Skip to contentSkip to conversation

rcu: rcu_read_lock_bh_held(): disabling irqs also disables bh

PATCH In review Paul E. McKenney 2 participants 3 messages Sep 23, 2010 RCU

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 rcu: rcu_read_lock_bh_held(): disabling irqs also disables bh

rcu_dereference_bh() doesnt know yet about hard irq being disabled, so lockdep can trigger in netpoll_rx() after commit f0f9deae9e7c4 (netpoll:
Disable IRQ around RCU dereference in netpoll_rx)

---
 include/linux/rcupdate.h |    3 ++-
 1 files changed, 2 insertions(+), 1 deletions(-)
include/linux/rcupdate.h +2 -1
index 9fbc54a..0dcc00e 100644
@@ -454,7 +454,8 @@ static inline notrace void rcu_read_unlock_sched_notrace(void)
  * Makes rcu_dereference_check() do the dirty work.
  */
 #define rcu_dereference_bh(p) \
-		rcu_dereference_check(p, rcu_read_lock_bh_held())
+		rcu_dereference_check(p, rcu_read_lock_bh_held() || \
+					 irqs_disabled())
 
 /**
  * rcu_dereference_sched - fetch RCU-protected pointer, checking for RCU-sched
Signature · 1 line
1.7.0.6
#1 lore raw
Quoted 17 lines · Paul E. McKenney

rcu_dereference_bh() doesnt know yet about hard irq being disabled, so lockdep can trigger in netpoll_rx() after commit f0f9deae9e7c4 (netpoll:
Disable IRQ around RCU dereference in netpoll_rx)

---
 include/linux/rcupdate.h |    3 ++-
 1 files changed, 2 insertions(+), 1 deletions(-)
include/linux/rcupdate.h +0 -0
index 9fbc54a..0dcc00e 100644

include/linux/rcupdate.h static inline notrace void rcu_read_unlock_sched_notrace(void) · line 461

Quoted 9 lines
include/linux/rcupdate.h +2 -1
@@ -454,7 +454,8 @@ static inline notrace void rcu_read_unlock_sched_notrace(void)
  * Makes rcu_dereference_check() do the dirty work.
  */
 #define rcu_dereference_bh(p) \
-		rcu_dereference_check(p, rcu_read_lock_bh_held())
+		rcu_dereference_check(p, rcu_read_lock_bh_held() || \
+					 irqs_disabled())
 
 /**
  * rcu_dereference_sched - fetch RCU-protected pointer, checking for RCU-sched

In -rt, rcu_read_lock_bh() is preemptible, but all RCU Implementation, PREEMPT_RCU, TREE_PREEMPT_RCU, TINY_PREEMPT_RCU, do not guarantee hard-irq/irq-disabled contexts are preeemptible-rcu-read-site critical regions, so it(disabling irqs also disables bh) is not true in -rt I think, To make the codes be still safe in the future(-rt), I strongly recommend to use rcu_read_lock_bh() as needed even irq disabled.

Current, preempt_disable()/rcu_read_lock_sched() do not guarantee that they implies rcu_read_lock() either.

But in future, the rcu may give these guarantees, it's a plan of mine.

Lai

#2 lore raw
Quoted 25 lines · Lai Jiangshan
Quoted 24 lines · Paul E. McKenney

rcu_dereference_bh() doesnt know yet about hard irq being disabled, so lockdep can trigger in netpoll_rx() after commit f0f9deae9e7c4 (netpoll:
Disable IRQ around RCU dereference in netpoll_rx)

---
 include/linux/rcupdate.h |    3 ++-
 1 files changed, 2 insertions(+), 1 deletions(-)
include/linux/rcupdate.h +2 -1
index 9fbc54a..0dcc00e 100644
@@ -454,7 +454,8 @@ static inline notrace void rcu_read_unlock_sched_notrace(void)
  * Makes rcu_dereference_check() do the dirty work.
  */
 #define rcu_dereference_bh(p) \
-		rcu_dereference_check(p, rcu_read_lock_bh_held())
+		rcu_dereference_check(p, rcu_read_lock_bh_held() || \
+					 irqs_disabled())
 
Quoted 9 lines

include/linux/rcupdate.h static inline notrace void rcu_read_unlock_sched_notrace(void) · line 461

Quoted 2 lines
include/linux/rcupdate.h +0 -0
 /**
  * rcu_dereference_sched - fetch RCU-protected pointer, checking for RCU-sched

In -rt, rcu_read_lock_bh() is preemptible, but all RCU Implementation, PREEMPT_RCU, TREE_PREEMPT_RCU, TINY_PREEMPT_RCU, do not guarantee hard-irq/irq-disabled contexts are preeemptible-rcu-read-site critical regions, so it(disabling irqs also disables bh) is not true in -rt I think, To make the codes be still safe in the future(-rt), I strongly recommend to use rcu_read_lock_bh() as needed even irq disabled.

We have been going back and forth on exactly how bh and irq should interact. The -rt case is indeed more interesting because local_bh_disable() is a no-op when CONFIG_PREEMPT_HARDIRQS, but the above patch is targeted only to mainline. Either way, the current implementations refuse to end a grace period until all CPUs have spent some time with irqs enabled. I don't want to export that guarantee, but it might be OK for internal RCU use, including for rcu_dereference_bh().

Within mainline, one might make rcu_read_lock_bh() do something like:

	if (!irqs_disabled())
		local_bh_disable();

with corresponding changes for rcu_read_unlock_bh(), or, alternatively have a separate API for use when the caller might have irqs disabled.

However, there is still a bit of contention about exactly how rcu_read_lock_bh() should be handled if the caller might have irqs disabled. In the past, all calls to rcu_read_lock_bh() have had irqs enabled, so this is a new situation. Some would prefer that rcu_read_lock_bh() require that irqs be enabled, for example.

I am sure that we will figure something out. ;-)

Quoted 2 lines

Current, preempt_disable()/rcu_read_lock_sched() do not guarantee that they implies rcu_read_lock() either.

Quite true!!!

Quoted 1 line

But in future, the rcu may give these guarantees, it's a plan of mine.

Back in PREEMPT_RCU days, I did put together patches for similar changes, but did not push them. Such guarantees are promises, and promises are often easier to make in the here and now than to keep in the longer term.

Thanx, Paul