In function `c4iw_dealloc_mw`, variable mhp's value is printed after
freed, which triggers undefined behavior according to this post:
https://trust-in-soft.com/dangling-pointer-indeterminate/.
This commit fixes it by swapping the order of `kfree` and `pr_debug`.
Signed-off-by: Shaobo He <[email protected]>
---
drivers/infiniband/hw/cxgb4/mem.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/infiniband/hw/cxgb4/mem.c b/drivers/infiniband/hw/cxgb4/mem.c
index 7b76e6f..bb8e0bc 100644
--- a/drivers/infiniband/hw/cxgb4/mem.c
+++ b/drivers/infiniband/hw/cxgb4/mem.c
@@ -684,8 +684,8 @@ int c4iw_dealloc_mw(struct ib_mw *mw)
mhp->wr_waitp);
kfree_skb(mhp->dereg_skb);
c4iw_put_wr_wait(mhp->wr_waitp);
- kfree(mhp);
pr_debug("ib_mw %p mmid 0x%x ptr %p\n", mw, mmid, mhp);
+ kfree(mhp);
return 0;
}
--
2.7.4
On Thu, 2019-02-28 at 15:38 -0700, Shaobo He wrote:
+AD4 In function +AGA-c4iw+AF8-dealloc+AF8-mw+AGA, variable mhp's value is printed after
+AD4 freed, which triggers undefined behavior according to this post:
+AD4 https://trust-in-soft.com/dangling-pointer-indeterminate/.
+AD4
+AD4 This commit fixes it by swapping the order of +AGA-kfree+AGA and +AGA-pr+AF8-debug+AGA.
+AD4
+AD4 Signed-off-by: Shaobo He +ADw-shaobo+AEA-cs.utah.edu+AD4
+AD4 ---
+AD4 drivers/infiniband/hw/cxgb4/mem.c +AHw 2 +--
+AD4 1 file changed, 1 insertion(+-), 1 deletion(-)
+AD4
+AD4 diff --git a/drivers/infiniband/hw/cxgb4/mem.c b/drivers/infiniband/hw/cxgb4/mem.c
+AD4 index 7b76e6f..bb8e0bc 100644
+AD4 --- a/drivers/infiniband/hw/cxgb4/mem.c
+AD4 +-+-+- b/drivers/infiniband/hw/cxgb4/mem.c
+AD4 +AEAAQA -684,8 +-684,8 +AEAAQA int c4iw+AF8-dealloc+AF8-mw(struct ib+AF8-mw +ACo-mw)
+AD4 mhp-+AD4-wr+AF8-waitp)+ADs
+AD4 kfree+AF8-skb(mhp-+AD4-dereg+AF8-skb)+ADs
+AD4 c4iw+AF8-put+AF8-wr+AF8-wait(mhp-+AD4-wr+AF8-waitp)+ADs
+AD4 - kfree(mhp)+ADs
+AD4 pr+AF8-debug(+ACI-ib+AF8-mw +ACU-p mmid 0x+ACU-x ptr +ACU-p+AFw-n+ACI, mw, mmid, mhp)+ADs
+AD4 +- kfree(mhp)+ADs
+AD4 return 0+ADs
+AD4 +AH0
Please quote the relevant paragraphs from the C standard. All I have found
about free() in ISO/IEC 9899:2017 is the following:
Description
The free function causes the space pointed to by ptr to be deallocated, that
is, made available for further allocation. If ptr is a null pointer, no
action occurs. Otherwise, if the argument does not match a pointer earlier
returned by a memory management function, or if the space has been
deallocated by a call to free or realloc, the behavior is undefined.
That is not sufficient to claim that the above code triggers undefined
behavior.
Bart.
On Thu, 2019-02-28 at 16:18 -0700, Shaobo He wrote:
+AD4 I can't afford a pdf version of the C standard. So I looked at the draft version
+AD4 used in the link I put in the commit message. It says (in 6.2.4:2),
+AD4
+AD4 +AGAAYABg
+AD4 The lifetime of an object is the portion of program execution during which
+AD4 storage is guaranteed to be reserved for it. An object exists, has a constant
+AD4 address, and retains its last-stored value throughout its lifetime. If an object
+AD4 is referred to outside of its lifetime, the behavior is undefined. The value of
+AD4 a pointer becomes indeterminate when the object it points to (or just past)
+AD4 reaches the end of its lifetime.
+AD4 +AGAAYABg
+AD4 I couldn't find the definition of lifetime over a dynamically allocated object
+AD4 in the draft of C standard. I refer to this link
+AD4 (https://en.cppreference.com/w/c/language/lifetime) which suggests that the
+AD4 lifetime of an allocated object ends after the deallocation function is called
+AD4 upon it.
+AD4
+AD4 I think maybe the more problematic issue is that the value of a freed pointer is
+AD4 intermediate.
In another section of the same draft I found the following:
J.2 Undefined behavior +AFs ... +AF0 The value of a pointer that refers to space
deallocated by a call to the free or realloc function is used (7.22.3).
Since the C standard explicitly refers to free() and realloc(), does that
mean that that statement about undefined behavior does not apply to munmap()
(for user space code) nor to kfree() (for kernel code)?
Bart.
I can't afford a pdf version of the C standard. So I looked at the draft version
used in the link I put in the commit message. It says (in 6.2.4:2),
```
The lifetime of an object is the portion of program execution during which
storage is guaranteed to be reserved for it. An object exists, has a constant
address, and retains its last-stored value throughout its lifetime. If an object
is referred to outside of its lifetime, the behavior is undefined. The value of
a pointer becomes indeterminate when the object it points to (or just past)
reaches the end of its lifetime.
```
I couldn't find the definition of lifetime over a dynamically allocated object
in the draft of C standard. I refer to this link
(https://en.cppreference.com/w/c/language/lifetime) which suggests that the
lifetime of an allocated object ends after the deallocation function is called
upon it.
I think maybe the more problematic issue is that the value of a freed pointer is
intermediate.
Shaobo
On 2/28/19 3:56 PM, Bart Van Assche wrote:
> On Thu, 2019-02-28 at 15:38 -0700, Shaobo He wrote:
>> In function `c4iw_dealloc_mw`, variable mhp's value is printed after
>> freed, which triggers undefined behavior according to this post:
>> https://trust-in-soft.com/dangling-pointer-indeterminate/.
>>
>> This commit fixes it by swapping the order of `kfree` and `pr_debug`.
>>
>> Signed-off-by: Shaobo He <[email protected]>
>> ---
>> drivers/infiniband/hw/cxgb4/mem.c | 2 +-
>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/drivers/infiniband/hw/cxgb4/mem.c b/drivers/infiniband/hw/cxgb4/mem.c
>> index 7b76e6f..bb8e0bc 100644
>> --- a/drivers/infiniband/hw/cxgb4/mem.c
>> +++ b/drivers/infiniband/hw/cxgb4/mem.c
>> @@ -684,8 +684,8 @@ int c4iw_dealloc_mw(struct ib_mw *mw)
>> mhp->wr_waitp);
>> kfree_skb(mhp->dereg_skb);
>> c4iw_put_wr_wait(mhp->wr_waitp);
>> - kfree(mhp);
>> pr_debug("ib_mw %p mmid 0x%x ptr %p\n", mw, mmid, mhp);
>> + kfree(mhp);
>> return 0;
>> }
>
> Please quote the relevant paragraphs from the C standard. All I have found
> about free() in ISO/IEC 9899:2017 is the following:
>
> Description
> The free function causes the space pointed to by ptr to be deallocated, that
> is, made available for further allocation. If ptr is a null pointer, no
> action occurs. Otherwise, if the argument does not match a pointer earlier
> returned by a memory management function, or if the space has been
> deallocated by a call to free or realloc, the behavior is undefined.
>
> That is not sufficient to claim that the above code triggers undefined
> behavior.
>
> Bart.
>
Good catch. But if we agree on that memory management functions are those
specified by the C standard, would it be OK to ignore so-called use after free
or double free bugs for the kernel as C standard does not apply to kfree?
On 2/28/19 4:33 PM, Bart Van Assche wrote:
> On Thu, 2019-02-28 at 16:18 -0700, Shaobo He wrote:
>> I can't afford a pdf version of the C standard. So I looked at the draft version
>> used in the link I put in the commit message. It says (in 6.2.4:2),
>>
>> ```
>> The lifetime of an object is the portion of program execution during which
>> storage is guaranteed to be reserved for it. An object exists, has a constant
>> address, and retains its last-stored value throughout its lifetime. If an object
>> is referred to outside of its lifetime, the behavior is undefined. The value of
>> a pointer becomes indeterminate when the object it points to (or just past)
>> reaches the end of its lifetime.
>> ```
>> I couldn't find the definition of lifetime over a dynamically allocated object
>> in the draft of C standard. I refer to this link
>> (https://en.cppreference.com/w/c/language/lifetime) which suggests that the
>> lifetime of an allocated object ends after the deallocation function is called
>> upon it.
>>
>> I think maybe the more problematic issue is that the value of a freed pointer is
>> intermediate.
>
> In another section of the same draft I found the following:
>
> J.2 Undefined behavior [ ... ] The value of a pointer that refers to space
> deallocated by a call to the free or realloc function is used (7.22.3).
>
> Since the C standard explicitly refers to free() and realloc(), does that
> mean that that statement about undefined behavior does not apply to munmap()
> (for user space code) nor to kfree() (for kernel code)?
>
> Bart.
>
On Thu, 2019-02-28 at 16:57 -0700, Shaobo He wrote:
> Good catch. But if we agree on that memory management functions are those
> specified by the C standard, would it be OK to ignore so-called use after free
> or double free bugs for the kernel as C standard does not apply to kfree?
No, most kernel use-after-free bugs are real bugs. This one might be
technically a bug by certain readings of the standard, but it's a non-
issue. Real use-after-free bugs don't just look at the value of a local
stack variable to get the memory's old address (which is what this does,
and the same could be achieved and be totally in spec by doing this:
old_ptr = mhp;
kfree(mhp);
pr_debug("%p\n", old_ptr);)
Real use after free things would actually dereference the pointer to
either read or write from the old memory region. That leads to data
corruption or kernel data leaks. Plus, in this case, the purpose of
printing the literal value of mhp is simply to provide a unique name for
tracing purposes. Since kfree() doesn't alter the local stack variable,
the name is still present in the local stack variable at the point you
call pr_debug().
It could be fixed. It's not like this patch is wrong. But I wouldn't
submit it this late in the -rc cycle, I'd just take it for next.
> On 2/28/19 4:33 PM, Bart Van Assche wrote:
> > On Thu, 2019-02-28 at 16:18 -0700, Shaobo He wrote:
> > > I can't afford a pdf version of the C standard. So I looked at the draft version
> > > used in the link I put in the commit message. It says (in 6.2.4:2),
> > >
> > > ```
> > > The lifetime of an object is the portion of program execution during which
> > > storage is guaranteed to be reserved for it. An object exists, has a constant
> > > address, and retains its last-stored value throughout its lifetime. If an object
> > > is referred to outside of its lifetime, the behavior is undefined. The value of
> > > a pointer becomes indeterminate when the object it points to (or just past)
> > > reaches the end of its lifetime.
> > > ```
> > > I couldn't find the definition of lifetime over a dynamically allocated object
> > > in the draft of C standard. I refer to this link
> > > (https://en.cppreference.com/w/c/language/lifetime) which suggests that the
> > > lifetime of an allocated object ends after the deallocation function is called
> > > upon it.
> > >
> > > I think maybe the more problematic issue is that the value of a freed pointer is
> > > intermediate.
> >
> > In another section of the same draft I found the following:
> >
> > J.2 Undefined behavior [ ... ] The value of a pointer that refers to space
> > deallocated by a call to the free or realloc function is used (7.22.3).
> >
> > Since the C standard explicitly refers to free() and realloc(), does that
> > mean that that statement about undefined behavior does not apply to munmap()
> > (for user space code) nor to kfree() (for kernel code)?
> >
> > Bart.
> >
--
Doug Ledford <[email protected]>
GPG KeyID: B826A3330E572FDD
Key fingerprint = AE6B 1BDA 122B 23B4 265B 1274 B826 A333 0E57 2FDD
On Thu, 28 Feb 2019, Shaobo He wrote:
> I think maybe the more problematic issue is that the value of a freed pointer
> is intermediate.
The pointer is not affected by freeing the data it points to. Thus it
definitely has the same value as before and is not indeterminate.
The pointer points now to an area of memory that could now be in use for
different purposes so maybe it could be taken as a dangerous situation.
But situations like that are common in code.
Yes, why wouldn't they be real bugs? I was simply pointing out the irrational
conclusion if the C standard is strictly applied to kernel code.
I think the spirit of the C standard is that one shouldn't rely on the
assumption that the value of a freed pointer does not change, even though in
practice any compiler developers probably won't bother to implement the logic to
change the pointer value or optimizations leveraging it even they are allowed
to. In other words, the original code may be a little bit problematic in the
spirit of the C standard whereas the patch simply makes it totally valid.
If it can be finally submitted, that would be great. If not, I'm totally fine.
Shaobo
On 2019/3/1 7:26, Doug Ledford wrote:
> On Thu, 2019-02-28 at 16:57 -0700, Shaobo He wrote:
>> Good catch. But if we agree on that memory management functions are those
>> specified by the C standard, would it be OK to ignore so-called use after free
>> or double free bugs for the kernel as C standard does not apply to kfree?
>
> No, most kernel use-after-free bugs are real bugs. This one might be
> technically a bug by certain readings of the standard, but it's a non-
> issue. Real use-after-free bugs don't just look at the value of a local
> stack variable to get the memory's old address (which is what this does,
> and the same could be achieved and be totally in spec by doing this:
>
> old_ptr = mhp;
> kfree(mhp);
> pr_debug("%p\n", old_ptr);)
>
> Real use after free things would actually dereference the pointer to
> either read or write from the old memory region. That leads to data
> corruption or kernel data leaks. Plus, in this case, the purpose of
> printing the literal value of mhp is simply to provide a unique name for
> tracing purposes. Since kfree() doesn't alter the local stack variable,
> the name is still present in the local stack variable at the point you
> call pr_debug().
>
> It could be fixed. It's not like this patch is wrong. But I wouldn't
> submit it this late in the -rc cycle, I'd just take it for next.
>
>> On 2/28/19 4:33 PM, Bart Van Assche wrote:
>>> On Thu, 2019-02-28 at 16:18 -0700, Shaobo He wrote:
>>>> I can't afford a pdf version of the C standard. So I looked at the draft version
>>>> used in the link I put in the commit message. It says (in 6.2.4:2),
>>>>
>>>> ```
>>>> The lifetime of an object is the portion of program execution during which
>>>> storage is guaranteed to be reserved for it. An object exists, has a constant
>>>> address, and retains its last-stored value throughout its lifetime. If an object
>>>> is referred to outside of its lifetime, the behavior is undefined. The value of
>>>> a pointer becomes indeterminate when the object it points to (or just past)
>>>> reaches the end of its lifetime.
>>>> ```
>>>> I couldn't find the definition of lifetime over a dynamically allocated object
>>>> in the draft of C standard. I refer to this link
>>>> (https://en.cppreference.com/w/c/language/lifetime) which suggests that the
>>>> lifetime of an allocated object ends after the deallocation function is called
>>>> upon it.
>>>>
>>>> I think maybe the more problematic issue is that the value of a freed pointer is
>>>> intermediate.
>>>
>>> In another section of the same draft I found the following:
>>>
>>> J.2 Undefined behavior [ ... ] The value of a pointer that refers to space
>>> deallocated by a call to the free or realloc function is used (7.22.3).
>>>
>>> Since the C standard explicitly refers to free() and realloc(), does that
>>> mean that that statement about undefined behavior does not apply to munmap()
>>> (for user space code) nor to kfree() (for kernel code)?
>>>
>>> Bart.
>>>
>
On Thu, Feb 28, 2019 at 03:38:38PM -0700, Shaobo He wrote:
> In function `c4iw_dealloc_mw`, variable mhp's value is printed after
> freed, which triggers undefined behavior according to this post:
> https://trust-in-soft.com/dangling-pointer-indeterminate/.
>
> This commit fixes it by swapping the order of `kfree` and `pr_debug`.
>
> Signed-off-by: Shaobo He <[email protected]>
> ---
> drivers/infiniband/hw/cxgb4/mem.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
Discussion aside, this is a worthwile fix. I rewrote the commit
message to avoid referencing 'undefined behavior' though, this is just
a straight up bug in the logging. Another thread could get the same
pointer value for the mhp before the print creating a confusing log.
cxgb4: kfree mhp after the debug print
In function `c4iw_dealloc_mw`, variable mhp's value is printed after
freed, it is clearer to have the print before the kfree.
Otherwise racing threads could allocate another mhp with the same pointer
value and create confusing tracing.
Jason