upgrade to linux 2.6.10-1.12_FC2

[linux-2.6.git] / mm / swapfile.c

diff --git a/mm/swapfile.c b/mm/swapfile.c
index 50c3a77..42288bb 100644 (file)
--- a/mm/swapfile.c
+++ b/mm/swapfile.c
@@ -25,6 +25,7 @@
 #include <linux/rmap.h>
 #include <linux/security.h>
 #include <linux/backing-dev.h>
+#include <linux/syscalls.h>
 
 #include <asm/pgtable.h>
 #include <asm/tlbflush.h>
@@ -650,11 +651,12 @@ static int try_to_unuse(unsigned int type)
         *
         * A simpler strategy would be to start at the last mm we
         * freed the previous entry from; but that would take less
-        * advantage of mmlist ordering (now preserved by swap_out()),
-        * which clusters forked address spaces together, most recent
-        * child immediately after parent.  If we race with dup_mmap(),
-        * we very much want to resolve parent before child, otherwise
-        * we may miss some entries: using last mm would invert that.
+        * advantage of mmlist ordering, which clusters forked mms
+        * together, child after parent.  If we race with dup_mmap(), we
+        * prefer to resolve parent before child, lest we miss entries
+        * duplicated after we scanned child: using last mm would invert
+        * that.  Though it's only a serious concern when an overflowed
+        * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
         */
        start_mm = &init_mm;
        atomic_inc(&init_mm.mm_users);
@@ -662,15 +664,7 @@ static int try_to_unuse(unsigned int type)
        /*
         * Keep on scanning until all entries have gone.  Usually,
         * one pass through swap_map is enough, but not necessarily:
-        * mmput() removes mm from mmlist before exit_mmap() and its
-        * zap_page_range().  That's not too bad, those entries are
-        * on their way out, and handled faster there than here.
-        * do_munmap() behaves similarly, taking the range out of mm's
-        * vma list before zap_page_range().  But unfortunately, when
-        * unmapping a part of a vma, it takes the whole out first,
-        * then reinserts what's left after (might even reschedule if
-        * open() method called) - so swap entries may be invisible
-        * to swapoff for a while, then reappear - but that is rare.
+        * there are races when an instance of an entry might be missed.
         */
        while ((i = find_next_to_unuse(si, i)) != 0) {
                if (signal_pending(current)) {
@@ -722,7 +716,7 @@ static int try_to_unuse(unsigned int type)
                wait_on_page_writeback(page);
 
                /*
-                * Remove all references to entry, without blocking.
+                * Remove all references to entry.
                 * Whenever we reach init_mm, there's no address space
                 * to search, but use it as a reminder to search shmem.
                 */
@@ -747,7 +741,10 @@ static int try_to_unuse(unsigned int type)
                        while (*swap_map > 1 && !retval &&
                                        (p = p->next) != &start_mm->mmlist) {
                                mm = list_entry(p, struct mm_struct, mmlist);
-                               atomic_inc(&mm->mm_users);
+                               if (atomic_inc_return(&mm->mm_users) == 1) {
+                                       atomic_dec(&mm->mm_users);
+                                       continue;
+                               }
                                spin_unlock(&mmlist_lock);
                                mmput(prev_mm);
                                prev_mm = mm;
@@ -860,6 +857,26 @@ static int try_to_unuse(unsigned int type)
        return retval;
 }
 
+/*
+ * After a successful try_to_unuse, if no swap is now in use, we know we
+ * can empty the mmlist.  swap_list_lock must be held on entry and exit.
+ * Note that mmlist_lock nests inside swap_list_lock, and an mm must be
+ * added to the mmlist just after page_duplicate - before would be racy.
+ */
+static void drain_mmlist(void)
+{
+       struct list_head *p, *next;
+       unsigned int i;
+
+       for (i = 0; i < nr_swapfiles; i++)
+               if (swap_info[i].inuse_pages)
+                       return;
+       spin_lock(&mmlist_lock);
+       list_for_each_safe(p, next, &init_mm.mmlist)
+               list_del_init(p);
+       spin_unlock(&mmlist_lock);
+}
+
 /*
  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
  * corresponds to page offset `offset'.
@@ -1174,6 +1191,7 @@ asmlinkage long sys_swapoff(const char __user * specialfile)
        }
        down(&swapon_sem);
        swap_list_lock();
+       drain_mmlist();
        swap_device_lock(p);
        swap_file = p->swap_file;
        p->swap_file = NULL;
@@ -1359,7 +1377,7 @@ asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
        p->highest_bit = 0;
        p->cluster_nr = 0;
        p->inuse_pages = 0;
-       p->sdev_lock = SPIN_LOCK_UNLOCKED;
+       spin_lock_init(&p->sdev_lock);
        p->next = -1;
        if (swap_flags & SWAP_FLAG_PREFER) {
                p->prio =