Re: [RFC PATCH v5 1/2] selftests/x86: sysret_rip: Handle syscall in a FRED system

[H. Peter Anvin]

On 1/24/23 19:49, Ammar Faizi wrote:
> +
> +	/*
> +	 * Test that:
> +	 *
> +	 * - "syscall" in a FRED system doesn't clobber %rcx and %r11.
> +	 * - "syscall" in a non-FRED system sets %rcx=%rip and %r11=%rflags.
> +	 */
> +	ret = check_regs_result(r11, rcx, rbx);
> +	assert(ret != REGS_ERROR);
> +
> +	/*
> +	 * Test that we don't get a mix of REGS_SAVED and REGS_SYSRET.
> +	 * It needs at least calling do_syscall() twice to assert.
> +	 */
> +	if (regs_ok_state == REGS_UNDEFINED) {
> +		/*
> +		 * First time calling do_syscall().
> +		 */
> +		regs_ok_state = ret;
> +	} else {
> +		assert(regs_ok_state == ret);
> +	}
> +

[...]

> +	ret = check_regs_result(ctx->uc_mcontext.gregs[REG_R11],
> +				ctx->uc_mcontext.gregs[REG_RCX],
> +				ctx->uc_mcontext.gregs[REG_RBX]);
> +
> +	assert(ret != REGS_ERROR);
> +

This instance, too, needs to be checked against regs_ok_result. It would 
make most sense to move that handling, and the assert() into 
check_regs_result() or into a separate function around it.

>   	/* Set IP and CX to match so that SYSRET can happen. */
>   	ctx->uc_mcontext.gregs[REG_RIP] = rip;
>   	ctx->uc_mcontext.gregs[REG_RCX] = rip;

It would be interesting to have the syscall handler try both with and 
without this (so it would end up doing both IRET and SYSCALL on legacy.) 
Perhaps SIGUSR1 versus SIGUSR2...

	-hpa