The core issue I've found is that there is no throttle for how many
deletes get assigned to one TXG. As a results when deleting large files
we end up filling consecutive TXGs with deletes/frees, then write
throttling other (more important) ops.

There is an easy test case for this problem. Try deleting several
large files (at least 1/2 TB) while you do write ops on the same
pool. What we've seen is performance of these write ops (let's
call it sideload I/O) would drop to zero.

More specifically the problem is that dmu_free_long_range_impl()
can/will fill up all of the dirty data in the pool "instantly",
before many of the sideload ops can get in. So sideload
performance will be impacted until all the files are freed.

The solution we have tested at Nexenta (with positive results)
creates a relatively simple throttle for how many "free" ops we let
into one TXG.

However this solution exposes other problems that should also be
addressed. If we are to slow down freeing of data that means one
has to wait even longer (assuming vnode ref count of 1) to get shell
back after an rm or for NFS thread to finish the free-ing op.
To avoid this the proposed solution is to call zfs_inactive() async
for "large" files. Async freeing then begs for the reclaimed space
to be accounted for in the zpool's "freeing" prop.

The other issue with having a longer delete is the inability to
export/unmount for a longer period of time. The proposed solution
is to interrupt freeing of blocks when a fs is unmounted.