Merge git://git.kernel.org/pub/scm/linux/kernel/git/wim/linux-2.6-watchdog
[linux-2.6] / fs / nfs / direct.c
1 /*
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  * 04 May 2005  support O_DIRECT with aio  --cel
38  *
39  */
40
41 #include <linux/errno.h>
42 #include <linux/sched.h>
43 #include <linux/kernel.h>
44 #include <linux/file.h>
45 #include <linux/pagemap.h>
46 #include <linux/kref.h>
47
48 #include <linux/nfs_fs.h>
49 #include <linux/nfs_page.h>
50 #include <linux/sunrpc/clnt.h>
51
52 #include <asm/system.h>
53 #include <asm/uaccess.h>
54 #include <asm/atomic.h>
55
56 #include "internal.h"
57 #include "iostat.h"
58
59 #define NFSDBG_FACILITY         NFSDBG_VFS
60
61 static struct kmem_cache *nfs_direct_cachep;
62
63 /*
64  * This represents a set of asynchronous requests that we're waiting on
65  */
66 struct nfs_direct_req {
67         struct kref             kref;           /* release manager */
68
69         /* I/O parameters */
70         struct nfs_open_context *ctx;           /* file open context info */
71         struct kiocb *          iocb;           /* controlling i/o request */
72         struct inode *          inode;          /* target file of i/o */
73
74         /* completion state */
75         atomic_t                io_count;       /* i/os we're waiting for */
76         spinlock_t              lock;           /* protect completion state */
77         ssize_t                 count,          /* bytes actually processed */
78                                 error;          /* any reported error */
79         struct completion       completion;     /* wait for i/o completion */
80
81         /* commit state */
82         struct list_head        rewrite_list;   /* saved nfs_write_data structs */
83         struct nfs_write_data * commit_data;    /* special write_data for commits */
84         int                     flags;
85 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
86 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
87         struct nfs_writeverf    verf;           /* unstable write verifier */
88 };
89
90 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
91 static const struct rpc_call_ops nfs_write_direct_ops;
92
93 static inline void get_dreq(struct nfs_direct_req *dreq)
94 {
95         atomic_inc(&dreq->io_count);
96 }
97
98 static inline int put_dreq(struct nfs_direct_req *dreq)
99 {
100         return atomic_dec_and_test(&dreq->io_count);
101 }
102
103 /**
104  * nfs_direct_IO - NFS address space operation for direct I/O
105  * @rw: direction (read or write)
106  * @iocb: target I/O control block
107  * @iov: array of vectors that define I/O buffer
108  * @pos: offset in file to begin the operation
109  * @nr_segs: size of iovec array
110  *
111  * The presence of this routine in the address space ops vector means
112  * the NFS client supports direct I/O.  However, we shunt off direct
113  * read and write requests before the VFS gets them, so this method
114  * should never be called.
115  */
116 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
117 {
118         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
119                         iocb->ki_filp->f_path.dentry->d_name.name,
120                         (long long) pos, nr_segs);
121
122         return -EINVAL;
123 }
124
125 static void nfs_direct_dirty_pages(struct page **pages, unsigned int pgbase, size_t count)
126 {
127         unsigned int npages;
128         unsigned int i;
129
130         if (count == 0)
131                 return;
132         pages += (pgbase >> PAGE_SHIFT);
133         npages = (count + (pgbase & ~PAGE_MASK) + PAGE_SIZE - 1) >> PAGE_SHIFT;
134         for (i = 0; i < npages; i++) {
135                 struct page *page = pages[i];
136                 if (!PageCompound(page))
137                         set_page_dirty(page);
138         }
139 }
140
141 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
142 {
143         unsigned int i;
144         for (i = 0; i < npages; i++)
145                 page_cache_release(pages[i]);
146 }
147
148 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
149 {
150         struct nfs_direct_req *dreq;
151
152         dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
153         if (!dreq)
154                 return NULL;
155
156         kref_init(&dreq->kref);
157         kref_get(&dreq->kref);
158         init_completion(&dreq->completion);
159         INIT_LIST_HEAD(&dreq->rewrite_list);
160         dreq->iocb = NULL;
161         dreq->ctx = NULL;
162         spin_lock_init(&dreq->lock);
163         atomic_set(&dreq->io_count, 0);
164         dreq->count = 0;
165         dreq->error = 0;
166         dreq->flags = 0;
167
168         return dreq;
169 }
170
171 static void nfs_direct_req_free(struct kref *kref)
172 {
173         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
174
175         if (dreq->ctx != NULL)
176                 put_nfs_open_context(dreq->ctx);
177         kmem_cache_free(nfs_direct_cachep, dreq);
178 }
179
180 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
181 {
182         kref_put(&dreq->kref, nfs_direct_req_free);
183 }
184
185 /*
186  * Collects and returns the final error value/byte-count.
187  */
188 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
189 {
190         ssize_t result = -EIOCBQUEUED;
191         struct rpc_clnt *clnt;
192         sigset_t oldset;
193
194         /* Async requests don't wait here */
195         if (dreq->iocb)
196                 goto out;
197
198         clnt = NFS_CLIENT(dreq->inode);
199         rpc_clnt_sigmask(clnt, &oldset);
200         result = wait_for_completion_interruptible(&dreq->completion);
201         rpc_clnt_sigunmask(clnt, &oldset);
202
203         if (!result)
204                 result = dreq->error;
205         if (!result)
206                 result = dreq->count;
207
208 out:
209         return (ssize_t) result;
210 }
211
212 /*
213  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
214  * the iocb is still valid here if this is a synchronous request.
215  */
216 static void nfs_direct_complete(struct nfs_direct_req *dreq)
217 {
218         if (dreq->iocb) {
219                 long res = (long) dreq->error;
220                 if (!res)
221                         res = (long) dreq->count;
222                 aio_complete(dreq->iocb, res, 0);
223         }
224         complete_all(&dreq->completion);
225
226         nfs_direct_req_release(dreq);
227 }
228
229 /*
230  * We must hold a reference to all the pages in this direct read request
231  * until the RPCs complete.  This could be long *after* we are woken up in
232  * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
233  */
234 static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
235 {
236         struct nfs_read_data *data = calldata;
237         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
238
239         if (nfs_readpage_result(task, data) != 0)
240                 return;
241
242         spin_lock(&dreq->lock);
243         if (unlikely(task->tk_status < 0)) {
244                 dreq->error = task->tk_status;
245                 spin_unlock(&dreq->lock);
246         } else {
247                 dreq->count += data->res.count;
248                 spin_unlock(&dreq->lock);
249                 nfs_direct_dirty_pages(data->pagevec,
250                                 data->args.pgbase,
251                                 data->res.count);
252         }
253         nfs_direct_release_pages(data->pagevec, data->npages);
254
255         if (put_dreq(dreq))
256                 nfs_direct_complete(dreq);
257 }
258
259 static const struct rpc_call_ops nfs_read_direct_ops = {
260         .rpc_call_done = nfs_direct_read_result,
261         .rpc_release = nfs_readdata_release,
262 };
263
264 /*
265  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
266  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
267  * bail and stop sending more reads.  Read length accounting is
268  * handled automatically by nfs_direct_read_result().  Otherwise, if
269  * no requests have been sent, just return an error.
270  */
271 static ssize_t nfs_direct_read_schedule_segment(struct nfs_direct_req *dreq,
272                                                 const struct iovec *iov,
273                                                 loff_t pos)
274 {
275         struct nfs_open_context *ctx = dreq->ctx;
276         struct inode *inode = ctx->path.dentry->d_inode;
277         unsigned long user_addr = (unsigned long)iov->iov_base;
278         size_t count = iov->iov_len;
279         size_t rsize = NFS_SERVER(inode)->rsize;
280         struct rpc_task *task;
281         struct rpc_message msg = {
282                 .rpc_cred = ctx->cred,
283         };
284         struct rpc_task_setup task_setup_data = {
285                 .rpc_client = NFS_CLIENT(inode),
286                 .rpc_message = &msg,
287                 .callback_ops = &nfs_read_direct_ops,
288                 .flags = RPC_TASK_ASYNC,
289         };
290         unsigned int pgbase;
291         int result;
292         ssize_t started = 0;
293
294         do {
295                 struct nfs_read_data *data;
296                 size_t bytes;
297
298                 pgbase = user_addr & ~PAGE_MASK;
299                 bytes = min(rsize,count);
300
301                 result = -ENOMEM;
302                 data = nfs_readdata_alloc(nfs_page_array_len(pgbase, bytes));
303                 if (unlikely(!data))
304                         break;
305
306                 down_read(&current->mm->mmap_sem);
307                 result = get_user_pages(current, current->mm, user_addr,
308                                         data->npages, 1, 0, data->pagevec, NULL);
309                 up_read(&current->mm->mmap_sem);
310                 if (result < 0) {
311                         nfs_readdata_release(data);
312                         break;
313                 }
314                 if ((unsigned)result < data->npages) {
315                         bytes = result * PAGE_SIZE;
316                         if (bytes <= pgbase) {
317                                 nfs_direct_release_pages(data->pagevec, result);
318                                 nfs_readdata_release(data);
319                                 break;
320                         }
321                         bytes -= pgbase;
322                         data->npages = result;
323                 }
324
325                 get_dreq(dreq);
326
327                 data->req = (struct nfs_page *) dreq;
328                 data->inode = inode;
329                 data->cred = msg.rpc_cred;
330                 data->args.fh = NFS_FH(inode);
331                 data->args.context = ctx;
332                 data->args.offset = pos;
333                 data->args.pgbase = pgbase;
334                 data->args.pages = data->pagevec;
335                 data->args.count = bytes;
336                 data->res.fattr = &data->fattr;
337                 data->res.eof = 0;
338                 data->res.count = bytes;
339                 msg.rpc_argp = &data->args;
340                 msg.rpc_resp = &data->res;
341
342                 task_setup_data.task = &data->task;
343                 task_setup_data.callback_data = data;
344                 NFS_PROTO(inode)->read_setup(data, &msg);
345
346                 task = rpc_run_task(&task_setup_data);
347                 if (!IS_ERR(task))
348                         rpc_put_task(task);
349
350                 dprintk("NFS: %5u initiated direct read call "
351                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
352                                 data->task.tk_pid,
353                                 inode->i_sb->s_id,
354                                 (long long)NFS_FILEID(inode),
355                                 bytes,
356                                 (unsigned long long)data->args.offset);
357
358                 started += bytes;
359                 user_addr += bytes;
360                 pos += bytes;
361                 /* FIXME: Remove this unnecessary math from final patch */
362                 pgbase += bytes;
363                 pgbase &= ~PAGE_MASK;
364                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
365
366                 count -= bytes;
367         } while (count != 0);
368
369         if (started)
370                 return started;
371         return result < 0 ? (ssize_t) result : -EFAULT;
372 }
373
374 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
375                                               const struct iovec *iov,
376                                               unsigned long nr_segs,
377                                               loff_t pos)
378 {
379         ssize_t result = -EINVAL;
380         size_t requested_bytes = 0;
381         unsigned long seg;
382
383         get_dreq(dreq);
384
385         for (seg = 0; seg < nr_segs; seg++) {
386                 const struct iovec *vec = &iov[seg];
387                 result = nfs_direct_read_schedule_segment(dreq, vec, pos);
388                 if (result < 0)
389                         break;
390                 requested_bytes += result;
391                 if ((size_t)result < vec->iov_len)
392                         break;
393                 pos += vec->iov_len;
394         }
395
396         if (put_dreq(dreq))
397                 nfs_direct_complete(dreq);
398
399         if (requested_bytes != 0)
400                 return 0;
401
402         if (result < 0)
403                 return result;
404         return -EIO;
405 }
406
407 static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
408                                unsigned long nr_segs, loff_t pos)
409 {
410         ssize_t result = 0;
411         struct inode *inode = iocb->ki_filp->f_mapping->host;
412         struct nfs_direct_req *dreq;
413
414         dreq = nfs_direct_req_alloc();
415         if (!dreq)
416                 return -ENOMEM;
417
418         dreq->inode = inode;
419         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
420         if (!is_sync_kiocb(iocb))
421                 dreq->iocb = iocb;
422
423         result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
424         if (!result)
425                 result = nfs_direct_wait(dreq);
426         nfs_direct_req_release(dreq);
427
428         return result;
429 }
430
431 static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
432 {
433         while (!list_empty(&dreq->rewrite_list)) {
434                 struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages);
435                 list_del(&data->pages);
436                 nfs_direct_release_pages(data->pagevec, data->npages);
437                 nfs_writedata_release(data);
438         }
439 }
440
441 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
442 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
443 {
444         struct inode *inode = dreq->inode;
445         struct list_head *p;
446         struct nfs_write_data *data;
447         struct rpc_task *task;
448         struct rpc_message msg = {
449                 .rpc_cred = dreq->ctx->cred,
450         };
451         struct rpc_task_setup task_setup_data = {
452                 .rpc_client = NFS_CLIENT(inode),
453                 .callback_ops = &nfs_write_direct_ops,
454                 .flags = RPC_TASK_ASYNC,
455         };
456
457         dreq->count = 0;
458         get_dreq(dreq);
459
460         list_for_each(p, &dreq->rewrite_list) {
461                 data = list_entry(p, struct nfs_write_data, pages);
462
463                 get_dreq(dreq);
464
465                 /* Use stable writes */
466                 data->args.stable = NFS_FILE_SYNC;
467
468                 /*
469                  * Reset data->res.
470                  */
471                 nfs_fattr_init(&data->fattr);
472                 data->res.count = data->args.count;
473                 memset(&data->verf, 0, sizeof(data->verf));
474
475                 /*
476                  * Reuse data->task; data->args should not have changed
477                  * since the original request was sent.
478                  */
479                 task_setup_data.task = &data->task;
480                 task_setup_data.callback_data = data;
481                 msg.rpc_argp = &data->args;
482                 msg.rpc_resp = &data->res;
483                 NFS_PROTO(inode)->write_setup(data, &msg);
484
485                 /*
486                  * We're called via an RPC callback, so BKL is already held.
487                  */
488                 task = rpc_run_task(&task_setup_data);
489                 if (!IS_ERR(task))
490                         rpc_put_task(task);
491
492                 dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
493                                 data->task.tk_pid,
494                                 inode->i_sb->s_id,
495                                 (long long)NFS_FILEID(inode),
496                                 data->args.count,
497                                 (unsigned long long)data->args.offset);
498         }
499
500         if (put_dreq(dreq))
501                 nfs_direct_write_complete(dreq, inode);
502 }
503
504 static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
505 {
506         struct nfs_write_data *data = calldata;
507         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
508
509         /* Call the NFS version-specific code */
510         if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
511                 return;
512         if (unlikely(task->tk_status < 0)) {
513                 dprintk("NFS: %5u commit failed with error %d.\n",
514                                 task->tk_pid, task->tk_status);
515                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
516         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
517                 dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
518                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
519         }
520
521         dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
522         nfs_direct_write_complete(dreq, data->inode);
523 }
524
525 static const struct rpc_call_ops nfs_commit_direct_ops = {
526         .rpc_call_done = nfs_direct_commit_result,
527         .rpc_release = nfs_commit_release,
528 };
529
530 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
531 {
532         struct nfs_write_data *data = dreq->commit_data;
533         struct rpc_task *task;
534         struct rpc_message msg = {
535                 .rpc_argp = &data->args,
536                 .rpc_resp = &data->res,
537                 .rpc_cred = dreq->ctx->cred,
538         };
539         struct rpc_task_setup task_setup_data = {
540                 .task = &data->task,
541                 .rpc_client = NFS_CLIENT(dreq->inode),
542                 .rpc_message = &msg,
543                 .callback_ops = &nfs_commit_direct_ops,
544                 .callback_data = data,
545                 .flags = RPC_TASK_ASYNC,
546         };
547
548         data->inode = dreq->inode;
549         data->cred = msg.rpc_cred;
550
551         data->args.fh = NFS_FH(data->inode);
552         data->args.offset = 0;
553         data->args.count = 0;
554         data->res.count = 0;
555         data->res.fattr = &data->fattr;
556         data->res.verf = &data->verf;
557
558         NFS_PROTO(data->inode)->commit_setup(data, &msg);
559
560         /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
561         dreq->commit_data = NULL;
562
563         dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
564
565         task = rpc_run_task(&task_setup_data);
566         if (!IS_ERR(task))
567                 rpc_put_task(task);
568 }
569
570 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
571 {
572         int flags = dreq->flags;
573
574         dreq->flags = 0;
575         switch (flags) {
576                 case NFS_ODIRECT_DO_COMMIT:
577                         nfs_direct_commit_schedule(dreq);
578                         break;
579                 case NFS_ODIRECT_RESCHED_WRITES:
580                         nfs_direct_write_reschedule(dreq);
581                         break;
582                 default:
583                         if (dreq->commit_data != NULL)
584                                 nfs_commit_free(dreq->commit_data);
585                         nfs_direct_free_writedata(dreq);
586                         nfs_zap_mapping(inode, inode->i_mapping);
587                         nfs_direct_complete(dreq);
588         }
589 }
590
591 static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
592 {
593         dreq->commit_data = nfs_commit_alloc();
594         if (dreq->commit_data != NULL)
595                 dreq->commit_data->req = (struct nfs_page *) dreq;
596 }
597 #else
598 static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
599 {
600         dreq->commit_data = NULL;
601 }
602
603 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
604 {
605         nfs_direct_free_writedata(dreq);
606         nfs_zap_mapping(inode, inode->i_mapping);
607         nfs_direct_complete(dreq);
608 }
609 #endif
610
611 static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
612 {
613         struct nfs_write_data *data = calldata;
614         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
615         int status = task->tk_status;
616
617         if (nfs_writeback_done(task, data) != 0)
618                 return;
619
620         spin_lock(&dreq->lock);
621
622         if (unlikely(status < 0)) {
623                 /* An error has occurred, so we should not commit */
624                 dreq->flags = 0;
625                 dreq->error = status;
626         }
627         if (unlikely(dreq->error != 0))
628                 goto out_unlock;
629
630         dreq->count += data->res.count;
631
632         if (data->res.verf->committed != NFS_FILE_SYNC) {
633                 switch (dreq->flags) {
634                         case 0:
635                                 memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
636                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
637                                 break;
638                         case NFS_ODIRECT_DO_COMMIT:
639                                 if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
640                                         dprintk("NFS: %5u write verify failed\n", task->tk_pid);
641                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
642                                 }
643                 }
644         }
645 out_unlock:
646         spin_unlock(&dreq->lock);
647 }
648
649 /*
650  * NB: Return the value of the first error return code.  Subsequent
651  *     errors after the first one are ignored.
652  */
653 static void nfs_direct_write_release(void *calldata)
654 {
655         struct nfs_write_data *data = calldata;
656         struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
657
658         if (put_dreq(dreq))
659                 nfs_direct_write_complete(dreq, data->inode);
660 }
661
662 static const struct rpc_call_ops nfs_write_direct_ops = {
663         .rpc_call_done = nfs_direct_write_result,
664         .rpc_release = nfs_direct_write_release,
665 };
666
667 /*
668  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
669  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
670  * bail and stop sending more writes.  Write length accounting is
671  * handled automatically by nfs_direct_write_result().  Otherwise, if
672  * no requests have been sent, just return an error.
673  */
674 static ssize_t nfs_direct_write_schedule_segment(struct nfs_direct_req *dreq,
675                                                  const struct iovec *iov,
676                                                  loff_t pos, int sync)
677 {
678         struct nfs_open_context *ctx = dreq->ctx;
679         struct inode *inode = ctx->path.dentry->d_inode;
680         unsigned long user_addr = (unsigned long)iov->iov_base;
681         size_t count = iov->iov_len;
682         struct rpc_task *task;
683         struct rpc_message msg = {
684                 .rpc_cred = ctx->cred,
685         };
686         struct rpc_task_setup task_setup_data = {
687                 .rpc_client = NFS_CLIENT(inode),
688                 .rpc_message = &msg,
689                 .callback_ops = &nfs_write_direct_ops,
690                 .flags = RPC_TASK_ASYNC,
691         };
692         size_t wsize = NFS_SERVER(inode)->wsize;
693         unsigned int pgbase;
694         int result;
695         ssize_t started = 0;
696
697         do {
698                 struct nfs_write_data *data;
699                 size_t bytes;
700
701                 pgbase = user_addr & ~PAGE_MASK;
702                 bytes = min(wsize,count);
703
704                 result = -ENOMEM;
705                 data = nfs_writedata_alloc(nfs_page_array_len(pgbase, bytes));
706                 if (unlikely(!data))
707                         break;
708
709                 down_read(&current->mm->mmap_sem);
710                 result = get_user_pages(current, current->mm, user_addr,
711                                         data->npages, 0, 0, data->pagevec, NULL);
712                 up_read(&current->mm->mmap_sem);
713                 if (result < 0) {
714                         nfs_writedata_release(data);
715                         break;
716                 }
717                 if ((unsigned)result < data->npages) {
718                         bytes = result * PAGE_SIZE;
719                         if (bytes <= pgbase) {
720                                 nfs_direct_release_pages(data->pagevec, result);
721                                 nfs_writedata_release(data);
722                                 break;
723                         }
724                         bytes -= pgbase;
725                         data->npages = result;
726                 }
727
728                 get_dreq(dreq);
729
730                 list_move_tail(&data->pages, &dreq->rewrite_list);
731
732                 data->req = (struct nfs_page *) dreq;
733                 data->inode = inode;
734                 data->cred = msg.rpc_cred;
735                 data->args.fh = NFS_FH(inode);
736                 data->args.context = ctx;
737                 data->args.offset = pos;
738                 data->args.pgbase = pgbase;
739                 data->args.pages = data->pagevec;
740                 data->args.count = bytes;
741                 data->args.stable = sync;
742                 data->res.fattr = &data->fattr;
743                 data->res.count = bytes;
744                 data->res.verf = &data->verf;
745
746                 task_setup_data.task = &data->task;
747                 task_setup_data.callback_data = data;
748                 msg.rpc_argp = &data->args;
749                 msg.rpc_resp = &data->res;
750                 NFS_PROTO(inode)->write_setup(data, &msg);
751
752                 task = rpc_run_task(&task_setup_data);
753                 if (!IS_ERR(task))
754                         rpc_put_task(task);
755
756                 dprintk("NFS: %5u initiated direct write call "
757                         "(req %s/%Ld, %zu bytes @ offset %Lu)\n",
758                                 data->task.tk_pid,
759                                 inode->i_sb->s_id,
760                                 (long long)NFS_FILEID(inode),
761                                 bytes,
762                                 (unsigned long long)data->args.offset);
763
764                 started += bytes;
765                 user_addr += bytes;
766                 pos += bytes;
767
768                 /* FIXME: Remove this useless math from the final patch */
769                 pgbase += bytes;
770                 pgbase &= ~PAGE_MASK;
771                 BUG_ON(pgbase != (user_addr & ~PAGE_MASK));
772
773                 count -= bytes;
774         } while (count != 0);
775
776         if (started)
777                 return started;
778         return result < 0 ? (ssize_t) result : -EFAULT;
779 }
780
781 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
782                                                const struct iovec *iov,
783                                                unsigned long nr_segs,
784                                                loff_t pos, int sync)
785 {
786         ssize_t result = 0;
787         size_t requested_bytes = 0;
788         unsigned long seg;
789
790         get_dreq(dreq);
791
792         for (seg = 0; seg < nr_segs; seg++) {
793                 const struct iovec *vec = &iov[seg];
794                 result = nfs_direct_write_schedule_segment(dreq, vec,
795                                                            pos, sync);
796                 if (result < 0)
797                         break;
798                 requested_bytes += result;
799                 if ((size_t)result < vec->iov_len)
800                         break;
801                 pos += vec->iov_len;
802         }
803
804         if (put_dreq(dreq))
805                 nfs_direct_write_complete(dreq, dreq->inode);
806
807         if (requested_bytes != 0)
808                 return 0;
809
810         if (result < 0)
811                 return result;
812         return -EIO;
813 }
814
815 static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
816                                 unsigned long nr_segs, loff_t pos,
817                                 size_t count)
818 {
819         ssize_t result = 0;
820         struct inode *inode = iocb->ki_filp->f_mapping->host;
821         struct nfs_direct_req *dreq;
822         size_t wsize = NFS_SERVER(inode)->wsize;
823         int sync = NFS_UNSTABLE;
824
825         dreq = nfs_direct_req_alloc();
826         if (!dreq)
827                 return -ENOMEM;
828         nfs_alloc_commit_data(dreq);
829
830         if (dreq->commit_data == NULL || count < wsize)
831                 sync = NFS_FILE_SYNC;
832
833         dreq->inode = inode;
834         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
835         if (!is_sync_kiocb(iocb))
836                 dreq->iocb = iocb;
837
838         result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, sync);
839         if (!result)
840                 result = nfs_direct_wait(dreq);
841         nfs_direct_req_release(dreq);
842
843         return result;
844 }
845
846 /**
847  * nfs_file_direct_read - file direct read operation for NFS files
848  * @iocb: target I/O control block
849  * @iov: vector of user buffers into which to read data
850  * @nr_segs: size of iov vector
851  * @pos: byte offset in file where reading starts
852  *
853  * We use this function for direct reads instead of calling
854  * generic_file_aio_read() in order to avoid gfar's check to see if
855  * the request starts before the end of the file.  For that check
856  * to work, we must generate a GETATTR before each direct read, and
857  * even then there is a window between the GETATTR and the subsequent
858  * READ where the file size could change.  Our preference is simply
859  * to do all reads the application wants, and the server will take
860  * care of managing the end of file boundary.
861  *
862  * This function also eliminates unnecessarily updating the file's
863  * atime locally, as the NFS server sets the file's atime, and this
864  * client must read the updated atime from the server back into its
865  * cache.
866  */
867 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
868                                 unsigned long nr_segs, loff_t pos)
869 {
870         ssize_t retval = -EINVAL;
871         struct file *file = iocb->ki_filp;
872         struct address_space *mapping = file->f_mapping;
873         size_t count;
874
875         count = iov_length(iov, nr_segs);
876         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
877
878         dprintk("nfs: direct read(%s/%s, %zd@%Ld)\n",
879                 file->f_path.dentry->d_parent->d_name.name,
880                 file->f_path.dentry->d_name.name,
881                 count, (long long) pos);
882
883         retval = 0;
884         if (!count)
885                 goto out;
886
887         retval = nfs_sync_mapping(mapping);
888         if (retval)
889                 goto out;
890
891         retval = nfs_direct_read(iocb, iov, nr_segs, pos);
892         if (retval > 0)
893                 iocb->ki_pos = pos + retval;
894
895 out:
896         return retval;
897 }
898
899 /**
900  * nfs_file_direct_write - file direct write operation for NFS files
901  * @iocb: target I/O control block
902  * @iov: vector of user buffers from which to write data
903  * @nr_segs: size of iov vector
904  * @pos: byte offset in file where writing starts
905  *
906  * We use this function for direct writes instead of calling
907  * generic_file_aio_write() in order to avoid taking the inode
908  * semaphore and updating the i_size.  The NFS server will set
909  * the new i_size and this client must read the updated size
910  * back into its cache.  We let the server do generic write
911  * parameter checking and report problems.
912  *
913  * We also avoid an unnecessary invocation of generic_osync_inode(),
914  * as it is fairly meaningless to sync the metadata of an NFS file.
915  *
916  * We eliminate local atime updates, see direct read above.
917  *
918  * We avoid unnecessary page cache invalidations for normal cached
919  * readers of this file.
920  *
921  * Note that O_APPEND is not supported for NFS direct writes, as there
922  * is no atomic O_APPEND write facility in the NFS protocol.
923  */
924 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
925                                 unsigned long nr_segs, loff_t pos)
926 {
927         ssize_t retval = -EINVAL;
928         struct file *file = iocb->ki_filp;
929         struct address_space *mapping = file->f_mapping;
930         size_t count;
931
932         count = iov_length(iov, nr_segs);
933         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
934
935         dfprintk(VFS, "nfs: direct write(%s/%s, %zd@%Ld)\n",
936                 file->f_path.dentry->d_parent->d_name.name,
937                 file->f_path.dentry->d_name.name,
938                 count, (long long) pos);
939
940         retval = generic_write_checks(file, &pos, &count, 0);
941         if (retval)
942                 goto out;
943
944         retval = -EINVAL;
945         if ((ssize_t) count < 0)
946                 goto out;
947         retval = 0;
948         if (!count)
949                 goto out;
950
951         retval = nfs_sync_mapping(mapping);
952         if (retval)
953                 goto out;
954
955         retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
956
957         if (retval > 0)
958                 iocb->ki_pos = pos + retval;
959
960 out:
961         return retval;
962 }
963
964 /**
965  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
966  *
967  */
968 int __init nfs_init_directcache(void)
969 {
970         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
971                                                 sizeof(struct nfs_direct_req),
972                                                 0, (SLAB_RECLAIM_ACCOUNT|
973                                                         SLAB_MEM_SPREAD),
974                                                 NULL);
975         if (nfs_direct_cachep == NULL)
976                 return -ENOMEM;
977
978         return 0;
979 }
980
981 /**
982  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
983  *
984  */
985 void nfs_destroy_directcache(void)
986 {
987         kmem_cache_destroy(nfs_direct_cachep);
988 }