Skip to contentSkip to conversation

fbdev: protect mode sysfs reads with lock_fb_info()

PATCH New Melbin K Mathew 3 participants 15 messages Jul 1 – Jul 18

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 fbdev: protect mode sysfs reads with lock_fb_info()

show_mode() dereferences fb_info->mode and show_modes() walks fb_info->modelist without holding lock_fb_info(). store_modes() takes lock_fb_info() while replacing the modelist and freeing the old one.

A concurrent reader can load a pointer to an old modelist entry before store_modes() frees it, then dereference freed memory in mode_string().

Take lock_fb_info() in both show_mode() and show_modes() to serialize with store_modes(). In show_mode(), copy the mode to the stack and format the stack copy after dropping the lock.

Impact: local kernel UAF read when a privileged writer races with sysfs readers of /sys/class/graphics/fb*/mode and modes.

---
A userspace reproducer triggering the race is available to maintainers on request.

 drivers/video/fbdev/core/fbsysfs.c | 16 ++++++++++++++--
 1 file changed, 14 insertions(+), 2 deletions(-)
drivers/video/fbdev/core/fbsysfs.c +14 -2
index ea196603c7..6bdb25f7be 100644
@@ -82,11 +82,20 @@ static ssize_t show_mode(struct device *device, struct device_attribute *attr,
 			 char *buf)
 {
 	struct fb_info *fb_info = dev_get_drvdata(device);
+	struct fb_videomode mode;
+	bool have_mode = false;
 
-	if (!fb_info->mode)
+	lock_fb_info(fb_info);
+	if (fb_info->mode) {
+		mode = *fb_info->mode;
+		have_mode = true;
+	}
+	unlock_fb_info(fb_info);
+
+	if (!have_mode)
 		return 0;
 
-	return mode_string(buf, 0, fb_info->mode);
+	return mode_string(buf, 0, &mode);
 }
 
 static ssize_t store_modes(struct device *device,
@@ -134,10 +143,13 @@ static ssize_t show_modes(struct device *device, struct device_attribute *attr,
 	const struct fb_videomode *mode;
 
 	i = 0;
+	lock_fb_info(fb_info);
 	list_for_each_entry(modelist, &fb_info->modelist, list) {
 		mode = &modelist->mode;
 		i += mode_string(buf, i, mode);
 	}
+	unlock_fb_info(fb_info);
+
 	return i;
 }
 
Signature · 1 line
2.39.5
#1 lore raw
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [High] The patch fails to resolve the local kernel Use-After-Free because it misses lockless traversal and assignment in store_mode().
- [High] Unprivileged local DoS via WARN_ON_ONCE triggered by snprintf size underflow in show_modes().
--

commit 781c0f45188865f8a46896496380e20ca20b9af8
Author: Melbin K Mathew <[email protected]>

fbdev: protect mode sysfs reads with lock_fb_info()

This commit adds lock_fb_info() around mode sysfs reads to protect against concurrent modifications to the modelist by store_modes().

Quoted 4 lines
drivers/video/fbdev/core/fbsysfs.c +0 -0
index ea196603c7a87..6bdb25f7be018 100644

[ ... ]

Quoted 4 lines
drivers/video/fbdev/core/fbsysfs.c +0 -0
@@ -82,11 +82,20 @@ static ssize_t show_mode(struct device *device, struct device_attribute *attr,
 			 char *buf)
 {
 	struct fb_info *fb_info = dev_get_drvdata(device);

drivers/video/fbdev/core/fbsysfs.c static ssize_t show_mode(struct device *device, struct device_attribute *attr, · line 93

Quoted 10 lines
drivers/video/fbdev/core/fbsysfs.c +8 -1
+	struct fb_videomode mode;
+	bool have_mode = false;
 
-	if (!fb_info->mode)
+	lock_fb_info(fb_info);
+	if (fb_info->mode) {
+		mode = *fb_info->mode;
+		have_mode = true;
+	}
+	unlock_fb_info(fb_info);

[Severity: High]
This is a pre-existing issue, but does store_mode() also need similar locking protection to fully prevent the use-after-free?

While this patch protects show_mode() and show_modes(), store_mode() still traverses the modelist locklessly:

drivers/video/fbdev/core/fbsysfs.c:store_mode() {
    ...
    list_for_each_entry(modelist, &fb_info->modelist, list) {
    ...
}

If store_modes() concurrently clears and frees the modelist, could store_mode() traverse freed list elements?

Additionally, store_mode() locklessly assigns the mode pointer:

drivers/video/fbdev/core/fbsysfs.c:store_mode() {
    ...
    fb_info->mode = mode;
    ...
}

If store_modes() freed the old modelist in the interim, could this resurrect a dangling pointer into fb_info->mode and bypass the fix for subsequent readers?

[ ... ]

drivers/video/fbdev/core/fbsysfs.c static ssize_t show_modes(struct device *device, struct device_attribute *attr, · line 150

Quoted 9 lines
drivers/video/fbdev/core/fbsysfs.c +1 -0
@@ -134,10 +143,13 @@ static ssize_t show_modes(struct device *device, struct device_attribute *attr,
 	const struct fb_videomode *mode;
 
 	i = 0;
+	lock_fb_info(fb_info);
 	list_for_each_entry(modelist, &fb_info->modelist, list) {
 		mode = &modelist->mode;
 		i += mode_string(buf, i, mode);
 	}

[Severity: High]
This is a pre-existing issue, but can this loop trigger a WARN_ON_ONCE due to integer underflow if the modelist is large?

If the accumulated length in i exceeds PAGE_SIZE (for example, with a very large modelist), the size calculation in mode_string() appears to underflow:

drivers/video/fbdev/core/fbsysfs.c:mode_string() {
    ...
    return snprintf(&buf[offset], PAGE_SIZE - offset, "%c:%dx%d%c-%d\n",
                    m, mode->xres, mode->yres, v, mode->refresh);
}

Since offset is unsigned, PAGE_SIZE - offset would become a massive positive value, which vsnprintf() detects as exceeding INT_MAX and triggers a WARN_ON_ONCE. Could an unprivileged local user trigger a kernel panic on systems with panic_on_warn=1 by repeatedly reading the modes sysfs file?

Quoted 5 lines
drivers/video/fbdev/core/fbsysfs.c +2 -0
+	unlock_fb_info(fb_info);
+
 	return i;
 }
 
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#2 lore raw

v2 cover fbdev: fix mode sysfs lifetime and bounds issues

In reply to Melbin K Mathew

This v2 addresses the remaining fbdev mode sysfs lifetime issues noted during review.

Patch 1 bounds mode sysfs output so show_modes() cannot advance the sysfs buffer offset past PAGE_SIZE. mode_string() is given a size parameter and switched to scnprintf().

Patch 2 serializes mode sysfs access with lock_fb_info(), including store_mode(), show_mode(), and show_modes(), so these paths cannot race with store_modes() while it replaces and frees the old modelist.

Changes in v2:
- Add bounds handling for mode_string()/show_modes().
- Extend locking to store_mode() via activate_locked().
- Keep show_mode() using a stack copy after dropping lock_fb_info().
- Avoid overclaiming unprivileged impact in the commit text.

Melbin K Mathew (2):
fbdev: bound mode sysfs output to the sysfs buffer
fbdev: serialize mode sysfs access with lock_fb_info()

 drivers/video/fbdev/core/fbsysfs.c | 59 ++++++++++++++++++++++++------
 1 file changed, 47 insertions(+), 12 deletions(-)
Signature · 1 line
2.39.5
#3 lore raw

v3 cover fbdev: fix mode sysfs lifetime and bounds issues

This v3 adds a patch to clear fb_info->mode before deleting a videomode through FBIOPUT_VSCREENINFO with FB_ACTIVATE_INV_MODE.

Patch 1 bounds mode sysfs output so show_modes() cannot advance the sysfs buffer offset past PAGE_SIZE.

Patch 2 clears fb_info->mode before fb_delete_videomode() when it matches the mode being removed via the FBIOPUT_VSCREENINFO ioctl.

Patch 3 serializes mode sysfs access with lock_fb_info(), including store_mode(), show_mode(), and show_modes().

Changes in v3:
- Add patch to clear fb_info->mode in fb_set_var() INV_MODE path.

Melbin K Mathew (3):
fbdev: bound mode sysfs output to the sysfs buffer
fbdev: clear fb_info->mode before deleting a videomode
fbdev: serialize mode sysfs access with lock_fb_info()

 drivers/video/fbdev/core/fbmem.c   |  5 ++-
 drivers/video/fbdev/core/fbsysfs.c | 59 ++++++++++++++++++++++++------
 2 files changed, 51 insertions(+), 13 deletions(-)
Signature · 1 line
2.39.5
#4 lore raw

PATCH v3 1/3 fbdev: bound mode sysfs output to the sysfs buffer

In reply to Melbin K Mathew

mode_string() uses snprintf() which can return a value larger than the remaining buffer space. show_modes() accumulates the return value into i without checking whether i has reached PAGE_SIZE, causing the offset to advance past the sysfs buffer if the modelist is long enough.

Add a size parameter to mode_string() and use scnprintf() to return only the bytes actually written. Add an early return when offset already exceeds the buffer. In show_modes(), stop accumulating once the buffer is full.

---
 drivers/video/fbdev/core/fbsysfs.c | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)
drivers/video/fbdev/core/fbsysfs.c +10 -5
index ea196603c7..af21dc5052 100644
@@ -27,12 +27,15 @@ static int activate(struct fb_info *fb_info, struct fb_var_screeninfo *var)
 	return 0;
 }
 
-static int mode_string(char *buf, unsigned int offset,
+static int mode_string(char *buf, size_t size, unsigned int offset,
 		       const struct fb_videomode *mode)
 {
 	char m = 'U';
 	char v = 'p';
 
+	if (offset >= size)
+		return 0;
+
 	if (mode->flag & FB_MODE_IS_DETAILED)
 		m = 'D';
 	if (mode->flag & FB_MODE_IS_VESA)
@@ -45,7 +48,7 @@ static int mode_string(char *buf, unsigned int offset,
 	if (mode->vmode & FB_VMODE_DOUBLE)
 		v = 'd';
 
-	return snprintf(&buf[offset], PAGE_SIZE - offset, "%c:%dx%d%c-%d\n",
+	return scnprintf(&buf[offset], size - offset, "%c:%dx%d%c-%d\n",
 	                m, mode->xres, mode->yres, v, mode->refresh);
 }
 
@@ -64,7 +67,7 @@ static ssize_t store_mode(struct device *device, struct device_attribute *attr,
 
 	list_for_each_entry(modelist, &fb_info->modelist, list) {
 		mode = &modelist->mode;
-		i = mode_string(mstr, 0, mode);
+		i = mode_string(mstr, sizeof(mstr), 0, mode);
 		if (strncmp(mstr, buf, max(count, i)) == 0) {
 
 			var = fb_info->var;
@@ -86,7 +89,7 @@ static ssize_t show_mode(struct device *device, struct device_attribute *attr,
 	if (!fb_info->mode)
 		return 0;
 
-	return mode_string(buf, 0, fb_info->mode);
+	return mode_string(buf, PAGE_SIZE, 0, fb_info->mode);
 }
 
 static ssize_t store_modes(struct device *device,
@@ -136,7 +139,9 @@ static ssize_t show_modes(struct device *device, struct device_attribute *attr,
 	i = 0;
 	list_for_each_entry(modelist, &fb_info->modelist, list) {
 		mode = &modelist->mode;
-		i += mode_string(buf, i, mode);
+		i += mode_string(buf, PAGE_SIZE, i, mode);
+		if (i >= PAGE_SIZE - 1)
+			break;
 	}
 	return i;
 }
Signature · 1 line
2.39.5
#5 lore raw

Re: [PATCH v3 1/3] fbdev: bound mode sysfs output to the sysfs buffer

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] Unprotected value copy of large fb_info->var struct leads to torn reads and state corruption in sysfs store functions. --

commit 03248859e392cd00678961d93e5c27717ac23498
Author: Melbin K Mathew <[email protected]>

fbdev: bound mode sysfs output to the sysfs buffer

This commit updates mode_string() to take a size parameter and use scnprintf() to prevent buffer overruns when generating sysfs output. It also modifies show_modes() to stop accumulating mode strings once the sysfs buffer is full.

Quoted 4 lines
drivers/video/fbdev/core/fbsysfs.c +0 -0
index ea196603c7a87..af21dc5052df3 100644

[ ... ]

drivers/video/fbdev/core/fbsysfs.c static ssize_t store_mode(struct device *device, struct device_attribute *attr, · line 73

Quoted 9 lines
drivers/video/fbdev/core/fbsysfs.c +1 -1
@@ -64,7 +67,7 @@ static ssize_t store_mode(struct device *device, struct device_attribute *attr,
 
 	list_for_each_entry(modelist, &fb_info->modelist, list) {
 		mode = &modelist->mode;
-		i = mode_string(mstr, 0, mode);
+		i = mode_string(mstr, sizeof(mstr), 0, mode);
 		if (strncmp(mstr, buf, max(count, i)) == 0) {
 
 			var = fb_info->var;

[Severity: High]
This is a pre-existing issue, but seeing the struct copy here prompted a question about other sysfs store functions in this file.

In store_mode(), var = fb_info->var; is protected by lock_fb_info(fb_info) acquired earlier in the function. However, other store functions in this file perform this same large struct copy locklessly before acquiring the lock:

drivers/video/fbdev/core/fbsysfs.c:store_bpp() {
    ...
    var = fb_info->var;
    ...
}

Similar unprotected copies occur in store_rotate(), store_virtual(), and store_pan().

Can these unprotected copies lead to torn reads if executed concurrently with an ioctl or another sysfs write that modifies the display state?

If a torn read occurs and a single field is updated before passing the torn structure to activate(), could this commit corrupted state back to the hardware?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
maintainer
#6 lore raw

Re: [PATCH v3 1/3] fbdev: bound mode sysfs output to the sysfs buffer

In reply to Melbin K Mathew
Quoted 10 lines · Melbin K Mathew

mode_string() uses snprintf() which can return a value larger than the remaining buffer space. show_modes() accumulates the return value into i without checking whether i has reached PAGE_SIZE, causing the offset to advance past the sysfs buffer if the modelist is long enough.

Add a size parameter to mode_string() and use scnprintf() to return only the bytes actually written. Add an early return when offset already exceeds the buffer. In show_modes(), stop accumulating once the buffer is full.

Quoted 5 lines
---
  drivers/video/fbdev/core/fbsysfs.c | 15 ++++++++++-----
  1 file changed, 10 insertions(+), 5 deletions(-)

I've applied the series to fbdev git tree for further testing.

Thanks!
Helge

#7 lore raw

PATCH v3 2/3 fbdev: clear fb_info->mode before deleting a videomode

In reply to Melbin K Mathew

fb_set_var() can delete a mode from info->modelist when userspace passes FB_ACTIVATE_INV_MODE through FBIOPUT_VSCREENINFO. The code checks that the mode being deleted is not the current info->var and that fbcon is not using it, but it does not check fb_info->mode.

fb_info->mode may still point into the modelist entry being deleted. If the entry is freed, later mode sysfs reads through show_mode() can dereference a stale pointer.

Clear fb_info->mode before calling fb_delete_videomode() when it matches the mode being removed.

---
 drivers/video/fbdev/core/fbmem.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
drivers/video/fbdev/core/fbmem.c +4 -1
index 2f1c56e5a7..c8aa163b0e 100644
@@ -246,8 +246,11 @@ fb_set_var(struct fb_info *info, struct fb_var_screeninfo *var)
 		ret = fb_mode_is_equal(&mode1, &mode2);
 		if (!ret) {
 			ret = fbcon_mode_deleted(info, &mode1);
-			if (!ret)
+			if (!ret) {
+				if (info->mode && fb_mode_is_equal(info->mode, &mode1))
+					info->mode = NULL;
 				fb_delete_videomode(&mode1, &info->modelist);
+			}
 		}
 
 		return ret ? -EINVAL : 0;
Signature · 1 line
2.39.5
#8 lore raw

PATCH v3 3/3 fbdev: serialize mode sysfs access with lock_fb_info()

In reply to Melbin K Mathew

show_mode(), show_modes(), and store_mode() access fb_info->modelist and fb_info->mode without holding lock_fb_info(). store_modes() takes lock_fb_info() while replacing the modelist and freeing the old one.

A concurrent reader or writer can load a pointer to an old modelist entry before store_modes() frees it, then dereference freed memory or store a stale freed pointer in fb_info->mode.

Take lock_fb_info() in show_mode(), show_modes(), and store_mode() to serialize with store_modes(). In show_mode(), copy the mode to the stack and format after dropping the lock. In store_mode(), split activate() into a _locked variant to avoid double-locking, and hold the locks for the modelist walk, mode conversion, activation, and fb_info->mode assignment together.

---
 drivers/video/fbdev/core/fbsysfs.c | 46 ++++++++++++++++++++++++------
 1 file changed, 38 insertions(+), 8 deletions(-)
drivers/video/fbdev/core/fbsysfs.c +38 -8
index af21dc5052..d3d60c555b 100644
@@ -12,19 +12,24 @@
 #include "fb_internal.h"
 #include "fbcon.h"
 
+static int activate_locked(struct fb_info *fb_info,
+			    struct fb_var_screeninfo *var)
+{
+	var->activate |= FB_ACTIVATE_FORCE;
+	return fb_set_var_from_user(fb_info, var);
+}
+
 static int activate(struct fb_info *fb_info, struct fb_var_screeninfo *var)
 {
 	int err;
 
-	var->activate |= FB_ACTIVATE_FORCE;
 	console_lock();
 	lock_fb_info(fb_info);
-	err = fb_set_var_from_user(fb_info, var);
+	err = activate_locked(fb_info, var);
 	unlock_fb_info(fb_info);
 	console_unlock();
-	if (err)
-		return err;
-	return 0;
+
+	return err;
 }
 
 static int mode_string(char *buf, size_t size, unsigned int offset,
@@ -65,6 +70,9 @@ static ssize_t store_mode(struct device *device, struct device_attribute *attr,
 
 	memset(&var, 0, sizeof(var));
 
+	console_lock();
+	lock_fb_info(fb_info);
+
 	list_for_each_entry(modelist, &fb_info->modelist, list) {
 		mode = &modelist->mode;
 		i = mode_string(mstr, sizeof(mstr), 0, mode);
@@ -72,12 +80,22 @@ static ssize_t store_mode(struct device *device, struct device_attribute *attr,
 
 			var = fb_info->var;
 			fb_videomode_to_var(&var, mode);
-			if ((err = activate(fb_info, &var)))
+			err = activate_locked(fb_info, &var);
+			if (err) {
+				unlock_fb_info(fb_info);
+				console_unlock();
 				return err;
+			}
 			fb_info->mode = mode;
+			unlock_fb_info(fb_info);
+			console_unlock();
 			return count;
 		}
 	}
+
+	unlock_fb_info(fb_info);
+	console_unlock();
+
 	return -EINVAL;
 }
 
@@ -85,11 +103,20 @@ static ssize_t show_mode(struct device *device, struct device_attribute *attr,
 			 char *buf)
 {
 	struct fb_info *fb_info = dev_get_drvdata(device);
+	struct fb_videomode mode;
+	bool have_mode = false;
 
-	if (!fb_info->mode)
+	lock_fb_info(fb_info);
+	if (fb_info->mode) {
+		mode = *fb_info->mode;
+		have_mode = true;
+	}
+	unlock_fb_info(fb_info);
+
+	if (!have_mode)
 		return 0;
 
-	return mode_string(buf, PAGE_SIZE, 0, fb_info->mode);
+	return mode_string(buf, PAGE_SIZE, 0, &mode);
 }
 
 static ssize_t store_modes(struct device *device,
@@ -137,12 +164,15 @@ static ssize_t show_modes(struct device *device, struct device_attribute *attr,
 	const struct fb_videomode *mode;
 
 	i = 0;
+	lock_fb_info(fb_info);
 	list_for_each_entry(modelist, &fb_info->modelist, list) {
 		mode = &modelist->mode;
 		i += mode_string(buf, PAGE_SIZE, i, mode);
 		if (i >= PAGE_SIZE - 1)
 			break;
 	}
+	unlock_fb_info(fb_info);
+
 	return i;
 }
 
Signature · 1 line
2.39.5
#9 lore raw

Re: [PATCH v3 3/3] fbdev: serialize mode sysfs access with lock_fb_info()

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [High] show_modes() iterates fb_info->modelist without holding console_lock(), which is required to protect against concurrent list modifications by fbcon.
- [High] Sysfs store handlers (store_bpp, store_rotate, store_virtual, store_pan, store_blank) bypass lock_fb_info(), risking torn reads and concurrent driver execution. --

commit 52388e2c8c0d0f6bb72f9bedf44a38f798ee348d
Author: Melbin K Mathew <[email protected]>

fbdev: serialize mode sysfs access with lock_fb_info()

show_mode(), show_modes(), and store_mode() access fb_info->modelist and fb_info->mode without holding lock_fb_info(). This commit adds lock_fb_info() in these functions to serialize with store_modes() and avoid data races.

Quoted 4 lines
drivers/video/fbdev/core/fbsysfs.c +0 -0
index af21dc5052df3..d3d60c555bb86 100644

[ ... ]

drivers/video/fbdev/core/fbsysfs.c static ssize_t show_modes(struct device *device, struct device_attribute *attr, · line 168

Quoted 6 lines
drivers/video/fbdev/core/fbsysfs.c +1 -0
@@ -137,12 +164,15 @@ static ssize_t show_modes(struct device *device, struct device_attribute *attr,
 	const struct fb_videomode *mode;
 
 	i = 0;
+	lock_fb_info(fb_info);
 	list_for_each_entry(modelist, &fb_info->modelist, list) {

[Severity: High]
This is a pre-existing issue, but show_modes() iterates fb_info->modelist without holding console_lock(), which is required to protect against concurrent list modifications by fbcon.

Can concurrent VT switches call fbcon_switch() -> fb_set_var() -> fb_add_videomode() and modify fb_info->modelist under console_lock() without holding lock_fb_info()? Would an unprivileged user reading this sysfs attribute risk a torn pointer or out-of-bounds access since console_lock() is missing here?

drivers/video/fbdev/core/fbsysfs.c static ssize_t show_modes(struct device *device, struct device_attribute *attr, · line 177

Quoted 9 lines
drivers/video/fbdev/core/fbsysfs.c +2 -0
 		mode = &modelist->mode;
 		i += mode_string(buf, PAGE_SIZE, i, mode);
 		if (i >= PAGE_SIZE - 1)
 			break;
 	}
+	unlock_fb_info(fb_info);
+
 	return i;
 }

[Severity: High]
This isn't a bug introduced by this patch, but other sysfs store handlers (store_bpp, store_rotate, store_virtual, store_pan, store_blank) bypass lock_fb_info(), risking torn reads and concurrent driver execution.

For example, in store_bpp(), store_rotate(), and store_virtual(), fb_info->var is read locklessly before activating it. In store_pan() and store_blank(), driver functions like fb_pan_display() or fb_blank_from_user() are invoked without taking lock_fb_info().

Does this bypass the mutual exclusion that drivers rely on in the equivalent ioctl paths? Could this result in applying corrupt configurations to hardware or concurrent execution of non-reentrant driver code?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=3
#10 lore raw

Re: [PATCH v3 3/3] fbdev: serialize mode sysfs access with lock_fb_info()

Thanks for the review.

I think these are broader pre-existing fbdev sysfs locking observations rather than blockers for this focused series.

This series addresses the mode/modelist lifetime issue and the mode sysfs output bounds issue. The modelist lifetime paths touched here are now serialized with lock_fb_info(), including store_mode(), show_mode(), and show_modes(), and the FB_ACTIVATE_INV_MODE deletion path clears fb_info->mode before deleting a matching mode.

For show_modes(), lock_fb_info() serializes against store_modes() replacing and freeing fb_info->modelist. I do not see fbcon itself mutating fb_info->modelist in the same way store_modes() does, so I am not adding console_lock() there unless maintainers prefer that locking discipline.

The other fb_info->var accesses in sysfs store handlers look like a separate wider cleanup, which I can look at as a follow-up if maintainers prefer.

Quoted 69 lines · On 02/07/2026 00:56, [email protected] wrote:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [High] show_modes() iterates fb_info->modelist without holding console_lock(), which is required to protect against concurrent list modifications by fbcon.
- [High] Sysfs store handlers (store_bpp, store_rotate, store_virtual, store_pan, store_blank) bypass lock_fb_info(), risking torn reads and concurrent driver execution. --

commit 52388e2c8c0d0f6bb72f9bedf44a38f798ee348d
Author: Melbin K Mathew <[email protected]>

fbdev: serialize mode sysfs access with lock_fb_info()

show_mode(), show_modes(), and store_mode() access fb_info->modelist and fb_info->mode without holding lock_fb_info(). This commit adds lock_fb_info() in these functions to serialize with store_modes() and avoid data races.

Quoted 4 lines
drivers/video/fbdev/core/fbsysfs.c +0 -0
index af21dc5052df3..d3d60c555bb86 100644

[ ... ]

drivers/video/fbdev/core/fbsysfs.c static ssize_t show_modes(struct device *device, struct device_attribute *attr, · line 168

Quoted 6 lines
drivers/video/fbdev/core/fbsysfs.c +1 -0
@@ -137,12 +164,15 @@ static ssize_t show_modes(struct device *device, struct device_attribute *attr,
  	const struct fb_videomode *mode;
 
  	i = 0;
+	lock_fb_info(fb_info);
  	list_for_each_entry(modelist, &fb_info->modelist, list) {

[Severity: High]
This is a pre-existing issue, but show_modes() iterates fb_info->modelist without holding console_lock(), which is required to protect against concurrent list modifications by fbcon.

Can concurrent VT switches call fbcon_switch() -> fb_set_var() -> fb_add_videomode() and modify fb_info->modelist under console_lock() without holding lock_fb_info()? Would an unprivileged user reading this sysfs attribute risk a torn pointer or out-of-bounds access since console_lock() is missing here?

drivers/video/fbdev/core/fbsysfs.c static ssize_t show_modes(struct device *device, struct device_attribute *attr, · line 177

Quoted 9 lines
drivers/video/fbdev/core/fbsysfs.c +2 -0
  		mode = &modelist->mode;
  		i += mode_string(buf, PAGE_SIZE, i, mode);
  		if (i >= PAGE_SIZE - 1)
  			break;
  	}
+	unlock_fb_info(fb_info);
+
  	return i;
  }

[Severity: High]
This isn't a bug introduced by this patch, but other sysfs store handlers (store_bpp, store_rotate, store_virtual, store_pan, store_blank) bypass lock_fb_info(), risking torn reads and concurrent driver execution.

For example, in store_bpp(), store_rotate(), and store_virtual(), fb_info->var is read locklessly before activating it. In store_pan() and store_blank(), driver functions like fb_pan_display() or fb_blank_from_user() are invoked without taking lock_fb_info().

Does this bypass the mutual exclusion that drivers rely on in the equivalent ioctl paths? Could this result in applying corrupt configurations to hardware or concurrent execution of non-reentrant driver code?

#11 lore raw

PATCH v2 1/2 fbdev: bound mode sysfs output to the sysfs buffer

In reply to Melbin K Mathew

mode_string() uses snprintf() which can return a value larger than the remaining buffer space. show_modes() accumulates the return value into i without checking whether i has reached PAGE_SIZE, causing the offset to advance past the sysfs buffer if the modelist is long enough.

Add a size parameter to mode_string() and use scnprintf() to return only the bytes actually written. Add an early return when offset already exceeds the buffer. In show_modes(), stop accumulating once the buffer is full.

---
 drivers/video/fbdev/core/fbsysfs.c | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)
drivers/video/fbdev/core/fbsysfs.c +10 -5
index ea196603c7..af21dc5052 100644
@@ -27,12 +27,15 @@ static int activate(struct fb_info *fb_info, struct fb_var_screeninfo *var)
 	return 0;
 }
 
-static int mode_string(char *buf, unsigned int offset,
+static int mode_string(char *buf, size_t size, unsigned int offset,
 		       const struct fb_videomode *mode)
 {
 	char m = 'U';
 	char v = 'p';
 
+	if (offset >= size)
+		return 0;
+
 	if (mode->flag & FB_MODE_IS_DETAILED)
 		m = 'D';
 	if (mode->flag & FB_MODE_IS_VESA)
@@ -45,7 +48,7 @@ static int mode_string(char *buf, unsigned int offset,
 	if (mode->vmode & FB_VMODE_DOUBLE)
 		v = 'd';
 
-	return snprintf(&buf[offset], PAGE_SIZE - offset, "%c:%dx%d%c-%d\n",
+	return scnprintf(&buf[offset], size - offset, "%c:%dx%d%c-%d\n",
 	                m, mode->xres, mode->yres, v, mode->refresh);
 }
 
@@ -64,7 +67,7 @@ static ssize_t store_mode(struct device *device, struct device_attribute *attr,
 
 	list_for_each_entry(modelist, &fb_info->modelist, list) {
 		mode = &modelist->mode;
-		i = mode_string(mstr, 0, mode);
+		i = mode_string(mstr, sizeof(mstr), 0, mode);
 		if (strncmp(mstr, buf, max(count, i)) == 0) {
 
 			var = fb_info->var;
@@ -86,7 +89,7 @@ static ssize_t show_mode(struct device *device, struct device_attribute *attr,
 	if (!fb_info->mode)
 		return 0;
 
-	return mode_string(buf, 0, fb_info->mode);
+	return mode_string(buf, PAGE_SIZE, 0, fb_info->mode);
 }
 
 static ssize_t store_modes(struct device *device,
@@ -136,7 +139,9 @@ static ssize_t show_modes(struct device *device, struct device_attribute *attr,
 	i = 0;
 	list_for_each_entry(modelist, &fb_info->modelist, list) {
 		mode = &modelist->mode;
-		i += mode_string(buf, i, mode);
+		i += mode_string(buf, PAGE_SIZE, i, mode);
+		if (i >= PAGE_SIZE - 1)
+			break;
 	}
 	return i;
 }
Signature · 1 line
2.39.5
#12 lore raw

Re: [PATCH v2 1/2] fbdev: bound mode sysfs output to the sysfs buffer

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] Lockless reads of fb_info->var cause torn reads and state corruption. --

commit b5317fd1a6886d886e344587e27b5ae1721b2542
Author: Melbin K Mathew <[email protected]>

fbdev: bound mode sysfs output to the sysfs buffer

mode_string() uses snprintf() which can return a value larger than the remaining buffer space. show_modes() accumulates the return value into i without checking whether i has reached PAGE_SIZE. This commit adds a size parameter to mode_string() and uses scnprintf() to return only the bytes actually written.

Quoted 4 lines
drivers/video/fbdev/core/fbsysfs.c +0 -0
index ea196603c7a87..af21dc5052df3 100644

[ ... ]

drivers/video/fbdev/core/fbsysfs.c static ssize_t store_mode(struct device *device, struct device_attribute *attr, · line 73

Quoted 9 lines
drivers/video/fbdev/core/fbsysfs.c +1 -1
@@ -64,7 +67,7 @@ static ssize_t store_mode(struct device *device, struct device_attribute *attr,
 
 	list_for_each_entry(modelist, &fb_info->modelist, list) {
 		mode = &modelist->mode;
-		i = mode_string(mstr, 0, mode);
+		i = mode_string(mstr, sizeof(mstr), 0, mode);
 		if (strncmp(mstr, buf, max(count, i)) == 0) {
 
 			var = fb_info->var;

[Severity: High]
This isn't a bug introduced by this patch, but while looking at fbsysfs.c, I noticed several other functions perform a lockless copy of the fb_info->var struct:

    var = fb_info->var;

This happens in store_bpp(), store_rotate(), store_virtual(), and store_pan().

If this lockless sysfs read races with an ioctl(FBIOPUT_VSCREENINFO) executing fb_set_var() (which modifies info->var under fb_info->lock), could it tear the large struct fb_var_screeninfo?

Modifying and writing this corrupted state back to the hardware via activate() or fb_pan_display() might configure the display controller with invalid settings, potentially causing display corruption or crashes in low-level graphics drivers.

Should these other functions also serialize their access with lock_fb_info()?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#13 lore raw

PATCH v2 2/2 fbdev: serialize mode sysfs access with lock_fb_info()

In reply to Melbin K Mathew

show_mode(), show_modes(), and store_mode() access fb_info->modelist and fb_info->mode without holding lock_fb_info(). store_modes() takes lock_fb_info() while replacing the modelist and freeing the old one.

A concurrent reader or writer can load a pointer to an old modelist entry before store_modes() frees it, then dereference freed memory or store a stale freed pointer in fb_info->mode.

Take lock_fb_info() in show_mode(), show_modes(), and store_mode() to serialize with store_modes(). In show_mode(), copy the mode to the stack and format after dropping the lock. In store_mode(), split activate() into a _locked variant to avoid double-locking, and hold the locks for the modelist walk, mode conversion, activation, and fb_info->mode assignment together.

---
 drivers/video/fbdev/core/fbsysfs.c | 46 ++++++++++++++++++++++++------
 1 file changed, 38 insertions(+), 8 deletions(-)
drivers/video/fbdev/core/fbsysfs.c +38 -8
index af21dc5052..d3d60c555b 100644
@@ -12,19 +12,24 @@
 #include "fb_internal.h"
 #include "fbcon.h"
 
+static int activate_locked(struct fb_info *fb_info,
+			    struct fb_var_screeninfo *var)
+{
+	var->activate |= FB_ACTIVATE_FORCE;
+	return fb_set_var_from_user(fb_info, var);
+}
+
 static int activate(struct fb_info *fb_info, struct fb_var_screeninfo *var)
 {
 	int err;
 
-	var->activate |= FB_ACTIVATE_FORCE;
 	console_lock();
 	lock_fb_info(fb_info);
-	err = fb_set_var_from_user(fb_info, var);
+	err = activate_locked(fb_info, var);
 	unlock_fb_info(fb_info);
 	console_unlock();
-	if (err)
-		return err;
-	return 0;
+
+	return err;
 }
 
 static int mode_string(char *buf, size_t size, unsigned int offset,
@@ -65,6 +70,9 @@ static ssize_t store_mode(struct device *device, struct device_attribute *attr,
 
 	memset(&var, 0, sizeof(var));
 
+	console_lock();
+	lock_fb_info(fb_info);
+
 	list_for_each_entry(modelist, &fb_info->modelist, list) {
 		mode = &modelist->mode;
 		i = mode_string(mstr, sizeof(mstr), 0, mode);
@@ -72,12 +80,22 @@ static ssize_t store_mode(struct device *device, struct device_attribute *attr,
 
 			var = fb_info->var;
 			fb_videomode_to_var(&var, mode);
-			if ((err = activate(fb_info, &var)))
+			err = activate_locked(fb_info, &var);
+			if (err) {
+				unlock_fb_info(fb_info);
+				console_unlock();
 				return err;
+			}
 			fb_info->mode = mode;
+			unlock_fb_info(fb_info);
+			console_unlock();
 			return count;
 		}
 	}
+
+	unlock_fb_info(fb_info);
+	console_unlock();
+
 	return -EINVAL;
 }
 
@@ -85,11 +103,20 @@ static ssize_t show_mode(struct device *device, struct device_attribute *attr,
 			 char *buf)
 {
 	struct fb_info *fb_info = dev_get_drvdata(device);
+	struct fb_videomode mode;
+	bool have_mode = false;
 
-	if (!fb_info->mode)
+	lock_fb_info(fb_info);
+	if (fb_info->mode) {
+		mode = *fb_info->mode;
+		have_mode = true;
+	}
+	unlock_fb_info(fb_info);
+
+	if (!have_mode)
 		return 0;
 
-	return mode_string(buf, PAGE_SIZE, 0, fb_info->mode);
+	return mode_string(buf, PAGE_SIZE, 0, &mode);
 }
 
 static ssize_t store_modes(struct device *device,
@@ -137,12 +164,15 @@ static ssize_t show_modes(struct device *device, struct device_attribute *attr,
 	const struct fb_videomode *mode;
 
 	i = 0;
+	lock_fb_info(fb_info);
 	list_for_each_entry(modelist, &fb_info->modelist, list) {
 		mode = &modelist->mode;
 		i += mode_string(buf, PAGE_SIZE, i, mode);
 		if (i >= PAGE_SIZE - 1)
 			break;
 	}
+	unlock_fb_info(fb_info);
+
 	return i;
 }
 
Signature · 1 line
2.39.5
#14 lore raw

Re: [PATCH v2 2/2] fbdev: serialize mode sysfs access with lock_fb_info()

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] fb_info->mode can become a dangling pointer if the active mode is deleted via ioctl(FBIOPUT_VSCREENINFO) with FB_ACTIVATE_INV_MODE, leading to a Use-After-Free in show_mode(). --

commit 243c82e865602245e21ae16b7cdd7738ccfa6a53
Author: Melbin K Mathew <[email protected]>

fbdev: serialize mode sysfs access with lock_fb_info()

Take lock_fb_info() in show_mode(), show_modes(), and store_mode() to serialize with store_modes(). In show_mode(), copy the mode to the stack and format after dropping the lock. In store_mode(), split activate() into a _locked variant to avoid double-locking.

Quoted 4 lines
drivers/video/fbdev/core/fbsysfs.c +0 -0
index af21dc5052df3..d3d60c555bb86 100644

[ ... ]

drivers/video/fbdev/core/fbsysfs.c static ssize_t show_mode(struct device *device, struct device_attribute *attr, · line 111

Quoted 11 lines
drivers/video/fbdev/core/fbsysfs.c +5 -1
@@ -85,11 +103,20 @@ static ssize_t show_mode(struct device *device, struct device_attribute *attr,
 			 char *buf)
 {
 	struct fb_info *fb_info = dev_get_drvdata(device);
+	struct fb_videomode mode;
+	bool have_mode = false;
 
-	if (!fb_info->mode)
+	lock_fb_info(fb_info);
+	if (fb_info->mode) {
+		mode = *fb_info->mode;

[Severity: High]
This is a pre-existing issue, but can fb_info->mode become a dangling pointer and lead to a use-after-free here?

If the active mode is deleted via ioctl(FBIOPUT_VSCREENINFO) with FB_ACTIVATE_INV_MODE, it appears fb_delete_videomode() unlinks and frees the mode node:

drivers/video/fbdev/core/modedb.c:fb_delete_videomode() {
    ...
    list_del(pos);
    kfree(pos);
    ...
}

However, it does not seem to clear the fb_info->mode reference. If a user subsequently reads /sys/class/graphics/fb0/mode, would this dereference the freed memory even while holding the lock?

Quoted 10 lines
drivers/video/fbdev/core/fbsysfs.c +6 -1
+		have_mode = true;
+	}
+	unlock_fb_info(fb_info);
+
+	if (!have_mode)
 		return 0;
 
-	return mode_string(buf, PAGE_SIZE, 0, fb_info->mode);
+	return mode_string(buf, PAGE_SIZE, 0, &mode);
 }
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=2