Merge tag 'ovl-update-5.1' of git://git.kernel.org/pub/scm/linux/kernel/git/mszeredi/vfs
[sfrench/cifs-2.6.git] / fs / nfs / pagelist.c
1 /*
2  * linux/fs/nfs/pagelist.c
3  *
4  * A set of helper functions for managing NFS read and write requests.
5  * The main purpose of these routines is to provide support for the
6  * coalescing of several requests into a single RPC call.
7  *
8  * Copyright 2000, 2001 (c) Trond Myklebust <trond.myklebust@fys.uio.no>
9  *
10  */
11
12 #include <linux/slab.h>
13 #include <linux/file.h>
14 #include <linux/sched.h>
15 #include <linux/sunrpc/clnt.h>
16 #include <linux/nfs.h>
17 #include <linux/nfs3.h>
18 #include <linux/nfs4.h>
19 #include <linux/nfs_page.h>
20 #include <linux/nfs_fs.h>
21 #include <linux/nfs_mount.h>
22 #include <linux/export.h>
23
24 #include "internal.h"
25 #include "pnfs.h"
26
27 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
28
29 static struct kmem_cache *nfs_page_cachep;
30 static const struct rpc_call_ops nfs_pgio_common_ops;
31
32 struct nfs_pgio_mirror *
33 nfs_pgio_current_mirror(struct nfs_pageio_descriptor *desc)
34 {
35         return nfs_pgio_has_mirroring(desc) ?
36                 &desc->pg_mirrors[desc->pg_mirror_idx] :
37                 &desc->pg_mirrors[0];
38 }
39 EXPORT_SYMBOL_GPL(nfs_pgio_current_mirror);
40
41 void nfs_pgheader_init(struct nfs_pageio_descriptor *desc,
42                        struct nfs_pgio_header *hdr,
43                        void (*release)(struct nfs_pgio_header *hdr))
44 {
45         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
46
47
48         hdr->req = nfs_list_entry(mirror->pg_list.next);
49         hdr->inode = desc->pg_inode;
50         hdr->cred = hdr->req->wb_context->cred;
51         hdr->io_start = req_offset(hdr->req);
52         hdr->good_bytes = mirror->pg_count;
53         hdr->io_completion = desc->pg_io_completion;
54         hdr->dreq = desc->pg_dreq;
55         hdr->release = release;
56         hdr->completion_ops = desc->pg_completion_ops;
57         if (hdr->completion_ops->init_hdr)
58                 hdr->completion_ops->init_hdr(hdr);
59
60         hdr->pgio_mirror_idx = desc->pg_mirror_idx;
61 }
62 EXPORT_SYMBOL_GPL(nfs_pgheader_init);
63
64 void nfs_set_pgio_error(struct nfs_pgio_header *hdr, int error, loff_t pos)
65 {
66         unsigned int new = pos - hdr->io_start;
67
68         if (hdr->good_bytes > new) {
69                 hdr->good_bytes = new;
70                 clear_bit(NFS_IOHDR_EOF, &hdr->flags);
71                 if (!test_and_set_bit(NFS_IOHDR_ERROR, &hdr->flags))
72                         hdr->error = error;
73         }
74 }
75
76 static inline struct nfs_page *
77 nfs_page_alloc(void)
78 {
79         struct nfs_page *p = kmem_cache_zalloc(nfs_page_cachep, GFP_NOIO);
80         if (p)
81                 INIT_LIST_HEAD(&p->wb_list);
82         return p;
83 }
84
85 static inline void
86 nfs_page_free(struct nfs_page *p)
87 {
88         kmem_cache_free(nfs_page_cachep, p);
89 }
90
91 /**
92  * nfs_iocounter_wait - wait for i/o to complete
93  * @l_ctx: nfs_lock_context with io_counter to use
94  *
95  * returns -ERESTARTSYS if interrupted by a fatal signal.
96  * Otherwise returns 0 once the io_count hits 0.
97  */
98 int
99 nfs_iocounter_wait(struct nfs_lock_context *l_ctx)
100 {
101         return wait_var_event_killable(&l_ctx->io_count,
102                                        !atomic_read(&l_ctx->io_count));
103 }
104
105 /**
106  * nfs_async_iocounter_wait - wait on a rpc_waitqueue for I/O
107  * to complete
108  * @task: the rpc_task that should wait
109  * @l_ctx: nfs_lock_context with io_counter to check
110  *
111  * Returns true if there is outstanding I/O to wait on and the
112  * task has been put to sleep.
113  */
114 bool
115 nfs_async_iocounter_wait(struct rpc_task *task, struct nfs_lock_context *l_ctx)
116 {
117         struct inode *inode = d_inode(l_ctx->open_context->dentry);
118         bool ret = false;
119
120         if (atomic_read(&l_ctx->io_count) > 0) {
121                 rpc_sleep_on(&NFS_SERVER(inode)->uoc_rpcwaitq, task, NULL);
122                 ret = true;
123         }
124
125         if (atomic_read(&l_ctx->io_count) == 0) {
126                 rpc_wake_up_queued_task(&NFS_SERVER(inode)->uoc_rpcwaitq, task);
127                 ret = false;
128         }
129
130         return ret;
131 }
132 EXPORT_SYMBOL_GPL(nfs_async_iocounter_wait);
133
134 /*
135  * nfs_page_group_lock - lock the head of the page group
136  * @req - request in group that is to be locked
137  *
138  * this lock must be held when traversing or modifying the page
139  * group list
140  *
141  * return 0 on success, < 0 on error
142  */
143 int
144 nfs_page_group_lock(struct nfs_page *req)
145 {
146         struct nfs_page *head = req->wb_head;
147
148         WARN_ON_ONCE(head != head->wb_head);
149
150         if (!test_and_set_bit(PG_HEADLOCK, &head->wb_flags))
151                 return 0;
152
153         set_bit(PG_CONTENDED1, &head->wb_flags);
154         smp_mb__after_atomic();
155         return wait_on_bit_lock(&head->wb_flags, PG_HEADLOCK,
156                                 TASK_UNINTERRUPTIBLE);
157 }
158
159 /*
160  * nfs_page_group_unlock - unlock the head of the page group
161  * @req - request in group that is to be unlocked
162  */
163 void
164 nfs_page_group_unlock(struct nfs_page *req)
165 {
166         struct nfs_page *head = req->wb_head;
167
168         WARN_ON_ONCE(head != head->wb_head);
169
170         smp_mb__before_atomic();
171         clear_bit(PG_HEADLOCK, &head->wb_flags);
172         smp_mb__after_atomic();
173         if (!test_bit(PG_CONTENDED1, &head->wb_flags))
174                 return;
175         wake_up_bit(&head->wb_flags, PG_HEADLOCK);
176 }
177
178 /*
179  * nfs_page_group_sync_on_bit_locked
180  *
181  * must be called with page group lock held
182  */
183 static bool
184 nfs_page_group_sync_on_bit_locked(struct nfs_page *req, unsigned int bit)
185 {
186         struct nfs_page *head = req->wb_head;
187         struct nfs_page *tmp;
188
189         WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_flags));
190         WARN_ON_ONCE(test_and_set_bit(bit, &req->wb_flags));
191
192         tmp = req->wb_this_page;
193         while (tmp != req) {
194                 if (!test_bit(bit, &tmp->wb_flags))
195                         return false;
196                 tmp = tmp->wb_this_page;
197         }
198
199         /* true! reset all bits */
200         tmp = req;
201         do {
202                 clear_bit(bit, &tmp->wb_flags);
203                 tmp = tmp->wb_this_page;
204         } while (tmp != req);
205
206         return true;
207 }
208
209 /*
210  * nfs_page_group_sync_on_bit - set bit on current request, but only
211  *   return true if the bit is set for all requests in page group
212  * @req - request in page group
213  * @bit - PG_* bit that is used to sync page group
214  */
215 bool nfs_page_group_sync_on_bit(struct nfs_page *req, unsigned int bit)
216 {
217         bool ret;
218
219         nfs_page_group_lock(req);
220         ret = nfs_page_group_sync_on_bit_locked(req, bit);
221         nfs_page_group_unlock(req);
222
223         return ret;
224 }
225
226 /*
227  * nfs_page_group_init - Initialize the page group linkage for @req
228  * @req - a new nfs request
229  * @prev - the previous request in page group, or NULL if @req is the first
230  *         or only request in the group (the head).
231  */
232 static inline void
233 nfs_page_group_init(struct nfs_page *req, struct nfs_page *prev)
234 {
235         struct inode *inode;
236         WARN_ON_ONCE(prev == req);
237
238         if (!prev) {
239                 /* a head request */
240                 req->wb_head = req;
241                 req->wb_this_page = req;
242         } else {
243                 /* a subrequest */
244                 WARN_ON_ONCE(prev->wb_this_page != prev->wb_head);
245                 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &prev->wb_head->wb_flags));
246                 req->wb_head = prev->wb_head;
247                 req->wb_this_page = prev->wb_this_page;
248                 prev->wb_this_page = req;
249
250                 /* All subrequests take a ref on the head request until
251                  * nfs_page_group_destroy is called */
252                 kref_get(&req->wb_head->wb_kref);
253
254                 /* grab extra ref and bump the request count if head request
255                  * has extra ref from the write/commit path to handle handoff
256                  * between write and commit lists. */
257                 if (test_bit(PG_INODE_REF, &prev->wb_head->wb_flags)) {
258                         inode = page_file_mapping(req->wb_page)->host;
259                         set_bit(PG_INODE_REF, &req->wb_flags);
260                         kref_get(&req->wb_kref);
261                         atomic_long_inc(&NFS_I(inode)->nrequests);
262                 }
263         }
264 }
265
266 /*
267  * nfs_page_group_destroy - sync the destruction of page groups
268  * @req - request that no longer needs the page group
269  *
270  * releases the page group reference from each member once all
271  * members have called this function.
272  */
273 static void
274 nfs_page_group_destroy(struct kref *kref)
275 {
276         struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
277         struct nfs_page *head = req->wb_head;
278         struct nfs_page *tmp, *next;
279
280         if (!nfs_page_group_sync_on_bit(req, PG_TEARDOWN))
281                 goto out;
282
283         tmp = req;
284         do {
285                 next = tmp->wb_this_page;
286                 /* unlink and free */
287                 tmp->wb_this_page = tmp;
288                 tmp->wb_head = tmp;
289                 nfs_free_request(tmp);
290                 tmp = next;
291         } while (tmp != req);
292 out:
293         /* subrequests must release the ref on the head request */
294         if (head != req)
295                 nfs_release_request(head);
296 }
297
298 /**
299  * nfs_create_request - Create an NFS read/write request.
300  * @ctx: open context to use
301  * @page: page to write
302  * @last: last nfs request created for this page group or NULL if head
303  * @offset: starting offset within the page for the write
304  * @count: number of bytes to read/write
305  *
306  * The page must be locked by the caller. This makes sure we never
307  * create two different requests for the same page.
308  * User should ensure it is safe to sleep in this function.
309  */
310 struct nfs_page *
311 nfs_create_request(struct nfs_open_context *ctx, struct page *page,
312                    struct nfs_page *last, unsigned int offset,
313                    unsigned int count)
314 {
315         struct nfs_page         *req;
316         struct nfs_lock_context *l_ctx;
317
318         if (test_bit(NFS_CONTEXT_BAD, &ctx->flags))
319                 return ERR_PTR(-EBADF);
320         /* try to allocate the request struct */
321         req = nfs_page_alloc();
322         if (req == NULL)
323                 return ERR_PTR(-ENOMEM);
324
325         /* get lock context early so we can deal with alloc failures */
326         l_ctx = nfs_get_lock_context(ctx);
327         if (IS_ERR(l_ctx)) {
328                 nfs_page_free(req);
329                 return ERR_CAST(l_ctx);
330         }
331         req->wb_lock_context = l_ctx;
332         atomic_inc(&l_ctx->io_count);
333
334         /* Initialize the request struct. Initially, we assume a
335          * long write-back delay. This will be adjusted in
336          * update_nfs_request below if the region is not locked. */
337         req->wb_page    = page;
338         if (page) {
339                 req->wb_index = page_index(page);
340                 get_page(page);
341         }
342         req->wb_offset  = offset;
343         req->wb_pgbase  = offset;
344         req->wb_bytes   = count;
345         req->wb_context = get_nfs_open_context(ctx);
346         kref_init(&req->wb_kref);
347         nfs_page_group_init(req, last);
348         return req;
349 }
350
351 /**
352  * nfs_unlock_request - Unlock request and wake up sleepers.
353  * @req:
354  */
355 void nfs_unlock_request(struct nfs_page *req)
356 {
357         if (!NFS_WBACK_BUSY(req)) {
358                 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
359                 BUG();
360         }
361         smp_mb__before_atomic();
362         clear_bit(PG_BUSY, &req->wb_flags);
363         smp_mb__after_atomic();
364         if (!test_bit(PG_CONTENDED2, &req->wb_flags))
365                 return;
366         wake_up_bit(&req->wb_flags, PG_BUSY);
367 }
368
369 /**
370  * nfs_unlock_and_release_request - Unlock request and release the nfs_page
371  * @req:
372  */
373 void nfs_unlock_and_release_request(struct nfs_page *req)
374 {
375         nfs_unlock_request(req);
376         nfs_release_request(req);
377 }
378
379 /*
380  * nfs_clear_request - Free up all resources allocated to the request
381  * @req:
382  *
383  * Release page and open context resources associated with a read/write
384  * request after it has completed.
385  */
386 static void nfs_clear_request(struct nfs_page *req)
387 {
388         struct page *page = req->wb_page;
389         struct nfs_open_context *ctx = req->wb_context;
390         struct nfs_lock_context *l_ctx = req->wb_lock_context;
391
392         if (page != NULL) {
393                 put_page(page);
394                 req->wb_page = NULL;
395         }
396         if (l_ctx != NULL) {
397                 if (atomic_dec_and_test(&l_ctx->io_count)) {
398                         wake_up_var(&l_ctx->io_count);
399                         if (test_bit(NFS_CONTEXT_UNLOCK, &ctx->flags))
400                                 rpc_wake_up(&NFS_SERVER(d_inode(ctx->dentry))->uoc_rpcwaitq);
401                 }
402                 nfs_put_lock_context(l_ctx);
403                 req->wb_lock_context = NULL;
404         }
405         if (ctx != NULL) {
406                 put_nfs_open_context(ctx);
407                 req->wb_context = NULL;
408         }
409 }
410
411 /**
412  * nfs_release_request - Release the count on an NFS read/write request
413  * @req: request to release
414  *
415  * Note: Should never be called with the spinlock held!
416  */
417 void nfs_free_request(struct nfs_page *req)
418 {
419         WARN_ON_ONCE(req->wb_this_page != req);
420
421         /* extra debug: make sure no sync bits are still set */
422         WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
423         WARN_ON_ONCE(test_bit(PG_UNLOCKPAGE, &req->wb_flags));
424         WARN_ON_ONCE(test_bit(PG_UPTODATE, &req->wb_flags));
425         WARN_ON_ONCE(test_bit(PG_WB_END, &req->wb_flags));
426         WARN_ON_ONCE(test_bit(PG_REMOVE, &req->wb_flags));
427
428         /* Release struct file and open context */
429         nfs_clear_request(req);
430         nfs_page_free(req);
431 }
432
433 void nfs_release_request(struct nfs_page *req)
434 {
435         kref_put(&req->wb_kref, nfs_page_group_destroy);
436 }
437 EXPORT_SYMBOL_GPL(nfs_release_request);
438
439 /**
440  * nfs_wait_on_request - Wait for a request to complete.
441  * @req: request to wait upon.
442  *
443  * Interruptible by fatal signals only.
444  * The user is responsible for holding a count on the request.
445  */
446 int
447 nfs_wait_on_request(struct nfs_page *req)
448 {
449         if (!test_bit(PG_BUSY, &req->wb_flags))
450                 return 0;
451         set_bit(PG_CONTENDED2, &req->wb_flags);
452         smp_mb__after_atomic();
453         return wait_on_bit_io(&req->wb_flags, PG_BUSY,
454                               TASK_UNINTERRUPTIBLE);
455 }
456 EXPORT_SYMBOL_GPL(nfs_wait_on_request);
457
458 /*
459  * nfs_generic_pg_test - determine if requests can be coalesced
460  * @desc: pointer to descriptor
461  * @prev: previous request in desc, or NULL
462  * @req: this request
463  *
464  * Returns zero if @req cannot be coalesced into @desc, otherwise it returns
465  * the size of the request.
466  */
467 size_t nfs_generic_pg_test(struct nfs_pageio_descriptor *desc,
468                            struct nfs_page *prev, struct nfs_page *req)
469 {
470         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
471
472
473         if (mirror->pg_count > mirror->pg_bsize) {
474                 /* should never happen */
475                 WARN_ON_ONCE(1);
476                 return 0;
477         }
478
479         /*
480          * Limit the request size so that we can still allocate a page array
481          * for it without upsetting the slab allocator.
482          */
483         if (((mirror->pg_count + req->wb_bytes) >> PAGE_SHIFT) *
484                         sizeof(struct page *) > PAGE_SIZE)
485                 return 0;
486
487         return min(mirror->pg_bsize - mirror->pg_count, (size_t)req->wb_bytes);
488 }
489 EXPORT_SYMBOL_GPL(nfs_generic_pg_test);
490
491 struct nfs_pgio_header *nfs_pgio_header_alloc(const struct nfs_rw_ops *ops)
492 {
493         struct nfs_pgio_header *hdr = ops->rw_alloc_header();
494
495         if (hdr) {
496                 INIT_LIST_HEAD(&hdr->pages);
497                 hdr->rw_ops = ops;
498         }
499         return hdr;
500 }
501 EXPORT_SYMBOL_GPL(nfs_pgio_header_alloc);
502
503 /**
504  * nfs_pgio_data_destroy - make @hdr suitable for reuse
505  *
506  * Frees memory and releases refs from nfs_generic_pgio, so that it may
507  * be called again.
508  *
509  * @hdr: A header that has had nfs_generic_pgio called
510  */
511 static void nfs_pgio_data_destroy(struct nfs_pgio_header *hdr)
512 {
513         if (hdr->args.context)
514                 put_nfs_open_context(hdr->args.context);
515         if (hdr->page_array.pagevec != hdr->page_array.page_array)
516                 kfree(hdr->page_array.pagevec);
517 }
518
519 /*
520  * nfs_pgio_header_free - Free a read or write header
521  * @hdr: The header to free
522  */
523 void nfs_pgio_header_free(struct nfs_pgio_header *hdr)
524 {
525         nfs_pgio_data_destroy(hdr);
526         hdr->rw_ops->rw_free_header(hdr);
527 }
528 EXPORT_SYMBOL_GPL(nfs_pgio_header_free);
529
530 /**
531  * nfs_pgio_rpcsetup - Set up arguments for a pageio call
532  * @hdr: The pageio hdr
533  * @count: Number of bytes to read
534  * @offset: Initial offset
535  * @how: How to commit data (writes only)
536  * @cinfo: Commit information for the call (writes only)
537  */
538 static void nfs_pgio_rpcsetup(struct nfs_pgio_header *hdr,
539                               unsigned int count,
540                               int how, struct nfs_commit_info *cinfo)
541 {
542         struct nfs_page *req = hdr->req;
543
544         /* Set up the RPC argument and reply structs
545          * NB: take care not to mess about with hdr->commit et al. */
546
547         hdr->args.fh     = NFS_FH(hdr->inode);
548         hdr->args.offset = req_offset(req);
549         /* pnfs_set_layoutcommit needs this */
550         hdr->mds_offset = hdr->args.offset;
551         hdr->args.pgbase = req->wb_pgbase;
552         hdr->args.pages  = hdr->page_array.pagevec;
553         hdr->args.count  = count;
554         hdr->args.context = get_nfs_open_context(req->wb_context);
555         hdr->args.lock_context = req->wb_lock_context;
556         hdr->args.stable  = NFS_UNSTABLE;
557         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
558         case 0:
559                 break;
560         case FLUSH_COND_STABLE:
561                 if (nfs_reqs_to_commit(cinfo))
562                         break;
563                 /* fall through */
564         default:
565                 hdr->args.stable = NFS_FILE_SYNC;
566         }
567
568         hdr->res.fattr   = &hdr->fattr;
569         hdr->res.count   = count;
570         hdr->res.eof     = 0;
571         hdr->res.verf    = &hdr->verf;
572         nfs_fattr_init(&hdr->fattr);
573 }
574
575 /**
576  * nfs_pgio_prepare - Prepare pageio hdr to go over the wire
577  * @task: The current task
578  * @calldata: pageio header to prepare
579  */
580 static void nfs_pgio_prepare(struct rpc_task *task, void *calldata)
581 {
582         struct nfs_pgio_header *hdr = calldata;
583         int err;
584         err = NFS_PROTO(hdr->inode)->pgio_rpc_prepare(task, hdr);
585         if (err)
586                 rpc_exit(task, err);
587 }
588
589 int nfs_initiate_pgio(struct rpc_clnt *clnt, struct nfs_pgio_header *hdr,
590                       const struct cred *cred, const struct nfs_rpc_ops *rpc_ops,
591                       const struct rpc_call_ops *call_ops, int how, int flags)
592 {
593         struct rpc_task *task;
594         struct rpc_message msg = {
595                 .rpc_argp = &hdr->args,
596                 .rpc_resp = &hdr->res,
597                 .rpc_cred = cred,
598         };
599         struct rpc_task_setup task_setup_data = {
600                 .rpc_client = clnt,
601                 .task = &hdr->task,
602                 .rpc_message = &msg,
603                 .callback_ops = call_ops,
604                 .callback_data = hdr,
605                 .workqueue = nfsiod_workqueue,
606                 .flags = RPC_TASK_ASYNC | flags,
607         };
608         int ret = 0;
609
610         hdr->rw_ops->rw_initiate(hdr, &msg, rpc_ops, &task_setup_data, how);
611
612         dprintk("NFS: initiated pgio call "
613                 "(req %s/%llu, %u bytes @ offset %llu)\n",
614                 hdr->inode->i_sb->s_id,
615                 (unsigned long long)NFS_FILEID(hdr->inode),
616                 hdr->args.count,
617                 (unsigned long long)hdr->args.offset);
618
619         task = rpc_run_task(&task_setup_data);
620         if (IS_ERR(task)) {
621                 ret = PTR_ERR(task);
622                 goto out;
623         }
624         if (how & FLUSH_SYNC) {
625                 ret = rpc_wait_for_completion_task(task);
626                 if (ret == 0)
627                         ret = task->tk_status;
628         }
629         rpc_put_task(task);
630 out:
631         return ret;
632 }
633 EXPORT_SYMBOL_GPL(nfs_initiate_pgio);
634
635 /**
636  * nfs_pgio_error - Clean up from a pageio error
637  * @desc: IO descriptor
638  * @hdr: pageio header
639  */
640 static void nfs_pgio_error(struct nfs_pgio_header *hdr)
641 {
642         set_bit(NFS_IOHDR_REDO, &hdr->flags);
643         hdr->completion_ops->completion(hdr);
644 }
645
646 /**
647  * nfs_pgio_release - Release pageio data
648  * @calldata: The pageio header to release
649  */
650 static void nfs_pgio_release(void *calldata)
651 {
652         struct nfs_pgio_header *hdr = calldata;
653         hdr->completion_ops->completion(hdr);
654 }
655
656 static void nfs_pageio_mirror_init(struct nfs_pgio_mirror *mirror,
657                                    unsigned int bsize)
658 {
659         INIT_LIST_HEAD(&mirror->pg_list);
660         mirror->pg_bytes_written = 0;
661         mirror->pg_count = 0;
662         mirror->pg_bsize = bsize;
663         mirror->pg_base = 0;
664         mirror->pg_recoalesce = 0;
665 }
666
667 /**
668  * nfs_pageio_init - initialise a page io descriptor
669  * @desc: pointer to descriptor
670  * @inode: pointer to inode
671  * @pg_ops: pointer to pageio operations
672  * @compl_ops: pointer to pageio completion operations
673  * @rw_ops: pointer to nfs read/write operations
674  * @bsize: io block size
675  * @io_flags: extra parameters for the io function
676  */
677 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
678                      struct inode *inode,
679                      const struct nfs_pageio_ops *pg_ops,
680                      const struct nfs_pgio_completion_ops *compl_ops,
681                      const struct nfs_rw_ops *rw_ops,
682                      size_t bsize,
683                      int io_flags)
684 {
685         desc->pg_moreio = 0;
686         desc->pg_inode = inode;
687         desc->pg_ops = pg_ops;
688         desc->pg_completion_ops = compl_ops;
689         desc->pg_rw_ops = rw_ops;
690         desc->pg_ioflags = io_flags;
691         desc->pg_error = 0;
692         desc->pg_lseg = NULL;
693         desc->pg_io_completion = NULL;
694         desc->pg_dreq = NULL;
695         desc->pg_bsize = bsize;
696
697         desc->pg_mirror_count = 1;
698         desc->pg_mirror_idx = 0;
699
700         desc->pg_mirrors_dynamic = NULL;
701         desc->pg_mirrors = desc->pg_mirrors_static;
702         nfs_pageio_mirror_init(&desc->pg_mirrors[0], bsize);
703 }
704
705 /**
706  * nfs_pgio_result - Basic pageio error handling
707  * @task: The task that ran
708  * @calldata: Pageio header to check
709  */
710 static void nfs_pgio_result(struct rpc_task *task, void *calldata)
711 {
712         struct nfs_pgio_header *hdr = calldata;
713         struct inode *inode = hdr->inode;
714
715         dprintk("NFS: %s: %5u, (status %d)\n", __func__,
716                 task->tk_pid, task->tk_status);
717
718         if (hdr->rw_ops->rw_done(task, hdr, inode) != 0)
719                 return;
720         if (task->tk_status < 0)
721                 nfs_set_pgio_error(hdr, task->tk_status, hdr->args.offset);
722         else
723                 hdr->rw_ops->rw_result(task, hdr);
724 }
725
726 /*
727  * Create an RPC task for the given read or write request and kick it.
728  * The page must have been locked by the caller.
729  *
730  * It may happen that the page we're passed is not marked dirty.
731  * This is the case if nfs_updatepage detects a conflicting request
732  * that has been written but not committed.
733  */
734 int nfs_generic_pgio(struct nfs_pageio_descriptor *desc,
735                      struct nfs_pgio_header *hdr)
736 {
737         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
738
739         struct nfs_page         *req;
740         struct page             **pages,
741                                 *last_page;
742         struct list_head *head = &mirror->pg_list;
743         struct nfs_commit_info cinfo;
744         struct nfs_page_array *pg_array = &hdr->page_array;
745         unsigned int pagecount, pageused;
746         gfp_t gfp_flags = GFP_KERNEL;
747
748         pagecount = nfs_page_array_len(mirror->pg_base, mirror->pg_count);
749         pg_array->npages = pagecount;
750
751         if (pagecount <= ARRAY_SIZE(pg_array->page_array))
752                 pg_array->pagevec = pg_array->page_array;
753         else {
754                 if (hdr->rw_mode == FMODE_WRITE)
755                         gfp_flags = GFP_NOIO;
756                 pg_array->pagevec = kcalloc(pagecount, sizeof(struct page *), gfp_flags);
757                 if (!pg_array->pagevec) {
758                         pg_array->npages = 0;
759                         nfs_pgio_error(hdr);
760                         desc->pg_error = -ENOMEM;
761                         return desc->pg_error;
762                 }
763         }
764
765         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
766         pages = hdr->page_array.pagevec;
767         last_page = NULL;
768         pageused = 0;
769         while (!list_empty(head)) {
770                 req = nfs_list_entry(head->next);
771                 nfs_list_remove_request(req);
772                 nfs_list_add_request(req, &hdr->pages);
773
774                 if (!last_page || last_page != req->wb_page) {
775                         pageused++;
776                         if (pageused > pagecount)
777                                 break;
778                         *pages++ = last_page = req->wb_page;
779                 }
780         }
781         if (WARN_ON_ONCE(pageused != pagecount)) {
782                 nfs_pgio_error(hdr);
783                 desc->pg_error = -EINVAL;
784                 return desc->pg_error;
785         }
786
787         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
788             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
789                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
790
791         /* Set up the argument struct */
792         nfs_pgio_rpcsetup(hdr, mirror->pg_count, desc->pg_ioflags, &cinfo);
793         desc->pg_rpc_callops = &nfs_pgio_common_ops;
794         return 0;
795 }
796 EXPORT_SYMBOL_GPL(nfs_generic_pgio);
797
798 static int nfs_generic_pg_pgios(struct nfs_pageio_descriptor *desc)
799 {
800         struct nfs_pgio_header *hdr;
801         int ret;
802
803         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
804         if (!hdr) {
805                 desc->pg_error = -ENOMEM;
806                 return desc->pg_error;
807         }
808         nfs_pgheader_init(desc, hdr, nfs_pgio_header_free);
809         ret = nfs_generic_pgio(desc, hdr);
810         if (ret == 0)
811                 ret = nfs_initiate_pgio(NFS_CLIENT(hdr->inode),
812                                         hdr,
813                                         hdr->cred,
814                                         NFS_PROTO(hdr->inode),
815                                         desc->pg_rpc_callops,
816                                         desc->pg_ioflags, 0);
817         return ret;
818 }
819
820 static struct nfs_pgio_mirror *
821 nfs_pageio_alloc_mirrors(struct nfs_pageio_descriptor *desc,
822                 unsigned int mirror_count)
823 {
824         struct nfs_pgio_mirror *ret;
825         unsigned int i;
826
827         kfree(desc->pg_mirrors_dynamic);
828         desc->pg_mirrors_dynamic = NULL;
829         if (mirror_count == 1)
830                 return desc->pg_mirrors_static;
831         ret = kmalloc_array(mirror_count, sizeof(*ret), GFP_NOFS);
832         if (ret != NULL) {
833                 for (i = 0; i < mirror_count; i++)
834                         nfs_pageio_mirror_init(&ret[i], desc->pg_bsize);
835                 desc->pg_mirrors_dynamic = ret;
836         }
837         return ret;
838 }
839
840 /*
841  * nfs_pageio_setup_mirroring - determine if mirroring is to be used
842  *                              by calling the pg_get_mirror_count op
843  */
844 static void nfs_pageio_setup_mirroring(struct nfs_pageio_descriptor *pgio,
845                                        struct nfs_page *req)
846 {
847         unsigned int mirror_count = 1;
848
849         if (pgio->pg_ops->pg_get_mirror_count)
850                 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
851         if (mirror_count == pgio->pg_mirror_count || pgio->pg_error < 0)
852                 return;
853
854         if (!mirror_count || mirror_count > NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX) {
855                 pgio->pg_error = -EINVAL;
856                 return;
857         }
858
859         pgio->pg_mirrors = nfs_pageio_alloc_mirrors(pgio, mirror_count);
860         if (pgio->pg_mirrors == NULL) {
861                 pgio->pg_error = -ENOMEM;
862                 pgio->pg_mirrors = pgio->pg_mirrors_static;
863                 mirror_count = 1;
864         }
865         pgio->pg_mirror_count = mirror_count;
866 }
867
868 /*
869  * nfs_pageio_stop_mirroring - stop using mirroring (set mirror count to 1)
870  */
871 void nfs_pageio_stop_mirroring(struct nfs_pageio_descriptor *pgio)
872 {
873         pgio->pg_mirror_count = 1;
874         pgio->pg_mirror_idx = 0;
875 }
876
877 static void nfs_pageio_cleanup_mirroring(struct nfs_pageio_descriptor *pgio)
878 {
879         pgio->pg_mirror_count = 1;
880         pgio->pg_mirror_idx = 0;
881         pgio->pg_mirrors = pgio->pg_mirrors_static;
882         kfree(pgio->pg_mirrors_dynamic);
883         pgio->pg_mirrors_dynamic = NULL;
884 }
885
886 static bool nfs_match_lock_context(const struct nfs_lock_context *l1,
887                 const struct nfs_lock_context *l2)
888 {
889         return l1->lockowner == l2->lockowner;
890 }
891
892 /**
893  * nfs_can_coalesce_requests - test two requests for compatibility
894  * @prev: pointer to nfs_page
895  * @req: pointer to nfs_page
896  *
897  * The nfs_page structures 'prev' and 'req' are compared to ensure that the
898  * page data area they describe is contiguous, and that their RPC
899  * credentials, NFSv4 open state, and lockowners are the same.
900  *
901  * Return 'true' if this is the case, else return 'false'.
902  */
903 static bool nfs_can_coalesce_requests(struct nfs_page *prev,
904                                       struct nfs_page *req,
905                                       struct nfs_pageio_descriptor *pgio)
906 {
907         size_t size;
908         struct file_lock_context *flctx;
909
910         if (prev) {
911                 if (!nfs_match_open_context(req->wb_context, prev->wb_context))
912                         return false;
913                 flctx = d_inode(req->wb_context->dentry)->i_flctx;
914                 if (flctx != NULL &&
915                     !(list_empty_careful(&flctx->flc_posix) &&
916                       list_empty_careful(&flctx->flc_flock)) &&
917                     !nfs_match_lock_context(req->wb_lock_context,
918                                             prev->wb_lock_context))
919                         return false;
920                 if (req_offset(req) != req_offset(prev) + prev->wb_bytes)
921                         return false;
922                 if (req->wb_page == prev->wb_page) {
923                         if (req->wb_pgbase != prev->wb_pgbase + prev->wb_bytes)
924                                 return false;
925                 } else {
926                         if (req->wb_pgbase != 0 ||
927                             prev->wb_pgbase + prev->wb_bytes != PAGE_SIZE)
928                                 return false;
929                 }
930         }
931         size = pgio->pg_ops->pg_test(pgio, prev, req);
932         WARN_ON_ONCE(size > req->wb_bytes);
933         if (size && size < req->wb_bytes)
934                 req->wb_bytes = size;
935         return size > 0;
936 }
937
938 /**
939  * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
940  * @desc: destination io descriptor
941  * @req: request
942  *
943  * Returns true if the request 'req' was successfully coalesced into the
944  * existing list of pages 'desc'.
945  */
946 static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
947                                      struct nfs_page *req)
948 {
949         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
950
951         struct nfs_page *prev = NULL;
952
953         if (mirror->pg_count != 0) {
954                 prev = nfs_list_entry(mirror->pg_list.prev);
955         } else {
956                 if (desc->pg_ops->pg_init)
957                         desc->pg_ops->pg_init(desc, req);
958                 if (desc->pg_error < 0)
959                         return 0;
960                 mirror->pg_base = req->wb_pgbase;
961         }
962         if (!nfs_can_coalesce_requests(prev, req, desc))
963                 return 0;
964         nfs_list_remove_request(req);
965         nfs_list_add_request(req, &mirror->pg_list);
966         mirror->pg_count += req->wb_bytes;
967         return 1;
968 }
969
970 /*
971  * Helper for nfs_pageio_add_request and nfs_pageio_complete
972  */
973 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
974 {
975         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
976
977
978         if (!list_empty(&mirror->pg_list)) {
979                 int error = desc->pg_ops->pg_doio(desc);
980                 if (error < 0)
981                         desc->pg_error = error;
982                 else
983                         mirror->pg_bytes_written += mirror->pg_count;
984         }
985         if (list_empty(&mirror->pg_list)) {
986                 mirror->pg_count = 0;
987                 mirror->pg_base = 0;
988         }
989 }
990
991 /**
992  * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
993  * @desc: destination io descriptor
994  * @req: request
995  *
996  * This may split a request into subrequests which are all part of the
997  * same page group.
998  *
999  * Returns true if the request 'req' was successfully coalesced into the
1000  * existing list of pages 'desc'.
1001  */
1002 static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
1003                            struct nfs_page *req)
1004 {
1005         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1006
1007         struct nfs_page *subreq;
1008         unsigned int bytes_left = 0;
1009         unsigned int offset, pgbase;
1010
1011         nfs_page_group_lock(req);
1012
1013         subreq = req;
1014         bytes_left = subreq->wb_bytes;
1015         offset = subreq->wb_offset;
1016         pgbase = subreq->wb_pgbase;
1017
1018         do {
1019                 if (!nfs_pageio_do_add_request(desc, subreq)) {
1020                         /* make sure pg_test call(s) did nothing */
1021                         WARN_ON_ONCE(subreq->wb_bytes != bytes_left);
1022                         WARN_ON_ONCE(subreq->wb_offset != offset);
1023                         WARN_ON_ONCE(subreq->wb_pgbase != pgbase);
1024
1025                         nfs_page_group_unlock(req);
1026                         desc->pg_moreio = 1;
1027                         nfs_pageio_doio(desc);
1028                         if (desc->pg_error < 0)
1029                                 return 0;
1030                         if (mirror->pg_recoalesce)
1031                                 return 0;
1032                         /* retry add_request for this subreq */
1033                         nfs_page_group_lock(req);
1034                         continue;
1035                 }
1036
1037                 /* check for buggy pg_test call(s) */
1038                 WARN_ON_ONCE(subreq->wb_bytes + subreq->wb_pgbase > PAGE_SIZE);
1039                 WARN_ON_ONCE(subreq->wb_bytes > bytes_left);
1040                 WARN_ON_ONCE(subreq->wb_bytes == 0);
1041
1042                 bytes_left -= subreq->wb_bytes;
1043                 offset += subreq->wb_bytes;
1044                 pgbase += subreq->wb_bytes;
1045
1046                 if (bytes_left) {
1047                         subreq = nfs_create_request(req->wb_context,
1048                                         req->wb_page,
1049                                         subreq, pgbase, bytes_left);
1050                         if (IS_ERR(subreq))
1051                                 goto err_ptr;
1052                         nfs_lock_request(subreq);
1053                         subreq->wb_offset  = offset;
1054                         subreq->wb_index = req->wb_index;
1055                 }
1056         } while (bytes_left > 0);
1057
1058         nfs_page_group_unlock(req);
1059         return 1;
1060 err_ptr:
1061         desc->pg_error = PTR_ERR(subreq);
1062         nfs_page_group_unlock(req);
1063         return 0;
1064 }
1065
1066 static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc)
1067 {
1068         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1069         LIST_HEAD(head);
1070
1071         do {
1072                 list_splice_init(&mirror->pg_list, &head);
1073                 mirror->pg_bytes_written -= mirror->pg_count;
1074                 mirror->pg_count = 0;
1075                 mirror->pg_base = 0;
1076                 mirror->pg_recoalesce = 0;
1077
1078                 while (!list_empty(&head)) {
1079                         struct nfs_page *req;
1080
1081                         req = list_first_entry(&head, struct nfs_page, wb_list);
1082                         nfs_list_remove_request(req);
1083                         if (__nfs_pageio_add_request(desc, req))
1084                                 continue;
1085                         if (desc->pg_error < 0) {
1086                                 list_splice_tail(&head, &mirror->pg_list);
1087                                 mirror->pg_recoalesce = 1;
1088                                 return 0;
1089                         }
1090                         break;
1091                 }
1092         } while (mirror->pg_recoalesce);
1093         return 1;
1094 }
1095
1096 static int nfs_pageio_add_request_mirror(struct nfs_pageio_descriptor *desc,
1097                 struct nfs_page *req)
1098 {
1099         int ret;
1100
1101         do {
1102                 ret = __nfs_pageio_add_request(desc, req);
1103                 if (ret)
1104                         break;
1105                 if (desc->pg_error < 0)
1106                         break;
1107                 ret = nfs_do_recoalesce(desc);
1108         } while (ret);
1109
1110         return ret;
1111 }
1112
1113 static void nfs_pageio_error_cleanup(struct nfs_pageio_descriptor *desc)
1114 {
1115         u32 midx;
1116         struct nfs_pgio_mirror *mirror;
1117
1118         if (!desc->pg_error)
1119                 return;
1120
1121         for (midx = 0; midx < desc->pg_mirror_count; midx++) {
1122                 mirror = &desc->pg_mirrors[midx];
1123                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
1124         }
1125 }
1126
1127 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
1128                            struct nfs_page *req)
1129 {
1130         u32 midx;
1131         unsigned int pgbase, offset, bytes;
1132         struct nfs_page *dupreq, *lastreq;
1133
1134         pgbase = req->wb_pgbase;
1135         offset = req->wb_offset;
1136         bytes = req->wb_bytes;
1137
1138         nfs_pageio_setup_mirroring(desc, req);
1139         if (desc->pg_error < 0)
1140                 goto out_failed;
1141
1142         for (midx = 0; midx < desc->pg_mirror_count; midx++) {
1143                 if (midx) {
1144                         nfs_page_group_lock(req);
1145
1146                         /* find the last request */
1147                         for (lastreq = req->wb_head;
1148                              lastreq->wb_this_page != req->wb_head;
1149                              lastreq = lastreq->wb_this_page)
1150                                 ;
1151
1152                         dupreq = nfs_create_request(req->wb_context,
1153                                         req->wb_page, lastreq, pgbase, bytes);
1154
1155                         if (IS_ERR(dupreq)) {
1156                                 nfs_page_group_unlock(req);
1157                                 desc->pg_error = PTR_ERR(dupreq);
1158                                 goto out_failed;
1159                         }
1160
1161                         nfs_lock_request(dupreq);
1162                         nfs_page_group_unlock(req);
1163                         dupreq->wb_offset = offset;
1164                         dupreq->wb_index = req->wb_index;
1165                 } else
1166                         dupreq = req;
1167
1168                 if (nfs_pgio_has_mirroring(desc))
1169                         desc->pg_mirror_idx = midx;
1170                 if (!nfs_pageio_add_request_mirror(desc, dupreq))
1171                         goto out_failed;
1172         }
1173
1174         return 1;
1175
1176 out_failed:
1177         nfs_pageio_error_cleanup(desc);
1178         return 0;
1179 }
1180
1181 /*
1182  * nfs_pageio_complete_mirror - Complete I/O on the current mirror of an
1183  *                              nfs_pageio_descriptor
1184  * @desc: pointer to io descriptor
1185  * @mirror_idx: pointer to mirror index
1186  */
1187 static void nfs_pageio_complete_mirror(struct nfs_pageio_descriptor *desc,
1188                                        u32 mirror_idx)
1189 {
1190         struct nfs_pgio_mirror *mirror = &desc->pg_mirrors[mirror_idx];
1191         u32 restore_idx = desc->pg_mirror_idx;
1192
1193         if (nfs_pgio_has_mirroring(desc))
1194                 desc->pg_mirror_idx = mirror_idx;
1195         for (;;) {
1196                 nfs_pageio_doio(desc);
1197                 if (!mirror->pg_recoalesce)
1198                         break;
1199                 if (!nfs_do_recoalesce(desc))
1200                         break;
1201         }
1202         desc->pg_mirror_idx = restore_idx;
1203 }
1204
1205 /*
1206  * nfs_pageio_resend - Transfer requests to new descriptor and resend
1207  * @hdr - the pgio header to move request from
1208  * @desc - the pageio descriptor to add requests to
1209  *
1210  * Try to move each request (nfs_page) from @hdr to @desc then attempt
1211  * to send them.
1212  *
1213  * Returns 0 on success and < 0 on error.
1214  */
1215 int nfs_pageio_resend(struct nfs_pageio_descriptor *desc,
1216                       struct nfs_pgio_header *hdr)
1217 {
1218         LIST_HEAD(failed);
1219
1220         desc->pg_io_completion = hdr->io_completion;
1221         desc->pg_dreq = hdr->dreq;
1222         while (!list_empty(&hdr->pages)) {
1223                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1224
1225                 nfs_list_remove_request(req);
1226                 if (!nfs_pageio_add_request(desc, req))
1227                         nfs_list_add_request(req, &failed);
1228         }
1229         nfs_pageio_complete(desc);
1230         if (!list_empty(&failed)) {
1231                 list_move(&failed, &hdr->pages);
1232                 return desc->pg_error < 0 ? desc->pg_error : -EIO;
1233         }
1234         return 0;
1235 }
1236 EXPORT_SYMBOL_GPL(nfs_pageio_resend);
1237
1238 /**
1239  * nfs_pageio_complete - Complete I/O then cleanup an nfs_pageio_descriptor
1240  * @desc: pointer to io descriptor
1241  */
1242 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
1243 {
1244         u32 midx;
1245
1246         for (midx = 0; midx < desc->pg_mirror_count; midx++)
1247                 nfs_pageio_complete_mirror(desc, midx);
1248
1249         if (desc->pg_error < 0)
1250                 nfs_pageio_error_cleanup(desc);
1251         if (desc->pg_ops->pg_cleanup)
1252                 desc->pg_ops->pg_cleanup(desc);
1253         nfs_pageio_cleanup_mirroring(desc);
1254 }
1255
1256 /**
1257  * nfs_pageio_cond_complete - Conditional I/O completion
1258  * @desc: pointer to io descriptor
1259  * @index: page index
1260  *
1261  * It is important to ensure that processes don't try to take locks
1262  * on non-contiguous ranges of pages as that might deadlock. This
1263  * function should be called before attempting to wait on a locked
1264  * nfs_page. It will complete the I/O if the page index 'index'
1265  * is not contiguous with the existing list of pages in 'desc'.
1266  */
1267 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
1268 {
1269         struct nfs_pgio_mirror *mirror;
1270         struct nfs_page *prev;
1271         u32 midx;
1272
1273         for (midx = 0; midx < desc->pg_mirror_count; midx++) {
1274                 mirror = &desc->pg_mirrors[midx];
1275                 if (!list_empty(&mirror->pg_list)) {
1276                         prev = nfs_list_entry(mirror->pg_list.prev);
1277                         if (index != prev->wb_index + 1) {
1278                                 nfs_pageio_complete(desc);
1279                                 break;
1280                         }
1281                 }
1282         }
1283 }
1284
1285 int __init nfs_init_nfspagecache(void)
1286 {
1287         nfs_page_cachep = kmem_cache_create("nfs_page",
1288                                             sizeof(struct nfs_page),
1289                                             0, SLAB_HWCACHE_ALIGN,
1290                                             NULL);
1291         if (nfs_page_cachep == NULL)
1292                 return -ENOMEM;
1293
1294         return 0;
1295 }
1296
1297 void nfs_destroy_nfspagecache(void)
1298 {
1299         kmem_cache_destroy(nfs_page_cachep);
1300 }
1301
1302 static const struct rpc_call_ops nfs_pgio_common_ops = {
1303         .rpc_call_prepare = nfs_pgio_prepare,
1304         .rpc_call_done = nfs_pgio_result,
1305         .rpc_release = nfs_pgio_release,
1306 };
1307
1308 const struct nfs_pageio_ops nfs_pgio_rw_ops = {
1309         .pg_test = nfs_generic_pg_test,
1310         .pg_doio = nfs_generic_pg_pgios,
1311 };