[PATCH] mm: fix rss and mmlist locking
Hugh Dickins [Sun, 30 Oct 2005 01:16:41 +0000 (18:16 -0700)]
A couple of oddities were guarded by page_table_lock, no longer properly
guarded when that is split.

The mm_counters of file_rss and anon_rss: make those an atomic_t, or an
atomic64_t if the architecture supports it, in such a case.  Definitions by
courtesy of Christoph Lameter: who spent considerable effort on more scalable
ways of counting, but found insufficient benefit in practice.

And adding an mm with swap to the mmlist for swapoff: the list is well-
guarded by its own lock, but the list_empty check now has to be repeated
inside it.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

include/linux/sched.h
mm/memory.c
mm/rmap.c

index 292cb57..1c30bc3 100644 (file)
@@ -249,13 +249,47 @@ arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
 extern void arch_unmap_area(struct mm_struct *, unsigned long);
 extern void arch_unmap_area_topdown(struct mm_struct *, unsigned long);
 
+#if NR_CPUS >= CONFIG_SPLIT_PTLOCK_CPUS
+/*
+ * The mm counters are not protected by its page_table_lock,
+ * so must be incremented atomically.
+ */
+#ifdef ATOMIC64_INIT
+#define set_mm_counter(mm, member, value) atomic64_set(&(mm)->_##member, value)
+#define get_mm_counter(mm, member) ((unsigned long)atomic64_read(&(mm)->_##member))
+#define add_mm_counter(mm, member, value) atomic64_add(value, &(mm)->_##member)
+#define inc_mm_counter(mm, member) atomic64_inc(&(mm)->_##member)
+#define dec_mm_counter(mm, member) atomic64_dec(&(mm)->_##member)
+typedef atomic64_t mm_counter_t;
+#else /* !ATOMIC64_INIT */
+/*
+ * The counters wrap back to 0 at 2^32 * PAGE_SIZE,
+ * that is, at 16TB if using 4kB page size.
+ */
+#define set_mm_counter(mm, member, value) atomic_set(&(mm)->_##member, value)
+#define get_mm_counter(mm, member) ((unsigned long)atomic_read(&(mm)->_##member))
+#define add_mm_counter(mm, member, value) atomic_add(value, &(mm)->_##member)
+#define inc_mm_counter(mm, member) atomic_inc(&(mm)->_##member)
+#define dec_mm_counter(mm, member) atomic_dec(&(mm)->_##member)
+typedef atomic_t mm_counter_t;
+#endif /* !ATOMIC64_INIT */
+
+#else  /* NR_CPUS < CONFIG_SPLIT_PTLOCK_CPUS */
+/*
+ * The mm counters are protected by its page_table_lock,
+ * so can be incremented directly.
+ */
 #define set_mm_counter(mm, member, value) (mm)->_##member = (value)
 #define get_mm_counter(mm, member) ((mm)->_##member)
 #define add_mm_counter(mm, member, value) (mm)->_##member += (value)
 #define inc_mm_counter(mm, member) (mm)->_##member++
 #define dec_mm_counter(mm, member) (mm)->_##member--
-#define get_mm_rss(mm) ((mm)->_file_rss + (mm)->_anon_rss)
+typedef unsigned long mm_counter_t;
+
+#endif /* NR_CPUS < CONFIG_SPLIT_PTLOCK_CPUS */
 
+#define get_mm_rss(mm)                                 \
+       (get_mm_counter(mm, file_rss) + get_mm_counter(mm, anon_rss))
 #define update_hiwater_rss(mm) do {                    \
        unsigned long _rss = get_mm_rss(mm);            \
        if ((mm)->hiwater_rss < _rss)                   \
@@ -266,8 +300,6 @@ extern void arch_unmap_area_topdown(struct mm_struct *, unsigned long);
                (mm)->hiwater_vm = (mm)->total_vm;      \
 } while (0)
 
-typedef unsigned long mm_counter_t;
-
 struct mm_struct {
        struct vm_area_struct * mmap;           /* list of VMAs */
        struct rb_root mm_rb;
@@ -291,7 +323,9 @@ struct mm_struct {
                                                 * by mmlist_lock
                                                 */
 
-       /* Special counters protected by the page_table_lock */
+       /* Special counters, in some configurations protected by the
+        * page_table_lock, in other configurations by being atomic.
+        */
        mm_counter_t _file_rss;
        mm_counter_t _anon_rss;
 
index e9ef599..d68421d 100644 (file)
@@ -372,7 +372,9 @@ copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
                        /* make sure dst_mm is on swapoff's mmlist. */
                        if (unlikely(list_empty(&dst_mm->mmlist))) {
                                spin_lock(&mmlist_lock);
-                               list_add(&dst_mm->mmlist, &src_mm->mmlist);
+                               if (list_empty(&dst_mm->mmlist))
+                                       list_add(&dst_mm->mmlist,
+                                                &src_mm->mmlist);
                                spin_unlock(&mmlist_lock);
                        }
                }
index a33e779..a7427bb 100644 (file)
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -559,7 +559,8 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
                swap_duplicate(entry);
                if (list_empty(&mm->mmlist)) {
                        spin_lock(&mmlist_lock);
-                       list_add(&mm->mmlist, &init_mm.mmlist);
+                       if (list_empty(&mm->mmlist))
+                               list_add(&mm->mmlist, &init_mm.mmlist);
                        spin_unlock(&mmlist_lock);
                }
                set_pte_at(mm, address, pte, swp_entry_to_pte(entry));