[dpdk-dev] [PATCH v2 0/3] vhost: MQ live-migration fixes

[Maxime Coquelin <maxime.coquelin@redhat.com>]

Sorry, posted the wrong version. Only patch 2 changes in the v2,
the log_lock is read-locked is moved after the VHOST_F_LOG_ALL
feature check, so that it does not degrade performance when
not doing the live-migration.

This 3 patches series fixes issues met when doing live-migration
with multiple queue pairs.

Patch 1 is theorical and unlikely to be reproduced in real use-cases,
so it may be safe not to pick it in stable trees.

Patch 2 reproduces quite often when lots of packets are being processed.
Easiest way to reproduce it is to run DPDK in guest and perform IO
loopback with testpmd. This patch targets both v16.11 & v17.11 stable
trees, and will require a rework for v16.11 as some dirty logging
functions moved from virtio-net.c to vhost.h. I'm not sure of the
process here, but I can provide the v16.11 backport if needed.

Patch 3 is a regression introduced in v17.11. For a reason I have
yet to understand, QEMU sends VHOST_USER_SET_VRING_ADDR requests
when live-migration is initiated. The problem is that the vhost-user
protocol thread has no way to be sure the PMD threads are accessing
the rings or not. As the new addresses sent by QEMU are the same
it sent intially, this patch just ignores them.

Regards,
Maxime

Maxime Coquelin (3):
  vhost: fix fd leak in VHOST_USER_SET_LOG_BASE
  vhost: protect dirty logging against logging base change
  vhost: don't invalidate vrings if new addresses are identical

 lib/librte_vhost/vhost.c      |  2 ++
 lib/librte_vhost/vhost.h      | 14 +++++++++++---
 lib/librte_vhost/vhost_user.c | 32 ++++++++++++++++++++++++++++----
 3 files changed, 41 insertions(+), 7 deletions(-)

-- 
2.14.3