On Sun, Dec 17, 2017 at 02:54:55PM +0200, Marcel Apfelbaum wrote: > Signed-off-by: Marcel Apfelbaum <mar...@redhat.com> > Signed-off-by: Yuval Shaia <yuval.sh...@oracle.com> > --- > docs/pvrdma.txt | 145 > ++++++++++++++++++++++++++++++++++++++++++++++++++++++++ > 1 file changed, 145 insertions(+) > create mode 100644 docs/pvrdma.txt > > diff --git a/docs/pvrdma.txt b/docs/pvrdma.txt > new file mode 100644 > index 0000000000..74c5cf2495 > --- /dev/null > +++ b/docs/pvrdma.txt > @@ -0,0 +1,145 @@ > +Paravirtualized RDMA Device (PVRDMA) > +==================================== > + > + > +1. Description > +=============== > +PVRDMA is the QEMU implementation of VMware's paravirtualized RDMA device. > +It works with its Linux Kernel driver AS IS, no need for any special guest > +modifications. > + > +While it complies with the VMware device, it can also communicate with bare > +metal RDMA-enabled machines and does not require an RDMA HCA in the host, it > +can work with Soft-RoCE (rxe). > + > +It does not require the whole guest RAM to be pinned allowing memory > +over-commit and, even if not implemented yet, migration support will be > +possible with some HW assistance. > + > +A project presentation accompany this document: > +- > http://events.linuxfoundation.org/sites/events/files/slides/lpc-2017-pvrdma-marcel-apfelbaum-yuval-shaia.pdf > + > + > + > +2. Setup > +======== > + > + > +2.1 Guest setup > +=============== > +Fedora 27+ kernels work out of the box, older distributions > +require updating the kernel to 4.14 to include the pvrdma driver. > + > +However the libpvrdma library needed by User Level Software is still > +not available as part of the distributions, so the rdma-core library > +needs to be compiled and optionally installed. > + > +Please follow the instructions at: > + https://github.com/linux-rdma/rdma-core.git > + > + > +2.2 Host Setup > +============== > +The pvrdma backend is an ibdevice interface that can be exposed > +either by a Soft-RoCE(rxe) device on machines with no RDMA device, > +or an HCA SRIOV function(VF/PF). > +Note that ibdevice interfaces can't be shared between pvrdma devices, > +each one requiring a separate instance (rxe or SRIOV VF). > + > + > +2.2.1 Soft-RoCE backend(rxe) > +=========================== > +A stable version of rxe is required, Fedora 27+ or a Linux > +Kernel 4.14+ is preferred. > + > +The rdma_rxe module is part of the Linux Kernel but not loaded by default. > +Install the User Level library (librxe) following the instructions from: > +https://github.com/SoftRoCE/rxe-dev/wiki/rxe-dev:-Home > + > +Associate an ETH interface with rxe by running: > + rxe_cfg add eth0 > +An rxe0 ibdevice interface will be created and can be used as pvrdma backend. > + > + > +2.2.2 RDMA device Virtual Function backend > +========================================== > +Nothing special is required, the pvrdma device can work not only with > +Ethernet Links, but also Infinibands Links. > +All is needed is an ibdevice with an active port, for Mellanox cards > +will be something like mlx5_6 which can be the backend. > + > + > +2.2.3 QEMU setup > +================ > +Configure QEMU with --enable-rdma flag, installing > +the required RDMA libraries. > + > + > +3. Usage > +======== > +Currently the device is working only with memory backed RAM > +and it must be mark as "shared": > + -m 1G \ > + -object memory-backend-ram,id=mb1,size=1G,share \ > + -numa node,memdev=mb1 \ > + > +The pvrdma device is composed of two functions: > + - Function 0 is a vmxnet Ethernet Device which is redundant in Guest > + but is required to pass the ibdevice GID using its MAC. > + Examples: > + For an rxe backend using eth0 interface it will use its mac: > + -device vmxnet3,addr=<slot>.0,multifunction=on,mac=<eth0 MAC> > + For an SRIOV VF, we take the Ethernet Interface exposed by it: > + -device vmxnet3,multifunction=on,mac=<RoCE eth MAC> > + - Function 1 is the actual device: > + -device > pvrdma,addr=<slot>.1,backend-dev=<ibdevice>,backend-gid-idx=<gid>,backend-port=<port> > + where the ibdevice can be rxe or RDMA VF (e.g. mlx5_4) > + Note: Pay special attention that the GID at backend-gid-idx matches > vmxnet's MAC. > + The rules of conversion are part of the RoCE spec, but since manual > conversion > + is not required, spotting problems is not hard: > + Example: GID: fe80:0000:0000:0000:7efe:90ff:fecb:743a > + MAC: 7c:fe:90:cb:74:3a > + Note the difference between the first byte of the MAC and the GID. > + > + > +4. Implementation details > +========================= > +The device acts like a proxy between the Guest Driver and the host > +ibdevice interface. > +On configuration path: > + - For every hardware resource request (PD/QP/CQ/...) the pvrdma will request > + a resource from the backend interface, maintaining a 1-1 mapping > + between the guest and host. > +On data path: > + - Every post_send/receive received from the guest will be converted into > + a post_send/receive for the backend. The buffers data will not be touched > + or copied resulting in near bare-metal performance for large enough > buffers. > + - Completions from the backend interface will result in completions for > + the pvrdma device.
Where's the host/guest interface documented? > + > + > +5. Limitations > +============== > +- The device obviously is limited by the Guest Linux Driver features > implementation > + of the VMware device API. > +- Memory registration mechanism requires mremap for every page in the buffer > in order > + to map it to a contiguous virtual address range. Since this is not the > data path > + it should not matter much. > +- QEMU cannot map guest RAM from a file descriptor if a pvrdma device is > attached, > + so it can't work with huge pages. The limitation will be addressed in the > future, > + however QEMU allocates Gust RAM with MADV_HUGEPAGE so if there are enough > huge > + pages available, QEMU will use them. > +- As previously stated, migration is not supported yet, however with some > hardware > + support can be done. > + > + > + > +6. Performance > +============== > +By design the pvrdma device exits on each post-send/receive, so for small > buffers > +the performance is affected; however for medium buffers it will became close > to > +bare metal and from 1MB buffers and up it reaches bare metal performance. > +(tested with 2 VMs, the pvrdma devices connected to 2 VFs of the same device) > + > +All the above assumes no memory registration is done on data path. > -- > 2.13.5
