how to get Linux to automatically release old pagecache pages?

Ever since upgrading to Linux kernel 5.8 and later, I've been having problems with my system freezing up from running out of RAM, and it's all going to the pagecache. I have a program that reorganizes the data from the OpenStreetMap planet_latest.osm.pbf file into a structure that's more efficient for my usage. However, because this file is larger than the amount of RAM on my system (60GB file versus 48GB RAM), the page cache fills up. Before kernel 5.8, the cache would reach full, and then keep chugging along (with an increase in disk thrashing). Since 5.8, the system freezes because it won't ever automatically release a page from the page cache (such as 30GB earlier in my sequential read of the planet_latest.osm.pbf file). I don't need to use my reorganizing program to hang the system; I found the following unprivileged command would do it: cat planet_latest.osm.pbf >/dev/null I have tried using the fadvise64() system service to manually force releases of pages in the planet file I have already passed; it helps, but doesn't entirely solve the problem with the various output files my program creates (especially when those temporary output files are randomly read back later). So, what does it take to get the 5.8 through 5.10 Linux kernel to actually automatically release old pages from the page cache when system RAM gets low? To work around the problem, I have been using a script to monitor cache size and write to /proc/sys/vm/drop_caches when the cache gets too large, but of course that also releases new pages I am currently using along with obsolete pages. while true ; do H=free | head -2 | tail -1 | awk '{print $6}' if [ $H -gt 35000000 ]; then echo -n $H " @ " ; date echo 1 >/proc/sys/vm/drop_caches sensors | grep '°C' H=free | head -2 | tail -1 | awk '{print $6}' echo -n $H " @ "; date fi sleep 30 done (the sensors stuff is to watch out for CPU overheating in stages of my program that are multi-threaded CPU-intensive rather than disk-intensive). I have filed a bug report at kernel.org, but they haven't looked at it yet.