OpenOCD and ndmreset signal reset issue

[imi415]

Hi, I'm trying to integrate Ibex with debug module, however I found that under current implementation of the debug module, both reset-to-halt and reset-to-run are not working correctly, which crashes the OpenOCD process.

Below are my attempts to locate the problem, However, I have little knowledge about the code or the specification, if there are anything I got wrong or missed, please point out. Any suggestions are appreciated. Thanks in advance.

---

OpenOCD sets NDMRESET field in dmcontrol during reset period, and expects to see ALLRUNNING field to be set in dmstatus indicating the end of reset operation. These behaviors are addressed below:
https://github.com/riscv/riscv-openocd/blob/4f9e2d7171f3cad8d1a99dff3eee5ec5e6d8ea2b/src/target/riscv/riscv-013.c#L2090

https://github.com/riscv/riscv-openocd/blob/4f9e2d7171f3cad8d1a99dff3eee5ec5e6d8ea2b/src/target/riscv/riscv-013.c#L2144

However, the debug module has its ALLRUNNING field wired to halted signal, which is set by a write operation to HALT address of the debug bus, and cleared by write to RESUME address of the debug bus.

riscv-dbg/src/dm_csrs.sv

Lines 249 to 250 in 3670267

	dmstatus.allrunning = ~halted_aligned[selected_hart] & ~unavailable_aligned[selected_hart];
	dmstatus.anyrunning = ~halted_aligned[selected_hart] & ~unavailable_aligned[selected_hart];

riscv-dbg/src/dm_mem.sv

Lines 242 to 255 in 3670267

	unique case (addr_i[DbgAddressBits-1:0]) inside
	HaltedAddr: begin
	halted_aligned[wdata_hartsel] = 1'b1;
	halted_d_aligned[wdata_hartsel] = 1'b1;
	end
	GoingAddr: begin
	going = 1'b1;
	end
	ResumingAddr: begin
	// clear the halted flag as the hart resumed execution
	halted_d_aligned[wdata_hartsel] = 1'b0;
	// set the resuming flag which needs to be cleared by the debugger
	resuming_d_aligned[wdata_hartsel] = 1'b1;
	end

When the core is already in parked-loop state and a reset-to-run command is issued by OpenOCD, the core may never has a chance to write to RESUME address, since it has been resetted by ndmreset signal. So the ALLRUNNING field will never be set and OpenOCD just get stucked in the loop until timeout reached.

This internal state (core is actually running while dmstatus has wrong value) also causes crash of OpenOCD, then a reset to dm will be needed for the dm to function properly.

[bluewww]

I guess they need to be put into the unavailable or running state after a ndmreset

[gxflying]

yeah， this is because when ndmreset is set, the current version DM doesn't clear the "halted flags".

and one more thing, the dmcontrol.dmactive may also not works well because it should do something with the registers in dm_mem.sv but it doesn't.
for example:
spec says dmactive can clear the abstractcs.busy status. but the current version does nothing with abstractcs.busy

[bluewww]

Can you open a separate issue for this, please?