[RFC PATCH v1 1/1] arm64: Forget syscall if different from execve*()

[Catalin Marinas]

On Tue, May 10, 2022 at 03:03:33PM +0100, Will Deacon wrote:
> On Tue, May 10, 2022 at 03:00:11PM +0100, Francis Laniel wrote:
> > Le mardi 10 mai 2022, 11:59:48 BST Will Deacon a �crit :
> > > On Mon, May 09, 2022 at 04:19:57PM +0100, Francis Laniel wrote:
> > > > diff --git a/arch/arm64/include/asm/processor.h
> > > > b/arch/arm64/include/asm/processor.h index 73e38d9a540c..e12ceb363d6a
> > > > 100644
> > > > --- a/arch/arm64/include/asm/processor.h
> > > > +++ b/arch/arm64/include/asm/processor.h
> > > > @@ -34,6 +34,8 @@
> > > > 
> > > >  #include <vdso/processor.h>
> > > > 
> > > > +#include <asm-generic/unistd.h>
> > > > +
> > > > 
> > > >  #include <asm/alternative.h>
> > > >  #include <asm/cpufeature.h>
> > > >  #include <asm/hw_breakpoint.h>
> > > > 
> > > > @@ -250,8 +252,12 @@ void tls_preserve_current_state(void);
> > > > 
> > > >  static inline void start_thread_common(struct pt_regs *regs, unsigned
> > > >  long pc) {
> > > > 
> > > > +	s32 previous_syscall = regs->syscallno;
> > > > 
> > > >  	memset(regs, 0, sizeof(*regs));
> > > > 
> > > > -	forget_syscall(regs);
> > > > +	if (previous_syscall == __NR_execve || previous_syscall ==
> > > > __NR_execveat)
> > > > +		regs->syscallno = previous_syscall;
> > > > +	else
> > > > +		forget_syscall(regs);
> > > 
> > > Hmm, this really looks like a bodge and it doesn't handle the compat case
> > > either.
> > > 
> > > How do other architectures handle this?
> > 
> > My understanding of others architectures is quite limited, but here are my 
> > findings and understanding of some of them:
> > * arm (32 bits) EABI: start_thread() sets r7 to previous r7 for ELF FDPIC  and 
> > to 0 for other binfmts [1].
> > * arm (32 bits) OABI: syscall number is set to -1 if 
> > ptrace_report_syscall_entry() failed [2].
> > * mips: start_thread() does not modify current_thread_info->syscall which is 
> > taken directly from v0 [3, 4].
> > * riscv: start_thread() does not modify a7 [5].
> > * x86_64: start_thread_common() does not touch orig_ax which seems to contain 
> > the syscall number [6].
> 
> Hmm, so the million dollar question is why on Earth we have that
> forget_syscall() call to start with. Amusingly I've, err, forgotten;
> forget_forget_syscall() perhaps?
> 
> Catalin? It's been there since day one afaict.

Looking at the old logs, this appeared somewhere between day 0 and 1. We
had a memset(regs, 0) with a pt_regs of 36 registers but moved to
dedicated names for syscallno etc. and that's where I changed it from 0
to ~0UL.

A few weeks ago James Morse (cc'ed) came across this issue and even sent
a patch internally to remove forget_syscall() here. But that looks like
a bodge and James' suggestion was that maybe the core code can preserve
the syscallno and this would fix it for all architectures.

-- 
Catalin