Without HTIF: 5000 instructions executed, then spike hangs

[neuschaefer]

Hi, I encountered this problem while debugging coreboot, which doesn't use the HTIF.

If a program doesn't provide the tohost ELF symbol, spike will execute 5000 instructions and then hang in a loop without executing any more instructions. I tested this with spike f5ecf65 and riscvarchive/riscv-fesvr@916191caf3.

One easy way to proproduce this bug is to:

• Write a program that spins a loop
• run spike -l pk ./program to verify that execution doesn't stop
• copy pk into the current directory
• find the tohost string in pk and patch it to something else (like nohost)
• run spike -l ./pk ./program |& tee trace-log; observe that execution stops at some point
• run wc -l trace-log and observe that 5000 instructions have been executed

I haven't fully debugged it, but I've seen the following happening:

• sim_t::run() calls step(INTERLEAVE), INTERLEAVE is defined as 5000
• sim_t::step(n) simulates n instrucions, does some bookkeeping and finally calls htif->tick()
• htif_isasim_t::tick() calls tick_once()
• tick_once() calls htif_pthread_t::recv which calls recv_nonblocking
• recv_nonblocking doesn't find any data, so it calls host.switch_to()
• the other thread (this is some kind of cooperative multitasking, right?) doesn't switch back, because it's stuck in the loop in htif_t::run()

It might not be related to this problem, but I noticed that both spike and libfesvr register their own handlers for SIGINT. As signal handlers don't automatically stack, this is probably a bad idea.

[aswaterman]

We'll deal with most of this when we get rid of the HTIF, very soon.

The SIGINT handler thing is probably benign, but is nevertheless a design flaw.