platform/kernel/linux-rpi.git
2 years agodrm/vc4: v3d: Rework the runtime_pm setup
Maxime Ripard [Fri, 10 Jun 2022 08:00:50 +0000 (10:00 +0200)]
drm/vc4: v3d: Rework the runtime_pm setup

At bind time, vc4_v3d_bind() will read a register to retrieve the v3d
version and make sure it's a version we're compatible with.

However, the v3d has an optional clock that is enabled only after the
register read-out and a power domain that wasn't enabled at all in the bind
implementation. This was working fine at boot because both were enabled,
but resulted in the version check failing if we were unbinding and
rebinding the driver because the unbinding would have turned them off.

The fix isn't as easy as calling pm_runtime_resume_and_get() prior to the
register access to power up the power domain though.

Indeed, the runtime_resume implementation will enable the clock mentioned
above, call vc4_v3d_init_hw() and then vc4_irq_enable().

Prior to the previous patch, vc4_irq_enable() needed to occur after our
call to platform_get_irq() and vc4_irq_install(), since vc4_irq_enable()
used to call enable_irq() and vc4_irq_install() will call request_irq().

vc4_irq_install() will also do some register access, so needs the power
domain to be on. So we ended up in a situation where
vc4_v3d_runtime_resume() needed vc4_irq_install() to have been called
before, and vc4_irq_install() needed vc4_v3d_runtime_resume().

The previous patch removed the enable_irq() call in vc4_irq_enable() and
thus removed the dependency of vc4_v3d_runtime_resume() on
vc4_irq_install().

Thus, we can now rework our bind implementation to call
pm_runtime_resume_and_get() before our register access to make sure the
power domain is on. vc4_v3d_runtime_resume() also takes care of turning the
clock on and calling vc4_v3d_init_hw() so we can remove them from bind.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: v3d: Stop disabling interrupts
Maxime Ripard [Thu, 9 Jun 2022 15:53:30 +0000 (17:53 +0200)]
drm/vc4: v3d: Stop disabling interrupts

The vc4_irq_disable(), among other things, will call disable_irq() to
complete any in-flight interrupts.

This requires its counterpart, vc4_irq_enable(), to call enable_irq() which
causes issues addressed in a later patch.

However, vc4_irq_disable() is called by two callees: vc4_irq_uninstall()
and vc4_v3d_runtime_suspend().

vc4_irq_uninstall() also calls free_irq() which already disables the
interrupt line. We thus don't require an explicit disable_irq() for that
call site.

vc4_v3d_runtime_suspend() doesn't have any other code. However, the rest of
vc4_irq_disable() masks the interrupts coming from the v3d, so explictly
disabling the interrupt line is also redundant.

The only thing we really care about is thus to make sure we don't have any
handler in-flight, as suggested by the comment. We can thus replace
disable_irq() by synchronize_irq().

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: perfmon: Add missing mutex_destroy
Maxime Ripard [Mon, 6 Jun 2022 14:41:53 +0000 (16:41 +0200)]
drm/vc4: perfmon: Add missing mutex_destroy

vc4_perfmon_open_file() will instantiate a mutex for that file instance,
but we never call mutex_destroy () in vc4_perfmon_close_file().

Let's add that missing call.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: Switch to drmm_mutex_init
Maxime Ripard [Mon, 6 Jun 2022 14:40:47 +0000 (16:40 +0200)]
drm/vc4: Switch to drmm_mutex_init

mutex_init is supposed to be balanced by a call to mutex_destroy that we
were never doing in the vc4 driver.

Since a DRM-managed mutex_init variant has been introduced, let's just
switch to it.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: hvs: Skip DebugFS Registration for FKMS
Maxime Ripard [Mon, 11 Jul 2022 08:38:25 +0000 (10:38 +0200)]
drm/vc4: hvs: Skip DebugFS Registration for FKMS

FKMS doesn't have an HVS and it's expected. Return from the debugfs init
function immediately if we're running with fkms.

Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: debugfs: Simplify debugfs registration
Maxime Ripard [Mon, 6 Jun 2022 14:37:53 +0000 (16:37 +0200)]
drm/vc4: debugfs: Simplify debugfs registration

The vc4 has a custom API to allow components to register a debugfs file
before the DRM driver has been registered and the debugfs_init hook has
been called.

However, the .late_register hook allows to have the debugfs file creation
deferred after that time already.

Let's remove our custom code to only register later our debugfs entries as
part of either debugfs_init or after it.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: debugfs: Return an error on failure
Maxime Ripard [Wed, 25 May 2022 15:21:41 +0000 (17:21 +0200)]
drm/vc4: debugfs: Return an error on failure

vc4_debugfs_add_file() can fail, so let's propagate its error code instead
of silencing it.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: debugfs: Protect device resources
Maxime Ripard [Wed, 11 May 2022 14:26:30 +0000 (16:26 +0200)]
drm/vc4: debugfs: Protect device resources

Our current code now mixes some resources whose lifetime are tied to the
device (clocks, IO mappings, etc.) and some that are tied to the DRM device
(encoder, bridge).

The device one will be freed at unbind time, but the DRM one will only be
freed when the last user of the DRM device closes its file handle.

So we end up with a time window during which we can call the encoder hooks,
but we don't have access to the underlying resources and device.

Let's protect all those sections with drm_dev_enter() and drm_dev_exit() so
that we bail out if we are during that window.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: vec: Switch to devm_pm_runtime_enable
Maxime Ripard [Fri, 10 Jun 2022 08:51:26 +0000 (10:51 +0200)]
drm/vc4: vec: Switch to devm_pm_runtime_enable

devm_pm_runtime_enable() simplifies the driver a bit since it will call
pm_runtime_disable() automatically through a device-managed action.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: vec: Protect device resources after removal
Maxime Ripard [Tue, 17 May 2022 16:19:19 +0000 (18:19 +0200)]
drm/vc4: vec: Protect device resources after removal

Whenever the device and driver are unbound, the main device and all the
subdevices will be removed by calling their unbind() method.

However, the DRM device itself will only be freed when the last user will
have closed it.

It means that there is a time window where the device and its resources
aren't there anymore, but the userspace can still call into our driver.

Fortunately, the DRM framework provides the drm_dev_enter() and
drm_dev_exit() functions to make sure our underlying device is still there
for the section protected by those calls. Let's add them to the VEC driver.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: vec: Switch to DRM-managed connector initialization
Maxime Ripard [Wed, 11 May 2022 09:05:36 +0000 (11:05 +0200)]
drm/vc4: vec: Switch to DRM-managed connector initialization

The current code will call drm_connector_unregister() and
drm_connector_cleanup() when the device is unbound. However, by then, there
might still be some references held to that connector, including by the
userspace that might still have the DRM device open.

Let's switch to a DRM-managed initialization to clean up after ourselves
only once the DRM device has been last closed.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: vec: Switch to DRM-managed encoder initialization
Maxime Ripard [Wed, 11 May 2022 09:05:36 +0000 (11:05 +0200)]
drm/vc4: vec: Switch to DRM-managed encoder initialization

The current code will call drm_encoder_cleanup() when the device is
unbound. However, by then, there might still be some references held to
that encoder, including by the userspace that might still have the DRM
device open.

Let's switch to a DRM-managed initialization to clean up after ourselves
only once the DRM device has been last closed.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: vec: Remove call to drm_connector_unregister()
Maxime Ripard [Wed, 22 Jun 2022 14:20:35 +0000 (16:20 +0200)]
drm/vc4: vec: Remove call to drm_connector_unregister()

drm_connector_unregister() is only to be used for connectors that have been
registered through drm_connector_register() after drm_dev_register() has
been called. This is our case here so let's remove the call.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: vec: Switch to drmm_kzalloc
Maxime Ripard [Mon, 23 May 2022 15:11:29 +0000 (17:11 +0200)]
drm/vc4: vec: Switch to drmm_kzalloc

Our internal structure that stores the DRM entities structure is allocated
through a device-managed kzalloc.

This means that this will eventually be freed whenever the device is
removed. In our case, the most likely source of removal is that the main
device is going to be unbound, and component_unbind_all() is being run.

However, it occurs while the DRM device is still registered, which will
create dangling pointers, eventually resulting in use-after-free.

Switch to a DRM-managed allocation to keep our structure until the DRM
driver doesn't need it anymore.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: vec: Embed DRM structures into the private structure
Maxime Ripard [Wed, 11 May 2022 09:03:23 +0000 (11:03 +0200)]
drm/vc4: vec: Embed DRM structures into the private structure

The VC4 VEC driver private structure contains only a pointer to the
encoder and connector it implements. This makes the overall structure
somewhat inconsistent with the rest of the driver, and complicates its
initialisation without any apparent gain.

Let's embed the drm_encoder structure (through the vc4_encoder one) and
drm_connector into struct vc4_vec to fix both issues.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: vec: Remove vc4_dev vec pointer
Maxime Ripard [Thu, 3 Sep 2020 08:01:20 +0000 (10:01 +0200)]
drm/vc4: vec: Remove vc4_dev vec pointer

There's no user for that pointer so let's just get rid of it.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: txp: Protect device resources
Maxime Ripard [Wed, 11 May 2022 14:26:30 +0000 (16:26 +0200)]
drm/vc4: txp: Protect device resources

Our current code now mixes some resources whose lifetime are tied to the
device (clocks, IO mappings, etc.) and some that are tied to the DRM device
(encoder, bridge).

The device one will be freed at unbind time, but the DRM one will only be
freed when the last user of the DRM device closes its file handle.

So we end up with a time window during which we can call the encoder hooks,
but we don't have access to the underlying resources and device.

Let's protect all those sections with drm_dev_enter() and drm_dev_exit() so
that we bail out if we are during that window.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: txp: Remove call to drm_connector_unregister()
Maxime Ripard [Wed, 22 Jun 2022 14:20:35 +0000 (16:20 +0200)]
drm/vc4: txp: Remove call to drm_connector_unregister()

drm_connector_unregister() is only to be used for connectors that have been
registered through drm_connector_register() after drm_dev_register() has
been called. This is our case here so let's remove the call.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: txp: Switch to drmm_kzalloc
Maxime Ripard [Mon, 23 May 2022 15:11:29 +0000 (17:11 +0200)]
drm/vc4: txp: Switch to drmm_kzalloc

Our internal structure that stores the DRM entities structure is allocated
through a device-managed kzalloc.

This means that this will eventually be freed whenever the device is
removed. In our case, the most likely source of removal is that the main
device is going to be unbound, and component_unbind_all() is being run.

However, it occurs while the DRM device is still registered, which will
create dangling pointers, eventually resulting in use-after-free.

Switch to a DRM-managed allocation to keep our structure until the DRM
driver doesn't need it anymore.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: txp: Remove duplicate regset
Maxime Ripard [Mon, 23 May 2022 13:44:50 +0000 (15:44 +0200)]
drm/vc4: txp: Remove duplicate regset

There's already a regset in the vc4_crtc structure so there's no need to
duplicate it in vc4_txp.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: txp: Remove vc4_dev txp pointer
Maxime Ripard [Thu, 3 Sep 2020 08:01:20 +0000 (10:01 +0200)]
drm/vc4: txp: Remove vc4_dev txp pointer

There's no user for that pointer so let's just get rid of it.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: hdmi: Switch to devm_pm_runtime_enable
Maxime Ripard [Fri, 10 Jun 2022 08:51:26 +0000 (10:51 +0200)]
drm/vc4: hdmi: Switch to devm_pm_runtime_enable

devm_pm_runtime_enable() simplifies the driver a bit since it will call
pm_runtime_disable() automatically through a device-managed action.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: hdmi: Protect device resources after removal
Maxime Ripard [Tue, 17 May 2022 16:19:19 +0000 (18:19 +0200)]
drm/vc4: hdmi: Protect device resources after removal

Whenever the device and driver are unbound, the main device and all the
subdevices will be removed by calling their unbind() method.

However, the DRM device itself will only be freed when the last user will
have closed it.

It means that there is a time window where the device and its resources
aren't there anymore, but the userspace can still call into our driver.

Fortunately, the DRM framework provides the drm_dev_enter() and
drm_dev_exit() functions to make sure our underlying device is still there
for the section protected by those calls. Let's add them to the HDMI driver.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: hdmi: Move audio structure offset checks
Maxime Ripard [Tue, 17 May 2022 10:38:06 +0000 (12:38 +0200)]
drm/vc4: hdmi: Move audio structure offset checks

The HDMI driver unbind hook doesn't have any ALSA-related code anymore, so
let's move the ALSA sanity checks and comments we have to some other part
of the driver dedicated to ALSA.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: hdmi: Use devm to register hotplug interrupts
Maxime Ripard [Tue, 17 May 2022 10:32:05 +0000 (12:32 +0200)]
drm/vc4: hdmi: Use devm to register hotplug interrupts

Commit 776efe800fed ("drm/vc4: hdmi: Drop devm interrupt handler for
hotplug interrupts") dropped the device-managed interrupt registration
because it was creating bugs and races whenever an interrupt was coming in
while the device was removed.

However, our latest patches to the HDMI controller driver fix this as well,
so we can use device-managed interrupt handlers again.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: hdmi: Switch to DRM-managed kfree to build regsets
Maxime Ripard [Tue, 17 May 2022 10:30:40 +0000 (12:30 +0200)]
drm/vc4: hdmi: Switch to DRM-managed kfree to build regsets

The current code to build the registers set later exposed in debugfs for
the HDMI controller relies on traditional allocations, that are later
free'd as part of the driver unbind hook.

Since krealloc doesn't have a DRM-managed equivalent, let's add an action
to free the buffer later on.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: hdmi: Use a device-managed action for DDC
Maxime Ripard [Tue, 17 May 2022 10:38:47 +0000 (12:38 +0200)]
drm/vc4: hdmi: Use a device-managed action for DDC

The reference to the DDC controller device needs to be put back when we're
done with it. Let's use a device-managed action to simplify the driver.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: hdmi: Switch to device-managed CEC initialization
Maxime Ripard [Fri, 13 May 2022 11:57:41 +0000 (13:57 +0200)]
drm/vc4: hdmi: Switch to device-managed CEC initialization

The current code to unregister our CEC device needs to be undone manually
when we remove the HDMI driver.

Since the CEC framework will allocate its main structure, and will defer
its deallocation to when the last user will have closed it, we don't really
need to take any particular measure to prevent any use-after-free and can
thus use any managed action.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: hdmi: Switch to device-managed ALSA initialization
Maxime Ripard [Fri, 13 May 2022 11:57:41 +0000 (13:57 +0200)]
drm/vc4: hdmi: Switch to device-managed ALSA initialization

The current code to unregister our ALSA device needs to be undone manually
when we remove the HDMI driver.

Since ALSA doesn't seem to support any mechanism to defer freeing something
until the last user of the ALSA device is gone, we can either use a
device-managed or a DRM-managed action.

The consistent way would be to use a DRM-managed one, just like pretty much
any framework-facing structure should be doing. However, ALSA does a lot of
allocation and registration using device-managed calls. Thus, if we're
going that way, by the time the DRM-managed action would run all of those
allocation would have been freed and we would end up with a use-after-free.

Thus, let's do a device-managed action. It's been tested with KASAN enabled
and doesn't seem to trigger any issue, so it's as good as anything.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: hdmi: Switch to DRM-managed connector initialization
Maxime Ripard [Wed, 11 May 2022 09:05:36 +0000 (11:05 +0200)]
drm/vc4: hdmi: Switch to DRM-managed connector initialization

The current code will call drm_connector_unregister() and
drm_connector_cleanup() when the device is unbound. However, by then, there
might still be some references held to that connector, including by the
userspace that might still have the DRM device open.

Let's switch to a DRM-managed initialization to clean up after ourselves
only once the DRM device has been last closed.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: hdmi: Switch to DRM-managed encoder initialization
Maxime Ripard [Wed, 11 May 2022 09:05:36 +0000 (11:05 +0200)]
drm/vc4: hdmi: Switch to DRM-managed encoder initialization

The current code will call drm_encoder_cleanup() when the device is
unbound. However, by then, there might still be some references held to
that encoder, including by the userspace that might still have the DRM
device open.

Let's switch to a DRM-managed initialization to clean up after ourselves
only once the DRM device has been last closed.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: hdmi: Remove call to drm_connector_unregister()
Maxime Ripard [Wed, 22 Jun 2022 14:20:35 +0000 (16:20 +0200)]
drm/vc4: hdmi: Remove call to drm_connector_unregister()

drm_connector_unregister() is only to be used for connectors that have been
registered through drm_connector_register() after drm_dev_register() has
been called. This is our case here so let's remove the call.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: hdmi: Switch to drmm_kzalloc
Maxime Ripard [Mon, 23 May 2022 15:11:29 +0000 (17:11 +0200)]
drm/vc4: hdmi: Switch to drmm_kzalloc

Our internal structure that stores the DRM entities structure is allocated
through a device-managed kzalloc.

This means that this will eventually be freed whenever the device is
removed. In our case, the most likely source of removal is that the main
device is going to be unbound, and component_unbind_all() is being run.

However, it occurs while the DRM device is still registered, which will
create dangling pointers, eventually resulting in use-after-free.

Switch to a DRM-managed allocation to keep our structure until the DRM
driver doesn't need it anymore.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: dsi: Switch to devm_pm_runtime_enable
Maxime Ripard [Fri, 10 Jun 2022 08:51:26 +0000 (10:51 +0200)]
drm/vc4: dsi: Switch to devm_pm_runtime_enable

devm_pm_runtime_enable() simplifies the driver a bit since it will call
pm_runtime_disable() automatically through a device-managed action.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: dsi: Fix the driver structure lifetime
Maxime Ripard [Mon, 23 May 2022 13:48:01 +0000 (15:48 +0200)]
drm/vc4: dsi: Fix the driver structure lifetime

The vc4_dsi structure is currently allocated through a device-managed
allocation. This can lead to use-after-free issues however in the unbinding
path since the DRM entities will stick around, but the underlying structure
has been freed.

However, we can't just fix it by using a DRM-managed allocation like we did
for the other drivers since the DSI case is a bit more intricate.

Indeed, the structure will be allocated at probe time, when we don't have a
DRM device yet, to be able to register the DSI bus driver. We will then
reuse it at bind time to register our KMS entities in the framework.

In order to work around both constraints, we can use a kref to track the
users of the structure (DSI host, and KMS), and then put our structure when
the DSI host will have been unregistered, and through a DRM-managed action
that will execute once we won't need the KMS entities anymore.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: dsi: Switch to drmm_of_get_bridge
Maxime Ripard [Wed, 11 May 2022 14:08:44 +0000 (16:08 +0200)]
drm/vc4: dsi: Switch to drmm_of_get_bridge

The current code uses a device-managed function to retrieve the next bridge
downstream.

However, that means that it will be removed at unbind time, where the DRM
device is still very much live and might still have some applications that
still have it open.

Switch to a DRM-managed variant to clean everything up once the DRM device
has been last closed.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: dsi: Switch to DRM-managed encoder initialization
Maxime Ripard [Wed, 11 May 2022 09:05:36 +0000 (11:05 +0200)]
drm/vc4: dsi: Switch to DRM-managed encoder initialization

The current code will call drm_encoder_cleanup() when the device is
unbound. However, by then, there might still be some references held to
that encoder, including by the userspace that might still have the DRM
device open.

Let's switch to a DRM-managed initialization to clean up after ourselves
only once the DRM device has been last closed.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: dsi: Embed DRM structures into the private structure
Maxime Ripard [Wed, 11 May 2022 09:03:23 +0000 (11:03 +0200)]
drm/vc4: dsi: Embed DRM structures into the private structure

The VC4 DSI driver private structure contains only a pointer to the
encoder it implements. This makes the overall structure somewhat
inconsistent with the rest of the driver, and complicates its
initialisation without any apparent gain.

Let's embed the drm_encoder structure (through the vc4_encoder one) into
struct vc4_dsi to fix both issues.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: dpi: Protect device resources
Maxime Ripard [Wed, 11 May 2022 14:26:30 +0000 (16:26 +0200)]
drm/vc4: dpi: Protect device resources

Our current code now mixes some resources whose lifetime are tied to the
device (clocks, IO mappings, etc.) and some that are tied to the DRM device
(encoder, bridge).

The device one will be freed at unbind time, but the DRM one will only be
freed when the last user of the DRM device closes its file handle.

So we end up with a time window during which we can call the encoder hooks,
but we don't have access to the underlying resources and device.

Let's protect all those sections with drm_dev_enter() and drm_dev_exit() so
that we bail out if we are during that window.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: dpi: Switch to drmm_of_get_bridge
Maxime Ripard [Wed, 11 May 2022 14:08:44 +0000 (16:08 +0200)]
drm/vc4: dpi: Switch to drmm_of_get_bridge

The current code uses a device-managed function to retrieve the next bridge
downstream.

However, that means that it will be removed at unbind time, where the DRM
device is still very much live and might still have some applications that
still have it open.

Switch to a DRM-managed variant to clean everything up once the DRM device
has been last closed.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: dpi: Switch to DRM-managed encoder initialization
Maxime Ripard [Wed, 11 May 2022 09:05:36 +0000 (11:05 +0200)]
drm/vc4: dpi: Switch to DRM-managed encoder initialization

The current code will call drm_encoder_cleanup() when the device is
unbound. However, by then, there might still be some references held to
that encoder, including by the userspace that might still have the DRM
device open.

Let's switch to a DRM-managed initialization to clean up after ourselves
only once the DRM device has been last closed.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: dpi: Add action to disable the clock
Maxime Ripard [Mon, 23 May 2022 15:04:25 +0000 (17:04 +0200)]
drm/vc4: dpi: Add action to disable the clock

The DPI controller has two clocks called core and pixel, the core clock
being enabled at bind time.

Adding a device-managed action will make the error path easier, so let's
create one to disable it.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: dpi: Remove unnecessary drm_of_panel_bridge_remove call
Maxime Ripard [Wed, 11 May 2022 09:19:22 +0000 (11:19 +0200)]
drm/vc4: dpi: Remove unnecessary drm_of_panel_bridge_remove call

Since we have a managed call to create our panel_bridge instance, the call
to drm_of_panel_bridge_remove() at unbind is both redundant and dangerous
since it might lead to a use-after-free.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: dpi: Return an error if we can't enable our clock
Maxime Ripard [Wed, 11 May 2022 09:09:16 +0000 (11:09 +0200)]
drm/vc4: dpi: Return an error if we can't enable our clock

If we fail to enable the DPI clock, we just ignore the error and moves
forward. Let's return an error instead.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: dpi: Switch to drmm_kzalloc
Maxime Ripard [Mon, 23 May 2022 15:11:29 +0000 (17:11 +0200)]
drm/vc4: dpi: Switch to drmm_kzalloc

Our internal structure that stores the DRM entities structure is allocated
through a device-managed kzalloc.

This means that this will eventually be freed whenever the device is
removed. In our case, the most likely source of removal is that the main
device is going to be unbound, and component_unbind_all() is being run.

However, it occurs while the DRM device is still registered, which will
create dangling pointers, eventually resulting in use-after-free.

Switch to a DRM-managed allocation to keep our structure until the DRM
driver doesn't need it anymore.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: dpi: Embed DRM structures into the private structure
Maxime Ripard [Wed, 11 May 2022 09:03:23 +0000 (11:03 +0200)]
drm/vc4: dpi: Embed DRM structures into the private structure

The VC4 DPI driver private structure contains only a pointer to the
encoder it implements. This makes the overall structure somewhat
inconsistent with the rest of the driver, and complicates its
initialisation without any apparent gain.

Let's embed the drm_encoder structure (through the vc4_encoder one) into
struct vc4_dpi to fix both issues.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: dpi: Remove vc4_dev dpi pointer
Maxime Ripard [Thu, 3 Sep 2020 08:01:20 +0000 (10:01 +0200)]
drm/vc4: dpi: Remove vc4_dev dpi pointer

There's no user for that pointer so let's just get rid of it.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: crtc: Switch to DRM-managed CRTC initialization
Maxime Ripard [Wed, 11 May 2022 09:05:36 +0000 (11:05 +0200)]
drm/vc4: crtc: Switch to DRM-managed CRTC initialization

The current code will call drm_crtc_cleanup() when the device is
unbound. However, by then, there might still be some references held to
that CRTC, including by the userspace that might still have the DRM
device open.

Let's switch to a DRM-managed initialization to clean up after ourselves
only once the DRM device has been last closed.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: crtc: Switch to drmm_kzalloc
Maxime Ripard [Mon, 23 May 2022 15:11:29 +0000 (17:11 +0200)]
drm/vc4: crtc: Switch to drmm_kzalloc

Our internal structure that stores the DRM entities structure is allocated
through a device-managed kzalloc.

This means that this will eventually be freed whenever the device is
removed. In our case, the most likely source of removal is that the main
device is going to be unbound, and component_unbind_all() is being run.

However, it occurs while the DRM device is still registered, which will
create dangling pointers, eventually resulting in use-after-free.

Switch to a DRM-managed allocation to keep our structure until the DRM
driver doesn't need it anymore.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: crtc: Move debugfs_name to crtc_data
Maxime Ripard [Mon, 23 May 2022 13:44:13 +0000 (15:44 +0200)]
drm/vc4: crtc: Move debugfs_name to crtc_data

All the CRTCs, including the TXP, have a debugfs file and name so we can
consolidate it into vc4_crtc_data.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: plane: Switch to drmm_universal_plane_alloc()
Maxime Ripard [Wed, 25 May 2022 09:05:30 +0000 (11:05 +0200)]
drm/vc4: plane: Switch to drmm_universal_plane_alloc()

Let's switch to drmm_universal_plane_alloc() for our plane allocation and
initialisation to make the driver a bit simpler.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: crtc: Remove manual plane removal on error
Maxime Ripard [Wed, 22 Jun 2022 14:15:11 +0000 (16:15 +0200)]
drm/vc4: crtc: Remove manual plane removal on error

When vc4_crtc_bind() fails after vc4_crtc_init() has been called, we have
a loop undoing the plane creation and calling destroy on each plane
registered and matching the possible_crtcs mask.

However, this is redundant with what drm_mode_config_cleanup() is doing, so
let's remove it.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: plane: Take possible_crtcs as an argument
Maxime Ripard [Wed, 25 May 2022 09:03:10 +0000 (11:03 +0200)]
drm/vc4: plane: Take possible_crtcs as an argument

vc4_plane_init() currently initialises the plane with no possible CRTCs,
and will expect the caller to set it up by itself.

Let's change that logic a bit to follow the syntax of
drm_universal_plane_init() and pass the possible CRTCs bitmask as an
argument to the function instead.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: hvs: Remove planes currently allocated before taking down
Maxime Ripard [Mon, 9 May 2022 11:46:54 +0000 (13:46 +0200)]
drm/vc4: hvs: Remove planes currently allocated before taking down

When the HVS driver is unbound, a lot of memory allocations in the LBM and
DLIST RAM are still assigned to planes that are still allocated.

Thus, we hit a warning when calling drm_mm_takedown() since the memory pool
is not completely free of allocations.

Let's free all the currently live entries before calling drm_mm_takedown().

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: hvs: Protect device resources after removal
Maxime Ripard [Tue, 17 May 2022 16:19:19 +0000 (18:19 +0200)]
drm/vc4: hvs: Protect device resources after removal

Whenever the device and driver are unbound, the main device and all the
subdevices will be removed by calling their unbind() method.

However, the DRM device itself will only be freed when the last user will
have closed it.

It means that there is a time window where the device and its resources
aren't there anymore, but the userspace can still call into our driver.

Fortunately, the DRM framework provides the drm_dev_enter() and
drm_dev_exit() functions to make sure our underlying device is still there
for the section protected by those calls. Let's add them to the HVS driver.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: crtc: Create vblank reporting function
Maxime Ripard [Mon, 20 Jun 2022 12:05:46 +0000 (14:05 +0200)]
drm/vc4: crtc: Create vblank reporting function

We'll need that code in the HVS driver, so let's create a shared function
to reuse it.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: drv: Use drm_dev_unplug
Maxime Ripard [Thu, 16 Jun 2022 11:39:23 +0000 (13:39 +0200)]
drm/vc4: drv: Use drm_dev_unplug

When our KMS driver is unbound, the device is no longer there but we might
still have users with an opened fd to the KMS device.

To avoid any issue in such a situation, every device access needs to be
protected by calls to drm_dev_enter() and drm_dev_exit(), and the driver
needs to call drm_dev_unplug().

We'll add calls to drm_dev_enter()/drm_dev_exit() in subsequent patches
changing the relevant drivers, but let's start by calling drm_dev_unplug().

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: drv: Call component_unbind_all()
Maxime Ripard [Mon, 9 May 2022 10:33:32 +0000 (12:33 +0200)]
drm/vc4: drv: Call component_unbind_all()

While we were using the component framework to deal with all the DRM
subdevices, we were not calling component_unbind_all().

This leads to none of the subdevices freeing up their resources as part of
their unbind() or device managed hooks.

Fixes: c8b75bca92cb ("drm/vc4: Add KMS support for Raspberry Pi.")
Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: Add module dependency on hdmi-codec
Maxime Ripard [Fri, 2 Sep 2022 11:37:22 +0000 (13:37 +0200)]
drm/vc4: Add module dependency on hdmi-codec

The VC4 HDMI controller driver relies on the HDMI codec ASoC driver. In
order to set it up properly, in vc4_hdmi_audio_init(), our HDMI driver
will register a device matching the HDMI codec driver, and then register
an ASoC card using that codec.

However, if vc4 is compiled as a module, chances are that the hdmi-codec
driver will be too. In such a case, the module loader will have a very
narrow window to load the module between the device registration and the
card registration.

If it fails to load the module in time, the card registration will fail
with EPROBE_DEFER, and we'll abort the audio initialisation,
unregistering the HDMI codec device in the process.

The next time the bind callback will be run, it's likely that we end up
missing that window again, effectively preventing vc4 to probe entirely.

In order to prevent this, we can create a soft dependency of the vc4
driver on the HDMI codec one so that we're sure the HDMI codec will be
loaded before the VC4 module is, and thus we'll never end up in the
previous situation.

Fixes: 91e99e113929 ("drm/vc4: hdmi: Register HDMI codec")
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/bridge: panel: Introduce drmm_of_get_bridge
Maxime Ripard [Mon, 23 May 2022 14:38:52 +0000 (16:38 +0200)]
drm/bridge: panel: Introduce drmm_of_get_bridge

Unlike what can be found for other DRM entities, we don't have a
DRM-managed function equivalent to devm_drm_of_get_bridge().

Let's create it.

Acked-by: Sam Ravnborg <sam@ravnborg.org>
Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/bridge: panel: Introduce drmm_panel_bridge_add
Maxime Ripard [Mon, 23 May 2022 14:27:12 +0000 (16:27 +0200)]
drm/bridge: panel: Introduce drmm_panel_bridge_add

Unlike what can be found for other entities, there's no DRM-managed
function to create a panel_bridge instance from a panel.

Let's introduce one.

Acked-by: Sam Ravnborg <sam@ravnborg.org>
Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/connector: Introduce drmm_connector_init
Maxime Ripard [Tue, 17 May 2022 13:09:33 +0000 (15:09 +0200)]
drm/connector: Introduce drmm_connector_init

Unlike other DRM entities, there's no helper to create a DRM-managed
initialisation of a connector.

Let's create an helper to initialise a connector that would be passed as an
argument, and handle the cleanup through a DRM-managed action.

Acked-by: Sam Ravnborg <sam@ravnborg.org>
Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/connector: Check for destroy implementation
Maxime Ripard [Wed, 29 Jun 2022 09:17:10 +0000 (11:17 +0200)]
drm/connector: Check for destroy implementation

Connectors need to be cleaned up with a call to drm_connector_cleanup()
in their drm_connector_funcs.destroy implementation.

Let's check for this and complain if there's no such function.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/connector: Consolidate Connector Initialization
Maxime Ripard [Wed, 29 Jun 2022 09:04:05 +0000 (11:04 +0200)]
drm/connector: Consolidate Connector Initialization

We're going to add a DRM-managed connector initialization function.
Since we'll need both the with and without the DDC pointer, having a
single function that takes an optional pointer is easier to maintain.

Let's create a static function that will back both existing variants,
and will be reused by the DRM-managed variant.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Suggested-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/connector: Clarify when drm_connector_unregister is needed
Maxime Ripard [Wed, 22 Jun 2022 08:33:59 +0000 (10:33 +0200)]
drm/connector: Clarify when drm_connector_unregister is needed

The current documentation for drm_connector_unregister() mentions that
it's needed for connectors that have been registered through
drm_dev_register().

However, this was a typo and was meant to be drm_connector_register(),
which only applies to connectors registered after drm_dev_register() has
been called.

In addition, it was also mentioning that connectors are unregistered
automatically when drm_dev_unregister() is called. This part is a bit
misleading, since it might make it appear that
drm_connector_unregister() applies either to all connectors, or none of
them.

After discussing it with Daniel, it appears that we always need to call
drm_connector_unregister() on connectors that have been registered with
drm_connector_register(), but only those.

drm_connector_init() already mentions that it only needs
drm_connector_cleanup(), so let's clarify the drm_connector_register()
and drm_connector_unregister() documentation to point at each other, and
remove the misleading part about drm_dev_unregister().

Acked-by: Sam Ravnborg <sam@ravnborg.org>
Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/connector: Mention the cleanup after drm_connector_init
Maxime Ripard [Tue, 17 May 2022 13:03:34 +0000 (15:03 +0200)]
drm/connector: Mention the cleanup after drm_connector_init

Unlike encoders and CRTCs, the drm_connector_init() and
drm_connector_init_with_ddc() don't mention how the cleanup is supposed to
be done. Let's add it.

Acked-by: Sam Ravnborg <sam@ravnborg.org>
Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/connector: Reorder headers
Maxime Ripard [Tue, 17 May 2022 11:22:54 +0000 (13:22 +0200)]
drm/connector: Reorder headers

Unlike most of the other files in DRM, and Linux in general, the headers in
drm_connector.c aren't sorted alphabetically. Let's fix that.

Acked-by: Sam Ravnborg <sam@ravnborg.org>
Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/encoder: Introduce drmm_encoder_init
Maxime Ripard [Tue, 17 May 2022 13:09:33 +0000 (15:09 +0200)]
drm/encoder: Introduce drmm_encoder_init

The DRM-managed function to register an encoder is
drmm_encoder_alloc() and its variants, which will allocate the underlying
structure and initialisation the encoder.

However, we might want to separate the structure creation and the encoder
initialisation, for example if the structure is shared across multiple DRM
entities, for example an encoder and a connector.

Let's create an helper to only initialise an encoder that would be passed
as an argument.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/crtc: Introduce drmm_crtc_init_with_planes
Maxime Ripard [Tue, 17 May 2022 13:09:33 +0000 (15:09 +0200)]
drm/crtc: Introduce drmm_crtc_init_with_planes

The DRM-managed function to register a CRTC is
drmm_crtc_alloc_with_planes(), which will allocate the underlying
structure and initialisation the CRTC.

However, we might want to separate the structure creation and the CRTC
initialisation, for example if the structure is shared across multiple
DRM entities, for example an encoder and a connector.

Let's create an helper to only initialise a CRTC that would be passed as
an argument.

Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/mipi-dsi: Detach devices when removing the host
Maxime Ripard [Mon, 23 May 2022 13:42:24 +0000 (15:42 +0200)]
drm/mipi-dsi: Detach devices when removing the host

Whenever the MIPI-DSI host is unregistered, the code of
mipi_dsi_host_unregister() loops over every device currently found on that
bus and will unregister it.

However, it doesn't detach it from the bus first, which leads to all kind
of resource leaks if the host wants to perform some clean up whenever a
device is detached.

Fixes: 068a00233969 ("drm: Add MIPI DSI bus support")
Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm: Add DRM-managed mutex_init()
Thomas Zimmermann [Mon, 2 May 2022 14:25:12 +0000 (16:25 +0200)]
drm: Add DRM-managed mutex_init()

Add drmm_mutex_init(), a helper that provides managed mutex cleanup. The
mutex will be destroyed with the final reference of the DRM device.

Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Jocelyn Falempe <jfalempe@redhat.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20220502142514.2174-2-tzimmermann@suse.de
2 years agodrm/vc4: dsi: Release workaround buffer and DMA
Dave Stevenson [Mon, 13 Jun 2022 14:47:38 +0000 (16:47 +0200)]
drm/vc4: dsi: Release workaround buffer and DMA

On Pi0-3 the driver allocates a buffer and requests a DMA channel
because the ARM can't write to DSI1's registers directly.

However, we never release that buffer or channel. Let's add a
device-managed action to release each.

Fixes: 4078f5757144 ("drm/vc4: Add DSI driver")
Signed-off-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
Link: https://lore.kernel.org/r/20220613144800.326124-12-maxime@cerno.tech
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
2 years agodrm/vc4: dsi: Switch to devm_drm_of_get_bridge
Maxime Ripard [Fri, 10 Sep 2021 13:09:41 +0000 (15:09 +0200)]
drm/vc4: dsi: Switch to devm_drm_of_get_bridge

The new devm_drm_of_get_bridge removes most of the boilerplate we
have to deal with. Let's switch to it.

Signed-off-by: Maxime Ripard <maxime@cerno.tech>
Acked-by: Sam Ravnborg <sam@ravnborg.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20210910130941.1740182-4-maxime@cerno.tech
2 years agodrm/vc4: dpi: Switch to devm_drm_of_get_bridge
Maxime Ripard [Fri, 10 Sep 2021 13:09:40 +0000 (15:09 +0200)]
drm/vc4: dpi: Switch to devm_drm_of_get_bridge

The new devm_drm_of_get_bridge removes most of the boilerplate we
have to deal with. Let's switch to it.

Signed-off-by: Maxime Ripard <maxime@cerno.tech>
Acked-by: Sam Ravnborg <sam@ravnborg.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20210910130941.1740182-3-maxime@cerno.tech
2 years agodrm/bridge: Move devm_drm_of_get_bridge to bridge/panel.c
Maxime Ripard [Fri, 17 Sep 2021 18:09:25 +0000 (20:09 +0200)]
drm/bridge: Move devm_drm_of_get_bridge to bridge/panel.c

By depending on devm_drm_panel_bridge_add(), devm_drm_of_get_bridge()
introduces a circular dependency between the modules drm (where
devm_drm_of_get_bridge() ends up) and drm_kms_helper (where
devm_drm_panel_bridge_add() is).

Fix this by moving devm_drm_of_get_bridge() to bridge/panel.c and thus
drm_kms_helper.

Fixes: 87ea95808d53 ("drm/bridge: Add a function to abstract away panels")
Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
Tested-by: Heiko Stuebner <heiko@sntech.de>
Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Link: https://patchwork.freedesktop.org/patch/msgid/20210917180925.2602266-1-maxime@cerno.tech
2 years agodrm/bridge: Add a function to abstract away panels
Maxime Ripard [Fri, 10 Sep 2021 13:09:39 +0000 (15:09 +0200)]
drm/bridge: Add a function to abstract away panels

Display drivers so far need to have a lot of boilerplate to first
retrieve either the panel or bridge that they are connected to using
drm_of_find_panel_or_bridge(), and then either deal with each with ad-hoc
functions or create a drm panel bridge through drm_panel_bridge_add.

In order to reduce the boilerplate and hopefully create a path of least
resistance towards using the DRM panel bridge layer, let's create the
function devm_drm_of_get_bridge() to reduce that boilerplate.

Signed-off-by: Maxime Ripard <maxime@cerno.tech>
Reviewed-by: Sam Ravnborg <sam@ravnborg.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20210910130941.1740182-2-maxime@cerno.tech
2 years agooverlays: Shorten overlay name to watterott-display
Phil Elwell [Wed, 7 Sep 2022 08:38:24 +0000 (09:38 +0100)]
overlays: Shorten overlay name to watterott-display

watterott-rpi-display, although accurate, is too cumbersome, leading to
some potentially very long lines, Shorten it to watterott-display.

Signed-off-by: Phil Elwell <phil@raspberrypi.com>
2 years agooverlays: Rename unofficial "rpi-" overlays
Phil Elwell [Tue, 6 Sep 2022 13:43:34 +0000 (14:43 +0100)]
overlays: Rename unofficial "rpi-" overlays

There are a number of overlays - some very old - with names beginning
with "rpi-", even though they aren't for products made by Raspberry Pi.
Rename them to make their status clearer. The entries in the overlay
map will translate the old names to the new ones, with a warning message
that may be displayed in the GUI.

Signed-off-by: Phil Elwell <phil@raspberrypi.com>
2 years agooverlays: Add the Raspberry Pi sound cards
Phil Elwell [Mon, 5 Sep 2022 11:16:32 +0000 (12:16 +0100)]
overlays: Add the Raspberry Pi sound cards

Signed-off-by: Phil Elwell <phil@raspberrypi.com>
2 years agomedia: i2c: imx290: Do not reset exposure time from set_fmt
Dave Stevenson [Mon, 5 Sep 2022 14:32:41 +0000 (15:32 +0100)]
media: i2c: imx290: Do not reset exposure time from set_fmt

V4L2 controls are meant to be persistent, and only altered if
the values break hardware constraints.

In imx290_set_fmt the CID_VBLANK parameter was being reset to
the mode default, and therefore the maximum exposure value
potentially also changed.

VBLANK is no longer reset, and even if it is, then the set_ctrl
that will be called will alter the range for the exposure
control. Therefore there is no need to reset the exposure time
from set_fmt.

Signed-off-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
2 years agomedia: i2c: imx290: Drop incorrect comment about pixelrate
Dave Stevenson [Mon, 5 Sep 2022 14:10:05 +0000 (15:10 +0100)]
media: i2c: imx290: Drop incorrect comment about pixelrate

commit b5eca6fd4284 ("media: i2c: imx290: Fix the pixel rate at 148.5Mpix/s")
corrected imx290_calc_pixel_rate to always return the correct
pixle rate of 148.5Mpix/s, but didn't remove the comment that said
it relied on some other parameters.

Drop that comment.

Signed-off-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
2 years agoRevert "media: i2c: imx290: Explicitly set v&h blank on mode change"
Dave Stevenson [Mon, 5 Sep 2022 10:50:30 +0000 (11:50 +0100)]
Revert "media: i2c: imx290: Explicitly set v&h blank on mode change"

This reverts commit 090d29f5854aa7b6511b0ae8a926929130fd3978.

Whilst the behaviour of not resetting the frame rate to the
"correct"/default value on mode change is useful, it is not correct
under the V4L2 API.

Revert the patch so that blanking values persist.

Signed-off-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
2 years agomedia: i2c: imx290: Updating VBLANK should update exposure in all states
Dave Stevenson [Mon, 5 Sep 2022 10:44:23 +0000 (11:44 +0100)]
media: i2c: imx290: Updating VBLANK should update exposure in all states

The code to update the range of the exposure control was after a
check that the sensor was powered up. This is incorrect as the
range should be updated under all conditions.

Move the range update code to the correct place.

Fixes: 52d076ea221e "media: i2c: imx290: Handle exposure correctly when vblank changes"

Signed-off-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
2 years agoMerge tag 'v5.15.65' into rpi-5.15.y
Dom Cobley [Mon, 5 Sep 2022 14:01:30 +0000 (15:01 +0100)]
Merge tag 'v5.15.65' into rpi-5.15.y

This is the 5.15.65 stable release

2 years agoMerge tag 'v5.15.64' into rpi-5.15.y
Dom Cobley [Mon, 5 Sep 2022 11:46:18 +0000 (12:46 +0100)]
Merge tag 'v5.15.64' into rpi-5.15.y

This is the 5.15.64 stable release

2 years agoRevert "Revert "usbnet: smsc95xx: Fix deadlock on runtime resume""
Dom Cobley [Mon, 5 Sep 2022 11:45:09 +0000 (12:45 +0100)]
Revert "Revert "usbnet: smsc95xx: Fix deadlock on runtime resume""

This reverts commit 2f82838ec69a86ffe786e57a37d4bcf746760017.

2 years agoRevert "Revert "usbnet: smsc95xx: Forward PHY interrupts to PHY driver to avoid polling""
Dom Cobley [Mon, 5 Sep 2022 11:45:08 +0000 (12:45 +0100)]
Revert "Revert "usbnet: smsc95xx: Forward PHY interrupts to PHY driver to avoid polling""

This reverts commit bb0401e1e9875b361aff49e75b9ef9e6688359bc.

2 years agoRevert "Revert "usbnet: smsc95xx: Avoid link settings race on interrupt reception""
Dom Cobley [Mon, 5 Sep 2022 11:45:07 +0000 (12:45 +0100)]
Revert "Revert "usbnet: smsc95xx: Avoid link settings race on interrupt reception""

This reverts commit c7cbb86f4763c5d9b52698e6f58b6d1719ade011.

2 years agoRevert "Revert "usbnet: smsc95xx: Don't clear read-only PHY interrupt""
Dom Cobley [Mon, 5 Sep 2022 11:45:06 +0000 (12:45 +0100)]
Revert "Revert "usbnet: smsc95xx: Don't clear read-only PHY interrupt""

This reverts commit b50b2815045cbc86e3b34178e9df69a988991a61.

2 years agoMerge tag 'v5.15.63' into rpi-5.15.y
Dom Cobley [Mon, 5 Sep 2022 11:43:58 +0000 (12:43 +0100)]
Merge tag 'v5.15.63' into rpi-5.15.y

This is the 5.15.63 stable release

2 years agoMerge tag 'v5.15.62' into rpi-5.15.y
Dom Cobley [Mon, 5 Sep 2022 11:43:55 +0000 (12:43 +0100)]
Merge tag 'v5.15.62' into rpi-5.15.y

This is the 5.15.62 stable release

2 years agoLinux 5.15.65
Greg Kroah-Hartman [Mon, 5 Sep 2022 08:30:13 +0000 (10:30 +0200)]
Linux 5.15.65

Link: https://lore.kernel.org/r/20220902121404.435662285@linuxfoundation.org
Tested-by: Jon Hunter <jonathanh@nvidia.com>
Tested-by: Florian Fainelli <f.fainelli@gmail.com>
Tested-by: Shuah Khan <skhan@linuxfoundation.org>
Tested-by: Guenter Roeck <linux@roeck-us.net>
Tested-by: Linux Kernel Functional Testing <lkft@linaro.org>
Tested-by: Bagas Sanjaya <bagasdotme@gmail.com>
Tested-by: Ron Economos <re@w6rz.net>
Tested-by: Sudip Mukherjee <sudip.mukherjee@codethink.co.uk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agonet: neigh: don't call kfree_skb() under spin_lock_irqsave()
Yang Yingliang [Mon, 22 Aug 2022 02:53:46 +0000 (10:53 +0800)]
net: neigh: don't call kfree_skb() under spin_lock_irqsave()

commit d5485d9dd24e1d04e5509916515260186eb1455c upstream.

It is not allowed to call kfree_skb() from hardware interrupt
context or with interrupts being disabled. So add all skb to
a tmp list, then free them after spin_unlock_irqrestore() at
once.

Fixes: 66ba215cb513 ("neigh: fix possible DoS due to net iface start/stop loop")
Suggested-by: Denis V. Lunev <den@openvz.org>
Signed-off-by: Yang Yingliang <yangyingliang@huawei.com>
Reviewed-by: Nikolay Aleksandrov <razor@blackwall.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agonet/af_packet: check len when min_header_len equals to 0
Zhengchao Shao [Wed, 27 Jul 2022 09:33:12 +0000 (17:33 +0800)]
net/af_packet: check len when min_header_len equals to 0

commit dc633700f00f726e027846a318c5ffeb8deaaeda upstream.

User can use AF_PACKET socket to send packets with the length of 0.
When min_header_len equals to 0, packet_snd will call __dev_queue_xmit
to send packets, and sock->type can be any type.

Reported-by: syzbot+5ea725c25d06fb9114c4@syzkaller.appspotmail.com
Fixes: fd1894224407 ("bpf: Don't redirect packets with invalid pkt_len")
Signed-off-by: Zhengchao Shao <shaozhengchao@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agoandroid: binder: fix lockdep check on clearing vma
Liam Howlett [Mon, 27 Jun 2022 15:18:59 +0000 (15:18 +0000)]
android: binder: fix lockdep check on clearing vma

commit b0cab80ecd54ae3b2356bb081af0bffd538c8265 upstream.

When munmapping a vma, the mmap_lock can be degraded to a write before
calling close() on the file handle.  The binder close() function calls
binder_alloc_set_vma() to clear the vma address, which now has a lock dep
check for writing on the mmap_lock.  Change the lockdep check to ensure
the reading lock is held while clearing and keep the write check while
writing.

Link: https://lkml.kernel.org/r/20220627151857.2316964-1-Liam.Howlett@oracle.com
Fixes: 472a68df605b ("android: binder: stop saving a pointer to the VMA")
Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Reported-by: syzbot+da54fa8d793ca89c741f@syzkaller.appspotmail.com
Acked-by: Todd Kjos <tkjos@google.com>
Cc: "Arve Hjønnevåg" <arve@android.com>
Cc: Christian Brauner (Microsoft) <brauner@kernel.org>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Hridya Valsaraju <hridya@google.com>
Cc: Joel Fernandes <joel@joelfernandes.org>
Cc: Martijn Coenen <maco@android.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agobtrfs: fix space cache corruption and potential double allocations
Omar Sandoval [Tue, 23 Aug 2022 18:28:13 +0000 (11:28 -0700)]
btrfs: fix space cache corruption and potential double allocations

commit ced8ecf026fd8084cf175530ff85c76d6085d715 upstream.

When testing space_cache v2 on a large set of machines, we encountered a
few symptoms:

1. "unable to add free space :-17" (EEXIST) errors.
2. Missing free space info items, sometimes caught with a "missing free
   space info for X" error.
3. Double-accounted space: ranges that were allocated in the extent tree
   and also marked as free in the free space tree, ranges that were
   marked as allocated twice in the extent tree, or ranges that were
   marked as free twice in the free space tree. If the latter made it
   onto disk, the next reboot would hit the BUG_ON() in
   add_new_free_space().
4. On some hosts with no on-disk corruption or error messages, the
   in-memory space cache (dumped with drgn) disagreed with the free
   space tree.

All of these symptoms have the same underlying cause: a race between
caching the free space for a block group and returning free space to the
in-memory space cache for pinned extents causes us to double-add a free
range to the space cache. This race exists when free space is cached
from the free space tree (space_cache=v2) or the extent tree
(nospace_cache, or space_cache=v1 if the cache needs to be regenerated).
struct btrfs_block_group::last_byte_to_unpin and struct
btrfs_block_group::progress are supposed to protect against this race,
but commit d0c2f4fa555e ("btrfs: make concurrent fsyncs wait less when
waiting for a transaction commit") subtly broke this by allowing
multiple transactions to be unpinning extents at the same time.

Specifically, the race is as follows:

1. An extent is deleted from an uncached block group in transaction A.
2. btrfs_commit_transaction() is called for transaction A.
3. btrfs_run_delayed_refs() -> __btrfs_free_extent() runs the delayed
   ref for the deleted extent.
4. __btrfs_free_extent() -> do_free_extent_accounting() ->
   add_to_free_space_tree() adds the deleted extent back to the free
   space tree.
5. do_free_extent_accounting() -> btrfs_update_block_group() ->
   btrfs_cache_block_group() queues up the block group to get cached.
   block_group->progress is set to block_group->start.
6. btrfs_commit_transaction() for transaction A calls
   switch_commit_roots(). It sets block_group->last_byte_to_unpin to
   block_group->progress, which is block_group->start because the block
   group hasn't been cached yet.
7. The caching thread gets to our block group. Since the commit roots
   were already switched, load_free_space_tree() sees the deleted extent
   as free and adds it to the space cache. It finishes caching and sets
   block_group->progress to U64_MAX.
8. btrfs_commit_transaction() advances transaction A to
   TRANS_STATE_SUPER_COMMITTED.
9. fsync calls btrfs_commit_transaction() for transaction B. Since
   transaction A is already in TRANS_STATE_SUPER_COMMITTED and the
   commit is for fsync, it advances.
10. btrfs_commit_transaction() for transaction B calls
    switch_commit_roots(). This time, the block group has already been
    cached, so it sets block_group->last_byte_to_unpin to U64_MAX.
11. btrfs_commit_transaction() for transaction A calls
    btrfs_finish_extent_commit(), which calls unpin_extent_range() for
    the deleted extent. It sees last_byte_to_unpin set to U64_MAX (by
    transaction B!), so it adds the deleted extent to the space cache
    again!

This explains all of our symptoms above:

* If the sequence of events is exactly as described above, when the free
  space is re-added in step 11, it will fail with EEXIST.
* If another thread reallocates the deleted extent in between steps 7
  and 11, then step 11 will silently re-add that space to the space
  cache as free even though it is actually allocated. Then, if that
  space is allocated *again*, the free space tree will be corrupted
  (namely, the wrong item will be deleted).
* If we don't catch this free space tree corruption, it will continue
  to get worse as extents are deleted and reallocated.

The v1 space_cache is synchronously loaded when an extent is deleted
(btrfs_update_block_group() with alloc=0 calls btrfs_cache_block_group()
with load_cache_only=1), so it is not normally affected by this bug.
However, as noted above, if we fail to load the space cache, we will
fall back to caching from the extent tree and may hit this bug.

The easiest fix for this race is to also make caching from the free
space tree or extent tree synchronous. Josef tested this and found no
performance regressions.

A few extra changes fall out of this change. Namely, this fix does the
following, with step 2 being the crucial fix:

1. Factor btrfs_caching_ctl_wait_done() out of
   btrfs_wait_block_group_cache_done() to allow waiting on a caching_ctl
   that we already hold a reference to.
2. Change the call in btrfs_cache_block_group() of
   btrfs_wait_space_cache_v1_finished() to
   btrfs_caching_ctl_wait_done(), which makes us wait regardless of the
   space_cache option.
3. Delete the now unused btrfs_wait_space_cache_v1_finished() and
   space_cache_v1_done().
4. Change btrfs_cache_block_group()'s `int load_cache_only` parameter to
   `bool wait` to more accurately describe its new meaning.
5. Change a few callers which had a separate call to
   btrfs_wait_block_group_cache_done() to use wait = true instead.
6. Make btrfs_wait_block_group_cache_done() static now that it's not
   used outside of block-group.c anymore.

Fixes: d0c2f4fa555e ("btrfs: make concurrent fsyncs wait less when waiting for a transaction commit")
CC: stable@vger.kernel.org # 5.12+
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agokprobes: don't call disarm_kprobe() for disabled kprobes
Kuniyuki Iwashima [Sat, 13 Aug 2022 02:05:09 +0000 (19:05 -0700)]
kprobes: don't call disarm_kprobe() for disabled kprobes

commit 9c80e79906b4ca440d09e7f116609262bb747909 upstream.

The assumption in __disable_kprobe() is wrong, and it could try to disarm
an already disarmed kprobe and fire the WARN_ONCE() below. [0]  We can
easily reproduce this issue.

1. Write 0 to /sys/kernel/debug/kprobes/enabled.

  # echo 0 > /sys/kernel/debug/kprobes/enabled

2. Run execsnoop.  At this time, one kprobe is disabled.

  # /usr/share/bcc/tools/execsnoop &
  [1] 2460
  PCOMM            PID    PPID   RET ARGS

  # cat /sys/kernel/debug/kprobes/list
  ffffffff91345650  r  __x64_sys_execve+0x0    [FTRACE]
  ffffffff91345650  k  __x64_sys_execve+0x0    [DISABLED][FTRACE]

3. Write 1 to /sys/kernel/debug/kprobes/enabled, which changes
   kprobes_all_disarmed to false but does not arm the disabled kprobe.

  # echo 1 > /sys/kernel/debug/kprobes/enabled

  # cat /sys/kernel/debug/kprobes/list
  ffffffff91345650  r  __x64_sys_execve+0x0    [FTRACE]
  ffffffff91345650  k  __x64_sys_execve+0x0    [DISABLED][FTRACE]

4. Kill execsnoop, when __disable_kprobe() calls disarm_kprobe() for the
   disabled kprobe and hits the WARN_ONCE() in __disarm_kprobe_ftrace().

  # fg
  /usr/share/bcc/tools/execsnoop
  ^C

Actually, WARN_ONCE() is fired twice, and __unregister_kprobe_top() misses
some cleanups and leaves the aggregated kprobe in the hash table.  Then,
__unregister_trace_kprobe() initialises tk->rp.kp.list and creates an
infinite loop like this.

  aggregated kprobe.list -> kprobe.list -.
                                     ^    |
                                     '.__.'

In this situation, these commands fall into the infinite loop and result
in RCU stall or soft lockup.

  cat /sys/kernel/debug/kprobes/list : show_kprobe_addr() enters into the
                                       infinite loop with RCU.

  /usr/share/bcc/tools/execsnoop : warn_kprobe_rereg() holds kprobe_mutex,
                                   and __get_valid_kprobe() is stuck in
   the loop.

To avoid the issue, make sure we don't call disarm_kprobe() for disabled
kprobes.

[0]
Failed to disarm kprobe-ftrace at __x64_sys_execve+0x0/0x40 (error -2)
WARNING: CPU: 6 PID: 2460 at kernel/kprobes.c:1130 __disarm_kprobe_ftrace.isra.19 (kernel/kprobes.c:1129)
Modules linked in: ena
CPU: 6 PID: 2460 Comm: execsnoop Not tainted 5.19.0+ #28
Hardware name: Amazon EC2 c5.2xlarge/, BIOS 1.0 10/16/2017
RIP: 0010:__disarm_kprobe_ftrace.isra.19 (kernel/kprobes.c:1129)
Code: 24 8b 02 eb c1 80 3d c4 83 f2 01 00 75 d4 48 8b 75 00 89 c2 48 c7 c7 90 fa 0f 92 89 04 24 c6 05 ab 83 01 e8 e4 94 f0 ff <0f> 0b 8b 04 24 eb b1 89 c6 48 c7 c7 60 fa 0f 92 89 04 24 e8 cc 94
RSP: 0018:ffff9e6ec154bd98 EFLAGS: 00010282
RAX: 0000000000000000 RBX: ffffffff930f7b00 RCX: 0000000000000001
RDX: 0000000080000001 RSI: ffffffff921461c5 RDI: 00000000ffffffff
RBP: ffff89c504286da8 R08: 0000000000000000 R09: c0000000fffeffff
R10: 0000000000000000 R11: ffff9e6ec154bc28 R12: ffff89c502394e40
R13: ffff89c502394c00 R14: ffff9e6ec154bc00 R15: 0000000000000000
FS:  00007fe800398740(0000) GS:ffff89c812d80000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 000000c00057f010 CR3: 0000000103b54006 CR4: 00000000007706e0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
PKRU: 55555554
Call Trace:
<TASK>
 __disable_kprobe (kernel/kprobes.c:1716)
 disable_kprobe (kernel/kprobes.c:2392)
 __disable_trace_kprobe (kernel/trace/trace_kprobe.c:340)
 disable_trace_kprobe (kernel/trace/trace_kprobe.c:429)
 perf_trace_event_unreg.isra.2 (./include/linux/tracepoint.h:93 kernel/trace/trace_event_perf.c:168)
 perf_kprobe_destroy (kernel/trace/trace_event_perf.c:295)
 _free_event (kernel/events/core.c:4971)
 perf_event_release_kernel (kernel/events/core.c:5176)
 perf_release (kernel/events/core.c:5186)
 __fput (fs/file_table.c:321)
 task_work_run (./include/linux/sched.h:2056 (discriminator 1) kernel/task_work.c:179 (discriminator 1))
 exit_to_user_mode_prepare (./include/linux/resume_user_mode.h:49 kernel/entry/common.c:169 kernel/entry/common.c:201)
 syscall_exit_to_user_mode (./arch/x86/include/asm/jump_label.h:55 ./arch/x86/include/asm/nospec-branch.h:384 ./arch/x86/include/asm/entry-common.h:94 kernel/entry/common.c:133 kernel/entry/common.c:296)
 do_syscall_64 (arch/x86/entry/common.c:87)
 entry_SYSCALL_64_after_hwframe (arch/x86/entry/entry_64.S:120)
RIP: 0033:0x7fe7ff210654
Code: 15 79 89 20 00 f7 d8 64 89 02 48 c7 c0 ff ff ff ff eb be 0f 1f 00 8b 05 9a cd 20 00 48 63 ff 85 c0 75 11 b8 03 00 00 00 0f 05 <48> 3d 00 f0 ff ff 77 3a f3 c3 48 83 ec 18 48 89 7c 24 08 e8 34 fc
RSP: 002b:00007ffdbd1d3538 EFLAGS: 00000246 ORIG_RAX: 0000000000000003
RAX: 0000000000000000 RBX: 0000000000000008 RCX: 00007fe7ff210654
RDX: 0000000000000000 RSI: 0000000000002401 RDI: 0000000000000008
RBP: 0000000000000000 R08: 94ae31d6fda838a4 R0900007fe8001c9d30
R10: 00007ffdbd1d34b0 R11: 0000000000000246 R12: 00007ffdbd1d3600
R13: 0000000000000000 R14: fffffffffffffffc R15: 00007ffdbd1d3560
</TASK>

Link: https://lkml.kernel.org/r/20220813020509.90805-1-kuniyu@amazon.com
Fixes: 69d54b916d83 ("kprobes: makes kprobes/enabled works correctly for optimized kprobes.")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reported-by: Ayushman Dutta <ayudutta@amazon.com>
Cc: "Naveen N. Rao" <naveen.n.rao@linux.ibm.com>
Cc: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Wang Nan <wangnan0@huawei.com>
Cc: Kuniyuki Iwashima <kuniyu@amazon.com>
Cc: Kuniyuki Iwashima <kuni1840@gmail.com>
Cc: Ayushman Dutta <ayudutta@amazon.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2 years agobtrfs: tree-checker: check for overlapping extent items
Josef Bacik [Wed, 3 Aug 2022 18:28:47 +0000 (14:28 -0400)]
btrfs: tree-checker: check for overlapping extent items

[ Upstream commit 899b7f69f244e539ea5df1b4d756046337de44a5 ]

We're seeing a weird problem in production where we have overlapping
extent items in the extent tree.  It's unclear where these are coming
from, and in debugging we realized there's no check in the tree checker
for this sort of problem.  Add a check to the tree-checker to make sure
that the extents do not overlap each other.

Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
2 years agobtrfs: fix lockdep splat with reloc root extent buffers
Josef Bacik [Tue, 26 Jul 2022 20:24:04 +0000 (16:24 -0400)]
btrfs: fix lockdep splat with reloc root extent buffers

[ Upstream commit b40130b23ca4a08c5785d5a3559805916bddba3c ]

We have been hitting the following lockdep splat with btrfs/187 recently

  WARNING: possible circular locking dependency detected
  5.19.0-rc8+ #775 Not tainted
  ------------------------------------------------------
  btrfs/752500 is trying to acquire lock:
  ffff97e1875a97b8 (btrfs-treloc-02#2){+.+.}-{3:3}, at: __btrfs_tree_lock+0x24/0x110

  but task is already holding lock:
  ffff97e1875a9278 (btrfs-tree-01/1){+.+.}-{3:3}, at: __btrfs_tree_lock+0x24/0x110

  which lock already depends on the new lock.

  the existing dependency chain (in reverse order) is:

  -> #2 (btrfs-tree-01/1){+.+.}-{3:3}:
 down_write_nested+0x41/0x80
 __btrfs_tree_lock+0x24/0x110
 btrfs_init_new_buffer+0x7d/0x2c0
 btrfs_alloc_tree_block+0x120/0x3b0
 __btrfs_cow_block+0x136/0x600
 btrfs_cow_block+0x10b/0x230
 btrfs_search_slot+0x53b/0xb70
 btrfs_lookup_inode+0x2a/0xa0
 __btrfs_update_delayed_inode+0x5f/0x280
 btrfs_async_run_delayed_root+0x24c/0x290
 btrfs_work_helper+0xf2/0x3e0
 process_one_work+0x271/0x590
 worker_thread+0x52/0x3b0
 kthread+0xf0/0x120
 ret_from_fork+0x1f/0x30

  -> #1 (btrfs-tree-01){++++}-{3:3}:
 down_write_nested+0x41/0x80
 __btrfs_tree_lock+0x24/0x110
 btrfs_search_slot+0x3c3/0xb70
 do_relocation+0x10c/0x6b0
 relocate_tree_blocks+0x317/0x6d0
 relocate_block_group+0x1f1/0x560
 btrfs_relocate_block_group+0x23e/0x400
 btrfs_relocate_chunk+0x4c/0x140
 btrfs_balance+0x755/0xe40
 btrfs_ioctl+0x1ea2/0x2c90
 __x64_sys_ioctl+0x88/0xc0
 do_syscall_64+0x38/0x90
 entry_SYSCALL_64_after_hwframe+0x63/0xcd

  -> #0 (btrfs-treloc-02#2){+.+.}-{3:3}:
 __lock_acquire+0x1122/0x1e10
 lock_acquire+0xc2/0x2d0
 down_write_nested+0x41/0x80
 __btrfs_tree_lock+0x24/0x110
 btrfs_lock_root_node+0x31/0x50
 btrfs_search_slot+0x1cb/0xb70
 replace_path+0x541/0x9f0
 merge_reloc_root+0x1d6/0x610
 merge_reloc_roots+0xe2/0x260
 relocate_block_group+0x2c8/0x560
 btrfs_relocate_block_group+0x23e/0x400
 btrfs_relocate_chunk+0x4c/0x140
 btrfs_balance+0x755/0xe40
 btrfs_ioctl+0x1ea2/0x2c90
 __x64_sys_ioctl+0x88/0xc0
 do_syscall_64+0x38/0x90
 entry_SYSCALL_64_after_hwframe+0x63/0xcd

  other info that might help us debug this:

  Chain exists of:
    btrfs-treloc-02#2 --> btrfs-tree-01 --> btrfs-tree-01/1

   Possible unsafe locking scenario:

 CPU0                    CPU1
 ----                    ----
    lock(btrfs-tree-01/1);
 lock(btrfs-tree-01);
 lock(btrfs-tree-01/1);
    lock(btrfs-treloc-02#2);

   *** DEADLOCK ***

  7 locks held by btrfs/752500:
   #0: ffff97e292fdf460 (sb_writers#12){.+.+}-{0:0}, at: btrfs_ioctl+0x208/0x2c90
   #1: ffff97e284c02050 (&fs_info->reclaim_bgs_lock){+.+.}-{3:3}, at: btrfs_balance+0x55f/0xe40
   #2: ffff97e284c00878 (&fs_info->cleaner_mutex){+.+.}-{3:3}, at: btrfs_relocate_block_group+0x236/0x400
   #3: ffff97e292fdf650 (sb_internal#2){.+.+}-{0:0}, at: merge_reloc_root+0xef/0x610
   #4: ffff97e284c02378 (btrfs_trans_num_writers){++++}-{0:0}, at: join_transaction+0x1a8/0x5a0
   #5: ffff97e284c023a0 (btrfs_trans_num_extwriters){++++}-{0:0}, at: join_transaction+0x1a8/0x5a0
   #6: ffff97e1875a9278 (btrfs-tree-01/1){+.+.}-{3:3}, at: __btrfs_tree_lock+0x24/0x110

  stack backtrace:
  CPU: 1 PID: 752500 Comm: btrfs Not tainted 5.19.0-rc8+ #775
  Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.13.0-2.fc32 04/01/2014
  Call Trace:

   dump_stack_lvl+0x56/0x73
   check_noncircular+0xd6/0x100
   ? lock_is_held_type+0xe2/0x140
   __lock_acquire+0x1122/0x1e10
   lock_acquire+0xc2/0x2d0
   ? __btrfs_tree_lock+0x24/0x110
   down_write_nested+0x41/0x80
   ? __btrfs_tree_lock+0x24/0x110
   __btrfs_tree_lock+0x24/0x110
   btrfs_lock_root_node+0x31/0x50
   btrfs_search_slot+0x1cb/0xb70
   ? lock_release+0x137/0x2d0
   ? _raw_spin_unlock+0x29/0x50
   ? release_extent_buffer+0x128/0x180
   replace_path+0x541/0x9f0
   merge_reloc_root+0x1d6/0x610
   merge_reloc_roots+0xe2/0x260
   relocate_block_group+0x2c8/0x560
   btrfs_relocate_block_group+0x23e/0x400
   btrfs_relocate_chunk+0x4c/0x140
   btrfs_balance+0x755/0xe40
   btrfs_ioctl+0x1ea2/0x2c90
   ? lock_is_held_type+0xe2/0x140
   ? lock_is_held_type+0xe2/0x140
   ? __x64_sys_ioctl+0x88/0xc0
   __x64_sys_ioctl+0x88/0xc0
   do_syscall_64+0x38/0x90
   entry_SYSCALL_64_after_hwframe+0x63/0xcd

This isn't necessarily new, it's just tricky to hit in practice.  There
are two competing things going on here.  With relocation we create a
snapshot of every fs tree with a reloc tree.  Any extent buffers that
get initialized here are initialized with the reloc root lockdep key.
However since it is a snapshot, any blocks that are currently in cache
that originally belonged to the fs tree will have the normal tree
lockdep key set.  This creates the lock dependency of

  reloc tree -> normal tree

for the extent buffer locking during the first phase of the relocation
as we walk down the reloc root to relocate blocks.

However this is problematic because the final phase of the relocation is
merging the reloc root into the original fs root.  This involves
searching down to any keys that exist in the original fs root and then
swapping the relocated block and the original fs root block.  We have to
search down to the fs root first, and then go search the reloc root for
the block we need to replace.  This creates the dependency of

  normal tree -> reloc tree

which is why lockdep complains.

Additionally even if we were to fix this particular mismatch with a
different nesting for the merge case, we're still slotting in a block
that has a owner of the reloc root objectid into a normal tree, so that
block will have its lockdep key set to the tree reloc root, and create a
lockdep splat later on when we wander into that block from the fs root.

Unfortunately the only solution here is to make sure we do not set the
lockdep key to the reloc tree lockdep key normally, and then reset any
blocks we wander into from the reloc root when we're doing the merged.

This solves the problem of having mixed tree reloc keys intermixed with
normal tree keys, and then allows us to make sure in the merge case we
maintain the lock order of

  normal tree -> reloc tree

We handle this by setting a bit on the reloc root when we do the search
for the block we want to relocate, and any block we search into or COW
at that point gets set to the reloc tree key.  This works correctly
because we only ever COW down to the parent node, so we aren't resetting
the key for the block we're linking into the fs root.

With this patch we no longer have the lockdep splat in btrfs/187.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
2 years agobtrfs: move lockdep class helpers to locking.c
Josef Bacik [Tue, 26 Jul 2022 20:24:03 +0000 (16:24 -0400)]
btrfs: move lockdep class helpers to locking.c

[ Upstream commit 0a27a0474d146eb79e09ec88bf0d4229f4cfc1b8 ]

These definitions exist in disk-io.c, which is not related to the
locking.  Move this over to locking.h/c where it makes more sense.

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>