7 KiB
Quickstart (docker)
make dockerbuild
make dockershell
Then, from inside the docker shell
make
./hosts/host/host apps/helloworld/helloworld.bin
host is the command line test VM for running samples. Run host -h for a full list of options.
Native build
The example VM hosts should all build with any C compiler. To build all of the examples in apps, you will need a RISC-V cross compiler, Zig 0.15.2 and Rust (stable). To build the example host-arduino you will need arduino-cli.
On mac
brew install arduino-cli riscv64-elf-gcc riscv64-elf-binutils sdl3
cd hosts/host-sdl
make
cd apps/zigdoom
make test
Quickstart API
uint8_t bytecode[] = { /* ... */ }; // some compiled bytecode
uvm32_state_t vmst; // execution state of the vm
uvm32_evt_t evt; // events passed from vm to host
uvm32_init(&vmst); // setup vm
uvm32_load(&vmst, bytecode, sizeof(bytecode)); // load the bytecode
uvm32_run(&vmst, &evt, 100); // run up to 100 instructions
switch(evt.typ) {
// check why the vm stopped executing
}
Operation
Once loaded with bytecode, uvm32's state is advanced by calling uvm32_run().
uint32_t uvm32_run(uvm32_state_t *vmst, uvm32_evt_t *evt, uint32_t instr_meter)
uvm32_run() will execute until the bytecode requests some IO activity from the host.
These IO activities are called "syscalls" and are the only way for bytecode to communicate with the host.
If the bytecode attempts to execute more instructions than the the passed value of instr_meter it is assumed to have crashed and an error is reported.
(As with a watchdog on an embedded system, the yield() bytecode function tells the host that the code requires more time to complete and has not hung)
uvm32_run() always returns an event. There are four possible events:
UVM32_EVT_ENDthe program has endedUVM32_EVT_ERRthe program has encountered an errorUVM32_EVT_SYSCALLthe program requests some IO via the host
Internals
uvm32 emulates a RISC-V 32bit CPU using mini-rv32ima. All IO from vm bytecode to the host is performed using ecall syscalls. Each syscall provided by the host requires a unique syscall value. A syscall passes two values and receives one on return.
uvm32 is always in one of 4 states, paused, running, ended or error.
stateDiagram
[*] --> UVM32_STATUS_PAUSED : uvm32_init()
UVM32_STATUS_PAUSED-->UVM32_STATUS_RUNNING : uvm32_run()
UVM32_STATUS_RUNNING --> UVM32_STATUS_PAUSED : syscall event
UVM32_STATUS_RUNNING --> UVM32_STATUS_ENDED : halt()
UVM32_STATUS_RUNNING --> UVM32_STATUS_ERROR
Boot
At boot, the whole memory is zeroed. The user program is placed at the start. The stack pointer is set to the end of memory and grows downwards. No heap region is setup and all code is in RAM.
ExtRAM
A single block of external RAM may be memory mapped into the VM at any time using:
uvm32_extram(uvm32_state_t *vmst, uint32_t *ram, uint32_t len)
The ram region must be 32bit aligned, as all accesses will be 32bit words. The len is given in bytes.
From inside the VM, the memory is available from address 0x10000000
uint32_t *p = (uint32_t *)UVM32_EXTRAM_BASE;
p[0] = 0xDEADBEEF;
When the external RAM is written to by the VM, the dirty flag will be set. The flag is automatically cleared on the next call to uvm32_run(). The flag can be checked using:
bool uvm32_extramDirty(uvm32_state_t *vmst)
syscall ABI
All communication between bytecode and the vm host is performed via syscalls.
To make a syscall, register a7 is set with the syscall number (a UVM32_SYSCALL_x) and a0, a1 are set with the syscall parameters. The response is returned in a2.
static uint32_t syscall(uint32_t id, uint32_t param1, uint32_t param2) {
register uint32_t a0 asm("a0") = (uint32_t)(param1);
register uint32_t a1 asm("a1") = (uint32_t)(param2);
register uint32_t a2 asm("a2");
register uint32_t a7 asm("a7") = (uint32_t)(id);
asm volatile (
"ecall"
: "=r"(a2) // output
: "r"(a7), "r"(a0), "r"(a1) // input
: "memory"
);
return a2;
}
The RISC-V SBI is not followed, a simpler approach is taken.
syscalls
There are two inbuilt syscalls used by uvm32, halt() and yield().
halt() tells the host that the program has ended normally. yield() tells the host that the program requires more instructions to be executed. Halt is handled internally and transitions the VM to UVM32_STATUS_ENDED, yield() is handled in the VM host like other syscalls.
Syscalls are handled in the host by reading the syscall identifier, then using the provided functions to get arguments and set a return response. Direct access to the VM's memory space is not allowed, to avoid memory corruption issues.
The following functions are used to access syscall parameters safely:
uint32_t uvm32_getval(uvm32_state_t *vmst, uvm32_evt_t *evt, uvm32_arg_t);
const char *uvm32_getcstr(uvm32_state_t *vmst, uvm32_evt_t *evt, uvm32_arg_t);
void uvm32_setval(uvm32_state_t *vmst, uvm32_evt_t *evt, uvm32_arg_t, uint32_t val);
uvm32_evt_syscall_buf_t uvm32_getbuf(uvm32_state_t *vmst, uvm32_evt_t *evt, uvm32_arg_t argPtr, uvm32_arg_t argLen);
Event driven operation
A useful pattern for code running in the VM is to be event-driven. In this setup the program requests blocks until woken up with a reason. This requires some support in the host, but can be implemented as follows.
In the VM code:
while(1) {
// ask host to suspend running until one of the following events
uint32_t wakeReason = yield(KEYPRESS_EVENT_MASK | MOUSE_EVENT_MASK | NETWORK_EVENT_MASK);
switch(wakeReason) {
...
}
}
In the host code:
uvm32_run(&vmst, &evt, 1000);
switch(evt.typ) {
case UVM32_EVT_SYSCALL:
switch(evt.data.syscall.code) {
case UVM32_SYSCALL_YIELD:
uint32_t events = uvm32_getval(&vmst, &evt, ARG0);
// do not call uvm32_run() again until something in the events set is triggered
break;
...
}
break;
...
}
Then, to wake the VM once an (eg. key) event has occurred
uvm32_setval(&vmst, &evt, RET, KEYPRESS_EVENT_MASK);
uvm32_run(&vmst, &evt, 1000);
Configuration
The uvm32 memory size is set at compile time with -DUVM32_MEMORY_SIZE=X (in bytes). A memory of 512 bytes will be sufficient for trivial programs.
Define UVM32_ERROR_STRINGS to add an errstr field to uvm32_evt_err_t giving a printable error string.
Define UVM32_STACK_PROTECTION to enable a basic stack canary, to cause an early crash when the stack grows too large. Without this, the VM will normally crash (safely) in some other way which is less easily detected.
Debugging
Binaries can be disassembled with
riscv64-elf-objdump -d -f hello-asm.elf
riscv64-elf-objdump -S -d -b binary -m riscv:rv32 -D -M no-aliases -f hello-asm.bin