QSPI-bringup stability enhancements

docs/ARCHITECTURE.md

@@ -769,7 +769,9 @@ Full command reference with constraints and remote-admin restrictions: `Repeater
 ### 7.2 DataStore (`adapters/datastore/`)
 
 - **Internal**: LittleFS on flash (`/lfs`), 256-byte cache for reduced flash I/O
-- **External**: Optional LittleFS on QSPI (`/ext`) with auto-migration
+- **External**: Optional LittleFS on QSPI (`/ext`) with auto-migration. The flash is `zephyr,deferred-init` and `/ext` is
+  not automounted (`boards/common/qspi-ext.dtsi`): `zephcore_fs_mount_ext()` brings it up on first use, because a
+  boot-time probe could beat a cold power rail (first boot after a UF2 update) and silently drop to internal flash
 - **BLE bonds**: NVS (`storage_partition`, 0xD0000 on nRF52) via Zephyr settings backend (≥1.16.2)
 - **Prefs**: `prefs.json` through upstream's `ConfigSerializer`, with upstream's key names for shared fields and ZephCore's own under `zc` (see §13); the older binary file is read once to migrate and kept
 - **Contacts**: 152-byte records, stored on external flash if available

docs/DESIGN.md

@@ -312,6 +312,9 @@ it polls external state. Repeater and room server share one loop (`server_main_c
 
 ### 8.5 Persistence
 - LittleFS `/lfs` (internal) and optional `/ext` (QSPI). BLE bonds in NVS, isolated from LittleFS on nRF52/ESP32.
+- `/ext` is never probed at boot: its flash is deferred-init and `zephcore_fs_mount_ext()` brings it up on first
+  use, since a boot-time probe can beat a switched rail that is still cold (first boot after a UF2 update). If it
+  still fails, the companion falls back to `/lfs` and the next boot that mounts `/ext` keeps the `/ext` copy.
 - Prefs are upstream's `prefs.json` (ZephCore fields under `zc`); the legacy binary layouts are migrated once and
   kept for downgrades.
 - Atomic replace for identity, prefs and channels; contacts on `/ext` only.

releasenotes/RELEASE_NOTES_1.17.6a-zephcore.md

@@ -100,6 +100,26 @@ write exactly what is already stored is skipped.
 Pending writes are now flushed before every clean reboot and power-off (CLI, UI, bootloader and OTA
 paths), not only before the companion app's reboot command.
 
+## Contacts and channels no longer disappear after an update
+
+Boards with external QSPI flash keep contacts, channels and the advert cache on it (`/ext`); identity and
+settings stay on the chip's internal flash. On some of these boards the flash is powered through a switched
+rail, and the firmware used to check for the flash very early in boot, just after switching that rail on.
+After a normal reboot the rail is still charged and the flash answers. After a UF2 update the rail has been
+off for the whole time the bootloader ran, the flash was not ready in time, and `/ext` was silently not
+used: the node came up with an empty contact list and default channels, while its name, radio settings
+and keys were all intact. Reported on a ThinkNode M1 upgrading to 1.17.5.
+
+The flash is now brought up when contacts are first loaded instead of during boot, on every board with
+`/ext`: ThinkNode M1, T-Echo, T-Impulse Plus, MeshTracker X1 (which already had this), SenseCAP Solar,
+Wio Tracker L1, Wio Tracker L1 Pro 1W and XIAO nRF52840.
+
+> [!NOTE]
+> **If this happened to you, your contacts were not erased.** They are still on the external flash, and
+> the next boot that finds it brings them back (a plain reboot is usually enough). Contacts and channels
+> added while the list was empty were stored on internal flash instead; they are dropped when the old
+> list returns, so re-add those few by hand.
+
 ## Remote CLI now executes on the device (protocol v14)
 
 The companion protocol's v14 additions — `CMD_RUN_CLI_COMMAND`/`PACKET_CLI_REPLY`, and

zephcore/adapters/datastore/ZephyrDataStore.cpp

@@ -51,52 +51,13 @@ bool ZephyrDataStore::mount()
 		return false;
 	}
 
-	/* Check if external QSPI was automounted */
-	if (zephcore_fs_is_mounted(extMountPoint())) {
-		ext_lfs_mounted = true;
-		LOG_INF("External QSPI LittleFS at %s (automounted, 100 blobs)", extMountPoint());
-	} else {
-#if DT_NODE_EXISTS(DT_NODELABEL(qspi_lfs))
-		/* Boot-time automount can miss the QSPI on the first boot after a
-		 * factory-erase (blank flash) or an early-boot timing race with QSPI
-		 * init.  Retry the mount explicitly (fs_mount auto-formats blank flash,
-		 * and mounts valid data without touching it) so contacts/channels land
-		 * on /ext.  Without this the store silently falls back to internal /lfs,
-		 * and the next boot that does mount /ext runs a needless contact
-		 * migration — the "Migrating contacts to external storage" churn. */
-#if DT_NODE_HAS_PROP(DT_COMPAT_GET_ANY_STATUS_OKAY(nordic_qspi_nor), zephyr_deferred_init)
-		/* The flash is deferred-init: probing it at boot raced its own
-		 * power rail and the JEDEC read came back 00 00 00, so the
-		 * driver failed and /ext could never mount. Initialising it here
-		 * instead means the part has had until first use to wake up —
-		 * over a second — rather than us guessing a settling delay. */
-		{
-			const struct device *qspi_dev =
-				DEVICE_DT_GET(DT_COMPAT_GET_ANY_STATUS_OKAY(nordic_qspi_nor));
-
-			if (!device_is_ready(qspi_dev)) {
-				int drc = device_init(qspi_dev);
-
-				LOG_INF("QSPI flash deferred init: rc=%d ready=%d",
-					drc, (int)device_is_ready(qspi_dev));
-			}
-		}
-#endif
-
-		FS_FSTAB_DECLARE_ENTRY(DT_NODELABEL(qspi_lfs));
-		int rc = fs_mount(&FS_FSTAB_ENTRY(DT_NODELABEL(qspi_lfs)));
-		if (zephcore_fs_is_mounted(extMountPoint())) {
-			ext_lfs_mounted = true;
-			LOG_INF("External QSPI LittleFS at %s (mounted on retry, rc=%d)",
-				extMountPoint(), rc);
-		} else {
-			ext_lfs_mounted = false;
-			LOG_WRN("External QSPI mount retry failed (rc=%d) - using internal only (20 blobs)", rc);
-		}
-#else
-		ext_lfs_mounted = false;
-		LOG_INF("External QSPI NOT mounted at %s - using internal only (20 blobs)", extMountPoint());
-#endif
+	/* External QSPI: deferred flash init + explicit mount, never at boot.
+	 * If it fails, contacts/channels fall back to internal /lfs, and the
+	 * next boot that does mount /ext keeps the /ext copy. */
+	ext_lfs_mounted = zephcore_fs_mount_ext();
+	if (!ext_lfs_mounted) {
+		LOG_INF("External QSPI not mounted at %s - using internal only (20 blobs)",
+			extMountPoint());
 	}
 
 	return true;

zephcore/adapters/datastore/ZephyrFsFormat.c

@@ -7,6 +7,7 @@
 #include "ZephyrFsUtil.h"
 
 #include <zephyr/devicetree.h>
+#include <zephyr/device.h>
 #include <zephyr/fs/fs.h>
 #include <zephyr/storage/flash_map.h>
 #include <zephyr/logging/log.h>
@@ -36,6 +37,53 @@ static void flatten(uint8_t id, const char *tag)
 	flash_area_close(fap);
 }
 
+#if FIXED_PARTITION_EXISTS(qspi_storage_partition)
+/* qspi-ext.dtsi makes the flash under /ext zephyr,deferred-init, so nothing
+ * probes it at boot. The flash sits behind a GPIO-switched rail on several
+ * boards, and the boot-time JEDEC read went out before a cold rail had come
+ * up: the driver failed, /ext stayed unmounted and contacts silently fell
+ * back to internal flash. A warm reboot keeps the rail charged; the first
+ * boot after a UF2 update (seconds in the bootloader, rail off) did not.
+ * Bringing the part up here, on first use, is long after the rail. Each role
+ * reaches this on its own path, so it is idempotent: -EALREADY means an
+ * earlier call already ran the init, whatever the result was. */
+static void ext_flash_init(void)
+{
+	const struct device *dev = FIXED_PARTITION_DEVICE(qspi_storage_partition);
+
+	if (!device_is_ready(dev)) {
+		int rc = device_init(dev);
+
+		LOG_INF("%s flash init: rc=%d ready=%d", EXT_MNT_POINT, rc,
+			(int)device_is_ready(dev));
+	}
+}
+#endif
+
+bool zephcore_fs_mount_ext(void)
+{
+#if DT_NODE_EXISTS(DT_NODELABEL(qspi_lfs))
+	if (zephcore_fs_is_mounted(EXT_MNT_POINT)) {
+		return true;
+	}
+	ext_flash_init();
+
+	/* fs_mount() auto-formats blank flash and mounts valid data untouched. */
+	FS_FSTAB_DECLARE_ENTRY(DT_NODELABEL(qspi_lfs));
+	int rc = fs_mount(&FS_FSTAB_ENTRY(DT_NODELABEL(qspi_lfs)));
+	bool mounted = zephcore_fs_is_mounted(EXT_MNT_POINT);
+
+	if (mounted) {
+		LOG_INF("%s mounted (rc=%d)", EXT_MNT_POINT, rc);
+	} else {
+		LOG_ERR("%s mount failed (rc=%d)", EXT_MNT_POINT, rc);
+	}
+	return mounted;
+#else
+	return false;
+#endif
+}
+
 bool zephcore_fs_format_all(bool *out_ext_mounted)
 {
 	LOG_INF("zephcore_fs_format_all: starting...");
@@ -71,6 +119,8 @@ bool zephcore_fs_format_all(bool *out_ext_mounted)
 #endif
 
 #if FIXED_PARTITION_EXISTS(qspi_storage_partition)
+	/* A repeater never mounts /ext, so its flash is still uninitialised. */
+	ext_flash_init();
 	flatten(PARTITION_ID(qspi_storage_partition), "qspi_storage_partition");
 #endif
 
@@ -93,14 +143,8 @@ bool zephcore_fs_format_all(bool *out_ext_mounted)
 	 * contacts/channels save to /lfs and get needlessly migrated back to
 	 * /ext on the next boot ("Migrating contacts to external storage"). */
 	{
-		int ext_rc = fs_mount(&FS_FSTAB_ENTRY(DT_NODELABEL(qspi_lfs)));
-		bool ext_mounted = zephcore_fs_is_mounted(EXT_MNT_POINT);
+		bool ext_mounted = zephcore_fs_mount_ext();
 
-		if (ext_mounted) {
-			LOG_INF("format: %s remounted (rc=%d)", EXT_MNT_POINT, ext_rc);
-		} else {
-			LOG_ERR("format: %s remount failed (rc=%d)", EXT_MNT_POINT, ext_rc);
-		}
 		if (out_ext_mounted) {
 			*out_ext_mounted = ext_mounted;
 		}

zephcore/adapters/datastore/ZephyrFsFormat.h

@@ -34,6 +34,16 @@ extern "C" {
  */
 bool zephcore_fs_format_all(bool *out_ext_mounted);
 
+/**
+ * @brief Mount /ext (QSPI) on first use, initialising its deferred flash.
+ *
+ * Idempotent. The flash is zephyr,deferred-init (qspi-ext.dtsi) and /ext is
+ * not automounted, so this is the only way /ext comes up.
+ *
+ * @return true if /ext is mounted; always false on boards with no QSPI.
+ */
+bool zephcore_fs_mount_ext(void);
+
 #ifdef __cplusplus
 }
 #endif

zephcore/boards/common/qspi-ext.dtsi

@@ -1,6 +1,6 @@
 /*
  * SPDX-License-Identifier: MIT
- * QSPI external flash /ext automount — optional add-on for any platform
+ * QSPI external flash /ext — optional add-on for any platform
  *
  * See nrf52_partitions_sdv7.dtsi for full include guide.
  *
@@ -8,7 +8,8 @@
  * and blobs to QSPI flash (more space, faster writes on native 4KB blocks).
  * Prefs and identity stay on internal flash.
  *
- * Requires board DTS to define: qspi_storage_partition
+ * Requires board DTS to define: qspi_flash (the flash device node) and
+ * qspi_storage_partition
  *
  * Usage in board overlay (in addition to partition or filesystem include):
  *   #include "../../common/qspi-ext.dtsi"
@@ -16,6 +17,26 @@
  * Works with all platforms using /lfs mount.
  */
 
+/*
+ * The flash is not probed at boot. On boards where it sits behind a switched
+ * rail, the boot-time JEDEC read could beat a cold rail (first boot after a
+ * UF2 update, or a power-up): the driver failed, /ext stayed unmounted and
+ * contacts/channels silently fell back to internal flash. A warm reboot keeps
+ * the rail charged, which is why only updates showed it. The store initialises
+ * the flash and mounts /ext on first use (zephcore_fs_mount_ext()), so /ext is
+ * not automounted either.
+ *
+ * t-exit-dpd (nanoseconds, so 50 us) is how long the driver waits after
+ * Release-from-Deep-Power-Down, which it sends once at init to wake a part a
+ * previous firmware left asleep. Boards used to carry 30, 35 or 50 us each; 50
+ * covers all of them, and the extra few us run once per flash init. It
+ * assumes a powered part, so it does nothing for the cold-rail race above.
+ */
+&qspi_flash {
+	zephyr,deferred-init;
+	t-exit-dpd = <50000>;
+};
+
 / {
 	fstab {
 		/* External QSPI for contacts, channels, blobs */
@@ -23,7 +44,6 @@
 			compatible = "zephyr,fstab,littlefs";
 			mount-point = "/ext";
 			partition = <&qspi_storage_partition>;
-			automount;
 			read-size = <16>;
 			prog-size = <16>;
 			cache-size = <64>;

zephcore/boards/esp32/heltec_wireless_tracker_v2/heltec_wireless_tracker_v2_procpu.dts

@@ -88,11 +88,11 @@
 		enable-gpios = <&gpio0 3 (GPIO_ACTIVE_HIGH | ESP32_GPIO_SLEEP_HOLD_EN)>;
 		regulator-boot-on;
 		regulator-always-on;
-		/* Documents the rail's requirement, but do not rely on it. It is
-		 * inert twice over: regulator-boot-on sends regulator_common_init()
-		 * down the refcount-only branch so regulator_delay() never runs, and
-		 * regulator-always-on makes regulator_enable() return at its first
-		 * line. The real settle time comes from the ST7735R driver's own
+		/* The rail's requirement. It runs once at boot (Zephyr
+		 * 3b4821f1ba6, pinned from 1.17.5; it was skipped before), and
+		 * never again: regulator-always-on makes regulator_enable() return
+		 * at its first line. Do not rely on it alone: the real settle time
+		 * comes from the ST7735R driver's own
 		 * mipi_dbi_reset() + exit-sleep delay at DISPLAY_INIT_PRIORITY 85;
 		 * the GNSS side is safe because gnss-nmea-generic never probes the
 		 * receiver, it just parses whatever the UART delivers. */
@@ -123,9 +123,9 @@
 		enable-gpios = <&gpio0 7 (GPIO_ACTIVE_HIGH | ESP32_GPIO_SLEEP_HOLD_EN)>;  /* held across light sleep */
 		regulator-boot-on;
 		regulator-always-on;
-		/* Documents the FEM's requirement, but do not rely on it — inert for
-		 * the same two reasons as vext above. Harmless because the FEM is a
-		 * passive gain block with no bus to probe. */
+		/* The FEM's requirement; runs once at boot, as vext above. Nothing
+		 * depends on it: the FEM is a passive gain block with no bus to
+		 * probe. */
 		startup-delay-us = <1000>;
 	};
 

zephcore/boards/example_board/board.overlay

@@ -571,7 +571,7 @@
  *         size = <16777216>;
  *         has-dpd;
  *         t-enter-dpd = <3000>;
- *         t-exit-dpd = <30000>;
+ *         (t-exit-dpd comes from qspi-ext.dtsi: 50 us)
  *         quad-enable-requirements = "S2B1v1";
  *     };
  * };

zephcore/boards/nrf52840/heltec_t096/heltec_t096_nrf52840.dts

@@ -104,10 +104,9 @@
 		regulator-name = "gps-power";
 		enable-gpios = <&gpio0 6 GPIO_ACTIVE_LOW>;
 		regulator-boot-on;
-		/* Documents the module's requirement, but do not rely on it:
-		 * regulator-boot-on sends regulator_common_init() down the
-		 * refcount-only branch, so regulator_delay() never runs and this
-		 * value is inert. Harmless here because gnss-nmea-generic never
+		/* The module's requirement. It runs at boot (Zephyr 3b4821f1ba6,
+		 * pinned from 1.17.5; it was skipped before), but nothing here
+		 * depends on it: gnss-nmea-generic never
 		 * probes the receiver — it attaches a UART and parses whatever
 		 * arrives, so an unpowered module means no sentences yet, not a
 		 * failed init. ZephyrGPSManager owns the rail at runtime via
@@ -130,11 +129,10 @@
 		regulator-name = "vfem-enable";
 		enable-gpios = <&gpio0 30 GPIO_ACTIVE_HIGH>;
 		regulator-boot-on;
-		/* Documents the FEM's requirement, but do not rely on it:
-		 * regulator-boot-on sends regulator_common_init() down the
-		 * refcount-only branch, so regulator_delay() never runs and this
-		 * value is inert. Harmless here because the FEM is a passive
-		 * gain block with no bus to probe — nothing addresses it at init. */
+		/* The FEM's requirement. It runs at boot (Zephyr 3b4821f1ba6,
+		 * pinned from 1.17.5; it was skipped before), but nothing depends
+		 * on it: the FEM is a passive gain block with no bus to probe —
+		 * nothing addresses it at init. */
 		startup-delay-us = <1000>;
 	};
 
@@ -154,10 +152,9 @@
 		regulator-name = "tft-pwr-enable";
 		enable-gpios = <&gpio0 26 GPIO_ACTIVE_HIGH>;
 		regulator-boot-on;
-		/* Documents the panel's requirement, but do not rely on it:
-		 * regulator-boot-on sends regulator_common_init() down the
-		 * refcount-only branch, so regulator_delay() never runs and this
-		 * value is inert. The real settle time comes from the ST7735R
+		/* The panel's requirement. It runs at boot (Zephyr 3b4821f1ba6,
+		 * pinned from 1.17.5; it was skipped before), but do not rely on
+		 * it alone: the real settle time comes from the ST7735R
 		 * driver's own mipi_dbi_reset() + exit-sleep delay, at
 		 * DISPLAY_INIT_PRIORITY 85 against this rail's 75. */
 		startup-delay-us = <10000>;

zephcore/boards/nrf52840/heltec_t1/heltec_t1_nrf52840.dts

@@ -168,10 +168,9 @@
 		regulator-name = "gps-power";
 		enable-gpios = <&gpio0 4 GPIO_ACTIVE_LOW>;
 		regulator-boot-on;
-		/* Documents the module's requirement, but do not rely on it:
-		 * regulator-boot-on sends regulator_common_init() down the
-		 * refcount-only branch, so regulator_delay() never runs and this
-		 * value is inert. Harmless here because gnss-nmea-generic never
+		/* The module's requirement. It runs at boot (Zephyr 3b4821f1ba6,
+		 * pinned from 1.17.5; it was skipped before), but nothing here
+		 * depends on it: gnss-nmea-generic never
 		 * probes the receiver — it attaches a UART and parses whatever
 		 * arrives, so an unpowered module means no sentences yet, not a
 		 * failed init. ZephyrGPSManager owns the rail at runtime via
@@ -234,10 +233,9 @@
 		regulator-name = "tft-pwr-enable";
 		enable-gpios = <&gpio0 13 GPIO_ACTIVE_LOW>;
 		regulator-boot-on;
-		/* Documents the panel's requirement, but do not rely on it:
-		 * regulator-boot-on sends regulator_common_init() down the
-		 * refcount-only branch, so regulator_delay() never runs and this
-		 * value is inert. The real settle time comes from the ST7735R
+		/* The panel's requirement. It runs at boot (Zephyr 3b4821f1ba6,
+		 * pinned from 1.17.5; it was skipped before), but do not rely on
+		 * it alone: the real settle time comes from the ST7735R
 		 * driver's own mipi_dbi_reset() + exit-sleep delay, at
 		 * DISPLAY_INIT_PRIORITY 85 against this rail's 75. */
 		startup-delay-us = <10000>;

zephcore/boards/nrf52840/lilygo_techo/lilygo_techo.dts

@@ -147,10 +147,10 @@
 	 * regulator framework brings them up before the SX1262 driver init —
 	 * REGULATOR_FIXED_INIT_PRIORITY 75 against LORA_INIT_PRIORITY 90.
 	 *
-	 * Ordering is all that guarantee covers. It does NOT mean a rail has
-	 * settled: regulator-boot-on sends regulator_common_init() down the
-	 * refcount-only branch, so regulator_delay() never runs and the
-	 * startup-delay-us below is inert. The SX1262 is safe because its driver
+	 * Ordering is all that guarantee covers. The startup-delay-us below does
+	 * run at boot (Zephyr 3b4821f1ba6, pinned from 1.17.5; it was skipped
+	 * before), but it is a settle floor, not a readiness check. The SX1262 is
+	 * safe because its driver
 	 * resets the part (5 ms NRESET pulse + 5 ms wait + BUSY poll) before any
 	 * SPI traffic. Anything put behind these rails that does NOT
 	 * reset-then-wait needs zephyr,deferred-init instead; see
@@ -385,7 +385,6 @@
 		size = <16777216>;            /* 2MB = 16Mbit */
 		has-dpd;
 		t-enter-dpd = <10000>;
-		t-exit-dpd = <35000>;
 	};
 };
 

zephcore/boards/nrf52840/lilygo_timpulse_plus/lilygo_timpulse_plus.dts

@@ -125,10 +125,9 @@
 		regulator-name = "pwr-enable";
 		enable-gpios = <&gpio0 14 GPIO_ACTIVE_HIGH>;
 		regulator-boot-on;
-		/* Documents the LDO's requirement, but do not rely on it:
-		 * regulator-boot-on sends regulator_common_init() down the
-		 * refcount-only branch, so regulator_delay() never runs and this
-		 * value is inert. What actually gives this rail time to settle is the
+		/* The LDO's requirement. It runs at boot (Zephyr 3b4821f1ba6,
+		 * pinned from 1.17.5; it was skipped before), but do not rely on
+		 * it alone. What also gives this rail time to settle is the
 		 * priority-30 pin above — every consumer here (QSPI 41, OLED 85,
 		 * SX1262 90) has real boot work between it and the rail, and the
 		 * SX1262 and SSD1315 both reset-then-wait besides. The GPS, which
@@ -357,7 +356,6 @@
 		size = <33554432>;            /* size is in BITS: 32Mbit = 4MB */
 		has-dpd;
 		t-enter-dpd = <10000>;        /* tDP 3us (conservative) */
-		t-exit-dpd = <35000>;         /* tRES1 8us (conservative) */
 	};
 };
 

zephcore/boards/nrf52840/meshtracker_x1/board.conf

@@ -21,8 +21,3 @@ CONFIG_ZEPHCORE_RADIO_LR2021=y
 # DRV2605L vibration motor
 CONFIG_HAPTICS=y
 CONFIG_HAPTICS_DRV2605=y
-
-# The flash sits behind the P0.15 power switch, modelled as a regulator so its
-# power-up time is a startup-delay rather than a race. Regulators come up at 75;
-# probe the flash after that.
-CONFIG_NORDIC_QSPI_NOR_INIT_PRIORITY=80

zephcore/boards/nrf52840/meshtracker_x1/meshtracker_x1_nrf52840.dts

@@ -113,12 +113,11 @@
 		regulator-name = "flash-power";
 		enable-gpios = <&gpio0 15 GPIO_ACTIVE_HIGH>;
 		regulator-boot-on;
-		/* Documents the part's requirement, but do not rely on it:
-		 * regulator_fixed_init() drives the enable GPIO active and then
-		 * passes is_enabled=true into regulator_common_init(), which
-		 * takes the refcount-only branch — regulator_delay() sits on the
-		 * other one and never runs. The flash gets its settling time
-		 * from deferred init instead (see the qspi_flash node). */
+		/* Runs at boot: since Zephyr 3b4821f1ba6 (pinned from 1.17.5)
+		 * regulator_fixed_init() sleeps this long after driving a
+		 * regulator-boot-on rail active. Before that it was skipped. The
+		 * flash's real margin is still its deferred init (qspi-ext.dtsi),
+		 * which runs when the store first mounts /ext. */
 		startup-delay-us = <100000>;
 	};
 
@@ -290,9 +289,10 @@
 	 *
 	 * sensor_power comes up at CONFIG_REGULATOR_FIXED_INIT_PRIORITY (75) and
 	 * both drivers run at 90 — microseconds later, with nothing scheduled in
-	 * between on this board. The node's startup-delay-us does not save them:
-	 * with regulator-boot-on, regulator_common_init() takes the refcount-only
-	 * branch and regulator_delay() never runs. Same trap as flash_power below.
+	 * between on this board. The node's startup-delay-us did not save them:
+	 * until Zephyr 3b4821f1ba6 (pinned from 1.17.5) it was skipped for a
+	 * regulator-boot-on rail. It now runs (5 ms), but the parts were never
+	 * validated against it, so the deferral below stays.
 	 * (vin-supply + regulator_enable() would not help either — that applies
 	 * off-on-delay-us, and only on the 0->1 refcount transition.)
 	 *
@@ -365,14 +365,11 @@
 	qspi_flash: qspi_flash@0 {
 		compatible = "nordic,qspi-nor";
 		reg = <0>;
-		/* Do not probe this at boot. The rail is switched on by
-		 * flash-power a few init priorities earlier, and the part is not
-		 * driving the bus yet when the driver's JEDEC read goes out —
-		 * it answers 00 00 00 at 0.4 s and a correct c8 40 17 by 5.9 s
-		 * on the very same boot. A fixed settling delay would only be a
-		 * guess at where in between it wakes; instead the driver is
-		 * initialised on demand, when the store first mounts /ext. */
-		zephyr,deferred-init;
+		/* Not probed at boot: qspi-ext.dtsi makes it zephyr,deferred-init.
+		 * This board is where that came from — probed a few init
+		 * priorities after flash-power switched the rail on, the part
+		 * answered 00 00 00 at 0.4 s and a correct c8 40 17 by 5.9 s on
+		 * the very same boot. */
 		/* Single-line, as every other QSPI board here uses.
 		 *
 		 * Quad needs the part's QE bit set, and the binding defaults
@@ -395,7 +392,6 @@
 		size = <67108864>;            /* 64Mbit = 8MB, in bits */
 		has-dpd;
 		t-enter-dpd = <3000>;
-		t-exit-dpd = <30000>;
 	};
 };
 

zephcore/boards/nrf52840/promicro_lr2021/promicro_lr2021.dts

@@ -62,10 +62,9 @@
 		regulator-name = "vcc3v3-enable";
 		enable-gpios = <&gpio0 13 GPIO_ACTIVE_HIGH>;
 		regulator-boot-on;
-		/* Documents the module's requirement, but do not rely on it:
-		 * regulator-boot-on sends regulator_common_init() down the
-		 * refcount-only branch, so regulator_delay() never runs and this
-		 * value is inert. The real settle time comes from lr20xx_hal_reset(),
+		/* The module's requirement. It runs at boot (Zephyr 3b4821f1ba6,
+		 * pinned from 1.17.5; it was skipped before), but do not rely on
+		 * it alone: the real settle time comes from lr20xx_hal_reset(),
 		 * which the driver runs (with two retries, 10 ms apart) before any
 		 * SPI traffic — and at LORA_INIT_PRIORITY 90 against this rail's 75.
 		 * Anything put behind this rail that does NOT reset-then-wait needs

zephcore/boards/nrf52840/promicro_sx1262/promicro_sx1262.dts

@@ -68,10 +68,9 @@
 		regulator-name = "vcc3v3-enable";
 		enable-gpios = <&gpio0 13 GPIO_ACTIVE_HIGH>;
 		regulator-boot-on;
-		/* Documents the module's requirement, but do not rely on it:
-		 * regulator-boot-on sends regulator_common_init() down the
-		 * refcount-only branch, so regulator_delay() never runs and this
-		 * value is inert. The real settle time comes from the SX126x driver's
+		/* The module's requirement. It runs at boot (Zephyr 3b4821f1ba6,
+		 * pinned from 1.17.5; it was skipped before), but do not rely on
+		 * it alone: the real settle time comes from the SX126x driver's
 		 * own reset (5 ms NRESET pulse + 5 ms wait + BUSY poll) before any
 		 * SPI traffic, at LORA_INIT_PRIORITY 90 against this rail's 75.
 		 * Anything put behind this rail that does NOT reset-then-wait needs

zephcore/boards/nrf52840/rak3401_1watt/rak3401_1watt.dts

@@ -93,10 +93,9 @@
 		regulator-name = "lora-pwr-enable";
 		enable-gpios = <&gpio0 21 GPIO_ACTIVE_HIGH>;
 		regulator-boot-on;
-		/* Documents the FEM's requirement, but do not rely on it:
-		 * regulator-boot-on sends regulator_common_init() down the
-		 * refcount-only branch, so regulator_delay() never runs and this
-		 * value is inert. The real settle time comes from the SX126x driver's
+		/* The FEM's requirement. It runs at boot (Zephyr 3b4821f1ba6,
+		 * pinned from 1.17.5; it was skipped before), but do not rely on
+		 * it alone: the real settle time comes from the SX126x driver's
 		 * own reset (5 ms NRESET pulse + 5 ms wait + BUSY poll) before any
 		 * SPI traffic, at LORA_INIT_PRIORITY 90 against this rail's 75.
 		 * Anything put behind this rail that does NOT reset-then-wait needs

zephcore/boards/nrf52840/rak_wismesh_tag/rak_wismesh_tag.dts

@@ -120,10 +120,9 @@
 		regulator-name = "lora-pwr-enable";
 		enable-gpios = <&gpio1 5 GPIO_ACTIVE_HIGH>;
 		regulator-boot-on;
-		/* Documents the module's requirement, but do not rely on it:
-		 * regulator-boot-on sends regulator_common_init() down the
-		 * refcount-only branch, so regulator_delay() never runs and this
-		 * value is inert. The real settle time comes from the SX126x driver's
+		/* The module's requirement. It runs at boot (Zephyr 3b4821f1ba6,
+		 * pinned from 1.17.5; it was skipped before), but do not rely on
+		 * it alone: the real settle time comes from the SX126x driver's
 		 * own reset (5 ms NRESET pulse + 5 ms wait + BUSY poll) before any
 		 * SPI traffic, at LORA_INIT_PRIORITY 90 against this rail's 75.
 		 * Anything put behind this rail that does NOT reset-then-wait needs

zephcore/boards/nrf52840/sensecap_solar/sensecap_solar_nrf52840.dts

@@ -218,7 +218,6 @@
 		size = <16777216>;
 		has-dpd;
 		t-enter-dpd = <3000>;
-		t-exit-dpd = <50000>;
 	};
 };
 

zephcore/boards/nrf52840/thinknode_m1/thinknode_m1.dts

@@ -359,7 +359,6 @@
 		size = <16777216>;            /* 2MB = 16Mbit */
 		has-dpd;
 		t-enter-dpd = <10000>;
-		t-exit-dpd = <35000>;
 	};
 };
 

zephcore/boards/nrf52840/wio_tracker_l1/board.overlay

@@ -331,7 +331,6 @@ sh1106: &oled {
 		size = <16777216>;        /* 2MB = 16Mbit */
 		has-dpd;
 		t-enter-dpd = <3000>;
-		t-exit-dpd = <50000>;     /* 50ms for reliable wakeup */
 		/* quad-enable not needed for single-bit mode */
 
 		partitions {

zephcore/boards/nrf52840/wio_tracker_l1_1w/wio_tracker_l1_1w_nrf52840.dts

@@ -152,10 +152,9 @@
 		regulator-name = "lora-pwr-enable";
 		enable-gpios = <&gpio0 14 GPIO_ACTIVE_HIGH>;
 		regulator-boot-on;
-		/* Documents the LDO + PA settle requirement but does not enforce
-		 * it: regulator-boot-on takes regulator_common_init()'s
-		 * refcount-only branch, so regulator_delay() never runs. The real
-		 * settle time comes from the SX126x driver's own reset (5 ms
+		/* The LDO + PA settle requirement. It runs at boot (Zephyr
+		 * 3b4821f1ba6, pinned from 1.17.5; it was skipped before), but do
+		 * not rely on it alone: the real settle time comes from the SX126x driver's own reset (5 ms
 		 * NRESET pulse + 5 ms wait + BUSY poll) 15 init priorities later.
 		 * See the identical note on rak3401_1watt. */
 		startup-delay-us = <10000>;
@@ -392,7 +391,6 @@
 		size = <16777216>;        /* 2MB = 16Mbit */
 		has-dpd;
 		t-enter-dpd = <3000>;
-		t-exit-dpd = <50000>;     /* 50ms for reliable wakeup */
 		/* quad-enable not needed for single-bit mode */
 	};
 };

zephcore/boards/nrf52840/xiao_nrf52840/xiao_nrf52840_nrf52840.dts

@@ -230,7 +230,6 @@
 		size = <16777216>;
 		has-dpd;
 		t-enter-dpd = <3000>;
-		t-exit-dpd = <50000>;
 	};
 };
 

zephcore/helpers/qspi_probe.c

@@ -67,6 +67,15 @@ static int qspi_rdid_probe(void)
 		return 0;
 	}
 
+	if (!flash->state->initialized) {
+		/* qspi-ext.dtsi defers the flash to first use, and a repeater
+		 * never mounts /ext. Stealing the pins now would break the init
+		 * a later factory reset runs, so there is nothing to probe. */
+		LOG_INF("qspi flash '%s' not initialised yet - RDID probe skipped",
+			flash->name);
+		return 0;
+	}
+
 	if (device_is_ready(flash)) {
 		/* Bit-banging steals the QSPI pins as GPIOs and leaves them
 		 * that way, which would break a flash that just came up fine.