Unionfs: avoid using drop_pagecache_sb in remount

Exporting drop_pagecache_sb to modules is somewhat risky because one cannot
sleep inside invalidate_mapping_pages.  This could cause a lot of latency in
the pre-emption code.  So don't export this symbol to minimize the risk that
others will use it.

Instead, unionfs will try to directly invalidate as many pages it can from
the unionfs_remount code.  Invalidating those inode pages is not strictly
required, but helpful in encouraging a revalidation of inodes sooner than
waiting for individual f/s ops to access the union.  Since a remount is
already an expensive but rare operation, this inode pages invalidation
shouldn't add too much overhead.

CC: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Erez Zadok <ezk@cs.sunysb.edu>

diff --git a/fs/drop_caches.c b/fs/drop_caches.c
index 51baa6d664a277d71e6dd999c01d2a0819ee8e9a..4e4762389bdc560d993dfb42b3db9b664396e77f 100644 (file)
--- a/fs/drop_caches.c
+++ b/fs/drop_caches.c
@@ -3,7 +3,6 @@
  */
 
 #include <linux/kernel.h>
-#include <linux/module.h>
 #include <linux/mm.h>
 #include <linux/fs.h>
 #include <linux/writeback.h>
@@ -13,7 +12,7 @@
 /* A global variable is a bit ugly, but it keeps the code simple */
 int sysctl_drop_caches;
 
-void drop_pagecache_sb(struct super_block *sb)
+static void drop_pagecache_sb(struct super_block *sb)
 {
        struct inode *inode;
 
@@ -25,7 +24,6 @@ void drop_pagecache_sb(struct super_block *sb)
        }
        spin_unlock(&inode_lock);
 }
-EXPORT_SYMBOL(drop_pagecache_sb);
 
 void drop_pagecache(void)
 {

diff --git a/fs/unionfs/dentry.c b/fs/unionfs/dentry.c
index 951ca073bea2387388a9575bfff9bb04d8adf21c..fd7fdff6d2190e8fde79ac02cb364ba515775bf1 100644 (file)
--- a/fs/unionfs/dentry.c
+++ b/fs/unionfs/dentry.c
@@ -269,6 +269,17 @@ static inline void purge_inode_data(struct inode *inode)
                truncate_inode_pages(&inode->i_data, 0);
 }
 
+void purge_sb_data(struct super_block *sb)
+{
+       struct inode *inode;
+
+       list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
+               if (inode->i_state & (I_FREEING|I_WILL_FREE))
+                       continue;
+               purge_inode_data(inode);
+       }
+}
+
 /*
  * Revalidate a parent chain of dentries, then the actual node.
  * Assumes that dentry is locked, but will lock all parents if/when needed.

diff --git a/fs/unionfs/super.c b/fs/unionfs/super.c
index 01dc7f067722189f85c88b754177b3626e000199..4902af6c222b0a3b7a2e02bdb45accd2b2078d71 100644 (file)
--- a/fs/unionfs/super.c
+++ b/fs/unionfs/super.c
@@ -704,29 +704,19 @@ out_no_change:
         */
 
        /*
-        * Now we call drop_pagecache_sb() to invalidate all pages in this
-        * super.  This function calls invalidate_inode_pages(mapping),
-        * which calls invalidate_mapping_pages(): the latter, however, will
-        * not invalidate pages which are dirty, locked, under writeback, or
-        * mapped into page tables.  We shouldn't have to worry about dirty
-        * or under-writeback pages, because do_remount_sb() called
-        * fsync_super() which would not have returned until all dirty pages
-        * were flushed.
-        *
-        * But do we have to worry about locked pages?  Is there any chance
-        * that in here we'll get locked pages?
-        *
-        * XXX: what about pages mapped into pagetables?  Are these pages
-        * which user processes may have mmap(2)'ed?  If so, then we need to
-        * invalidate those too, no?  Maybe we'll have to write our own
-        * version of invalidate_mapping_pages() which also handled mapped
-        * pages.
-        *
-        * XXX: Alternatively, maybe we should call truncate_inode_pages(),
-        * which use two passes over the pages list, and will truncate all
-        * pages.
+        * Once we finish the remounting successfully, our superblock
+        * generation number will have increased.  This will be detected by
+        * our dentry-revalidation code upon subsequent f/s operations
+        * through unionfs.  The revalidation code will rebuild the union of
+        * lower inodes for a given unionfs inode and invalidate any pages
+        * of such "stale" inodes (by calling our purge_inode_data
+        * function).  This revalidation will happen lazily and
+        * incrementally, as users perform operations on cached inodes.  We
+        * would like to encourage this revalidation to happen sooner if
+        * possible, so we try to invalidate as many other pages in our
+        * superblock as we can.
         */
-       drop_pagecache_sb(sb);
+       purge_sb_data(sb);
 
        /* copy new vectors into their correct place */
        tmp_data = UNIONFS_SB(sb)->data;

diff --git a/fs/unionfs/union.h b/fs/unionfs/union.h
index e701b39c3014ceeb8d2cf1a464a7255415f709ac..fe11a23431c505482b4d8d954adcad4f7ca30477 100644 (file)
--- a/fs/unionfs/union.h
+++ b/fs/unionfs/union.h
@@ -373,6 +373,7 @@ extern int unionfs_rmdir(struct inode *dir, struct dentry *dentry);
 extern bool __unionfs_d_revalidate_chain(struct dentry *dentry,
                                         struct nameidata *nd, bool willwrite);
 extern bool is_newer_lower(const struct dentry *dentry);
+extern void purge_sb_data(struct super_block *sb);
 
 /* The values for unionfs_interpose's flag. */
 #define INTERPOSE_DEFAULT      0

diff --git a/include/linux/mm.h b/include/linux/mm.h
index 7f58dc16eb8972b8d16cc440ff0528f6a9dc67a7..f0b135cd86da1e54789bce46eb7bb8f6487edcaa 100644 (file)
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -1062,7 +1062,6 @@ int drop_caches_sysctl_handler(struct ctl_table *, int, struct file *,
                                        void __user *, size_t *, loff_t *);
 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
                        unsigned long lru_pages);
-extern void drop_pagecache_sb(struct super_block *);
 void drop_pagecache(void);
 void drop_slab(void);