fix loop with disabled tasklets
Mathijs,
You may want to have a look at the following patches (tested by visual inspection):
softirq.c renamed from softirq.c.orig +12 -22
@@ -188,22 +188,17 @@ static void tasklet_action(struct softir list = list->next; +if (!test_and_clear_bit (TASKLET_STATE_SCHED , &t->state)) +BUG (); +if (tasklet_trylock (t)) { -if (!atomic_read (&t->count)) { -if (!test_and_clear_bit (TASKLET_STATE_SCHED , &t->state)) -BUG (); +if (!atomic_read (&t->count)) t->func (t->data); -tasklet_unlock (t); -continue ; - }tasklet_unlock (t); +continue ; } -local_irq_disable (); - t->next = tasklet_vec[cpu].list; - tasklet_vec[cpu].list = t; -__cpu_raise_softirq (cpu,TASKLET_SOFTIRQ ); -local_irq_enable (); +tasklet_schedule (t); } }
@@ -222,22 +217,17 @@ static void tasklet_hi_action(struct sof list = list->next; +if (!test_and_clear_bit (TASKLET_STATE_SCHED , &t->state)) +BUG (); +if (tasklet_trylock (t)) { -if (!atomic_read (&t->count)) { -if (!test_and_clear_bit (TASKLET_STATE_SCHED , &t->state)) -BUG (); +if (!atomic_read (&t->count)) t->func (t->data); -tasklet_unlock (t); -continue ; - }tasklet_unlock (t); +continue ; } -local_irq_disable (); - t->next = tasklet_hi_vec[cpu].list; - tasklet_hi_vec[cpu].list = t; -__cpu_raise_softirq (cpu,HI_SOFTIRQ ); -local_irq_enable (); +tasklet_hi_schedule (t); } }
interrupt.h renamed from interrupt.h.orig +4 -2
@@ -185,13 +185,15 @@ static inline void tasklet_disable(strucstatic inline void tasklet_enable (struct tasklet_struct *t) {smp_mb__before_atomic_dec (); -atomic_dec (&t->count); +if (atomic_dec_and_test (&t->count)) +tasklet_schedule (t); }static inline void tasklet_hi_enable (struct tasklet_struct *t) {smp_mb__before_atomic_dec (); -atomic_dec (&t->count); +if (atomic_dec_and_test (&t->count)) +tasklet_hi_schedule (t); }extern void tasklet_kill (struct tasklet_struct *t);
Regards,
-velco
Quoted 2 lines · Momchil Velikov
You may want to have a look at the following patches (tested by visual inspection):
I like this one. I think it is what Andrea was going for, without the changes to interrupt.h. If we are going for thisone we should add some comments to interrupt.h warning about deadlocks etc.
still working on a patch that adds a cpu field to the tasklet_struct. Looking for races takes me a long time :).
me
From: Mathijs Mohlmann <[email protected]>
Date: Mon, 12 Nov 2001 08:46:08 +0100 (CET)
Quoted 2 lines · Momchil Velikov
You may want to have a look at the following patches (tested by visual inspection):
I like this one.
I don't like these changes, just make DECLARE_TASKLET_DISABLED and long-term tasklet_disable()'s be illegal.
There is only one abuser of this (keyboard), and it is easily cured.
Franks a lot,
David S. Miller
[email protected]
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Mathijs" == Mathijs Mohlmann <[email protected]> writes:
Quoted 2 lines · Momchil Velikov
Quoted 2 lines
You may want to have a look at the following patches (tested by visual inspection):
Mathijs> I like this one. I think it is what Andrea was going for, Mathijs> without the changes to interrupt.h. If we are going for
In this patch, the first thing is to deschedule the tasklet. So, the changes to interrupt.h are needed in order to put back the tasklet in the queue.
Mathijs> thisone we should add some comments to interrupt.h warning Mathijs> about deadlocks etc.
What deadlocks ? ;)
Mathijs> still working on a patch that adds a cpu field to the Mathijs> tasklet_struct. Looking for races takes me a long time :).
I think we can ignore the cpu whenever a tasklet is scheduled and disabled, treating it as not scheduled at all.
Regards,
-velco
Quoted 3 lines · On 12-Nov-2001 Momchil Velikov wrote:
In this patch, the first thing is to deschedule the tasklet. So, the changes to
interrupt.hare needed in order to put back the tasklet in the queue.
I know, but Andrea suggested not to allow scheduling of disabled tasklets Also, enableing the tasklet will result in a scheduled tasklet, regardless whether it was scheduled. Plus, we are not sure if it is scheduled on the same cpu that did the tasklet_schedule (but i might be the only one who cares about this ;)
Quoted 3 lines
Mathijs> thisone we should add some comments to
interrupt.hwarning Mathijs> about deadlocks etc.
What deadlocks ? ;)
well, loops. Dont use tasklet_kill on disabled tasklet or on not scheduled tasklets.
me
Signature · 1 line
me
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Mathijs" == Mathijs Mohlmann <[email protected]> writes:
Quoted 3 lines · Mathijs> On 12-Nov-2001 Momchil Velikov wrote:
Quoted 3 lines
In this patch, the first thing is to deschedule the tasklet. So, the changes to
interrupt.hare needed in order to put back the tasklet in the queue.
Mathijs> I know, but Andrea suggested not to allow scheduling of Mathijs> disabled tasklets Also, enableing the tasklet will result in
Disabled tasklets are not scheduled by enable_tasklet (). A disabled tasklet may temporarily appear in the queue, but nevertheless tasklet_action will remove it. It seems gross to traverse the list in order to remove a tasklet at the first disable.
Mathijs> a scheduled tasklet, regardless whether it was Mathijs> scheduled.
Hmm, if it isn't scheduled, there is not much sense in disabling it at all.
Mathijs> Plus, we are not sure if it is scheduled on the Mathijs> same cpu that did the tasklet_schedule (but i might be the Mathijs> only one who cares about this ;)
Mathijs> thisone we should add some comments to interrupt.h warning Mathijs> about deadlocks etc.
Quoted 1 line
Quoted 1 line
What deadlocks ? ;)
Mathijs> well, loops. Dont use tasklet_kill on disabled tasklet or on Mathijs> not scheduled tasklets.
Hmm, if TASKLET_STATE_SCHED is not set tasklet_kill will not deadlock. And tasklet_kill yields (?). Doesn't that mean that tasklet_action will be called eventually ?
Regards,
-velco
Quoted 2 lines · On 12-Nov-2001 Momchil Velikov wrote:
Hmm, if it isn't scheduled, there is not much sense in disabling it at all.
DECLARE_TASKLET_DISABLEDtasklet_enable
Quoted 3 lines
Hmm, if
TASKLET_STATE_SCHEDis not settasklet_killwill not deadlock. Andtasklet_killyields (?). Doesn't that mean thattasklet_actionwill be called eventually ?
oops, i misread tasklet_unlock_wait. So sorry.
me
Signature · 1 line
me