]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/fs/nfs/write.c | |
3 | * | |
4 | * Writing file data over NFS. | |
5 | * | |
6 | * We do it like this: When a (user) process wishes to write data to an | |
7 | * NFS file, a write request is allocated that contains the RPC task data | |
8 | * plus some info on the page to be written, and added to the inode's | |
9 | * write chain. If the process writes past the end of the page, an async | |
10 | * RPC call to write the page is scheduled immediately; otherwise, the call | |
11 | * is delayed for a few seconds. | |
12 | * | |
13 | * Just like readahead, no async I/O is performed if wsize < PAGE_SIZE. | |
14 | * | |
15 | * Write requests are kept on the inode's writeback list. Each entry in | |
16 | * that list references the page (portion) to be written. When the | |
17 | * cache timeout has expired, the RPC task is woken up, and tries to | |
18 | * lock the page. As soon as it manages to do so, the request is moved | |
19 | * from the writeback list to the writelock list. | |
20 | * | |
21 | * Note: we must make sure never to confuse the inode passed in the | |
22 | * write_page request with the one in page->inode. As far as I understand | |
23 | * it, these are different when doing a swap-out. | |
24 | * | |
25 | * To understand everything that goes on here and in the NFS read code, | |
26 | * one should be aware that a page is locked in exactly one of the following | |
27 | * cases: | |
28 | * | |
29 | * - A write request is in progress. | |
30 | * - A user process is in generic_file_write/nfs_update_page | |
31 | * - A user process is in generic_file_read | |
32 | * | |
33 | * Also note that because of the way pages are invalidated in | |
34 | * nfs_revalidate_inode, the following assertions hold: | |
35 | * | |
36 | * - If a page is dirty, there will be no read requests (a page will | |
37 | * not be re-read unless invalidated by nfs_revalidate_inode). | |
38 | * - If the page is not uptodate, there will be no pending write | |
39 | * requests, and no process will be in nfs_update_page. | |
40 | * | |
41 | * FIXME: Interaction with the vmscan routines is not optimal yet. | |
42 | * Either vmscan must be made nfs-savvy, or we need a different page | |
43 | * reclaim concept that supports something like FS-independent | |
44 | * buffer_heads with a b_ops-> field. | |
45 | * | |
46 | * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de> | |
47 | */ | |
48 | ||
49 | #include <linux/config.h> | |
50 | #include <linux/types.h> | |
51 | #include <linux/slab.h> | |
52 | #include <linux/mm.h> | |
53 | #include <linux/pagemap.h> | |
54 | #include <linux/file.h> | |
55 | #include <linux/mpage.h> | |
56 | #include <linux/writeback.h> | |
57 | ||
58 | #include <linux/sunrpc/clnt.h> | |
59 | #include <linux/nfs_fs.h> | |
60 | #include <linux/nfs_mount.h> | |
61 | #include <linux/nfs_page.h> | |
62 | #include <asm/uaccess.h> | |
63 | #include <linux/smp_lock.h> | |
64 | ||
65 | #include "delegation.h" | |
91d5b470 | 66 | #include "iostat.h" |
1da177e4 LT |
67 | |
68 | #define NFSDBG_FACILITY NFSDBG_PAGECACHE | |
69 | ||
70 | #define MIN_POOL_WRITE (32) | |
71 | #define MIN_POOL_COMMIT (4) | |
72 | ||
73 | /* | |
74 | * Local function declarations | |
75 | */ | |
76 | static struct nfs_page * nfs_update_request(struct nfs_open_context*, | |
77 | struct inode *, | |
78 | struct page *, | |
79 | unsigned int, unsigned int); | |
1da177e4 LT |
80 | static int nfs_wait_on_write_congestion(struct address_space *, int); |
81 | static int nfs_wait_on_requests(struct inode *, unsigned long, unsigned int); | |
82 | static int nfs_flush_inode(struct inode *inode, unsigned long idx_start, | |
83 | unsigned int npages, int how); | |
788e7a89 TM |
84 | static const struct rpc_call_ops nfs_write_partial_ops; |
85 | static const struct rpc_call_ops nfs_write_full_ops; | |
86 | static const struct rpc_call_ops nfs_commit_ops; | |
1da177e4 LT |
87 | |
88 | static kmem_cache_t *nfs_wdata_cachep; | |
89 | mempool_t *nfs_wdata_mempool; | |
90 | static mempool_t *nfs_commit_mempool; | |
91 | ||
92 | static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion); | |
93 | ||
e17b1fc4 | 94 | struct nfs_write_data *nfs_commit_alloc(unsigned int pagecount) |
1da177e4 LT |
95 | { |
96 | struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, SLAB_NOFS); | |
40859d7e | 97 | |
1da177e4 LT |
98 | if (p) { |
99 | memset(p, 0, sizeof(*p)); | |
100 | INIT_LIST_HEAD(&p->pages); | |
40859d7e CL |
101 | if (pagecount < NFS_PAGEVEC_SIZE) |
102 | p->pagevec = &p->page_array[0]; | |
103 | else { | |
104 | size_t size = ++pagecount * sizeof(struct page *); | |
bd647545 ES |
105 | p->pagevec = kzalloc(size, GFP_NOFS); |
106 | if (!p->pagevec) { | |
40859d7e CL |
107 | mempool_free(p, nfs_commit_mempool); |
108 | p = NULL; | |
109 | } | |
110 | } | |
1da177e4 LT |
111 | } |
112 | return p; | |
113 | } | |
114 | ||
e17b1fc4 | 115 | void nfs_commit_free(struct nfs_write_data *p) |
1da177e4 | 116 | { |
40859d7e CL |
117 | if (p && (p->pagevec != &p->page_array[0])) |
118 | kfree(p->pagevec); | |
1da177e4 LT |
119 | mempool_free(p, nfs_commit_mempool); |
120 | } | |
121 | ||
963d8fe5 | 122 | void nfs_writedata_release(void *wdata) |
1da177e4 | 123 | { |
1da177e4 LT |
124 | nfs_writedata_free(wdata); |
125 | } | |
126 | ||
127 | /* Adjust the file length if we're writing beyond the end */ | |
128 | static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count) | |
129 | { | |
130 | struct inode *inode = page->mapping->host; | |
131 | loff_t end, i_size = i_size_read(inode); | |
132 | unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT; | |
133 | ||
134 | if (i_size > 0 && page->index < end_index) | |
135 | return; | |
136 | end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count); | |
137 | if (i_size >= end) | |
138 | return; | |
91d5b470 | 139 | nfs_inc_stats(inode, NFSIOS_EXTENDWRITE); |
1da177e4 LT |
140 | i_size_write(inode, end); |
141 | } | |
142 | ||
143 | /* We can set the PG_uptodate flag if we see that a write request | |
144 | * covers the full page. | |
145 | */ | |
146 | static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count) | |
147 | { | |
148 | loff_t end_offs; | |
149 | ||
150 | if (PageUptodate(page)) | |
151 | return; | |
152 | if (base != 0) | |
153 | return; | |
154 | if (count == PAGE_CACHE_SIZE) { | |
155 | SetPageUptodate(page); | |
156 | return; | |
157 | } | |
158 | ||
159 | end_offs = i_size_read(page->mapping->host) - 1; | |
160 | if (end_offs < 0) | |
161 | return; | |
162 | /* Is this the last page? */ | |
163 | if (page->index != (unsigned long)(end_offs >> PAGE_CACHE_SHIFT)) | |
164 | return; | |
165 | /* This is the last page: set PG_uptodate if we cover the entire | |
166 | * extent of the data, then zero the rest of the page. | |
167 | */ | |
168 | if (count == (unsigned int)(end_offs & (PAGE_CACHE_SIZE - 1)) + 1) { | |
169 | memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count); | |
170 | SetPageUptodate(page); | |
171 | } | |
172 | } | |
173 | ||
174 | /* | |
175 | * Write a page synchronously. | |
176 | * Offset is the data offset within the page. | |
177 | */ | |
178 | static int nfs_writepage_sync(struct nfs_open_context *ctx, struct inode *inode, | |
179 | struct page *page, unsigned int offset, unsigned int count, | |
180 | int how) | |
181 | { | |
182 | unsigned int wsize = NFS_SERVER(inode)->wsize; | |
183 | int result, written = 0; | |
184 | struct nfs_write_data *wdata; | |
185 | ||
40859d7e | 186 | wdata = nfs_writedata_alloc(1); |
1da177e4 LT |
187 | if (!wdata) |
188 | return -ENOMEM; | |
189 | ||
190 | wdata->flags = how; | |
191 | wdata->cred = ctx->cred; | |
192 | wdata->inode = inode; | |
193 | wdata->args.fh = NFS_FH(inode); | |
194 | wdata->args.context = ctx; | |
195 | wdata->args.pages = &page; | |
196 | wdata->args.stable = NFS_FILE_SYNC; | |
197 | wdata->args.pgbase = offset; | |
198 | wdata->args.count = wsize; | |
199 | wdata->res.fattr = &wdata->fattr; | |
200 | wdata->res.verf = &wdata->verf; | |
201 | ||
202 | dprintk("NFS: nfs_writepage_sync(%s/%Ld %d@%Ld)\n", | |
203 | inode->i_sb->s_id, | |
204 | (long long)NFS_FILEID(inode), | |
205 | count, (long long)(page_offset(page) + offset)); | |
206 | ||
bb713d6d | 207 | set_page_writeback(page); |
1da177e4 LT |
208 | nfs_begin_data_update(inode); |
209 | do { | |
210 | if (count < wsize) | |
211 | wdata->args.count = count; | |
212 | wdata->args.offset = page_offset(page) + wdata->args.pgbase; | |
213 | ||
214 | result = NFS_PROTO(inode)->write(wdata); | |
215 | ||
216 | if (result < 0) { | |
217 | /* Must mark the page invalid after I/O error */ | |
218 | ClearPageUptodate(page); | |
219 | goto io_error; | |
220 | } | |
221 | if (result < wdata->args.count) | |
222 | printk(KERN_WARNING "NFS: short write, count=%u, result=%d\n", | |
223 | wdata->args.count, result); | |
224 | ||
225 | wdata->args.offset += result; | |
226 | wdata->args.pgbase += result; | |
227 | written += result; | |
228 | count -= result; | |
91d5b470 | 229 | nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, result); |
1da177e4 LT |
230 | } while (count); |
231 | /* Update file length */ | |
232 | nfs_grow_file(page, offset, written); | |
233 | /* Set the PG_uptodate flag? */ | |
234 | nfs_mark_uptodate(page, offset, written); | |
235 | ||
236 | if (PageError(page)) | |
237 | ClearPageError(page); | |
238 | ||
239 | io_error: | |
951a143b | 240 | nfs_end_data_update(inode); |
bb713d6d | 241 | end_page_writeback(page); |
1da177e4 LT |
242 | nfs_writedata_free(wdata); |
243 | return written ? written : result; | |
244 | } | |
245 | ||
246 | static int nfs_writepage_async(struct nfs_open_context *ctx, | |
247 | struct inode *inode, struct page *page, | |
248 | unsigned int offset, unsigned int count) | |
249 | { | |
250 | struct nfs_page *req; | |
1da177e4 LT |
251 | |
252 | req = nfs_update_request(ctx, inode, page, offset, count); | |
abd3e641 TM |
253 | if (IS_ERR(req)) |
254 | return PTR_ERR(req); | |
1da177e4 LT |
255 | /* Update file length */ |
256 | nfs_grow_file(page, offset, count); | |
257 | /* Set the PG_uptodate flag? */ | |
258 | nfs_mark_uptodate(page, offset, count); | |
259 | nfs_unlock_request(req); | |
abd3e641 | 260 | return 0; |
1da177e4 LT |
261 | } |
262 | ||
263 | static int wb_priority(struct writeback_control *wbc) | |
264 | { | |
265 | if (wbc->for_reclaim) | |
266 | return FLUSH_HIGHPRI; | |
267 | if (wbc->for_kupdate) | |
268 | return FLUSH_LOWPRI; | |
269 | return 0; | |
270 | } | |
271 | ||
272 | /* | |
273 | * Write an mmapped page to the server. | |
274 | */ | |
275 | int nfs_writepage(struct page *page, struct writeback_control *wbc) | |
276 | { | |
277 | struct nfs_open_context *ctx; | |
278 | struct inode *inode = page->mapping->host; | |
279 | unsigned long end_index; | |
280 | unsigned offset = PAGE_CACHE_SIZE; | |
281 | loff_t i_size = i_size_read(inode); | |
282 | int inode_referenced = 0; | |
283 | int priority = wb_priority(wbc); | |
284 | int err; | |
285 | ||
91d5b470 CL |
286 | nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGE); |
287 | nfs_add_stats(inode, NFSIOS_WRITEPAGES, 1); | |
288 | ||
1da177e4 LT |
289 | /* |
290 | * Note: We need to ensure that we have a reference to the inode | |
291 | * if we are to do asynchronous writes. If not, waiting | |
292 | * in nfs_wait_on_request() may deadlock with clear_inode(). | |
293 | * | |
294 | * If igrab() fails here, then it is in any case safe to | |
295 | * call nfs_wb_page(), since there will be no pending writes. | |
296 | */ | |
297 | if (igrab(inode) != 0) | |
298 | inode_referenced = 1; | |
299 | end_index = i_size >> PAGE_CACHE_SHIFT; | |
300 | ||
301 | /* Ensure we've flushed out any previous writes */ | |
302 | nfs_wb_page_priority(inode, page, priority); | |
303 | ||
304 | /* easy case */ | |
305 | if (page->index < end_index) | |
306 | goto do_it; | |
307 | /* things got complicated... */ | |
308 | offset = i_size & (PAGE_CACHE_SIZE-1); | |
309 | ||
310 | /* OK, are we completely out? */ | |
311 | err = 0; /* potential race with truncate - ignore */ | |
312 | if (page->index >= end_index+1 || !offset) | |
313 | goto out; | |
314 | do_it: | |
d530838b | 315 | ctx = nfs_find_open_context(inode, NULL, FMODE_WRITE); |
1da177e4 LT |
316 | if (ctx == NULL) { |
317 | err = -EBADF; | |
318 | goto out; | |
319 | } | |
320 | lock_kernel(); | |
321 | if (!IS_SYNC(inode) && inode_referenced) { | |
322 | err = nfs_writepage_async(ctx, inode, page, 0, offset); | |
abd3e641 TM |
323 | if (!wbc->for_writepages) |
324 | nfs_flush_inode(inode, 0, 0, wb_priority(wbc)); | |
1da177e4 LT |
325 | } else { |
326 | err = nfs_writepage_sync(ctx, inode, page, 0, | |
327 | offset, priority); | |
328 | if (err >= 0) { | |
329 | if (err != offset) | |
330 | redirty_page_for_writepage(wbc, page); | |
331 | err = 0; | |
332 | } | |
333 | } | |
334 | unlock_kernel(); | |
335 | put_nfs_open_context(ctx); | |
336 | out: | |
337 | unlock_page(page); | |
338 | if (inode_referenced) | |
339 | iput(inode); | |
340 | return err; | |
341 | } | |
342 | ||
343 | /* | |
344 | * Note: causes nfs_update_request() to block on the assumption | |
345 | * that the writeback is generated due to memory pressure. | |
346 | */ | |
347 | int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc) | |
348 | { | |
349 | struct backing_dev_info *bdi = mapping->backing_dev_info; | |
350 | struct inode *inode = mapping->host; | |
351 | int err; | |
352 | ||
91d5b470 CL |
353 | nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES); |
354 | ||
1da177e4 LT |
355 | err = generic_writepages(mapping, wbc); |
356 | if (err) | |
357 | return err; | |
358 | while (test_and_set_bit(BDI_write_congested, &bdi->state) != 0) { | |
359 | if (wbc->nonblocking) | |
360 | return 0; | |
361 | nfs_wait_on_write_congestion(mapping, 0); | |
362 | } | |
363 | err = nfs_flush_inode(inode, 0, 0, wb_priority(wbc)); | |
364 | if (err < 0) | |
365 | goto out; | |
91d5b470 | 366 | nfs_add_stats(inode, NFSIOS_WRITEPAGES, err); |
1da177e4 LT |
367 | wbc->nr_to_write -= err; |
368 | if (!wbc->nonblocking && wbc->sync_mode == WB_SYNC_ALL) { | |
369 | err = nfs_wait_on_requests(inode, 0, 0); | |
370 | if (err < 0) | |
371 | goto out; | |
372 | } | |
3da28eb1 | 373 | err = nfs_commit_inode(inode, wb_priority(wbc)); |
1da177e4 LT |
374 | if (err > 0) { |
375 | wbc->nr_to_write -= err; | |
376 | err = 0; | |
377 | } | |
378 | out: | |
379 | clear_bit(BDI_write_congested, &bdi->state); | |
380 | wake_up_all(&nfs_write_congestion); | |
381 | return err; | |
382 | } | |
383 | ||
384 | /* | |
385 | * Insert a write request into an inode | |
386 | */ | |
387 | static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req) | |
388 | { | |
389 | struct nfs_inode *nfsi = NFS_I(inode); | |
390 | int error; | |
391 | ||
392 | error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req); | |
393 | BUG_ON(error == -EEXIST); | |
394 | if (error) | |
395 | return error; | |
396 | if (!nfsi->npages) { | |
397 | igrab(inode); | |
398 | nfs_begin_data_update(inode); | |
399 | if (nfs_have_delegation(inode, FMODE_WRITE)) | |
400 | nfsi->change_attr++; | |
401 | } | |
402 | nfsi->npages++; | |
403 | atomic_inc(&req->wb_count); | |
404 | return 0; | |
405 | } | |
406 | ||
407 | /* | |
408 | * Insert a write request into an inode | |
409 | */ | |
410 | static void nfs_inode_remove_request(struct nfs_page *req) | |
411 | { | |
412 | struct inode *inode = req->wb_context->dentry->d_inode; | |
413 | struct nfs_inode *nfsi = NFS_I(inode); | |
414 | ||
415 | BUG_ON (!NFS_WBACK_BUSY(req)); | |
416 | ||
417 | spin_lock(&nfsi->req_lock); | |
418 | radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index); | |
419 | nfsi->npages--; | |
420 | if (!nfsi->npages) { | |
421 | spin_unlock(&nfsi->req_lock); | |
951a143b | 422 | nfs_end_data_update(inode); |
1da177e4 LT |
423 | iput(inode); |
424 | } else | |
425 | spin_unlock(&nfsi->req_lock); | |
426 | nfs_clear_request(req); | |
427 | nfs_release_request(req); | |
428 | } | |
429 | ||
430 | /* | |
431 | * Find a request | |
432 | */ | |
433 | static inline struct nfs_page * | |
434 | _nfs_find_request(struct inode *inode, unsigned long index) | |
435 | { | |
436 | struct nfs_inode *nfsi = NFS_I(inode); | |
437 | struct nfs_page *req; | |
438 | ||
439 | req = (struct nfs_page*)radix_tree_lookup(&nfsi->nfs_page_tree, index); | |
440 | if (req) | |
441 | atomic_inc(&req->wb_count); | |
442 | return req; | |
443 | } | |
444 | ||
445 | static struct nfs_page * | |
446 | nfs_find_request(struct inode *inode, unsigned long index) | |
447 | { | |
448 | struct nfs_page *req; | |
449 | struct nfs_inode *nfsi = NFS_I(inode); | |
450 | ||
451 | spin_lock(&nfsi->req_lock); | |
452 | req = _nfs_find_request(inode, index); | |
453 | spin_unlock(&nfsi->req_lock); | |
454 | return req; | |
455 | } | |
456 | ||
457 | /* | |
458 | * Add a request to the inode's dirty list. | |
459 | */ | |
460 | static void | |
461 | nfs_mark_request_dirty(struct nfs_page *req) | |
462 | { | |
463 | struct inode *inode = req->wb_context->dentry->d_inode; | |
464 | struct nfs_inode *nfsi = NFS_I(inode); | |
465 | ||
466 | spin_lock(&nfsi->req_lock); | |
3da28eb1 TM |
467 | radix_tree_tag_set(&nfsi->nfs_page_tree, |
468 | req->wb_index, NFS_PAGE_TAG_DIRTY); | |
1da177e4 LT |
469 | nfs_list_add_request(req, &nfsi->dirty); |
470 | nfsi->ndirty++; | |
471 | spin_unlock(&nfsi->req_lock); | |
472 | inc_page_state(nr_dirty); | |
473 | mark_inode_dirty(inode); | |
474 | } | |
475 | ||
476 | /* | |
477 | * Check if a request is dirty | |
478 | */ | |
479 | static inline int | |
480 | nfs_dirty_request(struct nfs_page *req) | |
481 | { | |
482 | struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode); | |
483 | return !list_empty(&req->wb_list) && req->wb_list_head == &nfsi->dirty; | |
484 | } | |
485 | ||
486 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) | |
487 | /* | |
488 | * Add a request to the inode's commit list. | |
489 | */ | |
490 | static void | |
491 | nfs_mark_request_commit(struct nfs_page *req) | |
492 | { | |
493 | struct inode *inode = req->wb_context->dentry->d_inode; | |
494 | struct nfs_inode *nfsi = NFS_I(inode); | |
495 | ||
496 | spin_lock(&nfsi->req_lock); | |
497 | nfs_list_add_request(req, &nfsi->commit); | |
498 | nfsi->ncommit++; | |
499 | spin_unlock(&nfsi->req_lock); | |
500 | inc_page_state(nr_unstable); | |
501 | mark_inode_dirty(inode); | |
502 | } | |
503 | #endif | |
504 | ||
505 | /* | |
506 | * Wait for a request to complete. | |
507 | * | |
508 | * Interruptible by signals only if mounted with intr flag. | |
509 | */ | |
510 | static int | |
511 | nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages) | |
512 | { | |
513 | struct nfs_inode *nfsi = NFS_I(inode); | |
514 | struct nfs_page *req; | |
515 | unsigned long idx_end, next; | |
516 | unsigned int res = 0; | |
517 | int error; | |
518 | ||
519 | if (npages == 0) | |
520 | idx_end = ~0; | |
521 | else | |
522 | idx_end = idx_start + npages - 1; | |
523 | ||
524 | spin_lock(&nfsi->req_lock); | |
525 | next = idx_start; | |
c6a556b8 | 526 | while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) { |
1da177e4 LT |
527 | if (req->wb_index > idx_end) |
528 | break; | |
529 | ||
530 | next = req->wb_index + 1; | |
c6a556b8 | 531 | BUG_ON(!NFS_WBACK_BUSY(req)); |
1da177e4 LT |
532 | |
533 | atomic_inc(&req->wb_count); | |
534 | spin_unlock(&nfsi->req_lock); | |
535 | error = nfs_wait_on_request(req); | |
536 | nfs_release_request(req); | |
537 | if (error < 0) | |
538 | return error; | |
539 | spin_lock(&nfsi->req_lock); | |
540 | res++; | |
541 | } | |
542 | spin_unlock(&nfsi->req_lock); | |
543 | return res; | |
544 | } | |
545 | ||
546 | /* | |
547 | * nfs_scan_dirty - Scan an inode for dirty requests | |
548 | * @inode: NFS inode to scan | |
549 | * @dst: destination list | |
550 | * @idx_start: lower bound of page->index to scan. | |
551 | * @npages: idx_start + npages sets the upper bound to scan. | |
552 | * | |
553 | * Moves requests from the inode's dirty page list. | |
554 | * The requests are *not* checked to ensure that they form a contiguous set. | |
555 | */ | |
556 | static int | |
557 | nfs_scan_dirty(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages) | |
558 | { | |
559 | struct nfs_inode *nfsi = NFS_I(inode); | |
3da28eb1 TM |
560 | int res = 0; |
561 | ||
562 | if (nfsi->ndirty != 0) { | |
563 | res = nfs_scan_lock_dirty(nfsi, dst, idx_start, npages); | |
564 | nfsi->ndirty -= res; | |
565 | sub_page_state(nr_dirty,res); | |
566 | if ((nfsi->ndirty == 0) != list_empty(&nfsi->dirty)) | |
567 | printk(KERN_ERR "NFS: desynchronized value of nfs_i.ndirty.\n"); | |
568 | } | |
1da177e4 LT |
569 | return res; |
570 | } | |
571 | ||
572 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) | |
573 | /* | |
574 | * nfs_scan_commit - Scan an inode for commit requests | |
575 | * @inode: NFS inode to scan | |
576 | * @dst: destination list | |
577 | * @idx_start: lower bound of page->index to scan. | |
578 | * @npages: idx_start + npages sets the upper bound to scan. | |
579 | * | |
580 | * Moves requests from the inode's 'commit' request list. | |
581 | * The requests are *not* checked to ensure that they form a contiguous set. | |
582 | */ | |
583 | static int | |
584 | nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages) | |
585 | { | |
586 | struct nfs_inode *nfsi = NFS_I(inode); | |
3da28eb1 TM |
587 | int res = 0; |
588 | ||
589 | if (nfsi->ncommit != 0) { | |
590 | res = nfs_scan_list(&nfsi->commit, dst, idx_start, npages); | |
591 | nfsi->ncommit -= res; | |
592 | if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit)) | |
593 | printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n"); | |
594 | } | |
1da177e4 LT |
595 | return res; |
596 | } | |
597 | #endif | |
598 | ||
599 | static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr) | |
600 | { | |
601 | struct backing_dev_info *bdi = mapping->backing_dev_info; | |
602 | DEFINE_WAIT(wait); | |
603 | int ret = 0; | |
604 | ||
605 | might_sleep(); | |
606 | ||
607 | if (!bdi_write_congested(bdi)) | |
608 | return 0; | |
91d5b470 CL |
609 | |
610 | nfs_inc_stats(mapping->host, NFSIOS_CONGESTIONWAIT); | |
611 | ||
1da177e4 LT |
612 | if (intr) { |
613 | struct rpc_clnt *clnt = NFS_CLIENT(mapping->host); | |
614 | sigset_t oldset; | |
615 | ||
616 | rpc_clnt_sigmask(clnt, &oldset); | |
617 | prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE); | |
618 | if (bdi_write_congested(bdi)) { | |
619 | if (signalled()) | |
620 | ret = -ERESTARTSYS; | |
621 | else | |
622 | schedule(); | |
623 | } | |
624 | rpc_clnt_sigunmask(clnt, &oldset); | |
625 | } else { | |
626 | prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE); | |
627 | if (bdi_write_congested(bdi)) | |
628 | schedule(); | |
629 | } | |
630 | finish_wait(&nfs_write_congestion, &wait); | |
631 | return ret; | |
632 | } | |
633 | ||
634 | ||
635 | /* | |
636 | * Try to update any existing write request, or create one if there is none. | |
637 | * In order to match, the request's credentials must match those of | |
638 | * the calling process. | |
639 | * | |
640 | * Note: Should always be called with the Page Lock held! | |
641 | */ | |
642 | static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx, | |
643 | struct inode *inode, struct page *page, | |
644 | unsigned int offset, unsigned int bytes) | |
645 | { | |
646 | struct nfs_server *server = NFS_SERVER(inode); | |
647 | struct nfs_inode *nfsi = NFS_I(inode); | |
648 | struct nfs_page *req, *new = NULL; | |
649 | unsigned long rqend, end; | |
650 | ||
651 | end = offset + bytes; | |
652 | ||
653 | if (nfs_wait_on_write_congestion(page->mapping, server->flags & NFS_MOUNT_INTR)) | |
654 | return ERR_PTR(-ERESTARTSYS); | |
655 | for (;;) { | |
656 | /* Loop over all inode entries and see if we find | |
657 | * A request for the page we wish to update | |
658 | */ | |
659 | spin_lock(&nfsi->req_lock); | |
660 | req = _nfs_find_request(inode, page->index); | |
661 | if (req) { | |
662 | if (!nfs_lock_request_dontget(req)) { | |
663 | int error; | |
664 | spin_unlock(&nfsi->req_lock); | |
665 | error = nfs_wait_on_request(req); | |
666 | nfs_release_request(req); | |
1dd594b2 NB |
667 | if (error < 0) { |
668 | if (new) | |
669 | nfs_release_request(new); | |
1da177e4 | 670 | return ERR_PTR(error); |
1dd594b2 | 671 | } |
1da177e4 LT |
672 | continue; |
673 | } | |
674 | spin_unlock(&nfsi->req_lock); | |
675 | if (new) | |
676 | nfs_release_request(new); | |
677 | break; | |
678 | } | |
679 | ||
680 | if (new) { | |
681 | int error; | |
682 | nfs_lock_request_dontget(new); | |
683 | error = nfs_inode_add_request(inode, new); | |
684 | if (error) { | |
685 | spin_unlock(&nfsi->req_lock); | |
686 | nfs_unlock_request(new); | |
687 | return ERR_PTR(error); | |
688 | } | |
689 | spin_unlock(&nfsi->req_lock); | |
690 | nfs_mark_request_dirty(new); | |
691 | return new; | |
692 | } | |
693 | spin_unlock(&nfsi->req_lock); | |
694 | ||
695 | new = nfs_create_request(ctx, inode, page, offset, bytes); | |
696 | if (IS_ERR(new)) | |
697 | return new; | |
698 | } | |
699 | ||
700 | /* We have a request for our page. | |
701 | * If the creds don't match, or the | |
702 | * page addresses don't match, | |
703 | * tell the caller to wait on the conflicting | |
704 | * request. | |
705 | */ | |
706 | rqend = req->wb_offset + req->wb_bytes; | |
707 | if (req->wb_context != ctx | |
708 | || req->wb_page != page | |
709 | || !nfs_dirty_request(req) | |
710 | || offset > rqend || end < req->wb_offset) { | |
711 | nfs_unlock_request(req); | |
712 | return ERR_PTR(-EBUSY); | |
713 | } | |
714 | ||
715 | /* Okay, the request matches. Update the region */ | |
716 | if (offset < req->wb_offset) { | |
717 | req->wb_offset = offset; | |
718 | req->wb_pgbase = offset; | |
719 | req->wb_bytes = rqend - req->wb_offset; | |
720 | } | |
721 | ||
722 | if (end > rqend) | |
723 | req->wb_bytes = end - req->wb_offset; | |
724 | ||
725 | return req; | |
726 | } | |
727 | ||
728 | int nfs_flush_incompatible(struct file *file, struct page *page) | |
729 | { | |
730 | struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data; | |
731 | struct inode *inode = page->mapping->host; | |
732 | struct nfs_page *req; | |
733 | int status = 0; | |
734 | /* | |
735 | * Look for a request corresponding to this page. If there | |
736 | * is one, and it belongs to another file, we flush it out | |
737 | * before we try to copy anything into the page. Do this | |
738 | * due to the lack of an ACCESS-type call in NFSv2. | |
739 | * Also do the same if we find a request from an existing | |
740 | * dropped page. | |
741 | */ | |
742 | req = nfs_find_request(inode, page->index); | |
743 | if (req) { | |
744 | if (req->wb_page != page || ctx != req->wb_context) | |
745 | status = nfs_wb_page(inode, page); | |
746 | nfs_release_request(req); | |
747 | } | |
748 | return (status < 0) ? status : 0; | |
749 | } | |
750 | ||
751 | /* | |
752 | * Update and possibly write a cached page of an NFS file. | |
753 | * | |
754 | * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad | |
755 | * things with a page scheduled for an RPC call (e.g. invalidate it). | |
756 | */ | |
757 | int nfs_updatepage(struct file *file, struct page *page, | |
758 | unsigned int offset, unsigned int count) | |
759 | { | |
760 | struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data; | |
1da177e4 LT |
761 | struct inode *inode = page->mapping->host; |
762 | struct nfs_page *req; | |
763 | int status = 0; | |
764 | ||
91d5b470 CL |
765 | nfs_inc_stats(inode, NFSIOS_VFSUPDATEPAGE); |
766 | ||
1da177e4 | 767 | dprintk("NFS: nfs_updatepage(%s/%s %d@%Ld)\n", |
0bbacc40 CL |
768 | file->f_dentry->d_parent->d_name.name, |
769 | file->f_dentry->d_name.name, count, | |
770 | (long long)(page_offset(page) +offset)); | |
1da177e4 LT |
771 | |
772 | if (IS_SYNC(inode)) { | |
773 | status = nfs_writepage_sync(ctx, inode, page, offset, count, 0); | |
774 | if (status > 0) { | |
775 | if (offset == 0 && status == PAGE_CACHE_SIZE) | |
776 | SetPageUptodate(page); | |
777 | return 0; | |
778 | } | |
779 | return status; | |
780 | } | |
781 | ||
782 | /* If we're not using byte range locks, and we know the page | |
783 | * is entirely in cache, it may be more efficient to avoid | |
784 | * fragmenting write requests. | |
785 | */ | |
ab0a3dbe | 786 | if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) { |
1da177e4 LT |
787 | loff_t end_offs = i_size_read(inode) - 1; |
788 | unsigned long end_index = end_offs >> PAGE_CACHE_SHIFT; | |
789 | ||
790 | count += offset; | |
791 | offset = 0; | |
792 | if (unlikely(end_offs < 0)) { | |
793 | /* Do nothing */ | |
794 | } else if (page->index == end_index) { | |
795 | unsigned int pglen; | |
796 | pglen = (unsigned int)(end_offs & (PAGE_CACHE_SIZE-1)) + 1; | |
797 | if (count < pglen) | |
798 | count = pglen; | |
799 | } else if (page->index < end_index) | |
800 | count = PAGE_CACHE_SIZE; | |
801 | } | |
802 | ||
803 | /* | |
804 | * Try to find an NFS request corresponding to this page | |
805 | * and update it. | |
806 | * If the existing request cannot be updated, we must flush | |
807 | * it out now. | |
808 | */ | |
809 | do { | |
810 | req = nfs_update_request(ctx, inode, page, offset, count); | |
811 | status = (IS_ERR(req)) ? PTR_ERR(req) : 0; | |
812 | if (status != -EBUSY) | |
813 | break; | |
814 | /* Request could not be updated. Flush it out and try again */ | |
815 | status = nfs_wb_page(inode, page); | |
816 | } while (status >= 0); | |
817 | if (status < 0) | |
818 | goto done; | |
819 | ||
820 | status = 0; | |
821 | ||
822 | /* Update file length */ | |
823 | nfs_grow_file(page, offset, count); | |
824 | /* Set the PG_uptodate flag? */ | |
825 | nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes); | |
826 | nfs_unlock_request(req); | |
827 | done: | |
828 | dprintk("NFS: nfs_updatepage returns %d (isize %Ld)\n", | |
829 | status, (long long)i_size_read(inode)); | |
830 | if (status < 0) | |
831 | ClearPageUptodate(page); | |
832 | return status; | |
833 | } | |
834 | ||
835 | static void nfs_writepage_release(struct nfs_page *req) | |
836 | { | |
837 | end_page_writeback(req->wb_page); | |
838 | ||
839 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) | |
840 | if (!PageError(req->wb_page)) { | |
841 | if (NFS_NEED_RESCHED(req)) { | |
842 | nfs_mark_request_dirty(req); | |
843 | goto out; | |
844 | } else if (NFS_NEED_COMMIT(req)) { | |
845 | nfs_mark_request_commit(req); | |
846 | goto out; | |
847 | } | |
848 | } | |
849 | nfs_inode_remove_request(req); | |
850 | ||
851 | out: | |
852 | nfs_clear_commit(req); | |
853 | nfs_clear_reschedule(req); | |
854 | #else | |
855 | nfs_inode_remove_request(req); | |
856 | #endif | |
c6a556b8 | 857 | nfs_clear_page_writeback(req); |
1da177e4 LT |
858 | } |
859 | ||
860 | static inline int flush_task_priority(int how) | |
861 | { | |
862 | switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) { | |
863 | case FLUSH_HIGHPRI: | |
864 | return RPC_PRIORITY_HIGH; | |
865 | case FLUSH_LOWPRI: | |
866 | return RPC_PRIORITY_LOW; | |
867 | } | |
868 | return RPC_PRIORITY_NORMAL; | |
869 | } | |
870 | ||
871 | /* | |
872 | * Set up the argument/result storage required for the RPC call. | |
873 | */ | |
874 | static void nfs_write_rpcsetup(struct nfs_page *req, | |
875 | struct nfs_write_data *data, | |
788e7a89 | 876 | const struct rpc_call_ops *call_ops, |
1da177e4 LT |
877 | unsigned int count, unsigned int offset, |
878 | int how) | |
879 | { | |
1da177e4 | 880 | struct inode *inode; |
788e7a89 | 881 | int flags; |
1da177e4 LT |
882 | |
883 | /* Set up the RPC argument and reply structs | |
884 | * NB: take care not to mess about with data->commit et al. */ | |
885 | ||
886 | data->req = req; | |
887 | data->inode = inode = req->wb_context->dentry->d_inode; | |
888 | data->cred = req->wb_context->cred; | |
889 | ||
890 | data->args.fh = NFS_FH(inode); | |
891 | data->args.offset = req_offset(req) + offset; | |
892 | data->args.pgbase = req->wb_pgbase + offset; | |
893 | data->args.pages = data->pagevec; | |
894 | data->args.count = count; | |
895 | data->args.context = req->wb_context; | |
896 | ||
897 | data->res.fattr = &data->fattr; | |
898 | data->res.count = count; | |
899 | data->res.verf = &data->verf; | |
0e574af1 | 900 | nfs_fattr_init(&data->fattr); |
1da177e4 | 901 | |
788e7a89 TM |
902 | /* Set up the initial task struct. */ |
903 | flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC; | |
904 | rpc_init_task(&data->task, NFS_CLIENT(inode), flags, call_ops, data); | |
1da177e4 LT |
905 | NFS_PROTO(inode)->write_setup(data, how); |
906 | ||
907 | data->task.tk_priority = flush_task_priority(how); | |
908 | data->task.tk_cookie = (unsigned long)inode; | |
1da177e4 LT |
909 | |
910 | dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n", | |
0bbacc40 | 911 | data->task.tk_pid, |
1da177e4 LT |
912 | inode->i_sb->s_id, |
913 | (long long)NFS_FILEID(inode), | |
914 | count, | |
915 | (unsigned long long)data->args.offset); | |
916 | } | |
917 | ||
918 | static void nfs_execute_write(struct nfs_write_data *data) | |
919 | { | |
920 | struct rpc_clnt *clnt = NFS_CLIENT(data->inode); | |
921 | sigset_t oldset; | |
922 | ||
923 | rpc_clnt_sigmask(clnt, &oldset); | |
924 | lock_kernel(); | |
925 | rpc_execute(&data->task); | |
926 | unlock_kernel(); | |
927 | rpc_clnt_sigunmask(clnt, &oldset); | |
928 | } | |
929 | ||
930 | /* | |
931 | * Generate multiple small requests to write out a single | |
932 | * contiguous dirty area on one page. | |
933 | */ | |
934 | static int nfs_flush_multi(struct list_head *head, struct inode *inode, int how) | |
935 | { | |
936 | struct nfs_page *req = nfs_list_entry(head->next); | |
937 | struct page *page = req->wb_page; | |
938 | struct nfs_write_data *data; | |
939 | unsigned int wsize = NFS_SERVER(inode)->wsize; | |
940 | unsigned int nbytes, offset; | |
941 | int requests = 0; | |
942 | LIST_HEAD(list); | |
943 | ||
944 | nfs_list_remove_request(req); | |
945 | ||
946 | nbytes = req->wb_bytes; | |
947 | for (;;) { | |
40859d7e | 948 | data = nfs_writedata_alloc(1); |
1da177e4 LT |
949 | if (!data) |
950 | goto out_bad; | |
951 | list_add(&data->pages, &list); | |
952 | requests++; | |
953 | if (nbytes <= wsize) | |
954 | break; | |
955 | nbytes -= wsize; | |
956 | } | |
957 | atomic_set(&req->wb_complete, requests); | |
958 | ||
959 | ClearPageError(page); | |
bb713d6d | 960 | set_page_writeback(page); |
1da177e4 LT |
961 | offset = 0; |
962 | nbytes = req->wb_bytes; | |
963 | do { | |
964 | data = list_entry(list.next, struct nfs_write_data, pages); | |
965 | list_del_init(&data->pages); | |
966 | ||
967 | data->pagevec[0] = page; | |
1da177e4 LT |
968 | |
969 | if (nbytes > wsize) { | |
788e7a89 TM |
970 | nfs_write_rpcsetup(req, data, &nfs_write_partial_ops, |
971 | wsize, offset, how); | |
1da177e4 LT |
972 | offset += wsize; |
973 | nbytes -= wsize; | |
974 | } else { | |
788e7a89 TM |
975 | nfs_write_rpcsetup(req, data, &nfs_write_partial_ops, |
976 | nbytes, offset, how); | |
1da177e4 LT |
977 | nbytes = 0; |
978 | } | |
979 | nfs_execute_write(data); | |
980 | } while (nbytes != 0); | |
981 | ||
982 | return 0; | |
983 | ||
984 | out_bad: | |
985 | while (!list_empty(&list)) { | |
986 | data = list_entry(list.next, struct nfs_write_data, pages); | |
987 | list_del(&data->pages); | |
988 | nfs_writedata_free(data); | |
989 | } | |
990 | nfs_mark_request_dirty(req); | |
c6a556b8 | 991 | nfs_clear_page_writeback(req); |
1da177e4 LT |
992 | return -ENOMEM; |
993 | } | |
994 | ||
995 | /* | |
996 | * Create an RPC task for the given write request and kick it. | |
997 | * The page must have been locked by the caller. | |
998 | * | |
999 | * It may happen that the page we're passed is not marked dirty. | |
1000 | * This is the case if nfs_updatepage detects a conflicting request | |
1001 | * that has been written but not committed. | |
1002 | */ | |
1003 | static int nfs_flush_one(struct list_head *head, struct inode *inode, int how) | |
1004 | { | |
1005 | struct nfs_page *req; | |
1006 | struct page **pages; | |
1007 | struct nfs_write_data *data; | |
1008 | unsigned int count; | |
1009 | ||
1010 | if (NFS_SERVER(inode)->wsize < PAGE_CACHE_SIZE) | |
1011 | return nfs_flush_multi(head, inode, how); | |
1012 | ||
40859d7e | 1013 | data = nfs_writedata_alloc(NFS_SERVER(inode)->wpages); |
1da177e4 LT |
1014 | if (!data) |
1015 | goto out_bad; | |
1016 | ||
1017 | pages = data->pagevec; | |
1018 | count = 0; | |
1019 | while (!list_empty(head)) { | |
1020 | req = nfs_list_entry(head->next); | |
1021 | nfs_list_remove_request(req); | |
1022 | nfs_list_add_request(req, &data->pages); | |
1023 | ClearPageError(req->wb_page); | |
bb713d6d | 1024 | set_page_writeback(req->wb_page); |
1da177e4 LT |
1025 | *pages++ = req->wb_page; |
1026 | count += req->wb_bytes; | |
1027 | } | |
1028 | req = nfs_list_entry(data->pages.next); | |
1029 | ||
1da177e4 | 1030 | /* Set up the argument struct */ |
788e7a89 | 1031 | nfs_write_rpcsetup(req, data, &nfs_write_full_ops, count, 0, how); |
1da177e4 LT |
1032 | |
1033 | nfs_execute_write(data); | |
1034 | return 0; | |
1035 | out_bad: | |
1036 | while (!list_empty(head)) { | |
1037 | struct nfs_page *req = nfs_list_entry(head->next); | |
1038 | nfs_list_remove_request(req); | |
1039 | nfs_mark_request_dirty(req); | |
c6a556b8 | 1040 | nfs_clear_page_writeback(req); |
1da177e4 LT |
1041 | } |
1042 | return -ENOMEM; | |
1043 | } | |
1044 | ||
1045 | static int | |
1046 | nfs_flush_list(struct list_head *head, int wpages, int how) | |
1047 | { | |
1048 | LIST_HEAD(one_request); | |
1049 | struct nfs_page *req; | |
1050 | int error = 0; | |
1051 | unsigned int pages = 0; | |
1052 | ||
1053 | while (!list_empty(head)) { | |
1054 | pages += nfs_coalesce_requests(head, &one_request, wpages); | |
1055 | req = nfs_list_entry(one_request.next); | |
1056 | error = nfs_flush_one(&one_request, req->wb_context->dentry->d_inode, how); | |
1057 | if (error < 0) | |
1058 | break; | |
1059 | } | |
1060 | if (error >= 0) | |
1061 | return pages; | |
1062 | ||
1063 | while (!list_empty(head)) { | |
1064 | req = nfs_list_entry(head->next); | |
1065 | nfs_list_remove_request(req); | |
1066 | nfs_mark_request_dirty(req); | |
c6a556b8 | 1067 | nfs_clear_page_writeback(req); |
1da177e4 LT |
1068 | } |
1069 | return error; | |
1070 | } | |
1071 | ||
1072 | /* | |
1073 | * Handle a write reply that flushed part of a page. | |
1074 | */ | |
788e7a89 | 1075 | static void nfs_writeback_done_partial(struct rpc_task *task, void *calldata) |
1da177e4 | 1076 | { |
788e7a89 | 1077 | struct nfs_write_data *data = calldata; |
1da177e4 LT |
1078 | struct nfs_page *req = data->req; |
1079 | struct page *page = req->wb_page; | |
1080 | ||
1081 | dprintk("NFS: write (%s/%Ld %d@%Ld)", | |
1082 | req->wb_context->dentry->d_inode->i_sb->s_id, | |
1083 | (long long)NFS_FILEID(req->wb_context->dentry->d_inode), | |
1084 | req->wb_bytes, | |
1085 | (long long)req_offset(req)); | |
1086 | ||
788e7a89 TM |
1087 | if (nfs_writeback_done(task, data) != 0) |
1088 | return; | |
1089 | ||
1090 | if (task->tk_status < 0) { | |
1da177e4 LT |
1091 | ClearPageUptodate(page); |
1092 | SetPageError(page); | |
788e7a89 TM |
1093 | req->wb_context->error = task->tk_status; |
1094 | dprintk(", error = %d\n", task->tk_status); | |
1da177e4 LT |
1095 | } else { |
1096 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) | |
1097 | if (data->verf.committed < NFS_FILE_SYNC) { | |
1098 | if (!NFS_NEED_COMMIT(req)) { | |
1099 | nfs_defer_commit(req); | |
1100 | memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf)); | |
1101 | dprintk(" defer commit\n"); | |
1102 | } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) { | |
1103 | nfs_defer_reschedule(req); | |
1104 | dprintk(" server reboot detected\n"); | |
1105 | } | |
1106 | } else | |
1107 | #endif | |
1108 | dprintk(" OK\n"); | |
1109 | } | |
1110 | ||
1111 | if (atomic_dec_and_test(&req->wb_complete)) | |
1112 | nfs_writepage_release(req); | |
1113 | } | |
1114 | ||
788e7a89 TM |
1115 | static const struct rpc_call_ops nfs_write_partial_ops = { |
1116 | .rpc_call_done = nfs_writeback_done_partial, | |
1117 | .rpc_release = nfs_writedata_release, | |
1118 | }; | |
1119 | ||
1da177e4 LT |
1120 | /* |
1121 | * Handle a write reply that flushes a whole page. | |
1122 | * | |
1123 | * FIXME: There is an inherent race with invalidate_inode_pages and | |
1124 | * writebacks since the page->count is kept > 1 for as long | |
1125 | * as the page has a write request pending. | |
1126 | */ | |
788e7a89 | 1127 | static void nfs_writeback_done_full(struct rpc_task *task, void *calldata) |
1da177e4 | 1128 | { |
788e7a89 | 1129 | struct nfs_write_data *data = calldata; |
1da177e4 LT |
1130 | struct nfs_page *req; |
1131 | struct page *page; | |
1132 | ||
788e7a89 TM |
1133 | if (nfs_writeback_done(task, data) != 0) |
1134 | return; | |
1135 | ||
1da177e4 LT |
1136 | /* Update attributes as result of writeback. */ |
1137 | while (!list_empty(&data->pages)) { | |
1138 | req = nfs_list_entry(data->pages.next); | |
1139 | nfs_list_remove_request(req); | |
1140 | page = req->wb_page; | |
1141 | ||
1142 | dprintk("NFS: write (%s/%Ld %d@%Ld)", | |
1143 | req->wb_context->dentry->d_inode->i_sb->s_id, | |
1144 | (long long)NFS_FILEID(req->wb_context->dentry->d_inode), | |
1145 | req->wb_bytes, | |
1146 | (long long)req_offset(req)); | |
1147 | ||
788e7a89 | 1148 | if (task->tk_status < 0) { |
1da177e4 LT |
1149 | ClearPageUptodate(page); |
1150 | SetPageError(page); | |
788e7a89 | 1151 | req->wb_context->error = task->tk_status; |
1da177e4 LT |
1152 | end_page_writeback(page); |
1153 | nfs_inode_remove_request(req); | |
788e7a89 | 1154 | dprintk(", error = %d\n", task->tk_status); |
1da177e4 LT |
1155 | goto next; |
1156 | } | |
1157 | end_page_writeback(page); | |
1158 | ||
1159 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) | |
1160 | if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) { | |
1161 | nfs_inode_remove_request(req); | |
1162 | dprintk(" OK\n"); | |
1163 | goto next; | |
1164 | } | |
1165 | memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf)); | |
1166 | nfs_mark_request_commit(req); | |
1167 | dprintk(" marked for commit\n"); | |
1168 | #else | |
1169 | nfs_inode_remove_request(req); | |
1170 | #endif | |
1171 | next: | |
c6a556b8 | 1172 | nfs_clear_page_writeback(req); |
1da177e4 LT |
1173 | } |
1174 | } | |
1175 | ||
788e7a89 TM |
1176 | static const struct rpc_call_ops nfs_write_full_ops = { |
1177 | .rpc_call_done = nfs_writeback_done_full, | |
1178 | .rpc_release = nfs_writedata_release, | |
1179 | }; | |
1180 | ||
1181 | ||
1da177e4 LT |
1182 | /* |
1183 | * This function is called when the WRITE call is complete. | |
1184 | */ | |
462d5b32 | 1185 | int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data) |
1da177e4 | 1186 | { |
1da177e4 LT |
1187 | struct nfs_writeargs *argp = &data->args; |
1188 | struct nfs_writeres *resp = &data->res; | |
788e7a89 | 1189 | int status; |
1da177e4 LT |
1190 | |
1191 | dprintk("NFS: %4d nfs_writeback_done (status %d)\n", | |
1192 | task->tk_pid, task->tk_status); | |
1193 | ||
788e7a89 TM |
1194 | /* Call the NFS version-specific code */ |
1195 | status = NFS_PROTO(data->inode)->write_done(task, data); | |
1196 | if (status != 0) | |
1197 | return status; | |
91d5b470 CL |
1198 | nfs_add_stats(data->inode, NFSIOS_SERVERWRITTENBYTES, resp->count); |
1199 | ||
1da177e4 LT |
1200 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) |
1201 | if (resp->verf->committed < argp->stable && task->tk_status >= 0) { | |
1202 | /* We tried a write call, but the server did not | |
1203 | * commit data to stable storage even though we | |
1204 | * requested it. | |
1205 | * Note: There is a known bug in Tru64 < 5.0 in which | |
1206 | * the server reports NFS_DATA_SYNC, but performs | |
1207 | * NFS_FILE_SYNC. We therefore implement this checking | |
1208 | * as a dprintk() in order to avoid filling syslog. | |
1209 | */ | |
1210 | static unsigned long complain; | |
1211 | ||
1212 | if (time_before(complain, jiffies)) { | |
1213 | dprintk("NFS: faulty NFS server %s:" | |
1214 | " (committed = %d) != (stable = %d)\n", | |
1215 | NFS_SERVER(data->inode)->hostname, | |
1216 | resp->verf->committed, argp->stable); | |
1217 | complain = jiffies + 300 * HZ; | |
1218 | } | |
1219 | } | |
1220 | #endif | |
1221 | /* Is this a short write? */ | |
1222 | if (task->tk_status >= 0 && resp->count < argp->count) { | |
1223 | static unsigned long complain; | |
1224 | ||
91d5b470 CL |
1225 | nfs_inc_stats(data->inode, NFSIOS_SHORTWRITE); |
1226 | ||
1da177e4 LT |
1227 | /* Has the server at least made some progress? */ |
1228 | if (resp->count != 0) { | |
1229 | /* Was this an NFSv2 write or an NFSv3 stable write? */ | |
1230 | if (resp->verf->committed != NFS_UNSTABLE) { | |
1231 | /* Resend from where the server left off */ | |
1232 | argp->offset += resp->count; | |
1233 | argp->pgbase += resp->count; | |
1234 | argp->count -= resp->count; | |
1235 | } else { | |
1236 | /* Resend as a stable write in order to avoid | |
1237 | * headaches in the case of a server crash. | |
1238 | */ | |
1239 | argp->stable = NFS_FILE_SYNC; | |
1240 | } | |
1241 | rpc_restart_call(task); | |
788e7a89 | 1242 | return -EAGAIN; |
1da177e4 LT |
1243 | } |
1244 | if (time_before(complain, jiffies)) { | |
1245 | printk(KERN_WARNING | |
1246 | "NFS: Server wrote zero bytes, expected %u.\n", | |
1247 | argp->count); | |
1248 | complain = jiffies + 300 * HZ; | |
1249 | } | |
1250 | /* Can't do anything about it except throw an error. */ | |
1251 | task->tk_status = -EIO; | |
1252 | } | |
788e7a89 | 1253 | return 0; |
1da177e4 LT |
1254 | } |
1255 | ||
1256 | ||
1257 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) | |
963d8fe5 | 1258 | void nfs_commit_release(void *wdata) |
1da177e4 | 1259 | { |
1da177e4 LT |
1260 | nfs_commit_free(wdata); |
1261 | } | |
1262 | ||
1263 | /* | |
1264 | * Set up the argument/result storage required for the RPC call. | |
1265 | */ | |
1266 | static void nfs_commit_rpcsetup(struct list_head *head, | |
788e7a89 TM |
1267 | struct nfs_write_data *data, |
1268 | int how) | |
1da177e4 | 1269 | { |
3da28eb1 | 1270 | struct nfs_page *first; |
1da177e4 | 1271 | struct inode *inode; |
788e7a89 | 1272 | int flags; |
1da177e4 LT |
1273 | |
1274 | /* Set up the RPC argument and reply structs | |
1275 | * NB: take care not to mess about with data->commit et al. */ | |
1276 | ||
1277 | list_splice_init(head, &data->pages); | |
1278 | first = nfs_list_entry(data->pages.next); | |
1da177e4 LT |
1279 | inode = first->wb_context->dentry->d_inode; |
1280 | ||
1da177e4 LT |
1281 | data->inode = inode; |
1282 | data->cred = first->wb_context->cred; | |
1283 | ||
1284 | data->args.fh = NFS_FH(data->inode); | |
3da28eb1 TM |
1285 | /* Note: we always request a commit of the entire inode */ |
1286 | data->args.offset = 0; | |
1287 | data->args.count = 0; | |
1288 | data->res.count = 0; | |
1da177e4 LT |
1289 | data->res.fattr = &data->fattr; |
1290 | data->res.verf = &data->verf; | |
0e574af1 | 1291 | nfs_fattr_init(&data->fattr); |
788e7a89 TM |
1292 | |
1293 | /* Set up the initial task struct. */ | |
1294 | flags = (how & FLUSH_SYNC) ? 0 : RPC_TASK_ASYNC; | |
1295 | rpc_init_task(&data->task, NFS_CLIENT(inode), flags, &nfs_commit_ops, data); | |
1da177e4 LT |
1296 | NFS_PROTO(inode)->commit_setup(data, how); |
1297 | ||
1298 | data->task.tk_priority = flush_task_priority(how); | |
1299 | data->task.tk_cookie = (unsigned long)inode; | |
1da177e4 | 1300 | |
0bbacc40 | 1301 | dprintk("NFS: %4d initiated commit call\n", data->task.tk_pid); |
1da177e4 LT |
1302 | } |
1303 | ||
1304 | /* | |
1305 | * Commit dirty pages | |
1306 | */ | |
1307 | static int | |
40859d7e | 1308 | nfs_commit_list(struct inode *inode, struct list_head *head, int how) |
1da177e4 LT |
1309 | { |
1310 | struct nfs_write_data *data; | |
1311 | struct nfs_page *req; | |
1312 | ||
40859d7e | 1313 | data = nfs_commit_alloc(NFS_SERVER(inode)->wpages); |
1da177e4 LT |
1314 | |
1315 | if (!data) | |
1316 | goto out_bad; | |
1317 | ||
1318 | /* Set up the argument struct */ | |
1319 | nfs_commit_rpcsetup(head, data, how); | |
1320 | ||
1321 | nfs_execute_write(data); | |
1322 | return 0; | |
1323 | out_bad: | |
1324 | while (!list_empty(head)) { | |
1325 | req = nfs_list_entry(head->next); | |
1326 | nfs_list_remove_request(req); | |
1327 | nfs_mark_request_commit(req); | |
c6a556b8 | 1328 | nfs_clear_page_writeback(req); |
1da177e4 LT |
1329 | } |
1330 | return -ENOMEM; | |
1331 | } | |
1332 | ||
1333 | /* | |
1334 | * COMMIT call returned | |
1335 | */ | |
788e7a89 | 1336 | static void nfs_commit_done(struct rpc_task *task, void *calldata) |
1da177e4 | 1337 | { |
963d8fe5 | 1338 | struct nfs_write_data *data = calldata; |
1da177e4 LT |
1339 | struct nfs_page *req; |
1340 | int res = 0; | |
1341 | ||
1342 | dprintk("NFS: %4d nfs_commit_done (status %d)\n", | |
1343 | task->tk_pid, task->tk_status); | |
1344 | ||
788e7a89 TM |
1345 | /* Call the NFS version-specific code */ |
1346 | if (NFS_PROTO(data->inode)->commit_done(task, data) != 0) | |
1347 | return; | |
1348 | ||
1da177e4 LT |
1349 | while (!list_empty(&data->pages)) { |
1350 | req = nfs_list_entry(data->pages.next); | |
1351 | nfs_list_remove_request(req); | |
1352 | ||
1353 | dprintk("NFS: commit (%s/%Ld %d@%Ld)", | |
1354 | req->wb_context->dentry->d_inode->i_sb->s_id, | |
1355 | (long long)NFS_FILEID(req->wb_context->dentry->d_inode), | |
1356 | req->wb_bytes, | |
1357 | (long long)req_offset(req)); | |
1358 | if (task->tk_status < 0) { | |
1359 | req->wb_context->error = task->tk_status; | |
1360 | nfs_inode_remove_request(req); | |
1361 | dprintk(", error = %d\n", task->tk_status); | |
1362 | goto next; | |
1363 | } | |
1364 | ||
1365 | /* Okay, COMMIT succeeded, apparently. Check the verifier | |
1366 | * returned by the server against all stored verfs. */ | |
1367 | if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) { | |
1368 | /* We have a match */ | |
1369 | nfs_inode_remove_request(req); | |
1370 | dprintk(" OK\n"); | |
1371 | goto next; | |
1372 | } | |
1373 | /* We have a mismatch. Write the page again */ | |
1374 | dprintk(" mismatch\n"); | |
1375 | nfs_mark_request_dirty(req); | |
1376 | next: | |
c6a556b8 | 1377 | nfs_clear_page_writeback(req); |
1da177e4 LT |
1378 | res++; |
1379 | } | |
1380 | sub_page_state(nr_unstable,res); | |
1381 | } | |
788e7a89 TM |
1382 | |
1383 | static const struct rpc_call_ops nfs_commit_ops = { | |
1384 | .rpc_call_done = nfs_commit_done, | |
1385 | .rpc_release = nfs_commit_release, | |
1386 | }; | |
1da177e4 LT |
1387 | #endif |
1388 | ||
1389 | static int nfs_flush_inode(struct inode *inode, unsigned long idx_start, | |
1390 | unsigned int npages, int how) | |
1391 | { | |
1392 | struct nfs_inode *nfsi = NFS_I(inode); | |
1393 | LIST_HEAD(head); | |
1394 | int res, | |
1395 | error = 0; | |
1396 | ||
1397 | spin_lock(&nfsi->req_lock); | |
1398 | res = nfs_scan_dirty(inode, &head, idx_start, npages); | |
1399 | spin_unlock(&nfsi->req_lock); | |
ab0a3dbe TM |
1400 | if (res) { |
1401 | struct nfs_server *server = NFS_SERVER(inode); | |
1402 | ||
1403 | /* For single writes, FLUSH_STABLE is more efficient */ | |
1404 | if (res == nfsi->npages && nfsi->npages <= server->wpages) { | |
1405 | if (res > 1 || nfs_list_entry(head.next)->wb_bytes <= server->wsize) | |
1406 | how |= FLUSH_STABLE; | |
1407 | } | |
1408 | error = nfs_flush_list(&head, server->wpages, how); | |
1409 | } | |
1da177e4 LT |
1410 | if (error < 0) |
1411 | return error; | |
1412 | return res; | |
1413 | } | |
1414 | ||
1415 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) | |
3da28eb1 | 1416 | int nfs_commit_inode(struct inode *inode, int how) |
1da177e4 LT |
1417 | { |
1418 | struct nfs_inode *nfsi = NFS_I(inode); | |
1419 | LIST_HEAD(head); | |
1420 | int res, | |
1421 | error = 0; | |
1422 | ||
1423 | spin_lock(&nfsi->req_lock); | |
3da28eb1 TM |
1424 | res = nfs_scan_commit(inode, &head, 0, 0); |
1425 | spin_unlock(&nfsi->req_lock); | |
1da177e4 | 1426 | if (res) { |
40859d7e | 1427 | error = nfs_commit_list(inode, &head, how); |
3da28eb1 TM |
1428 | if (error < 0) |
1429 | return error; | |
1430 | } | |
1da177e4 LT |
1431 | return res; |
1432 | } | |
1433 | #endif | |
1434 | ||
1435 | int nfs_sync_inode(struct inode *inode, unsigned long idx_start, | |
1436 | unsigned int npages, int how) | |
1437 | { | |
70b9ecbd TM |
1438 | int nocommit = how & FLUSH_NOCOMMIT; |
1439 | int wait = how & FLUSH_WAIT; | |
1440 | int error; | |
1da177e4 | 1441 | |
70b9ecbd | 1442 | how &= ~(FLUSH_WAIT|FLUSH_NOCOMMIT); |
1da177e4 LT |
1443 | |
1444 | do { | |
70b9ecbd | 1445 | if (wait) { |
1da177e4 | 1446 | error = nfs_wait_on_requests(inode, idx_start, npages); |
70b9ecbd TM |
1447 | if (error != 0) |
1448 | continue; | |
1449 | } | |
1450 | error = nfs_flush_inode(inode, idx_start, npages, how); | |
1451 | if (error != 0) | |
1452 | continue; | |
1453 | if (!nocommit) | |
3da28eb1 | 1454 | error = nfs_commit_inode(inode, how); |
1da177e4 LT |
1455 | } while (error > 0); |
1456 | return error; | |
1457 | } | |
1458 | ||
1459 | int nfs_init_writepagecache(void) | |
1460 | { | |
1461 | nfs_wdata_cachep = kmem_cache_create("nfs_write_data", | |
1462 | sizeof(struct nfs_write_data), | |
1463 | 0, SLAB_HWCACHE_ALIGN, | |
1464 | NULL, NULL); | |
1465 | if (nfs_wdata_cachep == NULL) | |
1466 | return -ENOMEM; | |
1467 | ||
1468 | nfs_wdata_mempool = mempool_create(MIN_POOL_WRITE, | |
1469 | mempool_alloc_slab, | |
1470 | mempool_free_slab, | |
1471 | nfs_wdata_cachep); | |
1472 | if (nfs_wdata_mempool == NULL) | |
1473 | return -ENOMEM; | |
1474 | ||
1475 | nfs_commit_mempool = mempool_create(MIN_POOL_COMMIT, | |
1476 | mempool_alloc_slab, | |
1477 | mempool_free_slab, | |
1478 | nfs_wdata_cachep); | |
1479 | if (nfs_commit_mempool == NULL) | |
1480 | return -ENOMEM; | |
1481 | ||
1482 | return 0; | |
1483 | } | |
1484 | ||
1485 | void nfs_destroy_writepagecache(void) | |
1486 | { | |
1487 | mempool_destroy(nfs_commit_mempool); | |
1488 | mempool_destroy(nfs_wdata_mempool); | |
1489 | if (kmem_cache_destroy(nfs_wdata_cachep)) | |
1490 | printk(KERN_INFO "nfs_write_data: not all structures were freed\n"); | |
1491 | } | |
1492 |