makefiles: Add openocd-rtt

[chrysn]

Contribution description

The RTT mechanism, by which a debugger emulates a serial interface, has been invented and described by Segger (creator of JLink), but was so far only available through their own JLink program. As the same interface is also implemented by OpenOCD, it can be used with that and thus any debug adapter (eg. STLink).

Open issues

• Do we need any parametrization?
  • Channel 0 is hard-coded which is probably OK
  • No clue where the addresses come from.
  • The string "SEGGER JLINK" is the sentinel value also hardcoded in stdio_rtt.c, so that should be OK.
• Do we need better autoselection? (See future steps)

Testing procedure

• Pick any board that has an OpenOCD-style debugger and usually runs serial stdio; I picked microbit-v2, and an example that has nice shell (eg. saul)
• Add lines to the Makefile:

  USEMODULE += stdio_rtt
  RIOT_TERMINAL = openocd-rtt
  

• make -C examples/saul BOARD=microbit-v2 all flash term

Issues/PRs references

This is based on the existing code in jlink.sh, and notes from @nica-f.

Future steps

• There's a lot of code duplication both internally to openocd.sh and between it and jlink.sh.
• The process of picking a terminal happens inside these scripts, so they have their own JLINK_TERMPROG rather than having composition from the terminal selection.

Both of these are probably doable in shell and Makefile, but IMO too complex to be done there well and maintainably -- so some duplication it is for now, as it has been for some time.

• The default RIOT_TERMINAL is pyterm, and only overridden by board or application. It might make sense to set defaults based on the selected stdio_ modules. As this would also affect stdio_acm (in which case it'd influence MOST_RECENT_BOARD selectors as in makefiles/tools/serial.inc.mk: Improve PORT selection when RIOT's USB CDC ACM is used for stdio #18525), I think that's better done in a follow-up step.

[benpicco]

Nice, works well even with a particle debugger and just SWDIO/CLK!
(After bumping STDIO_TX_BUFSIZE, otherwise large output gets cut off.

I agree that

--- a/makefiles/tools/serial.inc.mk
+++ b/makefiles/tools/serial.inc.mk
@@ -19,6 +19,10 @@ PROG_DEV ?= $(PORT)
 
 export BAUD ?= 115200
 
+ifneq (,$(filter stdio_rtt,$(USEMODULE)))
+  RIOT_TERMINAL ?= openocd-rtt
+endif
+
 RIOT_TERMINAL ?= pyterm
 ifeq ($(RIOT_TERMINAL),pyterm)
   TERMPROG  ?= $(RIOTTOOLS)/pyterm/pyterm

would be more ergonomic - I had no idea about openocd-rtt.

[chrysn]

Once we go for the more ergonomic default sizes, I'd also introspect the PROGRAMMER -- if it's jlink, RIOT_TERMINAL should become jlink-rtt instead. (I'd start from the diff you showed in a next iteration after this is through).

In the meantime I've found out the numbers: They're a start and length of memory to search for the magic string (provided in the argument) in. I'm a bit conflicted about how to best fill them in: My default mode is pedantic, for which I'd pick the rtt_cb symbol out of the ELF file. But that creates dependencies we don't currently have, and easily breaks if the flashed binary doesn't exactly match what was last built. So let's do as JLink does (apparently -- it isn't given the ELF file or data from it) and use autodiscovery, which AFAICT relies on the string "SEGGER RTT" being present only at one location in memory. As JLink presumably knows something about the RAM setup that OpenOCD doesn't (or doesn't evaluate automatically), picking RAM_START_ADDR and RAM_LEN from the build system (these shouldn't drastically change between builds). This is reflected in the second commit added now, answering the open questions I've had.

[maribu]

I love this one. Other than the ergonomic defaults, this looks ready for merge, right?

[chrysn]

I think yes -- I didn't merge it because the second commit (which addresses the hardcoded numbers that were there before) wasn't part of the original ACK, and touches a few more places.

If you're happy with that, please bring it ahead :-)

[maribu]

ACK. Looks good to me and has already been tested. IMO you could squash the two commits. If you don't feel like squashing, you may hit merge instead.

[chrysn]

Squashed without code changes after a successful run, setting the "skip build" flag.