2 * linux/fs/nfs/direct.c
4 * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
6 * High-performance uncached I/O for the Linux NFS client
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.
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
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.
29 * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
30 * help from Andrew Morton.
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 * 04 May 2005 support O_DIRECT with aio --cel
41 #include <linux/config.h>
42 #include <linux/errno.h>
43 #include <linux/sched.h>
44 #include <linux/kernel.h>
45 #include <linux/smp_lock.h>
46 #include <linux/file.h>
47 #include <linux/pagemap.h>
48 #include <linux/kref.h>
50 #include <linux/nfs_fs.h>
51 #include <linux/nfs_page.h>
52 #include <linux/sunrpc/clnt.h>
54 #include <asm/system.h>
55 #include <asm/uaccess.h>
56 #include <asm/atomic.h>
60 #define NFSDBG_FACILITY NFSDBG_VFS
62 static void nfs_free_user_pages(struct page **pages, int npages, int do_dirty);
63 static kmem_cache_t *nfs_direct_cachep;
66 * This represents a set of asynchronous requests that we're waiting on
68 struct nfs_direct_req {
69 struct kref kref; /* release manager */
72 struct list_head list, /* nfs_read/write_data structs */
73 rewrite_list; /* saved nfs_write_data structs */
74 struct file * filp; /* file descriptor */
75 struct kiocb * iocb; /* controlling i/o request */
76 wait_queue_head_t wait; /* wait for i/o completion */
77 struct inode * inode; /* target file of i/o */
78 unsigned long user_addr; /* location of user's buffer */
79 size_t user_count; /* total bytes to move */
80 loff_t pos; /* starting offset in file */
81 struct page ** pages; /* pages in our buffer */
82 unsigned int npages; /* count of pages */
84 /* completion state */
85 spinlock_t lock; /* protect completion state */
86 int outstanding; /* i/os we're waiting for */
87 ssize_t count, /* bytes actually processed */
88 error; /* any reported error */
91 struct nfs_write_data * commit_data; /* special write_data for commits */
93 #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
94 #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
95 struct nfs_writeverf verf; /* unstable write verifier */
98 static void nfs_direct_write_schedule(struct nfs_direct_req *dreq, int sync);
99 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
102 * nfs_direct_IO - NFS address space operation for direct I/O
103 * @rw: direction (read or write)
104 * @iocb: target I/O control block
105 * @iov: array of vectors that define I/O buffer
106 * @pos: offset in file to begin the operation
107 * @nr_segs: size of iovec array
109 * The presence of this routine in the address space ops vector means
110 * the NFS client supports direct I/O. However, we shunt off direct
111 * read and write requests before the VFS gets them, so this method
112 * should never be called.
114 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
116 struct dentry *dentry = iocb->ki_filp->f_dentry;
118 dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
119 dentry->d_name.name, (long long) pos, nr_segs);
124 static inline int nfs_get_user_pages(int rw, unsigned long user_addr, size_t size, struct page ***pages)
126 int result = -ENOMEM;
127 unsigned long page_count;
130 page_count = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT;
131 page_count -= user_addr >> PAGE_SHIFT;
133 array_size = (page_count * sizeof(struct page *));
134 *pages = kmalloc(array_size, GFP_KERNEL);
136 down_read(¤t->mm->mmap_sem);
137 result = get_user_pages(current, current->mm, user_addr,
138 page_count, (rw == READ), 0,
140 up_read(¤t->mm->mmap_sem);
142 * If we got fewer pages than expected from get_user_pages(),
143 * the user buffer runs off the end of a mapping; return EFAULT.
145 if (result >= 0 && result < page_count) {
146 nfs_free_user_pages(*pages, result, 0);
154 static void nfs_free_user_pages(struct page **pages, int npages, int do_dirty)
157 for (i = 0; i < npages; i++) {
158 struct page *page = pages[i];
159 if (do_dirty && !PageCompound(page))
160 set_page_dirty_lock(page);
161 page_cache_release(page);
166 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
168 struct nfs_direct_req *dreq;
170 dreq = kmem_cache_alloc(nfs_direct_cachep, SLAB_KERNEL);
174 kref_init(&dreq->kref);
175 init_waitqueue_head(&dreq->wait);
176 INIT_LIST_HEAD(&dreq->list);
177 INIT_LIST_HEAD(&dreq->rewrite_list);
179 spin_lock_init(&dreq->lock);
180 dreq->outstanding = 0;
188 static void nfs_direct_req_release(struct kref *kref)
190 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
191 kmem_cache_free(nfs_direct_cachep, dreq);
195 * Collects and returns the final error value/byte-count.
197 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
199 ssize_t result = -EIOCBQUEUED;
201 /* Async requests don't wait here */
205 result = wait_event_interruptible(dreq->wait, (dreq->outstanding == 0));
208 result = dreq->error;
210 result = dreq->count;
213 kref_put(&dreq->kref, nfs_direct_req_release);
214 return (ssize_t) result;
218 * We must hold a reference to all the pages in this direct read request
219 * until the RPCs complete. This could be long *after* we are woken up in
220 * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
222 * In addition, synchronous I/O uses a stack-allocated iocb. Thus we
223 * can't trust the iocb is still valid here if this is a synchronous
224 * request. If the waiter is woken prematurely, the iocb is long gone.
226 static void nfs_direct_complete(struct nfs_direct_req *dreq)
228 nfs_free_user_pages(dreq->pages, dreq->npages, 1);
231 long res = (long) dreq->error;
233 res = (long) dreq->count;
234 aio_complete(dreq->iocb, res, 0);
236 wake_up(&dreq->wait);
239 kref_put(&dreq->kref, nfs_direct_req_release);
243 * Note we also set the number of requests we have in the dreq when we are
244 * done. This prevents races with I/O completion so we will always wait
245 * until all requests have been dispatched and completed.
247 static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, size_t rsize)
249 struct list_head *list;
250 struct nfs_direct_req *dreq;
251 unsigned int rpages = (rsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
253 dreq = nfs_direct_req_alloc();
259 struct nfs_read_data *data = nfs_readdata_alloc(rpages);
261 if (unlikely(!data)) {
262 while (!list_empty(list)) {
263 data = list_entry(list->next,
264 struct nfs_read_data, pages);
265 list_del(&data->pages);
266 nfs_readdata_free(data);
268 kref_put(&dreq->kref, nfs_direct_req_release);
272 INIT_LIST_HEAD(&data->pages);
273 list_add(&data->pages, list);
275 data->req = (struct nfs_page *) dreq;
281 kref_get(&dreq->kref);
285 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
287 struct nfs_read_data *data = calldata;
288 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
290 if (nfs_readpage_result(task, data) != 0)
293 spin_lock(&dreq->lock);
295 if (likely(task->tk_status >= 0))
296 dreq->count += data->res.count;
298 dreq->error = task->tk_status;
300 if (--dreq->outstanding) {
301 spin_unlock(&dreq->lock);
305 spin_unlock(&dreq->lock);
306 nfs_direct_complete(dreq);
309 static const struct rpc_call_ops nfs_read_direct_ops = {
310 .rpc_call_done = nfs_direct_read_result,
311 .rpc_release = nfs_readdata_release,
315 * For each nfs_read_data struct that was allocated on the list, dispatch
316 * an NFS READ operation
318 static void nfs_direct_read_schedule(struct nfs_direct_req *dreq)
320 struct file *file = dreq->filp;
321 struct inode *inode = file->f_mapping->host;
322 struct nfs_open_context *ctx = (struct nfs_open_context *)
324 struct list_head *list = &dreq->list;
325 struct page **pages = dreq->pages;
326 size_t count = dreq->user_count;
327 loff_t pos = dreq->pos;
328 size_t rsize = NFS_SERVER(inode)->rsize;
329 unsigned int curpage, pgbase;
332 pgbase = dreq->user_addr & ~PAGE_MASK;
334 struct nfs_read_data *data;
341 data = list_entry(list->next, struct nfs_read_data, pages);
342 list_del_init(&data->pages);
345 data->cred = ctx->cred;
346 data->args.fh = NFS_FH(inode);
347 data->args.context = ctx;
348 data->args.offset = pos;
349 data->args.pgbase = pgbase;
350 data->args.pages = &pages[curpage];
351 data->args.count = bytes;
352 data->res.fattr = &data->fattr;
354 data->res.count = bytes;
356 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
357 &nfs_read_direct_ops, data);
358 NFS_PROTO(inode)->read_setup(data);
360 data->task.tk_cookie = (unsigned long) inode;
363 rpc_execute(&data->task);
366 dfprintk(VFS, "NFS: %4d initiated direct read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
369 (long long)NFS_FILEID(inode),
371 (unsigned long long)data->args.offset);
375 curpage += pgbase >> PAGE_SHIFT;
376 pgbase &= ~PAGE_MASK;
379 } while (count != 0);
382 static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos, struct page **pages, unsigned int nr_pages)
386 struct inode *inode = iocb->ki_filp->f_mapping->host;
387 struct rpc_clnt *clnt = NFS_CLIENT(inode);
388 struct nfs_direct_req *dreq;
390 dreq = nfs_direct_read_alloc(count, NFS_SERVER(inode)->rsize);
394 dreq->user_addr = user_addr;
395 dreq->user_count = count;
398 dreq->npages = nr_pages;
401 dreq->filp = iocb->ki_filp;
402 if (!is_sync_kiocb(iocb))
405 nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
406 rpc_clnt_sigmask(clnt, &oldset);
407 nfs_direct_read_schedule(dreq);
408 result = nfs_direct_wait(dreq);
409 rpc_clnt_sigunmask(clnt, &oldset);
414 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
416 list_splice_init(&dreq->rewrite_list, &dreq->list);
417 while (!list_empty(&dreq->list)) {
418 struct nfs_write_data *data = list_entry(dreq->list.next, struct nfs_write_data, pages);
419 list_del(&data->pages);
420 nfs_writedata_release(data);
424 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
425 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
427 struct list_head *pos;
429 list_splice_init(&dreq->rewrite_list, &dreq->list);
430 list_for_each(pos, &dreq->list)
434 nfs_direct_write_schedule(dreq, FLUSH_STABLE);
437 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
439 struct nfs_write_data *data = calldata;
440 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
442 /* Call the NFS version-specific code */
443 if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
445 if (unlikely(task->tk_status < 0)) {
446 dreq->error = task->tk_status;
447 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
449 if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
450 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
451 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
454 dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
455 nfs_direct_write_complete(dreq, data->inode);
458 static const struct rpc_call_ops nfs_commit_direct_ops = {
459 .rpc_call_done = nfs_direct_commit_result,
460 .rpc_release = nfs_commit_release,
463 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
465 struct file *file = dreq->filp;
466 struct nfs_open_context *ctx = (struct nfs_open_context *)
468 struct nfs_write_data *data = dreq->commit_data;
469 struct rpc_task *task = &data->task;
471 data->inode = dreq->inode;
472 data->cred = ctx->cred;
474 data->args.fh = NFS_FH(data->inode);
475 data->args.offset = dreq->pos;
476 data->args.count = dreq->user_count;
478 data->res.fattr = &data->fattr;
479 data->res.verf = &data->verf;
481 rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
482 &nfs_commit_direct_ops, data);
483 NFS_PROTO(data->inode)->commit_setup(data, 0);
485 data->task.tk_priority = RPC_PRIORITY_NORMAL;
486 data->task.tk_cookie = (unsigned long)data->inode;
487 /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
488 dreq->commit_data = NULL;
490 dprintk("NFS: %5u initiated commit call\n", task->tk_pid);
493 rpc_execute(&data->task);
497 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
499 int flags = dreq->flags;
503 case NFS_ODIRECT_DO_COMMIT:
504 nfs_direct_commit_schedule(dreq);
506 case NFS_ODIRECT_RESCHED_WRITES:
507 nfs_direct_write_reschedule(dreq);
510 nfs_end_data_update(inode);
511 if (dreq->commit_data != NULL)
512 nfs_commit_free(dreq->commit_data);
513 nfs_direct_free_writedata(dreq);
514 nfs_direct_complete(dreq);
518 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
520 dreq->commit_data = nfs_commit_alloc(0);
521 if (dreq->commit_data != NULL)
522 dreq->commit_data->req = (struct nfs_page *) dreq;
525 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
527 dreq->commit_data = NULL;
530 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
532 nfs_end_data_update(inode);
533 nfs_direct_free_writedata(dreq);
534 nfs_direct_complete(dreq);
538 static struct nfs_direct_req *nfs_direct_write_alloc(size_t nbytes, size_t wsize)
540 struct list_head *list;
541 struct nfs_direct_req *dreq;
542 unsigned int wpages = (wsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
544 dreq = nfs_direct_req_alloc();
550 struct nfs_write_data *data = nfs_writedata_alloc(wpages);
552 if (unlikely(!data)) {
553 while (!list_empty(list)) {
554 data = list_entry(list->next,
555 struct nfs_write_data, pages);
556 list_del(&data->pages);
557 nfs_writedata_free(data);
559 kref_put(&dreq->kref, nfs_direct_req_release);
563 INIT_LIST_HEAD(&data->pages);
564 list_add(&data->pages, list);
566 data->req = (struct nfs_page *) dreq;
573 nfs_alloc_commit_data(dreq);
575 kref_get(&dreq->kref);
579 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
581 struct nfs_write_data *data = calldata;
582 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
583 int status = task->tk_status;
585 if (nfs_writeback_done(task, data) != 0)
588 spin_lock(&dreq->lock);
590 if (likely(status >= 0))
591 dreq->count += data->res.count;
593 dreq->error = task->tk_status;
595 if (data->res.verf->committed != NFS_FILE_SYNC) {
596 switch (dreq->flags) {
598 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
599 dreq->flags = NFS_ODIRECT_DO_COMMIT;
601 case NFS_ODIRECT_DO_COMMIT:
602 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
603 dprintk("NFS: %5u write verify failed\n", task->tk_pid);
604 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
608 /* In case we have to resend */
609 data->args.stable = NFS_FILE_SYNC;
611 spin_unlock(&dreq->lock);
615 * NB: Return the value of the first error return code. Subsequent
616 * errors after the first one are ignored.
618 static void nfs_direct_write_release(void *calldata)
620 struct nfs_write_data *data = calldata;
621 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
623 spin_lock(&dreq->lock);
624 if (--dreq->outstanding) {
625 spin_unlock(&dreq->lock);
628 spin_unlock(&dreq->lock);
630 nfs_direct_write_complete(dreq, data->inode);
633 static const struct rpc_call_ops nfs_write_direct_ops = {
634 .rpc_call_done = nfs_direct_write_result,
635 .rpc_release = nfs_direct_write_release,
639 * For each nfs_write_data struct that was allocated on the list, dispatch
640 * an NFS WRITE operation
642 static void nfs_direct_write_schedule(struct nfs_direct_req *dreq, int sync)
644 struct file *file = dreq->filp;
645 struct inode *inode = file->f_mapping->host;
646 struct nfs_open_context *ctx = (struct nfs_open_context *)
648 struct list_head *list = &dreq->list;
649 struct page **pages = dreq->pages;
650 size_t count = dreq->user_count;
651 loff_t pos = dreq->pos;
652 size_t wsize = NFS_SERVER(inode)->wsize;
653 unsigned int curpage, pgbase;
656 pgbase = dreq->user_addr & ~PAGE_MASK;
658 struct nfs_write_data *data;
665 data = list_entry(list->next, struct nfs_write_data, pages);
666 list_move_tail(&data->pages, &dreq->rewrite_list);
669 data->cred = ctx->cred;
670 data->args.fh = NFS_FH(inode);
671 data->args.context = ctx;
672 data->args.offset = pos;
673 data->args.pgbase = pgbase;
674 data->args.pages = &pages[curpage];
675 data->args.count = bytes;
676 data->res.fattr = &data->fattr;
677 data->res.count = bytes;
678 data->res.verf = &data->verf;
680 rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
681 &nfs_write_direct_ops, data);
682 NFS_PROTO(inode)->write_setup(data, sync);
684 data->task.tk_priority = RPC_PRIORITY_NORMAL;
685 data->task.tk_cookie = (unsigned long) inode;
688 rpc_execute(&data->task);
691 dfprintk(VFS, "NFS: %4d initiated direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
694 (long long)NFS_FILEID(inode),
696 (unsigned long long)data->args.offset);
700 curpage += pgbase >> PAGE_SHIFT;
701 pgbase &= ~PAGE_MASK;
704 } while (count != 0);
707 static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos, struct page **pages, int nr_pages)
711 struct inode *inode = iocb->ki_filp->f_mapping->host;
712 struct rpc_clnt *clnt = NFS_CLIENT(inode);
713 struct nfs_direct_req *dreq;
714 size_t wsize = NFS_SERVER(inode)->wsize;
717 dreq = nfs_direct_write_alloc(count, wsize);
720 if (dreq->commit_data == NULL || count < wsize)
723 dreq->user_addr = user_addr;
724 dreq->user_count = count;
727 dreq->npages = nr_pages;
730 dreq->filp = iocb->ki_filp;
731 if (!is_sync_kiocb(iocb))
734 nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
736 nfs_begin_data_update(inode);
738 rpc_clnt_sigmask(clnt, &oldset);
739 nfs_direct_write_schedule(dreq, sync);
740 result = nfs_direct_wait(dreq);
741 rpc_clnt_sigunmask(clnt, &oldset);
747 * nfs_file_direct_read - file direct read operation for NFS files
748 * @iocb: target I/O control block
749 * @buf: user's buffer into which to read data
750 * @count: number of bytes to read
751 * @pos: byte offset in file where reading starts
753 * We use this function for direct reads instead of calling
754 * generic_file_aio_read() in order to avoid gfar's check to see if
755 * the request starts before the end of the file. For that check
756 * to work, we must generate a GETATTR before each direct read, and
757 * even then there is a window between the GETATTR and the subsequent
758 * READ where the file size could change. Our preference is simply
759 * to do all reads the application wants, and the server will take
760 * care of managing the end of file boundary.
762 * This function also eliminates unnecessarily updating the file's
763 * atime locally, as the NFS server sets the file's atime, and this
764 * client must read the updated atime from the server back into its
767 ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
769 ssize_t retval = -EINVAL;
772 struct file *file = iocb->ki_filp;
773 struct address_space *mapping = file->f_mapping;
775 dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
776 file->f_dentry->d_parent->d_name.name,
777 file->f_dentry->d_name.name,
778 (unsigned long) count, (long long) pos);
783 if (!access_ok(VERIFY_WRITE, buf, count))
789 retval = nfs_sync_mapping(mapping);
793 page_count = nfs_get_user_pages(READ, (unsigned long) buf,
795 if (page_count < 0) {
796 nfs_free_user_pages(pages, 0, 0);
801 retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos,
804 iocb->ki_pos = pos + retval;
811 * nfs_file_direct_write - file direct write operation for NFS files
812 * @iocb: target I/O control block
813 * @buf: user's buffer from which to write data
814 * @count: number of bytes to write
815 * @pos: byte offset in file where writing starts
817 * We use this function for direct writes instead of calling
818 * generic_file_aio_write() in order to avoid taking the inode
819 * semaphore and updating the i_size. The NFS server will set
820 * the new i_size and this client must read the updated size
821 * back into its cache. We let the server do generic write
822 * parameter checking and report problems.
824 * We also avoid an unnecessary invocation of generic_osync_inode(),
825 * as it is fairly meaningless to sync the metadata of an NFS file.
827 * We eliminate local atime updates, see direct read above.
829 * We avoid unnecessary page cache invalidations for normal cached
830 * readers of this file.
832 * Note that O_APPEND is not supported for NFS direct writes, as there
833 * is no atomic O_APPEND write facility in the NFS protocol.
835 ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
840 struct file *file = iocb->ki_filp;
841 struct address_space *mapping = file->f_mapping;
843 dfprintk(VFS, "nfs: direct write(%s/%s, %lu@%Ld)\n",
844 file->f_dentry->d_parent->d_name.name,
845 file->f_dentry->d_name.name,
846 (unsigned long) count, (long long) pos);
848 retval = generic_write_checks(file, &pos, &count, 0);
853 if ((ssize_t) count < 0)
860 if (!access_ok(VERIFY_READ, buf, count))
863 retval = nfs_sync_mapping(mapping);
867 page_count = nfs_get_user_pages(WRITE, (unsigned long) buf,
869 if (page_count < 0) {
870 nfs_free_user_pages(pages, 0, 0);
875 retval = nfs_direct_write(iocb, (unsigned long) buf, count,
876 pos, pages, page_count);
879 * XXX: nfs_end_data_update() already ensures this file's
880 * cached data is subsequently invalidated. Do we really
881 * need to call invalidate_inode_pages2() again here?
883 * For aio writes, this invalidation will almost certainly
884 * occur before the writes complete. Kind of racey.
886 if (mapping->nrpages)
887 invalidate_inode_pages2(mapping);
890 iocb->ki_pos = pos + retval;
897 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
900 int nfs_init_directcache(void)
902 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
903 sizeof(struct nfs_direct_req),
904 0, SLAB_RECLAIM_ACCOUNT,
906 if (nfs_direct_cachep == NULL)
913 * nfs_init_directcache - destroy the slab cache for nfs_direct_req structures
916 void nfs_destroy_directcache(void)
918 if (kmem_cache_destroy(nfs_direct_cachep))
919 printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n");