Skip to contentSkip to conversation

fpga: dfl: afu: support Rev 2 of DFL Port feature

PATCH In review Matthew Gerlach 2 participants 12 messages Jan 25, 2024 – Mar 2, 2024

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 fpga: dfl: afu: support Rev 2 of DFL Port feature

Revision 2 of the Device Feature List (DFL) Port feature adds support for connecting the contents of the port to multiple PCIe Physical Functions (PF).

This new functionality requires changing the port reset behavior during FPGA and software initialization from revision 1 of the port feature. With revision 1, the initial state of the logic inside the port was not guaranteed to be valid until a port reset was performed by software during driver initialization. With revision 2, the initial state of the logic inside the port is guaranteed to be valid, and a port reset is not required during driver initialization.

This change in port reset behavior avoids a potential race condition during PCI enumeration when a port is connected to multiple PFs. Problems can occur if the driver attached to the PF managing the port asserts reset in its probe function when a driver attached to another PF accesses the port in its own probe function. The potential problems include failed or hung transaction layer packet (TLP) transactions and invalid data being returned.

---
v2:
- Update commit message for clarity
- Remove potentially confusing dev_info message.

---
 drivers/fpga/dfl-afu-main.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
drivers/fpga/dfl-afu-main.c +9 -1
index c0a75ca360d6..42fe27660ab7 100644
@@ -417,7 +417,15 @@ static const struct attribute_group port_hdr_group = {
 static int port_hdr_init(struct platform_device *pdev,
 			 struct dfl_feature *feature)
 {
-	port_reset(pdev);
+	void __iomem *base;
+	u8 rev;
+
+	base = dfl_get_feature_ioaddr_by_id(&pdev->dev, PORT_FEATURE_ID_HEADER);
+
+	rev = dfl_feature_revision(base);
+
+	if (rev < 2)
+		port_reset(pdev);
 
 	return 0;
 }
Signature · 1 line
2.34.1
#1 lore raw
Quoted 13 lines · Matthew Gerlach

Revision 2 of the Device Feature List (DFL) Port feature adds support for connecting the contents of the port to multiple PCIe Physical Functions (PF).

This new functionality requires changing the port reset behavior during FPGA and software initialization from revision 1 of the port feature. With revision 1, the initial state of the logic inside the port was not guaranteed to be valid until a port reset was performed by software during driver initialization. With revision 2, the initial state of the logic inside the port is guaranteed to be valid, and a port reset is not required during driver initialization.

Quoted 6 lines

This change in port reset behavior avoids a potential race condition during PCI enumeration when a port is connected to multiple PFs. Problems can occur if the driver attached to the PF managing the port asserts reset in its probe function when a driver attached to another PF accesses the port in its own probe function. The potential problems include failed or hung

Only racing during probe functions? I assume any time port_reset() would fail TLPs for the other PF. And port_reset() could be triggered at runtime by ioctl().

Thanks,
Yilun

Quoted 38 lines

transaction layer packet (TLP) transactions and invalid data being returned.

---
v2:
- Update commit message for clarity
- Remove potentially confusing dev_info message.

---
 drivers/fpga/dfl-afu-main.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
drivers/fpga/dfl-afu-main.c +9 -1
index c0a75ca360d6..42fe27660ab7 100644
@@ -417,7 +417,15 @@ static const struct attribute_group port_hdr_group = {
 static int port_hdr_init(struct platform_device *pdev,
 			 struct dfl_feature *feature)
 {
-	port_reset(pdev);
+	void __iomem *base;
+	u8 rev;
+
+	base = dfl_get_feature_ioaddr_by_id(&pdev->dev, PORT_FEATURE_ID_HEADER);
+
+	rev = dfl_feature_revision(base);
+
+	if (rev < 2)
+		port_reset(pdev);
 
 	return 0;
 }
Signature · 1 line
2.34.1
#2 lore raw
Quoted 21 lines · Xu Yilun
Quoted 13 lines · Matthew Gerlach

Revision 2 of the Device Feature List (DFL) Port feature adds support for connecting the contents of the port to multiple PCIe Physical Functions (PF).

This new functionality requires changing the port reset behavior during FPGA and software initialization from revision 1 of the port feature. With revision 1, the initial state of the logic inside the port was not guaranteed to be valid until a port reset was performed by software during driver initialization. With revision 2, the initial state of the logic inside the port is guaranteed to be valid, and a port reset is not required during driver initialization.

Quoted 6 lines

This change in port reset behavior avoids a potential race condition during PCI enumeration when a port is connected to multiple PFs. Problems can occur if the driver attached to the PF managing the port asserts reset in its probe function when a driver attached to another PF accesses the port in its own probe function. The potential problems include failed or hung

Quoted 3 lines

Only racing during probe functions? I assume any time port_reset() would fail TLPs for the other PF. And port_reset() could be triggered at runtime by ioctl().

Yes, a port_reset() triggered by ioctl could result in failed TLP for the other PFs. The user space SW performing the ioctl needs to ensure all PFs involved are properly quiesced before the port_reset is performed.

Do you want me to update the commit message with this information?

Thanks,
Matthew

Quoted 43 lines

Thanks,
Yilun

Quoted 38 lines

transaction layer packet (TLP) transactions and invalid data being returned.

---
v2:
- Update commit message for clarity
- Remove potentially confusing dev_info message.

---
 drivers/fpga/dfl-afu-main.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
drivers/fpga/dfl-afu-main.c +9 -2
index c0a75ca360d6..42fe27660ab7 100644
@@ -417,7 +417,15 @@ static const struct attribute_group port_hdr_group = {
 static int port_hdr_init(struct platform_device *pdev,
 			 struct dfl_feature *feature)
 {
-	port_reset(pdev);
+	void __iomem *base;
+	u8 rev;
+
+	base = dfl_get_feature_ioaddr_by_id(&pdev->dev, PORT_FEATURE_ID_HEADER);
+
+	rev = dfl_feature_revision(base);
+
+	if (rev < 2)
+		port_reset(pdev);
 
 	return 0;
 }
--

2.34.1

#3 lore raw
Quoted 25 lines · On Tue, Jan 30, 2024 at 10:00:16AM -0800, [email protected] wrote:
Quoted 21 lines · Xu Yilun
Quoted 19 lines · Matthew Gerlach

Revision 2 of the Device Feature List (DFL) Port feature adds support for connecting the contents of the port to multiple PCIe Physical Functions (PF).

This new functionality requires changing the port reset behavior during FPGA and software initialization from revision 1 of the port feature. With revision 1, the initial state of the logic inside the port was not guaranteed to be valid until a port reset was performed by software during driver initialization. With revision 2, the initial state of the logic inside the port is guaranteed to be valid, and a port reset is not required during driver initialization.

This change in port reset behavior avoids a potential race condition during PCI enumeration when a port is connected to multiple PFs. Problems can occur if the driver attached to the PF managing the port asserts reset in its probe function when a driver attached to another PF accesses the port in its own probe function. The potential problems include failed or hung

Quoted 7 lines
Quoted 3 lines

Only racing during probe functions? I assume any time port_reset() would fail TLPs for the other PF. And port_reset() could be triggered at runtime by ioctl().

Yes, a port_reset() triggered by ioctl could result in failed TLP for the other PFs. The user space SW performing the ioctl needs to ensure all PFs involved are properly quiesced before the port_reset is performed.

How would user get an insight into other PF drivers to know everything is quiesced? I mean do we need driver level management for this?

Thanks,
Yilun

Quoted 5 lines

Do you want me to update the commit message with this information?

Thanks,
Matthew

#4 lore raw
Quoted 30 lines · Xu Yilun
Quoted 29 lines · On Tue, Jan 30, 2024 at 10:00:16AM -0800, [email protected] wrote:
Quoted 24 lines · Xu Yilun
Quoted 13 lines · Matthew Gerlach

Revision 2 of the Device Feature List (DFL) Port feature adds support for connecting the contents of the port to multiple PCIe Physical Functions (PF).

This new functionality requires changing the port reset behavior during FPGA and software initialization from revision 1 of the port feature. With revision 1, the initial state of the logic inside the port was not guaranteed to be valid until a port reset was performed by software during driver initialization. With revision 2, the initial state of the logic inside the port is guaranteed to be valid, and a port reset is not required during driver initialization.

Quoted 6 lines

This change in port reset behavior avoids a potential race condition during PCI enumeration when a port is connected to multiple PFs. Problems can occur if the driver attached to the PF managing the port asserts reset in its probe function when a driver attached to another PF accesses the port in its own probe function. The potential problems include failed or hung

Only racing during probe functions? I assume any time port_reset() would fail TLPs for the other PF. And port_reset() could be triggered at runtime by ioctl().

Quoted 6 lines
Quoted 3 lines

Yes, a port_reset() triggered by ioctl could result in failed TLP for the other PFs. The user space SW performing the ioctl needs to ensure all PFs involved are properly quiesced before the port_reset is performed.

How would user get an insight into other PF drivers to know everything is quiesced? I mean do we need driver level management for this?

Since this is an FPGA, the number of other PFs and the drivers bound to those PFs depends on the FPGA image. There would also be user space software stacks involved with the other PFs as well. The user would have to ensure all the SW stacks and drivers are quiesced as appropriate for the FPGA image. I don't think the driver performing the port_reset() can know all the components to be able to provide any meaningful management.

Thanks,
Matthew

Quoted 11 lines

Thanks,
Yilun

Quoted 5 lines

Do you want me to update the commit message with this information?

Thanks,
Matthew

#5 lore raw
Quoted 38 lines · On Wed, Jan 31, 2024 at 04:26:27PM -0800, [email protected] wrote:
Quoted 34 lines · Xu Yilun
Quoted 32 lines · On Tue, Jan 30, 2024 at 10:00:16AM -0800, [email protected] wrote:
Quoted 21 lines · Xu Yilun
Quoted 13 lines · Matthew Gerlach

Revision 2 of the Device Feature List (DFL) Port feature adds support for connecting the contents of the port to multiple PCIe Physical Functions (PF).

This new functionality requires changing the port reset behavior during FPGA and software initialization from revision 1 of the port feature. With revision 1, the initial state of the logic inside the port was not guaranteed to be valid until a port reset was performed by software during driver initialization. With revision 2, the initial state of the logic inside the port is guaranteed to be valid, and a port reset is not required during driver initialization.

Quoted 6 lines

This change in port reset behavior avoids a potential race condition during PCI enumeration when a port is connected to multiple PFs. Problems can occur if the driver attached to the PF managing the port asserts reset in its probe function when a driver attached to another PF accesses the port in its own probe function. The potential problems include failed or hung

Quoted 3 lines

Only racing during probe functions? I assume any time port_reset() would fail TLPs for the other PF. And port_reset() could be triggered at runtime by ioctl().

Yes, a port_reset() triggered by ioctl could result in failed TLP for the other PFs. The user space SW performing the ioctl needs to ensure all PFs involved are properly quiesced before the port_reset is performed.

Quoted 7 lines
Quoted 2 lines

How would user get an insight into other PF drivers to know everything is quiesced? I mean do we need driver level management for this?

Since this is an FPGA, the number of other PFs and the drivers bound to those PFs depends on the FPGA image. There would also be user space software stacks involved with the other PFs as well. The user would have to ensure all the SW stacks and drivers are quiesced as appropriate for the FPGA

User may not know everything about the device, they only get part of the controls that drivers grant. This is still true for vfio + userspace drivers.

Quoted 1 line

image. I don't think the driver performing the port_reset() can know all the

Other PF drivers should know their own components. They should be aware that their devices are being reset.

Quoted 1 line

components to be able to provide any meaningful management.

If the reset provider and reset consumer are not in the same driver, they should interact with each other. IIRC, some reset controller class works for this purpose.

Thanks,
Yilun

Quoted 15 lines

Thanks,
Matthew

Quoted 11 lines

Thanks,
Yilun

Quoted 5 lines

Do you want me to update the commit message with this information?

Thanks,
Matthew

#6 lore raw
Quoted 42 lines · Xu Yilun
Quoted 41 lines · On Wed, Jan 31, 2024 at 04:26:27PM -0800, [email protected] wrote:
Quoted 36 lines · Xu Yilun
Quoted 25 lines · On Tue, Jan 30, 2024 at 10:00:16AM -0800, [email protected] wrote:
Quoted 21 lines · Xu Yilun
Quoted 19 lines · Matthew Gerlach

Revision 2 of the Device Feature List (DFL) Port feature adds support for connecting the contents of the port to multiple PCIe Physical Functions (PF).

This new functionality requires changing the port reset behavior during FPGA and software initialization from revision 1 of the port feature. With revision 1, the initial state of the logic inside the port was not guaranteed to be valid until a port reset was performed by software during driver initialization. With revision 2, the initial state of the logic inside the port is guaranteed to be valid, and a port reset is not required during driver initialization.

This change in port reset behavior avoids a potential race condition during PCI enumeration when a port is connected to multiple PFs. Problems can occur if the driver attached to the PF managing the port asserts reset in its probe function when a driver attached to another PF accesses the port in its own probe function. The potential problems include failed or hung

Quoted 7 lines
Quoted 3 lines

Only racing during probe functions? I assume any time port_reset() would fail TLPs for the other PF. And port_reset() could be triggered at runtime by ioctl().

Yes, a port_reset() triggered by ioctl could result in failed TLP for the other PFs. The user space SW performing the ioctl needs to ensure all PFs involved are properly quiesced before the port_reset is performed.

How would user get an insight into other PF drivers to know everything is quiesced? I mean do we need driver level management for this?

Quoted 8 lines
Quoted 4 lines

Since this is an FPGA, the number of other PFs and the drivers bound to those PFs depends on the FPGA image. There would also be user space software stacks involved with the other PFs as well. The user would have to ensure all the SW stacks and drivers are quiesced as appropriate for the FPGA

User may not know everything about the device, they only get part of the controls that drivers grant. This is still true for vfio + userspace drivers.

A user performing a port reset would have to know the impact to the specific FPGA image being run in order to ensure all SW stacks are ready for the reset.

Quoted 5 lines
Quoted 2 lines

image. I don't think the driver performing the port_reset() can know all the

Other PF drivers should know their own components. They should be aware that their devices are being reset.

The other PF drivers depend on the actual FPGA image being run.

Quoted 6 lines
Quoted 2 lines

components to be able to provide any meaningful management.

If the reset provider and reset consumer are not in the same driver, they should interact with each other. IIRC, some reset controller class works for this purpose.

The other PFs in many cases can present as standard devices with existing drivers like virtio-net or virtio-blk. It does not seem desireable to have to change existing drivers for a particular FPGA implementation

Thanks,
Matthew

Quoted 21 lines

Thanks,
Yilun

Quoted 15 lines

Thanks,
Matthew

Quoted 11 lines

Thanks,
Yilun

Quoted 5 lines

Do you want me to update the commit message with this information?

Thanks,
Matthew

#7 lore raw
Quoted 51 lines · On Wed, Feb 07, 2024 at 08:40:55AM -0800, [email protected] wrote:
Quoted 47 lines · Xu Yilun
Quoted 45 lines · On Wed, Jan 31, 2024 at 04:26:27PM -0800, [email protected] wrote:
Quoted 30 lines · Xu Yilun
Quoted 29 lines · On Tue, Jan 30, 2024 at 10:00:16AM -0800, [email protected] wrote:
Quoted 24 lines · Xu Yilun
Quoted 13 lines · Matthew Gerlach

Revision 2 of the Device Feature List (DFL) Port feature adds support for connecting the contents of the port to multiple PCIe Physical Functions (PF).

This new functionality requires changing the port reset behavior during FPGA and software initialization from revision 1 of the port feature. With revision 1, the initial state of the logic inside the port was not guaranteed to be valid until a port reset was performed by software during driver initialization. With revision 2, the initial state of the logic inside the port is guaranteed to be valid, and a port reset is not required during driver initialization.

Quoted 6 lines

This change in port reset behavior avoids a potential race condition during PCI enumeration when a port is connected to multiple PFs. Problems can occur if the driver attached to the PF managing the port asserts reset in its probe function when a driver attached to another PF accesses the port in its own probe function. The potential problems include failed or hung

Only racing during probe functions? I assume any time port_reset() would fail TLPs for the other PF. And port_reset() could be triggered at runtime by ioctl().

Quoted 6 lines
Quoted 3 lines

Yes, a port_reset() triggered by ioctl could result in failed TLP for the other PFs. The user space SW performing the ioctl needs to ensure all PFs involved are properly quiesced before the port_reset is performed.

How would user get an insight into other PF drivers to know everything is quiesced? I mean do we need driver level management for this?

Since this is an FPGA, the number of other PFs and the drivers bound to those PFs depends on the FPGA image. There would also be user space software stacks involved with the other PFs as well. The user would have to ensure all the SW stacks and drivers are quiesced as appropriate for the FPGA

Quoted 7 lines
Quoted 3 lines

User may not know everything about the device, they only get part of the controls that drivers grant. This is still true for vfio + userspace drivers.

A user performing a port reset would have to know the impact to the specific FPGA image being run in order to ensure all SW stacks are ready for the reset.

We are not going to change the logic of the whole driver model just because the device is backed up by an FPGA image. The *driver* should be fully responsible for matched devices. A HW reset unaware to the device driver is not wanted. Assuming that the userspace could control every access to device makes no sense.

For your case, there is no garantee userspace could block every access to "other PF" initiated by "other PF" driver. There is also no notification to "other PF" driver that userspace is doing reset to "other PF" via "management PF" interface. "other PF" driver just break on reset.

Quoted 12 lines
Quoted 6 lines
Quoted 3 lines

image. I don't think the driver performing the port_reset() can know all the

Other PF drivers should know their own components. They should be aware that their devices are being reset.

The other PF drivers depend on the actual FPGA image being run.

Quoted 3 lines
Quoted 2 lines

components to be able to provide any meaningful management.

Quoted 7 lines
Quoted 3 lines

If the reset provider and reset consumer are not in the same driver, they should interact with each other. IIRC, some reset controller class works for this purpose.

The other PFs in many cases can present as standard devices with existing drivers like virtio-net or virtio-blk. It does not seem desireable to have to change existing drivers for a particular FPGA implementation

If you have to use a specific method for reset, it is not a standard virtio pci device, and you have to make some change.

Thanks,
Yilun

Quoted 24 lines

Thanks,
Matthew

Quoted 21 lines

Thanks,
Yilun

Quoted 15 lines

Thanks,
Matthew

Quoted 11 lines

Thanks,
Yilun

Quoted 5 lines

Do you want me to update the commit message with this information?

Thanks,
Matthew

#8 lore raw
Quoted 66 lines · Xu Yilun
Quoted 51 lines · On Wed, Feb 07, 2024 at 08:40:55AM -0800, [email protected] wrote:
Quoted 47 lines · Xu Yilun
Quoted 45 lines · On Wed, Jan 31, 2024 at 04:26:27PM -0800, [email protected] wrote:
Quoted 30 lines · Xu Yilun
Quoted 29 lines · On Tue, Jan 30, 2024 at 10:00:16AM -0800, [email protected] wrote:
Quoted 24 lines · Xu Yilun
Quoted 13 lines · Matthew Gerlach

Revision 2 of the Device Feature List (DFL) Port feature adds support for connecting the contents of the port to multiple PCIe Physical Functions (PF).

This new functionality requires changing the port reset behavior during FPGA and software initialization from revision 1 of the port feature. With revision 1, the initial state of the logic inside the port was not guaranteed to be valid until a port reset was performed by software during driver initialization. With revision 2, the initial state of the logic inside the port is guaranteed to be valid, and a port reset is not required during driver initialization.

Quoted 6 lines

This change in port reset behavior avoids a potential race condition during PCI enumeration when a port is connected to multiple PFs. Problems can occur if the driver attached to the PF managing the port asserts reset in its probe function when a driver attached to another PF accesses the port in its own probe function. The potential problems include failed or hung

Only racing during probe functions? I assume any time port_reset() would fail TLPs for the other PF. And port_reset() could be triggered at runtime by ioctl().

Quoted 6 lines
Quoted 3 lines

Yes, a port_reset() triggered by ioctl could result in failed TLP for the other PFs. The user space SW performing the ioctl needs to ensure all PFs involved are properly quiesced before the port_reset is performed.

How would user get an insight into other PF drivers to know everything is quiesced? I mean do we need driver level management for this?

Since this is an FPGA, the number of other PFs and the drivers bound to those PFs depends on the FPGA image. There would also be user space software stacks involved with the other PFs as well. The user would have to ensure all the SW stacks and drivers are quiesced as appropriate for the FPGA

Quoted 7 lines
Quoted 3 lines

User may not know everything about the device, they only get part of the controls that drivers grant. This is still true for vfio + userspace drivers.

A user performing a port reset would have to know the impact to the specific FPGA image being run in order to ensure all SW stacks are ready for the reset.

We are not going to change the logic of the whole driver model just because the device is backed up by an FPGA image. The *driver* should be fully responsible for matched devices. A HW reset unaware to the device driver is not wanted. Assuming that the userspace could control every access to device makes no sense.

Quoted 5 lines

For your case, there is no garantee userspace could block every access to "other PF" initiated by "other PF" driver. There is also no notification to "other PF" driver that userspace is doing reset to "other PF" via "management PF" interface. "other PF" driver just break on reset.

Hi Yilun,

I think this conversation has gotten a little off track. This patch only changes the port reset behavior at driver initialization for revision 2 of the port IP. The behavior and the requirements of port reset during run time have not changed. The existing implementation requires the user performing the port reset to ensure appropriate SW was quiesced. This patch does not change this requirement.

Quoted 17 lines
Quoted 17 lines
Quoted 7 lines
Quoted 4 lines

image. I don't think the driver performing the port_reset() can know all the

Other PF drivers should know their own components. They should be aware that their devices are being reset.

The other PF drivers depend on the actual FPGA image being run.

Quoted 6 lines
Quoted 2 lines

components to be able to provide any meaningful management.

If the reset provider and reset consumer are not in the same driver, they should interact with each other. IIRC, some reset controller class works for this purpose.

Quoted 6 lines
Quoted 3 lines

The other PFs in many cases can present as standard devices with existing drivers like virtio-net or virtio-blk. It does not seem desireable to have to change existing drivers for a particular FPGA implementation

If you have to use a specific method for reset, it is not a standard virtio pci device, and you have to make some change.

From the perspective of the PCI PF/VF implementing the virtio or other standard device, the pci endpoint is completely compliant to the standard, and no driver changes should be required. As mentioned above, this patch does nothing to change any of this behavior. Consider a port reset that is part of a partial configuration flow. The virtio endpoint can become something completely different with a completely different driver. This patch does not affect this flow either.

Thanks,
Matthew

Quoted 29 lines

Thanks,
Yilun

Quoted 24 lines

Thanks,
Matthew

Quoted 21 lines

Thanks,
Yilun

Quoted 15 lines

Thanks,
Matthew

Quoted 11 lines

Thanks,
Yilun

Quoted 5 lines

Do you want me to update the commit message with this information?

Thanks,
Matthew

#9 lore raw
Quoted 76 lines · On Tue, Feb 20, 2024 at 05:49:04PM -0800, [email protected] wrote:
Quoted 66 lines · Xu Yilun
Quoted 51 lines · On Wed, Feb 07, 2024 at 08:40:55AM -0800, [email protected] wrote:
Quoted 47 lines · Xu Yilun
Quoted 45 lines · On Wed, Jan 31, 2024 at 04:26:27PM -0800, [email protected] wrote:
Quoted 30 lines · Xu Yilun
Quoted 29 lines · On Tue, Jan 30, 2024 at 10:00:16AM -0800, [email protected] wrote:
Quoted 24 lines · Xu Yilun
Quoted 13 lines · Matthew Gerlach

Revision 2 of the Device Feature List (DFL) Port feature adds support for connecting the contents of the port to multiple PCIe Physical Functions (PF).

This new functionality requires changing the port reset behavior during FPGA and software initialization from revision 1 of the port feature. With revision 1, the initial state of the logic inside the port was not guaranteed to be valid until a port reset was performed by software during driver initialization. With revision 2, the initial state of the logic inside the port is guaranteed to be valid, and a port reset is not required during driver initialization.

Quoted 6 lines

This change in port reset behavior avoids a potential race condition during PCI enumeration when a port is connected to multiple PFs. Problems can occur if the driver attached to the PF managing the port asserts reset in its probe function when a driver attached to another PF accesses the port in its own probe function. The potential problems include failed or hung

Only racing during probe functions? I assume any time port_reset() would fail TLPs for the other PF. And port_reset() could be triggered at runtime by ioctl().

Quoted 6 lines
Quoted 3 lines

Yes, a port_reset() triggered by ioctl could result in failed TLP for the other PFs. The user space SW performing the ioctl needs to ensure all PFs involved are properly quiesced before the port_reset is performed.

How would user get an insight into other PF drivers to know everything is quiesced? I mean do we need driver level management for this?

Since this is an FPGA, the number of other PFs and the drivers bound to those PFs depends on the FPGA image. There would also be user space software stacks involved with the other PFs as well. The user would have to ensure all the SW stacks and drivers are quiesced as appropriate for the FPGA

Quoted 7 lines
Quoted 3 lines

User may not know everything about the device, they only get part of the controls that drivers grant. This is still true for vfio + userspace drivers.

A user performing a port reset would have to know the impact to the specific FPGA image being run in order to ensure all SW stacks are ready for the reset.

We are not going to change the logic of the whole driver model just because the device is backed up by an FPGA image. The *driver* should be fully responsible for matched devices. A HW reset unaware to the device driver is not wanted. Assuming that the userspace could control every access to device makes no sense.

Quoted 5 lines

For your case, there is no garantee userspace could block every access to "other PF" initiated by "other PF" driver. There is also no notification to "other PF" driver that userspace is doing reset to "other PF" via "management PF" interface. "other PF" driver just break on reset.

Quoted 8 lines

Hi Yilun,

I think this conversation has gotten a little off track. This patch only changes the port reset behavior at driver initialization for revision 2 of the port IP. The behavior and the requirements of port reset during run time have not changed. The existing implementation requires the user performing the port reset to ensure appropriate SW was quiesced. This patch does not change this requirement.

You are actually adding support to the new devices behavior defined by revision 2. Previously the user requires the management PF driver to do port reset, and this only affects some logic in the PF itself and the impact could be handled inside the driver. The current PF driver doesn't actually do anything because it (or any kernel component) never touches the disabled logic, and the user accessing of the mmapped but disabled logic won't cause system issues.

But the new management PF does port reset and affect other PFs, and may cause disorder in other drivers. so you need extra code to support the behavior.

This patch does make some progress, as you said, "With revision 2,the initial state of the logic inside the port is guaranteed to be valid, and a port reset is not required during driver initialization". But it should not be the only patch to enable the new port reset behavior.

Quoted 25 lines
Quoted 18 lines
Quoted 18 lines
Quoted 8 lines
Quoted 5 lines

image. I don't think the driver performing the port_reset() can know all the

Other PF drivers should know their own components. They should be aware that their devices are being reset.

The other PF drivers depend on the actual FPGA image being run.

Quoted 6 lines
Quoted 2 lines

components to be able to provide any meaningful management.

If the reset provider and reset consumer are not in the same driver, they should interact with each other. IIRC, some reset controller class works for this purpose.

Quoted 6 lines
Quoted 3 lines

The other PFs in many cases can present as standard devices with existing drivers like virtio-net or virtio-blk. It does not seem desireable to have to change existing drivers for a particular FPGA implementation

If you have to use a specific method for reset, it is not a standard virtio pci device, and you have to make some change.

Quoted 6 lines

From the perspective of the PCI PF/VF implementing the virtio or other standard device, the pci endpoint is completely compliant to the standard, and no driver changes should be required. As mentioned above, this patch does nothing to change any of this behavior. Consider a port reset that is part of a partial configuration flow. The virtio endpoint can become something completely different with a completely different driver. This

Then how could the virtio driver stop and remove itself to avoid crashing the kernel, and how could the new driver be probed. If the answer is still userspace responsible for everything, that's not good to me.

Thanks,
Yilun

Quoted 34 lines

patch does not affect this flow either.

Thanks,
Matthew

Quoted 29 lines

Thanks,
Yilun

Quoted 24 lines

Thanks,
Matthew

Quoted 21 lines

Thanks,
Yilun

Quoted 15 lines

Thanks,
Matthew

Quoted 11 lines

Thanks,
Yilun

Quoted 5 lines

Do you want me to update the commit message with this information?

Thanks,
Matthew

#10 lore raw

[snip]

Quoted 10 lines
Quoted 9 lines

Hi Yilun,

I think this conversation has gotten a little off track. This patch only changes the port reset behavior at driver initialization for revision 2 of the port IP. The behavior and the requirements of port reset during run time have not changed. The existing implementation requires the user performing the port reset to ensure appropriate SW was quiesced. This patch does not change this requirement.

Quoted 6 lines

You are actually adding support to the new devices behavior defined by revision 2. Previously the user requires the management PF driver to do port reset, and this only affects some logic in the PF itself and the impact could be handled inside the driver. The current PF driver doesn't actually do anything because it (or any kernel component) never touches the disabled logic, and the user accessing of the mmapped but disabled logic won't cause system issues.

I'm wondering about the use case when Virtual Functions (VFs) are enabled with the current driver. If partial reconfiguration of the port occurs, which includes a reset, what happens to the user software attached to the VFs when the hardware has changed out from under it? While the current port implementation won't cause PCIe errors when the logic is disable, something has to notify the user software that the hardware changed. I think the user space software would be typically notified of the change by the orchestration software performing the partial reconfiguration. I think the use case of VFs with the original port implementation is logically equivalent to the new port implementation with multiple PFs/VFs.

Quoted 4 lines

But the new management PF does port reset and affect other PFs, and may cause disorder in other drivers. so you need extra code to support the behavior.

I think the real problem is that the new port is not doing as good of a job preventing PCIe errors during port reset as the previous version.

Matthew

Quoted 80 lines

This patch does make some progress, as you said, "With revision 2,the initial state of the logic inside the port is guaranteed to be valid, and a port reset is not required during driver initialization". But it should not be the only patch to enable the new port reset behavior.

Quoted 25 lines
Quoted 18 lines
Quoted 18 lines
Quoted 8 lines
Quoted 5 lines

image. I don't think the driver performing the port_reset() can know all the

Other PF drivers should know their own components. They should be aware that their devices are being reset.

The other PF drivers depend on the actual FPGA image being run.

Quoted 6 lines
Quoted 2 lines

components to be able to provide any meaningful management.

If the reset provider and reset consumer are not in the same driver, they should interact with each other. IIRC, some reset controller class works for this purpose.

Quoted 6 lines
Quoted 3 lines

The other PFs in many cases can present as standard devices with existing drivers like virtio-net or virtio-blk. It does not seem desireable to have to change existing drivers for a particular FPGA implementation

If you have to use a specific method for reset, it is not a standard virtio pci device, and you have to make some change.

Quoted 6 lines

From the perspective of the PCI PF/VF implementing the virtio or other standard device, the pci endpoint is completely compliant to the standard, and no driver changes should be required. As mentioned above, this patch does nothing to change any of this behavior. Consider a port reset that is part of a partial configuration flow. The virtio endpoint can become something completely different with a completely different driver. This

Then how could the virtio driver stop and remove itself to avoid crashing the kernel, and how could the new driver be probed. If the answer is still userspace responsible for everything, that's not good to me.

Thanks,
Yilun

Quoted 34 lines

patch does not affect this flow either.

Thanks,
Matthew

Quoted 29 lines

Thanks,
Yilun

Quoted 24 lines

Thanks,
Matthew

Quoted 21 lines

Thanks,
Yilun

Quoted 15 lines

Thanks,
Matthew

Quoted 11 lines

Thanks,
Yilun

Quoted 5 lines

Do you want me to update the commit message with this information?

Thanks,
Matthew

#11 lore raw
Quoted 21 lines · On Fri, Mar 01, 2024 at 12:39:24PM -0800, [email protected] wrote:

[snip]

Quoted 10 lines
Quoted 9 lines

Hi Yilun,

I think this conversation has gotten a little off track. This patch only changes the port reset behavior at driver initialization for revision 2 of the port IP. The behavior and the requirements of port reset during run time have not changed. The existing implementation requires the user performing the port reset to ensure appropriate SW was quiesced. This patch does not change this requirement.

Quoted 6 lines

You are actually adding support to the new devices behavior defined by revision 2. Previously the user requires the management PF driver to do port reset, and this only affects some logic in the PF itself and the impact could be handled inside the driver. The current PF driver doesn't actually do anything because it (or any kernel component) never touches the disabled logic, and the user accessing of the mmapped but disabled logic won't cause system issues.

Quoted 4 lines

I'm wondering about the use case when Virtual Functions (VFs) are enabled with the current driver. If partial reconfiguration of the port occurs, which includes a reset, what happens to the user software attached to the VFs when the hardware has changed out from under it? While the current port

Ah, yes. That may be the problem for user, you can fix it. But the fix could actually be easier, cause the port driver is fully aware of the reset request and the reset flow is under its control.

Quoted 6 lines

implementation won't cause PCIe errors when the logic is disable, something has to notify the user software that the hardware changed. I think the user space software would be typically notified of the change by the orchestration software performing the partial reconfiguration. I think the use case of VFs with the original port implementation is logically equivalent to the new port implementation with multiple PFs/VFs.

The difference is that, the original port driver is fully responsible for the reset flow and the AFU mapping, while the new port implementation has no control to the to-be-resetted logic, it blindly does the reset.

Quoted 8 lines
Quoted 5 lines

But the new management PF does port reset and affect other PFs, and may cause disorder in other drivers. so you need extra code to support the behavior.

I think the real problem is that the new port is not doing as good of a job preventing PCIe errors during port reset as the previous version.

Even if the reset won't cause PCIe errors, the other PF/VF drivers may still meet HW access issues not understandable by themselves and cause kernel level panic.

Thanks,
Yilun

Quoted 82 lines

Matthew

Quoted 80 lines

This patch does make some progress, as you said, "With revision 2,the initial state of the logic inside the port is guaranteed to be valid, and a port reset is not required during driver initialization". But it should not be the only patch to enable the new port reset behavior.

Quoted 25 lines
Quoted 18 lines
Quoted 18 lines
Quoted 8 lines
Quoted 5 lines

image. I don't think the driver performing the port_reset() can know all the

Other PF drivers should know their own components. They should be aware that their devices are being reset.

The other PF drivers depend on the actual FPGA image being run.

Quoted 6 lines
Quoted 2 lines

components to be able to provide any meaningful management.

If the reset provider and reset consumer are not in the same driver, they should interact with each other. IIRC, some reset controller class works for this purpose.

Quoted 6 lines
Quoted 3 lines

The other PFs in many cases can present as standard devices with existing drivers like virtio-net or virtio-blk. It does not seem desireable to have to change existing drivers for a particular FPGA implementation

If you have to use a specific method for reset, it is not a standard virtio pci device, and you have to make some change.

Quoted 6 lines

From the perspective of the PCI PF/VF implementing the virtio or other standard device, the pci endpoint is completely compliant to the standard, and no driver changes should be required. As mentioned above, this patch does nothing to change any of this behavior. Consider a port reset that is part of a partial configuration flow. The virtio endpoint can become something completely different with a completely different driver. This

Then how could the virtio driver stop and remove itself to avoid crashing the kernel, and how could the new driver be probed. If the answer is still userspace responsible for everything, that's not good to me.

Thanks,
Yilun

Quoted 34 lines

patch does not affect this flow either.

Thanks,
Matthew

Quoted 29 lines

Thanks,
Yilun

Quoted 24 lines

Thanks,
Matthew

Quoted 21 lines

Thanks,
Yilun

Quoted 15 lines

Thanks,
Matthew

Quoted 11 lines

Thanks,
Yilun

Quoted 5 lines

Do you want me to update the commit message with this information?

Thanks,
Matthew