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->task.tk_release = nfs_readdata_release;
273 data->complete = nfs_direct_read_result;
276 rpc_execute(&data->task);
279 dfprintk(VFS, "NFS: %4d initiated direct read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
282 (long long)NFS_FILEID(inode),
284 (unsigned long long)data->args.offset);
286 file_offset += bytes;
288 curpage += pgbase >> PAGE_SHIFT;
289 pgbase &= ~PAGE_MASK;
292 } while (count != 0);
296 * nfs_direct_read_wait - wait for I/O completion for direct reads
297 * @dreq: request on which we are to wait
298 * @intr: whether or not this wait can be interrupted
300 * Collects and returns the final error value/byte-count.
302 static ssize_t nfs_direct_read_wait(struct nfs_direct_req *dreq, int intr)
307 result = wait_event_interruptible(dreq->wait,
308 (atomic_read(&dreq->complete) == 0));
310 wait_event(dreq->wait, (atomic_read(&dreq->complete) == 0));
314 result = atomic_read(&dreq->error);
316 result = atomic_read(&dreq->count);
318 kref_put(&dreq->kref, nfs_direct_req_release);
319 return (ssize_t) result;
323 * nfs_direct_read_seg - Read in one iov segment. Generate separate
324 * read RPCs for each "rsize" bytes.
325 * @inode: target inode
326 * @ctx: target file open context
327 * @user_addr: starting address of this segment of user's buffer
328 * @count: size of this segment
329 * @file_offset: offset in file to begin the operation
330 * @pages: array of addresses of page structs defining user's buffer
331 * @nr_pages: number of pages in the array
334 static ssize_t nfs_direct_read_seg(struct inode *inode,
335 struct nfs_open_context *ctx, unsigned long user_addr,
336 size_t count, loff_t file_offset, struct page **pages,
337 unsigned int nr_pages)
341 struct rpc_clnt *clnt = NFS_CLIENT(inode);
342 struct nfs_direct_req *dreq;
344 dreq = nfs_direct_read_alloc(count, NFS_SERVER(inode)->rsize);
349 dreq->npages = nr_pages;
351 rpc_clnt_sigmask(clnt, &oldset);
352 nfs_direct_read_schedule(dreq, inode, ctx, user_addr, count,
354 result = nfs_direct_read_wait(dreq, clnt->cl_intr);
355 rpc_clnt_sigunmask(clnt, &oldset);
361 * nfs_direct_read - For each iov segment, map the user's buffer
362 * then generate read RPCs.
363 * @inode: target inode
364 * @ctx: target file open context
365 * @iov: array of vectors that define I/O buffer
366 * file_offset: offset in file to begin the operation
367 * nr_segs: size of iovec array
369 * We've already pushed out any non-direct writes so that this read
370 * will see them when we read from the server.
373 nfs_direct_read(struct inode *inode, struct nfs_open_context *ctx,
374 const struct iovec *iov, loff_t file_offset,
375 unsigned long nr_segs)
377 ssize_t tot_bytes = 0;
378 unsigned long seg = 0;
380 while ((seg < nr_segs) && (tot_bytes >= 0)) {
384 const struct iovec *vec = &iov[seg++];
385 unsigned long user_addr = (unsigned long) vec->iov_base;
386 size_t size = vec->iov_len;
388 page_count = nfs_get_user_pages(READ, user_addr, size, &pages);
389 if (page_count < 0) {
390 nfs_free_user_pages(pages, 0, 0);
396 result = nfs_direct_read_seg(inode, ctx, user_addr, size,
397 file_offset, pages, page_count);
405 file_offset += result;
414 * nfs_direct_write_seg - Write out one iov segment. Generate separate
415 * write RPCs for each "wsize" bytes, then commit.
416 * @inode: target inode
417 * @ctx: target file open context
418 * user_addr: starting address of this segment of user's buffer
419 * count: size of this segment
420 * file_offset: offset in file to begin the operation
421 * @pages: array of addresses of page structs defining user's buffer
422 * nr_pages: size of pages array
424 static ssize_t nfs_direct_write_seg(struct inode *inode,
425 struct nfs_open_context *ctx, unsigned long user_addr,
426 size_t count, loff_t file_offset, struct page **pages,
429 const unsigned int wsize = NFS_SERVER(inode)->wsize;
431 int curpage, need_commit;
432 ssize_t result, tot_bytes;
433 struct nfs_writeverf first_verf;
434 struct nfs_write_data *wdata;
436 wdata = nfs_writedata_alloc();
440 wdata->inode = inode;
441 wdata->cred = ctx->cred;
442 wdata->args.fh = NFS_FH(inode);
443 wdata->args.context = ctx;
444 wdata->args.stable = NFS_UNSTABLE;
445 if (IS_SYNC(inode) || NFS_PROTO(inode)->version == 2 || count <= wsize)
446 wdata->args.stable = NFS_FILE_SYNC;
447 wdata->res.fattr = &wdata->fattr;
448 wdata->res.verf = &wdata->verf;
450 nfs_begin_data_update(inode);
456 wdata->args.pgbase = user_addr & ~PAGE_MASK;
457 wdata->args.offset = file_offset;
459 wdata->args.count = request;
460 if (wdata->args.count > wsize)
461 wdata->args.count = wsize;
462 wdata->args.pages = &pages[curpage];
464 dprintk("NFS: direct write: c=%u o=%Ld ua=%lu, pb=%u, cp=%u\n",
465 wdata->args.count, (long long) wdata->args.offset,
466 user_addr + tot_bytes, wdata->args.pgbase, curpage);
469 result = NFS_PROTO(inode)->write(wdata);
479 memcpy(&first_verf.verifier, &wdata->verf.verifier,
480 sizeof(first_verf.verifier));
481 if (wdata->verf.committed != NFS_FILE_SYNC) {
483 if (memcmp(&first_verf.verifier, &wdata->verf.verifier,
484 sizeof(first_verf.verifier)));
490 /* in case of a short write: stop now, let the app recover */
491 if (result < wdata->args.count)
494 wdata->args.offset += result;
495 wdata->args.pgbase += result;
496 curpage += wdata->args.pgbase >> PAGE_SHIFT;
497 wdata->args.pgbase &= ~PAGE_MASK;
499 } while (request != 0);
502 * Commit data written so far, even in the event of an error
505 wdata->args.count = tot_bytes;
506 wdata->args.offset = file_offset;
509 result = NFS_PROTO(inode)->commit(wdata);
512 if (result < 0 || memcmp(&first_verf.verifier,
513 &wdata->verf.verifier,
514 sizeof(first_verf.verifier)) != 0)
520 nfs_end_data_update(inode);
521 nfs_writedata_free(wdata);
525 wdata->args.stable = NFS_FILE_SYNC;
530 * nfs_direct_write - For each iov segment, map the user's buffer
531 * then generate write and commit RPCs.
532 * @inode: target inode
533 * @ctx: target file open context
534 * @iov: array of vectors that define I/O buffer
535 * file_offset: offset in file to begin the operation
536 * nr_segs: size of iovec array
538 * Upon return, generic_file_direct_IO invalidates any cached pages
539 * that non-direct readers might access, so they will pick up these
540 * writes immediately.
542 static ssize_t nfs_direct_write(struct inode *inode,
543 struct nfs_open_context *ctx, const struct iovec *iov,
544 loff_t file_offset, unsigned long nr_segs)
546 ssize_t tot_bytes = 0;
547 unsigned long seg = 0;
549 while ((seg < nr_segs) && (tot_bytes >= 0)) {
553 const struct iovec *vec = &iov[seg++];
554 unsigned long user_addr = (unsigned long) vec->iov_base;
555 size_t size = vec->iov_len;
557 page_count = nfs_get_user_pages(WRITE, user_addr, size, &pages);
558 if (page_count < 0) {
559 nfs_free_user_pages(pages, 0, 0);
565 result = nfs_direct_write_seg(inode, ctx, user_addr, size,
566 file_offset, pages, page_count);
567 nfs_free_user_pages(pages, page_count, 0);
575 file_offset += result;
583 * nfs_direct_IO - NFS address space operation for direct I/O
584 * rw: direction (read or write)
585 * @iocb: target I/O control block
586 * @iov: array of vectors that define I/O buffer
587 * file_offset: offset in file to begin the operation
588 * nr_segs: size of iovec array
592 nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
593 loff_t file_offset, unsigned long nr_segs)
595 ssize_t result = -EINVAL;
596 struct file *file = iocb->ki_filp;
597 struct nfs_open_context *ctx;
598 struct dentry *dentry = file->f_dentry;
599 struct inode *inode = dentry->d_inode;
602 * No support for async yet
604 if (!is_sync_kiocb(iocb))
607 ctx = (struct nfs_open_context *)file->private_data;
610 dprintk("NFS: direct_IO(read) (%s) off/no(%Lu/%lu)\n",
611 dentry->d_name.name, file_offset, nr_segs);
613 result = nfs_direct_read(inode, ctx, iov,
614 file_offset, nr_segs);
617 dprintk("NFS: direct_IO(write) (%s) off/no(%Lu/%lu)\n",
618 dentry->d_name.name, file_offset, nr_segs);
620 result = nfs_direct_write(inode, ctx, iov,
621 file_offset, nr_segs);
630 * nfs_file_direct_read - file direct read operation for NFS files
631 * @iocb: target I/O control block
632 * @buf: user's buffer into which to read data
633 * count: number of bytes to read
634 * pos: byte offset in file where reading starts
636 * We use this function for direct reads instead of calling
637 * generic_file_aio_read() in order to avoid gfar's check to see if
638 * the request starts before the end of the file. For that check
639 * to work, we must generate a GETATTR before each direct read, and
640 * even then there is a window between the GETATTR and the subsequent
641 * READ where the file size could change. So our preference is simply
642 * to do all reads the application wants, and the server will take
643 * care of managing the end of file boundary.
645 * This function also eliminates unnecessarily updating the file's
646 * atime locally, as the NFS server sets the file's atime, and this
647 * client must read the updated atime from the server back into its
651 nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
653 ssize_t retval = -EINVAL;
654 loff_t *ppos = &iocb->ki_pos;
655 struct file *file = iocb->ki_filp;
656 struct nfs_open_context *ctx =
657 (struct nfs_open_context *) file->private_data;
658 struct address_space *mapping = file->f_mapping;
659 struct inode *inode = mapping->host;
665 dprintk("nfs: direct read(%s/%s, %lu@%lu)\n",
666 file->f_dentry->d_parent->d_name.name,
667 file->f_dentry->d_name.name,
668 (unsigned long) count, (unsigned long) pos);
670 if (!is_sync_kiocb(iocb))
675 if (!access_ok(VERIFY_WRITE, iov.iov_base, iov.iov_len))
681 if (mapping->nrpages) {
682 retval = filemap_fdatawrite(mapping);
684 retval = nfs_wb_all(inode);
686 retval = filemap_fdatawait(mapping);
691 retval = nfs_direct_read(inode, ctx, &iov, pos, 1);
693 *ppos = pos + retval;
700 * nfs_file_direct_write - file direct write operation for NFS files
701 * @iocb: target I/O control block
702 * @buf: user's buffer from which to write data
703 * count: number of bytes to write
704 * pos: byte offset in file where writing starts
706 * We use this function for direct writes instead of calling
707 * generic_file_aio_write() in order to avoid taking the inode
708 * semaphore and updating the i_size. The NFS server will set
709 * the new i_size and this client must read the updated size
710 * back into its cache. We let the server do generic write
711 * parameter checking and report problems.
713 * We also avoid an unnecessary invocation of generic_osync_inode(),
714 * as it is fairly meaningless to sync the metadata of an NFS file.
716 * We eliminate local atime updates, see direct read above.
718 * We avoid unnecessary page cache invalidations for normal cached
719 * readers of this file.
721 * Note that O_APPEND is not supported for NFS direct writes, as there
722 * is no atomic O_APPEND write facility in the NFS protocol.
725 nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
727 ssize_t retval = -EINVAL;
728 loff_t *ppos = &iocb->ki_pos;
729 unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
730 struct file *file = iocb->ki_filp;
731 struct nfs_open_context *ctx =
732 (struct nfs_open_context *) file->private_data;
733 struct address_space *mapping = file->f_mapping;
734 struct inode *inode = mapping->host;
736 .iov_base = (char __user *)buf,
740 dfprintk(VFS, "nfs: direct write(%s/%s(%ld), %lu@%lu)\n",
741 file->f_dentry->d_parent->d_name.name,
742 file->f_dentry->d_name.name, inode->i_ino,
743 (unsigned long) count, (unsigned long) pos);
745 if (!is_sync_kiocb(iocb))
752 if (!access_ok(VERIFY_READ, iov.iov_base, iov.iov_len))
755 if (limit != RLIM_INFINITY) {
757 send_sig(SIGXFSZ, current, 0);
760 if (count > limit - (unsigned long) pos)
761 count = limit - (unsigned long) pos;
767 if (mapping->nrpages) {
768 retval = filemap_fdatawrite(mapping);
770 retval = nfs_wb_all(inode);
772 retval = filemap_fdatawait(mapping);
777 retval = nfs_direct_write(inode, ctx, &iov, pos, 1);
778 if (mapping->nrpages)
779 invalidate_inode_pages2(mapping);
781 *ppos = pos + retval;
787 int nfs_init_directcache(void)
789 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
790 sizeof(struct nfs_direct_req),
791 0, SLAB_RECLAIM_ACCOUNT,
793 if (nfs_direct_cachep == NULL)
799 void nfs_destroy_directcache(void)
801 if (kmem_cache_destroy(nfs_direct_cachep))
802 printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n");