diff mbox series

[SRU,Xenial,1/1] UBUNTU: SAUCE: (no-up) s390: fix rwlock implementation

Message ID 5619f9200cc002e3947b24dec43eff74c1f8e34d.1523284497.git.joseph.salisbury@canonical.com
State New
Headers show
Series UBUNTU: SAUCE: (no-up) s390: fix rwlock implementation | expand

Commit Message

Joseph Salisbury May 7, 2018, 10:04 p.m. UTC 
From: Heiko Carstens <heiko.carstens@de.ibm.com>

BugLink: http://bugs.launchpad.net/bugs/1761674

Description:  kernel: fix rwlock implementation
Symptom:      Kernel hangs, due to deadlock on an rwlock.
Problem:      With upstream commit 94232a4332de ("s390/rwlock: improve writer
              fairness") rwlock writer fairness was supposed to be
              implemented. If a writer tries to take an rwlock it sets
              unconditionally the writer bit within the lock word and waits
              until all readers have released the lock. This however can lead
              to a deadlock since rwlocks can be taken recursively by readers.
              If e.g. CPU 0 holds the lock as a reader, and CPU 1 wants to
              write-lock the lock, then CPU 1 sets the writer bit and
              afterwards busy waits for CPU 0 to release the lock. If now CPU 0
              tries to read-lock the lock again (recursively) it will also busy
              wait until CPU 1 removes the writer bit, which will never happen,
              since it waits for the first reader on CPU 0 to release the lock.
Solution:     Revert the rwlock writer fairness semantics again.

Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Joseph Salisbury <joseph.salisbury@canonical.com>
---
 arch/s390/lib/spinlock.c | 48 +++++++++---------------------------------------
 1 file changed, 9 insertions(+), 39 deletions(-)

Comments

Kleber Sacilotto de Souza May 11, 2018, 10:53 a.m. UTC | #1 
On 05/08/18 00:04, Joseph Salisbury wrote:
> From: Heiko Carstens <heiko.carstens@de.ibm.com>
> 
> BugLink: http://bugs.launchpad.net/bugs/1761674
> 
> Description:  kernel: fix rwlock implementation
> Symptom:      Kernel hangs, due to deadlock on an rwlock.
> Problem:      With upstream commit 94232a4332de ("s390/rwlock: improve writer
>               fairness") rwlock writer fairness was supposed to be
>               implemented. If a writer tries to take an rwlock it sets
>               unconditionally the writer bit within the lock word and waits
>               until all readers have released the lock. This however can lead
>               to a deadlock since rwlocks can be taken recursively by readers.
>               If e.g. CPU 0 holds the lock as a reader, and CPU 1 wants to
>               write-lock the lock, then CPU 1 sets the writer bit and
>               afterwards busy waits for CPU 0 to release the lock. If now CPU 0
>               tries to read-lock the lock again (recursively) it will also busy
>               wait until CPU 1 removes the writer bit, which will never happen,
>               since it waits for the first reader on CPU 0 to release the lock.
> Solution:     Revert the rwlock writer fairness semantics again.
> 
> Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
> Signed-off-by: Joseph Salisbury <joseph.salisbury@canonical.com>

Scope limited to specific arch, tested by vendor.

Acked-by: Kleber Sacilotto de Souza <kleber.souza@canonical.com>

> ---
>  arch/s390/lib/spinlock.c | 48 +++++++++---------------------------------------
>  1 file changed, 9 insertions(+), 39 deletions(-)
> 
> diff --git a/arch/s390/lib/spinlock.c b/arch/s390/lib/spinlock.c
> index 427aa44..99da25a 100644
> --- a/arch/s390/lib/spinlock.c
> +++ b/arch/s390/lib/spinlock.c
> @@ -182,42 +182,18 @@ int _raw_read_trylock_retry(arch_rwlock_t *rw)
>  EXPORT_SYMBOL(_raw_read_trylock_retry);
>  
>  #ifdef CONFIG_HAVE_MARCH_Z196_FEATURES
> -
>  void _raw_write_lock_wait(arch_rwlock_t *rw, unsigned int prev)
> -{
> -	unsigned int owner, old;
> -	int count = spin_retry;
> -
> -	owner = 0;
> -	while (1) {
> -		if (count-- <= 0) {
> -			if (owner && !smp_vcpu_scheduled(~owner))
> -				smp_yield_cpu(~owner);
> -			count = spin_retry;
> -		}
> -		old = ACCESS_ONCE(rw->lock);
> -		owner = ACCESS_ONCE(rw->owner);
> -		smp_mb();
> -		if ((int) old >= 0) {
> -			prev = __RAW_LOCK(&rw->lock, 0x80000000, __RAW_OP_OR);
> -			old = prev;
> -		}
> -		if ((old & 0x7fffffff) == 0 && (int) prev >= 0)
> -			break;
> -		if (MACHINE_HAS_CAD)
> -			_raw_compare_and_delay(&rw->lock, old);
> -	}
> -}
> -EXPORT_SYMBOL(_raw_write_lock_wait);
> -
> -#else /* CONFIG_HAVE_MARCH_Z196_FEATURES */
> -
> +#else
>  void _raw_write_lock_wait(arch_rwlock_t *rw)
> +#endif
>  {
> -	unsigned int owner, old, prev;
> +	unsigned int owner, old;
>  	int count = spin_retry;
>  
> -	prev = 0x80000000;
> +#ifdef CONFIG_HAVE_MARCH_Z196_FEATURES
> +	if ((int) prev > 0)
> +		__RAW_UNLOCK(&rw->lock, 0x7fffffff, __RAW_OP_AND);
> +#endif
>  	owner = 0;
>  	while (1) {
>  		if (count-- <= 0) {
> @@ -227,12 +203,8 @@ void _raw_write_lock_wait(arch_rwlock_t *rw)
>  		}
>  		old = ACCESS_ONCE(rw->lock);
>  		owner = ACCESS_ONCE(rw->owner);
> -		if ((int) old >= 0 &&
> -		    _raw_compare_and_swap(&rw->lock, old, old | 0x80000000))
> -			prev = old;
> -		else
> -			smp_mb();
> -		if ((old & 0x7fffffff) == 0 && (int) prev >= 0)
> +		if (old == 0 &&
> +		    _raw_compare_and_swap(&rw->lock, 0, 0x80000000))
>  			break;
>  		if (MACHINE_HAS_CAD)
>  			_raw_compare_and_delay(&rw->lock, old);
> @@ -240,8 +212,6 @@ void _raw_write_lock_wait(arch_rwlock_t *rw)
>  }
>  EXPORT_SYMBOL(_raw_write_lock_wait);
>  
> -#endif /* CONFIG_HAVE_MARCH_Z196_FEATURES */
> -
>  int _raw_write_trylock_retry(arch_rwlock_t *rw)
>  {
>  	unsigned int old;
>
Stefan Bader May 23, 2018, 12:55 p.m. UTC | #2 
On 08.05.2018 00:04, Joseph Salisbury wrote:
> From: Heiko Carstens <heiko.carstens@de.ibm.com>
> 
> BugLink: http://bugs.launchpad.net/bugs/1761674
> 
> Description:  kernel: fix rwlock implementation
> Symptom:      Kernel hangs, due to deadlock on an rwlock.
> Problem:      With upstream commit 94232a4332de ("s390/rwlock: improve writer
>               fairness") rwlock writer fairness was supposed to be
>               implemented. If a writer tries to take an rwlock it sets
>               unconditionally the writer bit within the lock word and waits
>               until all readers have released the lock. This however can lead
>               to a deadlock since rwlocks can be taken recursively by readers.
>               If e.g. CPU 0 holds the lock as a reader, and CPU 1 wants to
>               write-lock the lock, then CPU 1 sets the writer bit and
>               afterwards busy waits for CPU 0 to release the lock. If now CPU 0
>               tries to read-lock the lock again (recursively) it will also busy
>               wait until CPU 1 removes the writer bit, which will never happen,
>               since it waits for the first reader on CPU 0 to release the lock.
> Solution:     Revert the rwlock writer fairness semantics again.
> 
> Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
> Signed-off-by: Joseph Salisbury <joseph.salisbury@canonical.com>
Acked-by: Stefan Bader <stefan.bader@canonical.com>

> ---
>  arch/s390/lib/spinlock.c | 48 +++++++++---------------------------------------
>  1 file changed, 9 insertions(+), 39 deletions(-)
> 
> diff --git a/arch/s390/lib/spinlock.c b/arch/s390/lib/spinlock.c
> index 427aa44..99da25a 100644
> --- a/arch/s390/lib/spinlock.c
> +++ b/arch/s390/lib/spinlock.c
> @@ -182,42 +182,18 @@ int _raw_read_trylock_retry(arch_rwlock_t *rw)
>  EXPORT_SYMBOL(_raw_read_trylock_retry);
>  
>  #ifdef CONFIG_HAVE_MARCH_Z196_FEATURES
> -
>  void _raw_write_lock_wait(arch_rwlock_t *rw, unsigned int prev)
> -{
> -	unsigned int owner, old;
> -	int count = spin_retry;
> -
> -	owner = 0;
> -	while (1) {
> -		if (count-- <= 0) {
> -			if (owner && !smp_vcpu_scheduled(~owner))
> -				smp_yield_cpu(~owner);
> -			count = spin_retry;
> -		}
> -		old = ACCESS_ONCE(rw->lock);
> -		owner = ACCESS_ONCE(rw->owner);
> -		smp_mb();
> -		if ((int) old >= 0) {
> -			prev = __RAW_LOCK(&rw->lock, 0x80000000, __RAW_OP_OR);
> -			old = prev;
> -		}
> -		if ((old & 0x7fffffff) == 0 && (int) prev >= 0)
> -			break;
> -		if (MACHINE_HAS_CAD)
> -			_raw_compare_and_delay(&rw->lock, old);
> -	}
> -}
> -EXPORT_SYMBOL(_raw_write_lock_wait);
> -
> -#else /* CONFIG_HAVE_MARCH_Z196_FEATURES */
> -
> +#else
>  void _raw_write_lock_wait(arch_rwlock_t *rw)
> +#endif
>  {
> -	unsigned int owner, old, prev;
> +	unsigned int owner, old;
>  	int count = spin_retry;
>  
> -	prev = 0x80000000;
> +#ifdef CONFIG_HAVE_MARCH_Z196_FEATURES
> +	if ((int) prev > 0)
> +		__RAW_UNLOCK(&rw->lock, 0x7fffffff, __RAW_OP_AND);
> +#endif
>  	owner = 0;
>  	while (1) {
>  		if (count-- <= 0) {
> @@ -227,12 +203,8 @@ void _raw_write_lock_wait(arch_rwlock_t *rw)
>  		}
>  		old = ACCESS_ONCE(rw->lock);
>  		owner = ACCESS_ONCE(rw->owner);
> -		if ((int) old >= 0 &&
> -		    _raw_compare_and_swap(&rw->lock, old, old | 0x80000000))
> -			prev = old;
> -		else
> -			smp_mb();
> -		if ((old & 0x7fffffff) == 0 && (int) prev >= 0)
> +		if (old == 0 &&
> +		    _raw_compare_and_swap(&rw->lock, 0, 0x80000000))
>  			break;
>  		if (MACHINE_HAS_CAD)
>  			_raw_compare_and_delay(&rw->lock, old);
> @@ -240,8 +212,6 @@ void _raw_write_lock_wait(arch_rwlock_t *rw)
>  }
>  EXPORT_SYMBOL(_raw_write_lock_wait);
>  
> -#endif /* CONFIG_HAVE_MARCH_Z196_FEATURES */
> -
>  int _raw_write_trylock_retry(arch_rwlock_t *rw)
>  {
>  	unsigned int old;
>
diff mbox series

Patch

diff --git a/arch/s390/lib/spinlock.c b/arch/s390/lib/spinlock.c
index 427aa44..99da25a 100644
--- a/arch/s390/lib/spinlock.c
+++ b/arch/s390/lib/spinlock.c
@@ -182,42 +182,18 @@  int _raw_read_trylock_retry(arch_rwlock_t *rw)
 EXPORT_SYMBOL(_raw_read_trylock_retry);
 
 #ifdef CONFIG_HAVE_MARCH_Z196_FEATURES
-
 void _raw_write_lock_wait(arch_rwlock_t *rw, unsigned int prev)
-{
-	unsigned int owner, old;
-	int count = spin_retry;
-
-	owner = 0;
-	while (1) {
-		if (count-- <= 0) {
-			if (owner && !smp_vcpu_scheduled(~owner))
-				smp_yield_cpu(~owner);
-			count = spin_retry;
-		}
-		old = ACCESS_ONCE(rw->lock);
-		owner = ACCESS_ONCE(rw->owner);
-		smp_mb();
-		if ((int) old >= 0) {
-			prev = __RAW_LOCK(&rw->lock, 0x80000000, __RAW_OP_OR);
-			old = prev;
-		}
-		if ((old & 0x7fffffff) == 0 && (int) prev >= 0)
-			break;
-		if (MACHINE_HAS_CAD)
-			_raw_compare_and_delay(&rw->lock, old);
-	}
-}
-EXPORT_SYMBOL(_raw_write_lock_wait);
-
-#else /* CONFIG_HAVE_MARCH_Z196_FEATURES */
-
+#else
 void _raw_write_lock_wait(arch_rwlock_t *rw)
+#endif
 {
-	unsigned int owner, old, prev;
+	unsigned int owner, old;
 	int count = spin_retry;
 
-	prev = 0x80000000;
+#ifdef CONFIG_HAVE_MARCH_Z196_FEATURES
+	if ((int) prev > 0)
+		__RAW_UNLOCK(&rw->lock, 0x7fffffff, __RAW_OP_AND);
+#endif
 	owner = 0;
 	while (1) {
 		if (count-- <= 0) {
@@ -227,12 +203,8 @@  void _raw_write_lock_wait(arch_rwlock_t *rw)
 		}
 		old = ACCESS_ONCE(rw->lock);
 		owner = ACCESS_ONCE(rw->owner);
-		if ((int) old >= 0 &&
-		    _raw_compare_and_swap(&rw->lock, old, old | 0x80000000))
-			prev = old;
-		else
-			smp_mb();
-		if ((old & 0x7fffffff) == 0 && (int) prev >= 0)
+		if (old == 0 &&
+		    _raw_compare_and_swap(&rw->lock, 0, 0x80000000))
 			break;
 		if (MACHINE_HAS_CAD)
 			_raw_compare_and_delay(&rw->lock, old);
@@ -240,8 +212,6 @@  void _raw_write_lock_wait(arch_rwlock_t *rw)
 }
 EXPORT_SYMBOL(_raw_write_lock_wait);
 
-#endif /* CONFIG_HAVE_MARCH_Z196_FEATURES */
-
 int _raw_write_trylock_retry(arch_rwlock_t *rw)
 {
 	unsigned int old;