Virtualization - Aug 2010 - [PATCH 6/6] staging: hv: Gracefully handle SCSI resets

From: Hank Janssen <hjanssen at microsoft.com>

If we get a SCSI host bus reset we now gracefully handle it, and we take the
device offline.
This before sometimes caused hangs.

Signed-off-by:Hank Janssen <hjanssen at microsoft.com>
Signed-off-by:Haiyang Zhang <haiyangz at microsoft.com>


---
 drivers/staging/hv/storvsc.c |   36 +++++++++++++++++++++++++++++++++++-
 1 files changed, 35 insertions(+), 1 deletions(-)

diff --git a/drivers/staging/hv/storvsc.c b/drivers/staging/hv/storvsc.c index
6bd2ff1..5f222cf 100644
--- a/drivers/staging/hv/storvsc.c
+++ b/drivers/staging/hv/storvsc.c
@@ -48,7 +48,9 @@ struct storvsc_device {
 
 	/* 0 indicates the device is being destroyed */
 	atomic_t RefCount;
-
+	
+	int reset;
+	spinlock_t lock;
 	atomic_t NumOutstandingRequests;
 
 	/*
@@ -93,6 +95,9 @@ static inline struct storvsc_device *AllocStorDevice(struct
hv_device *Device)
 	atomic_cmpxchg(&storDevice->RefCount, 0, 2);
 
 	storDevice->Device = Device;
+	storDevice->reset  = 0;
+	spin_lock_init(&storDevice->lock);
+
 	Device->Extension = storDevice;
 
 	return storDevice;
@@ -101,6 +106,7 @@ static inline struct storvsc_device *AllocStorDevice(struct
hv_device *Device)  static inline void FreeStorDevice(struct storvsc_device
*Device)  {
 	/* ASSERT(atomic_read(&Device->RefCount) == 0); */
+	/*kfree(Device->lock);*/
 	kfree(Device);
 }
 
@@ -108,13 +114,24 @@ static inline void FreeStorDevice(struct storvsc_device
*Device)  static inline struct storvsc_device *GetStorDevice(struct hv_device
*Device)  {
 	struct storvsc_device *storDevice;
+	unsigned long flags;
 
 	storDevice = (struct storvsc_device *)Device->Extension;
+
+	spin_lock_irqsave(&storDevice->lock, flags);
+
+	if (storDevice->reset == 1) {
+		spin_unlock_irqrestore(&storDevice->lock, flags);
+		return NULL;
+	}
+
 	if (storDevice && atomic_read(&storDevice->RefCount) > 1)
 		atomic_inc(&storDevice->RefCount);
 	else
 		storDevice = NULL;
 
+	spin_unlock_irqrestore(&storDevice->lock, flags);
+
 	return storDevice;
 }
 
@@ -122,13 +139,19 @@ static inline struct storvsc_device *GetStorDevice(struct
hv_device *Device)  static inline struct storvsc_device
*MustGetStorDevice(struct hv_device *Device)  {
 	struct storvsc_device *storDevice;
+	unsigned long flags;
 
 	storDevice = (struct storvsc_device *)Device->Extension;
+
+	spin_lock_irqsave(&storDevice->lock, flags);
+
 	if (storDevice && atomic_read(&storDevice->RefCount))
 		atomic_inc(&storDevice->RefCount);
 	else
 		storDevice = NULL;
 
+	spin_unlock_irqrestore(&storDevice->lock, flags);
+
 	return storDevice;
 }
 
@@ -614,6 +637,7 @@ int StorVscOnHostReset(struct hv_device *Device)
 	struct storvsc_device *storDevice;
 	struct storvsc_request_extension *request;
 	struct vstor_packet *vstorPacket;
+	unsigned long flags;
 	int ret;
 
 	DPRINT_INFO(STORVSC, "resetting host adapter..."); @@ -625,6 +649,16
@@ int StorVscOnHostReset(struct hv_device *Device)
 		return -1;
 	}
 
+	spin_lock_irqsave(&storDevice->lock, flags);
+	storDevice->reset = 1;
+	spin_unlock_irqrestore(&storDevice->lock, flags);
+
+	/*
+	 * Wait for traffic in transit to complete
+	 */
+	while (atomic_read(&storDevice->NumOutstandingRequests))
+		udelay(1000);
+
 	request = &storDevice->ResetRequest;
 	vstorPacket = &request->VStorPacket;
 
--
1.6.0.2

On Tue, Aug 03, 2010 at 05:31:56PM +0000, Hank Janssen
wrote:
> From: Hank Janssen <hjanssen at microsoft.com>
> 
> If we get a SCSI host bus reset we now gracefully handle it, and we take
the device offline.
> This before sometimes caused hangs.
Is this a problem for all older versions as well?  If so, should it be
backported to the -stable kernel releases?
> 
> Signed-off-by:Hank Janssen <hjanssen at microsoft.com>
> Signed-off-by:Haiyang Zhang <haiyangz at microsoft.com>
> 
> 
> ---
>  drivers/staging/hv/storvsc.c |   36 +++++++++++++++++++++++++++++++++++-
>  1 files changed, 35 insertions(+), 1 deletions(-)
> 
> diff --git a/drivers/staging/hv/storvsc.c b/drivers/staging/hv/storvsc.c
index 6bd2ff1..5f222cf 100644
> --- a/drivers/staging/hv/storvsc.c
> +++ b/drivers/staging/hv/storvsc.c
> @@ -48,7 +48,9 @@ struct storvsc_device {
>  
>  	/* 0 indicates the device is being destroyed */
>  	atomic_t RefCount;
> -
> +	
Trailing whitespace :(
> +	int reset;
Can't this be a bool?
> +	spinlock_t lock;
>  	atomic_t NumOutstandingRequests;
>  
>  	/*
> @@ -93,6 +95,9 @@ static inline struct storvsc_device
*AllocStorDevice(struct hv_device *Device)
>  	atomic_cmpxchg(&storDevice->RefCount, 0, 2);
>  
>  	storDevice->Device = Device;
> +	storDevice->reset  = 0;
> +	spin_lock_init(&storDevice->lock);
> +
>  	Device->Extension = storDevice;
>  
>  	return storDevice;
> @@ -101,6 +106,7 @@ static inline struct storvsc_device
*AllocStorDevice(struct hv_device *Device)  static inline void
FreeStorDevice(struct storvsc_device *Device)  {
>  	/* ASSERT(atomic_read(&Device->RefCount) == 0); */
> +	/*kfree(Device->lock);*/
Why add a commented out line?  Especially one that is incorrect?  :)
>  	kfree(Device);
>  }
>  
> @@ -108,13 +114,24 @@ static inline void FreeStorDevice(struct
storvsc_device *Device)  static inline struct storvsc_device
*GetStorDevice(struct hv_device *Device)  {
>  	struct storvsc_device *storDevice;
> +	unsigned long flags;
>  
>  	storDevice = (struct storvsc_device *)Device->Extension;
> +
> +	spin_lock_irqsave(&storDevice->lock, flags);
> +
> +	if (storDevice->reset == 1) {
> +		spin_unlock_irqrestore(&storDevice->lock, flags);
> +		return NULL;
Don't return here, jump to the end of the function and return there.
That way you only have one lock/unlock pair and it's much easier to
maintain and audit over time that you got everything correct.
> +	}
> +
>  	if (storDevice && atomic_read(&storDevice->RefCount) >
1)
>  		atomic_inc(&storDevice->RefCount);
>  	else
>  		storDevice = NULL;
>  
> +	spin_unlock_irqrestore(&storDevice->lock, flags);
> +
>  	return storDevice;
>  }
>  
> @@ -122,13 +139,19 @@ static inline struct storvsc_device
*GetStorDevice(struct hv_device *Device)  static inline struct storvsc_device
*MustGetStorDevice(struct hv_device *Device)  {
>  	struct storvsc_device *storDevice;
> +	unsigned long flags;
>  
>  	storDevice = (struct storvsc_device *)Device->Extension;
> +
> +	spin_lock_irqsave(&storDevice->lock, flags);
> +
>  	if (storDevice && atomic_read(&storDevice->RefCount))
>  		atomic_inc(&storDevice->RefCount);
>  	else
>  		storDevice = NULL;
>  
> +	spin_unlock_irqrestore(&storDevice->lock, flags);
> +
>  	return storDevice;
>  }
>  
> @@ -614,6 +637,7 @@ int StorVscOnHostReset(struct hv_device *Device)
>  	struct storvsc_device *storDevice;
>  	struct storvsc_request_extension *request;
>  	struct vstor_packet *vstorPacket;
> +	unsigned long flags;
>  	int ret;
>  
>  	DPRINT_INFO(STORVSC, "resetting host adapter..."); @@ -625,6
+649,16 @@ int StorVscOnHostReset(struct hv_device *Device)
>  		return -1;
>  	}
>  
> +	spin_lock_irqsave(&storDevice->lock, flags);
> +	storDevice->reset = 1;
> +	spin_unlock_irqrestore(&storDevice->lock, flags);
> +
> +	/*
> +	 * Wait for traffic in transit to complete
> +	 */
> +	while (atomic_read(&storDevice->NumOutstandingRequests))
> +		udelay(1000);
What's ever going to get us out of this loop?  You need a fall-back in
case this read never succeeds.

And why an atomic value if you have a lock protecting it?  That's major
overkill and is probably not needed.

thanks,

greg k-h

And then Greg spoke;
>>On Tue, Aug 03, 2010 at 05:31:56PM +0000, Hank Janssen wrote:
>> From: Hank Janssen <hjanssen at microsoft.com>
>> 
>> If we get a SCSI host bus reset we now gracefully handle it, and we
take the device offline.
>> This before sometimes caused hangs.
>
>Is this a problem for all older versions as well?  If so, should it be
backported to the -stable kernel releases?
Yes, this should be backported to all stable kernel releases that have the
Hyper-V drivers
In them.  This fixes a VM freeze.

All patches I submitted in the batch of 6 (Except for the TODO patch) fix bugs.
This one
Fixes the worst one.

Thanks,

Hank.