48-bit sequence arithmetic: proposed changes
Below is a write-up of some issues that I found with the sequence number arithmetic. This is mainly for the record.
The most important thing is that there are currently two problems:
1) before48(a, b) can not decide if the difference between a and b is 2^47
2) unsigned sequence number arithmetic
causes problems/headaches/curses. For instance, there are tests such as
if( delta_seqno(a, b) > NUM_DUPACKS)
in the code. What happens if a and b are reordered (e.g. (5, 3) instead of (3, 5))? When using unsigned difference, we get the two's complement and that is a large number. With signed arithmetic, the former case gives a negative result and we can then decide whether we want to be bothered with the reordering or not.
Again, the text below is for the record - I hope to be back with a patch to implement this before the end of the week.
Gerrit
---
Modulo-2^48 Sequence Number Operations in DCCP ----------------------------------------------
This text is a write-up of issues regarding sequence number comparisons and operations when 48 bits are wanted, but 64 are given. A good illustration is given in section 24.7 of Stevens, vol 2. In particular, this text defines
* ordering (the before' relation which implies the after' relation)
* signed difference/delta of sequence numbers
* derived relations
Ordering n-bit numbers in modulo 2^n arithmetic
-----------------------------------------------
The n-bits sequence numbers start at 0 and wrap around from 2^n-1 back to 0. For two n-bit sequence numbers, a and b, we want to determine whether a is before' b. This is a strict ordering, i.e. for all a != b exactly one of a before' b or b `before' must hold. When this condition is fulfilled, we obtain an implied relation:
a before' b <=> b after' a
(a) Assume first that a = 0
Then a `before' b for 1 <= b <= 2^n-1
But, since sequence numbers wrap, we also have that
2^(n-1)+1 ... 2^n-1 are `before' a = 0
To disambiguate, we define
* 0 is `before' 1 ... 2^(n-1)-1
* 2^(n-1) ... 2^n-1 are `before' 0
(b) The general case: if a > 0 then the above just cycles around. The following table shows this.
a | a `before' b | b `before' a
---+--------------------------+--------------------------------------------
0 | 1 ... (2^(n-1) - 1) | 2^(n-1) ... (2^n - 1)
1 | 2 ... 2^(n-1) | (2^(n-1) + 1) ... (2^n - 1) ... 0
2 | 3 ... (2^(n-1) + 1) | (2^(n-1) + 2) ... (2^n - 1) ... 0 ... 1
... ... ..... ... ... ...
k | k+1 ... (2^(n-1) -1 + k) | (2^(n-1) + k) ... (2^n - 1 + k)
---+--------------------------+--------------------------------------------
(0 < k <= 2^n-1, all additions and subtractions are modulo-2^n)(c) Using the difference instead
With differences we are independent of the value of k:
* a is `before' a number b = a+1 ... (2^(n-1) - 1 + a) whenever (b - a) = 1 ... (2^(n-1) - 1)
* this implies that the test "c `before' a" fails for all numbers c which are such that 2^(n-1) + a <= c <= 2^n - 1 + a
* using two's complement, we get another, equivalent condition:
a `before' b <=> 1 <= b-a <= 2^(n-1) - 1
<=> -2^(n-1) + 1 <= a-b <= -1
This is a necessary condition, and it also shows that testing whether "a-b < 0" is not sufficient for "a `before' b" since -2^(n-1) is also < 0
Modulo-2^48 arithmetic on 64 bit
--------------------------------
With the above, we can instantiate a relation `before48' as follows:
a `before48' b <=> 2^47 <= a-b <= 2^48-1
In [RFC 4340, sec. 3.1] it is suggested that
"It may make sense to store DCCP sequence numbers in the most significant 48 bits
of 64-bit integers and set the least significant 16 bits to zero, since this supports a common technique that implements circular comparison A < B by testing whether (A - B) < 0 using conventional two's-complement arithmetic."
Unfortunately this does not work.
Signed 64-bit numbers represent the range -2^63 ... 0 ... 2^63-1. When the difference between two left-shifted 48-bit numbers a and b is 2^63, we can not tell whether a is before' b or b is before' a - the difference will always be negative due to the constraints of the data type. This case occurs for all 48-bit numbers whose difference amounts to 2^47. Since there are 2^47 such numbers, the test based on the above test fails to produce correct results in 2^47 = 140.7375e12 cases. The suggestion from RFC 4340 can thus not be used is misleading.
We therefore develop a definition of `sequence number delta', inspired by dccp_delta_seqno().
Computing sequence number delta -------------------------------
The aim is to develop a function delta_seqno(u64 a, u64 b) which essentially performs subtraction modulo-48 of signed numbers, i.e. it shall satisfy
(a + delta_seqno(a, b)) % 2^48 = b
Hence delta_seqno(a, b) reduces to the subtraction b-a modulo-48, i.e.
delta_seqno(a, b) = (b + 2^48 - a) % 2^48
The main difference is that we want the return result to be a _signed_ 48-bit number, i.e. the range 2^47 ... 2^48-1 is to be mapped into the range -2^47 ... -1 (two's complement in 48 bit). The following table illustrates some cases; a, b are unsigned 48-bit numbers.
+--------+---------+----------+---------+----------------+
| | | b + 2^48 - a | |
| a | b | unsigned | signed | a `before' b |
+--------+---------+----------+---------+----------------+
| 0 | 1 | 1 | 1 | X |
| 0 | 2^47-1 | 2^47-1 | 2^47-1 | X |
| 0 | 2^47 | 2^47 | -2^47 | |
| 0 | 2^47+1 | 2^47+1 | -2^47-1 | |
| 0 | 2^48-1 | 2^48-1 | -1 | |
| 2^48-1 | 0 | 1 | 1 | X |
| 2^47+1 | 0 | 2^47-1 | 2^47-1 | X |
| 2^47 | 0 | 2^47 | -2^47 | |
| 2^48-3 | 5 | 8 | 8 | X |
| 2^47 | 1 | 2^47+1 | -2^47-1 | |
| 2^47+1 | 1 | 2^47 | -2^47 | |
| 5 | 2^48-3 | 2^48 - 8 | -8 | |
+--------+---------+----------+---------+----------------+The following is now the definition of the function delta_seqno:
#define COMPLEMENT48(x) (2^48 - x)
#define TO_SIGNED48(x) ((x < 2^47)? : -COMPLEMENT48(x))
/* returns:
> 0 if a `before' b
= 0 if a = b
< 0 which indicates that b `before' a
*/
s64 delta_seqno(u64 a, u64 b)
{
s64 delta = (b + COMPLEMENT48(x)) % 2^48;
return TO_SIGNED48(x);
}Derived comparisons for nH bit --------------------------------
The dccp_delta_seqno function saves a lot of work:
(i) `before'
a `before' b <=> 1 <= b-a <= 2^(n-1)-1 <=> delta_seqno(a, b) > 0
(ii) `after'
a after' b <=> b before' a <=> delta_seqno(b, a) > 0
(iii) `the same as'
a `the same as' b <=> (a-b) = 0 <=> delta_seqno(b, a) = 0
(iv) `follows' (b is the immediate successor of a)
b `follows' a <=> delta_seqno(b, a) = 1
(v) `precedes' (b is the immediate predecessor of a)
b `precedes' a <=> delta_seqno(b, a) = -1