Re: [PATCH] ACPI / video: driver must be registered before checking for keypresses

From: Hans de Goede
Date: Thu Jan 14 2016 - 03:00:24 EST

Hi All,

On 05-01-16 12:43, Hans de Goede wrote:
Hi,

On 04-01-16 23:22, Adrien Schildknecht wrote:
acpi_video_handles_brightness_key_presses() may use an uninitialized mutex.
The error has been reported by lockdep: DEBUG_LOCKS_WARN_ON(l->magic != l).
The function assumes that the video driver has been registered before being
called. As explained in the comment of acpi_video_init(), the registration
of the video class may be defered and thus may not take place in the init
function of the module.

Use completion mechanisms to make sure that
acpi_video_handles_brightness_key_presses() wait for the completion of
acpi_video_register() before using the mutex.
Also get rid of register_count since task completion can replace it.

Signed-off-by: Adrien Schildknecht <adrien+dev@xxxxxxxxxxx>

Thanks for the fix, my first instinct was that there should be an easier
fix, but thinking more about it this and how this function is used
in thinkpad_acpi. this is indeed the correct way to fix this:

Reviewed-by: Hans de Goede <hdegoede@xxxxxxxxxx>

Rafael, can you please add this fix to acpi-next ?

This morning my mind wandered back to this patch, and I've one worry about it:

---
drivers/acpi/acpi_video.c | 27 ++++++++++++++-------------
1 file changed, 14 insertions(+), 13 deletions(-)

diff --git a/drivers/acpi/acpi_video.c b/drivers/acpi/acpi_video.c
index 80b13d4..06a006f 100644
--- a/drivers/acpi/acpi_video.c
+++ b/drivers/acpi/acpi_video.c
@@ -90,8 +90,8 @@ module_param(device_id_scheme, bool, 0444);
static bool only_lcd = false;
module_param(only_lcd, bool, 0444);

-static int register_count;
-static DEFINE_MUTEX(register_count_mutex);
+static DECLARE_COMPLETION(register_done);
+static DEFINE_MUTEX(register_done_mutex);
static struct mutex video_list_lock;
static struct list_head video_bus_head;
static int acpi_video_bus_add(struct acpi_device *device);
@@ -2049,8 +2049,8 @@ int acpi_video_register(void)
{
int ret = 0;

- mutex_lock(&register_count_mutex);
- if (register_count) {
+ mutex_lock(&register_done_mutex);
+ if (completion_done(&register_done)) {
/*
* if the function of acpi_video_register is already called,
* don't register the acpi_vide_bus again and return no error.
@@ -2071,22 +2071,22 @@ int acpi_video_register(void)
* When the acpi_video_bus is loaded successfully, increase
* the counter reference.
*/
- register_count = 1;
+ complete(&register_done);

leave:
- mutex_unlock(&register_count_mutex);
+ mutex_unlock(&register_done_mutex);
return ret;
}
EXPORT_SYMBOL(acpi_video_register);

void acpi_video_unregister(void)
{
- mutex_lock(&register_count_mutex);
- if (register_count) {
+ mutex_lock(&register_done_mutex);
+ if (completion_done(&register_done)) {
acpi_bus_unregister_driver(&acpi_video_bus);
- register_count = 0;
+ reinit_completion(&register_done);
}
- mutex_unlock(&register_count_mutex);
+ mutex_unlock(&register_done_mutex);
}
EXPORT_SYMBOL(acpi_video_unregister);

@@ -2094,20 +2094,21 @@ void acpi_video_unregister_backlight(void)
{
struct acpi_video_bus *video;

- mutex_lock(&register_count_mutex);
- if (register_count) {
+ mutex_lock(&register_done_mutex);
+ if (completion_done(&register_done)) {
mutex_lock(&video_list_lock);
list_for_each_entry(video, &video_bus_head, entry)
acpi_video_bus_unregister_backlight(video);
mutex_unlock(&video_list_lock);
}
- mutex_unlock(&register_count_mutex);
+ mutex_unlock(&register_done_mutex);
}

bool acpi_video_handles_brightness_key_presses(void)
{
bool have_video_busses;

+ wait_for_completion(&register_done);
mutex_lock(&video_list_lock);
have_video_busses = !list_empty(&video_bus_head);
mutex_unlock(&video_list_lock);


What if register_done never completes? There are 2 scenarios where
this can happen:

a) The machine has an intel video bios opcode region, but the i915 driver
never loads

b) The i915 driver gets unloaded


b. We can fix by switching back to using register_count to check if
the driver is registered, adding the completion on top of register_count
rather then replacing it, and only checking the completion in
acpi_video_handles_brightness_key_presses(), dropping the reinit_completion()
from acpi_video_unregister().

This means that after a rmmod of the i915 driver
acpi_video_handles_brightness_key_presses() will always return false,
which is fine as after a rmmod of the i915 driver acpi-video is indeed
neverhandling brightness key presses.

a. I find more worry some, this means that if for some reason the i915
driver is not being loaded callers of acpi_video_handles_brightness_key_presses()
may get stuck indefinitely. Which IMHO is unacceptable.

I believe that the best way to fix this is to:

1) Revert this patch
2) Fix the original bug by doing:
--- a/drivers/acpi/acpi_video.c
+++ b/drivers/acpi/acpi_video.c
@@ -92,8 +92,8 @@ module_param(only_lcd, bool, 0444);

static int register_count;
static DEFINE_MUTEX(register_count_mutex);
-static struct mutex video_list_lock;
-static struct list_head video_bus_head;
+static DEFINE_MUTEX(video_list_lock);
+static LIST_HEAD(video_bus_head);
static int acpi_video_bus_add(struct acpi_device *device);
static int acpi_video_bus_remove(struct acpi_device *device);
static void acpi_video_bus_notify(struct acpi_device *device, u32 event);
3) Document that acpi_video_handles_brightness_key_presses()
return value may change over time and should not be cached
4) Fix thinkpad_acpi.c for 3.

Note that for the current kernel cycle we can replace 4 with
reverting the patch which switches thinkpad_apci over to using
acpi_video_handles_brightness_key_presses(), because that is
only a cleanup really and does not fix any bugs (*).

I will prepare a patch-set doing that, as the problem I've outlined
above as "a." is unacceptable IMHO.

Regards,

Hans



*) Unlike the same change in dell-wmi.c, which does fix an actual bug