GVisor bufferv2

GVisor discovered that memory copy slows down operations - and switched to a new reference counted model and lists of buffers. This is a fork, cleaned up of dependencies.

The intent is to use it in the io utils and h2 implementation.

The model:

Buffer is a linked list of "View"

• Prepend/Append
• PullUp - makes a range contiguous
• Flatten - one large buffer
• Clone - copy on write, chunks shared until written to
• Apply(fn)
• Merge
• WriteFromReader - multiple reads of 64k
• ReadToWriter - pushes chunks to writer
• AsBufferReader,

View:

• has Next/Prev
• a 'chunk' containing data
• a reference count.
• read and write
• NewView gets a view from the view pool and a chunk.
• NewViewSize - also calls Grow(n), which adds to 'write' (==end)
• Reader, Writer
• AsSlice - no copy but 'should not modify directly', instead Write(At)
• ToSlice - a copy
• Clone - reused chunk, different read/write. Most call Release()
• Reset - set read/write to 0

History

Forked from ce87948214909bcf7d274d9dc8f0772829d21d2a on Dec 12.

• use std log, context, atomic instead of the gvisor extensions
• still using MostSignificantOne64 in assembly for getChunkPool - can be rewritten
• removed ref leak checking and associated dependency