2007-11-15 18:05:48

by Dan Aloni

[permalink] [raw]
Subject: [PATCH] x86_64: fix a deadlock in set_rtc_mmss()

Patch is valid only for 2.6.23.x, guessing from the recent arch/ changes
in 2.6.24-rc.

set_rtc_mmss() was used to be called from interrupt context in 2.6.22,
however in 2.6.23, it is called from a timer context, where interrupts
are enabled. This patch ensures that rtc_interrupt() won't dead-lock
with set_rtc_mmss().

--
BUG: spinlock recursion on CPU#1, swapper/0

lock: ffffffff8063ffe0, .magic: dead4ead, .owner: swapper/0, .owner_cpu: 1

Call Trace:
<IRQ> [<ffffffff80361297>] spin_bug+0xa7/0x100
[<ffffffff803614d5>] _raw_spin_lock+0x145/0x150
[<ffffffff80228751>] inc_nr_running+0x31/0x40
[<ffffffff8024b4a0>] sync_cmos_clock+0x0/0xc0
[<ffffffff804fa839>] _spin_lock+0x9/0x10
[<ffffffff803ae770>] rtc_interrupt+0x10/0xe0
[<ffffffff80258975>] handle_IRQ_event+0x35/0x70
[<ffffffff80259f6c>] handle_edge_irq+0xcc/0x150
[<ffffffff8020e600>] do_IRQ+0x80/0x100
[<ffffffff8020c441>] ret_from_intr+0x0/0xa
[<ffffffff8020f4e7>] update_persistent_clock+0x47/0x1f0
[<ffffffff8024b54d>] sync_cmos_clock+0xad/0xc0
[<ffffffff8023aa58>] run_timer_softirq+0x178/0x1e0
[<ffffffff802370e4>] __do_softirq+0x74/0xe0
[<ffffffff8020d0bc>] call_softirq+0x1c/0x30
[<ffffffff8020e52d>] do_softirq+0x3d/0x90
[<ffffffff80237065>] irq_exit+0x45/0x50
[<ffffffff802182a5>] smp_apic_timer_interrupt+0x55/0x70
[<ffffffff8020b050>] default_idle+0x0/0x50
[<ffffffff8020cb66>] apic_timer_interrupt+0x66/0x70
<EOI> [<ffffffff8020b07d>] default_idle+0x2d/0x50
[<ffffffff8020a922>] enter_idle+0x22/0x30
[<ffffffff8020b0fc>] cpu_idle+0x5c/0x80
[<ffffffff806c0298>] start_secondary+0x258/0x360

Signed-off-by: Dan Aloni <[email protected]>

---
commit df9b0fba29fec2479d4c106ccdc3524fd8a61be8
tree 85b6a0f6285ab4f363291ca429ff86557a839378
parent 1915566e8113fc9cd1c6c0b9d1f5e7ee99080850
author Dan Aloni <[email protected]> Thu, 15 Nov 2007 15:11:39 +0200
committer Dan Aloni <[email protected]> Thu, 15 Nov 2007 15:11:39 +0200

arch/x86_64/kernel/time.c | 5 +++--
1 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/arch/x86_64/kernel/time.c b/arch/x86_64/kernel/time.c
index 6241b50..8414236 100644
--- a/arch/x86_64/kernel/time.c
+++ b/arch/x86_64/kernel/time.c
@@ -87,13 +87,14 @@ static int set_rtc_mmss(unsigned long nowtime)
int retval = 0;
int real_seconds, real_minutes, cmos_minutes;
unsigned char control, freq_select;
+ unsigned long flags;

/*
* IRQs are disabled when we're called from the timer interrupt,
* no need for spin_lock_irqsave()
*/

- spin_lock(&rtc_lock);
+ spin_lock_irqsave(&rtc_lock, flags);

/*
* Tell the clock it's being set and stop it.
@@ -143,7 +144,7 @@ static int set_rtc_mmss(unsigned long nowtime)
CMOS_WRITE(control, RTC_CONTROL);
CMOS_WRITE(freq_select, RTC_FREQ_SELECT);

- spin_unlock(&rtc_lock);
+ spin_unlock_irqrestore(&rtc_lock, flags);

return retval;
}

--
Dan Aloni
XIV LTD, http://www.xivstorage.com
da-x (at) monatomic.org, dan (at) xiv.co.il


2007-11-15 18:31:12

by Aristeu Rozanski

[permalink] [raw]
Subject: Re: [PATCH] x86_64: fix a deadlock in set_rtc_mmss()

Hello Dan,
> Patch is valid only for 2.6.23.x, guessing from the recent arch/ changes
> in 2.6.24-rc.
>
> set_rtc_mmss() was used to be called from interrupt context in 2.6.22,
> however in 2.6.23, it is called from a timer context, where interrupts
> are enabled. This patch ensures that rtc_interrupt() won't dead-lock
> with set_rtc_mmss().

> diff --git a/arch/x86_64/kernel/time.c b/arch/x86_64/kernel/time.c
> index 6241b50..8414236 100644
> --- a/arch/x86_64/kernel/time.c
> +++ b/arch/x86_64/kernel/time.c
> @@ -87,13 +87,14 @@ static int set_rtc_mmss(unsigned long nowtime)
> int retval = 0;
> int real_seconds, real_minutes, cmos_minutes;
> unsigned char control, freq_select;
> + unsigned long flags;
>
> /*
> * IRQs are disabled when we're called from the timer interrupt,
> * no need for spin_lock_irqsave()
> */
>
> - spin_lock(&rtc_lock);
> + spin_lock_irqsave(&rtc_lock, flags);
I think it's a good idea to update or just remove the comment above.

--
Aristeu

2007-11-16 03:29:00

by Li Zefan

[permalink] [raw]
Subject: Re: [PATCH] x86_64: fix a deadlock in set_rtc_mmss()

Aristeu Rozanski wrote:
> Hello Dan,
>> Patch is valid only for 2.6.23.x, guessing from the recent arch/ changes
>> in 2.6.24-rc.
>>
>> set_rtc_mmss() was used to be called from interrupt context in 2.6.22,
>> however in 2.6.23, it is called from a timer context, where interrupts
>> are enabled. This patch ensures that rtc_interrupt() won't dead-lock
>> with set_rtc_mmss().
>
>> diff --git a/arch/x86_64/kernel/time.c b/arch/x86_64/kernel/time.c
>> index 6241b50..8414236 100644
>> --- a/arch/x86_64/kernel/time.c
>> +++ b/arch/x86_64/kernel/time.c
>> @@ -87,13 +87,14 @@ static int set_rtc_mmss(unsigned long nowtime)
>> int retval = 0;
>> int real_seconds, real_minutes, cmos_minutes;
>> unsigned char control, freq_select;
>> + unsigned long flags;
>>
>> /*
>> * IRQs are disabled when we're called from the timer interrupt,
>> * no need for spin_lock_irqsave()
>> */
>>
>> - spin_lock(&rtc_lock);
>> + spin_lock_irqsave(&rtc_lock, flags);
> I think it's a good idea to update or just remove the comment above.
>

David P. Reed has sent a patch to fix this bug before you.

http://lkml.org/lkml/2007/11/14/435


Li Zefan

2007-11-17 08:35:12

by Dan Aloni

[permalink] [raw]
Subject: Re: [PATCH] x86_64: fix a deadlock in set_rtc_mmss()

On Fri, Nov 16, 2007 at 11:26:17AM +0800, Li Zefan wrote:
> Aristeu Rozanski wrote:
> > Hello Dan,
> >> Patch is valid only for 2.6.23.x, guessing from the recent arch/ changes
> >> in 2.6.24-rc.
> >>
> >> set_rtc_mmss() was used to be called from interrupt context in 2.6.22,
> >> however in 2.6.23, it is called from a timer context, where interrupts
> >> are enabled. This patch ensures that rtc_interrupt() won't dead-lock
> >> with set_rtc_mmss().
> >
> >> diff --git a/arch/x86_64/kernel/time.c b/arch/x86_64/kernel/time.c
> >> index 6241b50..8414236 100644
> >> --- a/arch/x86_64/kernel/time.c
> >> +++ b/arch/x86_64/kernel/time.c
> >> @@ -87,13 +87,14 @@ static int set_rtc_mmss(unsigned long nowtime)
> >> int retval = 0;
> >> int real_seconds, real_minutes, cmos_minutes;
> >> unsigned char control, freq_select;
> >> + unsigned long flags;
> >>
> >> /*
> >> * IRQs are disabled when we're called from the timer interrupt,
> >> * no need for spin_lock_irqsave()
> >> */
> >>
> >> - spin_lock(&rtc_lock);
> >> + spin_lock_irqsave(&rtc_lock, flags);
> > I think it's a good idea to update or just remove the comment above.
> >
>
> David P. Reed has sent a patch to fix this bug before you.
>
> http://lkml.org/lkml/2007/11/14/435

Yes, missed by a day. Although that patch was aimed for 2.6.24-rc,
I am currently more concerned about the 2.6.23.x tree.

Greg, can this be added to 2.6.23-stable?

Signed-off-by: Dan Aloni <[email protected]>

---
commit 26c05aa66d10e95de63d95a348023ada4827ceed
tree 7c12f2b35fcba31c9f62ba278b9d232406f80e83
parent ad1765c56c2178a09e87d123548a35f221e1fc05
author Dan Aloni <[email protected]> Sat, 17 Nov 2007 10:31:01 +0200
committer Dan Aloni <[email protected]> Sat, 17 Nov 2007 10:31:01 +0200

arch/x86_64/kernel/time.c | 10 +++-------
1 files changed, 3 insertions(+), 7 deletions(-)

diff --git a/arch/x86_64/kernel/time.c b/arch/x86_64/kernel/time.c
index 6d48a4e..fde5796 100644
--- a/arch/x86_64/kernel/time.c
+++ b/arch/x86_64/kernel/time.c
@@ -87,13 +87,9 @@ static int set_rtc_mmss(unsigned long nowtime)
int retval = 0;
int real_seconds, real_minutes, cmos_minutes;
unsigned char control, freq_select;
+ unsigned long flags;

-/*
- * IRQs are disabled when we're called from the timer interrupt,
- * no need for spin_lock_irqsave()
- */
-
- spin_lock(&rtc_lock);
+ spin_lock_irqsave(&rtc_lock, flags);

/*
* Tell the clock it's being set and stop it.
@@ -143,7 +139,7 @@ static int set_rtc_mmss(unsigned long nowtime)
CMOS_WRITE(control, RTC_CONTROL);
CMOS_WRITE(freq_select, RTC_FREQ_SELECT);

- spin_unlock(&rtc_lock);
+ spin_unlock_irqrestore(&rtc_lock, flags);

return retval;
}


--
Dan Aloni
XIV LTD, http://www.xivstorage.com
da-x (at) monatomic.org, dan (at) xiv.co.il

2007-11-20 08:00:19

by Thomas Gleixner

[permalink] [raw]
Subject: Re: [PATCH] x86_64: fix a deadlock in set_rtc_mmss()

On Thu, 15 Nov 2007, Dan Aloni wrote:

> Patch is valid only for 2.6.23.x, guessing from the recent arch/ changes
> in 2.6.24-rc.

There is a backported patch already queued in the stable branch.

Thanks,

tglx