blob: ae2be0744191089984d26a206193bea4917a7ae8 [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
10 * (multiple copies of the same instance running on separate hosts)
11 * implement their own cache coherency protocol that subsumes file
12 * 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
15 * 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
37 *
38 */
39
40#include <linux/config.h>
41#include <linux/errno.h>
42#include <linux/sched.h>
43#include <linux/kernel.h>
44#include <linux/smp_lock.h>
45#include <linux/file.h>
46#include <linux/pagemap.h>
47#include <linux/kref.h>
48
49#include <linux/nfs_fs.h>
50#include <linux/nfs_page.h>
51#include <linux/sunrpc/clnt.h>
52
53#include <asm/system.h>
54#include <asm/uaccess.h>
55#include <asm/atomic.h>
56
57#define NFSDBG_FACILITY NFSDBG_VFS
58#define MAX_DIRECTIO_SIZE (4096UL << PAGE_SHIFT)
59
60static kmem_cache_t *nfs_direct_cachep;
61
62/*
63 * This represents a set of asynchronous requests that we're waiting on
64 */
65struct nfs_direct_req {
66 struct kref kref; /* release manager */
67 struct list_head list; /* nfs_read_data structs */
68 wait_queue_head_t wait; /* wait for i/o completion */
69 struct page ** pages; /* pages in our buffer */
70 unsigned int npages; /* count of pages */
71 atomic_t complete, /* i/os we're waiting for */
72 count, /* bytes actually processed */
73 error; /* any reported error */
74};
75
76
77/**
78 * nfs_get_user_pages - find and set up pages underlying user's buffer
79 * rw: direction (read or write)
80 * user_addr: starting address of this segment of user's buffer
81 * count: size of this segment
82 * @pages: returned array of page struct pointers underlying user's buffer
83 */
84static inline int
85nfs_get_user_pages(int rw, unsigned long user_addr, size_t size,
86 struct page ***pages)
87{
88 int result = -ENOMEM;
89 unsigned long page_count;
90 size_t array_size;
91
92 /* set an arbitrary limit to prevent type overflow */
93 /* XXX: this can probably be as large as INT_MAX */
94 if (size > MAX_DIRECTIO_SIZE) {
95 *pages = NULL;
96 return -EFBIG;
97 }
98
99 page_count = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT;
100 page_count -= user_addr >> PAGE_SHIFT;
101
102 array_size = (page_count * sizeof(struct page *));
103 *pages = kmalloc(array_size, GFP_KERNEL);
104 if (*pages) {
105 down_read(&current->mm->mmap_sem);
106 result = get_user_pages(current, current->mm, user_addr,
107 page_count, (rw == READ), 0,
108 *pages, NULL);
109 up_read(&current->mm->mmap_sem);
110 }
111 return result;
112}
113
114/**
115 * nfs_free_user_pages - tear down page struct array
116 * @pages: array of page struct pointers underlying target buffer
117 * @npages: number of pages in the array
118 * @do_dirty: dirty the pages as we release them
119 */
120static void
121nfs_free_user_pages(struct page **pages, int npages, int do_dirty)
122{
123 int i;
124 for (i = 0; i < npages; i++) {
125 if (do_dirty)
126 set_page_dirty_lock(pages[i]);
127 page_cache_release(pages[i]);
128 }
129 kfree(pages);
130}
131
132/**
133 * nfs_direct_req_release - release nfs_direct_req structure for direct read
134 * @kref: kref object embedded in an nfs_direct_req structure
135 *
136 */
137static void nfs_direct_req_release(struct kref *kref)
138{
139 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
140 kmem_cache_free(nfs_direct_cachep, dreq);
141}
142
143/**
144 * nfs_direct_read_alloc - allocate nfs_read_data structures for direct read
145 * @count: count of bytes for the read request
146 * @rsize: local rsize setting
147 *
148 * Note we also set the number of requests we have in the dreq when we are
149 * done. This prevents races with I/O completion so we will always wait
150 * until all requests have been dispatched and completed.
151 */
152static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, unsigned int rsize)
153{
154 struct list_head *list;
155 struct nfs_direct_req *dreq;
156 unsigned int reads = 0;
157
158 dreq = kmem_cache_alloc(nfs_direct_cachep, SLAB_KERNEL);
159 if (!dreq)
160 return NULL;
161
162 kref_init(&dreq->kref);
163 init_waitqueue_head(&dreq->wait);
164 INIT_LIST_HEAD(&dreq->list);
165 atomic_set(&dreq->count, 0);
166 atomic_set(&dreq->error, 0);
167
168 list = &dreq->list;
169 for(;;) {
170 struct nfs_read_data *data = nfs_readdata_alloc();
171
172 if (unlikely(!data)) {
173 while (!list_empty(list)) {
174 data = list_entry(list->next,
175 struct nfs_read_data, pages);
176 list_del(&data->pages);
177 nfs_readdata_free(data);
178 }
179 kref_put(&dreq->kref, nfs_direct_req_release);
180 return NULL;
181 }
182
183 INIT_LIST_HEAD(&data->pages);
184 list_add(&data->pages, list);
185
186 data->req = (struct nfs_page *) dreq;
187 reads++;
188 if (nbytes <= rsize)
189 break;
190 nbytes -= rsize;
191 }
192 kref_get(&dreq->kref);
193 atomic_set(&dreq->complete, reads);
194 return dreq;
195}
196
197/**
198 * nfs_direct_read_result - handle a read reply for a direct read request
199 * @data: address of NFS READ operation control block
200 * @status: status of this NFS READ operation
201 *
202 * We must hold a reference to all the pages in this direct read request
203 * until the RPCs complete. This could be long *after* we are woken up in
204 * nfs_direct_read_wait (for instance, if someone hits ^C on a slow server).
205 */
206static void nfs_direct_read_result(struct nfs_read_data *data, int status)
207{
208 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
209
210 if (likely(status >= 0))
211 atomic_add(data->res.count, &dreq->count);
212 else
213 atomic_set(&dreq->error, status);
214
215 if (unlikely(atomic_dec_and_test(&dreq->complete))) {
216 nfs_free_user_pages(dreq->pages, dreq->npages, 1);
217 wake_up(&dreq->wait);
218 kref_put(&dreq->kref, nfs_direct_req_release);
219 }
220}
221
222/**
223 * nfs_direct_read_schedule - dispatch NFS READ operations for a direct read
224 * @dreq: address of nfs_direct_req struct for this request
225 * @inode: target inode
226 * @ctx: target file open context
227 * @user_addr: starting address of this segment of user's buffer
228 * @count: size of this segment
229 * @file_offset: offset in file to begin the operation
230 *
231 * For each nfs_read_data struct that was allocated on the list, dispatch
232 * an NFS READ operation
233 */
234static void nfs_direct_read_schedule(struct nfs_direct_req *dreq,
235 struct inode *inode, struct nfs_open_context *ctx,
236 unsigned long user_addr, size_t count, loff_t file_offset)
237{
238 struct list_head *list = &dreq->list;
239 struct page **pages = dreq->pages;
240 unsigned int curpage, pgbase;
241 unsigned int rsize = NFS_SERVER(inode)->rsize;
242
243 curpage = 0;
244 pgbase = user_addr & ~PAGE_MASK;
245 do {
246 struct nfs_read_data *data;
247 unsigned int bytes;
248
249 bytes = rsize;
250 if (count < rsize)
251 bytes = count;
252
253 data = list_entry(list->next, struct nfs_read_data, pages);
254 list_del_init(&data->pages);
255
256 data->inode = inode;
257 data->cred = ctx->cred;
258 data->args.fh = NFS_FH(inode);
259 data->args.context = ctx;
260 data->args.offset = file_offset;
261 data->args.pgbase = pgbase;
262 data->args.pages = &pages[curpage];
263 data->args.count = bytes;
264 data->res.fattr = &data->fattr;
265 data->res.eof = 0;
266 data->res.count = bytes;
267
268 NFS_PROTO(inode)->read_setup(data);
269
270 data->task.tk_cookie = (unsigned long) inode;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271 data->complete = nfs_direct_read_result;
272
273 lock_kernel();
274 rpc_execute(&data->task);
275 unlock_kernel();
276
277 dfprintk(VFS, "NFS: %4d initiated direct read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
278 data->task.tk_pid,
279 inode->i_sb->s_id,
280 (long long)NFS_FILEID(inode),
281 bytes,
282 (unsigned long long)data->args.offset);
283
284 file_offset += bytes;
285 pgbase += bytes;
286 curpage += pgbase >> PAGE_SHIFT;
287 pgbase &= ~PAGE_MASK;
288
289 count -= bytes;
290 } while (count != 0);
291}
292
293/**
294 * nfs_direct_read_wait - wait for I/O completion for direct reads
295 * @dreq: request on which we are to wait
296 * @intr: whether or not this wait can be interrupted
297 *
298 * Collects and returns the final error value/byte-count.
299 */
300static ssize_t nfs_direct_read_wait(struct nfs_direct_req *dreq, int intr)
301{
302 int result = 0;
303
304 if (intr) {
305 result = wait_event_interruptible(dreq->wait,
306 (atomic_read(&dreq->complete) == 0));
307 } else {
308 wait_event(dreq->wait, (atomic_read(&dreq->complete) == 0));
309 }
310
311 if (!result)
312 result = atomic_read(&dreq->error);
313 if (!result)
314 result = atomic_read(&dreq->count);
315
316 kref_put(&dreq->kref, nfs_direct_req_release);
317 return (ssize_t) result;
318}
319
320/**
321 * nfs_direct_read_seg - Read in one iov segment. Generate separate
322 * read RPCs for each "rsize" bytes.
323 * @inode: target inode
324 * @ctx: target file open context
325 * @user_addr: starting address of this segment of user's buffer
326 * @count: size of this segment
327 * @file_offset: offset in file to begin the operation
328 * @pages: array of addresses of page structs defining user's buffer
329 * @nr_pages: number of pages in the array
330 *
331 */
332static ssize_t nfs_direct_read_seg(struct inode *inode,
333 struct nfs_open_context *ctx, unsigned long user_addr,
334 size_t count, loff_t file_offset, struct page **pages,
335 unsigned int nr_pages)
336{
337 ssize_t result;
338 sigset_t oldset;
339 struct rpc_clnt *clnt = NFS_CLIENT(inode);
340 struct nfs_direct_req *dreq;
341
342 dreq = nfs_direct_read_alloc(count, NFS_SERVER(inode)->rsize);
343 if (!dreq)
344 return -ENOMEM;
345
346 dreq->pages = pages;
347 dreq->npages = nr_pages;
348
349 rpc_clnt_sigmask(clnt, &oldset);
350 nfs_direct_read_schedule(dreq, inode, ctx, user_addr, count,
351 file_offset);
352 result = nfs_direct_read_wait(dreq, clnt->cl_intr);
353 rpc_clnt_sigunmask(clnt, &oldset);
354
355 return result;
356}
357
358/**
359 * nfs_direct_read - For each iov segment, map the user's buffer
360 * then generate read RPCs.
361 * @inode: target inode
362 * @ctx: target file open context
363 * @iov: array of vectors that define I/O buffer
364 * file_offset: offset in file to begin the operation
365 * nr_segs: size of iovec array
366 *
367 * We've already pushed out any non-direct writes so that this read
368 * will see them when we read from the server.
369 */
370static ssize_t
371nfs_direct_read(struct inode *inode, struct nfs_open_context *ctx,
372 const struct iovec *iov, loff_t file_offset,
373 unsigned long nr_segs)
374{
375 ssize_t tot_bytes = 0;
376 unsigned long seg = 0;
377
378 while ((seg < nr_segs) && (tot_bytes >= 0)) {
379 ssize_t result;
380 int page_count;
381 struct page **pages;
382 const struct iovec *vec = &iov[seg++];
383 unsigned long user_addr = (unsigned long) vec->iov_base;
384 size_t size = vec->iov_len;
385
386 page_count = nfs_get_user_pages(READ, user_addr, size, &pages);
387 if (page_count < 0) {
388 nfs_free_user_pages(pages, 0, 0);
389 if (tot_bytes > 0)
390 break;
391 return page_count;
392 }
393
394 result = nfs_direct_read_seg(inode, ctx, user_addr, size,
395 file_offset, pages, page_count);
396
397 if (result <= 0) {
398 if (tot_bytes > 0)
399 break;
400 return result;
401 }
402 tot_bytes += result;
403 file_offset += result;
404 if (result < size)
405 break;
406 }
407
408 return tot_bytes;
409}
410
411/**
412 * nfs_direct_write_seg - Write out one iov segment. Generate separate
413 * write RPCs for each "wsize" bytes, then commit.
414 * @inode: target inode
415 * @ctx: target file open context
416 * user_addr: starting address of this segment of user's buffer
417 * count: size of this segment
418 * file_offset: offset in file to begin the operation
419 * @pages: array of addresses of page structs defining user's buffer
420 * nr_pages: size of pages array
421 */
422static ssize_t nfs_direct_write_seg(struct inode *inode,
423 struct nfs_open_context *ctx, unsigned long user_addr,
424 size_t count, loff_t file_offset, struct page **pages,
425 int nr_pages)
426{
427 const unsigned int wsize = NFS_SERVER(inode)->wsize;
428 size_t request;
429 int curpage, need_commit;
430 ssize_t result, tot_bytes;
431 struct nfs_writeverf first_verf;
432 struct nfs_write_data *wdata;
433
434 wdata = nfs_writedata_alloc();
435 if (!wdata)
436 return -ENOMEM;
437
438 wdata->inode = inode;
439 wdata->cred = ctx->cred;
440 wdata->args.fh = NFS_FH(inode);
441 wdata->args.context = ctx;
442 wdata->args.stable = NFS_UNSTABLE;
443 if (IS_SYNC(inode) || NFS_PROTO(inode)->version == 2 || count <= wsize)
444 wdata->args.stable = NFS_FILE_SYNC;
445 wdata->res.fattr = &wdata->fattr;
446 wdata->res.verf = &wdata->verf;
447
448 nfs_begin_data_update(inode);
449retry:
450 need_commit = 0;
451 tot_bytes = 0;
452 curpage = 0;
453 request = count;
454 wdata->args.pgbase = user_addr & ~PAGE_MASK;
455 wdata->args.offset = file_offset;
456 do {
457 wdata->args.count = request;
458 if (wdata->args.count > wsize)
459 wdata->args.count = wsize;
460 wdata->args.pages = &pages[curpage];
461
462 dprintk("NFS: direct write: c=%u o=%Ld ua=%lu, pb=%u, cp=%u\n",
463 wdata->args.count, (long long) wdata->args.offset,
464 user_addr + tot_bytes, wdata->args.pgbase, curpage);
465
466 lock_kernel();
467 result = NFS_PROTO(inode)->write(wdata);
468 unlock_kernel();
469
470 if (result <= 0) {
471 if (tot_bytes > 0)
472 break;
473 goto out;
474 }
475
476 if (tot_bytes == 0)
477 memcpy(&first_verf.verifier, &wdata->verf.verifier,
478 sizeof(first_verf.verifier));
479 if (wdata->verf.committed != NFS_FILE_SYNC) {
480 need_commit = 1;
481 if (memcmp(&first_verf.verifier, &wdata->verf.verifier,
482 sizeof(first_verf.verifier)));
483 goto sync_retry;
484 }
485
486 tot_bytes += result;
487
488 /* in case of a short write: stop now, let the app recover */
489 if (result < wdata->args.count)
490 break;
491
492 wdata->args.offset += result;
493 wdata->args.pgbase += result;
494 curpage += wdata->args.pgbase >> PAGE_SHIFT;
495 wdata->args.pgbase &= ~PAGE_MASK;
496 request -= result;
497 } while (request != 0);
498
499 /*
500 * Commit data written so far, even in the event of an error
501 */
502 if (need_commit) {
503 wdata->args.count = tot_bytes;
504 wdata->args.offset = file_offset;
505
506 lock_kernel();
507 result = NFS_PROTO(inode)->commit(wdata);
508 unlock_kernel();
509
510 if (result < 0 || memcmp(&first_verf.verifier,
511 &wdata->verf.verifier,
512 sizeof(first_verf.verifier)) != 0)
513 goto sync_retry;
514 }
515 result = tot_bytes;
516
517out:
Trond Myklebust951a1432005-06-22 17:16:30 +0000518 nfs_end_data_update(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 nfs_writedata_free(wdata);
520 return result;
521
522sync_retry:
523 wdata->args.stable = NFS_FILE_SYNC;
524 goto retry;
525}
526
527/**
528 * nfs_direct_write - For each iov segment, map the user's buffer
529 * then generate write and commit RPCs.
530 * @inode: target inode
531 * @ctx: target file open context
532 * @iov: array of vectors that define I/O buffer
533 * file_offset: offset in file to begin the operation
534 * nr_segs: size of iovec array
535 *
536 * Upon return, generic_file_direct_IO invalidates any cached pages
537 * that non-direct readers might access, so they will pick up these
538 * writes immediately.
539 */
540static ssize_t nfs_direct_write(struct inode *inode,
541 struct nfs_open_context *ctx, const struct iovec *iov,
542 loff_t file_offset, unsigned long nr_segs)
543{
544 ssize_t tot_bytes = 0;
545 unsigned long seg = 0;
546
547 while ((seg < nr_segs) && (tot_bytes >= 0)) {
548 ssize_t result;
549 int page_count;
550 struct page **pages;
551 const struct iovec *vec = &iov[seg++];
552 unsigned long user_addr = (unsigned long) vec->iov_base;
553 size_t size = vec->iov_len;
554
555 page_count = nfs_get_user_pages(WRITE, user_addr, size, &pages);
556 if (page_count < 0) {
557 nfs_free_user_pages(pages, 0, 0);
558 if (tot_bytes > 0)
559 break;
560 return page_count;
561 }
562
563 result = nfs_direct_write_seg(inode, ctx, user_addr, size,
564 file_offset, pages, page_count);
565 nfs_free_user_pages(pages, page_count, 0);
566
567 if (result <= 0) {
568 if (tot_bytes > 0)
569 break;
570 return result;
571 }
572 tot_bytes += result;
573 file_offset += result;
574 if (result < size)
575 break;
576 }
577 return tot_bytes;
578}
579
580/**
581 * nfs_direct_IO - NFS address space operation for direct I/O
582 * rw: direction (read or write)
583 * @iocb: target I/O control block
584 * @iov: array of vectors that define I/O buffer
585 * file_offset: offset in file to begin the operation
586 * nr_segs: size of iovec array
587 *
588 */
589ssize_t
590nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
591 loff_t file_offset, unsigned long nr_segs)
592{
593 ssize_t result = -EINVAL;
594 struct file *file = iocb->ki_filp;
595 struct nfs_open_context *ctx;
596 struct dentry *dentry = file->f_dentry;
597 struct inode *inode = dentry->d_inode;
598
599 /*
600 * No support for async yet
601 */
602 if (!is_sync_kiocb(iocb))
603 return result;
604
605 ctx = (struct nfs_open_context *)file->private_data;
606 switch (rw) {
607 case READ:
608 dprintk("NFS: direct_IO(read) (%s) off/no(%Lu/%lu)\n",
609 dentry->d_name.name, file_offset, nr_segs);
610
611 result = nfs_direct_read(inode, ctx, iov,
612 file_offset, nr_segs);
613 break;
614 case WRITE:
615 dprintk("NFS: direct_IO(write) (%s) off/no(%Lu/%lu)\n",
616 dentry->d_name.name, file_offset, nr_segs);
617
618 result = nfs_direct_write(inode, ctx, iov,
619 file_offset, nr_segs);
620 break;
621 default:
622 break;
623 }
624 return result;
625}
626
627/**
628 * nfs_file_direct_read - file direct read operation for NFS files
629 * @iocb: target I/O control block
630 * @buf: user's buffer into which to read data
631 * count: number of bytes to read
632 * pos: byte offset in file where reading starts
633 *
634 * We use this function for direct reads instead of calling
635 * generic_file_aio_read() in order to avoid gfar's check to see if
636 * the request starts before the end of the file. For that check
637 * to work, we must generate a GETATTR before each direct read, and
638 * even then there is a window between the GETATTR and the subsequent
639 * READ where the file size could change. So our preference is simply
640 * to do all reads the application wants, and the server will take
641 * care of managing the end of file boundary.
642 *
643 * This function also eliminates unnecessarily updating the file's
644 * atime locally, as the NFS server sets the file's atime, and this
645 * client must read the updated atime from the server back into its
646 * cache.
647 */
648ssize_t
649nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
650{
651 ssize_t retval = -EINVAL;
652 loff_t *ppos = &iocb->ki_pos;
653 struct file *file = iocb->ki_filp;
654 struct nfs_open_context *ctx =
655 (struct nfs_open_context *) file->private_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656 struct address_space *mapping = file->f_mapping;
657 struct inode *inode = mapping->host;
658 struct iovec iov = {
659 .iov_base = buf,
660 .iov_len = count,
661 };
662
663 dprintk("nfs: direct read(%s/%s, %lu@%lu)\n",
Chuck Lever0bbacc42005-11-01 16:53:32 -0500664 file->f_dentry->d_parent->d_name.name,
665 file->f_dentry->d_name.name,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666 (unsigned long) count, (unsigned long) pos);
667
668 if (!is_sync_kiocb(iocb))
669 goto out;
670 if (count < 0)
671 goto out;
672 retval = -EFAULT;
673 if (!access_ok(VERIFY_WRITE, iov.iov_base, iov.iov_len))
674 goto out;
675 retval = 0;
676 if (!count)
677 goto out;
678
Trond Myklebust29884df2005-12-13 16:13:54 -0500679 retval = nfs_sync_mapping(mapping);
680 if (retval)
681 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682
683 retval = nfs_direct_read(inode, ctx, &iov, pos, 1);
684 if (retval > 0)
685 *ppos = pos + retval;
686
687out:
688 return retval;
689}
690
691/**
692 * nfs_file_direct_write - file direct write operation for NFS files
693 * @iocb: target I/O control block
694 * @buf: user's buffer from which to write data
695 * count: number of bytes to write
696 * pos: byte offset in file where writing starts
697 *
698 * We use this function for direct writes instead of calling
699 * generic_file_aio_write() in order to avoid taking the inode
700 * semaphore and updating the i_size. The NFS server will set
701 * the new i_size and this client must read the updated size
702 * back into its cache. We let the server do generic write
703 * parameter checking and report problems.
704 *
705 * We also avoid an unnecessary invocation of generic_osync_inode(),
706 * as it is fairly meaningless to sync the metadata of an NFS file.
707 *
708 * We eliminate local atime updates, see direct read above.
709 *
710 * We avoid unnecessary page cache invalidations for normal cached
711 * readers of this file.
712 *
713 * Note that O_APPEND is not supported for NFS direct writes, as there
714 * is no atomic O_APPEND write facility in the NFS protocol.
715 */
716ssize_t
717nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
718{
719 ssize_t retval = -EINVAL;
720 loff_t *ppos = &iocb->ki_pos;
721 unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
722 struct file *file = iocb->ki_filp;
723 struct nfs_open_context *ctx =
724 (struct nfs_open_context *) file->private_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 struct address_space *mapping = file->f_mapping;
726 struct inode *inode = mapping->host;
727 struct iovec iov = {
728 .iov_base = (char __user *)buf,
729 .iov_len = count,
730 };
731
732 dfprintk(VFS, "nfs: direct write(%s/%s(%ld), %lu@%lu)\n",
Chuck Lever0bbacc42005-11-01 16:53:32 -0500733 file->f_dentry->d_parent->d_name.name,
734 file->f_dentry->d_name.name, inode->i_ino,
735 (unsigned long) count, (unsigned long) pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736
737 if (!is_sync_kiocb(iocb))
738 goto out;
739 if (count < 0)
740 goto out;
741 if (pos < 0)
742 goto out;
743 retval = -EFAULT;
744 if (!access_ok(VERIFY_READ, iov.iov_base, iov.iov_len))
745 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746 retval = -EFBIG;
747 if (limit != RLIM_INFINITY) {
748 if (pos >= limit) {
749 send_sig(SIGXFSZ, current, 0);
750 goto out;
751 }
752 if (count > limit - (unsigned long) pos)
753 count = limit - (unsigned long) pos;
754 }
755 retval = 0;
756 if (!count)
757 goto out;
758
Trond Myklebust29884df2005-12-13 16:13:54 -0500759 retval = nfs_sync_mapping(mapping);
760 if (retval)
761 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762
763 retval = nfs_direct_write(inode, ctx, &iov, pos, 1);
764 if (mapping->nrpages)
765 invalidate_inode_pages2(mapping);
766 if (retval > 0)
767 *ppos = pos + retval;
768
769out:
770 return retval;
771}
772
773int nfs_init_directcache(void)
774{
775 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
776 sizeof(struct nfs_direct_req),
777 0, SLAB_RECLAIM_ACCOUNT,
778 NULL, NULL);
779 if (nfs_direct_cachep == NULL)
780 return -ENOMEM;
781
782 return 0;
783}
784
785void nfs_destroy_directcache(void)
786{
787 if (kmem_cache_destroy(nfs_direct_cachep))
788 printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n");
789}