NFS: support large reads and writes on the wire

 Most NFS server implementations allow up to 64KB reads and writes on the
 wire.  The Solaris NFS server allows up to a megabyte, for instance.

 Now the Linux NFS client supports transfer sizes up to 1MB, too.  This will
 help reduce protocol and context switch overhead on read/write intensive NFS
 workloads, and support larger atomic read and write operations on servers
 that support them.

 Test-plan:
 Connectathon and iozone on mount point with wsize=rsize>32768 over TCP.
 Tests with NFS over UDP to verify the maximum RPC payload size cap.

 Signed-off-by: Chuck Lever <cel@netapp.com>
 Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>

8 files changed, 91 insertions, 53 deletions

diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c
index f69d95aa78b..fd7ac5e841c 100644
--- a/fs/nfs/direct.c
+++ b/fs/nfs/direct.c
@@ -154,6 +154,7 @@ static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, unsigned int
 	struct list_head *list;
 	struct nfs_direct_req *dreq;
 	unsigned int reads = 0;
+	unsigned int rpages = (rsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 
 	dreq = kmem_cache_alloc(nfs_direct_cachep, SLAB_KERNEL);
 	if (!dreq)
@@ -167,7 +168,7 @@ static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, unsigned int
 
 	list = &dreq->list;
 	for(;;) {
-		struct nfs_read_data *data = nfs_readdata_alloc();
+		struct nfs_read_data *data = nfs_readdata_alloc(rpages);
 
 		if (unlikely(!data)) {
 			while (!list_empty(list)) {
@@ -431,7 +432,7 @@ static ssize_t nfs_direct_write_seg(struct inode *inode,
 	struct nfs_writeverf first_verf;
 	struct nfs_write_data *wdata;
 
-	wdata = nfs_writedata_alloc();
+	wdata = nfs_writedata_alloc(NFS_SERVER(inode)->wpages);
 	if (!wdata)
 		return -ENOMEM;
 
diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
index 4e6558df54b..acde2c5725b 100644
--- a/fs/nfs/inode.c
+++ b/fs/nfs/inode.c
@@ -221,10 +221,10 @@ nfs_calc_block_size(u64 tsize)
 static inline unsigned long
 nfs_block_size(unsigned long bsize, unsigned char *nrbitsp)
 {
-	if (bsize < 1024)
-		bsize = NFS_DEF_FILE_IO_BUFFER_SIZE;
-	else if (bsize >= NFS_MAX_FILE_IO_BUFFER_SIZE)
-		bsize = NFS_MAX_FILE_IO_BUFFER_SIZE;
+	if (bsize < NFS_MIN_FILE_IO_SIZE)
+		bsize = NFS_DEF_FILE_IO_SIZE;
+	else if (bsize >= NFS_MAX_FILE_IO_SIZE)
+		bsize = NFS_MAX_FILE_IO_SIZE;
 
 	return nfs_block_bits(bsize, nrbitsp);
 }
@@ -307,20 +307,15 @@ nfs_sb_init(struct super_block *sb, rpc_authflavor_t authflavor)
 	max_rpc_payload = nfs_block_size(rpc_max_payload(server->client), NULL);
 	if (server->rsize > max_rpc_payload)
 		server->rsize = max_rpc_payload;
-	if (server->wsize > max_rpc_payload)
-		server->wsize = max_rpc_payload;
-
+	if (server->rsize > NFS_MAX_FILE_IO_SIZE)
+		server->rsize = NFS_MAX_FILE_IO_SIZE;
 	server->rpages = (server->rsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
-	if (server->rpages > NFS_READ_MAXIOV) {
-		server->rpages = NFS_READ_MAXIOV;
-		server->rsize = server->rpages << PAGE_CACHE_SHIFT;
-	}
 
+	if (server->wsize > max_rpc_payload)
+		server->wsize = max_rpc_payload;
+	if (server->wsize > NFS_MAX_FILE_IO_SIZE)
+		server->wsize = NFS_MAX_FILE_IO_SIZE;
 	server->wpages = (server->wsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
-        if (server->wpages > NFS_WRITE_MAXIOV) {
-		server->wpages = NFS_WRITE_MAXIOV;
-                server->wsize = server->wpages << PAGE_CACHE_SHIFT;
-	}
 
 	if (sb->s_blocksize == 0)
 		sb->s_blocksize = nfs_block_bits(server->wsize,
diff --git a/fs/nfs/nfsroot.c b/fs/nfs/nfsroot.c
index 1b272a135a3..985cc53b8dd 100644
--- a/fs/nfs/nfsroot.c
+++ b/fs/nfs/nfsroot.c
@@ -296,8 +296,8 @@ static int __init root_nfs_name(char *name)
 	nfs_port          = -1;
 	nfs_data.version  = NFS_MOUNT_VERSION;
 	nfs_data.flags    = NFS_MOUNT_NONLM;	/* No lockd in nfs root yet */
-	nfs_data.rsize    = NFS_DEF_FILE_IO_BUFFER_SIZE;
-	nfs_data.wsize    = NFS_DEF_FILE_IO_BUFFER_SIZE;
+	nfs_data.rsize    = NFS_DEF_FILE_IO_SIZE;
+	nfs_data.wsize    = NFS_DEF_FILE_IO_SIZE;
 	nfs_data.acregmin = 3;
 	nfs_data.acregmax = 60;
 	nfs_data.acdirmin = 30;
diff --git a/fs/nfs/read.c b/fs/nfs/read.c
index 21486242c3d..05eb43fadf8 100644
--- a/fs/nfs/read.c
+++ b/fs/nfs/read.c
@@ -83,7 +83,7 @@ static int nfs_readpage_sync(struct nfs_open_context *ctx, struct inode *inode,
 	int		result;
 	struct nfs_read_data *rdata;
 
-	rdata = nfs_readdata_alloc();
+	rdata = nfs_readdata_alloc(1);
 	if (!rdata)
 		return -ENOMEM;
 
@@ -283,7 +283,7 @@ static int nfs_pagein_multi(struct list_head *head, struct inode *inode)
 
 	nbytes = req->wb_bytes;
 	for(;;) {
-		data = nfs_readdata_alloc();
+		data = nfs_readdata_alloc(1);
 		if (!data)
 			goto out_bad;
 		INIT_LIST_HEAD(&data->pages);
@@ -339,7 +339,7 @@ static int nfs_pagein_one(struct list_head *head, struct inode *inode)
 	if (NFS_SERVER(inode)->rsize < PAGE_CACHE_SIZE)
 		return nfs_pagein_multi(head, inode);
 
-	data = nfs_readdata_alloc();
+	data = nfs_readdata_alloc(NFS_SERVER(inode)->rpages);
 	if (!data)
 		goto out_bad;
 
diff --git a/fs/nfs/write.c b/fs/nfs/write.c
index 80bc4ea1b82..1ce0c200df1 100644
--- a/fs/nfs/write.c
+++ b/fs/nfs/write.c
@@ -89,18 +89,33 @@ static mempool_t *nfs_commit_mempool;
 
 static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion);
 
-static inline struct nfs_write_data *nfs_commit_alloc(void)
+static inline struct nfs_write_data *nfs_commit_alloc(unsigned int pagecount)
 {
 	struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, SLAB_NOFS);
+
 	if (p) {
 		memset(p, 0, sizeof(*p));
 		INIT_LIST_HEAD(&p->pages);
+		if (pagecount < NFS_PAGEVEC_SIZE)
+			p->pagevec = &p->page_array[0];
+		else {
+			size_t size = ++pagecount * sizeof(struct page *);
+			p->pagevec = kmalloc(size, GFP_NOFS);
+			if (p->pagevec) {
+				memset(p->pagevec, 0, size);
+			} else {
+				mempool_free(p, nfs_commit_mempool);
+				p = NULL;
+			}
+		}
 	}
 	return p;
 }
 
 static inline void nfs_commit_free(struct nfs_write_data *p)
 {
+	if (p && (p->pagevec != &p->page_array[0]))
+		kfree(p->pagevec);
 	mempool_free(p, nfs_commit_mempool);
 }
 
@@ -167,7 +182,7 @@ static int nfs_writepage_sync(struct nfs_open_context *ctx, struct inode *inode,
 	int		result, written = 0;
 	struct nfs_write_data *wdata;
 
-	wdata = nfs_writedata_alloc();
+	wdata = nfs_writedata_alloc(1);
 	if (!wdata)
 		return -ENOMEM;
 
@@ -909,7 +924,7 @@ static int nfs_flush_multi(struct list_head *head, struct inode *inode, int how)
 
 	nbytes = req->wb_bytes;
 	for (;;) {
-		data = nfs_writedata_alloc();
+		data = nfs_writedata_alloc(1);
 		if (!data)
 			goto out_bad;
 		list_add(&data->pages, &list);
@@ -973,7 +988,7 @@ static int nfs_flush_one(struct list_head *head, struct inode *inode, int how)
 	if (NFS_SERVER(inode)->wsize < PAGE_CACHE_SIZE)
 		return nfs_flush_multi(head, inode, how);
 
-	data = nfs_writedata_alloc();
+	data = nfs_writedata_alloc(NFS_SERVER(inode)->wpages);
 	if (!data)
 		goto out_bad;
 
@@ -1241,12 +1256,12 @@ static void nfs_commit_rpcsetup(struct list_head *head,
  * Commit dirty pages
  */
 static int
-nfs_commit_list(struct list_head *head, int how)
+nfs_commit_list(struct inode *inode, struct list_head *head, int how)
 {
 	struct nfs_write_data	*data;
 	struct nfs_page         *req;
 
-	data = nfs_commit_alloc();
+	data = nfs_commit_alloc(NFS_SERVER(inode)->wpages);
 
 	if (!data)
 		goto out_bad;
@@ -1351,7 +1366,7 @@ int nfs_commit_inode(struct inode *inode, int how)
 	res = nfs_scan_commit(inode, &head, 0, 0);
 	spin_unlock(&nfsi->req_lock);
 	if (res) {
-		error = nfs_commit_list(&head, how);
+		error = nfs_commit_list(inode, &head, how);
 		if (error < 0)
 			return error;
 	}
diff --git a/include/linux/nfs_fs.h b/include/linux/nfs_fs.h
index 4dff705d2ff..d38010ba647 100644
--- a/include/linux/nfs_fs.h
+++ b/include/linux/nfs_fs.h
@@ -38,9 +38,6 @@
 # define NFS_DEBUG
 #endif
 
-#define NFS_MAX_FILE_IO_BUFFER_SIZE	32768
-#define NFS_DEF_FILE_IO_BUFFER_SIZE	4096
-
 /* Default timeout values */
 #define NFS_MAX_UDP_TIMEOUT	(60*HZ)
 #define NFS_MAX_TCP_TIMEOUT	(600*HZ)
@@ -462,18 +459,33 @@ static inline int nfs_wb_page(struct inode *inode, struct page* page)
  */
 extern mempool_t *nfs_wdata_mempool;
 
-static inline struct nfs_write_data *nfs_writedata_alloc(void)
+static inline struct nfs_write_data *nfs_writedata_alloc(unsigned int pagecount)
 {
 	struct nfs_write_data *p = mempool_alloc(nfs_wdata_mempool, SLAB_NOFS);
+
 	if (p) {
 		memset(p, 0, sizeof(*p));
 		INIT_LIST_HEAD(&p->pages);
+		if (pagecount < NFS_PAGEVEC_SIZE)
+			p->pagevec = &p->page_array[0];
+		else {
+			size_t size = ++pagecount * sizeof(struct page *);
+			p->pagevec = kmalloc(size, GFP_NOFS);
+			if (p->pagevec) {
+				memset(p->pagevec, 0, size);
+			} else {
+				mempool_free(p, nfs_wdata_mempool);
+				p = NULL;
+			}
+		}
 	}
 	return p;
 }
 
 static inline void nfs_writedata_free(struct nfs_write_data *p)
 {
+	if (p && (p->pagevec != &p->page_array[0]))
+		kfree(p->pagevec);
 	mempool_free(p, nfs_wdata_mempool);
 }
 
@@ -492,16 +504,33 @@ extern void  nfs_readdata_release(void *data);
  */
 extern mempool_t *nfs_rdata_mempool;
 
-static inline struct nfs_read_data *nfs_readdata_alloc(void)
+static inline struct nfs_read_data *nfs_readdata_alloc(unsigned int pagecount)
 {
 	struct nfs_read_data *p = mempool_alloc(nfs_rdata_mempool, SLAB_NOFS);
-	if (p)
+
+	if (p) {
 		memset(p, 0, sizeof(*p));
+		INIT_LIST_HEAD(&p->pages);
+		if (pagecount < NFS_PAGEVEC_SIZE)
+			p->pagevec = &p->page_array[0];
+		else {
+			size_t size = ++pagecount * sizeof(struct page *);
+			p->pagevec = kmalloc(size, GFP_NOFS);
+			if (p->pagevec) {
+				memset(p->pagevec, 0, size);
+			} else {
+				mempool_free(p, nfs_rdata_mempool);
+				p = NULL;
+			}
+		}
+	}
 	return p;
 }
 
 static inline void nfs_readdata_free(struct nfs_read_data *p)
 {
+	if (p && (p->pagevec != &p->page_array[0]))
+		kfree(p->pagevec);
 	mempool_free(p, nfs_rdata_mempool);
 }
 
diff --git a/include/linux/nfs_xdr.h b/include/linux/nfs_xdr.h
index b8b0eed98ec..9f422fd8767 100644
--- a/include/linux/nfs_xdr.h
+++ b/include/linux/nfs_xdr.h
@@ -4,6 +4,16 @@
 #include <linux/sunrpc/xprt.h>
 #include <linux/nfsacl.h>
 
+/*
+ * To change the maximum rsize and wsize supported by the NFS client, adjust
+ * NFS_MAX_FILE_IO_SIZE.  64KB is a typical maximum, but some servers can
+ * support a megabyte or more.  The default is left at 4096 bytes, which is
+ * reasonable for NFS over UDP.
+ */
+#define NFS_MAX_FILE_IO_SIZE	(1048576U)
+#define NFS_DEF_FILE_IO_SIZE	(4096U)
+#define NFS_MIN_FILE_IO_SIZE	(1024U)
+
 struct nfs4_fsid {
 	__u64 major;
 	__u64 minor;
@@ -215,12 +225,6 @@ struct nfs4_delegreturnargs {
 /*
  * Arguments to the read call.
  */
-
-#define NFS_READ_MAXIOV		(9U)
-#if (NFS_READ_MAXIOV > (MAX_IOVEC -2))
-#error "NFS_READ_MAXIOV is too large"
-#endif
-
 struct nfs_readargs {
 	struct nfs_fh *		fh;
 	struct nfs_open_context *context;
@@ -239,11 +243,6 @@ struct nfs_readres {
 /*
  * Arguments to the write call.
  */
-#define NFS_WRITE_MAXIOV	(9U)
-#if (NFS_WRITE_MAXIOV > (MAX_IOVEC -2))
-#error "NFS_WRITE_MAXIOV is too large"
-#endif
-
 struct nfs_writeargs {
 	struct nfs_fh *		fh;
 	struct nfs_open_context *context;
@@ -674,6 +673,8 @@ struct nfs4_server_caps_res {
 
 struct nfs_page;
 
+#define NFS_PAGEVEC_SIZE	(8U)
+
 struct nfs_read_data {
 	int			flags;
 	struct rpc_task		task;
@@ -682,13 +683,14 @@ struct nfs_read_data {
 	struct nfs_fattr	fattr;	/* fattr storage */
 	struct list_head	pages;	/* Coalesced read requests */
 	struct nfs_page		*req;	/* multi ops per nfs_page */
-	struct page		*pagevec[NFS_READ_MAXIOV];
+	struct page		**pagevec;
 	struct nfs_readargs args;
 	struct nfs_readres  res;
 #ifdef CONFIG_NFS_V4
 	unsigned long		timestamp;	/* For lease renewal */
 #endif
 	void (*complete) (struct nfs_read_data *, int);
+	struct page		*page_array[NFS_PAGEVEC_SIZE + 1];
 };
 
 struct nfs_write_data {
@@ -700,13 +702,14 @@ struct nfs_write_data {
 	struct nfs_writeverf	verf;
 	struct list_head	pages;		/* Coalesced requests we wish to flush */
 	struct nfs_page		*req;		/* multi ops per nfs_page */
-	struct page		*pagevec[NFS_WRITE_MAXIOV];
+	struct page		**pagevec;
 	struct nfs_writeargs	args;		/* argument struct */
 	struct nfs_writeres	res;		/* result struct */
 #ifdef CONFIG_NFS_V4
 	unsigned long		timestamp;	/* For lease renewal */
 #endif
 	void (*complete) (struct nfs_write_data *, int);
+	struct page		*page_array[NFS_PAGEVEC_SIZE + 1];
 };
 
 struct nfs_access_entry;
diff --git a/include/linux/sunrpc/xdr.h b/include/linux/sunrpc/xdr.h
index 5da968729cf..5676794ee34 100644
--- a/include/linux/sunrpc/xdr.h
+++ b/include/linux/sunrpc/xdr.h
@@ -135,11 +135,6 @@ xdr_adjust_iovec(struct kvec *iov, u32 *p)
 }
 
 /*
- * Maximum number of iov's we use.
- */
-#define MAX_IOVEC	(12)
-
-/*
  * XDR buffer helper functions
  */
 extern void xdr_shift_buf(struct xdr_buf *, size_t);