[PATCH 1/4] arm64: alternative: wait for other CPUs before patching

[Will Deacon]

On Mon, Dec 13, 2021 at 01:31:52PM +0000, Will Deacon wrote:
> On Fri, Dec 03, 2021 at 10:47:20AM +0000, Mark Rutland wrote:
> > In __apply_alternatives_multi_stop() we have a "really simple polling
> > protocol" to avoid patching code that is concurrently executed on other
> > CPUs. Secondary CPUs wait for the boot CPU to signal that patching is
> > complete, but the boot CPU doesn't wait for secondaries to enter the
> > polling loop, and it's possible that patching starts while secondaries
> > are still within the stop_machine logic.
> > 
> > Let's fix this by adding a vaguely simple polling protocol where the
> > boot CPU waits for secondaries to signal that they have entered the
> > unpatchable stop function. We can use the arch_atomic_*() functions for
> > this, as they are not patched with alternatives.
> > 
> > At the same time, let's make `all_alternatives_applied` local to
> > __apply_alternatives_multi_stop(), since it is only used there, and this
> > makes the code a little clearer.
> > 
> > Signed-off-by: Mark Rutland <mark.rutland at arm.com>
> > Cc: Andre Przywara <andre.przywara at arm.com>
> > Cc: Ard Biesheuvel <ardb at kernel.org>
> > Cc: Catalin Marinas <catalin.marinas at arm.com>
> > Cc: James Morse <james.morse at arm.com>
> > Cc: Joey Gouly <joey.gouly at arm.com>
> > Cc: Suzuki K Poulose <suzuki.poulose at arm.com>
> > Cc: Will Deacon <will at kernel.org>
> > ---
> >  arch/arm64/kernel/alternative.c | 17 ++++++++++++-----
> >  1 file changed, 12 insertions(+), 5 deletions(-)
> > 
> > diff --git a/arch/arm64/kernel/alternative.c b/arch/arm64/kernel/alternative.c
> > index 3fb79b76e9d9..4f32d4425aac 100644
> > --- a/arch/arm64/kernel/alternative.c
> > +++ b/arch/arm64/kernel/alternative.c
> > @@ -21,9 +21,6 @@
> >  #define ALT_ORIG_PTR(a)		__ALT_PTR(a, orig_offset)
> >  #define ALT_REPL_PTR(a)		__ALT_PTR(a, alt_offset)
> >  
> > -/* Volatile, as we may be patching the guts of READ_ONCE() */
> > -static volatile int all_alternatives_applied;
> > -
> >  static DECLARE_BITMAP(applied_alternatives, ARM64_NCAPS);
> >  
> >  struct alt_region {
> > @@ -193,11 +190,17 @@ static void __nocfi __apply_alternatives(struct alt_region *region, bool is_modu
> >  }
> >  
> >  /*
> > - * We might be patching the stop_machine state machine, so implement a
> > - * really simple polling protocol here.
> > + * Apply alternatives, ensuring that no CPUs are concurrently executing code
> > + * being patched.
> > + *
> > + * We might be patching the stop_machine state machine or READ_ONCE(), so
> > + * we implement a simple polling protocol.
> >   */
> >  static int __apply_alternatives_multi_stop(void *unused)
> >  {
> > +	/* Volatile, as we may be patching the guts of READ_ONCE() */
> > +	static volatile int all_alternatives_applied;
> > +	static atomic_t stopped_cpus = ATOMIC_INIT(0);
> >  	struct alt_region region = {
> >  		.begin	= (struct alt_instr *)__alt_instructions,
> >  		.end	= (struct alt_instr *)__alt_instructions_end,
> > @@ -205,12 +208,16 @@ static int __apply_alternatives_multi_stop(void *unused)
> >  
> >  	/* We always have a CPU 0 at this point (__init) */
> >  	if (smp_processor_id()) {
> > +		arch_atomic_inc(&stopped_cpus);
> 
> Why can't we use normal atomic_inc() here?

Ah, ok, this is to deal with instrumentation and you add 'noinstr' when you
factor this out later on. It does, however, mean that we need to be really
careful with this because we're relying on (a) our atomics patching using
static keys and (b) static key patching not requiring stop_machine().

In particular, we cannot backport this to kernels where the atomics were
patched directly.

> 
> >  		while (!all_alternatives_applied)
> >  			cpu_relax();
> >  		isb();
> >  	} else {
> >  		DECLARE_BITMAP(remaining_capabilities, ARM64_NPATCHABLE);
> >  
> > +		while (arch_atomic_read(&stopped_cpus) != num_online_cpus() - 1)
> 
> and normal atomic_read() here?

This one I'm still thinking doesn't need the arch_ prefix.

Will