[PATCH v2] hash: fix pointer alignment

[Radu Nicolau]

On 27-Feb-26 12:37 PM, Marat Khalili wrote:
>> -----Original Message-----
>> From: Radu Nicolau <radu.nicolau at intel.com>
>> Sent: Friday 27 February 2026 09:48
>> To: dev at dpdk.org
>> Cc: Marat Khalili <marat.khalili at huawei.com>; Radu Nicolau <radu.nicolau at intel.com>; stable at dpdk.org;
>> stephen at networkplumber.org; Yipeng Wang <yipeng1.wang at intel.com>; Sameh Gobriel
>> <sameh.gobriel at intel.com>; Bruce Richardson <bruce.richardson at intel.com>; Vladimir Medvedkin
>> <vladimir.medvedkin at intel.com>; Pablo de Lara <pablo.de.lara.guarch at intel.com>; Yerden Zhumabekov
>> <e_zhumabekov at sts.kz>
>> Subject: [PATCH v2] hash: fix pointer alignment
>>
>> rte_hash_crc assumes input pointer address is 8 byte aligned
>> which may not be always the case.
>> This fix aligns the input pointer before proceeding to process it
>> in 8 byte chunks.
>>
>> Fixes: 504a29af13a7 ("hash: fix strict-aliasing for CRC")
>> Cc: stable at dpdk.org
>> Cc: stephen at networkplumber.org
>>
>> Signed-off-by: Radu Nicolau <radu.nicolau at intel.com>
>> ---
>> v2: reverse the order of alignment adjustment calls
>>
>>   lib/hash/rte_hash_crc.h | 18 ++++++++++++++++++
>>   1 file changed, 18 insertions(+)
>>
>> diff --git a/lib/hash/rte_hash_crc.h b/lib/hash/rte_hash_crc.h
>> index fa07c97685..d61420868a 100644
>> --- a/lib/hash/rte_hash_crc.h
>> +++ b/lib/hash/rte_hash_crc.h
>> @@ -127,6 +127,24 @@ rte_hash_crc(const void *data, uint32_t data_len, uint32_t init_val)
>>   	unsigned i;
>>   	uintptr_t pd = (uintptr_t) data;
>>
>> +	/* align input to 8 byte boundary if needed */
>> +	if ((pd & 0x7) && data_len >= 8) {
>> +		uintptr_t unaligned_bytes = 8 - (pd & 0x7);
>> +		data_len -= unaligned_bytes;
>> +		if (unaligned_bytes & 0x1) {
>> +			init_val = rte_hash_crc_1byte(*(const uint8_t *)pd, init_val);
>> +			pd += 1;
>> +		}
>> +		if (unaligned_bytes & 0x2) {
>> +			init_val = rte_hash_crc_2byte(*(const uint16_t *)pd, init_val);
>> +			pd += 2;
>> +		}
>> +		if (unaligned_bytes & 0x4) {
>> +			init_val = rte_hash_crc_4byte(*(const uint32_t *)pd, init_val);
>> +			pd += 4;
>> +		}
>> +	}
>> +
>>   	for (i = 0; i < data_len / 8; i++) {
>>   		init_val = rte_hash_crc_8byte(*(const uint64_t *)pd, init_val);
>>   		pd += 8;
>> --
>> 2.52.0
>>
> Surprisingly, we do not seem to have any tests calling rte_hash_crc with misaligned data. I tried to tweak existing ones in test_hash.c and test_hash_functions.c, and found out that rte_hash_crc still fails (with ubsan) for sizes less than 8 in one of the last 3 if's, and also that jhash has the same problems. Still this commit is a step in the right direction IMO, so:
>
> Acked-by: Marat Khalili <marat.khalili at huawei.com>

Thanks for the ack!

The issue still persist with less that 8 bytes unaligned input because 
the 3 ifs after the loop assume the input is aligned, which is the case 
only if the loop was taken. I will fix the patch to address this case in 
the added section - you were right with the first comment.