blob: 1940f1a56a5fe059cac63144a4f17cdbf26ab1da [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>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090047#include <linux/slab.h>
Konstantin Khlebnikov7ec10f22011-02-22 00:28:34 +030048#include <linux/task_io_accounting_ops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070049
50#include <linux/nfs_fs.h>
51#include <linux/nfs_page.h>
52#include <linux/sunrpc/clnt.h>
53
54#include <asm/system.h>
55#include <asm/uaccess.h>
Arun Sharma600634972011-07-26 16:09:06 -070056#include <linux/atomic.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070057
Trond Myklebust8d5658c2007-04-10 09:26:35 -040058#include "internal.h"
Chuck Lever91d5b472006-03-20 13:44:14 -050059#include "iostat.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
Linus Torvalds1da177e2005-04-16 15:20:36 -070061#define NFSDBG_FACILITY NFSDBG_VFS
Linus Torvalds1da177e2005-04-16 15:20:36 -070062
Christoph Lametere18b8902006-12-06 20:33:20 -080063static struct kmem_cache *nfs_direct_cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -070064
65/*
66 * This represents a set of asynchronous requests that we're waiting on
67 */
68struct nfs_direct_req {
69 struct kref kref; /* release manager */
Chuck Lever15ce4a02006-03-20 13:44:34 -050070
71 /* I/O parameters */
Trond Myklebusta8881f52006-03-20 13:44:36 -050072 struct nfs_open_context *ctx; /* file open context info */
Trond Myklebustf11ac8d2010-06-25 16:35:53 -040073 struct nfs_lock_context *l_ctx; /* Lock context info */
Chuck Lever99514f82006-03-20 13:44:30 -050074 struct kiocb * iocb; /* controlling i/o request */
Chuck Lever88467052006-03-20 13:44:34 -050075 struct inode * inode; /* target file of i/o */
Chuck Lever15ce4a02006-03-20 13:44:34 -050076
77 /* completion state */
Trond Myklebust607f31e2006-06-28 16:52:45 -040078 atomic_t io_count; /* i/os we're waiting for */
Chuck Lever15ce4a02006-03-20 13:44:34 -050079 spinlock_t lock; /* protect completion state */
Chuck Lever15ce4a02006-03-20 13:44:34 -050080 ssize_t count, /* bytes actually processed */
Linus Torvalds1da177e2005-04-16 15:20:36 -070081 error; /* any reported error */
Trond Myklebustd72b7a62006-03-20 13:44:43 -050082 struct completion completion; /* wait for i/o completion */
Trond Myklebustfad61492006-03-20 13:44:36 -050083
84 /* commit state */
Trond Myklebust607f31e2006-06-28 16:52:45 -040085 struct list_head rewrite_list; /* saved nfs_write_data structs */
Trond Myklebustfad61492006-03-20 13:44:36 -050086 struct nfs_write_data * commit_data; /* special write_data for commits */
87 int flags;
88#define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
89#define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
90 struct nfs_writeverf verf; /* unstable write verifier */
Linus Torvalds1da177e2005-04-16 15:20:36 -070091};
92
Trond Myklebustfad61492006-03-20 13:44:36 -050093static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
Trond Myklebust607f31e2006-06-28 16:52:45 -040094static const struct rpc_call_ops nfs_write_direct_ops;
95
96static inline void get_dreq(struct nfs_direct_req *dreq)
97{
98 atomic_inc(&dreq->io_count);
99}
100
101static inline int put_dreq(struct nfs_direct_req *dreq)
102{
103 return atomic_dec_and_test(&dreq->io_count);
104}
105
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106/**
Chuck Leverb8a32e22006-03-20 13:44:28 -0500107 * nfs_direct_IO - NFS address space operation for direct I/O
108 * @rw: direction (read or write)
109 * @iocb: target I/O control block
110 * @iov: array of vectors that define I/O buffer
111 * @pos: offset in file to begin the operation
112 * @nr_segs: size of iovec array
113 *
114 * The presence of this routine in the address space ops vector means
115 * the NFS client supports direct I/O. However, we shunt off direct
116 * read and write requests before the VFS gets them, so this method
117 * should never be called.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118 */
Chuck Leverb8a32e22006-03-20 13:44:28 -0500119ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
120{
Chuck Leverb8a32e22006-03-20 13:44:28 -0500121 dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
Josef "Jeff" Sipek01cce932006-12-08 02:36:40 -0800122 iocb->ki_filp->f_path.dentry->d_name.name,
Trond Myklebuste99170f2006-04-18 13:21:42 -0400123 (long long) pos, nr_segs);
Chuck Leverb8a32e22006-03-20 13:44:28 -0500124
125 return -EINVAL;
126}
127
Trond Myklebustd4a8f362007-05-22 10:22:27 -0400128static void nfs_direct_dirty_pages(struct page **pages, unsigned int pgbase, size_t count)
Trond Myklebust6b45d852006-03-20 13:44:43 -0500129{
Trond Myklebustd4a8f362007-05-22 10:22:27 -0400130 unsigned int npages;
Chuck Lever749e1462007-05-19 17:22:46 -0400131 unsigned int i;
Trond Myklebustd4a8f362007-05-22 10:22:27 -0400132
133 if (count == 0)
134 return;
135 pages += (pgbase >> PAGE_SHIFT);
136 npages = (count + (pgbase & ~PAGE_MASK) + PAGE_SIZE - 1) >> PAGE_SHIFT;
Trond Myklebust6b45d852006-03-20 13:44:43 -0500137 for (i = 0; i < npages; i++) {
138 struct page *page = pages[i];
Trond Myklebust607f31e2006-06-28 16:52:45 -0400139 if (!PageCompound(page))
Trond Myklebustd4a8f362007-05-22 10:22:27 -0400140 set_page_dirty(page);
Trond Myklebust6b45d852006-03-20 13:44:43 -0500141 }
Chuck Lever9c93ab72006-06-20 12:56:31 -0400142}
143
Chuck Lever749e1462007-05-19 17:22:46 -0400144static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
Chuck Lever9c93ab72006-06-20 12:56:31 -0400145{
Chuck Lever749e1462007-05-19 17:22:46 -0400146 unsigned int i;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400147 for (i = 0; i < npages; i++)
148 page_cache_release(pages[i]);
Trond Myklebust6b45d852006-03-20 13:44:43 -0500149}
150
Chuck Lever93619e52006-03-20 13:44:31 -0500151static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153 struct nfs_direct_req *dreq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154
Christoph Lametere94b1762006-12-06 20:33:17 -0800155 dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156 if (!dreq)
157 return NULL;
158
159 kref_init(&dreq->kref);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400160 kref_get(&dreq->kref);
Trond Myklebustd72b7a62006-03-20 13:44:43 -0500161 init_completion(&dreq->completion);
Trond Myklebustfad61492006-03-20 13:44:36 -0500162 INIT_LIST_HEAD(&dreq->rewrite_list);
Chuck Lever93619e52006-03-20 13:44:31 -0500163 dreq->iocb = NULL;
Trond Myklebusta8881f52006-03-20 13:44:36 -0500164 dreq->ctx = NULL;
Trond Myklebustf11ac8d2010-06-25 16:35:53 -0400165 dreq->l_ctx = NULL;
Chuck Lever15ce4a02006-03-20 13:44:34 -0500166 spin_lock_init(&dreq->lock);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400167 atomic_set(&dreq->io_count, 0);
Chuck Lever15ce4a02006-03-20 13:44:34 -0500168 dreq->count = 0;
169 dreq->error = 0;
Trond Myklebustfad61492006-03-20 13:44:36 -0500170 dreq->flags = 0;
Chuck Lever93619e52006-03-20 13:44:31 -0500171
172 return dreq;
173}
174
Trond Myklebustb4946ff2007-05-30 12:58:00 -0400175static void nfs_direct_req_free(struct kref *kref)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176{
177 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
Trond Myklebusta8881f52006-03-20 13:44:36 -0500178
Trond Myklebustf11ac8d2010-06-25 16:35:53 -0400179 if (dreq->l_ctx != NULL)
180 nfs_put_lock_context(dreq->l_ctx);
Trond Myklebusta8881f52006-03-20 13:44:36 -0500181 if (dreq->ctx != NULL)
182 put_nfs_open_context(dreq->ctx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183 kmem_cache_free(nfs_direct_cachep, dreq);
184}
185
Trond Myklebustb4946ff2007-05-30 12:58:00 -0400186static void nfs_direct_req_release(struct nfs_direct_req *dreq)
187{
188 kref_put(&dreq->kref, nfs_direct_req_free);
189}
190
Chuck Leverd4cc9482006-03-20 13:44:28 -0500191/*
Chuck Leverbc0fb202006-03-20 13:44:31 -0500192 * Collects and returns the final error value/byte-count.
193 */
194static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
195{
Chuck Lever15ce4a02006-03-20 13:44:34 -0500196 ssize_t result = -EIOCBQUEUED;
Chuck Leverbc0fb202006-03-20 13:44:31 -0500197
198 /* Async requests don't wait here */
199 if (dreq->iocb)
200 goto out;
201
Matthew Wilcox150030b2007-12-06 16:24:39 -0500202 result = wait_for_completion_killable(&dreq->completion);
Chuck Leverbc0fb202006-03-20 13:44:31 -0500203
204 if (!result)
Chuck Lever15ce4a02006-03-20 13:44:34 -0500205 result = dreq->error;
Chuck Leverbc0fb202006-03-20 13:44:31 -0500206 if (!result)
Chuck Lever15ce4a02006-03-20 13:44:34 -0500207 result = dreq->count;
Chuck Leverbc0fb202006-03-20 13:44:31 -0500208
209out:
Chuck Leverbc0fb202006-03-20 13:44:31 -0500210 return (ssize_t) result;
211}
212
213/*
Trond Myklebust607f31e2006-06-28 16:52:45 -0400214 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
215 * the iocb is still valid here if this is a synchronous request.
Chuck Lever63ab46a2006-03-20 13:44:31 -0500216 */
217static void nfs_direct_complete(struct nfs_direct_req *dreq)
218{
Chuck Lever63ab46a2006-03-20 13:44:31 -0500219 if (dreq->iocb) {
Chuck Lever15ce4a02006-03-20 13:44:34 -0500220 long res = (long) dreq->error;
Chuck Lever63ab46a2006-03-20 13:44:31 -0500221 if (!res)
Chuck Lever15ce4a02006-03-20 13:44:34 -0500222 res = (long) dreq->count;
Chuck Lever63ab46a2006-03-20 13:44:31 -0500223 aio_complete(dreq->iocb, res, 0);
Trond Myklebustd72b7a62006-03-20 13:44:43 -0500224 }
225 complete_all(&dreq->completion);
Chuck Lever63ab46a2006-03-20 13:44:31 -0500226
Trond Myklebustb4946ff2007-05-30 12:58:00 -0400227 nfs_direct_req_release(dreq);
Chuck Lever63ab46a2006-03-20 13:44:31 -0500228}
229
230/*
Trond Myklebust607f31e2006-06-28 16:52:45 -0400231 * We must hold a reference to all the pages in this direct read request
232 * until the RPCs complete. This could be long *after* we are woken up in
233 * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
Chuck Lever06cf6f22006-06-20 12:56:49 -0400234 */
Trond Myklebustec06c092006-03-20 13:44:27 -0500235static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236{
Trond Myklebustec06c092006-03-20 13:44:27 -0500237 struct nfs_read_data *data = calldata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238
Trond Myklebustfdd1e742008-04-15 16:33:58 -0400239 nfs_readpage_result(task, data);
240}
241
242static void nfs_direct_read_release(void *calldata)
243{
244
245 struct nfs_read_data *data = calldata;
246 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
247 int status = data->task.tk_status;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248
Chuck Lever15ce4a02006-03-20 13:44:34 -0500249 spin_lock(&dreq->lock);
Trond Myklebustfdd1e742008-04-15 16:33:58 -0400250 if (unlikely(status < 0)) {
251 dreq->error = status;
Trond Myklebustd4a8f362007-05-22 10:22:27 -0400252 spin_unlock(&dreq->lock);
253 } else {
254 dreq->count += data->res.count;
255 spin_unlock(&dreq->lock);
256 nfs_direct_dirty_pages(data->pagevec,
257 data->args.pgbase,
258 data->res.count);
259 }
260 nfs_direct_release_pages(data->pagevec, data->npages);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400261
262 if (put_dreq(dreq))
263 nfs_direct_complete(dreq);
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400264 nfs_readdata_free(data);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265}
266
Trond Myklebustec06c092006-03-20 13:44:27 -0500267static const struct rpc_call_ops nfs_read_direct_ops = {
Andy Adamsonf11c88a2009-04-01 09:22:25 -0400268#if defined(CONFIG_NFS_V4_1)
269 .rpc_call_prepare = nfs_read_prepare,
270#endif /* CONFIG_NFS_V4_1 */
Trond Myklebustec06c092006-03-20 13:44:27 -0500271 .rpc_call_done = nfs_direct_read_result,
Trond Myklebustfdd1e742008-04-15 16:33:58 -0400272 .rpc_release = nfs_direct_read_release,
Trond Myklebustec06c092006-03-20 13:44:27 -0500273};
274
Chuck Leverd4cc9482006-03-20 13:44:28 -0500275/*
Trond Myklebust607f31e2006-06-28 16:52:45 -0400276 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
277 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
278 * bail and stop sending more reads. Read length accounting is
279 * handled automatically by nfs_direct_read_result(). Otherwise, if
280 * no requests have been sent, just return an error.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 */
Chuck Lever02fe4942007-11-12 12:17:03 -0500282static ssize_t nfs_direct_read_schedule_segment(struct nfs_direct_req *dreq,
283 const struct iovec *iov,
284 loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285{
Trond Myklebusta8881f52006-03-20 13:44:36 -0500286 struct nfs_open_context *ctx = dreq->ctx;
Al Viro3d4ff432011-06-22 18:40:12 -0400287 struct inode *inode = ctx->dentry->d_inode;
Chuck Lever02fe4942007-11-12 12:17:03 -0500288 unsigned long user_addr = (unsigned long)iov->iov_base;
289 size_t count = iov->iov_len;
Chuck Lever5dd602f2006-03-20 13:44:29 -0500290 size_t rsize = NFS_SERVER(inode)->rsize;
Trond Myklebust07737692007-10-25 18:42:54 -0400291 struct rpc_task *task;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400292 struct rpc_message msg = {
293 .rpc_cred = ctx->cred,
294 };
Trond Myklebust84115e12007-07-14 15:39:59 -0400295 struct rpc_task_setup task_setup_data = {
296 .rpc_client = NFS_CLIENT(inode),
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400297 .rpc_message = &msg,
Trond Myklebust84115e12007-07-14 15:39:59 -0400298 .callback_ops = &nfs_read_direct_ops,
Trond Myklebust101070c2008-02-19 20:04:23 -0500299 .workqueue = nfsiod_workqueue,
Trond Myklebust84115e12007-07-14 15:39:59 -0400300 .flags = RPC_TASK_ASYNC,
301 };
Trond Myklebust607f31e2006-06-28 16:52:45 -0400302 unsigned int pgbase;
303 int result;
304 ssize_t started = 0;
Chuck Lever82b145c2006-06-20 12:57:03 -0400305
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306 do {
Chuck Lever82b145c2006-06-20 12:57:03 -0400307 struct nfs_read_data *data;
Chuck Lever5dd602f2006-03-20 13:44:29 -0500308 size_t bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700310 pgbase = user_addr & ~PAGE_MASK;
311 bytes = min(rsize,count);
312
Trond Myklebust607f31e2006-06-28 16:52:45 -0400313 result = -ENOMEM;
Trond Myklebust8d5658c2007-04-10 09:26:35 -0400314 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
Trond Myklebust607f31e2006-06-28 16:52:45 -0400315 if (unlikely(!data))
316 break;
317
Trond Myklebust607f31e2006-06-28 16:52:45 -0400318 down_read(&current->mm->mmap_sem);
319 result = get_user_pages(current, current->mm, user_addr,
320 data->npages, 1, 0, data->pagevec, NULL);
321 up_read(&current->mm->mmap_sem);
Chuck Lever749e1462007-05-19 17:22:46 -0400322 if (result < 0) {
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400323 nfs_readdata_free(data);
Chuck Lever749e1462007-05-19 17:22:46 -0400324 break;
325 }
326 if ((unsigned)result < data->npages) {
Trond Myklebustd9df8d62007-05-22 10:22:20 -0400327 bytes = result * PAGE_SIZE;
328 if (bytes <= pgbase) {
329 nfs_direct_release_pages(data->pagevec, result);
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400330 nfs_readdata_free(data);
Trond Myklebustd9df8d62007-05-22 10:22:20 -0400331 break;
332 }
333 bytes -= pgbase;
334 data->npages = result;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400335 }
Chuck Lever06cf6f22006-06-20 12:56:49 -0400336
Trond Myklebust607f31e2006-06-28 16:52:45 -0400337 get_dreq(dreq);
338
339 data->req = (struct nfs_page *) dreq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 data->inode = inode;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400341 data->cred = msg.rpc_cred;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342 data->args.fh = NFS_FH(inode);
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400343 data->args.context = ctx;
Trond Myklebustf11ac8d2010-06-25 16:35:53 -0400344 data->args.lock_context = dreq->l_ctx;
Chuck Lever88467052006-03-20 13:44:34 -0500345 data->args.offset = pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346 data->args.pgbase = pgbase;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400347 data->args.pages = data->pagevec;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348 data->args.count = bytes;
349 data->res.fattr = &data->fattr;
350 data->res.eof = 0;
351 data->res.count = bytes;
Chuck Lever65d26952010-02-15 12:19:53 -0500352 nfs_fattr_init(&data->fattr);
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400353 msg.rpc_argp = &data->args;
354 msg.rpc_resp = &data->res;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355
Trond Myklebust07737692007-10-25 18:42:54 -0400356 task_setup_data.task = &data->task;
Trond Myklebust84115e12007-07-14 15:39:59 -0400357 task_setup_data.callback_data = data;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400358 NFS_PROTO(inode)->read_setup(data, &msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359
Trond Myklebust07737692007-10-25 18:42:54 -0400360 task = rpc_run_task(&task_setup_data);
Trond Myklebustdbae4c72008-04-14 14:54:53 -0400361 if (IS_ERR(task))
362 break;
363 rpc_put_task(task);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364
Chuck Levera3f565b2007-01-31 12:14:01 -0500365 dprintk("NFS: %5u initiated direct read call "
366 "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 data->task.tk_pid,
368 inode->i_sb->s_id,
369 (long long)NFS_FILEID(inode),
370 bytes,
371 (unsigned long long)data->args.offset);
372
Trond Myklebust607f31e2006-06-28 16:52:45 -0400373 started += bytes;
374 user_addr += bytes;
Chuck Lever88467052006-03-20 13:44:34 -0500375 pos += bytes;
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700376 /* FIXME: Remove this unnecessary math from final patch */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 pgbase += bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378 pgbase &= ~PAGE_MASK;
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700379 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380
381 count -= bytes;
382 } while (count != 0);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400383
Trond Myklebust607f31e2006-06-28 16:52:45 -0400384 if (started)
Chuck Leverc216fd72007-11-12 12:16:52 -0500385 return started;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400386 return result < 0 ? (ssize_t) result : -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387}
388
Chuck Lever19f73782007-11-12 12:16:47 -0500389static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
390 const struct iovec *iov,
391 unsigned long nr_segs,
392 loff_t pos)
393{
394 ssize_t result = -EINVAL;
395 size_t requested_bytes = 0;
396 unsigned long seg;
397
398 get_dreq(dreq);
399
400 for (seg = 0; seg < nr_segs; seg++) {
401 const struct iovec *vec = &iov[seg];
Chuck Lever02fe4942007-11-12 12:17:03 -0500402 result = nfs_direct_read_schedule_segment(dreq, vec, pos);
Chuck Lever19f73782007-11-12 12:16:47 -0500403 if (result < 0)
404 break;
405 requested_bytes += result;
406 if ((size_t)result < vec->iov_len)
407 break;
408 pos += vec->iov_len;
409 }
410
Chuck Lever839f7ad2011-01-21 15:54:57 +0000411 /*
412 * If no bytes were started, return the error, and let the
413 * generic layer handle the completion.
414 */
415 if (requested_bytes == 0) {
416 nfs_direct_req_release(dreq);
417 return result < 0 ? result : -EIO;
418 }
419
Chuck Lever19f73782007-11-12 12:16:47 -0500420 if (put_dreq(dreq))
421 nfs_direct_complete(dreq);
Chuck Lever839f7ad2011-01-21 15:54:57 +0000422 return 0;
Chuck Lever19f73782007-11-12 12:16:47 -0500423}
424
Chuck Leverc216fd72007-11-12 12:16:52 -0500425static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
426 unsigned long nr_segs, loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427{
Trond Myklebustf11ac8d2010-06-25 16:35:53 -0400428 ssize_t result = -ENOMEM;
Chuck Lever99514f82006-03-20 13:44:30 -0500429 struct inode *inode = iocb->ki_filp->f_mapping->host;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430 struct nfs_direct_req *dreq;
431
Trond Myklebust607f31e2006-06-28 16:52:45 -0400432 dreq = nfs_direct_req_alloc();
Trond Myklebustf11ac8d2010-06-25 16:35:53 -0400433 if (dreq == NULL)
434 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435
Chuck Lever91d5b472006-03-20 13:44:14 -0500436 dreq->inode = inode;
Trond Myklebustcd3758e2007-08-10 17:44:32 -0400437 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
Trond Myklebustf11ac8d2010-06-25 16:35:53 -0400438 dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
439 if (dreq->l_ctx == NULL)
440 goto out_release;
Chuck Lever487b8372006-03-20 13:44:30 -0500441 if (!is_sync_kiocb(iocb))
442 dreq->iocb = iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443
Chuck Leverc216fd72007-11-12 12:16:52 -0500444 result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400445 if (!result)
446 result = nfs_direct_wait(dreq);
Trond Myklebustf11ac8d2010-06-25 16:35:53 -0400447out_release:
Trond Myklebustb4946ff2007-05-30 12:58:00 -0400448 nfs_direct_req_release(dreq);
Trond Myklebustf11ac8d2010-06-25 16:35:53 -0400449out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 return result;
451}
452
Trond Myklebustfad61492006-03-20 13:44:36 -0500453static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454{
Trond Myklebust607f31e2006-06-28 16:52:45 -0400455 while (!list_empty(&dreq->rewrite_list)) {
456 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
Trond Myklebustfad61492006-03-20 13:44:36 -0500457 list_del(&data->pages);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400458 nfs_direct_release_pages(data->pagevec, data->npages);
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400459 nfs_writedata_free(data);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461}
462
Trond Myklebustfad61492006-03-20 13:44:36 -0500463#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
464static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465{
Trond Myklebust607f31e2006-06-28 16:52:45 -0400466 struct inode *inode = dreq->inode;
467 struct list_head *p;
468 struct nfs_write_data *data;
Trond Myklebust07737692007-10-25 18:42:54 -0400469 struct rpc_task *task;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400470 struct rpc_message msg = {
471 .rpc_cred = dreq->ctx->cred,
472 };
Trond Myklebust84115e12007-07-14 15:39:59 -0400473 struct rpc_task_setup task_setup_data = {
474 .rpc_client = NFS_CLIENT(inode),
Terry Loftina8b40bc2009-10-22 21:36:01 -0400475 .rpc_message = &msg,
Trond Myklebust84115e12007-07-14 15:39:59 -0400476 .callback_ops = &nfs_write_direct_ops,
Trond Myklebust101070c2008-02-19 20:04:23 -0500477 .workqueue = nfsiod_workqueue,
Trond Myklebust84115e12007-07-14 15:39:59 -0400478 .flags = RPC_TASK_ASYNC,
479 };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480
Trond Myklebustfad61492006-03-20 13:44:36 -0500481 dreq->count = 0;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400482 get_dreq(dreq);
Trond Myklebustfad61492006-03-20 13:44:36 -0500483
Trond Myklebust607f31e2006-06-28 16:52:45 -0400484 list_for_each(p, &dreq->rewrite_list) {
485 data = list_entry(p, struct nfs_write_data, pages);
486
487 get_dreq(dreq);
488
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400489 /* Use stable writes */
490 data->args.stable = NFS_FILE_SYNC;
491
Trond Myklebust607f31e2006-06-28 16:52:45 -0400492 /*
493 * Reset data->res.
494 */
495 nfs_fattr_init(&data->fattr);
496 data->res.count = data->args.count;
497 memset(&data->verf, 0, sizeof(data->verf));
498
499 /*
500 * Reuse data->task; data->args should not have changed
501 * since the original request was sent.
502 */
Trond Myklebust07737692007-10-25 18:42:54 -0400503 task_setup_data.task = &data->task;
Trond Myklebust84115e12007-07-14 15:39:59 -0400504 task_setup_data.callback_data = data;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400505 msg.rpc_argp = &data->args;
506 msg.rpc_resp = &data->res;
507 NFS_PROTO(inode)->write_setup(data, &msg);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400508
Trond Myklebust607f31e2006-06-28 16:52:45 -0400509 /*
510 * We're called via an RPC callback, so BKL is already held.
511 */
Trond Myklebust07737692007-10-25 18:42:54 -0400512 task = rpc_run_task(&task_setup_data);
513 if (!IS_ERR(task))
514 rpc_put_task(task);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400515
516 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
517 data->task.tk_pid,
518 inode->i_sb->s_id,
519 (long long)NFS_FILEID(inode),
520 data->args.count,
521 (unsigned long long)data->args.offset);
522 }
523
524 if (put_dreq(dreq))
525 nfs_direct_write_complete(dreq, inode);
Trond Myklebustfad61492006-03-20 13:44:36 -0500526}
527
528static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
529{
530 struct nfs_write_data *data = calldata;
Trond Myklebustfad61492006-03-20 13:44:36 -0500531
532 /* Call the NFS version-specific code */
Trond Myklebustc9d8f892008-04-15 16:56:39 -0400533 NFS_PROTO(data->inode)->commit_done(task, data);
534}
535
536static void nfs_direct_commit_release(void *calldata)
537{
538 struct nfs_write_data *data = calldata;
539 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
540 int status = data->task.tk_status;
541
542 if (status < 0) {
Trond Myklebust60fa3f72007-04-14 19:11:52 -0400543 dprintk("NFS: %5u commit failed with error %d.\n",
Trond Myklebustc9d8f892008-04-15 16:56:39 -0400544 data->task.tk_pid, status);
Trond Myklebustfad61492006-03-20 13:44:36 -0500545 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
Trond Myklebust60fa3f72007-04-14 19:11:52 -0400546 } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
Trond Myklebustc9d8f892008-04-15 16:56:39 -0400547 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
Trond Myklebustfad61492006-03-20 13:44:36 -0500548 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
549 }
550
Trond Myklebustc9d8f892008-04-15 16:56:39 -0400551 dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
Trond Myklebustfad61492006-03-20 13:44:36 -0500552 nfs_direct_write_complete(dreq, data->inode);
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400553 nfs_commit_free(data);
Trond Myklebustfad61492006-03-20 13:44:36 -0500554}
555
556static const struct rpc_call_ops nfs_commit_direct_ops = {
Andy Adamson21d9a852009-04-01 09:22:27 -0400557#if defined(CONFIG_NFS_V4_1)
558 .rpc_call_prepare = nfs_write_prepare,
559#endif /* CONFIG_NFS_V4_1 */
Trond Myklebustfad61492006-03-20 13:44:36 -0500560 .rpc_call_done = nfs_direct_commit_result,
Trond Myklebustc9d8f892008-04-15 16:56:39 -0400561 .rpc_release = nfs_direct_commit_release,
Trond Myklebustfad61492006-03-20 13:44:36 -0500562};
563
564static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
565{
Trond Myklebustfad61492006-03-20 13:44:36 -0500566 struct nfs_write_data *data = dreq->commit_data;
Trond Myklebust07737692007-10-25 18:42:54 -0400567 struct rpc_task *task;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400568 struct rpc_message msg = {
569 .rpc_argp = &data->args,
570 .rpc_resp = &data->res,
571 .rpc_cred = dreq->ctx->cred,
572 };
Trond Myklebust84115e12007-07-14 15:39:59 -0400573 struct rpc_task_setup task_setup_data = {
Trond Myklebust07737692007-10-25 18:42:54 -0400574 .task = &data->task,
Trond Myklebust84115e12007-07-14 15:39:59 -0400575 .rpc_client = NFS_CLIENT(dreq->inode),
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400576 .rpc_message = &msg,
Trond Myklebust84115e12007-07-14 15:39:59 -0400577 .callback_ops = &nfs_commit_direct_ops,
578 .callback_data = data,
Trond Myklebust101070c2008-02-19 20:04:23 -0500579 .workqueue = nfsiod_workqueue,
Trond Myklebust84115e12007-07-14 15:39:59 -0400580 .flags = RPC_TASK_ASYNC,
581 };
Trond Myklebustfad61492006-03-20 13:44:36 -0500582
583 data->inode = dreq->inode;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400584 data->cred = msg.rpc_cred;
Trond Myklebustfad61492006-03-20 13:44:36 -0500585
586 data->args.fh = NFS_FH(data->inode);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400587 data->args.offset = 0;
588 data->args.count = 0;
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400589 data->args.context = dreq->ctx;
Trond Myklebustf11ac8d2010-06-25 16:35:53 -0400590 data->args.lock_context = dreq->l_ctx;
Trond Myklebustfad61492006-03-20 13:44:36 -0500591 data->res.count = 0;
592 data->res.fattr = &data->fattr;
593 data->res.verf = &data->verf;
Chuck Lever65d26952010-02-15 12:19:53 -0500594 nfs_fattr_init(&data->fattr);
Trond Myklebustfad61492006-03-20 13:44:36 -0500595
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400596 NFS_PROTO(data->inode)->commit_setup(data, &msg);
Trond Myklebustfad61492006-03-20 13:44:36 -0500597
Trond Myklebustfad61492006-03-20 13:44:36 -0500598 /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
599 dreq->commit_data = NULL;
600
Trond Myklebuste99170f2006-04-18 13:21:42 -0400601 dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
Trond Myklebustfad61492006-03-20 13:44:36 -0500602
Trond Myklebust07737692007-10-25 18:42:54 -0400603 task = rpc_run_task(&task_setup_data);
604 if (!IS_ERR(task))
605 rpc_put_task(task);
Trond Myklebustfad61492006-03-20 13:44:36 -0500606}
607
608static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
609{
610 int flags = dreq->flags;
611
612 dreq->flags = 0;
613 switch (flags) {
614 case NFS_ODIRECT_DO_COMMIT:
615 nfs_direct_commit_schedule(dreq);
616 break;
617 case NFS_ODIRECT_RESCHED_WRITES:
618 nfs_direct_write_reschedule(dreq);
619 break;
620 default:
Trond Myklebustfad61492006-03-20 13:44:36 -0500621 if (dreq->commit_data != NULL)
622 nfs_commit_free(dreq->commit_data);
623 nfs_direct_free_writedata(dreq);
Trond Myklebustcd9ae2b2006-10-19 23:28:40 -0700624 nfs_zap_mapping(inode, inode->i_mapping);
Trond Myklebustfad61492006-03-20 13:44:36 -0500625 nfs_direct_complete(dreq);
626 }
627}
628
629static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
630{
Trond Myklebustc9d8f892008-04-15 16:56:39 -0400631 dreq->commit_data = nfs_commitdata_alloc();
Trond Myklebustfad61492006-03-20 13:44:36 -0500632 if (dreq->commit_data != NULL)
633 dreq->commit_data->req = (struct nfs_page *) dreq;
634}
635#else
636static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
637{
638 dreq->commit_data = NULL;
639}
640
641static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
642{
Trond Myklebustfad61492006-03-20 13:44:36 -0500643 nfs_direct_free_writedata(dreq);
Trond Myklebustcd9ae2b2006-10-19 23:28:40 -0700644 nfs_zap_mapping(inode, inode->i_mapping);
Trond Myklebustfad61492006-03-20 13:44:36 -0500645 nfs_direct_complete(dreq);
646}
647#endif
648
Chuck Lever462d5b32006-03-20 13:44:32 -0500649static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
650{
651 struct nfs_write_data *data = calldata;
Chuck Lever462d5b32006-03-20 13:44:32 -0500652
Fred Isaman83762c52011-02-11 15:42:37 +0000653 nfs_writeback_done(task, data);
Trond Myklebustc9d8f892008-04-15 16:56:39 -0400654}
655
656/*
657 * NB: Return the value of the first error return code. Subsequent
658 * errors after the first one are ignored.
659 */
660static void nfs_direct_write_release(void *calldata)
661{
662 struct nfs_write_data *data = calldata;
663 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
664 int status = data->task.tk_status;
Chuck Lever462d5b32006-03-20 13:44:32 -0500665
Chuck Lever15ce4a02006-03-20 13:44:34 -0500666 spin_lock(&dreq->lock);
Chuck Lever462d5b32006-03-20 13:44:32 -0500667
Trond Myklebust60fa3f72007-04-14 19:11:52 -0400668 if (unlikely(status < 0)) {
Neil Brown432409e2007-10-23 17:09:13 -0400669 /* An error has occurred, so we should not commit */
Trond Myklebust60fa3f72007-04-14 19:11:52 -0400670 dreq->flags = 0;
671 dreq->error = status;
Trond Myklebusteda3cef2006-10-19 23:28:38 -0700672 }
Neil Brown432409e2007-10-23 17:09:13 -0400673 if (unlikely(dreq->error != 0))
674 goto out_unlock;
Trond Myklebusteda3cef2006-10-19 23:28:38 -0700675
676 dreq->count += data->res.count;
Chuck Lever15ce4a02006-03-20 13:44:34 -0500677
Trond Myklebustfad61492006-03-20 13:44:36 -0500678 if (data->res.verf->committed != NFS_FILE_SYNC) {
679 switch (dreq->flags) {
680 case 0:
681 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
682 dreq->flags = NFS_ODIRECT_DO_COMMIT;
683 break;
684 case NFS_ODIRECT_DO_COMMIT:
685 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
Trond Myklebustc9d8f892008-04-15 16:56:39 -0400686 dprintk("NFS: %5u write verify failed\n", data->task.tk_pid);
Trond Myklebustfad61492006-03-20 13:44:36 -0500687 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
688 }
689 }
690 }
Trond Myklebusteda3cef2006-10-19 23:28:38 -0700691out_unlock:
Trond Myklebustfad61492006-03-20 13:44:36 -0500692 spin_unlock(&dreq->lock);
Trond Myklebustfad61492006-03-20 13:44:36 -0500693
Trond Myklebust607f31e2006-06-28 16:52:45 -0400694 if (put_dreq(dreq))
695 nfs_direct_write_complete(dreq, data->inode);
Chuck Lever462d5b32006-03-20 13:44:32 -0500696}
697
698static const struct rpc_call_ops nfs_write_direct_ops = {
Andy Adamsondef6ed72009-04-01 09:22:26 -0400699#if defined(CONFIG_NFS_V4_1)
700 .rpc_call_prepare = nfs_write_prepare,
701#endif /* CONFIG_NFS_V4_1 */
Chuck Lever462d5b32006-03-20 13:44:32 -0500702 .rpc_call_done = nfs_direct_write_result,
Trond Myklebustfad61492006-03-20 13:44:36 -0500703 .rpc_release = nfs_direct_write_release,
Chuck Lever462d5b32006-03-20 13:44:32 -0500704};
705
706/*
Trond Myklebust607f31e2006-06-28 16:52:45 -0400707 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
708 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
709 * bail and stop sending more writes. Write length accounting is
710 * handled automatically by nfs_direct_write_result(). Otherwise, if
711 * no requests have been sent, just return an error.
Chuck Lever462d5b32006-03-20 13:44:32 -0500712 */
Chuck Lever02fe4942007-11-12 12:17:03 -0500713static ssize_t nfs_direct_write_schedule_segment(struct nfs_direct_req *dreq,
714 const struct iovec *iov,
715 loff_t pos, int sync)
Chuck Lever462d5b32006-03-20 13:44:32 -0500716{
Trond Myklebusta8881f52006-03-20 13:44:36 -0500717 struct nfs_open_context *ctx = dreq->ctx;
Al Viro3d4ff432011-06-22 18:40:12 -0400718 struct inode *inode = ctx->dentry->d_inode;
Chuck Lever02fe4942007-11-12 12:17:03 -0500719 unsigned long user_addr = (unsigned long)iov->iov_base;
720 size_t count = iov->iov_len;
Trond Myklebust07737692007-10-25 18:42:54 -0400721 struct rpc_task *task;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400722 struct rpc_message msg = {
723 .rpc_cred = ctx->cred,
724 };
Trond Myklebust84115e12007-07-14 15:39:59 -0400725 struct rpc_task_setup task_setup_data = {
726 .rpc_client = NFS_CLIENT(inode),
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400727 .rpc_message = &msg,
Trond Myklebust84115e12007-07-14 15:39:59 -0400728 .callback_ops = &nfs_write_direct_ops,
Trond Myklebust101070c2008-02-19 20:04:23 -0500729 .workqueue = nfsiod_workqueue,
Trond Myklebust84115e12007-07-14 15:39:59 -0400730 .flags = RPC_TASK_ASYNC,
731 };
Chuck Lever462d5b32006-03-20 13:44:32 -0500732 size_t wsize = NFS_SERVER(inode)->wsize;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400733 unsigned int pgbase;
734 int result;
735 ssize_t started = 0;
Chuck Lever82b145c2006-06-20 12:57:03 -0400736
Chuck Lever462d5b32006-03-20 13:44:32 -0500737 do {
Chuck Lever82b145c2006-06-20 12:57:03 -0400738 struct nfs_write_data *data;
Chuck Lever462d5b32006-03-20 13:44:32 -0500739 size_t bytes;
740
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700741 pgbase = user_addr & ~PAGE_MASK;
742 bytes = min(wsize,count);
743
Trond Myklebust607f31e2006-06-28 16:52:45 -0400744 result = -ENOMEM;
Trond Myklebust8d5658c2007-04-10 09:26:35 -0400745 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
Trond Myklebust607f31e2006-06-28 16:52:45 -0400746 if (unlikely(!data))
747 break;
748
Trond Myklebust607f31e2006-06-28 16:52:45 -0400749 down_read(&current->mm->mmap_sem);
750 result = get_user_pages(current, current->mm, user_addr,
751 data->npages, 0, 0, data->pagevec, NULL);
752 up_read(&current->mm->mmap_sem);
Chuck Lever749e1462007-05-19 17:22:46 -0400753 if (result < 0) {
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400754 nfs_writedata_free(data);
Chuck Lever749e1462007-05-19 17:22:46 -0400755 break;
756 }
757 if ((unsigned)result < data->npages) {
Trond Myklebustd9df8d62007-05-22 10:22:20 -0400758 bytes = result * PAGE_SIZE;
759 if (bytes <= pgbase) {
760 nfs_direct_release_pages(data->pagevec, result);
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400761 nfs_writedata_free(data);
Trond Myklebustd9df8d62007-05-22 10:22:20 -0400762 break;
763 }
764 bytes -= pgbase;
765 data->npages = result;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400766 }
767
768 get_dreq(dreq);
769
Trond Myklebustfad61492006-03-20 13:44:36 -0500770 list_move_tail(&data->pages, &dreq->rewrite_list);
Chuck Lever462d5b32006-03-20 13:44:32 -0500771
Trond Myklebust607f31e2006-06-28 16:52:45 -0400772 data->req = (struct nfs_page *) dreq;
Chuck Lever462d5b32006-03-20 13:44:32 -0500773 data->inode = inode;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400774 data->cred = msg.rpc_cred;
Chuck Lever462d5b32006-03-20 13:44:32 -0500775 data->args.fh = NFS_FH(inode);
Trond Myklebust1ae88b22009-08-12 09:12:30 -0400776 data->args.context = ctx;
Trond Myklebustf11ac8d2010-06-25 16:35:53 -0400777 data->args.lock_context = dreq->l_ctx;
Chuck Lever88467052006-03-20 13:44:34 -0500778 data->args.offset = pos;
Chuck Lever462d5b32006-03-20 13:44:32 -0500779 data->args.pgbase = pgbase;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400780 data->args.pages = data->pagevec;
Chuck Lever462d5b32006-03-20 13:44:32 -0500781 data->args.count = bytes;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400782 data->args.stable = sync;
Chuck Lever462d5b32006-03-20 13:44:32 -0500783 data->res.fattr = &data->fattr;
784 data->res.count = bytes;
Chuck Lever47989d72006-03-20 13:44:32 -0500785 data->res.verf = &data->verf;
Chuck Lever65d26952010-02-15 12:19:53 -0500786 nfs_fattr_init(&data->fattr);
Chuck Lever462d5b32006-03-20 13:44:32 -0500787
Trond Myklebust07737692007-10-25 18:42:54 -0400788 task_setup_data.task = &data->task;
Trond Myklebust84115e12007-07-14 15:39:59 -0400789 task_setup_data.callback_data = data;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400790 msg.rpc_argp = &data->args;
791 msg.rpc_resp = &data->res;
792 NFS_PROTO(inode)->write_setup(data, &msg);
Chuck Lever462d5b32006-03-20 13:44:32 -0500793
Trond Myklebust07737692007-10-25 18:42:54 -0400794 task = rpc_run_task(&task_setup_data);
Trond Myklebustdbae4c72008-04-14 14:54:53 -0400795 if (IS_ERR(task))
796 break;
797 rpc_put_task(task);
Chuck Lever462d5b32006-03-20 13:44:32 -0500798
Chuck Levera3f565b2007-01-31 12:14:01 -0500799 dprintk("NFS: %5u initiated direct write call "
800 "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
Chuck Lever462d5b32006-03-20 13:44:32 -0500801 data->task.tk_pid,
802 inode->i_sb->s_id,
803 (long long)NFS_FILEID(inode),
804 bytes,
805 (unsigned long long)data->args.offset);
806
Trond Myklebust607f31e2006-06-28 16:52:45 -0400807 started += bytes;
808 user_addr += bytes;
Chuck Lever88467052006-03-20 13:44:34 -0500809 pos += bytes;
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700810
811 /* FIXME: Remove this useless math from the final patch */
Chuck Lever462d5b32006-03-20 13:44:32 -0500812 pgbase += bytes;
Chuck Lever462d5b32006-03-20 13:44:32 -0500813 pgbase &= ~PAGE_MASK;
Trond Myklebuste9f7bee2006-09-08 09:48:54 -0700814 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
Chuck Lever462d5b32006-03-20 13:44:32 -0500815
816 count -= bytes;
817 } while (count != 0);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400818
Trond Myklebust607f31e2006-06-28 16:52:45 -0400819 if (started)
Chuck Leverc216fd72007-11-12 12:16:52 -0500820 return started;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400821 return result < 0 ? (ssize_t) result : -EFAULT;
Chuck Lever462d5b32006-03-20 13:44:32 -0500822}
823
Chuck Lever19f73782007-11-12 12:16:47 -0500824static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
825 const struct iovec *iov,
826 unsigned long nr_segs,
827 loff_t pos, int sync)
828{
829 ssize_t result = 0;
830 size_t requested_bytes = 0;
831 unsigned long seg;
832
833 get_dreq(dreq);
834
835 for (seg = 0; seg < nr_segs; seg++) {
836 const struct iovec *vec = &iov[seg];
Chuck Lever02fe4942007-11-12 12:17:03 -0500837 result = nfs_direct_write_schedule_segment(dreq, vec,
838 pos, sync);
Chuck Lever19f73782007-11-12 12:16:47 -0500839 if (result < 0)
840 break;
841 requested_bytes += result;
842 if ((size_t)result < vec->iov_len)
843 break;
844 pos += vec->iov_len;
845 }
846
Chuck Lever839f7ad2011-01-21 15:54:57 +0000847 /*
848 * If no bytes were started, return the error, and let the
849 * generic layer handle the completion.
850 */
851 if (requested_bytes == 0) {
852 nfs_direct_req_release(dreq);
853 return result < 0 ? result : -EIO;
854 }
855
Chuck Lever19f73782007-11-12 12:16:47 -0500856 if (put_dreq(dreq))
857 nfs_direct_write_complete(dreq, dreq->inode);
Chuck Lever839f7ad2011-01-21 15:54:57 +0000858 return 0;
Chuck Lever19f73782007-11-12 12:16:47 -0500859}
860
Chuck Leverc216fd72007-11-12 12:16:52 -0500861static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
862 unsigned long nr_segs, loff_t pos,
863 size_t count)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864{
Trond Myklebustf11ac8d2010-06-25 16:35:53 -0400865 ssize_t result = -ENOMEM;
Chuck Leverc89f2ee2006-03-20 13:44:33 -0500866 struct inode *inode = iocb->ki_filp->f_mapping->host;
Chuck Lever462d5b32006-03-20 13:44:32 -0500867 struct nfs_direct_req *dreq;
Trond Myklebustfad61492006-03-20 13:44:36 -0500868 size_t wsize = NFS_SERVER(inode)->wsize;
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400869 int sync = NFS_UNSTABLE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870
Trond Myklebust607f31e2006-06-28 16:52:45 -0400871 dreq = nfs_direct_req_alloc();
Chuck Lever462d5b32006-03-20 13:44:32 -0500872 if (!dreq)
Trond Myklebustf11ac8d2010-06-25 16:35:53 -0400873 goto out;
Trond Myklebust607f31e2006-06-28 16:52:45 -0400874 nfs_alloc_commit_data(dreq);
875
Arun Bharadwajb47d19d2010-11-18 10:36:43 +0000876 if (dreq->commit_data == NULL || count <= wsize)
Trond Myklebustbdc7f022007-07-14 15:40:00 -0400877 sync = NFS_FILE_SYNC;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878
Chuck Leverc89f2ee2006-03-20 13:44:33 -0500879 dreq->inode = inode;
Trond Myklebustcd3758e2007-08-10 17:44:32 -0400880 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
Trond Myklebustf11ac8d2010-06-25 16:35:53 -0400881 dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
Steve Dickson568a810d2010-10-28 08:17:54 -0400882 if (dreq->l_ctx == NULL)
Trond Myklebustf11ac8d2010-06-25 16:35:53 -0400883 goto out_release;
Chuck Leverc89f2ee2006-03-20 13:44:33 -0500884 if (!is_sync_kiocb(iocb))
885 dreq->iocb = iocb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886
Chuck Leverc216fd72007-11-12 12:16:52 -0500887 result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync);
Trond Myklebust607f31e2006-06-28 16:52:45 -0400888 if (!result)
889 result = nfs_direct_wait(dreq);
Trond Myklebustf11ac8d2010-06-25 16:35:53 -0400890out_release:
Trond Myklebustb4946ff2007-05-30 12:58:00 -0400891 nfs_direct_req_release(dreq);
Trond Myklebustf11ac8d2010-06-25 16:35:53 -0400892out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 return result;
894}
895
896/**
897 * nfs_file_direct_read - file direct read operation for NFS files
898 * @iocb: target I/O control block
Badari Pulavarty027445c2006-09-30 23:28:46 -0700899 * @iov: vector of user buffers into which to read data
900 * @nr_segs: size of iov vector
Chuck Lever88467052006-03-20 13:44:34 -0500901 * @pos: byte offset in file where reading starts
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 *
903 * We use this function for direct reads instead of calling
904 * generic_file_aio_read() in order to avoid gfar's check to see if
905 * the request starts before the end of the file. For that check
906 * to work, we must generate a GETATTR before each direct read, and
907 * even then there is a window between the GETATTR and the subsequent
Chuck Lever88467052006-03-20 13:44:34 -0500908 * READ where the file size could change. Our preference is simply
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 * to do all reads the application wants, and the server will take
910 * care of managing the end of file boundary.
Chuck Lever88467052006-03-20 13:44:34 -0500911 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912 * This function also eliminates unnecessarily updating the file's
913 * atime locally, as the NFS server sets the file's atime, and this
914 * client must read the updated atime from the server back into its
915 * cache.
916 */
Badari Pulavarty027445c2006-09-30 23:28:46 -0700917ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
918 unsigned long nr_segs, loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919{
920 ssize_t retval = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 struct file *file = iocb->ki_filp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 struct address_space *mapping = file->f_mapping;
Chuck Leverc216fd72007-11-12 12:16:52 -0500923 size_t count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924
Chuck Leverc216fd72007-11-12 12:16:52 -0500925 count = iov_length(iov, nr_segs);
926 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
927
Chuck Lever6da24bc2008-06-11 17:55:58 -0400928 dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n",
Josef "Jeff" Sipek01cce932006-12-08 02:36:40 -0800929 file->f_path.dentry->d_parent->d_name.name,
930 file->f_path.dentry->d_name.name,
Chuck Leverc216fd72007-11-12 12:16:52 -0500931 count, (long long) pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933 retval = 0;
934 if (!count)
935 goto out;
936
Trond Myklebust29884df2005-12-13 16:13:54 -0500937 retval = nfs_sync_mapping(mapping);
938 if (retval)
939 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940
Konstantin Khlebnikov7ec10f22011-02-22 00:28:34 +0300941 task_io_account_read(count);
942
Chuck Leverc216fd72007-11-12 12:16:52 -0500943 retval = nfs_direct_read(iocb, iov, nr_segs, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944 if (retval > 0)
Chuck Lever0cdd80d2006-03-20 13:44:29 -0500945 iocb->ki_pos = pos + retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946
947out:
948 return retval;
949}
950
951/**
952 * nfs_file_direct_write - file direct write operation for NFS files
953 * @iocb: target I/O control block
Badari Pulavarty027445c2006-09-30 23:28:46 -0700954 * @iov: vector of user buffers from which to write data
955 * @nr_segs: size of iov vector
Chuck Lever88467052006-03-20 13:44:34 -0500956 * @pos: byte offset in file where writing starts
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957 *
958 * We use this function for direct writes instead of calling
959 * generic_file_aio_write() in order to avoid taking the inode
960 * semaphore and updating the i_size. The NFS server will set
961 * the new i_size and this client must read the updated size
962 * back into its cache. We let the server do generic write
963 * parameter checking and report problems.
964 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 * We eliminate local atime updates, see direct read above.
966 *
967 * We avoid unnecessary page cache invalidations for normal cached
968 * readers of this file.
969 *
970 * Note that O_APPEND is not supported for NFS direct writes, as there
971 * is no atomic O_APPEND write facility in the NFS protocol.
972 */
Badari Pulavarty027445c2006-09-30 23:28:46 -0700973ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
974 unsigned long nr_segs, loff_t pos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975{
Chuck Lever070ea602007-05-19 17:22:52 -0400976 ssize_t retval = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 struct file *file = iocb->ki_filp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978 struct address_space *mapping = file->f_mapping;
Chuck Leverc216fd72007-11-12 12:16:52 -0500979 size_t count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980
Chuck Leverc216fd72007-11-12 12:16:52 -0500981 count = iov_length(iov, nr_segs);
982 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
983
Chuck Lever6da24bc2008-06-11 17:55:58 -0400984 dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n",
Josef "Jeff" Sipek01cce932006-12-08 02:36:40 -0800985 file->f_path.dentry->d_parent->d_name.name,
986 file->f_path.dentry->d_name.name,
Chuck Leverc216fd72007-11-12 12:16:52 -0500987 count, (long long) pos);
Badari Pulavarty027445c2006-09-30 23:28:46 -0700988
Chuck Leverce1a8e62005-11-30 18:08:17 -0500989 retval = generic_write_checks(file, &pos, &count, 0);
990 if (retval)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991 goto out;
Chuck Leverce1a8e62005-11-30 18:08:17 -0500992
993 retval = -EINVAL;
994 if ((ssize_t) count < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700995 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 retval = 0;
997 if (!count)
998 goto out;
Chuck Leverce1a8e62005-11-30 18:08:17 -0500999
Trond Myklebust29884df2005-12-13 16:13:54 -05001000 retval = nfs_sync_mapping(mapping);
1001 if (retval)
1002 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003
Konstantin Khlebnikov7ec10f22011-02-22 00:28:34 +03001004 task_io_account_write(count);
1005
Chuck Leverc216fd72007-11-12 12:16:52 -05001006 retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
Chuck Lever9eafa8c2006-03-20 13:44:33 -05001007
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008 if (retval > 0)
Chuck Leverce1a8e62005-11-30 18:08:17 -05001009 iocb->ki_pos = pos + retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010
1011out:
1012 return retval;
1013}
1014
Chuck Lever88467052006-03-20 13:44:34 -05001015/**
1016 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1017 *
1018 */
David Howellsf7b422b2006-06-09 09:34:33 -04001019int __init nfs_init_directcache(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020{
1021 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1022 sizeof(struct nfs_direct_req),
Paul Jacksonfffb60f2006-03-24 03:16:06 -08001023 0, (SLAB_RECLAIM_ACCOUNT|
1024 SLAB_MEM_SPREAD),
Paul Mundt20c2df82007-07-20 10:11:58 +09001025 NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026 if (nfs_direct_cachep == NULL)
1027 return -ENOMEM;
1028
1029 return 0;
1030}
1031
Chuck Lever88467052006-03-20 13:44:34 -05001032/**
David Howellsf7b422b2006-06-09 09:34:33 -04001033 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
Chuck Lever88467052006-03-20 13:44:34 -05001034 *
1035 */
David Brownell266bee82006-06-27 12:59:15 -07001036void nfs_destroy_directcache(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037{
Alexey Dobriyan1a1d92c2006-09-27 01:49:40 -07001038 kmem_cache_destroy(nfs_direct_cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039}