Linux 6.5 Last Minute Fixes A Performance Regression - 34% Drop In A Micro-Benchmark
While Linux 6.5 is expected for release tomorrow, the flow of last minute fixes isn't over.
Sent out today were some x86 fixes for Linux 6.5 Git ahead of its final tagging -- or Linux 6.5-rc8 if Linus Torvalds decides to drag out the release by an extra week. Making the weekend pull request notable is that it has a fix for a performance regression due to a missing set of X86_FEATURE_OSXSAVE.
The patch addressing the issue explains that Intel's 0-day testing of the Linux kernel uncovered a 34.6% performance regression. The culprit comes down to missing a simple two lines of code for checking and setting X86_FEATURE_OSXSAVE earlier in the boot process.
This is a regression for just new code introduced during the Linux 6.5 merge window. Introducing the regression was the Linux 6.5 change for delaying x86 FPU initialization as part of cleaning up the boot process. Intel's kernel test robot reported the regression in mid-July but it wasn't until today that a fix worked its way to the mainline kernel.
This 34.6% performance regression was uncovered by the stress-ng af-alg test case (exercising the AF_ALG socket domain) as a micro-benchmark but nevertheless at least this was at least spotted and fixes ahead of the likely Linux 6.5 stable debut tomorrow.
Sent out today were some x86 fixes for Linux 6.5 Git ahead of its final tagging -- or Linux 6.5-rc8 if Linus Torvalds decides to drag out the release by an extra week. Making the weekend pull request notable is that it has a fix for a performance regression due to a missing set of X86_FEATURE_OSXSAVE.
The patch addressing the issue explains that Intel's 0-day testing of the Linux kernel uncovered a 34.6% performance regression. The culprit comes down to missing a simple two lines of code for checking and setting X86_FEATURE_OSXSAVE earlier in the boot process.
This is a regression for just new code introduced during the Linux 6.5 merge window. Introducing the regression was the Linux 6.5 change for delaying x86 FPU initialization as part of cleaning up the boot process. Intel's kernel test robot reported the regression in mid-July but it wasn't until today that a fix worked its way to the mainline kernel.
This 34.6% performance regression was uncovered by the stress-ng af-alg test case (exercising the AF_ALG socket domain) as a micro-benchmark but nevertheless at least this was at least spotted and fixes ahead of the likely Linux 6.5 stable debut tomorrow.
11 Comments