2008-08-21 18:58:42

by Anton Vorontsov

Subject: [PATCH] genirq: fix irq_desc->depth handling with DEBUG_SHIRQ

When DEBUG_SHIRQ is selected, a spurious IRQ is issued before
the setup_irq() initializes the desc->depth. An IRQ handler may
call disable_irq_nosync(), but then setup_irq() will overwrite
desc->depth, and upon enable_irq() we'll catch this WARN:

------------[ cut here ]------------
Badness at kernel/irq/manage.c:180
NIP: c0061ab8 LR: c0061f10 CTR: 00000000
REGS: cf83be50 TRAP: 0700 Not tainted (2.6.27-rc3-23450-g74919b0)
MSR: 00021032 <ME,IR,DR> CR: 22042022 XER: 20000000
TASK = cf829100[5] 'events/0' THREAD: cf83a000
GPR00: c0061f10 cf83bf00 cf829100 c038e674 00000016 00000000 cf83bef8 00000038
GPR08: c0298910 00000000 c0310d28 cf83a000 00000c9c 1001a1a8 0fffe000 00800000
GPR16: ffffffff 00000000 007fff00 00000000 007ffeb0 c03320a0 c031095c c0310924
GPR24: cf8292ec cf807190 cf83a000 00009032 c038e6a4 c038e674 cf99b1cc c038e674
NIP [c0061ab8] __enable_irq+0x20/0x80
LR [c0061f10] enable_irq+0x50/0x70
Call Trace:
[cf83bf00] [c038e674] irq_desc+0x630/0x9000 (unreliable)
[cf83bf10] [c0061f10] enable_irq+0x50/0x70
[cf83bf30] [c01abe94] phy_change+0x68/0x108
[cf83bf50] [c0046394] run_workqueue+0xc4/0x16c
[cf83bf90] [c0046834] worker_thread+0x74/0xd4
[cf83bfd0] [c004ab7c] kthread+0x48/0x84
[cf83bff0] [c00135e0] kernel_thread+0x44/0x60
Instruction dump:
4e800020 3d20c031 38a94214 4bffffcc 9421fff0 7c0802a6 93e1000c 7c7f1b78
90010014 8123001c 2f890000 409e001c <0fe00000> 80010014 83e1000c 38210010

That trace corresponds to this line:
WARN(1, KERN_WARNING "Unbalanced enable for IRQ %d\n", irq);

The patch fixes the problem by moving the SHIRQ code below the
setup_irq().

Unfortunately we can't easily move the SHIRQ code inside the
setup_irq(), since it grabs a spinlock, so to prvent a 'real'
IRQ from interfere us we should disable that IRQ.

p.s. The driver in question is drivers/net/phy/phy.c.

Signed-off-by: Anton Vorontsov <[email protected]>
---
kernel/irq/manage.c | 17 ++++++++++-------
1 files changed, 10 insertions(+), 7 deletions(-)

diff --git a/kernel/irq/manage.c b/kernel/irq/manage.c
index 0314074..ba03259 100644
--- a/kernel/irq/manage.c
+++ b/kernel/irq/manage.c
@@ -629,26 +629,29 @@ int request_irq(unsigned int irq, irq_handler_t handler,
action->next = NULL;
action->dev_id = dev_id;

+ retval = setup_irq(irq, action);
+ if (retval)
+ kfree(action);
+
#ifdef CONFIG_DEBUG_SHIRQ
if (irqflags & IRQF_SHARED) {
/*
* It's a shared IRQ -- the driver ought to be prepared for it
* to happen immediately, so let's make sure....
- * We do this before actually registering it, to make sure that
- * a 'real' IRQ doesn't run in parallel with our fake
+ * We disable the irq to make sure that a 'real' IRQ doesn't
+ * run in parallel with our fake.
*/
unsigned long flags;

+ disable_irq(irq);
local_irq_save(flags);
+
handler(irq, dev_id);
+
local_irq_restore(flags);
+ enable_irq(irq);
}
#endif
-
- retval = setup_irq(irq, action);
- if (retval)
- kfree(action);
-
return retval;
}
EXPORT_SYMBOL(request_irq);
--
1.5.6.3

2008-08-22 06:08:05

by Ingo Molnar

Subject: Re: [PATCH] genirq: fix irq_desc->depth handling with DEBUG_SHIRQ

* Anton Vorontsov <[email protected]> wrote:

> When DEBUG_SHIRQ is selected, a spurious IRQ is issued before
> the setup_irq() initializes the desc->depth. An IRQ handler may
> call disable_irq_nosync(), but then setup_irq() will overwrite
> desc->depth, and upon enable_irq() we'll catch this WARN:
>
> ------------[ cut here ]------------
> Badness at kernel/irq/manage.c:180
> NIP: c0061ab8 LR: c0061f10 CTR: 00000000
> REGS: cf83be50 TRAP: 0700 Not tainted (2.6.27-rc3-23450-g74919b0)
> MSR: 00021032 <ME,IR,DR> CR: 22042022 XER: 20000000
> TASK = cf829100[5] 'events/0' THREAD: cf83a000
> GPR00: c0061f10 cf83bf00 cf829100 c038e674 00000016 00000000 cf83bef8 00000038
> GPR08: c0298910 00000000 c0310d28 cf83a000 00000c9c 1001a1a8 0fffe000 00800000
> GPR16: ffffffff 00000000 007fff00 00000000 007ffeb0 c03320a0 c031095c c0310924
> GPR24: cf8292ec cf807190 cf83a000 00009032 c038e6a4 c038e674 cf99b1cc c038e674
> NIP [c0061ab8] __enable_irq+0x20/0x80
> LR [c0061f10] enable_irq+0x50/0x70
> Call Trace:
> [cf83bf00] [c038e674] irq_desc+0x630/0x9000 (unreliable)
> [cf83bf10] [c0061f10] enable_irq+0x50/0x70
> [cf83bf30] [c01abe94] phy_change+0x68/0x108
> [cf83bf50] [c0046394] run_workqueue+0xc4/0x16c
> [cf83bf90] [c0046834] worker_thread+0x74/0xd4
> [cf83bfd0] [c004ab7c] kthread+0x48/0x84
> [cf83bff0] [c00135e0] kernel_thread+0x44/0x60
> Instruction dump:
> 4e800020 3d20c031 38a94214 4bffffcc 9421fff0 7c0802a6 93e1000c 7c7f1b78
> 90010014 8123001c 2f890000 409e001c <0fe00000> 80010014 83e1000c 38210010
>
> That trace corresponds to this line:
> WARN(1, KERN_WARNING "Unbalanced enable for IRQ %d\n", irq);
>
> The patch fixes the problem by moving the SHIRQ code below the
> setup_irq().
>
> Unfortunately we can't easily move the SHIRQ code inside the
> setup_irq(), since it grabs a spinlock, so to prvent a 'real'
> IRQ from interfere us we should disable that IRQ.
>
> p.s. The driver in question is drivers/net/phy/phy.c.
>
> Signed-off-by: Anton Vorontsov <[email protected]>

applied to tip/genirq - thanks Anton.

Ingo