Comments on Linux Unleashed: Raw socket, Packet socket and Zero copy networking in Linux

Thanks alot Yusuf, really the expalnation are very...

2018-02-02T10:46:39.462-08:00

Thanks alot Yusuf, really the expalnation are very crystal clear

Hey Yusuf, Thanks for this amazing article. Helped...

2016-11-08T03:53:32.328-08:00

Hey Yusuf, Thanks for this amazing article. Helped me understand a lot of stuff.

I have worked on PF_RING for quite sometime as par...

2016-03-22T11:53:40.717-07:00

I have worked on PF_RING for quite sometime as part of network monitoring solutions. At the time, was not aware that AF_PACKET existed for the same purpose that almost seems to provide the same functionality as PF_RING.

There are a couple of doubts however:
1. Are there any test results to prove which one is better in performance ? Assume PACKET_MMAP is enabled
2. To use PACKET_MMAP, kernel has to be compiled with this flag. PF_RING does not require a kernel compile, just a kernel module compile. That makes it's more easy to use than AF_PACKET. Any drawback though ?
3. Setting BPF filters in PF_RING was simple, just enter the string containing the BPF expression and call the API. In AF_PACKET seems more complex to do

I want zero copy code any one have it ??

2016-03-08T01:40:34.242-08:00

I want zero copy code any one have it ??

Nice post and really informative. Is there any way...

2014-03-07T11:17:32.092-08:00

Nice post and really informative. Is there any way to filter pf_packets? Perhaps ebtables or is it a fact that packet socket cant be filtered at all?

What's about joins to specific multicast group...

2013-08-15T13:14:18.990-07:00

What's about joins to specific multicast groups?
I have tried this (setsockopt() succeeds) but no joins established.

Can you share your experience?

Thanks in advance.
Regards

hi this is a nice article...i am in need of help i...

2013-07-21T09:51:11.507-07:00

hi this is a nice article...i am in need of help in designing a zero copy ftp client and server using C in linux

Excellent article. Thanks.

2013-07-04T00:45:09.265-07:00

Excellent article. Thanks.

Great article! Thanks for posting it... However, ...

2013-06-04T00:46:43.178-07:00

Great article! Thanks for posting it...

However, I have a big issue when trying to implement a zero-copy socket following the instructions that you provide. The problem comes when I try to call twice to "setsockopt()" using the same socket, in order to get, with the first call, access to the TX_RING and, with the second one, access to the RX_RING (or viceversa). The first call to setsockopt() always works fine (either for accessing the TX_RING or the RX_RING), but the second call to setsockopt() always returns an error (-16 in accordance with perror(), i.e., EBUSY kernel signal). I could access both RINGS using TWO DIFFERENT SOCKETS, one for TX_RING and a different one for RX_RING.

Is this the normal approach or am I doing anything wrong?

I m trying to implement a very light weight protoc...

2011-02-18T07:02:20.771-08:00

I m trying to implement a very light weight protocol for transferring files between nodes in a cluster system.
I m using Zero copy concept as described above. The problem is that i m facing packet losses. I just want to add reliability to it without using any retransmission scheme. Is it Possible??

I have tried to map a larger buffer on receiver side to prevent packet loss..
Please HELP!!!!

Thanks

Steven sorry for very late reply, Yes PF_RING used...

2011-02-02T02:30:45.401-08:00

Steven sorry for very late reply, Yes PF_RING used for so called Zero Copy :-)

Great Article! In reference to Zero Copy, is this...

2010-12-23T17:47:42.251-08:00

Great Article!

In reference to Zero Copy, is this what is used by PF_RING?

HI Jeff, Thanks for pointing this out, Yes you ar...

2010-12-06T08:18:26.022-08:00

HI Jeff,

Thanks for pointing this out, Yes you are right inside the kernel its not zero copy and even if the packet can be dma directly to skb, still it requires 1 copy to pkt_rx_ring, but i wrote this article from user point of view ,for users the most time consuming copy is from the kernel to user memory and it is the major bottleneck for real-time protocols at application layer, With this feature we avoid this most expensive copy and that's the reason i call it zero copy.

Regarding zero-copy, it's not. PF_PACKET is d...

2010-12-04T18:59:20.894-08:00

Regarding zero-copy, it's not.

PF_PACKET is designed for more efficient bandwidth handling in network sniffers by removing the need for vigorous read/write system calls (as you noted). It works by installing an *additional* protocol handler on target network devices, meaning data packets will also be delivered to any other existing protocol handler (ETH_P_IP aka PF_INET) sockets in other applications (security issue). Furthermore, the NIC driver is unaware of the PF_PACKET framework and the kernel doesn't abstract skb memory allocation enough that the driver can DMA data directly into the PACKET_RX_RING. The NIC just allocates a random skb (and possibly DMAs data to it to avoid a copy), then passes the skb to the protocol handlers. If you review the PF_PACKET protocol handler implementation when a PACKET_RX_RING is installed, specifically net/packet/af_packet.c:tpacket_rcv(), you will notice that the protocol handler uses skb_copy_bits() to *copy* the data from the NIC allocated skb to the PACKET_RX_RING.

In order to achieve true zero copy using PF_PACKET, the NIC driver would need to be able to allocate skb's where the data segment was part of the PACKET_RX_RING naturally. It's probably not worth the effort to use DMA to copy an *already* allocated skb to the ring because it would inject latency into a framework designed for efficient sniffing, doing more harm than good.