Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/shaggy...
[linux-2.6] / 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/sunrpc/clnt.h>
15 #include <linux/nfs3.h>
16 #include <linux/nfs4.h>
17 #include <linux/nfs_page.h>
18 #include <linux/nfs_fs.h>
19 #include <linux/nfs_mount.h>
20 #include <linux/writeback.h>
21
22 #define NFS_PARANOIA 1
23
24 static struct kmem_cache *nfs_page_cachep;
25
26 static inline struct nfs_page *
27 nfs_page_alloc(void)
28 {
29         struct nfs_page *p;
30         p = kmem_cache_alloc(nfs_page_cachep, GFP_KERNEL);
31         if (p) {
32                 memset(p, 0, sizeof(*p));
33                 INIT_LIST_HEAD(&p->wb_list);
34         }
35         return p;
36 }
37
38 static inline void
39 nfs_page_free(struct nfs_page *p)
40 {
41         kmem_cache_free(nfs_page_cachep, p);
42 }
43
44 /**
45  * nfs_create_request - Create an NFS read/write request.
46  * @file: file descriptor to use
47  * @inode: inode to which the request is attached
48  * @page: page to write
49  * @offset: starting offset within the page for the write
50  * @count: number of bytes to read/write
51  *
52  * The page must be locked by the caller. This makes sure we never
53  * create two different requests for the same page, and avoids
54  * a possible deadlock when we reach the hard limit on the number
55  * of dirty pages.
56  * User should ensure it is safe to sleep in this function.
57  */
58 struct nfs_page *
59 nfs_create_request(struct nfs_open_context *ctx, struct inode *inode,
60                    struct page *page,
61                    unsigned int offset, unsigned int count)
62 {
63         struct nfs_server *server = NFS_SERVER(inode);
64         struct nfs_page         *req;
65
66         /* Deal with hard limits.  */
67         for (;;) {
68                 /* try to allocate the request struct */
69                 req = nfs_page_alloc();
70                 if (req != NULL)
71                         break;
72
73                 /* Try to free up at least one request in order to stay
74                  * below the hard limit
75                  */
76                 if (signalled() && (server->flags & NFS_MOUNT_INTR))
77                         return ERR_PTR(-ERESTARTSYS);
78                 yield();
79         }
80
81         /* Initialize the request struct. Initially, we assume a
82          * long write-back delay. This will be adjusted in
83          * update_nfs_request below if the region is not locked. */
84         req->wb_page    = page;
85         atomic_set(&req->wb_complete, 0);
86         req->wb_index   = page->index;
87         page_cache_get(page);
88         BUG_ON(PagePrivate(page));
89         BUG_ON(!PageLocked(page));
90         BUG_ON(page->mapping->host != inode);
91         req->wb_offset  = offset;
92         req->wb_pgbase  = offset;
93         req->wb_bytes   = count;
94         atomic_set(&req->wb_count, 1);
95         req->wb_context = get_nfs_open_context(ctx);
96
97         return req;
98 }
99
100 /**
101  * nfs_unlock_request - Unlock request and wake up sleepers.
102  * @req:
103  */
104 void nfs_unlock_request(struct nfs_page *req)
105 {
106         if (!NFS_WBACK_BUSY(req)) {
107                 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
108                 BUG();
109         }
110         smp_mb__before_clear_bit();
111         clear_bit(PG_BUSY, &req->wb_flags);
112         smp_mb__after_clear_bit();
113         wake_up_bit(&req->wb_flags, PG_BUSY);
114         nfs_release_request(req);
115 }
116
117 /**
118  * nfs_set_page_writeback_locked - Lock a request for writeback
119  * @req:
120  */
121 int nfs_set_page_writeback_locked(struct nfs_page *req)
122 {
123         struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
124
125         if (!nfs_lock_request(req))
126                 return 0;
127         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_WRITEBACK);
128         return 1;
129 }
130
131 /**
132  * nfs_clear_page_writeback - Unlock request and wake up sleepers
133  */
134 void nfs_clear_page_writeback(struct nfs_page *req)
135 {
136         struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
137
138         if (req->wb_page != NULL) {
139                 spin_lock(&nfsi->req_lock);
140                 radix_tree_tag_clear(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_WRITEBACK);
141                 spin_unlock(&nfsi->req_lock);
142         }
143         nfs_unlock_request(req);
144 }
145
146 /**
147  * nfs_clear_request - Free up all resources allocated to the request
148  * @req:
149  *
150  * Release page resources associated with a write request after it
151  * has completed.
152  */
153 void nfs_clear_request(struct nfs_page *req)
154 {
155         struct page *page = req->wb_page;
156         if (page != NULL) {
157                 page_cache_release(page);
158                 req->wb_page = NULL;
159         }
160 }
161
162
163 /**
164  * nfs_release_request - Release the count on an NFS read/write request
165  * @req: request to release
166  *
167  * Note: Should never be called with the spinlock held!
168  */
169 void
170 nfs_release_request(struct nfs_page *req)
171 {
172         if (!atomic_dec_and_test(&req->wb_count))
173                 return;
174
175 #ifdef NFS_PARANOIA
176         BUG_ON (!list_empty(&req->wb_list));
177         BUG_ON (NFS_WBACK_BUSY(req));
178 #endif
179
180         /* Release struct file or cached credential */
181         nfs_clear_request(req);
182         put_nfs_open_context(req->wb_context);
183         nfs_page_free(req);
184 }
185
186 static int nfs_wait_bit_interruptible(void *word)
187 {
188         int ret = 0;
189
190         if (signal_pending(current))
191                 ret = -ERESTARTSYS;
192         else
193                 schedule();
194         return ret;
195 }
196
197 /**
198  * nfs_wait_on_request - Wait for a request to complete.
199  * @req: request to wait upon.
200  *
201  * Interruptible by signals only if mounted with intr flag.
202  * The user is responsible for holding a count on the request.
203  */
204 int
205 nfs_wait_on_request(struct nfs_page *req)
206 {
207         struct rpc_clnt *clnt = NFS_CLIENT(req->wb_context->dentry->d_inode);
208         sigset_t oldmask;
209         int ret = 0;
210
211         if (!test_bit(PG_BUSY, &req->wb_flags))
212                 goto out;
213         /*
214          * Note: the call to rpc_clnt_sigmask() suffices to ensure that we
215          *       are not interrupted if intr flag is not set
216          */
217         rpc_clnt_sigmask(clnt, &oldmask);
218         ret = out_of_line_wait_on_bit(&req->wb_flags, PG_BUSY,
219                         nfs_wait_bit_interruptible, TASK_INTERRUPTIBLE);
220         rpc_clnt_sigunmask(clnt, &oldmask);
221 out:
222         return ret;
223 }
224
225 /**
226  * nfs_coalesce_requests - Split coalesced requests out from a list.
227  * @head: source list
228  * @dst: destination list
229  * @nmax: maximum number of requests to coalesce
230  *
231  * Moves a maximum of 'nmax' elements from one list to another.
232  * The elements are checked to ensure that they form a contiguous set
233  * of pages, and that the RPC credentials are the same.
234  */
235 int
236 nfs_coalesce_requests(struct list_head *head, struct list_head *dst,
237                       unsigned int nmax)
238 {
239         struct nfs_page         *req = NULL;
240         unsigned int            npages = 0;
241
242         while (!list_empty(head)) {
243                 struct nfs_page *prev = req;
244
245                 req = nfs_list_entry(head->next);
246                 if (prev) {
247                         if (req->wb_context->cred != prev->wb_context->cred)
248                                 break;
249                         if (req->wb_context->lockowner != prev->wb_context->lockowner)
250                                 break;
251                         if (req->wb_context->state != prev->wb_context->state)
252                                 break;
253                         if (req->wb_index != (prev->wb_index + 1))
254                                 break;
255
256                         if (req->wb_pgbase != 0)
257                                 break;
258                 }
259                 nfs_list_remove_request(req);
260                 nfs_list_add_request(req, dst);
261                 npages++;
262                 if (req->wb_pgbase + req->wb_bytes != PAGE_CACHE_SIZE)
263                         break;
264                 if (npages >= nmax)
265                         break;
266         }
267         return npages;
268 }
269
270 #define NFS_SCAN_MAXENTRIES 16
271 /**
272  * nfs_scan_dirty - Scan the radix tree for dirty requests
273  * @mapping: pointer to address space
274  * @wbc: writeback_control structure
275  * @dst: Destination list
276  *
277  * Moves elements from one of the inode request lists.
278  * If the number of requests is set to 0, the entire address_space
279  * starting at index idx_start, is scanned.
280  * The requests are *not* checked to ensure that they form a contiguous set.
281  * You must be holding the inode's req_lock when calling this function
282  */
283 long nfs_scan_dirty(struct address_space *mapping,
284                         struct writeback_control *wbc,
285                         struct list_head *dst)
286 {
287         struct nfs_inode *nfsi = NFS_I(mapping->host);
288         struct nfs_page *pgvec[NFS_SCAN_MAXENTRIES];
289         struct nfs_page *req;
290         pgoff_t idx_start, idx_end;
291         long res = 0;
292         int found, i;
293
294         if (nfsi->ndirty == 0)
295                 return 0;
296         if (wbc->range_cyclic) {
297                 idx_start = 0;
298                 idx_end = ULONG_MAX;
299         } else if (wbc->range_end == 0) {
300                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
301                 idx_end = ULONG_MAX;
302         } else {
303                 idx_start = wbc->range_start >> PAGE_CACHE_SHIFT;
304                 idx_end = wbc->range_end >> PAGE_CACHE_SHIFT;
305         }
306
307         for (;;) {
308                 unsigned int toscan = NFS_SCAN_MAXENTRIES;
309
310                 found = radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree,
311                                 (void **)&pgvec[0], idx_start, toscan,
312                                 NFS_PAGE_TAG_DIRTY);
313
314                 /* Did we make progress? */
315                 if (found <= 0)
316                         break;
317
318                 for (i = 0; i < found; i++) {
319                         req = pgvec[i];
320                         if (!wbc->range_cyclic && req->wb_index > idx_end)
321                                 goto out;
322
323                         /* Try to lock request and mark it for writeback */
324                         if (!nfs_set_page_writeback_locked(req))
325                                 goto next;
326                         radix_tree_tag_clear(&nfsi->nfs_page_tree,
327                                         req->wb_index, NFS_PAGE_TAG_DIRTY);
328                         nfsi->ndirty--;
329                         nfs_list_remove_request(req);
330                         nfs_list_add_request(req, dst);
331                         res++;
332                         if (res == LONG_MAX)
333                                 goto out;
334 next:
335                         idx_start = req->wb_index + 1;
336                 }
337         }
338 out:
339         WARN_ON ((nfsi->ndirty == 0) != list_empty(&nfsi->dirty));
340         return res;
341 }
342
343 /**
344  * nfs_scan_list - Scan a list for matching requests
345  * @nfsi: NFS inode
346  * @head: One of the NFS inode request lists
347  * @dst: Destination list
348  * @idx_start: lower bound of page->index to scan
349  * @npages: idx_start + npages sets the upper bound to scan.
350  *
351  * Moves elements from one of the inode request lists.
352  * If the number of requests is set to 0, the entire address_space
353  * starting at index idx_start, is scanned.
354  * The requests are *not* checked to ensure that they form a contiguous set.
355  * You must be holding the inode's req_lock when calling this function
356  */
357 int nfs_scan_list(struct nfs_inode *nfsi, struct list_head *head,
358                 struct list_head *dst, unsigned long idx_start,
359                 unsigned int npages)
360 {
361         struct nfs_page *pgvec[NFS_SCAN_MAXENTRIES];
362         struct nfs_page *req;
363         unsigned long idx_end;
364         int found, i;
365         int res;
366
367         res = 0;
368         if (npages == 0)
369                 idx_end = ~0;
370         else
371                 idx_end = idx_start + npages - 1;
372
373         for (;;) {
374                 found = radix_tree_gang_lookup(&nfsi->nfs_page_tree,
375                                 (void **)&pgvec[0], idx_start,
376                                 NFS_SCAN_MAXENTRIES);
377                 if (found <= 0)
378                         break;
379                 for (i = 0; i < found; i++) {
380                         req = pgvec[i];
381                         if (req->wb_index > idx_end)
382                                 goto out;
383                         idx_start = req->wb_index + 1;
384                         if (req->wb_list_head != head)
385                                 continue;
386                         if (nfs_set_page_writeback_locked(req)) {
387                                 nfs_list_remove_request(req);
388                                 nfs_list_add_request(req, dst);
389                                 res++;
390                         }
391                 }
392
393         }
394 out:
395         return res;
396 }
397
398 int __init nfs_init_nfspagecache(void)
399 {
400         nfs_page_cachep = kmem_cache_create("nfs_page",
401                                             sizeof(struct nfs_page),
402                                             0, SLAB_HWCACHE_ALIGN,
403                                             NULL, NULL);
404         if (nfs_page_cachep == NULL)
405                 return -ENOMEM;
406
407         return 0;
408 }
409
410 void nfs_destroy_nfspagecache(void)
411 {
412         kmem_cache_destroy(nfs_page_cachep);
413 }
414