[PATCH v1 08/13] test/bbdev: extend support for large TB

[Maxime Coquelin]

On 2/13/23 21:20, Vargas, Hernan wrote:
> Hi Maxime,
> 
>> -----Original Message-----
>> From: Maxime Coquelin <maxime.coquelin at redhat.com>
>> Sent: Tuesday, January 31, 2023 5:29 AM
>> To: Vargas, Hernan <hernan.vargas at intel.com>; dev at dpdk.org;
>> gakhil at marvell.com; Rix, Tom <trix at redhat.com>
>> Cc: Chautru, Nicolas <nicolas.chautru at intel.com>; Zhang, Qi Z
>> <qi.z.zhang at intel.com>
>> Subject: Re: [PATCH v1 08/13] test/bbdev: extend support for large TB
>>
>>
>>
>> On 1/17/23 17:50, Hernan Vargas wrote:
>>> Add support for large TB when it cannot fit into a true mbuf.
>>>
>>> Signed-off-by: Hernan Vargas <hernan.vargas at intel.com>
>>> ---
>>>    app/test-bbdev/test_bbdev_perf.c | 21 ++++++++++++---------
>>>    1 file changed, 12 insertions(+), 9 deletions(-)
>>>
>>> diff --git a/app/test-bbdev/test_bbdev_perf.c
>>> b/app/test-bbdev/test_bbdev_perf.c
>>> index 69b86cdeb1..fdf7a28ba2 100644
>>> --- a/app/test-bbdev/test_bbdev_perf.c
>>> +++ b/app/test-bbdev/test_bbdev_perf.c
>>> @@ -1072,8 +1072,6 @@ init_op_data_objs(struct rte_bbdev_op_data
>> *bufs,
>>>    			 * Special case when DPDK mbuf cannot handle
>>>    			 * the required input size
>>>    			 */
>>> -			printf("Warning: Larger input size than DPDK mbuf
>> %d\n",
>>> -					seg->length);
>>>    			large_input = true;
>>>    		}
>>>    		bufs[i].data = m_head;
>>> @@ -2030,6 +2028,7 @@ validate_op_chain(struct rte_bbdev_op_data
>> *op,
>>>    	struct rte_mbuf *m = op->data;
>>>    	uint8_t nb_dst_segments = orig_op->nb_segments;
>>>    	uint32_t total_data_size = 0;
>>> +	bool ignore_mbuf = false; /* ignore mbuf limitations */
>>>
>>>    	TEST_ASSERT(nb_dst_segments == m->nb_segs,
>>>    			"Number of segments differ in original (%u) and filled
>> (%u) op",
>>> @@ -2042,21 +2041,25 @@ validate_op_chain(struct rte_bbdev_op_data
>> *op,
>>>    		uint16_t data_len = rte_pktmbuf_data_len(m) - offset;
>>>    		total_data_size += orig_op->segments[i].length;
>>>
>>> -		TEST_ASSERT(orig_op->segments[i].length == data_len,
>>> -				"Length of segment differ in original (%u) and
>> filled (%u) op",
>>> -				orig_op->segments[i].length, data_len);
>>> +		if (orig_op->segments[i].length >
>> RTE_BBDEV_LDPC_E_MAX_MBUF)
>>> +			ignore_mbuf = true;
>>> +		if (!ignore_mbuf)
>>> +			TEST_ASSERT(orig_op->segments[i].length ==
>> data_len,
>>> +					"Length of segment differ in original
>> (%u) and filled (%u) op",
>>> +					orig_op->segments[i].length,
>> data_len);
>>>    		TEST_ASSERT_BUFFERS_ARE_EQUAL(orig_op-
>>> segments[i].addr,
>>>    				rte_pktmbuf_mtod_offset(m, uint32_t *,
>> offset),
>>> -				data_len,
>>> +				orig_op->segments[i].length,
>>
>> Isn't it ending up in performing out of bounds access in the mbuf?
> 
> No, in the case when ignore_mbuf is set to true, we use a "fake" mbuf allocated in memory with rte_malloc.
> The size allocated is segments[i].length.

Ok.

Thanks,
Maxime

> Thanks
> 
>>>    				"Output buffers (CB=%u) are not equal", i);
>>>    		m = m->next;
>>>    	}
>>>
>>>    	/* Validate total mbuf pkt length */
>>>    	uint32_t pkt_len = rte_pktmbuf_pkt_len(op->data) - op->offset;
>>> -	TEST_ASSERT(total_data_size == pkt_len,
>>> -			"Length of data differ in original (%u) and filled (%u)
>> op",
>>> -			total_data_size, pkt_len);
>>> +	if (!ignore_mbuf)
>>> +		TEST_ASSERT(total_data_size == pkt_len,
>>> +				"Length of data differ in original (%u) and
>> filled (%u) op",
>>> +				total_data_size, pkt_len);
>>>
>>>    	return TEST_SUCCESS;
>>>    }
>