Re: [PATCH 2/3] TPM: Close data_pending and data_buffer races

[Tim Gardner <tim.gardner@canonical.com>]

On 12/22/2011 10:42 AM, Rajiv Andrade wrote:
> On 20-12-2011 17:39, Tim Gardner wrote:
>> On 12/20/2011 09:38 AM, Rajiv Andrade wrote:
>>> On 06/12/11 16:29, Tim Gardner wrote:
>>>> There is a race betwen tpm_read() and tpm_write where both
>>>> chip->data_pending
>>>> and chip->data_buffer can be changed by tpm_write() when tpm_read()
>>>> clears chip->data_pending, but before tpm_read() grabs the mutex.
>>>>
>>>> Protect changes to chip->data_pending and chip->data_buffer by
>>>> expanding
>>>> the scope of chip->buffer_mutex.
>>>>
>>>> Reported-by: Seth Forshee<seth.forshee@canonical.com>
>>>> Cc: Debora Velarde<debora@linux.vnet.ibm.com>
>>>> Cc: Rajiv Andrade<srajiv@linux.vnet.ibm.com>
>>>> Cc: Marcel Selhorst<m.selhorst@sirrix.com>
>>>> Cc: tpmdd-devel@lists.sourceforge.net
>>>> Cc: stable@vger.kernel.org
>>>> Signed-off-by: Tim Gardner<tim.gardner@canonical.com>
>>>> ---
>>>> drivers/char/tpm/tpm.c | 17 +++++++++--------
>>>> 1 files changed, 9 insertions(+), 8 deletions(-)
>>>>
>>>> diff --git a/drivers/char/tpm/tpm.c b/drivers/char/tpm/tpm.c
>>>> index b366b34..70bf9e5 100644
>>>> --- a/drivers/char/tpm/tpm.c
>>>> +++ b/drivers/char/tpm/tpm.c
>>>> @@ -1074,12 +1074,15 @@ ssize_t tpm_write(struct file *file, const
>>>> char __user *buf,
>>>> struct tpm_chip *chip = file->private_data;
>>>> size_t in_size = size, out_size;
>>>>
>>>> + mutex_lock(&chip->buffer_mutex);
>>>> +
>>>> /* cannot perform a write until the read has cleared
>>>> either via tpm_read or a user_read_timer timeout */
>>>> - while (atomic_read(&chip->data_pending) != 0)
>>>> + while (atomic_read(&chip->data_pending) != 0) {
>>>> + mutex_unlock(&chip->buffer_mutex);
>>>> msleep(TPM_TIMEOUT);
>>>> -
>>>> - mutex_lock(&chip->buffer_mutex);
>>>> + mutex_lock(&chip->buffer_mutex);
>>>> + }
>>>>
>>>> if (in_size> TPM_BUFSIZE)
>>>> in_size = TPM_BUFSIZE;
>>>> @@ -1112,22 +1115,20 @@ ssize_t tpm_read(struct file *file, char
>>>> __user *buf,
>>>>
>>>> del_singleshot_timer_sync(&chip->user_read_timer);
>>>> flush_work_sync(&chip->work);
>>>> - ret_size = atomic_read(&chip->data_pending);
>>>> - atomic_set(&chip->data_pending, 0);
>>>> + mutex_lock(&chip->buffer_mutex);
>>>> + ret_size = atomic_xchg(&chip->data_pending, 0);
>>>> if (ret_size> 0) { /* relay data */
>>>> ssize_t orig_ret_size = ret_size;
>>>> if (size< ret_size)
>>>> ret_size = size;
>>>>
>>>> - mutex_lock(&chip->buffer_mutex);
>>>> rc = copy_to_user(buf, chip->data_buffer, ret_size);
>>>> memset(chip->data_buffer, 0, orig_ret_size);
>>>> if (rc)
>>>> ret_size = -EFAULT;
>>>
>>> What about just moving atomic_set(&chip->data_pending, 0); to here?
>>> If I'm not missing anything, this would be cleaner.
>>>
>>> Rajiv
>>
>> I'm not sure I agree. Moving just that statement doesn't close the
>> race. Perhaps you could send me your version of this patch so that its
>> clear what you are suggesting.
>>
>> rtg
> diff --git a/drivers/char/tpm/tpm.c b/drivers/char/tpm/tpm.c
> index 6a8771f..6a37212b 100644
> --- a/drivers/char/tpm/tpm.c
> +++ b/drivers/char/tpm/tpm.c
> @@ -1210,7 +1210,6 @@ ssize_t tpm_read(struct file *file, char __user *buf,
> del_singleshot_timer_sync(&chip->user_read_timer);
> flush_work_sync(&chip->work);
> ret_size = atomic_read(&chip->data_pending);
> - atomic_set(&chip->data_pending, 0);
> if (ret_size> 0) { /* relay data */
> if (size< ret_size)
> ret_size = size;
> @@ -1223,6 +1222,7 @@ ssize_t tpm_read(struct file *file, char __user *buf,
> mutex_unlock(&chip->buffer_mutex);
> } + atomic_set(&chip->data_pending, 0);
> return ret_size;
> }
>
> If we reset chip->data_pending after the buffer was copied to userspace,
> it's guaranteed that tpm_write() won't touch such buffer before tpm_read()
> handles it, given it polls chip->data_pending first.
>

NAK - this patch makes it worse (if I'm reading your email client 
garbled patch correctly). Now it races with tpm_write() as well as 
timeout_work(). You cannot futz with chip->data_pending outside of the 
exclusion zones. Consider what will happen if a user process just loops 
doing reads. chip->data_pending gets cleared every time tpm_read() is 
called, regardless of what tpm_write() or timeout_work() are doing at 
the time.

tpm_read() / tpm_write() is a simple producer consumer model. Just use 
mutexes in an uncomplicated way. There is no need for data_pending to be 
atomic_t.

rtg
-- 
Tim Gardner tim.gardner@canonical.com