blob: 5e8d82f6666be9e1cd4ec1d4d645c394e115d2f1 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/fs/nfs/direct.c
3 *
4 * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
5 *
6 * High-performance uncached I/O for the Linux NFS client
7 *
8 * There are important applications whose performance or correctness
9 * depends on uncached access to file data. Database clusters
Chuck Lever88467052006-03-20 13:44:34 -050010 * (multiple copies of the same instance running on separate hosts)
Linus Torvalds1da177e2005-04-16 15:20:36 -070011 * implement their own cache coherency protocol that subsumes file
Chuck Lever88467052006-03-20 13:44:34 -050012 * system cache protocols. Applications that process datasets
13 * considerably larger than the client's memory do not always benefit
14 * from a local cache. A streaming video server, for instance, has no
Linus Torvalds1da177e2005-04-16 15:20:36 -070015 * need to cache the contents of a file.
16 *
17 * When an application requests uncached I/O, all read and write requests
18 * are made directly to the server; data stored or fetched via these
19 * requests is not cached in the Linux page cache. The client does not
20 * correct unaligned requests from applications. All requested bytes are
21 * held on permanent storage before a direct write system call returns to
22 * an application.
23 *
24 * Solaris implements an uncached I/O facility called directio() that
25 * is used for backups and sequential I/O to very large files. Solaris
26 * also supports uncaching whole NFS partitions with "-o forcedirectio,"
27 * an undocumented mount option.
28 *
29 * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
30 * help from Andrew Morton.
31 *
32 * 18 Dec 2001 Initial implementation for 2.4 --cel
33 * 08 Jul 2002 Version for 2.4.19, with bug fixes --trondmy
34 * 08 Jun 2003 Port to 2.5 APIs --cel
35 * 31 Mar 2004 Handle direct I/O without VFS support --cel
36 * 15 Sep 2004 Parallel async reads --cel
Chuck Lever88467052006-03-20 13:44:34 -050037 * 04 May 2005 support O_DIRECT with aio --cel
Linus Torvalds1da177e2005-04-16 15:20:36 -070038 *
39 */
40
Linus Torvalds1da177e2005-04-16 15:20:36 -070041#include <linux/errno.h>
42#include <linux/sched.h>
43#include <linux/kernel.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070044#include <linux/file.h>
45#include <linux/pagemap.h>
46#include <linux/kref.h>
47
48#include <linux/nfs_fs.h>
49#include <linux/nfs_page.h>
50#include <linux/sunrpc/clnt.h>
51
52#include <asm/system.h>
53#include <asm/uaccess.h>
54#include <asm/atomic.h>
55
Trond Myklebust8d5658c2007-04-10 09:26:35 -040056#include "internal.h"
Chuck Lever91d5b472006-03-20 13:44:14 -050057#include "iostat.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070058
Linus Torvalds1da177e2005-04-16 15:20:36 -070059#define NFSDBG_FACILITY NFSDBG_VFS
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
Christoph Lametere18b8902006-12-06 20:33:20 -080061static struct kmem_cache *nfs_direct_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -070062
63/*
64 * This represents a set of asynchronous requests that we're waiting on
65 */
66struct nfs_direct_req {
67 struct kref kref; /* release manager */
Chuck Lever15ce4a02006-03-20 13:44:34 -050068
69 /* I/O parameters */
Trond Myklebusta8881f52006-03-20 13:44:36 -050070 struct nfs_open_context *ctx; /* file open context info */
Chuck Lever99514f82006-03-20 13:44:30 -050071 struct kiocb * iocb; /* controlling i/o request */
Chuck Lever88467052006-03-20 13:44:34 -050072 struct inode * inode; /* target file of i/o */
Chuck Lever15ce4a02006-03-20 13:44:34 -050073
74 /* completion state */
Trond Myklebust607f31e2006-06-28 16:52:45 -040075 atomic_t io_count; /* i/os we're waiting for */
Chuck Lever15ce4a02006-03-20 13:44:34 -050076 spinlock_t lock; /* protect completion state */
Chuck Lever15ce4a02006-03-20 13:44:34 -050077 ssize_t count, /* bytes actually processed */
Linus Torvalds1da177e2005-04-16 15:20:36 -070078 error; /* any reported error */
Trond Myklebustd72b7a62006-03-20 13:44:43 -050079 struct completion completion; /* wait for i/o completion */
Trond Myklebustfad61492006-03-20 13:44:36 -050080
81 /* commit state */
Trond Myklebust607f31e2006-06-28 16:52:45 -040082 struct list_head rewrite_list; /* saved nfs_write_data structs */
Trond Myklebustfad61492006-03-20 13:44:36 -050083 struct nfs_write_data * commit_data; /* special write_data for commits */
84 int flags;
85#define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
86#define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
87 struct nfs_writeverf verf; /* unstable write verifier */
Linus Torvalds1da177e2005-04-16 15:20:36 -070088};
89
Trond Myklebustfad61492006-03-20 13:44:36 -050090static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
Trond Myklebust607f31e2006-06-28 16:52:45 -040091static const struct rpc_call_ops nfs_write_direct_ops;
92
93static inline void get_dreq(struct nfs_direct_req *dreq)
94{
95 atomic_inc(&dreq->io_count);
96}
97
98static inline int put_dreq(struct nfs_direct_req *dreq)
99{
100 return atomic_dec_and_test(&dreq->io_count);
101}
102
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103/**
Chuck Leverb8a32e22006-03-20 13:44:28 -0500104 * nfs_direct_IO - NFS address space operation for direct I/O
105 * @rw: direction (read or write)
106 * @iocb: target I/O control block
107 * @iov: array of vectors that define I/O buffer
108 * @pos: offset in file to begin the operation
109 * @nr_segs: size of iovec array
110 *
111 * The presence of this routine in the address space ops vector means
112 * the NFS client supports direct I/O. However, we shunt off direct
113 * read and write requests before the VFS gets them, so this method
114 * should never be called.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115 */
Chuck Leverb8a32e22006-03-20 13:44:28 -0500116ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
117{
Chuck Leverb8a32e22006-03-20 13:44:28 -0500118 dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
Josef "Jeff" Sipek01cce932006-12-08 02:36:40 -0800119 iocb->ki_filp->f_path.dentry->d_name.name,
Trond Myklebuste99170f2006-04-18 13:21:42 -0400120 (long long) pos, nr_segs);
Chuck Leverb8a32e22006-03-20 13:44:28 -0500121
122 return -EINVAL;
123}
124
Trond Myklebustd4a8f362007-05-22 10:22:27 -0400125static void nfs_direct_dirty_pages(struct page **pages, unsigned int pgbase, size_t count)
Trond Myklebust6b45d852006-03-20 13:44:43 -0500126{
Trond Myklebustd4a8f362007-05-22 10:22:27 -0400127 unsigned int npages;
Chuck Lever749e1462007-05-19 17:22:46 -0400128 unsigned int i;
Trond Myklebustd4a8f362007-05-22 10:22:27 -0400129
130 if (count == 0)
131 return;
132 pages += (pgbase >> PAGE_SHIFT);
133 npages = (count + (pgbase & ~PAGE_MASK) + PAGE_SIZE - 1) >> PAGE_SHIFT;
Trond Myklebust6b45d852006-03-20 13:44:43 -0500134 for (i = 0; i < npages; i++) {
135 struct page *page = pages[i];
Trond Myklebust607f31e2006-06-28 16:52:45 -0400136 if (!PageCompound(page))
Trond Myklebustd4a8f362007-05-22 10:22:27 -0400137 set_page_dirty(page);
Trond Myklebust6b45d852006-03-20 13:44:43 -0500138 }
Chuck Lever9c93ab72006-06-20 12:56:31 -0400139}
140
Chuck Lever749e1462007-05-19 17:22:46 -0400141static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
Chuck Lever9c93ab72006-06-20 12:56:31 -0400142{
Chuck Lever749e1462007-05-19 17:22:46 -0400143 unsigned int i;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400144 for (i = 0; i < npages; i++)
145 page_cache_release(pages[i]);
Trond Myklebust6b45d852006-03-20 13:44:43 -0500146}
147
Chuck Lever93619e52006-03-20 13:44:31 -0500148static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150 struct nfs_direct_req *dreq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151
Christoph Lametere94b1762006-12-06 20:33:17 -0800152 dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153 if (!dreq)
154 return NULL;
155
156 kref_init(&dreq->kref);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400157 kref_get(&dreq->kref);
Trond Myklebustd72b7a62006-03-20 13:44:43 -0500158 init_completion(&dreq->completion);
Trond Myklebustfad61492006-03-20 13:44:36 -0500159 INIT_LIST_HEAD(&dreq->rewrite_list);
Chuck Lever93619e52006-03-20 13:44:31 -0500160 dreq->iocb = NULL;
Trond Myklebusta8881f52006-03-20 13:44:36 -0500161 dreq->ctx = NULL;
Chuck Lever15ce4a02006-03-20 13:44:34 -0500162 spin_lock_init(&dreq->lock);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400163 atomic_set(&dreq->io_count, 0);
Chuck Lever15ce4a02006-03-20 13:44:34 -0500164 dreq->count = 0;
165 dreq->error = 0;
Trond Myklebustfad61492006-03-20 13:44:36 -0500166 dreq->flags = 0;
Chuck Lever93619e52006-03-20 13:44:31 -0500167
168 return dreq;
169}
170
Trond Myklebustb4946ff2007-05-30 12:58:00 -0400171static void nfs_direct_req_free(struct kref *kref)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172{
173 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
Trond Myklebusta8881f52006-03-20 13:44:36 -0500174
175 if (dreq->ctx != NULL)
176 put_nfs_open_context(dreq->ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 kmem_cache_free(nfs_direct_cachep, dreq);
178}
179
Trond Myklebustb4946ff2007-05-30 12:58:00 -0400180static void nfs_direct_req_release(struct nfs_direct_req *dreq)
181{
182 kref_put(&dreq->kref, nfs_direct_req_free);
183}
184
Chuck Leverd4cc9482006-03-20 13:44:28 -0500185/*
Chuck Leverbc0fb202006-03-20 13:44:31 -0500186 * Collects and returns the final error value/byte-count.
187 */
188static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
189{
Chuck Lever15ce4a02006-03-20 13:44:34 -0500190 ssize_t result = -EIOCBQUEUED;
Chuck Leverbc0fb202006-03-20 13:44:31 -0500191
192 /* Async requests don't wait here */
193 if (dreq->iocb)
194 goto out;
195
Trond Myklebustd72b7a62006-03-20 13:44:43 -0500196 result = wait_for_completion_interruptible(&dreq->completion);
Chuck Leverbc0fb202006-03-20 13:44:31 -0500197
198 if (!result)
Chuck Lever15ce4a02006-03-20 13:44:34 -0500199 result = dreq->error;
Chuck Leverbc0fb202006-03-20 13:44:31 -0500200 if (!result)
Chuck Lever15ce4a02006-03-20 13:44:34 -0500201 result = dreq->count;
Chuck Leverbc0fb202006-03-20 13:44:31 -0500202
203out:
Chuck Leverbc0fb202006-03-20 13:44:31 -0500204 return (ssize_t) result;
205}
206
207/*
Trond Myklebust607f31e2006-06-28 16:52:45 -0400208 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
209 * the iocb is still valid here if this is a synchronous request.
Chuck Lever63ab46a2006-03-20 13:44:31 -0500210 */
211static void nfs_direct_complete(struct nfs_direct_req *dreq)
212{
Chuck Lever63ab46a2006-03-20 13:44:31 -0500213 if (dreq->iocb) {
Chuck Lever15ce4a02006-03-20 13:44:34 -0500214 long res = (long) dreq->error;
Chuck Lever63ab46a2006-03-20 13:44:31 -0500215 if (!res)
Chuck Lever15ce4a02006-03-20 13:44:34 -0500216 res = (long) dreq->count;
Chuck Lever63ab46a2006-03-20 13:44:31 -0500217 aio_complete(dreq->iocb, res, 0);
Trond Myklebustd72b7a62006-03-20 13:44:43 -0500218 }
219 complete_all(&dreq->completion);
Chuck Lever63ab46a2006-03-20 13:44:31 -0500220
Trond Myklebustb4946ff2007-05-30 12:58:00 -0400221 nfs_direct_req_release(dreq);
Chuck Lever63ab46a2006-03-20 13:44:31 -0500222}
223
224/*
Trond Myklebust607f31e2006-06-28 16:52:45 -0400225 * We must hold a reference to all the pages in this direct read request
226 * until the RPCs complete. This could be long *after* we are woken up in
227 * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
Chuck Lever06cf6f22006-06-20 12:56:49 -0400228 */
Trond Myklebustec06c092006-03-20 13:44:27 -0500229static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230{
Trond Myklebustec06c092006-03-20 13:44:27 -0500231 struct nfs_read_data *data = calldata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
233
Trond Myklebustec06c092006-03-20 13:44:27 -0500234 if (nfs_readpage_result(task, data) != 0)
235 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236
Chuck Lever15ce4a02006-03-20 13:44:34 -0500237 spin_lock(&dreq->lock);
Trond Myklebustd4a8f362007-05-22 10:22:27 -0400238 if (unlikely(task->tk_status < 0)) {
Chuck Lever15ce4a02006-03-20 13:44:34 -0500239 dreq->error = task->tk_status;
Trond Myklebustd4a8f362007-05-22 10:22:27 -0400240 spin_unlock(&dreq->lock);
241 } else {
242 dreq->count += data->res.count;
243 spin_unlock(&dreq->lock);
244 nfs_direct_dirty_pages(data->pagevec,
245 data->args.pgbase,
246 data->res.count);
247 }
248 nfs_direct_release_pages(data->pagevec, data->npages);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400249
250 if (put_dreq(dreq))
251 nfs_direct_complete(dreq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252}
253
Trond Myklebustec06c092006-03-20 13:44:27 -0500254static const struct rpc_call_ops nfs_read_direct_ops = {
255 .rpc_call_done = nfs_direct_read_result,
256 .rpc_release = nfs_readdata_release,
257};
258
Chuck Leverd4cc9482006-03-20 13:44:28 -0500259/*
Trond Myklebust607f31e2006-06-28 16:52:45 -0400260 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
261 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
262 * bail and stop sending more reads. Read length accounting is
263 * handled automatically by nfs_direct_read_result(). Otherwise, if
264 * no requests have been sent, just return an error.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265 */
Chuck Lever02fe4942007-11-12 12:17:03 -0500266static ssize_t nfs_direct_read_schedule_segment(struct nfs_direct_req *dreq,
267 const struct iovec *iov,
268 loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269{
Trond Myklebusta8881f52006-03-20 13:44:36 -0500270 struct nfs_open_context *ctx = dreq->ctx;
Trond Myklebust88be9f92007-06-05 10:42:27 -0400271 struct inode *inode = ctx->path.dentry->d_inode;
Chuck Lever02fe4942007-11-12 12:17:03 -0500272 unsigned long user_addr = (unsigned long)iov->iov_base;
273 size_t count = iov->iov_len;
Chuck Lever5dd602f2006-03-20 13:44:29 -0500274 size_t rsize = NFS_SERVER(inode)->rsize;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400275 unsigned int pgbase;
276 int result;
277 ssize_t started = 0;
Chuck Lever82b145c2006-06-20 12:57:03 -0400278
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279 do {
Chuck Lever82b145c2006-06-20 12:57:03 -0400280 struct nfs_read_data *data;
Chuck Lever5dd602f2006-03-20 13:44:29 -0500281 size_t bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700283 pgbase = user_addr & ~PAGE_MASK;
284 bytes = min(rsize,count);
285
Trond Myklebust607f31e2006-06-28 16:52:45 -0400286 result = -ENOMEM;
Trond Myklebust8d5658c2007-04-10 09:26:35 -0400287 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
Trond Myklebust607f31e2006-06-28 16:52:45 -0400288 if (unlikely(!data))
289 break;
290
Trond Myklebust607f31e2006-06-28 16:52:45 -0400291 down_read(&current->mm->mmap_sem);
292 result = get_user_pages(current, current->mm, user_addr,
293 data->npages, 1, 0, data->pagevec, NULL);
294 up_read(&current->mm->mmap_sem);
Chuck Lever749e1462007-05-19 17:22:46 -0400295 if (result < 0) {
296 nfs_readdata_release(data);
297 break;
298 }
299 if ((unsigned)result < data->npages) {
Trond Myklebustd9df8d62007-05-22 10:22:20 -0400300 bytes = result * PAGE_SIZE;
301 if (bytes <= pgbase) {
302 nfs_direct_release_pages(data->pagevec, result);
303 nfs_readdata_release(data);
304 break;
305 }
306 bytes -= pgbase;
307 data->npages = result;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400308 }
Chuck Lever06cf6f22006-06-20 12:56:49 -0400309
Trond Myklebust607f31e2006-06-28 16:52:45 -0400310 get_dreq(dreq);
311
312 data->req = (struct nfs_page *) dreq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313 data->inode = inode;
314 data->cred = ctx->cred;
315 data->args.fh = NFS_FH(inode);
316 data->args.context = ctx;
Chuck Lever88467052006-03-20 13:44:34 -0500317 data->args.offset = pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 data->args.pgbase = pgbase;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400319 data->args.pages = data->pagevec;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320 data->args.count = bytes;
321 data->res.fattr = &data->fattr;
322 data->res.eof = 0;
323 data->res.count = bytes;
324
Trond Myklebustec06c092006-03-20 13:44:27 -0500325 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
326 &nfs_read_direct_ops, data);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327 NFS_PROTO(inode)->read_setup(data);
328
329 data->task.tk_cookie = (unsigned long) inode;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331 rpc_execute(&data->task);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332
Chuck Levera3f565b2007-01-31 12:14:01 -0500333 dprintk("NFS: %5u initiated direct read call "
334 "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335 data->task.tk_pid,
336 inode->i_sb->s_id,
337 (long long)NFS_FILEID(inode),
338 bytes,
339 (unsigned long long)data->args.offset);
340
Trond Myklebust607f31e2006-06-28 16:52:45 -0400341 started += bytes;
342 user_addr += bytes;
Chuck Lever88467052006-03-20 13:44:34 -0500343 pos += bytes;
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700344 /* FIXME: Remove this unnecessary math from final patch */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345 pgbase += bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346 pgbase &= ~PAGE_MASK;
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700347 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348
349 count -= bytes;
350 } while (count != 0);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400351
Trond Myklebust607f31e2006-06-28 16:52:45 -0400352 if (started)
Chuck Leverc216fd72007-11-12 12:16:52 -0500353 return started;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400354 return result < 0 ? (ssize_t) result : -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355}
356
Chuck Lever19f73782007-11-12 12:16:47 -0500357static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
358 const struct iovec *iov,
359 unsigned long nr_segs,
360 loff_t pos)
361{
362 ssize_t result = -EINVAL;
363 size_t requested_bytes = 0;
364 unsigned long seg;
365
366 get_dreq(dreq);
367
368 for (seg = 0; seg < nr_segs; seg++) {
369 const struct iovec *vec = &iov[seg];
Chuck Lever02fe4942007-11-12 12:17:03 -0500370 result = nfs_direct_read_schedule_segment(dreq, vec, pos);
Chuck Lever19f73782007-11-12 12:16:47 -0500371 if (result < 0)
372 break;
373 requested_bytes += result;
374 if ((size_t)result < vec->iov_len)
375 break;
376 pos += vec->iov_len;
377 }
378
379 if (put_dreq(dreq))
380 nfs_direct_complete(dreq);
381
382 if (requested_bytes != 0)
383 return 0;
384
385 if (result < 0)
386 return result;
387 return -EIO;
388}
389
Chuck Leverc216fd72007-11-12 12:16:52 -0500390static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
391 unsigned long nr_segs, loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392{
Trond Myklebust607f31e2006-06-28 16:52:45 -0400393 ssize_t result = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394 sigset_t oldset;
Chuck Lever99514f82006-03-20 13:44:30 -0500395 struct inode *inode = iocb->ki_filp->f_mapping->host;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396 struct rpc_clnt *clnt = NFS_CLIENT(inode);
397 struct nfs_direct_req *dreq;
398
Trond Myklebust607f31e2006-06-28 16:52:45 -0400399 dreq = nfs_direct_req_alloc();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400 if (!dreq)
401 return -ENOMEM;
402
Chuck Lever91d5b472006-03-20 13:44:14 -0500403 dreq->inode = inode;
Trond Myklebustcd3758e2007-08-10 17:44:32 -0400404 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
Chuck Lever487b8372006-03-20 13:44:30 -0500405 if (!is_sync_kiocb(iocb))
406 dreq->iocb = iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407
408 rpc_clnt_sigmask(clnt, &oldset);
Chuck Leverc216fd72007-11-12 12:16:52 -0500409 result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400410 if (!result)
411 result = nfs_direct_wait(dreq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412 rpc_clnt_sigunmask(clnt, &oldset);
Trond Myklebustb4946ff2007-05-30 12:58:00 -0400413 nfs_direct_req_release(dreq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414
415 return result;
416}
417
Trond Myklebustfad61492006-03-20 13:44:36 -0500418static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419{
Trond Myklebust607f31e2006-06-28 16:52:45 -0400420 while (!list_empty(&dreq->rewrite_list)) {
421 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
Trond Myklebustfad61492006-03-20 13:44:36 -0500422 list_del(&data->pages);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400423 nfs_direct_release_pages(data->pagevec, data->npages);
Trond Myklebustfad61492006-03-20 13:44:36 -0500424 nfs_writedata_release(data);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426}
427
Trond Myklebustfad61492006-03-20 13:44:36 -0500428#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
429static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430{
Trond Myklebust607f31e2006-06-28 16:52:45 -0400431 struct inode *inode = dreq->inode;
432 struct list_head *p;
433 struct nfs_write_data *data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434
Trond Myklebustfad61492006-03-20 13:44:36 -0500435 dreq->count = 0;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400436 get_dreq(dreq);
Trond Myklebustfad61492006-03-20 13:44:36 -0500437
Trond Myklebust607f31e2006-06-28 16:52:45 -0400438 list_for_each(p, &dreq->rewrite_list) {
439 data = list_entry(p, struct nfs_write_data, pages);
440
441 get_dreq(dreq);
442
443 /*
444 * Reset data->res.
445 */
446 nfs_fattr_init(&data->fattr);
447 data->res.count = data->args.count;
448 memset(&data->verf, 0, sizeof(data->verf));
449
450 /*
451 * Reuse data->task; data->args should not have changed
452 * since the original request was sent.
453 */
454 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
455 &nfs_write_direct_ops, data);
456 NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE);
457
458 data->task.tk_priority = RPC_PRIORITY_NORMAL;
459 data->task.tk_cookie = (unsigned long) inode;
460
461 /*
462 * We're called via an RPC callback, so BKL is already held.
463 */
464 rpc_execute(&data->task);
465
466 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
467 data->task.tk_pid,
468 inode->i_sb->s_id,
469 (long long)NFS_FILEID(inode),
470 data->args.count,
471 (unsigned long long)data->args.offset);
472 }
473
474 if (put_dreq(dreq))
475 nfs_direct_write_complete(dreq, inode);
Trond Myklebustfad61492006-03-20 13:44:36 -0500476}
477
478static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
479{
480 struct nfs_write_data *data = calldata;
481 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
482
483 /* Call the NFS version-specific code */
484 if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
485 return;
486 if (unlikely(task->tk_status < 0)) {
Trond Myklebust60fa3f72007-04-14 19:11:52 -0400487 dprintk("NFS: %5u commit failed with error %d.\n",
488 task->tk_pid, task->tk_status);
Trond Myklebustfad61492006-03-20 13:44:36 -0500489 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
Trond Myklebust60fa3f72007-04-14 19:11:52 -0400490 } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
Trond Myklebustfad61492006-03-20 13:44:36 -0500491 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
492 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
493 }
494
495 dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
496 nfs_direct_write_complete(dreq, data->inode);
497}
498
499static const struct rpc_call_ops nfs_commit_direct_ops = {
500 .rpc_call_done = nfs_direct_commit_result,
501 .rpc_release = nfs_commit_release,
502};
503
504static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
505{
Trond Myklebustfad61492006-03-20 13:44:36 -0500506 struct nfs_write_data *data = dreq->commit_data;
Trond Myklebustfad61492006-03-20 13:44:36 -0500507
508 data->inode = dreq->inode;
Trond Myklebusta8881f52006-03-20 13:44:36 -0500509 data->cred = dreq->ctx->cred;
Trond Myklebustfad61492006-03-20 13:44:36 -0500510
511 data->args.fh = NFS_FH(data->inode);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400512 data->args.offset = 0;
513 data->args.count = 0;
Trond Myklebustfad61492006-03-20 13:44:36 -0500514 data->res.count = 0;
515 data->res.fattr = &data->fattr;
516 data->res.verf = &data->verf;
517
518 rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
519 &nfs_commit_direct_ops, data);
520 NFS_PROTO(data->inode)->commit_setup(data, 0);
521
522 data->task.tk_priority = RPC_PRIORITY_NORMAL;
523 data->task.tk_cookie = (unsigned long)data->inode;
524 /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
525 dreq->commit_data = NULL;
526
Trond Myklebuste99170f2006-04-18 13:21:42 -0400527 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
Trond Myklebustfad61492006-03-20 13:44:36 -0500528
Trond Myklebustfad61492006-03-20 13:44:36 -0500529 rpc_execute(&data->task);
Trond Myklebustfad61492006-03-20 13:44:36 -0500530}
531
532static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
533{
534 int flags = dreq->flags;
535
536 dreq->flags = 0;
537 switch (flags) {
538 case NFS_ODIRECT_DO_COMMIT:
539 nfs_direct_commit_schedule(dreq);
540 break;
541 case NFS_ODIRECT_RESCHED_WRITES:
542 nfs_direct_write_reschedule(dreq);
543 break;
544 default:
Trond Myklebustfad61492006-03-20 13:44:36 -0500545 if (dreq->commit_data != NULL)
546 nfs_commit_free(dreq->commit_data);
547 nfs_direct_free_writedata(dreq);
Trond Myklebustcd9ae2b2006-10-19 23:28:40 -0700548 nfs_zap_mapping(inode, inode->i_mapping);
Trond Myklebustfad61492006-03-20 13:44:36 -0500549 nfs_direct_complete(dreq);
550 }
551}
552
553static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
554{
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700555 dreq->commit_data = nfs_commit_alloc();
Trond Myklebustfad61492006-03-20 13:44:36 -0500556 if (dreq->commit_data != NULL)
557 dreq->commit_data->req = (struct nfs_page *) dreq;
558}
559#else
560static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
561{
562 dreq->commit_data = NULL;
563}
564
565static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
566{
Trond Myklebustfad61492006-03-20 13:44:36 -0500567 nfs_direct_free_writedata(dreq);
Trond Myklebustcd9ae2b2006-10-19 23:28:40 -0700568 nfs_zap_mapping(inode, inode->i_mapping);
Trond Myklebustfad61492006-03-20 13:44:36 -0500569 nfs_direct_complete(dreq);
570}
571#endif
572
Chuck Lever462d5b32006-03-20 13:44:32 -0500573static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
574{
575 struct nfs_write_data *data = calldata;
576 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
577 int status = task->tk_status;
578
579 if (nfs_writeback_done(task, data) != 0)
580 return;
Chuck Lever462d5b32006-03-20 13:44:32 -0500581
Chuck Lever15ce4a02006-03-20 13:44:34 -0500582 spin_lock(&dreq->lock);
Chuck Lever462d5b32006-03-20 13:44:32 -0500583
Trond Myklebust60fa3f72007-04-14 19:11:52 -0400584 if (unlikely(status < 0)) {
Neil Brown432409e2007-10-23 17:09:13 -0400585 /* An error has occurred, so we should not commit */
Trond Myklebust60fa3f72007-04-14 19:11:52 -0400586 dreq->flags = 0;
587 dreq->error = status;
Trond Myklebusteda3cef2006-10-19 23:28:38 -0700588 }
Neil Brown432409e2007-10-23 17:09:13 -0400589 if (unlikely(dreq->error != 0))
590 goto out_unlock;
Trond Myklebusteda3cef2006-10-19 23:28:38 -0700591
592 dreq->count += data->res.count;
Chuck Lever15ce4a02006-03-20 13:44:34 -0500593
Trond Myklebustfad61492006-03-20 13:44:36 -0500594 if (data->res.verf->committed != NFS_FILE_SYNC) {
595 switch (dreq->flags) {
596 case 0:
597 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
598 dreq->flags = NFS_ODIRECT_DO_COMMIT;
599 break;
600 case NFS_ODIRECT_DO_COMMIT:
601 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
602 dprintk("NFS: %5u write verify failed\n", task->tk_pid);
603 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
604 }
605 }
606 }
Trond Myklebusteda3cef2006-10-19 23:28:38 -0700607out_unlock:
Trond Myklebustfad61492006-03-20 13:44:36 -0500608 spin_unlock(&dreq->lock);
609}
610
611/*
612 * NB: Return the value of the first error return code. Subsequent
613 * errors after the first one are ignored.
614 */
615static void nfs_direct_write_release(void *calldata)
616{
617 struct nfs_write_data *data = calldata;
618 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
619
Trond Myklebust607f31e2006-06-28 16:52:45 -0400620 if (put_dreq(dreq))
621 nfs_direct_write_complete(dreq, data->inode);
Chuck Lever462d5b32006-03-20 13:44:32 -0500622}
623
624static const struct rpc_call_ops nfs_write_direct_ops = {
625 .rpc_call_done = nfs_direct_write_result,
Trond Myklebustfad61492006-03-20 13:44:36 -0500626 .rpc_release = nfs_direct_write_release,
Chuck Lever462d5b32006-03-20 13:44:32 -0500627};
628
629/*
Trond Myklebust607f31e2006-06-28 16:52:45 -0400630 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
631 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
632 * bail and stop sending more writes. Write length accounting is
633 * handled automatically by nfs_direct_write_result(). Otherwise, if
634 * no requests have been sent, just return an error.
Chuck Lever462d5b32006-03-20 13:44:32 -0500635 */
Chuck Lever02fe4942007-11-12 12:17:03 -0500636static ssize_t nfs_direct_write_schedule_segment(struct nfs_direct_req *dreq,
637 const struct iovec *iov,
638 loff_t pos, int sync)
Chuck Lever462d5b32006-03-20 13:44:32 -0500639{
Trond Myklebusta8881f52006-03-20 13:44:36 -0500640 struct nfs_open_context *ctx = dreq->ctx;
Trond Myklebust88be9f92007-06-05 10:42:27 -0400641 struct inode *inode = ctx->path.dentry->d_inode;
Chuck Lever02fe4942007-11-12 12:17:03 -0500642 unsigned long user_addr = (unsigned long)iov->iov_base;
643 size_t count = iov->iov_len;
Chuck Lever462d5b32006-03-20 13:44:32 -0500644 size_t wsize = NFS_SERVER(inode)->wsize;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400645 unsigned int pgbase;
646 int result;
647 ssize_t started = 0;
Chuck Lever82b145c2006-06-20 12:57:03 -0400648
Chuck Lever462d5b32006-03-20 13:44:32 -0500649 do {
Chuck Lever82b145c2006-06-20 12:57:03 -0400650 struct nfs_write_data *data;
Chuck Lever462d5b32006-03-20 13:44:32 -0500651 size_t bytes;
652
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700653 pgbase = user_addr & ~PAGE_MASK;
654 bytes = min(wsize,count);
655
Trond Myklebust607f31e2006-06-28 16:52:45 -0400656 result = -ENOMEM;
Trond Myklebust8d5658c2007-04-10 09:26:35 -0400657 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
Trond Myklebust607f31e2006-06-28 16:52:45 -0400658 if (unlikely(!data))
659 break;
660
Trond Myklebust607f31e2006-06-28 16:52:45 -0400661 down_read(&current->mm->mmap_sem);
662 result = get_user_pages(current, current->mm, user_addr,
663 data->npages, 0, 0, data->pagevec, NULL);
664 up_read(&current->mm->mmap_sem);
Chuck Lever749e1462007-05-19 17:22:46 -0400665 if (result < 0) {
666 nfs_writedata_release(data);
667 break;
668 }
669 if ((unsigned)result < data->npages) {
Trond Myklebustd9df8d62007-05-22 10:22:20 -0400670 bytes = result * PAGE_SIZE;
671 if (bytes <= pgbase) {
672 nfs_direct_release_pages(data->pagevec, result);
673 nfs_writedata_release(data);
674 break;
675 }
676 bytes -= pgbase;
677 data->npages = result;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400678 }
679
680 get_dreq(dreq);
681
Trond Myklebustfad61492006-03-20 13:44:36 -0500682 list_move_tail(&data->pages, &dreq->rewrite_list);
Chuck Lever462d5b32006-03-20 13:44:32 -0500683
Trond Myklebust607f31e2006-06-28 16:52:45 -0400684 data->req = (struct nfs_page *) dreq;
Chuck Lever462d5b32006-03-20 13:44:32 -0500685 data->inode = inode;
686 data->cred = ctx->cred;
687 data->args.fh = NFS_FH(inode);
688 data->args.context = ctx;
Chuck Lever88467052006-03-20 13:44:34 -0500689 data->args.offset = pos;
Chuck Lever462d5b32006-03-20 13:44:32 -0500690 data->args.pgbase = pgbase;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400691 data->args.pages = data->pagevec;
Chuck Lever462d5b32006-03-20 13:44:32 -0500692 data->args.count = bytes;
693 data->res.fattr = &data->fattr;
694 data->res.count = bytes;
Chuck Lever47989d72006-03-20 13:44:32 -0500695 data->res.verf = &data->verf;
Chuck Lever462d5b32006-03-20 13:44:32 -0500696
697 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
698 &nfs_write_direct_ops, data);
Trond Myklebustfad61492006-03-20 13:44:36 -0500699 NFS_PROTO(inode)->write_setup(data, sync);
Chuck Lever462d5b32006-03-20 13:44:32 -0500700
701 data->task.tk_priority = RPC_PRIORITY_NORMAL;
702 data->task.tk_cookie = (unsigned long) inode;
703
Chuck Lever462d5b32006-03-20 13:44:32 -0500704 rpc_execute(&data->task);
Chuck Lever462d5b32006-03-20 13:44:32 -0500705
Chuck Levera3f565b2007-01-31 12:14:01 -0500706 dprintk("NFS: %5u initiated direct write call "
707 "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
Chuck Lever462d5b32006-03-20 13:44:32 -0500708 data->task.tk_pid,
709 inode->i_sb->s_id,
710 (long long)NFS_FILEID(inode),
711 bytes,
712 (unsigned long long)data->args.offset);
713
Trond Myklebust607f31e2006-06-28 16:52:45 -0400714 started += bytes;
715 user_addr += bytes;
Chuck Lever88467052006-03-20 13:44:34 -0500716 pos += bytes;
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700717
718 /* FIXME: Remove this useless math from the final patch */
Chuck Lever462d5b32006-03-20 13:44:32 -0500719 pgbase += bytes;
Chuck Lever462d5b32006-03-20 13:44:32 -0500720 pgbase &= ~PAGE_MASK;
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700721 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
Chuck Lever462d5b32006-03-20 13:44:32 -0500722
723 count -= bytes;
724 } while (count != 0);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400725
Trond Myklebust607f31e2006-06-28 16:52:45 -0400726 if (started)
Chuck Leverc216fd72007-11-12 12:16:52 -0500727 return started;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400728 return result < 0 ? (ssize_t) result : -EFAULT;
Chuck Lever462d5b32006-03-20 13:44:32 -0500729}
730
Chuck Lever19f73782007-11-12 12:16:47 -0500731static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
732 const struct iovec *iov,
733 unsigned long nr_segs,
734 loff_t pos, int sync)
735{
736 ssize_t result = 0;
737 size_t requested_bytes = 0;
738 unsigned long seg;
739
740 get_dreq(dreq);
741
742 for (seg = 0; seg < nr_segs; seg++) {
743 const struct iovec *vec = &iov[seg];
Chuck Lever02fe4942007-11-12 12:17:03 -0500744 result = nfs_direct_write_schedule_segment(dreq, vec,
745 pos, sync);
Chuck Lever19f73782007-11-12 12:16:47 -0500746 if (result < 0)
747 break;
748 requested_bytes += result;
749 if ((size_t)result < vec->iov_len)
750 break;
751 pos += vec->iov_len;
752 }
753
754 if (put_dreq(dreq))
755 nfs_direct_write_complete(dreq, dreq->inode);
756
757 if (requested_bytes != 0)
758 return 0;
759
760 if (result < 0)
761 return result;
762 return -EIO;
763}
764
Chuck Leverc216fd72007-11-12 12:16:52 -0500765static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
766 unsigned long nr_segs, loff_t pos,
767 size_t count)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768{
Trond Myklebust607f31e2006-06-28 16:52:45 -0400769 ssize_t result = 0;
Chuck Lever462d5b32006-03-20 13:44:32 -0500770 sigset_t oldset;
Chuck Leverc89f2ee2006-03-20 13:44:33 -0500771 struct inode *inode = iocb->ki_filp->f_mapping->host;
Chuck Lever462d5b32006-03-20 13:44:32 -0500772 struct rpc_clnt *clnt = NFS_CLIENT(inode);
773 struct nfs_direct_req *dreq;
Trond Myklebustfad61492006-03-20 13:44:36 -0500774 size_t wsize = NFS_SERVER(inode)->wsize;
775 int sync = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776
Trond Myklebust607f31e2006-06-28 16:52:45 -0400777 dreq = nfs_direct_req_alloc();
Chuck Lever462d5b32006-03-20 13:44:32 -0500778 if (!dreq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779 return -ENOMEM;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400780 nfs_alloc_commit_data(dreq);
781
Trond Myklebustfad61492006-03-20 13:44:36 -0500782 if (dreq->commit_data == NULL || count < wsize)
783 sync = FLUSH_STABLE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784
Chuck Leverc89f2ee2006-03-20 13:44:33 -0500785 dreq->inode = inode;
Trond Myklebustcd3758e2007-08-10 17:44:32 -0400786 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
Chuck Leverc89f2ee2006-03-20 13:44:33 -0500787 if (!is_sync_kiocb(iocb))
788 dreq->iocb = iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789
Chuck Lever462d5b32006-03-20 13:44:32 -0500790 rpc_clnt_sigmask(clnt, &oldset);
Chuck Leverc216fd72007-11-12 12:16:52 -0500791 result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400792 if (!result)
793 result = nfs_direct_wait(dreq);
Chuck Lever462d5b32006-03-20 13:44:32 -0500794 rpc_clnt_sigunmask(clnt, &oldset);
Trond Myklebustb4946ff2007-05-30 12:58:00 -0400795 nfs_direct_req_release(dreq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797 return result;
798}
799
800/**
801 * nfs_file_direct_read - file direct read operation for NFS files
802 * @iocb: target I/O control block
Badari Pulavarty027445c2006-09-30 23:28:46 -0700803 * @iov: vector of user buffers into which to read data
804 * @nr_segs: size of iov vector
Chuck Lever88467052006-03-20 13:44:34 -0500805 * @pos: byte offset in file where reading starts
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806 *
807 * We use this function for direct reads instead of calling
808 * generic_file_aio_read() in order to avoid gfar's check to see if
809 * the request starts before the end of the file. For that check
810 * to work, we must generate a GETATTR before each direct read, and
811 * even then there is a window between the GETATTR and the subsequent
Chuck Lever88467052006-03-20 13:44:34 -0500812 * READ where the file size could change. Our preference is simply
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813 * to do all reads the application wants, and the server will take
814 * care of managing the end of file boundary.
Chuck Lever88467052006-03-20 13:44:34 -0500815 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816 * This function also eliminates unnecessarily updating the file's
817 * atime locally, as the NFS server sets the file's atime, and this
818 * client must read the updated atime from the server back into its
819 * cache.
820 */
Badari Pulavarty027445c2006-09-30 23:28:46 -0700821ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
822 unsigned long nr_segs, loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823{
824 ssize_t retval = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825 struct file *file = iocb->ki_filp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 struct address_space *mapping = file->f_mapping;
Chuck Leverc216fd72007-11-12 12:16:52 -0500827 size_t count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828
Chuck Leverc216fd72007-11-12 12:16:52 -0500829 count = iov_length(iov, nr_segs);
830 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
831
832 dprintk("nfs: direct read(%s/%s, %zd@%Ld)\n",
Josef "Jeff" Sipek01cce932006-12-08 02:36:40 -0800833 file->f_path.dentry->d_parent->d_name.name,
834 file->f_path.dentry->d_name.name,
Chuck Leverc216fd72007-11-12 12:16:52 -0500835 count, (long long) pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 retval = 0;
838 if (!count)
839 goto out;
840
Trond Myklebust29884df2005-12-13 16:13:54 -0500841 retval = nfs_sync_mapping(mapping);
842 if (retval)
843 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844
Chuck Leverc216fd72007-11-12 12:16:52 -0500845 retval = nfs_direct_read(iocb, iov, nr_segs, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846 if (retval > 0)
Chuck Lever0cdd80d2006-03-20 13:44:29 -0500847 iocb->ki_pos = pos + retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848
849out:
850 return retval;
851}
852
853/**
854 * nfs_file_direct_write - file direct write operation for NFS files
855 * @iocb: target I/O control block
Badari Pulavarty027445c2006-09-30 23:28:46 -0700856 * @iov: vector of user buffers from which to write data
857 * @nr_segs: size of iov vector
Chuck Lever88467052006-03-20 13:44:34 -0500858 * @pos: byte offset in file where writing starts
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 *
860 * We use this function for direct writes instead of calling
861 * generic_file_aio_write() in order to avoid taking the inode
862 * semaphore and updating the i_size. The NFS server will set
863 * the new i_size and this client must read the updated size
864 * back into its cache. We let the server do generic write
865 * parameter checking and report problems.
866 *
867 * We also avoid an unnecessary invocation of generic_osync_inode(),
868 * as it is fairly meaningless to sync the metadata of an NFS file.
869 *
870 * We eliminate local atime updates, see direct read above.
871 *
872 * We avoid unnecessary page cache invalidations for normal cached
873 * readers of this file.
874 *
875 * Note that O_APPEND is not supported for NFS direct writes, as there
876 * is no atomic O_APPEND write facility in the NFS protocol.
877 */
Badari Pulavarty027445c2006-09-30 23:28:46 -0700878ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
879 unsigned long nr_segs, loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880{
Chuck Lever070ea602007-05-19 17:22:52 -0400881 ssize_t retval = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882 struct file *file = iocb->ki_filp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883 struct address_space *mapping = file->f_mapping;
Chuck Leverc216fd72007-11-12 12:16:52 -0500884 size_t count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885
Chuck Leverc216fd72007-11-12 12:16:52 -0500886 count = iov_length(iov, nr_segs);
887 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
888
889 dfprintk(VFS, "nfs: direct write(%s/%s, %zd@%Ld)\n",
Josef "Jeff" Sipek01cce932006-12-08 02:36:40 -0800890 file->f_path.dentry->d_parent->d_name.name,
891 file->f_path.dentry->d_name.name,
Chuck Leverc216fd72007-11-12 12:16:52 -0500892 count, (long long) pos);
Badari Pulavarty027445c2006-09-30 23:28:46 -0700893
Chuck Leverce1a8e62005-11-30 18:08:17 -0500894 retval = generic_write_checks(file, &pos, &count, 0);
895 if (retval)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896 goto out;
Chuck Leverb9148c62007-11-12 12:16:58 -0500897 if (!count)
898 goto out; /* return 0 */
Chuck Leverce1a8e62005-11-30 18:08:17 -0500899
900 retval = -EINVAL;
901 if ((ssize_t) count < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903 retval = 0;
904 if (!count)
905 goto out;
Chuck Leverce1a8e62005-11-30 18:08:17 -0500906
Trond Myklebust29884df2005-12-13 16:13:54 -0500907 retval = nfs_sync_mapping(mapping);
908 if (retval)
909 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910
Chuck Leverc216fd72007-11-12 12:16:52 -0500911 retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
Chuck Lever9eafa8c2006-03-20 13:44:33 -0500912
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913 if (retval > 0)
Chuck Leverce1a8e62005-11-30 18:08:17 -0500914 iocb->ki_pos = pos + retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915
916out:
917 return retval;
918}
919
Chuck Lever88467052006-03-20 13:44:34 -0500920/**
921 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
922 *
923 */
David Howellsf7b422b2006-06-09 09:34:33 -0400924int __init nfs_init_directcache(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925{
926 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
927 sizeof(struct nfs_direct_req),
Paul Jacksonfffb60f2006-03-24 03:16:06 -0800928 0, (SLAB_RECLAIM_ACCOUNT|
929 SLAB_MEM_SPREAD),
Paul Mundt20c2df82007-07-20 10:11:58 +0900930 NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 if (nfs_direct_cachep == NULL)
932 return -ENOMEM;
933
934 return 0;
935}
936
Chuck Lever88467052006-03-20 13:44:34 -0500937/**
David Howellsf7b422b2006-06-09 09:34:33 -0400938 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
Chuck Lever88467052006-03-20 13:44:34 -0500939 *
940 */
David Brownell266bee82006-06-27 12:59:15 -0700941void nfs_destroy_directcache(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942{
Alexey Dobriyan1a1d92c2006-09-27 01:49:40 -0700943 kmem_cache_destroy(nfs_direct_cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944}