[RFC PATCH 27/27] powerpc/64s: system call support for scv/rfscv instructions

[Michal Suchánek]

On Sun, Sep 15, 2019 at 11:28:13AM +1000, Nicholas Piggin wrote:
> Add support for the scv instruction on POWER9 and later CPUs.
> 
> For now this implements the zeroth scv vector 'scv 0', as identical
> to 'sc' system calls, with the exception that lr is not preserved, and
> it is 64-bit only. There may yet be changes made to this ABI, so it's
> for testing only.
> 
> This also doesn't yet properly handle PR KVM, or the case where a guest
> is denied AIL=3 mode. I haven't added real mode entry points, so scv
> must not be used when AIL=0, but I haven't cleared the FSCR in this
> case.
> 
> This also implements a strange hack to handle the non-implemented
> vectors, scheduling a decrementer and expecting it to interrupt and
> replay pending soft masked interrupts. This is unlikely to be a good
> idea, and needs to actually do a proper handler and replay in case an
> interrupt is pending.
> 
> It may also require some errata handling before it can be safely used
> on all POWER9 CPUs, I have to look that up.
> 
> rfscv is implemented to return from scv type system calls. It can not
> be used to return from sc system calls because those are defined to
> preserve lr.
> 
> In a comparison of getpid syscall, the test program had scv taking
> about 3 more cycles in user mode (92 vs 89 for sc), due to lr handling.
> Total cycles taken for a getpid system call on POWER9 are improved from
> 436 to 345 (26.3% faster), mostly due to reducing mtmsr and mtspr.
...
> diff --git a/arch/powerpc/kernel/syscall_64.c b/arch/powerpc/kernel/syscall_64.c
> index 034b52d3d78c..3e8aa5ae8ec8 100644
> --- a/arch/powerpc/kernel/syscall_64.c
> +++ b/arch/powerpc/kernel/syscall_64.c
> @@ -15,6 +15,77 @@ extern void __noreturn tabort_syscall(void);
>  
>  typedef long (*syscall_fn)(long, long, long, long, long, long);
>  
> +#ifdef CONFIG_PPC_BOOK3S
> +long system_call_vectored_exception(long r3, long r4, long r5, long r6, long r7, long r8, unsigned long r0, struct pt_regs *regs)
> +{
> +	unsigned long ti_flags;
> +	syscall_fn f;
> +
> +	BUG_ON(!(regs->msr & MSR_RI));
> +	BUG_ON(!(regs->msr & MSR_PR));
> +	BUG_ON(!FULL_REGS(regs));
> +	BUG_ON(regs->softe != IRQS_ENABLED);
> +
> +	if (IS_ENABLED(CONFIG_PPC_TRANSACTIONAL_MEM) &&
> +			unlikely(regs->msr & MSR_TS_T))
> +		tabort_syscall();
> +
> +	account_cpu_user_entry();
> +
> +#ifdef CONFIG_PPC_SPLPAR
> +	if (IS_ENABLED(CONFIG_VIRT_CPU_ACCOUNTING_NATIVE) &&
> +			firmware_has_feature(FW_FEATURE_SPLPAR)) {
> +		struct lppaca *lp = get_lppaca();
> +
> +		if (unlikely(local_paca->dtl_ridx != lp->dtl_idx))
This adds another instance of the lack of endian conversion issue.
> +			accumulate_stolen_time();
> +	}
> +#endif

Thanks

Michal