Add %jinx hint to Vere.

[sigilante]

Resolves proposed UIP (unnumbered).

Adds a %jinx hint to Vere which permits a computation's timeout to be specified in Urbit time.

> ~>  %jinx.[~s4]  =|(i=@ |-(?:(=(10.000.000 i) i $(i +(i)))))
10.000.000

> ~>  %jinx.[~s3|  =|(i=@ |-(?:(=(10.000.000 i) i $(i +(i)))))
recover: dig: alrm
crud: %belt event failed
call: failed

[sigilante]

May want to improve the error message to provide guidance on why the crash takes place.

[joemfb]

You can replace this math with u3_time_msc_out(c3_d).

c3_d tim_d;
if ( c3y == u3r_safe_chub(*clu, &tim_d) ) {
  c3_w mic_w = u3_time_msc_out(tim_d);
  c3_w mil_w = mic_w / 1000;
  u3m_timer_set(mil_w);
}

(Or you could, if it was defined in pkg/vere. I'd be inclined to copy it into this file as a static function rather than use gmp.)

[sigilante]

Urbit time is bigger than that, though, so the u3r_safe_chub always fails.

[sigilante]

It could retrieve two chubs for the time, very unlikely to be too big for that.

[sigilante]

notes from discussion today, for posterity:

• mil_w was a functioning instance of this mechanism in u3m_soft for certain executions.
• u3m_soft → u3m_soft_top which processes trace for you
• signals send you to a top-level call
• we probably have to do different things depending on whether we're in a virtual computation or not.
• if this only ever happens in a road, and we only run u3m_soft from inside, then we can assume jump buffer is initialized and all we need to do is reset itimer_virtual in cm_signal_deep()
• in cm_signal_deep() if mil_ != 0
• so reset SIGVTALRM if we hit a new %jinx hint
• fn in manage.c that nock.c can call when it encounters this hint that does the same thing that cm_signal_deep() does with the interval timer, but it also needs to check somehow
• need some other global state to be able to see if we are prepared to handle this signal or not
• vere/serf.c → ll. 26, 28 mil=@
  • inner timer if computation is done
  • outer timer to see if still valid
• could build proof of concept that doesn't worry about safety and nesting
• if you call setitimer multiple times you clobber the previous value, so to nest you need global state (timer value and monotonic timestamp), then when comp finishes then restore parent with adjusted time
• proof of concept: explore hint callbacks in nock.c, whitelist in bytecode compiler, then check dynamic hint processing

    case c3__jinx: {
      if (c3y == u3a_is_cat(*clu)) {
        c3_w mil_w = *clu;
        // call fn in u3m passing it mil_w TODO
      }
      u3z(*clu);
      *clu = u3_nul;
    } break;

one to set ITIMER like cm_signal_deep, one to cancel it using zeroed itervaltimerstruct

[tacryt-socryp]

Extremely based functionality

[sigilante]

From discussion ~2024.8.12, some thoughts on the interface:

• subsecond syntax for @dr, like ~us1000? ~ms, ~us, ~ns (mibiseconds as powers of two?)
• (ms 1000) support function?
• use bloq size instead?

So write a UIP to implement subsecond timing syntax?

kiki-second, miki-second, niki-second, piki-second

[sigilante]

Probably just stick with @dr and add ++msec, ++usec arms for now.

[sigilante]

Ongoing support work for Urbit @dr time here: urbit/urbit#7057

[sigilante]

Also urbit/urbit#7067 for the parser.

[sigilante]

Should the nested timer be measured against the nominal outer timer or against the actual time remaining? I'm building the stack handler for the latter case, but worth thinking about.

[sigilante]

::  should fail (timeout)
~>  %jinx.[~s0..8000]  =|(i=@ |-(?:(=(10.000.000 i) i $(i +(i)))))

::  should succeed (normal)
~>  %jinx.[~s1]  =|(i=@ |-(?:(=(10.000.000 i) i $(i +(i)))))

::  should succeed (nested properly)
~>  %jinx.[~s2..8000]  ~>  %jinx.[~s1]  =|(i=@ |-(?:(=(10.000.000 i) i $(i +(i)))))

::  should fail (nested improperly)
~>  %jinx.[~s1..8000]  ~>  %jinx.[~s5]  =|(i=@ |-(?:(=(10.000.000 i) i $(i +(i)))))

::  should fail (nested improperly and timeout)
~>  %jinx.[~s0..8000]  ~>  %jinx.[~s1]  =|(i=@ |-(?:(=(10.000.000 i) i $(i +(i)))))

%jinx squats on the SIGVTALRM signal, and when this is caught in manage.c it always assumes that the cause is an expired timer. While this is correct for the current usage of Vere, I highlight it in case some future runtime engineer decides to utilize that signal in some other way.

[sigilante]

At this point, it would be helpful to have someone else verify the behavior. There seems to be an incorrect fall-through case on ~s0 still, and I suspect that the timer stack may not always clear correctly.

Test Cases

Since the failure of a %jinx hint results in a nondeterministic error, we cannot test them using in-Arvo means. However, the following should all work correctly:

::  should fail (timeout)
~>  %jinx.[~s1]  =|(i=@ |-(?:(=(10.000.000 i) i $(i +(i)))))

::  should succeed (normal)
~>  %jinx.[~s1]  =|(i=@ |-(?:(=(10.000 i) i $(i +(i)))))

::  should succeed (nested properly)
~>  %jinx.[~s3..8000]  ~>  %jinx.[~s3]  =|(i=@ |-(?:(=(10.000.000 i) i $(i +(i)))))

::  should warn (nested improperly; will succeed on outer hint)
~>  %jinx.[~s1..8000]  ~>  %jinx.[~s5]  =|(i=@ |-(?:(=(10.000.000 i) i $(i +(i)))))

::  should warn (nested improperly; will timeout on outer hint)
~>  %jinx.[~s0..8000]  ~>  %jinx.[~s1]  =|(i=@ |-(?:(=(10.000.000 i) i $(i +(i)))))

::  should fail (timeout)
~>  %jinx.[~s0]  =|(i=@ |-(?:(=(10.000.000 i) i $(i +(i)))))

Developer Notes

1. %jinx squats on the SIGVTALRM signal, and when this is caught in manage.c it always assumes that the cause is an expired timer. While this is correct for the current usage of Vere, I highlight it in case some future runtime engineer decides to utilize that signal in some other way.

2. %jinx simply returns on an invalid nesting, rather than clearing the timer stack.

3. The fast way to build this:

   bazel build :urbit --copt=-DU3_MEM_DEBUG
   ./bazel-bin/pkg/vere/urbit zod

[sigilante]

@pkova @mopfel-winrux wants me to bug you on this for 3.2

[joemfb]

Superseded by #910