Re: [PATCH] error-injection: Add prompt for function error injection

[Alexei Starovoitov]

On Fri, Dec 02, 2022 at 06:17:24PM -0500, Steven Rostedt wrote:
> On Fri, 2 Dec 2022 13:27:11 -0800
> Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote:
> 
> > --- a/lib/Kconfig.debug
> > +++ b/lib/Kconfig.debug
> > @@ -1874,8 +1874,12 @@ config NETDEV_NOTIFIER_ERROR_INJECT
> >  	  If unsure, say N.
> >  
> >  config FUNCTION_ERROR_INJECTION
> 
> Why not just call this "ERROR_INJECTION" having this be FUNCTION and the
> one for functions be FUNC is confusing.

That's what I had initially, but it causes plenty of churn to arch/*/Makefile
and a bunch of .h-s.
Keeping it as FUNCTION_ERROR_INJECTION removes all that noise from the diff.

> > +        bool
> > +
> > +config FUNC_ERROR_INJECTION
> >  	bool "Fault-injections of functions"
> >  	depends on HAVE_FUNCTION_ERROR_INJECTION && KPROBES
> > +	select FUNCTION_ERROR_INJECTION
> >  	help
> >  	  Add fault injections into various functions that are annotated with
> >  	  ALLOW_ERROR_INJECTION() in the kernel. BPF may also modify the return
> > @@ -1883,6 +1887,17 @@ config FUNCTION_ERROR_INJECTION
> >  
> >  	  If unsure, say N
> >  
> > +config SYSCALL_ERROR_INJECTION
> > +	bool "Error injections in syscalls"
> > +	depends on HAVE_FUNCTION_ERROR_INJECTION && KPROBES
> > +	select FUNCTION_ERROR_INJECTION
> > +	default y
> 
> IIUC, Linus prefers everything to be "default n" unless there's a really
> good reason for it. Like only making other options available, but not doing
> anything to the kernel. I do have DYNAMIC_FTRACE as "default y" but that's
> only because it depends on CONFIG_FUNCTION_TRACER and nobody that enables
> that should have DYNAMIC_FTRACE off (except for academia).

The FUNCTION_ERROR_INJECTION used to be "def_bool y" for ~5 years.
BTW the macro was called BPF_ALLOW_ERROR_INJECTION() when Josef initially implemented it.
Massami later renamed it ALLOW_ERROR_INJECTION() and allowed kprobes to use it.
Today there is a user expectation that this feature is available in the kernel.
We can do "default n" here, let distros decide and potentially upset users.
I don't feel strongly about that.

> 
> > +	help
> > +	  Allows error injection framework to return errors from syscalls.
> > +	  BPF may modify return values of syscalls as well.
> 
> And here's the thing. If BPF returns anything *but* an error, then this is
> a misnomer and incorrect. Name it something else like "HIJACK_SYSCALLS".

The bpf prog must return errno. No doubt about that.
Today the verifier validates return values whenever is necessary.
When original bpf_override_return was added the verifier wasn't that smart.
Since then we added return value checks pretty much everywhere.
Looks like the check is still missing bpf_override_return.
We will fix it asap.

> > +
> > +	  If unsure, say Y
> 
> And I'm curious, why Y if unsure?

Copy-paste. I can remove that line.