x86/mm/pat: Improve scaling of pat_pagerange_is_ram()
authorJohn Dykstra <jdykstra@cray.com>
Fri, 25 May 2012 21:12:46 +0000 (16:12 -0500)
committerIngo Molnar <mingo@kernel.org>
Wed, 30 May 2012 08:57:11 +0000 (10:57 +0200)
Function pat_pagerange_is_ram() scales poorly to large address
ranges, because it probes the resource tree for each page.

On a 2.6 GHz Opteron, this function consumes 34 ms for a 1 GB range.

It is called twice during untrack_pfn_vma(), slowing process
cleanup and handicapping the OOM killer.

This replacement consumes less than 1ms, under the same conditions.

Signed-off-by: John Dykstra <jdykstra@cray.com> on behalf of Cray Inc.
Acked-by: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1337980366.1979.6.camel@redwood
[ Small stylistic cleanups and renames ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
arch/x86/mm/pat.c

index f6ff57b..bea6e57 100644 (file)
@@ -158,31 +158,47 @@ static unsigned long pat_x_mtrr_type(u64 start, u64 end, unsigned long req_type)
        return req_type;
 }
 
+struct pagerange_state {
+       unsigned long           cur_pfn;
+       int                     ram;
+       int                     not_ram;
+};
+
+static int
+pagerange_is_ram_callback(unsigned long initial_pfn, unsigned long total_nr_pages, void *arg)
+{
+       struct pagerange_state *state = arg;
+
+       state->not_ram  |= initial_pfn > state->cur_pfn;
+       state->ram      |= total_nr_pages > 0;
+       state->cur_pfn   = initial_pfn + total_nr_pages;
+
+       return state->ram && state->not_ram;
+}
+
 static int pat_pagerange_is_ram(resource_size_t start, resource_size_t end)
 {
-       int ram_page = 0, not_rampage = 0;
-       unsigned long page_nr;
+       int ret = 0;
+       unsigned long start_pfn = start >> PAGE_SHIFT;
+       unsigned long end_pfn = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
+       struct pagerange_state state = {start_pfn, 0, 0};
 
-       for (page_nr = (start >> PAGE_SHIFT); page_nr < (end >> PAGE_SHIFT);
-            ++page_nr) {
-               /*
-                * For legacy reasons, physical address range in the legacy ISA
-                * region is tracked as non-RAM. This will allow users of
-                * /dev/mem to map portions of legacy ISA region, even when
-                * some of those portions are listed(or not even listed) with
-                * different e820 types(RAM/reserved/..)
-                */
-               if (page_nr >= (ISA_END_ADDRESS >> PAGE_SHIFT) &&
-                   page_is_ram(page_nr))
-                       ram_page = 1;
-               else
-                       not_rampage = 1;
-
-               if (ram_page == not_rampage)
-                       return -1;
+       /*
+        * For legacy reasons, physical address range in the legacy ISA
+        * region is tracked as non-RAM. This will allow users of
+        * /dev/mem to map portions of legacy ISA region, even when
+        * some of those portions are listed(or not even listed) with
+        * different e820 types(RAM/reserved/..)
+        */
+       if (start_pfn < ISA_END_ADDRESS >> PAGE_SHIFT)
+               start_pfn = ISA_END_ADDRESS >> PAGE_SHIFT;
+
+       if (start_pfn < end_pfn) {
+               ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn,
+                               &state, pagerange_is_ram_callback);
        }
 
-       return ram_page;
+       return (ret > 0) ? -1 : (state.ram ? 1 : 0);
 }
 
 /*