fbmem: prevent potential use-after-free issues with console_lock()
Subsystems (2)
PATCH 1/1 fbmem: prevent potential use-after-free issues with console_lock()
In do_fb_ioctl(), user specified "fb_info" can be freed in the callee fbcon_get_con2fb_map_ioctl() -> set_con2fb_map() -> con2fb_release_oldinfo(), this free operation is protected by console_lock() in fbcon_set_con2fb_map_ioctl(), it also results in the change of certain states such as "minfo->dead" in matroxfb_remove(), so that it can be checked to avoid use-after-free before the use sites (e.g., the check at the beginning of matroxfb_ioctl()). However, the problem is that the use site is not protected by the same locks as for the free operation, e.g., "default" case in do_fb_ioctl() can lead to "matroxfb_ioctl()" but it's not protected by console_lock(), which can invalidate the aforementioned state set and check in a concurrent setting.
Prevent the potential use-after-free issues by protecting the "default" case in do_fb_ioctl() with console_lock(), similarly as for many other cases like "case FBIOBLANK" and "case FBIOPAN_DISPLAY".
--- drivers/video/fbdev/core/fbmem.c | 2 ++ 1 file changed, 2 insertions(+)
drivers/video/fbdev/core/fbmem.c +2 -0
@@ -1182,6 +1182,7 @@ static long do_fb_ioctl(struct fb_info *info, unsigned int cmd,console_unlock ();break ;default : +console_lock ();lock_fb_info (info); fb = info->fbops;if (fb->fb_ioctl)
@@ -1189,6 +1190,7 @@ static long do_fb_ioctl(struct fb_info *info, unsigned int cmd,else ret = -ENOTTY ;unlock_fb_info (info); +console_unlock (); }return ret; }
Signature · 1 line
2.39.0
Quoted 13 lines · Hang Zhang
In
do_fb_ioctl(), user specified "fb_info" can be freed in the calleefbcon_get_con2fb_map_ioctl()->set_con2fb_map()->con2fb_release_oldinfo(), this free operation is protected byconsole_lock()infbcon_set_con2fb_map_ioctl(), it also results in the change of certain states such as "minfo->dead" inmatroxfb_remove(), so that it can be checked to avoid use-after-free before the use sites (e.g., the check at the beginning ofmatroxfb_ioctl()). However, the problem is that the use site is not protected by the same locks as for the free operation, e.g., "default" case indo_fb_ioctl()can lead to "matroxfb_ioctl()" but it's not protected byconsole_lock(), which can invalidate the aforementioned state set and check in a concurrent setting.
Quoted 5 lines
Prevent the potential use-after-free issues by protecting the "default" case in
do_fb_ioctl()withconsole_lock(), similarly as for many other cases like "case FBIOBLANK" and "caseFBIOPAN_DISPLAY".Signed-off-by Hang Zhang <[email protected]>
applied to fbdev git tree.
Thanks,
Helge
Quoted 24 lines
--- drivers/video/fbdev/core/fbmem.c | 2 ++ 1 file changed, 2 insertions(+)drivers/video/fbdev/core/fbmem.c +2
-0@@ -1182,6 +1182,7 @@ static long do_fb_ioctl(struct fb_info *info, unsigned int cmd,console_unlock ();break ;default : +console_lock ();lock_fb_info (info); fb = info->fbops;if (fb->fb_ioctl)@@ -1189,6 +1190,7 @@ static long do_fb_ioctl(struct fb_info *info, unsigned int cmd,else ret = -ENOTTY ;unlock_fb_info (info); +console_unlock (); }return ret; }
Hi Helge
Quoted 15 lines · Helge Deller
Quoted 13 lines · Hang Zhang
In
do_fb_ioctl(), user specified "fb_info" can be freed in the calleefbcon_get_con2fb_map_ioctl()->set_con2fb_map()->con2fb_release_oldinfo(), this free operation is protected byconsole_lock()infbcon_set_con2fb_map_ioctl(), it also results in the change of certain states such as "minfo->dead" inmatroxfb_remove(), so that it can be checked to avoid use-after-free before the use sites (e.g., the check at the beginning ofmatroxfb_ioctl()). However, the problem is that the use site is not protected by the same locks as for the free operation, e.g., "default" case indo_fb_ioctl()can lead to "matroxfb_ioctl()" but it's not protected byconsole_lock(), which can invalidate the aforementioned state set and check in a concurrent setting.
Quoted 7 lines
Quoted 5 lines
Prevent the potential use-after-free issues by protecting the "default" case in
do_fb_ioctl()withconsole_lock(), similarly as for many other cases like "case FBIOBLANK" and "caseFBIOPAN_DISPLAY".Signed-off-by Hang Zhang <[email protected]>applied to fbdev git tree.
The patch above makes no sense at all to me:
- fb_info is protected by lock_fb_info and
- the lifetime of fb_info is protected by the get/put functions
- yes there's the interaction with con2fb, which is protected by console_lock, but the lifetime guarantees are ensured by the device removal
- which means any stuff happening in matroxfb_remove is also not a concern here (unless matroxfb completely gets all the device lifetime stuff wrong, but it doesn't look like it's any worse than any of the other fbdev drivers that we haven't recently fixed up due to the takeover issues with firmware drivers
On the very clear downside this now means we take console_lock for the vblank ioctl (which is a device driver extension for reasons, despite that it's a standard fbdev ioctl), which is no good at all given how console_lock() is a really expensive lock.
Unless I'm massively missing something, can you pls push the revert before this lands in Linus' tree?
Thanks, Daniel
Quoted 28 lines
Thanks,
HelgeQuoted 24 lines
--- drivers/video/fbdev/core/fbmem.c | 2 ++ 1 file changed, 2 insertions(+)drivers/video/fbdev/core/fbmem.c +2
-0@@ -1182,6 +1182,7 @@ static long do_fb_ioctl(struct fb_info *info, unsigned int cmd,console_unlock ();break ;default : +console_lock ();lock_fb_info (info); fb = info->fbops;if (fb->fb_ioctl)@@ -1189,6 +1190,7 @@ static long do_fb_ioctl(struct fb_info *info, unsigned int cmd,else ret = -ENOTTY ;unlock_fb_info (info); +console_unlock (); }return ret; }
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
Quoted 29 lines · Daniel Vetter
Hi Helge
Quoted 15 lines · Helge Deller
Quoted 13 lines · Hang Zhang
In
do_fb_ioctl(), user specified "fb_info" can be freed in the calleefbcon_get_con2fb_map_ioctl()->set_con2fb_map()->con2fb_release_oldinfo(), this free operation is protected byconsole_lock()infbcon_set_con2fb_map_ioctl(), it also results in the change of certain states such as "minfo->dead" inmatroxfb_remove(), so that it can be checked to avoid use-after-free before the use sites (e.g., the check at the beginning ofmatroxfb_ioctl()). However, the problem is that the use site is not protected by the same locks as for the free operation, e.g., "default" case indo_fb_ioctl()can lead to "matroxfb_ioctl()" but it's not protected byconsole_lock(), which can invalidate the aforementioned state set and check in a concurrent setting.Quoted 7 lines
Quoted 5 lines
Prevent the potential use-after-free issues by protecting the "default" case in
do_fb_ioctl()withconsole_lock(), similarly as for many other cases like "case FBIOBLANK" and "caseFBIOPAN_DISPLAY".Signed-off-by Hang Zhang <[email protected]>applied to fbdev git tree.
The patch above makes no sense at all to me:
Quoted 10 lines
-
fb_infois protected bylock_fb_infoand
- the lifetime offb_infois protected by the get/put functions
- yes there's the interaction with con2fb, which is protected byconsole_lock, but the lifetime guarantees are ensured by the device removal
- which means any stuff happening inmatroxfb_removeis also not a concern here (unless matroxfb completely gets all the device lifetime stuff wrong, but it doesn't look like it's any worse than any of the other fbdev drivers that we haven't recently fixed up due to the takeover issues with firmware drivers
I have also a really hard timing finding the con2fb map use in the matroxfb ioctl code, but that just might be that I didn't look carefully enough. Maybe that would shed some light on this. -Daniel
Quoted 45 lines
On the very clear downside this now means we take
console_lockfor the vblank ioctl (which is a device driver extension for reasons, despite that it's a standard fbdev ioctl), which is no good at all given howconsole_lock()is a really expensive lock.Unless I'm massively missing something, can you pls push the revert before this lands in Linus' tree?
Thanks, Daniel
Quoted 19 lines
Thanks,
HelgeQuoted 16 lines
--- drivers/video/fbdev/core/fbmem.c | 2 ++ 1 file changed, 2 insertions(+)drivers/video/fbdev/core/fbmem.c +1
-0@@ -1182,6 +1182,7 @@ static long do_fb_ioctl(struct fb_info *info, unsigned int cmd,console_unlock ();break ;default : +console_lock ();lock_fb_info (info); fb = info->fbops;if (fb->fb_ioctl)Quoted 9 lines
drivers/video/fbdev/core/fbmem.c static long do_fb_ioctl(struct fb_info *info, unsigned int cmd, · line 1196
Quoted 8 lines
drivers/video/fbdev/core/fbmem.c +1
-0@@ -1189,6 +1190,7 @@ static long do_fb_ioctl(struct fb_info *info, unsigned int cmd,else ret = -ENOTTY ;unlock_fb_info (info); +console_unlock (); }return ret; }--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
Quoted 42 lines · Daniel Vetter
Quoted 29 lines · Daniel Vetter
Hi Helge
Quoted 15 lines · Helge Deller
Quoted 13 lines · Hang Zhang
In
do_fb_ioctl(), user specified "fb_info" can be freed in the calleefbcon_get_con2fb_map_ioctl()->set_con2fb_map()->con2fb_release_oldinfo(), this free operation is protected byconsole_lock()infbcon_set_con2fb_map_ioctl(), it also results in the change of certain states such as "minfo->dead" inmatroxfb_remove(), so that it can be checked to avoid use-after-free before the use sites (e.g., the check at the beginning ofmatroxfb_ioctl()). However, the problem is that the use site is not protected by the same locks as for the free operation, e.g., "default" case indo_fb_ioctl()can lead to "matroxfb_ioctl()" but it's not protected byconsole_lock(), which can invalidate the aforementioned state set and check in a concurrent setting.Quoted 7 lines
Quoted 5 lines
Prevent the potential use-after-free issues by protecting the "default" case in
do_fb_ioctl()withconsole_lock(), similarly as for many other cases like "case FBIOBLANK" and "caseFBIOPAN_DISPLAY".Signed-off-by Hang Zhang <[email protected]>applied to fbdev git tree.
The patch above makes no sense at all to me:
Quoted 10 lines
-
fb_infois protected bylock_fb_infoand
- the lifetime offb_infois protected by the get/put functions
- yes there's the interaction with con2fb, which is protected byconsole_lock, but the lifetime guarantees are ensured by the device removal
- which means any stuff happening inmatroxfb_removeis also not a concern here (unless matroxfb completely gets all the device lifetime stuff wrong, but it doesn't look like it's any worse than any of the other fbdev drivers that we haven't recently fixed up due to the takeover issues with firmware drivers
Quoted 16 lines
I have also a really hard timing finding the con2fb map use in the matroxfb ioctl code, but that just might be that I didn't look carefully enough. Maybe that would shed some light on this. -Daniel
Quoted 10 lines
On the very clear downside this now means we take
console_lockfor the vblank ioctl (which is a device driver extension for reasons, despite that it's a standard fbdev ioctl), which is no good at all given howconsole_lock()is a really expensive lock.Unless I'm massively missing something, can you pls push the revert before this lands in Linus' tree?
Thanks, Daniel
Hi, Daniel. Thank you for your feedback. We're not developers of the video subsystem and thus may be short in domain knowledge (e.g., the performance of console_lock() and the complex lifetime management). This patch initially intended to bring up the potential use-after-free issues here to the community - we have performed a best-effort code review and cannot exclude the possibility based on our understanding.
What we have observed is that the call chain leading to the free site (do_fb_ioctl()->fbcon_set_con2fb_map_ioctl()->set_con2fb_map()-> con2fb_release_oldinfo()-> ... ->matroxfb_remove()) is only protected by console_lock() but not lock_fb_info(), while the potential use site (call chain starts from the default case in do_fb_ioctl()) is only protected by lock_fb_info() but not console_lock(). We thus propose to add this extra console_lock() to the default case, which is inspired by the lock protection of many other existing switch-case terms in the same function.
Since we do not have deep domain knowledge of this subsystem, we will rely on the developers to make a decision regarding the patch. Thank you again for your review and help!
Best,
Hang
Quoted 42 lines
Quoted 31 lines
Quoted 20 lines
Thanks,
HelgeQuoted 16 lines
--- drivers/video/fbdev/core/fbmem.c | 2 ++ 1 file changed, 2 insertions(+)drivers/video/fbdev/core/fbmem.c +1
-0@@ -1182,6 +1182,7 @@ static long do_fb_ioctl(struct fb_info *info, unsigned int cmd,console_unlock ();break ;default : +console_lock ();lock_fb_info (info); fb = info->fbops;if (fb->fb_ioctl)Quoted 9 lines
drivers/video/fbdev/core/fbmem.c static long do_fb_ioctl(struct fb_info *info, unsigned int cmd, · line 1196
Quoted 8 lines
drivers/video/fbdev/core/fbmem.c +1
-0@@ -1189,6 +1190,7 @@ static long do_fb_ioctl(struct fb_info *info, unsigned int cmd,else ret = -ENOTTY ;unlock_fb_info (info); +console_unlock (); }return ret; }Quoted 4 lines
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
Quoted 77 lines · Hang Zhang
Quoted 42 lines · Daniel Vetter
Quoted 29 lines · Daniel Vetter
Hi Helge
Quoted 15 lines · Helge Deller
Quoted 13 lines · Hang Zhang
In
do_fb_ioctl(), user specified "fb_info" can be freed in the calleefbcon_get_con2fb_map_ioctl()->set_con2fb_map()->con2fb_release_oldinfo(), this free operation is protected byconsole_lock()infbcon_set_con2fb_map_ioctl(), it also results in the change of certain states such as "minfo->dead" inmatroxfb_remove(), so that it can be checked to avoid use-after-free before the use sites (e.g., the check at the beginning ofmatroxfb_ioctl()). However, the problem is that the use site is not protected by the same locks as for the free operation, e.g., "default" case indo_fb_ioctl()can lead to "matroxfb_ioctl()" but it's not protected byconsole_lock(), which can invalidate the aforementioned state set and check in a concurrent setting.Quoted 7 lines
Quoted 5 lines
Prevent the potential use-after-free issues by protecting the "default" case in
do_fb_ioctl()withconsole_lock(), similarly as for many other cases like "case FBIOBLANK" and "caseFBIOPAN_DISPLAY".Signed-off-by Hang Zhang <[email protected]>applied to fbdev git tree.
The patch above makes no sense at all to me:
Quoted 10 lines
-
fb_infois protected bylock_fb_infoand
- the lifetime offb_infois protected by the get/put functions
- yes there's the interaction with con2fb, which is protected byconsole_lock, but the lifetime guarantees are ensured by the device removal
- which means any stuff happening inmatroxfb_removeis also not a concern here (unless matroxfb completely gets all the device lifetime stuff wrong, but it doesn't look like it's any worse than any of the other fbdev drivers that we haven't recently fixed up due to the takeover issues with firmware driversQuoted 16 lines
I have also a really hard timing finding the con2fb map use in the matroxfb ioctl code, but that just might be that I didn't look carefully enough. Maybe that would shed some light on this. -Daniel
Quoted 10 lines
On the very clear downside this now means we take
console_lockfor the vblank ioctl (which is a device driver extension for reasons, despite that it's a standard fbdev ioctl), which is no good at all given howconsole_lock()is a really expensive lock.Unless I'm massively missing something, can you pls push the revert before this lands in Linus' tree?
Thanks, Daniel
Hi, Daniel. Thank you for your feedback. We're not developers of the video subsystem and thus may be short in domain knowledge (e.g., the performance of
console_lock()and the complex lifetime management). This patch initially intended to bring up the potential use-after-free issues here to the community - we have performed a best-effort code review and cannot exclude the possibility based on our understanding.What we have observed is that the call chain leading to the free site (do_fb_ioctl()->fbcon_set_con2fb_map_ioctl()->set_con2fb_map()-> con2fb_release_oldinfo()-> ... ->matroxfb_remove()) is only protected by
console_lock()but notlock_fb_info(), while the potential use site (call chain starts from the default case indo_fb_ioctl()) is only protected bylock_fb_info()but notconsole_lock(). We thus propose to add this extraconsole_lock()to the default case, which is inspired by the lock protection of many other existing switch-case terms in the same function.
Quoted 3 lines
Since we do not have deep domain knowledge of this subsystem, we will rely on the developers to make a decision regarding the patch. Thank you again for your review and help!
Can you please elaborate where you've found this use-after-free and how? I'm still not understanding how you even got here - this is orthogonal to whether the patch is the right fix or not.
-Daniel
Quoted 46 lines
Best,
HangQuoted 42 lines
Quoted 31 lines
Quoted 20 lines
Thanks,
HelgeQuoted 16 lines
--- drivers/video/fbdev/core/fbmem.c | 2 ++ 1 file changed, 2 insertions(+)drivers/video/fbdev/core/fbmem.c +1
-0@@ -1182,6 +1182,7 @@ static long do_fb_ioctl(struct fb_info *info, unsigned int cmd,console_unlock ();break ;default : +console_lock ();lock_fb_info (info); fb = info->fbops;if (fb->fb_ioctl)Quoted 9 lines
drivers/video/fbdev/core/fbmem.c static long do_fb_ioctl(struct fb_info *info, unsigned int cmd, · line 1196
Quoted 8 lines
drivers/video/fbdev/core/fbmem.c +1
-0@@ -1189,6 +1190,7 @@ static long do_fb_ioctl(struct fb_info *info, unsigned int cmd,else ret = -ENOTTY ;unlock_fb_info (info); +console_unlock (); }return ret; }Quoted 4 lines
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
Quoted 79 lines · Daniel Vetter
Quoted 77 lines · Hang Zhang
Quoted 42 lines · Daniel Vetter
Quoted 29 lines · Daniel Vetter
Hi Helge
Quoted 15 lines · Helge Deller
Quoted 13 lines · Hang Zhang
In
do_fb_ioctl(), user specified "fb_info" can be freed in the calleefbcon_get_con2fb_map_ioctl()->set_con2fb_map()->con2fb_release_oldinfo(), this free operation is protected byconsole_lock()infbcon_set_con2fb_map_ioctl(), it also results in the change of certain states such as "minfo->dead" inmatroxfb_remove(), so that it can be checked to avoid use-after-free before the use sites (e.g., the check at the beginning ofmatroxfb_ioctl()). However, the problem is that the use site is not protected by the same locks as for the free operation, e.g., "default" case indo_fb_ioctl()can lead to "matroxfb_ioctl()" but it's not protected byconsole_lock(), which can invalidate the aforementioned state set and check in a concurrent setting.Quoted 7 lines
Quoted 5 lines
Prevent the potential use-after-free issues by protecting the "default" case in
do_fb_ioctl()withconsole_lock(), similarly as for many other cases like "case FBIOBLANK" and "caseFBIOPAN_DISPLAY".Signed-off-by Hang Zhang <[email protected]>applied to fbdev git tree.
The patch above makes no sense at all to me:
Quoted 10 lines
-
fb_infois protected bylock_fb_infoand
- the lifetime offb_infois protected by the get/put functions
- yes there's the interaction with con2fb, which is protected byconsole_lock, but the lifetime guarantees are ensured by the device removal
- which means any stuff happening inmatroxfb_removeis also not a concern here (unless matroxfb completely gets all the device lifetime stuff wrong, but it doesn't look like it's any worse than any of the other fbdev drivers that we haven't recently fixed up due to the takeover issues with firmware driversQuoted 16 lines
I have also a really hard timing finding the con2fb map use in the matroxfb ioctl code, but that just might be that I didn't look carefully enough. Maybe that would shed some light on this. -Daniel
Quoted 10 lines
On the very clear downside this now means we take
console_lockfor the vblank ioctl (which is a device driver extension for reasons, despite that it's a standard fbdev ioctl), which is no good at all given howconsole_lock()is a really expensive lock.Unless I'm massively missing something, can you pls push the revert before this lands in Linus' tree?
Thanks, Daniel
Hi, Daniel. Thank you for your feedback. We're not developers of the video subsystem and thus may be short in domain knowledge (e.g., the performance of
console_lock()and the complex lifetime management). This patch initially intended to bring up the potential use-after-free issues here to the community - we have performed a best-effort code review and cannot exclude the possibility based on our understanding.What we have observed is that the call chain leading to the free site (do_fb_ioctl()->fbcon_set_con2fb_map_ioctl()->set_con2fb_map()-> con2fb_release_oldinfo()-> ... ->matroxfb_remove()) is only protected by
console_lock()but notlock_fb_info(), while the potential use site (call chain starts from the default case indo_fb_ioctl()) is only protected bylock_fb_info()but notconsole_lock(). We thus propose to add this extraconsole_lock()to the default case, which is inspired by the lock protection of many other existing switch-case terms in the same function.
Quoted 8 lines
Quoted 3 lines
Since we do not have deep domain knowledge of this subsystem, we will rely on the developers to make a decision regarding the patch. Thank you again for your review and help!
Can you please elaborate where you've found this use-after-free and how? I'm still not understanding how you even got here - this is orthogonal to whether the patch is the right fix or not.
-Daniel
Hi, Daniel. Sure. This issue was initially flagged by our experimental static code analyzer aiming to find use-after-free issues in the kernel - that's why we don't have PoC or execution traces here. We deeply understand that static analyzer can generate false alarms, so we have tried our best and spent a good amount of time carefully reviewing the related code. We eventually found that we could not exclude this potential issue based on our study, so we decided to report this to the community with this tentative fix. As mentioned, we may be short in domain knowledge, so your input is highly appreciated. We respect the developer's decision about whether this is really a problem and whether/how to fix it. However, if you think the use-after-free is actually not possible, it will be very helpful if you can elaborate on the reasoning since it will greatly help us improve our analyzer. Thank you very much for your help!
BTW, if this is worthed a fix and the performance of console_lock() is a major concern, then I think there may be alternative solutions like adding a lock_fb_info() to the free call chain - if that's better in performance, or maybe selectively protect the matroxfb ioctl but not vblank ioctl as you mentioned.
Thanks,
Hang
Quoted 52 lines
Quoted 47 lines
Best,
HangQuoted 42 lines
Quoted 31 lines
Quoted 20 lines
Thanks,
HelgeQuoted 16 lines
--- drivers/video/fbdev/core/fbmem.c | 2 ++ 1 file changed, 2 insertions(+)drivers/video/fbdev/core/fbmem.c +1
-0@@ -1182,6 +1182,7 @@ static long do_fb_ioctl(struct fb_info *info, unsigned int cmd,console_unlock ();break ;default : +console_lock ();lock_fb_info (info); fb = info->fbops;if (fb->fb_ioctl)Quoted 9 lines
drivers/video/fbdev/core/fbmem.c static long do_fb_ioctl(struct fb_info *info, unsigned int cmd, · line 1196
Quoted 8 lines
drivers/video/fbdev/core/fbmem.c +1
-0@@ -1189,6 +1190,7 @@ static long do_fb_ioctl(struct fb_info *info, unsigned int cmd,else ret = -ENOTTY ;unlock_fb_info (info); +console_unlock (); }return ret; }Quoted 4 lines
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
Quoted 89 lines · Hang Zhang
Quoted 79 lines · Daniel Vetter
Quoted 77 lines · Hang Zhang
Quoted 42 lines · Daniel Vetter
Quoted 29 lines · Daniel Vetter
Hi Helge
Quoted 15 lines · Helge Deller
Quoted 13 lines · Hang Zhang
In
do_fb_ioctl(), user specified "fb_info" can be freed in the calleefbcon_get_con2fb_map_ioctl()->set_con2fb_map()->con2fb_release_oldinfo(), this free operation is protected byconsole_lock()infbcon_set_con2fb_map_ioctl(), it also results in the change of certain states such as "minfo->dead" inmatroxfb_remove(), so that it can be checked to avoid use-after-free before the use sites (e.g., the check at the beginning ofmatroxfb_ioctl()). However, the problem is that the use site is not protected by the same locks as for the free operation, e.g., "default" case indo_fb_ioctl()can lead to "matroxfb_ioctl()" but it's not protected byconsole_lock(), which can invalidate the aforementioned state set and check in a concurrent setting.Quoted 7 lines
Quoted 5 lines
Prevent the potential use-after-free issues by protecting the "default" case in
do_fb_ioctl()withconsole_lock(), similarly as for many other cases like "case FBIOBLANK" and "caseFBIOPAN_DISPLAY".Signed-off-by Hang Zhang <[email protected]>applied to fbdev git tree.
The patch above makes no sense at all to me:
Quoted 10 lines
-
fb_infois protected bylock_fb_infoand
- the lifetime offb_infois protected by the get/put functions
- yes there's the interaction with con2fb, which is protected byconsole_lock, but the lifetime guarantees are ensured by the device removal
- which means any stuff happening inmatroxfb_removeis also not a concern here (unless matroxfb completely gets all the device lifetime stuff wrong, but it doesn't look like it's any worse than any of the other fbdev drivers that we haven't recently fixed up due to the takeover issues with firmware driversQuoted 16 lines
I have also a really hard timing finding the con2fb map use in the matroxfb ioctl code, but that just might be that I didn't look carefully enough. Maybe that would shed some light on this. -Daniel
Quoted 10 lines
On the very clear downside this now means we take
console_lockfor the vblank ioctl (which is a device driver extension for reasons, despite that it's a standard fbdev ioctl), which is no good at all given howconsole_lock()is a really expensive lock.Unless I'm massively missing something, can you pls push the revert before this lands in Linus' tree?
Thanks, Daniel
Hi, Daniel. Thank you for your feedback. We're not developers of the video subsystem and thus may be short in domain knowledge (e.g., the performance of
console_lock()and the complex lifetime management). This patch initially intended to bring up the potential use-after-free issues here to the community - we have performed a best-effort code review and cannot exclude the possibility based on our understanding.What we have observed is that the call chain leading to the free site (do_fb_ioctl()->fbcon_set_con2fb_map_ioctl()->set_con2fb_map()-> con2fb_release_oldinfo()-> ... ->matroxfb_remove()) is only protected by
console_lock()but notlock_fb_info(), while the potential use site (call chain starts from the default case indo_fb_ioctl()) is only protected bylock_fb_info()but notconsole_lock(). We thus propose to add this extraconsole_lock()to the default case, which is inspired by the lock protection of many other existing switch-case terms in the same function.Quoted 8 lines
Quoted 3 lines
Since we do not have deep domain knowledge of this subsystem, we will rely on the developers to make a decision regarding the patch. Thank you again for your review and help!
Can you please elaborate where you've found this use-after-free and how? I'm still not understanding how you even got here - this is orthogonal to whether the patch is the right fix or not.
-Daniel
Quoted 13 lines
Hi, Daniel. Sure. This issue was initially flagged by our experimental static code analyzer aiming to find use-after-free issues in the kernel - that's why we don't have PoC or execution traces here. We deeply understand that static analyzer can generate false alarms, so we have tried our best and spent a good amount of time carefully reviewing the related code. We eventually found that we could not exclude this potential issue based on our study, so we decided to report this to the community with this tentative fix. As mentioned, we may be short in domain knowledge, so your input is highly appreciated. We respect the developer's decision about whether this is really a problem and whether/how to fix it. However, if you think the use-after-free is actually not possible, it will be very helpful if you can elaborate on the reasoning since it will greatly help us improve our analyzer. Thank you very much for your help!
Please start out these patches with the fact that this is from an experimental checker.
Also do include _why_ your checker things something is going wrong. If you cannot follow why the checker complains about something, then don't report it as an issue until you do. Also, if you do not understand the code, please make it absolutely clear that you just guessed a possible fix, but not that it's been tested in any way or form.
If you don't do this, then we end up wasting a ton of time of people who don't have surplus time, because in this case the patch got review, applied, pull request made, I realized it looks funny, patch dropped, pull request remade, and then a fairly big thread here.
All for a bug that's likely in your checker and not in the kernel. This is not great.
Quoted 5 lines
BTW, if this is worthed a fix and the performance of
console_lock()is a major concern, then I think there may be alternative solutions like adding alock_fb_info()to the free call chain - if that's better in performance, or maybe selectively protect the matroxfb ioctl but not vblank ioctl as you mentioned.
Please start out with explaining what kind of bug your checker is seeing, and why. Not how you're trying to fix it. Because I'm pretty sure there isn't a bug, but since I've already spent a pile of time looking at this, I want to make sure.
Cheers, Daniel
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
Quoted 133 lines · Daniel Vetter
Quoted 89 lines · Hang Zhang
Quoted 79 lines · Daniel Vetter
Quoted 77 lines · Hang Zhang
Quoted 42 lines · Daniel Vetter
Quoted 29 lines · Daniel Vetter
Hi Helge
Quoted 15 lines · Helge Deller
Quoted 13 lines · Hang Zhang
In
do_fb_ioctl(), user specified "fb_info" can be freed in the calleefbcon_get_con2fb_map_ioctl()->set_con2fb_map()->con2fb_release_oldinfo(), this free operation is protected byconsole_lock()infbcon_set_con2fb_map_ioctl(), it also results in the change of certain states such as "minfo->dead" inmatroxfb_remove(), so that it can be checked to avoid use-after-free before the use sites (e.g., the check at the beginning ofmatroxfb_ioctl()). However, the problem is that the use site is not protected by the same locks as for the free operation, e.g., "default" case indo_fb_ioctl()can lead to "matroxfb_ioctl()" but it's not protected byconsole_lock(), which can invalidate the aforementioned state set and check in a concurrent setting.Quoted 7 lines
Quoted 5 lines
Prevent the potential use-after-free issues by protecting the "default" case in
do_fb_ioctl()withconsole_lock(), similarly as for many other cases like "case FBIOBLANK" and "caseFBIOPAN_DISPLAY".Signed-off-by Hang Zhang <[email protected]>applied to fbdev git tree.
The patch above makes no sense at all to me:
Quoted 10 lines
-
fb_infois protected bylock_fb_infoand
- the lifetime offb_infois protected by the get/put functions
- yes there's the interaction with con2fb, which is protected byconsole_lock, but the lifetime guarantees are ensured by the device removal
- which means any stuff happening inmatroxfb_removeis also not a concern here (unless matroxfb completely gets all the device lifetime stuff wrong, but it doesn't look like it's any worse than any of the other fbdev drivers that we haven't recently fixed up due to the takeover issues with firmware driversQuoted 16 lines
I have also a really hard timing finding the con2fb map use in the matroxfb ioctl code, but that just might be that I didn't look carefully enough. Maybe that would shed some light on this. -Daniel
Quoted 10 lines
On the very clear downside this now means we take
console_lockfor the vblank ioctl (which is a device driver extension for reasons, despite that it's a standard fbdev ioctl), which is no good at all given howconsole_lock()is a really expensive lock.Unless I'm massively missing something, can you pls push the revert before this lands in Linus' tree?
Thanks, Daniel
Hi, Daniel. Thank you for your feedback. We're not developers of the video subsystem and thus may be short in domain knowledge (e.g., the performance of
console_lock()and the complex lifetime management). This patch initially intended to bring up the potential use-after-free issues here to the community - we have performed a best-effort code review and cannot exclude the possibility based on our understanding.What we have observed is that the call chain leading to the free site (do_fb_ioctl()->fbcon_set_con2fb_map_ioctl()->set_con2fb_map()-> con2fb_release_oldinfo()-> ... ->matroxfb_remove()) is only protected by
console_lock()but notlock_fb_info(), while the potential use site (call chain starts from the default case indo_fb_ioctl()) is only protected bylock_fb_info()but notconsole_lock(). We thus propose to add this extraconsole_lock()to the default case, which is inspired by the lock protection of many other existing switch-case terms in the same function.Quoted 8 lines
Quoted 3 lines
Since we do not have deep domain knowledge of this subsystem, we will rely on the developers to make a decision regarding the patch. Thank you again for your review and help!
Can you please elaborate where you've found this use-after-free and how? I'm still not understanding how you even got here - this is orthogonal to whether the patch is the right fix or not.
-DanielQuoted 13 lines
Hi, Daniel. Sure. This issue was initially flagged by our experimental static code analyzer aiming to find use-after-free issues in the kernel - that's why we don't have PoC or execution traces here. We deeply understand that static analyzer can generate false alarms, so we have tried our best and spent a good amount of time carefully reviewing the related code. We eventually found that we could not exclude this potential issue based on our study, so we decided to report this to the community with this tentative fix. As mentioned, we may be short in domain knowledge, so your input is highly appreciated. We respect the developer's decision about whether this is really a problem and whether/how to fix it. However, if you think the use-after-free is actually not possible, it will be very helpful if you can elaborate on the reasoning since it will greatly help us improve our analyzer. Thank you very much for your help!
Please start out these patches with the fact that this is from an experimental checker.
Also do include _why_ your checker things something is going wrong. If you cannot follow why the checker complains about something, then don't report it as an issue until you do. Also, if you do not understand the code, please make it absolutely clear that you just guessed a possible fix, but not that it's been tested in any way or form.
If you don't do this, then we end up wasting a ton of time of people who don't have surplus time, because in this case the patch got review, applied, pull request made, I realized it looks funny, patch dropped, pull request remade, and then a fairly big thread here.
All for a bug that's likely in your checker and not in the kernel. This is not great.
Quoted 5 lines
BTW, if this is worthed a fix and the performance of
console_lock()is a major concern, then I think there may be alternative solutions like adding alock_fb_info()to the free call chain - if that's better in performance, or maybe selectively protect the matroxfb ioctl but not vblank ioctl as you mentioned.Please start out with explaining what kind of bug your checker is seeing, and why. Not how you're trying to fix it. Because I'm pretty sure there isn't a bug, but since I've already spent a pile of time looking at this, I want to make sure.
Quoted 5 lines
Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
We are sorry for the inconvenience caused, we'll follow these practices and guidelines in the future. Thank you!
Quoted 9 lines · Hang Zhang
Quoted 8 lines · Daniel Vetter
Quoted 6 lines · Hang Zhang
On Fri, Jan 6, 2023 at 1:59 PM Daniel Vetter <[email protected]> wrote:
BTW, if this is worthed a fix and the performance ofconsole_lock()is a major concern, then I think there may be alternative solutions like adding alock_fb_info()to the free call chain - if that's better in performance, or maybe selectively protect the matroxfb ioctl but not vblank ioctl as you mentioned.
Quoted 7 lines
Quoted 4 lines
Please start out with explaining what kind of bug your checker is seeing, and why. Not how you're trying to fix it. Because I'm pretty sure there isn't a bug, but since I've already spent a pile of time looking at this, I want to make sure.
We are sorry for the inconvenience caused, we'll follow these practices and guidelines in the future. Thank you!
Once more: Please explain what you're static checker is seeing. I want to understanding this, and I'm hoping at least someone involved in this static checker can explain what it thinks is going on.
Thanks, Daniel
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
Quoted 23 lines · Daniel Vetter
Quoted 9 lines · Hang Zhang
Quoted 8 lines · Daniel Vetter
Quoted 6 lines · Hang Zhang
On Fri, Jan 6, 2023 at 1:59 PM Daniel Vetter <[email protected]> wrote:
BTW, if this is worthed a fix and the performance ofconsole_lock()is a major concern, then I think there may be alternative solutions like adding alock_fb_info()to the free call chain - if that's better in performance, or maybe selectively protect the matroxfb ioctl but not vblank ioctl as you mentioned.Quoted 7 lines
Quoted 4 lines
Please start out with explaining what kind of bug your checker is seeing, and why. Not how you're trying to fix it. Because I'm pretty sure there isn't a bug, but since I've already spent a pile of time looking at this, I want to make sure.
We are sorry for the inconvenience caused, we'll follow these practices and guidelines in the future. Thank you!
Once more: Please explain what you're static checker is seeing. I want to understanding this, and I'm hoping at least someone involved in this static checker can explain what it thinks is going on.
Quoted 5 lines
Thanks, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
Thank you for your interest, Daniel. The checker tries first to find the free and
use sites of a certain object (in this case "fb_info"), then reason about whether
the use can actually happen after the free (e.g., taking into account factors like
state set/check, locks, etc.), if so, it will flag a potential use-after-free. As a static
checker, is doesn't execute a program or generate a PoC. We then manually review each flagged issue by inspecting all related code. In this case, the checker
(and us) are unaware of the lifetime management logic, which may cause problems.
Best,
Hang
Quoted 29 lines · Hang Zhang
Quoted 22 lines · Daniel Vetter
Quoted 9 lines · Hang Zhang
Quoted 8 lines · Daniel Vetter
Quoted 6 lines · Hang Zhang
On Fri, Jan 6, 2023 at 1:59 PM Daniel Vetter <[email protected]> wrote:
BTW, if this is worthed a fix and the performance ofconsole_lock()is a major concern, then I think there may be alternative solutions like adding alock_fb_info()to the free call chain - if that's better in performance, or maybe selectively protect the matroxfb ioctl but not vblank ioctl as you mentioned.Quoted 7 lines
Quoted 4 lines
Please start out with explaining what kind of bug your checker is seeing, and why. Not how you're trying to fix it. Because I'm pretty sure there isn't a bug, but since I've already spent a pile of time looking at this, I want to make sure.
We are sorry for the inconvenience caused, we'll follow these practices and guidelines in the future. Thank you!
Once more: Please explain what you're static checker is seeing. I want to understanding this, and I'm hoping at least someone involved in this static checker can explain what it thinks is going on.
Quoted 5 lines
Thanks, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
Quoted 13 lines
Thank you for your interest, Daniel. The checker tries first to find the free and
use sites of a certain object (in this case "fb_info"), then reason about whether
the use can actually happen after the free (e.g., taking into account factors like
state set/check, locks, etc.), if so, it will flag a potential use-after-free. As a static
checker, is doesn't execute a program or generate a PoC. We then manually review each flagged issue by inspecting all related code. In this case, the checker
(and us) are unaware of the lifetime management logic, which may cause problems.
Lifetime management is and absolute basic part in the linux kernel. So if your checker flags every free which isn't protected by a lock, then you'll creating endless amounts of false positives. Is this really what you're doing?
I'm still very confused ...
-Daniel
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
Quoted 50 lines · Daniel Vetter
Quoted 29 lines · Hang Zhang
Quoted 22 lines · Daniel Vetter
Quoted 9 lines · Hang Zhang
Quoted 8 lines · Daniel Vetter
Quoted 6 lines · Hang Zhang
On Fri, Jan 6, 2023 at 1:59 PM Daniel Vetter <[email protected]> wrote:
BTW, if this is worthed a fix and the performance ofconsole_lock()is a major concern, then I think there may be alternative solutions like adding alock_fb_info()to the free call chain - if that's better in performance, or maybe selectively protect the matroxfb ioctl but not vblank ioctl as you mentioned.Quoted 7 lines
Quoted 4 lines
Please start out with explaining what kind of bug your checker is seeing, and why. Not how you're trying to fix it. Because I'm pretty sure there isn't a bug, but since I've already spent a pile of time looking at this, I want to make sure.
We are sorry for the inconvenience caused, we'll follow these practices and guidelines in the future. Thank you!
Once more: Please explain what you're static checker is seeing. I want to understanding this, and I'm hoping at least someone involved in this static checker can explain what it thinks is going on.
Quoted 5 lines
Thanks, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.chQuoted 13 lines
Thank you for your interest, Daniel. The checker tries first to find the free and
use sites of a certain object (in this case "fb_info"), then reason about whether
the use can actually happen after the free (e.g., taking into account factors like
state set/check, locks, etc.), if so, it will flag a potential use-after-free. As a static
checker, is doesn't execute a program or generate a PoC. We then manually review each flagged issue by inspecting all related code. In this case, the checker
(and us) are unaware of the lifetime management logic, which may cause problems.Lifetime management is and absolute basic part in the linux kernel. So if your checker flags every free which isn't protected by a lock, then you'll creating endless amounts of false positives. Is this really what you're doing?
Quoted 6 lines
I'm still very confused ...
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
Hi, Daniel. Lock is only one of many factors we consider in the checker, so the actual workflow is certainly more complicated than "mark every free w/o lock". E.g., we also need to consider the data flow between use and free, the state check, etc. But as you know, it is very challenging for a static checker to automatically and accurately reason about everything in the code (automatic lifetime management analysis can also be tricky for a static analyzer), that's why we perform a manual review afterward. We will continue working on the checker to reduce its false alarms and submit higher-quality reports to the community following your guidelines. Thank you so much for your time!
Best,
Hang