A conceptually simpler way to do this is to assign one (or more) extra bits in the head an tail pointers. For example, for a 512 entry ring, use 16-bit indices. Whenever you index the ring, and the index with 511 before performing the index.
Write to ring: ring[511&(head++)] = data
Read from ring: data = ring[511&(tail++)]
Ring is empty: head == tail
Ring is full: tail + 512 == head
This post is literally about saving one byte, at the cost of being slower and not producer-consumer friendly. Not very interesting.
The flag could have been hidden in any other field as a bit or something. Then it could be at least masked with simple AND operation which is usually faster than branching, especially on pipelined CPUs.
This fix for concurrency issue is an ugly hack. I'm not sure if it's even correct in this particular scenario, and definitely not proper for anything that would aspire to be good reusable code. I'd advise this code to push atomicity requirement onto caller. Irqs should have been disabled by calling code.
"register" keyword is obsolete. There's no point in using it.
No, the author can use all entries in the buffer. The space cost is hidden in that one of the pointers need to hold X+1 values, ie for 256 entries it is not enough with 8-bit indices.
An unfortunate effect of this implementation is that both indices are modified by the consumer. It's not safe to write/read dats from different contexts, something that would be very useful in a driver.
If I understand correctly he's storing X items in a buffer of size X, but using an invalid value in the head pointer to indicate that the buffer is full.
It's easier to understand without the decrementing. Initialize head and tail to 0, keep incrementing and cycling through the end as usual. When removing an element check if head and tail are equal before removing. When adding an element check if head and tail are equal after adding, and set the head to a sentinel value, say MAX_INT. Now you can check for the sentinel before attempting to add an element.
[+] [-] jhallenworld|10 years ago|reply
[+] [-] dpc_pw|10 years ago|reply
The flag could have been hidden in any other field as a bit or something. Then it could be at least masked with simple AND operation which is usually faster than branching, especially on pipelined CPUs.
Update: Quick implementation: https://gist.github.com/dpc/a194b7784adfa150a450
This fix for concurrency issue is an ugly hack. I'm not sure if it's even correct in this particular scenario, and definitely not proper for anything that would aspire to be good reusable code. I'd advise this code to push atomicity requirement onto caller. Irqs should have been disabled by calling code.
"register" keyword is obsolete. There's no point in using it.
[+] [-] kstenerud|10 years ago|reply
[+] [-] rhaps0dy|10 years ago|reply
[+] [-] mungoman2|10 years ago|reply
An unfortunate effect of this implementation is that both indices are modified by the consumer. It's not safe to write/read dats from different contexts, something that would be very useful in a driver.
[+] [-] akkartik|10 years ago|reply
It's easier to understand without the decrementing. Initialize head and tail to 0, keep incrementing and cycling through the end as usual. When removing an element check if head and tail are equal before removing. When adding an element check if head and tail are equal after adding, and set the head to a sentinel value, say MAX_INT. Now you can check for the sentinel before attempting to add an element.
Neat trick!
[+] [-] cpr|10 years ago|reply
[+] [-] jhallenworld|10 years ago|reply
[+] [-] jonsen|10 years ago|reply