Fix problems with new write_record() interface:

- we were not incrementing the size properly
  - block > i_size >> blkbits was never true when it needed to be
  - new size was always 1 block too small
- when size was set, it didn't make it to disk, because
  ext3_get_block_handle() updated size after marking the inode dirty

Proper fix is to move i_disksize update into ext3_splice_branch(), so that
it is done before inode is marked dirty.  For now, we mark inode dirty
again in fsfilt, so we don't need a new kernel for Wed LLNL debug shot.

Make fsfilt_ext3 and fsfilt_extN identical again (cosmetic changes only).

diff --git a/lustre/include/linux/lustre_fsfilt.h b/lustre/include/linux/lustre_fsfilt.h
index 17d5f76..97c7d2e 100644 (file)
--- a/lustre/include/linux/lustre_fsfilt.h
+++ b/lustre/include/linux/lustre_fsfilt.h
@@ -61,8 +61,8 @@ struct fsfilt_operations {
         int     (* fs_sync)(struct super_block *sb);
         int     (* fs_prep_san_write)(struct inode *inode, long *blocks,
                                       int nblocks, loff_t newsize);
-        int     (* fs_write_record)(struct file *, char *, loff_t, loff_t *);
-        int     (* fs_read_record)(struct file *, char *, loff_t, loff_t *);
+        int     (* fs_write_record)(struct file *, char *, int size, loff_t *);
+        int     (* fs_read_record)(struct file *, char *, int size, loff_t *);
 };
 
 extern int fsfilt_register_ops(struct fsfilt_operations *fs_ops);

diff --git a/lustre/obdclass/fsfilt_ext3.c b/lustre/obdclass/fsfilt_ext3.c
index 0a4cb1e..82c2f58 100644 (file)
--- a/lustre/obdclass/fsfilt_ext3.c
+++ b/lustre/obdclass/fsfilt_ext3.c
@@ -501,8 +501,8 @@ static int fsfilt_ext3_prep_san_write(struct inode *inode, long *blocks,
         return ext3_prep_san_write(inode, blocks, nblocks, newsize);
 }
 
-static int fsfilt_ext3_read_record(struct file * file, char * buf,
-                loff_t size, loff_t *offs)
+static int fsfilt_ext3_read_record(struct file * file, char *buf,
+                                   int size, loff_t *offs)
 {
         struct buffer_head *bh;
         unsigned long block, boffs;
@@ -510,9 +510,10 @@ static int fsfilt_ext3_read_record(struct file * file, char * buf,
         int err;
 
         if (inode->i_size < *offs + size) {
-                CERROR("file is too short for this request\n");
+                CERROR("file size %llu is too short for read %u@%llu\n",
+                       inode->i_size, size, *offs);
                 return -EIO;
-        } 
+        }
 
         block = *offs >> inode->i_blkbits;
         bh = ext3_bread(NULL, inode, block, 0, &err);
@@ -535,12 +536,13 @@ static int fsfilt_ext3_read_record(struct file * file, char * buf,
         return size;
 }
 
-static int fsfilt_ext3_write_record(struct file * file, char * buf,
-                loff_t size, loff_t *offs)
+static int fsfilt_ext3_write_record(struct file * file, char *buf,
+                                    int size, loff_t *offs)
 {
         struct buffer_head *bh;
         unsigned long block, boffs;
         struct inode *inode = file->f_dentry->d_inode;
+        loff_t old_size = inode->i_size;
         journal_t *journal;
         handle_t *handle;
         int err;
@@ -553,18 +555,27 @@ static int fsfilt_ext3_write_record(struct file * file, char * buf,
         }
 
         block = *offs >> inode->i_blkbits;
-        if (block > inode->i_size >> inode->i_blkbits) {
+        if (*offs + size > inode->i_size) {
                 down(&inode->i_sem);
-                if (block > inode->i_size >> inode->i_blkbits) 
-                        inode->i_size = block << inode->i_blkbits;
+                if (*offs + size > inode->i_size)
+                        inode->i_size = ((loff_t)block + 1) << inode->i_blkbits;
                 up(&inode->i_sem);
         }
+
         bh = ext3_bread(handle, inode, block, 1, &err);
         if (!bh) {
                 CERROR("can't read/create block: %d\n", err);
                 goto out;
         }
 
+        /* This is a hack only needed because ext3_get_block_handle() updates
+         * i_disksize after marking the inode dirty in ext3_splice_branch().
+         * We will fix that when we get a chance, as ext3_mark_inode_dirty()
+         * is not without cost, nor is it even exported.
+         */
+        if (inode->i_size > old_size)
+                mark_inode_dirty(inode);
+
         boffs = (unsigned)*offs % bh->b_size;
         if (boffs + size > bh->b_size) {
                 CERROR("request crosses block's border. offset %lu, size %lu\n",
@@ -582,7 +593,7 @@ static int fsfilt_ext3_write_record(struct file * file, char * buf,
         err = ext3_journal_dirty_metadata(handle, bh);
         if (err) {
                 CERROR("journal_dirty_metadata() returned error %d\n", err);
-                goto out; 
+                goto out;
         }
         err = size;
 out: