[RFC,v2,0/2] Add a PMD for DMA-accelerated vhost-user

  In vhost-user enqueue and dequeue operations, where data movement is
heavily involved, performing large memory copies usually takes up a
major part of CPU cycles and becomes the hot spot. To offload expensive
memory operations from the CPU, this patch set proposes to leverage DMA
engines, e.g., I/OAT, a DMA engine in the Intel's processor, to accelerate
large copies for vhost-user.

We implement a new PMD for the DMA accelerated vhost-user, called
vhost-dma. This PMD leverages librte_vhost to handle vhost messages,
but implements own vring's enqueue and dequeue operations. It offloads
large memory copies to the DMA in an asynchronous mode; that is, the CPU
just submits copy jobs to the DMA but without waiting for its
completion. Thus, there is no CPU intervention during data transfer;
we can save precious CPU cycles and improve the overall throughput for
vhost-user based applications, like OVS.

The PMD provides basic functionality of packet reception and
transmission. During packet reception and transmission, it offloads
large copies to the DMA and performs small copies by the CPU, due to
startup overheads associated with the DMA.

The PMD is able to support various DMA engines to accelrate data
movements in enqueue and dequeue operations; currently, the supported
DMA engine is I/OAT. The PMD just supports I/OAT acceleration in the
PMD's transmit data path (i.e. vring's enqueue operation); it still
uses the CPU to perform all copies in the PMD's receive data path (i.e.
vring's dequeue operation). In addition, the PMD just supports split ring.

Users can explicitly assign a DMA device to a TX queue by the
parameter 'dmas'. But currently, one DMA device can only be used by
one queue and a queue can use one DMA device at a time. In addition,
the PMD supports multiqueue and both client and server modes. Users can
specify the queue number and client/server mode by 'queues' and 'client'
parameters.

We measure the performance of vhost-dma in testpmd. With 1024 bytes
packets, compared with vhost-user PMD, vhost-dma can improve the throughput
for host testpmd around 20%~30% in the VM2VM and PVP cases; with larger
packets, the throughput improvement will be higher.

Change log
==========
v2:
- provide DMA-accelerated vhost-user PMD to support various DMA engines.

Jiayu Hu (2):
  vhost: populate guest memory for DMA-accelerated vhost-user
  net/vhost_dma: add vHost DMA driver

 config/common_base                                 |    2 +
 config/common_linux                                |    1 +
 drivers/Makefile                                   |    2 +-
 drivers/net/Makefile                               |    1 +
 drivers/net/vhost_dma/Makefile                     |   31 +
 drivers/net/vhost_dma/eth_vhost.c                  | 1495 ++++++++++++++++++++
 drivers/net/vhost_dma/eth_vhost.h                  |  264 ++++
 drivers/net/vhost_dma/internal.h                   |  225 +++
 .../net/vhost_dma/rte_pmd_vhost_dma_version.map    |    4 +
 drivers/net/vhost_dma/virtio_net.c                 | 1234 ++++++++++++++++
 lib/librte_vhost/rte_vhost.h                       |    1 +
 lib/librte_vhost/socket.c                          |   11 +
 lib/librte_vhost/vhost.h                           |    2 +
 lib/librte_vhost/vhost_user.c                      |    3 +-
 mk/rte.app.mk                                      |    1 +
 15 files changed, 3275 insertions(+), 2 deletions(-)
 create mode 100644 drivers/net/vhost_dma/Makefile
 create mode 100644 drivers/net/vhost_dma/eth_vhost.c
 create mode 100644 drivers/net/vhost_dma/eth_vhost.h
 create mode 100644 drivers/net/vhost_dma/internal.h
 create mode 100644 drivers/net/vhost_dma/rte_pmd_vhost_dma_version.map
 create mode 100644 drivers/net/vhost_dma/virtio_net.c