[PATCH v3 0/1] rust: helpers: Avoid raw_spin_lock initialization for PREEMPT_RT

[PATCH v3 0/1] rust: helpers: Avoid raw_spin_lock initialization for PREEMPT_RT

[Eder Zulian]

Hello!

When PREEMPT_RT=y, spin locks are mapped to rt_mutex types, so using
spinlock_check() + __raw_spin_lock_init() to initialize spin locks is
incorrect, and would cause build errors. 

This v3 patch introduces __spin_lock_init() to initialize a spin lock with
lockdep rquired information for PREEMPT_RT builds, and use it in the Rust
helper.

This patch was developed on top of linux-next/master.

As a note, at the time of writing, RUST support for x86_64 depends on
!(MITIGATION_RETHUNK && KASAN) || RUSTC_VERSION >= 108300. Miguel Ojeda
pointed out that this can be avoided with Rust 1.83, to be released in 3
weeks (2024-11-28).

In order to reproduce the problem rust must be available on the system.

  $ make LLVM=1 rustavailable

With CONFIG_PREEMPT_RT=y, CONFIG_RUST=y, and CONFIG_DEBUG_SPINLOCK=y a
x86_64 kernel can be built with

  $ make LLVM=1 -j$(nproc) bzImage 

The problem was reported at least in:
https://lore.kernel.org/oe-kbuild-all/202409251238.vetlgXE9-lkp@intel.com/
https://lore.kernel.org/all/20241107182411.57e2b418@canb.auug.org.au/

Links to v1 and v2 where improvement suggestions were made:
https://lore.kernel.org/all/20241014195253.1704625-1-ezulian@redhat.com/
https://lore.kernel.org/all/20241106211215.2005909-1-ezulian@redhat.com/

Version 2 changes:

- Cleaned up style and incorporated feedback from reviewers Boqun Feng and
  Miguel Ojeda.

Version 3 changes:

- Addressed review comments from Boqun Feng. Improved commit title and
  description and used a proper 'Fixed:' tag.

Thanks,

Eder Zulian (1):
  rust: helpers: Avoid raw_spin_lock initialization for PREEMPT_RT

 include/linux/spinlock_rt.h | 15 +++++++--------
 rust/helpers/spinlock.c     |  8 ++++++--
 2 files changed, 13 insertions(+), 10 deletions(-)

-- 
2.47.0

[PATCH v3 1/1] rust: helpers: Avoid raw_spin_lock initialization for PREEMPT_RT

[Eder Zulian]

When PREEMPT_RT=y, spin locks are mapped to rt_mutex types, so using
spinlock_check() + __raw_spin_lock_init() to initialize spin locks is
incorrect, and would cause build errors.

Introduce __spin_lock_init() to initialize a spin lock with lockdep
rquired information for PREEMPT_RT builds, and use it in the Rust
helper.

Fixes: d2d6422f8bd1 ("x86: Allow to enable PREEMPT_RT.")
Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/oe-kbuild-all/202409251238.vetlgXE9-lkp@intel.com/
Signed-off-by: Eder Zulian <ezulian@redhat.com>
---
V1 -> V2: Cleaned up style and addressed review comments
V2 -> V3: Improved commit title and description and corrected the 'Fixed:'
tag as per reviewer's suggestion

 include/linux/spinlock_rt.h | 15 +++++++--------
 rust/helpers/spinlock.c     |  8 ++++++--
 2 files changed, 13 insertions(+), 10 deletions(-)

diff --git a/include/linux/spinlock_rt.h b/include/linux/spinlock_rt.h
index f9f14e135be7..f6499c37157d 100644
--- a/include/linux/spinlock_rt.h
+++ b/include/linux/spinlock_rt.h
@@ -16,22 +16,21 @@ static inline void __rt_spin_lock_init(spinlock_t *lock, const char *name,
 }
 #endif
 
-#define spin_lock_init(slock)					\
+#define __spin_lock_init(slock, name, key, percpu)		\
 do {								\
-	static struct lock_class_key __key;			\
-								\
 	rt_mutex_base_init(&(slock)->lock);			\
-	__rt_spin_lock_init(slock, #slock, &__key, false);	\
+	__rt_spin_lock_init(slock, name, key, percpu);		\
 } while (0)
 
-#define local_spin_lock_init(slock)				\
+#define _spin_lock_init(slock, percpu)				\
 do {								\
 	static struct lock_class_key __key;			\
-								\
-	rt_mutex_base_init(&(slock)->lock);			\
-	__rt_spin_lock_init(slock, #slock, &__key, true);	\
+	__spin_lock_init(slock, #slock, &__key, percpu);	\
 } while (0)
 
+#define spin_lock_init(slock)		_spin_lock_init(slock, false)
+#define local_spin_lock_init(slock)	_spin_lock_init(slock, true)
+
 extern void rt_spin_lock(spinlock_t *lock) __acquires(lock);
 extern void rt_spin_lock_nested(spinlock_t *lock, int subclass)	__acquires(lock);
 extern void rt_spin_lock_nest_lock(spinlock_t *lock, struct lockdep_map *nest_lock) __acquires(lock);
diff --git a/rust/helpers/spinlock.c b/rust/helpers/spinlock.c
index b7b0945e8b3c..5971fdf6f755 100644
--- a/rust/helpers/spinlock.c
+++ b/rust/helpers/spinlock.c
@@ -6,10 +6,14 @@ void rust_helper___spin_lock_init(spinlock_t *lock, const char *name,
 				  struct lock_class_key *key)
 {
 #ifdef CONFIG_DEBUG_SPINLOCK
+# if defined(CONFIG_PREEMPT_RT)
+	__spin_lock_init(lock, name, key, false);
+# else /*!CONFIG_PREEMPT_RT */
 	__raw_spin_lock_init(spinlock_check(lock), name, key, LD_WAIT_CONFIG);
-#else
+# endif /* CONFIG_PREEMPT_RT */
+#else /* !CONFIG_DEBUG_SPINLOCK */
 	spin_lock_init(lock);
-#endif
+#endif /* CONFIG_DEBUG_SPINLOCK */
 }
 
 void rust_helper_spin_lock(spinlock_t *lock)
-- 
2.47.0

Re: [PATCH v3 1/1] rust: helpers: Avoid raw_spin_lock initialization for PREEMPT_RT

[Boqun Feng]

On Thu, Nov 07, 2024 at 05:32:23PM +0100, Eder Zulian wrote:
> When PREEMPT_RT=y, spin locks are mapped to rt_mutex types, so using
> spinlock_check() + __raw_spin_lock_init() to initialize spin locks is
> incorrect, and would cause build errors.
> 
> Introduce __spin_lock_init() to initialize a spin lock with lockdep
> rquired information for PREEMPT_RT builds, and use it in the Rust
> helper.
> 
> Fixes: d2d6422f8bd1 ("x86: Allow to enable PREEMPT_RT.")
> Reported-by: kernel test robot <lkp@intel.com>
> Closes: https://lore.kernel.org/oe-kbuild-all/202409251238.vetlgXE9-lkp@intel.com/
> Signed-off-by: Eder Zulian <ezulian@redhat.com>

Reviewed-by: Boqun Feng <boqun.feng@gmail.com>

For testing, I used the following command as a reproducer:

ARM64:

./tools/testing/kunit/kunit.py run --make_options LLVM=1 --arch arm64 --kconfig_add CONFIG_RUST=y --kconfig_add CONFIG_SMP=y  rust_doctests_kernel --kconfig_add CONFIG_WERROR=y --kconfig_add CONFIG_EXPERT=y --kconfig_add CONFIG_PREEMPT_RT=y --kconfig_add CONFIG_PROVE_LOCKING=y

X86_64:

./tools/testing/kunit/kunit.py run --make_options LLVM=1 --arch arm64 --kconfig_add CONFIG_RUST=y --kconfig_add CONFIG_SMP=y  rust_doctests_kernel --kconfig_add CONFIG_WERROR=y --kconfig_add CONFIG_EXPERT=y --kconfig_add CONFIG_PREEMPT_RT=y --kconfig_add CONFIG_PROVE_LOCKING=y

And I applied this onto the tip/locking/core, can confirm the build
errors are gone. So

Tested-by: Boqun Feng <boqun.feng@gmail.com>

Regards,
Boqun

> ---
> V1 -> V2: Cleaned up style and addressed review comments
> V2 -> V3: Improved commit title and description and corrected the 'Fixed:'
> tag as per reviewer's suggestion
> 
>  include/linux/spinlock_rt.h | 15 +++++++--------
>  rust/helpers/spinlock.c     |  8 ++++++--
>  2 files changed, 13 insertions(+), 10 deletions(-)
> 
> diff --git a/include/linux/spinlock_rt.h b/include/linux/spinlock_rt.h
> index f9f14e135be7..f6499c37157d 100644
> --- a/include/linux/spinlock_rt.h
> +++ b/include/linux/spinlock_rt.h
> @@ -16,22 +16,21 @@ static inline void __rt_spin_lock_init(spinlock_t *lock, const char *name,
>  }
>  #endif
>  
> -#define spin_lock_init(slock)					\
> +#define __spin_lock_init(slock, name, key, percpu)		\
>  do {								\
> -	static struct lock_class_key __key;			\
> -								\
>  	rt_mutex_base_init(&(slock)->lock);			\
> -	__rt_spin_lock_init(slock, #slock, &__key, false);	\
> +	__rt_spin_lock_init(slock, name, key, percpu);		\
>  } while (0)
>  
> -#define local_spin_lock_init(slock)				\
> +#define _spin_lock_init(slock, percpu)				\
>  do {								\
>  	static struct lock_class_key __key;			\
> -								\
> -	rt_mutex_base_init(&(slock)->lock);			\
> -	__rt_spin_lock_init(slock, #slock, &__key, true);	\
> +	__spin_lock_init(slock, #slock, &__key, percpu);	\
>  } while (0)
>  
> +#define spin_lock_init(slock)		_spin_lock_init(slock, false)
> +#define local_spin_lock_init(slock)	_spin_lock_init(slock, true)
> +
>  extern void rt_spin_lock(spinlock_t *lock) __acquires(lock);
>  extern void rt_spin_lock_nested(spinlock_t *lock, int subclass)	__acquires(lock);
>  extern void rt_spin_lock_nest_lock(spinlock_t *lock, struct lockdep_map *nest_lock) __acquires(lock);
> diff --git a/rust/helpers/spinlock.c b/rust/helpers/spinlock.c
> index b7b0945e8b3c..5971fdf6f755 100644
> --- a/rust/helpers/spinlock.c
> +++ b/rust/helpers/spinlock.c
> @@ -6,10 +6,14 @@ void rust_helper___spin_lock_init(spinlock_t *lock, const char *name,
>  				  struct lock_class_key *key)
>  {
>  #ifdef CONFIG_DEBUG_SPINLOCK
> +# if defined(CONFIG_PREEMPT_RT)
> +	__spin_lock_init(lock, name, key, false);
> +# else /*!CONFIG_PREEMPT_RT */
>  	__raw_spin_lock_init(spinlock_check(lock), name, key, LD_WAIT_CONFIG);
> -#else
> +# endif /* CONFIG_PREEMPT_RT */
> +#else /* !CONFIG_DEBUG_SPINLOCK */
>  	spin_lock_init(lock);
> -#endif
> +#endif /* CONFIG_DEBUG_SPINLOCK */
>  }
>  
>  void rust_helper_spin_lock(spinlock_t *lock)
> -- 
> 2.47.0
> 

Re: [PATCH v3 0/1] rust: helpers: Avoid raw_spin_lock initialization for PREEMPT_RT

[Miguel Ojeda]

On Thu, Nov 7, 2024 at 5:33 PM Eder Zulian <ezulian@redhat.com> wrote:
>
> As a note, at the time of writing, RUST support for x86_64 depends on
> !(MITIGATION_RETHUNK && KASAN) || RUSTC_VERSION >= 108300. Miguel Ojeda
> pointed out that this can be avoided with Rust 1.83, to be released in 3
> weeks (2024-11-28).

I was referring there to the "or" in that condition, i.e. the "||
RUSTC_VERSION >= 108300" part. In other words, it was just a comment I
made to explain in the other thread that disabling KASAN or RETHUNK is
not needed anymore when you use 1.83 in the future. :)

But that seems unrelated to the patch here, so normally you wouldn't
add it to the cover letter. Or am I missing something? Same for the
`make rustavailable` note below (i.e. `RUST=y` already implies that).

(Of course, no need to resend anything for this -- it is just a note
to clarify, and anyway the cover letter does not go into the
repository :)

Thanks!

Cheers,
Miguel

Re: [PATCH v3 0/1] rust: helpers: Avoid raw_spin_lock initialization for PREEMPT_RT

[Eder Zulian]

Hi Miguel,

On Thu, Nov 07, 2024 at 05:50:50PM +0100, Miguel Ojeda wrote:
> On Thu, Nov 7, 2024 at 5:33 PM Eder Zulian <ezulian@redhat.com> wrote:
> >
> > As a note, at the time of writing, RUST support for x86_64 depends on
> > !(MITIGATION_RETHUNK && KASAN) || RUSTC_VERSION >= 108300. Miguel Ojeda
> > pointed out that this can be avoided with Rust 1.83, to be released in 3
> > weeks (2024-11-28).
> 
> I was referring there to the "or" in that condition, i.e. the "||
> RUSTC_VERSION >= 108300" part. In other words, it was just a comment I
> made to explain in the other thread that disabling KASAN or RETHUNK is
> not needed anymore when you use 1.83 in the future. :)
> 

Yes, I thought that was clear all along.

> But that seems unrelated to the patch here, so normally you wouldn't
> add it to the cover letter. Or am I missing something? Same for the
> `make rustavailable` note below (i.e. `RUST=y` already implies that).
> 

Noted. I don't think you're missing anything. Thank you for the hints.

> (Of course, no need to resend anything for this -- it is just a note
> to clarify, and anyway the cover letter does not go into the
> repository :)
> 
> Thanks!
> 
> Cheers,
> Miguel
> 

Thank you,
Eder