[dpdk-dev] Is rte_mempool library is multi-thread safe ???

[ankit kumar]

Hi,

I was testing rte_ring in DPDK as it provides multi-consumer &
multi-produce queue(lock-free but not wait-free).

It was working fine but i am not sure that memory allocation &
deallocation in rte_mempool library, is multi-thread safe or not.
Because its giving me error during allocation(from
multi-thread) as "Allocation failed(returned pointer is NULL in
allocation function rte_pktmbuf_alloc())" & during rte_pktmbuf_free()
it generates segmentation fault and when i traced out, core file
generated, it indicated that the problem is in rte_pktmbuf_free().

Any help !!!

[Thomas Monjalon]

Hello,

Post by ankit kumar
It was working fine but i am not sure that memory allocation &
deallocation in rte_mempool library, is multi-thread safe or not.

From http://dpdk.org/doc/api/rte__mempool_8h.html:
"the mempool implementation is not preemptable"
There is more explanations in this email:
http://dpdk.org/ml/archives/dev/2013-August/000402.html

[Peter Chen]

does that mean that on the same core, we can't do rte_eth_rx_burst in one
thread (I assume this function allocates from mempool for storing mbufs
everytime it receives a packet), while another thread calls
rte_pktmbuf_alloc from the same mem_pool?


On Thu, Dec 19, 2013 at 9:04 AM, Thomas Monjalon

Post by Thomas Monjalon
Hello,

Post by ankit kumar
It was working fine but i am not sure that memory allocation &
deallocation in rte_mempool library, is multi-thread safe or not.

[Olivier MATZ]

Hi Peter,

Post by Peter Chen
does that mean that on the same core, we can't do rte_eth_rx_burst in one
thread (I assume this function allocates from mempool for storing mbufs
everytime it receives a packet), while another thread calls
rte_pktmbuf_alloc from the same mem_pool?

That's correct. In the rte_mempool code, there is a per-lcore cache:
see the local_cache field of struct rte_mempool.

If you are running several pthreads per lcore, they will share the
same cache if they have the same lcore_id and the mempool is
not designed for that. Therefore it can return wrong results.
The cache can be disabled (at run-time or compile-time), but you will
loose a lot of performance.

Even if you solve the problem of the cache, as the mempool uses a ring
internally, you would still experiment performance issues (see links
from Thomas' previous email).

By the way, why would you need to have several pthreads on one lcore?

Regards,
Olivier