I've switched to BUFQ_PRIOCSCAN with the suggested queue sizings, and I
can report its working at least as well for me as the previous state,
which was BUFQ_READPRIO.

My load-case is a laptop with 1GB of memory, doing Xorg, and running
csup and/or cvsup which usually stresses the file buffer cache
extremely, and makes interactive response U/S.

So far, no crashes, and what feels like tolerable interactivity.

I don't think one respondent is a proof, but I'd certainly back
removing the air of 'experimental' from this one, unless somebody has a
strongly negative experience: and the default settings in
kern/bufq_priocscan.c would be fine at priocscan_burst[512 128 32]

cheers

-george