]> git.karo-electronics.de Git - karo-tx-linux.git/commitdiff
pageattr: prevent PSE and GLOABL leftovers to confuse pmd/pte_present and pmd_huge
authorAndrea Arcangeli <aarcange@redhat.com>
Thu, 7 Feb 2013 01:25:50 +0000 (12:25 +1100)
committerStephen Rothwell <sfr@canb.auug.org.au>
Mon, 18 Feb 2013 05:45:21 +0000 (16:45 +1100)
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>
arch/x86/mm/pageattr.c

index a1b1c88f9caf5f6c1eaf6a8b6a95477dad36a5f8..4a3de7ce0bf2382e050efc9a516c27bd719206aa 100644 (file)
@@ -472,6 +472,19 @@ try_preserve_large_page(pte_t *kpte, unsigned long address,
        pgprot_val(req_prot) &= ~pgprot_val(cpa->mask_clr);
        pgprot_val(req_prot) |= pgprot_val(cpa->mask_set);
 
+       /*
+        * Set the PSE and GLOBAL flags only if the PRESENT flag is
+        * set otherwise pmd_present/pmd_huge will return true even on
+        * a non present pmd. The canon_pgprot will clear _PAGE_GLOBAL
+        * for the ancient hardware that doesn't support it.
+        */
+       if (pgprot_val(new_prot) & _PAGE_PRESENT)
+               pgprot_val(new_prot) |= _PAGE_PSE | _PAGE_GLOBAL;
+       else
+               pgprot_val(new_prot) &= ~(_PAGE_PSE | _PAGE_GLOBAL);
+
+       new_prot = canon_pgprot(new_prot);
+
        /*
         * old_pte points to the large page base address. So we need
         * to add the offset of the virtual address:
@@ -517,7 +530,7 @@ try_preserve_large_page(pte_t *kpte, unsigned long address,
                 * The address is aligned and the number of pages
                 * covers the full page.
                 */
-               new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
+               new_pte = pfn_pte(pte_pfn(old_pte), new_prot);
                __set_pmd_pte(kpte, address, new_pte);
                cpa->flags |= CPA_FLUSHTLB;
                do_split = 0;
@@ -568,16 +581,35 @@ static int split_large_page(pte_t *kpte, unsigned long address)
 #ifdef CONFIG_X86_64
        if (level == PG_LEVEL_1G) {
                pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
-               pgprot_val(ref_prot) |= _PAGE_PSE;
+               /*
+                * Set the PSE flags only if the PRESENT flag is set
+                * otherwise pmd_present/pmd_huge will return true
+                * even on a non present pmd.
+                */
+               if (pgprot_val(ref_prot) & _PAGE_PRESENT)
+                       pgprot_val(ref_prot) |= _PAGE_PSE;
+               else
+                       pgprot_val(ref_prot) &= ~_PAGE_PSE;
        }
 #endif
 
+       /*
+        * Set the GLOBAL flags only if the PRESENT flag is set
+        * otherwise pmd/pte_present will return true even on a non
+        * present pmd/pte. The canon_pgprot will clear _PAGE_GLOBAL
+        * for the ancient hardware that doesn't support it.
+        */
+       if (pgprot_val(ref_prot) & _PAGE_PRESENT)
+               pgprot_val(ref_prot) |= _PAGE_GLOBAL;
+       else
+               pgprot_val(ref_prot) &= ~_PAGE_GLOBAL;
+
        /*
         * Get the target pfn from the original entry:
         */
        pfn = pte_pfn(*kpte);
        for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
-               set_pte(&pbase[i], pfn_pte(pfn, ref_prot));
+               set_pte(&pbase[i], pfn_pte(pfn, canon_pgprot(ref_prot)));
 
        if (pfn_range_is_mapped(PFN_DOWN(__pa(address)),
                                PFN_DOWN(__pa(address)) + 1))
@@ -681,6 +713,18 @@ repeat:
 
                new_prot = static_protections(new_prot, address, pfn);
 
+               /*
+                * Set the GLOBAL flags only if the PRESENT flag is
+                * set otherwise pte_present will return true even on
+                * a non present pte. The canon_pgprot will clear
+                * _PAGE_GLOBAL for the ancient hardware that doesn't
+                * support it.
+                */
+               if (pgprot_val(new_prot) & _PAGE_PRESENT)
+                       pgprot_val(new_prot) |= _PAGE_GLOBAL;
+               else
+                       pgprot_val(new_prot) &= ~_PAGE_GLOBAL;
+
                /*
                 * We need to keep the pfn from the existing PTE,
                 * after all we're only going to change it's attributes