[PATCH] firmware: arm_scmi: Set device ready before register scmi_virtio_driver

[PATCH] firmware: arm_scmi: Set device ready before register scmi_virtio_driver

[Junnan Wu]

Fix irregularities in scmi driver probe.

After 'commit 20bda12a0ea0 ("firmware: arm_scmi: Make VirtIO transport a standalone driver")'
making SCMI VirtIO transport a standalone driver,
`scmi_virtio_probe` will follow `scmi_vio_probe`,
and during `scmi_virtio_probe`, it will do `scmi_probe` which will invoke
function `scmi_protocol_acquire(handle, SCMI_PROTOCOL_BASE)`,
and this function will send message via virtqueue and wait device's reply

Before this patch, only after `scmi_vio_probe`,
frontend will mark device DRIVER_OK by function `virtio_dev_probe`
in 'drivers/virtio/virtio.c' itself.

Accroding to chapter '3.1 Device Initialization' in virtio-spec,
```
The driver MUST NOT send any buffer available notifications
to the device before setting DRIVER_OK.
```

In some type1-hypervisor, available buffer notifications
will be intercepted until virtio driver is okay. In this case,
`scmi_vio_probe` will be blocked at `scmi_wait_for_reply`
and probe can never complete.

Therefore, the operation to set DRIVER_OK must follow
perform device-specific setup immediately,
then available buffer notifications can be sent in function `scmi_probe`.

Fixes: 20bda12a0ea0 ("firmware: arm_scmi: Make VirtIO transport a standalone driver")
Signed-off-by: Junnan Wu <junnan01.wu@samsung.com>
---
 drivers/firmware/arm_scmi/transports/virtio.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/drivers/firmware/arm_scmi/transports/virtio.c b/drivers/firmware/arm_scmi/transports/virtio.c
index cb934db9b2b4..326c4a93e44b 100644
--- a/drivers/firmware/arm_scmi/transports/virtio.c
+++ b/drivers/firmware/arm_scmi/transports/virtio.c
@@ -871,6 +871,9 @@ static int scmi_vio_probe(struct virtio_device *vdev)
 	/* Ensure initialized scmi_vdev is visible */
 	smp_store_mb(scmi_vdev, vdev);
 
+	/* Set device ready */
+	virtio_device_ready(vdev);
+
 	ret = platform_driver_register(&scmi_virtio_driver);
 	if (ret) {
 		vdev->priv = NULL;
-- 
2.34.1

回复: [PATCH] firmware: arm_scmi: Set device ready before register scmi_virtio_driver

[yizhijiao2025]

-----邮件原件-----
发件人: arm-scmi+bounces-2017-yizhijiao2025=163.com@vger.kernel.org <arm-scmi+bounces-2017-yizhijiao2025=163.com@vger.kernel.org> 代表 Junnan Wu
发送时间: 2025年8月12日 15:54
收件人: sudeep.holla@arm.com; cristian.marussi@arm.com
抄送: florian.fainelli@broadcom.com; arm-scmi@vger.kernel.org; linux-arm-kernel@lists.infradead.org; linux-kernel@vger.kernel.org; lei19.wang@samsung.com; q1.huang@samsung.com; Junnan Wu <junnan01.wu@samsung.com>
主题: [PATCH] firmware: arm_scmi: Set device ready before register scmi_virtio_driver

Fix irregularities in scmi driver probe.

After 'commit 20bda12a0ea0 ("firmware: arm_scmi: Make VirtIO transport a standalone driver")'
making SCMI VirtIO transport a standalone driver, `scmi_virtio_probe` will follow `scmi_vio_probe`, and during `scmi_virtio_probe`, it will do `scmi_probe` which will invoke function `scmi_protocol_acquire(handle, SCMI_PROTOCOL_BASE)`, and this function will send message via virtqueue and wait device's reply

Before this patch, only after `scmi_vio_probe`, frontend will mark device DRIVER_OK by function `virtio_dev_probe` in 'drivers/virtio/virtio.c' itself.

Accroding to chapter '3.1 Device Initialization' in virtio-spec, ``` The driver MUST NOT send any buffer available notifications to the device before setting DRIVER_OK.
```

In some type1-hypervisor, available buffer notifications will be intercepted until virtio driver is okay. In this case, `scmi_vio_probe` will be blocked at `scmi_wait_for_reply` and probe can never complete.

Therefore, the operation to set DRIVER_OK must follow perform device-specific setup immediately, then available buffer notifications can be sent in function `scmi_probe`.

Fixes: 20bda12a0ea0 ("firmware: arm_scmi: Make VirtIO transport a standalone driver")
Signed-off-by: Junnan Wu <junnan01.wu@samsung.com>
---
 drivers/firmware/arm_scmi/transports/virtio.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/drivers/firmware/arm_scmi/transports/virtio.c b/drivers/firmware/arm_scmi/transports/virtio.c
index cb934db9b2b4..326c4a93e44b 100644
--- a/drivers/firmware/arm_scmi/transports/virtio.c
+++ b/drivers/firmware/arm_scmi/transports/virtio.c
@@ -871,6 +871,9 @@ static int scmi_vio_probe(struct virtio_device *vdev)
 	/* Ensure initialized scmi_vdev is visible */
 	smp_store_mb(scmi_vdev, vdev);
 
+	/* Set device ready */
+	virtio_device_ready(vdev);
+
 	ret = platform_driver_register(&scmi_virtio_driver);
 	if (ret) {
 		vdev->priv = NULL;
--
2.34.1

Re: [PATCH] firmware: arm_scmi: Set device ready before register scmi_virtio_driver

[Cristian Marussi]

On Tue, Aug 12, 2025 at 03:53:43PM +0800, Junnan Wu wrote:
> Fix irregularities in scmi driver probe.
> 

Good catch, thanks for this: indeed after the transports have been
converted to modules the probe sequence changed a bit and we cannot
leverage anymore the virtio_device_ready() called by the core in
virtio_dev_probe() since the scmi_vio_probe() does not complete until
the core SCMI stack has probed with scmi_probe() which in turn transmits
the first BASE protocol exchanges.

I did not know about this behaviour of type_1 hyp..

> After 'commit 20bda12a0ea0 ("firmware: arm_scmi: Make VirtIO transport a standalone driver")'
> making SCMI VirtIO transport a standalone driver,
> `scmi_virtio_probe` will follow `scmi_vio_probe`,
> and during `scmi_virtio_probe`, it will do `scmi_probe` which will invoke
> function `scmi_protocol_acquire(handle, SCMI_PROTOCOL_BASE)`,
> and this function will send message via virtqueue and wait device's reply
> 
> Before this patch, only after `scmi_vio_probe`,
> frontend will mark device DRIVER_OK by function `virtio_dev_probe`
> in 'drivers/virtio/virtio.c' itself.
> 
> Accroding to chapter '3.1 Device Initialization' in virtio-spec,
> ```
> The driver MUST NOT send any buffer available notifications
> to the device before setting DRIVER_OK.
> ```
> 
> In some type1-hypervisor, available buffer notifications
> will be intercepted until virtio driver is okay. In this case,
> `scmi_vio_probe` will be blocked at `scmi_wait_for_reply`
> and probe can never complete.
> 
> Therefore, the operation to set DRIVER_OK must follow
> perform device-specific setup immediately,
> then available buffer notifications can be sent in function `scmi_probe`.
> 
> Fixes: 20bda12a0ea0 ("firmware: arm_scmi: Make VirtIO transport a standalone driver")
> Signed-off-by: Junnan Wu <junnan01.wu@samsung.com>
> ---
>  drivers/firmware/arm_scmi/transports/virtio.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/drivers/firmware/arm_scmi/transports/virtio.c b/drivers/firmware/arm_scmi/transports/virtio.c
> index cb934db9b2b4..326c4a93e44b 100644
> --- a/drivers/firmware/arm_scmi/transports/virtio.c
> +++ b/drivers/firmware/arm_scmi/transports/virtio.c
> @@ -871,6 +871,9 @@ static int scmi_vio_probe(struct virtio_device *vdev)
>  	/* Ensure initialized scmi_vdev is visible */
>  	smp_store_mb(scmi_vdev, vdev);
>  
> +	/* Set device ready */
> +	virtio_device_ready(vdev);
> +
>  	ret = platform_driver_register(&scmi_virtio_driver);
>  	if (ret) {
>  		vdev->priv = NULL;

...also, I suppose, that here nothing is to be done on the error path
regarding the vdedv status, becasue is already taken care in the error
path of virtio_dev_probe()...am I right ? (just confirming..)

Maybe you could simplify a bit the commit message but other than this

LGTM.

Reviewed-by: Cristian Marussi <cristian.marussi@arm.com>
Tested-by: Cristian Marussi <cristian.marussi@arm.com>

Thanks,
Cristian

Re: [PATCH] firmware: arm_scmi: Set device ready before register scmi_virtio_driver

[Junnan Wu]

On Tue, 19 Aug 2025 13:33:10 +0100, Cristian Marussi wrote:
> On Tue, Aug 12, 2025 at 03:53:43PM +0800, Junnan Wu wrote:
> > Fix irregularities in scmi driver probe.
> > 
> 
> Good catch, thanks for this: indeed after the transports have been
> converted to modules the probe sequence changed a bit and we cannot
> leverage anymore the virtio_device_ready() called by the core in
> virtio_dev_probe() since the scmi_vio_probe() does not complete until
> the core SCMI stack has probed with scmi_probe() which in turn transmits
> the first BASE protocol exchanges.
> 
> I did not know about this behaviour of type_1 hyp..

It's our local hyp, which intercepts notification until driver is ready.
That helps us to find this irregularity. I think you can ignore this.

> 
> > After 'commit 20bda12a0ea0 ("firmware: arm_scmi: Make VirtIO transport a standalone driver")'
> > making SCMI VirtIO transport a standalone driver,
> > `scmi_virtio_probe` will follow `scmi_vio_probe`,
> > and during `scmi_virtio_probe`, it will do `scmi_probe` which will invoke
> > function `scmi_protocol_acquire(handle, SCMI_PROTOCOL_BASE)`,
> > and this function will send message via virtqueue and wait device's reply
> > 
> > Before this patch, only after `scmi_vio_probe`,
> > frontend will mark device DRIVER_OK by function `virtio_dev_probe`
> > in 'drivers/virtio/virtio.c' itself.
> > 
> > Accroding to chapter '3.1 Device Initialization' in virtio-spec,
> > ```
> > The driver MUST NOT send any buffer available notifications
> > to the device before setting DRIVER_OK.
> > ```
> > 
> > In some type1-hypervisor, available buffer notifications
> > will be intercepted until virtio driver is okay. In this case,
> > `scmi_vio_probe` will be blocked at `scmi_wait_for_reply`
> > and probe can never complete.
> > 
> > Therefore, the operation to set DRIVER_OK must follow
> > perform device-specific setup immediately,
> > then available buffer notifications can be sent in function `scmi_probe`.
> > 
> > Fixes: 20bda12a0ea0 ("firmware: arm_scmi: Make VirtIO transport a standalone driver")
> > Signed-off-by: Junnan Wu <junnan01.wu@samsung.com>
> > ---
> >  drivers/firmware/arm_scmi/transports/virtio.c | 3 +++
> >  1 file changed, 3 insertions(+)
> > 
> > diff --git a/drivers/firmware/arm_scmi/transports/virtio.c b/drivers/firmware/arm_scmi/transports/virtio.c
> > index cb934db9b2b4..326c4a93e44b 100644
> > --- a/drivers/firmware/arm_scmi/transports/virtio.c
> > +++ b/drivers/firmware/arm_scmi/transports/virtio.c
> > @@ -871,6 +871,9 @@ static int scmi_vio_probe(struct virtio_device *vdev)
> >  	/* Ensure initialized scmi_vdev is visible */
> >  	smp_store_mb(scmi_vdev, vdev);
> >  
> > +	/* Set device ready */
> > +	virtio_device_ready(vdev);
> > +
> >  	ret = platform_driver_register(&scmi_virtio_driver);
> >  	if (ret) {
> >  		vdev->priv = NULL;
> 
> ...also, I suppose, that here nothing is to be done on the error path
> regarding the vdedv status, becasue is already taken care in the error
> path of virtio_dev_probe()...am I right ? (just confirming..)
> 

Yes you are right, I also checked it, error is handled by `virtio_dev_probe`

> Maybe you could simplify a bit the commit message but other than this
> 
> LGTM.
> 
> Reviewed-by: Cristian Marussi <cristian.marussi@arm.com>
> Tested-by: Cristian Marussi <cristian.marussi@arm.com>
> 
> Thanks,
> Cristian

Thanks for your kindly review!
BRs
Junnan Wu

Re: [PATCH] firmware: arm_scmi: Set device ready before register scmi_virtio_driver

[Sudeep Holla]

On Tue, 12 Aug 2025 15:53:43 +0800, Junnan Wu wrote:
> Fix irregularities in scmi driver probe.
> 
> After 'commit 20bda12a0ea0 ("firmware: arm_scmi: Make VirtIO transport a standalone driver")'
> making SCMI VirtIO transport a standalone driver,
> `scmi_virtio_probe` will follow `scmi_vio_probe`,
> and during `scmi_virtio_probe`, it will do `scmi_probe` which will invoke
> function `scmi_protocol_acquire(handle, SCMI_PROTOCOL_BASE)`,
> and this function will send message via virtqueue and wait device's reply
> 
> [...]

Applied to sudeep.holla/linux (for-next/scmi/updates), thanks!

[1/1] firmware: arm_scmi: Set device ready before register scmi_virtio_driver
      https://git.kernel.org/sudeep.holla/c/e8faa8a466f6
--
Regards,
Sudeep