2009-09-10 14:36:08

by Alexander Strakh

[permalink] [raw]
Subject: [PATCH] fix lock imbalances in /drivers/media/video/cafe_ccic.c

In ./drivers/media/video/cafe_ccic.c, in function cafe_pci_probe:
Mutex must be unlocked before exit
1. On paths starting with mutex lock in line 1912, then continuing in lines:
1929, 1936 (goto unreg) and 1940 (goto iounmap) .
2. On path starting in line 1971 mutex lock, and then continuing in line 1978
(goto out_smbus) mutex.

Fix lock imbalances in function cafe_pci_probe.
Found by Linux Driver Verification project.

Signed-off-by: Alexander Strakh <[email protected]>

---
diff --git a/./a/drivers/media/video/cafe_ccic.c
b/./b/drivers/media/video/cafe_ccic.c
index c4d181d..2987433 100644
--- a/./a/drivers/media/video/cafe_ccic.c
+++ b/./b/drivers/media/video/cafe_ccic.c
@@ -1925,19 +1925,24 @@ static int cafe_pci_probe(struct pci_dev *pdev,
* Get set up on the PCI bus.
*/
ret = pci_enable_device(pdev);
- if (ret)
+ if (ret) {
+ mutex_unlock(&cam->s_mutex);
goto out_unreg;
+ }
pci_set_master(pdev);

ret = -EIO;
cam->regs = pci_iomap(pdev, 0, 0);
if (! cam->regs) {
printk(KERN_ERR "Unable to ioremap cafe-ccic regs\n");
+ mutex_unlock(&cam->s_mutex);
goto out_unreg;
}
ret = request_irq(pdev->irq, cafe_irq, IRQF_SHARED, "cafe-ccic", cam);
- if (ret)
+ if (ret) {
+ mutex_unlock(&cam->s_mutex);
goto out_iounmap;
+ }
/*
* Initialize the controller and leave it powered up. It will
* stay that way until the sensor driver shows up.
@@ -1974,8 +1979,10 @@ static int cafe_pci_probe(struct pci_dev *pdev,
/* cam->vdev.debug = V4L2_DEBUG_IOCTL_ARG;*/
cam->vdev.v4l2_dev = &cam->v4l2_dev;
ret = video_register_device(&cam->vdev, VFL_TYPE_GRABBER, -1);
- if (ret)
+ if (ret) {
+ mutex_unlock(&cam->s_mutex);
goto out_smbus;
+ }
video_set_drvdata(&cam->vdev, cam);

/*


2009-09-10 14:48:20

by Daniel Walker

[permalink] [raw]
Subject: Re: [PATCH] fix lock imbalances in /drivers/media/video/cafe_ccic.c

On Thu, 2009-09-10 at 18:37 +0000, iceberg wrote:
> In ./drivers/media/video/cafe_ccic.c, in function cafe_pci_probe:
> Mutex must be unlocked before exit
> 1. On paths starting with mutex lock in line 1912, then continuing in lines:
> 1929, 1936 (goto unreg) and 1940 (goto iounmap) .
> 2. On path starting in line 1971 mutex lock, and then continuing in line 1978
> (goto out_smbus) mutex.
>
> Fix lock imbalances in function cafe_pci_probe.
> Found by Linux Driver Verification project.

Could you run this through checkpatch and fix any issues you find? It
looks like you need to use tabs for indentation in a couple places.

Daniel

2009-09-10 15:30:04

by Jonathan Corbet

[permalink] [raw]
Subject: Re: [PATCH] fix lock imbalances in /drivers/media/video/cafe_ccic.c

On Thu, 10 Sep 2009 18:37:34 +0000
iceberg <[email protected]> wrote:

> In ./drivers/media/video/cafe_ccic.c, in function cafe_pci_probe:
> Mutex must be unlocked before exit
> 1. On paths starting with mutex lock in line 1912, then continuing in lines:
> 1929, 1936 (goto unreg) and 1940 (goto iounmap) .
> 2. On path starting in line 1971 mutex lock, and then continuing in line 1978
> (goto out_smbus) mutex.

That's a definite bug, but I hate all those unlocks in the error
branches. As it happens, we don't really need the mutex until the
device has been exposed to the rest of the kernel, so I propose the
following as a better patch.

Thanks for pointing this out,

jon

---
Fix a mutex leak

Certain error exits from cafe_pci_probe() can leave the camera mutex
locked. For much of the time, we didn't need the mutex anyway; take it out
and add an unlock in the path where it is needed.

Reported-by: Alexander Strakh <[email protected]>
Signed-off-by: Jonathan Corbet <[email protected]>
---
drivers/media/video/cafe_ccic.c | 3 +--
1 files changed, 1 insertions(+), 2 deletions(-)

diff --git a/drivers/media/video/cafe_ccic.c b/drivers/media/video/cafe_ccic.c
index c4d181d..0f62b5e 100644
--- a/drivers/media/video/cafe_ccic.c
+++ b/drivers/media/video/cafe_ccic.c
@@ -1909,7 +1909,6 @@ static int cafe_pci_probe(struct pci_dev *pdev,
goto out_free;

mutex_init(&cam->s_mutex);
- mutex_lock(&cam->s_mutex);
spin_lock_init(&cam->dev_lock);
cam->state = S_NOTREADY;
cafe_set_config_needed(cam, 1);
@@ -1949,7 +1948,6 @@ static int cafe_pci_probe(struct pci_dev *pdev,
* because the sensor could attach in this call chain, leading to
* unsightly deadlocks.
*/
- mutex_unlock(&cam->s_mutex); /* attach can deadlock */
ret = cafe_smbus_setup(cam);
if (ret)
goto out_freeirq;
@@ -1991,6 +1989,7 @@ static int cafe_pci_probe(struct pci_dev *pdev,
return 0;

out_smbus:
+ mutex_unlock(&cam->s_mutex);
cafe_smbus_shutdown(cam);
out_freeirq:
cafe_ctlr_power_down(cam);
--
1.6.2.5

2009-09-10 19:05:37

by Steven Rostedt

[permalink] [raw]
Subject: Re: [PATCH] fix lock imbalances in /drivers/media/video/cafe_ccic.c

On Thu, Sep 10, 2009 at 09:30:03AM -0600, Jonathan Corbet wrote:
> On Thu, 10 Sep 2009 18:37:34 +0000
> iceberg <[email protected]> wrote:
>
> > In ./drivers/media/video/cafe_ccic.c, in function cafe_pci_probe:
> > Mutex must be unlocked before exit
> > 1. On paths starting with mutex lock in line 1912, then continuing in lines:
> > 1929, 1936 (goto unreg) and 1940 (goto iounmap) .
> > 2. On path starting in line 1971 mutex lock, and then continuing in line 1978
> > (goto out_smbus) mutex.
>
> That's a definite bug, but I hate all those unlocks in the error
> branches. As it happens, we don't really need the mutex until the
> device has been exposed to the rest of the kernel, so I propose the
> following as a better patch.
>
> Thanks for pointing this out,

Actually, for something like this, I would put the mutex_unlock in the error path,
and just add a local variable to tell that it is locked.

int is_locked = 0;

[...]

mutex_lock(&cam->s_mutex);
is_locked = 1;

[...]

mutex_unlock(&cam->s_mutex);
is_locked = 0;

[...]

out_iounmap:
pci_iounmap(pdev, cam->regs);
if (is_locked)
mutex_unlock(&cam->s_mutex);
out_free:

[...]

Or something similar. I hate the multiple unlocks too.

-- Steve

>
> jon
>
> ---
> Fix a mutex leak
>
> Certain error exits from cafe_pci_probe() can leave the camera mutex
> locked. For much of the time, we didn't need the mutex anyway; take it out
> and add an unlock in the path where it is needed.
>
> Reported-by: Alexander Strakh <[email protected]>
> Signed-off-by: Jonathan Corbet <[email protected]>
> ---
> drivers/media/video/cafe_ccic.c | 3 +--
> 1 files changed, 1 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/media/video/cafe_ccic.c b/drivers/media/video/cafe_ccic.c
> index c4d181d..0f62b5e 100644
> --- a/drivers/media/video/cafe_ccic.c
> +++ b/drivers/media/video/cafe_ccic.c
> @@ -1909,7 +1909,6 @@ static int cafe_pci_probe(struct pci_dev *pdev,
> goto out_free;
>
> mutex_init(&cam->s_mutex);
> - mutex_lock(&cam->s_mutex);
> spin_lock_init(&cam->dev_lock);
> cam->state = S_NOTREADY;
> cafe_set_config_needed(cam, 1);
> @@ -1949,7 +1948,6 @@ static int cafe_pci_probe(struct pci_dev *pdev,
> * because the sensor could attach in this call chain, leading to
> * unsightly deadlocks.
> */
> - mutex_unlock(&cam->s_mutex); /* attach can deadlock */
> ret = cafe_smbus_setup(cam);
> if (ret)
> goto out_freeirq;
> @@ -1991,6 +1989,7 @@ static int cafe_pci_probe(struct pci_dev *pdev,
> return 0;
>
> out_smbus:
> + mutex_unlock(&cam->s_mutex);
> cafe_smbus_shutdown(cam);
> out_freeirq:
> cafe_ctlr_power_down(cam);
> --
> 1.6.2.5
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/

2009-09-10 15:27:40

by Alexander Strakh

[permalink] [raw]
Subject: Re: Re: [PATCH] fix lock imbalances in /drivers/media/video/cafe_ccic.c

In ./drivers/media/video/cafe_ccic.c, in function cafe_pci_probe:
Mutex must be unlocked before exit
1. On paths starting with mutex lock in line 1912, then continuing in lines:
1929, 1936 (goto unreg) and 1940 (goto iounmap) .
2. On path starting in line 1971 mutex lock, and then continuing in line 1978
(goto out_smbus) mutex.

Fix lock imbalances in function cafe_pci_probe.
Found by Linux Driver Verification project.

Signed-off-by: Alexander Strakh <[email protected]>

---
diff --git a/./a/drivers/media/video/cafe_ccic.c
b/./b/drivers/media/video/cafe_ccic.c
index c4d181d..2987433 100644
--- a/./a/drivers/media/video/cafe_ccic.c
+++ b/./b/drivers/media/video/cafe_ccic.c
@@ -1925,19 +1925,24 @@ static int cafe_pci_probe(struct pci_dev *pdev,
* Get set up on the PCI bus.
*/
ret = pci_enable_device(pdev);
- if (ret)
+ if (ret) {
+ mutex_unlock(&cam->s_mutex);
goto out_unreg;
+ }
pci_set_master(pdev);

ret = -EIO;
cam->regs = pci_iomap(pdev, 0, 0);
if (! cam->regs) {
printk(KERN_ERR "Unable to ioremap cafe-ccic regs\n");
+ mutex_unlock(&cam->s_mutex);
goto out_unreg;
}
ret = request_irq(pdev->irq, cafe_irq, IRQF_SHARED, "cafe-ccic", cam);
- if (ret)
+ if (ret) {
+ mutex_unlock(&cam->s_mutex);
goto out_iounmap;
+ }
/*
* Initialize the controller and leave it powered up. It will
* stay that way until the sensor driver shows up.
@@ -1974,8 +1979,10 @@ static int cafe_pci_probe(struct pci_dev *pdev,
/* cam->vdev.debug = V4L2_DEBUG_IOCTL_ARG;*/
cam->vdev.v4l2_dev = &cam->v4l2_dev;
ret = video_register_device(&cam->vdev, VFL_TYPE_GRABBER, -1);
- if (ret)
+ if (ret) {
+ mutex_unlock(&cam->s_mutex);
goto out_smbus;
+ }
video_set_drvdata(&cam->vdev, cam);

/*

2009-09-11 13:46:10

by Alexander Strakh

[permalink] [raw]
Subject: Re: [PATCH] fix lock imbalances in /drivers/media/video/cafe_ccic.c

On Thursday 10 September 2009 15:30:03 you wrote:

Incorrect patch. See path starting with "if (cam->sensor == null) {" in line
1960. In this case we goto out_smbs and try mutex_unlock on unlocking mutex.

> On Thu, 10 Sep 2009 18:37:34 +0000
>
> iceberg <[email protected]> wrote:
> > In ./drivers/media/video/cafe_ccic.c, in function cafe_pci_probe:
> > Mutex must be unlocked before exit
> > 1. On paths starting with mutex lock in line 1912, then continuing in
> > lines: 1929, 1936 (goto unreg) and 1940 (goto iounmap) .
> > 2. On path starting in line 1971 mutex lock, and then continuing in line
> > 1978 (goto out_smbus) mutex.
>
> That's a definite bug, but I hate all those unlocks in the error
> branches. As it happens, we don't really need the mutex until the
> device has been exposed to the rest of the kernel, so I propose the
> following as a better patch.
>
> Thanks for pointing this out,
>
> jon
>
> ---
> Fix a mutex leak
>
> Certain error exits from cafe_pci_probe() can leave the camera mutex
> locked. For much of the time, we didn't need the mutex anyway; take it out
> and add an unlock in the path where it is needed.
>
> Reported-by: Alexander Strakh <[email protected]>
> Signed-off-by: Jonathan Corbet <[email protected]>
> ---
> drivers/media/video/cafe_ccic.c | 3 +--
> 1 files changed, 1 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/media/video/cafe_ccic.c
> b/drivers/media/video/cafe_ccic.c index c4d181d..0f62b5e 100644
> --- a/drivers/media/video/cafe_ccic.c
> +++ b/drivers/media/video/cafe_ccic.c
> @@ -1909,7 +1909,6 @@ static int cafe_pci_probe(struct pci_dev *pdev,
> goto out_free;
>
> mutex_init(&cam->s_mutex);
> - mutex_lock(&cam->s_mutex);
> spin_lock_init(&cam->dev_lock);
> cam->state = S_NOTREADY;
> cafe_set_config_needed(cam, 1);
> @@ -1949,7 +1948,6 @@ static int cafe_pci_probe(struct pci_dev *pdev,
> * because the sensor could attach in this call chain, leading to
> * unsightly deadlocks.
> */
> - mutex_unlock(&cam->s_mutex); /* attach can deadlock */
> ret = cafe_smbus_setup(cam);
> if (ret)
> goto out_freeirq;
> @@ -1991,6 +1989,7 @@ static int cafe_pci_probe(struct pci_dev *pdev,
> return 0;
>
> out_smbus:
> + mutex_unlock(&cam->s_mutex);
> cafe_smbus_shutdown(cam);
> out_freeirq:
> cafe_ctlr_power_down(cam);