x86/mm/ptdump: Optimize check for W+X mappings for CONFIG_KASAN=y

Enabling both DEBUG_WX=y and KASAN=y options significantly increases
boot time (dozens of seconds at least).
KASAN fills kernel page tables with repeated values to map several
TBs of the virtual memory to the single kasan_zero_page:

    kasan_zero_pud ->
        kasan_zero_pmd->
            kasan_zero_pte->
                kasan_zero_page

So, the page table walker used to find W+X mapping check the same
kasan_zero_p?d page table entries a lot more than once.
With patch pud walker will skip the pud if it has the same value as
the previous one . Skipping done iff we search for W+X mappings,
so this optimization won't affect the page table dump via debugfs.

This dropped time spend in W+X check from ~30 sec to reasonable 0.1 sec:

Before:
[    4.579991] Freeing unused kernel memory: 1000K
[   35.257523] x86/mm: Checked W+X mappings: passed, no W+X pages found.

After:
[    5.138756] Freeing unused kernel memory: 1000K
[    5.266496] x86/mm: Checked W+X mappings: passed, no W+X pages found.

Signed-off-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Reviewed-by: Alexander Potapenko <glider@google.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: kasan-dev@googlegroups.com
Cc: Tobias Regnery <tobias.regnery@gmail.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Link: http://lkml.kernel.org/r/20170214100839.17186-1-aryabinin@virtuozzo.com
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>

diff --git a/arch/x86/mm/dump_pagetables.c b/arch/x86/mm/dump_pagetables.c
index 8aa6bea1cd6cc467cd5d39dbcde7d625c009495c..08135341798cbb2fc05097d1bc8bdfed7eff5498 100644 (file)
--- a/arch/x86/mm/dump_pagetables.c
+++ b/arch/x86/mm/dump_pagetables.c
@@ -327,18 +327,31 @@ static void walk_pmd_level(struct seq_file *m, struct pg_state *st, pud_t addr,
 
 #if PTRS_PER_PUD > 1
 
+/*
+ * This is an optimization for CONFIG_DEBUG_WX=y + CONFIG_KASAN=y
+ * KASAN fills page tables with the same values. Since there is no
+ * point in checking page table more than once we just skip repeated
+ * entries. This saves us dozens of seconds during boot.
+ */
+static bool pud_already_checked(pud_t *prev_pud, pud_t *pud, bool checkwx)
+{
+       return checkwx && prev_pud && (pud_val(*prev_pud) == pud_val(*pud));
+}
+
 static void walk_pud_level(struct seq_file *m, struct pg_state *st, pgd_t addr,
                                                        unsigned long P)
 {
        int i;
        pud_t *start;
        pgprotval_t prot;
+       pud_t *prev_pud = NULL;
 
        start = (pud_t *) pgd_page_vaddr(addr);
 
        for (i = 0; i < PTRS_PER_PUD; i++) {
                st->current_address = normalize_addr(P + i * PUD_LEVEL_MULT);
-               if (!pud_none(*start)) {
+               if (!pud_none(*start) &&
+                   !pud_already_checked(prev_pud, start, st->check_wx)) {
                        if (pud_large(*start) || !pud_present(*start)) {
                                prot = pud_flags(*start);
                                note_page(m, st, __pgprot(prot), 2);
@@ -349,6 +362,7 @@ static void walk_pud_level(struct seq_file *m, struct pg_state *st, pgd_t addr,
                } else
                        note_page(m, st, __pgprot(0), 2);
 
+               prev_pud = start;
                start++;
        }
 }