Hi, Damjan, My concern is even the cache for transmit thread is full, then buffer will be return to the main pool, but the buffer in transmit thread cache will not be freed or allocated any more. So it will be deposit in the transmit thread cache.
BR/Lollita Liu From: Damjan Marion (damarion) <damar...@cisco.com> Sent: Friday, March 15, 2019 7:06 PM To: Lollita Liu <lollita....@ericsson.com> Cc: vpp-dev@lists.fd.io Subject: Re: possible buffer deposit if allocating buffer via CPU on one NUMA and freeing buffer via CPU on another NUMA -- Damjan On 15 Mar 2019, at 11:42, Lollita Liu <lollita....@ericsson.com<mailto:lollita....@ericsson.com>> wrote: Hi, Damjan. After reading the buffer management mechanism of master’s latest code, I think there is possibility that some buffers can’t be used anymore. Let me an example to explain the special case: The VPP is running on a server with 2 socket, CPU 0/2 in socket 1 and CPU 1/3 in socket 2. Worker 0 on CPU 0 will receive packet from an interface, say memif, then handoff packet to worker 1/2/3 on CPU 1/2/3 for further processing. Finally worker 1/2/3 will send the packet out and free buffers. Looks the buffers will be allocated from buffer_pool default_numa_0, and firstly freed to default_num_0->threads[1/2/3]-> cached_buffers. But there is no one will allocated buffer from default_num_0->threads[1/3]-> cached_buffers. The buffers will be left on thread 1 and 2’s cached buffers then. I know handoff across NUMA is highly not recommended. Just want make sure my understanding is right and nothing in buffer management mechanism has ignored by me. Hmm, nut sure I understood the question. In case of handoff buffers will be freed into different cache, but then at some point, if cache is full, they will be returned to the main pool.
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#12566): https://lists.fd.io/g/vpp-dev/message/12566 Mute This Topic: https://lists.fd.io/mt/30439106/21656 Group Owner: vpp-dev+ow...@lists.fd.io Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com] -=-=-=-=-=-=-=-=-=-=-=-