[syzbot] KASAN: vmalloc-out-of-bounds Write in imageblit (2)

[Khalid Masum]

On 7/30/22 23:25, Helge Deller wrote:
> On 7/29/22 08:51, Khalid Masum wrote:
>> Here is a simplified reproducer for the issue:
>>
>> https://gist.githubusercontent.com/Labnann/923d6b9b3a19848fc129637b839b8a55/raw/a68271fcc724569735fe27f80817e561b3ff629a/reproducer.c
> 
> The reproducer does this:

Thanks for Looking into this. Being to this, so I have some questions.
> ioctl(3, TIOCLINUX, TIOCL_SETSEL, selection: xs:3  ys:0  xe:0 ye:0 mode:0)  = 0

How did you find out the selection values? From strace and man pages I 
know the third argument is an address.

> -> sets the text selection area
> ioctl(4, KDFONTOP)  with op=0 (con_font_set), charcount=512  width=8  height=32, 0x20000000) = 0

Same here, It would be very helpful if you could tell me how.

> -> changes the font size.
> 
> It does not crash with current Linus' head (v5.19-rc8).

I tested in 5.19-rc8 in Qemu x86_64 and it crashed for me.
> Kernel v5.16, which was used by this KASAN report, hasn't received backports
> since months, so I tried stable kernel v5.15.58 instead, and this
> kernel crashed with the reproducer.
> 
> The reproducer brings up two issues with current code:
> 1. The reproducer uses ioctl(TIOCLINUX, TIOCL_SETSEL) and hands over (invalid)
> zero-values for ys and ye for the starting lines.
> This is wrong, since the API seems to expect a "1" as the very first line for the selection.
> This can be easily fixed by adding checks for zero-values and return -EINVAL if found.
> 
> But this bug isn't critical itself and is not the reason for the kernel crash.
> Without the checks, the ioctl handler simply wraps the coordinate values and converts them
> from:
> input selection: xs:3  ys:0  xe:0   ye:0  mode:0    to the new:
> vc_selection =   xs:2  ys:23 xe:127 ye:23 mode:0
> which is the current maximum coordinates for the screen.
> 
> Those higher values now trigger issue #2:
> After the TIOCL_SETSEL the last line on the screen is now selected. The KDFONTOP ioctl
> then sets a 8x32 console font, and replaces the former 8x16 console font.
> With the bigger font the current screen selection is now outside the visible screen
> and this finally triggeres this backtrace, because vc_do_resize() calls clear_selection()
> to unhighlight the selection (which starts to render chars outside of the screen):

That makes sense.

>   drm_fb_helper_sys_imageblit drivers/gpu/drm/drm_fb_helper.c:794 [inline]
>   drm_fbdev_fb_imageblit+0x15c/0x350 drivers/gpu/drm/drm_fb_helper.c:2288
>   bit_putcs_unaligned drivers/video/fbdev/core/bitblit.c:124 [inline]
>   bit_putcs+0x6e1/0xd20 drivers/video/fbdev/core/bitblit.c:173
>   fbcon_putcs+0x353/0x440 drivers/video/fbdev/core/fbcon.c:1277
>   do_update_region+0x399/0x630 drivers/tty/vt/vt.c:676
>   invert_screen+0x1d4/0x600 drivers/tty/vt/vt.c:800
>   highlight drivers/tty/vt/selection.c:57 [inline]
>   clear_selection drivers/tty/vt/selection.c:84 [inline]
>   clear_selection+0x55/0x70 drivers/tty/vt/selection.c:80
>   vc_do_resize+0xe6e/0x1180 drivers/tty/vt/vt.c:1257
> 
> IMHO the easiest way to prevent this crash is to simply clear the
> selection before the various con_font_set() console handlers are called.
> Otherwise every console driver needs to add checks and verify if the current
> selection still fits with the selected font, which gets tricky because some
> of those drivers fiddle with the screen width&height before calling vc_do_resize().
> 
> I'll follow up to this mail with patches for both issues shortly.

I tested the patches. The crash no longer occurs with the reproducer.

> Helge

Thanks,
  -- Khalid Masum