Re: [dpdk-dev] [PATCH] net: fix the way how L4 checksum choice is tested

[Andrew Rybchenko <arybchenko@solarflare.com>]

On 6/24/19 3:01 PM, Ananyev, Konstantin wrote:
>
>> -----Original Message-----
>> From: Ananyev, Konstantin
>> Sent: Monday, June 24, 2019 12:52 PM
>> To: 'Ivan Malov' <ivan.malov@oktetlabs.ru>; Olivier Matz <olivier.matz@6wind.com>
>> Cc: dev@dpdk.org; Kulasek, TomaszX <tomaszx.kulasek@intel.com>; stable@dpdk.org
>> Subject: RE: [dpdk-dev] [PATCH] net: fix the way how L4 checksum choice is tested
>>
>>
>>> The API to prepare checksum offloads mistreats L4
>>> checksum type enum values as self-contained flags.
>>>
>>> Turning these flag checks into enum checks causes
>>> warnings by GCC about possibly uninitialised IPv4
>>> header pointer. The issue was found to show up in
>>> the case of GCC versions 4.8.5 and 5.4.0, however,
>>> it might be the case for a wider variety of other
>>> versions. As GCC version 7.4.0 is not susceptible
>>> to the said false positive assessment, this patch
>>> maintains a compiler barrier for earlier versions.
>>>
>>> Fixes: 4fb7e803eb1a ("ethdev: add Tx preparation")
>>> Cc: Tomasz Kulasek <tomaszx.kulasek@intel.com>
>>> Cc: stable@dpdk.org
>>>
>>> Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
>>> ---
>>>   lib/librte_net/rte_net.h | 16 ++++++++++++++--
>>>   1 file changed, 14 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/lib/librte_net/rte_net.h b/lib/librte_net/rte_net.h
>>> index 7088584..fb09431 100644
>>> --- a/lib/librte_net/rte_net.h
>>> +++ b/lib/librte_net/rte_net.h
>>> @@ -151,7 +151,19 @@ uint32_t rte_net_get_ptype(const struct rte_mbuf *m,
>>>   			ipv4_hdr->hdr_checksum = 0;
>>>   	}
>>>
>> As I remember, saw something similar before...
>> Probably the eaiser way to overcome it, is just to always initialize ipv4_hdr above,
>> something like:
>>
>> +ipv4_hdr = NULL;
>> if (ol_flags & PKT_TX_IPV4) {
>>                  ipv4_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *,
>>                                  inner_l3_offset);
>>
>>                  if (ol_flags & PKT_TX_IP_CKSUM)
>>                          ipv4_hdr->hdr_checksum = 0;
>>   }
> As another possible option  always initialisze both, and then use either one or another,
> depending on flags:
>
> ipv4_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *, inner_l3_offset);
> ipv6_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv6_hdr *, inner_l3_offset);
> ....

We have discussed the solution locally and rejected it since it just kills
the compiler help. It makes it possible to use invalid header and
compiler will not save you.

Suggested solution looks ugly, but at least it does not kill help from
compiler.

>>> -	if ((ol_flags & PKT_TX_UDP_CKSUM) == PKT_TX_UDP_CKSUM) {
>>> +#ifdef GCC_VERSION
>>> +#if GCC_VERSION < 70400
>>> +	/*
>>> +	 * Earlier versions of GCC suspect access to possibly
>>> +	 * uninitialised ipv4_hdr in the code below, although
>>> +	 * the said variable is properly initialised above.
>>> +	 * Use compiler barrier to cope with the problem.
>>> +	 */
>>> +	rte_compiler_barrier();
>>> +#endif
>>> +#endif
>>> +
>>> +	if ((ol_flags & PKT_TX_L4_MASK) == PKT_TX_UDP_CKSUM) {
>>>   		if (ol_flags & PKT_TX_IPV4) {
>>>   			udp_hdr = (struct rte_udp_hdr *)((char *)ipv4_hdr +
>>>   					m->l3_len);
>>> @@ -167,7 +179,7 @@ uint32_t rte_net_get_ptype(const struct rte_mbuf *m,
>>>   			udp_hdr->dgram_cksum = rte_ipv6_phdr_cksum(ipv6_hdr,
>>>   					ol_flags);
>>>   		}
>>> -	} else if ((ol_flags & PKT_TX_TCP_CKSUM) ||
>>> +	} else if ((ol_flags & PKT_TX_L4_MASK) == PKT_TX_TCP_CKSUM ||
>>>   			(ol_flags & PKT_TX_TCP_SEG)) {
>>>   		if (ol_flags & PKT_TX_IPV4) {
>>>   			/* non-TSO tcp or TSO */
>>> --
>>> 1.8.3.1