x86/mm: Handle physical-virtual alignment mismatch in phys_p4d_init()

x86/mm: Handle physical-virtual alignment mismatch in phys_p4d_init()

Kyle has reported occasional crashes when booting a kernel in 5-level
paging mode with KASLR enabled:

  WARNING: CPU: 0 PID: 0 at arch/x86/mm/init_64.c:87 phys_p4d_init+0x1d4/0x1ea
  RIP: 0010:phys_p4d_init+0x1d4/0x1ea
  Call Trace:
   __kernel_physical_mapping_init+0x10a/0x35c
   kernel_physical_mapping_init+0xe/0x10
   init_memory_mapping+0x1aa/0x3b0
   init_range_memory_mapping+0xc8/0x116
   init_mem_mapping+0x225/0x2eb
   setup_arch+0x6ff/0xcf5
   start_kernel+0x64/0x53b
   ? copy_bootdata+0x1f/0xce
   x86_64_start_reservations+0x24/0x26
   x86_64_start_kernel+0x8a/0x8d
   secondary_startup_64+0xb6/0xc0

which causes later:

  BUG: unable to handle page fault for address: ff484d019580eff8
  #PF: supervisor read access in kernel mode
  #PF: error_code(0x0000) - not-present page
  BAD
  Oops: 0000 [#1] SMP NOPTI
  RIP: 0010:fill_pud+0x13/0x130
  Call Trace:
   set_pte_vaddr_p4d+0x2e/0x50
   set_pte_vaddr+0x6f/0xb0
   __native_set_fixmap+0x28/0x40
   native_set_fixmap+0x39/0x70
   register_lapic_address+0x49/0xb6
   early_acpi_boot_init+0xa5/0xde
   setup_arch+0x944/0xcf5
   start_kernel+0x64/0x53b

Kyle bisected the issue to commit b569c1843498 ("x86/mm/KASLR: Reduce
randomization granularity for 5-level paging to 1GB")

Before this commit PAGE_OFFSET was always aligned to P4D_SIZE when booting
5-level paging mode. But now only PUD_SIZE alignment is guaranteed.

In the case I was able to reproduce the following vaddr/paddr values were
observed in phys_p4d_init():

Iteration     vaddr			paddr
   1 	      0xff4228027fe00000 	0x033fe00000
   2	      0xff42287f40000000	0x8000000000

'vaddr' in both cases belongs to the same p4d entry.

But due to the original assumption that PAGE_OFFSET is aligned to P4D_SIZE
this overlap cannot be handled correctly. The code assumes strictly aligned
entries and unconditionally increments the index into the P4D table, which
creates false duplicate entries. Once the index reaches the end, the last
entry in the page table is missing.

Aside of that the 'paddr >= paddr_end' condition can evaluate wrong which
causes an P4D entry to be cleared incorrectly.

Change the loop in phys_p4d_init() to walk purely based on virtual
addresses like __kernel_physical_mapping_init() does. This makes it work
correctly with unaligned virtual addresses.

Fixes: b569c1843498 ("x86/mm/KASLR: Reduce randomization granularity for 5-level paging to 1GB")
Reported-by: Kyle Pelton <kyle.d.pelton@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Kyle Pelton <kyle.d.pelton@intel.com>
Acked-by: Baoquan He <bhe@redhat.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20190624123150.920-1-kirill.shutemov@linux.intel.com