[RFC] AFFS: Trying to fix fsx/O_DIRECT

[Fabian Frederick <fabf@skynet.be>]

xfstests/ltp/fsx version (fsx file -d -Z -r 4096 -w 4096)
always ends up with filesystem being remounted read-only due to a
write operation done beyond truncate offset.

affs debug:
affs: error (device sdd1): get_block(): strange block request 136
affs: Remounting filesystem read-only

Fsx report:
"
129(129 mod 256): TRUNCATE DOWN from 0x3ff01 to 0xb3f6
130(130 mod 256): WRITE    0x22000 thru 0x2dfff (0xc000 bytes) HOLE
"

Error in affs_get_block:
if (block >= AFFS_I(inode)->i_blkcnt) {
                if (block > AFFS_I(inode)->i_blkcnt || !create)
                        goto err_big;
        } else
                create = 0;

When I display values there, block is 136; i_blkcnt: 45

It seems operations are mixed up and especially truncate/get_block.
I tried the patch below to add truncate mutex featuring in some other
filesystems but problem remains the same.
Is there something else I could do to avoid such problem ?

Regards,
Fabian

---
 fs/affs/affs.h  |  1 +
 fs/affs/file.c  | 10 +++++++++-
 fs/affs/super.c |  1 +
 3 files changed, 11 insertions(+), 1 deletion(-)

diff --git a/fs/affs/affs.h b/fs/affs/affs.h
index ff44ff3..c7942d9 100644
--- a/fs/affs/affs.h
+++ b/fs/affs/affs.h
@@ -56,6 +56,7 @@ struct affs_inode_info {
 	u32	 i_protect;			/* unused attribute bits */
 	u32	 i_lastalloc;			/* last allocated block */
 	int	 i_pa_cnt;			/* number of preallocated blocks */
+	struct mutex truncate_mutex;
 	struct inode vfs_inode;
 };
 
diff --git a/fs/affs/file.c b/fs/affs/file.c
index 8faa659..7e1ab3e 100644
--- a/fs/affs/file.c
+++ b/fs/affs/file.c
@@ -321,7 +321,10 @@ affs_get_block(struct inode *inode, sector_t block, struct buffer_head *bh_resul
 	map_bh(bh_result, sb, (sector_t)be32_to_cpu(AFFS_BLOCK(sb, ext_bh, block)));
 
 	if (create) {
-		u32 blocknr = affs_alloc_block(inode, ext_bh->b_blocknr);
+		u32 blocknr;
+
+		mutex_lock(&AFFS_I(inode)->truncate_mutex);
+		blocknr = affs_alloc_block(inode, ext_bh->b_blocknr);
 		if (!blocknr)
 			goto err_alloc;
 		set_buffer_new(bh_result);
@@ -349,6 +352,7 @@ affs_get_block(struct inode *inode, sector_t block, struct buffer_head *bh_resul
 
 	affs_brelse(ext_bh);
 	//unlock cache
+	mutex_unlock(&AFFS_I(inode)->truncate_mutex);
 	affs_unlock_ext(inode);
 	return 0;
 
@@ -365,6 +369,7 @@ err_alloc:
 	clear_buffer_mapped(bh_result);
 	bh_result->b_bdev = NULL;
 	// unlock cache
+	mutex_unlock(&AFFS_I(inode)->truncate_mutex);
 	affs_unlock_ext(inode);
 	return -ENOSPC;
 }
@@ -860,6 +865,7 @@ affs_truncate(struct inode *inode)
 
 	// lock cache
 	ext_bh = affs_get_extblock(inode, ext);
+	mutex_lock(&AFFS_I(inode)->truncate_mutex);
 	if (IS_ERR(ext_bh)) {
 		affs_warning(sb, "truncate",
 			     "unexpected read error for ext block %u (%ld)",
@@ -912,6 +918,7 @@ affs_truncate(struct inode *inode)
 				affs_warning(sb, "truncate",
 					     "unexpected read error for last block %u (%ld)",
 					     (unsigned int)ext, PTR_ERR(bh));
+				mutex_unlock(&AFFS_I(inode)->truncate_mutex);
 				return;
 			}
 			tmp = be32_to_cpu(AFFS_DATA_HEAD(bh)->next);
@@ -938,6 +945,7 @@ affs_truncate(struct inode *inode)
 		affs_brelse(ext_bh);
 	}
 	affs_free_prealloc(inode);
+	mutex_unlock(&AFFS_I(inode)->truncate_mutex);
 }
 
 int affs_file_fsync(struct file *filp, loff_t start, loff_t end, int datasync)
diff --git a/fs/affs/super.c b/fs/affs/super.c
index f754ab6..66c74df 100644
--- a/fs/affs/super.c
+++ b/fs/affs/super.c
@@ -123,6 +123,7 @@ static void init_once(void *foo)
 
 	sema_init(&ei->i_link_lock, 1);
 	sema_init(&ei->i_ext_lock, 1);
+	mutex_init(&ei->truncate_mutex);
 	inode_init_once(&ei->vfs_inode);
 }
 
-- 
2.1.0