Re: [dpdk-dev] [PATCH v9] ethdev: add sanity checks in control APIs

[Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>]

On 4/21/21 4:40 PM, Ferruh Yigit wrote:
> On 4/21/2021 2:19 PM, Ferruh Yigit wrote:
>> On 4/21/2021 12:28 PM, Andrew Rybchenko wrote:
>>> On 4/21/21 5:36 AM, Ferruh Yigit wrote:
>>>> From: "Min Hu (Connor)" <humin29@huawei.com>
>>>>
>>>> This patch adds more sanity checks in control path APIs.
>>>>
>>>> Fixes: 214ed1acd125 ("ethdev: add iterator to match devargs input")
>>>> Fixes: 3d98f921fbe9 ("ethdev: unify prefix for static functions and
>>>> variables")
>>>> Fixes: 0366137722a0 ("ethdev: check for invalid device name")
>>>> Fixes: d948f596fee2 ("ethdev: fix port data mismatched in multiple
>>>> process model")
>>>> Fixes: 5b7ba31148a8 ("ethdev: add port ownership")
>>>> Fixes: f8244c6399d9 ("ethdev: increase port id range")
>>>> Cc: stable@dpdk.org
>>>>
>>>> Signed-off-by: Min Hu (Connor) <humin29@huawei.com>
>>>> Signed-off-by: Ferruh Yigit <ferruh.yigit@intel.com>
>>>> Reviewed-by: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
>>>> Acked-by: Kevin Traynor <ktraynor@redhat.com>
>>>
>>> Few nits below.
>>> Other than that I confirm my "Reviewed-by".
>>>
>>> The patch is really long. It would be better to split it into
>>> few:
>>>   - relocate dev assignment
>>>   - empty lines mangling (when it is unrelated to previous item)
>>>   - ops check before usage (combined with related style checks)
>>>   - error logs refinement
>>>
>>> However, since the patch is already reviewed this way, may
>>> be it is better to push as is after review notes processing.
>>>
>>>> @@ -817,7 +859,12 @@ rte_eth_dev_get_port_by_name(const char *name,
>>>> uint16_t *port_id)
>>>>       uint16_t pid;
>>>>       if (name == NULL) {
>>>> -        RTE_ETHDEV_LOG(ERR, "Null pointer is specified\n");
>>>> +        RTE_ETHDEV_LOG(ERR, "Cannot get port ID from NULL name");
>>>> +        return -EINVAL;
>>>> +    }
>>>> +
>>>> +    if (port_id == NULL) {
>>>> +        RTE_ETHDEV_LOG(ERR, "Cannot get port ID to NULL\n");
>>>
>>> Since name is already checked above, I think it would be useful
>>> to log 'name' here to provide context.
>>>
>>>>           return -EINVAL;
>>>>       }
>>>
>>> [snip]
>>>
>>>> @@ -3256,6 +3370,20 @@ rte_eth_dev_fw_version_get(uint16_t port_id,
>>>> char *fw_version, size_t fw_size)
>>>>       RTE_ETH_VALID_PORTID_OR_ERR_RET(port_id, -ENODEV);
>>>>       dev = &rte_eth_devices[port_id];
>>>> +    if (fw_version == NULL) {
>>>> +        RTE_ETHDEV_LOG(ERR,
>>>> +            "Cannot get ethdev port %u FW version to NULL\n",
>>>> +            port_id);
>>>> +        return -EINVAL;
>>>> +    }
>>>> +
>>>> +    if (fw_size == 0) {
>>>> +        RTE_ETHDEV_LOG(ERR,
>>>> +            "Cannot get ethdev port %u FW version to buffer with
>>>> zero size\n",
>>>> +            port_id);
>>>> +        return -EINVAL;
>>>> +    }
>>>> +
>>>
>>> The only error condition is NULL fw_version with positive
>>> fw_size. Othwerwise, it could be just a call to get required
>>> size of buffer for FW version.
>>>
>>
>> Right, above is wrong.
>>
>> Agree that "fw_version == NULL && fw_size > 0" is error condition,
>> but is it clear if how this API should behave on
>> "fw_version == NULL && fw_size == 0"?
>>
>> Like sfc has following,
>> if ((fw_version == NULL) || (fw_size == 0))
>>      return -EINVAL;
> 
> axgbe, qede also returns error when fw_version is NULL, independent from
> fw_size.
> 
> But I think taking "fw_version == NULL && fw_size > 0" as only error
> condition is reasonable, although some PMDs will be behaving wrong.
> I can send a separate patch to unify the behavior.

Many thanks, please, let me know if you need help with net/sfc.