I/O degradation in recent kernels

[shuras]

Hi there!
I've just found there is a serious degradation of SATA SSD disk performance over recent kernels on TF307-MB-A-0.
We tried 5.10.60 initially and it was working fine. Still working fine actually. Upgraded to 5.10.101 and things get more complicated.
At the same system with SSD attached I've got 4 times lower linear (kinda linear as it one big file operations) read/write performance. 5.10.60 able to achieve 440 MB/s at read and 87 MB/s at write as 5.10.101 produce only 120 MB/s at read and 17 MB/s at write.
At the same time I've also got ~100 times slower performance for multicore simultaneous files check-summing.
Nothing interesting can be found in dmesg.
After this unfortunate discovery I've tried other branches. baikalm-5.15.y happens to have exactly same problem.
At the same time baikalm-5.10.y-next, baikalm-5.15.y-next and baikalm-5.18.y-next performing better. All of those *-next branches still have same around 4 times slower performance but there is no such catastrophic degradation with multicore simultaneous files check-summing. With *-next branches it just same around 4 times slower. And 5.18.9 maybe has slightly slower I/O than 5.10.123.
Probably something should be done to kernel config. I just using default baikalm configs here.