| Message ID | 20230113065955.815667-2-boqun.feng@gmail.com (mailing list archive) |
|---|---|
| State | Accepted |
| Commit | bb5abfaa476b7329d875566e3bcc9cc6e1519a12 |
| Headers | show |
| Series | Detect SRCU related deadlocks | expand |
On 1/13/23 01:59, Boqun Feng wrote: > Currently, in order to annonate functions like synchronize_srcu() for > lockdep, a trick as follow can be used: > > lock_acquire(); > lock_release(); > > , which indicates synchronize_srcu() acts like an empty critical section > that waits for other (read-side) critical sections to finish. This > surely can catch some deadlock, but as discussion brought up by Paul > Mckenney [1], this could introduce false positives because of > irq-safe/unsafe detection. Extra tricks might help this: > > local_irq_disable(..); > lock_acquire(); > lock_release(); > local_irq_enable(...); > > But it's better that lockdep could provide an annonation for > synchronize_srcu() like functions, so that people won't need to repeat > the ugly tricks above. Therefore introduce lock_sync(). It's simply an > lock+unlock pair with no irq safe/unsafe deadlock check, since the > to-be-annontated functions don't create real critical sections therefore > there is no way that irq can create extra dependencies. > > [1]: https://lore.kernel.org/lkml/20180412021233.ewncg5jjuzjw3x62@tardis/ > > Signed-off-by: Boqun Feng <boqun.feng@gmail.com> > --- > include/linux/lockdep.h | 5 +++++ > kernel/locking/lockdep.c | 34 ++++++++++++++++++++++++++++++++++ > 2 files changed, 39 insertions(+) > > diff --git a/include/linux/lockdep.h b/include/linux/lockdep.h > index 1f1099dac3f0..ba09df6a0872 100644 > --- a/include/linux/lockdep.h > +++ b/include/linux/lockdep.h > @@ -268,6 +268,10 @@ extern void lock_acquire(struct lockdep_map *lock, unsigned int subclass, > > extern void lock_release(struct lockdep_map *lock, unsigned long ip); > > +extern void lock_sync(struct lockdep_map *lock, unsigned int subclass, > + int read, int check, struct lockdep_map *nest_lock, > + unsigned long ip); > + > /* lock_is_held_type() returns */ > #define LOCK_STATE_UNKNOWN -1 > #define LOCK_STATE_NOT_HELD 0 > @@ -555,6 +559,7 @@ do { \ > #define lock_map_acquire_read(l) lock_acquire_shared_recursive(l, 0, 0, NULL, _THIS_IP_) > #define lock_map_acquire_tryread(l) lock_acquire_shared_recursive(l, 0, 1, NULL, _THIS_IP_) > #define lock_map_release(l) lock_release(l, _THIS_IP_) > +#define lock_map_sync(l) lock_sync(l, 0, 0, 1, NULL, _THIS_IP_) > > #ifdef CONFIG_PROVE_LOCKING > # define might_lock(lock) \ > diff --git a/kernel/locking/lockdep.c b/kernel/locking/lockdep.c > index e3375bc40dad..cffa026a765f 100644 > --- a/kernel/locking/lockdep.c > +++ b/kernel/locking/lockdep.c > @@ -5692,6 +5692,40 @@ void lock_release(struct lockdep_map *lock, unsigned long ip) > } > EXPORT_SYMBOL_GPL(lock_release); > > +/* > + * lock_sync() - A special annotation for synchronize_{s,}rcu()-like API. > + * > + * No actual critical section is created by the APIs annotated with this: these > + * APIs are used to wait for one or multiple critical sections (on other CPUs > + * or threads), and it means that calling these APIs inside these critical > + * sections is potential deadlock. > + * > + * This annotation acts as an acqurie+release anontation pair with hardirqoff > + * being 1. Since there's no critical section, no interrupt can create extra > + * dependencies "inside" the annotation, hardirqoff == 1 allows us to avoid > + * false positives. > + */ > +void lock_sync(struct lockdep_map *lock, unsigned subclass, int read, > + int check, struct lockdep_map *nest_lock, unsigned long ip) > +{ > + unsigned long flags; > + > + if (unlikely(!lockdep_enabled())) > + return; > + > + raw_local_irq_save(flags); > + check_flags(flags); > + > + lockdep_recursion_inc(); > + __lock_acquire(lock, subclass, 0, read, check, 1, nest_lock, ip, 0, 0); > + > + if (__lock_release(lock, ip)) > + check_chain_key(current); > + lockdep_recursion_finish(); > + raw_local_irq_restore(flags); > +} > +EXPORT_SYMBOL_GPL(lock_sync); > + > noinstr int lock_is_held_type(const struct lockdep_map *lock, int read) > { > unsigned long flags; This patch looks good to me. Acked-by: Waiman Long <longman@redhat.com>
diff --git a/include/linux/lockdep.h b/include/linux/lockdep.h index 1f1099dac3f0..ba09df6a0872 100644 --- a/include/linux/lockdep.h +++ b/include/linux/lockdep.h @@ -268,6 +268,10 @@ extern void lock_acquire(struct lockdep_map *lock, unsigned int subclass, extern void lock_release(struct lockdep_map *lock, unsigned long ip); +extern void lock_sync(struct lockdep_map *lock, unsigned int subclass, + int read, int check, struct lockdep_map *nest_lock, + unsigned long ip); + /* lock_is_held_type() returns */ #define LOCK_STATE_UNKNOWN -1 #define LOCK_STATE_NOT_HELD 0 @@ -555,6 +559,7 @@ do { \ #define lock_map_acquire_read(l) lock_acquire_shared_recursive(l, 0, 0, NULL, _THIS_IP_) #define lock_map_acquire_tryread(l) lock_acquire_shared_recursive(l, 0, 1, NULL, _THIS_IP_) #define lock_map_release(l) lock_release(l, _THIS_IP_) +#define lock_map_sync(l) lock_sync(l, 0, 0, 1, NULL, _THIS_IP_) #ifdef CONFIG_PROVE_LOCKING # define might_lock(lock) \ diff --git a/kernel/locking/lockdep.c b/kernel/locking/lockdep.c index e3375bc40dad..cffa026a765f 100644 --- a/kernel/locking/lockdep.c +++ b/kernel/locking/lockdep.c @@ -5692,6 +5692,40 @@ void lock_release(struct lockdep_map *lock, unsigned long ip) } EXPORT_SYMBOL_GPL(lock_release); +/* + * lock_sync() - A special annotation for synchronize_{s,}rcu()-like API. + * + * No actual critical section is created by the APIs annotated with this: these + * APIs are used to wait for one or multiple critical sections (on other CPUs + * or threads), and it means that calling these APIs inside these critical + * sections is potential deadlock. + * + * This annotation acts as an acqurie+release anontation pair with hardirqoff + * being 1. Since there's no critical section, no interrupt can create extra + * dependencies "inside" the annotation, hardirqoff == 1 allows us to avoid + * false positives. + */ +void lock_sync(struct lockdep_map *lock, unsigned subclass, int read, + int check, struct lockdep_map *nest_lock, unsigned long ip) +{ + unsigned long flags; + + if (unlikely(!lockdep_enabled())) + return; + + raw_local_irq_save(flags); + check_flags(flags); + + lockdep_recursion_inc(); + __lock_acquire(lock, subclass, 0, read, check, 1, nest_lock, ip, 0, 0); + + if (__lock_release(lock, ip)) + check_chain_key(current); + lockdep_recursion_finish(); + raw_local_irq_restore(flags); +} +EXPORT_SYMBOL_GPL(lock_sync); + noinstr int lock_is_held_type(const struct lockdep_map *lock, int read) { unsigned long flags;
Currently, in order to annonate functions like synchronize_srcu() for lockdep, a trick as follow can be used: lock_acquire(); lock_release(); , which indicates synchronize_srcu() acts like an empty critical section that waits for other (read-side) critical sections to finish. This surely can catch some deadlock, but as discussion brought up by Paul Mckenney [1], this could introduce false positives because of irq-safe/unsafe detection. Extra tricks might help this: local_irq_disable(..); lock_acquire(); lock_release(); local_irq_enable(...); But it's better that lockdep could provide an annonation for synchronize_srcu() like functions, so that people won't need to repeat the ugly tricks above. Therefore introduce lock_sync(). It's simply an lock+unlock pair with no irq safe/unsafe deadlock check, since the to-be-annontated functions don't create real critical sections therefore there is no way that irq can create extra dependencies. [1]: https://lore.kernel.org/lkml/20180412021233.ewncg5jjuzjw3x62@tardis/ Signed-off-by: Boqun Feng <boqun.feng@gmail.com> --- include/linux/lockdep.h | 5 +++++ kernel/locking/lockdep.c | 34 ++++++++++++++++++++++++++++++++++ 2 files changed, 39 insertions(+)