[PATCH] Fix CacheFS barrier handling and other kernel discrepancies

[David Howells]

The attached patch fixes some discrepancies between the latest version of the
kernel and CacheFS; namely: BIO_RW_BARRIER may now cause an error if the
underlying blockdev doesn't support it and i_op->write_inode() now returns an
error.

Signed-Off-By: David Howells <dhowells@xxxxxxxxxx>
---

diffstat cachefs-269rc2mm1.diff
 cachefs-int.h |    8 +++--
 inode.c       |    3 +-
 journal.c     |   83 +++++++++++++++++++++++++++++++++++++++++++++++++++++++---
 super.c       |    9 ++++--
 4 files changed, 94 insertions(+), 9 deletions(-)

diff -uNrp linux-2.6.9-rc2-mm1/fs/cachefs/cachefs-int.h linux-2.6.9-rc2-mm1-afskey/fs/cachefs/cachefs-int.h
--- linux-2.6.9-rc2-mm1/fs/cachefs/cachefs-int.h	2004-09-16 12:09:51.000000000 +0100
+++ linux-2.6.9-rc2-mm1-afskey/fs/cachefs/cachefs-int.h	2004-09-17 11:15:29.000000000 +0100
@@ -73,6 +73,8 @@ struct cachefs_super
 #define CACHEFS_SUPER_WITHDRAWN		4		/* T if cache has been withdrawn */
 #define CACHEFS_SUPER_REPLAYING_UJNL	5		/* T if replaying u-journal */
 
+	int				bio_wr_barrier;	/* command to submit a write barrier BIO */
+
 	/* index management */
 	struct list_head		ino_list;	/* list of data/index inodes */
 	spinlock_t			ino_list_lock;
@@ -154,8 +156,6 @@ struct cachefs_super
 	wait_queue_head_t		batch_timer_wq;	/* batch timer wait queue */
 	wait_queue_head_t		batch_sync_wq;	/* batch sync wait queue */
 
-	struct list_head		jnld_link;	/* journalling daemon list */
-
 	/* validity journal tracking */
 	unsigned long			*vjnl_map;	/* bitmap of free entries (1 page) */
 	unsigned			vjnl_count;	/* number of free entries */
@@ -200,6 +200,8 @@ extern void cachefs_recycle_transfer_sta
 extern void cachefs_recycle_reclaim(struct cachefs_super *super);
 extern void cachefs_recycle_unallocate_data_block(struct cachefs_super *super);
 
+extern int cachefs_ujnl_check_barrier_cap(struct cachefs_super *super);
+
 /*****************************************************************************/
 /*
  * block management record
@@ -384,7 +386,7 @@ extern struct file_operations cachefs_st
 
 extern struct cachefs_inode *cachefs_iget(struct cachefs_super *super,
 					  ino_t ino);
-extern void cachefs_write_inode(struct inode *_inode, int sync);
+extern int cachefs_write_inode(struct inode *_inode, int sync);
 extern void cachefs_clear_inode(struct inode *vfs_inode);
 
 static inline struct cachefs_inode *cachefs_igrab(struct cachefs_inode *iinode)
diff -uNrp linux-2.6.9-rc2-mm1/fs/cachefs/inode.c linux-2.6.9-rc2-mm1-afskey/fs/cachefs/inode.c
--- linux-2.6.9-rc2-mm1/fs/cachefs/inode.c	2004-09-16 12:09:51.000000000 +0100
+++ linux-2.6.9-rc2-mm1-afskey/fs/cachefs/inode.c	2004-09-17 11:16:50.000000000 +0100
@@ -366,9 +366,10 @@ struct cachefs_inode *cachefs_iget(struc
  * - don't use generic_file_write() to write out the meta-data file's meta-data
  *   as it updates the mtime & ctime and marks the inode dirty again
  */
-void cachefs_write_inode(struct inode *vfs_inode, int sync)
+int cachefs_write_inode(struct inode *vfs_inode, int sync)
 {
 	_enter("{sb=%p ino=%lu},%d", vfs_inode->i_sb, vfs_inode->i_ino, sync);
+	return 0;
 
 } /* end cachefs_write_inode() */
 
diff -uNrp linux-2.6.9-rc2-mm1/fs/cachefs/journal.c linux-2.6.9-rc2-mm1-afskey/fs/cachefs/journal.c
--- linux-2.6.9-rc2-mm1/fs/cachefs/journal.c	2004-09-16 12:09:51.000000000 +0100
+++ linux-2.6.9-rc2-mm1-afskey/fs/cachefs/journal.c	2004-09-17 11:14:24.000000000 +0100
@@ -1040,7 +1040,7 @@ static void cachefs_trans_batch_write_uj
 		/* submit for write (with barrier on last chunk) */
 		rw = WRITE;
 		if (super->ujnl_tail == jstop)
-			rw |= 1 << BIO_RW_BARRIER;
+			rw = super->bio_wr_barrier;
 		submit_bio(rw, bio);
 
 	} while (super->ujnl_tail != jstop);
@@ -1237,7 +1237,7 @@ static void cachefs_trans_batch_write_ma
 
 	/* and send to disc */
 	//dump_bio(bio,1);
-	submit_bio(WRITE | (1 << BIO_RW_BARRIER), bio);
+	submit_bio(super->bio_wr_barrier, bio);
 
 	cachefs_block_put(jblock);
 	_leave("");
@@ -1317,7 +1317,7 @@ static void cachefs_trans_batch_write_ac
 
 	/* and send to disc */
 	//dump_bio(bio,1);
-	submit_bio(WRITE | (1 << BIO_RW_BARRIER), bio);
+	submit_bio(super->bio_wr_barrier, bio);
 
 	/* wait for I/O completion */
 	set_current_state(TASK_UNINTERRUPTIBLE);
@@ -1669,3 +1669,80 @@ void cachefs_trans_sync(struct cachefs_s
 	_leave("");
 
 } /* end cachefs_trans_sync() */
+
+/*****************************************************************************/
+/*
+ * barrier capabability check complete
+ */
+static int cachefs_ujnl_barrier_cap_checked(struct bio *bio,
+					    unsigned int bytes_done,
+					    int error)
+{
+	kenter("%p{%lx},%u,%d", bio, bio->bi_flags, bytes_done, error);
+
+	/* we're only interested in completion */
+	if (bio->bi_size > 0) {
+		kleave(" = 1");
+		return 1;
+	}
+
+	*(int*) bio->bi_private = error;
+	end_page_writeback(bio->bi_io_vec[0].bv_page);
+
+	bio_put(bio);
+	kleave(" = 0");
+	return 0;
+
+} /* end cachefs_trans_ack_written() */
+
+/*****************************************************************************/
+/*
+ * determine whether the block device supports barriers
+ * - need barriers to be able to journal properly
+ */
+int cachefs_ujnl_check_barrier_cap(struct cachefs_super *super)
+{
+	struct page *superpage;
+	struct bio *bio;
+	int ret;
+
+	super->bio_wr_barrier = WRITE | (1 << BIO_RW_BARRIER);
+
+	/* attempt a barriered write on the superblock */
+	superpage = virt_to_page(super->layout);
+	wait_on_page_writeback(superpage);
+	SetPageWriteback(superpage);
+
+	ret = -ENOMEM;
+	bio = bio_alloc(GFP_KERNEL, 1);
+	if (!bio)
+		goto error;
+
+	bio->bi_bdev	= super->sb->s_bdev;
+	bio->bi_private	= &ret;
+	bio->bi_end_io	= cachefs_ujnl_barrier_cap_checked;
+	bio->bi_sector	= 0;
+
+	if (!bio_add_page(bio, superpage, PAGE_SIZE, 0))
+		BUG();
+
+	/* and send to disc */
+	//dump_bio(bio,1);
+	ret = 0;
+	submit_bio(super->bio_wr_barrier, bio);
+	wait_on_page_writeback(superpage);
+
+	if (ret < 0) {
+		if (ret == -EOPNOTSUPP) {
+			/* it appears barriers are not supported */
+			printk("CacheFS:"
+			       " The blockdev does not support barriers,"
+			       " so the journal may not be reliable\n");
+			super->bio_wr_barrier = WRITE;
+			ret = 0;
+		}
+	}
+
+ error:
+	return ret;
+} /* end cachefs_ujnl_check_barrier_cap() */
diff -uNrp linux-2.6.9-rc2-mm1/fs/cachefs/super.c linux-2.6.9-rc2-mm1-afskey/fs/cachefs/super.c
--- linux-2.6.9-rc2-mm1/fs/cachefs/super.c	2004-09-16 12:09:51.000000000 +0100
+++ linux-2.6.9-rc2-mm1-afskey/fs/cachefs/super.c	2004-09-17 11:19:04.000000000 +0100
@@ -297,7 +297,6 @@ static int cachefs_fill_super(struct sup
 	INIT_LIST_HEAD(&super->ujnl_replayq);
 
 	spin_lock_init(&super->njalt_lock);
-	INIT_LIST_HEAD(&super->jnld_link);
 
 	init_MUTEX(&super->batch_sem);
 	init_MUTEX(&super->batch_uj_sem);
@@ -396,12 +395,19 @@ static int cachefs_fill_super(struct sup
 	ret = -EROFS;
 	if (bdev_read_only(sb->s_bdev)) {
 		printk("CacheFS: blockdev read-only\n");
+		goto error;
 	}
 
 	if (sb->s_flags & MS_RDONLY) {
 		printk("CacheFS: filesystem mounted read-only\n");
+		goto error;
 	}
 
+	/* check for barrier capability on the blockdev */
+	ret = cachefs_ujnl_check_barrier_cap(super);
+	if (ret < 0)
+		goto error;
+
 	/* replay the journal if the cache was initialised */
 	super->ujnl_jsof = super->layout->bix_ujournal;
 	super->ujnl_jsof <<= (PAGE_SHIFT - super->sb->s_blocksize_bits);
@@ -794,7 +800,6 @@ static void cachefs_put_super(struct sup
 
 	DECLARE_WAITQUEUE(myself, current);
 
-	printk("\n\n");
 	_enter("{%p}", super);
 
 	BUG_ON(!super);
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/