mbox series

[v3,0/9] Register non-EAL threads as lcore

Message ID 20200622132531.21857-1-david.marchand@redhat.com (mailing list archive)
Headers show
Series Register non-EAL threads as lcore | expand


David Marchand June 22, 2020, 1:25 p.m. UTC
OVS and some other applications have been hacking into DPDK internals to
fake EAL threads and avoid performance penalty of only having non-EAL

This series proposes to add a new type of lcores and maps those threads
to such lcores.
non-EAL threads won't run the DPDK eal mainloop.
As a consequence, part of the EAL threads API cannot work.

Having new lcores appearing during the process lifetime is not expected
by some DPDK components. This is addressed by introducing init/uninit
callacks invoked when hotplugging of such lcore.

There is still some work/discussion:
- refuse new lcore role in incompatible EAL threads API (or document it
  only as those API were already incompatible?),
- think about deprecation notices for existing RTE_FOREACH_LCORE macros
  and consorts, it is probably worth discussing on how to iterate over

For the interested parties, I have a patch [1] against dpdk-latest OVS
branch that makes use of this series.

1: https://github.com/david-marchand/ovs/commit/dpdk-latest-external-lcores

Changes since v2:
- fixed windows build error due to missing trace stub,
- fixed bug when rolling back on lcore register,

Changes since v1:
- rebased on master (conflicts on merged Windows series),
- separated lcore role code cleanup in a patch,
- tried to use a single naming, so kept non-EAL threads as the main
  notion. non-EAL threads are then distinguished between registered and
  unregistered non-EAL threads,
- added unit tests (still missing some coverage, marked with a FIXME),
- reworked callbacks call under a common rwlock lock which protects
  lcores allocations and callbacks registration,
- introduced lcore iterators and converted the bucket mempool driver,