.Fx kernel includes a facility for eliminating data copies on socket reads and writes.
This code is collectively known as the zero copy sockets code, because during normal network I/O, data will not be copied by the CPU at all. Rather it will be DMAed from the users buffer to the NIC (for sends), or DMAed from the NIC to a buffer that will then be given to the user (receives).
The zero copy sockets code uses the standard socket read and write semantics, and therefore has some limitations and restrictions that programmers should be aware of when trying to take advantage of this functionality.
For sending data, there are no special requirements or capabilities that the sending NIC must have. The data written to the socket, though, must be at least a page in size and page aligned in order to be mapped into the kernel. If it does not meet the page size and alignment constraints, it will be copied into the kernel, as is normally the case with socket I/O.
The user should be careful not to overwrite buffers that have been written to the socket before the data has been freed by the kernel, and the copy-on-write mapping cleared. If a buffer is overwritten before it has been given up by the kernel, the data will be copied, and no savings in CPU utilization and memory bandwidth utilization will be realized.
The socket(2) API does not really give the user any indication of when his data has actually been sent over the wire, or when the data has been freed from kernel buffers. For protocols like TCP, the data will be kept around in the kernel until it has been acknowledged by the other side; it must be kept until the acknowledgement is received in case retransmission is required.
From an application standpoint, the best way to guarantee that the data has been sent out over the wire and freed by the kernel (for TCP-based sockets) is to set a socket buffer size (see the SO_SNDBUF socket option in the setsockopt(2) manual page) appropriate for the application and network environment and then make sure you have sent out twice as much data as the socket buffer size before reusing a buffer. For TCP, the send and receive socket buffer sizes generally directly correspond to the TCP window size.
For receiving data, in order to take advantage of the zero copy receive code, the user must have a NIC that is configured for an MTU greater than the architecture page size. (E.g., for alpha this would be 8KB, for i386, it would be 4KB.) Additionally, in order for zero copy receive to work, packet payloads must be at least a page in size and page aligned.
Achieving page aligned payloads requires a NIC that can split an incoming packet into multiple buffers. It also generally requires some sort of intelligence on the NIC to make sure that the payload starts in its own buffer. This is called ""header splitting"". Currently the only NICs with support for header splitting are Alteon Tigon 2 based boards running slightly modified firmware. The
.Fx ti(4) driver includes modified firmware for Tigon 2 boards only. Header splitting code can be written, however, for any NIC that allows putting received packets into multiple buffers and that has enough programmability to determine that the header should go into one buffer and the payload into another.
You can also do a form of header splitting that does not require any NIC modifications if your NIC is at least capable of splitting packets into multiple buffers. This requires that you optimize the NIC driver for your most common packet header size. If that size (ethernet + IP + TCP headers) is generally 66 bytes, for instance, you would set the first buffer in a set for a particular packet to be 66 bytes long, and then subsequent buffers would be a page in size. For packets that have headers that are exactly 66 bytes long, your payload will be page aligned.
The other requirement for zero copy receive to work is that the buffer that is the destination for the data read from a socket must be at least a page in size and page aligned.
Obviously the requirements for receive side zero copy are impossible to meet without NIC hardware that is programmable enough to do header splitting of some sort. Since most NICs are not that programmable, or their manufacturers will not share the source code to their firmware, this approach to zero copy receive is not widely useful.
There are other approaches, such as RDMA and TCP Offload, that may potentially help alleviate the CPU overhead associated with copying data out of the kernel. Most known techniques require some sort of support at the NIC level to work, and describing such techniques is beyond the scope of this manual page.
The zero copy send and zero copy receive code can be individually turned off via the kern.ipc.zero_copy.send and kern.ipc.zero_copy.receive sysctl variables respectively.
sendfile(2), socket(2), ti(4)