[dpdk-dev] [PATCH v3 00/19] Vhost-user: Implement device IOTLB support

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

This v3 lists the feature in the release note, and fixes the bug in
is_vring_iotlb_update() reported by Yuanhan.

The purpose of this series is to add support for
VIRTIO_F_IOMMU_PLATFORM feature, by implementing device IOTLB in the
vhost-user backend. It improves the guest safety by enabling the
possibility to isolate the Virtio device.

It makes possible to use Virtio PMD in guest with using VFIO driver
without enable_unsafe_noiommu_mode parameter set, so that the DPDK
application on guest can only access memory its has been allowed to,
and preventing malicious/buggy DPDK application in guest to make
vhost-user backend write random guest memory. Note that Virtio-net
Kernel driver also support IOMMU.

The series depends on Qemu's "vhost-user: Specify and implement
device IOTLB support" [0], available upstream and which will be part
of Qemu v2.10 release.

Performance-wise, even if this RFC has still room for optimizations,
no performance degradation is noticed with static mappings (i.e. DPDK
on guest) with PVP benchmark:
	Traffic Generator: Moongen (lua-trafficgen)
	Acceptable Loss: 0.005%
	Validation run time: 1 min
	Guest DPDK version/commit: v17.05
	QEMU version/commit: master (6db174aed1fd)
	Virtio features: default
	CPU: Intel(R) Xeon(R) CPU E5-2667 v4 @ 3.20GHz
	NIC: 2 x X710
	Page size: 1G host/1G guest
	Results (bidirectional, total of the two flows):
	 - base: 18.8Mpps
	 - base + IOTLB series, IOMMU OFF: 18.8Mpps
	 - base + IOTLB series, IOMMU ON: 18.8Mpps (14.5Mpps w/o PATCH 21/21)

This is explained because IOTLB misses, which are very costly, only
happen at startup time. Indeed, once used, the buffers are not
invalidated, so if the IOTLB cache is large enough, there will be only
cache hits. Also, the use of 1G huge pages improves the IOTLB cache
searching time by reducing the number of entries.

With 2M hugepages, a performance degradation is seen with IOMMU on:
	Traffic Generator: Moongen (lua-trafficgen)
	Acceptable Loss: 0.005%
	Validation run time: 1 min
	Guest DPDK version/commit: v17.05
	QEMU version/commit: master (6db174aed1fd)
	Virtio features: default
	CPU: Intel(R) Xeon(R) CPU E5-2667 v4 @ 3.20GHz
	NIC: 2 x X710
	Page size: 2M host/2M guest
	Results (bidirectional, total of the two flows):
	 - base: 18.8Mpps
	 - base + IOTLB series, IOMMU OFF: 18.8Mpps
	 - base + IOTLB series, IOMMU ON: 13.5Mpps (12.4Mpps wo PATCH 21/21)

A possible improvement would be to merge contiguous IOTLB entries sharing
the same permissions. A very rough patch implementing this idea fixes
the performance degradation (18.8Mpps), but the required work to clean
it would delay this series after v17.11.

With dynamic mappings (i.e. Virtio-net kernel driver), this is another
story. The performance is so poor it makes it almost unusable. Indeed,
since the Kernel driver unmaps the buffers as soon as they are handled,
almost all descriptors buffers addresses translations result in an IOTLB
miss. There is not much that can be done on DPDK side, except maybe
batching IOTLB miss requests no to break bursts, but it would require
a big rework. In Qemu, we may consider enabling IOMMU MAP notifications,
so that DPDK receives the IOTLB updates without having to send IOTLB miss
request.

Regarding the design choices:
 - I initially intended to use userspace RCU library[1] for the cache
implementation, but it would have added an external dependency, and the
lib is not available in all distros. Qemu for example got rid of this
dependency by copying some of the userspace RCU lib parts into Qemu tree,
but this is not possible with DPDK due to licensing issues (RCU lib is
LGPL v2). Thanks to Jason advice, I implemented the cache using rd/wr
locks.
 - I initially implemented a per-device IOTLB cache, but the concurrent
acccesses on the IOTLB lock had huge impact on performance (~-40% in
bidirectionnal, expect even worse with multiqueue). I move to a per-
virtqueue IOTLB design, which prevents this concurrency.
 - The slave IOTLB miss request supports reply-ack feature in spec, but
this version doesn't block or busy-wait for the corresponding update so
that other queues sharing the same lcore can be processed in the meantime.

For those who would like to test the series, I made it available on
gitlab[2] (vhost_user_iotlb_v1 tag). The guest kernel command line requires
the intel_iommu=on parameter, and the guest should be started with and
iommu device attached to the virtio-net device. For example:

./qemu-system-x86_64 \
  -enable-kvm -m 4096 -smp 2 \
  -M q35,kernel-irqchip=split \
  -cpu host \
  -device intel-iommu,device-iotlb=on,intremap \
  -device ioh3420,id=root.1,chassis=1 \
  -chardev socket,id=char0,path=/tmp/vhost-user1 \
  -netdev type=vhost-user,id=hn2,chardev=char0 \
  -device virtio-net-pci,netdev=hn2,id=v0,mq=off,mac=$MAC,bus=root.1,disable-modern=off,disable-legacy=on,iommu_platform=on,ats=on \
...

[0]: https://lists.gnu.org/archive/html/qemu-devel/2017-06/msg00520.html
[1]: http://liburcu.org/
[2]: https://gitlab.com/mcoquelin/dpdk-next-virtio/commits/vhost_user_iotlb_v2

Changes since v1:
=================
- Add feature support in release note (Yuanhan)
- Fix return value in is_vring_iotlb_update() (Yuanhan)

Changes since v1:
=================
- Implement random cache entry eviction instead of removing all entries
  when cache is full (Yuanhan)
- Adds bounds check on vring_idx (Remy)
- Initialize slave_req_fd to -1 (Tiwei)
- Remove virtio-net device lock (Tiwei)
- Simplify vhost_user_iotlb_cache_remove() (Tiwei)
- Squash iotlb lock usage optimization patch
- Inline noiommu part of vhost_iova_to_vva to remove performance
  regressions with IOMMU=off
- Reworked iotlb files copyrights

Changes since RFC:
==================
- Fix memory leak in error patch reported by Jens
- Rework wait for IOTLB update by stopping the burst to let other
  queues to be processed, if any. It implies the introduction an
  iotlb_pending_list, so that iotlb miss requests aren't sent multiple
  times for a same address.
- Optimize iotlb lock usage to recover to same as IOMMU off performance
- Fix device locking issue in rte_vhost_dequeue_burst() error path
- Change virtio_dev_rx error handling for consistency with mergeable rx,
  and to ease returning in case of IOTLB misses.
- Fix checkpatch warnings reported by checkpatch@dpdk.org

Maxime Coquelin (19):
  Revert "vhost: workaround MQ fails to startup"
  vhost: make error handling consistent in rx path
  vhost: prepare send_vhost_message() to slave requests
  vhost: add support to slave requests channel
  vhost: declare missing IOMMU-related definitions for old kernels
  vhost: add iotlb helper functions
  vhost: iotlb: add pending miss request list and helpers
  vhost-user: add support to IOTLB miss slave requests
  vhost: initialize vrings IOTLB caches
  vhost-user: handle IOTLB update and invalidate requests
  vhost: introduce guest IOVA to backend VA helper
  vhost: use the guest IOVA to host VA helper
  vhost: enable rings at the right time
  vhost: don't dereference invalid dev pointer after its reallocation
  vhost: postpone rings addresses translation
  vhost-user: translate ring addresses when IOMMU enabled
  vhost-user: iommu: postpone device creation until ring are mapped
  vhost: iommu: Invalidate vring in case of matching IOTLB invalidate
  vhost: enable IOMMU support

 doc/guides/rel_notes/release_17_11.rst |   4 +
 lib/librte_vhost/Makefile              |   4 +-
 lib/librte_vhost/iotlb.c               | 352 +++++++++++++++++++++++++++++++++
 lib/librte_vhost/iotlb.h               |  76 +++++++
 lib/librte_vhost/vhost.c               | 115 +++++++++--
 lib/librte_vhost/vhost.h               |  61 +++++-
 lib/librte_vhost/vhost_user.c          | 312 +++++++++++++++++++++++++----
 lib/librte_vhost/vhost_user.h          |  20 +-
 lib/librte_vhost/virtio_net.c          |  96 +++++++--
 9 files changed, 962 insertions(+), 78 deletions(-)
 create mode 100644 lib/librte_vhost/iotlb.c
 create mode 100644 lib/librte_vhost/iotlb.h

-- 
2.13.6