ibv_poll_cq on peer-direct CQ does not have proper memory fencing/flushing

[drossetti]

see

libmlx5/src/cq.c

Line 157 in 9d34626

if (cq->peer_peek_table[idx]->busy) {

cq->peer_peek_table[idx]->busy could be asynchronously updated by a PCIe device, so it needs a platform-dependent way of forcing a memory read (a la READ_ONCE or more).

[drossetti]

a more serious problem might be present in mlx5_exp_peer_abort_peek_cq():
((struct mlx5_peek_entry *)(uintptr_t)peek_ctx->peek_id)->busy = 0;