Thread (90 messages) flat view 90 messages, 8 authors, 13h ago

Re: [PATCH v2 04/40] mm: consistently validate VMA state after mmap[_prepare] hooks

From: sashiko-bot@kernel.org
Date: 2026-09-14 16:18:07
Also in: bpf, dri-devel, kvm, kvmarm, linux-perf-users, linux-rdma, linux-s390, linux-scsi, selinux

Thank you for your contribution! Sashiko AI review found 4 potential issue(s) to consider:

New issues:
- [High] Premature execution of vma_close() before unmap_region() during validation failure creates a Use-After-Free race window.
- [High] Restoring vma->vm_start before calling unmap_region() on the validation error path leaves orphaned PTEs, causing permanent UAF.

Pre-existing issues:
- [High] mmap_validate() fails to check if a driver erroneously sets VMA_WRITE_BIT on a mapping without VMA_MAYWRITE_BIT, allowing read-only files to be mapped writable.
- [High] mmap_validate() fails to check if a driver alters vma->vm_end, which silently corrupts the maple tree.
--

commit d07b85544000595b057fc1ff43289af0359a7b6c
Author: Lorenzo Stoakes (ARM) [off-list ref]

mm: consistently validate VMA state after mmap[_prepare] hooks

This patch generalizes VMA state validation checks for both mmap_prepare
and legacy mmap hooks. It aims to ensure that drivers do not violate
invariants, such as altering the VMA start address or creating writable
mappings on read-only files, and adds validations for insert_vm_struct().
quoted hunk ↗ jump to hunk
diff --git a/mm/internal.h b/mm/internal.h
index fe576d468af45..1ed11eaf276a7 100644
--- a/mm/internal.h
+++ b/mm/internal.h
[ ... ]
quoted hunk ↗ jump to hunk
@@ -239,26 +260,16 @@ static inline int mmap_file(struct file *file, struct vm_area_struct *vma)
 	 */
 	if (unlikely(err || !vma->vm_ops))
 		vma->vm_ops = &vma_dummy_vm_ops;
+	if (unlikely(err))
+		return err;
 
-	return err;
-}
-
-/*
- * If the VMA has a close hook then close it, and since closing it might leave
- * it in an inconsistent state which makes the use of any hooks suspect, clear
- * them down by installing dummy empty hooks.
- */
-static inline void vma_close(struct vm_area_struct *vma)
-{
-	if (vma->vm_ops && vma->vm_ops->close) {
-		vma->vm_ops->close(vma);
-
-		/*
-		 * The mapping is in an inconsistent state, and no further hooks
-		 * may be invoked upon it.
-		 */
-		vma->vm_ops = &vma_dummy_vm_ops;
+	err = mmap_hook_validate(prev_start, &prev_flags, vma);
+	if (unlikely(err)) {
+		vma->vm_start = prev_start;
[Severity: High]
If a driver's legacy mmap hook erroneously decreases vma->vm_start (expanding
the VMA downwards) and populates PTEs in that expanded range,
mmap_hook_validate() catches the violation. By forcing vma->vm_start back to
prev_start here before __mmap_new_file_vma() invokes unmap_region(), doesn't
that initialize the unmap operation with the smaller, restored address range?
This would cause unmap_region() to skip the expanded PTEs, leaving them
orphaned and causing a permanent use-after-free since the backing memory gets
freed. Should the unmap occur before restoring the start address?
+		vma_close(vma);
[Severity: High]
Is it safe to invoke vma_close() here immediately on validation failure?
If a driver successfully populates page tables (for example, via
remap_pfn_range) but fails the mmap_hook_validate() checks, calling
vma_close() frees driver memory prematurely. Since unmap_region() is not
called until mmap_file() returns to __mmap_new_file_vma(), hardware PTEs
remain populated while the backing resources are freed, creating a
use-after-free race window.
quoted hunk ↗ jump to hunk
 	}
+
+	return err;
 }
 
 /* unmap_vmas is in mm/memory.c */
diff --git a/mm/vma.c b/mm/vma.c
index dc74b3d721cc5..626a18c08a269 100644
--- a/mm/vma.c
+++ b/mm/vma.c
[ ... ]
quoted hunk ↗ jump to hunk
@@ -2777,6 +2762,76 @@ static void __mmap_complete(struct mmap_state *map, struct vm_area_struct *vma)
 	vma_set_page_prot(vma);
 }
 
+/* Check to ensure that the VMA flags of a newly mapped VMA are sane. */
+static int mmap_validate_vma_flags(const vma_flags_t *flags)
+{
[ ... ]
+/* Check to ensure a driver hasn't done something crazy. */
+static int mmap_validate(unsigned long prev_start,
+			 unsigned long curr_start,
+			 const vma_flags_t *prev_flags,
+			 const vma_flags_t *curr_flags)
+{
+	bool was_maywrite, is_maywrite;
+
+	/* Drivers cannot alter the address of the VMA. */
+	if (WARN_ON_ONCE(prev_start != curr_start))
+		return -EINVAL;
[Severity: High]
This isn't a bug introduced by this patch, but since the patch aims to
implement comprehensive VMA state validation after mmap hooks, shouldn't it
also verify that the end address (vma->vm_end) remains unmodified? If a
buggy driver alters vm_end, the VMA object's bounds will mismatch the maple
tree node's pre-allocated bounds. This structural inconsistency corrupts
the maple tree and can lead to kernel panics later.
+
+	was_maywrite = vma_flags_test(prev_flags, VMA_MAYWRITE_BIT);
+	is_maywrite = vma_flags_test(curr_flags, VMA_MAYWRITE_BIT);
+
+	/* A driver may not make a previously unwritable mapping writable. */
+	if (WARN_ON_ONCE(!was_maywrite && is_maywrite))
+		return -EINVAL;
[Severity: High]
This is a pre-existing issue, but this check appears to bypass a critical
condition. It only evaluates VMA_MAYWRITE_BIT. If a malicious or buggy
driver bitwise-ORs VMA_WRITE_BIT into vma->flags in its mmap hook without
setting VMA_MAYWRITE_BIT, wouldn't it bypass this check completely? This
could allow a read-only file mapping to be committed with VM_WRITE set,
granting hardware write permissions and allowing an unprivileged user to
overwrite read-only files.
+
+	return mmap_validate_vma_flags(curr_flags);
+}
-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260914-b4-mmap-prepare-vma-flag-sanify-v2-0-7d9781ed5361@kernel.org?part=4
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help