fs: kill i_alloc_sem

i_alloc_sem is a rather special rw_semaphore.  It's the last one that may
be released by a non-owner, and it's write side is always mirrored by
real exclusion.  It's intended use it to wait for all pending direct I/O
requests to finish before starting a truncate.

Replace it with a hand-grown construct:

 - exclusion for truncates is already guaranteed by i_mutex, so it can
   simply fall way
 - the reader side is replaced by an i_dio_count member in struct inode
   that counts the number of pending direct I/O requests.  Truncate can't
   proceed as long as it's non-zero
 - when i_dio_count reaches non-zero we wake up a pending truncate using
   wake_up_bit on a new bit in i_flags
 - new references to i_dio_count can't appear while we are waiting for
   it to read zero because the direct I/O count always needs i_mutex
   (or an equivalent like XFS's i_iolock) for starting a new operation.

This scheme is much simpler, and saves the space of a spinlock_t and a
struct list_head in struct inode (typically 160 bits on a non-debug 64-bit
system).

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

fs/attr.c

@@ -233,16 +233,13 @@ int notify_change(struct dentry * dentry, struct iattr * attr)
 		return error;
 
 	if (ia_valid & ATTR_SIZE)
-		down_write(&dentry->d_inode->i_alloc_sem);
+		inode_dio_wait(inode);
 
 	if (inode->i_op->setattr)
 		error = inode->i_op->setattr(dentry, attr);
 	else
 		error = simple_setattr(dentry, attr);
 
-	if (ia_valid & ATTR_SIZE)
-		up_write(&dentry->d_inode->i_alloc_sem);
-
 	if (!error)
 		fsnotify_change(dentry, ia_valid);
 

fs/direct-io.c

@@ -135,6 +135,50 @@ struct dio {
 	struct page *pages[DIO_PAGES];	/* page buffer */
 };
 
+static void __inode_dio_wait(struct inode *inode)
+{
+	wait_queue_head_t *wq = bit_waitqueue(&inode->i_state, __I_DIO_WAKEUP);
+	DEFINE_WAIT_BIT(q, &inode->i_state, __I_DIO_WAKEUP);
+
+	do {
+		prepare_to_wait(wq, &q.wait, TASK_UNINTERRUPTIBLE);
+		if (atomic_read(&inode->i_dio_count))
+			schedule();
+	} while (atomic_read(&inode->i_dio_count));
+	finish_wait(wq, &q.wait);
+}
+
+/**
+ * inode_dio_wait - wait for outstanding DIO requests to finish
+ * @inode: inode to wait for
+ *
+ * Waits for all pending direct I/O requests to finish so that we can
+ * proceed with a truncate or equivalent operation.
+ *
+ * Must be called under a lock that serializes taking new references
+ * to i_dio_count, usually by inode->i_mutex.
+ */
+void inode_dio_wait(struct inode *inode)
+{
+	if (atomic_read(&inode->i_dio_count))
+		__inode_dio_wait(inode);
+}
+EXPORT_SYMBOL_GPL(inode_dio_wait);
+
+/*
+ * inode_dio_done - signal finish of a direct I/O requests
+ * @inode: inode the direct I/O happens on
+ *
+ * This is called once we've finished processing a direct I/O request,
+ * and is used to wake up callers waiting for direct I/O to be quiesced.
+ */
+void inode_dio_done(struct inode *inode)
+{
+	if (atomic_dec_and_test(&inode->i_dio_count))
+		wake_up_bit(&inode->i_state, __I_DIO_WAKEUP);
+}
+EXPORT_SYMBOL_GPL(inode_dio_done);
+
 /*
  * How many pages are in the queue?
  */
@@ -254,9 +298,7 @@ static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is
 	}
 
 	if (dio->flags & DIO_LOCKING)
-		/* lockdep: non-owner release */
+		inode_dio_done(dio->inode);
-		up_read_non_owner(&dio->inode->i_alloc_sem);
-
 	return ret;
 }
 
@@ -980,9 +1022,6 @@ static int do_direct_IO(struct dio *dio)
 	return ret;
 }
 
-/*
- * Releases both i_mutex and i_alloc_sem
- */
 static ssize_t
 direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode, 
 	const struct iovec *iov, loff_t offset, unsigned long nr_segs, 
@@ -1146,15 +1185,14 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
  *    For writes this function is called under i_mutex and returns with
  *    i_mutex held, for reads, i_mutex is not held on entry, but it is
  *    taken and dropped again before returning.
- *    For reads and writes i_alloc_sem is taken in shared mode and released
+ *    The i_dio_count counter keeps track of the number of outstanding
- *    on I/O completion (which may happen asynchronously after returning to
+ *    direct I/O requests, and truncate waits for it to reach zero.
- *    the caller).
+ *    New references to i_dio_count must only be grabbed with i_mutex
+ *    held.
  *
  *  - if the flags value does NOT contain DIO_LOCKING we don't use any
  *    internal locking but rather rely on the filesystem to synchronize
  *    direct I/O reads/writes versus each other and truncate.
- *    For reads and writes both i_mutex and i_alloc_sem are not held on
- *    entry and are never taken.
  */
 ssize_t
 __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
@@ -1234,10 +1272,9 @@ __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
 		}
 
 		/*
-		 * Will be released at I/O completion, possibly in a
+		 * Will be decremented at I/O completion time.
-		 * different thread.
 		 */
-		down_read_non_owner(&inode->i_alloc_sem);
+		atomic_inc(&inode->i_dio_count);
 	}
 
 	/*

fs/inode.c

@@ -168,8 +168,7 @@ int inode_init_always(struct super_block *sb, struct inode *inode)
 	mutex_init(&inode->i_mutex);
 	lockdep_set_class(&inode->i_mutex, &sb->s_type->i_mutex_key);
 
-	init_rwsem(&inode->i_alloc_sem);
+	atomic_set(&inode->i_dio_count, 0);
-	lockdep_set_class(&inode->i_alloc_sem, &sb->s_type->i_alloc_sem_key);
 
 	mapping->a_ops = &empty_aops;
 	mapping->host = inode;

fs/ntfs/file.c

@@ -1832,9 +1832,8 @@ static ssize_t ntfs_file_buffered_write(struct kiocb *iocb,
 	 * fails again.
 	 */
 	if (unlikely(NInoTruncateFailed(ni))) {
-		down_write(&vi->i_alloc_sem);
+		inode_dio_wait(vi);
 		err = ntfs_truncate(vi);
-		up_write(&vi->i_alloc_sem);
 		if (err || NInoTruncateFailed(ni)) {
 			if (!err)
 				err = -EIO;

fs/ntfs/inode.c

@@ -2357,12 +2357,7 @@ static const char *es = "  Leaving inconsistent metadata.  Unmount and run "
  *
  * Returns 0 on success or -errno on error.
  *
- * Called with ->i_mutex held.  In all but one case ->i_alloc_sem is held for
+ * Called with ->i_mutex held.
- * writing.  The only case in the kernel where ->i_alloc_sem is not held is
- * mm/filemap.c::generic_file_buffered_write() where vmtruncate() is called
- * with the current i_size as the offset.  The analogous place in NTFS is in
- * fs/ntfs/file.c::ntfs_file_buffered_write() where we call vmtruncate() again
- * without holding ->i_alloc_sem.
  */
 int ntfs_truncate(struct inode *vi)
 {
@@ -2887,8 +2882,7 @@ void ntfs_truncate_vfs(struct inode *vi) {
  * We also abort all changes of user, group, and mode as we do not implement
  * the NTFS ACLs yet.
  *
- * Called with ->i_mutex held.  For the ATTR_SIZE (i.e. ->truncate) case, also
+ * Called with ->i_mutex held.
- * called with ->i_alloc_sem held for writing.
  */
 int ntfs_setattr(struct dentry *dentry, struct iattr *attr)
 {

fs/ocfs2/aops.c

@@ -551,9 +551,8 @@ static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
 
 /*
  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
- * particularly interested in the aio/dio case.  Like the core uses
+ * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
- * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
+ * to protect io on one node from truncation on another.
- * truncation on another.
  */
 static void ocfs2_dio_end_io(struct kiocb *iocb,
 			     loff_t offset,
@@ -569,7 +568,7 @@ static void ocfs2_dio_end_io(struct kiocb *iocb,
 	BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
 
 	if (ocfs2_iocb_is_sem_locked(iocb)) {
-		up_read(&inode->i_alloc_sem);
+		inode_dio_done(inode);
 		ocfs2_iocb_clear_sem_locked(iocb);
 	}
 

fs/ocfs2/file.c

@@ -2236,9 +2236,9 @@ static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
 	ocfs2_iocb_clear_sem_locked(iocb);
 
 relock:
-	/* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
+	/* to match setattr's i_mutex -> rw_lock ordering */
 	if (direct_io) {
-		down_read(&inode->i_alloc_sem);
+		atomic_inc(&inode->i_dio_count);
 		have_alloc_sem = 1;
 		/* communicate with ocfs2_dio_end_io */
 		ocfs2_iocb_set_sem_locked(iocb);
@@ -2290,7 +2290,7 @@ static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
 	 */
 	if (direct_io && !can_do_direct) {
 		ocfs2_rw_unlock(inode, rw_level);
-		up_read(&inode->i_alloc_sem);
+		inode_dio_done(inode);
 
 		have_alloc_sem = 0;
 		rw_level = -1;
@@ -2361,8 +2361,7 @@ static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
 	/*
 	 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
 	 * function pointer which is called when o_direct io completes so that
-	 * it can unlock our rw lock.  (it's the clustered equivalent of
+	 * it can unlock our rw lock.
-	 * i_alloc_sem; protects truncate from racing with pending ios).
 	 * Unfortunately there are error cases which call end_io and others
 	 * that don't.  so we don't have to unlock the rw_lock if either an
 	 * async dio is going to do it in the future or an end_io after an
@@ -2379,7 +2378,7 @@ static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
 
 out_sems:
 	if (have_alloc_sem) {
-		up_read(&inode->i_alloc_sem);
+		inode_dio_done(inode);
 		ocfs2_iocb_clear_sem_locked(iocb);
 	}
 
@@ -2531,8 +2530,8 @@ static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
 	 * need locks to protect pending reads from racing with truncate.
 	 */
 	if (filp->f_flags & O_DIRECT) {
-		down_read(&inode->i_alloc_sem);
 		have_alloc_sem = 1;
+		atomic_inc(&inode->i_dio_count);
 		ocfs2_iocb_set_sem_locked(iocb);
 
 		ret = ocfs2_rw_lock(inode, 0);
@@ -2575,7 +2574,7 @@ static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
 
 bail:
 	if (have_alloc_sem) {
-		up_read(&inode->i_alloc_sem);
+		inode_dio_done(inode);
 		ocfs2_iocb_clear_sem_locked(iocb);
 	}
 	if (rw_level != -1)

fs/reiserfs/xattr.c

@@ -555,11 +555,10 @@ reiserfs_xattr_set_handle(struct reiserfs_transaction_handle *th,
 
 		reiserfs_write_unlock(inode->i_sb);
 		mutex_lock_nested(&dentry->d_inode->i_mutex, I_MUTEX_XATTR);
-		down_write(&dentry->d_inode->i_alloc_sem);
+		inode_dio_wait(dentry->d_inode);
 		reiserfs_write_lock(inode->i_sb);
 
 		err = reiserfs_setattr(dentry, &newattrs);
-		up_write(&dentry->d_inode->i_alloc_sem);
 		mutex_unlock(&dentry->d_inode->i_mutex);
 	} else
 		update_ctime(inode);

include/linux/fs.h

@@ -779,7 +779,7 @@ struct inode {
 	struct timespec		i_ctime;
 	blkcnt_t		i_blocks;
 	unsigned short          i_bytes;
-	struct rw_semaphore	i_alloc_sem;
+	atomic_t		i_dio_count;
 	const struct file_operations	*i_fop;	/* former ->i_op->default_file_ops */
 	struct file_lock	*i_flock;
 	struct address_space	*i_mapping;
@@ -1705,6 +1705,10 @@ struct super_operations {
  *			set during data writeback, and cleared with a wakeup
  *			on the bit address once it is done.
  *
+ * I_REFERENCED		Marks the inode as recently references on the LRU list.
+ *
+ * I_DIO_WAKEUP		Never set.  Only used as a key for wait_on_bit().
+ *
  * Q: What is the difference between I_WILL_FREE and I_FREEING?
  */
 #define I_DIRTY_SYNC		(1 << 0)
@@ -1718,6 +1722,8 @@ struct super_operations {
 #define __I_SYNC		7
 #define I_SYNC			(1 << __I_SYNC)
 #define I_REFERENCED		(1 << 8)
+#define __I_DIO_WAKEUP		9
+#define I_DIO_WAKEUP		(1 << I_DIO_WAKEUP)
 
 #define I_DIRTY (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_PAGES)
 
@@ -1828,7 +1834,6 @@ struct file_system_type {
 	struct lock_class_key i_lock_key;
 	struct lock_class_key i_mutex_key;
 	struct lock_class_key i_mutex_dir_key;
-	struct lock_class_key i_alloc_sem_key;
 };
 
 extern struct dentry *mount_ns(struct file_system_type *fs_type, int flags,
@@ -2404,6 +2409,8 @@ enum {
 };
 
 void dio_end_io(struct bio *bio, int error);
+void inode_dio_wait(struct inode *inode);
+void inode_dio_done(struct inode *inode);
 
 ssize_t __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
 	struct block_device *bdev, const struct iovec *iov, loff_t offset,

mm/filemap.c

@@ -78,9 +78,6 @@
  *  ->i_mutex			(generic_file_buffered_write)
  *    ->mmap_sem		(fault_in_pages_readable->do_page_fault)
  *
- *  ->i_mutex
- *    ->i_alloc_sem             (various)
- *
  *  inode_wb_list_lock
  *    sb_lock			(fs/fs-writeback.c)
  *    ->mapping->tree_lock	(__sync_single_inode)

mm/madvise.c

@@ -218,7 +218,7 @@ static long madvise_remove(struct vm_area_struct *vma,
 	endoff = (loff_t)(end - vma->vm_start - 1)
 			+ ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 
-	/* vmtruncate_range needs to take i_mutex and i_alloc_sem */
+	/* vmtruncate_range needs to take i_mutex */
 	up_read(&current->mm->mmap_sem);
 	error = vmtruncate_range(mapping->host, offset, endoff);
 	down_read(&current->mm->mmap_sem);

mm/rmap.c

@@ -21,7 +21,6 @@
  * Lock ordering in mm:
  *
  * inode->i_mutex	(while writing or truncating, not reading or faulting)
- *   inode->i_alloc_sem (vmtruncate_range)
  *   mm->mmap_sem
  *     page->flags PG_locked (lock_page)
  *       mapping->i_mmap_mutex

mm/truncate.c

@@ -622,12 +622,11 @@ int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
 		return -ENOSYS;
 
 	mutex_lock(&inode->i_mutex);
-	down_write(&inode->i_alloc_sem);
+	inode_dio_wait(inode);
 	unmap_mapping_range(mapping, offset, (end - offset), 1);
 	inode->i_op->truncate_range(inode, offset, end);
 	/* unmap again to remove racily COWed private pages */
 	unmap_mapping_range(mapping, offset, (end - offset), 1);
-	up_write(&inode->i_alloc_sem);
 	mutex_unlock(&inode->i_mutex);
 
 	return 0;