blob: e4e089a8f294343875f730b655a18f156b4839e3 [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
Matthew Wilcox150030b2007-12-06 16:24:39 -0500196 result = wait_for_completion_killable(&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
Trond Myklebustfdd1e742008-04-15 16:33:58 -0400233 nfs_readpage_result(task, data);
234}
235
236static void nfs_direct_read_release(void *calldata)
237{
238
239 struct nfs_read_data *data = calldata;
240 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
241 int status = data->task.tk_status;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242
Chuck Lever15ce4a02006-03-20 13:44:34 -0500243 spin_lock(&dreq->lock);
Trond Myklebustfdd1e742008-04-15 16:33:58 -0400244 if (unlikely(status < 0)) {
245 dreq->error = status;
Trond Myklebustd4a8f362007-05-22 10:22:27 -0400246 spin_unlock(&dreq->lock);
247 } else {
248 dreq->count += data->res.count;
249 spin_unlock(&dreq->lock);
250 nfs_direct_dirty_pages(data->pagevec,
251 data->args.pgbase,
252 data->res.count);
253 }
254 nfs_direct_release_pages(data->pagevec, data->npages);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400255
256 if (put_dreq(dreq))
257 nfs_direct_complete(dreq);
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400258 nfs_readdata_free(data);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259}
260
Trond Myklebustec06c092006-03-20 13:44:27 -0500261static const struct rpc_call_ops nfs_read_direct_ops = {
Andy Adamsonf11c88a2009-04-01 09:22:25 -0400262#if defined(CONFIG_NFS_V4_1)
263 .rpc_call_prepare = nfs_read_prepare,
264#endif /* CONFIG_NFS_V4_1 */
Trond Myklebustec06c092006-03-20 13:44:27 -0500265 .rpc_call_done = nfs_direct_read_result,
Trond Myklebustfdd1e742008-04-15 16:33:58 -0400266 .rpc_release = nfs_direct_read_release,
Trond Myklebustec06c092006-03-20 13:44:27 -0500267};
268
Chuck Leverd4cc9482006-03-20 13:44:28 -0500269/*
Trond Myklebust607f31e2006-06-28 16:52:45 -0400270 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
271 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
272 * bail and stop sending more reads. Read length accounting is
273 * handled automatically by nfs_direct_read_result(). Otherwise, if
274 * no requests have been sent, just return an error.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275 */
Chuck Lever02fe4942007-11-12 12:17:03 -0500276static ssize_t nfs_direct_read_schedule_segment(struct nfs_direct_req *dreq,
277 const struct iovec *iov,
278 loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279{
Trond Myklebusta8881f52006-03-20 13:44:36 -0500280 struct nfs_open_context *ctx = dreq->ctx;
Trond Myklebust88be9f92007-06-05 10:42:27 -0400281 struct inode *inode = ctx->path.dentry->d_inode;
Chuck Lever02fe4942007-11-12 12:17:03 -0500282 unsigned long user_addr = (unsigned long)iov->iov_base;
283 size_t count = iov->iov_len;
Chuck Lever5dd602f2006-03-20 13:44:29 -0500284 size_t rsize = NFS_SERVER(inode)->rsize;
Trond Myklebust07737692007-10-25 18:42:54 -0400285 struct rpc_task *task;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400286 struct rpc_message msg = {
287 .rpc_cred = ctx->cred,
288 };
Trond Myklebust84115e12007-07-14 15:39:59 -0400289 struct rpc_task_setup task_setup_data = {
290 .rpc_client = NFS_CLIENT(inode),
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400291 .rpc_message = &msg,
Trond Myklebust84115e12007-07-14 15:39:59 -0400292 .callback_ops = &nfs_read_direct_ops,
Trond Myklebust101070c2008-02-19 20:04:23 -0500293 .workqueue = nfsiod_workqueue,
Trond Myklebust84115e12007-07-14 15:39:59 -0400294 .flags = RPC_TASK_ASYNC,
295 };
Trond Myklebust607f31e2006-06-28 16:52:45 -0400296 unsigned int pgbase;
297 int result;
298 ssize_t started = 0;
Chuck Lever82b145c2006-06-20 12:57:03 -0400299
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300 do {
Chuck Lever82b145c2006-06-20 12:57:03 -0400301 struct nfs_read_data *data;
Chuck Lever5dd602f2006-03-20 13:44:29 -0500302 size_t bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700304 pgbase = user_addr & ~PAGE_MASK;
305 bytes = min(rsize,count);
306
Trond Myklebust607f31e2006-06-28 16:52:45 -0400307 result = -ENOMEM;
Trond Myklebust8d5658c2007-04-10 09:26:35 -0400308 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
Trond Myklebust607f31e2006-06-28 16:52:45 -0400309 if (unlikely(!data))
310 break;
311
Trond Myklebust607f31e2006-06-28 16:52:45 -0400312 down_read(&current->mm->mmap_sem);
313 result = get_user_pages(current, current->mm, user_addr,
314 data->npages, 1, 0, data->pagevec, NULL);
315 up_read(&current->mm->mmap_sem);
Chuck Lever749e1462007-05-19 17:22:46 -0400316 if (result < 0) {
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400317 nfs_readdata_free(data);
Chuck Lever749e1462007-05-19 17:22:46 -0400318 break;
319 }
320 if ((unsigned)result < data->npages) {
Trond Myklebustd9df8d62007-05-22 10:22:20 -0400321 bytes = result * PAGE_SIZE;
322 if (bytes <= pgbase) {
323 nfs_direct_release_pages(data->pagevec, result);
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400324 nfs_readdata_free(data);
Trond Myklebustd9df8d62007-05-22 10:22:20 -0400325 break;
326 }
327 bytes -= pgbase;
328 data->npages = result;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400329 }
Chuck Lever06cf6f22006-06-20 12:56:49 -0400330
Trond Myklebust607f31e2006-06-28 16:52:45 -0400331 get_dreq(dreq);
332
333 data->req = (struct nfs_page *) dreq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334 data->inode = inode;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400335 data->cred = msg.rpc_cred;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336 data->args.fh = NFS_FH(inode);
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400337 data->args.context = ctx;
Chuck Lever88467052006-03-20 13:44:34 -0500338 data->args.offset = pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339 data->args.pgbase = pgbase;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400340 data->args.pages = data->pagevec;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341 data->args.count = bytes;
342 data->res.fattr = &data->fattr;
343 data->res.eof = 0;
344 data->res.count = bytes;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400345 msg.rpc_argp = &data->args;
346 msg.rpc_resp = &data->res;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347
Trond Myklebust07737692007-10-25 18:42:54 -0400348 task_setup_data.task = &data->task;
Trond Myklebust84115e12007-07-14 15:39:59 -0400349 task_setup_data.callback_data = data;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400350 NFS_PROTO(inode)->read_setup(data, &msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700351
Trond Myklebust07737692007-10-25 18:42:54 -0400352 task = rpc_run_task(&task_setup_data);
Trond Myklebustdbae4c72008-04-14 14:54:53 -0400353 if (IS_ERR(task))
354 break;
355 rpc_put_task(task);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356
Chuck Levera3f565b2007-01-31 12:14:01 -0500357 dprintk("NFS: %5u initiated direct read call "
358 "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 data->task.tk_pid,
360 inode->i_sb->s_id,
361 (long long)NFS_FILEID(inode),
362 bytes,
363 (unsigned long long)data->args.offset);
364
Trond Myklebust607f31e2006-06-28 16:52:45 -0400365 started += bytes;
366 user_addr += bytes;
Chuck Lever88467052006-03-20 13:44:34 -0500367 pos += bytes;
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700368 /* FIXME: Remove this unnecessary math from final patch */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 pgbase += bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370 pgbase &= ~PAGE_MASK;
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700371 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372
373 count -= bytes;
374 } while (count != 0);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400375
Trond Myklebust607f31e2006-06-28 16:52:45 -0400376 if (started)
Chuck Leverc216fd72007-11-12 12:16:52 -0500377 return started;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400378 return result < 0 ? (ssize_t) result : -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379}
380
Chuck Lever19f73782007-11-12 12:16:47 -0500381static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
382 const struct iovec *iov,
383 unsigned long nr_segs,
384 loff_t pos)
385{
386 ssize_t result = -EINVAL;
387 size_t requested_bytes = 0;
388 unsigned long seg;
389
390 get_dreq(dreq);
391
392 for (seg = 0; seg < nr_segs; seg++) {
393 const struct iovec *vec = &iov[seg];
Chuck Lever02fe4942007-11-12 12:17:03 -0500394 result = nfs_direct_read_schedule_segment(dreq, vec, pos);
Chuck Lever19f73782007-11-12 12:16:47 -0500395 if (result < 0)
396 break;
397 requested_bytes += result;
398 if ((size_t)result < vec->iov_len)
399 break;
400 pos += vec->iov_len;
401 }
402
403 if (put_dreq(dreq))
404 nfs_direct_complete(dreq);
405
406 if (requested_bytes != 0)
407 return 0;
408
409 if (result < 0)
410 return result;
411 return -EIO;
412}
413
Chuck Leverc216fd72007-11-12 12:16:52 -0500414static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
415 unsigned long nr_segs, loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416{
Trond Myklebust607f31e2006-06-28 16:52:45 -0400417 ssize_t result = 0;
Chuck Lever99514f82006-03-20 13:44:30 -0500418 struct inode *inode = iocb->ki_filp->f_mapping->host;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 struct nfs_direct_req *dreq;
420
Trond Myklebust607f31e2006-06-28 16:52:45 -0400421 dreq = nfs_direct_req_alloc();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 if (!dreq)
423 return -ENOMEM;
424
Chuck Lever91d5b472006-03-20 13:44:14 -0500425 dreq->inode = inode;
Trond Myklebustcd3758e2007-08-10 17:44:32 -0400426 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
Chuck Lever487b8372006-03-20 13:44:30 -0500427 if (!is_sync_kiocb(iocb))
428 dreq->iocb = iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429
Chuck Leverc216fd72007-11-12 12:16:52 -0500430 result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400431 if (!result)
432 result = nfs_direct_wait(dreq);
Trond Myklebustb4946ff2007-05-30 12:58:00 -0400433 nfs_direct_req_release(dreq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434
435 return result;
436}
437
Trond Myklebustfad61492006-03-20 13:44:36 -0500438static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439{
Trond Myklebust607f31e2006-06-28 16:52:45 -0400440 while (!list_empty(&dreq->rewrite_list)) {
441 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
Trond Myklebustfad61492006-03-20 13:44:36 -0500442 list_del(&data->pages);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400443 nfs_direct_release_pages(data->pagevec, data->npages);
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400444 nfs_writedata_free(data);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446}
447
Trond Myklebustfad61492006-03-20 13:44:36 -0500448#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
449static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450{
Trond Myklebust607f31e2006-06-28 16:52:45 -0400451 struct inode *inode = dreq->inode;
452 struct list_head *p;
453 struct nfs_write_data *data;
Trond Myklebust07737692007-10-25 18:42:54 -0400454 struct rpc_task *task;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400455 struct rpc_message msg = {
456 .rpc_cred = dreq->ctx->cred,
457 };
Trond Myklebust84115e12007-07-14 15:39:59 -0400458 struct rpc_task_setup task_setup_data = {
459 .rpc_client = NFS_CLIENT(inode),
460 .callback_ops = &nfs_write_direct_ops,
Trond Myklebust101070c2008-02-19 20:04:23 -0500461 .workqueue = nfsiod_workqueue,
Trond Myklebust84115e12007-07-14 15:39:59 -0400462 .flags = RPC_TASK_ASYNC,
463 };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464
Trond Myklebustfad61492006-03-20 13:44:36 -0500465 dreq->count = 0;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400466 get_dreq(dreq);
Trond Myklebustfad61492006-03-20 13:44:36 -0500467
Trond Myklebust607f31e2006-06-28 16:52:45 -0400468 list_for_each(p, &dreq->rewrite_list) {
469 data = list_entry(p, struct nfs_write_data, pages);
470
471 get_dreq(dreq);
472
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400473 /* Use stable writes */
474 data->args.stable = NFS_FILE_SYNC;
475
Trond Myklebust607f31e2006-06-28 16:52:45 -0400476 /*
477 * Reset data->res.
478 */
479 nfs_fattr_init(&data->fattr);
480 data->res.count = data->args.count;
481 memset(&data->verf, 0, sizeof(data->verf));
482
483 /*
484 * Reuse data->task; data->args should not have changed
485 * since the original request was sent.
486 */
Trond Myklebust07737692007-10-25 18:42:54 -0400487 task_setup_data.task = &data->task;
Trond Myklebust84115e12007-07-14 15:39:59 -0400488 task_setup_data.callback_data = data;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400489 msg.rpc_argp = &data->args;
490 msg.rpc_resp = &data->res;
491 NFS_PROTO(inode)->write_setup(data, &msg);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400492
Trond Myklebust607f31e2006-06-28 16:52:45 -0400493 /*
494 * We're called via an RPC callback, so BKL is already held.
495 */
Trond Myklebust07737692007-10-25 18:42:54 -0400496 task = rpc_run_task(&task_setup_data);
497 if (!IS_ERR(task))
498 rpc_put_task(task);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400499
500 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
501 data->task.tk_pid,
502 inode->i_sb->s_id,
503 (long long)NFS_FILEID(inode),
504 data->args.count,
505 (unsigned long long)data->args.offset);
506 }
507
508 if (put_dreq(dreq))
509 nfs_direct_write_complete(dreq, inode);
Trond Myklebustfad61492006-03-20 13:44:36 -0500510}
511
512static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
513{
514 struct nfs_write_data *data = calldata;
Trond Myklebustfad61492006-03-20 13:44:36 -0500515
516 /* Call the NFS version-specific code */
Trond Myklebustc9d8f892008-04-15 16:56:39 -0400517 NFS_PROTO(data->inode)->commit_done(task, data);
518}
519
520static void nfs_direct_commit_release(void *calldata)
521{
522 struct nfs_write_data *data = calldata;
523 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
524 int status = data->task.tk_status;
525
526 if (status < 0) {
Trond Myklebust60fa3f72007-04-14 19:11:52 -0400527 dprintk("NFS: %5u commit failed with error %d.\n",
Trond Myklebustc9d8f892008-04-15 16:56:39 -0400528 data->task.tk_pid, status);
Trond Myklebustfad61492006-03-20 13:44:36 -0500529 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
Trond Myklebust60fa3f72007-04-14 19:11:52 -0400530 } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
Trond Myklebustc9d8f892008-04-15 16:56:39 -0400531 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
Trond Myklebustfad61492006-03-20 13:44:36 -0500532 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
533 }
534
Trond Myklebustc9d8f892008-04-15 16:56:39 -0400535 dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
Trond Myklebustfad61492006-03-20 13:44:36 -0500536 nfs_direct_write_complete(dreq, data->inode);
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400537 nfs_commit_free(data);
Trond Myklebustfad61492006-03-20 13:44:36 -0500538}
539
540static const struct rpc_call_ops nfs_commit_direct_ops = {
Andy Adamson21d9a852009-04-01 09:22:27 -0400541#if defined(CONFIG_NFS_V4_1)
542 .rpc_call_prepare = nfs_write_prepare,
543#endif /* CONFIG_NFS_V4_1 */
Trond Myklebustfad61492006-03-20 13:44:36 -0500544 .rpc_call_done = nfs_direct_commit_result,
Trond Myklebustc9d8f892008-04-15 16:56:39 -0400545 .rpc_release = nfs_direct_commit_release,
Trond Myklebustfad61492006-03-20 13:44:36 -0500546};
547
548static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
549{
Trond Myklebustfad61492006-03-20 13:44:36 -0500550 struct nfs_write_data *data = dreq->commit_data;
Trond Myklebust07737692007-10-25 18:42:54 -0400551 struct rpc_task *task;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400552 struct rpc_message msg = {
553 .rpc_argp = &data->args,
554 .rpc_resp = &data->res,
555 .rpc_cred = dreq->ctx->cred,
556 };
Trond Myklebust84115e12007-07-14 15:39:59 -0400557 struct rpc_task_setup task_setup_data = {
Trond Myklebust07737692007-10-25 18:42:54 -0400558 .task = &data->task,
Trond Myklebust84115e12007-07-14 15:39:59 -0400559 .rpc_client = NFS_CLIENT(dreq->inode),
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400560 .rpc_message = &msg,
Trond Myklebust84115e12007-07-14 15:39:59 -0400561 .callback_ops = &nfs_commit_direct_ops,
562 .callback_data = data,
Trond Myklebust101070c2008-02-19 20:04:23 -0500563 .workqueue = nfsiod_workqueue,
Trond Myklebust84115e12007-07-14 15:39:59 -0400564 .flags = RPC_TASK_ASYNC,
565 };
Trond Myklebustfad61492006-03-20 13:44:36 -0500566
567 data->inode = dreq->inode;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400568 data->cred = msg.rpc_cred;
Trond Myklebustfad61492006-03-20 13:44:36 -0500569
570 data->args.fh = NFS_FH(data->inode);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400571 data->args.offset = 0;
572 data->args.count = 0;
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400573 data->args.context = dreq->ctx;
Trond Myklebustfad61492006-03-20 13:44:36 -0500574 data->res.count = 0;
575 data->res.fattr = &data->fattr;
576 data->res.verf = &data->verf;
577
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400578 NFS_PROTO(data->inode)->commit_setup(data, &msg);
Trond Myklebustfad61492006-03-20 13:44:36 -0500579
Trond Myklebustfad61492006-03-20 13:44:36 -0500580 /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
581 dreq->commit_data = NULL;
582
Trond Myklebuste99170f2006-04-18 13:21:42 -0400583 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
Trond Myklebustfad61492006-03-20 13:44:36 -0500584
Trond Myklebust07737692007-10-25 18:42:54 -0400585 task = rpc_run_task(&task_setup_data);
586 if (!IS_ERR(task))
587 rpc_put_task(task);
Trond Myklebustfad61492006-03-20 13:44:36 -0500588}
589
590static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
591{
592 int flags = dreq->flags;
593
594 dreq->flags = 0;
595 switch (flags) {
596 case NFS_ODIRECT_DO_COMMIT:
597 nfs_direct_commit_schedule(dreq);
598 break;
599 case NFS_ODIRECT_RESCHED_WRITES:
600 nfs_direct_write_reschedule(dreq);
601 break;
602 default:
Trond Myklebustfad61492006-03-20 13:44:36 -0500603 if (dreq->commit_data != NULL)
604 nfs_commit_free(dreq->commit_data);
605 nfs_direct_free_writedata(dreq);
Trond Myklebustcd9ae2b2006-10-19 23:28:40 -0700606 nfs_zap_mapping(inode, inode->i_mapping);
Trond Myklebustfad61492006-03-20 13:44:36 -0500607 nfs_direct_complete(dreq);
608 }
609}
610
611static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
612{
Trond Myklebustc9d8f892008-04-15 16:56:39 -0400613 dreq->commit_data = nfs_commitdata_alloc();
Trond Myklebustfad61492006-03-20 13:44:36 -0500614 if (dreq->commit_data != NULL)
615 dreq->commit_data->req = (struct nfs_page *) dreq;
616}
617#else
618static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
619{
620 dreq->commit_data = NULL;
621}
622
623static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
624{
Trond Myklebustfad61492006-03-20 13:44:36 -0500625 nfs_direct_free_writedata(dreq);
Trond Myklebustcd9ae2b2006-10-19 23:28:40 -0700626 nfs_zap_mapping(inode, inode->i_mapping);
Trond Myklebustfad61492006-03-20 13:44:36 -0500627 nfs_direct_complete(dreq);
628}
629#endif
630
Chuck Lever462d5b32006-03-20 13:44:32 -0500631static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
632{
633 struct nfs_write_data *data = calldata;
Chuck Lever462d5b32006-03-20 13:44:32 -0500634
635 if (nfs_writeback_done(task, data) != 0)
636 return;
Trond Myklebustc9d8f892008-04-15 16:56:39 -0400637}
638
639/*
640 * NB: Return the value of the first error return code. Subsequent
641 * errors after the first one are ignored.
642 */
643static void nfs_direct_write_release(void *calldata)
644{
645 struct nfs_write_data *data = calldata;
646 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
647 int status = data->task.tk_status;
Chuck Lever462d5b32006-03-20 13:44:32 -0500648
Chuck Lever15ce4a02006-03-20 13:44:34 -0500649 spin_lock(&dreq->lock);
Chuck Lever462d5b32006-03-20 13:44:32 -0500650
Trond Myklebust60fa3f72007-04-14 19:11:52 -0400651 if (unlikely(status < 0)) {
Neil Brown432409e2007-10-23 17:09:13 -0400652 /* An error has occurred, so we should not commit */
Trond Myklebust60fa3f72007-04-14 19:11:52 -0400653 dreq->flags = 0;
654 dreq->error = status;
Trond Myklebusteda3cef2006-10-19 23:28:38 -0700655 }
Neil Brown432409e2007-10-23 17:09:13 -0400656 if (unlikely(dreq->error != 0))
657 goto out_unlock;
Trond Myklebusteda3cef2006-10-19 23:28:38 -0700658
659 dreq->count += data->res.count;
Chuck Lever15ce4a02006-03-20 13:44:34 -0500660
Trond Myklebustfad61492006-03-20 13:44:36 -0500661 if (data->res.verf->committed != NFS_FILE_SYNC) {
662 switch (dreq->flags) {
663 case 0:
664 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
665 dreq->flags = NFS_ODIRECT_DO_COMMIT;
666 break;
667 case NFS_ODIRECT_DO_COMMIT:
668 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
Trond Myklebustc9d8f892008-04-15 16:56:39 -0400669 dprintk("NFS: %5u write verify failed\n", data->task.tk_pid);
Trond Myklebustfad61492006-03-20 13:44:36 -0500670 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
671 }
672 }
673 }
Trond Myklebusteda3cef2006-10-19 23:28:38 -0700674out_unlock:
Trond Myklebustfad61492006-03-20 13:44:36 -0500675 spin_unlock(&dreq->lock);
Trond Myklebustfad61492006-03-20 13:44:36 -0500676
Trond Myklebust607f31e2006-06-28 16:52:45 -0400677 if (put_dreq(dreq))
678 nfs_direct_write_complete(dreq, data->inode);
Chuck Lever462d5b32006-03-20 13:44:32 -0500679}
680
681static const struct rpc_call_ops nfs_write_direct_ops = {
Andy Adamsondef6ed72009-04-01 09:22:26 -0400682#if defined(CONFIG_NFS_V4_1)
683 .rpc_call_prepare = nfs_write_prepare,
684#endif /* CONFIG_NFS_V4_1 */
Chuck Lever462d5b32006-03-20 13:44:32 -0500685 .rpc_call_done = nfs_direct_write_result,
Trond Myklebustfad61492006-03-20 13:44:36 -0500686 .rpc_release = nfs_direct_write_release,
Chuck Lever462d5b32006-03-20 13:44:32 -0500687};
688
689/*
Trond Myklebust607f31e2006-06-28 16:52:45 -0400690 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
691 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
692 * bail and stop sending more writes. Write length accounting is
693 * handled automatically by nfs_direct_write_result(). Otherwise, if
694 * no requests have been sent, just return an error.
Chuck Lever462d5b32006-03-20 13:44:32 -0500695 */
Chuck Lever02fe4942007-11-12 12:17:03 -0500696static ssize_t nfs_direct_write_schedule_segment(struct nfs_direct_req *dreq,
697 const struct iovec *iov,
698 loff_t pos, int sync)
Chuck Lever462d5b32006-03-20 13:44:32 -0500699{
Trond Myklebusta8881f52006-03-20 13:44:36 -0500700 struct nfs_open_context *ctx = dreq->ctx;
Trond Myklebust88be9f92007-06-05 10:42:27 -0400701 struct inode *inode = ctx->path.dentry->d_inode;
Chuck Lever02fe4942007-11-12 12:17:03 -0500702 unsigned long user_addr = (unsigned long)iov->iov_base;
703 size_t count = iov->iov_len;
Trond Myklebust07737692007-10-25 18:42:54 -0400704 struct rpc_task *task;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400705 struct rpc_message msg = {
706 .rpc_cred = ctx->cred,
707 };
Trond Myklebust84115e12007-07-14 15:39:59 -0400708 struct rpc_task_setup task_setup_data = {
709 .rpc_client = NFS_CLIENT(inode),
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400710 .rpc_message = &msg,
Trond Myklebust84115e12007-07-14 15:39:59 -0400711 .callback_ops = &nfs_write_direct_ops,
Trond Myklebust101070c2008-02-19 20:04:23 -0500712 .workqueue = nfsiod_workqueue,
Trond Myklebust84115e12007-07-14 15:39:59 -0400713 .flags = RPC_TASK_ASYNC,
714 };
Chuck Lever462d5b32006-03-20 13:44:32 -0500715 size_t wsize = NFS_SERVER(inode)->wsize;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400716 unsigned int pgbase;
717 int result;
718 ssize_t started = 0;
Chuck Lever82b145c2006-06-20 12:57:03 -0400719
Chuck Lever462d5b32006-03-20 13:44:32 -0500720 do {
Chuck Lever82b145c2006-06-20 12:57:03 -0400721 struct nfs_write_data *data;
Chuck Lever462d5b32006-03-20 13:44:32 -0500722 size_t bytes;
723
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700724 pgbase = user_addr & ~PAGE_MASK;
725 bytes = min(wsize,count);
726
Trond Myklebust607f31e2006-06-28 16:52:45 -0400727 result = -ENOMEM;
Trond Myklebust8d5658c2007-04-10 09:26:35 -0400728 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
Trond Myklebust607f31e2006-06-28 16:52:45 -0400729 if (unlikely(!data))
730 break;
731
Trond Myklebust607f31e2006-06-28 16:52:45 -0400732 down_read(&current->mm->mmap_sem);
733 result = get_user_pages(current, current->mm, user_addr,
734 data->npages, 0, 0, data->pagevec, NULL);
735 up_read(&current->mm->mmap_sem);
Chuck Lever749e1462007-05-19 17:22:46 -0400736 if (result < 0) {
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400737 nfs_writedata_free(data);
Chuck Lever749e1462007-05-19 17:22:46 -0400738 break;
739 }
740 if ((unsigned)result < data->npages) {
Trond Myklebustd9df8d62007-05-22 10:22:20 -0400741 bytes = result * PAGE_SIZE;
742 if (bytes <= pgbase) {
743 nfs_direct_release_pages(data->pagevec, result);
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400744 nfs_writedata_free(data);
Trond Myklebustd9df8d62007-05-22 10:22:20 -0400745 break;
746 }
747 bytes -= pgbase;
748 data->npages = result;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400749 }
750
751 get_dreq(dreq);
752
Trond Myklebustfad61492006-03-20 13:44:36 -0500753 list_move_tail(&data->pages, &dreq->rewrite_list);
Chuck Lever462d5b32006-03-20 13:44:32 -0500754
Trond Myklebust607f31e2006-06-28 16:52:45 -0400755 data->req = (struct nfs_page *) dreq;
Chuck Lever462d5b32006-03-20 13:44:32 -0500756 data->inode = inode;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400757 data->cred = msg.rpc_cred;
Chuck Lever462d5b32006-03-20 13:44:32 -0500758 data->args.fh = NFS_FH(inode);
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400759 data->args.context = ctx;
Chuck Lever88467052006-03-20 13:44:34 -0500760 data->args.offset = pos;
Chuck Lever462d5b32006-03-20 13:44:32 -0500761 data->args.pgbase = pgbase;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400762 data->args.pages = data->pagevec;
Chuck Lever462d5b32006-03-20 13:44:32 -0500763 data->args.count = bytes;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400764 data->args.stable = sync;
Chuck Lever462d5b32006-03-20 13:44:32 -0500765 data->res.fattr = &data->fattr;
766 data->res.count = bytes;
Chuck Lever47989d72006-03-20 13:44:32 -0500767 data->res.verf = &data->verf;
Chuck Lever462d5b32006-03-20 13:44:32 -0500768
Trond Myklebust07737692007-10-25 18:42:54 -0400769 task_setup_data.task = &data->task;
Trond Myklebust84115e12007-07-14 15:39:59 -0400770 task_setup_data.callback_data = data;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400771 msg.rpc_argp = &data->args;
772 msg.rpc_resp = &data->res;
773 NFS_PROTO(inode)->write_setup(data, &msg);
Chuck Lever462d5b32006-03-20 13:44:32 -0500774
Trond Myklebust07737692007-10-25 18:42:54 -0400775 task = rpc_run_task(&task_setup_data);
Trond Myklebustdbae4c72008-04-14 14:54:53 -0400776 if (IS_ERR(task))
777 break;
778 rpc_put_task(task);
Chuck Lever462d5b32006-03-20 13:44:32 -0500779
Chuck Levera3f565b2007-01-31 12:14:01 -0500780 dprintk("NFS: %5u initiated direct write call "
781 "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
Chuck Lever462d5b32006-03-20 13:44:32 -0500782 data->task.tk_pid,
783 inode->i_sb->s_id,
784 (long long)NFS_FILEID(inode),
785 bytes,
786 (unsigned long long)data->args.offset);
787
Trond Myklebust607f31e2006-06-28 16:52:45 -0400788 started += bytes;
789 user_addr += bytes;
Chuck Lever88467052006-03-20 13:44:34 -0500790 pos += bytes;
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700791
792 /* FIXME: Remove this useless math from the final patch */
Chuck Lever462d5b32006-03-20 13:44:32 -0500793 pgbase += bytes;
Chuck Lever462d5b32006-03-20 13:44:32 -0500794 pgbase &= ~PAGE_MASK;
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700795 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
Chuck Lever462d5b32006-03-20 13:44:32 -0500796
797 count -= bytes;
798 } while (count != 0);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400799
Trond Myklebust607f31e2006-06-28 16:52:45 -0400800 if (started)
Chuck Leverc216fd72007-11-12 12:16:52 -0500801 return started;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400802 return result < 0 ? (ssize_t) result : -EFAULT;
Chuck Lever462d5b32006-03-20 13:44:32 -0500803}
804
Chuck Lever19f73782007-11-12 12:16:47 -0500805static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
806 const struct iovec *iov,
807 unsigned long nr_segs,
808 loff_t pos, int sync)
809{
810 ssize_t result = 0;
811 size_t requested_bytes = 0;
812 unsigned long seg;
813
814 get_dreq(dreq);
815
816 for (seg = 0; seg < nr_segs; seg++) {
817 const struct iovec *vec = &iov[seg];
Chuck Lever02fe4942007-11-12 12:17:03 -0500818 result = nfs_direct_write_schedule_segment(dreq, vec,
819 pos, sync);
Chuck Lever19f73782007-11-12 12:16:47 -0500820 if (result < 0)
821 break;
822 requested_bytes += result;
823 if ((size_t)result < vec->iov_len)
824 break;
825 pos += vec->iov_len;
826 }
827
828 if (put_dreq(dreq))
829 nfs_direct_write_complete(dreq, dreq->inode);
830
831 if (requested_bytes != 0)
832 return 0;
833
834 if (result < 0)
835 return result;
836 return -EIO;
837}
838
Chuck Leverc216fd72007-11-12 12:16:52 -0500839static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
840 unsigned long nr_segs, loff_t pos,
841 size_t count)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842{
Trond Myklebust607f31e2006-06-28 16:52:45 -0400843 ssize_t result = 0;
Chuck Leverc89f2ee2006-03-20 13:44:33 -0500844 struct inode *inode = iocb->ki_filp->f_mapping->host;
Chuck Lever462d5b32006-03-20 13:44:32 -0500845 struct nfs_direct_req *dreq;
Trond Myklebustfad61492006-03-20 13:44:36 -0500846 size_t wsize = NFS_SERVER(inode)->wsize;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400847 int sync = NFS_UNSTABLE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848
Trond Myklebust607f31e2006-06-28 16:52:45 -0400849 dreq = nfs_direct_req_alloc();
Chuck Lever462d5b32006-03-20 13:44:32 -0500850 if (!dreq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 return -ENOMEM;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400852 nfs_alloc_commit_data(dreq);
853
Trond Myklebustfad61492006-03-20 13:44:36 -0500854 if (dreq->commit_data == NULL || count < wsize)
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400855 sync = NFS_FILE_SYNC;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856
Chuck Leverc89f2ee2006-03-20 13:44:33 -0500857 dreq->inode = inode;
Trond Myklebustcd3758e2007-08-10 17:44:32 -0400858 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
Chuck Leverc89f2ee2006-03-20 13:44:33 -0500859 if (!is_sync_kiocb(iocb))
860 dreq->iocb = iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861
Chuck Leverc216fd72007-11-12 12:16:52 -0500862 result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400863 if (!result)
864 result = nfs_direct_wait(dreq);
Trond Myklebustb4946ff2007-05-30 12:58:00 -0400865 nfs_direct_req_release(dreq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867 return result;
868}
869
870/**
871 * nfs_file_direct_read - file direct read operation for NFS files
872 * @iocb: target I/O control block
Badari Pulavarty027445c2006-09-30 23:28:46 -0700873 * @iov: vector of user buffers into which to read data
874 * @nr_segs: size of iov vector
Chuck Lever88467052006-03-20 13:44:34 -0500875 * @pos: byte offset in file where reading starts
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 *
877 * We use this function for direct reads instead of calling
878 * generic_file_aio_read() in order to avoid gfar's check to see if
879 * the request starts before the end of the file. For that check
880 * to work, we must generate a GETATTR before each direct read, and
881 * even then there is a window between the GETATTR and the subsequent
Chuck Lever88467052006-03-20 13:44:34 -0500882 * READ where the file size could change. Our preference is simply
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883 * to do all reads the application wants, and the server will take
884 * care of managing the end of file boundary.
Chuck Lever88467052006-03-20 13:44:34 -0500885 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 * This function also eliminates unnecessarily updating the file's
887 * atime locally, as the NFS server sets the file's atime, and this
888 * client must read the updated atime from the server back into its
889 * cache.
890 */
Badari Pulavarty027445c2006-09-30 23:28:46 -0700891ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
892 unsigned long nr_segs, loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893{
894 ssize_t retval = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895 struct file *file = iocb->ki_filp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896 struct address_space *mapping = file->f_mapping;
Chuck Leverc216fd72007-11-12 12:16:52 -0500897 size_t count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898
Chuck Leverc216fd72007-11-12 12:16:52 -0500899 count = iov_length(iov, nr_segs);
900 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
901
Chuck Lever6da24bc2008-06-11 17:55:58 -0400902 dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n",
Josef "Jeff" Sipek01cce932006-12-08 02:36:40 -0800903 file->f_path.dentry->d_parent->d_name.name,
904 file->f_path.dentry->d_name.name,
Chuck Leverc216fd72007-11-12 12:16:52 -0500905 count, (long long) pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907 retval = 0;
908 if (!count)
909 goto out;
910
Trond Myklebust29884df2005-12-13 16:13:54 -0500911 retval = nfs_sync_mapping(mapping);
912 if (retval)
913 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914
Chuck Leverc216fd72007-11-12 12:16:52 -0500915 retval = nfs_direct_read(iocb, iov, nr_segs, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 if (retval > 0)
Chuck Lever0cdd80d2006-03-20 13:44:29 -0500917 iocb->ki_pos = pos + retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918
919out:
920 return retval;
921}
922
923/**
924 * nfs_file_direct_write - file direct write operation for NFS files
925 * @iocb: target I/O control block
Badari Pulavarty027445c2006-09-30 23:28:46 -0700926 * @iov: vector of user buffers from which to write data
927 * @nr_segs: size of iov vector
Chuck Lever88467052006-03-20 13:44:34 -0500928 * @pos: byte offset in file where writing starts
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 *
930 * We use this function for direct writes instead of calling
931 * generic_file_aio_write() in order to avoid taking the inode
932 * semaphore and updating the i_size. The NFS server will set
933 * the new i_size and this client must read the updated size
934 * back into its cache. We let the server do generic write
935 * parameter checking and report problems.
936 *
937 * We also avoid an unnecessary invocation of generic_osync_inode(),
938 * as it is fairly meaningless to sync the metadata of an NFS file.
939 *
940 * We eliminate local atime updates, see direct read above.
941 *
942 * We avoid unnecessary page cache invalidations for normal cached
943 * readers of this file.
944 *
945 * Note that O_APPEND is not supported for NFS direct writes, as there
946 * is no atomic O_APPEND write facility in the NFS protocol.
947 */
Badari Pulavarty027445c2006-09-30 23:28:46 -0700948ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
949 unsigned long nr_segs, loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950{
Chuck Lever070ea602007-05-19 17:22:52 -0400951 ssize_t retval = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952 struct file *file = iocb->ki_filp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953 struct address_space *mapping = file->f_mapping;
Chuck Leverc216fd72007-11-12 12:16:52 -0500954 size_t count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955
Chuck Leverc216fd72007-11-12 12:16:52 -0500956 count = iov_length(iov, nr_segs);
957 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
958
Chuck Lever6da24bc2008-06-11 17:55:58 -0400959 dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n",
Josef "Jeff" Sipek01cce932006-12-08 02:36:40 -0800960 file->f_path.dentry->d_parent->d_name.name,
961 file->f_path.dentry->d_name.name,
Chuck Leverc216fd72007-11-12 12:16:52 -0500962 count, (long long) pos);
Badari Pulavarty027445c2006-09-30 23:28:46 -0700963
Chuck Leverce1a8e62005-11-30 18:08:17 -0500964 retval = generic_write_checks(file, &pos, &count, 0);
965 if (retval)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966 goto out;
Chuck Leverce1a8e62005-11-30 18:08:17 -0500967
968 retval = -EINVAL;
969 if ((ssize_t) count < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971 retval = 0;
972 if (!count)
973 goto out;
Chuck Leverce1a8e62005-11-30 18:08:17 -0500974
Trond Myklebust29884df2005-12-13 16:13:54 -0500975 retval = nfs_sync_mapping(mapping);
976 if (retval)
977 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978
Chuck Leverc216fd72007-11-12 12:16:52 -0500979 retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
Chuck Lever9eafa8c2006-03-20 13:44:33 -0500980
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981 if (retval > 0)
Chuck Leverce1a8e62005-11-30 18:08:17 -0500982 iocb->ki_pos = pos + retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983
984out:
985 return retval;
986}
987
Chuck Lever88467052006-03-20 13:44:34 -0500988/**
989 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
990 *
991 */
David Howellsf7b422b2006-06-09 09:34:33 -0400992int __init nfs_init_directcache(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993{
994 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
995 sizeof(struct nfs_direct_req),
Paul Jacksonfffb60f2006-03-24 03:16:06 -0800996 0, (SLAB_RECLAIM_ACCOUNT|
997 SLAB_MEM_SPREAD),
Paul Mundt20c2df82007-07-20 10:11:58 +0900998 NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999 if (nfs_direct_cachep == NULL)
1000 return -ENOMEM;
1001
1002 return 0;
1003}
1004
Chuck Lever88467052006-03-20 13:44:34 -05001005/**
David Howellsf7b422b2006-06-09 09:34:33 -04001006 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
Chuck Lever88467052006-03-20 13:44:34 -05001007 *
1008 */
David Brownell266bee82006-06-27 12:59:15 -07001009void nfs_destroy_directcache(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010{
Alexey Dobriyan1a1d92c2006-09-27 01:49:40 -07001011 kmem_cache_destroy(nfs_direct_cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012}