Hacker News Reader

A thread can still deadlock with itself with non-recursive mutexes or spinlocks. Also assuming a single CPU system is no longer relevant today.

"Thread halting is a not problem as long as it resumes"

the whole point of a non blocking algorithm is that it gives guarantees if a thread doesn't resume for whatever reason.

View on HN

The whole point of a non-blocking algorithm is system-wide forward progress. Yes, this is much harder to do if a thread can suspend for an unbounded amount of time, but that's kind of the point of the article: being in the kernel allows you to pull the kind of trick that makes that behaviour a non-issue.

(edit: s/per-thread/system-wide, since we're talking about lock-freedom).

Technically non-blocking only guarantees system wide progress. Only wait-free algorithms guarantee per thread progress.

Anyway, even controlling preemption and disabling interrupts very much does not make a spinlocked critical section a lock free algorithm. That's not just a theoretical issue. We are dealing right now with a couple of machines where periodically the kflush kernel thread livelocks hard preventing the rest of the system from ever writing a page to disk requiring a hard powercycle.

Reply on news.ycombinator.com