When tcache is enabled, memory tends to not shrink very well, but
be fragmented in long running threads with several different malloc
allocation patterns.

Closest similar issue I did find was
https://bugs.linaro.org/show_bug.cgi?id=3950
[LKFT: libhugetlbfs: heapshrink-2M-64 failed]

While talking with the user experimenting the problem the first idea
I have, but just heuristics, could be to change _int_free():

if (tcache->counts[tc_idx] < mp_.tcache_count)

to swap the 'p' argument with the oldest entry in the tcache, and not
'return'. Checking if any other local variable needs an adjust as well,
at least the 'size' variable.