The Latest On The Linux 5.9 Kernel Regression Stemming From Page Lock Fairness

After looking through the perf data collected during some of the benchmark runs, Torvalds ended up stumbling upon some EXT4 code that does some "fairly crazy things" and code that does "all kinds of bad". More details and the patch via this mailing list post in roping in EXT4 developers.

Unfortunately, that EXT4 patch didn't solve the regression. However, in separate workloads not regressed by Linux 5.9, it does help offer better performance in areas like MariaDB/MySQL. In particular, for lower concurrent users/queries, the EXT4 patch does seem to provide measurable improvements. I am still testing that EXT4 patch as time allows though it's not clear yet if that patch will still be pursued for upstream or the like to address that locking code spotted by Torvalds.

The belief is that the new page lock fairness may just be really bad for performance.