4. IFPGA Rawdev Driver
FPGA is used more and more widely in Cloud and NFV, one primary reason is that FPGA not only provides ASIC performance but also it’s more flexible than ASIC.
FPGA uses Partial Reconfigure (PR) Parts of Bit Stream to achieve its flexibility. That means one FPGA Device Bit Stream is divided into many Parts of Bit Stream(each Part of Bit Stream is defined as AFU-Accelerated Function Unit), and each AFU is a hardware acceleration unit which can be dynamically reloaded respectively.
By PR (Partial Reconfiguration) AFUs, one FPGA resources can be time-shared by different users. FPGA hot upgrade and fault tolerance can be provided easily.
The SW IFPGA Rawdev Driver (ifpga_rawdev) provides a Rawdev driver that utilizes Intel FPGA Software Stack OPAE(Open Programmable Acceleration Engine) for FPGA management.
4.1. Implementation details
Each instance of IFPGA Rawdev Driver is probed by Intel FpgaDev. In coordination with OPAE share code IFPGA Rawdev Driver provides common FPGA management ops for FPGA operation, OPAE provides all following operations: - FPGA PR (Partial Reconfiguration) management - FPGA AFUs Identifying - FPGA Thermal Management - FPGA Power Management - FPGA Performance reporting - FPGA Remote Debug
All configuration parameters are taken by vdev_ifpga_cfg driver. Besides configuration, vdev_ifpga_cfg driver also hot plugs in IFPGA Bus.
All of the AFUs of one FPGA may share same PCI BDF and AFUs scan depend on IFPGA Rawdev Driver so IFPGA Bus takes AFU device scan and AFU drivers probe. All AFU device driver bind to AFU device by its UUID (Universally Unique Identifier).
To avoid unnecessary code duplication and ensure maximum performance, handling of AFU devices is left to different PMDs; all the design as summarized by the following block diagram:
+---------------------------------------------------------------+
| Application(s) |
+----------------------------.----------------------------------+
|
|
+----------------------------'----------------------------------+
| DPDK Framework (APIs) |
+----------|------------|--------.---------------------|--------+
/ \ |
/ \ |
+-------'-------+ +-------'-------+ +--------'--------+
| Eth PMD | | Crypto PMD | | |
+-------.-------+ +-------.-------+ | |
| | | |
| | | |
+-------'-------+ +-------'-------+ | IFPGA |
| Eth AFU Dev | |Crypto AFU Dev | | Rawdev Driver |
+-------.-------+ +-------.-------+ |(OPAE Share Code)|
| | | |
| | Rawdev | |
+-------'------------------'-------+ Ops | |
| IFPGA Bus | -------->| |
+-----------------.----------------+ +--------.--------+
| |
Hot-plugin -->| |
| |
+-----------------'------------------+ +--------'--------+
| vdev_ifpga_cfg driver | | Intel FpgaDev |
+------------------------------------+ +-----------------+
4.2. Run-time parameters
This driver is invoked automatically in systems added with Intel FPGA,
but PR and IFPGA Bus scan is triggered by command line using
--vdev 'ifpga_rawdev_cfg EAL option.
The following device parameters are supported:
ifpga[string]Provide a specific Intel FPGA device PCI BDF. Can be provided multiple times for additional instances.
port[int]Each FPGA can provide many channels to PR AFU by software, each channels is identified by this parameter.
afu_bts[string]If null, the AFU Bit Stream has been PR in FPGA, if not forces PR and identifies AFU Bit Stream file.
5. IFPGA AFU Driver
AFU (Acceleration Function Unit) is a function or set of functions that perform various acceleration task on FPGA platform. The image of AFU is called as GBS (Green Bit Stream) which can be used by PR (Partial Reconfigure) tool to load into the FPGA, different AFUs can be dynamically reloaded respectively.
AFU has two main communication paths between the host:
FPGA to host transactions
The FPGA accesses host memory using a 512 bits data path. This data path has separate channels for read and write traffic allowing for simultaneous read and write to occur. The read and write channels support bursts of 1, 2, and 4 cache lines.
Host to FPGA (MMIO) transactions
The host can access a 256 KB address space within the FPGA. This address space contains Device Feature Header (DFHs) and the control and status registers of the AFU hardware.
AFU must implement the following registers:
AFU DFH - a 64-bit header at MMIO address offset 0x0
AFU ID - a 128-bit UUID at MMIO address offset 0x2
The AFU is enumerated and recorded by IFPGA Rawdev Driver. Then AFU devices are created with the help of IFPGA Bus Driver, AFU driver probe these AFU devices and expose them as standard raw devices for application to access.
5.1. Implementation details
IFPGA Rawdev Driver identifies AFU in FPGA, AFU location (PF/VF address)
and UUID are taken by ifpga_rawdev_cfg vdev driver
which hot plug AFU into IFPGA Bus.
IFPGA Bus takes AFU device scan and AFU driver probe. All AFU device driver bind to AFU device by its dedicated UUID. To avoid unnecessary code duplication and ensure maximum performance, AFU driver implements the common part of raw device driver. Several specific AFU drivers are provided for reference. The design is summarized by the following block diagram:
+---------------------------------------------------------------+
| Application(s) |
+----------------------------.----------------------------------+
|
+----------------------------'----------------------------------+
| DPDK Framework (Rawdev APIs) |
+-----------------+------------------------------------+--------+
| |
+-----------------'----------------+ |
| IFPGA AFU Driver | +--------'--------+
| | | |
|+---------------+ +--------------+| | |
|| AFU Dev1 PMD | | AFU Dev2 PMD || | |
|+-------+-------+ +-------+------+| | |
+--------|-----------------|-------+ | |
| | | |
+--------'------+ +-------'-------+ | IFPGA |
| AFU Dev1 | | AFU Dev2 | | Rawdev Driver |
+-------.-------+ +-------.-------+ | |
| | Rawdev | |
+-------'------------------'-------+ Ops | |
| IFPGA Bus |--------->| |
+-----------------.----------------+ +--------.--------+
| |
Hot-plugin -->| |
| |
+-----------------'------------------+ +--------'--------+
| ifpga_rawdev_cfg vdev driver | | Intel FpgaDev |
+------------------------------------+ +-----------------+
5.2. How to test AFU function
Suppose AFU is found in FPGA at PCI address 31:00.0, then you can create and test a AFU device by following steps in application.
rte_vdev_init(“ifpga_rawdev_cfg0”, “ifpga=31:00.0,port=0”)
rawdev = rte_rawdev_pmd_get_named_dev(“afu_0|31:00.0”)
rte_rawdev_configure(rawdev->dev_id, &cfg, sizeof(cfg))
rte_rawdev_selftest(rawdev->dev_id)
rte_vdev_uninit(“ifpga_rawdev_cfg0”)
AFU device name format used in rte_rawdev_pmd_get_named_dev is afu_[port]|[BDF].
Please refer to OPAE documentation for the meaning of port.
Each AFU device has specific configuration data, they are defined in rte_pmd_afu.h.
6. Open FPGA Stack
Open FPGA Stack (OFS) is a collection of RTL and open source software providing interfaces to access the instantiated RTL easily in an FPGA. OFS leverages the DFL for the implementation of the FPGA RTL design.
OFS designs allow for the arrangement of software interfaces across multiple PCIe endpoints. Some of these interfaces may be PFs defined in the static region that connect to interfaces in an IP that is loaded via Partial Reconfiguration (PR). And some of these interfaces may be VFs defined in the PR region that can be reconfigured by the end-user. Furthermore, these PFs/VFs may use DFLs such that features may be discovered and accessed in user space with the aid of a generic kernel driver like vfio-pci. The diagram below depicts an example design with one PF and two VFs. In this example, it will export the management functions via PF0 and acceleration functions via VF0 and VF1, leverage VFIO to export the MMIO space to an application.:
+-----------------+ +-------------+ +------------+
| FPGA Management | | DPDK App | | User App |
| App | | | | |
+--------+--------+ +------+------+ +-----+------+
| | |
+--------+--------+ +------+------+ |
| IFPGA PMD | | AFU PMD | |
+--------+--------+ +------+------+ |
| | |
+--------+------------------+---------------+------+
| VFIO-PCI |
+--------+------------------+---------------+------+
| | |
+--------+--------+ +------+------+ +-----+------+
| PF0 | | PF0_VF0 | | PF0_VF1 |
+-----------------+ +-------------+ +------------+
As accelerators are specialized hardware, they are typically limited in the number installed in a given system. Many use cases require them to be shared across multiple software contexts or threads of software execution, either through partitioning of individual dedicated resources, or virtualization of shared resources. OFS provides several models to share the AFU resources via PR mechanism and hardware-based virtualization schemes.
Legacy model. With legacy model FPGA cards like Intel PAC N3000 or N5000, there is a notion that the boundary between the AFU and the shell is also the unit of PR for those FPGA platforms. This model is only able to handle a single context, because it only has one PR engine, and one PR region which has an associated Port device.
Multiple VFs per PR slot. In this model, available AFU resources may allow instantiation of many VFs which have a dedicated PCIe function with their own dedicated MMIO space, or partition a region of MMIO space on a single PCIe function. Intel PAC N6000 card has implemented this model. In this model, the AFU/PR slot was not connected to port device. For DFL’s view, the Next_AFU pointer in FIU feature header of port device points to NULL in this model. On the other hand, each VF can start with an AFU feature header without being connected to a FIU Port feature header.
The VFs are created through the Linux kernel driver before we use them in DPDK.
OFS provides the diversity for accessing the AFU resource to RTL developer. An IP designer may choose to add more than one PF for interfacing with IP on the FPGA and choose different model to access the AFU resource.
There is one reference architecture design using the “Multiple VFs per PR slot” model for OFS as illustrated below. In this reference design, it exports the FPGA management functions via PF0. PF1 will bind with DPDK virtio driver presenting itself as a network interface to the application. PF2 will bind to the vfio-pci driver allowing the user space software to discover and interface with the specific workload like diagnostic test. It leverages AFU PMD driver to access the AFU resources in DPDK.:
+----------------------+
| PF/VF mux/demux |
+--+--+-----+------+-+-+
| | | | |
+------------------------+ | | | |
PF0 | +---------+ +-+ | |
+---+---+ | +---+----+ | |
| DFH | | | DFH | | |
+-------+ +-----+----+ +--------+ | |
| FME | | VirtIO | | Test | | |
+---+---+ +----------+ +--------+ | |
| PF1 PF2 | |
| | |
| +----------+ |
| | ++
| | |
| | PF0_VF0 | PF0_VF1
| +-----------------+-----------+------------+
| | +-----+-----------+--------+ |
| | | | | | |
| | +------+ | +--+ -+ +--+---+ | |
| | | Port | | | DFH | | DFH | | |
+-----------+ +------+ | +-----+ +------+ | |
| | | DEV | | DEV | | |
| | +-----+ +------+ | |
| | PR Slot | |
| +--------------------------+ |
| Port Gasket |
+------------------------------------------+