arch: pgtable: define MAX_POSSIBLE_PHYSMEM_BITS where needed

arch: pgtable: define MAX_POSSIBLE_PHYSMEM_BITS where needed

Stefan Agner reported a bug when using zsram on 32-bit Arm machines
with RAM above the 4GB address boundary:

  Unable to handle kernel NULL pointer dereference at virtual address 00000000
  pgd = a27bd01c
  [00000000] *pgd=236a0003, *pmd=1ffa64003
  Internal error: Oops: 207 [#1] SMP ARM
  Modules linked in: mdio_bcm_unimac(+) brcmfmac cfg80211 brcmutil raspberrypi_hwmon hci_uart crc32_arm_ce bcm2711_thermal phy_generic genet
  CPU: 0 PID: 123 Comm: mkfs.ext4 Not tainted 5.9.6 #1
  Hardware name: BCM2711
  PC is at zs_map_object+0x94/0x338
  LR is at zram_bvec_rw.constprop.0+0x330/0xa64
  pc : [<c0602b38>]    lr : [<c0bda6a0>]    psr: 60000013
  sp : e376bbe0  ip : 00000000  fp : c1e2921c
  r10: 00000002  r9 : c1dda730  r8 : 00000000
  r7 : e8ff7a00  r6 : 00000000  r5 : 02f9ffa0  r4 : e3710000
  r3 : 000fdffe  r2 : c1e0ce80  r1 : ebf979a0  r0 : 00000000
  Flags: nZCv  IRQs on  FIQs on  Mode SVC_32  ISA ARM  Segment user
  Control: 30c5383d  Table: 235c2a80  DAC: fffffffd
  Process mkfs.ext4 (pid: 123, stack limit = 0x495a22e6)
  Stack: (0xe376bbe0 to 0xe376c000)

As it turns out, zsram needs to know the maximum memory size, which
is defined in MAX_PHYSMEM_BITS when CONFIG_SPARSEMEM is set, or in
MAX_POSSIBLE_PHYSMEM_BITS on the x86 architecture.

The same problem will be hit on all 32-bit architectures that have a
physical address space larger than 4GB and happen to not enable sparsemem
and include asm/sparsemem.h from asm/pgtable.h.

After the initial discussion, I suggested just always defining
MAX_POSSIBLE_PHYSMEM_BITS whenever CONFIG_PHYS_ADDR_T_64BIT is
set, or provoking a build error otherwise. This addresses all
configurations that can currently have this runtime bug, but
leaves all other configurations unchanged.

I looked up the possible number of bits in source code and
datasheets, here is what I found:

 - on ARC, CONFIG_ARC_HAS_PAE40 controls whether 32 or 40 bits are used
 - on ARM, CONFIG_LPAE enables 40 bit addressing, without it we never
   support more than 32 bits, even though supersections in theory allow
   up to 40 bits as well.
 - on MIPS, some MIPS32r1 or later chips support 36 bits, and MIPS32r5
   XPA supports up to 60 bits in theory, but 40 bits are more than
   anyone will ever ship
 - On PowerPC, there are three different implementations of 36 bit
   addressing, but 32-bit is used without CONFIG_PTE_64BIT
 - On RISC-V, the normal page table format can support 34 bit
   addressing. There is no highmem support on RISC-V, so anything
   above 2GB is unused, but it might be useful to eventually support
   CONFIG_ZRAM for high pages.

Fixes: 61989a80fb3a ("staging: zsmalloc: zsmalloc memory allocation library")
Fixes: 02390b87a945 ("mm/zsmalloc: Prepare to variable MAX_PHYSMEM_BITS")
Acked-by: Thomas Bogendoerfer <tsbogend@alpha.franken.de>
Reviewed-by: Stefan Agner <stefan@agner.ch>
Tested-by: Stefan Agner <stefan@agner.ch>
Acked-by: Mike Rapoport <rppt@linux.ibm.com>
Link: https://lore.kernel.org/linux-mm/bdfa44bf1c570b05d6c70898e2bbb0acf234ecdf.1604762181.git.stefan@agner.ch/
Signed-off-by: Arnd Bergmann <arnd@arndb.de>