[ovs-dev] [PATCH v8 04/13] netdev-dpdk: Serialise non-pmds mbufs' alloc/free.

[Ilya Maximets]

On 26.06.2018 12:19, Eelco Chaudron wrote:
> 
> 
> On 22 Jun 2018, at 21:03, Lam, Tiago wrote:
> 
>> On 18/06/2018 12:28, Eelco Chaudron wrote:
>>>
>>>
>>> On 11 Jun 2018, at 18:21, Tiago Lam wrote:
>>>
>>>> A new mutex, 'nonpmd_mp_mutex', has been introduced to serialise
>>>> allocation and free operations by non-pmd threads on a given mempool.
>>>>
>>>
>>> Can you explain why we need the mutex here? Can't see any reason why
>>> rte_pktmbuf_free() needs to be protected for non-pmd threads?
>>
>> My understanding is that each PMD has a thread-local cache of its own,
>> which it can access without locking. However, all non-PMD threads have
>> direct access to the mempool, and thus need to lock in order to access
>> it. Otherwise memory corruption could ensue.
>>
>> For OvS this means all these functions which can be called outside of a
>> PMD context, or without holding the `non_pmd_mutex` mutex, in struct
>> dp_netdev, need to lock before issue operations that modify a mempool.
>>
>> Hopefully I've put it succinctly enough without botching up the concept
>> too much.
> 
> I do not think this is a problem, as DPDK uses rte_mempool_default_cache() to figure out which cache to use.
> As the non PMD threads have a higher than RTE_MAX_LCORE id the cache will not be used.


It's not the problem. The main reason is that implementation of ring
buffers on which the mempools based is not preemptible:

  lib/librte_mempool/rte_mempool.h:
  * Note: the mempool implementation is not preemptible. An lcore must not be
  * interrupted by another task that uses the same mempool (because it uses a
  * ring which is not preemptible).

As soon as non-PMD threads are not pinned, they could be scheduled on the same
core and preempted while executing mempool operations.

> 
>>>
>>>> free_dpdk_buf() has been modified to make use of the introduced mutex.
>>>>
>>>> Signed-off-by: Tiago Lam <tiago.lam at intel.com>
>>>> ---
>>>>  lib/netdev-dpdk.c | 21 ++++++++++++++++++++-
>>>>  1 file changed, 20 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/lib/netdev-dpdk.c b/lib/netdev-dpdk.c
>>>> index f546507..efd7c20 100644
>>>> --- a/lib/netdev-dpdk.c
>>>> +++ b/lib/netdev-dpdk.c
>>>> @@ -294,6 +294,10 @@ static struct ovs_mutex dpdk_mp_mutex
>>>> OVS_ACQ_AFTER(dpdk_mutex)
>>>>  static struct ovs_list dpdk_mp_free_list
>>>> OVS_GUARDED_BY(dpdk_mp_mutex)
>>>>      = OVS_LIST_INITIALIZER(&dpdk_mp_free_list);
>>>>
>>>> +/* This mutex must be used by non pmd threads when allocating or
>>>> freeing
>>>> + * mbufs through mempools. */
>>>> +static struct ovs_mutex nonpmd_mp_mutex = OVS_MUTEX_INITIALIZER;
>>>> +
>>>>  /* Wrapper for a mempool released but not yet freed. */
>>>>  struct dpdk_mp {
>>>>       struct rte_mempool *mp;
>>>> @@ -461,6 +465,8 @@ struct netdev_rxq_dpdk {
>>>>      dpdk_port_t port_id;
>>>>  };
>>>>
>>>> +static bool dpdk_thread_is_pmd(void);
>>>> +
>>>>  static void netdev_dpdk_destruct(struct netdev *netdev);
>>>>  static void netdev_dpdk_vhost_destruct(struct netdev *netdev);
>>>>
>>>> @@ -494,6 +500,12 @@ dpdk_buf_size(int mtu)
>>>>                       NETDEV_DPDK_MBUF_ALIGN);
>>>>  }
>>>>
>>>> +static bool
>>>> +dpdk_thread_is_pmd(void)
>>>> +{
>>>> +     return rte_lcore_id() != NON_PMD_CORE_ID;
>>>
>>> Will this continue to work with newer DPDK versions? I've not looked at
>>> it in detail, but I did notice that the vhost threads in the newer DPDK
>>> now get created with rte_ctrl_thread_create() and does some lcore
>>> mangling.
>>>
>>
>> I had a look.  This seems to be a wrapper for creating management
>> threads only, to standardize pthread creation around that in DPDK. Usual
>> EAL thread creation and assigning of LCORE_ID_ANY to non-EAL threads
>> seems to still be the same.
> 
> Thanks for checking.
>>
>>>> +}
>>>> +
>>>>  /* Allocates an area of 'sz' bytes from DPDK.  The memory is zero'ed.
>>>>   *
>>>>   * Unlike xmalloc(), this function can return NULL on failure. */
>>>> @@ -506,9 +518,16 @@ dpdk_rte_mzalloc(size_t sz)
>>>>  void
>>>>  free_dpdk_buf(struct dp_packet *p)
>>>>  {
>>>> -    struct rte_mbuf *pkt = (struct rte_mbuf *) p;
>>>> +    if (!dpdk_thread_is_pmd()) {
>>>> +        ovs_mutex_lock(&nonpmd_mp_mutex);
>>>> +    }
>>>>
>>>> +    struct rte_mbuf *pkt = (struct rte_mbuf *) p;
>>>
>>> Should we not use a container_of macro here?
>>>
>>
>> I'll use it for the next iteration.
>>
>> Tiago.
> 
> 
>