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