Re: [Qemu-stable] [Qemu-devel] [PATCH] fw_cfg: fix memory corruption when all fw_cfg slots are used

[Laszlo Ersek]

On 01/09/18 14:18, Marcel Apfelbaum wrote:
> On 09/01/2018 15:09, Laszlo Ersek wrote:
> 
> Hi Laszlo,
> 
> I'll respond first to this mail' I'll take my time with the rest :)
> 
>> On 01/08/18 22:50, Marcel Apfelbaum wrote:
>>> When all the fw_cfg slots are used, a write is made outside the
>>> bounds of the fw_cfg files array as part of the sort algorithm.
>>>
>>> Fix it by avoiding an unnecessary array element move.
>>> Fix also an assert while at it.
>>>
>>> Signed-off-by: Marcel Apfelbaum <address@hidden>
>>> ---
>>>   hw/nvram/fw_cfg.c | 6 ++++--
>>>   1 file changed, 4 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/hw/nvram/fw_cfg.c b/hw/nvram/fw_cfg.c
>>> index 753ac0e4ea..4313484b21 100644
>>> --- a/hw/nvram/fw_cfg.c
>>> +++ b/hw/nvram/fw_cfg.c
>>> @@ -784,7 +784,7 @@ void fw_cfg_add_file_callback(FWCfgState *s, 
>>> const char *filename,
>>>        * index and "i - 1" is the one being copied from, thus the
>>>        * unusual start and end in the for statement.
>>>        */
>>> -    for (i = count + 1; i > index; i--) {
>>> +    for (i = count; i > index; i--) {
>>>           s->files->f[i] = s->files->f[i - 1];
>>>           s->files->f[i].select = cpu_to_be16(FW_CFG_FILE_FIRST + i);
>>>           s->entries[0][FW_CFG_FILE_FIRST + i] =
>>
>> This hunk looks correct to me.
> 
> After my change or before?

Well, the source code doesn't have "hunks", patches have hunks. :)

So, I meant, this part of your patch was correct, IMO.

> 
> I think I am right.
> At this point we have "count" elements in the array.
> That means the last element in the array is at arr[count - 1].
> We want to make room for the new element at index, so we move
> all the elements from index to index + 1.
> 
> The first element we should move is arr[count - 1] to arr[count].
> But the code moved arr[count] to arr [count + 1].
> This move is not needed.
> 
> 
>  We currently have count elements in the
>> array, so we cannot normally access the element *at* count. However, we
>> are extending the array right now, therefore we can assign (store) the
>> element at count (and then we'll increment count later). But accessing
>> an element at (count+1) is wrong.
>>
>>> @@ -833,7 +833,6 @@ void *fw_cfg_modify_file(FWCfgState *s, const
>>> char *filename,
>>>       assert(s->files);
>>>         index = be32_to_cpu(s->files->count);
>>> -    assert(index < fw_cfg_file_slots(s));
>>>         for (i = 0; i < index; i++) {
>>>           if (strcmp(filename, s->files->f[i].name) == 0) {
>>> @@ -843,6 +842,9 @@ void *fw_cfg_modify_file(FWCfgState *s, const
>>> char *filename,
>>>               return ptr;
>>>           }
>>>       }
>>> +
>>> +    assert(index < fw_cfg_file_slots(s));
>>> +
>>>       /* add new one */
>>>       fw_cfg_add_file_callback(s, filename, NULL, NULL, NULL, data,
>>> len, true);
>>>       return NULL;
>>>
>>
>> I think I agree with Marc-André here, when I say, replace the assert
>> with a comment instead? (About the fact that fw_cfg_add_file_callback()
>> will assert(), *if* we reach that far.)
> 
> Hmm, what should we add to the comment? "We lost, brace for impact :)"
> 
> My point, if we are going to abort, let's abort as early as we can.
> But if is a consensus, I'll get rid of it.

No, it's going to be another assert, just later. Assume that at this
point we have (index == fw_cfg_file_slots(s)), because the function
didn't find the element to modify, so it decides to add a new one, but
also we do not have room for the new one. So, with the suggested removal
of the assert, we call fw_cfg_add_file_callback().

Then, fw_cfg_add_file_callback() does:

    if (!s->files) {
        dsize = sizeof(uint32_t) + sizeof(FWCfgFile) * fw_cfg_file_slots(s);
        s->files = g_malloc0(dsize);
        fw_cfg_add_bytes(s, FW_CFG_FILE_DIR, s->files, dsize);
    }

    count = be32_to_cpu(s->files->count);
    assert(count < fw_cfg_file_slots(s));

The (!s->files) condition is expected to eval to false (our table is
full, so we do have a table).

And then, the assert() below the "if" will fire.

Am I missing something?

Thanks!
Laszlo