Head-of-line blocking
Head-of-line blocking occurs when earlier blocked work prevents later work from progressing even though that later work could otherwise be served.
Identify the shared queue or execution path, the blocked work and the resource available to later work. A long wait behind a busy single server alone does not demonstrate spare usable capacity or avoidable head-of-line blocking. [1]
Ruan and Pai's 2006 study traced blocking in filesystem-related calls in the Flash web server and Apache. Their FreeBSD 4.6 example held an exclusive parent-directory vnode lock while fetching a child inode from disk, delaying lookups sharing that path even when those lookups needed otherwise resident data. Read-only mounts and disabled access-time updates did not remove the blocking. Flash here names a web server, not NAND memory. These are historical observations, not a claim about current kernels. [1]
Separate queues or bypass can help only when work is actually independent and resources are usable. Reordering cannot discard data dependencies, and multiple host queues alone do not prove independent device service or completion order. [2] [3]
Relations
Citations
- Ruan and Pai. Understanding and Addressing Blocking-Induced Network Server Latency. USENIX ATC 2006, sections 2.2, 3.1 and 3.3.
- Linux 6.16: Multi-Queue Block IO Queueing Mechanism, hardware dispatch queues and tag-based completion.
- Hesela: queue order, blocked work and completion boundaries.
Sources & caveats
- Source review of a historical storage-stack blocking mechanism; no replication or general diagnosis of contemporary SSDs.