You can use TLS 1.0 as guidance: it is the direct successor of SSL 3.0, so many things are quite similar, and in some respects TLS 1.0 is a bit clearer. In section 6.3 you will find the key generation process, with the exact sentence:
To generate the key material, compute [...]
until enough output has been generated. Then the key_block is
partitioned as follows:
the important word being "partitioned".
For instance, if you are using 3DES as symmetric cipher and SHA-1 for the MAC, the "write keys" are 24-byte long each, the IV are 8-byte long each, and the MAC keys are 20-byte long each. So, a total of 104 bytes. The key generation function repeatedly invokes SHA-1 and MD5 on various elements; each round produces 16 additional bytes (that's the output size of MD5). You need 104 bytes, hence, you will need 7 rounds.
At the first round, you call SHA-1 over the concatenation of 'A' (a single byte of value 65), the master secret, the server random and the client random, in that order ('+' indeed denotes concatenation). This SHA-1 invocation yields 20 bytes. The concatenation of the master secret and the SHA-1 output is then hashed with MD5, which yields 16 bytes. These are the first 16 bytes of the key block.
For the second round, the processing is identical except that you use 'BB' instead of 'A' (two bytes of value 66 each). This produces the next 16 bytes of the key block.
You continue like this. Ultimately, you can potentially do 26 rounds (up to 'ZZZ...Z'; the SSL 3.0 specification does not define how to go beyond that). This would yield a total of 26*16 = 416 bytes. But if you need only the first 104 bytes, there is no need to compute the whole 26 rounds; just compute enough rounds to get the number of bytes you need.
Once you have your key block (104 bytes with 3DES and SHA-1), you split it into the needed key elements. The first 20 bytes go to the client write MAC key, the next 20 bytes for the server write MAC key, then the next 24 bytes for the client write key, and so on.