ESP32-S3 SDR: hardware quirks and how they are handled

Behaviour of the ESP32-S3 radio, its undocumented capture/DAC engine and the toolchain that matters when you build on the esp_sdr library, found while developing the apps. Each entry says what happens, how it shows up, and where the code deals with it. Link measurements: two XIAO ESP32S3 boards through a 30 dB attenuator; C6 entries: one XIAO ESP32C6.

Radio

The forced receive gain gets lost

What happens. A manual receive gain (esp_sdr_rx_set_gain(), forced through the vendor blob’s force_rx_gain()) is occasionally lost after minutes of traffic, on either board. The Wi-Fi AGC then takes over. With nothing on air it winds the gain up, and the quiet noise floor rises by up to 30 dB.

How it shows up. Carrier sense reads a busy channel all the time. A sender waiting for an ACK never transmits again. A receiver clips strong frames and decodes bursts of corrupted headers. It looks like hardware damage, but re-applying the gain cures it at once. The cause is unknown, and comparing the AGC force register (0x6001c02c) does not detect it: force_rx_gain() writes other registers too.

How it is handled. The library re-forces the gain at a capture when the last force is more than 20 ms old (esp_sdr_rx_gain_refresh() in lib/esp_sdr/src/esp_sdr_rx.c, GAIN_REFRESH_MS). Each re-force costs 14 to 51 us. esp_sdr_get_stats() counts them, and link status prints them on its engine line. A turnaround back to receive also forces the gain.

When debugging this kind of fault: any link set in apps/link re-applies the whole radio configuration, which hides it.

The C6 keeps the previous firmware’s modem state across a chip reset

What happens. A chip reset of the ESP32-C6 (through USB-Serial-JTAG, as esptool and every port open do) leaves its modem subsystem as the previous firmware left it. ESP-IDF resets it on a software restart (esp_system_reset_modules_on_exit()), a chip reset does not. After apps/wpan had run the C6’s native 802.15.4 radio, esp_sdr on the next firmware received with a noise floor about 25 dB up (rms 40 to 50 instead of 2 at fixed gain 40, the same at every gain index and frequency), and the vendor’s channel tuning put the PLL MHz off (2404 instead of 2412 MHz in its fractional divider). A power cycle or a deep sleep cleared it.

How it shows up. Tones from another board land at the wrong offset or not at all at Wi-Fi channel frequencies, and captures look like a much noisier receiver. tests/regression/lo_tuning failed its link check whenever it ran after tests/regression/ieee802154_link.

How it is handled. On the C6 the library resets the modem subsystem at PRE_KERNEL_1, before the Wi-Fi driver brings the radio up, as ESP-IDF’s restart path does (esp_sdr_modem_reset() in lib/esp_sdr/src/esp_sdr.c).

40 MHz mode widens the analog filters

The PHY’s channel bandwidth mode (esp_sdr_set_channel_bw(), the second argument of the blob’s set_chanfreq()) selects the analog filter. In its 40 MHz mode (2) the TX and RX path passes +-18 MHz within +-3 dB end to end: a 36 MHz OFDM frame decodes error free. The 20 MHz mode (0) cuts it off. apps/link defaults to CONFIG_APP_CBW=2. The spectrum edge of a single-carrier signal is a misleading measure of this, since it mixes the TX filter, the RX filter and the pulse shaping. Measure the channel per subcarrier instead (ofdm rx -v).

The transmitter compresses before the DAC words clip

Wide multicarrier signals distort in the analog TX path long before the 10-bit DAC words reach full scale. At an RMS amplitude of 120 DAC units, 116 subcarriers drop to 16 dB MER; at 60 to 90 they reach 26 to 29 dB. The digital clip counter stays at 0, and lowering the RX gain does not help, which places it in the transmitter. The OFDM default is amp 90.

Receive gain headroom

Through the 30 dB attenuator at TX step 4, RX gain 32 is the working point and 44 saturates the receiver (MER 4 dB).

MER ceilings

Signal

MER

Single carrier QAM, 20 Mbaud, 80 MS/s

about 31.5 dB, with or without attenuation

OFDM 16 MHz, DAC at 80 MS/s

about 31 dB

OFDM, DAC at 40 MS/s, 16 / 22 / 26 / 30 MHz

27 / 25 / 24 / 23 dB

The MER ceiling rules out 256-QAM on OFDM (about 33 dB needed) and 1024-QAM on a single carrier.

Carrier and sample clock offset

The two boards differ by about 19 to 21 kHz at 2412 MHz (about 8 ppm), and the sample clocks by the same ratio. Within a frame the estimate can be off by a few kHz, so phase tracking must follow a frequency error, not only a phase. apps/link/src/ofdm.c takes the residual offset from the phase step between its two pilot symbols and tracks it with a second order loop (track(), RATE_GAIN). The QAM receiver derives the sample clock offset from the carrier offset.

Looped frames are not phase coherent across copies

Phase noise over one loop period makes copies of a looped frame differ in phase. Decode one contiguous copy (OFDM), or measure the phase step at the wrap from the overlap of two copies (QAM, j->r.wrap in qam.c).

Receiving at 40 MS/s

ESP_SDR_RATE_40MSPS captures go through the radio’s decimation filter, with the same MER as 80 MS/s and half the samples to process. The filter spreads each symbol by about +-8 samples, so an OFDM FFT window that starts only a quarter of the prefix after its beginning picks up the previous symbol. ofdm_rx_begin_period() measures the window position from the pilots’ phase slope across the subcarriers and moves the window to the middle of the prefix (“fine timing”).

The DAC also runs at 40 MS/s (esp_sdr_tx_loop_begin(ESP_SDR_RATE_40MSPS)), at about 4 dB less MER than 80 MS/s.

Capture and DAC engine

The capture and DAC playback use the Wi-Fi MAC’s undocumented dump engine (layout from upstream esp-sdr, see lib/esp_sdr/src/esp_sdr_priv.h).

  • A bank is off limits while the engine owns it. The engine writes or reads one 64 KiB SRAM bank, selected one-hot in MAC_DUMP_USAGE. While it does, the CPUs cannot reach any of that bank, so nothing else may live there. The linker snippets reserve the banks (see Memory); a link error “DRAM image overlaps the … bank(s)” means the image grew into one.

  • Short captures go unreported. The engine can stop early without an error. capture_locked() writes a sentinel into the first and last word and fails the capture if either survives. Checking every word cost 1.2 ms for a full bank.

  • Capture length. At most 16380 samples (ESP_SDR_SAMPLES_MAX, a 14-bit count). A looped frame must fit twice into one window so that a whole copy is always present. That caps the bytes per received window: what matters is the payload bits per captured sample.

  • Circular capture. With DUMP_CTRL 0x24000 the engine writes a 16384-pair ring continuously and exposes its write index at 0x60033d60; switching the bank select while it runs splits the stream into gapless units. lib/esp_sdr/src/esp_sdr_ring.c (from upstream) rotates through banks 0 to 2 and proves continuity with sentinel windows: the switch must come within 2000 pairs (125 us at 16 MS/s) of its threshold, and a bank must be filtered before the engine comes round to it again.

  • The Wi-Fi MAC interrupt must not run inside the ring. The ring masks interrupts on its CPU and opens short windows for the kernel timer and IPIs (irq_window()). CPU interrupt line 0 is the Wi-Fi MAC’s (ETS_WMAC_INUM), and it fires for every packet the Wi-Fi receiver decodes at the tuned frequency, hundreds of times a second on busy air. Its handler is far longer than a window, so the ring abandoned units until the run failed: the receive path then lost whole frames, in bursts that came and went with the neighbours’ traffic and flipped with unrelated code changes. esp_sdr_ring_run() now disables line 0 on its CPU for the run and serves the interrupt afterwards. ring irq windows and lines in wpan status show how many windows ran and which lines were pending (0x40 is the tick timer; 0x41 would be the MAC). With the fix, abandoned units stay at 0 even with the C6 sending back to back.

  • Loop playback wraps only while the trigger is held. Loop playback (esp_sdr_tx_loop_*()) holds the trigger bit; pulsing it plays one pass. Retriggered bursts (esp_sdr_tx_play_for()) leave a hole of about 1.35 us each.

  • Turnaround. The default TX/RX switch with a retune takes about 3 ms. esp_sdr_set_turnaround() with 50 us and no retune works for packet links.

  • Analog I2C must take the PHY’s lock. The vendor PHY serializes its analog I2C traffic with regi2c_enter_critical(). Every rom_chip_i2c_* sequence in the library takes it too, because on two cores the Wi-Fi PHY code may run one at the same time.

Memory

On the S3 the engine’s 64 KiB banks and the ESP_SDR_HIGH_RAM region sit at the top of internal DRAM, laid out by the linker snippets lib/esp_sdr/esp_sdr_bank*.ld (picked in lib/esp_sdr/CMakeLists.txt). Bank 2 (0x3fcd0000) is always the capture bank. CONFIG_ESP_SDR_BANK1, which ESP_SDR_TX_DAC and ESP_SDR_RING select, adds bank 1 (0x3fcc0000) as the second DAC and capture bank, and the ring takes banks 0 to 2 (from 0x3fcb0000). Each snippet lowers _heap_sentry to its lowest bank, so the libc heap stays below it, and fails the link if the DRAM image grows into a bank (see “A bank is off limits” above). ESP_SDR_HIGH_RAM places buffers from the capture bank’s end up to DRAM_USER_END (0x3fce9704).

  • Internal RAM is the scarce resource. Besides the banks, the region above the capture bank (.esp_sdr_high, ESP_SDR_HIGH_RAM) holds apps/link’s QAM transmit context and OFDM receive context, with about 128 bytes left. “high buffers beyond user DRAM” means it is full. Large buffers that are not timing critical go to PSRAM (EXT_RAM_BSS_ATTR).

  • Why past user_dram_end. Zephyr’s user_dram_end (0x3fce4f00) only protects MCUboot’s loader while the image loads; the ROM’s download-mode buffers above it are free at runtime, so NOLOAD buffers can use the whole range (about 38 KiB).

  • Three banks for the ring move the DRAM limit to 0x3fcb0000. apps/osmosdr and apps/wpan fit by trimming the network stack and moving buffers and stacks into the high region (their READMEs). The Wi-Fi driver peaks at about 15.5 KiB of kernel heap with its default receive buffers, 12 KiB with 2 (doc/multiapp.md).

  • Not zeroed at boot. Neither .esp_sdr_high nor PSRAM .bss is cleared. apps/link clears what needs it at init (link_mac_init()).

  • PSRAM is slow for strided or per-sample work. RS units and payload bytes may live there. Hot buffers (FFT work buffers, the RS codeword being decoded, which goes on the stack) should not.

  • Compare-and-set is not atomic in PSRAM. S32C1I on external RAM lets both CPUs succeed: 2 x 200000 atomic_inc() from the two CPUs on one PSRAM counter lost counts. A spinlock, k_msgq, k_mem_slab or atomic_t in PSRAM breaks under SMP. Espressif’s HAL emulates it (esp_cpu_compare_and_set()), and so does the Zephyr branch west.yml pins (“soc: espressif: esp32s3: emulate atomics in PSRAM”): every atomic in the PSRAM range goes through a lock in internal RAM.

  • No flash access from a thread whose stack is PSRAM. A flash read or write (the flash API, so also MCUboot’s boot_is_img_confirmed()) parks the other CPU and suspends the cache, and PSRAM goes with it: the thread’s register window spills read back as 0xbad00add, and it double-faults with the other CPU still parked. Before the watchdog is armed (at boot, before USB starts) the board then hangs for good. A multiapp app’s main() runs on a PSRAM stack, which is why the prelude turns its app_usb_init() into a call that does nothing.

ESP32-C5

Measured on one Waveshare ESP32-C5-Zero (ESP32-C5HF4 rev v1.0, 48 MHz crystal, 4 MB flash, no PSRAM), board waveshare_esp32c5_zero/esp32c5/hpcore (boards/waveshare/esp32c5_zero/: console on USB-Serial-JTAG, 4 MB flash with Espressif’s partitions at the C5’s 0x2000 boot offset, PSRAM disabled, simple boot). The receive items were measured without transmitting; the transmit items with a XIAO ESP32-C6 on the other end of a 30 dB coax attenuator.

  • The capture engine is the C6’s with a rate field. Raw source 15 in 0x600a9008 bits 20:17, rate index in bits 23:21, a 17-bit count, lanes in 0x600a9018 (as librftest’s adctrig()). The six indices 0 to 5 give 80, 40, 20, 10, 8 and 4 MS/s: 16380 samples took 208, 412, 822, 1640, 2050 and 4098 us, and the count field (0x600a9008 bits 15:0) advances by 10 samples per 2.5 us at index 5. ESP_SDR_RATE_20MSPS to _4MSPS are C5 only.

  • The dump gate register stalls the bus. Any access to 0x600a20b4 (the C6’s dump source gate, in the Bluetooth baseband) takes 1.64 ms and reads 0. Captures work without it, so the C5 build skips it; with it, every capture took 1.7 ms longer.

  • The owner bit takes a whole 128 KiB block. SRAM_USAGE (0x60095004 bits 11:8) = 2 with MAC_DUMP_ALLOC (bit 16) puts the samples at 0x40830000, 64 KiB into block 1, but the CPU loses all of block 1 (0x40820000 to 0x4083ffff): reads return the last word it read there, writes are dropped, nothing faults. Blocks 0 and 2 stay with the CPU. Block 2 holds the ROM’s data, so block 1 is the only capture bank.

  • The image is split around the bank. esp_sdr_bank_c5.ld is a RAM_SECTIONS snippet: code and initialized data link below 0x40820000, the bank takes block 1, and every section linked after the snippet (logging and kernel object data, the rest of .noinit, .bss, the libc heap) moves to block 2. The network stack’s and the Wi-Fi driver’s .noinit go below the bank to balance the halves. Margins with the Wi-Fi driver’s defaults: apps/capture leaves 8.3 KiB in block 0 and 21 KiB of libc heap in block 2; apps/rfprobe 4.9 KiB and 20 KiB. “code and data overlap the capture bank” means block 0 is full; “will not fit in region sram0_0_seg” means block 2 is.

  • Tuning. phy_set_chanfreq() takes any MHz above 14 and runs a full channel setup, including 5 GHz. Wi-Fi bursts received with the LO off the channel grid (2433, 2444, 5190, 5200 MHz) centre on 2437, 5180 and 5220 MHz within one 0.31 MHz bin. The PHY takes its 5 GHz path from 4500 MHz up. Below about 1100 MHz the PLL’s SDM word goes negative, so those requests tune nothing. No PLL lock or VCO window flag is known on the C5, so esp_sdr_set_freq() never returns -ERANGE there.

  • Receive DC grows with gain, and the AGC idles high. The largest calibrated gain index is 77 at 2412 MHz and 83 at 5500 MHz. With a forced index up to 60 the DC stays within about 10 of 512. At 65 to 77 it reaches 20 to 50, and at 5500 MHz with index 77 it reaches 78. Under the AGC (ESP_SDR_RX_GAIN_AUTO) it measured +8 to +230 on I at 2412 MHz, varying from boot to boot. Force a gain for measurements.

  • The antenna switch. GPIO26 drives the board’s RF switch (200 kOhm pull-down, so the chip antenna while the pin floats). CONFIG_WAVESHARE_ESP32C5_ZERO_EXT_ANTENNA drives it high at boot (board.c), which selects the U.FL, as CONFIG_XIAO_ESP32C6_EXT_ANTENNA does on the XIAO C6; capture and rfprobe set it for the coax bench. At gain 60 over 2412 to 2472 MHz, low gave ambient Wi-Fi up to -1 dBFS (clipping), high at most -16 dBFS with the U.FL on coax and -30 dBFS with it open. The library does not drive the pin.

  • Gapless receive through the modem diagnostic bus. ESP_SDR_DIAG (esp_sdr_diag.h; the method is ESPsoup’s, github.com/pit711/espsoup-firmware, itself after ESP-SDR’s S31 streaming and Twotoz/C5VRX) streams 40 MS/s without a gap. The pieces, all measured:

    • The dump engine runs free: source 15, rate index 0, count 0x1ffff (the counter is 17 bits and counts above 65535 never reach DONE, so it never stops), started with RUN and a trigger pulse, then the SRAM usage register is written back so that the CPU owns block 1 again. The engine writes nothing into it afterwards, but the diagnostic bus keeps carrying the dump word.

    • MODEM_DIAG[9:0] (signal index 106 + lane) is the Q field of the dump word and [19:10] the I field, as in the capture’s words. Eight lanes (a window of 4 bits of each, lsb to lsb + 3 and 10 + lsb to 10 + lsb + 3) go through the GPIO matrix to eight pads (output enabled by esp_rom_gpio_connect_out_signal(), input enabled) and straight back into PARL_RX_DATA0..7 (signal index 48). No wire; the pads really toggle at up to 40 MHz, so nothing may be connected to GPIO 0, 1, 3, 4, 5, 7, 10 and 23 (ESPsoup’s pads; the board files here use none of them, the C5-Zero’s schematic was not checked).

    • PARLIO RX takes PLL_F240M / 6 = 40 MHz, 8 bits wide, rising edge, LSB first, a soft delimiter with EOF at the ring length: every second sample of the 80 MS/s bus, as a Q nibble (bits 3:0) and an I nibble (bits 7:4) per byte. The lane order was settled against the 10-bit capture: I + jQ gives a tone at +f, as the capture does. No filter before the sampling: narrow the low-pass to 20 MHz or less.

    • A GDMA RX channel (AHB, channel 0, trigger 9, owner check off) runs over a circular ring of 16 descriptors of 3584 bytes (the length field holds 4095 at most) in the lower half of block 1 (57344 bytes at 0x40820000). The hardware writes the received length back into every finished descriptor; the interrupt walks the ring from the last one seen and clears the length, so interrupts that merge lose nothing. The last 8 KiB of that half hold the descriptors and the consumer’s buffers (the osmosdr_serial image has 144 bytes of RAM margin without the mode).

    • Measured: 40 MB/s for 40 s, 0 skipped blocks, +0.15 ppm against the system timer (a byte lost per ring wrap would show as -17 ppm). box16 (a boxcar of 16, SWAR on four words) and a third-order CIC cost 4.1 to 4.4 CPU cycles per byte at 240 MHz (budget 6.0); with the USB service of osmosdr_serial, 5.7.

    • The nibble window needs the level: a signal that exceeds 8 << lsb counts wraps. At RX gain 60 on a hot link the front end clips and every window is useless; from index 48 up the C5 clips a tone that sits 10 dB below full scale at index 40.

    • Only the lane word matters (an ablation with a tone from the C6 present): 0x600a9018 = lanes 24 to 27 with bit 24 set. Without bit 24, or with other lane numbers, the bus carries no I/Q. RUN cleared after the start, source 0 or 3 instead of 15, either clock force (0x600a9c04, 0x600a0800 bit 2) cleared, or 0x600a901c zeroed: the I/Q stays. The selector pair 0x600a9404/0x600a9408 reads 0 on the working path.

    • USB: the decimated samples reach the sink in chunks every 51 us. The USB-Serial-JTAG FIFO takes one 64-byte packet per visit; serving it once per 8 KiB ring block (205 us) capped the stream near 310 kB/s and lost a quarter of the samples at 250 kS/s. osmosdr_serial takes the mode with -DEXTRA_CONF_FILE=diag.conf (CONFIG_APP_RX_DIAG); its 250 kS/s stream ran 60 s with no index jump, and tests/regression/stream_gapless runs both receive paths against a C6 transmitter.

  • Transmit. Two paths, both after the vendor CW turned the chain on and its tone stopped:

    • The TX DC value register (0x600a0c08, esp_sdr_tx_paced_*()), written by the CPU. A store costs 15 CPU cycles, a load 18. The library’s poll costs about 177 cycles per sample (its loads from .bss, code from flash), so it keeps the rate up to 1.35 MS/s; a loop that only stores one prepacked word per sample keeps it exactly up to 8 MS/s (ceiling about 8.8). Both see stalls of 3 to 36 us now and then even with interrupts locked (cause unknown); pacing on absolute targets turns them into jitter, not lost samples.

    • The DAC engine (0x600a900c, esp_sdr_replay_*(), ESP_SDR_REPLAY): plays up to 16383 words (I 9:0, Q 19:10) out of the dump bank, at 80 MS/s, or 40 MS/s with bit 17 (upstream’s layout; the C6 uses bit 15 instead). Bits 14 to 16 are no rate field on the C5: with them the pass time is not proportional to the count, and some codes never set DONE. A repeat count (bits 27:20, with LOOP) plays the passes back to back: 255 passes of 16000 words took 12240047 cycles against 12240000 for the samples. Repeat count 0 with LOOP plays until RUN clears, and its seam is invisible (a C6 capture folded at the loop period shows no step above 1 degree where one lost sample would make 31.5). Restarting a finished pass from the CPU leaves a hole of 35 to 37 samples at 80 MS/s (about 0.45 us: the poll, and about 0.25 us until the engine plays), so periodic content uses the repeat count. The engine owns all of block 1 while it plays: no diagnostic-bus ring meanwhile.

    • The level follows the samples times the DAC scale (0x600a0c04 bits 23:8, one byte per branch; the tone stop leaves full scale, the library sets 15 as upstream). The power step of the vendor tone does nothing on either path: steps 0 to 15 gave the same level. The chain compresses above amplitude times scale of about 1000 to 1500 (third harmonic at -16 dBc with 400 x 15); at 50 x 15 the image is 32 dB down and the third harmonic 46 dB. Once the chain came up 28 dB louder with the same settings, a clean tone, not reproduced since.

    • Half duplex only: while the chain transmits, the receive output freezes at a constant. Turning the receive power and clock back on (phy_pbus_xpd_rx_on(), phy_set_rxclk_en()) revives the receiver and stops the transmitter.

    • Host-fed transmit at 250 kS/s starves on the C5 too (USB OUT delivers about 239 kS/s): 125 kS/s is the reliable rate on both chips. ESP_SDR_RFTEST links the C5 librftest for force_rx_gain().

CPU and toolchain

  • No hardware float divide. A float / calls __divsf3. In hot loops, compare ratios by cross multiplication (sc_scan() in ofdm.c). An integer % by a constant that is not a power of two is a real division too; use a power-of-two size and a mask.

  • lrintf() is a library call, out of line in flash. Two per sample in the OFDM transmit path cost about 7.5 ms per frame. Round inline (emit() in ofdm.c).

  • Many accumulators spill. Loops with four or more running sums get them spilled to the stack (about 100 cycles per step for the sliding Schmidl-Cox sums). Do less work per step instead: the OFDM search scans coarsely first.

  • Contraction is off. Zephyr’s ISO C mode disables fused multiply-adds. apps/link turns them on for its DSP files (-ffp-contract=fast).

  • Hot code belongs in IRAM (IRAM_ATTR): flash code goes through the cache, which the other core and PSRAM traffic contend for.

  • esp-dsp’s S3 FFT kernel assembles empty without the target macro. apps/link builds only dsps_fft2r_fc32_aes3_.S from the esp-dsp west project, with -DCONFIG_IDF_TARGET_ESP32S3=1 (hal_espressif’s sdkconfig.h lacks it), and its own twiddle table in esp-dsp’s layout (cos, sin, bit reversed; one table built for N serves every smaller size). From flash it is still 1.76 times faster than the same algorithm in C from IRAM (85.7 against 150.6 us for 256 points). The S3 mask ROM has the twiddle table, but no FFT code.

  • The ring CPU runs with interrupts masked except for short windows. The ring runs on CPU 1 under arch_irq_lock(): on SMP, irq_lock() takes a global lock and would stall the other CPU. Nothing may be pinned to that CPU meanwhile, no flash writes may happen (code from flash would stall), and the vector registers it uses are not part of any thread context.

  • The ring filter’s budget at 16 MS/s is 15 cycles per pair (240 MHz): unpack 3.0, the fs/4 mix 1.6, stage 1 4.5, stage 2 2.6 (decimation 64), the app’s packing about 1. The multiply-accumulate chain on the single accx accumulator, not the instruction count, limits stage 1.

  • RS(255,223) decoding cost. About 49 us for a clean codeword, 220 us with 2 errors and 400 us with 10. With many corrections, RS decoding dominates a receiver.

Building and flashing

  • Every S3 app except apps/capture needs west build --sysbuild (MCUboot, USB DFU). A plain build of an existing sysbuild directory fails with a cached-source mismatch.

  • Kconfig values passed with -D stay in the CMake cache. An old override (for example -Dlink_CONFIG_APP_SPS=8) persists until set again.

  • Update over USB DFU with west dfu -d <build dir> [-s <usb serial>]. The signed image grows in steps of 64 KiB (flash MMU page alignment), so its size is no evidence of a new build.

  • Host udev rules match USB product IDs: a build with a new PID cannot be opened (or DFU-detached) until the rule covers it. The apps share 2fe3:0005; west dfu finds only 2fe3:0005 (app) and 2fe3:ffff (DFU).

  • A shell port serves one process: a host tool holding it (such as a running link_perf.py run) makes another tool’s open of the same board fail.

  • An S3 image put on by DFU must start USB OTG (app_usb_init() in apps/common, which also arms the trial watchdog) or clear RTC_CNTL_SW_HW_USB_PHY_SEL and RTC_CNTL_SW_USB_PHY_SEL early in boot. Otherwise it stays off USB until a power cycle: the OTG driver sets both bits and a software reset keeps them.

  • esptool reaches an S3 whose app owns OTG once the app’s shell command usj has handed the port to the USB-Serial-JTAG (apps/common); its reset at the end boots the app, which takes the port back. Hence west dfu --esptool writes MCUboot, the image and multiapp’s volume in one esptool run: a sysbuild west flash runs esptool once per image, and the first one’s reset leaves the second without a port.

  • Gate an update on the build’s exit status (a pipe into grep hides it), else west dfu flashes the previous image.

Debugging

  • JTAG on an image that owns OTG. On a running app, usj takes the USB device off the bus and selects the USB-Serial-JTAG: the PC sees the app go and the USB-Serial-JTAG come, and OpenOCD and esptool work until the next boot of the app. On a hung one, a shutdown handler (or a PRE_KERNEL_1 hook in a diagnostic build) that clears the two PHY select bits above hands the PHY over at the next restart; the PC sees no disconnect then, and a USBDEVFS_RESET ioctl on the port’s /dev/bus/usb node makes it enumerate.

  • A reset with no crash record (app crash, link status) is the TG0 watchdog’s first stage: a CPU wedged with interrupts masked.

  • Stacks. With CONFIG_INIT_STACKS, kernel thread stacks after exercising every path; a measured peak up to 80 % of the size is enough margin. An overflowing shell stack shows as 100 % on the logging stack below it. Xtensa frame sizes are in the disassembly as entry a1, N.