…
|
||
---|---|---|
.. | ||
README.txt | ||
ocdconsole.py |
README.txt
Console over Lightweight Link ============================= LWL is a Lightweight bidirectional communication between target and debug host without any need for additional hardware. It works with openOCD and other debuggers that are capable of reading and writing memory while the target is running...it should run with JLink for example, if you've got the SDK and modify this file accordingly. Principle of operation is simple; An 'upword' of 32 bits communicates from the target to the host, a 'downword' of the same size runs in the opposite direction. These two words can be in any memory that is read/write access for both the target and the debug host. A simple ping pong handshake protocol over these words allows up/down link communication. On the upside no additional integration is needed. On the downside it may be necessary to feed lwl with cycles to poll for changes in the downword, depending on the use case. For the case of a simple console, that's not needed. For convenience these communication locations are automatically discovered from the RAM by searching through it. Just define downwordaddr and upwordaddr if you want to work with fixed locations. Bit configuration ----------------- Downword (Host to target); A D U VV XXX O2 O1 O0 A 31 1 - Service Active (Set by host) D 30 1 - Downsense (Toggled when there is data) U 29 1 - Upsense ack (Toggled to acknowledge receipt of uplink data) VV 28-27 2 - Valid Octets (Number of octets valid in the message) XXX 26-24 3 - Port in use (Type of the message) O2 23-16 8 - Octet 2 O1 15-08 8 - Octet 1 O0 07-00 8 - Octet 0 Upword (Target to Host); A 31 1 - Service Active (Set by device) D 30 1 - Downsense ack (Toggled to acknowledge receipt of downlink data) U 29 1 - Upsense (Toggled when there is data) VV 28-27 2 - Valid upword octets XXX 26-24 3 - Port in use (Type of the message) O2 23-16 8 - Octet 2 O1 15-08 8 - Octet 1 O0 07-00 8 - Octet 0 Port 1 is used for Console. No other ports are currently defined. Usage ===== No special python modules are needed, it should be possible to run the application simply as shown below: In the first terminal execute the openocd command to connect to the board. Assuming that you already flashed to firmware (nuttx.bin) with the LWL console support. For stm32f4discovery board I use this command: ------------------------------------------ $ sudo openocd -f board/stm32f4discovery.cfg Open On-Chip Debugger v0.10.0-esp32-20200526-6-g4c41a632 (2020-06-23-10:12) Licensed under GNU GPL v2 For bug reports, read http://openocd.org/doc/doxygen/bugs.html Info : The selected transport took over low-level target control. The results might differ compared to plain JTAG/SWD srst_only separate srst_nogate srst_open_drain connect_deassert_srst Info : Listening on port 6666 for tcl connections Info : Listening on port 4444 for telnet connections Info : clock speed 2000 kHz Info : STLINK V2J17S0 (API v2) VID:PID 0483:3748 Info : Target voltage: 3.216252 Info : stm32f4x.cpu: hardware has 6 breakpoints, 4 watchpoints Info : Listening on port 3333 for gdb connections Info : accepting 'tcl' connection on tcp/6666 invalid command name "ocd_mdw" 0x20000000: 000000ff 0x20000000: 000000ff 0x20000004: 7216a318 0x2000000c: 994b5b1b 0x2000000c: 994b5b1b 0x2000000c: 994b5b1b ... The "0x2000000c:..." will repeat all the time. Now in another terminal execute: ------------------------------------------ $ ./ocdconsole.py ==Link Activated nsh> nsh> help help usage: help [-v] [<cmd>] ? echo exit hexdump ls mh sleep xd cat exec help kill mb mw usleep nsh> ------------------------------------------ This code is designed to be 'hardy' and will survive a shutdown and restart of the openocd process. When your target application changes then the location of the upword and downword may change, so they are re-searched for again. To speed up the start process consider putting those words at fixed locations (e.g. via the linker file) and referencing them directly. Future work/Improvements ======================== Currently the lwl driver on NuttX side is doing polling, but for better performance it could use interrupts to detect when the memory position was modified to read the data. It also will avoid using busy waiting inside the driver, look at nuttx/arch/arm/src/common/arm_lwl_console.c for more information.