[dpdk-users] ojectes from the mempool are disorderd by rte_mempool_generic_get

[Van Haaren, Harry]

Hi Jie,

Mempools don't provide any promise of dequeuing objects in a specific order. The mempool data-structure actually has a per-lcore cache which allows recycling buffers on a per-lcore basis, providing better performance (and certainly breaking your ordering hopes).

If you need strict I/O ordering from a datastructure, I suggest using a rte_ring. The ring datastructure is a FIFO, so the first packets you enqueue will be the first to dequeue too. Regardless of the rte_ring_dequeue_burst() size parameter, calling 1x 10 times, or 10x calls with 1 packet, the mbufs returned will be in the same order.

Hope that helps, -Harry

From: 宋捷 <songj at zctt.com>
Sent: Tuesday, March 24, 2020 5:16 AM
To: 'users' <users at dpdk.org>; 'Stephen Hemminger' <stephen at networkplumber.org>; Van Haaren, Harry <harry.van.haaren at intel.com>
Subject: ojectes from the mempool are disorderd by rte_mempool_generic_get

Hi all,


I created a packet mbuf pool (rte_pktmbuf_pool_create) for saving some packets, And using rte_mempool_obj_iter for put hundreds of packets (from a pcap file) into the packet pool.

Finally I need to get ten packets from the packet pool.
If I call rte_mempool_generic_get for one packet each time, and try 10 times, then I can get ten packets in correct order;
But if call rte_mempool_generic_get for ten packets one time, then these ten packets are disordered.

So using rte_mempool_generic_get for getting multi objects, it will return disorder objects from the mbuf pool?
If there’s any other way for can get objects from the mempool orderly

I used DPDK18.05

Thanks

Jie