Zack's Kernel News
Zack's Kernel News

Chronicler Zack Brown reports on the latest news, views, dilemmas, and developments within the Linux kernel community.
Speeding Up Futexes…Maybe
Thomas Gleixner pointed out that the kernel's 'futex' locks used a global hash value to keep track of state changes, but that the hash values were not guaranteed to be unique. The kernel could handle these collisions, but it would slow things down. He posted a patch to provide a reference counting mechanism to allow threads to guarantee no hash collisions. He said, "this creates a futex private state that avoids hash collisions and on NUMA systems also cross-node memory accesses."
To make use of the mechanism, he said, "each thread has to attach to the futex before any other operations on that futex." He explained the usage of the sys_futex()
call to increment and decrement the futex reference count.
Thomas added that, although the user interface could be simplified by having threads automatically attach to the futex, he didn't want to implement it that way. Apart from increasing memory usage, the futex attachment code would have its own performance issues if all threads on the system made use of it. As long as it was used only by threads that actually needed to avoid those hash collisions – in other words, for real-time software – the code would provide a measurable performance increase.
[...]
Buy this article as PDF
(incl. VAT)