Juraj's Blog

22 Apr 2021

(Mis)understanding RISC-V ecalls and syscalls

After spending some time with emuriscv and attempting to boot Linux into a shell I realized that I’m doing something really wrong regarding system calls.

RISC-V offers an ecall (Environment Call) instruction to implement system calls. These are basically requests made by a lower privileged code (user mode) to execute higher privileged code (kernel). Or, in some other case, the kernel itself can be the lower privileged code and it would invoke the machine mode code with an ecall.

On the RISC-V platform this call can also act as a convenient way of providing input/output, either from bare metal code or the kernel.

Enter the SBI for console output

According to the SBI (supervisor binary interface), which I take for a BIOS equivalent in the RISC-V world, there’s is a legacy “console” interface with two functions:

void sbi_console_putchar(int ch)
int sbi_console_getchar(void)

This getchar/putchar pair interacts with a debug console. Linux happens to provide a RISC-V SBI console driver , that’s enabled with the HVC_RISCV_SBI configuration option and calls into functions implemented in sbi.c .

The implementation calls a function named sbi_ecall, that generates the assembly code invoking the ecall and collects the return code (if any).

void sbi_console_putchar(int ch)
{
	sbi_ecall(SBI_EXT_0_1_CONSOLE_PUTCHAR, 0, ch, 0, 0, 0, 0, 0);
}

which gets compiled into the following assembly:

li a7, 1
li a0, ch
ecall

Doing it the wrong way

I misunderstood this and implemented an ecall handler in my emulator that did capture the arguments and printed out the parameter into the standard output:

//ecall_callback
	if (state->x[SYSCALL_REG] == SBI_CONSOLE_PUTCHAR) {
		char c = (char)state->x[SBI_ARG0_REG];
		fprintf(stdout, "%c", c);
		state->x[SBI_RETURN_REG] = SBI_SUCCESS;
	}
    else if (state->x[SYSCALL_REG] == SBI_CONSOLE_GETCHAR) {
		//write invalid return value in the register a0
		state->x[SBI_RETURN_REG] = -1;
	}

While this technically did work and I got the Linux boot console output, I realized quite late that the same ecall instruction is also used for something else than printing out characters to the screen. The reason it took so long was that I discarded all the other ecall invocations in my implementation.

As ecall is also used by the user-mode programs to call into the kernel, it also meant that no system calls whatsoever would get through to the kernel.

Doing it the right way

RISC-V defines the following mechanism for actually handling the ecalls - they should be an atomic jump to a controlled location, handled by an exception handler.

image

There are the following exception causes that correspond to bits that can be set in the CSR_MEDELEG register. If the delegation bit at the specified index is set, then exception gets delegated to S mode, otherwise it’s handled in M mode.

name code
CAUSE_USER_ECALL 0x8
CAUSE_SUPERVISOR_ECALL 0x9
CAUSE_HYPERVISOR_ECALL 0xa
CAUSE_MACHINE_ECALL 0xb

Now, if we have a simple binary that contains these instructions:

addi    a0, x0, 0   # Return code 0
addi    a7, x0, 93  # Syscall 93 terminates
ecall               # Call OS to terminate the program

The ecall raises an exception with the CAUSE_USER_ECALL, which gets trapped by the kernel trap handler and handled properly.

Interactive MEDELEG CSR decoder

Enter the medeleg value to see which exception bits it contains.

Bit 0 1 2 3 4 5 6
Field MISALIGNED
FETCH
FETCH
ACCESS
ILLEGAL
INSTRUCTION
BREAKPOINT MISALIGNED
LOAD
LOAD
ACCESS
MISALIGNED
STORE
Value
Bit 7 8 9 a b c d f
Field STORE
ACCESS
USER
ECALL
SUPERVISOR
ECALL
HYPERVISOR
ECALL
MACHINE
ECALL
FETCH
PAGE
FAULT
LOAD
PAGE
FAULT
STORE
PAGE
FAULT
Value

MISA CSR register and platform capabilities identification

To allow BBL/OpenSBI to set up trap handlers we should tell it which extensions we support by setting up the CSR_MISA register as follows, as it queries supports_extension('S') before setting up supervisor mode traps:

//RV32 IMAS  -> bits 0, 8, 12, 18, XLEN32 (bit 30)
state->csr[CSR_MISA] = 1 << 0 | 1 << 8 | 1 << 12 | 1 << 18 | 1 << 30;

Now we have syscalls, but no console again :(

Getting console back, which output driver to support?

As the ecall is now properly handled either by the OS or the OpenSBI, we need some way to produce output again.

The sbi_console_putchar call from kernel is now trapped by OpenSBI/BBL by its sbi_console module, that dispatches the character into a specific console driver.

As I am using both BBL and OpenSBI I was looking for something that’s easy to implement and available in both loaders.

Driver BBL OpenSBI
SiFive UART ✔️ ✔️
8250/16550 UART ✔️ ✔️
LiteX UART ✔️
HTIF ✔️ ✔️
Shakti UART ✔️

Because HTIF is not really supported on 32-bit (RV32) architecture, it’s out. SiFive UART seemed simpler than the 8250/16550 UART, with a couple of FIFO registers and flags.

Implementing SiFive UART in emuriscv

There are two parts to a virtual device - its implementation and making it discoverable.

The standard RISC-V mechanism of hardware discovery is via the device tree, so I had to create an entry for this new UART device:

uart@10000000 {
    compatible = "sifive,uart0";
    reg = <0x00 0x10000000 0x00 0x100>;
};

ℹ️ We can turn binary FDT into a text-based representation by dtc -I dtb -O dts binary_fdt.dtb

This tells the OpenSBI/BBL to initialize the SiFive UART driver, pointing to the address 0x10000000. The counterpart on the emulator side is a memory-mapped device, which has a following write handler:

int32_t uart_reg[7];
#define UART_REG_TXFIFO		0
#define UART_REG_RXFIFO		1
...
static void uart_write(void* opaque, uint32_t offset, uint32_t val,
	int size_log2)
{
	int offset_words = offset >> 2;
	uart_reg[offset_words] = val;
	if (offset_words == UART_REG_TXFIFO) {
		if(val != 0) //skip null characters
			fputc(val, stderr);
	}
}

There’s a similar uart_read() function that just returns whatever is inside the UART registers.

I liked this memory-mapped device mechanism from Bellard’s TinyEmu and could not think of anything more reasonable.

How about console input?

I’ve not implemented it yet as I’m still stuck in a phase where I didn’t get usermode code to print out a single character. However, a reasonable implementation would send a character from the standard input of the emulator with something like getch().

I’m still not sure whether this is all that’s needed as I see many calls to sbi_console_getchar around the boot time, which could be just polling for input. I’m also getting a lot of @^@^@^ sequences, which could actually mean something - an escape sequence that a terminal is supposed to respond to, I’m not exactly sure at this point in time.