2 * linux/fs/nfs/pagelist.c
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.
8 * Copyright 2000, 2001 (c) Trond Myklebust <trond.myklebust@fys.uio.no>
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>
27 #define NFSDBG_FACILITY NFSDBG_PAGECACHE
29 static struct kmem_cache *nfs_page_cachep;
30 static const struct rpc_call_ops nfs_pgio_common_ops;
32 static bool nfs_pgarray_set(struct nfs_page_array *p, unsigned int pagecount,
35 p->npages = pagecount;
36 if (pagecount <= ARRAY_SIZE(p->page_array))
37 p->pagevec = p->page_array;
39 p->pagevec = kcalloc(pagecount, sizeof(struct page *), gfp_flags);
43 return p->pagevec != NULL;
46 struct nfs_pgio_mirror *
47 nfs_pgio_current_mirror(struct nfs_pageio_descriptor *desc)
49 return nfs_pgio_has_mirroring(desc) ?
50 &desc->pg_mirrors[desc->pg_mirror_idx] :
53 EXPORT_SYMBOL_GPL(nfs_pgio_current_mirror);
55 void nfs_pgheader_init(struct nfs_pageio_descriptor *desc,
56 struct nfs_pgio_header *hdr,
57 void (*release)(struct nfs_pgio_header *hdr))
59 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
62 hdr->req = nfs_list_entry(mirror->pg_list.next);
63 hdr->inode = desc->pg_inode;
64 hdr->cred = hdr->req->wb_context->cred;
65 hdr->io_start = req_offset(hdr->req);
66 hdr->good_bytes = mirror->pg_count;
67 hdr->dreq = desc->pg_dreq;
68 hdr->layout_private = desc->pg_layout_private;
69 hdr->release = release;
70 hdr->completion_ops = desc->pg_completion_ops;
71 if (hdr->completion_ops->init_hdr)
72 hdr->completion_ops->init_hdr(hdr);
74 hdr->pgio_mirror_idx = desc->pg_mirror_idx;
76 EXPORT_SYMBOL_GPL(nfs_pgheader_init);
78 void nfs_set_pgio_error(struct nfs_pgio_header *hdr, int error, loff_t pos)
80 spin_lock(&hdr->lock);
81 if (!test_and_set_bit(NFS_IOHDR_ERROR, &hdr->flags)
82 || pos < hdr->io_start + hdr->good_bytes) {
83 clear_bit(NFS_IOHDR_EOF, &hdr->flags);
84 hdr->good_bytes = pos - hdr->io_start;
87 spin_unlock(&hdr->lock);
90 static inline struct nfs_page *
93 struct nfs_page *p = kmem_cache_zalloc(nfs_page_cachep, GFP_NOIO);
95 INIT_LIST_HEAD(&p->wb_list);
100 nfs_page_free(struct nfs_page *p)
102 kmem_cache_free(nfs_page_cachep, p);
106 * nfs_iocounter_wait - wait for i/o to complete
107 * @l_ctx: nfs_lock_context with io_counter to use
109 * returns -ERESTARTSYS if interrupted by a fatal signal.
110 * Otherwise returns 0 once the io_count hits 0.
113 nfs_iocounter_wait(struct nfs_lock_context *l_ctx)
115 return wait_on_atomic_t(&l_ctx->io_count, nfs_wait_atomic_killable,
120 * nfs_page_group_lock - lock the head of the page group
121 * @req - request in group that is to be locked
122 * @nonblock - if true don't block waiting for lock
124 * this lock must be held if modifying the page group list
126 * return 0 on success, < 0 on error: -EDELAY if nonblocking or the
127 * result from wait_on_bit_lock
129 * NOTE: calling with nonblock=false should always have set the
130 * lock bit (see fs/buffer.c and other uses of wait_on_bit_lock
131 * with TASK_UNINTERRUPTIBLE), so there is no need to check the result.
134 nfs_page_group_lock(struct nfs_page *req, bool nonblock)
136 struct nfs_page *head = req->wb_head;
138 WARN_ON_ONCE(head != head->wb_head);
140 if (!test_and_set_bit(PG_HEADLOCK, &head->wb_flags))
144 return wait_on_bit_lock(&head->wb_flags, PG_HEADLOCK,
145 TASK_UNINTERRUPTIBLE);
151 * nfs_page_group_lock_wait - wait for the lock to clear, but don't grab it
152 * @req - a request in the group
154 * This is a blocking call to wait for the group lock to be cleared.
157 nfs_page_group_lock_wait(struct nfs_page *req)
159 struct nfs_page *head = req->wb_head;
161 WARN_ON_ONCE(head != head->wb_head);
163 wait_on_bit(&head->wb_flags, PG_HEADLOCK,
164 TASK_UNINTERRUPTIBLE);
168 * nfs_page_group_unlock - unlock the head of the page group
169 * @req - request in group that is to be unlocked
172 nfs_page_group_unlock(struct nfs_page *req)
174 struct nfs_page *head = req->wb_head;
176 WARN_ON_ONCE(head != head->wb_head);
178 smp_mb__before_atomic();
179 clear_bit(PG_HEADLOCK, &head->wb_flags);
180 smp_mb__after_atomic();
181 wake_up_bit(&head->wb_flags, PG_HEADLOCK);
185 * nfs_page_group_sync_on_bit_locked
187 * must be called with page group lock held
190 nfs_page_group_sync_on_bit_locked(struct nfs_page *req, unsigned int bit)
192 struct nfs_page *head = req->wb_head;
193 struct nfs_page *tmp;
195 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_flags));
196 WARN_ON_ONCE(test_and_set_bit(bit, &req->wb_flags));
198 tmp = req->wb_this_page;
200 if (!test_bit(bit, &tmp->wb_flags))
202 tmp = tmp->wb_this_page;
205 /* true! reset all bits */
208 clear_bit(bit, &tmp->wb_flags);
209 tmp = tmp->wb_this_page;
210 } while (tmp != req);
216 * nfs_page_group_sync_on_bit - set bit on current request, but only
217 * return true if the bit is set for all requests in page group
218 * @req - request in page group
219 * @bit - PG_* bit that is used to sync page group
221 bool nfs_page_group_sync_on_bit(struct nfs_page *req, unsigned int bit)
225 nfs_page_group_lock(req, false);
226 ret = nfs_page_group_sync_on_bit_locked(req, bit);
227 nfs_page_group_unlock(req);
233 * nfs_page_group_init - Initialize the page group linkage for @req
234 * @req - a new nfs request
235 * @prev - the previous request in page group, or NULL if @req is the first
236 * or only request in the group (the head).
239 nfs_page_group_init(struct nfs_page *req, struct nfs_page *prev)
242 WARN_ON_ONCE(prev == req);
247 req->wb_this_page = req;
250 WARN_ON_ONCE(prev->wb_this_page != prev->wb_head);
251 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &prev->wb_head->wb_flags));
252 req->wb_head = prev->wb_head;
253 req->wb_this_page = prev->wb_this_page;
254 prev->wb_this_page = req;
256 /* All subrequests take a ref on the head request until
257 * nfs_page_group_destroy is called */
258 kref_get(&req->wb_head->wb_kref);
260 /* grab extra ref and bump the request count if head request
261 * has extra ref from the write/commit path to handle handoff
262 * between write and commit lists. */
263 if (test_bit(PG_INODE_REF, &prev->wb_head->wb_flags)) {
264 inode = page_file_mapping(req->wb_page)->host;
265 set_bit(PG_INODE_REF, &req->wb_flags);
266 kref_get(&req->wb_kref);
267 spin_lock(&inode->i_lock);
268 NFS_I(inode)->nrequests++;
269 spin_unlock(&inode->i_lock);
275 * nfs_page_group_destroy - sync the destruction of page groups
276 * @req - request that no longer needs the page group
278 * releases the page group reference from each member once all
279 * members have called this function.
282 nfs_page_group_destroy(struct kref *kref)
284 struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
285 struct nfs_page *tmp, *next;
287 /* subrequests must release the ref on the head request */
288 if (req->wb_head != req)
289 nfs_release_request(req->wb_head);
291 if (!nfs_page_group_sync_on_bit(req, PG_TEARDOWN))
296 next = tmp->wb_this_page;
297 /* unlink and free */
298 tmp->wb_this_page = tmp;
300 nfs_free_request(tmp);
302 } while (tmp != req);
306 * nfs_create_request - Create an NFS read/write request.
307 * @ctx: open context to use
308 * @page: page to write
309 * @last: last nfs request created for this page group or NULL if head
310 * @offset: starting offset within the page for the write
311 * @count: number of bytes to read/write
313 * The page must be locked by the caller. This makes sure we never
314 * create two different requests for the same page.
315 * User should ensure it is safe to sleep in this function.
318 nfs_create_request(struct nfs_open_context *ctx, struct page *page,
319 struct nfs_page *last, unsigned int offset,
322 struct nfs_page *req;
323 struct nfs_lock_context *l_ctx;
325 if (test_bit(NFS_CONTEXT_BAD, &ctx->flags))
326 return ERR_PTR(-EBADF);
327 /* try to allocate the request struct */
328 req = nfs_page_alloc();
330 return ERR_PTR(-ENOMEM);
332 /* get lock context early so we can deal with alloc failures */
333 l_ctx = nfs_get_lock_context(ctx);
336 return ERR_CAST(l_ctx);
338 req->wb_lock_context = l_ctx;
339 atomic_inc(&l_ctx->io_count);
341 /* Initialize the request struct. Initially, we assume a
342 * long write-back delay. This will be adjusted in
343 * update_nfs_request below if the region is not locked. */
346 req->wb_index = page_index(page);
349 req->wb_offset = offset;
350 req->wb_pgbase = offset;
351 req->wb_bytes = count;
352 req->wb_context = get_nfs_open_context(ctx);
353 kref_init(&req->wb_kref);
354 nfs_page_group_init(req, last);
359 * nfs_unlock_request - Unlock request and wake up sleepers.
362 void nfs_unlock_request(struct nfs_page *req)
364 if (!NFS_WBACK_BUSY(req)) {
365 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
368 smp_mb__before_atomic();
369 clear_bit(PG_BUSY, &req->wb_flags);
370 smp_mb__after_atomic();
371 wake_up_bit(&req->wb_flags, PG_BUSY);
375 * nfs_unlock_and_release_request - Unlock request and release the nfs_page
378 void nfs_unlock_and_release_request(struct nfs_page *req)
380 nfs_unlock_request(req);
381 nfs_release_request(req);
385 * nfs_clear_request - Free up all resources allocated to the request
388 * Release page and open context resources associated with a read/write
389 * request after it has completed.
391 static void nfs_clear_request(struct nfs_page *req)
393 struct page *page = req->wb_page;
394 struct nfs_open_context *ctx = req->wb_context;
395 struct nfs_lock_context *l_ctx = req->wb_lock_context;
402 if (atomic_dec_and_test(&l_ctx->io_count))
403 wake_up_atomic_t(&l_ctx->io_count);
404 nfs_put_lock_context(l_ctx);
405 req->wb_lock_context = NULL;
408 put_nfs_open_context(ctx);
409 req->wb_context = NULL;
414 * nfs_release_request - Release the count on an NFS read/write request
415 * @req: request to release
417 * Note: Should never be called with the spinlock held!
419 void nfs_free_request(struct nfs_page *req)
421 WARN_ON_ONCE(req->wb_this_page != req);
423 /* extra debug: make sure no sync bits are still set */
424 WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
425 WARN_ON_ONCE(test_bit(PG_UNLOCKPAGE, &req->wb_flags));
426 WARN_ON_ONCE(test_bit(PG_UPTODATE, &req->wb_flags));
427 WARN_ON_ONCE(test_bit(PG_WB_END, &req->wb_flags));
428 WARN_ON_ONCE(test_bit(PG_REMOVE, &req->wb_flags));
430 /* Release struct file and open context */
431 nfs_clear_request(req);
435 void nfs_release_request(struct nfs_page *req)
437 kref_put(&req->wb_kref, nfs_page_group_destroy);
441 * nfs_wait_on_request - Wait for a request to complete.
442 * @req: request to wait upon.
444 * Interruptible by fatal signals only.
445 * The user is responsible for holding a count on the request.
448 nfs_wait_on_request(struct nfs_page *req)
450 return wait_on_bit_io(&req->wb_flags, PG_BUSY,
451 TASK_UNINTERRUPTIBLE);
455 * nfs_generic_pg_test - determine if requests can be coalesced
456 * @desc: pointer to descriptor
457 * @prev: previous request in desc, or NULL
460 * Returns zero if @req can be coalesced into @desc, otherwise it returns
461 * the size of the request.
463 size_t nfs_generic_pg_test(struct nfs_pageio_descriptor *desc,
464 struct nfs_page *prev, struct nfs_page *req)
466 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
469 if (mirror->pg_count > mirror->pg_bsize) {
470 /* should never happen */
476 * Limit the request size so that we can still allocate a page array
477 * for it without upsetting the slab allocator.
479 if (((mirror->pg_count + req->wb_bytes) >> PAGE_SHIFT) *
480 sizeof(struct page *) > PAGE_SIZE)
483 return min(mirror->pg_bsize - mirror->pg_count, (size_t)req->wb_bytes);
485 EXPORT_SYMBOL_GPL(nfs_generic_pg_test);
487 struct nfs_pgio_header *nfs_pgio_header_alloc(const struct nfs_rw_ops *ops)
489 struct nfs_pgio_header *hdr = ops->rw_alloc_header();
492 INIT_LIST_HEAD(&hdr->pages);
493 spin_lock_init(&hdr->lock);
498 EXPORT_SYMBOL_GPL(nfs_pgio_header_alloc);
501 * nfs_pgio_data_destroy - make @hdr suitable for reuse
503 * Frees memory and releases refs from nfs_generic_pgio, so that it may
506 * @hdr: A header that has had nfs_generic_pgio called
508 static void nfs_pgio_data_destroy(struct nfs_pgio_header *hdr)
510 if (hdr->args.context)
511 put_nfs_open_context(hdr->args.context);
512 if (hdr->page_array.pagevec != hdr->page_array.page_array)
513 kfree(hdr->page_array.pagevec);
517 * nfs_pgio_header_free - Free a read or write header
518 * @hdr: The header to free
520 void nfs_pgio_header_free(struct nfs_pgio_header *hdr)
522 nfs_pgio_data_destroy(hdr);
523 hdr->rw_ops->rw_free_header(hdr);
525 EXPORT_SYMBOL_GPL(nfs_pgio_header_free);
528 * nfs_pgio_rpcsetup - Set up arguments for a pageio call
529 * @hdr: The pageio hdr
530 * @count: Number of bytes to read
531 * @offset: Initial offset
532 * @how: How to commit data (writes only)
533 * @cinfo: Commit information for the call (writes only)
535 static void nfs_pgio_rpcsetup(struct nfs_pgio_header *hdr,
536 unsigned int count, unsigned int offset,
537 int how, struct nfs_commit_info *cinfo)
539 struct nfs_page *req = hdr->req;
541 /* Set up the RPC argument and reply structs
542 * NB: take care not to mess about with hdr->commit et al. */
544 hdr->args.fh = NFS_FH(hdr->inode);
545 hdr->args.offset = req_offset(req) + offset;
546 /* pnfs_set_layoutcommit needs this */
547 hdr->mds_offset = hdr->args.offset;
548 hdr->args.pgbase = req->wb_pgbase + offset;
549 hdr->args.pages = hdr->page_array.pagevec;
550 hdr->args.count = count;
551 hdr->args.context = get_nfs_open_context(req->wb_context);
552 hdr->args.lock_context = req->wb_lock_context;
553 hdr->args.stable = NFS_UNSTABLE;
554 switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
557 case FLUSH_COND_STABLE:
558 if (nfs_reqs_to_commit(cinfo))
561 hdr->args.stable = NFS_FILE_SYNC;
564 hdr->res.fattr = &hdr->fattr;
567 hdr->res.verf = &hdr->verf;
568 nfs_fattr_init(&hdr->fattr);
572 * nfs_pgio_prepare - Prepare pageio hdr to go over the wire
573 * @task: The current task
574 * @calldata: pageio header to prepare
576 static void nfs_pgio_prepare(struct rpc_task *task, void *calldata)
578 struct nfs_pgio_header *hdr = calldata;
580 err = NFS_PROTO(hdr->inode)->pgio_rpc_prepare(task, hdr);
585 int nfs_initiate_pgio(struct rpc_clnt *clnt, struct nfs_pgio_header *hdr,
586 struct rpc_cred *cred, const struct nfs_rpc_ops *rpc_ops,
587 const struct rpc_call_ops *call_ops, int how, int flags)
589 struct rpc_task *task;
590 struct rpc_message msg = {
591 .rpc_argp = &hdr->args,
592 .rpc_resp = &hdr->res,
595 struct rpc_task_setup task_setup_data = {
599 .callback_ops = call_ops,
600 .callback_data = hdr,
601 .workqueue = nfsiod_workqueue,
602 .flags = RPC_TASK_ASYNC | flags,
606 hdr->rw_ops->rw_initiate(hdr, &msg, rpc_ops, &task_setup_data, how);
608 dprintk("NFS: initiated pgio call "
609 "(req %s/%llu, %u bytes @ offset %llu)\n",
610 hdr->inode->i_sb->s_id,
611 (unsigned long long)NFS_FILEID(hdr->inode),
613 (unsigned long long)hdr->args.offset);
615 task = rpc_run_task(&task_setup_data);
620 if (how & FLUSH_SYNC) {
621 ret = rpc_wait_for_completion_task(task);
623 ret = task->tk_status;
629 EXPORT_SYMBOL_GPL(nfs_initiate_pgio);
632 * nfs_pgio_error - Clean up from a pageio error
633 * @desc: IO descriptor
634 * @hdr: pageio header
636 static void nfs_pgio_error(struct nfs_pgio_header *hdr)
638 set_bit(NFS_IOHDR_REDO, &hdr->flags);
639 hdr->completion_ops->completion(hdr);
643 * nfs_pgio_release - Release pageio data
644 * @calldata: The pageio header to release
646 static void nfs_pgio_release(void *calldata)
648 struct nfs_pgio_header *hdr = calldata;
649 hdr->completion_ops->completion(hdr);
652 static void nfs_pageio_mirror_init(struct nfs_pgio_mirror *mirror,
655 INIT_LIST_HEAD(&mirror->pg_list);
656 mirror->pg_bytes_written = 0;
657 mirror->pg_count = 0;
658 mirror->pg_bsize = bsize;
660 mirror->pg_recoalesce = 0;
664 * nfs_pageio_init - initialise a page io descriptor
665 * @desc: pointer to descriptor
666 * @inode: pointer to inode
667 * @pg_ops: pointer to pageio operations
668 * @compl_ops: pointer to pageio completion operations
669 * @rw_ops: pointer to nfs read/write operations
670 * @bsize: io block size
671 * @io_flags: extra parameters for the io function
673 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
675 const struct nfs_pageio_ops *pg_ops,
676 const struct nfs_pgio_completion_ops *compl_ops,
677 const struct nfs_rw_ops *rw_ops,
681 struct nfs_pgio_mirror *new;
683 gfp_t gfp_flags = GFP_KERNEL;
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;
692 desc->pg_lseg = NULL;
693 desc->pg_dreq = NULL;
694 desc->pg_layout_private = NULL;
695 desc->pg_bsize = bsize;
697 desc->pg_mirror_count = 1;
698 desc->pg_mirror_idx = 0;
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 if (desc->pg_rw_ops->rw_mode == FMODE_WRITE)
704 gfp_flags = GFP_NOIO;
705 new = kcalloc(NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX,
706 sizeof(struct nfs_pgio_mirror), gfp_flags);
707 desc->pg_mirrors_dynamic = new;
708 desc->pg_mirrors = new;
710 for (i = 0; i < NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX; i++)
711 nfs_pageio_mirror_init(&desc->pg_mirrors[i], bsize);
713 desc->pg_mirrors_dynamic = NULL;
714 desc->pg_mirrors = desc->pg_mirrors_static;
715 nfs_pageio_mirror_init(&desc->pg_mirrors[0], bsize);
718 EXPORT_SYMBOL_GPL(nfs_pageio_init);
721 * nfs_pgio_result - Basic pageio error handling
722 * @task: The task that ran
723 * @calldata: Pageio header to check
725 static void nfs_pgio_result(struct rpc_task *task, void *calldata)
727 struct nfs_pgio_header *hdr = calldata;
728 struct inode *inode = hdr->inode;
730 dprintk("NFS: %s: %5u, (status %d)\n", __func__,
731 task->tk_pid, task->tk_status);
733 if (hdr->rw_ops->rw_done(task, hdr, inode) != 0)
735 if (task->tk_status < 0)
736 nfs_set_pgio_error(hdr, task->tk_status, hdr->args.offset);
738 hdr->rw_ops->rw_result(task, hdr);
742 * Create an RPC task for the given read or write request and kick it.
743 * The page must have been locked by the caller.
745 * It may happen that the page we're passed is not marked dirty.
746 * This is the case if nfs_updatepage detects a conflicting request
747 * that has been written but not committed.
749 int nfs_generic_pgio(struct nfs_pageio_descriptor *desc,
750 struct nfs_pgio_header *hdr)
752 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
754 struct nfs_page *req;
757 struct list_head *head = &mirror->pg_list;
758 struct nfs_commit_info cinfo;
759 unsigned int pagecount, pageused;
760 gfp_t gfp_flags = GFP_KERNEL;
762 pagecount = nfs_page_array_len(mirror->pg_base, mirror->pg_count);
763 if (desc->pg_rw_ops->rw_mode == FMODE_WRITE)
764 gfp_flags = GFP_NOIO;
765 if (!nfs_pgarray_set(&hdr->page_array, pagecount, gfp_flags)) {
767 desc->pg_error = -ENOMEM;
768 return desc->pg_error;
771 nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
772 pages = hdr->page_array.pagevec;
775 while (!list_empty(head)) {
776 req = nfs_list_entry(head->next);
777 nfs_list_remove_request(req);
778 nfs_list_add_request(req, &hdr->pages);
780 if (!last_page || last_page != req->wb_page) {
782 if (pageused > pagecount)
784 *pages++ = last_page = req->wb_page;
787 if (WARN_ON_ONCE(pageused != pagecount)) {
789 desc->pg_error = -EINVAL;
790 return desc->pg_error;
793 if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
794 (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
795 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
797 /* Set up the argument struct */
798 nfs_pgio_rpcsetup(hdr, mirror->pg_count, 0, desc->pg_ioflags, &cinfo);
799 desc->pg_rpc_callops = &nfs_pgio_common_ops;
802 EXPORT_SYMBOL_GPL(nfs_generic_pgio);
804 static int nfs_generic_pg_pgios(struct nfs_pageio_descriptor *desc)
806 struct nfs_pgio_header *hdr;
809 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
811 desc->pg_error = -ENOMEM;
812 return desc->pg_error;
814 nfs_pgheader_init(desc, hdr, nfs_pgio_header_free);
815 ret = nfs_generic_pgio(desc, hdr);
817 ret = nfs_initiate_pgio(NFS_CLIENT(hdr->inode),
820 NFS_PROTO(hdr->inode),
821 desc->pg_rpc_callops,
822 desc->pg_ioflags, 0);
827 * nfs_pageio_setup_mirroring - determine if mirroring is to be used
828 * by calling the pg_get_mirror_count op
830 static int nfs_pageio_setup_mirroring(struct nfs_pageio_descriptor *pgio,
831 struct nfs_page *req)
833 int mirror_count = 1;
835 if (!pgio->pg_ops->pg_get_mirror_count)
838 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
840 if (pgio->pg_error < 0)
841 return pgio->pg_error;
843 if (!mirror_count || mirror_count > NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX)
846 if (WARN_ON_ONCE(!pgio->pg_mirrors_dynamic))
849 pgio->pg_mirror_count = mirror_count;
854 static void nfs_pageio_cleanup_mirroring(struct nfs_pageio_descriptor *pgio)
856 pgio->pg_mirror_count = 1;
857 pgio->pg_mirror_idx = 0;
858 pgio->pg_mirrors = pgio->pg_mirrors_static;
859 kfree(pgio->pg_mirrors_dynamic);
860 pgio->pg_mirrors_dynamic = NULL;
863 static bool nfs_match_lock_context(const struct nfs_lock_context *l1,
864 const struct nfs_lock_context *l2)
866 return l1->lockowner.l_owner == l2->lockowner.l_owner
867 && l1->lockowner.l_pid == l2->lockowner.l_pid;
871 * nfs_can_coalesce_requests - test two requests for compatibility
872 * @prev: pointer to nfs_page
873 * @req: pointer to nfs_page
875 * The nfs_page structures 'prev' and 'req' are compared to ensure that the
876 * page data area they describe is contiguous, and that their RPC
877 * credentials, NFSv4 open state, and lockowners are the same.
879 * Return 'true' if this is the case, else return 'false'.
881 static bool nfs_can_coalesce_requests(struct nfs_page *prev,
882 struct nfs_page *req,
883 struct nfs_pageio_descriptor *pgio)
886 struct file_lock_context *flctx;
889 if (!nfs_match_open_context(req->wb_context, prev->wb_context))
891 flctx = d_inode(req->wb_context->dentry)->i_flctx;
893 !(list_empty_careful(&flctx->flc_posix) &&
894 list_empty_careful(&flctx->flc_flock)) &&
895 !nfs_match_lock_context(req->wb_lock_context,
896 prev->wb_lock_context))
898 if (req_offset(req) != req_offset(prev) + prev->wb_bytes)
900 if (req->wb_page == prev->wb_page) {
901 if (req->wb_pgbase != prev->wb_pgbase + prev->wb_bytes)
904 if (req->wb_pgbase != 0 ||
905 prev->wb_pgbase + prev->wb_bytes != PAGE_SIZE)
909 size = pgio->pg_ops->pg_test(pgio, prev, req);
910 WARN_ON_ONCE(size > req->wb_bytes);
911 if (size && size < req->wb_bytes)
912 req->wb_bytes = size;
917 * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
918 * @desc: destination io descriptor
921 * Returns true if the request 'req' was successfully coalesced into the
922 * existing list of pages 'desc'.
924 static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
925 struct nfs_page *req)
927 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
929 struct nfs_page *prev = NULL;
931 if (mirror->pg_count != 0) {
932 prev = nfs_list_entry(mirror->pg_list.prev);
934 if (desc->pg_ops->pg_init)
935 desc->pg_ops->pg_init(desc, req);
936 if (desc->pg_error < 0)
938 mirror->pg_base = req->wb_pgbase;
940 if (!nfs_can_coalesce_requests(prev, req, desc))
942 nfs_list_remove_request(req);
943 nfs_list_add_request(req, &mirror->pg_list);
944 mirror->pg_count += req->wb_bytes;
949 * Helper for nfs_pageio_add_request and nfs_pageio_complete
951 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
953 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
955 if (!list_empty(&mirror->pg_list)) {
956 int error = desc->pg_ops->pg_doio(desc);
958 desc->pg_error = error;
959 if (list_empty(&mirror->pg_list)) {
960 mirror->pg_bytes_written += mirror->pg_count;
961 mirror->pg_count = 0;
963 mirror->pg_recoalesce = 0;
969 nfs_pageio_cleanup_request(struct nfs_pageio_descriptor *desc,
970 struct nfs_page *req)
974 nfs_list_remove_request(req);
975 nfs_list_add_request(req, &head);
976 desc->pg_completion_ops->error_cleanup(&head);
980 * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
981 * @desc: destination io descriptor
984 * This may split a request into subrequests which are all part of the
987 * Returns true if the request 'req' was successfully coalesced into the
988 * existing list of pages 'desc'.
990 static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
991 struct nfs_page *req)
993 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
995 struct nfs_page *subreq;
996 unsigned int bytes_left = 0;
997 unsigned int offset, pgbase;
999 nfs_page_group_lock(req, false);
1002 bytes_left = subreq->wb_bytes;
1003 offset = subreq->wb_offset;
1004 pgbase = subreq->wb_pgbase;
1007 if (!nfs_pageio_do_add_request(desc, subreq)) {
1008 /* make sure pg_test call(s) did nothing */
1009 WARN_ON_ONCE(subreq->wb_bytes != bytes_left);
1010 WARN_ON_ONCE(subreq->wb_offset != offset);
1011 WARN_ON_ONCE(subreq->wb_pgbase != pgbase);
1013 nfs_page_group_unlock(req);
1014 desc->pg_moreio = 1;
1015 nfs_pageio_doio(desc);
1016 if (desc->pg_error < 0 || mirror->pg_recoalesce)
1017 goto out_cleanup_subreq;
1018 /* retry add_request for this subreq */
1019 nfs_page_group_lock(req, false);
1023 /* check for buggy pg_test call(s) */
1024 WARN_ON_ONCE(subreq->wb_bytes + subreq->wb_pgbase > PAGE_SIZE);
1025 WARN_ON_ONCE(subreq->wb_bytes > bytes_left);
1026 WARN_ON_ONCE(subreq->wb_bytes == 0);
1028 bytes_left -= subreq->wb_bytes;
1029 offset += subreq->wb_bytes;
1030 pgbase += subreq->wb_bytes;
1033 subreq = nfs_create_request(req->wb_context,
1035 subreq, pgbase, bytes_left);
1038 nfs_lock_request(subreq);
1039 subreq->wb_offset = offset;
1040 subreq->wb_index = req->wb_index;
1042 } while (bytes_left > 0);
1044 nfs_page_group_unlock(req);
1047 desc->pg_error = PTR_ERR(subreq);
1048 nfs_page_group_unlock(req);
1052 nfs_pageio_cleanup_request(desc, subreq);
1056 static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc)
1058 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1062 list_splice_init(&mirror->pg_list, &head);
1063 mirror->pg_count = 0;
1064 mirror->pg_base = 0;
1065 mirror->pg_recoalesce = 0;
1067 while (!list_empty(&head)) {
1068 struct nfs_page *req;
1070 req = list_first_entry(&head, struct nfs_page, wb_list);
1071 if (__nfs_pageio_add_request(desc, req))
1073 if (desc->pg_error < 0) {
1074 list_splice_tail(&head, &mirror->pg_list);
1075 mirror->pg_recoalesce = 1;
1080 } while (mirror->pg_recoalesce);
1084 static int nfs_pageio_add_request_mirror(struct nfs_pageio_descriptor *desc,
1085 struct nfs_page *req)
1090 ret = __nfs_pageio_add_request(desc, req);
1093 if (desc->pg_error < 0)
1095 ret = nfs_do_recoalesce(desc);
1101 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
1102 struct nfs_page *req)
1105 unsigned int pgbase, offset, bytes;
1106 struct nfs_page *dupreq, *lastreq;
1108 pgbase = req->wb_pgbase;
1109 offset = req->wb_offset;
1110 bytes = req->wb_bytes;
1112 nfs_pageio_setup_mirroring(desc, req);
1113 if (desc->pg_error < 0)
1116 for (midx = 0; midx < desc->pg_mirror_count; midx++) {
1118 nfs_page_group_lock(req, false);
1120 /* find the last request */
1121 for (lastreq = req->wb_head;
1122 lastreq->wb_this_page != req->wb_head;
1123 lastreq = lastreq->wb_this_page)
1126 dupreq = nfs_create_request(req->wb_context,
1127 req->wb_page, lastreq, pgbase, bytes);
1129 if (IS_ERR(dupreq)) {
1130 nfs_page_group_unlock(req);
1131 desc->pg_error = PTR_ERR(dupreq);
1135 nfs_lock_request(dupreq);
1136 nfs_page_group_unlock(req);
1137 dupreq->wb_offset = offset;
1138 dupreq->wb_index = req->wb_index;
1142 if (nfs_pgio_has_mirroring(desc))
1143 desc->pg_mirror_idx = midx;
1144 if (!nfs_pageio_add_request_mirror(desc, dupreq))
1145 goto out_cleanup_subreq;
1152 nfs_pageio_cleanup_request(desc, dupreq);
1155 * We might have failed before sending any reqs over wire.
1156 * Clean up rest of the reqs in mirror pg_list.
1158 if (desc->pg_error) {
1159 struct nfs_pgio_mirror *mirror;
1160 void (*func)(struct list_head *);
1162 /* remember fatal errors */
1163 if (nfs_error_is_fatal(desc->pg_error))
1164 mapping_set_error(desc->pg_inode->i_mapping,
1167 func = desc->pg_completion_ops->error_cleanup;
1168 for (midx = 0; midx < desc->pg_mirror_count; midx++) {
1169 mirror = &desc->pg_mirrors[midx];
1170 func(&mirror->pg_list);
1177 * nfs_pageio_complete_mirror - Complete I/O on the current mirror of an
1178 * nfs_pageio_descriptor
1179 * @desc: pointer to io descriptor
1180 * @mirror_idx: pointer to mirror index
1182 static void nfs_pageio_complete_mirror(struct nfs_pageio_descriptor *desc,
1185 struct nfs_pgio_mirror *mirror = &desc->pg_mirrors[mirror_idx];
1186 u32 restore_idx = desc->pg_mirror_idx;
1188 if (nfs_pgio_has_mirroring(desc))
1189 desc->pg_mirror_idx = mirror_idx;
1191 nfs_pageio_doio(desc);
1192 if (desc->pg_error < 0 || !mirror->pg_recoalesce)
1194 if (!nfs_do_recoalesce(desc))
1197 desc->pg_mirror_idx = restore_idx;
1201 * nfs_pageio_resend - Transfer requests to new descriptor and resend
1202 * @hdr - the pgio header to move request from
1203 * @desc - the pageio descriptor to add requests to
1205 * Try to move each request (nfs_page) from @hdr to @desc then attempt
1208 * Returns 0 on success and < 0 on error.
1210 int nfs_pageio_resend(struct nfs_pageio_descriptor *desc,
1211 struct nfs_pgio_header *hdr)
1215 desc->pg_dreq = hdr->dreq;
1216 while (!list_empty(&hdr->pages)) {
1217 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
1219 nfs_list_remove_request(req);
1220 if (!nfs_pageio_add_request(desc, req))
1221 nfs_list_add_request(req, &failed);
1223 nfs_pageio_complete(desc);
1224 if (!list_empty(&failed)) {
1225 list_move(&failed, &hdr->pages);
1226 return desc->pg_error < 0 ? desc->pg_error : -EIO;
1230 EXPORT_SYMBOL_GPL(nfs_pageio_resend);
1233 * nfs_pageio_complete - Complete I/O then cleanup an nfs_pageio_descriptor
1234 * @desc: pointer to io descriptor
1236 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
1240 for (midx = 0; midx < desc->pg_mirror_count; midx++)
1241 nfs_pageio_complete_mirror(desc, midx);
1243 if (desc->pg_ops->pg_cleanup)
1244 desc->pg_ops->pg_cleanup(desc);
1245 nfs_pageio_cleanup_mirroring(desc);
1249 * nfs_pageio_cond_complete - Conditional I/O completion
1250 * @desc: pointer to io descriptor
1251 * @index: page index
1253 * It is important to ensure that processes don't try to take locks
1254 * on non-contiguous ranges of pages as that might deadlock. This
1255 * function should be called before attempting to wait on a locked
1256 * nfs_page. It will complete the I/O if the page index 'index'
1257 * is not contiguous with the existing list of pages in 'desc'.
1259 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
1261 struct nfs_pgio_mirror *mirror;
1262 struct nfs_page *prev;
1265 for (midx = 0; midx < desc->pg_mirror_count; midx++) {
1266 mirror = &desc->pg_mirrors[midx];
1267 if (!list_empty(&mirror->pg_list)) {
1268 prev = nfs_list_entry(mirror->pg_list.prev);
1269 if (index != prev->wb_index + 1) {
1270 nfs_pageio_complete(desc);
1278 * nfs_pageio_stop_mirroring - stop using mirroring (set mirror count to 1)
1280 void nfs_pageio_stop_mirroring(struct nfs_pageio_descriptor *pgio)
1282 nfs_pageio_complete(pgio);
1285 int __init nfs_init_nfspagecache(void)
1287 nfs_page_cachep = kmem_cache_create("nfs_page",
1288 sizeof(struct nfs_page),
1289 0, SLAB_HWCACHE_ALIGN,
1291 if (nfs_page_cachep == NULL)
1297 void nfs_destroy_nfspagecache(void)
1299 kmem_cache_destroy(nfs_page_cachep);
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,
1308 const struct nfs_pageio_ops nfs_pgio_rw_ops = {
1309 .pg_test = nfs_generic_pg_test,
1310 .pg_doio = nfs_generic_pg_pgios,