[PATCH] mm: disallow ioremap'ing ordinary RAM in generic_access_phys()

[PATCH] mm: disallow ioremap'ing ordinary RAM in generic_access_phys()

[David Hildenbrand (Arm)]

generic_access_phys() can end up getting called on CoW PFNMAP mappings,
resulting in it trying to ioremap anonymous folios, which especially
the x86 code doesn't enjoy, and which is in general just odd/wrong.

E.g., with a CoW mapping of a PCI BAR:

ioremap on RAM at 0x0000000045623000 - 0x0000000045623fff
  WARNING: arch/x86/mm/ioremap.c:216 at __ioremap_caller.isra.0+0x4c2/0x5f0
  Call Trace:
   generic_access_phys+0x130/0x4d0 mm/memory.c:7178
   kernfs_vma_access+0x1ce/0x280 fs/kernfs/file.c:437
   __access_remote_vm+0x58f/0x890 mm/memory.c:7256
   mem_rw+0x2a1/0x670 fs/proc/base.c:912

But more generically, we should disallow access to anything that looks
like ordinary RAM in the system: including allocating memory (e.g.,
alloc_pages) and mapping them in a VM_PFNMAP.

For pages with an online memmap, allow only pages marked as reserved
(PageReserved). This also includes memblock allocations, however,
distinguishing them from other MMIO is tricky, and nobody should really
be mapping them in a PFNMAP (except someone abusing /dev/mem, which we
do not care about).

So only allow access if there is no online memmap, or if the pages are
marked reserved. Reject other pages with -EINVAL.

Note that, just like other pfn_to_online_page() users, we won't care about
concurrent memory offlining; mapping memory that can concurrently get
offlined is highly dubious.

Reported-by: syzbot+49b1021becba70c1f3f6@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=49b1021becba70c1f3f6
Reported-by: Nguyen Ngoc Thang <ngocthang2710.1999@gmail.com>
Closes: https://lore.kernel.org/r/20261001152524.171115-1-ngocthang2710.1999@gmail.com/
Fixes: 28b2ee20c7cb ("access_process_vm device memory infrastructure")
Cc: stable@vger.kernel.org
Signed-off-by: David Hildenbrand (Arm) <david@kernel.org>
---
I went back and forth 10 times and concluded that using PageReserved
should be the right thing to check for now.
---
 mm/memory.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/mm/memory.c b/mm/memory.c
index 1f5d5f7d39cd2..5d9f757a00073 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -7226,6 +7226,7 @@ int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
 	resource_size_t phys_addr;
 	pgprot_t prot = __pgprot(0);
 	void __iomem *maddr;
+	struct page *page;
 	int offset = offset_in_page(addr);
 	int ret = -EINVAL;
 	bool writable;
@@ -7240,6 +7241,19 @@ int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
 retry:
 	if (follow_pfnmap_start(&args))
 		return -EINVAL;
+
+	/*
+	 * Reject anything that looks like ordinary RAM, including anonymous
+	 * folios in CoW mappings or allocated kernel memory. Note that we
+	 * cannot distinguish memblock allocations from non-RAM reliably, so
+	 * we cannot easily reject them here.
+	 */
+	page = pfn_to_online_page(args.pfn);
+	if (page && !PageReserved(page)) {
+		follow_pfnmap_end(&args);
+		return -EINVAL;
+	}
+
 	prot = args.pgprot;
 	phys_addr = (resource_size_t)args.pfn << PAGE_SHIFT;
 	writable = args.writable;

---

base-commit: 634cbf5b794b734c78631b92e381cfc85f357bbe

change-id: 20261009-generic_access_phys-0f3e0806305d

--

Cheers,

David

Re: [PATCH] mm: disallow ioremap'ing ordinary RAM in generic_access_phys()

[Lorenzo Stoakes (ARM)]

On Tue, Oct 13, 2026 at 01:41:24PM +0200, David Hildenbrand (Arm) wrote:
> generic_access_phys() can end up getting called on CoW PFNMAP mappings,
> resulting in it trying to ioremap anonymous folios, which especially
> the x86 code doesn't enjoy, and which is in general just odd/wrong.

Ugh god.

>
> E.g., with a CoW mapping of a PCI BAR:
>
> ioremap on RAM at 0x0000000045623000 - 0x0000000045623fff
>   WARNING: arch/x86/mm/ioremap.c:216 at __ioremap_caller.isra.0+0x4c2/0x5f0
>   Call Trace:
>    generic_access_phys+0x130/0x4d0 mm/memory.c:7178
>    kernfs_vma_access+0x1ce/0x280 fs/kernfs/file.c:437
>    __access_remote_vm+0x58f/0x890 mm/memory.c:7256
>    mem_rw+0x2a1/0x670 fs/proc/base.c:912
>
> But more generically, we should disallow access to anything that looks
> like ordinary RAM in the system: including allocating memory (e.g.,
> alloc_pages) and mapping them in a VM_PFNMAP.

Yes.

>
> For pages with an online memmap, allow only pages marked as reserved
> (PageReserved). This also includes memblock allocations, however,
> distinguishing them from other MMIO is tricky, and nobody should really
> be mapping them in a PFNMAP (except someone abusing /dev/mem, which we
> do not care about).

Yeah I guess we shouldn't try too hard for weird edge cases, or at least
can revisit it later potentially.

>
> So only allow access if there is no online memmap, or if the pages are
> marked reserved. Reject other pages with -EINVAL.

Seems reasonable overall.

>
> Note that, just like other pfn_to_online_page() users, we won't care about
> concurrent memory offlining; mapping memory that can concurrently get
> offlined is highly dubious.

Yeah sounds like the user would be doing something crazy in this case,
therefore their problem!

>
> Reported-by: syzbot+49b1021becba70c1f3f6@syzkaller.appspotmail.com
> Closes: https://syzkaller.appspot.com/bug?extid=49b1021becba70c1f3f6
> Reported-by: Nguyen Ngoc Thang <ngocthang2710.1999@gmail.com>
> Closes: https://lore.kernel.org/r/20261001152524.171115-1-ngocthang2710.1999@gmail.com/
> Fixes: 28b2ee20c7cb ("access_process_vm device memory infrastructure")
> Cc: stable@vger.kernel.org
> Signed-off-by: David Hildenbrand (Arm) <david@kernel.org>

LGTM so:

Reviewed-by: Lorenzo Stoakes (ARM) <ljs@kernel.org>

> ---
> I went back and forth 10 times and concluded that using PageReserved
> should be the right thing to check for now.

It seems the sensible choice, yes. And locally, AI review agrees with you, FWIW!

--
Cheers, Lorenzo

Re: [PATCH] mm: disallow ioremap'ing ordinary RAM in generic_access_phys()

[Mike Rapoport]

On Tue, Oct 13, 2026 at 01:41:24PM +0200, David Hildenbrand (Arm) wrote:
> generic_access_phys() can end up getting called on CoW PFNMAP mappings,
> resulting in it trying to ioremap anonymous folios, which especially
> the x86 code doesn't enjoy, and which is in general just odd/wrong.
> 
> E.g., with a CoW mapping of a PCI BAR:
> 
> ioremap on RAM at 0x0000000045623000 - 0x0000000045623fff
>   WARNING: arch/x86/mm/ioremap.c:216 at __ioremap_caller.isra.0+0x4c2/0x5f0
>   Call Trace:
>    generic_access_phys+0x130/0x4d0 mm/memory.c:7178
>    kernfs_vma_access+0x1ce/0x280 fs/kernfs/file.c:437
>    __access_remote_vm+0x58f/0x890 mm/memory.c:7256
>    mem_rw+0x2a1/0x670 fs/proc/base.c:912
> 
> But more generically, we should disallow access to anything that looks
> like ordinary RAM in the system: including allocating memory (e.g.,
> alloc_pages) and mapping them in a VM_PFNMAP.
> 
> For pages with an online memmap, allow only pages marked as reserved
> (PageReserved). This also includes memblock allocations, however,
> distinguishing them from other MMIO is tricky, and nobody should really

With ARCH_KEEP_MEMBLOCK it's actually easy ...

> be mapping them in a PFNMAP (except someone abusing /dev/mem, which we
> do not care about).

... but yeah, it's not worth it.  

> So only allow access if there is no online memmap, or if the pages are
> marked reserved. Reject other pages with -EINVAL.
> 
> Note that, just like other pfn_to_online_page() users, we won't care about
> concurrent memory offlining; mapping memory that can concurrently get
> offlined is highly dubious.
> 
> Reported-by: syzbot+49b1021becba70c1f3f6@syzkaller.appspotmail.com
> Closes: https://syzkaller.appspot.com/bug?extid=49b1021becba70c1f3f6
> Reported-by: Nguyen Ngoc Thang <ngocthang2710.1999@gmail.com>
> Closes: https://lore.kernel.org/r/20261001152524.171115-1-ngocthang2710.1999@gmail.com/
> Fixes: 28b2ee20c7cb ("access_process_vm device memory infrastructure")
> Cc: stable@vger.kernel.org
> Signed-off-by: David Hildenbrand (Arm) <david@kernel.org>

Acked-by: Mike Rapoport (Microsoft) <rppt@kernel.org>

> ---
>  mm/memory.c | 14 ++++++++++++++
>  1 file changed, 14 insertions(+)

-- 
Sincerely yours,
Mike.