cmd/geth, console: support interrupting the js console

[holiman]

This PR changes the behavior of Ctrl-C (i.e. SIGINT) when the console is active.
Previously, this signal was ignored during JS execution, so it was not possible
to get out of infinite loops in the console. With this change, Ctrl-C now
interrupts JS.

Fixes #23344

[holiman]

Not sure what happens there...

Regarding that...

Local console, eternal loop, hit ctrl-c:

• With this PR, geth (almost fully) shuts down and the operation is aborted. User is put back into the console (with a dead blockchain) and can ctrl-d to exit.
• Master: geth shuts down, is prevented from doing so, and eventually needs to be panic:ed
• Desired: operation is aborted, geth does not shut down

So, even for that case, this PR is better than master.

What actually happens, though, is that since we're not in console reading mode any more, the ctrl-c handling is done in the geth normal handler, so that's what triggers the shutdown.

[holiman]

With this PR, geth (almost fully) shuts down and the operation is aborted. User is put back into the console (with a dead blockchain) and can ctrl-d to exit.

This has now been improved.

Local console, eternal loop, hit ctrl-c:

• With this PR, geth shuts down, the operation is aborted and geth exits gracefully.
• Master: geth shuts down, is prevented from doing so, and eventually needs to be panic:ed
• Desired behaviour: operation is aborted, geth does not shut down

To exit, press ctrl-d
> while(true){}
^C> INFO [08-26|13:53:51.952] Got interrupt, shutting down... 
aborted by caller at <eval>:1:1(3)
INFO [08-26|13:53:51.953] IPC endpoint closed                      url=/tmp/geth.ipc
INFO [08-26|13:53:51.953] Ethereum protocol stopped 
INFO [08-26|13:53:51.953] Transaction pool stopped 
INFO [08-26|13:53:51.953] Writing snapshot state to disk           root=2c6528..bde715
INFO [08-26|13:53:51.953] Persisted trie from memory database      nodes=0  size=0.00B   time="5.323µs" gcnodes=0 gcsize=0.00B gctime=0s livenodes=1 livesize=0.00B
INFO [08-26|13:53:51.953] Blockchain stopped 
exiting console

[holiman]

ping @gballet PTAL?

[fjl]

I have some ideas for this, please don't merge it yet.

[s1na]

@fjl was curious what your idea was regarding this PR

[fjl]

I tried to implement the proper semantics where geth doesn't shut down on Ctrl-C. It's a bit complicated though.
My idea was refactoring the geth interrupt handler so that it wouldn't be active in console mode.

[s1na]

I was getting a crash immediately during start. I couldn't figure out where it came from but after a rebase against master the crash is gone

[s1na]

My idea was refactoring the geth interrupt handler so that it wouldn't be active in console mode.

I tried your idea here. It works. I thought sigterm should still cause shutdown. Do you agree with that? if so I need to change the code a bit because right now if first a sigint arrives and we're in console mode any future sigterm is ignored.

Update: fixed sigterm handling here. So now ctrl+c only stops while(true) but kill <pid> stops while(true) and shuts geth down gracefully.

[s1na]

@holiman can you please take a look at my changes whenever you get a chance and tell me if its okay for me to push them to your branch?

[holiman]

I find it a bit hard to follow, can't compare across forks, since you rebased mine before you committed yours. Do you think you can push the rebased to my repo for easier comparison?

Edit: or squash your two commits into one?

[s1na]

Yes squashed: s1na@0e99d3f

[holiman]

LGTM, push it here why dontcha?

[holiman]

Yup, with your fixes, this now works for me:

[user@work go-ethereum]$ ./build/bin/geth --dev console
...
Welcome to the Geth JavaScript console!

instance: Geth/v1.10.14-unstable-46891a6b-20211203/linux-amd64/go1.17.3
coinbase: 0xc9493a99d64c05dbfcf6635a52aa7a990f0250bc
at block: 0 (Thu Jan 01 1970 01:00:00 GMT+0100 (CET))
 datadir: 
 modules: admin:1.0 clique:1.0 debug:1.0 eth:1.0 miner:1.0 net:1.0 personal:1.0 rpc:1.0 txpool:1.0 web3:1.0

To exit, press ctrl-d or type exit
> while(true){}
^C> aborted by caller at <eval>:1:1(1)

[fjl]

I will re-check and merge later today!

[fjl]

I've tested this extensively and decided to rework the interrupt logic a bit more to
really handle all the cases. In some of the cases listed below, the JS interrupt did not
work or caused an unclean shutdown of the node.

./geth console --dev

> partialInput^
# when SIGTERM received, geth should exit
# when SIGINT received, it should not exit
# when Ctrl-C, it should clear the line

./geth console --dev

> while(true) { admin.sleep(1); console.log("sleeping"); }
sleeping
sleeping
^
# when SIGTERM received, geth should exit
# when SIGINT received or Ctrl-C, it should return to JS prompt

TERM=dumb ./geth console --dev

> partialInput^
# special case: when Ctrl-C/SIGINT, geth should exit

TERM=dumb ./geth console --dev

> while(true) { admin.sleep(1); console.log("sleeping"); }
sleeping
sleeping
^
# when SIGTERM received, geth should exit
# when SIGINT received or Ctrl-C, it should return to JS prompt

./geth console --dev --exec 'while(true) { admin.sleep(1); console.log("sleeping"); }'
sleeping
sleeping
^
# SIGTERM/SIGINT/Ctrl-C should interrupt JS and shut down the node

echo 'while(true) { admin.sleep(1); console.log("sleeping"); }' > loop1.js
./geth --dev js loop1.js
sleeping
sleeping
^
# SIGTERM/SIGINT/Ctrl-C should interrupt JS and shut down the node

[holiman]

LGTM. While testing this, I had some issues with the history, things didn't seem to be added to it. Then I cleared out the history (there were 1923 lines in it), and tried to repro it, but suddenly things started to work after that, so not sure what it was.

[fjl]

I need to recheck my change to the history write. Maybe it doesn't work properly now.

[fjl]

I checked the history behavior and it seems to work.