Seeing this behaviour on 10.1-RELEASE and 10.1-STABLE (May 26th) amd64 systems.
Systems are low-memory (2GB), but have also seen this behaviour on high (256GB) memory systems all without swap.
10.0-RELEASE seems to NOT be affected.
Run a system without swap and a ZFS filesystem (eg: 1TB to 2TB used).
Appears to be triggered when deleting (either through snapshots, or via regular files) a large of amount data (hundreds of GB) (even with few files).
Wired memory skyrockets and is never reclaimed until hitting vm.v_free_min, after which the kernel randomly kills processes to stay up.
Once this has been done, sometimes the wired memory is reclaimed, sometimes not.
Exactly the same config on 10.0-RELEASE is not affected.
Have tried setting:
vfs.zfs.arc_max=128M (and lower)
vfs.zfs.zio.use_uma=0 (no effect)
increasing v_free_min and v_free_target = system just kills processes faster
my test case on 10.2-RELEASE amd64
1. create 1G file
# dd if=/dev/urandom bs=1024 count=1024 of=blockout.bin
2. while : ; do cat blockout.bin ; done | pv -brt | dd bs=1024k count=16000 of=blockout2.bin
3. see Wired and ARC counts on 'top' output
4. remove blockout2.bin
5. see Wired and ARC counts on 'top' output
my solution: disable UMA for ZFS