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

[Dmitry Osipenko <digetx@gmail.com>]

11.06.2019 10:38, Bitan Biswas пишет:
> 
> 
> 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.

When word_to_transfer is more than rx_fifo_avail, then the rx_fifo_avail
becomes zero and hence the nibbles won't be copied. Please take a closer
look, the current code is correct, but the buf_remaining > 3 is unneeded
because it can't ever happen.

The code is structured the way that it's difficult to follow, apparently
the person who added the BUG_ON check in the first place couldn't follow
it either. Maybe it's worth to invest some more effort into refactoring
at least that part of the code. At minimum a clarifying comments would
be helpful.

[snip]

>>>> 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.

There are two possible error conditions:

1) Underflow: the XFER_COMPLETE happens before message is fully sent.

2) Overflow: message is fully sent, but there is no XFER_COMPLETE and
then hardware asks to transfer more.

We are addressing the second case here, while you seems are confusing it
with the first case.