Repository navigation
Description
Summary
Changing $100/$101/$102 (steps/mm) while the machine is at a non-zero position leaves gc_state.position (the parser/planner position) inconsistent with sys.position and the reported MPos. The next move is planned from the corrupted value and travels the wrong distance — the error grows with distance from machine zero.
sync_position() is called before the new steps_per_mm is installed, so it converts the freshly rescaled step count using the old calibration.
Affected code
settings.c, set_axis_setting() — current master:
case Setting_AxisStepsPerMM:
if((status = validate_pulse_width(settings.axis[idx].max_rate, value, settings.steppers.pulse_microseconds)) == Status_OK) {
if(settings.axis[idx].steps_per_mm > 0.0f && settings.axis[idx].steps_per_mm != value) {
float comp = value / settings.axis[idx].steps_per_mm;
sys.position[idx] *= comp;
sys.home_position[idx] *= comp;
sys.probe_position[idx] *= comp;
sys.tlo_reference[idx] *= comp;
sync_position(); // <-- still the OLD steps_per_mm
}
settings.axis[idx].steps_per_mm = value; // <-- installed only here
}
break;sync_position() (gcode.h) expands to plan_sync_position(); system_convert_array_steps_to_mpos(gc_state.position, sys.position), and system_convert_array_steps_to_mpos() (system.c) does:
position[idx] = steps[idx] / settings.axis[idx].steps_per_mm;plan_sync_position() only memcpys the step array, so it is unaffected by the ordering — the problem is entirely in the mm conversion.
Why the result is wrong
Let S₀ = old steps/mm, Sₙ = new steps/mm, and P = the true step position.
Scaling sys.position by comp = Sₙ/S₀ is clearly deliberate: it keeps MPos unchanged across a recalibration, since MPos = (P·Sₙ/S₀)/Sₙ = P/S₀. That part works.
But sync_position() runs one line too early, so it computes
gc_state.position = (P·Sₙ/S₀)/S₀ = MPos_before · (Sₙ/S₀)
instead of MPos_before. sys.position and the reported MPos are correct; the parser position is off by a factor of Sₙ/S₀.
A subsequent G91 move of d mm therefore changes MPos by
MPos_before · (1 − Sₙ/S₀) + d
instead of d. At machine zero the term vanishes, which is why this is easy to miss.
Reproduction
No special hardware needed — it is visible purely in grblHAL's own reported position. On a 2000 steps/mm axis:
(move to Y = 8.000)
? -> <Idle|MPos:0.000,8.000,...>
$101=1000 -> ok
? -> <Idle|MPos:0.000,8.000,...> (MPos correctly unchanged)
$J=G91 G21 Y-3.000 F1000
? -> <Idle|MPos:0.000,1.000,...> (moved 7 mm, not 3 mm)
Predicted by the formula: 8·(1 − 1000/2000) + (−3) = −7.
Measured on my bench (SAME5x driver, ClearPath servos): 7000 steps were emitted, which at the motor's true 2000 steps/mm is 3.5 mm of physical rotation for a 3 mm command. In a longer randomized test the same defect produced 25,227 steps ≈ 12.6 mm of rotation for a nominal 5.903 mm jog from MPos.Y = 38.648.
On a machine with real travel and no soft limits armed, a large enough MPos_before can drive an axis into a hard stop.
Suggested fix
Install the new value before converting, keeping the existing behaviour for the first-set / unchanged-value path:
case Setting_AxisStepsPerMM:
if((status = validate_pulse_width(settings.axis[idx].max_rate, value, settings.steppers.pulse_microseconds)) == Status_OK) {
if(settings.axis[idx].steps_per_mm > 0.0f && settings.axis[idx].steps_per_mm != value) {
float comp = value / settings.axis[idx].steps_per_mm;
sys.position[idx] *= comp;
sys.home_position[idx] *= comp;
sys.probe_position[idx] *= comp;
sys.tlo_reference[idx] *= comp;
+ settings.axis[idx].steps_per_mm = value; // sync_position() reads this
sync_position();
- }
- settings.axis[idx].steps_per_mm = value;
+ } else
+ settings.axis[idx].steps_per_mm = value;
}
break;I'm happy to open a PR if the approach looks right.
While here
sync_position() in gcode.h is an unbraced two-statement macro:
#define sync_position() plan_sync_position(); system_convert_array_steps_to_mpos (gc_state.position, sys.position)so if(cond) sync_position(); guards only the first statement. All current call sites appear to be plain statements, so nothing is broken today, but wrapping it in do { ... } while(0) would remove the trap.
Environment
- grblHAL core: current master (verified against the
settings.c,gcode.h,system.con master at time of filing); also reproduces on the pinned306bf68d, build 20260726. - Driver: SAME5x / Teknic ClearCore (out-of-tree), 4 axes, no kinematics (
KINEMATICS_APInot defined). $20=0, unhomed, no limit switches fitted.
The defect is in core settings.c and independent of the driver — any driver without KINEMATICS_API should show it.
Activity