xen/rcu: let rcu work better with core scheduling
v7 cover xen/rcu: let rcu work better with core scheduling
Today the RCU handling in Xen is affecting scheduling in several ways. It is raising sched softirqs without any real need and it requires tasklets for rcu_barrier(), which interacts badly with core scheduling.
This small series repairs those issues.
Additionally some ASSERT()s are added for verification of sane rcu handling. In order to avoid those triggering right away the obvious violations are fixed. This includes making rcu locking functions type safe.
Changes in V7:
- new patch 1
- added some barriers in patch 1
Changes in V6:
- added memory barrier in patch 1
- drop cpu_map_lock only at the end of rcu_barrier()
- re-add prempt_disable() in patch 3
Changes in V5:
- dropped already committed patches 1 and 4
- fixed race
- rework blocking of rcu processing with held rcu locks
Changes in V4:
- patch 5: use barrier()
Changes in V3:
- type safe locking functions (functions instead of macros)
- per-lock debug additions
- new patches 4 and 6
- fixed races
Changes in V2:
- use get_cpu_maps() in rcu_barrier() handling
- avoid recursion in rcu_barrier() handling
- new patches 3 and 4
Juergen Gross (5):
xen: introduce smp_mb__[after|before]_atomic() barriers xen/rcu: don't use stop_machine_run() for rcu_barrier()
xen: don't process rcu callbacks when holding a rcu_read_lock() xen/rcu: add assertions to debug build
xen/rcu: add per-lock counter in debug builds
xen/common/rcupdate.c | 102 +++++++++++++++++++++++++++++++------------ xen/common/softirq.c | 14 +++++- xen/include/asm-arm/system.h | 3 ++ xen/include/asm-x86/system.h | 3 ++ xen/include/xen/rcupdate.h | 77 ++++++++++++++++++++++++++------ 5 files changed, 157 insertions(+), 42 deletions(-)
Signature · 1 line
2.16.4
PATCH v7 1/5 xen: introduce smp_mb__[after|before]_atomic() barriers
When using atomic variables for synchronization barriers are needed to ensure proper data serialization. Introduce smp_mb__before_atomic() and smp_mb__after_atomic() as in the Linux kernel for that purpose.
Use the same definitions as in the Linux kernel.
---
V7:
- new patch
--- xen/include/asm-arm/system.h | 3 +++ xen/include/asm-x86/system.h | 3 +++ 2 files changed, 6 insertions(+)
xen/include/asm-arm/system.h +3 -0
@@ -30,6 +30,9 @@#define smp_wmb () dmb(ishst) +#define smp_mb__before_atomic () smp_mb() +#define smp_mb__after_atomic () smp_mb() +/* * This is used to ensure the compiler did actually allocate the register we * asked it for some inline assembly sequences. Apparently we can't trust
xen/include/asm-x86/system.h +3 -0
@@ -233,6 +233,9 @@ static always_inline unsigned long __xadd(#define set_mb (var, value) do { xchg(&var, value); } while (0)#define set_wmb (var, value) do { var = value; smp_wmb(); } while (0) +#define smp_mb__before_atomic () do { } while (0) +#define smp_mb__after_atomic () do { } while (0) +/** * array_index_mask_nospec() - generate a mask that is ~0UL when the * bounds check succeeds and 0 otherwise
Signature · 1 line
2.16.4
Re: [Xen-devel] [PATCH v7 1/5] xen: introduce smp_mb__[after|before]_atomic() barriers
Quoted 8 lines · Juergen Gross
When using atomic variables for synchronization barriers are needed to ensure proper data serialization. Introduce
smp_mb__before_atomic()andsmp_mb__after_atomic()as in the Linux kernel for that purpose.Use the same definitions as in the Linux kernel.
Suggested-by Jan Beulich <[email protected]>Signed-off-by Juergen Gross <[email protected]>
Re: [Xen-devel] [PATCH v7 1/5] xen: introduce smp_mb__[after|before]_atomic() barriers
In reply to Juergen GrossHi Juergen,
Quoted 8 lines · Juergen Gross
When using atomic variables for synchronization barriers are needed to ensure proper data serialization. Introduce
smp_mb__before_atomic()andsmp_mb__after_atomic()as in the Linux kernel for that purpose.Use the same definitions as in the Linux kernel.
Suggested-by Jan Beulich <[email protected]>Signed-off-by Juergen Gross <[email protected]>
Cheers,
Quoted 37 lines
---
V7:
- new patch--- xen/include/asm-arm/system.h | 3 +++ xen/include/asm-x86/system.h | 3 +++ 2 files changed, 6 insertions(+)xen/include/asm-arm/system.h +3
-0@@ -30,6 +30,9 @@#define smp_wmb () dmb(ishst) +#define smp_mb__before_atomic () smp_mb() +#define smp_mb__after_atomic () smp_mb() +/* * This is used to ensure the compiler did actually allocate the register we * asked it for some inline assembly sequences. Apparently we can't trust xen/include/asm-x86/system.h +3
-0@@ -233,6 +233,9 @@ static always_inline unsigned long __xadd(#define set_mb (var, value) do { xchg(&var, value); } while (0)#define set_wmb (var, value) do { var = value; smp_wmb(); } while (0) +#define smp_mb__before_atomic () do { } while (0) +#define smp_mb__after_atomic () do { } while (0) +/** * array_index_mask_nospec() - generate a mask that is ~0UL when the * bounds check succeeds and 0 otherwise
Signature · 1 line
Julien Grall
PATCH v7 2/5 xen/rcu: don't use stop_machine_run() for rcu_barrier()
In reply to Juergen GrossToday rcu_barrier() is calling stop_machine_run() to synchronize all physical cpus in order to ensure all pending rcu calls have finished when returning.
As stop_machine_run() is using tasklets this requires scheduling of idle vcpus on all cpus imposing the need to call rcu_barrier() on idle cpus only in case of core scheduling being active, as otherwise a scheduling deadlock would occur.
There is no need at all to do the syncing of the cpus in tasklets, as rcu activity is started in __do_softirq() called whenever softirq activity is allowed. So rcu_barrier() can easily be modified to use softirq for synchronization of the cpus no longer requiring any scheduling activity.
As there already is a rcu softirq reuse that for the synchronization.
Remove the barrier element from struct rcu_data as it isn't used.
Finally switch rcu_barrier() to return void as it now can never fail.
---
V2:
- add recursion detection
V3:
- fix races (Igor Druzhinin)
V5:
- rename done_count to pending_count (Jan Beulich)
- fix race (Jan Beulich)
V6:
- add barrier (Julien Grall)
- add ASSERT() (Julien Grall)
- hold cpu_map lock until end of rcu_barrier() (Julien Grall)
V7:
- update comment (Jan Beulich)
- add barriers (Jan Beulich)
--- xen/common/rcupdate.c | 100 +++++++++++++++++++++++++++++++++------------ xen/include/xen/rcupdate.h | 2 +- 2 files changed, 74 insertions(+), 28 deletions(-)
xen/common/rcupdate.c +73 -27
@@ -83,7 +83,6 @@ struct rcu_data {struct rcu_head **donetail;long blimit;/* Upper limit on a processed batch */ int cpu; -struct rcu_head barrier;long last_rs_qlen;/* qlen during the last resched */ /* 3) idle CPUs handling */
@@ -91,6 +90,7 @@ struct rcu_data {bool idle_timer_active;bool process_callbacks; +bool barrier_active; };/*
@@ -143,51 +143,90 @@ static int qhimark = 10000;static int qlowmark =100 ;static int rsinterval =1000 ; -struct rcu_barrier_data { -struct rcu_head head; -atomic_t *cpu_count; -}; +/* +* rcu_barrier() handling: +* Two counters are used to synchronize rcu_barrier() work: +* - cpu_count holds the number of cpus required to finish barrier handling. +* It is decremented by each cpu when it has performed all pending rcu calls. +* - pending_count shows whether any rcu_barrier() activity is running and +* it is used to synchronize leaving rcu_barrier() only after all cpus +* have finished their processing. pending_count is initialized to nr_cpus + 1 +* and it is decremented by each cpu when it has seen that cpu_count has +* reached 0. The cpu where rcu_barrier() has been called will wait until +* pending_count has been decremented to 1 (so all cpus have seen cpu_count +* reaching 0) and will then set pending_count to 0 indicating there is no +* rcu_barrier() running. +* Cpus are synchronized via softirq mechanism. rcu_barrier() is regarded to +* be active if pending_count is not zero. In case rcu_barrier() is called on +* multiple cpus it is enough to check for pending_count being not zero on entry +* and to call process_pending_softirqs() in a loop until pending_count drops to +* zero, before starting the new rcu_barrier() processing. +*/ +static atomic_t cpu_count =ATOMIC_INIT (0 ); +static atomic_t pending_count =ATOMIC_INIT (0 );static void rcu_barrier_callback (struct rcu_head *head) { -struct rcu_barrier_data *data =container_of ( - head, struct rcu_barrier_data, head); -atomic_inc (data->cpu_count); +smp_mb__before_atomic ();/* Make all writes visible to other cpus. */ +atomic_dec (&cpu_count); } -static int rcu_barrier_action (void *_cpu_count) +static void rcu_barrier_action (void ) { -struct rcu_barrier_data data = { .cpu_count = _cpu_count }; - -ASSERT (!local_irq_is_enabled ()); -local_irq_enable (); +struct rcu_head head;/* * When callback is executed, all previously-queued RCU work on this CPU -* is completed. When all CPUs have executed their callback, data.cpu_count -* will have been incremented to include every online CPU. +* is completed. When all CPUs have executed their callback, cpu_count +* will have been decremented to 0. */ -call_rcu (&data.head, rcu_barrier_callback); +call_rcu (&head, rcu_barrier_callback); -while (atomic_read (data.cpu_count) !=num_online_cpus () ) +while (atomic_read (&cpu_count) ) {process_pending_softirqs ();cpu_relax (); } -local_irq_disable (); - -return 0 ; +smp_mb__before_atomic (); +atomic_dec (&pending_count); } -/* -* As rcu_barrier() is using stop_machine_run() it is allowed to be used in -* idle context only (see comment for stop_machine_run()). -*/ -int rcu_barrier (void ) +void rcu_barrier (void ) { -atomic_t cpu_count =ATOMIC_INIT (0 ); -return stop_machine_run (rcu_barrier_action, &cpu_count,NR_CPUS ); +unsigned int n_cpus; + +ASSERT (!in_irq () &&local_irq_is_enabled ()); + +for ( ; ; ) + { +if ( !atomic_read (&pending_count) &&get_cpu_maps () ) + { + n_cpus =num_online_cpus (); + +if (atomic_cmpxchg (&pending_count,0 , n_cpus +1 ) ==0 ) +break ; + +put_cpu_maps (); + } + +process_pending_softirqs (); +cpu_relax (); + } + +smp_mb__before_atomic (); +atomic_set (&cpu_count, n_cpus); +cpumask_raise_softirq (&cpu_online_map,RCU_SOFTIRQ ); + +while (atomic_read (&pending_count) !=1 ) + { +process_pending_softirqs (); +cpu_relax (); + } + +atomic_set (&pending_count,0 ); + +put_cpu_maps (); }/* Is batch a before batch b ? */
@@ -426,6 +465,13 @@ static void rcu_process_callbacks(void) rdp->process_callbacks = false;__rcu_process_callbacks (&rcu_ctrlblk, rdp); } + +if (atomic_read (&cpu_count) && !rdp->barrier_active ) + { + rdp->barrier_active = true; +rcu_barrier_action (); + rdp->barrier_active = false; + } }static int __rcu_pending (struct rcu_ctrlblk *rcp,struct rcu_data *rdp)
xen/include/xen/rcupdate.h +1 -1
@@ -144,7 +144,7 @@ void rcu_check_callbacks(int cpu);void call_rcu (struct rcu_head *head,void (*func)(struct rcu_head *head)); -int rcu_barrier (void ); +void rcu_barrier (void );void rcu_idle_enter (unsigned int cpu);void rcu_idle_exit (unsigned int cpu);
Signature · 1 line
2.16.4
Re: [Xen-devel] [PATCH v7 2/5] xen/rcu: don't use stop_machine_run() for rcu_barrier()
Quoted 15 lines · Juergen Gross
Today
rcu_barrier()is callingstop_machine_run()to synchronize all physical cpus in order to ensure all pending rcu calls have finished when returning.As
stop_machine_run()is using tasklets this requires scheduling of idle vcpus on all cpus imposing the need to callrcu_barrier()on idle cpus only in case of core scheduling being active, as otherwise a scheduling deadlock would occur.There is no need at all to do the syncing of the cpus in tasklets, as rcu activity is started in
__do_softirq()called whenever softirq activity is allowed. Sorcu_barrier()can easily be modified to use softirq for synchronization of the cpus no longer requiring any scheduling activity.
Quoted 8 lines
As there already is a rcu softirq reuse that for the synchronization.
Remove the barrier element from struct
rcu_dataas it isn't used.Finally switch
rcu_barrier()to return void as it now can never fail.Partially-based-on-patch-by Igor Druzhinin <[email protected]>Signed-off-by Juergen Gross <[email protected]>
Re: [Xen-devel] [PATCH v7 2/5] xen/rcu: don't use stop_machine_run() for rcu_barrier()
In reply to Juergen GrossHi Juergen,
Quoted 93 lines · Juergen Gross
Today
rcu_barrier()is callingstop_machine_run()to synchronize all physical cpus in order to ensure all pending rcu calls have finished when returning.As
stop_machine_run()is using tasklets this requires scheduling of idle vcpus on all cpus imposing the need to callrcu_barrier()on idle cpus only in case of core scheduling being active, as otherwise a scheduling deadlock would occur.There is no need at all to do the syncing of the cpus in tasklets, as rcu activity is started in
__do_softirq()called whenever softirq activity is allowed. Sorcu_barrier()can easily be modified to use softirq for synchronization of the cpus no longer requiring any scheduling activity.As there already is a rcu softirq reuse that for the synchronization.
Remove the barrier element from struct
rcu_dataas it isn't used.Finally switch
rcu_barrier()to return void as it now can never fail.Partially-based-on-patch-by Igor Druzhinin <[email protected]>Signed-off-by Juergen Gross <[email protected]>---
V2:
- add recursion detectionV3:
- fix races (Igor Druzhinin)V5:
- renamedone_counttopending_count(Jan Beulich)
- fix race (Jan Beulich)V6:
- add barrier (Julien Grall)
- addASSERT()(Julien Grall)
- holdcpu_maplock until end ofrcu_barrier()(Julien Grall)V7:
- update comment (Jan Beulich)
- add barriers (Jan Beulich)--- xen/common/rcupdate.c | 100 +++++++++++++++++++++++++++++++++------------ xen/include/xen/rcupdate.h | 2 +- 2 files changed, 74 insertions(+), 28 deletions(-)xen/common/rcupdate.c +19
-5@@ -83,7 +83,6 @@ struct rcu_data {struct rcu_head **donetail;long blimit;/* Upper limit on a processed batch */ int cpu; -struct rcu_head barrier;long last_rs_qlen;/* qlen during the last resched */ /* 3) idle CPUs handling */ @@ -91,6 +90,7 @@ struct rcu_data {bool idle_timer_active;bool process_callbacks; +bool barrier_active; };/* @@ -143,51 +143,90 @@ static int qhimark = 10000;static int qlowmark =100 ;static int rsinterval =1000 ; -struct rcu_barrier_data { -struct rcu_head head; -atomic_t *cpu_count; -}; +/* +* rcu_barrier() handling: +* Two counters are used to synchronize rcu_barrier() work: +* - cpu_count holds the number of cpus required to finish barrier handling. +* It is decremented by each cpu when it has performed all pending rcu calls. +* - pending_count shows whether any rcu_barrier() activity is running and +* it is used to synchronize leaving rcu_barrier() only after all cpus +* have finished their processing. pending_count is initialized to nr_cpus + 1 +* and it is decremented by each cpu when it has seen that cpu_count has +* reached 0. The cpu where rcu_barrier() has been called will wait until +* pending_count has been decremented to 1 (so all cpus have seen cpu_count +* reaching 0) and will then set pending_count to 0 indicating there is no +* rcu_barrier() running. +* Cpus are synchronized via softirq mechanism. rcu_barrier() is regarded to +* be active if pending_count is not zero. In case rcu_barrier() is called on +* multiple cpus it is enough to check for pending_count being not zero on entry +* and to call process_pending_softirqs() in a loop until pending_count drops to +* zero, before starting the new rcu_barrier() processing.
xen/common/rcupdate.c static int qhimark = 10000; · line 170
Quoted 10 lines
xen/common/rcupdate.c +4
-3+*/ +static atomic_t cpu_count =ATOMIC_INIT (0 ); +static atomic_t pending_count =ATOMIC_INIT (0 );static void rcu_barrier_callback (struct rcu_head *head) { -struct rcu_barrier_data *data =container_of ( - head, struct rcu_barrier_data, head); -atomic_inc (data->cpu_count); +smp_mb__before_atomic ();/* Make all writes visible to other cpus. */
smp_mb__before_atomic() will order both read and write. However, the comment suggest only the write are required to be ordered.
So either the barrier is too strong or the comment is incorrect. Can you clarify it?
Quoted 56 lines
xen/common/rcupdate.c +21
-19+atomic_dec (&cpu_count); } -static int rcu_barrier_action (void *_cpu_count) +static void rcu_barrier_action (void ) { -struct rcu_barrier_data data = { .cpu_count = _cpu_count }; - -ASSERT (!local_irq_is_enabled ()); -local_irq_enable (); +struct rcu_head head;/* * When callback is executed, all previously-queued RCU work on this CPU -* is completed. When all CPUs have executed their callback, data.cpu_count -* will have been incremented to include every online CPU. +* is completed. When all CPUs have executed their callback, cpu_count +* will have been decremented to 0. */ -call_rcu (&data.head, rcu_barrier_callback); +call_rcu (&head, rcu_barrier_callback); -while (atomic_read (data.cpu_count) !=num_online_cpus () ) +while (atomic_read (&cpu_count) ) {process_pending_softirqs ();cpu_relax (); } -local_irq_disable (); - -return 0 ; +smp_mb__before_atomic (); +atomic_dec (&pending_count); } -/* -* As rcu_barrier() is using stop_machine_run() it is allowed to be used in -* idle context only (see comment for stop_machine_run()). -*/ -int rcu_barrier (void ) +void rcu_barrier (void ) { -atomic_t cpu_count =ATOMIC_INIT (0 ); -return stop_machine_run (rcu_barrier_action, &cpu_count,NR_CPUS ); +unsigned int n_cpus; + +ASSERT (!in_irq () &&local_irq_is_enabled ()); + +for ( ; ; ) + { +if ( !atomic_read (&pending_count) &&get_cpu_maps () ) + { + n_cpus =num_online_cpus (); + +if (atomic_cmpxchg (&pending_count,0 , n_cpus +1 ) ==0 )
xen/common/rcupdate.c static int qhimark = 10000; · line 217
Quoted 10 lines
xen/common/rcupdate.c +10
-0+break ; + +put_cpu_maps (); + } + +process_pending_softirqs (); +cpu_relax (); + } + +smp_mb__before_atomic ();
Our semantic of atomic_cmpxchg() is exactly the same as Linux. I.e it will contain a full barrier when the cmpxchg succeed. So why do you need this barrier?
Quoted 35 lines
xen/common/rcupdate.c +19
-0+atomic_set (&cpu_count, n_cpus); +cpumask_raise_softirq (&cpu_online_map,RCU_SOFTIRQ ); + +while (atomic_read (&pending_count) !=1 ) + { +process_pending_softirqs (); +cpu_relax (); + } + +atomic_set (&pending_count,0 ); + +put_cpu_maps (); }/* Is batch a before batch b ? */ @@ -426,6 +465,13 @@ static void rcu_process_callbacks(void) rdp->process_callbacks = false;__rcu_process_callbacks (&rcu_ctrlblk, rdp); } + +if (atomic_read (&cpu_count) && !rdp->barrier_active ) + { + rdp->barrier_active = true; +rcu_barrier_action (); + rdp->barrier_active = false; + } }static int __rcu_pending (struct rcu_ctrlblk *rcp,struct rcu_data *rdp)xen/include/xen/rcupdate.h +0
-0
xen/include/xen/rcupdate.h void rcu_check_callbacks(int cpu); · line 150
Quoted 8 lines
xen/include/xen/rcupdate.h +1
-1@@ -144,7 +144,7 @@ void rcu_check_callbacks(int cpu);void call_rcu (struct rcu_head *head,void (*func)(struct rcu_head *head)); -int rcu_barrier (void ); +void rcu_barrier (void );void rcu_idle_enter (unsigned int cpu);void rcu_idle_exit (unsigned int cpu);
Cheers,
Signature · 1 line
Julien Grall
Re: [Xen-devel] [PATCH v7 2/5] xen/rcu: don't use stop_machine_run() for rcu_barrier()
Quoted 39 lines · Julien Grall
Quoted 24 lines · Juergen Gross
xen/common/rcupdate.c +17
-3@@ -143,51 +143,90 @@ static int qhimark = 10000; static int qlowmark = 100; static int rsinterval = 1000; -struct rcu_barrier_data { - struct rcu_head head; - atomic_t *cpu_count; -}; +/* +* rcu_barrier() handling: +* Two counters are used to synchronize rcu_barrier() work: +* - cpu_count holds the number of cpus required to finish barrier handling. +* It is decremented by each cpu when it has performed all pending rcu calls. +* - pending_count shows whether any rcu_barrier() activity is running and +* it is used to synchronize leaving rcu_barrier() only after all cpus +* have finished their processing. pending_count is initialized to nr_cpus + 1 +* and it is decremented by each cpu when it has seen that cpu_count has +* reached 0. The cpu where rcu_barrier() has been called will wait until +* pending_count has been decremented to 1 (so all cpus have seen cpu_count +* reaching 0) and will then set pending_count to 0 indicating there is no +* rcu_barrier() running. +* Cpus are synchronized via softirq mechanism. rcu_barrier() is regarded to +* be active if pending_count is not zero. In case rcu_barrier() is called on +* multiple cpus it is enough to check for pending_count being not zero on entry +* and to call process_pending_softirqs() in a loop until pending_count drops to xen/common/rcupdate.c static int qhimark = 10000; · line 169
Quoted 10 lines
xen/common/rcupdate.c +5
-3+* zero, before starting the new rcu_barrier() processing. +*/ +static atomic_t cpu_count =ATOMIC_INIT (0 ); +static atomic_t pending_count =ATOMIC_INIT (0 ); static void rcu_barrier_callback(struct rcu_head *head) { - struct rcu_barrier_data *data = container_of( - head, struct rcu_barrier_data, head); - atomic_inc(data->cpu_count); + smp_mb__before_atomic(); /* Make all writes visible to other cpus. */
smp_mb__before_atomic()will order both read and write. However, the comment suggest only the write are required to be ordered.
Quoted 2 lines
So either the barrier is too strong or the comment is incorrect. Can you clarify it?
Neither is the case, I guess: There simply is no smp_wmb__before_atomic() in Linux, and if we want to follow their model we shouldn't have one either. I'd rather take the comment to indicate that if one appeared, it could be used here.
Quoted 56 lines
Quoted 56 lines
xen/common/rcupdate.c +9
-9+ atomic_dec(&cpu_count); } -static int rcu_barrier_action(void *_cpu_count) +static voidrcu_barrier_action (void) { - struct rcu_barrier_data data = { .cpu_count = _cpu_count }; - - ASSERT(!local_irq_is_enabled()); - local_irq_enable(); + struct rcu_head head; /* * When callback is executed, all previously-queued RCU work on this CPU - * is completed. When all CPUs have executed their callback, data.cpu_count - * will have been incremented to include every online CPU. + * is completed. When all CPUs have executed their callback, cpu_count + * will have been decremented to 0. */ - call_rcu(&data.head, rcu_barrier_callback); + call_rcu(&head, rcu_barrier_callback); - while ( atomic_read(data.cpu_count) != num_online_cpus() ) + while ( atomic_read(&cpu_count) ) { process_pending_softirqs(); cpu_relax(); } - local_irq_disable(); - - return 0; + smp_mb__before_atomic(); + atomic_dec(&pending_count); }-/*
- * Asrcu_barrier()is usingstop_machine_run()it is allowed to be used in
- * idle context only (see comment forstop_machine_run()).
- */xen/common/rcupdate.c +1
-1-int rcu_barrier (void ) +void rcu_barrier (void ){
-atomic_tcpu_count=ATOMIC_INIT(0);
- returnstop_machine_run(rcu_barrier_action, &cpu_count, NR_CPUS); + unsigned intn_cpus;xen/common/rcupdate.c +15
-0+ + ASSERT(!in_irq() && local_irq_is_enabled()); + + for ( ; ; ) + { + if ( !atomic_read(&pending_count) && get_cpu_maps() ) + { + n_cpus = num_online_cpus(); + + if ( atomic_cmpxchg(&pending_count, 0, n_cpus + 1) == 0 ) + break; + + put_cpu_maps(); + } +
Quoted 8 lines
Quoted 5 lines
+
process_pending_softirqs();
+cpu_relax();
+ }
+
+smp_mb__before_atomic();Our semantic of
atomic_cmpxchg()is exactly the same as Linux. I.e
it will contain a full barrier when the cmpxchg succeed. So why do you need this barrier?
I was me I think to have (wrongly) suggested a barrier was missing here, sorry.
Jan
Re: [Xen-devel] [PATCH v7 2/5] xen/rcu: don't use stop_machine_run() for rcu_barrier()
Quoted 40 lines · Jan Beulich
Quoted 39 lines · Julien Grall
Quoted 24 lines · Juergen Gross
xen/common/rcupdate.c +17
-3@@ -143,51 +143,90 @@ static int qhimark = 10000; static int qlowmark = 100; static int rsinterval = 1000; -struct rcu_barrier_data { - struct rcu_head head; - atomic_t *cpu_count; -}; +/* +* rcu_barrier() handling: +* Two counters are used to synchronize rcu_barrier() work: +* - cpu_count holds the number of cpus required to finish barrier handling. +* It is decremented by each cpu when it has performed all pending rcu calls. +* - pending_count shows whether any rcu_barrier() activity is running and +* it is used to synchronize leaving rcu_barrier() only after all cpus +* have finished their processing. pending_count is initialized to nr_cpus + 1 +* and it is decremented by each cpu when it has seen that cpu_count has +* reached 0. The cpu where rcu_barrier() has been called will wait until +* pending_count has been decremented to 1 (so all cpus have seen cpu_count +* reaching 0) and will then set pending_count to 0 indicating there is no +* rcu_barrier() running. +* Cpus are synchronized via softirq mechanism. rcu_barrier() is regarded to +* be active if pending_count is not zero. In case rcu_barrier() is called on +* multiple cpus it is enough to check for pending_count being not zero on entry +* and to call process_pending_softirqs() in a loop until pending_count drops to xen/common/rcupdate.c static int qhimark = 10000; · line 169
Quoted 10 lines
xen/common/rcupdate.c +5
-3+* zero, before starting the new rcu_barrier() processing. +*/ +static atomic_t cpu_count =ATOMIC_INIT (0 ); +static atomic_t pending_count =ATOMIC_INIT (0 ); static void rcu_barrier_callback(struct rcu_head *head) { - struct rcu_barrier_data *data = container_of( - head, struct rcu_barrier_data, head); - atomic_inc(data->cpu_count); + smp_mb__before_atomic(); /* Make all writes visible to other cpus. */
smp_mb__before_atomic()will order both read and write. However, the comment suggest only the write are required to be ordered.
Quoted 7 lines
Quoted 2 lines
So either the barrier is too strong or the comment is incorrect. Can you clarify it?
Neither is the case, I guess: There simply is no
smp_wmb__before_atomic()in Linux, and if we want to follow their model we shouldn't have one either. I'd rather take the comment to indicate that if one appeared, it could be used here.
Right. Currently we have the choice of either using smp_mb__before_atomic() which is too strong for Arm, or smp_wmb() which is too strong for x86.
Quoted 66 lines
Quoted 57 lines
Quoted 57 lines
xen/common/rcupdate.c +26
-15+ atomic_dec(&cpu_count); } -static int rcu_barrier_action(void *_cpu_count) +static voidrcu_barrier_action (void) { - struct rcu_barrier_data data = { .cpu_count = _cpu_count }; - - ASSERT(!local_irq_is_enabled()); - local_irq_enable(); + struct rcu_head head; /* * When callback is executed, all previously-queued RCU work on this CPU - * is completed. When all CPUs have executed their callback, data.cpu_count - * will have been incremented to include every online CPU. + * is completed. When all CPUs have executed their callback, cpu_count + * will have been decremented to 0. */ - call_rcu(&data.head, rcu_barrier_callback); + call_rcu(&head, rcu_barrier_callback); - while ( atomic_read(data.cpu_count) != num_online_cpus() ) + while ( atomic_read(&cpu_count) ) { process_pending_softirqs(); cpu_relax(); } - local_irq_disable(); - - return 0; + smp_mb__before_atomic(); + atomic_dec(&pending_count); } -/* - * As rcu_barrier() is using stop_machine_run() it is allowed to be used in - * idle context only (see comment for stop_machine_run()). - */ -int rcu_barrier(void) +void rcu_barrier(void) { - atomic_t cpu_count = ATOMIC_INIT(0); - return stop_machine_run(rcu_barrier_action, &cpu_count, NR_CPUS); +unsigned int n_cpus; + +ASSERT(!in_irq() && local_irq_is_enabled()); + +for ( ; ; ) +{ +if ( !atomic_read(&pending_count) && get_cpu_maps() ) +{ +n_cpus = num_online_cpus(); + +if ( atomic_cmpxchg(&pending_count, 0, n_cpus + 1) == 0 ) +break; + +put_cpu_maps(); +} +Quoted 8 lines
xen/common/rcupdate.c static int qhimark = 10000; · line 217
Quoted 5 lines
xen/common/rcupdate.c +5
-0+ process_pending_softirqs(); + cpu_relax(); + } + + smp_mb__before_atomic();Our semantic of
atomic_cmpxchg()is exactly the same as Linux. I.e
it will contain a full barrier when the cmpxchg succeed. So why do you need this barrier?
Quoted 2 lines
I was me I think to have (wrongly) suggested a barrier was missing here, sorry.
I'll update the patch dropping the barrier.
Juergen
Re: [Xen-devel] [PATCH v7 2/5] xen/rcu: don't use stop_machine_run() for rcu_barrier()
Quoted 44 lines · Jürgen Groß
Quoted 43 lines · Jan Beulich
Quoted 41 lines · Julien Grall
Quoted 24 lines · Juergen Gross
xen/common/rcupdate.c +17
-3@@ -143,51 +143,90 @@ static int qhimark = 10000; static int qlowmark = 100; static int rsinterval = 1000; -struct rcu_barrier_data { - struct rcu_head head; - atomic_t *cpu_count; -}; +/* +* rcu_barrier() handling: +* Two counters are used to synchronize rcu_barrier() work: +* - cpu_count holds the number of cpus required to finish barrier handling. +* It is decremented by each cpu when it has performed all pending rcu calls. +* - pending_count shows whether any rcu_barrier() activity is running and +* it is used to synchronize leaving rcu_barrier() only after all cpus +* have finished their processing. pending_count is initialized to nr_cpus + 1 +* and it is decremented by each cpu when it has seen that cpu_count has +* reached 0. The cpu where rcu_barrier() has been called will wait until +* pending_count has been decremented to 1 (so all cpus have seen cpu_count +* reaching 0) and will then set pending_count to 0 indicating there is no +* rcu_barrier() running. +* Cpus are synchronized via softirq mechanism. rcu_barrier() is regarded to +* be active if pending_count is not zero. In case rcu_barrier() is called on +* multiple cpus it is enough to check for pending_count being not zero on entry +* and to call process_pending_softirqs() in a loop until pending_count drops to xen/common/rcupdate.c static int qhimark = 10000; · line 169
Quoted 10 lines
xen/common/rcupdate.c +5
-3+* zero, before starting the new rcu_barrier() processing. +*/ +static atomic_t cpu_count =ATOMIC_INIT (0 ); +static atomic_t pending_count =ATOMIC_INIT (0 ); static void rcu_barrier_callback(struct rcu_head *head) { - struct rcu_barrier_data *data = container_of( - head, struct rcu_barrier_data, head); - atomic_inc(data->cpu_count); + smp_mb__before_atomic(); /* Make all writes visible to other cpus. */
smp_mb__before_atomic()will order both read and write. However, the comment suggest only the write are required to be ordered.So either the barrier is too strong or the comment is incorrect. Can you clarify it?
Quoted 8 lines
Quoted 4 lines
Neither is the case, I guess: There simply is no
smp_wmb__before_atomic()in Linux, and if we want to follow their model we shouldn't have one either. I'd rather take the comment to indicate that if one appeared, it could be used here.Right. Currently we have the choice of either using
smp_mb__before_atomic()which is too strong for Arm, orsmp_wmb()which is too strong for x86.
For x86 smp_wmb() is actually only very slightly too strong - it expands to just barrier(), after all. So overall perhaps that's the better choice here (with a suitable comment)?
Jan
Re: [Xen-devel] [PATCH v7 2/5] xen/rcu: don't use stop_machine_run() for rcu_barrier()
Quoted 50 lines · Jan Beulich
Quoted 49 lines · Jürgen Groß
Quoted 47 lines · Jan Beulich
Quoted 39 lines · Julien Grall
Quoted 24 lines · Juergen Gross
xen/common/rcupdate.c +17
-3@@ -143,51 +143,90 @@ static int qhimark = 10000; static int qlowmark = 100; static int rsinterval = 1000; -struct rcu_barrier_data { - struct rcu_head head; - atomic_t *cpu_count; -}; +/* +* rcu_barrier() handling: +* Two counters are used to synchronize rcu_barrier() work: +* - cpu_count holds the number of cpus required to finish barrier handling. +* It is decremented by each cpu when it has performed all pending rcu calls. +* - pending_count shows whether any rcu_barrier() activity is running and +* it is used to synchronize leaving rcu_barrier() only after all cpus +* have finished their processing. pending_count is initialized to nr_cpus + 1 +* and it is decremented by each cpu when it has seen that cpu_count has +* reached 0. The cpu where rcu_barrier() has been called will wait until +* pending_count has been decremented to 1 (so all cpus have seen cpu_count +* reaching 0) and will then set pending_count to 0 indicating there is no +* rcu_barrier() running. +* Cpus are synchronized via softirq mechanism. rcu_barrier() is regarded to +* be active if pending_count is not zero. In case rcu_barrier() is called on +* multiple cpus it is enough to check for pending_count being not zero on entry +* and to call process_pending_softirqs() in a loop until pending_count drops to xen/common/rcupdate.c static int qhimark = 10000; · line 169
Quoted 10 lines
xen/common/rcupdate.c +5
-3+* zero, before starting the new rcu_barrier() processing. +*/ +static atomic_t cpu_count =ATOMIC_INIT (0 ); +static atomic_t pending_count =ATOMIC_INIT (0 ); static void rcu_barrier_callback(struct rcu_head *head) { - struct rcu_barrier_data *data = container_of( - head, struct rcu_barrier_data, head); - atomic_inc(data->cpu_count); + smp_mb__before_atomic(); /* Make all writes visible to other cpus. */
smp_mb__before_atomic()will order both read and write. However, the comment suggest only the write are required to be ordered.Quoted 2 lines
So either the barrier is too strong or the comment is incorrect. Can you clarify it?
Neither is the case, I guess: There simply is no
smp_wmb__before_atomic()in Linux, and if we want to follow their model we shouldn't have one either. I'd rather take the comment to indicate that if one appeared, it could be used here.
Quoted 7 lines
Quoted 3 lines
Right. Currently we have the choice of either using
smp_mb__before_atomic()which is too strong for Arm, orsmp_wmb()which is too strong for x86.For x86
smp_wmb()is actually only very slightly too strong - it expands to justbarrier(), after all. So overall perhaps that's the better choice here (with a suitable comment)?
Fine with me.
Juergen
Re: [Xen-devel] [PATCH v7 2/5] xen/rcu: don't use stop_machine_run() for rcu_barrier()
Quoted 69 lines · Jürgen Groß
Quoted 68 lines · Jan Beulich
Quoted 66 lines · Jürgen Groß
Quoted 54 lines · Jan Beulich
Quoted 53 lines · Julien Grall
Quoted 41 lines · Juergen Gross
xen/common/rcupdate.c +4
-3@@ -143,51 +143,90 @@ static int qhimark = 10000; static int qlowmark = 100; static int rsinterval = 1000; -struct rcu_barrier_data { - struct rcu_head head; - atomic_t *cpu_count; -}; +/* +* rcu_barrier() handling: +* Two counters are used to synchronize rcu_barrier() work: +* - cpu_count holds the number of cpus required to finish barrier handling.
+ * It is decremented by each cpu when it has performed all pending rcu calls.
+ * -pending_countshows whether anyrcu_barrier()activity is running and
+ * it is used to synchronize leavingrcu_barrier()only after all cpus
+ * have finished their processing.pending_countis initialized tonr_cpus+ 1
+ * and it is decremented by each cpu when it has seen thatcpu_counthas
+ * reached 0. The cpu wherercu_barrier()has been called will wait until
+ *pending_counthas been decremented to 1 (so all cpus have seencpu_count
+ * reaching 0) and will then setpending_countto 0 indicating there is no
+ *rcu_barrier()running.
+ * Cpus are synchronized via softirq mechanism.rcu_barrier()is regarded to
+ * be active ifpending_countis not zero. In casercu_barrier()is called on
+ * multiple cpus it is enough to check forpending_countbeing not zero on entry
+ * and to callprocess_pending_softirqs()in a loop untilpending_countdrops to
+ * zero, before starting the newrcu_barrier()processing.
+ */xen/common/rcupdate.c +2
-0+static atomic_t cpu_count =ATOMIC_INIT (0 ); +static atomic_t pending_count =ATOMIC_INIT (0 );Quoted 7 lines
static void
rcu_barrier_callback(struct rcu_head *head){
- structrcu_barrier_data*data = container_of(
- head, structrcu_barrier_data, head);
-atomic_inc(data->cpu_count);
+ smp_mb__before_atomic(); /* Make all writes visible to other cpus. */
smp_mb__before_atomic()will order both read and write. However, the comment suggest only the write are required to be ordered.Quoted 7 lines
Quoted 2 lines
So either the barrier is too strong or the comment is incorrect. Can you clarify it?
Neither is the case, I guess: There simply is no
smp_wmb__before_atomic()in Linux, and if we want to follow their model we shouldn't have one either. I'd rather take the comment to indicate that if one appeared, it could be used here.Right. Currently we have the choice of either using
smp_mb__before_atomic()which is too strong for Arm, orsmp_wmb()which is too strong for x86.
Quoted 5 lines
Quoted 3 lines
For x86
smp_wmb()is actually only very slightly too strong - it expands to justbarrier(), after all. So overall perhaps that's the better choice here (with a suitable comment)?Fine with me.
I am happy with that.
Cheers,
Signature · 1 line
Julien Grall
PATCH v7 3/5 xen: don't process rcu callbacks when holding a rcu_read_lock()
In reply to Juergen GrossSome keyhandlers are calling process_pending_softirqs() while holding a rcu_read_lock(). This is wrong, as process_pending_softirqs() might activate rcu calls which should not happen inside a rcu_read_lock().
For that purpose modify process_pending_softirqs() to not allow rcu callback processing when a rcu_read_lock() is being held.
---
V3:
- add RCU_SOFTIRQ to ignore in process_pending_softirqs_norcu() (Roger Pau Monné)
V5:
- block rcu processing depending on rch_read_lock() being held or not (Jan Beulich)
--- xen/common/softirq.c | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-)
xen/common/softirq.c +10 -2
@@ -29,6 +29,7 @@ static void __do_softirq(unsigned long ignore_mask) {unsigned int i, cpu;unsigned long pending; +bool rcu_allowed = !(ignore_mask & (1ul <<RCU_SOFTIRQ ));for ( ; ; ) {
@@ -38,7 +39,7 @@ static void __do_softirq(unsigned long ignore_mask)*/ cpu =smp_processor_id (); -if (rcu_pending (cpu) ) +if ( rcu_allowed &&rcu_pending (cpu) )rcu_check_callbacks (cpu); if ( ((pending = (softirq_pending(cpu) & ~ignore_mask)) == 0)
@@ -53,9 +54,16 @@ static void __do_softirq(unsigned long ignore_mask)void process_pending_softirqs (void ) { +unsigned long ignore_mask = (1ul <<SCHEDULE_SOFTIRQ ) | + (1ul <<SCHED_SLAVE_SOFTIRQ ); + +/* Block RCU processing in case of rcu_read_lock() held. */ +if (preempt_count () ) + ignore_mask |=1ul <<RCU_SOFTIRQ ; +ASSERT (!in_irq () &&local_irq_is_enabled ());/* Do not enter scheduler as it can preempt the calling context. */ -__do_softirq ((1ul <<SCHEDULE_SOFTIRQ ) | (1ul <<SCHED_SLAVE_SOFTIRQ )); +__do_softirq (ignore_mask); }void do_softirq (void )
Signature · 1 line
2.16.4
PATCH v7 4/5 xen/rcu: add assertions to debug build
In reply to Juergen GrossXen's RCU implementation relies on no softirq handling taking place while being in a RCU critical section. Add ASSERT()s in debug builds in order to catch any violations.
For that purpose modify rcu_read_[un]lock() to use a dedicated percpu counter additional to preempt_[en|dis]able() as this enables to test that condition in __do_softirq() (ASSERT_NOT_IN_ATOMIC() is not usable there due to __cpu_up() calling process_pending_softirqs() while holding the cpu hotplug lock).
While at it switch the rcu_read_[un]lock() implementation to static inline functions instead of macros.
---
V3:
- add barriers to rcu_[en|dis]able() (Roger Pau Monné)
- add rcu_quiesce_allowed() to ASSERT_NOT_IN_ATOMIC (Roger Pau Monné)
- convert macros to static inline functions
- add sanity check in rcu_read_unlock()
V4:
- use barrier() in rcu_[en|dis]able() (Julien Grall)
V5:
- use rcu counter even if not using a debug build
V6:
- keep preempt_[dis|en]able() calls
--- xen/common/rcupdate.c | 2 ++ xen/common/softirq.c | 4 +++- xen/include/xen/rcupdate.h | 37 ++++++++++++++++++++++++++++++++++--- 3 files changed, 39 insertions(+), 4 deletions(-)
xen/common/rcupdate.c +2 -0
@@ -46,6 +46,8 @@#include <xen/cpu.h> #include <xen/stop_machine.h> +DEFINE_PER_CPU(unsigned int, rcu_lock_cnt); +/* Global control variables for rcupdate callback mechanism. */ static struct rcu_ctrlblk { long cur; /* Current batch number. */
xen/common/softirq.c +3 -1
@@ -31,6 +31,8 @@ static void __do_softirq(unsigned long ignore_mask)unsigned long pending;bool rcu_allowed = !(ignore_mask & (1ul <<RCU_SOFTIRQ )); +ASSERT (!rcu_allowed ||rcu_quiesce_allowed ()); +for ( ; ; ) {/*
@@ -58,7 +60,7 @@ void process_pending_softirqs(void) (1ul <<SCHED_SLAVE_SOFTIRQ );/* Block RCU processing in case of rcu_read_lock() held. */ -if (preempt_count () ) +if ( !rcu_quiesce_allowed () ) ignore_mask |=1ul <<RCU_SOFTIRQ ;ASSERT (!in_irq () &&local_irq_is_enabled ());
xen/include/xen/rcupdate.h +34 -3
@@ -32,12 +32,35 @@#define __XEN_RCUPDATE_H#include <xen/cache.h> +#include <xen/compiler.h> #include <xen/spinlock.h> #include <xen/cpumask.h> +#include <xen/percpu.h> #include <xen/preempt.h> #define __rcu +DECLARE_PER_CPU(unsigned int, rcu_lock_cnt); + +static inline void rcu_quiesce_disable (void ) +{ +preempt_disable (); +this_cpu (rcu_lock_cnt)++; +barrier (); +} + +static inline void rcu_quiesce_enable (void ) +{ +barrier (); +this_cpu (rcu_lock_cnt)--; +preempt_enable (); +} + +static inline bool rcu_quiesce_allowed (void ) +{ +return !this_cpu (rcu_lock_cnt); +} +/** * struct rcu_head - callback structure for use with RCU * @next: next update requests in a list
@@ -91,16 +114,24 @@ typedef struct _rcu_read_lock rcu_read_lock_t;* will be deferred until the outermost RCU read-side critical section * completes. * -* It is illegal to block while in an RCU read-side critical section. +* It is illegal to process softirqs or block while in an RCU read-side +* critical section. */ -#define rcu_read_lock (x) ({ ((void)(x)); preempt_disable(); }) +static inline void rcu_read_lock (rcu_read_lock_t *lock) +{ +rcu_quiesce_disable (); +}/** * rcu_read_unlock - marks the end of an RCU read-side critical section. * * See rcu_read_lock() for more information. */ -#define rcu_read_unlock (x) ({ ((void)(x)); preempt_enable(); }) +static inline void rcu_read_unlock (rcu_read_lock_t *lock) +{ +ASSERT (!rcu_quiesce_allowed ()); +rcu_quiesce_enable (); +}/* * So where is rcu_write_lock()? It does not exist, as there is no
Signature · 1 line
2.16.4
PATCH v7 5/5 xen/rcu: add per-lock counter in debug builds
In reply to Juergen GrossAdd a lock specific counter to rcu read locks in debug builds. This allows to test for matching lock/unlock calls.
This will help to avoid cases like the one fixed by commit 98ed1f43cc2c89 where different rcu read locks were referenced in the lock and unlock calls.
---
V5:
- updated commit message (Jan Beulich)
--- xen/include/xen/rcupdate.h | 46 +++++++++++++++++++++++++++++++++------------- 1 file changed, 33 insertions(+), 13 deletions(-)
xen/include/xen/rcupdate.h +33 -13
@@ -37,21 +37,50 @@#include <xen/cpumask.h> #include <xen/percpu.h> #include <xen/preempt.h> +#include <asm/atomic.h> #define __rcu +#ifndef NDEBUG +/* * Lock type for passing to rcu_read_{lock,unlock}. */ +struct _rcu_read_lock { +atomic_t cnt; +}; +typedef struct _rcu_read_lock rcu_read_lock_t ; +#define DEFINE_RCU_READ_LOCK (x) rcu_read_lock_t x = { .cnt = ATOMIC_INIT(0) } +#define RCU_READ_LOCK_INIT (x) atomic_set(&(x)->cnt, 0) + +#else +/* +* Dummy lock type for passing to rcu_read_{lock,unlock}. Currently exists +* only to document the reason for rcu_read_lock() critical sections. +*/ +struct _rcu_read_lock {}; +typedef struct _rcu_read_lock rcu_read_lock_t ; +#define DEFINE_RCU_READ_LOCK (x) rcu_read_lock_t x +#define RCU_READ_LOCK_INIT (x) + +#endif + DECLARE_PER_CPU(unsigned int, rcu_lock_cnt); -static inline void rcu_quiesce_disable (void ) +static inline void rcu_quiesce_disable (rcu_read_lock_t *lock) {preempt_disable ();this_cpu (rcu_lock_cnt)++; +#ifndef NDEBUG +atomic_inc (&lock->cnt); +#endif barrier (); } -static inline void rcu_quiesce_enable (void ) +static inline void rcu_quiesce_enable (rcu_read_lock_t *lock) {barrier (); +#ifndef NDEBUG +ASSERT (atomic_read (&lock->cnt)); +atomic_dec (&lock->cnt); +#endif this_cpu (rcu_lock_cnt)--;preempt_enable (); }
@@ -81,15 +110,6 @@ struct rcu_head {int rcu_pending (int cpu);int rcu_needs_cpu (int cpu); -/* -* Dummy lock type for passing to rcu_read_{lock,unlock}. Currently exists -* only to document the reason for rcu_read_lock() critical sections. -*/ -struct _rcu_read_lock {}; -typedef struct _rcu_read_lock rcu_read_lock_t ; -#define DEFINE_RCU_READ_LOCK (x) rcu_read_lock_t x -#define RCU_READ_LOCK_INIT (x) -/** * rcu_read_lock - mark the beginning of an RCU read-side critical section. *
@@ -119,7 +139,7 @@ typedef struct _rcu_read_lock rcu_read_lock_t;*/ static inline void rcu_read_lock (rcu_read_lock_t *lock) { -rcu_quiesce_disable (); +rcu_quiesce_disable (lock); }/**
@@ -130,7 +150,7 @@ static inline void rcu_read_lock(rcu_read_lock_t *lock)static inline void rcu_read_unlock (rcu_read_lock_t *lock) {ASSERT (!rcu_quiesce_allowed ()); -rcu_quiesce_enable (); +rcu_quiesce_enable (lock); }/*
Signature · 1 line
2.16.4