pageattr: prevent PSE and GLOABL leftovers to confuse pmd/pte_present and pmd_huge
Without this patch any kernel code that reads kernel memory in non
present kernel pte/pmds (as set by pageattr.c) will crash.
With this kernel code:
static struct page *crash_page;
static unsigned long *crash_address;
[..]
crash_page = alloc_pages(GFP_KERNEL, 9);
crash_address = page_address(crash_page);
if (set_memory_np((unsigned long)crash_address, 1))
printk("set_memory_np failure\n");
[..]
The kernel will crash if inside the "crash tool" one would try to read
the memory at the not present address.
crash> p crash_address
crash_address = $8 = (long unsigned int *) 0xffff88023c000000
crash> rd 0xffff88023c000000
[ *lockup* ]
The lockup happens because _PAGE_GLOBAL and _PAGE_PROTNONE shares the
same bit, and pageattr leaves _PAGE_GLOBAL set on a kernel pte which
is then mistaken as _PAGE_PROTNONE (so pte_present returns true by
mistake and the kernel fault then gets confused and loops).
With THP the same can happen after we taught pmd_present to check
_PAGE_PROTNONE and _PAGE_PSE in commit
027ef6c87853b0a9df5317 ("mm: thp:
fix pmd_present for split_huge_page and PROT_NONE with THP"). THP has the
same problem with _PAGE_GLOBAL as the 4k pages, but it also has a problem
with _PAGE_PSE, which must be cleared too.
After the patch is applied copy_user correctly returns -EFAULT and
doesn't lockup anymore.
crash> p crash_address
crash_address = $9 = (long unsigned int *) 0xffff88023c000000
crash> rd 0xffff88023c000000
rd: read error: kernel virtual address:
ffff88023c000000 type: "64-bit KVADDR"
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: "H. Peter Anvin" <hpa@linux.intel.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>