NFS: alloc nfs_read/write_data as direct I/O is scheduled

Re-arrange the logic in the NFS direct I/O path so that nfs_read/write_data
structs are allocated just before they are scheduled, rather than
allocating them all at once before we start scheduling requests.

Signed-off-by: Chuck Lever <cel@netapp.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c
index b1630d5..e25b759 100644
--- a/fs/nfs/direct.c
+++ b/fs/nfs/direct.c
@@ -68,8 +68,6 @@
 	struct kref		kref;		/* release manager */
 
 	/* I/O parameters */
-	struct list_head	list,		/* nfs_read/write_data structs */
-				rewrite_list;	/* saved nfs_write_data structs */
 	struct nfs_open_context	*ctx;		/* file open context info */
 	struct kiocb *		iocb;		/* controlling i/o request */
 	struct inode *		inode;		/* target file of i/o */
@@ -82,6 +80,7 @@
 	struct completion	completion;	/* wait for i/o completion */
 
 	/* commit state */
+	struct list_head	rewrite_list;	/* saved nfs_write_data structs */
 	struct nfs_write_data *	commit_data;	/* special write_data for commits */
 	int			flags;
 #define NFS_ODIRECT_DO_COMMIT		(1)	/* an unstable reply was received */
@@ -116,6 +115,11 @@
 	return page_count;
 }
 
+static inline unsigned int nfs_max_pages(unsigned int size)
+{
+	return (size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
+}
+
 /**
  * nfs_direct_IO - NFS address space operation for direct I/O
  * @rw: direction (read or write)
@@ -164,8 +168,8 @@
 		return NULL;
 
 	kref_init(&dreq->kref);
+	kref_get(&dreq->kref);
 	init_completion(&dreq->completion);
-	INIT_LIST_HEAD(&dreq->list);
 	INIT_LIST_HEAD(&dreq->rewrite_list);
 	dreq->iocb = NULL;
 	dreq->ctx = NULL;
@@ -228,49 +232,6 @@
 }
 
 /*
- * Note we also set the number of requests we have in the dreq when we are
- * done.  This prevents races with I/O completion so we will always wait
- * until all requests have been dispatched and completed.
- */
-static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, size_t rsize)
-{
-	struct list_head *list;
-	struct nfs_direct_req *dreq;
-	unsigned int rpages = (rsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
-
-	dreq = nfs_direct_req_alloc();
-	if (!dreq)
-		return NULL;
-
-	list = &dreq->list;
-	for(;;) {
-		struct nfs_read_data *data = nfs_readdata_alloc(rpages);
-
-		if (unlikely(!data)) {
-			while (!list_empty(list)) {
-				data = list_entry(list->next,
-						  struct nfs_read_data, pages);
-				list_del(&data->pages);
-				nfs_readdata_free(data);
-			}
-			kref_put(&dreq->kref, nfs_direct_req_release);
-			return NULL;
-		}
-
-		INIT_LIST_HEAD(&data->pages);
-		list_add(&data->pages, list);
-
-		data->req = (struct nfs_page *) dreq;
-		get_dreq(dreq);
-		if (nbytes <= rsize)
-			break;
-		nbytes -= rsize;
-	}
-	kref_get(&dreq->kref);
-	return dreq;
-}
-
-/*
  * We must hold a reference to all the pages in this direct read request
  * until the RPCs complete.  This could be long *after* we are woken up in
  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
@@ -305,42 +266,53 @@
 };
 
 /*
- * For each nfs_read_data struct that was allocated on the list, dispatch
- * an NFS READ operation.  If get_user_pages() fails, we stop sending reads.
- * Read length accounting is handled by nfs_direct_read_result().
- * Otherwise, if no requests have been sent, just return an error.
+ * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
+ * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
+ * bail and stop sending more reads.  Read length accounting is
+ * handled automatically by nfs_direct_read_result().  Otherwise, if
+ * no requests have been sent, just return an error.
  */
 static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos)
 {
 	struct nfs_open_context *ctx = dreq->ctx;
 	struct inode *inode = ctx->dentry->d_inode;
-	struct list_head *list = &dreq->list;
 	size_t rsize = NFS_SERVER(inode)->rsize;
+	unsigned int rpages = nfs_max_pages(rsize);
 	unsigned int pgbase;
 	int result;
 	ssize_t started = 0;
-	struct nfs_read_data *data;
+
+	get_dreq(dreq);
 
 	pgbase = user_addr & ~PAGE_MASK;
 	do {
+		struct nfs_read_data *data;
 		size_t bytes;
 
+		result = -ENOMEM;
+		data = nfs_readdata_alloc(rpages);
+		if (unlikely(!data))
+			break;
+
 		bytes = rsize;
 		if (count < rsize)
 			bytes = count;
 
-		BUG_ON(list_empty(list));
-		data = list_entry(list->next, struct nfs_read_data, pages);
-		list_del_init(&data->pages);
-
 		data->npages = nfs_direct_count_pages(user_addr, bytes);
 		down_read(&current->mm->mmap_sem);
 		result = get_user_pages(current, current->mm, user_addr,
 					data->npages, 1, 0, data->pagevec, NULL);
 		up_read(&current->mm->mmap_sem);
-		if (unlikely(result < data->npages))
-			goto out_err;
+		if (unlikely(result < data->npages)) {
+			if (result > 0)
+				nfs_direct_release_pages(data->pagevec, result);
+			nfs_readdata_release(data);
+			break;
+		}
 
+		get_dreq(dreq);
+
+		data->req = (struct nfs_page *) dreq;
 		data->inode = inode;
 		data->cred = ctx->cred;
 		data->args.fh = NFS_FH(inode);
@@ -378,21 +350,9 @@
 
 		count -= bytes;
 	} while (count != 0);
-	BUG_ON(!list_empty(list));
-	return 0;
 
-out_err:
-	if (result > 0)
-		nfs_direct_release_pages(data->pagevec, result);
-
-	list_add(&data->pages, list);
-	while (!list_empty(list)) {
-		data = list_entry(list->next, struct nfs_read_data, pages);
-		list_del(&data->pages);
-		nfs_readdata_free(data);
-		if (put_dreq(dreq))
-			nfs_direct_complete(dreq);
-	}
+	if (put_dreq(dreq))
+		nfs_direct_complete(dreq);
 
 	if (started)
 		return 0;
@@ -401,13 +361,13 @@
 
 static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
 {
-	ssize_t result;
+	ssize_t result = 0;
 	sigset_t oldset;
 	struct inode *inode = iocb->ki_filp->f_mapping->host;
 	struct rpc_clnt *clnt = NFS_CLIENT(inode);
 	struct nfs_direct_req *dreq;
 
-	dreq = nfs_direct_read_alloc(count, NFS_SERVER(inode)->rsize);
+	dreq = nfs_direct_req_alloc();
 	if (!dreq)
 		return -ENOMEM;
 
@@ -428,9 +388,8 @@
 
 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
 {
-	list_splice_init(&dreq->rewrite_list, &dreq->list);
-	while (!list_empty(&dreq->list)) {
-		struct nfs_write_data *data = list_entry(dreq->list.next, struct nfs_write_data, pages);
+	while (!list_empty(&dreq->rewrite_list)) {
+		struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
 		list_del(&data->pages);
 		nfs_direct_release_pages(data->pagevec, data->npages);
 		nfs_writedata_release(data);
@@ -584,47 +543,6 @@
 }
 #endif
 
-static struct nfs_direct_req *nfs_direct_write_alloc(size_t nbytes, size_t wsize)
-{
-	struct list_head *list;
-	struct nfs_direct_req *dreq;
-	unsigned int wpages = (wsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
-
-	dreq = nfs_direct_req_alloc();
-	if (!dreq)
-		return NULL;
-
-	list = &dreq->list;
-	for(;;) {
-		struct nfs_write_data *data = nfs_writedata_alloc(wpages);
-
-		if (unlikely(!data)) {
-			while (!list_empty(list)) {
-				data = list_entry(list->next,
-						  struct nfs_write_data, pages);
-				list_del(&data->pages);
-				nfs_writedata_free(data);
-			}
-			kref_put(&dreq->kref, nfs_direct_req_release);
-			return NULL;
-		}
-
-		INIT_LIST_HEAD(&data->pages);
-		list_add(&data->pages, list);
-
-		data->req = (struct nfs_page *) dreq;
-		get_dreq(dreq);
-		if (nbytes <= wsize)
-			break;
-		nbytes -= wsize;
-	}
-
-	nfs_alloc_commit_data(dreq);
-
-	kref_get(&dreq->kref);
-	return dreq;
-}
-
 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
 {
 	struct nfs_write_data *data = calldata;
@@ -677,43 +595,55 @@
 };
 
 /*
- * For each nfs_write_data struct that was allocated on the list, dispatch
- * an NFS WRITE operation.  If get_user_pages() fails, we stop sending writes.
- * Write length accounting is handled by nfs_direct_write_result().
- * Otherwise, if no requests have been sent, just return an error.
+ * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
+ * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
+ * bail and stop sending more writes.  Write length accounting is
+ * handled automatically by nfs_direct_write_result().  Otherwise, if
+ * no requests have been sent, just return an error.
  */
 static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync)
 {
 	struct nfs_open_context *ctx = dreq->ctx;
 	struct inode *inode = ctx->dentry->d_inode;
-	struct list_head *list = &dreq->list;
 	size_t wsize = NFS_SERVER(inode)->wsize;
+	unsigned int wpages = nfs_max_pages(wsize);
 	unsigned int pgbase;
 	int result;
 	ssize_t started = 0;
-	struct nfs_write_data *data;
+
+	get_dreq(dreq);
 
 	pgbase = user_addr & ~PAGE_MASK;
 	do {
+		struct nfs_write_data *data;
 		size_t bytes;
 
+		result = -ENOMEM;
+		data = nfs_writedata_alloc(wpages);
+		if (unlikely(!data))
+			break;
+
 		bytes = wsize;
 		if (count < wsize)
 			bytes = count;
 
-		BUG_ON(list_empty(list));
-		data = list_entry(list->next, struct nfs_write_data, pages);
-
 		data->npages = nfs_direct_count_pages(user_addr, bytes);
 		down_read(&current->mm->mmap_sem);
 		result = get_user_pages(current, current->mm, user_addr,
 					data->npages, 0, 0, data->pagevec, NULL);
 		up_read(&current->mm->mmap_sem);
-		if (unlikely(result < data->npages))
-			goto out_err;
+		if (unlikely(result < data->npages)) {
+			if (result > 0)
+				nfs_direct_release_pages(data->pagevec, result);
+			nfs_writedata_release(data);
+			break;
+		}
+
+		get_dreq(dreq);
 
 		list_move_tail(&data->pages, &dreq->rewrite_list);
 
+		data->req = (struct nfs_page *) dreq;
 		data->inode = inode;
 		data->cred = ctx->cred;
 		data->args.fh = NFS_FH(inode);
@@ -752,21 +682,9 @@
 
 		count -= bytes;
 	} while (count != 0);
-	BUG_ON(!list_empty(list));
-	return 0;
 
-out_err:
-	if (result > 0)
-		nfs_direct_release_pages(data->pagevec, result);
-
-	list_add(&data->pages, list);
-	while (!list_empty(list)) {
-		data = list_entry(list->next, struct nfs_write_data, pages);
-		list_del(&data->pages);
-		nfs_writedata_free(data);
-		if (put_dreq(dreq))
-			nfs_direct_write_complete(dreq, inode);
-	}
+	if (put_dreq(dreq))
+		nfs_direct_write_complete(dreq, inode);
 
 	if (started)
 		return 0;
@@ -775,7 +693,7 @@
 
 static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
 {
-	ssize_t result;
+	ssize_t result = 0;
 	sigset_t oldset;
 	struct inode *inode = iocb->ki_filp->f_mapping->host;
 	struct rpc_clnt *clnt = NFS_CLIENT(inode);
@@ -783,9 +701,11 @@
 	size_t wsize = NFS_SERVER(inode)->wsize;
 	int sync = 0;
 
-	dreq = nfs_direct_write_alloc(count, wsize);
+	dreq = nfs_direct_req_alloc();
 	if (!dreq)
 		return -ENOMEM;
+	nfs_alloc_commit_data(dreq);
+
 	if (dreq->commit_data == NULL || count < wsize)
 		sync = FLUSH_STABLE;