fbcon: Do not takeover the console from atomic context

Taking over the console involves allocating mem with GFP_KERNEL, talking
to drm drivers, etc. So this should not be done from an atomic context.

But the console-output trigger deferred console takeover may happen from an
atomic context, which leads to "BUG: sleeping function called from invalid
context" errors.

This commit fixes these errors by doing the deferred takeover from a
workqueue when the notifier runs from an atomic context.

Signed-off-by: Hans de Goede <redacted>
---
 drivers/video/fbdev/core/fbcon.c | 21 +++++++++++++++++++--
 1 file changed, 19 insertions(+), 2 deletions(-)

diff --git a/drivers/video/fbdev/core/fbcon.c b/drivers/video/fbdev/core/fbcon.c
index ef8b2d0b7071..4e5997d53fc4 100644
--- a/drivers/video/fbdev/core/fbcon.c
+++ b/drivers/video/fbdev/core/fbcon.c

@@ -3592,7 +3592,20 @@ static int fbcon_init_device(void)
 }
 
 #ifdef CONFIG_FRAMEBUFFER_CONSOLE_DEFERRED_TAKEOVER
+static void fbcon_register_existing_fbs(struct work_struct *work)
+{
+	int i;
+
+	console_lock();
+
+	for_each_registered_fb(i)
+		fbcon_fb_registered(registered_fb[i]);
+
+	console_unlock();
+}
+
 static struct notifier_block fbcon_output_nb;
+static DECLARE_WORK(fbcon_deferred_takeover_work, fbcon_register_existing_fbs);
 
 static int fbcon_output_notifier(struct notifier_block *nb,
 				 unsigned long action, void *data)

@@ -3607,8 +3620,12 @@ static int fbcon_output_notifier(struct notifier_block *nb,
 	deferred_takeover = false;
 	logo_shown = FBCON_LOGO_DONTSHOW;
 
-	for_each_registered_fb(i)
-		fbcon_fb_registered(registered_fb[i]);
+	if (in_atomic() || irqs_disabled()) {
+		schedule_work(&fbcon_deferred_takeover_work);
+	} else {
+		for_each_registered_fb(i)
+			fbcon_fb_registered(registered_fb[i]);
+	}
 
 	return NOTIFY_OK;
 }

-- 
2.18.0

On Monday, August 06, 2018 05:54:16 PM Hans de Goede wrote:

Taking over the console involves allocating mem with GFP_KERNEL, talking
to drm drivers, etc. So this should not be done from an atomic context.

But the console-output trigger deferred console takeover may happen from an
atomic context, which leads to "BUG: sleeping function called from invalid
context" errors.

This commit fixes these errors by doing the deferred takeover from a
workqueue when the notifier runs from an atomic context.

Signed-off-by: Hans de Goede <redacted>

checkpatch.pl complains about in_atomic use:

ERROR: do not use in_atomic in drivers
#51: FILE: drivers/video/fbdev/core/fbcon.c:3623:
+       if (in_atomic() || irqs_disabled()) {

please also see the comment in preempt.h:

/*
 * Are we running in atomic context?  WARNING: this macro cannot
 * always detect atomic context; in particular, it cannot know about
 * held spinlocks in non-preemptible kernels.  Thus it should not be
 * used in the general case to determine whether sleeping is possible.
 * Do not use in_atomic() in driver code.
 */
#define in_atomic()	(preempt_count() != 0)

Therefore please explain why it is fine to use in_atomic in fbcon's case.

4 more quoted lines

Best regards,
--
Bartlomiej Zolnierkiewicz
Samsung R&D Institute Poland
Samsung Electronics

Hi,

On 09-08-18 12:03, Bartlomiej Zolnierkiewicz wrote:

On Monday, August 06, 2018 05:54:16 PM Hans de Goede wrote:

quoted

Taking over the console involves allocating mem with GFP_KERNEL, talking
to drm drivers, etc. So this should not be done from an atomic context.

But the console-output trigger deferred console takeover may happen from an
atomic context, which leads to "BUG: sleeping function called from invalid
context" errors.

This commit fixes these errors by doing the deferred takeover from a
workqueue when the notifier runs from an atomic context.

Signed-off-by: Hans de Goede <redacted>

checkpatch.pl complains about in_atomic use:

ERROR: do not use in_atomic in drivers
#51: FILE: drivers/video/fbdev/core/fbcon.c:3623:
+       if (in_atomic() || irqs_disabled()) {

please also see the comment in preempt.h:

/*
  * Are we running in atomic context?  WARNING: this macro cannot
  * always detect atomic context; in particular, it cannot know about
  * held spinlocks in non-preemptible kernels.  Thus it should not be
  * used in the general case to determine whether sleeping is possible.
  * Do not use in_atomic() in driver code.
  */
#define in_atomic()	(preempt_count() != 0)

Therefore please explain why it is fine to use in_atomic in fbcon's case.

Ok I will send a v2 which adds a comment why we need to use in_atomic
here.

Regards,

Hans

quoted

---
  drivers/video/fbdev/core/fbcon.c | 21 +++++++++++++++++++--
  1 file changed, 19 insertions(+), 2 deletions(-)

diff --git a/drivers/video/fbdev/core/fbcon.c b/drivers/video/fbdev/core/fbcon.c
index ef8b2d0b7071..4e5997d53fc4 100644
--- a/drivers/video/fbdev/core/fbcon.c
+++ b/drivers/video/fbdev/core/fbcon.c

@@ -3592,7 +3592,20 @@ static int fbcon_init_device(void)
  }
  
  #ifdef CONFIG_FRAMEBUFFER_CONSOLE_DEFERRED_TAKEOVER
+static void fbcon_register_existing_fbs(struct work_struct *work)
+{
+	int i;
+
+	console_lock();
+
+	for_each_registered_fb(i)
+		fbcon_fb_registered(registered_fb[i]);
+
+	console_unlock();
+}
+
  static struct notifier_block fbcon_output_nb;
+static DECLARE_WORK(fbcon_deferred_takeover_work, fbcon_register_existing_fbs);
  
  static int fbcon_output_notifier(struct notifier_block *nb,
  				 unsigned long action, void *data)

@@ -3607,8 +3620,12 @@ static int fbcon_output_notifier(struct notifier_block *nb,
  	deferred_takeover = false;
  	logo_shown = FBCON_LOGO_DONTSHOW;
  
-	for_each_registered_fb(i)
-		fbcon_fb_registered(registered_fb[i]);
+	if (in_atomic() || irqs_disabled()) {
+		schedule_work(&fbcon_deferred_takeover_work);
+	} else {
+		for_each_registered_fb(i)

12 more quoted lines