Re: [RFC PATCH 1/1] x86/traps: Get rid of exception handlers' second argument error code

[Peter Zijlstra]

On Fri, Aug 04, 2023 at 05:35:11PM +0000, Li, Xin3 wrote:
> > > The IDT event delivery of X86_TRAP_DF, X86_TRAP_TS, X86_TRAP_NP,
> > > X86_TRAP_SS, X86_TRAP_GP, X86_TRAP_AC and X86_TRAP_CP pushes an error
> > > code into the orig_ax member of the pt_regs structure, and the error
> > > code is passed as the second argument of their C-handlers, although
> > > the pt_regs structure is already passed as the first argument.
> > >
> > > The asm entry code of such faults does the following
> > >
> > >   movq ORIG_RAX(%rsp), %rsi	/* get error code into 2nd argument*/
> > >   movq $-1, ORIG_RAX(%rsp)	/* no syscall to restart */
> > >
> > > to set the orig_ax member to -1 just before calling the C-handler.
> > >
> > > In addition, the IRQ entry code uses the second error code argument as
> > > its IRQ vector, as the IRQ asm entry code pushes its IRQ vector into
> > > the orig_ax member.
> > >
> > > The commit d99015b1abbad ("x86: move entry_64.S register saving out of
> > > the macros") introduced the changes to set orig_ax to -1, but I can't
> > > see why it's required. Our tests on x86_64 and x86_32 seem fine if
> > > orig_ax is left unchanged instead of set to -1.
> > 
> > That means that SYSCALL_NUM(regs) get to be garbage; or something like that.
> 
> I find SYSCALL_NUM(regs) in tools/testing/selftests/seccomp/seccomp_bpf.c,
> but nothing obvious to me.
> 
> I think it's clear that once exceptions and IRQs are handled, the original
> context will be fully recovered in a normal case.
> 
> Is it related to preemption after such a event?
> 
> I must have missed something; can you please elaborate it?

arch/x86/include/asm/syscall.h

syscall_get_nr() syscall_rollback()