[PATCH] framebuffer: fix screen corruption when copying

The function bitcpy_rev has a bug that may result in screen corruption.
The bug happens under these conditions:
* the end of the destination area of a copy operation is aligned on a long
  word boundary
* the end of the source area is not aligned on a long word boundary
* we are copying more than one long word

In this case, the variable shift is non-zero and the variable first is
zero. The statements FB_WRITEL(comp(d0, FB_READL(dst), first), dst) reads
the last long word of the destination and writes it back unchanged
(because first is zero). Correctly, we should write the variable d0 to the
last word of the destination in this case.

This patch fixes the bug by introducing and extra test if first is zero.

The patch also removes the references to fb_memmove in the code that is
commented out because fb_memmove was removed from framebuffer subsystem.

Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Cc: stable@vger.kernel.org

Index: linux-3.16.1/drivers/video/fbdev/core/cfbcopyarea.c
=================================--- linux-3.16.1.orig/drivers/video/cfbcopyarea.c

+++ linux-3.16.1/drivers/video/fbdev/core/cfbcopyarea.c

@@ -55,8 +55,8 @@ bitcpy(struct fb_info *p, unsigned long
 	 * If you suspect bug in this function, compare it with this simple
 	 * memmove implementation.
 	 */
-	fb_memmove((char *)dst + ((dst_idx & (bits - 1))) / 8,
-		   (char *)src + ((src_idx & (bits - 1))) / 8, n / 8);
+	memmove((char *)dst + ((dst_idx & (bits - 1))) / 8,
+		(char *)src + ((src_idx & (bits - 1))) / 8, n / 8);
 	return;
 #endif
 

@@ -221,8 +221,8 @@ bitcpy_rev(struct fb_info *p, unsigned l
 	 * If you suspect bug in this function, compare it with this simple
 	 * memmove implementation.
 	 */
-	fb_memmove((char *)dst + ((dst_idx & (bits - 1))) / 8,
-		   (char *)src + ((src_idx & (bits - 1))) / 8, n / 8);
+	memmove((char *)dst + ((dst_idx & (bits - 1))) / 8,
+		(char *)src + ((src_idx & (bits - 1))) / 8, n / 8);
 	return;
 #endif
 

@@ -324,7 +324,10 @@ bitcpy_rev(struct fb_info *p, unsigned l
 				d0 = d0 << left | d1 >> right;
 			}
 			d0 = fb_rev_pixels_in_long(d0, bswapmask);
-			FB_WRITEL(comp(d0, FB_READL(dst), first), dst);
+			if (!first)
+				FB_WRITEL(d0, dst);
+			else
+				FB_WRITEL(comp(d0, FB_READL(dst), first), dst);
 			d0 = d1;
 			dst--;
 			n -= dst_idx+1;

Hi Mikulas,

On Tue, Sep 16, 2014 at 6:38 PM, Mikulas Patocka [off-list ref] wrote:

The patch also removes the references to fb_memmove in the code that is
commented out because fb_memmove was removed from framebuffer subsystem.

I'd leave it as-is. Using plain memmove() is not correct, as it should operate
on MMIO space, not RAM. On x86 it works with memmove(), though, but
not on all architectures.

Gr{oetje,eeting}s,

                        Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds

On Wed, 17 Sep 2014, Geert Uytterhoeven wrote:

Hi Mikulas,

On Tue, Sep 16, 2014 at 6:38 PM, Mikulas Patocka [off-list ref] wrote:

quoted

The patch also removes the references to fb_memmove in the code that is
commented out because fb_memmove was removed from framebuffer subsystem.

I'd leave it as-is. Using plain memmove() is not correct, as it should operate
on MMIO space, not RAM. On x86 it works with memmove(), though, but
not on all architectures.

Gr{oetje,eeting}s,

                        Geert

That code is disabled with #if 0. So it doesn't matter if it's incorrect 
in some specific situation.

The purpose of that disabled code is to be able to verify that the bug is 
in bitcpy_rev. I get screen corruption - so I enable that "memmove" 
statement - the screen corruption goes away - so I know that the bug is in 
bitcpy_rev.

Mikulas

On 16/09/14 19:38, Mikulas Patocka wrote:

The function bitcpy_rev has a bug that may result in screen corruption.
The bug happens under these conditions:
* the end of the destination area of a copy operation is aligned on a long
  word boundary
* the end of the source area is not aligned on a long word boundary
* we are copying more than one long word

In this case, the variable shift is non-zero and the variable first is
zero. The statements FB_WRITEL(comp(d0, FB_READL(dst), first), dst) reads
the last long word of the destination and writes it back unchanged
(because first is zero). Correctly, we should write the variable d0 to the
last word of the destination in this case.

This patch fixes the bug by introducing and extra test if first is zero.

The patch also removes the references to fb_memmove in the code that is
commented out because fb_memmove was removed from framebuffer subsystem.

Thanks, queued for 3.18.

 Tomi