If a process dumps core while owning a cxl file descriptor obtained
from an AFU driver (e.g. cxlflash) through the cxl_get_fd() API, the
following error occurs:

  [  868.027591] Unable to handle kernel paging request for data at address ...
  [  868.027778] Faulting instruction address: 0xc00000000035edb0
  cpu 0x8c: Vector: 300 (Data Access) at [c000003c688275e0]
      pc: c00000000035edb0: elf_core_dump+0xd60/0x1300
      lr: c00000000035ed80: elf_core_dump+0xd30/0x1300
      sp: c000003c68827860
     msr: 9000000100009033
     dar: c
  dsisr: 40000000
   current = 0xc000003c68780000
   paca    = 0xc000000001b73200   softe: 0        irq_happened: 0x01
      pid   = 46725, comm = hxesurelock
  enter ? for help
  [c000003c68827a60] c00000000036948c do_coredump+0xcec/0x11e0
  [c000003c68827c20] c0000000000ce9e0 get_signal+0x540/0x7b0
  [c000003c68827d10] c000000000017354 do_signal+0x54/0x2b0
  [c000003c68827e00] c00000000001777c do_notify_resume+0xbc/0xd0
  [c000003c68827e30] c000000000009838 ret_from_except_lite+0x64/0x68
  --- Exception: 300 (Data Access) at 00003fff98ad2918

The root cause is that the address_space structure for the file
doesn't define a 'host' member.

When cxl allocates a file descriptor, it's using the anonymous inode
to back the file, but allocates a private address_space for each
context. The private address_space allows to track memory allocation
for each context. cxl doesn't define the 'host' member of the address
space, i.e. the inode. We don't want to define it as the anonymous
inode, since there's no longer a 1-to-1 relation between address_space
and inode.

To fix it, instead of using the anonymous inode, we introduce a simple
pseudo filesystem so that cxl can allocate its own inodes. So we now
have one inode for each file and address_space. The pseudo filesystem
is only mounted on the first allocation of a file descriptor by
cxl_get_fd().

Tested with cxlflash.
Signed-off-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
Reviewed-by: Matthew R. Ochs <mrochs@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>