b555efca01d204adc66380831279162100d58141
[releases.git] / direct.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * linux/fs/nfs/direct.c
4  *
5  * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
6  *
7  * High-performance uncached I/O for the Linux NFS client
8  *
9  * There are important applications whose performance or correctness
10  * depends on uncached access to file data.  Database clusters
11  * (multiple copies of the same instance running on separate hosts)
12  * implement their own cache coherency protocol that subsumes file
13  * system cache protocols.  Applications that process datasets
14  * considerably larger than the client's memory do not always benefit
15  * from a local cache.  A streaming video server, for instance, has no
16  * need to cache the contents of a file.
17  *
18  * When an application requests uncached I/O, all read and write requests
19  * are made directly to the server; data stored or fetched via these
20  * requests is not cached in the Linux page cache.  The client does not
21  * correct unaligned requests from applications.  All requested bytes are
22  * held on permanent storage before a direct write system call returns to
23  * an application.
24  *
25  * Solaris implements an uncached I/O facility called directio() that
26  * is used for backups and sequential I/O to very large files.  Solaris
27  * also supports uncaching whole NFS partitions with "-o forcedirectio,"
28  * an undocumented mount option.
29  *
30  * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
31  * help from Andrew Morton.
32  *
33  * 18 Dec 2001  Initial implementation for 2.4  --cel
34  * 08 Jul 2002  Version for 2.4.19, with bug fixes --trondmy
35  * 08 Jun 2003  Port to 2.5 APIs  --cel
36  * 31 Mar 2004  Handle direct I/O without VFS support  --cel
37  * 15 Sep 2004  Parallel async reads  --cel
38  * 04 May 2005  support O_DIRECT with aio  --cel
39  *
40  */
41
42 #include <linux/errno.h>
43 #include <linux/sched.h>
44 #include <linux/kernel.h>
45 #include <linux/file.h>
46 #include <linux/pagemap.h>
47 #include <linux/kref.h>
48 #include <linux/slab.h>
49 #include <linux/task_io_accounting_ops.h>
50 #include <linux/module.h>
51
52 #include <linux/nfs_fs.h>
53 #include <linux/nfs_page.h>
54 #include <linux/sunrpc/clnt.h>
55
56 #include <linux/uaccess.h>
57 #include <linux/atomic.h>
58
59 #include "internal.h"
60 #include "iostat.h"
61 #include "pnfs.h"
62 #include "fscache.h"
63 #include "nfstrace.h"
64
65 #define NFSDBG_FACILITY         NFSDBG_VFS
66
67 static struct kmem_cache *nfs_direct_cachep;
68
69 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
70 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
71 static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
72 static void nfs_direct_write_schedule_work(struct work_struct *work);
73
74 static inline void get_dreq(struct nfs_direct_req *dreq)
75 {
76         atomic_inc(&dreq->io_count);
77 }
78
79 static inline int put_dreq(struct nfs_direct_req *dreq)
80 {
81         return atomic_dec_and_test(&dreq->io_count);
82 }
83
84 static void
85 nfs_direct_handle_truncated(struct nfs_direct_req *dreq,
86                             const struct nfs_pgio_header *hdr,
87                             ssize_t dreq_len)
88 {
89         if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) ||
90               test_bit(NFS_IOHDR_EOF, &hdr->flags)))
91                 return;
92         if (dreq->max_count >= dreq_len) {
93                 dreq->max_count = dreq_len;
94                 if (dreq->count > dreq_len)
95                         dreq->count = dreq_len;
96
97                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
98                         dreq->error = hdr->error;
99                 else /* Clear outstanding error if this is EOF */
100                         dreq->error = 0;
101         }
102 }
103
104 static void
105 nfs_direct_count_bytes(struct nfs_direct_req *dreq,
106                        const struct nfs_pgio_header *hdr)
107 {
108         loff_t hdr_end = hdr->io_start + hdr->good_bytes;
109         ssize_t dreq_len = 0;
110
111         if (hdr_end > dreq->io_start)
112                 dreq_len = hdr_end - dreq->io_start;
113
114         nfs_direct_handle_truncated(dreq, hdr, dreq_len);
115
116         if (dreq_len > dreq->max_count)
117                 dreq_len = dreq->max_count;
118
119         if (dreq->count < dreq_len)
120                 dreq->count = dreq_len;
121 }
122
123 /**
124  * nfs_swap_rw - NFS address space operation for swap I/O
125  * @iocb: target I/O control block
126  * @iter: I/O buffer
127  *
128  * Perform IO to the swap-file.  This is much like direct IO.
129  */
130 int nfs_swap_rw(struct kiocb *iocb, struct iov_iter *iter)
131 {
132         ssize_t ret;
133
134         VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
135
136         if (iov_iter_rw(iter) == READ)
137                 ret = nfs_file_direct_read(iocb, iter, true);
138         else
139                 ret = nfs_file_direct_write(iocb, iter, true);
140         if (ret < 0)
141                 return ret;
142         return 0;
143 }
144
145 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
146 {
147         unsigned int i;
148         for (i = 0; i < npages; i++)
149                 put_page(pages[i]);
150 }
151
152 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
153                               struct nfs_direct_req *dreq)
154 {
155         cinfo->inode = dreq->inode;
156         cinfo->mds = &dreq->mds_cinfo;
157         cinfo->ds = &dreq->ds_cinfo;
158         cinfo->dreq = dreq;
159         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
160 }
161
162 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
163 {
164         struct nfs_direct_req *dreq;
165
166         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
167         if (!dreq)
168                 return NULL;
169
170         kref_init(&dreq->kref);
171         kref_get(&dreq->kref);
172         init_completion(&dreq->completion);
173         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
174         pnfs_init_ds_commit_info(&dreq->ds_cinfo);
175         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
176         spin_lock_init(&dreq->lock);
177
178         return dreq;
179 }
180
181 static void nfs_direct_req_free(struct kref *kref)
182 {
183         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
184
185         pnfs_release_ds_info(&dreq->ds_cinfo, dreq->inode);
186         if (dreq->l_ctx != NULL)
187                 nfs_put_lock_context(dreq->l_ctx);
188         if (dreq->ctx != NULL)
189                 put_nfs_open_context(dreq->ctx);
190         kmem_cache_free(nfs_direct_cachep, dreq);
191 }
192
193 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
194 {
195         kref_put(&dreq->kref, nfs_direct_req_free);
196 }
197
198 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq, loff_t offset)
199 {
200         loff_t start = offset - dreq->io_start;
201         return dreq->max_count - start;
202 }
203 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
204
205 /*
206  * Collects and returns the final error value/byte-count.
207  */
208 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
209 {
210         ssize_t result = -EIOCBQUEUED;
211
212         /* Async requests don't wait here */
213         if (dreq->iocb)
214                 goto out;
215
216         result = wait_for_completion_killable(&dreq->completion);
217
218         if (!result) {
219                 result = dreq->count;
220                 WARN_ON_ONCE(dreq->count < 0);
221         }
222         if (!result)
223                 result = dreq->error;
224
225 out:
226         return (ssize_t) result;
227 }
228
229 /*
230  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
231  * the iocb is still valid here if this is a synchronous request.
232  */
233 static void nfs_direct_complete(struct nfs_direct_req *dreq)
234 {
235         struct inode *inode = dreq->inode;
236
237         inode_dio_end(inode);
238
239         if (dreq->iocb) {
240                 long res = (long) dreq->error;
241                 if (dreq->count != 0) {
242                         res = (long) dreq->count;
243                         WARN_ON_ONCE(dreq->count < 0);
244                 }
245                 dreq->iocb->ki_complete(dreq->iocb, res);
246         }
247
248         complete(&dreq->completion);
249
250         nfs_direct_req_release(dreq);
251 }
252
253 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
254 {
255         unsigned long bytes = 0;
256         struct nfs_direct_req *dreq = hdr->dreq;
257
258         spin_lock(&dreq->lock);
259         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
260                 spin_unlock(&dreq->lock);
261                 goto out_put;
262         }
263
264         nfs_direct_count_bytes(dreq, hdr);
265         spin_unlock(&dreq->lock);
266
267         while (!list_empty(&hdr->pages)) {
268                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
269                 struct page *page = req->wb_page;
270
271                 if (!PageCompound(page) && bytes < hdr->good_bytes &&
272                     (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
273                         set_page_dirty(page);
274                 bytes += req->wb_bytes;
275                 nfs_list_remove_request(req);
276                 nfs_release_request(req);
277         }
278 out_put:
279         if (put_dreq(dreq))
280                 nfs_direct_complete(dreq);
281         hdr->release(hdr);
282 }
283
284 static void nfs_read_sync_pgio_error(struct list_head *head, int error)
285 {
286         struct nfs_page *req;
287
288         while (!list_empty(head)) {
289                 req = nfs_list_entry(head->next);
290                 nfs_list_remove_request(req);
291                 nfs_release_request(req);
292         }
293 }
294
295 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
296 {
297         get_dreq(hdr->dreq);
298 }
299
300 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
301         .error_cleanup = nfs_read_sync_pgio_error,
302         .init_hdr = nfs_direct_pgio_init,
303         .completion = nfs_direct_read_completion,
304 };
305
306 /*
307  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
308  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
309  * bail and stop sending more reads.  Read length accounting is
310  * handled automatically by nfs_direct_read_result().  Otherwise, if
311  * no requests have been sent, just return an error.
312  */
313
314 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
315                                               struct iov_iter *iter,
316                                               loff_t pos)
317 {
318         struct nfs_pageio_descriptor desc;
319         struct inode *inode = dreq->inode;
320         ssize_t result = -EINVAL;
321         size_t requested_bytes = 0;
322         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
323
324         nfs_pageio_init_read(&desc, dreq->inode, false,
325                              &nfs_direct_read_completion_ops);
326         get_dreq(dreq);
327         desc.pg_dreq = dreq;
328         inode_dio_begin(inode);
329
330         while (iov_iter_count(iter)) {
331                 struct page **pagevec;
332                 size_t bytes;
333                 size_t pgbase;
334                 unsigned npages, i;
335
336                 result = iov_iter_get_pages_alloc2(iter, &pagevec,
337                                                   rsize, &pgbase);
338                 if (result < 0)
339                         break;
340         
341                 bytes = result;
342                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
343                 for (i = 0; i < npages; i++) {
344                         struct nfs_page *req;
345                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
346                         /* XXX do we need to do the eof zeroing found in async_filler? */
347                         req = nfs_create_request(dreq->ctx, pagevec[i],
348                                                  pgbase, req_len);
349                         if (IS_ERR(req)) {
350                                 result = PTR_ERR(req);
351                                 break;
352                         }
353                         req->wb_index = pos >> PAGE_SHIFT;
354                         req->wb_offset = pos & ~PAGE_MASK;
355                         if (!nfs_pageio_add_request(&desc, req)) {
356                                 result = desc.pg_error;
357                                 nfs_release_request(req);
358                                 break;
359                         }
360                         pgbase = 0;
361                         bytes -= req_len;
362                         requested_bytes += req_len;
363                         pos += req_len;
364                         dreq->bytes_left -= req_len;
365                 }
366                 nfs_direct_release_pages(pagevec, npages);
367                 kvfree(pagevec);
368                 if (result < 0)
369                         break;
370         }
371
372         nfs_pageio_complete(&desc);
373
374         /*
375          * If no bytes were started, return the error, and let the
376          * generic layer handle the completion.
377          */
378         if (requested_bytes == 0) {
379                 inode_dio_end(inode);
380                 nfs_direct_req_release(dreq);
381                 return result < 0 ? result : -EIO;
382         }
383
384         if (put_dreq(dreq))
385                 nfs_direct_complete(dreq);
386         return requested_bytes;
387 }
388
389 /**
390  * nfs_file_direct_read - file direct read operation for NFS files
391  * @iocb: target I/O control block
392  * @iter: vector of user buffers into which to read data
393  * @swap: flag indicating this is swap IO, not O_DIRECT IO
394  *
395  * We use this function for direct reads instead of calling
396  * generic_file_aio_read() in order to avoid gfar's check to see if
397  * the request starts before the end of the file.  For that check
398  * to work, we must generate a GETATTR before each direct read, and
399  * even then there is a window between the GETATTR and the subsequent
400  * READ where the file size could change.  Our preference is simply
401  * to do all reads the application wants, and the server will take
402  * care of managing the end of file boundary.
403  *
404  * This function also eliminates unnecessarily updating the file's
405  * atime locally, as the NFS server sets the file's atime, and this
406  * client must read the updated atime from the server back into its
407  * cache.
408  */
409 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
410                              bool swap)
411 {
412         struct file *file = iocb->ki_filp;
413         struct address_space *mapping = file->f_mapping;
414         struct inode *inode = mapping->host;
415         struct nfs_direct_req *dreq;
416         struct nfs_lock_context *l_ctx;
417         ssize_t result, requested;
418         size_t count = iov_iter_count(iter);
419         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
420
421         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
422                 file, count, (long long) iocb->ki_pos);
423
424         result = 0;
425         if (!count)
426                 goto out;
427
428         task_io_account_read(count);
429
430         result = -ENOMEM;
431         dreq = nfs_direct_req_alloc();
432         if (dreq == NULL)
433                 goto out;
434
435         dreq->inode = inode;
436         dreq->bytes_left = dreq->max_count = count;
437         dreq->io_start = iocb->ki_pos;
438         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
439         l_ctx = nfs_get_lock_context(dreq->ctx);
440         if (IS_ERR(l_ctx)) {
441                 result = PTR_ERR(l_ctx);
442                 nfs_direct_req_release(dreq);
443                 goto out_release;
444         }
445         dreq->l_ctx = l_ctx;
446         if (!is_sync_kiocb(iocb))
447                 dreq->iocb = iocb;
448
449         if (user_backed_iter(iter))
450                 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
451
452         if (!swap)
453                 nfs_start_io_direct(inode);
454
455         NFS_I(inode)->read_io += count;
456         requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
457
458         if (!swap)
459                 nfs_end_io_direct(inode);
460
461         if (requested > 0) {
462                 result = nfs_direct_wait(dreq);
463                 if (result > 0) {
464                         requested -= result;
465                         iocb->ki_pos += result;
466                 }
467                 iov_iter_revert(iter, requested);
468         } else {
469                 result = requested;
470         }
471
472 out_release:
473         nfs_direct_req_release(dreq);
474 out:
475         return result;
476 }
477
478 static void nfs_direct_add_page_head(struct list_head *list,
479                                      struct nfs_page *req)
480 {
481         struct nfs_page *head = req->wb_head;
482
483         if (!list_empty(&head->wb_list) || !nfs_lock_request(head))
484                 return;
485         if (!list_empty(&head->wb_list)) {
486                 nfs_unlock_request(head);
487                 return;
488         }
489         list_add(&head->wb_list, list);
490         kref_get(&head->wb_kref);
491         kref_get(&head->wb_kref);
492 }
493
494 static void nfs_direct_join_group(struct list_head *list, struct inode *inode)
495 {
496         struct nfs_page *req, *subreq;
497
498         list_for_each_entry(req, list, wb_list) {
499                 if (req->wb_head != req) {
500                         nfs_direct_add_page_head(&req->wb_list, req);
501                         continue;
502                 }
503                 subreq = req->wb_this_page;
504                 if (subreq == req)
505                         continue;
506                 do {
507                         /*
508                          * Remove subrequests from this list before freeing
509                          * them in the call to nfs_join_page_group().
510                          */
511                         if (!list_empty(&subreq->wb_list)) {
512                                 nfs_list_remove_request(subreq);
513                                 nfs_release_request(subreq);
514                         }
515                 } while ((subreq = subreq->wb_this_page) != req);
516                 nfs_join_page_group(req, inode);
517         }
518 }
519
520 static void
521 nfs_direct_write_scan_commit_list(struct inode *inode,
522                                   struct list_head *list,
523                                   struct nfs_commit_info *cinfo)
524 {
525         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
526         pnfs_recover_commit_reqs(list, cinfo);
527         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
528         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
529 }
530
531 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
532 {
533         struct nfs_pageio_descriptor desc;
534         struct nfs_page *req, *tmp;
535         LIST_HEAD(reqs);
536         struct nfs_commit_info cinfo;
537         LIST_HEAD(failed);
538
539         nfs_init_cinfo_from_dreq(&cinfo, dreq);
540         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
541
542         nfs_direct_join_group(&reqs, dreq->inode);
543
544         dreq->count = 0;
545         dreq->max_count = 0;
546         list_for_each_entry(req, &reqs, wb_list)
547                 dreq->max_count += req->wb_bytes;
548         nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
549         get_dreq(dreq);
550
551         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
552                               &nfs_direct_write_completion_ops);
553         desc.pg_dreq = dreq;
554
555         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
556                 /* Bump the transmission count */
557                 req->wb_nio++;
558                 if (!nfs_pageio_add_request(&desc, req)) {
559                         nfs_list_move_request(req, &failed);
560                         spin_lock(&cinfo.inode->i_lock);
561                         dreq->flags = 0;
562                         if (desc.pg_error < 0)
563                                 dreq->error = desc.pg_error;
564                         else
565                                 dreq->error = -EIO;
566                         spin_unlock(&cinfo.inode->i_lock);
567                 }
568                 nfs_release_request(req);
569         }
570         nfs_pageio_complete(&desc);
571
572         while (!list_empty(&failed)) {
573                 req = nfs_list_entry(failed.next);
574                 nfs_list_remove_request(req);
575                 nfs_unlock_and_release_request(req);
576         }
577
578         if (put_dreq(dreq))
579                 nfs_direct_write_complete(dreq);
580 }
581
582 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
583 {
584         const struct nfs_writeverf *verf = data->res.verf;
585         struct nfs_direct_req *dreq = data->dreq;
586         struct nfs_commit_info cinfo;
587         struct nfs_page *req;
588         int status = data->task.tk_status;
589
590         trace_nfs_direct_commit_complete(dreq);
591
592         if (status < 0) {
593                 /* Errors in commit are fatal */
594                 dreq->error = status;
595                 dreq->max_count = 0;
596                 dreq->count = 0;
597                 dreq->flags = NFS_ODIRECT_DONE;
598         } else {
599                 status = dreq->error;
600         }
601
602         nfs_init_cinfo_from_dreq(&cinfo, dreq);
603
604         while (!list_empty(&data->pages)) {
605                 req = nfs_list_entry(data->pages.next);
606                 nfs_list_remove_request(req);
607                 if (status >= 0 && !nfs_write_match_verf(verf, req)) {
608                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
609                         /*
610                          * Despite the reboot, the write was successful,
611                          * so reset wb_nio.
612                          */
613                         req->wb_nio = 0;
614                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
615                 } else /* Error or match */
616                         nfs_release_request(req);
617                 nfs_unlock_and_release_request(req);
618         }
619
620         if (nfs_commit_end(cinfo.mds))
621                 nfs_direct_write_complete(dreq);
622 }
623
624 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
625                 struct nfs_page *req)
626 {
627         struct nfs_direct_req *dreq = cinfo->dreq;
628
629         trace_nfs_direct_resched_write(dreq);
630
631         spin_lock(&dreq->lock);
632         if (dreq->flags != NFS_ODIRECT_DONE)
633                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
634         spin_unlock(&dreq->lock);
635         nfs_mark_request_commit(req, NULL, cinfo, 0);
636 }
637
638 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
639         .completion = nfs_direct_commit_complete,
640         .resched_write = nfs_direct_resched_write,
641 };
642
643 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
644 {
645         int res;
646         struct nfs_commit_info cinfo;
647         LIST_HEAD(mds_list);
648
649         nfs_init_cinfo_from_dreq(&cinfo, dreq);
650         nfs_commit_begin(cinfo.mds);
651         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
652         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
653         if (res < 0) { /* res == -ENOMEM */
654                 spin_lock(&dreq->lock);
655                 if (dreq->flags == 0)
656                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
657                 spin_unlock(&dreq->lock);
658         }
659         if (nfs_commit_end(cinfo.mds))
660                 nfs_direct_write_complete(dreq);
661 }
662
663 static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq)
664 {
665         struct nfs_commit_info cinfo;
666         struct nfs_page *req;
667         LIST_HEAD(reqs);
668
669         nfs_init_cinfo_from_dreq(&cinfo, dreq);
670         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
671
672         while (!list_empty(&reqs)) {
673                 req = nfs_list_entry(reqs.next);
674                 nfs_list_remove_request(req);
675                 nfs_release_request(req);
676                 nfs_unlock_and_release_request(req);
677         }
678 }
679
680 static void nfs_direct_write_schedule_work(struct work_struct *work)
681 {
682         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
683         int flags = dreq->flags;
684
685         dreq->flags = 0;
686         switch (flags) {
687                 case NFS_ODIRECT_DO_COMMIT:
688                         nfs_direct_commit_schedule(dreq);
689                         break;
690                 case NFS_ODIRECT_RESCHED_WRITES:
691                         nfs_direct_write_reschedule(dreq);
692                         break;
693                 default:
694                         nfs_direct_write_clear_reqs(dreq);
695                         nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
696                         nfs_direct_complete(dreq);
697         }
698 }
699
700 static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
701 {
702         trace_nfs_direct_write_complete(dreq);
703         queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
704 }
705
706 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
707 {
708         struct nfs_direct_req *dreq = hdr->dreq;
709         struct nfs_commit_info cinfo;
710         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
711         int flags = NFS_ODIRECT_DONE;
712
713         trace_nfs_direct_write_completion(dreq);
714
715         nfs_init_cinfo_from_dreq(&cinfo, dreq);
716
717         spin_lock(&dreq->lock);
718         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
719                 spin_unlock(&dreq->lock);
720                 goto out_put;
721         }
722
723         nfs_direct_count_bytes(dreq, hdr);
724         if (test_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags)) {
725                 if (!dreq->flags)
726                         dreq->flags = NFS_ODIRECT_DO_COMMIT;
727                 flags = dreq->flags;
728         }
729         spin_unlock(&dreq->lock);
730
731         while (!list_empty(&hdr->pages)) {
732
733                 req = nfs_list_entry(hdr->pages.next);
734                 nfs_list_remove_request(req);
735                 if (flags == NFS_ODIRECT_DO_COMMIT) {
736                         kref_get(&req->wb_kref);
737                         memcpy(&req->wb_verf, &hdr->verf.verifier,
738                                sizeof(req->wb_verf));
739                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
740                                 hdr->ds_commit_idx);
741                 } else if (flags == NFS_ODIRECT_RESCHED_WRITES) {
742                         kref_get(&req->wb_kref);
743                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
744                 }
745                 nfs_unlock_and_release_request(req);
746         }
747
748 out_put:
749         if (put_dreq(dreq))
750                 nfs_direct_write_complete(dreq);
751         hdr->release(hdr);
752 }
753
754 static void nfs_write_sync_pgio_error(struct list_head *head, int error)
755 {
756         struct nfs_page *req;
757
758         while (!list_empty(head)) {
759                 req = nfs_list_entry(head->next);
760                 nfs_list_remove_request(req);
761                 nfs_unlock_and_release_request(req);
762         }
763 }
764
765 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
766 {
767         struct nfs_direct_req *dreq = hdr->dreq;
768
769         trace_nfs_direct_write_reschedule_io(dreq);
770
771         spin_lock(&dreq->lock);
772         if (dreq->error == 0) {
773                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
774                 /* fake unstable write to let common nfs resend pages */
775                 hdr->verf.committed = NFS_UNSTABLE;
776                 hdr->good_bytes = hdr->args.offset + hdr->args.count -
777                         hdr->io_start;
778         }
779         spin_unlock(&dreq->lock);
780 }
781
782 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
783         .error_cleanup = nfs_write_sync_pgio_error,
784         .init_hdr = nfs_direct_pgio_init,
785         .completion = nfs_direct_write_completion,
786         .reschedule_io = nfs_direct_write_reschedule_io,
787 };
788
789
790 /*
791  * NB: Return the value of the first error return code.  Subsequent
792  *     errors after the first one are ignored.
793  */
794 /*
795  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
796  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
797  * bail and stop sending more writes.  Write length accounting is
798  * handled automatically by nfs_direct_write_result().  Otherwise, if
799  * no requests have been sent, just return an error.
800  */
801 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
802                                                struct iov_iter *iter,
803                                                loff_t pos, int ioflags)
804 {
805         struct nfs_pageio_descriptor desc;
806         struct inode *inode = dreq->inode;
807         ssize_t result = 0;
808         size_t requested_bytes = 0;
809         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
810
811         trace_nfs_direct_write_schedule_iovec(dreq);
812
813         nfs_pageio_init_write(&desc, inode, ioflags, false,
814                               &nfs_direct_write_completion_ops);
815         desc.pg_dreq = dreq;
816         get_dreq(dreq);
817         inode_dio_begin(inode);
818
819         NFS_I(inode)->write_io += iov_iter_count(iter);
820         while (iov_iter_count(iter)) {
821                 struct page **pagevec;
822                 size_t bytes;
823                 size_t pgbase;
824                 unsigned npages, i;
825
826                 result = iov_iter_get_pages_alloc2(iter, &pagevec,
827                                                   wsize, &pgbase);
828                 if (result < 0)
829                         break;
830
831                 bytes = result;
832                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
833                 for (i = 0; i < npages; i++) {
834                         struct nfs_page *req;
835                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
836
837                         req = nfs_create_request(dreq->ctx, pagevec[i],
838                                                  pgbase, req_len);
839                         if (IS_ERR(req)) {
840                                 result = PTR_ERR(req);
841                                 break;
842                         }
843
844                         if (desc.pg_error < 0) {
845                                 nfs_free_request(req);
846                                 result = desc.pg_error;
847                                 break;
848                         }
849
850                         nfs_lock_request(req);
851                         req->wb_index = pos >> PAGE_SHIFT;
852                         req->wb_offset = pos & ~PAGE_MASK;
853                         if (!nfs_pageio_add_request(&desc, req)) {
854                                 result = desc.pg_error;
855                                 nfs_unlock_and_release_request(req);
856                                 break;
857                         }
858                         pgbase = 0;
859                         bytes -= req_len;
860                         requested_bytes += req_len;
861                         pos += req_len;
862                         dreq->bytes_left -= req_len;
863                 }
864                 nfs_direct_release_pages(pagevec, npages);
865                 kvfree(pagevec);
866                 if (result < 0)
867                         break;
868         }
869         nfs_pageio_complete(&desc);
870
871         /*
872          * If no bytes were started, return the error, and let the
873          * generic layer handle the completion.
874          */
875         if (requested_bytes == 0) {
876                 inode_dio_end(inode);
877                 nfs_direct_req_release(dreq);
878                 return result < 0 ? result : -EIO;
879         }
880
881         if (put_dreq(dreq))
882                 nfs_direct_write_complete(dreq);
883         return requested_bytes;
884 }
885
886 /**
887  * nfs_file_direct_write - file direct write operation for NFS files
888  * @iocb: target I/O control block
889  * @iter: vector of user buffers from which to write data
890  * @swap: flag indicating this is swap IO, not O_DIRECT IO
891  *
892  * We use this function for direct writes instead of calling
893  * generic_file_aio_write() in order to avoid taking the inode
894  * semaphore and updating the i_size.  The NFS server will set
895  * the new i_size and this client must read the updated size
896  * back into its cache.  We let the server do generic write
897  * parameter checking and report problems.
898  *
899  * We eliminate local atime updates, see direct read above.
900  *
901  * We avoid unnecessary page cache invalidations for normal cached
902  * readers of this file.
903  *
904  * Note that O_APPEND is not supported for NFS direct writes, as there
905  * is no atomic O_APPEND write facility in the NFS protocol.
906  */
907 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
908                               bool swap)
909 {
910         ssize_t result, requested;
911         size_t count;
912         struct file *file = iocb->ki_filp;
913         struct address_space *mapping = file->f_mapping;
914         struct inode *inode = mapping->host;
915         struct nfs_direct_req *dreq;
916         struct nfs_lock_context *l_ctx;
917         loff_t pos, end;
918
919         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
920                 file, iov_iter_count(iter), (long long) iocb->ki_pos);
921
922         if (swap)
923                 /* bypass generic checks */
924                 result =  iov_iter_count(iter);
925         else
926                 result = generic_write_checks(iocb, iter);
927         if (result <= 0)
928                 return result;
929         count = result;
930         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
931
932         pos = iocb->ki_pos;
933         end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
934
935         task_io_account_write(count);
936
937         result = -ENOMEM;
938         dreq = nfs_direct_req_alloc();
939         if (!dreq)
940                 goto out;
941
942         dreq->inode = inode;
943         dreq->bytes_left = dreq->max_count = count;
944         dreq->io_start = pos;
945         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
946         l_ctx = nfs_get_lock_context(dreq->ctx);
947         if (IS_ERR(l_ctx)) {
948                 result = PTR_ERR(l_ctx);
949                 nfs_direct_req_release(dreq);
950                 goto out_release;
951         }
952         dreq->l_ctx = l_ctx;
953         if (!is_sync_kiocb(iocb))
954                 dreq->iocb = iocb;
955         pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode);
956
957         if (swap) {
958                 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
959                                                             FLUSH_STABLE);
960         } else {
961                 nfs_start_io_direct(inode);
962
963                 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
964                                                             FLUSH_COND_STABLE);
965
966                 if (mapping->nrpages) {
967                         invalidate_inode_pages2_range(mapping,
968                                                       pos >> PAGE_SHIFT, end);
969                 }
970
971                 nfs_end_io_direct(inode);
972         }
973
974         if (requested > 0) {
975                 result = nfs_direct_wait(dreq);
976                 if (result > 0) {
977                         requested -= result;
978                         iocb->ki_pos = pos + result;
979                         /* XXX: should check the generic_write_sync retval */
980                         generic_write_sync(iocb, result);
981                 }
982                 iov_iter_revert(iter, requested);
983         } else {
984                 result = requested;
985         }
986         nfs_fscache_invalidate(inode, FSCACHE_INVAL_DIO_WRITE);
987 out_release:
988         nfs_direct_req_release(dreq);
989 out:
990         return result;
991 }
992
993 /**
994  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
995  *
996  */
997 int __init nfs_init_directcache(void)
998 {
999         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1000                                                 sizeof(struct nfs_direct_req),
1001                                                 0, (SLAB_RECLAIM_ACCOUNT|
1002                                                         SLAB_MEM_SPREAD),
1003                                                 NULL);
1004         if (nfs_direct_cachep == NULL)
1005                 return -ENOMEM;
1006
1007         return 0;
1008 }
1009
1010 /**
1011  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1012  *
1013  */
1014 void nfs_destroy_directcache(void)
1015 {
1016         kmem_cache_destroy(nfs_direct_cachep);
1017 }