[PATCH RESEND 2] libata: if T_LENGTH is zero, dma direction should be DMA_NONE

[PATCH RESEND 2] libata: if T_LENGTH is zero, dma direction should be DMA_NONE

[George Kennedy]

Avoid data corruption by rejecting pass-through commands where
T_LENGTH is zero (No data is transferred) and the dma direction
is not DMA_NONE.

Cc:<stable@vger.kernel.org>  # 5.4.y
Reported-by: syzkaller<syzkaller@googlegroups.com>
Signed-off-by: George Kennedy<george.kennedy@oracle.com>
---
  drivers/ata/libata-scsi.c | 6 ++++++
  1 file changed, 6 insertions(+)

diff --git a/drivers/ata/libata-scsi.c b/drivers/ata/libata-scsi.c
index 1b84d55..d428392 100644
--- a/drivers/ata/libata-scsi.c
+++ b/drivers/ata/libata-scsi.c
@@ -2859,6 +2859,12 @@ static unsigned int ata_scsi_pass_thru(struct ata_queued_cmd *qc)
  		goto invalid_fld;
  	}
  
+	/* if T_LENGTH is zero (No data is transferred), then dir should be DMA_NONE */
+	if ((cdb[2 + cdb_offset] & 3) == 0 && scmd->sc_data_direction != DMA_NONE) {
+		fp = 2 + cdb_offset;
+		goto invalid_fld;
+	}
+
  	if (ata_is_ncq(tf->protocol) && (cdb[2 + cdb_offset] & 0x3) == 0)
  		tf->protocol = ATA_PROT_NCQ_NODATA;
  
-- 1.8.3.1

Re: [PATCH RESEND 2] libata: if T_LENGTH is zero, dma direction should be DMA_NONE

[Damien Le Moal]

On 2021/12/14 0:33, George Kennedy wrote:
> Avoid data corruption by rejecting pass-through commands where
> T_LENGTH is zero (No data is transferred) and the dma direction
> is not DMA_NONE.
> 
> Cc:<stable@vger.kernel.org>  # 5.4.y

5.4 only ? What about other LTS versions ? They do not have that bug ?

> Reported-by: syzkaller<syzkaller@googlegroups.com>
> Signed-off-by: George Kennedy<george.kennedy@oracle.com>
> ---
>   drivers/ata/libata-scsi.c | 6 ++++++
>   1 file changed, 6 insertions(+)
> 
> diff --git a/drivers/ata/libata-scsi.c b/drivers/ata/libata-scsi.c
> index 1b84d55..d428392 100644
> --- a/drivers/ata/libata-scsi.c
> +++ b/drivers/ata/libata-scsi.c
> @@ -2859,6 +2859,12 @@ static unsigned int ata_scsi_pass_thru(struct ata_queued_cmd *qc)
>   		goto invalid_fld;
>   	}
>   
> +	/* if T_LENGTH is zero (No data is transferred), then dir should be DMA_NONE */
> +	if ((cdb[2 + cdb_offset] & 3) == 0 && scmd->sc_data_direction != DMA_NONE) {
> +		fp = 2 + cdb_offset;
> +		goto invalid_fld;
> +	}
> +
>   	if (ata_is_ncq(tf->protocol) && (cdb[2 + cdb_offset] & 0x3) == 0)
>   		tf->protocol = ATA_PROT_NCQ_NODATA;
>   
> -- 1.8.3.1
> 


-- 
Damien Le Moal
Western Digital Research

Re: [PATCH RESEND 2] libata: if T_LENGTH is zero, dma direction should be DMA_NONE

[George Kennedy]

On 12/13/2021 4:37 PM, Damien Le Moal wrote:
> On 2021/12/14 0:33, George Kennedy wrote:
>> Avoid data corruption by rejecting pass-through commands where
>> T_LENGTH is zero (No data is transferred) and the dma direction
>> is not DMA_NONE.
>>
>> Cc:<stable@vger.kernel.org>  # 5.4.y
> 5.4 only ? What about other LTS versions ? They do not have that bug ?
Our distros are based on 5.4.y, 5.15.y and 4.14.y, so the fix should go 
there,
but it looks like the fix is also needed in: 5.14.y, 5.13.y, 5.10.y, and 
4.19.y.

Thank you,
George
>
>> Reported-by: syzkaller<syzkaller@googlegroups.com>
>> Signed-off-by: George Kennedy<george.kennedy@oracle.com>
>> ---
>>    drivers/ata/libata-scsi.c | 6 ++++++
>>    1 file changed, 6 insertions(+)
>>
>> diff --git a/drivers/ata/libata-scsi.c b/drivers/ata/libata-scsi.c
>> index 1b84d55..d428392 100644
>> --- a/drivers/ata/libata-scsi.c
>> +++ b/drivers/ata/libata-scsi.c
>> @@ -2859,6 +2859,12 @@ static unsigned int ata_scsi_pass_thru(struct ata_queued_cmd *qc)
>>    		goto invalid_fld;
>>    	}
>>    
>> +	/* if T_LENGTH is zero (No data is transferred), then dir should be DMA_NONE */
>> +	if ((cdb[2 + cdb_offset] & 3) == 0 && scmd->sc_data_direction != DMA_NONE) {
>> +		fp = 2 + cdb_offset;
>> +		goto invalid_fld;
>> +	}
>> +
>>    	if (ata_is_ncq(tf->protocol) && (cdb[2 + cdb_offset] & 0x3) == 0)
>>    		tf->protocol = ATA_PROT_NCQ_NODATA;
>>    
>> -- 1.8.3.1
>>
>

Re: [PATCH RESEND 2] libata: if T_LENGTH is zero, dma direction should be DMA_NONE

[Damien Le Moal]

On 2021/12/14 7:15, George Kennedy wrote:
> 
> 
> On 12/13/2021 4:37 PM, Damien Le Moal wrote:
>> On 2021/12/14 0:33, George Kennedy wrote:
>>> Avoid data corruption by rejecting pass-through commands where
>>> T_LENGTH is zero (No data is transferred) and the dma direction
>>> is not DMA_NONE.
>>>
>>> Cc:<stable@vger.kernel.org>  # 5.4.y
>> 5.4 only ? What about other LTS versions ? They do not have that bug ?
> Our distros are based on 5.4.y, 5.15.y and 4.14.y, so the fix should go 
> there,
> but it looks like the fix is also needed in: 5.14.y, 5.13.y, 5.10.y, and 
> 4.19.y.

Then please resend with:

Cc: stable@vger.kernel.org

No version specified.

> 
> Thank you,
> George
>>
>>> Reported-by: syzkaller<syzkaller@googlegroups.com>
>>> Signed-off-by: George Kennedy<george.kennedy@oracle.com>
>>> ---
>>>    drivers/ata/libata-scsi.c | 6 ++++++
>>>    1 file changed, 6 insertions(+)
>>>
>>> diff --git a/drivers/ata/libata-scsi.c b/drivers/ata/libata-scsi.c
>>> index 1b84d55..d428392 100644
>>> --- a/drivers/ata/libata-scsi.c
>>> +++ b/drivers/ata/libata-scsi.c
>>> @@ -2859,6 +2859,12 @@ static unsigned int ata_scsi_pass_thru(struct ata_queued_cmd *qc)
>>>    		goto invalid_fld;
>>>    	}
>>>    
>>> +	/* if T_LENGTH is zero (No data is transferred), then dir should be DMA_NONE */
>>> +	if ((cdb[2 + cdb_offset] & 3) == 0 && scmd->sc_data_direction != DMA_NONE) {
>>> +		fp = 2 + cdb_offset;
>>> +		goto invalid_fld;
>>> +	}
>>> +
>>>    	if (ata_is_ncq(tf->protocol) && (cdb[2 + cdb_offset] & 0x3) == 0)

cdb[2 + cdb_offset] & 3) is used again for the ncq test, so you can group things:

	if (!(cdb[2 + cdb_offset] & 0x3)) {
		/*
		 * When T_LENGTH is zero (no data is transferred),
		 * then dir should be DMA_NONE.
		 */
		if (scmd->sc_data_direction != DMA_NONE) {
			fp = 2 + cdb_offset;
			goto invalid_fld;
		}

		if (ata_is_ncq(tf->protocol))
			tf->protocol = ATA_PROT_NCQ_NODATA;
	}

>>>    		tf->protocol = ATA_PROT_NCQ_NODATA;
>>>    
>>> -- 1.8.3.1
>>>
>>
> 


-- 
Damien Le Moal
Western Digital Research