View source for Kernel: STREAMS

The IRIX [[STREAMS]] framework provides a modular, message-passing I/O subsystem for character devices, terminals, network protocols, and pseudo-devices. It uses queues (read/write pairs per module/driver), message blocks (mblk_t/dblk_t), and buffers for data transfer. The implementation is highly optimized for performance with:

Per-CPU free lists (NUMA-aware) for message blocks, data blocks, and buffers of various sizes.
Cache coloring and alignment support.
Zone-based allocation for small structures.
Page-based buffering with coalescing and reclaim.
Global limits and dynamic reclamation via strgiveback.
MP-safe design using spinlocks per CPU.

The core data structures are managed through list vectors (lf_listvec) per CPU, with separate pools for different buffer sizes (64B to full page).
== Key Functions ==
Initialization

str_init_master: Initializes global structures for master CPU.
str_init_slave: Sets up per-CPU structures for slave CPUs.
str_init_memsize: Configures page limits based on physical memory.
str_init_alloc_bufs: Preallocates initial buffers per CPU.
strinit_giveback: Starts periodic page reclamation timer.

Message Block Allocation

allocb: Allocates mblk_t + dblk_t + buffer (various sizes).
allocb_nobuffer: Allocates only mblk_t/dblk_t (no data buffer).
esballoc: Attaches external buffer (zero-copy).
bufcall / esbbcall: Schedule callback when memory available.

Queue Operations

putq / putbq / insq: Insert message into queue (priority-ordered).
getq: Remove next message from queue.
rmvq: Remove specific message.
flushq / flushband: Discard messages (all or by band).
qenable: Schedule queue service routine.
canput / bcanput: Flow control checks.

Buffer Management

lf_get / lf_free: Node allocation/deallocation from per-CPU lists.
lf_addpage / lf_getpage_tiled: Page tiling and addition to free lists.
lf_freepage: Return full page to kernel VM.
str_preallocate: Initial buffer preallocation.

Special Cases

msgpullup: Coalesce message into single contiguous buffer.
copyb / copymsg / dupb / dupmsg: Copy or reference-count messages.
linkb / unlinkb: Chain management.
adjmsg: Trim bytes from head/tail.

== Undocumented or IRIX-Specific Interfaces and Behaviors ==
Structures

mblk_t: Message descriptor (pointers, type, flags, band).
dblk_t: Data block (buffer pointers, refcount, type, size, free routine).
queue_t: Read/write queue pair with flow control parameters.
qband_t: Per-priority-band flow control (for non-zero bands).
lf_listvec / listvec: Per-CPU free list vectors with multiple size classes.
strstat: Per-CPU statistics.

Buffer Pools

Fixed sizes: 64, 256, 512, 2048 bytes + full page.
Combined mblk_t/dblk_t pool.
Separate message header pool.
Page reclamation: strgiveback returns excess pages to kernel.

NUMA Awareness

Per-node, per-CPU list vectors.
Allocation prefers local node.

Flow Control

High/low water marks per queue and band.
QFULL / QB_FULL flags.
Back-enable nearest upstream service routine.

Extended Free Routine

esballoc supports external buffers with custom free function.

== Similarities to illumos and BSD Kernel Implementations ==
illumos (Solaris-derived)
Strong similarity:

Message block (mblk_t/dblk_t) and queue model nearly identical.
allocb, esballoc, putq, getq, canput, qenable.
Per-CPU caches and buffer pools.
Flow control with high/low water.

Differences: illumos uses kmem_cache extensively; more modular.
BSD (FreeBSD, etc.)
Less similar:

No native STREAMS (removed long ago).
Character I/O via cdevsw or tty.
Some ports emulate via compatibility layers.

Overall, IRIX STREAMS is classic SVR4 with NUMA/per-CPU optimizations. illumos provides closest modern equivalent; BSD lacks native support. For replication: preserve per-CPU lists, buffer size classes, and queue scheduling semantics. Page reclamation and tiling logic are IRIX-specific.
[[Category: Kernel Documentation]]