Merge branch 'writeback'
[linux-2.6-microblaze.git] / 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 #include <linux/slab.h>
48 #include <linux/task_io_accounting_ops.h>
49 #include <linux/module.h>
50
51 #include <linux/nfs_fs.h>
52 #include <linux/nfs_page.h>
53 #include <linux/sunrpc/clnt.h>
54
55 #include <asm/uaccess.h>
56 #include <linux/atomic.h>
57
58 #include "internal.h"
59 #include "iostat.h"
60 #include "pnfs.h"
61
62 #define NFSDBG_FACILITY         NFSDBG_VFS
63
64 static struct kmem_cache *nfs_direct_cachep;
65
66 /*
67  * This represents a set of asynchronous requests that we're waiting on
68  */
69 struct nfs_direct_mirror {
70         ssize_t count;
71 };
72
73 struct nfs_direct_req {
74         struct kref             kref;           /* release manager */
75
76         /* I/O parameters */
77         struct nfs_open_context *ctx;           /* file open context info */
78         struct nfs_lock_context *l_ctx;         /* Lock context info */
79         struct kiocb *          iocb;           /* controlling i/o request */
80         struct inode *          inode;          /* target file of i/o */
81
82         /* completion state */
83         atomic_t                io_count;       /* i/os we're waiting for */
84         spinlock_t              lock;           /* protect completion state */
85
86         struct nfs_direct_mirror mirrors[NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX];
87         int                     mirror_count;
88
89         ssize_t                 count,          /* bytes actually processed */
90                                 max_count,      /* max expected count */
91                                 bytes_left,     /* bytes left to be sent */
92                                 io_start,       /* start of IO */
93                                 error;          /* any reported error */
94         struct completion       completion;     /* wait for i/o completion */
95
96         /* commit state */
97         struct nfs_mds_commit_info mds_cinfo;   /* Storage for cinfo */
98         struct pnfs_ds_commit_info ds_cinfo;    /* Storage for cinfo */
99         struct work_struct      work;
100         int                     flags;
101 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
102 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
103         struct nfs_writeverf    verf;           /* unstable write verifier */
104 };
105
106 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
107 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
108 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
109 static void nfs_direct_write_schedule_work(struct work_struct *work);
110
111 static inline void get_dreq(struct nfs_direct_req *dreq)
112 {
113         atomic_inc(&dreq->io_count);
114 }
115
116 static inline int put_dreq(struct nfs_direct_req *dreq)
117 {
118         return atomic_dec_and_test(&dreq->io_count);
119 }
120
121 static void
122 nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
123 {
124         int i;
125         ssize_t count;
126
127         WARN_ON_ONCE(dreq->count >= dreq->max_count);
128
129         if (dreq->mirror_count == 1) {
130                 dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
131                 dreq->count += hdr->good_bytes;
132         } else {
133                 /* mirrored writes */
134                 count = dreq->mirrors[hdr->pgio_mirror_idx].count;
135                 if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) {
136                         count = hdr->io_start + hdr->good_bytes - dreq->io_start;
137                         dreq->mirrors[hdr->pgio_mirror_idx].count = count;
138                 }
139                 /* update the dreq->count by finding the minimum agreed count from all
140                  * mirrors */
141                 count = dreq->mirrors[0].count;
142
143                 for (i = 1; i < dreq->mirror_count; i++)
144                         count = min(count, dreq->mirrors[i].count);
145
146                 dreq->count = count;
147         }
148 }
149
150 /*
151  * nfs_direct_select_verf - select the right verifier
152  * @dreq - direct request possibly spanning multiple servers
153  * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
154  * @commit_idx - commit bucket index for the DS
155  *
156  * returns the correct verifier to use given the role of the server
157  */
158 static struct nfs_writeverf *
159 nfs_direct_select_verf(struct nfs_direct_req *dreq,
160                        struct nfs_client *ds_clp,
161                        int commit_idx)
162 {
163         struct nfs_writeverf *verfp = &dreq->verf;
164
165 #ifdef CONFIG_NFS_V4_1
166         /*
167          * pNFS is in use, use the DS verf except commit_through_mds is set
168          * for layout segment where nbuckets is zero.
169          */
170         if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
171                 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
172                         verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
173                 else
174                         WARN_ON_ONCE(1);
175         }
176 #endif
177         return verfp;
178 }
179
180
181 /*
182  * nfs_direct_set_hdr_verf - set the write/commit verifier
183  * @dreq - direct request possibly spanning multiple servers
184  * @hdr - pageio header to validate against previously seen verfs
185  *
186  * Set the server's (MDS or DS) "seen" verifier
187  */
188 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
189                                     struct nfs_pgio_header *hdr)
190 {
191         struct nfs_writeverf *verfp;
192
193         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
194         WARN_ON_ONCE(verfp->committed >= 0);
195         memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
196         WARN_ON_ONCE(verfp->committed < 0);
197 }
198
199 static int nfs_direct_cmp_verf(const struct nfs_writeverf *v1,
200                 const struct nfs_writeverf *v2)
201 {
202         return nfs_write_verifier_cmp(&v1->verifier, &v2->verifier);
203 }
204
205 /*
206  * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
207  * @dreq - direct request possibly spanning multiple servers
208  * @hdr - pageio header to validate against previously seen verf
209  *
210  * set the server's "seen" verf if not initialized.
211  * returns result of comparison between @hdr->verf and the "seen"
212  * verf of the server used by @hdr (DS or MDS)
213  */
214 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
215                                           struct nfs_pgio_header *hdr)
216 {
217         struct nfs_writeverf *verfp;
218
219         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
220         if (verfp->committed < 0) {
221                 nfs_direct_set_hdr_verf(dreq, hdr);
222                 return 0;
223         }
224         return nfs_direct_cmp_verf(verfp, &hdr->verf);
225 }
226
227 /*
228  * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
229  * @dreq - direct request possibly spanning multiple servers
230  * @data - commit data to validate against previously seen verf
231  *
232  * returns result of comparison between @data->verf and the verf of
233  * the server used by @data (DS or MDS)
234  */
235 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
236                                            struct nfs_commit_data *data)
237 {
238         struct nfs_writeverf *verfp;
239
240         verfp = nfs_direct_select_verf(dreq, data->ds_clp,
241                                          data->ds_commit_index);
242
243         /* verifier not set so always fail */
244         if (verfp->committed < 0)
245                 return 1;
246
247         return nfs_direct_cmp_verf(verfp, &data->verf);
248 }
249
250 /**
251  * nfs_direct_IO - NFS address space operation for direct I/O
252  * @iocb: target I/O control block
253  * @iov: array of vectors that define I/O buffer
254  * @pos: offset in file to begin the operation
255  * @nr_segs: size of iovec array
256  *
257  * The presence of this routine in the address space ops vector means
258  * the NFS client supports direct I/O. However, for most direct IO, we
259  * shunt off direct read and write requests before the VFS gets them,
260  * so this method is only ever called for swap.
261  */
262 ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
263 {
264         struct inode *inode = iocb->ki_filp->f_mapping->host;
265
266         /* we only support swap file calling nfs_direct_IO */
267         if (!IS_SWAPFILE(inode))
268                 return 0;
269
270         VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
271
272         if (iov_iter_rw(iter) == READ)
273                 return nfs_file_direct_read(iocb, iter);
274         return nfs_file_direct_write(iocb, iter);
275 }
276
277 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
278 {
279         unsigned int i;
280         for (i = 0; i < npages; i++)
281                 put_page(pages[i]);
282 }
283
284 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
285                               struct nfs_direct_req *dreq)
286 {
287         cinfo->inode = dreq->inode;
288         cinfo->mds = &dreq->mds_cinfo;
289         cinfo->ds = &dreq->ds_cinfo;
290         cinfo->dreq = dreq;
291         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
292 }
293
294 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
295                                              struct nfs_pageio_descriptor *pgio,
296                                              struct nfs_page *req)
297 {
298         int mirror_count = 1;
299
300         if (pgio->pg_ops->pg_get_mirror_count)
301                 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
302
303         dreq->mirror_count = mirror_count;
304 }
305
306 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
307 {
308         struct nfs_direct_req *dreq;
309
310         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
311         if (!dreq)
312                 return NULL;
313
314         kref_init(&dreq->kref);
315         kref_get(&dreq->kref);
316         init_completion(&dreq->completion);
317         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
318         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
319         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
320         dreq->mirror_count = 1;
321         spin_lock_init(&dreq->lock);
322
323         return dreq;
324 }
325
326 static void nfs_direct_req_free(struct kref *kref)
327 {
328         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
329
330         nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
331         if (dreq->l_ctx != NULL)
332                 nfs_put_lock_context(dreq->l_ctx);
333         if (dreq->ctx != NULL)
334                 put_nfs_open_context(dreq->ctx);
335         kmem_cache_free(nfs_direct_cachep, dreq);
336 }
337
338 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
339 {
340         kref_put(&dreq->kref, nfs_direct_req_free);
341 }
342
343 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
344 {
345         return dreq->bytes_left;
346 }
347 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
348
349 /*
350  * Collects and returns the final error value/byte-count.
351  */
352 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
353 {
354         ssize_t result = -EIOCBQUEUED;
355
356         /* Async requests don't wait here */
357         if (dreq->iocb)
358                 goto out;
359
360         result = wait_for_completion_killable(&dreq->completion);
361
362         if (!result) {
363                 result = dreq->count;
364                 WARN_ON_ONCE(dreq->count < 0);
365         }
366         if (!result)
367                 result = dreq->error;
368
369 out:
370         return (ssize_t) result;
371 }
372
373 /*
374  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
375  * the iocb is still valid here if this is a synchronous request.
376  */
377 static void nfs_direct_complete(struct nfs_direct_req *dreq)
378 {
379         struct inode *inode = dreq->inode;
380
381         inode_dio_end(inode);
382
383         if (dreq->iocb) {
384                 long res = (long) dreq->error;
385                 if (dreq->count != 0) {
386                         res = (long) dreq->count;
387                         WARN_ON_ONCE(dreq->count < 0);
388                 }
389                 dreq->iocb->ki_complete(dreq->iocb, res, 0);
390         }
391
392         complete_all(&dreq->completion);
393
394         nfs_direct_req_release(dreq);
395 }
396
397 static void nfs_direct_readpage_release(struct nfs_page *req)
398 {
399         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
400                 req->wb_context->dentry->d_sb->s_id,
401                 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
402                 req->wb_bytes,
403                 (long long)req_offset(req));
404         nfs_release_request(req);
405 }
406
407 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
408 {
409         unsigned long bytes = 0;
410         struct nfs_direct_req *dreq = hdr->dreq;
411
412         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
413                 goto out_put;
414
415         spin_lock(&dreq->lock);
416         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
417                 dreq->error = hdr->error;
418         else
419                 nfs_direct_good_bytes(dreq, hdr);
420
421         spin_unlock(&dreq->lock);
422
423         while (!list_empty(&hdr->pages)) {
424                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
425                 struct page *page = req->wb_page;
426
427                 if (!PageCompound(page) && bytes < hdr->good_bytes)
428                         set_page_dirty(page);
429                 bytes += req->wb_bytes;
430                 nfs_list_remove_request(req);
431                 nfs_direct_readpage_release(req);
432         }
433 out_put:
434         if (put_dreq(dreq))
435                 nfs_direct_complete(dreq);
436         hdr->release(hdr);
437 }
438
439 static void nfs_read_sync_pgio_error(struct list_head *head)
440 {
441         struct nfs_page *req;
442
443         while (!list_empty(head)) {
444                 req = nfs_list_entry(head->next);
445                 nfs_list_remove_request(req);
446                 nfs_release_request(req);
447         }
448 }
449
450 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
451 {
452         get_dreq(hdr->dreq);
453 }
454
455 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
456         .error_cleanup = nfs_read_sync_pgio_error,
457         .init_hdr = nfs_direct_pgio_init,
458         .completion = nfs_direct_read_completion,
459 };
460
461 /*
462  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
463  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
464  * bail and stop sending more reads.  Read length accounting is
465  * handled automatically by nfs_direct_read_result().  Otherwise, if
466  * no requests have been sent, just return an error.
467  */
468
469 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
470                                               struct iov_iter *iter,
471                                               loff_t pos)
472 {
473         struct nfs_pageio_descriptor desc;
474         struct inode *inode = dreq->inode;
475         ssize_t result = -EINVAL;
476         size_t requested_bytes = 0;
477         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
478
479         nfs_pageio_init_read(&desc, dreq->inode, false,
480                              &nfs_direct_read_completion_ops);
481         get_dreq(dreq);
482         desc.pg_dreq = dreq;
483         inode_dio_begin(inode);
484
485         while (iov_iter_count(iter)) {
486                 struct page **pagevec;
487                 size_t bytes;
488                 size_t pgbase;
489                 unsigned npages, i;
490
491                 result = iov_iter_get_pages_alloc(iter, &pagevec, 
492                                                   rsize, &pgbase);
493                 if (result < 0)
494                         break;
495         
496                 bytes = result;
497                 iov_iter_advance(iter, bytes);
498                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
499                 for (i = 0; i < npages; i++) {
500                         struct nfs_page *req;
501                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
502                         /* XXX do we need to do the eof zeroing found in async_filler? */
503                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
504                                                  pgbase, req_len);
505                         if (IS_ERR(req)) {
506                                 result = PTR_ERR(req);
507                                 break;
508                         }
509                         req->wb_index = pos >> PAGE_SHIFT;
510                         req->wb_offset = pos & ~PAGE_MASK;
511                         if (!nfs_pageio_add_request(&desc, req)) {
512                                 result = desc.pg_error;
513                                 nfs_release_request(req);
514                                 break;
515                         }
516                         pgbase = 0;
517                         bytes -= req_len;
518                         requested_bytes += req_len;
519                         pos += req_len;
520                         dreq->bytes_left -= req_len;
521                 }
522                 nfs_direct_release_pages(pagevec, npages);
523                 kvfree(pagevec);
524                 if (result < 0)
525                         break;
526         }
527
528         nfs_pageio_complete(&desc);
529
530         /*
531          * If no bytes were started, return the error, and let the
532          * generic layer handle the completion.
533          */
534         if (requested_bytes == 0) {
535                 inode_dio_end(inode);
536                 nfs_direct_req_release(dreq);
537                 return result < 0 ? result : -EIO;
538         }
539
540         if (put_dreq(dreq))
541                 nfs_direct_complete(dreq);
542         return 0;
543 }
544
545 /**
546  * nfs_file_direct_read - file direct read operation for NFS files
547  * @iocb: target I/O control block
548  * @iter: vector of user buffers into which to read data
549  *
550  * We use this function for direct reads instead of calling
551  * generic_file_aio_read() in order to avoid gfar's check to see if
552  * the request starts before the end of the file.  For that check
553  * to work, we must generate a GETATTR before each direct read, and
554  * even then there is a window between the GETATTR and the subsequent
555  * READ where the file size could change.  Our preference is simply
556  * to do all reads the application wants, and the server will take
557  * care of managing the end of file boundary.
558  *
559  * This function also eliminates unnecessarily updating the file's
560  * atime locally, as the NFS server sets the file's atime, and this
561  * client must read the updated atime from the server back into its
562  * cache.
563  */
564 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter)
565 {
566         struct file *file = iocb->ki_filp;
567         struct address_space *mapping = file->f_mapping;
568         struct inode *inode = mapping->host;
569         struct nfs_direct_req *dreq;
570         struct nfs_lock_context *l_ctx;
571         ssize_t result = -EINVAL;
572         size_t count = iov_iter_count(iter);
573         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
574
575         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
576                 file, count, (long long) iocb->ki_pos);
577
578         result = 0;
579         if (!count)
580                 goto out;
581
582         task_io_account_read(count);
583
584         result = -ENOMEM;
585         dreq = nfs_direct_req_alloc();
586         if (dreq == NULL)
587                 goto out;
588
589         dreq->inode = inode;
590         dreq->bytes_left = dreq->max_count = count;
591         dreq->io_start = iocb->ki_pos;
592         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
593         l_ctx = nfs_get_lock_context(dreq->ctx);
594         if (IS_ERR(l_ctx)) {
595                 result = PTR_ERR(l_ctx);
596                 goto out_release;
597         }
598         dreq->l_ctx = l_ctx;
599         if (!is_sync_kiocb(iocb))
600                 dreq->iocb = iocb;
601
602         nfs_start_io_direct(inode);
603
604         NFS_I(inode)->read_io += count;
605         result = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
606
607         nfs_end_io_direct(inode);
608
609         if (!result) {
610                 result = nfs_direct_wait(dreq);
611                 if (result > 0)
612                         iocb->ki_pos += result;
613         }
614
615 out_release:
616         nfs_direct_req_release(dreq);
617 out:
618         return result;
619 }
620
621 static void
622 nfs_direct_write_scan_commit_list(struct inode *inode,
623                                   struct list_head *list,
624                                   struct nfs_commit_info *cinfo)
625 {
626         spin_lock(&cinfo->inode->i_lock);
627 #ifdef CONFIG_NFS_V4_1
628         if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
629                 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
630 #endif
631         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
632         spin_unlock(&cinfo->inode->i_lock);
633 }
634
635 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
636 {
637         struct nfs_pageio_descriptor desc;
638         struct nfs_page *req, *tmp;
639         LIST_HEAD(reqs);
640         struct nfs_commit_info cinfo;
641         LIST_HEAD(failed);
642         int i;
643
644         nfs_init_cinfo_from_dreq(&cinfo, dreq);
645         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
646
647         dreq->count = 0;
648         dreq->verf.committed = NFS_INVALID_STABLE_HOW;
649         nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
650         for (i = 0; i < dreq->mirror_count; i++)
651                 dreq->mirrors[i].count = 0;
652         get_dreq(dreq);
653
654         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
655                               &nfs_direct_write_completion_ops);
656         desc.pg_dreq = dreq;
657
658         req = nfs_list_entry(reqs.next);
659         nfs_direct_setup_mirroring(dreq, &desc, req);
660         if (desc.pg_error < 0) {
661                 list_splice_init(&reqs, &failed);
662                 goto out_failed;
663         }
664
665         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
666                 if (!nfs_pageio_add_request(&desc, req)) {
667                         nfs_list_remove_request(req);
668                         nfs_list_add_request(req, &failed);
669                         spin_lock(&cinfo.inode->i_lock);
670                         dreq->flags = 0;
671                         if (desc.pg_error < 0)
672                                 dreq->error = desc.pg_error;
673                         else
674                                 dreq->error = -EIO;
675                         spin_unlock(&cinfo.inode->i_lock);
676                 }
677                 nfs_release_request(req);
678         }
679         nfs_pageio_complete(&desc);
680
681 out_failed:
682         while (!list_empty(&failed)) {
683                 req = nfs_list_entry(failed.next);
684                 nfs_list_remove_request(req);
685                 nfs_unlock_and_release_request(req);
686         }
687
688         if (put_dreq(dreq))
689                 nfs_direct_write_complete(dreq, dreq->inode);
690 }
691
692 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
693 {
694         struct nfs_direct_req *dreq = data->dreq;
695         struct nfs_commit_info cinfo;
696         struct nfs_page *req;
697         int status = data->task.tk_status;
698
699         nfs_init_cinfo_from_dreq(&cinfo, dreq);
700         if (status < 0) {
701                 dprintk("NFS: %5u commit failed with error %d.\n",
702                         data->task.tk_pid, status);
703                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
704         } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
705                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
706                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
707         }
708
709         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
710         while (!list_empty(&data->pages)) {
711                 req = nfs_list_entry(data->pages.next);
712                 nfs_list_remove_request(req);
713                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
714                         /* Note the rewrite will go through mds */
715                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
716                 } else
717                         nfs_release_request(req);
718                 nfs_unlock_and_release_request(req);
719         }
720
721         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
722                 nfs_direct_write_complete(dreq, data->inode);
723 }
724
725 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
726                 struct nfs_page *req)
727 {
728         struct nfs_direct_req *dreq = cinfo->dreq;
729
730         spin_lock(&dreq->lock);
731         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
732         spin_unlock(&dreq->lock);
733         nfs_mark_request_commit(req, NULL, cinfo, 0);
734 }
735
736 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
737         .completion = nfs_direct_commit_complete,
738         .resched_write = nfs_direct_resched_write,
739 };
740
741 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
742 {
743         int res;
744         struct nfs_commit_info cinfo;
745         LIST_HEAD(mds_list);
746
747         nfs_init_cinfo_from_dreq(&cinfo, dreq);
748         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
749         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
750         if (res < 0) /* res == -ENOMEM */
751                 nfs_direct_write_reschedule(dreq);
752 }
753
754 static void nfs_direct_write_schedule_work(struct work_struct *work)
755 {
756         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
757         int flags = dreq->flags;
758
759         dreq->flags = 0;
760         switch (flags) {
761                 case NFS_ODIRECT_DO_COMMIT:
762                         nfs_direct_commit_schedule(dreq);
763                         break;
764                 case NFS_ODIRECT_RESCHED_WRITES:
765                         nfs_direct_write_reschedule(dreq);
766                         break;
767                 default:
768                         nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
769                         nfs_direct_complete(dreq);
770         }
771 }
772
773 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
774 {
775         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
776 }
777
778 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
779 {
780         struct nfs_direct_req *dreq = hdr->dreq;
781         struct nfs_commit_info cinfo;
782         bool request_commit = false;
783         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
784
785         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
786                 goto out_put;
787
788         nfs_init_cinfo_from_dreq(&cinfo, dreq);
789
790         spin_lock(&dreq->lock);
791
792         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
793                 dreq->flags = 0;
794                 dreq->error = hdr->error;
795         }
796         if (dreq->error == 0) {
797                 nfs_direct_good_bytes(dreq, hdr);
798                 if (nfs_write_need_commit(hdr)) {
799                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
800                                 request_commit = true;
801                         else if (dreq->flags == 0) {
802                                 nfs_direct_set_hdr_verf(dreq, hdr);
803                                 request_commit = true;
804                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
805                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
806                                 request_commit = true;
807                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
808                                         dreq->flags =
809                                                 NFS_ODIRECT_RESCHED_WRITES;
810                         }
811                 }
812         }
813         spin_unlock(&dreq->lock);
814
815         while (!list_empty(&hdr->pages)) {
816
817                 req = nfs_list_entry(hdr->pages.next);
818                 nfs_list_remove_request(req);
819                 if (request_commit) {
820                         kref_get(&req->wb_kref);
821                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
822                                 hdr->ds_commit_idx);
823                 }
824                 nfs_unlock_and_release_request(req);
825         }
826
827 out_put:
828         if (put_dreq(dreq))
829                 nfs_direct_write_complete(dreq, hdr->inode);
830         hdr->release(hdr);
831 }
832
833 static void nfs_write_sync_pgio_error(struct list_head *head)
834 {
835         struct nfs_page *req;
836
837         while (!list_empty(head)) {
838                 req = nfs_list_entry(head->next);
839                 nfs_list_remove_request(req);
840                 nfs_unlock_and_release_request(req);
841         }
842 }
843
844 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
845 {
846         struct nfs_direct_req *dreq = hdr->dreq;
847
848         spin_lock(&dreq->lock);
849         if (dreq->error == 0) {
850                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
851                 /* fake unstable write to let common nfs resend pages */
852                 hdr->verf.committed = NFS_UNSTABLE;
853                 hdr->good_bytes = hdr->args.count;
854         }
855         spin_unlock(&dreq->lock);
856 }
857
858 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
859         .error_cleanup = nfs_write_sync_pgio_error,
860         .init_hdr = nfs_direct_pgio_init,
861         .completion = nfs_direct_write_completion,
862         .reschedule_io = nfs_direct_write_reschedule_io,
863 };
864
865
866 /*
867  * NB: Return the value of the first error return code.  Subsequent
868  *     errors after the first one are ignored.
869  */
870 /*
871  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
872  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
873  * bail and stop sending more writes.  Write length accounting is
874  * handled automatically by nfs_direct_write_result().  Otherwise, if
875  * no requests have been sent, just return an error.
876  */
877 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
878                                                struct iov_iter *iter,
879                                                loff_t pos)
880 {
881         struct nfs_pageio_descriptor desc;
882         struct inode *inode = dreq->inode;
883         ssize_t result = 0;
884         size_t requested_bytes = 0;
885         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
886
887         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
888                               &nfs_direct_write_completion_ops);
889         desc.pg_dreq = dreq;
890         get_dreq(dreq);
891         inode_dio_begin(inode);
892
893         NFS_I(inode)->write_io += iov_iter_count(iter);
894         while (iov_iter_count(iter)) {
895                 struct page **pagevec;
896                 size_t bytes;
897                 size_t pgbase;
898                 unsigned npages, i;
899
900                 result = iov_iter_get_pages_alloc(iter, &pagevec, 
901                                                   wsize, &pgbase);
902                 if (result < 0)
903                         break;
904
905                 bytes = result;
906                 iov_iter_advance(iter, bytes);
907                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
908                 for (i = 0; i < npages; i++) {
909                         struct nfs_page *req;
910                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
911
912                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
913                                                  pgbase, req_len);
914                         if (IS_ERR(req)) {
915                                 result = PTR_ERR(req);
916                                 break;
917                         }
918
919                         nfs_direct_setup_mirroring(dreq, &desc, req);
920                         if (desc.pg_error < 0) {
921                                 nfs_free_request(req);
922                                 result = desc.pg_error;
923                                 break;
924                         }
925
926                         nfs_lock_request(req);
927                         req->wb_index = pos >> PAGE_SHIFT;
928                         req->wb_offset = pos & ~PAGE_MASK;
929                         if (!nfs_pageio_add_request(&desc, req)) {
930                                 result = desc.pg_error;
931                                 nfs_unlock_and_release_request(req);
932                                 break;
933                         }
934                         pgbase = 0;
935                         bytes -= req_len;
936                         requested_bytes += req_len;
937                         pos += req_len;
938                         dreq->bytes_left -= req_len;
939                 }
940                 nfs_direct_release_pages(pagevec, npages);
941                 kvfree(pagevec);
942                 if (result < 0)
943                         break;
944         }
945         nfs_pageio_complete(&desc);
946
947         /*
948          * If no bytes were started, return the error, and let the
949          * generic layer handle the completion.
950          */
951         if (requested_bytes == 0) {
952                 inode_dio_end(inode);
953                 nfs_direct_req_release(dreq);
954                 return result < 0 ? result : -EIO;
955         }
956
957         if (put_dreq(dreq))
958                 nfs_direct_write_complete(dreq, dreq->inode);
959         return 0;
960 }
961
962 /**
963  * nfs_file_direct_write - file direct write operation for NFS files
964  * @iocb: target I/O control block
965  * @iter: vector of user buffers from which to write data
966  *
967  * We use this function for direct writes instead of calling
968  * generic_file_aio_write() in order to avoid taking the inode
969  * semaphore and updating the i_size.  The NFS server will set
970  * the new i_size and this client must read the updated size
971  * back into its cache.  We let the server do generic write
972  * parameter checking and report problems.
973  *
974  * We eliminate local atime updates, see direct read above.
975  *
976  * We avoid unnecessary page cache invalidations for normal cached
977  * readers of this file.
978  *
979  * Note that O_APPEND is not supported for NFS direct writes, as there
980  * is no atomic O_APPEND write facility in the NFS protocol.
981  */
982 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
983 {
984         ssize_t result = -EINVAL;
985         size_t count;
986         struct file *file = iocb->ki_filp;
987         struct address_space *mapping = file->f_mapping;
988         struct inode *inode = mapping->host;
989         struct nfs_direct_req *dreq;
990         struct nfs_lock_context *l_ctx;
991         loff_t pos, end;
992
993         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
994                 file, iov_iter_count(iter), (long long) iocb->ki_pos);
995
996         result = generic_write_checks(iocb, iter);
997         if (result <= 0)
998                 return result;
999         count = result;
1000         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
1001
1002         pos = iocb->ki_pos;
1003         end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
1004
1005         task_io_account_write(count);
1006
1007         result = -ENOMEM;
1008         dreq = nfs_direct_req_alloc();
1009         if (!dreq)
1010                 goto out;
1011
1012         dreq->inode = inode;
1013         dreq->bytes_left = dreq->max_count = count;
1014         dreq->io_start = pos;
1015         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1016         l_ctx = nfs_get_lock_context(dreq->ctx);
1017         if (IS_ERR(l_ctx)) {
1018                 result = PTR_ERR(l_ctx);
1019                 goto out_release;
1020         }
1021         dreq->l_ctx = l_ctx;
1022         if (!is_sync_kiocb(iocb))
1023                 dreq->iocb = iocb;
1024
1025         nfs_start_io_direct(inode);
1026
1027         result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1028
1029         if (mapping->nrpages) {
1030                 invalidate_inode_pages2_range(mapping,
1031                                               pos >> PAGE_SHIFT, end);
1032         }
1033
1034         nfs_end_io_direct(inode);
1035
1036         if (!result) {
1037                 result = nfs_direct_wait(dreq);
1038                 if (result > 0) {
1039                         iocb->ki_pos = pos + result;
1040                         /* XXX: should check the generic_write_sync retval */
1041                         generic_write_sync(iocb, result);
1042                 }
1043         }
1044 out_release:
1045         nfs_direct_req_release(dreq);
1046 out:
1047         return result;
1048 }
1049
1050 /**
1051  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1052  *
1053  */
1054 int __init nfs_init_directcache(void)
1055 {
1056         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1057                                                 sizeof(struct nfs_direct_req),
1058                                                 0, (SLAB_RECLAIM_ACCOUNT|
1059                                                         SLAB_MEM_SPREAD),
1060                                                 NULL);
1061         if (nfs_direct_cachep == NULL)
1062                 return -ENOMEM;
1063
1064         return 0;
1065 }
1066
1067 /**
1068  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1069  *
1070  */
1071 void nfs_destroy_directcache(void)
1072 {
1073         kmem_cache_destroy(nfs_direct_cachep);
1074 }