Thread (13 messages) 13 messages, 5 authors, 2011-09-02

Re: [PATCH] fb: avoid possible deadlock caused by fb_set_suspend

From: Guennadi Liakhovetski <hidden>
Date: 2011-09-02 16:07:29
Also in: lkml
Subsystem: framebuffer layer, the rest · Maintainers: Helge Deller, Linus Torvalds

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/
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help