Re: [PATCH] fbcon: Set fb_display[i]->mode to NULL when the mode is released

[Quanmin Yan <yanquanmin1@huawei.com>]

Hi Thomas,

在 2025/9/24 17:17, Thomas Zimmermann 写道:
> Hi,
>
> thanks for the report.
>
> Am 23.09.25 um 13:06 schrieb Quanmin Yan:
>> Recently, we discovered the following issue through syzkaller:
>>
>> BUG: KASAN: slab-use-after-free in fb_mode_is_equal+0x285/0x2f0
>> Read of size 4 at addr ff11000001b3c69c by task syz.xxx
>> ...
>> Call Trace:
>>   <TASK>
>>   dump_stack_lvl+0xab/0xe0
>>   print_address_description.constprop.0+0x2c/0x390
>>   print_report+0xb9/0x280
>>   kasan_report+0xb8/0xf0
>>   fb_mode_is_equal+0x285/0x2f0
>>   fbcon_mode_deleted+0x129/0x180
>>   fb_set_var+0xe7f/0x11d0
>>   do_fb_ioctl+0x6a0/0x750
>>   fb_ioctl+0xe0/0x140
>>   __x64_sys_ioctl+0x193/0x210
>>   do_syscall_64+0x5f/0x9c0
>>   entry_SYSCALL_64_after_hwframe+0x76/0x7e
>>
>> The issue occurs in the function fb_mode_is_equal(p->mode, mode), I also
>> noticed that when freeing the memory related to fb_info->modelist, 
>> there's
>> no attempt to set the corresponding fb_display[i]->mode to NULL after
>> freeing. Based on analysis, the root cause of this bug appears to be 
>> that
>> a certain p->mode has become a wild pointer.
>>
>> I've identified two code paths for freeing modelist->mode:
>> 1. fb_delete_videomode - removes videomode entry from modelist.
>> 2. fb_destroy_modelist - destroys the entire modelist.
>
> What about fb_new_modelist()? [1] It's called from store_modes() and 
> the whole logic in the caller seems fragile to me. This could leave 
> p->mode set when it should be cleared; and vice versa.
>
> [1] 
> https://elixir.bootlin.com/linux/v6.16.8/source/drivers/video/fbdev/core/fbmem.c#L712
> [2] 
> https://elixir.bootlin.com/linux/v6.16.8/source/drivers/video/fbdev/core/fbsysfs.c#L113 
>
>
After testing, this path does not exhibit the issue because in 
fbcon_new_modelist [1], the
modes in fb_display[] get refreshed and won't leave wild pointers. Even 
if fb_new_modelist()
fails, the original modelist will be restored [2].

[1] 
https://elixir.bootlin.com/linux/v6.16.8/source/drivers/video/fbdev/core/fbmem.c#L736
[2] 
https://elixir.bootlin.com/linux/v6.16.8/source/drivers/video/fbdev/core/fbsysfs.c#L115 

>>
>> Analysis shows that fb_delete_videomode path should have been fixed in
>> a previous patch[1]. Therefore, the current bug is likely triggered
>> through the fb_destroy_modelist path. I've found a reproducible test 
>> case:
>> 1. With /dev/fb0 already registered in the system, load a kernel module
>>     to register a new device /dev/fb1;
>> 2. Set fb1's mode to the global fb_display[] array (via 
>> FBIOPUT_CON2FBMAP);
>> 3. Switch console from fb to VGA (to allow normal rmmod of the ko);
>> 4. Unload the kernel module - at this point fb1's modelist is freed, 
>> leaving
>>     a wild pointer in fb_display[];
>> 5. Trigger the bug via system calls through fb0 attempting to delete 
>> a mode
>>     from fb0.
>>
>> To prevent similar issues from recurring, consider traversing 
>> fb_display[]
>> whenever releasing a mode from fb_info. If the corresponding mode exists
>> in fb_display[], set its pointer to NULL.
>>
>> [1] 
>> https://lore.kernel.org/all/20210712085544.2828-1-thunder.leizhen@huawei.com/
>>
>> Signed-off-by: Quanmin Yan <yanquanmin1@huawei.com>
>> ---
>> This is my first time working on fb issues. If there are any 
>> misunderstandings
>> in my analysis, I would appreciate corrections from the community.
>>
>>   drivers/video/fbdev/core/fbcon.c  | 11 +++++++++++
>>   drivers/video/fbdev/core/modedb.c |  7 +++++++
>>   include/linux/fbcon.h             |  2 ++
>>   3 files changed, 20 insertions(+)
>>
>> diff --git a/drivers/video/fbdev/core/fbcon.c 
>> b/drivers/video/fbdev/core/fbcon.c
>> index b062b05f4128..bfbf79d6cd05 100644
>> --- a/drivers/video/fbdev/core/fbcon.c
>> +++ b/drivers/video/fbdev/core/fbcon.c
>> @@ -2803,6 +2803,17 @@ int fbcon_mode_deleted(struct fb_info *info,
>>       return found;
>>   }
>>   +void fb_display_clean_videomode(struct fb_videomode *m)
>
> Rather fbcon_delete_mode
>
>> +{
>> +    struct fbcon_display *p;
>> +
>> +    for (int i = first_fb_vc; i <= last_fb_vc; i++) {
>
> I think this code also needs to test the fb_info, or it might clear 
> another display's mode. See [3] for an example
>
> [3] 
> https://elixir.bootlin.com/linux/v6.16.8/source/drivers/video/fbdev/core/fbcon.c#L2786
>
>> +        p = &fb_display[i];
>> +        if (p->mode == m)
>> +            p->mode = NULL;
>> +    }
>> +}
>> +
>>   #ifdef CONFIG_VT_HW_CONSOLE_BINDING
>>   static void fbcon_unbind(void)
>>   {
>> diff --git a/drivers/video/fbdev/core/modedb.c 
>> b/drivers/video/fbdev/core/modedb.c
>> index 53a610948c4a..5a0ee96ebefa 100644
>> --- a/drivers/video/fbdev/core/modedb.c
>> +++ b/drivers/video/fbdev/core/modedb.c
>> @@ -16,6 +16,7 @@
>>   #include <linux/slab.h>
>>   #include <linux/fb.h>
>>   #include <linux/kernel.h>
>> +#include <linux/fbcon.h>
>>     #undef DEBUG
>>   @@ -1100,6 +1101,7 @@ void fb_delete_videomode(const struct 
>> fb_videomode *mode,
>>           modelist = list_entry(pos, struct fb_modelist, list);
>>           m = &modelist->mode;
>>           if (fb_mode_is_equal(m, mode)) {
>> +            fb_display_clean_videomode(m);
>
> There's only one caller of fb_delete_videomode(). I think this call 
> should be right before fb_delete_videomode().
This path has already been fixed [3], so this modification is no longer 
necessary and will be
removed in the next version of the patch.

[3] 
https://lore.kernel.org/all/20210712085544.2828-1-thunder.leizhen@huawei.com/

Additionally, in the next version of the patch, I have incorporated your 
other suggestions
and made corresponding adjustments. I will send the updated patch set 
shortly.

Best regards,
Quanmin Yan
>
>>               list_del(pos);
>>               kfree(pos);
>>           }
>> @@ -1113,8 +1115,13 @@ void fb_delete_videomode(const struct 
>> fb_videomode *mode,
>>   void fb_destroy_modelist(struct list_head *head)
>>   {
>>       struct list_head *pos, *n;
>> +    struct fb_modelist *modelist;
>> +    struct fb_videomode *m;
>>         list_for_each_safe(pos, n, head) {
>> +        modelist = list_entry(pos, struct fb_modelist, list);
>> +        m = &modelist->mode;
>> +        fb_display_clean_videomode(m);
>
> Same here: I think fb_destroy_modelist() should only release the mode. 
> Clearing the fbcon mode should be done by the caller.
>
> Best regards
> Thomas
>
>>           list_del(pos);
>>           kfree(pos);
>>       }
>> diff --git a/include/linux/fbcon.h b/include/linux/fbcon.h
>> index 81f0e698acbf..2b5e93aeaaff 100644
>> --- a/include/linux/fbcon.h
>> +++ b/include/linux/fbcon.h
>> @@ -18,6 +18,7 @@ void fbcon_suspended(struct fb_info *info);
>>   void fbcon_resumed(struct fb_info *info);
>>   int fbcon_mode_deleted(struct fb_info *info,
>>                  struct fb_videomode *mode);
>> +void fb_display_clean_videomode(struct fb_videomode *m);
>>   void fbcon_new_modelist(struct fb_info *info);
>>   void fbcon_get_requirement(struct fb_info *info,
>>                  struct fb_blit_caps *caps);
>> @@ -38,6 +39,7 @@ static inline void fbcon_suspended(struct fb_info 
>> *info) {}
>>   static inline void fbcon_resumed(struct fb_info *info) {}
>>   static inline int fbcon_mode_deleted(struct fb_info *info,
>>                        struct fb_videomode *mode) { return 0; }
>> +static inline void fb_display_clean_videomode(struct fb_videomode 
>> *m) {}
>>   static inline void fbcon_new_modelist(struct fb_info *info) {}
>>   static inline void fbcon_get_requirement(struct fb_info *info,
>>                        struct fb_blit_caps *caps) {}
>