xref: /openbmc/linux/fs/nfs/pagelist.c (revision 110e6f26)
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 	req->wb_index	= page_file_index(page);
345 	get_page(page);
346 	req->wb_offset  = offset;
347 	req->wb_pgbase	= offset;
348 	req->wb_bytes   = count;
349 	req->wb_context = get_nfs_open_context(ctx);
350 	kref_init(&req->wb_kref);
351 	nfs_page_group_init(req, last);
352 	return req;
353 }
354 
355 /**
356  * nfs_unlock_request - Unlock request and wake up sleepers.
357  * @req:
358  */
359 void nfs_unlock_request(struct nfs_page *req)
360 {
361 	if (!NFS_WBACK_BUSY(req)) {
362 		printk(KERN_ERR "NFS: Invalid unlock attempted\n");
363 		BUG();
364 	}
365 	smp_mb__before_atomic();
366 	clear_bit(PG_BUSY, &req->wb_flags);
367 	smp_mb__after_atomic();
368 	wake_up_bit(&req->wb_flags, PG_BUSY);
369 }
370 
371 /**
372  * nfs_unlock_and_release_request - Unlock request and release the nfs_page
373  * @req:
374  */
375 void nfs_unlock_and_release_request(struct nfs_page *req)
376 {
377 	nfs_unlock_request(req);
378 	nfs_release_request(req);
379 }
380 
381 /*
382  * nfs_clear_request - Free up all resources allocated to the request
383  * @req:
384  *
385  * Release page and open context resources associated with a read/write
386  * request after it has completed.
387  */
388 static void nfs_clear_request(struct nfs_page *req)
389 {
390 	struct page *page = req->wb_page;
391 	struct nfs_open_context *ctx = req->wb_context;
392 	struct nfs_lock_context *l_ctx = req->wb_lock_context;
393 
394 	if (page != NULL) {
395 		put_page(page);
396 		req->wb_page = NULL;
397 	}
398 	if (l_ctx != NULL) {
399 		if (atomic_dec_and_test(&l_ctx->io_count))
400 			wake_up_atomic_t(&l_ctx->io_count);
401 		nfs_put_lock_context(l_ctx);
402 		req->wb_lock_context = NULL;
403 	}
404 	if (ctx != NULL) {
405 		put_nfs_open_context(ctx);
406 		req->wb_context = NULL;
407 	}
408 }
409 
410 /**
411  * nfs_release_request - Release the count on an NFS read/write request
412  * @req: request to release
413  *
414  * Note: Should never be called with the spinlock held!
415  */
416 void nfs_free_request(struct nfs_page *req)
417 {
418 	WARN_ON_ONCE(req->wb_this_page != req);
419 
420 	/* extra debug: make sure no sync bits are still set */
421 	WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
422 	WARN_ON_ONCE(test_bit(PG_UNLOCKPAGE, &req->wb_flags));
423 	WARN_ON_ONCE(test_bit(PG_UPTODATE, &req->wb_flags));
424 	WARN_ON_ONCE(test_bit(PG_WB_END, &req->wb_flags));
425 	WARN_ON_ONCE(test_bit(PG_REMOVE, &req->wb_flags));
426 
427 	/* Release struct file and open context */
428 	nfs_clear_request(req);
429 	nfs_page_free(req);
430 }
431 
432 void nfs_release_request(struct nfs_page *req)
433 {
434 	kref_put(&req->wb_kref, nfs_page_group_destroy);
435 }
436 
437 /**
438  * nfs_wait_on_request - Wait for a request to complete.
439  * @req: request to wait upon.
440  *
441  * Interruptible by fatal signals only.
442  * The user is responsible for holding a count on the request.
443  */
444 int
445 nfs_wait_on_request(struct nfs_page *req)
446 {
447 	return wait_on_bit_io(&req->wb_flags, PG_BUSY,
448 			      TASK_UNINTERRUPTIBLE);
449 }
450 
451 /*
452  * nfs_generic_pg_test - determine if requests can be coalesced
453  * @desc: pointer to descriptor
454  * @prev: previous request in desc, or NULL
455  * @req: this request
456  *
457  * Returns zero if @req can be coalesced into @desc, otherwise it returns
458  * the size of the request.
459  */
460 size_t nfs_generic_pg_test(struct nfs_pageio_descriptor *desc,
461 			   struct nfs_page *prev, struct nfs_page *req)
462 {
463 	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
464 
465 
466 	if (mirror->pg_count > mirror->pg_bsize) {
467 		/* should never happen */
468 		WARN_ON_ONCE(1);
469 		return 0;
470 	}
471 
472 	/*
473 	 * Limit the request size so that we can still allocate a page array
474 	 * for it without upsetting the slab allocator.
475 	 */
476 	if (((mirror->pg_count + req->wb_bytes) >> PAGE_SHIFT) *
477 			sizeof(struct page *) > PAGE_SIZE)
478 		return 0;
479 
480 	return min(mirror->pg_bsize - mirror->pg_count, (size_t)req->wb_bytes);
481 }
482 EXPORT_SYMBOL_GPL(nfs_generic_pg_test);
483 
484 struct nfs_pgio_header *nfs_pgio_header_alloc(const struct nfs_rw_ops *ops)
485 {
486 	struct nfs_pgio_header *hdr = ops->rw_alloc_header();
487 
488 	if (hdr) {
489 		INIT_LIST_HEAD(&hdr->pages);
490 		spin_lock_init(&hdr->lock);
491 		hdr->rw_ops = ops;
492 	}
493 	return hdr;
494 }
495 EXPORT_SYMBOL_GPL(nfs_pgio_header_alloc);
496 
497 /*
498  * nfs_pgio_header_free - Free a read or write header
499  * @hdr: The header to free
500  */
501 void nfs_pgio_header_free(struct nfs_pgio_header *hdr)
502 {
503 	hdr->rw_ops->rw_free_header(hdr);
504 }
505 EXPORT_SYMBOL_GPL(nfs_pgio_header_free);
506 
507 /**
508  * nfs_pgio_data_destroy - make @hdr suitable for reuse
509  *
510  * Frees memory and releases refs from nfs_generic_pgio, so that it may
511  * be called again.
512  *
513  * @hdr: A header that has had nfs_generic_pgio called
514  */
515 void nfs_pgio_data_destroy(struct nfs_pgio_header *hdr)
516 {
517 	if (hdr->args.context)
518 		put_nfs_open_context(hdr->args.context);
519 	if (hdr->page_array.pagevec != hdr->page_array.page_array)
520 		kfree(hdr->page_array.pagevec);
521 }
522 EXPORT_SYMBOL_GPL(nfs_pgio_data_destroy);
523 
524 /**
525  * nfs_pgio_rpcsetup - Set up arguments for a pageio call
526  * @hdr: The pageio hdr
527  * @count: Number of bytes to read
528  * @offset: Initial offset
529  * @how: How to commit data (writes only)
530  * @cinfo: Commit information for the call (writes only)
531  */
532 static void nfs_pgio_rpcsetup(struct nfs_pgio_header *hdr,
533 			      unsigned int count, unsigned int offset,
534 			      int how, struct nfs_commit_info *cinfo)
535 {
536 	struct nfs_page *req = hdr->req;
537 
538 	/* Set up the RPC argument and reply structs
539 	 * NB: take care not to mess about with hdr->commit et al. */
540 
541 	hdr->args.fh     = NFS_FH(hdr->inode);
542 	hdr->args.offset = req_offset(req) + offset;
543 	/* pnfs_set_layoutcommit needs this */
544 	hdr->mds_offset = hdr->args.offset;
545 	hdr->args.pgbase = req->wb_pgbase + offset;
546 	hdr->args.pages  = hdr->page_array.pagevec;
547 	hdr->args.count  = count;
548 	hdr->args.context = get_nfs_open_context(req->wb_context);
549 	hdr->args.lock_context = req->wb_lock_context;
550 	hdr->args.stable  = NFS_UNSTABLE;
551 	switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
552 	case 0:
553 		break;
554 	case FLUSH_COND_STABLE:
555 		if (nfs_reqs_to_commit(cinfo))
556 			break;
557 	default:
558 		hdr->args.stable = NFS_FILE_SYNC;
559 	}
560 
561 	hdr->res.fattr   = &hdr->fattr;
562 	hdr->res.count   = count;
563 	hdr->res.eof     = 0;
564 	hdr->res.verf    = &hdr->verf;
565 	nfs_fattr_init(&hdr->fattr);
566 }
567 
568 /**
569  * nfs_pgio_prepare - Prepare pageio hdr to go over the wire
570  * @task: The current task
571  * @calldata: pageio header to prepare
572  */
573 static void nfs_pgio_prepare(struct rpc_task *task, void *calldata)
574 {
575 	struct nfs_pgio_header *hdr = calldata;
576 	int err;
577 	err = NFS_PROTO(hdr->inode)->pgio_rpc_prepare(task, hdr);
578 	if (err)
579 		rpc_exit(task, err);
580 }
581 
582 int nfs_initiate_pgio(struct rpc_clnt *clnt, struct nfs_pgio_header *hdr,
583 		      struct rpc_cred *cred, const struct nfs_rpc_ops *rpc_ops,
584 		      const struct rpc_call_ops *call_ops, int how, int flags)
585 {
586 	struct rpc_task *task;
587 	struct rpc_message msg = {
588 		.rpc_argp = &hdr->args,
589 		.rpc_resp = &hdr->res,
590 		.rpc_cred = cred,
591 	};
592 	struct rpc_task_setup task_setup_data = {
593 		.rpc_client = clnt,
594 		.task = &hdr->task,
595 		.rpc_message = &msg,
596 		.callback_ops = call_ops,
597 		.callback_data = hdr,
598 		.workqueue = nfsiod_workqueue,
599 		.flags = RPC_TASK_ASYNC | flags,
600 	};
601 	int ret = 0;
602 
603 	hdr->rw_ops->rw_initiate(hdr, &msg, rpc_ops, &task_setup_data, how);
604 
605 	dprintk("NFS: initiated pgio call "
606 		"(req %s/%llu, %u bytes @ offset %llu)\n",
607 		hdr->inode->i_sb->s_id,
608 		(unsigned long long)NFS_FILEID(hdr->inode),
609 		hdr->args.count,
610 		(unsigned long long)hdr->args.offset);
611 
612 	task = rpc_run_task(&task_setup_data);
613 	if (IS_ERR(task)) {
614 		ret = PTR_ERR(task);
615 		goto out;
616 	}
617 	if (how & FLUSH_SYNC) {
618 		ret = rpc_wait_for_completion_task(task);
619 		if (ret == 0)
620 			ret = task->tk_status;
621 	}
622 	rpc_put_task(task);
623 out:
624 	return ret;
625 }
626 EXPORT_SYMBOL_GPL(nfs_initiate_pgio);
627 
628 /**
629  * nfs_pgio_error - Clean up from a pageio error
630  * @desc: IO descriptor
631  * @hdr: pageio header
632  */
633 static void nfs_pgio_error(struct nfs_pgio_header *hdr)
634 {
635 	set_bit(NFS_IOHDR_REDO, &hdr->flags);
636 	nfs_pgio_data_destroy(hdr);
637 	hdr->completion_ops->completion(hdr);
638 }
639 
640 /**
641  * nfs_pgio_release - Release pageio data
642  * @calldata: The pageio header to release
643  */
644 static void nfs_pgio_release(void *calldata)
645 {
646 	struct nfs_pgio_header *hdr = calldata;
647 	nfs_pgio_data_destroy(hdr);
648 	hdr->completion_ops->completion(hdr);
649 }
650 
651 static void nfs_pageio_mirror_init(struct nfs_pgio_mirror *mirror,
652 				   unsigned int bsize)
653 {
654 	INIT_LIST_HEAD(&mirror->pg_list);
655 	mirror->pg_bytes_written = 0;
656 	mirror->pg_count = 0;
657 	mirror->pg_bsize = bsize;
658 	mirror->pg_base = 0;
659 	mirror->pg_recoalesce = 0;
660 }
661 
662 /**
663  * nfs_pageio_init - initialise a page io descriptor
664  * @desc: pointer to descriptor
665  * @inode: pointer to inode
666  * @pg_ops: pointer to pageio operations
667  * @compl_ops: pointer to pageio completion operations
668  * @rw_ops: pointer to nfs read/write operations
669  * @bsize: io block size
670  * @io_flags: extra parameters for the io function
671  */
672 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
673 		     struct inode *inode,
674 		     const struct nfs_pageio_ops *pg_ops,
675 		     const struct nfs_pgio_completion_ops *compl_ops,
676 		     const struct nfs_rw_ops *rw_ops,
677 		     size_t bsize,
678 		     int io_flags)
679 {
680 	struct nfs_pgio_mirror *new;
681 	int i;
682 
683 	desc->pg_moreio = 0;
684 	desc->pg_inode = inode;
685 	desc->pg_ops = pg_ops;
686 	desc->pg_completion_ops = compl_ops;
687 	desc->pg_rw_ops = rw_ops;
688 	desc->pg_ioflags = io_flags;
689 	desc->pg_error = 0;
690 	desc->pg_lseg = NULL;
691 	desc->pg_dreq = NULL;
692 	desc->pg_layout_private = NULL;
693 	desc->pg_bsize = bsize;
694 
695 	desc->pg_mirror_count = 1;
696 	desc->pg_mirror_idx = 0;
697 
698 	if (pg_ops->pg_get_mirror_count) {
699 		/* until we have a request, we don't have an lseg and no
700 		 * idea how many mirrors there will be */
701 		new = kcalloc(NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX,
702 			      sizeof(struct nfs_pgio_mirror), GFP_KERNEL);
703 		desc->pg_mirrors_dynamic = new;
704 		desc->pg_mirrors = new;
705 
706 		for (i = 0; i < NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX; i++)
707 			nfs_pageio_mirror_init(&desc->pg_mirrors[i], bsize);
708 	} else {
709 		desc->pg_mirrors_dynamic = NULL;
710 		desc->pg_mirrors = desc->pg_mirrors_static;
711 		nfs_pageio_mirror_init(&desc->pg_mirrors[0], bsize);
712 	}
713 }
714 EXPORT_SYMBOL_GPL(nfs_pageio_init);
715 
716 /**
717  * nfs_pgio_result - Basic pageio error handling
718  * @task: The task that ran
719  * @calldata: Pageio header to check
720  */
721 static void nfs_pgio_result(struct rpc_task *task, void *calldata)
722 {
723 	struct nfs_pgio_header *hdr = calldata;
724 	struct inode *inode = hdr->inode;
725 
726 	dprintk("NFS: %s: %5u, (status %d)\n", __func__,
727 		task->tk_pid, task->tk_status);
728 
729 	if (hdr->rw_ops->rw_done(task, hdr, inode) != 0)
730 		return;
731 	if (task->tk_status < 0)
732 		nfs_set_pgio_error(hdr, task->tk_status, hdr->args.offset);
733 	else
734 		hdr->rw_ops->rw_result(task, hdr);
735 }
736 
737 /*
738  * Create an RPC task for the given read or write request and kick it.
739  * The page must have been locked by the caller.
740  *
741  * It may happen that the page we're passed is not marked dirty.
742  * This is the case if nfs_updatepage detects a conflicting request
743  * that has been written but not committed.
744  */
745 int nfs_generic_pgio(struct nfs_pageio_descriptor *desc,
746 		     struct nfs_pgio_header *hdr)
747 {
748 	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
749 
750 	struct nfs_page		*req;
751 	struct page		**pages,
752 				*last_page;
753 	struct list_head *head = &mirror->pg_list;
754 	struct nfs_commit_info cinfo;
755 	unsigned int pagecount, pageused;
756 
757 	pagecount = nfs_page_array_len(mirror->pg_base, mirror->pg_count);
758 	if (!nfs_pgarray_set(&hdr->page_array, pagecount)) {
759 		nfs_pgio_error(hdr);
760 		desc->pg_error = -ENOMEM;
761 		return desc->pg_error;
762 	}
763 
764 	nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
765 	pages = hdr->page_array.pagevec;
766 	last_page = NULL;
767 	pageused = 0;
768 	while (!list_empty(head)) {
769 		req = nfs_list_entry(head->next);
770 		nfs_list_remove_request(req);
771 		nfs_list_add_request(req, &hdr->pages);
772 
773 		if (!last_page || last_page != req->wb_page) {
774 			pageused++;
775 			if (pageused > pagecount)
776 				break;
777 			*pages++ = last_page = req->wb_page;
778 		}
779 	}
780 	if (WARN_ON_ONCE(pageused != pagecount)) {
781 		nfs_pgio_error(hdr);
782 		desc->pg_error = -EINVAL;
783 		return desc->pg_error;
784 	}
785 
786 	if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
787 	    (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
788 		desc->pg_ioflags &= ~FLUSH_COND_STABLE;
789 
790 	/* Set up the argument struct */
791 	nfs_pgio_rpcsetup(hdr, mirror->pg_count, 0, desc->pg_ioflags, &cinfo);
792 	desc->pg_rpc_callops = &nfs_pgio_common_ops;
793 	return 0;
794 }
795 EXPORT_SYMBOL_GPL(nfs_generic_pgio);
796 
797 static int nfs_generic_pg_pgios(struct nfs_pageio_descriptor *desc)
798 {
799 	struct nfs_pgio_header *hdr;
800 	int ret;
801 
802 	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
803 	if (!hdr) {
804 		desc->pg_error = -ENOMEM;
805 		return desc->pg_error;
806 	}
807 	nfs_pgheader_init(desc, hdr, nfs_pgio_header_free);
808 	ret = nfs_generic_pgio(desc, hdr);
809 	if (ret == 0)
810 		ret = nfs_initiate_pgio(NFS_CLIENT(hdr->inode),
811 					hdr,
812 					hdr->cred,
813 					NFS_PROTO(hdr->inode),
814 					desc->pg_rpc_callops,
815 					desc->pg_ioflags, 0);
816 	return ret;
817 }
818 
819 /*
820  * nfs_pageio_setup_mirroring - determine if mirroring is to be used
821  *				by calling the pg_get_mirror_count op
822  */
823 static int nfs_pageio_setup_mirroring(struct nfs_pageio_descriptor *pgio,
824 				       struct nfs_page *req)
825 {
826 	int mirror_count = 1;
827 
828 	if (!pgio->pg_ops->pg_get_mirror_count)
829 		return 0;
830 
831 	mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
832 
833 	if (pgio->pg_error < 0)
834 		return pgio->pg_error;
835 
836 	if (!mirror_count || mirror_count > NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX)
837 		return -EINVAL;
838 
839 	if (WARN_ON_ONCE(!pgio->pg_mirrors_dynamic))
840 		return -EINVAL;
841 
842 	pgio->pg_mirror_count = mirror_count;
843 
844 	return 0;
845 }
846 
847 /*
848  * nfs_pageio_stop_mirroring - stop using mirroring (set mirror count to 1)
849  */
850 void nfs_pageio_stop_mirroring(struct nfs_pageio_descriptor *pgio)
851 {
852 	pgio->pg_mirror_count = 1;
853 	pgio->pg_mirror_idx = 0;
854 }
855 
856 static void nfs_pageio_cleanup_mirroring(struct nfs_pageio_descriptor *pgio)
857 {
858 	pgio->pg_mirror_count = 1;
859 	pgio->pg_mirror_idx = 0;
860 	pgio->pg_mirrors = pgio->pg_mirrors_static;
861 	kfree(pgio->pg_mirrors_dynamic);
862 	pgio->pg_mirrors_dynamic = NULL;
863 }
864 
865 static bool nfs_match_lock_context(const struct nfs_lock_context *l1,
866 		const struct nfs_lock_context *l2)
867 {
868 	return l1->lockowner.l_owner == l2->lockowner.l_owner
869 		&& l1->lockowner.l_pid == l2->lockowner.l_pid;
870 }
871 
872 /**
873  * nfs_can_coalesce_requests - test two requests for compatibility
874  * @prev: pointer to nfs_page
875  * @req: pointer to nfs_page
876  *
877  * The nfs_page structures 'prev' and 'req' are compared to ensure that the
878  * page data area they describe is contiguous, and that their RPC
879  * credentials, NFSv4 open state, and lockowners are the same.
880  *
881  * Return 'true' if this is the case, else return 'false'.
882  */
883 static bool nfs_can_coalesce_requests(struct nfs_page *prev,
884 				      struct nfs_page *req,
885 				      struct nfs_pageio_descriptor *pgio)
886 {
887 	size_t size;
888 	struct file_lock_context *flctx;
889 
890 	if (prev) {
891 		if (!nfs_match_open_context(req->wb_context, prev->wb_context))
892 			return false;
893 		flctx = d_inode(req->wb_context->dentry)->i_flctx;
894 		if (flctx != NULL &&
895 		    !(list_empty_careful(&flctx->flc_posix) &&
896 		      list_empty_careful(&flctx->flc_flock)) &&
897 		    !nfs_match_lock_context(req->wb_lock_context,
898 					    prev->wb_lock_context))
899 			return false;
900 		if (req_offset(req) != req_offset(prev) + prev->wb_bytes)
901 			return false;
902 		if (req->wb_page == prev->wb_page) {
903 			if (req->wb_pgbase != prev->wb_pgbase + prev->wb_bytes)
904 				return false;
905 		} else {
906 			if (req->wb_pgbase != 0 ||
907 			    prev->wb_pgbase + prev->wb_bytes != PAGE_SIZE)
908 				return false;
909 		}
910 	}
911 	size = pgio->pg_ops->pg_test(pgio, prev, req);
912 	WARN_ON_ONCE(size > req->wb_bytes);
913 	if (size && size < req->wb_bytes)
914 		req->wb_bytes = size;
915 	return size > 0;
916 }
917 
918 /**
919  * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
920  * @desc: destination io descriptor
921  * @req: request
922  *
923  * Returns true if the request 'req' was successfully coalesced into the
924  * existing list of pages 'desc'.
925  */
926 static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
927 				     struct nfs_page *req)
928 {
929 	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
930 
931 	struct nfs_page *prev = NULL;
932 
933 	if (mirror->pg_count != 0) {
934 		prev = nfs_list_entry(mirror->pg_list.prev);
935 	} else {
936 		if (desc->pg_ops->pg_init)
937 			desc->pg_ops->pg_init(desc, req);
938 		if (desc->pg_error < 0)
939 			return 0;
940 		mirror->pg_base = req->wb_pgbase;
941 	}
942 	if (!nfs_can_coalesce_requests(prev, req, desc))
943 		return 0;
944 	nfs_list_remove_request(req);
945 	nfs_list_add_request(req, &mirror->pg_list);
946 	mirror->pg_count += req->wb_bytes;
947 	return 1;
948 }
949 
950 /*
951  * Helper for nfs_pageio_add_request and nfs_pageio_complete
952  */
953 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
954 {
955 	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
956 
957 
958 	if (!list_empty(&mirror->pg_list)) {
959 		int error = desc->pg_ops->pg_doio(desc);
960 		if (error < 0)
961 			desc->pg_error = error;
962 		else
963 			mirror->pg_bytes_written += mirror->pg_count;
964 	}
965 	if (list_empty(&mirror->pg_list)) {
966 		mirror->pg_count = 0;
967 		mirror->pg_base = 0;
968 	}
969 }
970 
971 /**
972  * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
973  * @desc: destination io descriptor
974  * @req: request
975  *
976  * This may split a request into subrequests which are all part of the
977  * same page group.
978  *
979  * Returns true if the request 'req' was successfully coalesced into the
980  * existing list of pages 'desc'.
981  */
982 static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
983 			   struct nfs_page *req)
984 {
985 	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
986 
987 	struct nfs_page *subreq;
988 	unsigned int bytes_left = 0;
989 	unsigned int offset, pgbase;
990 
991 	nfs_page_group_lock(req, false);
992 
993 	subreq = req;
994 	bytes_left = subreq->wb_bytes;
995 	offset = subreq->wb_offset;
996 	pgbase = subreq->wb_pgbase;
997 
998 	do {
999 		if (!nfs_pageio_do_add_request(desc, subreq)) {
1000 			/* make sure pg_test call(s) did nothing */
1001 			WARN_ON_ONCE(subreq->wb_bytes != bytes_left);
1002 			WARN_ON_ONCE(subreq->wb_offset != offset);
1003 			WARN_ON_ONCE(subreq->wb_pgbase != pgbase);
1004 
1005 			nfs_page_group_unlock(req);
1006 			desc->pg_moreio = 1;
1007 			nfs_pageio_doio(desc);
1008 			if (desc->pg_error < 0)
1009 				return 0;
1010 			if (mirror->pg_recoalesce)
1011 				return 0;
1012 			/* retry add_request for this subreq */
1013 			nfs_page_group_lock(req, false);
1014 			continue;
1015 		}
1016 
1017 		/* check for buggy pg_test call(s) */
1018 		WARN_ON_ONCE(subreq->wb_bytes + subreq->wb_pgbase > PAGE_SIZE);
1019 		WARN_ON_ONCE(subreq->wb_bytes > bytes_left);
1020 		WARN_ON_ONCE(subreq->wb_bytes == 0);
1021 
1022 		bytes_left -= subreq->wb_bytes;
1023 		offset += subreq->wb_bytes;
1024 		pgbase += subreq->wb_bytes;
1025 
1026 		if (bytes_left) {
1027 			subreq = nfs_create_request(req->wb_context,
1028 					req->wb_page,
1029 					subreq, pgbase, bytes_left);
1030 			if (IS_ERR(subreq))
1031 				goto err_ptr;
1032 			nfs_lock_request(subreq);
1033 			subreq->wb_offset  = offset;
1034 			subreq->wb_index = req->wb_index;
1035 		}
1036 	} while (bytes_left > 0);
1037 
1038 	nfs_page_group_unlock(req);
1039 	return 1;
1040 err_ptr:
1041 	desc->pg_error = PTR_ERR(subreq);
1042 	nfs_page_group_unlock(req);
1043 	return 0;
1044 }
1045 
1046 static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc)
1047 {
1048 	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1049 	LIST_HEAD(head);
1050 
1051 	do {
1052 		list_splice_init(&mirror->pg_list, &head);
1053 		mirror->pg_bytes_written -= mirror->pg_count;
1054 		mirror->pg_count = 0;
1055 		mirror->pg_base = 0;
1056 		mirror->pg_recoalesce = 0;
1057 
1058 		while (!list_empty(&head)) {
1059 			struct nfs_page *req;
1060 
1061 			req = list_first_entry(&head, struct nfs_page, wb_list);
1062 			nfs_list_remove_request(req);
1063 			if (__nfs_pageio_add_request(desc, req))
1064 				continue;
1065 			if (desc->pg_error < 0) {
1066 				list_splice_tail(&head, &mirror->pg_list);
1067 				mirror->pg_recoalesce = 1;
1068 				return 0;
1069 			}
1070 			break;
1071 		}
1072 	} while (mirror->pg_recoalesce);
1073 	return 1;
1074 }
1075 
1076 static int nfs_pageio_add_request_mirror(struct nfs_pageio_descriptor *desc,
1077 		struct nfs_page *req)
1078 {
1079 	int ret;
1080 
1081 	do {
1082 		ret = __nfs_pageio_add_request(desc, req);
1083 		if (ret)
1084 			break;
1085 		if (desc->pg_error < 0)
1086 			break;
1087 		ret = nfs_do_recoalesce(desc);
1088 	} while (ret);
1089 
1090 	return ret;
1091 }
1092 
1093 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
1094 			   struct nfs_page *req)
1095 {
1096 	u32 midx;
1097 	unsigned int pgbase, offset, bytes;
1098 	struct nfs_page *dupreq, *lastreq;
1099 
1100 	pgbase = req->wb_pgbase;
1101 	offset = req->wb_offset;
1102 	bytes = req->wb_bytes;
1103 
1104 	nfs_pageio_setup_mirroring(desc, req);
1105 	if (desc->pg_error < 0)
1106 		goto out_failed;
1107 
1108 	for (midx = 0; midx < desc->pg_mirror_count; midx++) {
1109 		if (midx) {
1110 			nfs_page_group_lock(req, false);
1111 
1112 			/* find the last request */
1113 			for (lastreq = req->wb_head;
1114 			     lastreq->wb_this_page != req->wb_head;
1115 			     lastreq = lastreq->wb_this_page)
1116 				;
1117 
1118 			dupreq = nfs_create_request(req->wb_context,
1119 					req->wb_page, lastreq, pgbase, bytes);
1120 
1121 			if (IS_ERR(dupreq)) {
1122 				nfs_page_group_unlock(req);
1123 				desc->pg_error = PTR_ERR(dupreq);
1124 				goto out_failed;
1125 			}
1126 
1127 			nfs_lock_request(dupreq);
1128 			nfs_page_group_unlock(req);
1129 			dupreq->wb_offset = offset;
1130 			dupreq->wb_index = req->wb_index;
1131 		} else
1132 			dupreq = req;
1133 
1134 		if (nfs_pgio_has_mirroring(desc))
1135 			desc->pg_mirror_idx = midx;
1136 		if (!nfs_pageio_add_request_mirror(desc, dupreq))
1137 			goto out_failed;
1138 	}
1139 
1140 	return 1;
1141 
1142 out_failed:
1143 	/*
1144 	 * We might have failed before sending any reqs over wire.
1145 	 * Clean up rest of the reqs in mirror pg_list.
1146 	 */
1147 	if (desc->pg_error) {
1148 		struct nfs_pgio_mirror *mirror;
1149 		void (*func)(struct list_head *);
1150 
1151 		/* remember fatal errors */
1152 		if (nfs_error_is_fatal(desc->pg_error))
1153 			mapping_set_error(desc->pg_inode->i_mapping,
1154 					  desc->pg_error);
1155 
1156 		func = desc->pg_completion_ops->error_cleanup;
1157 		for (midx = 0; midx < desc->pg_mirror_count; midx++) {
1158 			mirror = &desc->pg_mirrors[midx];
1159 			func(&mirror->pg_list);
1160 		}
1161 	}
1162 	return 0;
1163 }
1164 
1165 /*
1166  * nfs_pageio_complete_mirror - Complete I/O on the current mirror of an
1167  *				nfs_pageio_descriptor
1168  * @desc: pointer to io descriptor
1169  * @mirror_idx: pointer to mirror index
1170  */
1171 static void nfs_pageio_complete_mirror(struct nfs_pageio_descriptor *desc,
1172 				       u32 mirror_idx)
1173 {
1174 	struct nfs_pgio_mirror *mirror = &desc->pg_mirrors[mirror_idx];
1175 	u32 restore_idx = desc->pg_mirror_idx;
1176 
1177 	if (nfs_pgio_has_mirroring(desc))
1178 		desc->pg_mirror_idx = mirror_idx;
1179 	for (;;) {
1180 		nfs_pageio_doio(desc);
1181 		if (!mirror->pg_recoalesce)
1182 			break;
1183 		if (!nfs_do_recoalesce(desc))
1184 			break;
1185 	}
1186 	desc->pg_mirror_idx = restore_idx;
1187 }
1188 
1189 /*
1190  * nfs_pageio_resend - Transfer requests to new descriptor and resend
1191  * @hdr - the pgio header to move request from
1192  * @desc - the pageio descriptor to add requests to
1193  *
1194  * Try to move each request (nfs_page) from @hdr to @desc then attempt
1195  * to send them.
1196  *
1197  * Returns 0 on success and < 0 on error.
1198  */
1199 int nfs_pageio_resend(struct nfs_pageio_descriptor *desc,
1200 		      struct nfs_pgio_header *hdr)
1201 {
1202 	LIST_HEAD(failed);
1203 
1204 	desc->pg_dreq = hdr->dreq;
1205 	while (!list_empty(&hdr->pages)) {
1206 		struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1207 
1208 		nfs_list_remove_request(req);
1209 		if (!nfs_pageio_add_request(desc, req))
1210 			nfs_list_add_request(req, &failed);
1211 	}
1212 	nfs_pageio_complete(desc);
1213 	if (!list_empty(&failed)) {
1214 		list_move(&failed, &hdr->pages);
1215 		return desc->pg_error < 0 ? desc->pg_error : -EIO;
1216 	}
1217 	return 0;
1218 }
1219 EXPORT_SYMBOL_GPL(nfs_pageio_resend);
1220 
1221 /**
1222  * nfs_pageio_complete - Complete I/O then cleanup an nfs_pageio_descriptor
1223  * @desc: pointer to io descriptor
1224  */
1225 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
1226 {
1227 	u32 midx;
1228 
1229 	for (midx = 0; midx < desc->pg_mirror_count; midx++)
1230 		nfs_pageio_complete_mirror(desc, midx);
1231 
1232 	if (desc->pg_ops->pg_cleanup)
1233 		desc->pg_ops->pg_cleanup(desc);
1234 	nfs_pageio_cleanup_mirroring(desc);
1235 }
1236 
1237 /**
1238  * nfs_pageio_cond_complete - Conditional I/O completion
1239  * @desc: pointer to io descriptor
1240  * @index: page index
1241  *
1242  * It is important to ensure that processes don't try to take locks
1243  * on non-contiguous ranges of pages as that might deadlock. This
1244  * function should be called before attempting to wait on a locked
1245  * nfs_page. It will complete the I/O if the page index 'index'
1246  * is not contiguous with the existing list of pages in 'desc'.
1247  */
1248 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
1249 {
1250 	struct nfs_pgio_mirror *mirror;
1251 	struct nfs_page *prev;
1252 	u32 midx;
1253 
1254 	for (midx = 0; midx < desc->pg_mirror_count; midx++) {
1255 		mirror = &desc->pg_mirrors[midx];
1256 		if (!list_empty(&mirror->pg_list)) {
1257 			prev = nfs_list_entry(mirror->pg_list.prev);
1258 			if (index != prev->wb_index + 1)
1259 				nfs_pageio_complete_mirror(desc, midx);
1260 		}
1261 	}
1262 }
1263 
1264 int __init nfs_init_nfspagecache(void)
1265 {
1266 	nfs_page_cachep = kmem_cache_create("nfs_page",
1267 					    sizeof(struct nfs_page),
1268 					    0, SLAB_HWCACHE_ALIGN,
1269 					    NULL);
1270 	if (nfs_page_cachep == NULL)
1271 		return -ENOMEM;
1272 
1273 	return 0;
1274 }
1275 
1276 void nfs_destroy_nfspagecache(void)
1277 {
1278 	kmem_cache_destroy(nfs_page_cachep);
1279 }
1280 
1281 static const struct rpc_call_ops nfs_pgio_common_ops = {
1282 	.rpc_call_prepare = nfs_pgio_prepare,
1283 	.rpc_call_done = nfs_pgio_result,
1284 	.rpc_release = nfs_pgio_release,
1285 };
1286 
1287 const struct nfs_pageio_ops nfs_pgio_rw_ops = {
1288 	.pg_test = nfs_generic_pg_test,
1289 	.pg_doio = nfs_generic_pg_pgios,
1290 };
1291