blob: a5c82b6f3b4556918aa169da26ce43c45c4a6f00 [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 */
Trond Myklebust607f31e2006-06-28 16:52:45 -0400266static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267{
Trond Myklebusta8881f52006-03-20 13:44:36 -0500268 struct nfs_open_context *ctx = dreq->ctx;
Trond Myklebust88be9f92007-06-05 10:42:27 -0400269 struct inode *inode = ctx->path.dentry->d_inode;
Chuck Lever5dd602f2006-03-20 13:44:29 -0500270 size_t rsize = NFS_SERVER(inode)->rsize;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400271 unsigned int pgbase;
272 int result;
273 ssize_t started = 0;
Chuck Lever82b145c2006-06-20 12:57:03 -0400274
Trond Myklebust607f31e2006-06-28 16:52:45 -0400275 get_dreq(dreq);
276
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277 do {
Chuck Lever82b145c2006-06-20 12:57:03 -0400278 struct nfs_read_data *data;
Chuck Lever5dd602f2006-03-20 13:44:29 -0500279 size_t bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700281 pgbase = user_addr & ~PAGE_MASK;
282 bytes = min(rsize,count);
283
Trond Myklebust607f31e2006-06-28 16:52:45 -0400284 result = -ENOMEM;
Trond Myklebust8d5658c2007-04-10 09:26:35 -0400285 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
Trond Myklebust607f31e2006-06-28 16:52:45 -0400286 if (unlikely(!data))
287 break;
288
Trond Myklebust607f31e2006-06-28 16:52:45 -0400289 down_read(&current->mm->mmap_sem);
290 result = get_user_pages(current, current->mm, user_addr,
291 data->npages, 1, 0, data->pagevec, NULL);
292 up_read(&current->mm->mmap_sem);
Chuck Lever749e1462007-05-19 17:22:46 -0400293 if (result < 0) {
294 nfs_readdata_release(data);
295 break;
296 }
297 if ((unsigned)result < data->npages) {
Trond Myklebustd9df8d62007-05-22 10:22:20 -0400298 bytes = result * PAGE_SIZE;
299 if (bytes <= pgbase) {
300 nfs_direct_release_pages(data->pagevec, result);
301 nfs_readdata_release(data);
302 break;
303 }
304 bytes -= pgbase;
305 data->npages = result;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400306 }
Chuck Lever06cf6f22006-06-20 12:56:49 -0400307
Trond Myklebust607f31e2006-06-28 16:52:45 -0400308 get_dreq(dreq);
309
310 data->req = (struct nfs_page *) dreq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311 data->inode = inode;
312 data->cred = ctx->cred;
313 data->args.fh = NFS_FH(inode);
314 data->args.context = ctx;
Chuck Lever88467052006-03-20 13:44:34 -0500315 data->args.offset = pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316 data->args.pgbase = pgbase;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400317 data->args.pages = data->pagevec;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 data->args.count = bytes;
319 data->res.fattr = &data->fattr;
320 data->res.eof = 0;
321 data->res.count = bytes;
322
Trond Myklebustec06c092006-03-20 13:44:27 -0500323 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
324 &nfs_read_direct_ops, data);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325 NFS_PROTO(inode)->read_setup(data);
326
327 data->task.tk_cookie = (unsigned long) inode;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329 rpc_execute(&data->task);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330
Chuck Levera3f565b2007-01-31 12:14:01 -0500331 dprintk("NFS: %5u initiated direct read call "
332 "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333 data->task.tk_pid,
334 inode->i_sb->s_id,
335 (long long)NFS_FILEID(inode),
336 bytes,
337 (unsigned long long)data->args.offset);
338
Trond Myklebust607f31e2006-06-28 16:52:45 -0400339 started += bytes;
340 user_addr += bytes;
Chuck Lever88467052006-03-20 13:44:34 -0500341 pos += bytes;
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700342 /* FIXME: Remove this unnecessary math from final patch */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343 pgbase += bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344 pgbase &= ~PAGE_MASK;
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700345 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346
347 count -= bytes;
348 } while (count != 0);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400349
350 if (put_dreq(dreq))
351 nfs_direct_complete(dreq);
352
353 if (started)
354 return 0;
355 return result < 0 ? (ssize_t) result : -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356}
357
Trond Myklebust607f31e2006-06-28 16:52:45 -0400358static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359{
Trond Myklebust607f31e2006-06-28 16:52:45 -0400360 ssize_t result = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 sigset_t oldset;
Chuck Lever99514f82006-03-20 13:44:30 -0500362 struct inode *inode = iocb->ki_filp->f_mapping->host;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363 struct rpc_clnt *clnt = NFS_CLIENT(inode);
364 struct nfs_direct_req *dreq;
365
Trond Myklebust607f31e2006-06-28 16:52:45 -0400366 dreq = nfs_direct_req_alloc();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 if (!dreq)
368 return -ENOMEM;
369
Chuck Lever91d5b472006-03-20 13:44:14 -0500370 dreq->inode = inode;
Trond Myklebusta8881f52006-03-20 13:44:36 -0500371 dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
Chuck Lever487b8372006-03-20 13:44:30 -0500372 if (!is_sync_kiocb(iocb))
373 dreq->iocb = iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374
Chuck Lever91d5b472006-03-20 13:44:14 -0500375 nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 rpc_clnt_sigmask(clnt, &oldset);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400377 result = nfs_direct_read_schedule(dreq, user_addr, count, pos);
378 if (!result)
379 result = nfs_direct_wait(dreq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 rpc_clnt_sigunmask(clnt, &oldset);
Trond Myklebustb4946ff2007-05-30 12:58:00 -0400381 nfs_direct_req_release(dreq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382
383 return result;
384}
385
Trond Myklebustfad61492006-03-20 13:44:36 -0500386static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387{
Trond Myklebust607f31e2006-06-28 16:52:45 -0400388 while (!list_empty(&dreq->rewrite_list)) {
389 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
Trond Myklebustfad61492006-03-20 13:44:36 -0500390 list_del(&data->pages);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400391 nfs_direct_release_pages(data->pagevec, data->npages);
Trond Myklebustfad61492006-03-20 13:44:36 -0500392 nfs_writedata_release(data);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394}
395
Trond Myklebustfad61492006-03-20 13:44:36 -0500396#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
397static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398{
Trond Myklebust607f31e2006-06-28 16:52:45 -0400399 struct inode *inode = dreq->inode;
400 struct list_head *p;
401 struct nfs_write_data *data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402
Trond Myklebustfad61492006-03-20 13:44:36 -0500403 dreq->count = 0;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400404 get_dreq(dreq);
Trond Myklebustfad61492006-03-20 13:44:36 -0500405
Trond Myklebust607f31e2006-06-28 16:52:45 -0400406 list_for_each(p, &dreq->rewrite_list) {
407 data = list_entry(p, struct nfs_write_data, pages);
408
409 get_dreq(dreq);
410
411 /*
412 * Reset data->res.
413 */
414 nfs_fattr_init(&data->fattr);
415 data->res.count = data->args.count;
416 memset(&data->verf, 0, sizeof(data->verf));
417
418 /*
419 * Reuse data->task; data->args should not have changed
420 * since the original request was sent.
421 */
422 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
423 &nfs_write_direct_ops, data);
424 NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE);
425
426 data->task.tk_priority = RPC_PRIORITY_NORMAL;
427 data->task.tk_cookie = (unsigned long) inode;
428
429 /*
430 * We're called via an RPC callback, so BKL is already held.
431 */
432 rpc_execute(&data->task);
433
434 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
435 data->task.tk_pid,
436 inode->i_sb->s_id,
437 (long long)NFS_FILEID(inode),
438 data->args.count,
439 (unsigned long long)data->args.offset);
440 }
441
442 if (put_dreq(dreq))
443 nfs_direct_write_complete(dreq, inode);
Trond Myklebustfad61492006-03-20 13:44:36 -0500444}
445
446static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
447{
448 struct nfs_write_data *data = calldata;
449 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
450
451 /* Call the NFS version-specific code */
452 if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
453 return;
454 if (unlikely(task->tk_status < 0)) {
Trond Myklebust60fa3f72007-04-14 19:11:52 -0400455 dprintk("NFS: %5u commit failed with error %d.\n",
456 task->tk_pid, task->tk_status);
Trond Myklebustfad61492006-03-20 13:44:36 -0500457 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
Trond Myklebust60fa3f72007-04-14 19:11:52 -0400458 } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
Trond Myklebustfad61492006-03-20 13:44:36 -0500459 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
460 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
461 }
462
463 dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
464 nfs_direct_write_complete(dreq, data->inode);
465}
466
467static const struct rpc_call_ops nfs_commit_direct_ops = {
468 .rpc_call_done = nfs_direct_commit_result,
469 .rpc_release = nfs_commit_release,
470};
471
472static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
473{
Trond Myklebustfad61492006-03-20 13:44:36 -0500474 struct nfs_write_data *data = dreq->commit_data;
Trond Myklebustfad61492006-03-20 13:44:36 -0500475
476 data->inode = dreq->inode;
Trond Myklebusta8881f52006-03-20 13:44:36 -0500477 data->cred = dreq->ctx->cred;
Trond Myklebustfad61492006-03-20 13:44:36 -0500478
479 data->args.fh = NFS_FH(data->inode);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400480 data->args.offset = 0;
481 data->args.count = 0;
Trond Myklebustfad61492006-03-20 13:44:36 -0500482 data->res.count = 0;
483 data->res.fattr = &data->fattr;
484 data->res.verf = &data->verf;
485
486 rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
487 &nfs_commit_direct_ops, data);
488 NFS_PROTO(data->inode)->commit_setup(data, 0);
489
490 data->task.tk_priority = RPC_PRIORITY_NORMAL;
491 data->task.tk_cookie = (unsigned long)data->inode;
492 /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
493 dreq->commit_data = NULL;
494
Trond Myklebuste99170f2006-04-18 13:21:42 -0400495 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
Trond Myklebustfad61492006-03-20 13:44:36 -0500496
Trond Myklebustfad61492006-03-20 13:44:36 -0500497 rpc_execute(&data->task);
Trond Myklebustfad61492006-03-20 13:44:36 -0500498}
499
500static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
501{
502 int flags = dreq->flags;
503
504 dreq->flags = 0;
505 switch (flags) {
506 case NFS_ODIRECT_DO_COMMIT:
507 nfs_direct_commit_schedule(dreq);
508 break;
509 case NFS_ODIRECT_RESCHED_WRITES:
510 nfs_direct_write_reschedule(dreq);
511 break;
512 default:
513 nfs_end_data_update(inode);
514 if (dreq->commit_data != NULL)
515 nfs_commit_free(dreq->commit_data);
516 nfs_direct_free_writedata(dreq);
Trond Myklebustcd9ae2b2006-10-19 23:28:40 -0700517 nfs_zap_mapping(inode, inode->i_mapping);
Trond Myklebustfad61492006-03-20 13:44:36 -0500518 nfs_direct_complete(dreq);
519 }
520}
521
522static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
523{
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700524 dreq->commit_data = nfs_commit_alloc();
Trond Myklebustfad61492006-03-20 13:44:36 -0500525 if (dreq->commit_data != NULL)
526 dreq->commit_data->req = (struct nfs_page *) dreq;
527}
528#else
529static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
530{
531 dreq->commit_data = NULL;
532}
533
534static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
535{
536 nfs_end_data_update(inode);
537 nfs_direct_free_writedata(dreq);
Trond Myklebustcd9ae2b2006-10-19 23:28:40 -0700538 nfs_zap_mapping(inode, inode->i_mapping);
Trond Myklebustfad61492006-03-20 13:44:36 -0500539 nfs_direct_complete(dreq);
540}
541#endif
542
Chuck Lever462d5b32006-03-20 13:44:32 -0500543static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
544{
545 struct nfs_write_data *data = calldata;
546 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
547 int status = task->tk_status;
548
549 if (nfs_writeback_done(task, data) != 0)
550 return;
Chuck Lever462d5b32006-03-20 13:44:32 -0500551
Chuck Lever15ce4a02006-03-20 13:44:34 -0500552 spin_lock(&dreq->lock);
Chuck Lever462d5b32006-03-20 13:44:32 -0500553
Trond Myklebust60fa3f72007-04-14 19:11:52 -0400554 if (unlikely(dreq->error != 0))
Trond Myklebusteda3cef2006-10-19 23:28:38 -0700555 goto out_unlock;
Trond Myklebust60fa3f72007-04-14 19:11:52 -0400556 if (unlikely(status < 0)) {
557 /* An error has occured, so we should not commit */
558 dreq->flags = 0;
559 dreq->error = status;
Trond Myklebusteda3cef2006-10-19 23:28:38 -0700560 }
561
562 dreq->count += data->res.count;
Chuck Lever15ce4a02006-03-20 13:44:34 -0500563
Trond Myklebustfad61492006-03-20 13:44:36 -0500564 if (data->res.verf->committed != NFS_FILE_SYNC) {
565 switch (dreq->flags) {
566 case 0:
567 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
568 dreq->flags = NFS_ODIRECT_DO_COMMIT;
569 break;
570 case NFS_ODIRECT_DO_COMMIT:
571 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
572 dprintk("NFS: %5u write verify failed\n", task->tk_pid);
573 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
574 }
575 }
576 }
Trond Myklebusteda3cef2006-10-19 23:28:38 -0700577out_unlock:
Trond Myklebustfad61492006-03-20 13:44:36 -0500578 spin_unlock(&dreq->lock);
579}
580
581/*
582 * NB: Return the value of the first error return code. Subsequent
583 * errors after the first one are ignored.
584 */
585static void nfs_direct_write_release(void *calldata)
586{
587 struct nfs_write_data *data = calldata;
588 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
589
Trond Myklebust607f31e2006-06-28 16:52:45 -0400590 if (put_dreq(dreq))
591 nfs_direct_write_complete(dreq, data->inode);
Chuck Lever462d5b32006-03-20 13:44:32 -0500592}
593
594static const struct rpc_call_ops nfs_write_direct_ops = {
595 .rpc_call_done = nfs_direct_write_result,
Trond Myklebustfad61492006-03-20 13:44:36 -0500596 .rpc_release = nfs_direct_write_release,
Chuck Lever462d5b32006-03-20 13:44:32 -0500597};
598
599/*
Trond Myklebust607f31e2006-06-28 16:52:45 -0400600 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
601 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
602 * bail and stop sending more writes. Write length accounting is
603 * handled automatically by nfs_direct_write_result(). Otherwise, if
604 * no requests have been sent, just return an error.
Chuck Lever462d5b32006-03-20 13:44:32 -0500605 */
Trond Myklebust607f31e2006-06-28 16:52:45 -0400606static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync)
Chuck Lever462d5b32006-03-20 13:44:32 -0500607{
Trond Myklebusta8881f52006-03-20 13:44:36 -0500608 struct nfs_open_context *ctx = dreq->ctx;
Trond Myklebust88be9f92007-06-05 10:42:27 -0400609 struct inode *inode = ctx->path.dentry->d_inode;
Chuck Lever462d5b32006-03-20 13:44:32 -0500610 size_t wsize = NFS_SERVER(inode)->wsize;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400611 unsigned int pgbase;
612 int result;
613 ssize_t started = 0;
Chuck Lever82b145c2006-06-20 12:57:03 -0400614
Trond Myklebust607f31e2006-06-28 16:52:45 -0400615 get_dreq(dreq);
616
Chuck Lever462d5b32006-03-20 13:44:32 -0500617 do {
Chuck Lever82b145c2006-06-20 12:57:03 -0400618 struct nfs_write_data *data;
Chuck Lever462d5b32006-03-20 13:44:32 -0500619 size_t bytes;
620
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700621 pgbase = user_addr & ~PAGE_MASK;
622 bytes = min(wsize,count);
623
Trond Myklebust607f31e2006-06-28 16:52:45 -0400624 result = -ENOMEM;
Trond Myklebust8d5658c2007-04-10 09:26:35 -0400625 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
Trond Myklebust607f31e2006-06-28 16:52:45 -0400626 if (unlikely(!data))
627 break;
628
Trond Myklebust607f31e2006-06-28 16:52:45 -0400629 down_read(&current->mm->mmap_sem);
630 result = get_user_pages(current, current->mm, user_addr,
631 data->npages, 0, 0, data->pagevec, NULL);
632 up_read(&current->mm->mmap_sem);
Chuck Lever749e1462007-05-19 17:22:46 -0400633 if (result < 0) {
634 nfs_writedata_release(data);
635 break;
636 }
637 if ((unsigned)result < data->npages) {
Trond Myklebustd9df8d62007-05-22 10:22:20 -0400638 bytes = result * PAGE_SIZE;
639 if (bytes <= pgbase) {
640 nfs_direct_release_pages(data->pagevec, result);
641 nfs_writedata_release(data);
642 break;
643 }
644 bytes -= pgbase;
645 data->npages = result;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400646 }
647
648 get_dreq(dreq);
649
Trond Myklebustfad61492006-03-20 13:44:36 -0500650 list_move_tail(&data->pages, &dreq->rewrite_list);
Chuck Lever462d5b32006-03-20 13:44:32 -0500651
Trond Myklebust607f31e2006-06-28 16:52:45 -0400652 data->req = (struct nfs_page *) dreq;
Chuck Lever462d5b32006-03-20 13:44:32 -0500653 data->inode = inode;
654 data->cred = ctx->cred;
655 data->args.fh = NFS_FH(inode);
656 data->args.context = ctx;
Chuck Lever88467052006-03-20 13:44:34 -0500657 data->args.offset = pos;
Chuck Lever462d5b32006-03-20 13:44:32 -0500658 data->args.pgbase = pgbase;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400659 data->args.pages = data->pagevec;
Chuck Lever462d5b32006-03-20 13:44:32 -0500660 data->args.count = bytes;
661 data->res.fattr = &data->fattr;
662 data->res.count = bytes;
Chuck Lever47989d72006-03-20 13:44:32 -0500663 data->res.verf = &data->verf;
Chuck Lever462d5b32006-03-20 13:44:32 -0500664
665 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
666 &nfs_write_direct_ops, data);
Trond Myklebustfad61492006-03-20 13:44:36 -0500667 NFS_PROTO(inode)->write_setup(data, sync);
Chuck Lever462d5b32006-03-20 13:44:32 -0500668
669 data->task.tk_priority = RPC_PRIORITY_NORMAL;
670 data->task.tk_cookie = (unsigned long) inode;
671
Chuck Lever462d5b32006-03-20 13:44:32 -0500672 rpc_execute(&data->task);
Chuck Lever462d5b32006-03-20 13:44:32 -0500673
Chuck Levera3f565b2007-01-31 12:14:01 -0500674 dprintk("NFS: %5u initiated direct write call "
675 "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
Chuck Lever462d5b32006-03-20 13:44:32 -0500676 data->task.tk_pid,
677 inode->i_sb->s_id,
678 (long long)NFS_FILEID(inode),
679 bytes,
680 (unsigned long long)data->args.offset);
681
Trond Myklebust607f31e2006-06-28 16:52:45 -0400682 started += bytes;
683 user_addr += bytes;
Chuck Lever88467052006-03-20 13:44:34 -0500684 pos += bytes;
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700685
686 /* FIXME: Remove this useless math from the final patch */
Chuck Lever462d5b32006-03-20 13:44:32 -0500687 pgbase += bytes;
Chuck Lever462d5b32006-03-20 13:44:32 -0500688 pgbase &= ~PAGE_MASK;
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700689 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
Chuck Lever462d5b32006-03-20 13:44:32 -0500690
691 count -= bytes;
692 } while (count != 0);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400693
694 if (put_dreq(dreq))
695 nfs_direct_write_complete(dreq, inode);
696
697 if (started)
698 return 0;
699 return result < 0 ? (ssize_t) result : -EFAULT;
Chuck Lever462d5b32006-03-20 13:44:32 -0500700}
701
Trond Myklebust607f31e2006-06-28 16:52:45 -0400702static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703{
Trond Myklebust607f31e2006-06-28 16:52:45 -0400704 ssize_t result = 0;
Chuck Lever462d5b32006-03-20 13:44:32 -0500705 sigset_t oldset;
Chuck Leverc89f2ee2006-03-20 13:44:33 -0500706 struct inode *inode = iocb->ki_filp->f_mapping->host;
Chuck Lever462d5b32006-03-20 13:44:32 -0500707 struct rpc_clnt *clnt = NFS_CLIENT(inode);
708 struct nfs_direct_req *dreq;
Trond Myklebustfad61492006-03-20 13:44:36 -0500709 size_t wsize = NFS_SERVER(inode)->wsize;
710 int sync = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711
Trond Myklebust607f31e2006-06-28 16:52:45 -0400712 dreq = nfs_direct_req_alloc();
Chuck Lever462d5b32006-03-20 13:44:32 -0500713 if (!dreq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714 return -ENOMEM;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400715 nfs_alloc_commit_data(dreq);
716
Trond Myklebustfad61492006-03-20 13:44:36 -0500717 if (dreq->commit_data == NULL || count < wsize)
718 sync = FLUSH_STABLE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719
Chuck Leverc89f2ee2006-03-20 13:44:33 -0500720 dreq->inode = inode;
Trond Myklebusta8881f52006-03-20 13:44:36 -0500721 dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
Chuck Leverc89f2ee2006-03-20 13:44:33 -0500722 if (!is_sync_kiocb(iocb))
723 dreq->iocb = iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724
Chuck Lever47989d72006-03-20 13:44:32 -0500725 nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726
727 nfs_begin_data_update(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728
Chuck Lever462d5b32006-03-20 13:44:32 -0500729 rpc_clnt_sigmask(clnt, &oldset);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400730 result = nfs_direct_write_schedule(dreq, user_addr, count, pos, sync);
731 if (!result)
732 result = nfs_direct_wait(dreq);
Chuck Lever462d5b32006-03-20 13:44:32 -0500733 rpc_clnt_sigunmask(clnt, &oldset);
Trond Myklebustb4946ff2007-05-30 12:58:00 -0400734 nfs_direct_req_release(dreq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736 return result;
737}
738
739/**
740 * nfs_file_direct_read - file direct read operation for NFS files
741 * @iocb: target I/O control block
Badari Pulavarty027445c2006-09-30 23:28:46 -0700742 * @iov: vector of user buffers into which to read data
743 * @nr_segs: size of iov vector
Chuck Lever88467052006-03-20 13:44:34 -0500744 * @pos: byte offset in file where reading starts
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 *
746 * We use this function for direct reads instead of calling
747 * generic_file_aio_read() in order to avoid gfar's check to see if
748 * the request starts before the end of the file. For that check
749 * to work, we must generate a GETATTR before each direct read, and
750 * even then there is a window between the GETATTR and the subsequent
Chuck Lever88467052006-03-20 13:44:34 -0500751 * READ where the file size could change. Our preference is simply
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 * to do all reads the application wants, and the server will take
753 * care of managing the end of file boundary.
Chuck Lever88467052006-03-20 13:44:34 -0500754 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755 * This function also eliminates unnecessarily updating the file's
756 * atime locally, as the NFS server sets the file's atime, and this
757 * client must read the updated atime from the server back into its
758 * cache.
759 */
Badari Pulavarty027445c2006-09-30 23:28:46 -0700760ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
761 unsigned long nr_segs, loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762{
763 ssize_t retval = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764 struct file *file = iocb->ki_filp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 struct address_space *mapping = file->f_mapping;
Badari Pulavarty027445c2006-09-30 23:28:46 -0700766 /* XXX: temporary */
767 const char __user *buf = iov[0].iov_base;
768 size_t count = iov[0].iov_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769
Chuck Leverce1a8e62005-11-30 18:08:17 -0500770 dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
Josef "Jeff" Sipek01cce932006-12-08 02:36:40 -0800771 file->f_path.dentry->d_parent->d_name.name,
772 file->f_path.dentry->d_name.name,
Chuck Leverce1a8e62005-11-30 18:08:17 -0500773 (unsigned long) count, (long long) pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774
Badari Pulavarty027445c2006-09-30 23:28:46 -0700775 if (nr_segs != 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 goto out;
Chuck Lever070ea602007-05-19 17:22:52 -0400777
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 retval = -EFAULT;
Chuck Lever0cdd80d2006-03-20 13:44:29 -0500779 if (!access_ok(VERIFY_WRITE, buf, count))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 goto out;
781 retval = 0;
782 if (!count)
783 goto out;
784
Trond Myklebust29884df2005-12-13 16:13:54 -0500785 retval = nfs_sync_mapping(mapping);
786 if (retval)
787 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788
Trond Myklebust607f31e2006-06-28 16:52:45 -0400789 retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790 if (retval > 0)
Chuck Lever0cdd80d2006-03-20 13:44:29 -0500791 iocb->ki_pos = pos + retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792
793out:
794 return retval;
795}
796
797/**
798 * nfs_file_direct_write - file direct write operation for NFS files
799 * @iocb: target I/O control block
Badari Pulavarty027445c2006-09-30 23:28:46 -0700800 * @iov: vector of user buffers from which to write data
801 * @nr_segs: size of iov vector
Chuck Lever88467052006-03-20 13:44:34 -0500802 * @pos: byte offset in file where writing starts
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 *
804 * We use this function for direct writes instead of calling
805 * generic_file_aio_write() in order to avoid taking the inode
806 * semaphore and updating the i_size. The NFS server will set
807 * the new i_size and this client must read the updated size
808 * back into its cache. We let the server do generic write
809 * parameter checking and report problems.
810 *
811 * We also avoid an unnecessary invocation of generic_osync_inode(),
812 * as it is fairly meaningless to sync the metadata of an NFS file.
813 *
814 * We eliminate local atime updates, see direct read above.
815 *
816 * We avoid unnecessary page cache invalidations for normal cached
817 * readers of this file.
818 *
819 * Note that O_APPEND is not supported for NFS direct writes, as there
820 * is no atomic O_APPEND write facility in the NFS protocol.
821 */
Badari Pulavarty027445c2006-09-30 23:28:46 -0700822ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
823 unsigned long nr_segs, loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824{
Chuck Lever070ea602007-05-19 17:22:52 -0400825 ssize_t retval = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 struct file *file = iocb->ki_filp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 struct address_space *mapping = file->f_mapping;
Badari Pulavarty027445c2006-09-30 23:28:46 -0700828 /* XXX: temporary */
829 const char __user *buf = iov[0].iov_base;
830 size_t count = iov[0].iov_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831
Chuck Levera3f565b2007-01-31 12:14:01 -0500832 dprintk("nfs: direct write(%s/%s, %lu@%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 Leverce1a8e62005-11-30 18:08:17 -0500835 (unsigned long) count, (long long) pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836
Badari Pulavarty027445c2006-09-30 23:28:46 -0700837 if (nr_segs != 1)
Chuck Lever070ea602007-05-19 17:22:52 -0400838 goto out;
Badari Pulavarty027445c2006-09-30 23:28:46 -0700839
Chuck Leverce1a8e62005-11-30 18:08:17 -0500840 retval = generic_write_checks(file, &pos, &count, 0);
841 if (retval)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842 goto out;
Chuck Leverce1a8e62005-11-30 18:08:17 -0500843
844 retval = -EINVAL;
845 if ((ssize_t) count < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700847 retval = 0;
848 if (!count)
849 goto out;
Chuck Leverce1a8e62005-11-30 18:08:17 -0500850
851 retval = -EFAULT;
Chuck Lever47989d72006-03-20 13:44:32 -0500852 if (!access_ok(VERIFY_READ, buf, count))
Chuck Leverce1a8e62005-11-30 18:08:17 -0500853 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854
Trond Myklebust29884df2005-12-13 16:13:54 -0500855 retval = nfs_sync_mapping(mapping);
856 if (retval)
857 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700858
Trond Myklebust607f31e2006-06-28 16:52:45 -0400859 retval = nfs_direct_write(iocb, (unsigned long) buf, count, pos);
Chuck Lever9eafa8c2006-03-20 13:44:33 -0500860
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861 if (retval > 0)
Chuck Leverce1a8e62005-11-30 18:08:17 -0500862 iocb->ki_pos = pos + retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863
864out:
865 return retval;
866}
867
Chuck Lever88467052006-03-20 13:44:34 -0500868/**
869 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
870 *
871 */
David Howellsf7b422b2006-06-09 09:34:33 -0400872int __init nfs_init_directcache(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873{
874 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
875 sizeof(struct nfs_direct_req),
Paul Jacksonfffb60f2006-03-24 03:16:06 -0800876 0, (SLAB_RECLAIM_ACCOUNT|
877 SLAB_MEM_SPREAD),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878 NULL, NULL);
879 if (nfs_direct_cachep == NULL)
880 return -ENOMEM;
881
882 return 0;
883}
884
Chuck Lever88467052006-03-20 13:44:34 -0500885/**
David Howellsf7b422b2006-06-09 09:34:33 -0400886 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
Chuck Lever88467052006-03-20 13:44:34 -0500887 *
888 */
David Brownell266bee82006-06-27 12:59:15 -0700889void nfs_destroy_directcache(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890{
Alexey Dobriyan1a1d92c2006-09-27 01:49:40 -0700891 kmem_cache_destroy(nfs_direct_cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892}