Re: Possible deadlock when suspending framebuffer

Paul, fbdev people.. Comments? This was sent to me and lkml, the right
people probably didn't see it.

I doubt it's a big problem in practice, but..

                      Linus

On Tue, Jun 14, 2011 at 6:10 AM, Francis Moreau [off-list ref] wrote:

Hello,

I noticed that a possible deadlock can happen when the current frame
buffering is being suspended and a new frambuffer device is being
registred at the same time.

When suspending the current frambuffer by doing : echo 1

quoted

/sys/class/graphics/fb0/state, the kernel actually takes the

following locks in that order: console_lock, lock_fb_info (see
store_fbstate()).

However when a new framebuffer is coming in, the lock sequence is:
lock_fb_info (taken by do_remove_conflicting_framebuffer()),
console_lock() (taken by unbind_console).

I don't know how this should be fixed though...

Thanks
--
Francis

Hi Linus,

On 06/14/2011 06:15 PM, Linus Torvalds wrote:

Paul, fbdev people.. Comments? This was sent to me and lkml, the right
people probably didn't see it.

Sounds very familiar. Indeed a quick glance at my archive revealed 2 
approaches/patches dealing with similar issues
http://marc.info/?l=linux-fbdev&m9539210207429&w=2
http://marc.info/?l=linux-fbdev&m9700789632450&w=2

I did not have time to verify that those do actually get those things right 
everywhere but the first one looks good and should also solve this issue I think.


Regards,

Florian Tobias Schandinat

I doubt it's a big problem in practice, but..

                       Linus

On Tue, Jun 14, 2011 at 6:10 AM, Francis Moreau[off-list ref]  wrote:

quoted

Hello,

I noticed that a possible deadlock can happen when the current frame
buffering is being suspended and a new frambuffer device is being
registred at the same time.

When suspending the current frambuffer by doing : echo 1

quoted

/sys/class/graphics/fb0/state, the kernel actually takes the

following locks in that order: console_lock, lock_fb_info (see
store_fbstate()).

However when a new framebuffer is coming in, the lock sequence is:
lock_fb_info (taken by do_remove_conflicting_framebuffer()),
console_lock() (taken by unbind_console).

I don't know how this should be fixed though...

Thanks
--
Francis

--
To unsubscribe from this list: send the line "unsubscribe linux-fbdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

From: Herton Ronaldo Krzesinski <redacted>

A lock ordering issue can cause deadlocks: in framebuffer/console code,
all needed struct fb_info locks are taken before acquire_console_sem(),
in places which need to take console semaphore.

But fb_set_suspend is always called with console semaphore held, and
inside it we call lock_fb_info which gets the fb_info lock, inverse
locking order of what the rest of the code does. This causes a real
deadlock issue, when we write to state fb sysfs attribute (which calls
fb_set_suspend) while a framebuffer is being unregistered by
remove_conflicting_framebuffers, as can be shown by following show
blocked state trace on a test program which loads i915 and runs another
forked processes writing to state attribute:

Test process with semaphore held and trying to get fb_info lock:
..
fb-test2      D 0000000000000000     0   237    228 0x00000000
 ffff8800774f3d68 0000000000000082 00000000000135c0 00000000000135c0
 ffff880000000000 ffff8800774f3fd8 ffff8800774f3fd8 ffff880076ee4530
 00000000000135c0 ffff8800774f3fd8 ffff8800774f2000 00000000000135c0
Call Trace:
 [<ffffffff8141287a>] __mutex_lock_slowpath+0x11a/0x1e0
 [<ffffffff814142f2>] ? _raw_spin_lock_irq+0x22/0x40
 [<ffffffff814123d3>] mutex_lock+0x23/0x50
 [<ffffffff8125dfc5>] lock_fb_info+0x25/0x60
 [<ffffffff8125e3f0>] fb_set_suspend+0x20/0x80
 [<ffffffff81263e2f>] store_fbstate+0x4f/0x70
 [<ffffffff812e7f70>] dev_attr_store+0x20/0x30
 [<ffffffff811c46b4>] sysfs_write_file+0xd4/0x160
 [<ffffffff81155a26>] vfs_write+0xc6/0x190
 [<ffffffff81155d51>] sys_write+0x51/0x90
 [<ffffffff8100c012>] system_call_fastpath+0x16/0x1b
..
modprobe process stalled because has the fb_info lock (got inside
unregister_framebuffer) but waiting for the semaphore held by the
test process which is waiting to get the fb_info lock:
..
modprobe      D 0000000000000000     0   230    218 0x00000000
 ffff880077a4d618 0000000000000082 0000000000000001 0000000000000001
 ffff880000000000 ffff880077a4dfd8 ffff880077a4dfd8 ffff8800775a2e20
 00000000000135c0 ffff880077a4dfd8 ffff880077a4c000 00000000000135c0
Call Trace:
 [<ffffffff81411fe5>] schedule_timeout+0x215/0x310
 [<ffffffff81058051>] ? get_parent_ip+0x11/0x50
 [<ffffffff814130dd>] __down+0x6d/0xb0
 [<ffffffff81089f71>] down+0x41/0x50
 [<ffffffff810629ac>] acquire_console_sem+0x2c/0x50
 [<ffffffff812ca53d>] unbind_con_driver+0xad/0x2d0
 [<ffffffff8126f5f7>] fbcon_event_notify+0x457/0x890
 [<ffffffff814144ff>] ? _raw_spin_unlock_irqrestore+0x1f/0x50
 [<ffffffff81058051>] ? get_parent_ip+0x11/0x50
 [<ffffffff8141836d>] notifier_call_chain+0x4d/0x70
 [<ffffffff8108a3b8>] __blocking_notifier_call_chain+0x58/0x80
 [<ffffffff8108a3f6>] blocking_notifier_call_chain+0x16/0x20
 [<ffffffff8125dabb>] fb_notifier_call_chain+0x1b/0x20
 [<ffffffff8125e6ac>] unregister_framebuffer+0x7c/0x130
 [<ffffffff8125e8b3>] remove_conflicting_framebuffers+0x153/0x180
 [<ffffffff8125eef3>] register_framebuffer+0x93/0x2c0
 [<ffffffffa0331112>] drm_fb_helper_single_fb_probe+0x252/0x2f0 [drm_kms_helper]
 [<ffffffffa03314a3>] drm_fb_helper_initial_config+0x2f3/0x6d0 [drm_kms_helper]
 [<ffffffffa03318dd>] ? drm_fb_helper_single_add_all_connectors+0x5d/0x1c0 [drm_kms_helper]
 [<ffffffffa037b588>] intel_fbdev_init+0xa8/0x160 [i915]
 [<ffffffffa0343d74>] i915_driver_load+0x854/0x12b0 [i915]
 [<ffffffffa02f0e7e>] drm_get_pci_dev+0x19e/0x360 [drm]
 [<ffffffff8141821d>] ? sub_preempt_count+0x9d/0xd0
 [<ffffffffa0386f91>] i915_pci_probe+0x15/0x17 [i915]
 [<ffffffff8124481f>] local_pci_probe+0x5f/0xd0
 [<ffffffff81244f89>] pci_device_probe+0x119/0x120
 [<ffffffff812eccaa>] ? driver_sysfs_add+0x7a/0xb0
 [<ffffffff812ed003>] driver_probe_device+0xa3/0x290
 [<ffffffff812ed1f0>] ? __driver_attach+0x0/0xb0
 [<ffffffff812ed29b>] __driver_attach+0xab/0xb0
 [<ffffffff812ed1f0>] ? __driver_attach+0x0/0xb0
 [<ffffffff812ebd3e>] bus_for_each_dev+0x5e/0x90
 [<ffffffff812ecc2e>] driver_attach+0x1e/0x20
 [<ffffffff812ec6f2>] bus_add_driver+0xe2/0x320
 [<ffffffffa03aa000>] ? i915_init+0x0/0x96 [i915]
 [<ffffffff812ed536>] driver_register+0x76/0x140
 [<ffffffffa03aa000>] ? i915_init+0x0/0x96 [i915]
 [<ffffffff81245216>] __pci_register_driver+0x56/0xd0
 [<ffffffffa02f1264>] drm_pci_init+0xe4/0xf0 [drm]
 [<ffffffffa03aa000>] ? i915_init+0x0/0x96 [i915]
 [<ffffffffa02e84a8>] drm_init+0x58/0x70 [drm]
 [<ffffffffa03aa094>] i915_init+0x94/0x96 [i915]
 [<ffffffff81002194>] do_one_initcall+0x44/0x190
 [<ffffffff810a066b>] sys_init_module+0xcb/0x210
 [<ffffffff8100c012>] system_call_fastpath+0x16/0x1b
..

fb-test2 which reproduces above is available on kernel.org bug #26232.
To solve this issue, avoid calling lock_fb_info inside fb_set_suspend,
and move it out to where needed (callers of fb_set_suspend must call
lock_fb_info before if needed). So far, the only place which needs to
call lock_fb_info is store_fbstate, all other places which calls
fb_set_suspend are suspend/resume hooks that should not need the lock as
they should be run only when processes are already frozen in
suspend/resume.

References: https://bugzilla.kernel.org/show_bug.cgi?id&232
Signed-off-by: Herton Ronaldo Krzesinski <redacted>
Signed-off-by: Florian Tobias Schandinat <FlorianSchandinat@gmx.de>
Cc: stable@kernel.org
---
 drivers/video/fbmem.c   |    3 ---
 drivers/video/fbsysfs.c |    3 +++
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/drivers/video/fbmem.c b/drivers/video/fbmem.c
index 5aac00e..ad93629 100644
--- a/drivers/video/fbmem.c
+++ b/drivers/video/fbmem.c

@@ -1738,8 +1738,6 @@ void fb_set_suspend(struct fb_info *info, int state)
 {
 	struct fb_event event;
 
-	if (!lock_fb_info(info))
-		return;
 	event.info = info;
 	if (state) {
 		fb_notifier_call_chain(FB_EVENT_SUSPEND, &event);

@@ -1748,7 +1746,6 @@ void fb_set_suspend(struct fb_info *info, int state)
 		info->state = FBINFO_STATE_RUNNING;
 		fb_notifier_call_chain(FB_EVENT_RESUME, &event);
 	}
-	unlock_fb_info(info);
 }
 
 /**

diff --git a/drivers/video/fbsysfs.c b/drivers/video/fbsysfs.c
index 04251ce..67afa9c 100644
--- a/drivers/video/fbsysfs.c
+++ b/drivers/video/fbsysfs.c

@@ -399,9 +399,12 @@ static ssize_t store_fbstate(struct device *device,
 
 	state = simple_strtoul(buf, &last, 0);
 
+	if (!lock_fb_info(fb_info))
+		return -ENODEV;
 	console_lock();
 	fb_set_suspend(fb_info, (int)state);
 	console_unlock();
+	unlock_fb_info(fb_info);
 
 	return count;
 }

-- 
1.6.3.2

On Fri, 17 June 2011 Florian Tobias Schandinat [off-list ref] wrote:

From: Herton Ronaldo Krzesinski <redacted>

A lock ordering issue can cause deadlocks: in framebuffer/console code,
all needed struct fb_info locks are taken before acquire_console_sem(),
in places which need to take console semaphore.

But fb_set_suspend is always called with console semaphore held, and
inside it we call lock_fb_info which gets the fb_info lock, inverse
locking order of what the rest of the code does. This causes a real
deadlock issue, when we write to state fb sysfs attribute (which calls
fb_set_suspend) while a framebuffer is being unregistered by
remove_conflicting_framebuffers, as can be shown by following show
blocked state trace on a test program which loads i915 and runs another
forked processes writing to state attribute:

Test process with semaphore held and trying to get fb_info lock:

...

fb-test2 which reproduces above is available on kernel.org bug #26232.
To solve this issue, avoid calling lock_fb_info inside fb_set_suspend,
and move it out to where needed (callers of fb_set_suspend must call
lock_fb_info before if needed). So far, the only place which needs to
call lock_fb_info is store_fbstate, all other places which calls
fb_set_suspend are suspend/resume hooks that should not need the lock as
they should be run only when processes are already frozen in
suspend/resume.

From a quick look through FB drivers in 2.6.39 I've found one that would need
more work:
- drivers/video/sh_mobile_hdmi.c: sh_hdmi_edid_work_fn()
  Should get changed to
  a) right locking order in case (hdmi->hp_state = HDMI_HOTPLUG_CONNECTED)
  b) lock fb_info in the other case
  For this one fb_set_suspend() does get call in a hotplug worker,
  thus independently on suspend/resume process.

The rest does match the suspend/resume hook pattern mentioned.

Bruno

Guennadi, could you have a look at (completely untested) patch which avoids
possible deadlock due to inverted lock taking order on hotplug as well
as "readding" lock_fb_info() for fb_set_suspend() call after Herton's
patch to fb_set_suspend().

Thanks,
Bruno


On Sat, 18 June 2011 Bruno Prémont [off-list ref] wrote:

On Fri, 17 June 2011 Florian Tobias Schandinat [off-list ref] wrote:

quoted

From: Herton Ronaldo Krzesinski <redacted>

A lock ordering issue can cause deadlocks: in framebuffer/console code,
all needed struct fb_info locks are taken before acquire_console_sem(),
in places which need to take console semaphore.

But fb_set_suspend is always called with console semaphore held, and
inside it we call lock_fb_info which gets the fb_info lock, inverse
locking order of what the rest of the code does. This causes a real
deadlock issue, when we write to state fb sysfs attribute (which calls
fb_set_suspend) while a framebuffer is being unregistered by
remove_conflicting_framebuffers, as can be shown by following show
blocked state trace on a test program which loads i915 and runs another
forked processes writing to state attribute:

Test process with semaphore held and trying to get fb_info lock:

...

quoted

fb-test2 which reproduces above is available on kernel.org bug #26232.
To solve this issue, avoid calling lock_fb_info inside fb_set_suspend,
and move it out to where needed (callers of fb_set_suspend must call
lock_fb_info before if needed). So far, the only place which needs to
call lock_fb_info is store_fbstate, all other places which calls
fb_set_suspend are suspend/resume hooks that should not need the lock as
they should be run only when processes are already frozen in
suspend/resume.

From a quick look through FB drivers in 2.6.39 I've found one that would need
more work:
- drivers/video/sh_mobile_hdmi.c: sh_hdmi_edid_work_fn()
  Should get changed to
  a) right locking order in case (hdmi->hp_state = HDMI_HOTPLUG_CONNECTED)
  b) lock fb_info in the other case
  For this one fb_set_suspend() does get call in a hotplug worker,
  thus independently on suspend/resume process.

The rest does match the suspend/resume hook pattern mentioned.

Bruno

quoted

References: https://bugzilla.kernel.org/show_bug.cgi?id&232
Signed-off-by: Herton Ronaldo Krzesinski <redacted>
Signed-off-by: Florian Tobias Schandinat <FlorianSchandinat@gmx.de>
Cc: stable@kernel.org
---
 drivers/video/fbmem.c   |    3 ---
 drivers/video/fbsysfs.c |    3 +++
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/drivers/video/fbmem.c b/drivers/video/fbmem.c
index 5aac00e..ad93629 100644
--- a/drivers/video/fbmem.c
+++ b/drivers/video/fbmem.c

@@ -1738,8 +1738,6 @@ void fb_set_suspend(struct fb_info *info, int state)
 {
 	struct fb_event event;
 
-	if (!lock_fb_info(info))
-		return;
 	event.info = info;
 	if (state) {
 		fb_notifier_call_chain(FB_EVENT_SUSPEND, &event);

@@ -1748,7 +1746,6 @@ void fb_set_suspend(struct fb_info *info, int state)
 		info->state = FBINFO_STATE_RUNNING;
 		fb_notifier_call_chain(FB_EVENT_RESUME, &event);
 	}
-	unlock_fb_info(info);
 }
 
 /**

diff --git a/drivers/video/fbsysfs.c b/drivers/video/fbsysfs.c
index 04251ce..67afa9c 100644
--- a/drivers/video/fbsysfs.c
+++ b/drivers/video/fbsysfs.c

@@ -399,9 +399,12 @@ static ssize_t store_fbstate(struct device *device,
 
 	state = simple_strtoul(buf, &last, 0);
 
+	if (!lock_fb_info(fb_info))
+		return -ENODEV;
 	console_lock();
 	fb_set_suspend(fb_info, (int)state);
 	console_unlock();
+	unlock_fb_info(fb_info);
 
 	return count;
 }

diff --git a/drivers/video/sh_mobile_hdmi.c b/drivers/video/sh_mobile_hdmi.c
index 2b9e56a..b1a13ab 100644
--- a/drivers/video/sh_mobile_hdmi.c
+++ b/drivers/video/sh_mobile_hdmi.c

@@ -1151,27 +1151,27 @@ static void sh_hdmi_edid_work_fn(struct work_struct *work)
 
 		ch = info->par;
 
-		console_lock();
+		if (lock_fb_info(info)) {
+			console_lock();
 
-		/* HDMI plug in */
-		if (!sh_hdmi_must_reconfigure(hdmi) &&
-		    info->state = FBINFO_STATE_RUNNING) {
-			/*
-			 * First activation with the default monitor - just turn
-			 * on, if we run a resume here, the logo disappears
-			 */
-			if (lock_fb_info(info)) {
+			/* HDMI plug in */
+			if (!sh_hdmi_must_reconfigure(hdmi) &&
+			    info->state = FBINFO_STATE_RUNNING) {
+				/*
+				 * First activation with the default monitor - just turn
+				 * on, if we run a resume here, the logo disappears
+				 */
 				info->var.width = hdmi->var.width;
 				info->var.height = hdmi->var.height;
 				sh_hdmi_display_on(hdmi, info);
-				unlock_fb_info(info);
+			} else {
+				/* New monitor or have to wake up */
+				fb_set_suspend(info, 0);
 			}
-		} else {
-			/* New monitor or have to wake up */
-			fb_set_suspend(info, 0);
-		}
 
-		console_unlock();
+			console_unlock();
+			unlock_fb_info(info);
+		}
 	} else {
 		ret = 0;
 		if (!hdmi->info)

@@ -1181,12 +1181,15 @@ static void sh_hdmi_edid_work_fn(struct work_struct *work)
 		fb_destroy_modedb(hdmi->monspec.modedb);
 		hdmi->monspec.modedb = NULL;
 
-		console_lock();
+		if (lock_fb_info(info)) {
+			console_lock();
 
-		/* HDMI disconnect */
-		fb_set_suspend(hdmi->info, 1);
+			/* HDMI disconnect */
+			fb_set_suspend(hdmi->info, 1);
 
-		console_unlock();
+			console_unlock();
+			unlock_fb_info(info);
+		}
 		pm_runtime_put(hdmi->dev);
 	}
 

Hi Paul,

what are your plans for the patch below? Do you queue it for 3.1?
The only feedback so far was that one driver (sh_mobile_hdmi) would require more 
work with a patch for that proposed. Without any objections to the patch I think 
we should take it as fixing the subsystem is more important than causing 
potential issues with a single driver.


Thanks,

Florian Tobias Schandinat

On 06/17/2011 07:02 PM, Florian Tobias Schandinat wrote:

On Wed, Jul 20, 2011 at 8:16 PM, Florian Tobias Schandinat
[off-list ref] wrote:

what are your plans for the patch below? Do you queue it for 3.1?

Looks like someone else fixed it slightly differently:

http://permalink.gmane.org/gmane.linux.kernel.stable/15187

-- 
Francis

ping

Guennadi, I really want this issue fixed. Please have a look at Bruno's patch
otherwise your driver might remain or get even more broken...

I am scheduling Herton's patch for the next merge window.


Regards,

Florian Tobias Schandinat


On 06/18/2011 09:19 AM, Bruno Prémont wrote:

Hi Florian

On Thu, 1 Sep 2011, Florian Tobias Schandinat wrote:

ping

Guennadi, I really want this issue fixed. Please have a look at Bruno's patch
otherwise your driver might remain or get even more broken...

Sorry, somehow this patch excaped my attention. Will test tomorrow.

Thanks
Guennadi

I am scheduling Herton's patch for the next merge window.


Regards,

Florian Tobias Schandinat


On 06/18/2011 09:19 AM, Bruno Prémont wrote:

quoted

Guennadi, could you have a look at (completely untested) patch which avoids
possible deadlock due to inverted lock taking order on hotplug as well
as "readding" lock_fb_info() for fb_set_suspend() call after Herton's
patch to fb_set_suspend().

Thanks,
Bruno


On Sat, 18 June 2011 Bruno Prémont [off-list ref] wrote:

quoted

On Fri, 17 June 2011 Florian Tobias Schandinat [off-list ref] wrote:

quoted

From: Herton Ronaldo Krzesinski <redacted>

A lock ordering issue can cause deadlocks: in framebuffer/console code,
all needed struct fb_info locks are taken before acquire_console_sem(),
in places which need to take console semaphore.

But fb_set_suspend is always called with console semaphore held, and
inside it we call lock_fb_info which gets the fb_info lock, inverse
locking order of what the rest of the code does. This causes a real
deadlock issue, when we write to state fb sysfs attribute (which calls
fb_set_suspend) while a framebuffer is being unregistered by
remove_conflicting_framebuffers, as can be shown by following show
blocked state trace on a test program which loads i915 and runs another
forked processes writing to state attribute:

Test process with semaphore held and trying to get fb_info lock:

...

quoted

fb-test2 which reproduces above is available on kernel.org bug #26232.
To solve this issue, avoid calling lock_fb_info inside fb_set_suspend,
and move it out to where needed (callers of fb_set_suspend must call
lock_fb_info before if needed). So far, the only place which needs to
call lock_fb_info is store_fbstate, all other places which calls
fb_set_suspend are suspend/resume hooks that should not need the lock as
they should be run only when processes are already frozen in
suspend/resume.

From a quick look through FB drivers in 2.6.39 I've found one that would need
more work:
- drivers/video/sh_mobile_hdmi.c: sh_hdmi_edid_work_fn()
  Should get changed to
  a) right locking order in case (hdmi->hp_state = HDMI_HOTPLUG_CONNECTED)
  b) lock fb_info in the other case
  For this one fb_set_suspend() does get call in a hotplug worker,
  thus independently on suspend/resume process.

The rest does match the suspend/resume hook pattern mentioned.

Bruno

quoted

References: https://bugzilla.kernel.org/show_bug.cgi?id&232
Signed-off-by: Herton Ronaldo Krzesinski <redacted>
Signed-off-by: Florian Tobias Schandinat <FlorianSchandinat@gmx.de>
Cc: stable@kernel.org
---
 drivers/video/fbmem.c   |    3 ---
 drivers/video/fbsysfs.c |    3 +++
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/drivers/video/fbmem.c b/drivers/video/fbmem.c
index 5aac00e..ad93629 100644
--- a/drivers/video/fbmem.c
+++ b/drivers/video/fbmem.c

@@ -1738,8 +1738,6 @@ void fb_set_suspend(struct fb_info *info, int state)
 {
 	struct fb_event event;
 
-	if (!lock_fb_info(info))
-		return;
 	event.info = info;
 	if (state) {
 		fb_notifier_call_chain(FB_EVENT_SUSPEND, &event);

@@ -1748,7 +1746,6 @@ void fb_set_suspend(struct fb_info *info, int state)
 		info->state = FBINFO_STATE_RUNNING;
 		fb_notifier_call_chain(FB_EVENT_RESUME, &event);
 	}
-	unlock_fb_info(info);
 }
 
 /**

diff --git a/drivers/video/fbsysfs.c b/drivers/video/fbsysfs.c
index 04251ce..67afa9c 100644
--- a/drivers/video/fbsysfs.c
+++ b/drivers/video/fbsysfs.c

@@ -399,9 +399,12 @@ static ssize_t store_fbstate(struct device *device,
 
 	state = simple_strtoul(buf, &last, 0);
 
+	if (!lock_fb_info(fb_info))
+		return -ENODEV;
 	console_lock();
 	fb_set_suspend(fb_info, (int)state);
 	console_unlock();
+	unlock_fb_info(fb_info);
 
 	return count;
 }

diff --git a/drivers/video/sh_mobile_hdmi.c b/drivers/video/sh_mobile_hdmi.c
index 2b9e56a..b1a13ab 100644
--- a/drivers/video/sh_mobile_hdmi.c
+++ b/drivers/video/sh_mobile_hdmi.c

@@ -1151,27 +1151,27 @@ static void sh_hdmi_edid_work_fn(struct work_struct *work)
 
 		ch = info->par;
 
-		console_lock();
+		if (lock_fb_info(info)) {
+			console_lock();
 
-		/* HDMI plug in */
-		if (!sh_hdmi_must_reconfigure(hdmi) &&
-		    info->state = FBINFO_STATE_RUNNING) {
-			/*
-			 * First activation with the default monitor - just turn
-			 * on, if we run a resume here, the logo disappears
-			 */
-			if (lock_fb_info(info)) {
+			/* HDMI plug in */
+			if (!sh_hdmi_must_reconfigure(hdmi) &&
+			    info->state = FBINFO_STATE_RUNNING) {
+				/*
+				 * First activation with the default monitor - just turn
+				 * on, if we run a resume here, the logo disappears
+				 */
 				info->var.width = hdmi->var.width;
 				info->var.height = hdmi->var.height;
 				sh_hdmi_display_on(hdmi, info);
-				unlock_fb_info(info);
+			} else {
+				/* New monitor or have to wake up */
+				fb_set_suspend(info, 0);
 			}
-		} else {
-			/* New monitor or have to wake up */
-			fb_set_suspend(info, 0);
-		}
 
-		console_unlock();
+			console_unlock();
+			unlock_fb_info(info);
+		}
 	} else {
 		ret = 0;
 		if (!hdmi->info)

@@ -1181,12 +1181,15 @@ static void sh_hdmi_edid_work_fn(struct work_struct *work)
 		fb_destroy_modedb(hdmi->monspec.modedb);
 		hdmi->monspec.modedb = NULL;
 
-		console_lock();
+		if (lock_fb_info(info)) {
+			console_lock();
 
-		/* HDMI disconnect */
-		fb_set_suspend(hdmi->info, 1);
+			/* HDMI disconnect */
+			fb_set_suspend(hdmi->info, 1);
 
-		console_unlock();
+			console_unlock();
+			unlock_fb_info(info);
+		}
 		pm_runtime_put(hdmi->dev);
 	}
 
--

To unsubscribe from this list: send the line "unsubscribe linux-fbdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

---
Guennadi Liakhovetski, Ph.D.
Freelance Open-Source Software Developer
http://www.open-technology.de/

Hi Florian

On Thu, 1 Sep 2011, Florian Tobias Schandinat wrote:

ping

Guennadi, I really want this issue fixed. Please have a look at Bruno's patch
otherwise your driver might remain or get even more broken...

I am scheduling Herton's patch for the next merge window.

So, the patch looks simple and correct, when applied on top of 
http://marc.info/?l=linux-kernel&m0833638508657&w=2
But it doesn't apply anymore and after fixing a trivial merge conflict, it 
fails to build, which is also trivially fixed. Please, add my

Signed-off-by: Guennadi Liakhovetski <redacted>

to the set and use this updated version:

diff --git a/drivers/video/sh_mobile_hdmi.c b/drivers/video/sh_mobile_hdmi.c
index 7d54e2c..647ba98 100644
--- a/drivers/video/sh_mobile_hdmi.c
+++ b/drivers/video/sh_mobile_hdmi.c

@@ -1111,6 +1111,7 @@ static long sh_hdmi_clk_configure(struct sh_hdmi *hdmi, unsigned long hdmi_rate,
 static void sh_hdmi_edid_work_fn(struct work_struct *work)
 {
 	struct sh_hdmi *hdmi = container_of(work, struct sh_hdmi, edid_work.work);
+	struct fb_info *info;
 	struct sh_mobile_hdmi_info *pdata = hdmi->dev->platform_data;
 	struct sh_mobile_lcdc_chan *ch;
 	int ret;

@@ -1123,8 +1124,9 @@ static void sh_hdmi_edid_work_fn(struct work_struct *work)
 
 	mutex_lock(&hdmi->mutex);
 
+	info = hdmi->info;
+
 	if (hdmi->hp_state = HDMI_HOTPLUG_CONNECTED) {
-		struct fb_info *info = hdmi->info;
 		unsigned long parent_rate = 0, hdmi_rate;
 
 		ret = sh_hdmi_read_edid(hdmi, &hdmi_rate, &parent_rate);

@@ -1148,42 +1150,45 @@ static void sh_hdmi_edid_work_fn(struct work_struct *work)
 
 		ch = info->par;
 
-		console_lock();
+		if (lock_fb_info(info)) {
+			console_lock();
 
-		/* HDMI plug in */
-		if (!sh_hdmi_must_reconfigure(hdmi) &&
-		    info->state = FBINFO_STATE_RUNNING) {
-			/*
-			 * First activation with the default monitor - just turn
-			 * on, if we run a resume here, the logo disappears
-			 */
-			if (lock_fb_info(info)) {
+			/* HDMI plug in */
+			if (!sh_hdmi_must_reconfigure(hdmi) &&
+			    info->state = FBINFO_STATE_RUNNING) {
+				/*
+				 * First activation with the default monitor - just turn
+				 * on, if we run a resume here, the logo disappears
+				 */
 				info->var.width = hdmi->var.width;
 				info->var.height = hdmi->var.height;
 				sh_hdmi_display_on(hdmi, info);
-				unlock_fb_info(info);
+			} else {
+				/* New monitor or have to wake up */
+				fb_set_suspend(info, 0);
 			}
-		} else {
-			/* New monitor or have to wake up */
-			fb_set_suspend(info, 0);
-		}
 
-		console_unlock();
+			console_unlock();
+			unlock_fb_info(info);
+		}
 	} else {
 		ret = 0;
-		if (!hdmi->info)
+		if (!info)
 			goto out;
 
 		hdmi->monspec.modedb_len = 0;
 		fb_destroy_modedb(hdmi->monspec.modedb);
 		hdmi->monspec.modedb = NULL;
 
-		console_lock();
+		if (lock_fb_info(info)) {
+			console_lock();
 
-		/* HDMI disconnect */
-		fb_set_suspend(hdmi->info, 1);
+			/* HDMI disconnect */
+			fb_set_suspend(info, 1);
 
-		console_unlock();
+			console_unlock();
+			unlock_fb_info(info);
+		}
 	}
 
 out:

Thanks
Guennadi

Regards,

Florian Tobias Schandinat


On 06/18/2011 09:19 AM, Bruno Prémont wrote:

quoted

Guennadi, could you have a look at (completely untested) patch which avoids
possible deadlock due to inverted lock taking order on hotplug as well
as "readding" lock_fb_info() for fb_set_suspend() call after Herton's
patch to fb_set_suspend().

Thanks,
Bruno


On Sat, 18 June 2011 Bruno Prémont [off-list ref] wrote:

quoted

On Fri, 17 June 2011 Florian Tobias Schandinat [off-list ref] wrote:

quoted

From: Herton Ronaldo Krzesinski <redacted>

A lock ordering issue can cause deadlocks: in framebuffer/console code,
all needed struct fb_info locks are taken before acquire_console_sem(),
in places which need to take console semaphore.

But fb_set_suspend is always called with console semaphore held, and
inside it we call lock_fb_info which gets the fb_info lock, inverse
locking order of what the rest of the code does. This causes a real
deadlock issue, when we write to state fb sysfs attribute (which calls
fb_set_suspend) while a framebuffer is being unregistered by
remove_conflicting_framebuffers, as can be shown by following show
blocked state trace on a test program which loads i915 and runs another
forked processes writing to state attribute:

Test process with semaphore held and trying to get fb_info lock:

...

quoted

fb-test2 which reproduces above is available on kernel.org bug #26232.
To solve this issue, avoid calling lock_fb_info inside fb_set_suspend,
and move it out to where needed (callers of fb_set_suspend must call
lock_fb_info before if needed). So far, the only place which needs to
call lock_fb_info is store_fbstate, all other places which calls
fb_set_suspend are suspend/resume hooks that should not need the lock as
they should be run only when processes are already frozen in
suspend/resume.

From a quick look through FB drivers in 2.6.39 I've found one that would need
more work:
- drivers/video/sh_mobile_hdmi.c: sh_hdmi_edid_work_fn()
  Should get changed to
  a) right locking order in case (hdmi->hp_state = HDMI_HOTPLUG_CONNECTED)
  b) lock fb_info in the other case
  For this one fb_set_suspend() does get call in a hotplug worker,
  thus independently on suspend/resume process.

The rest does match the suspend/resume hook pattern mentioned.

Bruno

quoted

References: https://bugzilla.kernel.org/show_bug.cgi?id&232
Signed-off-by: Herton Ronaldo Krzesinski <redacted>
Signed-off-by: Florian Tobias Schandinat <FlorianSchandinat@gmx.de>
Cc: stable@kernel.org
---
 drivers/video/fbmem.c   |    3 ---
 drivers/video/fbsysfs.c |    3 +++
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/drivers/video/fbmem.c b/drivers/video/fbmem.c
index 5aac00e..ad93629 100644
--- a/drivers/video/fbmem.c
+++ b/drivers/video/fbmem.c

@@ -1738,8 +1738,6 @@ void fb_set_suspend(struct fb_info *info, int state)
 {
 	struct fb_event event;
 
-	if (!lock_fb_info(info))
-		return;
 	event.info = info;
 	if (state) {
 		fb_notifier_call_chain(FB_EVENT_SUSPEND, &event);

@@ -1748,7 +1746,6 @@ void fb_set_suspend(struct fb_info *info, int state)
 		info->state = FBINFO_STATE_RUNNING;
 		fb_notifier_call_chain(FB_EVENT_RESUME, &event);
 	}
-	unlock_fb_info(info);
 }
 
 /**

diff --git a/drivers/video/fbsysfs.c b/drivers/video/fbsysfs.c
index 04251ce..67afa9c 100644
--- a/drivers/video/fbsysfs.c
+++ b/drivers/video/fbsysfs.c

@@ -399,9 +399,12 @@ static ssize_t store_fbstate(struct device *device,
 
 	state = simple_strtoul(buf, &last, 0);
 
+	if (!lock_fb_info(fb_info))
+		return -ENODEV;
 	console_lock();
 	fb_set_suspend(fb_info, (int)state);
 	console_unlock();
+	unlock_fb_info(fb_info);
 
 	return count;
 }

diff --git a/drivers/video/sh_mobile_hdmi.c b/drivers/video/sh_mobile_hdmi.c
index 2b9e56a..b1a13ab 100644
--- a/drivers/video/sh_mobile_hdmi.c
+++ b/drivers/video/sh_mobile_hdmi.c

@@ -1151,27 +1151,27 @@ static void sh_hdmi_edid_work_fn(struct work_struct *work)
 
 		ch = info->par;
 
-		console_lock();
+		if (lock_fb_info(info)) {
+			console_lock();
 
-		/* HDMI plug in */
-		if (!sh_hdmi_must_reconfigure(hdmi) &&
-		    info->state = FBINFO_STATE_RUNNING) {
-			/*
-			 * First activation with the default monitor - just turn
-			 * on, if we run a resume here, the logo disappears
-			 */
-			if (lock_fb_info(info)) {
+			/* HDMI plug in */
+			if (!sh_hdmi_must_reconfigure(hdmi) &&
+			    info->state = FBINFO_STATE_RUNNING) {
+				/*
+				 * First activation with the default monitor - just turn
+				 * on, if we run a resume here, the logo disappears
+				 */
 				info->var.width = hdmi->var.width;
 				info->var.height = hdmi->var.height;
 				sh_hdmi_display_on(hdmi, info);
-				unlock_fb_info(info);
+			} else {
+				/* New monitor or have to wake up */
+				fb_set_suspend(info, 0);
 			}
-		} else {
-			/* New monitor or have to wake up */
-			fb_set_suspend(info, 0);
-		}
 
-		console_unlock();
+			console_unlock();
+			unlock_fb_info(info);
+		}
 	} else {
 		ret = 0;
 		if (!hdmi->info)

@@ -1181,12 +1181,15 @@ static void sh_hdmi_edid_work_fn(struct work_struct *work)
 		fb_destroy_modedb(hdmi->monspec.modedb);
 		hdmi->monspec.modedb = NULL;
 
-		console_lock();
+		if (lock_fb_info(info)) {
+			console_lock();
 
-		/* HDMI disconnect */
-		fb_set_suspend(hdmi->info, 1);
+			/* HDMI disconnect */
+			fb_set_suspend(hdmi->info, 1);
 
-		console_unlock();
+			console_unlock();
+			unlock_fb_info(info);
+		}
 		pm_runtime_put(hdmi->dev);
 	}
 
--

To unsubscribe from this list: send the line "unsubscribe linux-fbdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

---
Guennadi Liakhovetski, Ph.D.
Freelance Open-Source Software Developer
http://www.open-technology.de/

Hi Bruno, Guennadi,

On 09/02/2011 04:06 PM, Guennadi Liakhovetski wrote:

Hi Florian

On Thu, 1 Sep 2011, Florian Tobias Schandinat wrote:

quoted

ping

Guennadi, I really want this issue fixed. Please have a look at Bruno's patch
otherwise your driver might remain or get even more broken...

I am scheduling Herton's patch for the next merge window.

So, the patch looks simple and correct, when applied on top of 
http://marc.info/?l=linux-kernel&m0833638508657&w=2
But it doesn't apply anymore and after fixing a trivial merge conflict, it 
fails to build, which is also trivially fixed. Please, add my

Signed-off-by: Guennadi Liakhovetski <redacted>

So I guess we can finally solve this issue, thanks.

On the other hand I just noticed the original patch didn't have Bruno's
Signed-off and a commit message was missing. So may I ask you, Bruno, to add
your Signed-off or preferably resend the version tested by Guennadi in the
appropriate patch format.


Thanks,

Florian Tobias Schandinat

Following on Herton's patch "fb: avoid possible deadlock caused by
fb_set_suspend" which moves lock_fb_info() out of fb_set_suspend()
to its callers, correct sh-mobile's locking around call to
fb_set_suspend() and the same sort of deaklocks with console_lock()
due to order of taking the lock.

console_lock() must be taken while fb_info is already locked and fb_info
must be locked while calling fb_set_suspend().

Signed-off-by: Bruno Prémont <bonbons@linux-vserver.org>
Signed-off-by: Guennadi Liakhovetski <redacted>
---

Hi Florian,

On the other hand I just noticed the original patch didn't have
Bruno's Signed-off and a commit message was missing. So may I ask you,
Bruno, to add your Signed-off or preferably resend the version tested
by Guennadi in the appropriate patch format.

Here it is with commit message and the Signed-off-by's.

Thanks,
Bruno




 drivers/video/sh_mobile_hdmi.c |   47 ++++++++++++++++++++++-----------------
 1 files changed, 26 insertions(+), 21 deletions(-)

diff --git a/drivers/video/sh_mobile_hdmi.c b/drivers/video/sh_mobile_hdmi.c
index 7d54e2c..647ba98 100644
--- a/drivers/video/sh_mobile_hdmi.c
+++ b/drivers/video/sh_mobile_hdmi.c

@@ -1111,6 +1111,7 @@ static long sh_hdmi_clk_configure(struct sh_hdmi *hdmi, unsigned long hdmi_rate,
 static void sh_hdmi_edid_work_fn(struct work_struct *work)
 {
 	struct sh_hdmi *hdmi = container_of(work, struct sh_hdmi, edid_work.work);
+	struct fb_info *info;
 	struct sh_mobile_hdmi_info *pdata = hdmi->dev->platform_data;
 	struct sh_mobile_lcdc_chan *ch;
 	int ret;

@@ -1123,8 +1124,9 @@ static void sh_hdmi_edid_work_fn(struct work_struct *work)
 
 	mutex_lock(&hdmi->mutex);
 
+	info = hdmi->info;
+
 	if (hdmi->hp_state = HDMI_HOTPLUG_CONNECTED) {
-		struct fb_info *info = hdmi->info;
 		unsigned long parent_rate = 0, hdmi_rate;
 
 		ret = sh_hdmi_read_edid(hdmi, &hdmi_rate, &parent_rate);

@@ -1148,42 +1150,45 @@ static void sh_hdmi_edid_work_fn(struct work_struct *work)
 
 		ch = info->par;
 
-		console_lock();
+		if (lock_fb_info(info)) {
+			console_lock();
 
-		/* HDMI plug in */
-		if (!sh_hdmi_must_reconfigure(hdmi) &&
-		    info->state = FBINFO_STATE_RUNNING) {
-			/*
-			 * First activation with the default monitor - just turn
-			 * on, if we run a resume here, the logo disappears
-			 */
-			if (lock_fb_info(info)) {
+			/* HDMI plug in */
+			if (!sh_hdmi_must_reconfigure(hdmi) &&
+			    info->state = FBINFO_STATE_RUNNING) {
+				/*
+				 * First activation with the default monitor - just turn
+				 * on, if we run a resume here, the logo disappears
+				 */
 				info->var.width = hdmi->var.width;
 				info->var.height = hdmi->var.height;
 				sh_hdmi_display_on(hdmi, info);
-				unlock_fb_info(info);
+			} else {
+				/* New monitor or have to wake up */
+				fb_set_suspend(info, 0);
 			}
-		} else {
-			/* New monitor or have to wake up */
-			fb_set_suspend(info, 0);
-		}
 
-		console_unlock();
+			console_unlock();
+			unlock_fb_info(info);
+		}
 	} else {
 		ret = 0;
-		if (!hdmi->info)
+		if (!info)
 			goto out;
 
 		hdmi->monspec.modedb_len = 0;
 		fb_destroy_modedb(hdmi->monspec.modedb);
 		hdmi->monspec.modedb = NULL;
 
-		console_lock();
+		if (lock_fb_info(info)) {
+			console_lock();
 
-		/* HDMI disconnect */
-		fb_set_suspend(hdmi->info, 1);
+			/* HDMI disconnect */
+			fb_set_suspend(info, 1);
 
-		console_unlock();
+			console_unlock();
+			unlock_fb_info(info);
+		}
 	}
 
 out:

On 09/02/2011 05:24 PM, Bruno Prémont wrote:

Following on Herton's patch "fb: avoid possible deadlock caused by
fb_set_suspend" which moves lock_fb_info() out of fb_set_suspend()
to its callers, correct sh-mobile's locking around call to
fb_set_suspend() and the same sort of deaklocks with console_lock()
due to order of taking the lock.

console_lock() must be taken while fb_info is already locked and fb_info
must be locked while calling fb_set_suspend().

Signed-off-by: Bruno Prémont <bonbons@linux-vserver.org>
Signed-off-by: Guennadi Liakhovetski <redacted>
---

Hi Florian,

quoted

On the other hand I just noticed the original patch didn't have
Bruno's Signed-off and a commit message was missing. So may I ask you,
Bruno, to add your Signed-off or preferably resend the version tested
by Guennadi in the appropriate patch format.

Here it is with commit message and the Signed-off-by's.

Applied and added stable as Cc as I think this should be applied as well
whenever Herton's patch is applied.


Thanks,

Florian Tobias Schandinat