drm/vmwgfx: vmw_cmd_draw header.size lower-bound missing - guest-local OOB-read/write

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA512

Hi all,

Following the drm/qxl apply_reloc report from 14 May 2026
(prior thread: https://lore.kernel.org/virtualization/[email protected]/T/),
I audited the DRM cousin-drivers for the same bug-class. vmwgfx exposes a
stronger instance.

Primary site: vmw_cmd_draw in drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c:1558-1607. The verifier does

    maxnum = (header->size - sizeof(cmd->body)) / sizeof(*decl);
    if (cmd->body.numVertexDecls > maxnum) return -EINVAL;

Commit 32b415a9dc2c ("drm/vmwgfx: Validate command header size against SVGA_CMD_MAX_DATASIZE", Oct 2025) added the upper bound but not the lower bound. With header->size < sizeof(cmd->body) the size_t subtraction wraps, maxnum becomes ~SIZE_MAX, and the loop walks attacker-chosen numVertexDecls entries past the cmd-buffer. Each iteration is a 4-byte OOB-read via vmw_cmd_res_check; on a handle collision vmw_resource_relocation_add records the OOB address as rel->offset (29-bit), and vmw_resource_relocations_apply later does a 32-bit kernel write at cb + rel->offset within a 512 MiB window.

Exploitability split:
- OOB-read at first iteration is deterministic.
- OOB-write requires id_loc to resolve to a valid surface handle; surface-spray makes this practical but the write side is probabilistic.

Two cousin sites share the same root cause (details in BUG.md):
- vmw_cmd_dma (line 1495): pointer underflow -> 4-byte OOB-read of

    `suffix->suffixSize`.

- vmw_cmd_shader_define (line 1898): size = header.size - sizeof(body) wraps and reaches vmw_compat_shader_add; ENOMEM-bounded DoS.

Reachability: DRM_VMW_EXECBUF, DRM_RENDER_ALLOW. Any /dev/dri/renderD128 user (default video group) on a VMware or KVM SVGA-II guest. No DRM_AUTH, no DRM_MASTER, no root.

CVSS 3.1: CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H -> 7.8 High (deterministic OOB-read, probabilistic 32-bit kernel write within 512 MiB; guest-local LPE, no host escape claimed).

Suggested fix: per-site lower-bound check on header->size before the subtraction, or the safe pattern using check_mul_overflow / check_add_overflow. A draft patch is sent as the follow-up RFC in this thread.

Attachments: BUG.md (full trace) and repro.c (KASAN-trigger PoC, no exploitation).

Happy to follow your standard timeline per Documentation/process/security-bugs.rst (typically <=7 days post-fix-merge).

Berkant Koc <[email protected]>
GPG: 0C58 8DFD 7620 4987 2842 13EA 0AC5 29C4 1F8A A5D6 -----BEGIN PGP SIGNATURE-----

iHUEARYKAB0WIQQMWI39diBJhyhCE+oKxSnEH4ql1gUCagm/bwAKCRAKxSnEH4ql 1i1PAQCXhrf6CdxpwCifgpjYZJdzspd9dckht9le4AeUJpKRpwD/csbOYVl+oMDY mekPtFhNq+TujSRVzvRZ0MKzOawzAwU=
=7oh3
-----END PGP SIGNATURE-----

You sent this to a public list, there is no "timeline" needed.

Just submit a fix for this now please.

thanks,

greg k-h

Commit 32b415a9dc2c ("drm/vmwgfx: Validate command header size against SVGA_CMD_MAX_DATASIZE") added an upper bound on the user-supplied SVGA3dCmdHeader.size field but no matching lower bound. When header->size is smaller than sizeof(cmd->body), the size_t subtraction in expressions like

  maxnum = (header->size - sizeof(cmd->body)) / sizeof(*decl);

underflows. The subsequent bound check

  if (cmd->body.numVertexDecls > maxnum) return -EINVAL;

is bypassed because maxnum is ~SIZE_MAX, and the loop walks attacker-chosen entries past the command buffer.

In vmw_cmd_draw this leads to a 4-byte OOB-read per iteration via vmw_cmd_res_check(&decl[i].array.surfaceId, ...); on a surface-handle collision, vmw_resource_relocation_add records the OOB address as rel->offset (29-bit bitfield), and vmw_resource_relocations_apply later performs a 32-bit kernel write at cb + rel->offset.

The same root cause is present in vmw_cmd_dma (suffix pointer-arith underflow leading to OOB-read of suffix->suffixSize) and vmw_cmd_shader_define (size_t wraparound passed to vmw_compat_shader_add).

Reachable via DRM_VMW_EXECBUF (DRM_RENDER_ALLOW). Reject undersized headers at all three sites before the subtraction.

---
 drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)

index e1f18020170a..6f9c7d61cc66 100644
@@ -1506,6 +1506,12 @@ static int vmw_cmd_dma(struct vmw_private *dev_priv,
 	bool dirty;
 
 	cmd = container_of(header, typeof(*cmd), header);
+
+	if (unlikely(header->size < sizeof(cmd->body) + sizeof(*suffix))) {
+		VMW_DEBUG_USER("DMA cmd header.size too small.\n");
+		return -EINVAL;
+	}
+
 	suffix = (SVGA3dCmdSurfaceDMASuffix *)((unsigned long) &cmd->body +
 					       header->size - sizeof(*suffix));
 

@@ -1572,6 +1578,12 @@ static int vmw_cmd_draw(struct vmw_private *dev_priv,
 		return ret;
 
 	cmd = container_of(header, typeof(*cmd), header);
+
+	if (unlikely(header->size < sizeof(cmd->body))) {
+		VMW_DEBUG_USER("Draw cmd header.size smaller than body.\n");
+		return -EINVAL;
+	}
+
 	maxnum = (header->size - sizeof(cmd->body)) / sizeof(*decl);
 
 	if (unlikely(cmd->body.numVertexDecls > maxnum)) {

@@ -1915,6 +1927,11 @@ static int vmw_cmd_shader_define(struct vmw_private *dev_priv,
 	if (unlikely(!dev_priv->has_mob))
 		return 0;
 
+	if (unlikely(cmd->header.size < sizeof(cmd->body))) {
+		VMW_DEBUG_USER("Shader define cmd header.size smaller than body.\n");
+		return -EINVAL;
+	}
+
 	size = cmd->header.size - sizeof(cmd->body);
 	ret = vmw_compat_shader_add(dev_priv, vmw_context_res_man(ctx),
 				    cmd->body.shid, cmd + 1, cmd->body.type,

2.47.3

I think you might have forgotten to disclose the tool/llm you've used for this. Using Claude Opus 4.7 I've found the same issues and the series fixing those bugs is available on dri-devel at
https://patchwork.freedesktop.org/series/166024/ . If you've used a different llm I'd love to know which one. And if you have a completely different tool for detecting those issues that'd be great to know as well.

z

Thanks for the catch, you're right, I should have flagged the tooling in the patch body. And cross-pollination with your own series on patchwork is a good outcome.

Tooling: berkoc-pipeline, a custom RAG framework on Claude Opus 4.7 (Anthropic CVP cohort, May 2026). Full agentic stack: multi-tool execution (filesystem, web fetch, code execution), parallel subagent orchestration with adaptive task decomposition, extended-thinking integration, retrieval-augmented context over a file-based semantic knowledge base, MCP-style integration patterns. 7-step pre-disclosure validation gate, manual verification on every finding before submit.

v2 of this patch will include the formal trailer:

Happy to send v2 with the trailer formalised, or keep the methodology disclosure to follow-up comments if you prefer.

Berkant