This week, the Linux kernel community learned that when it comes to maintaining a schedule, sometimes it’s best to cut your losses to get everything back to running normally. 

Sunday, Linus Torvalds released version 7.2 of the Linux kernel. As per the norm, the majority of changes relate to drivers, many others addressed networking, and there were plenty of commits to prune legacy code. And at least one group of commits added old code back into the kernel. 

Overall, the new release incorporates about 200 subsystem pull requests, according to the KernelNewbies page. They incorporated approximately 16,000 individual commits. In terms of changes, this release had about the same number of changes ushered in by Linux 7.1, which had 218 pull requests.

Pulling the Plug on GPU Scheduling

Torvalds singled out last-minute work on the scheduler for the Direct Rendering Manager, a scheduling subsystem for GPUs and hardware accelerators. This group of changes fixed a regression caused by moving to a new “Fair” scheduling policy, from the standard multi-queue First-in-First-Out (FIFO) and round-robin scheduling as a backup. 

Ostensibly, the DRM Fair Scheduler would improve performance for interactive applications competing for resources in busy environments on GPUs. When tested, however, this approach of dividing up the GPU’s attention resulted in slow screen refreshes and spiking GPU usage when a user played games, watched video, or undertook other graphics-intensive activities. 

This group of patches temporarily pulls the plug on the project, reverting the scheduler to FIFO, and the changes back to experimental status. 

The reversions “may not be pretty, but it’s the correct way to deal with ‘oh, that code wasn’t ready and caused problems’,” Torvalds wrote in the announcement of the 7.2 release.

Scheduling for the Cache

A more successful scheduling advance, this time for the CPU, comes with the inclusion in the 7.2 kernel of code to support cache-aware scheduling. This set of changes were initially submitted last December by engineers at Intel and other companies. 

This is the first scheduling algorithm that takes into consideration not only which cores are underutilized but also which caches are already prepped with supporting material.

This work addresses the overhead associated with moving a thread from one processor core to another, overhead previously thought to be minimal. Each core of a multi-core processor gets its own L1 and L2 cache, where a thread running on that core can “build a nest” of supporting data and functions, explained Qualcomm engineer Vishnu Santhosh in a LinkedIn post on the technology. 

When a thread is moved from one core to another, the thread must rebuild its nest, so to speak, which can incur a few microseconds of lag-time cache-rebuilding, offsetting the benefit of moving the thread in the first place.

“An idle core advertised itself as free. It was because the CPU came at no extra charge. The cache did not,” Santhosh wrote. 

The Intel engineers implemented Cache Aware Scheduling (“CONFIG_SCHED_CACHE”) within the Linux default CPU scheduler, instructing it to aggregate tasks that share data to the same cache domain. 

“Every migration is a bet about what matters more: the core that is free this instant, or the cache that is warm this instant,” Santhosh wrote. “For most of Linux’s history that bet undervalued every cache beyond the core right in front of the scheduler. Linux 7.2 is the release where the kernel started pricing both sides of it.”

The Scheduling of Linux

As Microsoft has found with Windows, AI testing tools are unearthing bugs and regressions at a previously unimagined speed.  

“This last week of the release was – once again – bigger than I would have wished for,” Torvalds wrote, concluding the increased numbers are now the “new normal,” thanks to AI aid.  

In fact, Linux 7.2 was the second-largest release in the kernel’s history, according to Linux Weekly News (subscription required). Linux 6.7 brought in more commits overall, but they were dominated by the bcachefs development history, whereas 7.2 is more broad-based.

Nonetheless, Torvalds is set on maintaining Linux’s rapid release schedule. With 7.2 out of the way, the merge window for Linux 7.3 has commenced, and the crew has until mid-October to go through the entire process again. Torvalds noted that 40 pull requests are already sitting in the queue. Back to work.