[dpdk-dev] [PATCH v3 0/2] app/testpmd: improve multiprocess support

Marcin Kerlin marcinx.kerlin at intel.com
Mon Sep 26 16:53:04 CEST 2016


This patch ensure not overwrite device data in the multiprocess application.

1)Changes in the library introduces continuity in array rte_eth_dev_data[]
shared between all processes. Secondary process adds new entries in free
space instead of overwriting existing entries.

2)Changes in application testpmd allow secondary process to attach the
mempool created by primary process rather than create new and in the case of
quit or force quit to free devices data from shared array rte_eth_dev_data[].

-------------------------
How to reproduce the bug:

1) Run primary process:
-c 0xf -n 4 --socket-mem='512,0' -w 03:00.1 -w 03:00.0 --proc-type=primary 
--file-prefix=xz1 -- -i

(gdb) print rte_eth_devices[0].data.name
$52 = "3:0.1"
(gdb) print rte_eth_devices[1].data.name
$53 = "3:0.0"

2) Run secondary process:
-c 0xf0 --socket-mem='512,0' -n 4 -v -b 03:00.1 -b 03:00.0 -b 01:00.0 -b 01:00.0
--vdev 'eth_pcap0,rx_pcap=/var/log/device1.pcap,tx_pcap=/var/log/device2.pcap'
--proc-type=secondary --file-prefix=xz1 -- -i

(gdb) print rte_eth_devices[0].data.name
$52 = "eth_pcap0"
(gdb) print rte_eth_devices[1].data.name
$53 = "eth_pcap1"

3) Go back to the primary and re-check:
(gdb) print rte_eth_devices[0].data.name
$54 = "eth_pcap0"
(gdb) print rte_eth_devices[1].data.name
$55 = "eth_pcap1"

It means that secondary process overwrite data of primary process.

This patch fix it and now if we go back to the primary and re-check then 
everything is fine:
(gdb) print rte_eth_devices[0].data.name
$56 = "3:0.1"
(gdb) print rte_eth_devices[1].data.name
$57 = "3:0.0"

So after this fix structure rte_eth_dev_data[] will keep all data one 
after the other instead of overwriting:

(gdb) print rte_eth_dev_data[0].name
$52 = "3:0.1"
(gdb) print rte_eth_dev_data[1].name
$53 = "3:0.0"
(gdb) print rte_eth_dev_data[2].name
$54 = "eth_pcap0"
(gdb) print rte_eth_dev_data[3].name
$55 = "eth_pcap1"
and so on will be append in the next indexes

If secondary process will be turned off then also will be deleted from array:
(gdb) print rte_eth_dev_data[0].name
$52 = "3:0.1"
(gdb) print rte_eth_dev_data[1].name
$53 = "3:0.0"
(gdb) print rte_eth_dev_data[2].name
$54 = ""
(gdb) print rte_eth_dev_data[3].name
$55 = ""
this also allows re-use index 2 and 3 for next another process
-------------------------

Breaking ABI:
Changes in the library librte_ether causes extending existing structure
rte_eth_dev_data with a new field lock. The reason is that this structure is
sharing between all the processes so it should be protected against attempting
to write from two different processes.

Tomasz Kulasek sent announce ABI change in librte_ether on 21 July 2016.
I would like to join to this breaking ABI, if it is possible.

v2:
* fix syntax error in version script
v3:
* changed scope of function
* improved description

Marcin Kerlin (2):
  librte_ether: ensure not overwrite device data in mp app
  app/testpmd: improve handling of multiprocess

 app/test-pmd/testpmd.c                 | 36 +++++++++++++-
 app/test-pmd/testpmd.h                 |  1 +
 lib/librte_ether/rte_ethdev.c          | 90 +++++++++++++++++++++++++++++++---
 lib/librte_ether/rte_ethdev.h          | 24 +++++++++
 lib/librte_ether/rte_ether_version.map |  7 +++
 5 files changed, 148 insertions(+), 10 deletions(-)

-- 
1.9.1



More information about the dev mailing list