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
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>
49 #include <linux/nfs_fs.h>
50 #include <linux/nfs_page.h>
51 #include <linux/sunrpc/clnt.h>
53 #include <asm/system.h>
54 #include <asm/uaccess.h>
55 #include <asm/atomic.h>
57 #define NFSDBG_FACILITY NFSDBG_VFS
58 #define MAX_DIRECTIO_SIZE (4096UL << PAGE_SHIFT)
60 static kmem_cache_t *nfs_direct_cachep;
63 * This represents a set of asynchronous requests that we're waiting on
65 struct 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 */
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
85 nfs_get_user_pages(int rw, unsigned long user_addr, size_t size,
89 unsigned long page_count;
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) {
99 page_count = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT;
100 page_count -= user_addr >> PAGE_SHIFT;
102 array_size = (page_count * sizeof(struct page *));
103 *pages = kmalloc(array_size, GFP_KERNEL);
105 down_read(¤t->mm->mmap_sem);
106 result = get_user_pages(current, current->mm, user_addr,
107 page_count, (rw == READ), 0,
109 up_read(¤t->mm->mmap_sem);
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
121 nfs_free_user_pages(struct page **pages, int npages, int do_dirty)
124 for (i = 0; i < npages; i++) {
126 set_page_dirty_lock(pages[i]);
127 page_cache_release(pages[i]);
133 * nfs_direct_req_release - release nfs_direct_req structure for direct read
134 * @kref: kref object embedded in an nfs_direct_req structure
137 static void nfs_direct_req_release(struct kref *kref)
139 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
140 kmem_cache_free(nfs_direct_cachep, dreq);
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
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.
152 static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, unsigned int rsize)
154 struct list_head *list;
155 struct nfs_direct_req *dreq;
156 unsigned int reads = 0;
158 dreq = kmem_cache_alloc(nfs_direct_cachep, SLAB_KERNEL);
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);
170 struct nfs_read_data *data = nfs_readdata_alloc();
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);
179 kref_put(&dreq->kref, nfs_direct_req_release);
183 INIT_LIST_HEAD(&data->pages);
184 list_add(&data->pages, list);
186 data->req = (struct nfs_page *) dreq;
192 kref_get(&dreq->kref);
193 atomic_set(&dreq->complete, reads);
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
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).
206 static void nfs_direct_read_result(struct nfs_read_data *data, int status)
208 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
210 if (likely(status >= 0))
211 atomic_add(data->res.count, &dreq->count);
213 atomic_set(&dreq->error, status);
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);
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
231 * For each nfs_read_data struct that was allocated on the list, dispatch
232 * an NFS READ operation
234 static 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)
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;
244 pgbase = user_addr & ~PAGE_MASK;
246 struct nfs_read_data *data;
253 data = list_entry(list->next, struct nfs_read_data, pages);
254 list_del_init(&data->pages);
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;
266 data->res.count = bytes;
268 NFS_PROTO(inode)->read_setup(data);
270 data->task.tk_cookie = (unsigned long) inode;
271 data->task.tk_calldata = data;
272 data->complete = nfs_direct_read_result;
275 rpc_execute(&data->task);
278 dfprintk(VFS, "NFS: %4d initiated direct read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
281 (long long)NFS_FILEID(inode),
283 (unsigned long long)data->args.offset);
285 file_offset += bytes;
287 curpage += pgbase >> PAGE_SHIFT;
288 pgbase &= ~PAGE_MASK;
291 } while (count != 0);
295 * nfs_direct_read_wait - wait for I/O completion for direct reads
296 * @dreq: request on which we are to wait
297 * @intr: whether or not this wait can be interrupted
299 * Collects and returns the final error value/byte-count.
301 static ssize_t nfs_direct_read_wait(struct nfs_direct_req *dreq, int intr)
306 result = wait_event_interruptible(dreq->wait,
307 (atomic_read(&dreq->complete) == 0));
309 wait_event(dreq->wait, (atomic_read(&dreq->complete) == 0));
313 result = atomic_read(&dreq->error);
315 result = atomic_read(&dreq->count);
317 kref_put(&dreq->kref, nfs_direct_req_release);
318 return (ssize_t) result;
322 * nfs_direct_read_seg - Read in one iov segment. Generate separate
323 * read RPCs for each "rsize" bytes.
324 * @inode: target inode
325 * @ctx: target file open context
326 * @user_addr: starting address of this segment of user's buffer
327 * @count: size of this segment
328 * @file_offset: offset in file to begin the operation
329 * @pages: array of addresses of page structs defining user's buffer
330 * @nr_pages: number of pages in the array
333 static ssize_t nfs_direct_read_seg(struct inode *inode,
334 struct nfs_open_context *ctx, unsigned long user_addr,
335 size_t count, loff_t file_offset, struct page **pages,
336 unsigned int nr_pages)
340 struct rpc_clnt *clnt = NFS_CLIENT(inode);
341 struct nfs_direct_req *dreq;
343 dreq = nfs_direct_read_alloc(count, NFS_SERVER(inode)->rsize);
348 dreq->npages = nr_pages;
350 rpc_clnt_sigmask(clnt, &oldset);
351 nfs_direct_read_schedule(dreq, inode, ctx, user_addr, count,
353 result = nfs_direct_read_wait(dreq, clnt->cl_intr);
354 rpc_clnt_sigunmask(clnt, &oldset);
360 * nfs_direct_read - For each iov segment, map the user's buffer
361 * then generate read RPCs.
362 * @inode: target inode
363 * @ctx: target file open context
364 * @iov: array of vectors that define I/O buffer
365 * file_offset: offset in file to begin the operation
366 * nr_segs: size of iovec array
368 * We've already pushed out any non-direct writes so that this read
369 * will see them when we read from the server.
372 nfs_direct_read(struct inode *inode, struct nfs_open_context *ctx,
373 const struct iovec *iov, loff_t file_offset,
374 unsigned long nr_segs)
376 ssize_t tot_bytes = 0;
377 unsigned long seg = 0;
379 while ((seg < nr_segs) && (tot_bytes >= 0)) {
383 const struct iovec *vec = &iov[seg++];
384 unsigned long user_addr = (unsigned long) vec->iov_base;
385 size_t size = vec->iov_len;
387 page_count = nfs_get_user_pages(READ, user_addr, size, &pages);
388 if (page_count < 0) {
389 nfs_free_user_pages(pages, 0, 0);
395 result = nfs_direct_read_seg(inode, ctx, user_addr, size,
396 file_offset, pages, page_count);
404 file_offset += result;
413 * nfs_direct_write_seg - Write out one iov segment. Generate separate
414 * write RPCs for each "wsize" bytes, then commit.
415 * @inode: target inode
416 * @ctx: target file open context
417 * user_addr: starting address of this segment of user's buffer
418 * count: size of this segment
419 * file_offset: offset in file to begin the operation
420 * @pages: array of addresses of page structs defining user's buffer
421 * nr_pages: size of pages array
423 static ssize_t nfs_direct_write_seg(struct inode *inode,
424 struct nfs_open_context *ctx, unsigned long user_addr,
425 size_t count, loff_t file_offset, struct page **pages,
428 const unsigned int wsize = NFS_SERVER(inode)->wsize;
430 int curpage, need_commit;
431 ssize_t result, tot_bytes;
432 struct nfs_writeverf first_verf;
433 struct nfs_write_data *wdata;
435 wdata = nfs_writedata_alloc();
439 wdata->inode = inode;
440 wdata->cred = ctx->cred;
441 wdata->args.fh = NFS_FH(inode);
442 wdata->args.context = ctx;
443 wdata->args.stable = NFS_UNSTABLE;
444 if (IS_SYNC(inode) || NFS_PROTO(inode)->version == 2 || count <= wsize)
445 wdata->args.stable = NFS_FILE_SYNC;
446 wdata->res.fattr = &wdata->fattr;
447 wdata->res.verf = &wdata->verf;
449 nfs_begin_data_update(inode);
455 wdata->args.pgbase = user_addr & ~PAGE_MASK;
456 wdata->args.offset = file_offset;
458 wdata->args.count = request;
459 if (wdata->args.count > wsize)
460 wdata->args.count = wsize;
461 wdata->args.pages = &pages[curpage];
463 dprintk("NFS: direct write: c=%u o=%Ld ua=%lu, pb=%u, cp=%u\n",
464 wdata->args.count, (long long) wdata->args.offset,
465 user_addr + tot_bytes, wdata->args.pgbase, curpage);
468 result = NFS_PROTO(inode)->write(wdata);
478 memcpy(&first_verf.verifier, &wdata->verf.verifier,
479 sizeof(first_verf.verifier));
480 if (wdata->verf.committed != NFS_FILE_SYNC) {
482 if (memcmp(&first_verf.verifier, &wdata->verf.verifier,
483 sizeof(first_verf.verifier)));
489 /* in case of a short write: stop now, let the app recover */
490 if (result < wdata->args.count)
493 wdata->args.offset += result;
494 wdata->args.pgbase += result;
495 curpage += wdata->args.pgbase >> PAGE_SHIFT;
496 wdata->args.pgbase &= ~PAGE_MASK;
498 } while (request != 0);
501 * Commit data written so far, even in the event of an error
504 wdata->args.count = tot_bytes;
505 wdata->args.offset = file_offset;
508 result = NFS_PROTO(inode)->commit(wdata);
511 if (result < 0 || memcmp(&first_verf.verifier,
512 &wdata->verf.verifier,
513 sizeof(first_verf.verifier)) != 0)
519 nfs_end_data_update(inode);
520 nfs_writedata_free(wdata);
524 wdata->args.stable = NFS_FILE_SYNC;
529 * nfs_direct_write - For each iov segment, map the user's buffer
530 * then generate write and commit RPCs.
531 * @inode: target inode
532 * @ctx: target file open context
533 * @iov: array of vectors that define I/O buffer
534 * file_offset: offset in file to begin the operation
535 * nr_segs: size of iovec array
537 * Upon return, generic_file_direct_IO invalidates any cached pages
538 * that non-direct readers might access, so they will pick up these
539 * writes immediately.
541 static ssize_t nfs_direct_write(struct inode *inode,
542 struct nfs_open_context *ctx, const struct iovec *iov,
543 loff_t file_offset, unsigned long nr_segs)
545 ssize_t tot_bytes = 0;
546 unsigned long seg = 0;
548 while ((seg < nr_segs) && (tot_bytes >= 0)) {
552 const struct iovec *vec = &iov[seg++];
553 unsigned long user_addr = (unsigned long) vec->iov_base;
554 size_t size = vec->iov_len;
556 page_count = nfs_get_user_pages(WRITE, user_addr, size, &pages);
557 if (page_count < 0) {
558 nfs_free_user_pages(pages, 0, 0);
564 result = nfs_direct_write_seg(inode, ctx, user_addr, size,
565 file_offset, pages, page_count);
566 nfs_free_user_pages(pages, page_count, 0);
574 file_offset += result;
582 * nfs_direct_IO - NFS address space operation for direct I/O
583 * rw: direction (read or write)
584 * @iocb: target I/O control block
585 * @iov: array of vectors that define I/O buffer
586 * file_offset: offset in file to begin the operation
587 * nr_segs: size of iovec array
591 nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
592 loff_t file_offset, unsigned long nr_segs)
594 ssize_t result = -EINVAL;
595 struct file *file = iocb->ki_filp;
596 struct nfs_open_context *ctx;
597 struct dentry *dentry = file->f_dentry;
598 struct inode *inode = dentry->d_inode;
601 * No support for async yet
603 if (!is_sync_kiocb(iocb))
606 ctx = (struct nfs_open_context *)file->private_data;
609 dprintk("NFS: direct_IO(read) (%s) off/no(%Lu/%lu)\n",
610 dentry->d_name.name, file_offset, nr_segs);
612 result = nfs_direct_read(inode, ctx, iov,
613 file_offset, nr_segs);
616 dprintk("NFS: direct_IO(write) (%s) off/no(%Lu/%lu)\n",
617 dentry->d_name.name, file_offset, nr_segs);
619 result = nfs_direct_write(inode, ctx, iov,
620 file_offset, nr_segs);
629 * nfs_file_direct_read - file direct read operation for NFS files
630 * @iocb: target I/O control block
631 * @buf: user's buffer into which to read data
632 * count: number of bytes to read
633 * pos: byte offset in file where reading starts
635 * We use this function for direct reads instead of calling
636 * generic_file_aio_read() in order to avoid gfar's check to see if
637 * the request starts before the end of the file. For that check
638 * to work, we must generate a GETATTR before each direct read, and
639 * even then there is a window between the GETATTR and the subsequent
640 * READ where the file size could change. So our preference is simply
641 * to do all reads the application wants, and the server will take
642 * care of managing the end of file boundary.
644 * This function also eliminates unnecessarily updating the file's
645 * atime locally, as the NFS server sets the file's atime, and this
646 * client must read the updated atime from the server back into its
650 nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
652 ssize_t retval = -EINVAL;
653 loff_t *ppos = &iocb->ki_pos;
654 struct file *file = iocb->ki_filp;
655 struct nfs_open_context *ctx =
656 (struct nfs_open_context *) file->private_data;
657 struct address_space *mapping = file->f_mapping;
658 struct inode *inode = mapping->host;
664 dprintk("nfs: direct read(%s/%s, %lu@%lu)\n",
665 file->f_dentry->d_parent->d_name.name,
666 file->f_dentry->d_name.name,
667 (unsigned long) count, (unsigned long) pos);
669 if (!is_sync_kiocb(iocb))
674 if (!access_ok(VERIFY_WRITE, iov.iov_base, iov.iov_len))
680 retval = nfs_sync_mapping(mapping);
684 retval = nfs_direct_read(inode, ctx, &iov, pos, 1);
686 *ppos = pos + retval;
693 * nfs_file_direct_write - file direct write operation for NFS files
694 * @iocb: target I/O control block
695 * @buf: user's buffer from which to write data
696 * count: number of bytes to write
697 * pos: byte offset in file where writing starts
699 * We use this function for direct writes instead of calling
700 * generic_file_aio_write() in order to avoid taking the inode
701 * semaphore and updating the i_size. The NFS server will set
702 * the new i_size and this client must read the updated size
703 * back into its cache. We let the server do generic write
704 * parameter checking and report problems.
706 * We also avoid an unnecessary invocation of generic_osync_inode(),
707 * as it is fairly meaningless to sync the metadata of an NFS file.
709 * We eliminate local atime updates, see direct read above.
711 * We avoid unnecessary page cache invalidations for normal cached
712 * readers of this file.
714 * Note that O_APPEND is not supported for NFS direct writes, as there
715 * is no atomic O_APPEND write facility in the NFS protocol.
718 nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
720 ssize_t retval = -EINVAL;
721 loff_t *ppos = &iocb->ki_pos;
722 unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
723 struct file *file = iocb->ki_filp;
724 struct nfs_open_context *ctx =
725 (struct nfs_open_context *) file->private_data;
726 struct address_space *mapping = file->f_mapping;
727 struct inode *inode = mapping->host;
729 .iov_base = (char __user *)buf,
733 dfprintk(VFS, "nfs: direct write(%s/%s(%ld), %lu@%lu)\n",
734 file->f_dentry->d_parent->d_name.name,
735 file->f_dentry->d_name.name, inode->i_ino,
736 (unsigned long) count, (unsigned long) pos);
738 if (!is_sync_kiocb(iocb))
745 if (!access_ok(VERIFY_READ, iov.iov_base, iov.iov_len))
748 if (limit != RLIM_INFINITY) {
750 send_sig(SIGXFSZ, current, 0);
753 if (count > limit - (unsigned long) pos)
754 count = limit - (unsigned long) pos;
760 retval = nfs_sync_mapping(mapping);
764 retval = nfs_direct_write(inode, ctx, &iov, pos, 1);
765 if (mapping->nrpages)
766 invalidate_inode_pages2(mapping);
768 *ppos = pos + retval;
774 int nfs_init_directcache(void)
776 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
777 sizeof(struct nfs_direct_req),
778 0, SLAB_RECLAIM_ACCOUNT,
780 if (nfs_direct_cachep == NULL)
786 void nfs_destroy_directcache(void)
788 if (kmem_cache_destroy(nfs_direct_cachep))
789 printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n");