[dpdk-dev] [PATCH v9 0/2] Add VHOST PMD

[Tetsuya Mukawa]

On 2016/03/01 11:00, Qiu, Michael wrote:
> On 2/26/2016 4:36 PM, Tetsuya Mukawa wrote:
>> On 2016/02/26 13:29, Tetsuya Mukawa wrote:
>>> On 2016/02/25 16:51, Qiu, Michael wrote:
>>>> On 2/24/2016 1:10 PM, Tetsuya Mukawa wrote:
>>>>> On 2016/02/24 11:45, Qiu, Michael wrote:
>>>>>> Hi,  Tetsuya
>>>>>>
>>>>>> When I applied your v6 patch, I could reach 9.5Mpps with 64B packet.
>>>>>>
>>>>>> But when apply v9 only 8.4 Mpps, could you figure out why has
>>>>>> performance drop?
>>>>> Hi Michael,
>>>>>
>>>>> Thanks for checking it.
>>>>> I tried to re-produce it, but I don't see the drop on my environment.
>>>>> (My cpu is Xeon E5-2697-v2, and the performances of v6 and v9 patch are
>>>>> almost 5.9Mpps)
>>>>> Did you use totally same code except for vhost PMD?
>>>> Yes, totally same code and same platform, only difference is versions of
>>>> vhost PMD.
>>>>
>>>> BTW, I have set the frontend mergeable off.
>>> I have checked below cases.
>>>  - Case1: Disable mergeable feature in virtio-net PMD.
>>>  - Case2: Disable mergeable feature in virtio-net PMD and use
>>> '--txqflags=0xf01' option to use simple ring deploying.
>>> Both cases,  I still cannot see the drop.
>>>
>>> Anyway, I will send a few patch-series to determine the cause of drop.
>>> So, could you please apply them and check the performance to determine
>>> which cause the drop?
>> Hi Michael,
>>
>> I may find what causes the drop.
>> Could you please restart testpmd on guest when you see the drop, then
>> check performance again?
>>
>> I guess the drop will occur only first time when testpmd on guest and
>> host is connected.
>> Here are rough steps.
>>
>> 1. Start testpmd on host
>> 2. Start QEMU
>> 3. Start testpmd on guest
>>
>> Then you will see the drop.
>> Probably, if testpmd on guest is restarted, then you don't see the drop
>> again.
>>
>> 4. Type 'quit' on guest.
>> 5. Start testpmd on guest again.

Hi Michael,

I am sorry that above was caused by my miss configuration.
So please ignore it.
If you can have time today, could you please check v7 and v8 performance?

Thanks,
Tetsuya

> OK, I will help to tested today.
>
> Thanks,
> Michael
>> If so, I guess the drop is caused by queue notifying.
>> Could you please let me know whether your issue is above case?
>>
>> Thanks,
>> Tetsuya
>>
>>> Thanks,
>>> Tetsuya
>>>
>>>> Thanks,
>>>> Michael
>>>>> Thanks,
>>>>> Tetsuya
>>>>>
>>>>>> Thanks,
>>>>>> Michael
>>>>>> On 2/9/2016 5:38 PM, Tetsuya Mukawa wrote:
>>>>>>> The patch introduces a new PMD. This PMD is implemented as thin wrapper
>>>>>>> of librte_vhost.
>>>>>>>
>>>>>>>
>>>>>>> PATCH v9 changes:
>>>>>>>  - Fix a null pointer access issue implemented in v8 patch.
>>>>>>>
>>>>>>> PATCH v8 changes:
>>>>>>>  - Manage ether devices list instead of internal structures list.
>>>>>>>  - Remove needless NULL checking.
>>>>>>>  - Replace "pthread_exit" to "return NULL".
>>>>>>>  - Replace rte_panic to RTE_LOG, also add error handling.
>>>>>>>  - Remove duplicated lines.
>>>>>>>  - Remove needless casting.
>>>>>>>  - Follow coding style.
>>>>>>>  - Remove needless parenthesis.
>>>>>>>
>>>>>>> PATCH v7 changes:
>>>>>>>  - Remove needless parenthesis.
>>>>>>>  - Add release note.
>>>>>>>  - Remove needless line wraps.
>>>>>>>  - Add null pointer check in vring_state_changed().
>>>>>>>  - Free queue memory in eth_queue_release().
>>>>>>>  - Fix wrong variable name.
>>>>>>>  - Fix error handling code of eth_dev_vhost_create() and
>>>>>>>    rte_pmd_vhost_devuninit().
>>>>>>>  - Remove needless null checking from rte_pmd_vhost_devinit/devuninit().
>>>>>>>  - Use port id to create mac address.
>>>>>>>  - Add doxygen style comments in "rte_eth_vhost.h".
>>>>>>>  - Fix wrong comment in "mk/rte.app.mk".
>>>>>>>
>>>>>>> PATCH v6 changes:
>>>>>>>  - Remove rte_vhost_driver_pmd_callback_registe().
>>>>>>>  - Support link status interrupt.
>>>>>>>  - Support queue state changed interrupt.
>>>>>>>  - Add rte_eth_vhost_get_queue_event().
>>>>>>>  - Support numa node detection when new device is connected.
>>>>>>>
>>>>>>> PATCH v5 changes:
>>>>>>>  - Rebase on latest master.
>>>>>>>  - Fix RX/TX routine to count RX/TX bytes.
>>>>>>>  - Fix RX/TX routine not to count as error packets if enqueue/dequeue
>>>>>>>    cannot send all packets.
>>>>>>>  - Fix if-condition checking for multiqueues.
>>>>>>>  - Add "static" to pthread variable.
>>>>>>>  - Fix format.
>>>>>>>  - Change default behavior not to receive queueing event from driver.
>>>>>>>  - Split the patch to separate rte_eth_vhost_portid2vdev().
>>>>>>>
>>>>>>> PATCH v4 changes:
>>>>>>>  - Rebase on latest DPDK tree.
>>>>>>>  - Fix cording style.
>>>>>>>  - Fix code not to invoke multiple messaging handling threads.
>>>>>>>  - Fix code to handle vdev parameters correctly.
>>>>>>>  - Remove needless cast.
>>>>>>>  - Remove needless if-condition before rt_free().
>>>>>>>
>>>>>>> PATCH v3 changes:
>>>>>>>  - Rebase on latest matser
>>>>>>>  - Specify correct queue_id in RX/TX function.
>>>>>>>
>>>>>>> PATCH v2 changes:
>>>>>>>  - Remove a below patch that fixes vhost library.
>>>>>>>    The patch was applied as a separate patch.
>>>>>>>    - vhost: fix crash with multiqueue enabled
>>>>>>>  - Fix typos.
>>>>>>>    (Thanks to Thomas, Monjalon)
>>>>>>>  - Rebase on latest tree with above bernard's patches.
>>>>>>>
>>>>>>> PATCH v1 changes:
>>>>>>>  - Support vhost multiple queues.
>>>>>>>  - Rebase on "remove pci driver from vdevs".
>>>>>>>  - Optimize RX/TX functions.
>>>>>>>  - Fix resource leaks.
>>>>>>>  - Fix compile issue.
>>>>>>>  - Add patch to fix vhost library.
>>>>>>>
>>>>>>> RFC PATCH v3 changes:
>>>>>>>  - Optimize performance.
>>>>>>>    In RX/TX functions, change code to access only per core data.
>>>>>>>  - Add below API to allow user to use vhost library APIs for a port managed
>>>>>>>    by vhost PMD. There are a few limitations. See "rte_eth_vhost.h".
>>>>>>>     - rte_eth_vhost_portid2vdev()
>>>>>>>    To support this functionality, vhost library is also changed.
>>>>>>>    Anyway, if users doesn't use vhost PMD, can fully use vhost library APIs.
>>>>>>>  - Add code to support vhost multiple queues.
>>>>>>>    Actually, multiple queues functionality is not enabled so far.
>>>>>>>
>>>>>>> RFC PATCH v2 changes:
>>>>>>>  - Fix issues reported by checkpatch.pl
>>>>>>>    (Thanks to Stephen Hemminger)
>>>>>>>
>>>>>>>
>>>>>>> Tetsuya Mukawa (2):
>>>>>>>   ethdev: Add a new event type to notify a queue state changed event
>>>>>>>   vhost: Add VHOST PMD
>>>>>>>
>>>>>>>  MAINTAINERS                                 |   4 +
>>>>>>>  config/common_linuxapp                      |   6 +
>>>>>>>  doc/guides/nics/index.rst                   |   1 +
>>>>>>>  doc/guides/rel_notes/release_2_3.rst        |   4 +
>>>>>>>  drivers/net/Makefile                        |   4 +
>>>>>>>  drivers/net/vhost/Makefile                  |  62 ++
>>>>>>>  drivers/net/vhost/rte_eth_vhost.c           | 911 ++++++++++++++++++++++++++++
>>>>>>>  drivers/net/vhost/rte_eth_vhost.h           | 109 ++++
>>>>>>>  drivers/net/vhost/rte_pmd_vhost_version.map |  11 +
>>>>>>>  lib/librte_ether/rte_ethdev.h               |   2 +
>>>>>>>  mk/rte.app.mk                               |   6 +
>>>>>>>  11 files changed, 1120 insertions(+)
>>>>>>>  create mode 100644 drivers/net/vhost/Makefile
>>>>>>>  create mode 100644 drivers/net/vhost/rte_eth_vhost.c
>>>>>>>  create mode 100644 drivers/net/vhost/rte_eth_vhost.h
>>>>>>>  create mode 100644 drivers/net/vhost/rte_pmd_vhost_version.map
>>>>>>>