How can sk_buff (SKB) have more than one user?

Question

I have come across that skb->users holds the numbers of entities using this particular buffer.

But Since SKB's belong to a particular socket and corresponds to a particular process, How can it have more than one user?

Answer 1

sk_buffs typically represent network packets. They can be generated as a result of a write to a particular socket by a particular process but they can also come into existence in other ways.

When a packet is received from a network device, for example, it is placed in an sk_buff. As it is passed up the stack, it may be processed by various modules and layers (IP, ARP, UDP/TCP, as well as things like "tap" devices). Some of those modules may need to do something with the packet that is not instantaneous (example: retransmit on a tap device). And of course, data from the packet might also be delivered to a process reading from a socket.

A component needing to process the packet data could of course simply create a copy of the packet buffer, but if no other component will be modifying it, that would be wasteful. So, incrementing the users count allows the module to put a "hold" on the packet so that it won't be freed -- and hence the underlying memory won't get re-used for some other purpose -- as long as the module still needs it. As the users of the sk_buff finish using it one-by-one, they decrement the users count and when it becomes zero then the memory is released.

Answer 2

Normally, in the case of single core machine, you only can have one "user" using the network buffer at any time, this "user" is referring to the different parts or functions in the linux networking infrastructure. they cannot be concurrently running at the same time, so this "user" reference count is not useful and I guessed not used.

But in SMP machine, where different CPU can be running the different parts of the kernel all at the same time, and concurrent reading the network buffer simultaneously, this reference count becomes important. You would not want to modify it when it is still in use but another part of the system, or free it either. But for read-only operation, it is possible to be concurrent, and so this reference count served the purpose of counting the number of "user" (or areas in the kernel), that uses the network buffer.