On 2021/01/25 22:32, Michal Hocko wrote:
> On Fri 22-01-21 19:47:42, Tetsuo Handa wrote:
>> On 2021/01/22 10:35, Andrew Morton wrote:
>>> On Wed, 20 Jan 2021 19:34:36 +0900 Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx> wrote:
>>>
>>>> syzbot is reporting that memdup_user_nul() which receives user-controlled
>>>> size (which can be up to (INT_MAX & PAGE_MASK)) via vfs_write() will hit
>>>> order >= MAX_ORDER path [1].
>
> That is nasty!
That's because -EFAULT will not be detected before memory allocation succeeds.
Fuzzer is passing huge size without corresponding valid buffer.
syscall(__NR_write, r[0], 0x200000c0ul, 0x200000cbul);
>
>>>> Let's add __GFP_NOWARN to memdup_user_nul() as with commit 6c8fcc096be9d02f
>>>> ("mm: don't let userspace spam allocations warnings"). Also use GFP_USER as
>>>> with commit 6c2c97a24f096e32 ("memdup_user(): switch to GFP_USER").
>
> No, this is papering over a more troubling underlying problem. Userspace
> shouldn't be able to trigger an aribitrary higher order allocations.
That requires inserting max size checking before calling memdup_user_nul().
Oh, scattering around such checking is not nice. Add max length argument
into memdup_user_nul() like strndup_user() ?
> Those users with a large size to copy should be really using kvmalloc
> based (e.g vmemdup_user).
No. The caller in this case (writing an entry to smackfs) is not expecting
such large allocations. Sane allocation size would be always less than PAGE_SIZE.
>
>>> That commit failed to explain why a switch to GFP_USER was performed,
>>> so that commit isn't a good substitute for an explanation of this
>>> change.
>>
>> For example, commit 2f77d107050abc14 ("Fix incorrect user space access locking
>> in mincore()") silently converted GFP_KERNEL to GFP_USER.
>>
>> #define GFP_KERNEL (__GFP_RECLAIM | __GFP_IO | __GFP_FS)
>> #define GFP_USER (__GFP_RECLAIM | __GFP_IO | __GFP_FS | __GFP_HARDWALL)
>>
>> * %GFP_KERNEL is typical for kernel-internal allocations. The caller requires
>> * %ZONE_NORMAL or a lower zone for direct access but can direct reclaim.
>>
>> * %GFP_USER is for userspace allocations that also need to be directly
>> * accessibly by the kernel or hardware. It is typically used by hardware
>> * for buffers that are mapped to userspace (e.g. graphics) that hardware
>> * still must DMA to. cpuset limits are enforced for these allocations.
>>
>> * %__GFP_HARDWALL enforces the cpuset memory allocation policy.
>>
>>>
>>> So... please fully describe the reason for this change right here in
>>> this patch's changelog.
>>
>> I guess that GFP_USER is chosen by cautious developers when memory is
>> allocated by userspace request. Is there a guideline for when to use GFP_USER ?
>
> I do not think we have anything better than the above. GFP_USER is
> indeed used for userspace controlable allocations. So they can be a
> subject to a more strict cpu policy. memdup_user_nul looks like a good
> fit for GFP_USER to me. memdup_user and other variant already does this.
>
Hmm, Sabyrzhan already proposed a patch that adds size check to the caller, but it seems
that that patch missed smk_write_ambient()/smk_write_onlycap()/smk_write_unconfined() etc.
Oh, bug-prone approach. Why not handle at memdup_user_nul() side?
