If CONFIG_ARCH_DMA_ADDR_T_64BIT enabled for x86 systems and physical
memory is more than 4GB, dma_map_page may return a valid memory
address which greater than 0xffffffff. As a result, the mlx5 device page
allocator RB tree will be initialized with valid addresses greater than
0xfffffff.

However, (addr & PAGE_MASK) set the high four bytes to zeros. So, it's
impossible for the function, free_4k, to release the pages whose
addresses greater than 4GB. Memory leaks. And mlx5_ib module can't
release the pages when user try to remove the module, as a result,
system hang.

[root@rdma05 root]# dmesg  | grep addr | head
addr             = 3fe384000
addr & PAGE_MASK =  fe384000
[root@rdma05 root]# rmmod mlx5_ib   <---- hang on

---------------------- cosnole log -----------------
mlx5_ib 0000:04:00.0: irq 138 for MSI/MSI-X
  alloc irq_desc for 139 on node -1
  alloc kstat_irqs on node -1
mlx5_ib 0000:04:00.0: irq 139 for MSI/MSI-X
0000:04:00.0:free_4k:221:(pid 1519): page not found
0000:04:00.0:free_4k:221:(pid 1519): page not found
0000:04:00.0:free_4k:221:(pid 1519): page not found
0000:04:00.0:free_4k:221:(pid 1519): page not found
---------------------- cosnole log -----------------

Fixes: bf0bf77f ('mlx5: Support communicating arbitrary host page size to firmware')
Signed-off-by: Honggang Li <honli@redhat.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>