Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1530866 > unrolled thread

[PATCH] vfio/pci: Support error recovery

Back to article view | Back to linux.kernel

Contents

  [PATCH] vfio/pci: Support error recovery Cao jin <caoj.fnst@cn.fujitsu.com> - 2016-11-27 12:40 +0100
    Re: [PATCH] vfio/pci: Support error recovery "Michael S. Tsirkin" <mst@redhat.com> - 2016-11-28 04:10 +0100
      Re: [PATCH] vfio/pci: Support error recovery Cao jin <caoj.fnst@cn.fujitsu.com> - 2016-11-28 10:30 +0100
        Re: [PATCH] vfio/pci: Support error recovery "Michael S. Tsirkin" <mst@redhat.com> - 2016-11-30 02:50 +0100
          Re: [PATCH] vfio/pci: Support error recovery Cao jin <caoj.fnst@cn.fujitsu.com> - 2016-12-01 14:40 +0100
    Re: [PATCH] vfio/pci: Support error recovery Alex Williamson <alex.williamson@redhat.com> - 2016-12-01 05:10 +0100
      Re: [PATCH] vfio/pci: Support error recovery "Michael S. Tsirkin" <mst@redhat.com> - 2016-12-01 06:10 +0100
        Re: [PATCH] vfio/pci: Support error recovery Cao jin <caoj.fnst@cn.fujitsu.com> - 2016-12-01 14:40 +0100
          Re: [PATCH] vfio/pci: Support error recovery "Michael S. Tsirkin" <mst@redhat.com> - 2016-12-06 04:50 +0100
            Re: [PATCH] vfio/pci: Support error recovery Cao jin <caoj.fnst@cn.fujitsu.com> - 2016-12-06 07:50 +0100
      Re: [PATCH] vfio/pci: Support error recovery Cao jin <caoj.fnst@cn.fujitsu.com> - 2016-12-01 14:40 +0100
        Re: [PATCH] vfio/pci: Support error recovery Alex Williamson <alex.williamson@redhat.com> - 2016-12-01 16:00 +0100
          Re: [PATCH] vfio/pci: Support error recovery Cao jin <caoj.fnst@cn.fujitsu.com> - 2016-12-04 13:20 +0100
            Re: [PATCH] vfio/pci: Support error recovery Alex Williamson <alex.williamson@redhat.com> - 2016-12-04 16:40 +0100
              Re: [PATCH] vfio/pci: Support error recovery Cao jin <caoj.fnst@cn.fujitsu.com> - 2016-12-05 06:50 +0100
                Re: [PATCH] vfio/pci: Support error recovery Alex Williamson <alex.williamson@redhat.com> - 2016-12-05 17:20 +0100
                  Re: [PATCH] vfio/pci: Support error recovery "Michael S. Tsirkin" <mst@redhat.com> - 2016-12-06 05:00 +0100
                    Re: [PATCH] vfio/pci: Support error recovery Alex Williamson <alex.williamson@redhat.com> - 2016-12-06 06:10 +0100
                      Re: [PATCH] vfio/pci: Support error recovery Cao jin <caoj.fnst@cn.fujitsu.com> - 2016-12-06 11:50 +0100
                        Re: [PATCH] vfio/pci: Support error recovery Alex Williamson <alex.williamson@redhat.com> - 2016-12-06 16:40 +0100
                          Re: [PATCH] vfio/pci: Support error recovery Cao jin <caoj.fnst@cn.fujitsu.com> - 2016-12-07 03:50 +0100
                          Re: [PATCH] vfio/pci: Support error recovery Cao jin <caoj.fnst@cn.fujitsu.com> - 2016-12-08 15:50 +0100
                            Re: [PATCH] vfio/pci: Support error recovery "Michael S. Tsirkin" <mst@redhat.com> - 2016-12-08 17:40 +0100
                              Re: [PATCH] vfio/pci: Support error recovery Cao jin <caoj.fnst@cn.fujitsu.com> - 2016-12-09 08:50 +0100
                            Re: [PATCH] vfio/pci: Support error recovery Cao jin <caoj.fnst@cn.fujitsu.com> - 2016-12-09 08:50 +0100
                  Re: [PATCH] vfio/pci: Support error recovery Cao jin <caoj.fnst@cn.fujitsu.com> - 2016-12-06 07:10 +0100
                    Re: [PATCH] vfio/pci: Support error recovery Alex Williamson <alex.williamson@redhat.com> - 2016-12-06 16:30 +0100
                      Re: [PATCH] vfio/pci: Support error recovery Cao jin <caoj.fnst@cn.fujitsu.com> - 2016-12-07 04:00 +0100

Page 1 of 2  [1] 2  Next page →

#1530866 — [PATCH] vfio/pci: Support error recovery

It is user space driver's or device-specific driver's(in guest) responsbility
to do a serious recovery when error happened. Link-reset is one part of
recovery, when pci device is assigned to VM via vfio, link-reset will do
twice in host & guest separately, which will cause many trouble for a
successful recovery, so, disable the vfio-pci's link-reset in aer driver
in host, this is a keypoint for guest to do error recovery successfully.

CC: alex.williamson@redhat.com
CC: mst@redhat.com
Signed-off-by: Cao jin <caoj.fnst@cn.fujitsu.com>
---
This is actually a RFC version(has debug lines left), and has minor changes in
aer driver, so I think maybe it is better not to CC pci guys in this round.
Later will do.

 drivers/pci/pcie/aer/aerdrv_core.c  | 12 ++++++-
 drivers/vfio/pci/vfio_pci.c         | 63 +++++++++++++++++++++++++++++++++++--
 drivers/vfio/pci/vfio_pci_private.h |  2 ++
 3 files changed, 74 insertions(+), 3 deletions(-)

diff --git a/drivers/pci/pcie/aer/aerdrv_core.c b/drivers/pci/pcie/aer/aerdrv_core.c
index 521e39c..289fb8e 100644
--- a/drivers/pci/pcie/aer/aerdrv_core.c
+++ b/drivers/pci/pcie/aer/aerdrv_core.c
@@ -496,7 +496,17 @@ static void do_recovery(struct pci_dev *dev, int severity)
 			"error_detected",
 			report_error_detected);
 
-	if (severity == AER_FATAL) {
+	/* vfio-pci as a general meta driver, it actually couldn't do any real
+	 * recovery for device. It is user space driver, or device-specific
+	 * driver in guest who should take care of the serious error recovery,
+	 * link reset actually is one part of whole recovery. Doing reset_link
+	 * in aer driver of host kernel for vfio-pci devices will cause many
+	 * trouble for user space driver or guest's device-specific driver,
+	 * for example: the serious recovery often need to read register in
+	 * config space, but if register reading happens during link-resetting,
+	 * it is quite possible to return invalid value like all F's, which
+	 * will result in unpredictable error. */
+	if (severity == AER_FATAL && strcmp(dev->driver->name, "vfio-pci")) {
 		result = reset_link(dev);
 		if (result != PCI_ERS_RESULT_RECOVERED)
 			goto failed;
diff --git a/drivers/vfio/pci/vfio_pci.c b/drivers/vfio/pci/vfio_pci.c
index 712a849..aefd751 100644
--- a/drivers/vfio/pci/vfio_pci.c
+++ b/drivers/vfio/pci/vfio_pci.c
@@ -535,6 +535,15 @@ static long vfio_pci_ioctl(void *device_data,
 	struct vfio_pci_device *vdev = device_data;
 	unsigned long minsz;
 
+	if (vdev->aer_recovering && (cmd == VFIO_DEVICE_SET_IRQS ||
+	    cmd == VFIO_DEVICE_RESET || cmd == VFIO_DEVICE_PCI_HOT_RESET)) {
+		int ret;
+		ret = wait_for_completion_interruptible(
+			&vdev->aer_error_completion);
+		if (ret)
+			return ret;
+	}
+
 	if (cmd == VFIO_DEVICE_GET_INFO) {
 		struct vfio_device_info info;
 
@@ -1117,6 +1126,7 @@ static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 	vdev->irq_type = VFIO_PCI_NUM_IRQS;
 	mutex_init(&vdev->igate);
 	spin_lock_init(&vdev->irqlock);
+	init_completion(&vdev->aer_error_completion);
 
 	ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
 	if (ret) {
@@ -1176,6 +1186,9 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
 {
 	struct vfio_pci_device *vdev;
 	struct vfio_device *device;
+	u32 uncor_status = 0;
+	unsigned int aer_cap_offset = 0;
+	int ret;
 
 	device = vfio_device_get_from_dev(&pdev->dev);
 	if (device == NULL)
@@ -1187,10 +1200,30 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
 		return PCI_ERS_RESULT_DISCONNECT;
 	}
 
+	/* get device's uncorrectable error status as soon as possible,
+	 * and signal it to user space. The later we read it, the possibility
+	 * the register value is mangled grows. */
+	aer_cap_offset = pci_find_ext_capability(vdev->pdev, PCI_EXT_CAP_ID_ERR);
+	ret = pci_read_config_dword(vdev->pdev, aer_cap_offset + 
+                                    PCI_ERR_UNCOR_STATUS, &uncor_status);
+        if (ret)
+                return PCI_ERS_RESULT_DISCONNECT;
+
+	pr_err("device %d got AER detect notification. uncorrectable error status = 0x%x\n", pdev->devfn, uncor_status);//to be removed
 	mutex_lock(&vdev->igate);
+    
+	vdev->aer_recovering = true;
+	reinit_completion(&vdev->aer_error_completion);
+
+	/* suspend config space access from user space,
+	 * when vfio-pci's error recovery process is on */
+	pci_cfg_access_trylock(vdev->pdev);
 
-	if (vdev->err_trigger)
-		eventfd_signal(vdev->err_trigger, 1);
+	if (vdev->err_trigger && uncor_status) {
+		pr_err("device %d signal uncor status to user space", pdev->devfn);//may be removed
+		/* signal uncorrectable error status to user space */
+		eventfd_signal(vdev->err_trigger, uncor_status);
+        }
 
 	mutex_unlock(&vdev->igate);
 
@@ -1199,8 +1232,34 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
 	return PCI_ERS_RESULT_CAN_RECOVER;
 }
 
+static void vfio_pci_aer_resume(struct pci_dev *pdev)
+{
+	struct vfio_pci_device *vdev;
+	struct vfio_device *device;
+
+	device = vfio_device_get_from_dev(&pdev->dev);
+	if (device == NULL)
+		return;
+
+	vdev = vfio_device_data(device);
+	if (vdev == NULL) {
+		vfio_device_put(device);
+		return;
+	}
+
+	/* vfio-pci's error recovery is done, time to
+	 * resume pci config space's accesses */
+	pci_cfg_access_unlock(vdev->pdev);
+
+	vdev->aer_recovering = false;
+	complete_all(&vdev->aer_error_completion);
+
+	vfio_device_put(device);
+}
+
 static const struct pci_error_handlers vfio_err_handlers = {
 	.error_detected = vfio_pci_aer_err_detected,
+	.resume         = vfio_pci_aer_resume,
 };
 
 static struct pci_driver vfio_pci_driver = {
diff --git a/drivers/vfio/pci/vfio_pci_private.h b/drivers/vfio/pci/vfio_pci_private.h
index 8a7d546..ebf1041 100644
--- a/drivers/vfio/pci/vfio_pci_private.h
+++ b/drivers/vfio/pci/vfio_pci_private.h
@@ -83,6 +83,8 @@ struct vfio_pci_device {
 	bool			bardirty;
 	bool			has_vga;
 	bool			needs_reset;
+	bool			aer_recovering;
+	struct completion	aer_error_completion;
 	struct pci_saved_state	*pci_saved_state;
 	int			refcnt;
 	struct eventfd_ctx	*err_trigger;
-- 
1.8.3.1

[toc] | [next] | [standalone]

#1531013

On Sun, Nov 27, 2016 at 07:34:17PM +0800, Cao jin wrote:
> It is user space driver's or device-specific driver's(in guest) responsbility
> to do a serious recovery when error happened. Link-reset is one part of
> recovery, when pci device is assigned to VM via vfio, link-reset will do
> twice in host & guest separately, which will cause many trouble for a
> successful recovery, so, disable the vfio-pci's link-reset in aer driver
> in host, this is a keypoint for guest to do error recovery successfully.
> 
> CC: alex.williamson@redhat.com
> CC: mst@redhat.com
> Signed-off-by: Cao jin <caoj.fnst@cn.fujitsu.com>
> ---
> This is actually a RFC version(has debug lines left), and has minor changes in
> aer driver, so I think maybe it is better not to CC pci guys in this round.
> Later will do.
> 
>  drivers/pci/pcie/aer/aerdrv_core.c  | 12 ++++++-
>  drivers/vfio/pci/vfio_pci.c         | 63 +++++++++++++++++++++++++++++++++++--
>  drivers/vfio/pci/vfio_pci_private.h |  2 ++
>  3 files changed, 74 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/pci/pcie/aer/aerdrv_core.c b/drivers/pci/pcie/aer/aerdrv_core.c
> index 521e39c..289fb8e 100644
> --- a/drivers/pci/pcie/aer/aerdrv_core.c
> +++ b/drivers/pci/pcie/aer/aerdrv_core.c
> @@ -496,7 +496,17 @@ static void do_recovery(struct pci_dev *dev, int severity)
>  			"error_detected",
>  			report_error_detected);
>  
> -	if (severity == AER_FATAL) {
> +	/* vfio-pci as a general meta driver, it actually couldn't do any real
> +	 * recovery for device. It is user space driver, or device-specific
> +	 * driver in guest who should take care of the serious error recovery,
> +	 * link reset actually is one part of whole recovery. Doing reset_link
> +	 * in aer driver of host kernel for vfio-pci devices will cause many
> +	 * trouble for user space driver or guest's device-specific driver,
> +	 * for example: the serious recovery often need to read register in
> +	 * config space, but if register reading happens during link-resetting,
> +	 * it is quite possible to return invalid value like all F's, which
> +	 * will result in unpredictable error. */

Fix multi-comment style please.

> +	if (severity == AER_FATAL && strcmp(dev->driver->name, "vfio-pci")) {

You really want some flag in the device, or something similar.
Also, how do we know driver is not going away at this point?


>  		result = reset_link(dev);
>  		if (result != PCI_ERS_RESULT_RECOVERED)
>  			goto failed;
> diff --git a/drivers/vfio/pci/vfio_pci.c b/drivers/vfio/pci/vfio_pci.c
> index 712a849..aefd751 100644
> --- a/drivers/vfio/pci/vfio_pci.c
> +++ b/drivers/vfio/pci/vfio_pci.c
> @@ -535,6 +535,15 @@ static long vfio_pci_ioctl(void *device_data,
>  	struct vfio_pci_device *vdev = device_data;
>  	unsigned long minsz;
>  
> +	if (vdev->aer_recovering && (cmd == VFIO_DEVICE_SET_IRQS ||
> +	    cmd == VFIO_DEVICE_RESET || cmd == VFIO_DEVICE_PCI_HOT_RESET)) {
> +		int ret;
> +		ret = wait_for_completion_interruptible(
> +			&vdev->aer_error_completion);
> +		if (ret)
> +			return ret;
> +	}
> +
>  	if (cmd == VFIO_DEVICE_GET_INFO) {
>  		struct vfio_device_info info;
>  
> @@ -1117,6 +1126,7 @@ static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
>  	vdev->irq_type = VFIO_PCI_NUM_IRQS;
>  	mutex_init(&vdev->igate);
>  	spin_lock_init(&vdev->irqlock);
> +	init_completion(&vdev->aer_error_completion);
>  
>  	ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
>  	if (ret) {
> @@ -1176,6 +1186,9 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
>  {
>  	struct vfio_pci_device *vdev;
>  	struct vfio_device *device;
> +	u32 uncor_status = 0;
> +	unsigned int aer_cap_offset = 0;
> +	int ret;
>  
>  	device = vfio_device_get_from_dev(&pdev->dev);
>  	if (device == NULL)
> @@ -1187,10 +1200,30 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
>  		return PCI_ERS_RESULT_DISCONNECT;
>  	}
>  
> +	/* get device's uncorrectable error status as soon as possible,
> +	 * and signal it to user space. The later we read it, the possibility
> +	 * the register value is mangled grows. */
> +	aer_cap_offset = pci_find_ext_capability(vdev->pdev, PCI_EXT_CAP_ID_ERR);
> +	ret = pci_read_config_dword(vdev->pdev, aer_cap_offset + 
> +                                    PCI_ERR_UNCOR_STATUS, &uncor_status);
> +        if (ret)
> +                return PCI_ERS_RESULT_DISCONNECT;
> +
> +	pr_err("device %d got AER detect notification. uncorrectable error status = 0x%x\n", pdev->devfn, uncor_status);//to be removed
>  	mutex_lock(&vdev->igate);
> +    
> +	vdev->aer_recovering = true;
> +	reinit_completion(&vdev->aer_error_completion);
> +
> +	/* suspend config space access from user space,
> +	 * when vfio-pci's error recovery process is on */

what about access to memory etc? Do you need to suspend this as well?

> +	pci_cfg_access_trylock(vdev->pdev);

If you trylock, you need to handle failure.

>  
> -	if (vdev->err_trigger)
> -		eventfd_signal(vdev->err_trigger, 1);
> +	if (vdev->err_trigger && uncor_status) {
> +		pr_err("device %d signal uncor status to user space", pdev->devfn);//may be removed
> +		/* signal uncorrectable error status to user space */
> +		eventfd_signal(vdev->err_trigger, uncor_status);
> +        }
>  
>  	mutex_unlock(&vdev->igate);
>  
> @@ -1199,8 +1232,34 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
>  	return PCI_ERS_RESULT_CAN_RECOVER;
>  }
>  
> +static void vfio_pci_aer_resume(struct pci_dev *pdev)
> +{
> +	struct vfio_pci_device *vdev;
> +	struct vfio_device *device;
> +
> +	device = vfio_device_get_from_dev(&pdev->dev);
> +	if (device == NULL)
> +		return;
> +
> +	vdev = vfio_device_data(device);
> +	if (vdev == NULL) {
> +		vfio_device_put(device);
> +		return;
> +	}
> +
> +	/* vfio-pci's error recovery is done, time to
> +	 * resume pci config space's accesses */
> +	pci_cfg_access_unlock(vdev->pdev);

pci_cfg_access blocks accesses through sysfs.
I would not bother, focus on accesses through vfio.


> +
> +	vdev->aer_recovering = false;

don't you need locks when changing aer_recovering?

> +	complete_all(&vdev->aer_error_completion);
> +
> +	vfio_device_put(device);
> +}
> +
>  static const struct pci_error_handlers vfio_err_handlers = {
>  	.error_detected = vfio_pci_aer_err_detected,
> +	.resume         = vfio_pci_aer_resume,
>  };
>  
>  static struct pci_driver vfio_pci_driver = {
> diff --git a/drivers/vfio/pci/vfio_pci_private.h b/drivers/vfio/pci/vfio_pci_private.h
> index 8a7d546..ebf1041 100644
> --- a/drivers/vfio/pci/vfio_pci_private.h
> +++ b/drivers/vfio/pci/vfio_pci_private.h
> @@ -83,6 +83,8 @@ struct vfio_pci_device {
>  	bool			bardirty;
>  	bool			has_vga;
>  	bool			needs_reset;
> +	bool			aer_recovering;
> +	struct completion	aer_error_completion;
>  	struct pci_saved_state	*pci_saved_state;
>  	int			refcnt;
>  	struct eventfd_ctx	*err_trigger;
> -- 
> 1.8.3.1
> 
> 

[toc] | [prev] | [next] | [standalone]

#1531145

On 11/28/2016 11:00 AM, Michael S. Tsirkin wrote:
> On Sun, Nov 27, 2016 at 07:34:17PM +0800, Cao jin wrote:
>> It is user space driver's or device-specific driver's(in guest) responsbility
>> to do a serious recovery when error happened. Link-reset is one part of
>> recovery, when pci device is assigned to VM via vfio, link-reset will do
>> twice in host & guest separately, which will cause many trouble for a
>> successful recovery, so, disable the vfio-pci's link-reset in aer driver
>> in host, this is a keypoint for guest to do error recovery successfully.
>>
>> CC: alex.williamson@redhat.com
>> CC: mst@redhat.com
>> Signed-off-by: Cao jin <caoj.fnst@cn.fujitsu.com>
>> ---
>> This is actually a RFC version(has debug lines left), and has minor changes in
>> aer driver, so I think maybe it is better not to CC pci guys in this round.
>> Later will do.
>>
>>   drivers/pci/pcie/aer/aerdrv_core.c  | 12 ++++++-
>>   drivers/vfio/pci/vfio_pci.c         | 63 +++++++++++++++++++++++++++++++++++--
>>   drivers/vfio/pci/vfio_pci_private.h |  2 ++
>>   3 files changed, 74 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/pci/pcie/aer/aerdrv_core.c b/drivers/pci/pcie/aer/aerdrv_core.c
>> index 521e39c..289fb8e 100644
>> --- a/drivers/pci/pcie/aer/aerdrv_core.c
>> +++ b/drivers/pci/pcie/aer/aerdrv_core.c
>> @@ -496,7 +496,17 @@ static void do_recovery(struct pci_dev *dev, int severity)
>>   			"error_detected",
>>   			report_error_detected);
>>
>> -	if (severity == AER_FATAL) {
>> +	/* vfio-pci as a general meta driver, it actually couldn't do any real
>> +	 * recovery for device. It is user space driver, or device-specific
>> +	 * driver in guest who should take care of the serious error recovery,
>> +	 * link reset actually is one part of whole recovery. Doing reset_link
>> +	 * in aer driver of host kernel for vfio-pci devices will cause many
>> +	 * trouble for user space driver or guest's device-specific driver,
>> +	 * for example: the serious recovery often need to read register in
>> +	 * config space, but if register reading happens during link-resetting,
>> +	 * it is quite possible to return invalid value like all F's, which
>> +	 * will result in unpredictable error. */
>
> Fix multi-comment style please.
>
>> +	if (severity == AER_FATAL && strcmp(dev->driver->name, "vfio-pci")) {
>
> You really want some flag in the device, or something similar.
> Also, how do we know driver is not going away at this point?
>

I didn't think of this condition, and I don't quite follow how would 
driver go away?(device has error happened, then is removed?)

>>   		result = reset_link(dev);
>>   		if (result != PCI_ERS_RESULT_RECOVERED)
>>   			goto failed;

>> @@ -1187,10 +1200,30 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
>>   		return PCI_ERS_RESULT_DISCONNECT;
>>   	}
>>
>> +	/* get device's uncorrectable error status as soon as possible,
>> +	 * and signal it to user space. The later we read it, the possibility
>> +	 * the register value is mangled grows. */
>> +	aer_cap_offset = pci_find_ext_capability(vdev->pdev, PCI_EXT_CAP_ID_ERR);
>> +	ret = pci_read_config_dword(vdev->pdev, aer_cap_offset +
>> +                                    PCI_ERR_UNCOR_STATUS, &uncor_status);
>> +        if (ret)
>> +                return PCI_ERS_RESULT_DISCONNECT;
>> +
>> +	pr_err("device %d got AER detect notification. uncorrectable error status = 0x%x\n", pdev->devfn, uncor_status);//to be removed
>>   	mutex_lock(&vdev->igate);
>> +
>> +	vdev->aer_recovering = true;
>> +	reinit_completion(&vdev->aer_error_completion);
>> +
>> +	/* suspend config space access from user space,
>> +	 * when vfio-pci's error recovery process is on */
>
> what about access to memory etc? Do you need to suspend this as well?
>

Yes, this question came into my mind a little bit, but I didn't see some 
existing APIs like pci_cfg_access_xxx which can help to do this.(I am 
still not familiar with kernel)

>> +	pci_cfg_access_trylock(vdev->pdev);
>
> If you trylock, you need to handle failure.

try lock returns 0 if access is already locked, 1 otherwise. Is it 
necessary to check its return value?


-- 
Sincerely,
Cao jin

[toc] | [prev] | [next] | [standalone]

#1532907

On Mon, Nov 28, 2016 at 05:32:15PM +0800, Cao jin wrote:
> 
> 
> On 11/28/2016 11:00 AM, Michael S. Tsirkin wrote:
> > On Sun, Nov 27, 2016 at 07:34:17PM +0800, Cao jin wrote:
> > > It is user space driver's or device-specific driver's(in guest) responsbility
> > > to do a serious recovery when error happened. Link-reset is one part of
> > > recovery, when pci device is assigned to VM via vfio, link-reset will do
> > > twice in host & guest separately, which will cause many trouble for a
> > > successful recovery, so, disable the vfio-pci's link-reset in aer driver
> > > in host, this is a keypoint for guest to do error recovery successfully.
> > > 
> > > CC: alex.williamson@redhat.com
> > > CC: mst@redhat.com
> > > Signed-off-by: Cao jin <caoj.fnst@cn.fujitsu.com>
> > > ---
> > > This is actually a RFC version(has debug lines left), and has minor changes in
> > > aer driver, so I think maybe it is better not to CC pci guys in this round.
> > > Later will do.
> > > 
> > >   drivers/pci/pcie/aer/aerdrv_core.c  | 12 ++++++-
> > >   drivers/vfio/pci/vfio_pci.c         | 63 +++++++++++++++++++++++++++++++++++--
> > >   drivers/vfio/pci/vfio_pci_private.h |  2 ++
> > >   3 files changed, 74 insertions(+), 3 deletions(-)
> > > 
> > > diff --git a/drivers/pci/pcie/aer/aerdrv_core.c b/drivers/pci/pcie/aer/aerdrv_core.c
> > > index 521e39c..289fb8e 100644
> > > --- a/drivers/pci/pcie/aer/aerdrv_core.c
> > > +++ b/drivers/pci/pcie/aer/aerdrv_core.c
> > > @@ -496,7 +496,17 @@ static void do_recovery(struct pci_dev *dev, int severity)
> > >   			"error_detected",
> > >   			report_error_detected);
> > > 
> > > -	if (severity == AER_FATAL) {
> > > +	/* vfio-pci as a general meta driver, it actually couldn't do any real
> > > +	 * recovery for device. It is user space driver, or device-specific
> > > +	 * driver in guest who should take care of the serious error recovery,
> > > +	 * link reset actually is one part of whole recovery. Doing reset_link
> > > +	 * in aer driver of host kernel for vfio-pci devices will cause many
> > > +	 * trouble for user space driver or guest's device-specific driver,
> > > +	 * for example: the serious recovery often need to read register in
> > > +	 * config space, but if register reading happens during link-resetting,
> > > +	 * it is quite possible to return invalid value like all F's, which
> > > +	 * will result in unpredictable error. */
> > 
> > Fix multi-comment style please.
> > 
> > > +	if (severity == AER_FATAL && strcmp(dev->driver->name, "vfio-pci")) {
> > 
> > You really want some flag in the device, or something similar.
> > Also, how do we know driver is not going away at this point?
> > 
> 
> I didn't think of this condition, and I don't quite follow how would driver
> go away?(device has error happened, then is removed?)

Yes - hotplug request detected. Does something prevent this?

> > >   		result = reset_link(dev);
> > >   		if (result != PCI_ERS_RESULT_RECOVERED)
> > >   			goto failed;
> 
> > > @@ -1187,10 +1200,30 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
> > >   		return PCI_ERS_RESULT_DISCONNECT;
> > >   	}
> > > 
> > > +	/* get device's uncorrectable error status as soon as possible,
> > > +	 * and signal it to user space. The later we read it, the possibility
> > > +	 * the register value is mangled grows. */
> > > +	aer_cap_offset = pci_find_ext_capability(vdev->pdev, PCI_EXT_CAP_ID_ERR);
> > > +	ret = pci_read_config_dword(vdev->pdev, aer_cap_offset +
> > > +                                    PCI_ERR_UNCOR_STATUS, &uncor_status);
> > > +        if (ret)
> > > +                return PCI_ERS_RESULT_DISCONNECT;
> > > +
> > > +	pr_err("device %d got AER detect notification. uncorrectable error status = 0x%x\n", pdev->devfn, uncor_status);//to be removed
> > >   	mutex_lock(&vdev->igate);
> > > +
> > > +	vdev->aer_recovering = true;
> > > +	reinit_completion(&vdev->aer_error_completion);
> > > +
> > > +	/* suspend config space access from user space,
> > > +	 * when vfio-pci's error recovery process is on */
> > 
> > what about access to memory etc? Do you need to suspend this as well?
> > 
> 
> Yes, this question came into my mind a little bit, but I didn't see some
> existing APIs like pci_cfg_access_xxx which can help to do this.(I am still
> not familiar with kernel)

This isn't easy to do at all.


> > > +	pci_cfg_access_trylock(vdev->pdev);
> > 
> > If you trylock, you need to handle failure.
> 
> try lock returns 0 if access is already locked, 1 otherwise. Is it necessary
> to check its return value?

Locked by whom? You blissfully access as if it's locked by you.

> 
> -- 
> Sincerely,
> Cao jin
> 

[toc] | [prev] | [next] | [standalone]

#1534096

On 11/30/2016 09:46 AM, Michael S. Tsirkin wrote:
> On Mon, Nov 28, 2016 at 05:32:15PM +0800, Cao jin wrote:
>>
>>

>>>
>>>> +	if (severity == AER_FATAL && strcmp(dev->driver->name, "vfio-pci")) {
>>>
>>> You really want some flag in the device, or something similar.
>>> Also, how do we know driver is not going away at this point?
>>>
>>
>> I didn't think of this condition, and I don't quite follow how would driver
>> go away?(device has error happened, then is removed?)
> 
> Yes - hotplug request detected. Does something prevent this?
> 

I wasn't realized there would be possible to have hotplug happened
during error recovery. But, it seems is the aer driver's thing to
consider it?

>>>>   		result = reset_link(dev);
>>>>   		if (result != PCI_ERS_RESULT_RECOVERED)
>>>>   			goto failed;
>>
>>>> @@ -1187,10 +1200,30 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
>>>>   		return PCI_ERS_RESULT_DISCONNECT;
>>>>   	}
>>>>
>>>> +	/* get device's uncorrectable error status as soon as possible,
>>>> +	 * and signal it to user space. The later we read it, the possibility
>>>> +	 * the register value is mangled grows. */
>>>> +	aer_cap_offset = pci_find_ext_capability(vdev->pdev, PCI_EXT_CAP_ID_ERR);
>>>> +	ret = pci_read_config_dword(vdev->pdev, aer_cap_offset +
>>>> +                                    PCI_ERR_UNCOR_STATUS, &uncor_status);
>>>> +        if (ret)
>>>> +                return PCI_ERS_RESULT_DISCONNECT;
>>>> +
>>>> +	pr_err("device %d got AER detect notification. uncorrectable error status = 0x%x\n", pdev->devfn, uncor_status);//to be removed
>>>>   	mutex_lock(&vdev->igate);
>>>> +
>>>> +	vdev->aer_recovering = true;
>>>> +	reinit_completion(&vdev->aer_error_completion);
>>>> +
>>>> +	/* suspend config space access from user space,
>>>> +	 * when vfio-pci's error recovery process is on */
>>>
>>> what about access to memory etc? Do you need to suspend this as well?
>>>
>>
>> Yes, this question came into my mind a little bit, but I didn't see some
>> existing APIs like pci_cfg_access_xxx which can help to do this.(I am still
>> not familiar with kernel)
> 
> This isn't easy to do at all.
> 

I guess we can use completion in vfio_pci_rw to block all kinds of
access during vfio's error recovery. But from another perspective, now
that we have disabled reset_link in host, I think the access to device
could be performed normally. To me, the benefit of using pci_cfg_access
is: we won't bother to do "wait 3s to do guest's link reset"

> 
>>>> +	pci_cfg_access_trylock(vdev->pdev);
>>>
>>> If you trylock, you need to handle failure.
>>
>> try lock returns 0 if access is already locked, 1 otherwise. Is it necessary
>> to check its return value?
> 
> Locked by whom? You blissfully access as if it's locked by you.
> 
>>
>> -- 
>> Sincerely,
>> Cao jin
>>
> 
> 
> .
> 

-- 
Sincerely,
Cao jin

[toc] | [prev] | [next] | [standalone]

#1533788

On Sun, 27 Nov 2016 19:34:17 +0800
Cao jin <caoj.fnst@cn.fujitsu.com> wrote:

> It is user space driver's or device-specific driver's(in guest) responsbility
> to do a serious recovery when error happened. Link-reset is one part of
> recovery, when pci device is assigned to VM via vfio, link-reset will do
> twice in host & guest separately, which will cause many trouble for a
> successful recovery, so, disable the vfio-pci's link-reset in aer driver
> in host, this is a keypoint for guest to do error recovery successfully.
> 
> CC: alex.williamson@redhat.com
> CC: mst@redhat.com
> Signed-off-by: Cao jin <caoj.fnst@cn.fujitsu.com>
> ---
> This is actually a RFC version(has debug lines left), and has minor changes in
> aer driver, so I think maybe it is better not to CC pci guys in this round.
> Later will do.
> 
>  drivers/pci/pcie/aer/aerdrv_core.c  | 12 ++++++-
>  drivers/vfio/pci/vfio_pci.c         | 63 +++++++++++++++++++++++++++++++++++--
>  drivers/vfio/pci/vfio_pci_private.h |  2 ++
>  3 files changed, 74 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/pci/pcie/aer/aerdrv_core.c b/drivers/pci/pcie/aer/aerdrv_core.c
> index 521e39c..289fb8e 100644
> --- a/drivers/pci/pcie/aer/aerdrv_core.c
> +++ b/drivers/pci/pcie/aer/aerdrv_core.c
> @@ -496,7 +496,17 @@ static void do_recovery(struct pci_dev *dev, int severity)
>  			"error_detected",
>  			report_error_detected);
>  
> -	if (severity == AER_FATAL) {
> +	/* vfio-pci as a general meta driver, it actually couldn't do any real
> +	 * recovery for device. It is user space driver, or device-specific
> +	 * driver in guest who should take care of the serious error recovery,
> +	 * link reset actually is one part of whole recovery. Doing reset_link
> +	 * in aer driver of host kernel for vfio-pci devices will cause many
> +	 * trouble for user space driver or guest's device-specific driver,
> +	 * for example: the serious recovery often need to read register in
> +	 * config space, but if register reading happens during link-resetting,
> +	 * it is quite possible to return invalid value like all F's, which
> +	 * will result in unpredictable error. */
> +	if (severity == AER_FATAL && strcmp(dev->driver->name, "vfio-pci")) {
>  		result = reset_link(dev);
>  		if (result != PCI_ERS_RESULT_RECOVERED)
>  			goto failed;

This is not acceptable.  If we want to make a path through AER recovery
that does not reset the link, there should be a way for the driver to
request it.  Testing the driver name is a horrible hack.  The other
question is what guarantees does vfio make that the device does get
reset?  If an AER fault occurs and the user doesn't do a reset, what
happens when that device is released and a host driver tries to make
use of it?  The user makes no commitment to do a reset and there are
only limited configurations where we even allow the user to perform a
reset.

> diff --git a/drivers/vfio/pci/vfio_pci.c b/drivers/vfio/pci/vfio_pci.c
> index 712a849..aefd751 100644
> --- a/drivers/vfio/pci/vfio_pci.c
> +++ b/drivers/vfio/pci/vfio_pci.c
> @@ -535,6 +535,15 @@ static long vfio_pci_ioctl(void *device_data,
>  	struct vfio_pci_device *vdev = device_data;
>  	unsigned long minsz;
>  
> +	if (vdev->aer_recovering && (cmd == VFIO_DEVICE_SET_IRQS ||
> +	    cmd == VFIO_DEVICE_RESET || cmd == VFIO_DEVICE_PCI_HOT_RESET)) {
> +		int ret;
> +		ret = wait_for_completion_interruptible(
> +			&vdev->aer_error_completion);
> +		if (ret)
> +			return ret;
> +	}
> +
>  	if (cmd == VFIO_DEVICE_GET_INFO) {
>  		struct vfio_device_info info;
>  
> @@ -1117,6 +1126,7 @@ static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
>  	vdev->irq_type = VFIO_PCI_NUM_IRQS;
>  	mutex_init(&vdev->igate);
>  	spin_lock_init(&vdev->irqlock);
> +	init_completion(&vdev->aer_error_completion);
>  
>  	ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
>  	if (ret) {
> @@ -1176,6 +1186,9 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
>  {
>  	struct vfio_pci_device *vdev;
>  	struct vfio_device *device;
> +	u32 uncor_status = 0;
> +	unsigned int aer_cap_offset = 0;

Unnecessary initialization.

> +	int ret;
>  
>  	device = vfio_device_get_from_dev(&pdev->dev);
>  	if (device == NULL)
> @@ -1187,10 +1200,30 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
>  		return PCI_ERS_RESULT_DISCONNECT;
>  	}
>  
> +	/* get device's uncorrectable error status as soon as possible,
> +	 * and signal it to user space. The later we read it, the possibility
> +	 * the register value is mangled grows. */

Bad comment style (throughout).

> +	aer_cap_offset = pci_find_ext_capability(vdev->pdev, PCI_EXT_CAP_ID_ERR);
> +	ret = pci_read_config_dword(vdev->pdev, aer_cap_offset + 
> +                                    PCI_ERR_UNCOR_STATUS, &uncor_status);
> +        if (ret)
> +                return PCI_ERS_RESULT_DISCONNECT;
> +
> +	pr_err("device %d got AER detect notification. uncorrectable error status = 0x%x\n", pdev->devfn, uncor_status);//to be removed
>  	mutex_lock(&vdev->igate);
> +    
> +	vdev->aer_recovering = true;
> +	reinit_completion(&vdev->aer_error_completion);
> +
> +	/* suspend config space access from user space,
> +	 * when vfio-pci's error recovery process is on */
> +	pci_cfg_access_trylock(vdev->pdev);
>  
> -	if (vdev->err_trigger)
> -		eventfd_signal(vdev->err_trigger, 1);
> +	if (vdev->err_trigger && uncor_status) {
> +		pr_err("device %d signal uncor status to user space", pdev->devfn);//may be removed
> +		/* signal uncorrectable error status to user space */
> +		eventfd_signal(vdev->err_trigger, uncor_status);
> +        }

} else... what?  By bypassing the AER link reset we've assumed
responsibility for resetting the device.  Even if we signal the user,
what guarantee do we have that the device is recovered?

>  
>  	mutex_unlock(&vdev->igate);
>  
> @@ -1199,8 +1232,34 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
>  	return PCI_ERS_RESULT_CAN_RECOVER;
>  }
>  
> +static void vfio_pci_aer_resume(struct pci_dev *pdev)
> +{
> +	struct vfio_pci_device *vdev;
> +	struct vfio_device *device;
> +
> +	device = vfio_device_get_from_dev(&pdev->dev);
> +	if (device == NULL)
> +		return;
> +
> +	vdev = vfio_device_data(device);
> +	if (vdev == NULL) {
> +		vfio_device_put(device);
> +		return;
> +	}
> +
> +	/* vfio-pci's error recovery is done, time to
> +	 * resume pci config space's accesses */
> +	pci_cfg_access_unlock(vdev->pdev);
> +
> +	vdev->aer_recovering = false;
> +	complete_all(&vdev->aer_error_completion);
> +
> +	vfio_device_put(device);
> +}
> +
>  static const struct pci_error_handlers vfio_err_handlers = {
>  	.error_detected = vfio_pci_aer_err_detected,
> +	.resume         = vfio_pci_aer_resume,
>  };
>  
>  static struct pci_driver vfio_pci_driver = {
> diff --git a/drivers/vfio/pci/vfio_pci_private.h b/drivers/vfio/pci/vfio_pci_private.h
> index 8a7d546..ebf1041 100644
> --- a/drivers/vfio/pci/vfio_pci_private.h
> +++ b/drivers/vfio/pci/vfio_pci_private.h
> @@ -83,6 +83,8 @@ struct vfio_pci_device {
>  	bool			bardirty;
>  	bool			has_vga;
>  	bool			needs_reset;
> +	bool			aer_recovering;
> +	struct completion	aer_error_completion;
>  	struct pci_saved_state	*pci_saved_state;
>  	int			refcnt;
>  	struct eventfd_ctx	*err_trigger;

[toc] | [prev] | [next] | [standalone]

#1533809

On Wed, Nov 30, 2016 at 09:04:13PM -0700, Alex Williamson wrote:
> On Sun, 27 Nov 2016 19:34:17 +0800
> Cao jin <caoj.fnst@cn.fujitsu.com> wrote:
> 
> > It is user space driver's or device-specific driver's(in guest) responsbility
> > to do a serious recovery when error happened. Link-reset is one part of
> > recovery, when pci device is assigned to VM via vfio, link-reset will do
> > twice in host & guest separately, which will cause many trouble for a
> > successful recovery, so, disable the vfio-pci's link-reset in aer driver
> > in host, this is a keypoint for guest to do error recovery successfully.
> > 
> > CC: alex.williamson@redhat.com
> > CC: mst@redhat.com
> > Signed-off-by: Cao jin <caoj.fnst@cn.fujitsu.com>
> > ---
> > This is actually a RFC version(has debug lines left), and has minor changes in
> > aer driver, so I think maybe it is better not to CC pci guys in this round.
> > Later will do.
> > 
> >  drivers/pci/pcie/aer/aerdrv_core.c  | 12 ++++++-
> >  drivers/vfio/pci/vfio_pci.c         | 63 +++++++++++++++++++++++++++++++++++--
> >  drivers/vfio/pci/vfio_pci_private.h |  2 ++
> >  3 files changed, 74 insertions(+), 3 deletions(-)
> > 
> > diff --git a/drivers/pci/pcie/aer/aerdrv_core.c b/drivers/pci/pcie/aer/aerdrv_core.c
> > index 521e39c..289fb8e 100644
> > --- a/drivers/pci/pcie/aer/aerdrv_core.c
> > +++ b/drivers/pci/pcie/aer/aerdrv_core.c
> > @@ -496,7 +496,17 @@ static void do_recovery(struct pci_dev *dev, int severity)
> >  			"error_detected",
> >  			report_error_detected);
> >  
> > -	if (severity == AER_FATAL) {
> > +	/* vfio-pci as a general meta driver, it actually couldn't do any real
> > +	 * recovery for device. It is user space driver, or device-specific
> > +	 * driver in guest who should take care of the serious error recovery,
> > +	 * link reset actually is one part of whole recovery. Doing reset_link
> > +	 * in aer driver of host kernel for vfio-pci devices will cause many
> > +	 * trouble for user space driver or guest's device-specific driver,
> > +	 * for example: the serious recovery often need to read register in
> > +	 * config space, but if register reading happens during link-resetting,
> > +	 * it is quite possible to return invalid value like all F's, which
> > +	 * will result in unpredictable error. */
> > +	if (severity == AER_FATAL && strcmp(dev->driver->name, "vfio-pci")) {
> >  		result = reset_link(dev);
> >  		if (result != PCI_ERS_RESULT_RECOVERED)
> >  			goto failed;
> 
> This is not acceptable.  If we want to make a path through AER recovery
> that does not reset the link, there should be a way for the driver to
> request it.  Testing the driver name is a horrible hack.

Right, just set a flag or something.

> The other
> question is what guarantees does vfio make that the device does get
> reset?  If an AER fault occurs and the user doesn't do a reset, what
> happens when that device is released and a host driver tries to make
> use of it?  The user makes no commitment to do a reset and there are
> only limited configurations where we even allow the user to perform a
> reset.

Let's try to work out what can be done a bit more constructively.
Doesn't vfio already reset device on release?
How about detecting there was an unhandled AER event
and doing link reset at that point?



> > diff --git a/drivers/vfio/pci/vfio_pci.c b/drivers/vfio/pci/vfio_pci.c
> > index 712a849..aefd751 100644
> > --- a/drivers/vfio/pci/vfio_pci.c
> > +++ b/drivers/vfio/pci/vfio_pci.c
> > @@ -535,6 +535,15 @@ static long vfio_pci_ioctl(void *device_data,
> >  	struct vfio_pci_device *vdev = device_data;
> >  	unsigned long minsz;
> >  
> > +	if (vdev->aer_recovering && (cmd == VFIO_DEVICE_SET_IRQS ||
> > +	    cmd == VFIO_DEVICE_RESET || cmd == VFIO_DEVICE_PCI_HOT_RESET)) {
> > +		int ret;
> > +		ret = wait_for_completion_interruptible(
> > +			&vdev->aer_error_completion);
> > +		if (ret)
> > +			return ret;
> > +	}
> > +
> >  	if (cmd == VFIO_DEVICE_GET_INFO) {
> >  		struct vfio_device_info info;
> >  
> > @@ -1117,6 +1126,7 @@ static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
> >  	vdev->irq_type = VFIO_PCI_NUM_IRQS;
> >  	mutex_init(&vdev->igate);
> >  	spin_lock_init(&vdev->irqlock);
> > +	init_completion(&vdev->aer_error_completion);
> >  
> >  	ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
> >  	if (ret) {
> > @@ -1176,6 +1186,9 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
> >  {
> >  	struct vfio_pci_device *vdev;
> >  	struct vfio_device *device;
> > +	u32 uncor_status = 0;
> > +	unsigned int aer_cap_offset = 0;
> 
> Unnecessary initialization.
> 
> > +	int ret;
> >  
> >  	device = vfio_device_get_from_dev(&pdev->dev);
> >  	if (device == NULL)
> > @@ -1187,10 +1200,30 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
> >  		return PCI_ERS_RESULT_DISCONNECT;
> >  	}
> >  
> > +	/* get device's uncorrectable error status as soon as possible,
> > +	 * and signal it to user space. The later we read it, the possibility
> > +	 * the register value is mangled grows. */
> 
> Bad comment style (throughout).
> 
> > +	aer_cap_offset = pci_find_ext_capability(vdev->pdev, PCI_EXT_CAP_ID_ERR);
> > +	ret = pci_read_config_dword(vdev->pdev, aer_cap_offset + 
> > +                                    PCI_ERR_UNCOR_STATUS, &uncor_status);
> > +        if (ret)
> > +                return PCI_ERS_RESULT_DISCONNECT;
> > +
> > +	pr_err("device %d got AER detect notification. uncorrectable error status = 0x%x\n", pdev->devfn, uncor_status);//to be removed
> >  	mutex_lock(&vdev->igate);
> > +    
> > +	vdev->aer_recovering = true;
> > +	reinit_completion(&vdev->aer_error_completion);
> > +
> > +	/* suspend config space access from user space,
> > +	 * when vfio-pci's error recovery process is on */
> > +	pci_cfg_access_trylock(vdev->pdev);
> >  
> > -	if (vdev->err_trigger)
> > -		eventfd_signal(vdev->err_trigger, 1);
> > +	if (vdev->err_trigger && uncor_status) {
> > +		pr_err("device %d signal uncor status to user space", pdev->devfn);//may be removed
> > +		/* signal uncorrectable error status to user space */
> > +		eventfd_signal(vdev->err_trigger, uncor_status);

What does this mean that we trigger with uncor_status as opposed to 1?

> > +        }
> 
> } else... what?  By bypassing the AER link reset we've assumed
> responsibility for resetting the device.  Even if we signal the user,
> what guarantee do we have that the device is recovered?

So set flag and recover ?

> >  
> >  	mutex_unlock(&vdev->igate);
> >  
> > @@ -1199,8 +1232,34 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
> >  	return PCI_ERS_RESULT_CAN_RECOVER;
> >  }
> >  
> > +static void vfio_pci_aer_resume(struct pci_dev *pdev)
> > +{
> > +	struct vfio_pci_device *vdev;
> > +	struct vfio_device *device;
> > +
> > +	device = vfio_device_get_from_dev(&pdev->dev);
> > +	if (device == NULL)
> > +		return;
> > +
> > +	vdev = vfio_device_data(device);
> > +	if (vdev == NULL) {
> > +		vfio_device_put(device);
> > +		return;
> > +	}
> > +
> > +	/* vfio-pci's error recovery is done, time to
> > +	 * resume pci config space's accesses */
> > +	pci_cfg_access_unlock(vdev->pdev);
> > +
> > +	vdev->aer_recovering = false;
> > +	complete_all(&vdev->aer_error_completion);
> > +
> > +	vfio_device_put(device);
> > +}
> > +
> >  static const struct pci_error_handlers vfio_err_handlers = {
> >  	.error_detected = vfio_pci_aer_err_detected,
> > +	.resume         = vfio_pci_aer_resume,
> >  };
> >  
> >  static struct pci_driver vfio_pci_driver = {
> > diff --git a/drivers/vfio/pci/vfio_pci_private.h b/drivers/vfio/pci/vfio_pci_private.h
> > index 8a7d546..ebf1041 100644
> > --- a/drivers/vfio/pci/vfio_pci_private.h
> > +++ b/drivers/vfio/pci/vfio_pci_private.h
> > @@ -83,6 +83,8 @@ struct vfio_pci_device {
> >  	bool			bardirty;
> >  	bool			has_vga;
> >  	bool			needs_reset;
> > +	bool			aer_recovering;
> > +	struct completion	aer_error_completion;
> >  	struct pci_saved_state	*pci_saved_state;
> >  	int			refcnt;
> >  	struct eventfd_ctx	*err_trigger;

[toc] | [prev] | [next] | [standalone]

#1534094

On 12/01/2016 12:51 PM, Michael S. Tsirkin wrote:
> On Wed, Nov 30, 2016 at 09:04:13PM -0700, Alex Williamson wrote:
>> On Sun, 27 Nov 2016 19:34:17 +0800
>> Cao jin <caoj.fnst@cn.fujitsu.com> wrote:
>>

>>> @@ -1187,10 +1200,30 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
>>>  		return PCI_ERS_RESULT_DISCONNECT;
>>>  	}
>>>  
>>> +	/* get device's uncorrectable error status as soon as possible,
>>> +	 * and signal it to user space. The later we read it, the possibility
>>> +	 * the register value is mangled grows. */
>>
>> Bad comment style (throughout).
>>
>>> +	aer_cap_offset = pci_find_ext_capability(vdev->pdev, PCI_EXT_CAP_ID_ERR);
>>> +	ret = pci_read_config_dword(vdev->pdev, aer_cap_offset + 
>>> +                                    PCI_ERR_UNCOR_STATUS, &uncor_status);
>>> +        if (ret)
>>> +                return PCI_ERS_RESULT_DISCONNECT;
>>> +
>>> +	pr_err("device %d got AER detect notification. uncorrectable error status = 0x%x\n", pdev->devfn, uncor_status);//to be removed
>>>  	mutex_lock(&vdev->igate);
>>> +    
>>> +	vdev->aer_recovering = true;
>>> +	reinit_completion(&vdev->aer_error_completion);
>>> +
>>> +	/* suspend config space access from user space,
>>> +	 * when vfio-pci's error recovery process is on */
>>> +	pci_cfg_access_trylock(vdev->pdev);
>>>  
>>> -	if (vdev->err_trigger)
>>> -		eventfd_signal(vdev->err_trigger, 1);
>>> +	if (vdev->err_trigger && uncor_status) {
>>> +		pr_err("device %d signal uncor status to user space", pdev->devfn);//may be removed
>>> +		/* signal uncorrectable error status to user space */
>>> +		eventfd_signal(vdev->err_trigger, uncor_status);
> 
> What does this mean that we trigger with uncor_status as opposed to 1?
> 

I guess you missed the changelog in qemu patchset's cover letter, see if
it helps(copy from cover letter):

5. Change what eventfd signals(check vfio driver patch). Before,
vfio-pci driver add 1 to the counter, which doesn't have meaning, just
for notification. Now, vfio-pci's error detected handler read the
uncorrectable error status and signal it to qemu.

Why? When testing previous version(host aer driver still reset link),
found that there is quite possibility that register reading returns the
invalid value 0xFFFFFFFF, which results in all(2 in my environment)
qemu's vfio-pci function send AER to root port while I only inject error
into one function. This is unreasonable, and I think it is the result of
reading during device reset. Previous patch does considered to find the
real function who has error happened, but won't work under the situation
I found. So move the register reading as early as possible would be the
nature thoughts, and it is moved into vfio-pci driver. Although now
reset_link is disabled in aer driver, get the uncor error status as
early as possible still make sense, for example: if user space driver
does device reset once receiving the error notification, and then read
register.

-- 
Sincerely,
Cao jin

[toc] | [prev] | [next] | [standalone]

#1536637

On Thu, Dec 01, 2016 at 09:40:23PM +0800, Cao jin wrote:
> 
> 
> On 12/01/2016 12:51 PM, Michael S. Tsirkin wrote:
> > On Wed, Nov 30, 2016 at 09:04:13PM -0700, Alex Williamson wrote:
> >> On Sun, 27 Nov 2016 19:34:17 +0800
> >> Cao jin <caoj.fnst@cn.fujitsu.com> wrote:
> >>
> 
> >>> @@ -1187,10 +1200,30 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
> >>>  		return PCI_ERS_RESULT_DISCONNECT;
> >>>  	}
> >>>  
> >>> +	/* get device's uncorrectable error status as soon as possible,
> >>> +	 * and signal it to user space. The later we read it, the possibility
> >>> +	 * the register value is mangled grows. */
> >>
> >> Bad comment style (throughout).
> >>
> >>> +	aer_cap_offset = pci_find_ext_capability(vdev->pdev, PCI_EXT_CAP_ID_ERR);
> >>> +	ret = pci_read_config_dword(vdev->pdev, aer_cap_offset + 
> >>> +                                    PCI_ERR_UNCOR_STATUS, &uncor_status);
> >>> +        if (ret)
> >>> +                return PCI_ERS_RESULT_DISCONNECT;
> >>> +
> >>> +	pr_err("device %d got AER detect notification. uncorrectable error status = 0x%x\n", pdev->devfn, uncor_status);//to be removed
> >>>  	mutex_lock(&vdev->igate);
> >>> +    
> >>> +	vdev->aer_recovering = true;
> >>> +	reinit_completion(&vdev->aer_error_completion);
> >>> +
> >>> +	/* suspend config space access from user space,
> >>> +	 * when vfio-pci's error recovery process is on */
> >>> +	pci_cfg_access_trylock(vdev->pdev);
> >>>  
> >>> -	if (vdev->err_trigger)
> >>> -		eventfd_signal(vdev->err_trigger, 1);
> >>> +	if (vdev->err_trigger && uncor_status) {
> >>> +		pr_err("device %d signal uncor status to user space", pdev->devfn);//may be removed
> >>> +		/* signal uncorrectable error status to user space */
> >>> +		eventfd_signal(vdev->err_trigger, uncor_status);
> > 
> > What does this mean that we trigger with uncor_status as opposed to 1?
> > 
> 
> I guess you missed the changelog in qemu patchset's cover letter, see if
> it helps(copy from cover letter):
> 
> 5. Change what eventfd signals(check vfio driver patch). Before,
> vfio-pci driver add 1 to the counter, which doesn't have meaning, just
> for notification. Now, vfio-pci's error detected handler read the
> uncorrectable error status and signal it to qemu.

I don't think you can use an eventfd to send values like this.

eventfd does a sum of these values, so sending e.g.
value 2 will look the same as sending value 1 twice.



> Why? When testing previous version(host aer driver still reset link),
> found that there is quite possibility that register reading returns the
> invalid value 0xFFFFFFFF, which results in all(2 in my environment)
> qemu's vfio-pci function send AER to root port while I only inject error
> into one function. This is unreasonable, and I think it is the result of
> reading during device reset.
>
> Previous patch does considered to find the
> real function who has error happened, but won't work under the situation
> I found. So move the register reading as early as possible would be the
> nature thoughts, and it is moved into vfio-pci driver. Although now
> reset_link is disabled in aer driver, get the uncor error status as
> early as possible still make sense, for example: if user space driver
> does device reset once receiving the error notification, and then read
> register.

I had trouble understanding the above. Let me ask you this:
should we try to avoid triggering uncorrectable errors?
Aren't any such errors likely to crash guests?

> -- 
> Sincerely,
> Cao jin
> 

[toc] | [prev] | [next] | [standalone]

#1536686

On 12/06/2016 11:46 AM, Michael S. Tsirkin wrote:
> On Thu, Dec 01, 2016 at 09:40:23PM +0800, Cao jin wrote:
>>
>>
>> On 12/01/2016 12:51 PM, Michael S. Tsirkin wrote:
>>> On Wed, Nov 30, 2016 at 09:04:13PM -0700, Alex Williamson wrote:
>>>> On Sun, 27 Nov 2016 19:34:17 +0800
>>>> Cao jin <caoj.fnst@cn.fujitsu.com> wrote:
>>>>
>>
>>>>> @@ -1187,10 +1200,30 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
>>>>>  		return PCI_ERS_RESULT_DISCONNECT;
>>>>>  	}
>>>>>  
>>>>> +	/* get device's uncorrectable error status as soon as possible,
>>>>> +	 * and signal it to user space. The later we read it, the possibility
>>>>> +	 * the register value is mangled grows. */
>>>>
>>>> Bad comment style (throughout).
>>>>
>>>>> +	aer_cap_offset = pci_find_ext_capability(vdev->pdev, PCI_EXT_CAP_ID_ERR);
>>>>> +	ret = pci_read_config_dword(vdev->pdev, aer_cap_offset + 
>>>>> +                                    PCI_ERR_UNCOR_STATUS, &uncor_status);
>>>>> +        if (ret)
>>>>> +                return PCI_ERS_RESULT_DISCONNECT;
>>>>> +
>>>>> +	pr_err("device %d got AER detect notification. uncorrectable error status = 0x%x\n", pdev->devfn, uncor_status);//to be removed
>>>>>  	mutex_lock(&vdev->igate);
>>>>> +    
>>>>> +	vdev->aer_recovering = true;
>>>>> +	reinit_completion(&vdev->aer_error_completion);
>>>>> +
>>>>> +	/* suspend config space access from user space,
>>>>> +	 * when vfio-pci's error recovery process is on */
>>>>> +	pci_cfg_access_trylock(vdev->pdev);
>>>>>  
>>>>> -	if (vdev->err_trigger)
>>>>> -		eventfd_signal(vdev->err_trigger, 1);
>>>>> +	if (vdev->err_trigger && uncor_status) {
>>>>> +		pr_err("device %d signal uncor status to user space", pdev->devfn);//may be removed
>>>>> +		/* signal uncorrectable error status to user space */
>>>>> +		eventfd_signal(vdev->err_trigger, uncor_status);
>>>
>>> What does this mean that we trigger with uncor_status as opposed to 1?
>>>
>>
>> I guess you missed the changelog in qemu patchset's cover letter, see if
>> it helps(copy from cover letter):
>>
>> 5. Change what eventfd signals(check vfio driver patch). Before,
>> vfio-pci driver add 1 to the counter, which doesn't have meaning, just
>> for notification. Now, vfio-pci's error detected handler read the
>> uncorrectable error status and signal it to qemu.
> 
> I don't think you can use an eventfd to send values like this.
> 
> eventfd does a sum of these values, so sending e.g.
> value 2 will look the same as sending value 1 twice.
> 

Yes, eventfd has a uint_64 counter, and eventfd_signal does a sum.

manpage of "eventfd" says:

The semantics of read(2) depend on whether the eventfd counter currently
has a nonzero value and whether the EFD_SEMAPHORE flag was specified
when creating the eventfd file descriptor:

  *  If EFD_SEMAPHORE was not specified and the eventfd counter has a
nonzero value, then a read(2) returns 8 bytes containing that value, and
the counter's value is reset to zero.

......

  *  If the eventfd counter is zero at the time of the call to  read(2),
 then the call either blocks until the counter becomes  nonzero (at
which time, the read(2) proceeds as described above) or fails with the
error EAGAIN if the file descriptor has been made nonblocking.

And in actual tests, the debug line I add in vfio_err_notifier_handler
shows[*]: qemu read this value right every time, the same as in vfio-pci
driver.

quick reference:
[*]https://lists.gnu.org/archive/html/qemu-devel/2016-11/msg04831.html

> 
>> Why? When testing previous version(host aer driver still reset link),
>> found that there is quite possibility that register reading returns the
>> invalid value 0xFFFFFFFF, which results in all(2 in my environment)
>> qemu's vfio-pci function send AER to root port while I only inject error
>> into one function. This is unreasonable, and I think it is the result of
>> reading during device reset.
>>
>> Previous patch does considered to find the
>> real function who has error happened, but won't work under the situation
>> I found. So move the register reading as early as possible would be the
>> nature thoughts, and it is moved into vfio-pci driver. Although now
>> reset_link is disabled in aer driver, get the uncor error status as
>> early as possible still make sense, for example: if user space driver
>> does device reset once receiving the error notification, and then read
>> register.
> 
> I had trouble understanding the above. Let me ask you this:
> should we try to avoid triggering uncorrectable errors?
> Aren't any such errors likely to crash guests?
> 

Sorry I don't quite understand why your questions are raised, it seems
just describing the current situation without my AER patch: when assign
device via vfio, we don't handle uncorrectable error, just
vm_stop()(this is what you mean "crash guests", right?), this is the
contents of current vfio_err_notifier_handler().

Without our AER patches, the answer to your questions are YES. But the
purpose of our patch is to fix your questions. Not sure my answer helps
you, but feel free to ask if still any questions.


-- 
Sincerely,
Cao jin

[toc] | [prev] | [next] | [standalone]

#1534095

On 12/01/2016 12:04 PM, Alex Williamson wrote:
> On Sun, 27 Nov 2016 19:34:17 +0800
> Cao jin <caoj.fnst@cn.fujitsu.com> wrote:
> 
>> It is user space driver's or device-specific driver's(in guest) responsbility
>> to do a serious recovery when error happened. Link-reset is one part of
>> recovery, when pci device is assigned to VM via vfio, link-reset will do
>> twice in host & guest separately, which will cause many trouble for a
>> successful recovery, so, disable the vfio-pci's link-reset in aer driver
>> in host, this is a keypoint for guest to do error recovery successfully.
>>
>> CC: alex.williamson@redhat.com
>> CC: mst@redhat.com
>> Signed-off-by: Cao jin <caoj.fnst@cn.fujitsu.com>
>> ---
>> This is actually a RFC version(has debug lines left), and has minor changes in
>> aer driver, so I think maybe it is better not to CC pci guys in this round.
>> Later will do.
>>
>>  drivers/pci/pcie/aer/aerdrv_core.c  | 12 ++++++-
>>  drivers/vfio/pci/vfio_pci.c         | 63 +++++++++++++++++++++++++++++++++++--
>>  drivers/vfio/pci/vfio_pci_private.h |  2 ++
>>  3 files changed, 74 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/pci/pcie/aer/aerdrv_core.c b/drivers/pci/pcie/aer/aerdrv_core.c
>> index 521e39c..289fb8e 100644
>> --- a/drivers/pci/pcie/aer/aerdrv_core.c
>> +++ b/drivers/pci/pcie/aer/aerdrv_core.c
>> @@ -496,7 +496,17 @@ static void do_recovery(struct pci_dev *dev, int severity)
>>  			"error_detected",
>>  			report_error_detected);
>>  
>> -	if (severity == AER_FATAL) {
>> +	/* vfio-pci as a general meta driver, it actually couldn't do any real
>> +	 * recovery for device. It is user space driver, or device-specific
>> +	 * driver in guest who should take care of the serious error recovery,
>> +	 * link reset actually is one part of whole recovery. Doing reset_link
>> +	 * in aer driver of host kernel for vfio-pci devices will cause many
>> +	 * trouble for user space driver or guest's device-specific driver,
>> +	 * for example: the serious recovery often need to read register in
>> +	 * config space, but if register reading happens during link-resetting,
>> +	 * it is quite possible to return invalid value like all F's, which
>> +	 * will result in unpredictable error. */
>> +	if (severity == AER_FATAL && strcmp(dev->driver->name, "vfio-pci")) {
>>  		result = reset_link(dev);
>>  		if (result != PCI_ERS_RESULT_RECOVERED)
>>  			goto failed;
> 
> This is not acceptable.  If we want to make a path through AER recovery
> that does not reset the link, there should be a way for the driver to
> request it.  Testing the driver name is a horrible hack.  The other
> question is what guarantees does vfio make that the device does get
> reset?  

I am not sure how vfio guarantee that...When device is assigned to VM,
we have that guarantees(aer driver in guest driver will do that), so I
think it is a well-behaved user space driver's responsibility to do link
reset?  And I think if there is a user space driver, it is surely its
responsibility to consider how to do serious error recovery, like I said
before, vfio, as a general meta driver, it surely don't know how each
device does its specific recovery, except bus/slot reset

> If an AER fault occurs and the user doesn't do a reset, what
> happens when that device is released and a host driver tries to make
> use of it?  The user makes no commitment to do a reset and there are
> only limited configurations where we even allow the user to perform a
> reset.
> 

Limited? Do you mean the things __pci_dev_reset() can do?

...
> 
>> +	aer_cap_offset = pci_find_ext_capability(vdev->pdev, PCI_EXT_CAP_ID_ERR);
>> +	ret = pci_read_config_dword(vdev->pdev, aer_cap_offset + 
>> +                                    PCI_ERR_UNCOR_STATUS, &uncor_status);
>> +        if (ret)
>> +                return PCI_ERS_RESULT_DISCONNECT;
>> +
>> +	pr_err("device %d got AER detect notification. uncorrectable error status = 0x%x\n", pdev->devfn, uncor_status);//to be removed
>>  	mutex_lock(&vdev->igate);
>> +    
>> +	vdev->aer_recovering = true;
>> +	reinit_completion(&vdev->aer_error_completion);
>> +
>> +	/* suspend config space access from user space,
>> +	 * when vfio-pci's error recovery process is on */
>> +	pci_cfg_access_trylock(vdev->pdev);
>>  
>> -	if (vdev->err_trigger)
>> -		eventfd_signal(vdev->err_trigger, 1);
>> +	if (vdev->err_trigger && uncor_status) {
>> +		pr_err("device %d signal uncor status to user space", pdev->devfn);//may be removed
>> +		/* signal uncorrectable error status to user space */
>> +		eventfd_signal(vdev->err_trigger, uncor_status);
>> +        }
> 
> } else... what?  By bypassing the AER link reset we've assumed
> responsibility for resetting the device.  Even if we signal the user,
> what guarantee do we have that the device is recovered?
> 

else...consider it as a fake error notification and ignore?

I am not sure I understand your thoughts totally, but it seems my
previous comments apply, that is: it is well-behaved user space driver's
responsibility to do a serious recovery.

In my understanding, user space driver has 2 category: one is VM(has
guest OS running inside), the other is ordinary user space program.

When device is assigned to a VM, (qemu + guest OS) will do fully steps
to do recovery(roughly is what struct pci_error_handlers has). So,
equally, if it is a ordinary user space program acting as the driver,
the responsibility belongs to it.


Sincerely,
Cao jin
>>  
>>  	mutex_unlock(&vdev->igate);
>>  
>> @@ -1199,8 +1232,34 @@ static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
>>  	return PCI_ERS_RESULT_CAN_RECOVER;
>>  }
>>  
>> +static void vfio_pci_aer_resume(struct pci_dev *pdev)
>> +{
>> +	struct vfio_pci_device *vdev;
>> +	struct vfio_device *device;
>> +
>> +	device = vfio_device_get_from_dev(&pdev->dev);
>> +	if (device == NULL)
>> +		return;
>> +
>> +	vdev = vfio_device_data(device);
>> +	if (vdev == NULL) {
>> +		vfio_device_put(device);
>> +		return;
>> +	}
>> +
>> +	/* vfio-pci's error recovery is done, time to
>> +	 * resume pci config space's accesses */
>> +	pci_cfg_access_unlock(vdev->pdev);
>> +
>> +	vdev->aer_recovering = false;
>> +	complete_all(&vdev->aer_error_completion);
>> +
>> +	vfio_device_put(device);
>> +}
>> +
>>  static const struct pci_error_handlers vfio_err_handlers = {
>>  	.error_detected = vfio_pci_aer_err_detected,
>> +	.resume         = vfio_pci_aer_resume,
>>  };
>>  
>>  static struct pci_driver vfio_pci_driver = {
>> diff --git a/drivers/vfio/pci/vfio_pci_private.h b/drivers/vfio/pci/vfio_pci_private.h
>> index 8a7d546..ebf1041 100644
>> --- a/drivers/vfio/pci/vfio_pci_private.h
>> +++ b/drivers/vfio/pci/vfio_pci_private.h
>> @@ -83,6 +83,8 @@ struct vfio_pci_device {
>>  	bool			bardirty;
>>  	bool			has_vga;
>>  	bool			needs_reset;
>> +	bool			aer_recovering;
>> +	struct completion	aer_error_completion;
>>  	struct pci_saved_state	*pci_saved_state;
>>  	int			refcnt;
>>  	struct eventfd_ctx	*err_trigger;
> 
> 
> 
> .
> 

[toc] | [prev] | [next] | [standalone]

#1534170

On Thu, 1 Dec 2016 21:40:00 +0800
Cao jin <caoj.fnst@cn.fujitsu.com> wrote:

> On 12/01/2016 12:04 PM, Alex Williamson wrote:
> > On Sun, 27 Nov 2016 19:34:17 +0800
> > Cao jin <caoj.fnst@cn.fujitsu.com> wrote:
> >   
> >> It is user space driver's or device-specific driver's(in guest) responsbility
> >> to do a serious recovery when error happened. Link-reset is one part of
> >> recovery, when pci device is assigned to VM via vfio, link-reset will do
> >> twice in host & guest separately, which will cause many trouble for a
> >> successful recovery, so, disable the vfio-pci's link-reset in aer driver
> >> in host, this is a keypoint for guest to do error recovery successfully.
> >>
> >> CC: alex.williamson@redhat.com
> >> CC: mst@redhat.com
> >> Signed-off-by: Cao jin <caoj.fnst@cn.fujitsu.com>
> >> ---
> >> This is actually a RFC version(has debug lines left), and has minor changes in
> >> aer driver, so I think maybe it is better not to CC pci guys in this round.
> >> Later will do.
> >>
> >>  drivers/pci/pcie/aer/aerdrv_core.c  | 12 ++++++-
> >>  drivers/vfio/pci/vfio_pci.c         | 63 +++++++++++++++++++++++++++++++++++--
> >>  drivers/vfio/pci/vfio_pci_private.h |  2 ++
> >>  3 files changed, 74 insertions(+), 3 deletions(-)
> >>
> >> diff --git a/drivers/pci/pcie/aer/aerdrv_core.c b/drivers/pci/pcie/aer/aerdrv_core.c
> >> index 521e39c..289fb8e 100644
> >> --- a/drivers/pci/pcie/aer/aerdrv_core.c
> >> +++ b/drivers/pci/pcie/aer/aerdrv_core.c
> >> @@ -496,7 +496,17 @@ static void do_recovery(struct pci_dev *dev, int severity)
> >>  			"error_detected",
> >>  			report_error_detected);
> >>  
> >> -	if (severity == AER_FATAL) {
> >> +	/* vfio-pci as a general meta driver, it actually couldn't do any real
> >> +	 * recovery for device. It is user space driver, or device-specific
> >> +	 * driver in guest who should take care of the serious error recovery,
> >> +	 * link reset actually is one part of whole recovery. Doing reset_link
> >> +	 * in aer driver of host kernel for vfio-pci devices will cause many
> >> +	 * trouble for user space driver or guest's device-specific driver,
> >> +	 * for example: the serious recovery often need to read register in
> >> +	 * config space, but if register reading happens during link-resetting,
> >> +	 * it is quite possible to return invalid value like all F's, which
> >> +	 * will result in unpredictable error. */
> >> +	if (severity == AER_FATAL && strcmp(dev->driver->name, "vfio-pci")) {
> >>  		result = reset_link(dev);
> >>  		if (result != PCI_ERS_RESULT_RECOVERED)
> >>  			goto failed;  
> > 
> > This is not acceptable.  If we want to make a path through AER recovery
> > that does not reset the link, there should be a way for the driver to
> > request it.  Testing the driver name is a horrible hack.  The other
> > question is what guarantees does vfio make that the device does get
> > reset?    
> 
> I am not sure how vfio guarantee that...When device is assigned to VM,
> we have that guarantees(aer driver in guest driver will do that), so I
> think it is a well-behaved user space driver's responsibility to do link
> reset?  And I think if there is a user space driver, it is surely its
> responsibility to consider how to do serious error recovery, like I said
> before, vfio, as a general meta driver, it surely don't know how each
> device does its specific recovery, except bus/slot reset

We can't assume the user it a VM and we certainly cannot assume that
the user is well behaved.  By bypassing the link reset here, you've
taken responsibility for assuring that at least that basic level of
recovery is performed.  We cannot assume the user will do it though.

> > If an AER fault occurs and the user doesn't do a reset, what
> > happens when that device is released and a host driver tries to make
> > use of it?  The user makes no commitment to do a reset and there are
> > only limited configurations where we even allow the user to perform a
> > reset.
> >   
> 
> Limited? Do you mean the things __pci_dev_reset() can do?

I mean that there are significant device and guest configuration
restrictions in order to support AER.  For instance, all the functions
of the slot need to appear in a PCI-e topology in the guest with all
the functions in the right place such that a guest bus reset translates
into a host bus reset.  The physical functions cannot be split between
guests even if IOMMU isolation would otherwise allow it.  The user
needs to explicitly enable AER support for the devices.  A VM need to
be specifically configured for AER support in order to set any sort of
expectations of a guest directed bus reset, let alone a guarantee that
it will happen.  So all the existing VMs, where functions are split
between guests, or the topology isn't exactly right, or AER isn't
enabled see a regression from the above change as the device is no
longer reset.

> ...
> >   
> >> +	aer_cap_offset = pci_find_ext_capability(vdev->pdev, PCI_EXT_CAP_ID_ERR);
> >> +	ret = pci_read_config_dword(vdev->pdev, aer_cap_offset + 
> >> +                                    PCI_ERR_UNCOR_STATUS, &uncor_status);
> >> +        if (ret)
> >> +                return PCI_ERS_RESULT_DISCONNECT;
> >> +
> >> +	pr_err("device %d got AER detect notification. uncorrectable error status = 0x%x\n", pdev->devfn, uncor_status);//to be removed
> >>  	mutex_lock(&vdev->igate);
> >> +    
> >> +	vdev->aer_recovering = true;
> >> +	reinit_completion(&vdev->aer_error_completion);
> >> +
> >> +	/* suspend config space access from user space,
> >> +	 * when vfio-pci's error recovery process is on */
> >> +	pci_cfg_access_trylock(vdev->pdev);
> >>  
> >> -	if (vdev->err_trigger)
> >> -		eventfd_signal(vdev->err_trigger, 1);
> >> +	if (vdev->err_trigger && uncor_status) {
> >> +		pr_err("device %d signal uncor status to user space", pdev->devfn);//may be removed
> >> +		/* signal uncorrectable error status to user space */
> >> +		eventfd_signal(vdev->err_trigger, uncor_status);
> >> +        }  
> > 
> > } else... what?  By bypassing the AER link reset we've assumed
> > responsibility for resetting the device.  Even if we signal the user,
> > what guarantee do we have that the device is recovered?
> >   
> 
> else...consider it as a fake error notification and ignore?

Nope.  Registering the error eventfd is optional, you've just
introduced a regression where the device was previously reset by the
AER core code and now it's in some unknown state because it generated
an error and we told the core not to reset it.

> I am not sure I understand your thoughts totally, but it seems my
> previous comments apply, that is: it is well-behaved user space driver's
> responsibility to do a serious recovery.

We can *never* assume a well behaved userspace.  That's an invalid
starting point.

> In my understanding, user space driver has 2 category: one is VM(has
> guest OS running inside), the other is ordinary user space program.
> 
> When device is assigned to a VM, (qemu + guest OS) will do fully steps
> to do recovery(roughly is what struct pci_error_handlers has). So,
> equally, if it is a ordinary user space program acting as the driver,
> the responsibility belongs to it.

This is simply not valid.  If vfio is going to assume responsibility
for resetting the device, it can't simply push that responsibility to
userspace.  There are only very limited user configurations where a
guest bus reset translating directly to a host bus reset are even
valid.  Also keep in mind that we must support old userspace on newer
kernels.  There currently does not exist AER recovery on existing
userspace VMs.  Thanks,

Alex

[toc] | [prev] | [next] | [standalone]

#1535647

On 12/01/2016 10:55 PM, Alex Williamson wrote:
> On Thu, 1 Dec 2016 21:40:00 +0800

>>> If an AER fault occurs and the user doesn't do a reset, what
>>> happens when that device is released and a host driver tries to make
>>> use of it?  The user makes no commitment to do a reset and there are
>>> only limited configurations where we even allow the user to perform a
>>> reset.
>>>   
>>
>> Limited? Do you mean the things __pci_dev_reset() can do?
> 
> I mean that there are significant device and guest configuration
> restrictions in order to support AER.  For instance, all the functions
> of the slot need to appear in a PCI-e topology in the guest with all
> the functions in the right place such that a guest bus reset translates
> into a host bus reset.  The physical functions cannot be split between
> guests even if IOMMU isolation would otherwise allow it.  The user
> needs to explicitly enable AER support for the devices.  A VM need to
> be specifically configured for AER support in order to set any sort of
> expectations of a guest directed bus reset, let alone a guarantee that
> it will happen.  So all the existing VMs, where functions are split
> between guests, or the topology isn't exactly right, or AER isn't
> enabled see a regression from the above change as the device is no
> longer reset.
> 

I am not clear why set these restrictions in the current design. I take
a glance at older versions of qemu's patchset, their thoughts is:
translate a guest bus reset into a host bus reset(Which is
unreasonable[*] to me). And I guess, that's the *cause* of these
restrictions?  Is there any other stories behind these restrictions?

[*] In physical world, set bridge's secondary bus reset would send
hot-reset TLP to all functions below, trigger every device's reset
separately. Emulated device should behave the same, means just using
each device's DeviceClass->reset method.
-- 
Sincerely,
Cao jin

[toc] | [prev] | [next] | [standalone]

#1535694

On Sun, 4 Dec 2016 20:16:42 +0800
Cao jin <caoj.fnst@cn.fujitsu.com> wrote:

> On 12/01/2016 10:55 PM, Alex Williamson wrote:
> > On Thu, 1 Dec 2016 21:40:00 +0800  
> 
> >>> If an AER fault occurs and the user doesn't do a reset, what
> >>> happens when that device is released and a host driver tries to make
> >>> use of it?  The user makes no commitment to do a reset and there are
> >>> only limited configurations where we even allow the user to perform a
> >>> reset.
> >>>     
> >>
> >> Limited? Do you mean the things __pci_dev_reset() can do?  
> > 
> > I mean that there are significant device and guest configuration
> > restrictions in order to support AER.  For instance, all the functions
> > of the slot need to appear in a PCI-e topology in the guest with all
> > the functions in the right place such that a guest bus reset translates
> > into a host bus reset.  The physical functions cannot be split between
> > guests even if IOMMU isolation would otherwise allow it.  The user
> > needs to explicitly enable AER support for the devices.  A VM need to
> > be specifically configured for AER support in order to set any sort of
> > expectations of a guest directed bus reset, let alone a guarantee that
> > it will happen.  So all the existing VMs, where functions are split
> > between guests, or the topology isn't exactly right, or AER isn't
> > enabled see a regression from the above change as the device is no
> > longer reset.
> >   
> 
> I am not clear why set these restrictions in the current design. I take
> a glance at older versions of qemu's patchset, their thoughts is:
> translate a guest bus reset into a host bus reset(Which is
> unreasonable[*] to me). And I guess, that's the *cause* of these
> restrictions?  Is there any other stories behind these restrictions?
> 
> [*] In physical world, set bridge's secondary bus reset would send
> hot-reset TLP to all functions below, trigger every device's reset
> separately. Emulated device should behave the same, means just using
> each device's DeviceClass->reset method.

Are you trying to say that an FLR is equivalent to a link reset?
Please go read the previous discussions, especially if you're sending
patches you don't believe in.  Thanks,

Alex

[toc] | [prev] | [next] | [standalone]

#1535817

On 12/04/2016 11:30 PM, Alex Williamson wrote:
> On Sun, 4 Dec 2016 20:16:42 +0800
> Cao jin <caoj.fnst@cn.fujitsu.com> wrote:
> 
>> On 12/01/2016 10:55 PM, Alex Williamson wrote:
>>> On Thu, 1 Dec 2016 21:40:00 +0800  
>>
>>>>> If an AER fault occurs and the user doesn't do a reset, what
>>>>> happens when that device is released and a host driver tries to make
>>>>> use of it?  The user makes no commitment to do a reset and there are
>>>>> only limited configurations where we even allow the user to perform a
>>>>> reset.
>>>>>     
>>>>
>>>> Limited? Do you mean the things __pci_dev_reset() can do?  
>>>
>>> I mean that there are significant device and guest configuration
>>> restrictions in order to support AER.  For instance, all the functions
>>> of the slot need to appear in a PCI-e topology in the guest with all
>>> the functions in the right place such that a guest bus reset translates
>>> into a host bus reset.  The physical functions cannot be split between
>>> guests even if IOMMU isolation would otherwise allow it.  The user
>>> needs to explicitly enable AER support for the devices.  A VM need to
>>> be specifically configured for AER support in order to set any sort of
>>> expectations of a guest directed bus reset, let alone a guarantee that
>>> it will happen.  So all the existing VMs, where functions are split
>>> between guests, or the topology isn't exactly right, or AER isn't
>>> enabled see a regression from the above change as the device is no
>>> longer reset.
>>>   
>>
>> I am not clear why set these restrictions in the current design. I take
>> a glance at older versions of qemu's patchset, their thoughts is:
>> translate a guest bus reset into a host bus reset(Which is
>> unreasonable[*] to me). And I guess, that's the *cause* of these
>> restrictions?  Is there any other stories behind these restrictions?
>>
>> [*] In physical world, set bridge's secondary bus reset would send
>> hot-reset TLP to all functions below, trigger every device's reset
>> separately. Emulated device should behave the same, means just using
>> each device's DeviceClass->reset method.
> 
> Are you trying to say that an FLR is equivalent to a link reset?

No.  Look at old versions patchset, there is one names "vote the
function 0 to do host bus reset when aer occurred"[1], that is what I
called "translate guest link reset to host link reset", and what I think
unreasonable(and I think it also does it wrongly).  So in v10 version of
mine, I dropped it.

[1]https://lists.gnu.org/archive/html/qemu-devel/2016-05/msg02987.html

If "translate guest link reset to host link reset" is right, I can
understand these restrictions[2][3].

[2]. All physical functions in a single card must be assigned to the VM
     with AER enabled on each and configured on the same virtual bus.
[3]. Don't place other devices under the virtual bus in [2], no matter
     physical, emulated, or paravirtual, even if other device
     supporting AER signaling

Certain device's FLR calls its DeviceClass->reset method; link reset
calls DeviceClass->reset of each device which on the bus. So, apparently
they have difference.  But if there is only 1 vfio-pci device under the
virtual pci bus,  I think FLR can be equivalent to a link reset, right?

> Please go read the previous discussions, especially if you're sending
> patches you don't believe in.  Thanks,
> 

I does not read ALL version's discussion thoroughly, but these
restrictions exist for a long time, so I guess it is a result of
previous discussions.  If it is not, I am thinking of the possibility of
dropping these restrictions[2][3], and drop the "aer" property,
automatically enable this functionality or not according to device's
capability.

-- 
Sincerely,
Cao jin

[toc] | [prev] | [next] | [standalone]

#1536203

On Mon, 5 Dec 2016 13:52:03 +0800
Cao jin <caoj.fnst@cn.fujitsu.com> wrote:

> On 12/04/2016 11:30 PM, Alex Williamson wrote:
> > On Sun, 4 Dec 2016 20:16:42 +0800
> > Cao jin <caoj.fnst@cn.fujitsu.com> wrote:
> >   
> >> On 12/01/2016 10:55 PM, Alex Williamson wrote:  
> >>> On Thu, 1 Dec 2016 21:40:00 +0800    
> >>  
> >>>>> If an AER fault occurs and the user doesn't do a reset, what
> >>>>> happens when that device is released and a host driver tries to make
> >>>>> use of it?  The user makes no commitment to do a reset and there are
> >>>>> only limited configurations where we even allow the user to perform a
> >>>>> reset.
> >>>>>       
> >>>>
> >>>> Limited? Do you mean the things __pci_dev_reset() can do?    
> >>>
> >>> I mean that there are significant device and guest configuration
> >>> restrictions in order to support AER.  For instance, all the functions
> >>> of the slot need to appear in a PCI-e topology in the guest with all
> >>> the functions in the right place such that a guest bus reset translates
> >>> into a host bus reset.  The physical functions cannot be split between
> >>> guests even if IOMMU isolation would otherwise allow it.  The user
> >>> needs to explicitly enable AER support for the devices.  A VM need to
> >>> be specifically configured for AER support in order to set any sort of
> >>> expectations of a guest directed bus reset, let alone a guarantee that
> >>> it will happen.  So all the existing VMs, where functions are split
> >>> between guests, or the topology isn't exactly right, or AER isn't
> >>> enabled see a regression from the above change as the device is no
> >>> longer reset.
> >>>     
> >>
> >> I am not clear why set these restrictions in the current design. I take
> >> a glance at older versions of qemu's patchset, their thoughts is:
> >> translate a guest bus reset into a host bus reset(Which is
> >> unreasonable[*] to me). And I guess, that's the *cause* of these
> >> restrictions?  Is there any other stories behind these restrictions?
> >>
> >> [*] In physical world, set bridge's secondary bus reset would send
> >> hot-reset TLP to all functions below, trigger every device's reset
> >> separately. Emulated device should behave the same, means just using
> >> each device's DeviceClass->reset method.  
> > 
> > Are you trying to say that an FLR is equivalent to a link reset?  
> 
> No.  Look at old versions patchset, there is one names "vote the
> function 0 to do host bus reset when aer occurred"[1], that is what I
> called "translate guest link reset to host link reset", and what I think
> unreasonable(and I think it also does it wrongly).  So in v10 version of
> mine, I dropped it.
> 
> [1]https://lists.gnu.org/archive/html/qemu-devel/2016-05/msg02987.html
> 
> If "translate guest link reset to host link reset" is right, I can
> understand these restrictions[2][3].
> 
> [2]. All physical functions in a single card must be assigned to the VM
>      with AER enabled on each and configured on the same virtual bus.
> [3]. Don't place other devices under the virtual bus in [2], no matter
>      physical, emulated, or paravirtual, even if other device
>      supporting AER signaling
> 
> Certain device's FLR calls its DeviceClass->reset method; link reset
> calls DeviceClass->reset of each device which on the bus. So, apparently
> they have difference.  But if there is only 1 vfio-pci device under the
> virtual pci bus,  I think FLR can be equivalent to a link reset, right?

No.  An FLR resets the device while a secondary bus reset does a reset
of the link and the device.  AER errors are sometimes issues with the
link, not the device.  If we were to perform only an FLR, we're not
performing the same reset as would be done on bare metal.
 
> > Please go read the previous discussions, especially if you're sending
> > patches you don't believe in.  Thanks,
> >   
> 
> I does not read ALL version's discussion thoroughly, but these
> restrictions exist for a long time, so I guess it is a result of
> previous discussions.  If it is not, I am thinking of the possibility of
> dropping these restrictions[2][3], and drop the "aer" property,
> automatically enable this functionality or not according to device's
> capability.

If you're going to take the lead for these AER patches, I would
certainly suggest that understanding the reasoning behind the bus reset
behavior is a central aspect to this series.  This effort has dragged
out for nearly two years and I apologize, but I don't really have a lot
of patience for rehashing some of these issues if you're not going to
read the previous discussions or consult with your colleagues to
understand how we got to this point.  If you want to challenge some of
the design points, that's great, it could use some new eyes, but please
understand how we got here first.  If we could replace a bus reset with
an FLR that would make things far more simple, but I don't see that as
a valid option.  Thanks,

Alex

[toc] | [prev] | [next] | [standalone]

#1536647

On Mon, Dec 05, 2016 at 09:17:30AM -0700, Alex Williamson wrote:
> If you're going to take the lead for these AER patches, I would
> certainly suggest that understanding the reasoning behind the bus reset
> behavior is a central aspect to this series.  This effort has dragged
> out for nearly two years and I apologize, but I don't really have a lot
> of patience for rehashing some of these issues if you're not going to
> read the previous discussions or consult with your colleagues to
> understand how we got to this point.  If you want to challenge some of
> the design points, that's great, it could use some new eyes, but please
> understand how we got here first.

Well I'm guessing Cao jin here isn't the only one not
willing to plough through all historical versions of the patchset
just to figure out the motivation for some code.

Including a summary of a high level architecture couldn't hurt.

Any chance of writing such?  Alternatively, we can try to build it as
part of this thread.  Shouldn't be hard as it seems somewhat
straight-forward on the surface:

- detect link error on the host, don't reset link as we would normally do
- report link error to guest
- detect link reset request from guest
- reset link on host

Since link reset will reset all devices behind it, for this to work we
need same set of devices behind the link in host and guest.  Enforcing
this would be nice to have.

- as link now might end up in bad state, reset
  it when device is unassigned

Any details I missed?

-- 
MST

[toc] | [prev] | [next] | [standalone]

#1536670

On Tue, 6 Dec 2016 05:55:28 +0200
"Michael S. Tsirkin" <mst@redhat.com> wrote:

> On Mon, Dec 05, 2016 at 09:17:30AM -0700, Alex Williamson wrote:
> > If you're going to take the lead for these AER patches, I would
> > certainly suggest that understanding the reasoning behind the bus reset
> > behavior is a central aspect to this series.  This effort has dragged
> > out for nearly two years and I apologize, but I don't really have a lot
> > of patience for rehashing some of these issues if you're not going to
> > read the previous discussions or consult with your colleagues to
> > understand how we got to this point.  If you want to challenge some of
> > the design points, that's great, it could use some new eyes, but please
> > understand how we got here first.  
> 
> Well I'm guessing Cao jin here isn't the only one not
> willing to plough through all historical versions of the patchset
> just to figure out the motivation for some code.
> 
> Including a summary of a high level architecture couldn't hurt.
> 
> Any chance of writing such?  Alternatively, we can try to build it as
> part of this thread.  Shouldn't be hard as it seems somewhat
> straight-forward on the surface:
> 
> - detect link error on the host, don't reset link as we would normally do

This is actually a new approach that I'm not sure I agree with.  By
skipping the host directed link reset, vfio is taking responsibility
for doing this, but then we just assume the user will do it.  I have
issues with this.

The previous approach was to use the error detected notifier to block
access to the device, allowing the host to perform the link reset.  A
subsequent notification in the AER process released the user access
which allowed the user AER process to proceed.  This did result in both
a host directed and a guest directed link reset, but other than
coordinating the blocking of the user process during host reset, that
hasn't been brought up as an issue previously.

> - report link error to guest
> - detect link reset request from guest
> - reset link on host
> 
> Since link reset will reset all devices behind it, for this to work we
> need same set of devices behind the link in host and guest.  Enforcing
> this would be nice to have.

This is a pretty significant complication and I think it's a
requirement.  This is the heart of why we have an AER vfio-pci device
option and why we require that QEMU should fail to initialize the
device if AER is enabled in an incompatible configuration.  If a user
cannot be sure that AER is enabled on a device, it's pointless to
bother implementing it, though honestly I question the value of it in
the VM altogether given configuration requirements (ie. are users
going to accept the reason that all the ports of a device need to be
assigned to a single guest for guest-based AER recovery when they were
previously able to assign each port to a separate VM?).
 
> - as link now might end up in bad state, reset
>   it when device is unassigned

This is also a new aspect for the approach here, previously we allowed
the host directed link reset so we could assume the device was
sufficiently recovered.  In the proposal here, the AER core skips any
devices bound to vfio-pci, but vfio can't guarantee that we can do a
bus reset on them.  PCI bus isolation is not accounted for in DMA
isolation, which is the basis for IOMMU groups.  A bus can host
multiple IOMMU groups, each of which may have a different user.  Only
in a very specific configuration can vfio do a bus reset.
 
> Any details I missed?

AIUI, the critical feature is that the guest needs to be able to reset
the device link, all the other design elements play out from the
logical expansion of that feature.  It means that a guest bus reset
needs to translate to a host bus reset, which means that all of the
affected host devices need to be accounted for and those that are
assigned to the guest need to be affected in the guest in the same
way.  QEMU must enforce this configuration or else a user cannot know
the result of a AER fault, ie. will it cause a VMSTOP condition or is
the fault forwarded to the guest.  The required configuration
restrictions are quite involved, therefore we can't simply require this
of all configurations, so a vfio-pci device option is added.  The
coordination of the host directed reset with the guest directed reset
was only a complications discovered within the last few revisions of
the series.  As noted above, the previous solution to this was to
attempt to block access to the device while the host reset proceeds.

Clearly it's a little disconcerting if we throw all of that away and
simply assume that an FLR is sufficient to reset the device when it
seems like link issues might be a nontrivial source of AER faults.  If
FLR is sufficient, then why does the core AER handling code in the
kernel not simply do this?  Thanks,

Alex

[toc] | [prev] | [next] | [standalone]

#1536842

On 12/06/2016 12:59 PM, Alex Williamson wrote:
> On Tue, 6 Dec 2016 05:55:28 +0200
> "Michael S. Tsirkin" <mst@redhat.com> wrote:
> 
>> On Mon, Dec 05, 2016 at 09:17:30AM -0700, Alex Williamson wrote:
>>> If you're going to take the lead for these AER patches, I would
>>> certainly suggest that understanding the reasoning behind the bus reset
>>> behavior is a central aspect to this series.  This effort has dragged
>>> out for nearly two years and I apologize, but I don't really have a lot
>>> of patience for rehashing some of these issues if you're not going to
>>> read the previous discussions or consult with your colleagues to
>>> understand how we got to this point.  If you want to challenge some of
>>> the design points, that's great, it could use some new eyes, but please
>>> understand how we got here first.  
>>
>> Well I'm guessing Cao jin here isn't the only one not
>> willing to plough through all historical versions of the patchset
>> just to figure out the motivation for some code.
>>
>> Including a summary of a high level architecture couldn't hurt.
>>
>> Any chance of writing such?  Alternatively, we can try to build it as
>> part of this thread.  Shouldn't be hard as it seems somewhat
>> straight-forward on the surface:
>>
>> - detect link error on the host, don't reset link as we would normally do
> 
> This is actually a new approach that I'm not sure I agree with.  By
> skipping the host directed link reset, vfio is taking responsibility
> for doing this, but then we just assume the user will do it.  I have
> issues with this.
> 
> The previous approach was to use the error detected notifier to block
> access to the device, allowing the host to perform the link reset.  A
> subsequent notification in the AER process released the user access
> which allowed the user AER process to proceed.  This did result in both
> a host directed and a guest directed link reset, but other than
> coordinating the blocking of the user process during host reset, that
> hasn't been brought up as an issue previously.
> 

Tests on previous versions didn't bring up issues as I find, I think
that is because we didn't test it completely. As I know, before August
of this year, we didn't have cable connected to NIC, let alone
connecting NIC to gateway.

Even if I fixed the guest oops issue in igb driver that Alex found in
v9, v9 still cannot work in my test. And in my test, disable link
reset(in host) in aer core for vfio-pci is the most significant step to
get my test passed.

>> - report link error to guest
>> - detect link reset request from guest
>> - reset link on host
>>
>> Since link reset will reset all devices behind it, for this to work we
>> need same set of devices behind the link in host and guest.  Enforcing
>> this would be nice to have.
> 
> This is a pretty significant complication and I think it's a
> requirement.  This is the heart of why we have an AER vfio-pci device
> option and why we require that QEMU should fail to initialize the
> device if AER is enabled in an incompatible configuration.  If a user
> cannot be sure that AER is enabled on a device, it's pointless to
> bother implementing it, though honestly I question the value of it in
> the VM altogether given configuration requirements (ie. are users
> going to accept the reason that all the ports of a device need to be
> assigned to a single guest for guest-based AER recovery when they were
> previously able to assign each port to a separate VM?).
>  
>> - as link now might end up in bad state, reset
>>   it when device is unassigned
> 
> This is also a new aspect for the approach here, previously we allowed
> the host directed link reset so we could assume the device was
> sufficiently recovered.  In the proposal here, the AER core skips any
> devices bound to vfio-pci, but vfio can't guarantee that we can do a
> bus reset on them.  PCI bus isolation is not accounted for in DMA
> isolation, which is the basis for IOMMU groups.  A bus can host
> multiple IOMMU groups, each of which may have a different user.  Only
> in a very specific configuration can vfio do a bus reset.
>  
>> Any details I missed?
> 
> AIUI, the critical feature is that the guest needs to be able to reset
> the device link, all the other design elements play out from the
> logical expansion of that feature.  It means that a guest bus reset
> needs to translate to a host bus reset, which means that all of the
> affected host devices need to be accounted for and those that are
> assigned to the guest need to be affected in the guest in the same
> way.  QEMU must enforce this configuration or else a user cannot know
> the result of a AER fault, ie. will it cause a VMSTOP condition or is
> the fault forwarded to the guest.  The required configuration
> restrictions are quite involved, therefore we can't simply require this
> of all configurations, so a vfio-pci device option is added.  The
> coordination of the host directed reset with the guest directed reset
> was only a complications discovered within the last few revisions of
> the series.  As noted above, the previous solution to this was to
> attempt to block access to the device while the host reset proceeds.
> 
> Clearly it's a little disconcerting if we throw all of that away and
> simply assume that an FLR is sufficient to reset the device when it
> seems like link issues might be a nontrivial source of AER faults.  If
> FLR is sufficient, then why does the core AER handling code in the
> kernel not simply do this?  Thanks,
> 

I agree with the points here.  Now I understand why  translate a guest
link reset to host link reset[*], and FLR shouldn't be equivalent to
link reset, but the PITY is, we can't trigger a real fatal error to see
if a FLR is sufficient to reset the device.

[*]https://lists.gnu.org/archive/html/qemu-devel/2016-07/msg04100.html
-- 
Sincerely,
Cao jin

[toc] | [prev] | [next] | [standalone]

#1537049

On Tue, 6 Dec 2016 18:46:04 +0800
Cao jin <caoj.fnst@cn.fujitsu.com> wrote:

> On 12/06/2016 12:59 PM, Alex Williamson wrote:
> > On Tue, 6 Dec 2016 05:55:28 +0200
> > "Michael S. Tsirkin" <mst@redhat.com> wrote:
> >   
> >> On Mon, Dec 05, 2016 at 09:17:30AM -0700, Alex Williamson wrote:  
> >>> If you're going to take the lead for these AER patches, I would
> >>> certainly suggest that understanding the reasoning behind the bus reset
> >>> behavior is a central aspect to this series.  This effort has dragged
> >>> out for nearly two years and I apologize, but I don't really have a lot
> >>> of patience for rehashing some of these issues if you're not going to
> >>> read the previous discussions or consult with your colleagues to
> >>> understand how we got to this point.  If you want to challenge some of
> >>> the design points, that's great, it could use some new eyes, but please
> >>> understand how we got here first.    
> >>
> >> Well I'm guessing Cao jin here isn't the only one not
> >> willing to plough through all historical versions of the patchset
> >> just to figure out the motivation for some code.
> >>
> >> Including a summary of a high level architecture couldn't hurt.
> >>
> >> Any chance of writing such?  Alternatively, we can try to build it as
> >> part of this thread.  Shouldn't be hard as it seems somewhat
> >> straight-forward on the surface:
> >>
> >> - detect link error on the host, don't reset link as we would normally do  
> > 
> > This is actually a new approach that I'm not sure I agree with.  By
> > skipping the host directed link reset, vfio is taking responsibility
> > for doing this, but then we just assume the user will do it.  I have
> > issues with this.
> > 
> > The previous approach was to use the error detected notifier to block
> > access to the device, allowing the host to perform the link reset.  A
> > subsequent notification in the AER process released the user access
> > which allowed the user AER process to proceed.  This did result in both
> > a host directed and a guest directed link reset, but other than
> > coordinating the blocking of the user process during host reset, that
> > hasn't been brought up as an issue previously.
> >   
> 
> Tests on previous versions didn't bring up issues as I find, I think
> that is because we didn't test it completely. As I know, before August
> of this year, we didn't have cable connected to NIC, let alone
> connecting NIC to gateway.

Lack of testing has been a significant issue throughout the development
of this series.

> Even if I fixed the guest oops issue in igb driver that Alex found in
> v9, v9 still cannot work in my test. And in my test, disable link
> reset(in host) in aer core for vfio-pci is the most significant step to
> get my test passed.

But is it the correct step?  I'm not convinced.  Why did blocking guest
access not work?  How do you plan to manage vfio taking the
responsibility to perform a bus reset when you don't know whether QEMU
is the user of the device or whether the user supports AER recovery?
 
> >> - report link error to guest
> >> - detect link reset request from guest
> >> - reset link on host
> >>
> >> Since link reset will reset all devices behind it, for this to work we
> >> need same set of devices behind the link in host and guest.  Enforcing
> >> this would be nice to have.  
> > 
> > This is a pretty significant complication and I think it's a
> > requirement.  This is the heart of why we have an AER vfio-pci device
> > option and why we require that QEMU should fail to initialize the
> > device if AER is enabled in an incompatible configuration.  If a user
> > cannot be sure that AER is enabled on a device, it's pointless to
> > bother implementing it, though honestly I question the value of it in
> > the VM altogether given configuration requirements (ie. are users
> > going to accept the reason that all the ports of a device need to be
> > assigned to a single guest for guest-based AER recovery when they were
> > previously able to assign each port to a separate VM?).
> >    
> >> - as link now might end up in bad state, reset
> >>   it when device is unassigned  
> > 
> > This is also a new aspect for the approach here, previously we allowed
> > the host directed link reset so we could assume the device was
> > sufficiently recovered.  In the proposal here, the AER core skips any
> > devices bound to vfio-pci, but vfio can't guarantee that we can do a
> > bus reset on them.  PCI bus isolation is not accounted for in DMA
> > isolation, which is the basis for IOMMU groups.  A bus can host
> > multiple IOMMU groups, each of which may have a different user.  Only
> > in a very specific configuration can vfio do a bus reset.
> >    
> >> Any details I missed?  
> > 
> > AIUI, the critical feature is that the guest needs to be able to reset
> > the device link, all the other design elements play out from the
> > logical expansion of that feature.  It means that a guest bus reset
> > needs to translate to a host bus reset, which means that all of the
> > affected host devices need to be accounted for and those that are
> > assigned to the guest need to be affected in the guest in the same
> > way.  QEMU must enforce this configuration or else a user cannot know
> > the result of a AER fault, ie. will it cause a VMSTOP condition or is
> > the fault forwarded to the guest.  The required configuration
> > restrictions are quite involved, therefore we can't simply require this
> > of all configurations, so a vfio-pci device option is added.  The
> > coordination of the host directed reset with the guest directed reset
> > was only a complications discovered within the last few revisions of
> > the series.  As noted above, the previous solution to this was to
> > attempt to block access to the device while the host reset proceeds.
> > 
> > Clearly it's a little disconcerting if we throw all of that away and
> > simply assume that an FLR is sufficient to reset the device when it
> > seems like link issues might be a nontrivial source of AER faults.  If
> > FLR is sufficient, then why does the core AER handling code in the
> > kernel not simply do this?  Thanks,
> >   
> 
> I agree with the points here.  Now I understand why  translate a guest
> link reset to host link reset[*], and FLR shouldn't be equivalent to
> link reset, but the PITY is, we can't trigger a real fatal error to see
> if a FLR is sufficient to reset the device.

In this case, it's not a matter of finding a scenario where it works.
Using FLR to recover from a fatal error would need to be supported by
verbiage in a specification.  I'm not interested in cases where it
happens to work on single device for a certain type of error.  The link
reset is the bare metal mechanism for recovering from a fatal error and
it should be our mechanism as well unless there's spec wording to
support another approach.  Thanks,

Alex

[toc] | [prev] | [next] | [standalone]

Page 1 of 2  [1] 2  Next page →

Back to top | Article view | linux.kernel

csiph-web