Xillybus driver for generic FPGA interface
:Author: Eli Billauer, Xillybus Ltd. (http://xillybus.com)
:Email: email@example.com or as advertised on Xillybus' site.
-- Xillybus Overview
-- User interface
-- Seekable pipes
-- Source code organization
-- Pipe attributes
-- Host never reads from the FPGA
-- Channels, pipes, and the message channel
-- Data streaming
-- Data granularity
-- Buffer allocation
-- The "nonempty" message (supporting poll)
An FPGA (Field Programmable Gate Array) is a piece of logic hardware, which
can be programmed to become virtually anything that is usually found as a
dedicated chipset: For instance, a display adapter, network interface card,
or even a processor with its peripherals. FPGAs are the LEGO of hardware:
Based upon certain building blocks, you make your own toys the way you like
them. It's usually pointless to reimplement something that is already
available on the market as a chipset, so FPGAs are mostly used when some
special functionality is needed, and the production volume is relatively low
(hence not justifying the development of an ASIC).
The challenge with FPGAs is that everything is implemented at a very low
level, even lower than assembly language. In order to allow FPGA designers to
focus on their specific project, and not reinvent the wheel over and over
again, pre-designed building blocks, IP cores, are often used. These are the
FPGA parallels of library functions. IP cores may implement certain
mathematical functions, a functional unit (e.g. a USB interface), an entire
processor (e.g. ARM) or anything that might come handy. Think of them as a
building block, with electrical wires dangling on the sides for connection to
One of the daunting tasks in FPGA design is communicating with a fullblown
operating system (actually, with the processor running it): Implementing the
low-level bus protocol and the somewhat higher-level interface with the host
(registers, interrupts, DMA etc.) is a project in itself. When the FPGA's
function is a well-known one (e.g. a video adapter card, or a NIC), it can
make sense to design the FPGA's interface logic specifically for the project.
A special driver is then written to present the FPGA as a well-known interface
to the kernel and/or user space. In that case, there is no reason to treat the
FPGA differently than any device on the bus.
It's however common that the desired data communication doesn't fit any well-
known peripheral function. Also, the effort of designing an elegant
abstraction for the data exchange is often considered too big. In those cases,
a quicker and possibly less elegant solution is sought: The driver is
effectively written as a user space program, leaving the kernel space part
with just elementary data transport. This still requires designing some
interface logic for the FPGA, and write a simple ad-hoc driver for the kernel.
Xillybus is an IP core and a Linux driver. Together, they form a kit for
elementary data transport between an FPGA and the host, providing pipe-like
data streams with a straightforward user interface. It's intended as a low-
effort solution for mixed FPGA-host projects, for which it makes sense to
have the project-specific part of the driver running in a user-space program.
Since the communication requirements may vary significantly from one FPGA
project to another (the number of data pipes needed in each direction and
their attributes), there isn't one specific chunk of logic being the Xillybus
IP core. Rather, the IP core is configured and built based upon a
specification given by its end user.
Xillybus presents independent data streams, which resemble pipes or TCP/IP
communication to the user. At the host side, a character device file is used
just like any pipe file. On the FPGA side, hardware FIFOs are used to stream
the data. This is contrary to a common method of communicating through fixed-
sized buffers (even though such buffers are used by Xillybus under the hood).
There may be more than a hundred of these streams on a single IP core, but
also no more than one, depending on the configuration.
In order to ease the deployment of the Xillybus IP core, it contains a simple
data structure which completely defines the core's configuration. The Linux
driver fetches this data structure during its initialization process, and sets
up the DMA buffers and character devices accordingly. As a result, a single
driver is used to work out of the box with any Xillybus IP core.
The data structure just mentioned should not be confused with PCI's
configuration space or the Flattened Device Tree.
On the host, all interface with Xillybus is done through /dev/xillybus_*
device files, which are generated automatically as the drivers loads. The
names of these files depend on the IP core that is loaded in the FPGA (see
Probing below). To communicate with the FPGA, open the device file that
corresponds to the hardware FIFO you want to send data or receive data from,
and use plain write() or read() calls, just like with a regular pipe. In
particular, it makes perfect sense to go::
$ cat mydata > /dev/xillybus_thisfifo
$ cat /dev/xillybus_thatfifo > hisdata
possibly pressing CTRL-C as some stage, even though the xillybus_* pipes have
the capability to send an EOF (but may not use it).
The driver and hardware are designed to behave sensibly as pipes, including:
* Supporting non-blocking I/O (by setting O_NONBLOCK on open() ).
* Supporting poll() and select().
* Being bandwidth efficient under load (using DMA) but also handle small
pieces of data sent across (like TCP/IP) by autoflushing.
A device file can be read only, write only or bidirectional. Bidirectional
device files are treated like two independent pipes (except for sharing a
"channel" structure in the implementation code).
Xillybus pipes are configured (on the IP core) to be either synchronous or
asynchronous. For a synchronous pipe, write() returns successfully only after
some data has been submitted and acknowledged by the FPGA. This slows down
bulk data transfers, and is nearly impossible for use with streams that
require data at a constant rate: There is no data transmitted to the FPGA
between write() calls, in particular when the process loses the CPU.
When a pipe is configured asynchronous, write() returns if there was enough
room in the buffers to store any of the data in the buffers.
For FPGA to host pipes, asynchronous pipes allow data transfer from the FPGA
as soon as the respective device file is opened, regardless of if the data
has been requested by a read() call. On synchronous pipes, only the amount
of data requested by a read() call is transmitted.
In summary, for synchronous pipes, data between the host and FPGA is
transmitted only to satisfy the read() or write() call currently handled
by the driver, and those calls wait for the transmission to complete before
Note that the synchronization attribute has nothing to do with the possibility
that read() or write() completes less bytes than requested. There is a
separate configuration flag ("allowpartial") that determines whether such a
partial completion is allowed.
A synchronous pipe can be configured to have the stream's position exposed
to the user logic at the FPGA. Such a pipe is also seekable on the host API.
With this feature, a memory or register interface can be attached on the
FPGA side to the seekable stream. Reading or writing to a certain address in
the attached memory is done by seeking to the desired address, and calling
read() or write() as required.
Source code organization
The Xillybus driver consists of a core module, xillybus_core.c, and modules
that depend on the specific bus interface (xillybus_of.c and xillybus_pcie.c).
The bus specific modules are those probed when a suitable device is found by
the kernel. Since the DMA mapping and synchronization functions, which are bus
dependent by their nature, are used by the core module, a
xilly_endpoint_hardware structure is passed to the core module on
initialization. This structure is populated with pointers to wrapper functions
which execute the DMA-related operations on the bus.
Each pipe has a number of attributes which are set when the FPGA component
(IP core) is built. They are fetched from the IDT (the data structure which
defines the core's configuration, see Probing below) by xilly_setupchannels()
in xillybus_core.c as follows:
* is_writebuf: The pipe's direction. A non-zero value means it's an FPGA to
host pipe (the FPGA "writes").
* channelnum: The pipe's identification number in communication between the
host and FPGA.
* format: The underlying data width. See Data Granularity below.
* allowpartial: A non-zero value means that a read() or write() (whichever
applies) may return with less than the requested number of bytes. The common
choice is a non-zero value, to match standard UNIX behavior.
* synchronous: A non-zero value means that the pipe is synchronous. See
* bufsize: Each DMA buffer's size. Always a power of two.
* bufnum: The number of buffers allocated for this pipe. Always a power of two.
* exclusive_open: A non-zero value forces exclusive opening of the associated
device file. If the device file is bidirectional, and already opened only in
one direction, the opposite direction may be opened once.
* seekable: A non-zero value indicates that the pipe is seekable. See
Seekable pipes above.
* supports_nonempty: A non-zero value (which is typical) indicates that the
hardware will send the messages that are necessary to support select() and
poll() for this pipe.
Host never reads from the FPGA
Even though PCI Express is hotpluggable in general, a typical motherboard
doesn't expect a card to go away all of the sudden. But since the PCIe card
is based upon reprogrammable logic, a sudden disappearance from the bus is
quite likely as a result of an accidental reprogramming of the FPGA while the
host is up. In practice, nothing happens immediately in such a situation. But
if the host attempts to read from an address that is mapped to the PCI Express
device, that leads to an immediate freeze of the system on some motherboards,
even though the PCIe standard requires a graceful recovery.
In order to avoid these freezes, the Xillybus driver refrains completely from
reading from the device's register space. All communication from the FPGA to
the host is done through DMA. In particular, the Interrupt Service Routine
doesn't follow the common practice of checking a status register when it's
invoked. Rather, the FPGA prepares a small buffer which contains short
messages, which inform the host what the interrupt was about.
This mechanism is used on non-PCIe buses as well for the sake of uniformity.
Channels, pipes, and the message channel
Each of the (possibly bidirectional) pipes presented to the user is allocated
a data channel between the FPGA and the host. The distinction between channels
and pipes is necessary only because of channel 0, which is used for interrupt-
related messages from the FPGA, and has no pipe attached to it.
Even though a non-segmented data stream is presented to the user at both
sides, the implementation relies on a set of DMA buffers which is allocated
for each channel. For the sake of illustration, let's take the FPGA to host
direction: As data streams into the respective channel's interface in the
FPGA, the Xillybus IP core writes it to one of the DMA buffers. When the
buffer is full, the FPGA informs the host about that (appending a
XILLYMSG_OPCODE_RELEASEBUF message channel 0 and sending an interrupt if
necessary). The host responds by making the data available for reading through
the character device. When all data has been read, the host writes on the
the FPGA's buffer control register, allowing the buffer's overwriting. Flow
control mechanisms exist on both sides to prevent underflows and overflows.
This is not good enough for creating a TCP/IP-like stream: If the data flow
stops momentarily before a DMA buffer is filled, the intuitive expectation is
that the partial data in buffer will arrive anyhow, despite the buffer not
being completed. This is implemented by adding a field in the
XILLYMSG_OPCODE_RELEASEBUF message, through which the FPGA informs not just
which buffer is submitted, but how much data it contains.
But the FPGA will submit a partially filled buffer only if directed to do so
by the host. This situation occurs when the read() method has been blocking
for XILLY_RX_TIMEOUT jiffies (currently 10 ms), after which the host commands
the FPGA to submit a DMA buffer as soon as it can. This timeout mechanism
balances between bus bandwidth efficiency (preventing a lot of partially
filled buffers being sent) and a latency held fairly low for tails of data.
A similar setting is used in the host to FPGA direction. The handling of
partial DMA buffers is somewhat different, though. The user can tell the
driver to submit all data it has in the buffers to the FPGA, by issuing a
write() with the byte count set to zero. This is similar to a flush request,
but it doesn't block. There is also an autoflushing mechanism, which triggers
an equivalent flush roughly XILLY_RX_TIMEOUT jiffies after the last write().
This allows the user to be oblivious about the underlying buffering mechanism
and yet enjoy a stream-like interface.
Note that the issue of partial buffer flushing is irrelevant for pipes having
the "synchronous" attribute nonzero, since synchronous pipes don't allow data
to lay around in the DMA buffers between read() and write() anyhow.
The data arrives or is sent at the FPGA as 8, 16 or 32 bit wide words, as
configured by the "format" attribute. Whenever possible, the driver attempts
to hide this when the pipe is accessed differently from its natural alignment.
For example, reading single bytes from a pipe with 32 bit granularity works
with no issues. Writing single bytes to pipes with 16 or 32 bit granularity
will also work, but the driver can't send partially completed words to the
FPGA, so the transmission of up to one word may be held until it's fully
occupied with user data.
This somewhat complicates the handling of host to FPGA streams, because
when a buffer is flushed, it may contain up to 3 bytes don't form a word in
the FPGA, and hence can't be sent. To prevent loss of data, these leftover
bytes need to be moved to the next buffer. The parts in xillybus_core.c
that mention "leftovers" in some way are related to this complication.
As mentioned earlier, the number of pipes that are created when the driver
loads and their attributes depend on the Xillybus IP core in the FPGA. During
the driver's initialization, a blob containing configuration info, the
Interface Description Table (IDT), is sent from the FPGA to the host. The
bootstrap process is done in three phases:
1. Acquire the length of the IDT, so a buffer can be allocated for it. This
is done by sending a quiesce command to the device, since the acknowledge
for this command contains the IDT's buffer length.
2. Acquire the IDT itself.
3. Create the interfaces according to the IDT.
In order to simplify the logic that prevents illegal boundary crossings of
PCIe packets, the following rule applies: If a buffer is smaller than 4kB,
it must not cross a 4kB boundary. Otherwise, it must be 4kB aligned. The
xilly_setupchannels() functions allocates these buffers by requesting whole
pages from the kernel, and diving them into DMA buffers as necessary. Since
all buffers' sizes are powers of two, it's possible to pack any set of such
buffers, with a maximal waste of one page of memory.
All buffers are allocated when the driver is loaded. This is necessary,
since large continuous physical memory segments are sometimes requested,
which are more likely to be available when the system is freshly booted.
The allocation of buffer memory takes place in the same order they appear in
the IDT. The driver relies on a rule that the pipes are sorted with decreasing
buffer size in the IDT. If a requested buffer is larger or equal to a page,
the necessary number of pages is requested from the kernel, and these are
used for this buffer. If the requested buffer is smaller than a page, one
single page is requested from the kernel, and that page is partially used.
Or, if there already is a partially used page at hand, the buffer is packed
into that page. It can be shown that all pages requested from the kernel
(except possibly for the last) are 100% utilized this way.
The "nonempty" message (supporting poll)
In order to support the "poll" method (and hence select() ), there is a small
catch regarding the FPGA to host direction: The FPGA may have filled a DMA
buffer with some data, but not submitted that buffer. If the host waited for
the buffer's submission by the FPGA, there would be a possibility that the
FPGA side has sent data, but a select() call would still block, because the
host has not received any notification about this. This is solved with
XILLYMSG_OPCODE_NONEMPTY messages sent by the FPGA when a channel goes from
completely empty to containing some data.
These messages are used only to support poll() and select(). The IP core can
be configured not to send them for a slight reduction of bandwidth.