[PATCH v5 0/7] Fix some races between sysfb device registration and drivers probe

[PATCH v5 0/7] Fix some races between sysfb device registration and drivers probe

[Javier Martinez Canillas]

Hello,

The patches in this series contain mostly changes suggested by Daniel Vetter
Thomas Zimmermann. They aim to fix existing races between the Generic System
Framebuffer (sysfb) infrastructure and the fbdev and DRM device registration.

For example, it is currently possible for sysfb to register a platform
device after a real DRM driver was registered and requested to remove the
conflicting framebuffers. Or is possible for a simple{fb,drm} to match with
a device previously registered by sysfb, even after a real driver is present.

A symptom of this issue, was worked around with the commit fb561bf9abde
("fbdev: Prevent probing generic drivers if a FB is already registered")
but that's really a hack and should be reverted instead.

This series attempt to fix it more correctly and revert the mentioned hack.
That will also allow to make the num_registered_fb variable not visible to
drivers anymore, since that's internal to fbdev core.

Pach 1 is just a simple cleanup in preparation for later patches.

Patch 2 add a sysfb_disable() and sysfb_try_unregister() helpers to allow
disabling sysfb and attempt to unregister registered devices respectively.

Patch 3 changes how is dealt with conflicting framebuffers unregistering,
rather than having a variable to determine if a lock should be take, it
just drops the lock before unregistering the platform device.

Patch 4 changes the fbdev core to not attempt to unregister devices that
were registered by sysfb, let the same code doing the registration to also
handle the unregistration.

Patch 5 fixes the race that exists between sysfb devices registration and
fbdev framebuffer devices registration, by disabling the sysfb when a DRM
or fbdev driver requests to remove conflicting framebuffers.

Patch 6 is the revert patch that was posted by Daniel before but dropped
from his set and finally patch 7 is the one that makes num_registered_fb
private to fbmem.c, to not allow drivers to use it anymore.

The patches were tested on a rpi4 with the vc4, simpledrm and simplefb
drivers, using different combinations of built-in and as a module.

For example, having simpledrm as a module and loading it after the vc4
driver probed would not register a DRM device, which happens now without
the patches from this series.

Best regards,
Javier

Changes in v5:
- Move the sysfb_disable() call at conflicting framebuffers again to
  avoid the need of a DRIVER_FIRMWARE capability flag.
- Add Daniel Vetter's Reviewed-by tag again since reverted to the old
  patch that he already reviewed in v2.
- Drop patches that added a DRM_FIRMWARE capability and use them
  since the case those prevented could be ignored (Daniel Vetter).

Changes in v4:
- Make sysfb_disable() to also attempt to unregister a device.
- Drop call to sysfb_disable() in fbmem since is done in other places now.
- Add patch to make registered_fb[] private.
- Add patches that introduce the DRM_FIRMWARE capability and usage.

Changes in v3:
- Add Thomas Zimmermann's Reviewed-by tag to patch #1.
- Call sysfb_disable() when a DRM dev and a fbdev are registered rather
  than when conflicting framebuffers are removed (Thomas Zimmermann).
- Call sysfb_disable() when a fbdev framebuffer is registered rather
  than when conflicting framebuffers are removed (Thomas Zimmermann).
- Drop Daniel Vetter's Reviewed-by tag since patch changed a lot.
- Rebase on top of latest drm-misc-next branch.

Changes in v2:
- Rebase on top of latest drm-misc-next and fix conflicts (Daniel Vetter).
- Add kernel-doc comments and include in other_interfaces.rst (Daniel Vetter).
- Explain in the commit message that fbmem has to unregister the device
  as fallback if a driver registered the device itself (Daniel Vetter).
- Also explain that fallback in a comment in the code (Daniel Vetter).
- Don't encode in fbmem the assumption that sysfb will always register
  platform devices (Daniel Vetter).
- Add a FIXME comment about drivers registering devices (Daniel Vetter).
- Explain in the commit message that fbmem has to unregister the device
  as fallback if a driver registered the device itself (Daniel Vetter).
- Also explain that fallback in a comment in the code (Daniel Vetter).
- Don't encode in fbmem the assumption that sysfb will always register
  platform devices (Daniel Vetter).
- Add a FIXME comment about drivers registering devices (Daniel Vetter).
- Drop RFC prefix since patches were already reviewed by Daniel Vetter.
- Add Daniel Reviewed-by tags to the patches.

Daniel Vetter (2):
  Revert "fbdev: Prevent probing generic drivers if a FB is already
    registered"
  fbdev: Make registered_fb[] private to fbmem.c

Javier Martinez Canillas (5):
  firmware: sysfb: Make sysfb_create_simplefb() return a pdev pointer
  firmware: sysfb: Add helpers to unregister a pdev and disable
    registration
  fbdev: Restart conflicting fb removal loop when unregistering devices
  fbdev: Make sysfb to unregister its own registered devices
  fbdev: Disable sysfb device registration when removing conflicting FBs

 .../driver-api/firmware/other_interfaces.rst  |  6 ++
 drivers/firmware/sysfb.c                      | 91 +++++++++++++++++--
 drivers/firmware/sysfb_simplefb.c             | 16 ++--
 drivers/video/fbdev/core/fbmem.c              | 67 +++++++++++---
 drivers/video/fbdev/efifb.c                   | 11 ---
 drivers/video/fbdev/simplefb.c                | 11 ---
 include/linux/fb.h                            |  8 +-
 include/linux/sysfb.h                         | 29 +++++-
 8 files changed, 178 insertions(+), 61 deletions(-)

-- 
2.35.1

[PATCH v5 2/7] firmware: sysfb: Add helpers to unregister a pdev and disable registration

[Javier Martinez Canillas]

These can be used by subsystems to unregister a platform device registered
by sysfb and also to disable future platform device registration in sysfb.

Suggested-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Javier Martinez Canillas <javierm@redhat.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
---

(no changes since v4)

Changes in v4:
- Make sysfb_disable() to also attempt to unregister a device.

Changes in v2:
- Add kernel-doc comments and include in other_interfaces.rst (Daniel Vetter).

 .../driver-api/firmware/other_interfaces.rst  |  6 ++
 drivers/firmware/sysfb.c                      | 87 +++++++++++++++++--
 include/linux/sysfb.h                         | 19 ++++
 3 files changed, 106 insertions(+), 6 deletions(-)

diff --git a/Documentation/driver-api/firmware/other_interfaces.rst b/Documentation/driver-api/firmware/other_interfaces.rst
index b81794e0cfbb..06ac89adaafb 100644
--- a/Documentation/driver-api/firmware/other_interfaces.rst
+++ b/Documentation/driver-api/firmware/other_interfaces.rst
@@ -13,6 +13,12 @@ EDD Interfaces
 .. kernel-doc:: drivers/firmware/edd.c
    :internal:
 
+Generic System Framebuffers Interface
+-------------------------------------
+
+.. kernel-doc:: drivers/firmware/sysfb.c
+   :export:
+
 Intel Stratix10 SoC Service Layer
 ---------------------------------
 Some features of the Intel Stratix10 SoC require a level of privilege
diff --git a/drivers/firmware/sysfb.c b/drivers/firmware/sysfb.c
index b032f40a92de..6768968949e6 100644
--- a/drivers/firmware/sysfb.c
+++ b/drivers/firmware/sysfb.c
@@ -34,21 +34,92 @@
 #include <linux/screen_info.h>
 #include <linux/sysfb.h>
 
+static struct platform_device *pd;
+static DEFINE_MUTEX(disable_lock);
+static bool disabled;
+
+static bool sysfb_unregister(void)
+{
+	if (IS_ERR_OR_NULL(pd))
+		return false;
+
+	platform_device_unregister(pd);
+	pd = NULL;
+
+	return true;
+}
+
+/**
+ * sysfb_disable() - disable the Generic System Framebuffers support
+ *
+ * This disables the registration of system framebuffer devices that match the
+ * generic drivers that make use of the system framebuffer set up by firmware.
+ *
+ * It also unregisters a device if this was already registered by sysfb_init().
+ *
+ * Context: The function can sleep. A @disable_lock mutex is acquired to serialize
+ *          against sysfb_init(), that registers a system framebuffer device and
+ *          sysfb_try_unregister(), that tries to unregister a framebuffer device.
+ */
+void sysfb_disable(void)
+{
+	mutex_lock(&disable_lock);
+	sysfb_unregister();
+	disabled = true;
+	mutex_unlock(&disable_lock);
+}
+EXPORT_SYMBOL_GPL(sysfb_disable);
+
+/**
+ * sysfb_try_unregister() - attempt to unregister a system framebuffer device
+ * @dev: device to unregister
+ *
+ * This tries to unregister a system framebuffer device if this was registered
+ * by the Generic System Framebuffers. The device will only be unregistered if
+ * it was registered by sysfb_init(), otherwise it will not be unregistered.
+ *
+ * Context: The function can sleep. a @load_lock mutex is acquired to serialize
+ *          against sysfb_init(), that registers a simple framebuffer device and
+ *          sysfb_disable(), that disables the Generic System Framebuffers support.
+ *
+ * Return:
+ * * true          - the device was unregistered successfully
+ * * false         - the device was not unregistered
+ */
+bool sysfb_try_unregister(struct device *dev)
+{
+	bool ret = false;
+
+	mutex_lock(&disable_lock);
+	if (IS_ERR_OR_NULL(pd) || pd != to_platform_device(dev))
+		goto unlock_mutex;
+
+	ret = sysfb_unregister();
+
+unlock_mutex:
+	mutex_unlock(&disable_lock);
+	return ret;
+}
+EXPORT_SYMBOL_GPL(sysfb_try_unregister);
+
 static __init int sysfb_init(void)
 {
 	struct screen_info *si = &screen_info;
 	struct simplefb_platform_data mode;
-	struct platform_device *pd;
 	const char *name;
 	bool compatible;
-	int ret;
+	int ret = 0;
+
+	mutex_lock(&disable_lock);
+	if (disabled)
+		goto unlock_mutex;
 
 	/* try to create a simple-framebuffer device */
 	compatible = sysfb_parse_mode(si, &mode);
 	if (compatible) {
 		pd = sysfb_create_simplefb(si, &mode);
 		if (!IS_ERR(pd))
-			return 0;
+			goto unlock_mutex;
 	}
 
 	/* if the FB is incompatible, create a legacy framebuffer device */
@@ -60,8 +131,10 @@ static __init int sysfb_init(void)
 		name = "platform-framebuffer";
 
 	pd = platform_device_alloc(name, 0);
-	if (!pd)
-		return -ENOMEM;
+	if (!pd) {
+		ret = -ENOMEM;
+		goto unlock_mutex;
+	}
 
 	sysfb_apply_efi_quirks(pd);
 
@@ -73,9 +146,11 @@ static __init int sysfb_init(void)
 	if (ret)
 		goto err;
 
-	return 0;
+	goto unlock_mutex;
 err:
 	platform_device_put(pd);
+unlock_mutex:
+	mutex_unlock(&disable_lock);
 	return ret;
 }
 
diff --git a/include/linux/sysfb.h b/include/linux/sysfb.h
index 708152e9037b..e8c0313fac8f 100644
--- a/include/linux/sysfb.h
+++ b/include/linux/sysfb.h
@@ -55,6 +55,25 @@ struct efifb_dmi_info {
 	int flags;
 };
 
+#ifdef CONFIG_SYSFB
+
+void sysfb_disable(void);
+bool sysfb_try_unregister(struct device *dev);
+
+#else /* CONFIG_SYSFB */
+
+static inline void sysfb_disable(void)
+{
+
+}
+
+static inline bool sysfb_try_unregister(struct device *dev)
+{
+	return false;
+}
+
+#endif /* CONFIG_SYSFB */
+
 #ifdef CONFIG_EFI
 
 extern struct efifb_dmi_info efifb_dmi_list[];
-- 
2.35.1

Re: [PATCH v5 2/7] firmware: sysfb: Add helpers to unregister a pdev and disable registration

[Thomas Zimmermann]

[-- Attachment #1.1: Type: text/plain, Size: 6439 bytes --]

Hi

Am 11.05.22 um 13:24 schrieb Javier Martinez Canillas:
> These can be used by subsystems to unregister a platform device registered
> by sysfb and also to disable future platform device registration in sysfb.
>

I find it very hard to review these patches, as related things are put 
into separate patches.

I suggest to put the sysfb_disable() stuff into patch 5 and the rest 
into patch 4.

Best regards
Thomas


> Suggested-by: Daniel Vetter <daniel.vetter@ffwll.ch>
> Signed-off-by: Javier Martinez Canillas <javierm@redhat.com>
> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
> ---
> 
> (no changes since v4)
> 
> Changes in v4:
> - Make sysfb_disable() to also attempt to unregister a device.
> 
> Changes in v2:
> - Add kernel-doc comments and include in other_interfaces.rst (Daniel Vetter).
> 
>   .../driver-api/firmware/other_interfaces.rst  |  6 ++
>   drivers/firmware/sysfb.c                      | 87 +++++++++++++++++--
>   include/linux/sysfb.h                         | 19 ++++
>   3 files changed, 106 insertions(+), 6 deletions(-)
> 
> diff --git a/Documentation/driver-api/firmware/other_interfaces.rst b/Documentation/driver-api/firmware/other_interfaces.rst
> index b81794e0cfbb..06ac89adaafb 100644
> --- a/Documentation/driver-api/firmware/other_interfaces.rst
> +++ b/Documentation/driver-api/firmware/other_interfaces.rst
> @@ -13,6 +13,12 @@ EDD Interfaces
>   .. kernel-doc:: drivers/firmware/edd.c
>      :internal:
>   
> +Generic System Framebuffers Interface
> +-------------------------------------
> +
> +.. kernel-doc:: drivers/firmware/sysfb.c
> +   :export:
> +
>   Intel Stratix10 SoC Service Layer
>   ---------------------------------
>   Some features of the Intel Stratix10 SoC require a level of privilege
> diff --git a/drivers/firmware/sysfb.c b/drivers/firmware/sysfb.c
> index b032f40a92de..6768968949e6 100644
> --- a/drivers/firmware/sysfb.c
> +++ b/drivers/firmware/sysfb.c
> @@ -34,21 +34,92 @@
>   #include <linux/screen_info.h>
>   #include <linux/sysfb.h>
>   
> +static struct platform_device *pd;
> +static DEFINE_MUTEX(disable_lock);
> +static bool disabled;
> +
> +static bool sysfb_unregister(void)
> +{
> +	if (IS_ERR_OR_NULL(pd))
> +		return false;
> +
> +	platform_device_unregister(pd);
> +	pd = NULL;
> +
> +	return true;
> +}
> +
> +/**
> + * sysfb_disable() - disable the Generic System Framebuffers support
> + *
> + * This disables the registration of system framebuffer devices that match the
> + * generic drivers that make use of the system framebuffer set up by firmware.
> + *
> + * It also unregisters a device if this was already registered by sysfb_init().
> + *
> + * Context: The function can sleep. A @disable_lock mutex is acquired to serialize
> + *          against sysfb_init(), that registers a system framebuffer device and
> + *          sysfb_try_unregister(), that tries to unregister a framebuffer device.
> + */
> +void sysfb_disable(void)
> +{
> +	mutex_lock(&disable_lock);
> +	sysfb_unregister();
> +	disabled = true;
> +	mutex_unlock(&disable_lock);
> +}
> +EXPORT_SYMBOL_GPL(sysfb_disable);
> +
> +/**
> + * sysfb_try_unregister() - attempt to unregister a system framebuffer device
> + * @dev: device to unregister
> + *
> + * This tries to unregister a system framebuffer device if this was registered
> + * by the Generic System Framebuffers. The device will only be unregistered if
> + * it was registered by sysfb_init(), otherwise it will not be unregistered.
> + *
> + * Context: The function can sleep. a @load_lock mutex is acquired to serialize
> + *          against sysfb_init(), that registers a simple framebuffer device and
> + *          sysfb_disable(), that disables the Generic System Framebuffers support.
> + *
> + * Return:
> + * * true          - the device was unregistered successfully
> + * * false         - the device was not unregistered
> + */
> +bool sysfb_try_unregister(struct device *dev)
> +{
> +	bool ret = false;
> +
> +	mutex_lock(&disable_lock);
> +	if (IS_ERR_OR_NULL(pd) || pd != to_platform_device(dev))
> +		goto unlock_mutex;
> +
> +	ret = sysfb_unregister();
> +
> +unlock_mutex:
> +	mutex_unlock(&disable_lock);
> +	return ret;
> +}
> +EXPORT_SYMBOL_GPL(sysfb_try_unregister);
> +
>   static __init int sysfb_init(void)
>   {
>   	struct screen_info *si = &screen_info;
>   	struct simplefb_platform_data mode;
> -	struct platform_device *pd;
>   	const char *name;
>   	bool compatible;
> -	int ret;
> +	int ret = 0;
> +
> +	mutex_lock(&disable_lock);
> +	if (disabled)
> +		goto unlock_mutex;
>   
>   	/* try to create a simple-framebuffer device */
>   	compatible = sysfb_parse_mode(si, &mode);
>   	if (compatible) {
>   		pd = sysfb_create_simplefb(si, &mode);
>   		if (!IS_ERR(pd))
> -			return 0;
> +			goto unlock_mutex;
>   	}
>   
>   	/* if the FB is incompatible, create a legacy framebuffer device */
> @@ -60,8 +131,10 @@ static __init int sysfb_init(void)
>   		name = "platform-framebuffer";
>   
>   	pd = platform_device_alloc(name, 0);
> -	if (!pd)
> -		return -ENOMEM;
> +	if (!pd) {
> +		ret = -ENOMEM;
> +		goto unlock_mutex;
> +	}
>   
>   	sysfb_apply_efi_quirks(pd);
>   
> @@ -73,9 +146,11 @@ static __init int sysfb_init(void)
>   	if (ret)
>   		goto err;
>   
> -	return 0;
> +	goto unlock_mutex;
>   err:
>   	platform_device_put(pd);
> +unlock_mutex:
> +	mutex_unlock(&disable_lock);
>   	return ret;
>   }
>   
> diff --git a/include/linux/sysfb.h b/include/linux/sysfb.h
> index 708152e9037b..e8c0313fac8f 100644
> --- a/include/linux/sysfb.h
> +++ b/include/linux/sysfb.h
> @@ -55,6 +55,25 @@ struct efifb_dmi_info {
>   	int flags;
>   };
>   
> +#ifdef CONFIG_SYSFB
> +
> +void sysfb_disable(void);
> +bool sysfb_try_unregister(struct device *dev);
> +
> +#else /* CONFIG_SYSFB */
> +
> +static inline void sysfb_disable(void)
> +{
> +
> +}
> +
> +static inline bool sysfb_try_unregister(struct device *dev)
> +{
> +	return false;
> +}
> +
> +#endif /* CONFIG_SYSFB */
> +
>   #ifdef CONFIG_EFI
>   
>   extern struct efifb_dmi_info efifb_dmi_list[];

-- 
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 840 bytes --]

Re: [PATCH v5 2/7] firmware: sysfb: Add helpers to unregister a pdev and disable registration

[Javier Martinez Canillas]

Hello Thomas,

On 5/11/22 13:54, Thomas Zimmermann wrote:
> Hi
> 
> Am 11.05.22 um 13:24 schrieb Javier Martinez Canillas:
>> These can be used by subsystems to unregister a platform device registered
>> by sysfb and also to disable future platform device registration in sysfb.
>>
> 
> I find it very hard to review these patches, as related things are put 
> into separate patches.
> 
> I suggest to put the sysfb_disable() stuff into patch 5 and the rest 
> into patch 4.
>

Ok, you then want in the same patch to have both the helper definition
and first usage.

Other subsystems ask you to do the opposite, to split the definition and
usage in separate patches. But I'm fine with merging those if you prefer.
 
> Best regards
> Thomas
> 
> 
-- 
Best regards,

Javier Martinez Canillas
Linux Engineering
Red Hat

Re: [PATCH v5 2/7] firmware: sysfb: Add helpers to unregister a pdev and disable registration

[Thomas Zimmermann]

[-- Attachment #1.1: Type: text/plain, Size: 6615 bytes --]



Am 11.05.22 um 13:24 schrieb Javier Martinez Canillas:
> These can be used by subsystems to unregister a platform device registered
> by sysfb and also to disable future platform device registration in sysfb.
> 
> Suggested-by: Daniel Vetter <daniel.vetter@ffwll.ch>
> Signed-off-by: Javier Martinez Canillas <javierm@redhat.com>
> Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
> ---
> 
> (no changes since v4)
> 
> Changes in v4:
> - Make sysfb_disable() to also attempt to unregister a device.
> 
> Changes in v2:
> - Add kernel-doc comments and include in other_interfaces.rst (Daniel Vetter).
> 
>   .../driver-api/firmware/other_interfaces.rst  |  6 ++
>   drivers/firmware/sysfb.c                      | 87 +++++++++++++++++--
>   include/linux/sysfb.h                         | 19 ++++
>   3 files changed, 106 insertions(+), 6 deletions(-)
> 
> diff --git a/Documentation/driver-api/firmware/other_interfaces.rst b/Documentation/driver-api/firmware/other_interfaces.rst
> index b81794e0cfbb..06ac89adaafb 100644
> --- a/Documentation/driver-api/firmware/other_interfaces.rst
> +++ b/Documentation/driver-api/firmware/other_interfaces.rst
> @@ -13,6 +13,12 @@ EDD Interfaces
>   .. kernel-doc:: drivers/firmware/edd.c
>      :internal:
>   
> +Generic System Framebuffers Interface
> +-------------------------------------
> +
> +.. kernel-doc:: drivers/firmware/sysfb.c
> +   :export:
> +
>   Intel Stratix10 SoC Service Layer
>   ---------------------------------
>   Some features of the Intel Stratix10 SoC require a level of privilege
> diff --git a/drivers/firmware/sysfb.c b/drivers/firmware/sysfb.c
> index b032f40a92de..6768968949e6 100644
> --- a/drivers/firmware/sysfb.c
> +++ b/drivers/firmware/sysfb.c
> @@ -34,21 +34,92 @@
>   #include <linux/screen_info.h>
>   #include <linux/sysfb.h>
>   
> +static struct platform_device *pd;
> +static DEFINE_MUTEX(disable_lock);
> +static bool disabled;
> +
> +static bool sysfb_unregister(void)
> +{
> +	if (IS_ERR_OR_NULL(pd))
> +		return false;
> +
> +	platform_device_unregister(pd);
> +	pd = NULL;
> +
> +	return true;
> +}
> +
> +/**
> + * sysfb_disable() - disable the Generic System Framebuffers support
> + *
> + * This disables the registration of system framebuffer devices that match the
> + * generic drivers that make use of the system framebuffer set up by firmware.
> + *
> + * It also unregisters a device if this was already registered by sysfb_init().

Why? I still cannot wrap my mind around, why we need to store *pd at all 
and use sysfb_try_unregister() for unregistering.

> + *
> + * Context: The function can sleep. A @disable_lock mutex is acquired to serialize
> + *          against sysfb_init(), that registers a system framebuffer device and
> + *          sysfb_try_unregister(), that tries to unregister a framebuffer device.
> + */
> +void sysfb_disable(void)
> +{
> +	mutex_lock(&disable_lock);
> +	sysfb_unregister();
> +	disabled = true;
> +	mutex_unlock(&disable_lock);
> +}
> +EXPORT_SYMBOL_GPL(sysfb_disable);
> +
> +/**
> + * sysfb_try_unregister() - attempt to unregister a system framebuffer device
> + * @dev: device to unregister
> + *
> + * This tries to unregister a system framebuffer device if this was registered
> + * by the Generic System Framebuffers. The device will only be unregistered if
> + * it was registered by sysfb_init(), otherwise it will not be unregistered.
> + *
> + * Context: The function can sleep. a @load_lock mutex is acquired to serialize
> + *          against sysfb_init(), that registers a simple framebuffer device and
> + *          sysfb_disable(), that disables the Generic System Framebuffers support.
> + *
> + * Return:
> + * * true          - the device was unregistered successfully
> + * * false         - the device was not unregistered
> + */
> +bool sysfb_try_unregister(struct device *dev)

As it stands, I strongly object the use of this function as still don't 
really get the purpose. It looks like a glorified wrapper around 
platform_device_unregister(). Do we need disable_lock to serialize with 
something else?

Best regards
Thomas


> +{
> +	bool ret = false;
> +
> +	mutex_lock(&disable_lock);
> +	if (IS_ERR_OR_NULL(pd) || pd != to_platform_device(dev))
> +		goto unlock_mutex;
> +
> +	ret = sysfb_unregister();
> +
> +unlock_mutex:
> +	mutex_unlock(&disable_lock);
> +	return ret;
> +}
> +EXPORT_SYMBOL_GPL(sysfb_try_unregister);
> +
>   static __init int sysfb_init(void)
>   {
>   	struct screen_info *si = &screen_info;
>   	struct simplefb_platform_data mode;
> -	struct platform_device *pd;
>   	const char *name;
>   	bool compatible;
> -	int ret;
> +	int ret = 0;
> +
> +	mutex_lock(&disable_lock);
> +	if (disabled)
> +		goto unlock_mutex;
>   
>   	/* try to create a simple-framebuffer device */
>   	compatible = sysfb_parse_mode(si, &mode);
>   	if (compatible) {
>   		pd = sysfb_create_simplefb(si, &mode);
>   		if (!IS_ERR(pd))
> -			return 0;
> +			goto unlock_mutex;
>   	}
>   
>   	/* if the FB is incompatible, create a legacy framebuffer device */
> @@ -60,8 +131,10 @@ static __init int sysfb_init(void)
>   		name = "platform-framebuffer";
>   
>   	pd = platform_device_alloc(name, 0);
> -	if (!pd)
> -		return -ENOMEM;
> +	if (!pd) {
> +		ret = -ENOMEM;
> +		goto unlock_mutex;
> +	}
>   
>   	sysfb_apply_efi_quirks(pd);
>   
> @@ -73,9 +146,11 @@ static __init int sysfb_init(void)
>   	if (ret)
>   		goto err;
>   
> -	return 0;
> +	goto unlock_mutex;
>   err:
>   	platform_device_put(pd);
> +unlock_mutex:
> +	mutex_unlock(&disable_lock);
>   	return ret;
>   }
>   
> diff --git a/include/linux/sysfb.h b/include/linux/sysfb.h
> index 708152e9037b..e8c0313fac8f 100644
> --- a/include/linux/sysfb.h
> +++ b/include/linux/sysfb.h
> @@ -55,6 +55,25 @@ struct efifb_dmi_info {
>   	int flags;
>   };
>   
> +#ifdef CONFIG_SYSFB
> +
> +void sysfb_disable(void);
> +bool sysfb_try_unregister(struct device *dev);
> +
> +#else /* CONFIG_SYSFB */
> +
> +static inline void sysfb_disable(void)
> +{
> +
> +}
> +
> +static inline bool sysfb_try_unregister(struct device *dev)
> +{
> +	return false;
> +}
> +
> +#endif /* CONFIG_SYSFB */
> +
>   #ifdef CONFIG_EFI
>   
>   extern struct efifb_dmi_info efifb_dmi_list[];

-- 
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 840 bytes --]

Re: [PATCH v5 2/7] firmware: sysfb: Add helpers to unregister a pdev and disable registration

[Thomas Zimmermann]

[-- Attachment #1.1: Type: text/plain, Size: 1396 bytes --]

Hi

Am 11.05.22 um 14:01 schrieb Javier Martinez Canillas:
> Hello Thomas,
> 
> On 5/11/22 13:54, Thomas Zimmermann wrote:
>> Hi
>>
>> Am 11.05.22 um 13:24 schrieb Javier Martinez Canillas:
>>> These can be used by subsystems to unregister a platform device registered
>>> by sysfb and also to disable future platform device registration in sysfb.
>>>
>>
>> I find it very hard to review these patches, as related things are put
>> into separate patches.
>>
>> I suggest to put the sysfb_disable() stuff into patch 5 and the rest
>> into patch 4.
>>
> 
> Ok, you then want in the same patch to have both the helper definition
> and first usage.
> 
> Other subsystems ask you to do the opposite, to split the definition and
> usage in separate patches. But I'm fine with merging those if you prefer.

Usually, I have no strong opinion on this. But in the case of this 
specific patchset, I have the feeling that I'm missing some important 
point because call and implementation are separate.  See my other 
replies for that.  Putting them next to each other will hopefully help. 
Sorry for the inconvenience.

Best regards
Thomas

>   
>> Best regards
>> Thomas
>>
>>

-- 
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 840 bytes --]

Re: [PATCH v5 2/7] firmware: sysfb: Add helpers to unregister a pdev and disable registration

[Javier Martinez Canillas]

Hello Thomas,

On 5/11/22 14:02, Thomas Zimmermann wrote:

[snip]

>> +
>> +/**
>> + * sysfb_disable() - disable the Generic System Framebuffers support
>> + *
>> + * This disables the registration of system framebuffer devices that match the
>> + * generic drivers that make use of the system framebuffer set up by firmware.
>> + *
>> + * It also unregisters a device if this was already registered by sysfb_init().
> 
> Why? I still cannot wrap my mind around, why we need to store *pd at all 
> and use sysfb_try_unregister() for unregistering.
>

Because on sysfb_disable(), the registered platform device has to unregistered.

And sysfb has no way to know if it was unregistered already or not unless that
stage is maintained in sysfb itself.

Let's have some examples assuming that we don't have this helper in sysfb
(will use the vc4 DRM driver just to avoid typing "a real DRM driver).

a) simplefb probed and then vc4

   1) "simple-framebuffer" pdev is registered by sysfb
   2) simplefb is registered and matches "simple-framebuffer"
   3) a vc4 device is registered by OF when parsing the DTB
   4) vc4 driver is registered, matches vc4 and probes
   5) vc4 requests the conflicting framebuffers to be removed
      and fbmem unregisters "simple-framebuffer"
   6) fbmem calls sysfb_disable()
   7) sysfb_disable() should unregister the pdev but can't
      because has no way to know that fbmem already did that.
 
b) vc4 probed and then simplefb.ko module is loaded

   1) "simple-framebuffer" pdev is registered by sysfb
   2) a vc4 device is registered by OF when parsing the DTB
   3) vc4 driver is registered, matches vc4 and probes
   4) vc4 requests the conflicting framebuffers to be removed
      and fbmem unregisters "simple-framebuffer"
   5) fbmem calls sysfb_disable()
   6) sysfb_disable() should unregister the pdev but can't
      because has no way to know that fbmem already did that.
   7) simplefb.ko is loaded and simplefb driver registered
   8) simplefb matches the registered "simple-framebuffer"
      and will wrongly probe and register a DRM device.

In option (a), making sysfb_disable() to attempt to unregister the device
that register in sysfb_init() will lead to a use-after-free if this was
already unregistered by fbmem in remove_conflicting_framebuffers(), so
it can't attempt to do that.

Same for option (b), but sysfb_disable() can't rely on fbmem to do the
unregistration because it only does for devices that are associated with
an already registered fbdev.

[snip]

>> + * Return:
>> + * * true          - the device was unregistered successfully
>> + * * false         - the device was not unregistered
>> + */
>> +bool sysfb_try_unregister(struct device *dev)
> 
> As it stands, I strongly object the use of this function as still don't 

No worries, it's my bad since I clearly failed to explain the rationale in
the commit message and comments.

> really get the purpose. It looks like a glorified wrapper around 
> platform_device_unregister(). Do we need disable_lock to serialize with 
> something else?
>

Yes, it has to serialize with sysfb_init() and sysfb_disable().
 
> Best regards
> Thomas
> 
> 
-- 
Best regards,

Javier Martinez Canillas
Linux Engineering
Red Hat

Re: [PATCH v5 2/7] firmware: sysfb: Add helpers to unregister a pdev and disable registration

[Javier Martinez Canillas]

Hello Thomas,

On 5/11/22 14:05, Thomas Zimmermann wrote:

[snip]

>>
>> Other subsystems ask you to do the opposite, to split the definition and
>> usage in separate patches. But I'm fine with merging those if you prefer.
> 
> Usually, I have no strong opinion on this. But in the case of this 
> specific patchset, I have the feeling that I'm missing some important 
> point because call and implementation are separate.  See my other 
> replies for that.  Putting them next to each other will hopefully help. 
> Sorry for the inconvenience.
>

No worries at all. Happy to do that change if the patches are easy to
understand. It took me some time as well to wrap my head around all
the race conditions and needed locking.

Same for patch 3/7, but I'm convinced that dropping the lock is the
correct thing to do than calling to drivers' .remove callbacks with
a lock held.

-- 
Best regards,

Javier Martinez Canillas
Linux Engineering
Red Hat

Re: [PATCH v5 0/7] Fix some races between sysfb device registration and drivers probe

[Thomas Zimmermann]

[-- Attachment #1.1: Type: text/plain, Size: 7952 bytes --]

Hi Javier,

thanks again for providing the examples. I think I now better get what 
you're trying to solve.

First of all let's merge patch 3, as it seems unrelated.

For the other patches, I'd like to take a step back and try to solve the 
broader problem. IIRC we talked about this briefly already.

 From my understanding, the problem of the current code is that removal 
of the firmware device is build around drivers (either DRM or fbdev). 
Everything works fine if a driver is bound to the firmware device and 
the native driver can remove it.  In other case, we have to manually 
disable sysfb and force-remove unbound firmware devices.  And the 
patchset doesn't even cover firmware devices that come from DT nodes.

But what we really want is to track resource ownership based on devices. 
When we add a firmware device (via sysfb or DT), we immediately add it 
to a list of firmware devices. When the native driver arrives, it can 
remove the firmware device even if no driver has been bound.

We can also operate in the other way if the native drivers implicitly 
reserves the framebuffer range. If sysfb would try to register a 
firmware device in that range, he can let it fail. No more need to fully 
disable sysfb on the first arriving native device.

We already track the memory ranges in drm aperture helpers. We'd need to 
move the code to a more prominent location (e.g., <linux/aperture.h>) 
and change fbdev to use it. Sysfb and DT code needs to insert platform 
devices upon creation. We can then implement the more fancy stuff, such 
as tracking native-device memory.  (And if we ever get to fix all usage 
of Linux' request-mem-region, we can even move all the functionality there).

Best regards
Thomas


Am 11.05.22 um 13:24 schrieb Javier Martinez Canillas:
> Hello,
> 
> The patches in this series contain mostly changes suggested by Daniel Vetter
> Thomas Zimmermann. They aim to fix existing races between the Generic System
> Framebuffer (sysfb) infrastructure and the fbdev and DRM device registration.
> 
> For example, it is currently possible for sysfb to register a platform
> device after a real DRM driver was registered and requested to remove the
> conflicting framebuffers. Or is possible for a simple{fb,drm} to match with
> a device previously registered by sysfb, even after a real driver is present.
> 
> A symptom of this issue, was worked around with the commit fb561bf9abde
> ("fbdev: Prevent probing generic drivers if a FB is already registered")
> but that's really a hack and should be reverted instead.
> 
> This series attempt to fix it more correctly and revert the mentioned hack.
> That will also allow to make the num_registered_fb variable not visible to
> drivers anymore, since that's internal to fbdev core.
> 
> Pach 1 is just a simple cleanup in preparation for later patches.
> 
> Patch 2 add a sysfb_disable() and sysfb_try_unregister() helpers to allow
> disabling sysfb and attempt to unregister registered devices respectively.
> 
> Patch 3 changes how is dealt with conflicting framebuffers unregistering,
> rather than having a variable to determine if a lock should be take, it
> just drops the lock before unregistering the platform device.
> 
> Patch 4 changes the fbdev core to not attempt to unregister devices that
> were registered by sysfb, let the same code doing the registration to also
> handle the unregistration.
> 
> Patch 5 fixes the race that exists between sysfb devices registration and
> fbdev framebuffer devices registration, by disabling the sysfb when a DRM
> or fbdev driver requests to remove conflicting framebuffers.
> 
> Patch 6 is the revert patch that was posted by Daniel before but dropped
> from his set and finally patch 7 is the one that makes num_registered_fb
> private to fbmem.c, to not allow drivers to use it anymore.
> 
> The patches were tested on a rpi4 with the vc4, simpledrm and simplefb
> drivers, using different combinations of built-in and as a module.
> 
> For example, having simpledrm as a module and loading it after the vc4
> driver probed would not register a DRM device, which happens now without
> the patches from this series.
> 
> Best regards,
> Javier
> 
> Changes in v5:
> - Move the sysfb_disable() call at conflicting framebuffers again to
>    avoid the need of a DRIVER_FIRMWARE capability flag.
> - Add Daniel Vetter's Reviewed-by tag again since reverted to the old
>    patch that he already reviewed in v2.
> - Drop patches that added a DRM_FIRMWARE capability and use them
>    since the case those prevented could be ignored (Daniel Vetter).
> 
> Changes in v4:
> - Make sysfb_disable() to also attempt to unregister a device.
> - Drop call to sysfb_disable() in fbmem since is done in other places now.
> - Add patch to make registered_fb[] private.
> - Add patches that introduce the DRM_FIRMWARE capability and usage.
> 
> Changes in v3:
> - Add Thomas Zimmermann's Reviewed-by tag to patch #1.
> - Call sysfb_disable() when a DRM dev and a fbdev are registered rather
>    than when conflicting framebuffers are removed (Thomas Zimmermann).
> - Call sysfb_disable() when a fbdev framebuffer is registered rather
>    than when conflicting framebuffers are removed (Thomas Zimmermann).
> - Drop Daniel Vetter's Reviewed-by tag since patch changed a lot.
> - Rebase on top of latest drm-misc-next branch.
> 
> Changes in v2:
> - Rebase on top of latest drm-misc-next and fix conflicts (Daniel Vetter).
> - Add kernel-doc comments and include in other_interfaces.rst (Daniel Vetter).
> - Explain in the commit message that fbmem has to unregister the device
>    as fallback if a driver registered the device itself (Daniel Vetter).
> - Also explain that fallback in a comment in the code (Daniel Vetter).
> - Don't encode in fbmem the assumption that sysfb will always register
>    platform devices (Daniel Vetter).
> - Add a FIXME comment about drivers registering devices (Daniel Vetter).
> - Explain in the commit message that fbmem has to unregister the device
>    as fallback if a driver registered the device itself (Daniel Vetter).
> - Also explain that fallback in a comment in the code (Daniel Vetter).
> - Don't encode in fbmem the assumption that sysfb will always register
>    platform devices (Daniel Vetter).
> - Add a FIXME comment about drivers registering devices (Daniel Vetter).
> - Drop RFC prefix since patches were already reviewed by Daniel Vetter.
> - Add Daniel Reviewed-by tags to the patches.
> 
> Daniel Vetter (2):
>    Revert "fbdev: Prevent probing generic drivers if a FB is already
>      registered"
>    fbdev: Make registered_fb[] private to fbmem.c
> 
> Javier Martinez Canillas (5):
>    firmware: sysfb: Make sysfb_create_simplefb() return a pdev pointer
>    firmware: sysfb: Add helpers to unregister a pdev and disable
>      registration
>    fbdev: Restart conflicting fb removal loop when unregistering devices
>    fbdev: Make sysfb to unregister its own registered devices
>    fbdev: Disable sysfb device registration when removing conflicting FBs
> 
>   .../driver-api/firmware/other_interfaces.rst  |  6 ++
>   drivers/firmware/sysfb.c                      | 91 +++++++++++++++++--
>   drivers/firmware/sysfb_simplefb.c             | 16 ++--
>   drivers/video/fbdev/core/fbmem.c              | 67 +++++++++++---
>   drivers/video/fbdev/efifb.c                   | 11 ---
>   drivers/video/fbdev/simplefb.c                | 11 ---
>   include/linux/fb.h                            |  8 +-
>   include/linux/sysfb.h                         | 29 +++++-
>   8 files changed, 178 insertions(+), 61 deletions(-)
> 

-- 
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 840 bytes --]

Re: [PATCH v5 0/7] Fix some races between sysfb device registration and drivers probe

[Javier Martinez Canillas]

Hello Thomas,

Thanks for your feedback and comments.

On 5/13/22 12:48, Thomas Zimmermann wrote:
> Hi Javier,
> 
> thanks again for providing the examples. I think I now better get what 
> you're trying to solve.
>

You are welcome.
 
> First of all let's merge patch 3, as it seems unrelated.
>

Agreed. I can push it to drm-misc-next.
 
> For the other patches, I'd like to take a step back and try to solve the 
> broader problem. IIRC we talked about this briefly already.
> 

Yes, that's what we discussed on IRC some time ago.

>  From my understanding, the problem of the current code is that removal 
> of the firmware device is build around drivers (either DRM or fbdev). 
> Everything works fine if a driver is bound to the firmware device and 
> the native driver can remove it.  In other case, we have to manually 

And this is the common case, since most kernels will have some driver
that binds to a platform device registered to provide the firmware FB.

> disable sysfb and force-remove unbound firmware devices.  And the 
> patchset doesn't even cover firmware devices that come from DT nodes.
>

Indeed.
 
> But what we really want is to track resource ownership based on devices. 
> When we add a firmware device (via sysfb or DT), we immediately add it 
> to a list of firmware devices. When the native driver arrives, it can 
> remove the firmware device even if no driver has been bound.
> 
> We can also operate in the other way if the native drivers implicitly 
> reserves the framebuffer range. If sysfb would try to register a 
> firmware device in that range, he can let it fail. No more need to fully 
> disable sysfb on the first arriving native device.
> 
> We already track the memory ranges in drm aperture helpers. We'd need to 
> move the code to a more prominent location (e.g., <linux/aperture.h>) 
> and change fbdev to use it. Sysfb and DT code needs to insert platform 
> devices upon creation. We can then implement the more fancy stuff, such 
> as tracking native-device memory.  (And if we ever get to fix all usage 
> of Linux' request-mem-region, we can even move all the functionality there).
>

Agreed. And as mentioned, the race that these patches attempt to fix are for
the less common case when a native driver probes but either no generic driver
registered a framebuffer yet or the platform device wasn't registered yet.

But this can only happen if for example a native driver is built-in but the
generic driver is build as a module, which is not the common configuration.

What most distros do is the opposite, to have {simple,of,efi,vesa}fb or
simpledrm built-in and the native drivers built as modules.

So there's no rush to fix this by piling more hacks on top of the ones we
already have and instead try to fix it more properly as you suggested.
 
-- 
Best regards,

Javier Martinez Canillas
Linux Engineering
Red Hat

Re: [PATCH v5 0/7] Fix some races between sysfb device registration and drivers probe

[Thomas Zimmermann]

[-- Attachment #1.1: Type: text/plain, Size: 2235 bytes --]

Hi Javier

Am 13.05.22 um 13:10 schrieb Javier Martinez Canillas:
...
>> We already track the memory ranges in drm aperture helpers. We'd need to
>> move the code to a more prominent location (e.g., <linux/aperture.h>)
>> and change fbdev to use it. Sysfb and DT code needs to insert platform
>> devices upon creation. We can then implement the more fancy stuff, such
>> as tracking native-device memory.  (And if we ever get to fix all usage
>> of Linux' request-mem-region, we can even move all the functionality there).
>>
> 
> Agreed. And as mentioned, the race that these patches attempt to fix are for
> the less common case when a native driver probes but either no generic driver
> registered a framebuffer yet or the platform device wasn't registered yet.
> 
> But this can only happen if for example a native driver is built-in but the
> generic driver is build as a module, which is not the common configuration.
> 
> What most distros do is the opposite, to have {simple,of,efi,vesa}fb or
> simpledrm built-in and the native drivers built as modules.
> 
> So there's no rush to fix this by piling more hacks on top of the ones we
> already have and instead try to fix it more properly as you suggested.
>   

A first step would be to use DRM's aperture helpers in fbdev. That would 
be a good idea anyway, as it would simplify both. You already mentioned 
some API changes to make aperture helpers DRM-independent.

The affected fbdev drivers use platform devices, so this should be easy.

Aperture helpers have something like the registration_lock. [1] I don't 
know if we need to recreate patch 3 for this as well.

If we absolutely need some special detachment handling for fbdev, we can 
make devm_aperture_aquire() a public interface. The detach helper is 
provided by the caller.

Best regards
Thomas


[1] 
https://elixir.bootlin.com/linux/v5.17.6/source/drivers/gpu/drm/drm_aperture.c#L254
[2] 
https://elixir.bootlin.com/linux/v5.17.6/source/drivers/gpu/drm/drm_aperture.c#L159

-- 
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 840 bytes --]