Linux on ESP32

Linux port with Sv32 virtual memory, Supervisor mode, XIP, and a Buildroot userspace, running natively on an ESP32-S31 microcontroller.

Module tested: ESP32-S31-WROOM-3 E1H16R16V (ESP32-S31 Core Board/Korvo).

Experimental hardware bring-up project. Definitely not something you want for production.

Install esptool, Espressif's tool for flashing ESP32s:

Then download the binaries in Release, connect your board through USB-UART, and flash the board per the provided command below (change /dev/ttyUSB0 to your actual serial device):

$ esptool -p /dev/ttyUSB0 -b 2000000 erase-flash
$ esptool -p /dev/ttyUSB0 -b 2000000 write-flash \
    --flash-mode dio --flash-freq 80m --flash-size 16MB \
    0x2000 bootloader.bin \
    0x8000 partition-table.bin \
    0x17000 ota_data_initial.bin \
    0x20000 hello_world.bin \
    0x220000 fw_payload.bin \
    0x2A0000 xipImage \
    0xA20000 rootfs.sqfs

🟢 Stable — Fully tested and working | 🟡 Untested — Seems working; not throughly tested | 🟠 WIP - Functions not fully implemented

Refer to the Build Instructions.

(For more hardware references, see docs/ folder)

This port was done before S31 TRM is available, therefore these guessworks were made:

S31 uses CLIC and CLINT similar to P4. Linux expects PLIC. Therefore a custom CLIC driver is needed. I referenced this CLIC patch from disdi to get the CLIC working.

Also, standard RISC-V interrupt CSRs are not usable, presumably because, from P4's TRM, CLINT interrupts are routed to CLIC and mtvec.MODE is hardwired to 0x3 (CLIC mode). Patches needed to make OpenSBI interrupts work.

S31's supervisor mode is not standard and has absolutely no usage in ESP-IDF so a lot of these CSR uses were mostly guessed from either P4's TRM or CSR probing (see docs/). For example, the use of sclicbase(?) and the lack of sie.

S31 implemented SCLIC (Supervisor CLIC?) which is confusing since there is no known standardization; According to all laws of esp-idf, mcliccfg.NMBITS is not writable. IT IS WRITABLE! And setting it to 0b01 enables writes to the clicintattr[i].MODE field and thus enabling the use of S-mode interrupts.

To save the precious 16MB PSRAM memory, OpenSBI was modified to use XIP in flash and internal SRAM (hence the 3915901 KB firmware size in OpenSBI banner, since flash and SRAM mappings are not continuous).

In mainline linux, XIP support on RISC-V was removed, so 6.12 was used instead which has proper XIP support.

For several reasons:

• Espressif's radio firmware blobs are closed source, and must run within ESP-IDF's FreeRTOS framework. It's near impossible to reverse-engineer them (not to mention legal risks.)
• S31's two cores are kinda heterogeneous already: SIMD path only on hart 1. SMP makes scheduling things on the right core harder.
• PSRAM is already slow enough (compared to SRAM); two cores would share the same, tiny 32KiB D-cache.
• Cache maintenance, IPC, Interrupt routing, etc.
• I like having an RTOS for other tasks. If you want absolute performance, a low-end MPU (like Allwinner T113-S3) would be a far better choice

For this port, think S31 as a reincarnated Bouffallo BL808.¹