Skip to contentSkip to conversation

workqueue: Use raise_softirq() to trigger softirq in irq_work handler

PATCH Applied Zqiang 3 participants 5 messages Jul 16 – Jul 17 Workqueue

lore
Share & tools
lore

Zqiang

author
#0 lore raw

PATCH 1/1 workqueue: Use raise_softirq() to trigger softirq in irq_work handler

The bh_pool_kick_normal() and bh_pool_kick_highpri() are registered via init_irq_work() without IRQ_WORK_HARD_IRQ flag. On PREEMPT_RT, such work items are processed by the per-CPU irq_workd kthread in preemptible task context with IRQs enabled. However calling raise_softirq_irqoff() requires IRQS must be disabled, This commit therefore replace raise_softirq_irqoff() with raise_softirq() in irq_work handler.

---
 kernel/workqueue.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
kernel/workqueue.c +2 -2
index 78068ae8f28a..0bb978df0200 100644
@@ -7923,12 +7923,12 @@ static inline void wq_watchdog_init(void) { }
 
 static void bh_pool_kick_normal(struct irq_work *irq_work)
 {
-	raise_softirq_irqoff(TASKLET_SOFTIRQ);
+	raise_softirq(TASKLET_SOFTIRQ);
 }
 
 static void bh_pool_kick_highpri(struct irq_work *irq_work)
 {
-	raise_softirq_irqoff(HI_SOFTIRQ);
+	raise_softirq(HI_SOFTIRQ);
 }
 
 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
Signature · 1 line
2.17.1
Reviewed-by on v1 1/1
#1 lore raw
Quoted 6 lines

The bh_pool_kick_normal() and bh_pool_kick_highpri() are registered via init_irq_work() without IRQ_WORK_HARD_IRQ flag. On PREEMPT_RT, such work items are processed by the per-CPU irq_workd kthread in preemptible task context with IRQs enabled. However calling raise_softirq_irqoff() requires IRQS must be disabled, This commit therefore replace raise_softirq_irqoff() with raise_softirq() in irq_work handler.

Hi zhang!

Could you add

Fixes 4cb1ef64609f ("workqueue: Implement BH workqueues to eventually replace tasklets")

?

Also, if you would like this to go to stable, could you write a small part of the description about the user-space effects and why this is bad?

After all changes.

Quoted 26 lines
---
 kernel/workqueue.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
kernel/workqueue.c +2 -2
index 78068ae8f28a..0bb978df0200 100644
@@ -7923,12 +7923,12 @@ static inline void wq_watchdog_init(void) { }
 
 static void bh_pool_kick_normal(struct irq_work *irq_work)
 {
-	raise_softirq_irqoff(TASKLET_SOFTIRQ);
+	raise_softirq(TASKLET_SOFTIRQ);
 }
 
 static void bh_pool_kick_highpri(struct irq_work *irq_work)
 {
-	raise_softirq_irqoff(HI_SOFTIRQ);
+	raise_softirq(HI_SOFTIRQ);
 }
 
 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
Signature · 1 line
2.17.1

Thanks!

maintainer Applied
#3 lore raw
In reply to Zqiang

Hello,

Applied to wq/for-7.1-fixes with the description reworded and the following tags added:

Fixes 2f34d7337d98 ("workqueue: Fix queue_work_on() with BH workqueues")

The Fixes target differs from the suggested 4cb1ef64609f. There, the raises ran under pool->lock with IRQs disabled. The irq_work handlers were added later by 2f34d7337d98 which moved the remote-CPU raise out of kick_pool(). Also picked up Bradley's Reviewed-by.

Thanks.

Signature · 1 line
tejun
author
#2 lore raw
In reply to Bradley Morgan
Quoted 13 lines
Quoted 9 lines

The bh_pool_kick_normal() and bh_pool_kick_highpri() are registered via
init_irq_work() without IRQ_WORK_HARD_IRQ flag. On PREEMPT_RT, such work items are processed by the per-CPU irq_workd kthread in preemptible task context with IRQs enabled. However calling raise_softirq_irqoff() requires IRQS must be disabled, This commit therefore replace raise_softirq_irqoff() with raise_softirq() in irq_work handler.

Hi zhang!

Could you add

Quoted 7 lines
Fixes 4cb1ef64609f ("workqueue: Implement BH workqueues to eventually replace tasklets")

?

Also, if you would like this to go to stable, could you write a small part of the description about the user-space effects and why this is bad?

The or_softirq_pending(nr) be translated into the following three assembly instructions(arm64 platform):

ldr w0, [x18, #softirq_pending_offset]

by interrupt:
calling __raise_softirq_irqoff(NET_RX_SOFTIRQ)

orr w0, w0, 1<< nr
str w0, [x18, #softirq_pending_offset] <-- still use old value, the NET_RX_SOFTIRQ maybe lose.

Thanks
Zqiang

Quoted 35 lines

After all changes.

Quoted 24 lines

---

 kernel/workqueue.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/kernel/workqueue.c b/kernel/workqueue.c
index 78068ae8f28a..0bb978df0200 100644
--- a/kernel/workqueue.c
+++ b/kernel/workqueue.c
@@ -7923,12 +7923,12 @@ static inline void wq_watchdog_init(void) { }

static void bh_pool_kick_normal(struct irq_work *irq_work) {
- raise_softirq_irqoff(TASKLET_SOFTIRQ);
+ raise_softirq(TASKLET_SOFTIRQ);
}

static void bh_pool_kick_highpri(struct irq_work *irq_work) {
- raise_softirq_irqoff(HI_SOFTIRQ);
+ raise_softirq(HI_SOFTIRQ);
}

Quoted 4 lines

static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask) --
2.17.1

Thanks!

author
#4 lore raw
In reply to Tejun Heo
Quoted 6 lines

Hello,

Applied to wq/for-7.1-fixes with the description reworded and the following tags added:

Quoted 7 lines
Fixes 2f34d7337d98 ("workqueue: Fix queue_work_on() with BH workqueues")

The Fixes target differs from the suggested 4cb1ef64609f. There, the raises ran under pool->lock with IRQs disabled. The irq_work handlers were added later by 2f34d7337d98 which moved the remote-CPU raise out of kick_pool(). Also picked up Bradley's Reviewed-by.

Hello, Tejun

Thank you very much for reword description!

Quoted 6 lines

Thanks.

Signature · 1 line
tejun