osc/rdma: bring bug/threading fixes into v3.1.x from master

[hjelmn]

This commit contains the contents of:

45db363
7f4872d

to the v3.1.x branch. These commits fix a couple of bugs and improve
the threading support (reference #2530).

To keep the code mostly in sync with master I added code to
osc_rdma_types.h to convert between the atomics support on master and
v3.1.x.

Signed-off-by: Nathan Hjelm [email protected]

[hjelmn]

@bwbarrett If you don't have time to review this I can look for someone else.

[hjelmn]

With this change MTT finally passes all RMA tests:

PASS: c_create
PASS: c_create_size
PASS: c_create_disp
PASS: c_create_info
PASS: c_create_info_half
PASS: c_create_no_free
PASS: c_fence_lock
PASS: c_win_attr
PASS: c_win_errhandler
PASS: c_fence_simple
PASS: c_fence_asserts
PASS: c_fence_put_1
PASS: c_lock_illegal
PASS: c_lock_negative_rank
PASS: c_put
PASS: c_put_big
PASS: c_accumulate
PASS: c_accumulate_atomic
PASS: c_get
PASS: c_get_big
PASS: c_transpose1
PASS: c_transpose2
PASS: c_transpose3
PASS: c_transpose6
PASS: c_transpose5
PASS: c_transpose7
PASS: c_strided_acc_indexed
PASS: c_strided_acc_subarray
PASS: c_strided_acc_onelock
PASS: c_strided_get_indexed
PASS: c_strided_putget_indexed
PASS: c_post_start
PASS: win_dup_fn
PASS: pp_1sided
PASS: 1sided
PASS: mt_1sided
PASS: small_1sided
PASS: multiple_locks2
PASS: c_fetch_and_op
PASS: c_flush
PASS: c_get_accumulate
PASS: c_get_accumulate_strided
PASS: c_reqops
PASS: c_win_shared
PASS: c_win_shared_noncontig
PASS: c_win_shared_noncontig_put
PASS: c_strided_getacc_indexed
PASS: c_strided_getacc_indexed_shared
PASS: c_strided_putget_indexed_shared
PASS: halo_1sided_put_alloc_mem
PASS: win_allocate_shared
PASS: win_allocate_two_shared
PASS: c_create_dynamic
PASS: compare_and_swap