Skip to contentSkip to conversation

drm/gma500: bound the fbdev fault loop to the framebuffer

PATCH New Baul Lee 2 participants 2 messages Aug 6

Subsystems (4)
lore
Share & tools
lore

Baul Lee

author
#0 lore raw

PATCH 1/1 drm/gma500: bound the fbdev fault loop to the framebuffer

psb_fbdev_vm_fault() maps as many frames as the caller asked for:

	unsigned long pfn = info->fix.smem_start >> PAGE_SHIFT;
	unsigned long page_num = vma_pages(vma);

	for (i = 0; i < page_num; ++i) {
		err = vmf_insert_mixed(vma, address, pfn);
		...
		address += PAGE_SIZE;
		++pfn;
	}

The trip count is the VMA length and the pfn walks up from the framebuffer base. info->fix.smem_len, the extent of the framebuffer, is not read anywhere in the handler, so a mapping longer than the framebuffer installs present PTEs for the physical frames above it.

Nothing else on the path bounds the length. psb_fbdev_fb_mmap() only requires vm_pgoff to be zero, and fb_mmap() passes the length through.

On x86_64 an unprivileged process in group video mapped 4800 pages of a 600-page framebuffer; every page was present and writable, and the 4200 past the object were the frames above it in physical memory. Nothing is
logged: the kernel installs valid PTEs and every access through them comes from userspace.

Reject a fault whose page offset is already past the framebuffer and clamp the loop to the frames the framebuffer owns.

Discovered by XBOW, triaged by Baul Lee <[email protected]>

Fixes 0867b42113ec ("staging: gma500: Intel GMA500 staging driver")
---
 drivers/gpu/drm/gma500/fbdev.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/gpu/drm/gma500/fbdev.c +6 -0
index d1e93588234f..393a939a0c37 100644
@@ -26,9 +26,15 @@ static vm_fault_t psb_fbdev_vm_fault(struct vm_fault *vmf)
 	unsigned long address = vmf->address - (vmf->pgoff << PAGE_SHIFT);
 	unsigned long pfn = info->fix.smem_start >> PAGE_SHIFT;
 	vm_fault_t err = VM_FAULT_SIGBUS;
+	unsigned long obj_pages = info->fix.smem_len >> PAGE_SHIFT;
 	unsigned long page_num = vma_pages(vma);
 	unsigned long i;
 
+	if (vmf->pgoff >= obj_pages)
+		return VM_FAULT_SIGBUS;
+	if (page_num > obj_pages)
+		page_num = obj_pages;
+
 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 
 	for (i = 0; i < page_num; ++i) {
Signature · 1 line
2.50.1 (Apple Git-155)
#1 lore raw
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] Modifying vma->vm_page_prot within the page fault handler violates mm locking rules and creates a data race. --

--- Patch [1]: [PATCH] drm/gma500: bound the fbdev fault loop to the framebuffer --- commit e85cf59eaca6225d308fd9d0f4608078a7b0c51e
Author: Baul Lee <[email protected]>

drm/gma500: bound the fbdev fault loop to the framebuffer

This commit limits the number of pages mapped in the page fault handler to the actual size of the framebuffer object. It rejects faults that start past the end of the object and clamps requested page counts, preventing userspace from mapping physical frames beyond the framebuffer boundary.

Quoted 8 lines
drivers/gpu/drm/gma500/fbdev.c +0 -0
index d1e93588234f5..393a939a0c376 100644
@@ -26,9 +26,15 @@ static vm_fault_t psb_fbdev_vm_fault(struct vm_fault *vmf)
 	unsigned long address = vmf->address - (vmf->pgoff << PAGE_SHIFT);
 	unsigned long pfn = info->fix.smem_start >> PAGE_SHIFT;
 	vm_fault_t err = VM_FAULT_SIGBUS;

drivers/gpu/drm/gma500/fbdev.c static vm_fault_t psb_fbdev_vm_fault(struct vm_fault *vmf) · line 38

Quoted 10 lines
drivers/gpu/drm/gma500/fbdev.c +6 -0
+	unsigned long obj_pages = info->fix.smem_len >> PAGE_SHIFT;
 	unsigned long page_num = vma_pages(vma);
 	unsigned long i;
 
+	if (vmf->pgoff >= obj_pages)
+		return VM_FAULT_SIGBUS;
+	if (page_num > obj_pages)
+		page_num = obj_pages;
+
 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);

[Severity: High]
This is a pre-existing issue, but could this concurrent modification corrupt page protection bits?

Modifying vma->vm_page_prot dynamically in the psb_fbdev_vm_fault handler happens under a read lock, such as mmap_read_lock or the per-VMA read lock.

On 32-bit systems with PAE where pgprot_t is 64-bit, writing to it without exclusive locks can cause torn writes. It also appears to race with other concurrent faults or mprotect calls, potentially reverting security boundaries or cache attributes since mm locking guidelines require holding mmap_write_lock when modifying vma->vm_page_prot.

Should this assignment be moved to the mmap callback where mmap_write_lock is held, rather than modifying it directly in the fault handler?

Quoted 3 lines
drivers/gpu/drm/gma500/fbdev.c +0 -0
 	for (i = 0; i < page_num; ++i) {
 		err = vmf_insert_mixed(vma, address, pfn);

if (unlikely(err & VM_FAULT_ERROR))

[ ... ]

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1