[PATCH v3] ethdev: fix strict aliasing lead to link cannot be up

[fengchengwen]

Hi Morten,

On 2024/4/11 20:44, Morten Brørup wrote:
>> From: Chengwen Feng [mailto:fengchengwen at huawei.com]
>> Sent: Thursday, 11 April 2024 14.04
>>
>> Fix a problem introduced by a compiler upgrade (from gcc10 to gcc12.3),
>> which will lead the hns3 NIC can't link up. The root cause is strict
>> aliasing violation in rte_eth_linkstatus_set() with hns3 driver, see
>> [1] for more details.
>>
>> This commit use union to avoid such aliasing violation.
>>
>> [1] Strict aliasing problem with rte_eth_linkstatus_set()
>>     https://marc.info/?l=dpdk-dev&m=171274148514777&w=3
>>
>> Cc: stable at dpdk.org
>>
>> Signed-off-by: Chengwen Feng <fengchengwen at huawei.com>
>> Signed-off-by: Dengdui Huang <huangdengdui at huawei.com>
>>
>> ---
>> v3: fix checkpatch warning "missing --in-reply-to".
>> v2: add RTE_ATOMIC(uint64_t) wrap which address Morten's comment.
>>
>> ---
>>  lib/ethdev/ethdev_driver.h | 23 +++++++----------------
>>  lib/ethdev/rte_ethdev.h    | 16 ++++++++++------
>>  2 files changed, 17 insertions(+), 22 deletions(-)
>>
>> diff --git a/lib/ethdev/ethdev_driver.h b/lib/ethdev/ethdev_driver.h
>> index 0dbf2dd6a2..9d831d5c84 100644
>> --- a/lib/ethdev/ethdev_driver.h
>> +++ b/lib/ethdev/ethdev_driver.h
>> @@ -1674,18 +1674,13 @@ static inline int
>>  rte_eth_linkstatus_set(struct rte_eth_dev *dev,
>>  		       const struct rte_eth_link *new_link)
>>  {
>> -	RTE_ATOMIC(uint64_t) *dev_link = (uint64_t __rte_atomic *)&(dev-
>>> data->dev_link);
>> -	union {
>> -		uint64_t val64;
>> -		struct rte_eth_link link;
>> -	} orig;
>> -
>> -	RTE_BUILD_BUG_ON(sizeof(*new_link) != sizeof(uint64_t));
>> +	struct rte_eth_link old_link;
>>
>> -	orig.val64 = rte_atomic_exchange_explicit(dev_link, *(const
>> uint64_t *)new_link,
>> -					rte_memory_order_seq_cst);
>> +	old_link.val64 = rte_atomic_exchange_explicit(&dev->data-
>>> dev_link.val64,
> 
> You are right; old_link has local scope and is on the stack, so atomic store is not required.
> 
> And since rte_eth_linkstatus_set() is an internal function called from the driver only, it is probably safe to assume that *new_link is on the caller's stack and doesn't change while being accessed by this function.
> I guess that new_link is passed by reference for performance and future-proofing reasons; it could have been passed by value instead. If it was passed by value, atomic access would certainly not be required.
> In other words: You are right here too; new_link does not require atomic load.
> 
>> +						      new_link->val64,
>> +						      rte_memory_order_seq_cst);
>>
>> -	return (orig.link.link_status == new_link->link_status) ? -1 : 0;
>> +	return (old_link.link_status == new_link->link_status) ? -1 : 0;
>>  }
>>
>>  /**
>> @@ -1701,12 +1696,8 @@ static inline void
>>  rte_eth_linkstatus_get(const struct rte_eth_dev *dev,
>>  		       struct rte_eth_link *link)
>>  {
>> -	RTE_ATOMIC(uint64_t) *src = (uint64_t __rte_atomic *)&(dev->data-
>>> dev_link);
>> -	uint64_t *dst = (uint64_t *)link;
>> -
>> -	RTE_BUILD_BUG_ON(sizeof(*link) != sizeof(uint64_t));
>> -
>> -	*dst = rte_atomic_load_explicit(src, rte_memory_order_seq_cst);
>> +	link->val64 = rte_atomic_load_explicit(&dev->data->dev_link.val64,
>> +					       rte_memory_order_seq_cst);
> 
> It is not safe to assume that the link pointer points to local memory on the caller's stack.
> The link pointer might point to a shared memory area, used by multiple threads/processes, so it needs to be stored atomically using rte_atomic_store_explicit(&link->val64, ..., rte_memory_order_seq_cst).

I checked every call of rte_eth_linkstatus_get in DPDK, and all the link parameters are local variables.
The dev->data->dev_link is placed in shared memory which could access from different threads/processes, it seems no need maintain another link struct which act the same role.

So I think we should keep current impl, and not using rte_atomic_store_explicit(&link->val64,...

Thanks

> 
>>  }
>>
>>  /**
>> diff --git a/lib/ethdev/rte_ethdev.h b/lib/ethdev/rte_ethdev.h
>> index 147257d6a2..ccf43e468a 100644
>> --- a/lib/ethdev/rte_ethdev.h
>> +++ b/lib/ethdev/rte_ethdev.h
>> @@ -332,12 +332,16 @@ struct rte_eth_stats {
>>  /**
>>   * A structure used to retrieve link-level information of an Ethernet
>> port.
>>   */
>> -__extension__
>> -struct __rte_aligned(8) rte_eth_link { /**< aligned for atomic64
>> read/write */
>> -	uint32_t link_speed;        /**< RTE_ETH_SPEED_NUM_ */
>> -	uint16_t link_duplex  : 1;  /**< RTE_ETH_LINK_[HALF/FULL]_DUPLEX
>> */
>> -	uint16_t link_autoneg : 1;  /**< RTE_ETH_LINK_[AUTONEG/FIXED] */
>> -	uint16_t link_status  : 1;  /**< RTE_ETH_LINK_[DOWN/UP] */
>> +struct rte_eth_link {
>> +	union {
>> +		RTE_ATOMIC(uint64_t) val64; /**< used for atomic64
>> read/write */
>> +		struct {
>> +			uint32_t link_speed;	    /**< RTE_ETH_SPEED_NUM_
>> */
>> +			uint16_t link_duplex  : 1;  /**<
>> RTE_ETH_LINK_[HALF/FULL]_DUPLEX */
>> +			uint16_t link_autoneg : 1;  /**<
>> RTE_ETH_LINK_[AUTONEG/FIXED] */
>> +			uint16_t link_status  : 1;  /**<
>> RTE_ETH_LINK_[DOWN/UP] */
>> +		};
>> +	};
>>  };
>>
>>  /**@{@name Link negotiation
>> --
>> 2.17.1
> 
> .
>