[Bug]: VirtualBoxVM crashes inside the host AMD graphics driver when the guest desktop renders images
维护者通常 6 天内回复
还没有人认领这个 Issue。
评估
- 难度
- 5/5
- 预计耗时
- 一周以上
- 新手友好度
- 18/100
- Issue 类型
- 缺陷
- 描述清晰度
- 基本清楚
- 活跃度
- 活跃
- 技术栈
- c
调研方向
Start by reading the VMSVGA 3D support code and the VBox.log exception record described in the report, then inspect how guest image surfaces are sent through the host graphics path. The AMD driver is identified as the faulting code, and the payload does not name a VirtualBox source file or failing test. Done means identifying whether VirtualBox can prevent the out-of-bounds driver read or otherwise keep the VM responsive.
由索引模型根据 Issue 内容生成。
描述
Version
7.2.22
Host OS Type
Windows
Host OS name + version
Windows 10 Home 22H2 (build 19045.6466), 64-bit
Host Architecture
x86
Guest OS Type
Linux
Guest Architecture
x86
Guest OS name + version
Kali Linux 2026.2 (64-bit, Debian-based), XFCE 4.20, Xorg at depth 24 (32 bpp effective)
Component
3D Support
What happened?
While a VM runs with VMSVGA + 3D on a Windows 10 host with an AMD Radeon RX 5700 XT (driver 26.10.2 /
32.0.21045.20006; the same fault was first captured on 32.0.21045.16010 at +0xF8B892),
VirtualBoxVM.exe takes a read access violation inside the host AMD user-mode driver, not inside VirtualBox code,
and the VM becomes unresponsive. The Windows dialog shows:
VirtualBoxVM.exe - Application Error
The instruction at 0x00007FFA7623B892 referenced memory at 0x00000021E80F42000.
The memory could not be read.
VirtualBox's own exception filter logs the exception and keeps the process alive but hung, so the guest display
freezes and the VM has to be killed. VBox.log contains the rtR3WinUnhandledXcptFilter record with a full
register dump and a resolved module/offset stack.
The faulting instruction is movzx r8d, byte ptr [rdx+2] at amdxx64.dll+0xF8B892 (a second compiled form of the
same copy exists at +0xF8B88F, movzx ecx, word ptr [rdx]). Both are 24-bit pixel copy helpers called from a row
loop at amdxx64.dll+0x6BD51. In every captured crash the containing private allocation ends exactly at the
faulting address, i.e. an out-of-bounds read of 1-2 bytes past a private PAGE_READWRITE allocation; the page after
it is MEM_RESERVE, which is why the read faults instead of silently succeeding. The same fault reproduces on the
current optional release 32.0.21045.20006 (AMD Software 26.10.2) at amdxx64.dll+0xF8BA12 (word form
+0xF8BA0F), so it is present in the newest driver available for this card.
The two row strides are related in every capture (58 of 58): r15 = ceil4(r14), i.e. the destination row is the
source row rounded up to a 4-byte boundary, so the difference is always exactly 3 bytes. The loop copies the source
row and then a second time for that 3-byte gap, reading from just past the source row - harmless on every row but
the last, where those bytes are past the allocation. Why the source row length itself has that value is not
identified from the logs; that does not change the defect.
Control: with 3D acceleration disabled the same workload produced no fault in 343 cycles, while with 3D enabled the
fault appeared from 49 s (GPU loaded) to 54 min (VM idle, one dialog open) across 58 captures.
Filing here because it is VirtualBox's process that dies and VirtualBox's handler that surfaces it; the faulting
code belongs to the AMD driver. The same report was submitted to AMD through the AMD Bug Report Tool on
2026-11-17, with the same evidence package.
Expected behavior: The guest desktop renders its dialogs and thumbnails and the VM keeps running.
Actual behavior: VirtualBoxVM.exe takes a read access violation inside the host GPU driver; VirtualBox's exception filter hangs the
process and the VM has to be killed.
Attached: amd-vbox-crash-evidence.zip (62 logs, 67 address-space maps, screenshots, scripts) and evidence-crash-video.mp4 - a 33-second recording of one fault with the host VBox.log live on the left and the VM window on the right.
How can we reproduce this?
- Host: Windows 10 19045.6466, AMD RX 5700 XT, driver 32.0.21045.20006 (AMD Software 26.10.2). No integrated GPU is present, so the display path cannot fall back.
- VirtualBox 7.2.22; VM with display controller VMSVGA, "Enable 3D Acceleration" on. Video memory is not critical: the captures were made with 141 MB (51), 179 MB (9) and 256 MB (2), so any value that lets 3D acceleration work is fine. Guest: Kali Linux 2026.2 with XFCE (Xorg at depth 24). The Extension Pack is not required.
- Log in to the guest; the X11 desktop must be up (XFCE, Xorg at depth 24 in our captures). A bare desktop with no windows open is not enough in our experience - the fault needs windows that make the guest generate image surfaces, which step 5 covers.
- Optional, only to get a hit quickly: load the host GPU with a WebGL page (two browser windows, any heavy WebGL demo; measured 76-78% utilisation). CPU load alone also reaches the fault, just slower.
- In the guest, repeatedly open views that force thumbnail/image-surface generation: the file chooser on a folder of images, or the desktop settings -> background folder selector (
Folder: -> Other location...) pointed at a directory of images. The selector/dialog is the reliable trigger; opening the settings window alone is not. - Typically within 1-3 minutes the host shows the Application Error dialog. Timing measured on this machine: with GPU load, 3 of 5 runs faulted within 180 s (68, 75, 76 s), fastest 49 s and 2 min 40 s on the 26.10.2 capture; with CPU-only load, 2-8 minutes; with no host load at all, left running with one dialog open, captures at 51 and 54 minutes.
https://github.com/user-attachments/assets/2fef91c0-75ed-44df-b394-06d301e3ba3f
Did you upload all of your necessary log files, screenshots, etc.?
- Yes, I've uploaded all pertinent files to this issue.
- 主要语言
- C
- 星标
- 1.4k
- 派生
- 219
- PR 合并指标
- 30 天内没有已合并 PR
环境准备
- 没有 Dockerfile 或 Docker Compose 文件
- 没有 Pull Request 模板
- 阅读贡献指南
从这里开始
- 先读完整个 Issue,再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库,在一个分支上完成修改。
- 提交 Pull Request,并在描述里引用这个 Issue 编号。
VirtualBox/virtualbox 的其他 Issue
-
bug
难度 1/5 1 小时以内 新手友好度 84/100
VirtualBox/virtualbox#996 ·
维护者通常 6 天内回复
-
enhancement
难度 2/5 1-3 小时 新手友好度 68/100
VirtualBox/virtualbox#979 ·
维护者通常 6 天内回复
-
[Bug]: Guest Additions 7.2.20 vboxvideo fails to build on Linux kernel 6.12.113 (drm_fb_helper_restore_fbdev_mode_unlocked: too few arguments)可能已有人在做 @Berbe 于 18 天前认领。 未关闭bug
难度 2/5 1-3 小时 新手友好度 76/100
VirtualBox/virtualbox#918 · 9 条评论 ·
维护者通常 6 天内回复
-
bug
难度 4/5 3-5 天 新手友好度 25/100
VirtualBox/virtualbox#995 ·
维护者通常 6 天内回复
-
bug
难度 3/5 1-2 天 新手友好度 45/100
VirtualBox/virtualbox#994 ·
维护者通常 6 天内回复
查看 VirtualBox/virtualbox 的全部 Issue
相似的 Issue
-
难度 2/5 1-3 小时 新手友好度 80/100
ESPresense/ESPresense#2680 · 2 条评论 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 74/100
ImageMagick/ImageMagick#9134 ·
维护者通常 1 天内回复
-
Builder: a structure opened without calculation data shows the Diffraction (fibre) view by default未关闭bug
难度 2/5 1-3 小时 新手友好度 65/100
FriendsofECCE/ECCE#1245 · 1 条评论 ·
-
Arch: ARM/AArch64 Doc
难度 1/5 1 小时以内 新手友好度 90/100
RT-Thread/rt-thread#12082 · 1 条评论 ·
维护者通常 2 天内回复
-
难度 2/5 1-3 小时 新手友好度 72/100
apache/nuttx-apps#4108 · 4 条评论 ·
维护者通常 1 天内回复