builds green; trunk clean.
CI matrix will exercise all S3 envs (the sdkconfig change triggers a framework rebuild per env).

Hardware checklist before undrafting

nRF52 BT-off: boot, config-save loops, reboot cycles, deep-sleep enter/wake, DFU entry (GPREGRET fallback)
nRF52 BT-off multi-day soak (the Workaround to disable bluetooth on NRF52 #4055 ghost)
nRF52 BT-on regression: advertise, pair (FIXED_PIN + RANDOM_PIN), reconnect, clearBonds
S3+PSRAM (t-deck or station-g2): pair/bond/reconnect/stale-bond matrix, confirm ESP.getFreeHeap() delta and no purgeIncompatibleBleBonds on same-version reboot

…AM on S3 nRF52: when config.bluetooth.enabled is false, leave the SoftDevice entirely off instead of fully initializing SoftDevice + Bluefruit and then muting advertising (the #4055-era workaround). Being "off" was costing ~3.5 KB of Bluefruit heap plus ~6 KB of BLE/SOC task stacks on a ~115 KB arena. With the SD disabled the BSP flash driver stays on its synchronous path (the same path every boot already uses before Bluefruit.begin()), checkSDEvents() keys off useSoftDevice, and the deep-sleep GPREGRET writes now mirror lfs_assert()'s direct-register fallback. clearBonds() still lazily brings the stack up for explicit bond maintenance; re-enabling bluetooth applies via reboot, so no runtime re-init path is needed. NRF52Bluetooth::startDisabled() is removed. ESP32-S3 with PSRAM: new esp32s3_psram_base sets CONFIG_BT_NIMBLE_MEM_ALLOC_MODE_EXTERNAL=y so NimBLE host pools (MSYS mbufs, GATT/CCCD tables, bond storage) allocate from PSRAM, freeing ~5-8 KB of internal DRAM on the WiFi+BLE coexistence path. Controller memory (CONFIG_BT_CTRL_*) stays internal. All PSRAM S3 envs re-parent to the new base; non-PSRAM envs are unchanged (verified by resolving the merged custom_sdkconfig per env). Native suite 535/535; rak4631 builds green. Hardware validation (pairing matrix, BT-off soak for the #4055 ghost) tracked in the PR.

coderabbitai · 2026-07-06T14:15:28Z

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: a01a3a59-9119-42b8-964c-01f6a0742004

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

_{Comment @coderabbitai help to get the list of available commands.}

github-actions · 2026-07-06T14:15:29Z

⚡ Try this PR in the Web Flasher

Warning

This is an automated, unreviewed CI test build. Back up your device configuration
before flashing, and only flash devices you are able to recover.

Supported boards built by this PR (26)

Device	Board	Platform
Crowpanel Adv 3.5 TFT	`elecrow-adv-35-tft`	esp32-s3
Heltec HT62	`heltec-ht62-esp32c3-sx1262`	esp32-c3
Heltec Mesh Node 096	`heltec-mesh-node-t096`	nrf52840
Heltec Mesh Node T1	`heltec-mesh-node-t1`	nrf52840
Heltec Mesh Node T114	`heltec-mesh-node-t114`	nrf52840
Heltec V3	`heltec-v3`	esp32-s3
Heltec V4	`heltec-v4`	esp32-s3
Raspberry Pi Pico	`pico`	rp2040
Raspberry Pi Pico W	`picow`	rp2040
RAK WisMesh Tag	`rak_wismeshtag`	nrf52840
RAK WisBlock 11200	`rak11200`	esp32
RAK WisBlock 11310	`rak11310`	rp2040
RAK3312	`rak3312`	esp32-s3
RAK WisBlock 4631	`rak4631`	nrf52840
Seeed SenseCAP Mesh-Tracker-X1	`seeed_mesh_tracker_X1`	nrf52840
Seeed Wio Tracker L1	`seeed_wio_tracker_L1`	nrf52840
Seeed Xiao NRF52840 Kit	`seeed_xiao_nrf52840_kit`	nrf52840
Seeed Xiao ESP32-S3	`seeed-xiao-s3`	esp32-s3
Station G2	`station-g2`	esp32-s3
Station G3	`station-g3`	esp32-s3
LILYGO T-Deck	`t-deck-tft`	esp32-s3
LILYGO T-Echo	`t-echo`	nrf52840
LILYGO T-Echo Plus	`t-echo-plus`	nrf52840
LILYGO T-Impulse Plus	`t-impulse-plus`	nrf52840
LilyGo T3-C6	`tlora-c6`	esp32-c6
Seeed SenseCAP T1000-E	`tracker-t1000-e`	nrf52840

Build artifacts expire on 2026-08-05. Updated for b1cbc9b.

github-actions · 2026-07-06T15:12:46Z

Firmware Size Report

22 targets | vs develop: 21 increased, 1 decreased, net +387,732 (+378.6 KB)

Target	Size	vs `develop`
`picow`	1,271,976	📈 +27,096 (+26.5 KB)
`pico2w`	1,246,660	📈 +26,144 (+25.5 KB)
`rak11310`	831,256	📈 +25,832 (+25.2 KB)
`pico`	808,504	📈 +25,816 (+25.2 KB)
`seeed_xiao_rp2040`	806,704	📈 +25,800 (+25.2 KB)

Show 17 more target(s)

Target	Size	vs `develop`
`pico2`	794,776	📈 +25,000 (+24.4 KB)
`seeed_xiao_rp2350`	792,920	📈 +24,984 (+24.4 KB)
`t-deck-tft`	3,832,064	📈 +21,392 (+20.9 KB)
`heltec-vision-master-e213-inkhud`	2,247,392	📈 +20,032 (+19.6 KB)
`elecrow-adv-35-tft`	3,435,520	📈 +18,896 (+18.5 KB)
`seeed-xiao-s3`	2,292,416	📈 +16,272 (+15.9 KB)
`heltec-ht62-esp32c3-sx1262`	2,150,576	📈 +15,584 (+15.2 KB)
`heltec-v3`	2,279,632	📈 +15,552 (+15.2 KB)
`rak3312`	2,287,600	📈 +15,360 (+15.0 KB)
`rak11200`	1,875,728	📈 +15,120 (+14.8 KB)
`station-g2`	2,281,072	📈 +14,720 (+14.4 KB)
`station-g3`	2,281,072	📈 +14,720 (+14.4 KB)
`t-eth-elite`	2,506,064	📈 +14,576 (+14.2 KB)
`tlora-c6`	2,382,272	📈 +13,968 (+13.6 KB)
`heltec-v4`	2,291,152	📈 +13,856 (+13.5 KB)
`rak3172`	182,492	📉 -3,804 (-3.7 KB)
`wio-e5`	239,444	📈 +816

Updated for 8a8d4f4

thebentern · 2026-07-06T16:53:46Z

Hardware validation on rak4631 (integration build with the full heap PR stack, 120-node NodeDB):

✅ BT-off boot logs Bluetooth disabled by config: leaving SoftDevice off; no Bluefruit init; node fully operational (LoRa TX/RX, PhoneAPI over serial, TMM, GPS probe).
✅ +9,560 B free heap with BT off: 53,564 → 63,124 (reproducible across 2 boots) — top of the predicted 7–9 KB band.
✅ Filesystem writes with the SoftDevice never enabled: set_owner (devicestate) + config write both persisted and survived reboot — the synchronous flash path works as designed.
✅ Multiple clean reboot cycles in BT-off mode.
✅ shutdown → deep sleep reaches system-off with the GPREGRET direct-register fallback (SD disabled, sd_power_gpregret_* unavailable); device woke by USB power-cycle and booted straight into the app (DFU correctly skipped).
✅ Re-enabling bluetooth via config + reboot brings SoftDevice + Bluefruit back (heap returns to exactly the BT-on figure).

Updated checklist state — still open: multi-day BT-off soak (the #4055 ghost; the test rig can run it), phone pairing matrix (needs a phone), and the ESP32-S3 PSRAM half (needs an S3+PSRAM board).

The first sweep only caught envs whose platformio.ini declares BOARD_HAS_PSRAM; another 23 variant dirs (42 envs) get the flag from their board JSON instead (EoRa-S3/Hub, t-beam-1w, t-watch-s3, tbeam-s3-core, unphone, vision-masters, mesh-tab, sensecap-indicator, t-deck-pro, wismesh-tap-v2, and friends). Re-parent them all to esp32s3_psram_base; inheritance verified by statically resolving the merged custom_sdkconfig per env (all 42 gain CONFIG_BT_NIMBLE_MEM_ALLOC_MODE_EXTERNAL=y; non-PSRAM controls heltec-v3 / heltec-wsl-v3 stay clean). Hardware-validated on a CDEBYTE EoRa-S3 (embedded 2 MB quad PSRAM): NimBLE initializes and advertises normally with host pools in PSRAM, and free internal DRAM in the client-attached steady state rises from 161,608 B to 179,296/178,828 B across boots - about 17.5 KB reclaimed, comfortably above the 5-8 KB estimate.

thebentern · 2026-07-06T17:58:00Z

ESP32-S3 PSRAM half validated on hardware (CDEBYTE EoRa-S3, ESP32-S3 with embedded 2 MB quad PSRAM AP_3v3):

Steady state (serial client, uptime ~80 s)	Free internal DRAM
`esp32s3_base` (NimBLE pools internal)	161,608 B
`esp32s3_psram_base` (this PR)	179,296 / 178,828 B (two boots)

≈17.5 KB of internal DRAM reclaimed — above the 5–8 KB estimate (host pools, mbufs, GATT/CCCD and bond storage all move; only the host task stack stays internal).
✅ NimBLE initializes and advertises normally with CONFIG_BT_NIMBLE_MEM_ALLOC_MODE_EXTERNAL=y: host task starts, MTU 517 set, advertising up, zero allocation errors across multiple boots.
Boot log confirms PSRAM residency for the rest of the LARGE-class profile too ([TM] NodeInfo cache … PSRAM flat array, warm store via ps_calloc).

Coverage fix pushed (3d0b7cc): the original sweep only re-parented envs whose platformio.ini declares BOARD_HAS_PSRAM; another 23 variant dirs (42 envs) get it from their board JSON (EoRa-S3/Hub, t-beam-1w, t-watch-s3, tbeam-s3-core, unphone, vision-masters, mesh-tab, sensecap-indicator, t-deck-pro, wismesh-tap-v2, …). All re-parented; merged custom_sdkconfig statically verified per env, with heltec-v3/wsl-v3 (no PSRAM) confirmed unaffected.

Still open before undraft: nRF52 BT-off multi-day soak, and a phone pairing/bond matrix on both platforms.

thebentern added the enhancement New feature or request label Jul 6, 2026

github-actions Bot added the needs-review Needs human review label Jul 6, 2026

thebentern mentioned this pull request Jul 6, 2026

nrf52840: right-size the SoftDevice RAM reservation (+8 KB heap arena) #10903

Merged

thebentern mentioned this pull request Jul 6, 2026

Right-size nRF52 heap tiers after 2.8.0 heap-exhaustion field reports #10898

Merged

Merge branch 'develop' into feat/ble-lifecycle-memory

2598444

github-actions Bot added the Stale Issues that will be closed if not triaged. label Aug 21, 2026

github-actions Bot closed this Sep 3, 2026

Navigation Menu

Uh oh!

Uh oh!

BLE lifecycle memory: true BLE-off on nRF52, NimBLE host pools to PSRAM on S3 - #10902

BLE lifecycle memory: true BLE-off on nRF52, NimBLE host pools to PSRAM on S3#10902
thebentern wants to merge 3 commits into
developfrom
feat/ble-lifecycle-memory

thebentern commented Jul 6, 2026

Uh oh!

coderabbitai Bot commented Jul 6, 2026 •

edited

Loading

Review skipped

Uh oh!

github-actions Bot commented Jul 6, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Jul 6, 2026

Uh oh!

thebentern commented Jul 6, 2026

Uh oh!

thebentern commented Jul 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Uh oh!

Conversation

thebentern commented Jul 6, 2026

nRF52: being "off" should be free

ESP32-S3 + PSRAM: NimBLE host pools out of internal DRAM

Verification so far

Hardware checklist before undrafting

Uh oh!

coderabbitai Bot commented Jul 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review skipped

Uh oh!

github-actions Bot commented Jul 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

⚡ Try this PR in the Web Flasher

Uh oh!

github-actions Bot commented Jul 6, 2026

Firmware Size Report

Uh oh!

thebentern commented Jul 6, 2026

Uh oh!

thebentern commented Jul 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

coderabbitai Bot commented Jul 6, 2026 •

edited

Loading

github-actions Bot commented Jul 6, 2026 •

edited

Loading