Process freeze when calling UNSAFE.freeMemory of large chunks.

[Ivan Daschinsky]

Hi! I've faced recently with strange issue with UNSAFE.freeMemory. When in
some thread(threads) we try to deallocate huge chunks of memory, all
process freezes completely.

I've wrote reproducer in C and Java with relatively same functionality.

In few words, java reproducer will create N threads, each of therm
allocated 2^M chunk
of memory, sleeps T, than wait for barrier's broke. So all N threads will
simultaneously deallocate N*2^M total bytes of memory.

When we run java code on machine with 1.5 Tb of RAM with params N=40, M=33,
T=30, we see strange freeze on memory deallocation, when I run strace -F
-p, I see freeze when on munmap calls,
main thread freezes and stop printing "Tick" messages. However, perf record
doesn't handle these events and complains about too low --proc-map-timeout.
Even encrease of this timeout doesn't help. -XX:+PerfDisableSharedMem also
doesn't help. I will attach some excerpts from perf top during freeze

PerfTop:    7898 irqs/sec  kernel:97.2%  exact:  0.0% [4000Hz cycles],
(all, 64 CPUs)
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------

    18.84%  [kernel]          [k] osq_lock
    16.53%  [kernel]          [k] down_read_trylock
     8.23%  [kernel]          [k] up_read
     7.41%  [kernel]          [k] handle_mm_fault
     4.02%  libjvm.so         [.]
GenericTaskQueueSet<OverflowTaskQueue<StarTask, (MemoryType)5, 131072u>,
(MemoryType)5>::steal_best_of_2
     3.90%  [kernel]          [k] page_add_new_anon_rmap
     3.10%  [kernel]          [k] find_vma
     2.80%  libjvm.so         [.] Copy::fill_to_memory_atomic
     2.74%  [kernel]          [k] mem_cgroup_charge_common
     2.39%  [kernel]          [k] clear_page_c_e
     1.76%  [kernel]          [k] free_pcppages_bulk
     1.53%  [kernel]          [k] page_fault
     1.46%  [kernel]          [k] unmap_page_range
     1.21%  [kernel]          [k] release_pages
     1.18%  [kernel]          [k] __mem_cgroup_commit_charge
     1.15%  [kernel]          [k] __do_page_fault
     1.08%  [kernel]          [k] __mem_cgroup_uncharge_common
     1.00%  [kernel]          [k] __list_del_entry
     0.89%  [kernel]          [k] get_pageblock_flags_group
     0.76%  [kernel]          [k] rwsem_spin_on_owner
     0.61%  libjvm.so         [.] G1ParScanThreadState::trim_queue
     0.58%  [kernel]          [k] mem_cgroup_page_lruvec

Java code was run on OpenJDK 1.8.0_131-b12 on RHEL 7.4. and also on Oracle
JDK 1.8.0_121-b13.

However, C reproducer works like a charm without any freezes, it does I
suppose simply the same as Java one.

Could you please explain why this happens? Is it expected behaviour?

Code of reproducers you can download from these gists.

https://gist.github.com/ivandasch/491227e74b6536a6ba67090844192300
https://gist.github.com/ivandasch/4647528b8b0dc8324c370b86b9799ece
https://gist.github.com/ivandasch/60a649e872089602461ac72fd17508b7

-- 
Sincerely yours, Ivan Daschinskiy