PR #20883 opened by thomasdullien
URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/20883
Patch URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/20883.patch

For transparency: I am experimenting with an AI-assisted patch process, where 
the AI agent attempts to help
root-cause analyze a crash by means of reproducing the crash with ASAN, making 
a recording with 'rr', and
documenting the analysis at a granular level with verbatim quotes from the 'rr' 
trace. This root-cause analysis
is then iteratively reviewed (e.g. as a human reviewer I check it for accuracy 
and plausibility) before a patch is
generated. The process generates a detailed analysis report, an 'rr' trace that 
can be shared with other to help
with the verification, and a patch. Given that it is unlear how to best share 
the 'rr' trace, I have only attached the
detailed root-cause analysis document that was at the end of the iterative 
process.

Tests have been run and pass.

=== Description ===

The YUV422 conversion functions process 2 rows at once but did not
check whether a second row actually exists when srcSliceH is odd.
With bottom-to-top processing (negative strides), this caused
pu_2/pv_2 pointers to be set before the buffer start, leading to
out-of-bounds memory access when accessing pu_2[0] or pv_2[0].

This patch adds a check to skip row 2 processing in the final
remainder section when srcSliceH is odd, preventing access to
non-existent rows while still processing all available source lines.

Performance impact is minimal: one bitwise AND operation only in the
remainder section (when width is not divisible by 4), so performance
regression is unlikely to be severe.

Fixes: https://trac.ffmpeg.org/ticket/11691

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>


From 95b96efa7bc7b587e857b3e780d72b1192cb6a49 Mon Sep 17 00:00:00 2001
From: Thomas Dullien <[email protected]>
Date: Sun, 9 Nov 2025 12:16:27 +0100
Subject: [PATCH] swscale/yuv2rgb: fix out-of-bounds access with odd srcSliceH
 in YUV422
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The YUV422 conversion functions process 2 rows at once but did not
check whether a second row actually exists when srcSliceH is odd.
With bottom-to-top processing (negative strides), this caused
pu_2/pv_2 pointers to be set before the buffer start, leading to
out-of-bounds memory access when accessing pu_2[0] or pv_2[0].

This patch adds a check to skip row 2 processing in the final
remainder section when srcSliceH is odd, preventing access to
non-existent rows while still processing all available source lines.

Performance impact is minimal: one bitwise AND operation only in the
remainder section (when width is not divisible by 4), so performance
regression is unlikely to be severe.

Fixes: https://trac.ffmpeg.org/ticket/11691

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>
---
 libswscale/yuv2rgb.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/libswscale/yuv2rgb.c b/libswscale/yuv2rgb.c
index 48089760f5..cb5bcfaac2 100644
--- a/libswscale/yuv2rgb.c
+++ b/libswscale/yuv2rgb.c
@@ -364,8 +364,10 @@ const int *sws_getCoefficients(int colorspace)
         LOADCHROMA(1, 0);                                               \
         PUTFUNC(1, 0, 0);                                               \
                                                                         \
-        LOADCHROMA(2, 0);                                               \
-        PUTFUNC(2, 0, 0 + 8);                                           \
+        if (!(srcSliceH & 1)) {                                         \
+            LOADCHROMA(2, 0);                                           \
+            PUTFUNC(2, 0, 0 + 8);                                       \
+        }                                                               \
     ENDYUV2RGBFUNC()
 
 #define LOADDITHER16                                    \
-- 
2.49.1

_______________________________________________
ffmpeg-devel mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to