[PATCH] O(1) sb list traversing on syncs
authorKirill Korotaev <dev@sw.ru>
Thu, 23 Jun 2005 07:09:54 +0000 (00:09 -0700)
committerLinus Torvalds <torvalds@ppc970.osdl.org>
Thu, 23 Jun 2005 16:45:27 +0000 (09:45 -0700)
This patch removes O(n^2) super block loops in sync_inodes(),
sync_filesystems() etc.  in favour of using __put_super_and_need_restart()
which I introduced earlier.  We faced a noticably long freezes on sb
syncing when there are thousands of super blocks in the system.

Signed-Off-By: Kirill Korotaev <dev@sw.ru>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
fs/fs-writeback.c
fs/quota.c
fs/super.c

index 8e050fa58218cdac75ad6a49771a224e3d7b7c61..e94ab398b71763b3caf598dfe8fbeb66e818789d 100644 (file)
@@ -485,32 +485,6 @@ static void set_sb_syncing(int val)
        spin_unlock(&sb_lock);
 }
 
-/*
- * Find a superblock with inodes that need to be synced
- */
-static struct super_block *get_super_to_sync(void)
-{
-       struct super_block *sb;
-restart:
-       spin_lock(&sb_lock);
-       sb = sb_entry(super_blocks.prev);
-       for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
-               if (sb->s_syncing)
-                       continue;
-               sb->s_syncing = 1;
-               sb->s_count++;
-               spin_unlock(&sb_lock);
-               down_read(&sb->s_umount);
-               if (!sb->s_root) {
-                       drop_super(sb);
-                       goto restart;
-               }
-               return sb;
-       }
-       spin_unlock(&sb_lock);
-       return NULL;
-}
-
 /**
  * sync_inodes - writes all inodes to disk
  * @wait: wait for completion
@@ -530,23 +504,39 @@ restart:
  * outstanding dirty inodes, the writeback goes block-at-a-time within the
  * filesystem's write_inode().  This is extremely slow.
  */
-void sync_inodes(int wait)
+static void __sync_inodes(int wait)
 {
        struct super_block *sb;
 
-       set_sb_syncing(0);
-       while ((sb = get_super_to_sync()) != NULL) {
-               sync_inodes_sb(sb, 0);
-               sync_blockdev(sb->s_bdev);
-               drop_super(sb);
+       spin_lock(&sb_lock);
+restart:
+       list_for_each_entry(sb, &super_blocks, s_list) {
+               if (sb->s_syncing)
+                       continue;
+               sb->s_syncing = 1;
+               sb->s_count++;
+               spin_unlock(&sb_lock);
+               down_read(&sb->s_umount);
+               if (sb->s_root) {
+                       sync_inodes_sb(sb, wait);
+                       sync_blockdev(sb->s_bdev);
+               }
+               up_read(&sb->s_umount);
+               spin_lock(&sb_lock);
+               if (__put_super_and_need_restart(sb))
+                       goto restart;
        }
+       spin_unlock(&sb_lock);
+}
+
+void sync_inodes(int wait)
+{
+       set_sb_syncing(0);
+       __sync_inodes(0);
+
        if (wait) {
                set_sb_syncing(0);
-               while ((sb = get_super_to_sync()) != NULL) {
-                       sync_inodes_sb(sb, 1);
-                       sync_blockdev(sb->s_bdev);
-                       drop_super(sb);
-               }
+               __sync_inodes(1);
        }
 }
 
index 3f0333a51a23d122b71ac6d62d811f301ff7bb29..f5d1cff551965ba30fa88d12d50d3a597c9a1245 100644 (file)
@@ -149,36 +149,6 @@ static int check_quotactl_valid(struct super_block *sb, int type, int cmd, qid_t
        return error;
 }
 
-static struct super_block *get_super_to_sync(int type)
-{
-       struct list_head *head;
-       int cnt, dirty;
-
-restart:
-       spin_lock(&sb_lock);
-       list_for_each(head, &super_blocks) {
-               struct super_block *sb = list_entry(head, struct super_block, s_list);
-
-               /* This test just improves performance so it needn't be reliable... */
-               for (cnt = 0, dirty = 0; cnt < MAXQUOTAS; cnt++)
-                       if ((type == cnt || type == -1) && sb_has_quota_enabled(sb, cnt)
-                           && info_any_dirty(&sb_dqopt(sb)->info[cnt]))
-                               dirty = 1;
-               if (!dirty)
-                       continue;
-               sb->s_count++;
-               spin_unlock(&sb_lock);
-               down_read(&sb->s_umount);
-               if (!sb->s_root) {
-                       drop_super(sb);
-                       goto restart;
-               }
-               return sb;
-       }
-       spin_unlock(&sb_lock);
-       return NULL;
-}
-
 static void quota_sync_sb(struct super_block *sb, int type)
 {
        int cnt;
@@ -219,17 +189,35 @@ static void quota_sync_sb(struct super_block *sb, int type)
 
 void sync_dquots(struct super_block *sb, int type)
 {
+       int cnt, dirty;
+
        if (sb) {
                if (sb->s_qcop->quota_sync)
                        quota_sync_sb(sb, type);
+               return;
        }
-       else {
-               while ((sb = get_super_to_sync(type)) != NULL) {
-                       if (sb->s_qcop->quota_sync)
-                               quota_sync_sb(sb, type);
-                       drop_super(sb);
-               }
+
+       spin_lock(&sb_lock);
+restart:
+       list_for_each_entry(sb, &super_blocks, s_list) {
+               /* This test just improves performance so it needn't be reliable... */
+               for (cnt = 0, dirty = 0; cnt < MAXQUOTAS; cnt++)
+                       if ((type == cnt || type == -1) && sb_has_quota_enabled(sb, cnt)
+                           && info_any_dirty(&sb_dqopt(sb)->info[cnt]))
+                               dirty = 1;
+               if (!dirty)
+                       continue;
+               sb->s_count++;
+               spin_unlock(&sb_lock);
+               down_read(&sb->s_umount);
+               if (sb->s_root && sb->s_qcop->quota_sync)
+                       quota_sync_sb(sb, type);
+               up_read(&sb->s_umount);
+               spin_lock(&sb_lock);
+               if (__put_super_and_need_restart(sb))
+                       goto restart;
        }
+       spin_unlock(&sb_lock);
 }
 
 /* Copy parameters and call proper function */
index 573bcc81bb82ae25caa4de0d0b779d562c87535b..25bc1ec6bc5dfd0e8348b4c12d6c82acd25246fb 100644 (file)
@@ -341,20 +341,22 @@ static inline void write_super(struct super_block *sb)
  */
 void sync_supers(void)
 {
-       struct super_block * sb;
-restart:
+       struct super_block *sb;
+
        spin_lock(&sb_lock);
-       sb = sb_entry(super_blocks.next);
-       while (sb != sb_entry(&super_blocks))
+restart:
+       list_for_each_entry(sb, &super_blocks, s_list) {
                if (sb->s_dirt) {
                        sb->s_count++;
                        spin_unlock(&sb_lock);
                        down_read(&sb->s_umount);
                        write_super(sb);
-                       drop_super(sb);
-                       goto restart;
-               } else
-                       sb = sb_entry(sb->s_list.next);
+                       up_read(&sb->s_umount);
+                       spin_lock(&sb_lock);
+                       if (__put_super_and_need_restart(sb))
+                               goto restart;
+               }
+       }
        spin_unlock(&sb_lock);
 }
 
@@ -381,20 +383,16 @@ void sync_filesystems(int wait)
 
        down(&mutex);           /* Could be down_interruptible */
        spin_lock(&sb_lock);
-       for (sb = sb_entry(super_blocks.next); sb != sb_entry(&super_blocks);
-                       sb = sb_entry(sb->s_list.next)) {
+       list_for_each_entry(sb, &super_blocks, s_list) {
                if (!sb->s_op->sync_fs)
                        continue;
                if (sb->s_flags & MS_RDONLY)
                        continue;
                sb->s_need_sync_fs = 1;
        }
-       spin_unlock(&sb_lock);
 
 restart:
-       spin_lock(&sb_lock);
-       for (sb = sb_entry(super_blocks.next); sb != sb_entry(&super_blocks);
-                       sb = sb_entry(sb->s_list.next)) {
+       list_for_each_entry(sb, &super_blocks, s_list) {
                if (!sb->s_need_sync_fs)
                        continue;
                sb->s_need_sync_fs = 0;
@@ -405,8 +403,11 @@ restart:
                down_read(&sb->s_umount);
                if (sb->s_root && (wait || sb->s_dirt))
                        sb->s_op->sync_fs(sb, wait);
-               drop_super(sb);
-               goto restart;
+               up_read(&sb->s_umount);
+               /* restart only when sb is no longer on the list */
+               spin_lock(&sb_lock);
+               if (__put_super_and_need_restart(sb))
+                       goto restart;
        }
        spin_unlock(&sb_lock);
        up(&mutex);
@@ -422,21 +423,25 @@ restart:
 
 struct super_block * get_super(struct block_device *bdev)
 {
-       struct list_head *p;
+       struct super_block *sb;
+
        if (!bdev)
                return NULL;
-rescan:
+
        spin_lock(&sb_lock);
-       list_for_each(p, &super_blocks) {
-               struct super_block *s = sb_entry(p);
-               if (s->s_bdev == bdev) {
-                       s->s_count++;
+rescan:
+       list_for_each_entry(sb, &super_blocks, s_list) {
+               if (sb->s_bdev == bdev) {
+                       sb->s_count++;
                        spin_unlock(&sb_lock);
-                       down_read(&s->s_umount);
-                       if (s->s_root)
-                               return s;
-                       drop_super(s);
-                       goto rescan;
+                       down_read(&sb->s_umount);
+                       if (sb->s_root)
+                               return sb;
+                       up_read(&sb->s_umount);
+                       /* restart only when sb is no longer on the list */
+                       spin_lock(&sb_lock);
+                       if (__put_super_and_need_restart(sb))
+                               goto rescan;
                }
        }
        spin_unlock(&sb_lock);
@@ -447,20 +452,22 @@ EXPORT_SYMBOL(get_super);
  
 struct super_block * user_get_super(dev_t dev)
 {
-       struct list_head *p;
+       struct super_block *sb;
 
-rescan:
        spin_lock(&sb_lock);
-       list_for_each(p, &super_blocks) {
-               struct super_block *s = sb_entry(p);
-               if (s->s_dev ==  dev) {
-                       s->s_count++;
+rescan:
+       list_for_each_entry(sb, &super_blocks, s_list) {
+               if (sb->s_dev ==  dev) {
+                       sb->s_count++;
                        spin_unlock(&sb_lock);
-                       down_read(&s->s_umount);
-                       if (s->s_root)
-                               return s;
-                       drop_super(s);
-                       goto rescan;
+                       down_read(&sb->s_umount);
+                       if (sb->s_root)
+                               return sb;
+                       up_read(&sb->s_umount);
+                       /* restart only when sb is no longer on the list */
+                       spin_lock(&sb_lock);
+                       if (__put_super_and_need_restart(sb))
+                               goto rescan;
                }
        }
        spin_unlock(&sb_lock);