Re: [PATCH] mm: fix COW mapping handing in generic_access_phys

[David Hildenbrand <david@redhat.com>]

On 28.05.25 16:54, Peter Xu wrote:
> [Add Jason]
> 
> On Wed, May 28, 2025 at 11:59:56AM +0200, David Hildenbrand wrote:
>> On 28.05.25 10:59, David Hildenbrand wrote:
>>> On 28.05.25 03:56, Jinjiang Tu wrote:
>>>> Syzkaller reports a below BUG:
>>>>     ioremap on RAM at 0x0000000022727000 - 0x0000000022727fff
>>>>     WARNING: CPU: 3 PID: 3609 at arch/x86/mm/ioremap.c:216 __ioremap_caller+0x644/0x7f0 arch/x86/mm/ioremap.c:216
>>>>     Modules linked in:
>>>>     CPU: 3 PID: 3609 Comm: syz.2.577 Not tainted 6.6.0+ #63
>>>>     Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.14.0-0-g155821a1990b-prebuilt.qemu.org 04/01/2014
>>>>     RIP: 0010:__ioremap_caller+0x644/0x7f0 arch/x86/mm/ioremap.c:216
>>>>     Call Trace:
>>>>      <TASK>
>>>>      generic_access_phys+0x241/0x480 mm/memory.c:6458
>>>>      __access_remote_vm+0x6af/0x970 mm/memory.c:6535
>>>>      access_process_vm+0x53/0x80 mm/memory.c:6600
>>>>      get_cmdline+0x192/0x380 mm/util.c:1041
>>>>      audit_log_proctitle kernel/auditsc.c:1620 [inline]
>>>>      audit_log_exit+0x1424/0x18c0 kernel/auditsc.c:1811
>>>>      __audit_syscall_exit+0x252/0x2f0 kernel/auditsc.c:2079
>>>>      audit_syscall_exit include/linux/audit.h:356 [inline]
>>>>      syscall_exit_work+0x10f/0x130 kernel/entry/common.c:166
>>>>      __syscall_exit_to_user_mode_work kernel/entry/common.c:205 [inline]
>>>>      syscall_exit_to_user_mode+0x10/0x1e0 kernel/entry/common.c:218
>>>>      do_syscall_64+0x66/0x110 arch/x86/entry/common.c:87
>>>>      entry_SYSCALL_64_after_hwframe+0x78/0xe2
>>>>
>>>> The /dev/mem is mapped with COW mapping, and mremap at the mm->args_start.
>>>> The special pfn mapping is replaced by anon folios due to COW.
>>>> generic_access_phys() is supposed to handle iomem, instead of RAM pfn,
>>>> thus trigger a WARN_ON.
>>>>
>>>> Similar to commit 04c35ab3bdae ("x86/mm/pat: fix VM_PAT handling in
>>>> COW mappings"). check if the pte is special to reject Cowed anon folios.
>>>>
>>>> Signed-off-by: Jinjiang Tu <tujinjiang@huawei.com>
>>>> ---
>>>>     mm/memory.c | 7 +++++++
>>>>     1 file changed, 7 insertions(+)
>>>>
>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>> index 49199410805c..e1dac84536ee 100644
>>>> --- a/mm/memory.c
>>>> +++ b/mm/memory.c
>>>> @@ -6840,6 +6840,13 @@ int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
>>>>     retry:
>>>>     	if (follow_pfnmap_start(&args))
>>>>     		return -EINVAL;
>>>> +
>>>> +	/* Never return PFNs of anon folios in COW mappings. */
>>>> +	if (!args.special) {
>>>> +		follow_pfnmap_end(&args);
>>>> +		return -EINVAL;
>>>> +	}
>>>> +
>>>>     	prot = args.pgprot;
>>>>     	phys_addr = (resource_size_t)args.pfn << PAGE_SHIFT;
>>>>     	writable = args.writable;
>>>
>>> I assume we trigger this through vma->vm_ops->access, when the vm_ops have generic_access_phys set.
>>>
>>> I still dislike exposing the "special" bit here, as it is absolutely not what we should care about in the caller.
>>>
>>> In case our arch does not support pte_special, you fix will not catch that case ...
>>>
>>> The following might be better:
>>>
>>> diff --git a/mm/memory.c b/mm/memory.c
>>> index 37d8738f5e12e..810adb8d1a53b 100644
>>> --- a/mm/memory.c
>>> +++ b/mm/memory.c
>>> @@ -6681,6 +6681,14 @@ int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
>>>            prot = args.pgprot;
>>>            phys_addr = (resource_size_t)args.pfn << PAGE_SHIFT;
>>>            writable = args.writable;
>>> +
>>> +       /* Refuse (refcounted) anonymous pages in CoW mappings. */
>>> +       if (is_cow_mapping(vma->vm_flags) &&
>>> +           vm_normal_page(vma, addr, ptep_get(args.ptep))) {
>>> +               follow_pfnmap_end(&args);
>>> +               return -EINVAL;
>>> +       }
>>> +
>>
>> Thinking again, we might have a PMD/PUD mapping, so maybe
>> follow_pfnmap_start() should really just refuse any refcounted pages.
> 
> We may want to be careful on this.
> 
> I feel like we can still potentially break drivers that
> follow_pfnmap_start() used to work on debateable things like RAM page
> injections, unless breaking them is the intention.

Yes, that all needs a cleanup likely; it's all very confusing and 
inconsistent.

> 
> OTOH, I also see at least two in-tree drivers set VM_IO|VM_MIXEDMAP:
> 
> *** drivers/gpu/drm/gma500/fbdev.c:
> psb_fbdev_fb_mmap[110]         vm_flags_set(vma, VM_IO | VM_MIXEDMAP | VM_DONTEXPAND | VM_DONTDUMP);
> 
> *** drivers/gpu/drm/omapdrm/omap_gem.c:
> omap_gem_object_mmap[538]      vm_flags_set(vma, VM_DONTEXPAND | VM_DONTDUMP | VM_IO | VM_MIXEDMAP);
> 
> AFAIU, these MIXEDMAP users will still rely on follow_pfnmap_start() to
> work on e.g. RAM pages, because GUP will simply fail them..

Right.

VM_IO essentially tells us "don't touch this memory, it might have side 
effects", such as MMIO, that's why GUP outright refuses VM_IO VMAs.

I am not sure why generic_access_phys() should be allowed to ... touch 
that memory instead?

Confusing.

> 
> Sligtly off-topic: it's also a bit confusing to me on whether the driver
> should set VM_IO for VM_MIXEDMAP.  I think it should because VM_IO says:

Depends on the driver. Some VM_MIXEDMAP users only map ordinary kernel 
allocations, where VM_IO is not required. For these drivers, I really 
don't know.

-- 
Cheers,

David / dhildenb