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 void nfs_free_user_pages(struct page **pages, int npages, int do_dirty);
61 static kmem_cache_t *nfs_direct_cachep;
64 * This represents a set of asynchronous requests that we're waiting on
66 struct nfs_direct_req {
67 struct kref kref; /* release manager */
68 struct list_head list; /* nfs_read_data structs */
69 wait_queue_head_t wait; /* wait for i/o completion */
70 struct page ** pages; /* pages in our buffer */
71 unsigned int npages; /* count of pages */
72 atomic_t complete, /* i/os we're waiting for */
73 count, /* bytes actually processed */
74 error; /* any reported error */
79 * nfs_get_user_pages - find and set up pages underlying user's buffer
80 * rw: direction (read or write)
81 * user_addr: starting address of this segment of user's buffer
82 * count: size of this segment
83 * @pages: returned array of page struct pointers underlying user's buffer
86 nfs_get_user_pages(int rw, unsigned long user_addr, size_t size,
90 unsigned long page_count;
93 /* set an arbitrary limit to prevent type overflow */
94 /* XXX: this can probably be as large as INT_MAX */
95 if (size > MAX_DIRECTIO_SIZE) {
100 page_count = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT;
101 page_count -= user_addr >> PAGE_SHIFT;
103 array_size = (page_count * sizeof(struct page *));
104 *pages = kmalloc(array_size, GFP_KERNEL);
106 down_read(¤t->mm->mmap_sem);
107 result = get_user_pages(current, current->mm, user_addr,
108 page_count, (rw == READ), 0,
110 up_read(¤t->mm->mmap_sem);
112 * If we got fewer pages than expected from get_user_pages(),
113 * the user buffer runs off the end of a mapping; return EFAULT.
115 if (result >= 0 && result < page_count) {
116 nfs_free_user_pages(*pages, result, 0);
125 * nfs_free_user_pages - tear down page struct array
126 * @pages: array of page struct pointers underlying target buffer
127 * @npages: number of pages in the array
128 * @do_dirty: dirty the pages as we release them
131 nfs_free_user_pages(struct page **pages, int npages, int do_dirty)
134 for (i = 0; i < npages; i++) {
135 struct page *page = pages[i];
136 if (do_dirty && !PageCompound(page))
137 set_page_dirty_lock(page);
138 page_cache_release(page);
144 * nfs_direct_req_release - release nfs_direct_req structure for direct read
145 * @kref: kref object embedded in an nfs_direct_req structure
148 static void nfs_direct_req_release(struct kref *kref)
150 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
151 kmem_cache_free(nfs_direct_cachep, dreq);
155 * nfs_direct_read_alloc - allocate nfs_read_data structures for direct read
156 * @count: count of bytes for the read request
157 * @rsize: local rsize setting
159 * Note we also set the number of requests we have in the dreq when we are
160 * done. This prevents races with I/O completion so we will always wait
161 * until all requests have been dispatched and completed.
163 static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, unsigned int rsize)
165 struct list_head *list;
166 struct nfs_direct_req *dreq;
167 unsigned int reads = 0;
168 unsigned int rpages = (rsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
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 atomic_set(&dreq->count, 0);
178 atomic_set(&dreq->error, 0);
182 struct nfs_read_data *data = nfs_readdata_alloc(rpages);
184 if (unlikely(!data)) {
185 while (!list_empty(list)) {
186 data = list_entry(list->next,
187 struct nfs_read_data, pages);
188 list_del(&data->pages);
189 nfs_readdata_free(data);
191 kref_put(&dreq->kref, nfs_direct_req_release);
195 INIT_LIST_HEAD(&data->pages);
196 list_add(&data->pages, list);
198 data->req = (struct nfs_page *) dreq;
204 kref_get(&dreq->kref);
205 atomic_set(&dreq->complete, reads);
210 * nfs_direct_read_result - handle a read reply for a direct read request
211 * @data: address of NFS READ operation control block
212 * @status: status of this NFS READ operation
214 * We must hold a reference to all the pages in this direct read request
215 * until the RPCs complete. This could be long *after* we are woken up in
216 * nfs_direct_read_wait (for instance, if someone hits ^C on a slow server).
218 static void nfs_direct_read_result(struct nfs_read_data *data, int status)
220 struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
222 if (likely(status >= 0))
223 atomic_add(data->res.count, &dreq->count);
225 atomic_set(&dreq->error, status);
227 if (unlikely(atomic_dec_and_test(&dreq->complete))) {
228 nfs_free_user_pages(dreq->pages, dreq->npages, 1);
229 wake_up(&dreq->wait);
230 kref_put(&dreq->kref, nfs_direct_req_release);
235 * nfs_direct_read_schedule - dispatch NFS READ operations for a direct read
236 * @dreq: address of nfs_direct_req struct for this request
237 * @inode: target inode
238 * @ctx: target file open context
239 * @user_addr: starting address of this segment of user's buffer
240 * @count: size of this segment
241 * @file_offset: offset in file to begin the operation
243 * For each nfs_read_data struct that was allocated on the list, dispatch
244 * an NFS READ operation
246 static void nfs_direct_read_schedule(struct nfs_direct_req *dreq,
247 struct inode *inode, struct nfs_open_context *ctx,
248 unsigned long user_addr, size_t count, loff_t file_offset)
250 struct list_head *list = &dreq->list;
251 struct page **pages = dreq->pages;
252 unsigned int curpage, pgbase;
253 unsigned int rsize = NFS_SERVER(inode)->rsize;
256 pgbase = user_addr & ~PAGE_MASK;
258 struct nfs_read_data *data;
265 data = list_entry(list->next, struct nfs_read_data, pages);
266 list_del_init(&data->pages);
269 data->cred = ctx->cred;
270 data->args.fh = NFS_FH(inode);
271 data->args.context = ctx;
272 data->args.offset = file_offset;
273 data->args.pgbase = pgbase;
274 data->args.pages = &pages[curpage];
275 data->args.count = bytes;
276 data->res.fattr = &data->fattr;
278 data->res.count = bytes;
280 NFS_PROTO(inode)->read_setup(data);
282 data->task.tk_cookie = (unsigned long) inode;
283 data->complete = nfs_direct_read_result;
286 rpc_execute(&data->task);
289 dfprintk(VFS, "NFS: %4d initiated direct read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
292 (long long)NFS_FILEID(inode),
294 (unsigned long long)data->args.offset);
296 file_offset += bytes;
298 curpage += pgbase >> PAGE_SHIFT;
299 pgbase &= ~PAGE_MASK;
302 } while (count != 0);
306 * nfs_direct_read_wait - wait for I/O completion for direct reads
307 * @dreq: request on which we are to wait
308 * @intr: whether or not this wait can be interrupted
310 * Collects and returns the final error value/byte-count.
312 static ssize_t nfs_direct_read_wait(struct nfs_direct_req *dreq, int intr)
317 result = wait_event_interruptible(dreq->wait,
318 (atomic_read(&dreq->complete) == 0));
320 wait_event(dreq->wait, (atomic_read(&dreq->complete) == 0));
324 result = atomic_read(&dreq->error);
326 result = atomic_read(&dreq->count);
328 kref_put(&dreq->kref, nfs_direct_req_release);
329 return (ssize_t) result;
333 * nfs_direct_read_seg - Read in one iov segment. Generate separate
334 * read RPCs for each "rsize" bytes.
335 * @inode: target inode
336 * @ctx: target file open context
337 * @user_addr: starting address of this segment of user's buffer
338 * @count: size of this segment
339 * @file_offset: offset in file to begin the operation
340 * @pages: array of addresses of page structs defining user's buffer
341 * @nr_pages: number of pages in the array
344 static ssize_t nfs_direct_read_seg(struct inode *inode,
345 struct nfs_open_context *ctx, unsigned long user_addr,
346 size_t count, loff_t file_offset, struct page **pages,
347 unsigned int nr_pages)
351 struct rpc_clnt *clnt = NFS_CLIENT(inode);
352 struct nfs_direct_req *dreq;
354 dreq = nfs_direct_read_alloc(count, NFS_SERVER(inode)->rsize);
359 dreq->npages = nr_pages;
361 rpc_clnt_sigmask(clnt, &oldset);
362 nfs_direct_read_schedule(dreq, inode, ctx, user_addr, count,
364 result = nfs_direct_read_wait(dreq, clnt->cl_intr);
365 rpc_clnt_sigunmask(clnt, &oldset);
371 * nfs_direct_read - For each iov segment, map the user's buffer
372 * then generate read RPCs.
373 * @inode: target inode
374 * @ctx: target file open context
375 * @iov: array of vectors that define I/O buffer
376 * file_offset: offset in file to begin the operation
377 * nr_segs: size of iovec array
379 * We've already pushed out any non-direct writes so that this read
380 * will see them when we read from the server.
383 nfs_direct_read(struct inode *inode, struct nfs_open_context *ctx,
384 const struct iovec *iov, loff_t file_offset,
385 unsigned long nr_segs)
387 ssize_t tot_bytes = 0;
388 unsigned long seg = 0;
390 while ((seg < nr_segs) && (tot_bytes >= 0)) {
394 const struct iovec *vec = &iov[seg++];
395 unsigned long user_addr = (unsigned long) vec->iov_base;
396 size_t size = vec->iov_len;
398 page_count = nfs_get_user_pages(READ, user_addr, size, &pages);
399 if (page_count < 0) {
400 nfs_free_user_pages(pages, 0, 0);
406 result = nfs_direct_read_seg(inode, ctx, user_addr, size,
407 file_offset, pages, page_count);
415 file_offset += result;
424 * nfs_direct_write_seg - Write out one iov segment. Generate separate
425 * write RPCs for each "wsize" bytes, then commit.
426 * @inode: target inode
427 * @ctx: target file open context
428 * user_addr: starting address of this segment of user's buffer
429 * count: size of this segment
430 * file_offset: offset in file to begin the operation
431 * @pages: array of addresses of page structs defining user's buffer
432 * nr_pages: size of pages array
434 static ssize_t nfs_direct_write_seg(struct inode *inode,
435 struct nfs_open_context *ctx, unsigned long user_addr,
436 size_t count, loff_t file_offset, struct page **pages,
439 const unsigned int wsize = NFS_SERVER(inode)->wsize;
441 int curpage, need_commit;
442 ssize_t result, tot_bytes;
443 struct nfs_writeverf first_verf;
444 struct nfs_write_data *wdata;
446 wdata = nfs_writedata_alloc(NFS_SERVER(inode)->wpages);
450 wdata->inode = inode;
451 wdata->cred = ctx->cred;
452 wdata->args.fh = NFS_FH(inode);
453 wdata->args.context = ctx;
454 wdata->args.stable = NFS_UNSTABLE;
455 if (IS_SYNC(inode) || NFS_PROTO(inode)->version == 2 || count <= wsize)
456 wdata->args.stable = NFS_FILE_SYNC;
457 wdata->res.fattr = &wdata->fattr;
458 wdata->res.verf = &wdata->verf;
460 nfs_begin_data_update(inode);
466 wdata->args.pgbase = user_addr & ~PAGE_MASK;
467 wdata->args.offset = file_offset;
469 wdata->args.count = request;
470 if (wdata->args.count > wsize)
471 wdata->args.count = wsize;
472 wdata->args.pages = &pages[curpage];
474 dprintk("NFS: direct write: c=%u o=%Ld ua=%lu, pb=%u, cp=%u\n",
475 wdata->args.count, (long long) wdata->args.offset,
476 user_addr + tot_bytes, wdata->args.pgbase, curpage);
479 result = NFS_PROTO(inode)->write(wdata);
489 memcpy(&first_verf.verifier, &wdata->verf.verifier,
490 sizeof(first_verf.verifier));
491 if (wdata->verf.committed != NFS_FILE_SYNC) {
493 if (memcmp(&first_verf.verifier, &wdata->verf.verifier,
494 sizeof(first_verf.verifier)))
500 /* in case of a short write: stop now, let the app recover */
501 if (result < wdata->args.count)
504 wdata->args.offset += result;
505 wdata->args.pgbase += result;
506 curpage += wdata->args.pgbase >> PAGE_SHIFT;
507 wdata->args.pgbase &= ~PAGE_MASK;
509 } while (request != 0);
512 * Commit data written so far, even in the event of an error
515 wdata->args.count = tot_bytes;
516 wdata->args.offset = file_offset;
519 result = NFS_PROTO(inode)->commit(wdata);
522 if (result < 0 || memcmp(&first_verf.verifier,
523 &wdata->verf.verifier,
524 sizeof(first_verf.verifier)) != 0)
530 nfs_end_data_update(inode);
531 nfs_writedata_free(wdata);
535 wdata->args.stable = NFS_FILE_SYNC;
540 * nfs_direct_write - For each iov segment, map the user's buffer
541 * then generate write and commit RPCs.
542 * @inode: target inode
543 * @ctx: target file open context
544 * @iov: array of vectors that define I/O buffer
545 * file_offset: offset in file to begin the operation
546 * nr_segs: size of iovec array
548 * Upon return, generic_file_direct_IO invalidates any cached pages
549 * that non-direct readers might access, so they will pick up these
550 * writes immediately.
552 static ssize_t nfs_direct_write(struct inode *inode,
553 struct nfs_open_context *ctx, const struct iovec *iov,
554 loff_t file_offset, unsigned long nr_segs)
556 ssize_t tot_bytes = 0;
557 unsigned long seg = 0;
559 while ((seg < nr_segs) && (tot_bytes >= 0)) {
563 const struct iovec *vec = &iov[seg++];
564 unsigned long user_addr = (unsigned long) vec->iov_base;
565 size_t size = vec->iov_len;
567 page_count = nfs_get_user_pages(WRITE, user_addr, size, &pages);
568 if (page_count < 0) {
569 nfs_free_user_pages(pages, 0, 0);
575 result = nfs_direct_write_seg(inode, ctx, user_addr, size,
576 file_offset, pages, page_count);
577 nfs_free_user_pages(pages, page_count, 0);
585 file_offset += result;
593 * nfs_direct_IO - NFS address space operation for direct I/O
594 * rw: direction (read or write)
595 * @iocb: target I/O control block
596 * @iov: array of vectors that define I/O buffer
597 * file_offset: offset in file to begin the operation
598 * nr_segs: size of iovec array
602 nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
603 loff_t file_offset, unsigned long nr_segs)
605 ssize_t result = -EINVAL;
606 struct file *file = iocb->ki_filp;
607 struct nfs_open_context *ctx;
608 struct dentry *dentry = file->f_dentry;
609 struct inode *inode = dentry->d_inode;
612 * No support for async yet
614 if (!is_sync_kiocb(iocb))
617 ctx = (struct nfs_open_context *)file->private_data;
620 dprintk("NFS: direct_IO(read) (%s) off/no(%Lu/%lu)\n",
621 dentry->d_name.name, file_offset, nr_segs);
623 result = nfs_direct_read(inode, ctx, iov,
624 file_offset, nr_segs);
627 dprintk("NFS: direct_IO(write) (%s) off/no(%Lu/%lu)\n",
628 dentry->d_name.name, file_offset, nr_segs);
630 result = nfs_direct_write(inode, ctx, iov,
631 file_offset, nr_segs);
640 * nfs_file_direct_read - file direct read operation for NFS files
641 * @iocb: target I/O control block
642 * @buf: user's buffer into which to read data
643 * count: number of bytes to read
644 * pos: byte offset in file where reading starts
646 * We use this function for direct reads instead of calling
647 * generic_file_aio_read() in order to avoid gfar's check to see if
648 * the request starts before the end of the file. For that check
649 * to work, we must generate a GETATTR before each direct read, and
650 * even then there is a window between the GETATTR and the subsequent
651 * READ where the file size could change. So our preference is simply
652 * to do all reads the application wants, and the server will take
653 * care of managing the end of file boundary.
655 * This function also eliminates unnecessarily updating the file's
656 * atime locally, as the NFS server sets the file's atime, and this
657 * client must read the updated atime from the server back into its
661 nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
663 ssize_t retval = -EINVAL;
664 loff_t *ppos = &iocb->ki_pos;
665 struct file *file = iocb->ki_filp;
666 struct nfs_open_context *ctx =
667 (struct nfs_open_context *) file->private_data;
668 struct address_space *mapping = file->f_mapping;
669 struct inode *inode = mapping->host;
675 dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
676 file->f_dentry->d_parent->d_name.name,
677 file->f_dentry->d_name.name,
678 (unsigned long) count, (long long) pos);
680 if (!is_sync_kiocb(iocb))
685 if (!access_ok(VERIFY_WRITE, iov.iov_base, iov.iov_len))
691 retval = nfs_sync_mapping(mapping);
695 retval = nfs_direct_read(inode, ctx, &iov, pos, 1);
697 *ppos = pos + retval;
704 * nfs_file_direct_write - file direct write operation for NFS files
705 * @iocb: target I/O control block
706 * @buf: user's buffer from which to write data
707 * count: number of bytes to write
708 * pos: byte offset in file where writing starts
710 * We use this function for direct writes instead of calling
711 * generic_file_aio_write() in order to avoid taking the inode
712 * semaphore and updating the i_size. The NFS server will set
713 * the new i_size and this client must read the updated size
714 * back into its cache. We let the server do generic write
715 * parameter checking and report problems.
717 * We also avoid an unnecessary invocation of generic_osync_inode(),
718 * as it is fairly meaningless to sync the metadata of an NFS file.
720 * We eliminate local atime updates, see direct read above.
722 * We avoid unnecessary page cache invalidations for normal cached
723 * readers of this file.
725 * Note that O_APPEND is not supported for NFS direct writes, as there
726 * is no atomic O_APPEND write facility in the NFS protocol.
729 nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
732 struct file *file = iocb->ki_filp;
733 struct nfs_open_context *ctx =
734 (struct nfs_open_context *) file->private_data;
735 struct address_space *mapping = file->f_mapping;
736 struct inode *inode = mapping->host;
738 .iov_base = (char __user *)buf,
741 dfprintk(VFS, "nfs: direct write(%s/%s, %lu@%Ld)\n",
742 file->f_dentry->d_parent->d_name.name,
743 file->f_dentry->d_name.name,
744 (unsigned long) count, (long long) pos);
747 if (!is_sync_kiocb(iocb))
750 retval = generic_write_checks(file, &pos, &count, 0);
755 if ((ssize_t) count < 0)
763 if (!access_ok(VERIFY_READ, iov.iov_base, iov.iov_len))
766 retval = nfs_sync_mapping(mapping);
770 retval = nfs_direct_write(inode, ctx, &iov, pos, 1);
771 if (mapping->nrpages)
772 invalidate_inode_pages2(mapping);
774 iocb->ki_pos = pos + retval;
780 int nfs_init_directcache(void)
782 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
783 sizeof(struct nfs_direct_req),
784 0, SLAB_RECLAIM_ACCOUNT,
786 if (nfs_direct_cachep == NULL)
792 void nfs_destroy_directcache(void)
794 if (kmem_cache_destroy(nfs_direct_cachep))
795 printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n");