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/nfs3.h>
17 #include <linux/nfs4.h>
18 #include <linux/nfs_page.h>
19 #include <linux/nfs_fs.h>
20 #include <linux/nfs_mount.h>
25 static struct kmem_cache *nfs_page_cachep;
27 static inline struct nfs_page *
30 struct nfs_page *p = kmem_cache_zalloc(nfs_page_cachep, GFP_KERNEL);
32 INIT_LIST_HEAD(&p->wb_list);
37 nfs_page_free(struct nfs_page *p)
39 kmem_cache_free(nfs_page_cachep, p);
43 * nfs_create_request - Create an NFS read/write request.
44 * @file: file descriptor to use
45 * @inode: inode to which the request is attached
46 * @page: page to write
47 * @offset: starting offset within the page for the write
48 * @count: number of bytes to read/write
50 * The page must be locked by the caller. This makes sure we never
51 * create two different requests for the same page.
52 * User should ensure it is safe to sleep in this function.
55 nfs_create_request(struct nfs_open_context *ctx, struct inode *inode,
57 unsigned int offset, unsigned int count)
61 /* try to allocate the request struct */
62 req = nfs_page_alloc();
64 return ERR_PTR(-ENOMEM);
66 /* get lock context early so we can deal with alloc failures */
67 req->wb_lock_context = nfs_get_lock_context(ctx);
68 if (req->wb_lock_context == NULL) {
70 return ERR_PTR(-ENOMEM);
73 /* Initialize the request struct. Initially, we assume a
74 * long write-back delay. This will be adjusted in
75 * update_nfs_request below if the region is not locked. */
77 atomic_set(&req->wb_complete, 0);
78 req->wb_index = page->index;
80 BUG_ON(PagePrivate(page));
81 BUG_ON(!PageLocked(page));
82 BUG_ON(page->mapping->host != inode);
83 req->wb_offset = offset;
84 req->wb_pgbase = offset;
85 req->wb_bytes = count;
86 req->wb_context = get_nfs_open_context(ctx);
87 kref_init(&req->wb_kref);
92 * nfs_unlock_request - Unlock request and wake up sleepers.
95 void nfs_unlock_request(struct nfs_page *req)
97 if (!NFS_WBACK_BUSY(req)) {
98 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
101 smp_mb__before_clear_bit();
102 clear_bit(PG_BUSY, &req->wb_flags);
103 smp_mb__after_clear_bit();
104 wake_up_bit(&req->wb_flags, PG_BUSY);
105 nfs_release_request(req);
109 * nfs_set_page_tag_locked - Tag a request as locked
112 int nfs_set_page_tag_locked(struct nfs_page *req)
114 if (!nfs_lock_request_dontget(req))
116 if (test_bit(PG_MAPPED, &req->wb_flags))
117 radix_tree_tag_set(&NFS_I(req->wb_context->dentry->d_inode)->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
122 * nfs_clear_page_tag_locked - Clear request tag and wake up sleepers
124 void nfs_clear_page_tag_locked(struct nfs_page *req)
126 if (test_bit(PG_MAPPED, &req->wb_flags)) {
127 struct inode *inode = req->wb_context->dentry->d_inode;
128 struct nfs_inode *nfsi = NFS_I(inode);
130 spin_lock(&inode->i_lock);
131 radix_tree_tag_clear(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
132 nfs_unlock_request(req);
133 spin_unlock(&inode->i_lock);
135 nfs_unlock_request(req);
139 * nfs_clear_request - Free up all resources allocated to the request
142 * Release page and open context resources associated with a read/write
143 * request after it has completed.
145 static void nfs_clear_request(struct nfs_page *req)
147 struct page *page = req->wb_page;
148 struct nfs_open_context *ctx = req->wb_context;
149 struct nfs_lock_context *l_ctx = req->wb_lock_context;
152 page_cache_release(page);
156 nfs_put_lock_context(l_ctx);
157 req->wb_lock_context = NULL;
160 put_nfs_open_context(ctx);
161 req->wb_context = NULL;
167 * nfs_release_request - Release the count on an NFS read/write request
168 * @req: request to release
170 * Note: Should never be called with the spinlock held!
172 static void nfs_free_request(struct kref *kref)
174 struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
176 /* Release struct file and open context */
177 nfs_clear_request(req);
181 void nfs_release_request(struct nfs_page *req)
183 kref_put(&req->wb_kref, nfs_free_request);
186 static int nfs_wait_bit_uninterruptible(void *word)
193 * nfs_wait_on_request - Wait for a request to complete.
194 * @req: request to wait upon.
196 * Interruptible by fatal signals only.
197 * The user is responsible for holding a count on the request.
200 nfs_wait_on_request(struct nfs_page *req)
202 return wait_on_bit(&req->wb_flags, PG_BUSY,
203 nfs_wait_bit_uninterruptible,
204 TASK_UNINTERRUPTIBLE);
207 bool nfs_generic_pg_test(struct nfs_pageio_descriptor *desc, struct nfs_page *prev, struct nfs_page *req)
210 * FIXME: ideally we should be able to coalesce all requests
211 * that are not block boundary aligned, but currently this
212 * is problematic for the case of bsize < PAGE_CACHE_SIZE,
213 * since nfs_flush_multi and nfs_pagein_multi assume you
214 * can have only one struct nfs_page.
216 if (desc->pg_bsize < PAGE_SIZE)
219 return desc->pg_count + req->wb_bytes <= desc->pg_bsize;
221 EXPORT_SYMBOL_GPL(nfs_generic_pg_test);
224 * nfs_pageio_init - initialise a page io descriptor
225 * @desc: pointer to descriptor
226 * @inode: pointer to inode
227 * @doio: pointer to io function
228 * @bsize: io block size
229 * @io_flags: extra parameters for the io function
231 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
233 const struct nfs_pageio_ops *pg_ops,
237 INIT_LIST_HEAD(&desc->pg_list);
238 desc->pg_bytes_written = 0;
240 desc->pg_bsize = bsize;
243 desc->pg_recoalesce = 0;
244 desc->pg_inode = inode;
245 desc->pg_ops = pg_ops;
246 desc->pg_ioflags = io_flags;
248 desc->pg_lseg = NULL;
252 * nfs_can_coalesce_requests - test two requests for compatibility
253 * @prev: pointer to nfs_page
254 * @req: pointer to nfs_page
256 * The nfs_page structures 'prev' and 'req' are compared to ensure that the
257 * page data area they describe is contiguous, and that their RPC
258 * credentials, NFSv4 open state, and lockowners are the same.
260 * Return 'true' if this is the case, else return 'false'.
262 static bool nfs_can_coalesce_requests(struct nfs_page *prev,
263 struct nfs_page *req,
264 struct nfs_pageio_descriptor *pgio)
266 if (req->wb_context->cred != prev->wb_context->cred)
268 if (req->wb_lock_context->lockowner != prev->wb_lock_context->lockowner)
270 if (req->wb_context->state != prev->wb_context->state)
272 if (req->wb_index != (prev->wb_index + 1))
274 if (req->wb_pgbase != 0)
276 if (prev->wb_pgbase + prev->wb_bytes != PAGE_CACHE_SIZE)
278 return pgio->pg_ops->pg_test(pgio, prev, req);
282 * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
283 * @desc: destination io descriptor
286 * Returns true if the request 'req' was successfully coalesced into the
287 * existing list of pages 'desc'.
289 static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
290 struct nfs_page *req)
292 if (desc->pg_count != 0) {
293 struct nfs_page *prev;
295 prev = nfs_list_entry(desc->pg_list.prev);
296 if (!nfs_can_coalesce_requests(prev, req, desc))
299 if (desc->pg_ops->pg_init)
300 desc->pg_ops->pg_init(desc, req);
301 desc->pg_base = req->wb_pgbase;
303 nfs_list_remove_request(req);
304 nfs_list_add_request(req, &desc->pg_list);
305 desc->pg_count += req->wb_bytes;
310 * Helper for nfs_pageio_add_request and nfs_pageio_complete
312 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
314 if (!list_empty(&desc->pg_list)) {
315 int error = desc->pg_ops->pg_doio(desc);
317 desc->pg_error = error;
319 desc->pg_bytes_written += desc->pg_count;
321 if (list_empty(&desc->pg_list)) {
328 * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
329 * @desc: destination io descriptor
332 * Returns true if the request 'req' was successfully coalesced into the
333 * existing list of pages 'desc'.
335 static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
336 struct nfs_page *req)
338 while (!nfs_pageio_do_add_request(desc, req)) {
340 nfs_pageio_doio(desc);
341 if (desc->pg_error < 0)
344 if (desc->pg_recoalesce)
350 static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc)
355 list_splice_init(&desc->pg_list, &head);
356 desc->pg_bytes_written -= desc->pg_count;
359 desc->pg_recoalesce = 0;
361 while (!list_empty(&head)) {
362 struct nfs_page *req;
364 req = list_first_entry(&head, struct nfs_page, wb_list);
365 nfs_list_remove_request(req);
366 if (__nfs_pageio_add_request(desc, req))
368 if (desc->pg_error < 0)
372 } while (desc->pg_recoalesce);
376 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
377 struct nfs_page *req)
382 ret = __nfs_pageio_add_request(desc, req);
385 if (desc->pg_error < 0)
387 ret = nfs_do_recoalesce(desc);
393 * nfs_pageio_complete - Complete I/O on an nfs_pageio_descriptor
394 * @desc: pointer to io descriptor
396 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
399 nfs_pageio_doio(desc);
400 if (!desc->pg_recoalesce)
402 if (!nfs_do_recoalesce(desc))
408 * nfs_pageio_cond_complete - Conditional I/O completion
409 * @desc: pointer to io descriptor
412 * It is important to ensure that processes don't try to take locks
413 * on non-contiguous ranges of pages as that might deadlock. This
414 * function should be called before attempting to wait on a locked
415 * nfs_page. It will complete the I/O if the page index 'index'
416 * is not contiguous with the existing list of pages in 'desc'.
418 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
420 if (!list_empty(&desc->pg_list)) {
421 struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
422 if (index != prev->wb_index + 1)
423 nfs_pageio_complete(desc);
427 #define NFS_SCAN_MAXENTRIES 16
429 * nfs_scan_list - Scan a list for matching requests
431 * @dst: Destination list
432 * @idx_start: lower bound of page->index to scan
433 * @npages: idx_start + npages sets the upper bound to scan.
434 * @tag: tag to scan for
436 * Moves elements from one of the inode request lists.
437 * If the number of requests is set to 0, the entire address_space
438 * starting at index idx_start, is scanned.
439 * The requests are *not* checked to ensure that they form a contiguous set.
440 * You must be holding the inode's i_lock when calling this function
442 int nfs_scan_list(struct nfs_inode *nfsi,
443 struct list_head *dst, pgoff_t idx_start,
444 unsigned int npages, int tag)
446 struct nfs_page *pgvec[NFS_SCAN_MAXENTRIES];
447 struct nfs_page *req;
451 struct list_head *list;
457 idx_end = idx_start + npages - 1;
460 found = radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree,
461 (void **)&pgvec[0], idx_start,
462 NFS_SCAN_MAXENTRIES, tag);
465 for (i = 0; i < found; i++) {
467 if (req->wb_index > idx_end)
469 idx_start = req->wb_index + 1;
470 if (nfs_set_page_tag_locked(req)) {
471 kref_get(&req->wb_kref);
472 radix_tree_tag_clear(&nfsi->nfs_page_tree,
474 list = pnfs_choose_commit_list(req, dst);
475 nfs_list_add_request(req, list);
481 /* for latency reduction */
482 cond_resched_lock(&nfsi->vfs_inode.i_lock);
488 int __init nfs_init_nfspagecache(void)
490 nfs_page_cachep = kmem_cache_create("nfs_page",
491 sizeof(struct nfs_page),
492 0, SLAB_HWCACHE_ALIGN,
494 if (nfs_page_cachep == NULL)
500 void nfs_destroy_nfspagecache(void)
502 kmem_cache_destroy(nfs_page_cachep);