mbox series

[v4,bpf-next,00/10] AF_XDP ZC selftests

Message ID 20220616180609.905015-1-maciej.fijalkowski@intel.com (mailing list archive)
Headers show
Series AF_XDP ZC selftests | expand

Message

Maciej Fijalkowski June 16, 2022, 6:05 p.m. UTC 
v3->v4:
* use ice_{down,up} rather than ice_{stop,open} and check retvals
  when toggling loopback mode (Jakub)
* Remove patch that was throwing away xsk->outstanding_tx (Magnus)

v2->v3:
* split loopback patch to ice_set_features() refactor part and other
  part with actual loopback toggling support (Alexandr)
* collect more acks from Magnus

v1->v2:
* collect acks from Magnus
* drop redundant 'ret' variable in patch 4 (Magnus)
* drop redundant assignments to ifobject->xdp_flags in patch 10 (Magnus)
* use NETIF_F_LOOPBACK instead of introducing priv-flag (Alexandr)

Hi!

This set makes it possible to test ice driver with test suite that
xdpxceiver provides. In order to do it, device needs to be turned on to
loopback mode, so in first patch ice_set_features() is refactored and
then in second patch NETIF_F_LOOPBACK support is added to it.
Furthermore, ethtool's loopback test is fixed for ice (patch 3 and 4).
Besides turning device into loopback mode, I was limiting queue count to
1, so that none of the steering filters are needed.

Rest of the work is related to xdpxceiver. Currently it is assumed that
underlying device supports native XDP. It is true for veth, but might
not be for other device that might be used with xdpxceiver once this
patch set land. So, patch 4 adds a logic to find out if underlying
device supports XDP so that TEST_MODE_DRV can be used for test suite.
Patch 6 restores close() on netns fd that was removed by accident.

In patch 7, default Rx pkt stream is added so physical device testing
will be able to use shared UMEM in a way that Tx will utilize first half
of buffer space and Rx second one. Then, patch 8 adds support for running
xdpxceiver on physical devices.

Patch 9 prepares xdpxceiver to address ZC drivers that clean lazily Tx
descriptors (and therefore produce XSK descs to CQ) so that pacing
algorithm works fine.

Patch 10 finally adds new TEST_MODE_ZC for testing zero copy AF_XDP
driver support.

This work already allowed us to spot and fix two bugs in AF_XDP kernel
side ([0], [1]).

v1 is here [2].
v2 is here [3].
v3 is here [4].

[0]: https://lore.kernel.org/bpf/20220425153745.481322-1-maciej.fijalkowski@intel.com/
[1]: https://lore.kernel.org/bpf/20220607142200.576735-1-maciej.fijalkowski@intel.com/
[2]: https://lore.kernel.org/bpf/20220610150923.583202-1-maciej.fijalkowski@intel.com/
[3]: https://lore.kernel.org/bpf/20220614174749.901044-1-maciej.fijalkowski@intel.com/
[4]: https://lore.kernel.org/bpf/20220615161041.902916-1-maciej.fijalkowski@intel.com/

Thank you.


Maciej Fijalkowski (10):
  ice: compress branches in ice_set_features()
  ice: allow toggling loopback mode via ndo_set_features callback
  ice: check DD bit on Rx descriptor rather than (EOP | RS)
  ice: do not setup vlan for loopback VSI
  selftests: xsk: query for native XDP support
  selftests: xsk: add missing close() on netns fd
  selftests: xsk: introduce default Rx pkt stream
  selftests: xsk: add support for executing tests on physical device
  selftests: xsk: rely on pkts_in_flight in wait_for_tx_completion()
  selftests: xsk: add support for zero copy testing

 drivers/net/ethernet/intel/ice/ice_ethtool.c |   2 +-
 drivers/net/ethernet/intel/ice/ice_main.c    |  81 ++--
 tools/testing/selftests/bpf/test_xsk.sh      |  52 ++-
 tools/testing/selftests/bpf/xdpxceiver.c     | 380 ++++++++++++++-----
 tools/testing/selftests/bpf/xdpxceiver.h     |   9 +-
 5 files changed, 378 insertions(+), 146 deletions(-)