PBlock protocol and toolset

PBlock is minimalistic protocol for serving random-access files using streams.

Tools

pblock-srv - serve a file using pblock stream
- should be renamed to pblock-file?
pblock-read - read range from pblock stream and print to output
pblock-write - write to pblock connection at specified offset
pblock-bd - attach pblock stream as local blok device

Tools in progress:

pblock-mix - mix many files like in overlay-fs

Tools intended:

pblock-mem/pblock-exec/pblock-ld - map pblock into memory of child process
pblock-cp - copy data from one pblock to another
pblock-swap - swap data between to pblocks
pblock-server - serve pblock connection over TCP or UNIX domain socket (This will allow for connecting multiple times to single pblock)
pblock-hub - like pblock-server but works on pipes and client count is fixed in single run (Possibly may be integrated with pblock-mix?)
pblock-mount - mount stream using fuse interface (Possibly not needed as there is pblock-bd?)
pblock-http - http(s) (using byte ranges) to pblock stream interface
compression/decompression commnand maybe?
pblock-sniff - (sniffs only one side of a connection?)
pblock-mirror - parallel usage of multiple underlying pblocks like in RAID-1. (Possibly may be integrated with pblock-mix?)
pblock-cache - ??? Not sure if possible, necessary and in-good-taste. Don't know how to inalidate cache.
pblock-rw - route READ and WRITE segments to different connections (Possibly may be integrated with pblock-mix?)

Protocol

v0.2

Axioms

PBlock connection allows for acces to single file.
PBlock interface works on top of pair of ordered, reliable, unidirectional byte streams.
PBlock should be minimal and simple. Simple both to understand its operation and to code its implementations.
Files are infinite, 0-based byte sequences. We have no notion of file size nor of "missing/unset blocks".
Access to files is transactional. In each transaction one can atomicaly write and read file in many locations, not necessarily coherent.

Nice-to-haves

PBlock interface should integrate well with zero-copy kernel interfaces like sendfile() system call. This will ensure space for performance improvement, especialy in large file-mangling pipelines.
Implementation of multiple readers/single writer locking for transactions should be possible. (This is currently not satisfied, I think.)
Some obvious things:
- For large transfers PBlock interface shouldn't add much overhead.

Implementation

Startup

No handshaking - this step should be performed by other, speclialized protocol (for example multistream).

No options negotiation - because flags are just unstructural approach to use one codebase for many things.

Just establish your connection and roll.

Transmission

Each half of connection is composed of segment objects.
Each segment has (in order)
- type (byte r, w or c),
  - r segments are used to request data ranges
  - w segments are used to transmit data
  - c segments commits pending operations
- if type is r or w: offset varint
- if type is r or w: length varint
- if type is w: payload - blob with size length
Client requests atomic reads and writes with r and w segments, then commits transaction with c segment.
Server processes segments in order and confirms finished atomic actions with c segment.
- For every r segment it must respond with w segment (with exact same offset and length).
- For every w segment it performs writes according to is internal semantics.

Shutdown

There is no such thing as pblock protocol shutdown. Pblock, once initiated, will remain forever.

Yea, thats theory. Adding termination steps will make protocol unnecesarily complex, and we can't rely on underlying streams termination, because they don't have to be closable. Also, adding "stream closability" to axioms is unnecessary assumption.

In practice many streams are closable (UNIX pipes, TCP, ...). When writing implementation for such streams one can do this logic:

Client can close it's writing stream after each sent segment. This is ok because server can't request any reaction from client, so client won't be needing to write any data.
Server, after handling all segments, can surely close its writing stream, because client won't be sending anything more. Server knows, that client won't be sending anything more when EOS is detected on client->server stream.

SupraSummus / pblock