Comments
Loading Dream Comments...
You must be logged in to write a comment - Log In
ArtistSeveral beautiful purple-haired women each wearing a long purple velvet ankle-length dress with flouncy sleeves and purple suede high heels, are dancing hand-in-hand around a maypole.
When a Buf contains more than 32K bytes of data (including the header and dope vectors), its datastore segment must occupy more than one page. therefore all its internal pointers must be "far" (32bit) pointers, because "near" (16bit) pointers have a maximum span of 64K. (32bit pointers can span 2^32 = 4 GB). The first page of a multipage segment will contain all the LBA's of the remaining pages making up that segment. It will also try to contain all the Buf headers and dope vectors for that segment, in order to simplify garbage collection. The data blocks will occupy the remaining pages of the segment. The system should try not to split a data block across page boundaries, in order that the pages need not occupy contiguous LBA's or page frames. The free space pointers within the segment header (see ASM code) will have slightly different semantics : the first page will have only the "front" space, (which may extend to subsequent pages if "really necessary"), and the remaining pages will have the "mid" space and "back" space (see ASM code). I am tempted to withdraw 16bit (near) addressing altogether, and make all Buf header pointers (and dope vector pointers) 32bit (far), in order to avoid having two sets of processing algorithms in each FBI op. (The high-order word of singlepage segments will always be zero, so all it costs us is an extra two zeroed-bytes per pointer).
The high-order word of the 32bit address will be used to point into the list of LBA's to select a page within the segment, and the low-order word will be the offset into that segment. Not all pages in a multipage segment need to be swapped-in, normal virtual storage algorithms apply (see ASM code).
More code will follow soon :)
