Re: [PATCH V4 6/6] i2c: tegra: remove BUG, BUG_ON

[Bitan Biswas <bbiswas@nvidia.com>]

On 6/10/19 2:00 PM, Dmitry Osipenko wrote:
> 10.06.2019 22:41, Bitan Biswas пишет:
>>
>>
>> On 6/10/19 11:12 AM, Dmitry Osipenko wrote:
>>> 10.06.2019 20:08, Bitan Biswas пишет:
>>>> Remove redundant BUG_ON calls or replace with WARN_ON_ONCE
>>>> as needed. Remove BUG() and make Rx and Tx case handling
>>>> similar.
>>>>
>>>> Signed-off-by: Bitan Biswas <bbiswas@nvidia.com>
>>>> ---
>>>>    drivers/i2c/busses/i2c-tegra.c | 11 ++++++-----
>>>>    1 file changed, 6 insertions(+), 5 deletions(-)
>>>
>>> Looks that this is still not correct. What if it transfer-complete flag
>>> is set and buffer is full on RX? In this case the transfer will succeed
>>> while it was a failure.
>>>
>>>> diff --git a/drivers/i2c/busses/i2c-tegra.c
>>>> b/drivers/i2c/busses/i2c-tegra.c
>>>> index 4dfb4c1..30619d6 100644
>>>> --- a/drivers/i2c/busses/i2c-tegra.c
>>>> +++ b/drivers/i2c/busses/i2c-tegra.c
>>>> @@ -515,7 +515,6 @@ static int tegra_i2c_empty_rx_fifo(struct
>>>> tegra_i2c_dev *i2c_dev)
>>>>         * prevent overwriting past the end of buf
>>>>         */
>>>>        if (rx_fifo_avail > 0 && buf_remaining > 0) {
>>>> -        BUG_ON(buf_remaining > 3);
>>>
>>> Actually error should be returned here since out-of-bounds memory
>>> accesses must be avoided, hence:
>>>
>>>      if (WARN_ON_ONCE(buf_remaining > 3))
>>>          return -EINVAL;
>> buf_remaining will be less than equal to 3 because of the expression
>> earlier
>> https://elixir.bootlin.com/linux/v5.2-rc4/source/drivers/i2c/busses/i2c-tegra.c#L520
>>
> 
> Ah yes, indeed!
> 
I see that I am wrong and buf_remaining > 3 needs to be prevented at

https://elixir.bootlin.com/linux/v5.2-rc4/source/drivers/i2c/busses/i2c-tegra.c#L528

because of word_to_transfer is limited to rx_fifo_avail:

https://elixir.bootlin.com/linux/v5.2-rc4/source/drivers/i2c/busses/i2c-tegra.c#L515

I shall add the check for less than 3 in both RX and TX cases in a 
separate patch in this series.

>>
>>>
>>>>            val = i2c_readl(i2c_dev, I2C_RX_FIFO);
>>>>            val = cpu_to_le32(val);
>>>>            memcpy(buf, &val, buf_remaining);
>>>> @@ -523,7 +522,6 @@ static int tegra_i2c_empty_rx_fifo(struct
>>>> tegra_i2c_dev *i2c_dev)
>>>>            rx_fifo_avail--;
>>>>        }
>>>>    -    BUG_ON(rx_fifo_avail > 0 && buf_remaining > 0);
>>>
>>> Better not to ignore this as well:
>>>
>>>      if (WARN_ON_ONCE(rx_fifo_avail > 0 &&
>>>               buf_remaining > 0))
>>>          return -EINVAL;
>>>
>> Please check below line.
>> https://elixir.bootlin.com/linux/v5.2-rc4/source/drivers/i2c/busses/i2c-tegra.c#L532
>>
>>
>> It ensures that buf_remaining will be 0 and we never hit the BUG_ON as
>> follows:
> 
> [1] Okay, but it doesn't ensure about rx_fifo_avail. So it could be:
> 
> 	if (WARN_ON_ONCE(rx_fifo_avail))
> 		return -EINVAL;
I shall add the WARN_ON_ONCE.


> 
>>>> -    BUG_ON(rx_fifo_avail > 0 && buf_remaining > 0);
>>
>>>>        i2c_dev->msg_buf_remaining = buf_remaining;
>>>>        i2c_dev->msg_buf = buf;
>>>>    @@ -581,7 +579,6 @@ static int tegra_i2c_fill_tx_fifo(struct
>>>> tegra_i2c_dev *i2c_dev)
>>>>         * boundary and fault.
>>>>         */
>>>>        if (tx_fifo_avail > 0 && buf_remaining > 0) {
>>>> -        BUG_ON(buf_remaining > 3);
>>>
>>> And here, cause this will corrupt stack:
>>>
>>>          if (WARN_ON_ONCE(buf_remaining > 3))
>>>              return -EINVAL;
>>>
>> Please check the line
>> https://elixir.bootlin.com/linux/v5.2-rc4/source/drivers/i2c/busses/i2c-tegra.c#L576
>>
>>
>> It ensures buf_remaining will be less or equal to 3.
> 
> Okay, agree here.
> 
>>>>            memcpy(&val, buf, buf_remaining);
>>>>            val = le32_to_cpu(val);
>>>>    @@ -850,7 +847,8 @@ static irqreturn_t tegra_i2c_isr(int irq, void
>>>> *dev_id)
>>>>                if (i2c_dev->msg_buf_remaining)
>>>>                    tegra_i2c_empty_rx_fifo(i2c_dev);
>>>>                else
>>>> -                BUG();
>>>> +                tegra_i2c_mask_irq(i2c_dev,
>>>> +                           I2C_INT_RX_FIFO_DATA_REQ);
>>>
>>> Then here:
>>>
>>>      if (WARN_ON_ONCE(!i2c_dev->msg_buf_remaining) ||
>>>          tegra_i2c_empty_rx_fifo(i2c_dev)) {
>>>          i2c_dev->msg_err |= I2C_ERR_UNKNOWN_INTERRUPT;
>>>          goto err;
>>>      }
>>>
>> Can you please elaborate why the condition needs to be as follows
>> instead of " if (WARN_ON_ONCE(i2c_dev->msg_buf_remaining)) " ?
>>
>>>       if (WARN_ON_ONCE(!i2c_dev->msg_buf_remaining) ||
>>>           tegra_i2c_empty_rx_fifo(i2c_dev)) {
> 
> Because this is a "receive" transfer and hence it is a error condition
> if the data-message was already fully received and then there is another
> request from hardware to receive more data. So
> "!i2c_dev->msg_buf_remaining" is the error condition here because there
> is no more space in the buffer.
> 
> Looking at this again, seems checking for "if
> (WARN_ON_ONCE(rx_fifo_avail))" in the above hunk [1] will be already
> enough since a not fully drained RX FIFO means that there is no enough
> space in the buffer. Then it could be:
> 
>          if (tegra_i2c_empty_rx_fifo(i2c_dev)) {
>                  i2c_dev->msg_err |= I2C_ERR_UNKNOWN_INTERRUPT;
>                  goto err;
> 	}
> 
In the case "if (status & I2C_INT_PACKET_XFER_COMPLETE) {" , we do not 
have any tegra_i2c_empty_rx_fifo call today. In this current driver I do 
not see any code that checks for the buffer space and prevents RX FIFO 
from being drained. The transfer complete when seen must have already 
consumed all bytes of msg_buf_remaining in the call at the line

https://elixir.bootlin.com/linux/v5.2-rc4/source/drivers/i2c/busses/i2c-tegra.c#L860

So we can put the "WARN_ON_ONCE(i2c_dev->msg_buf_remaining) with msg_err 
assignment and goto err" to confirm if some corner case is not handled.

Planning to share updated patch.

-Thanks,
  Bitan