]>
Commit | Line | Data |
---|---|---|
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" | |
66 | ||
67 | #define NFSDBG_FACILITY NFSDBG_PAGECACHE | |
68 | ||
69 | #define MIN_POOL_WRITE (32) | |
70 | #define MIN_POOL_COMMIT (4) | |
71 | ||
72 | /* | |
73 | * Local function declarations | |
74 | */ | |
75 | static struct nfs_page * nfs_update_request(struct nfs_open_context*, | |
76 | struct inode *, | |
77 | struct page *, | |
78 | unsigned int, unsigned int); | |
79 | static void nfs_writeback_done_partial(struct nfs_write_data *, int); | |
80 | static void nfs_writeback_done_full(struct nfs_write_data *, int); | |
81 | static int nfs_wait_on_write_congestion(struct address_space *, int); | |
82 | static int nfs_wait_on_requests(struct inode *, unsigned long, unsigned int); | |
83 | static int nfs_flush_inode(struct inode *inode, unsigned long idx_start, | |
84 | unsigned int npages, int how); | |
85 | ||
86 | static kmem_cache_t *nfs_wdata_cachep; | |
87 | mempool_t *nfs_wdata_mempool; | |
88 | static mempool_t *nfs_commit_mempool; | |
89 | ||
90 | static DECLARE_WAIT_QUEUE_HEAD(nfs_write_congestion); | |
91 | ||
92 | static inline struct nfs_write_data *nfs_commit_alloc(void) | |
93 | { | |
94 | struct nfs_write_data *p = mempool_alloc(nfs_commit_mempool, SLAB_NOFS); | |
95 | if (p) { | |
96 | memset(p, 0, sizeof(*p)); | |
97 | INIT_LIST_HEAD(&p->pages); | |
98 | } | |
99 | return p; | |
100 | } | |
101 | ||
102 | static inline void nfs_commit_free(struct nfs_write_data *p) | |
103 | { | |
104 | mempool_free(p, nfs_commit_mempool); | |
105 | } | |
106 | ||
107 | static void nfs_writedata_release(struct rpc_task *task) | |
108 | { | |
109 | struct nfs_write_data *wdata = (struct nfs_write_data *)task->tk_calldata; | |
110 | nfs_writedata_free(wdata); | |
111 | } | |
112 | ||
113 | /* Adjust the file length if we're writing beyond the end */ | |
114 | static void nfs_grow_file(struct page *page, unsigned int offset, unsigned int count) | |
115 | { | |
116 | struct inode *inode = page->mapping->host; | |
117 | loff_t end, i_size = i_size_read(inode); | |
118 | unsigned long end_index = (i_size - 1) >> PAGE_CACHE_SHIFT; | |
119 | ||
120 | if (i_size > 0 && page->index < end_index) | |
121 | return; | |
122 | end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + ((loff_t)offset+count); | |
123 | if (i_size >= end) | |
124 | return; | |
125 | i_size_write(inode, end); | |
126 | } | |
127 | ||
128 | /* We can set the PG_uptodate flag if we see that a write request | |
129 | * covers the full page. | |
130 | */ | |
131 | static void nfs_mark_uptodate(struct page *page, unsigned int base, unsigned int count) | |
132 | { | |
133 | loff_t end_offs; | |
134 | ||
135 | if (PageUptodate(page)) | |
136 | return; | |
137 | if (base != 0) | |
138 | return; | |
139 | if (count == PAGE_CACHE_SIZE) { | |
140 | SetPageUptodate(page); | |
141 | return; | |
142 | } | |
143 | ||
144 | end_offs = i_size_read(page->mapping->host) - 1; | |
145 | if (end_offs < 0) | |
146 | return; | |
147 | /* Is this the last page? */ | |
148 | if (page->index != (unsigned long)(end_offs >> PAGE_CACHE_SHIFT)) | |
149 | return; | |
150 | /* This is the last page: set PG_uptodate if we cover the entire | |
151 | * extent of the data, then zero the rest of the page. | |
152 | */ | |
153 | if (count == (unsigned int)(end_offs & (PAGE_CACHE_SIZE - 1)) + 1) { | |
154 | memclear_highpage_flush(page, count, PAGE_CACHE_SIZE - count); | |
155 | SetPageUptodate(page); | |
156 | } | |
157 | } | |
158 | ||
159 | /* | |
160 | * Write a page synchronously. | |
161 | * Offset is the data offset within the page. | |
162 | */ | |
163 | static int nfs_writepage_sync(struct nfs_open_context *ctx, struct inode *inode, | |
164 | struct page *page, unsigned int offset, unsigned int count, | |
165 | int how) | |
166 | { | |
167 | unsigned int wsize = NFS_SERVER(inode)->wsize; | |
168 | int result, written = 0; | |
169 | struct nfs_write_data *wdata; | |
170 | ||
171 | wdata = nfs_writedata_alloc(); | |
172 | if (!wdata) | |
173 | return -ENOMEM; | |
174 | ||
175 | wdata->flags = how; | |
176 | wdata->cred = ctx->cred; | |
177 | wdata->inode = inode; | |
178 | wdata->args.fh = NFS_FH(inode); | |
179 | wdata->args.context = ctx; | |
180 | wdata->args.pages = &page; | |
181 | wdata->args.stable = NFS_FILE_SYNC; | |
182 | wdata->args.pgbase = offset; | |
183 | wdata->args.count = wsize; | |
184 | wdata->res.fattr = &wdata->fattr; | |
185 | wdata->res.verf = &wdata->verf; | |
186 | ||
187 | dprintk("NFS: nfs_writepage_sync(%s/%Ld %d@%Ld)\n", | |
188 | inode->i_sb->s_id, | |
189 | (long long)NFS_FILEID(inode), | |
190 | count, (long long)(page_offset(page) + offset)); | |
191 | ||
192 | nfs_begin_data_update(inode); | |
193 | do { | |
194 | if (count < wsize) | |
195 | wdata->args.count = count; | |
196 | wdata->args.offset = page_offset(page) + wdata->args.pgbase; | |
197 | ||
198 | result = NFS_PROTO(inode)->write(wdata); | |
199 | ||
200 | if (result < 0) { | |
201 | /* Must mark the page invalid after I/O error */ | |
202 | ClearPageUptodate(page); | |
203 | goto io_error; | |
204 | } | |
205 | if (result < wdata->args.count) | |
206 | printk(KERN_WARNING "NFS: short write, count=%u, result=%d\n", | |
207 | wdata->args.count, result); | |
208 | ||
209 | wdata->args.offset += result; | |
210 | wdata->args.pgbase += result; | |
211 | written += result; | |
212 | count -= result; | |
213 | } while (count); | |
214 | /* Update file length */ | |
215 | nfs_grow_file(page, offset, written); | |
216 | /* Set the PG_uptodate flag? */ | |
217 | nfs_mark_uptodate(page, offset, written); | |
218 | ||
219 | if (PageError(page)) | |
220 | ClearPageError(page); | |
221 | ||
222 | io_error: | |
223 | nfs_end_data_update(inode); | |
224 | nfs_writedata_free(wdata); | |
225 | return written ? written : result; | |
226 | } | |
227 | ||
228 | static int nfs_writepage_async(struct nfs_open_context *ctx, | |
229 | struct inode *inode, struct page *page, | |
230 | unsigned int offset, unsigned int count) | |
231 | { | |
232 | struct nfs_page *req; | |
233 | int status; | |
234 | ||
235 | req = nfs_update_request(ctx, inode, page, offset, count); | |
236 | status = (IS_ERR(req)) ? PTR_ERR(req) : 0; | |
237 | if (status < 0) | |
238 | goto out; | |
239 | /* Update file length */ | |
240 | nfs_grow_file(page, offset, count); | |
241 | /* Set the PG_uptodate flag? */ | |
242 | nfs_mark_uptodate(page, offset, count); | |
243 | nfs_unlock_request(req); | |
244 | out: | |
245 | return status; | |
246 | } | |
247 | ||
248 | static int wb_priority(struct writeback_control *wbc) | |
249 | { | |
250 | if (wbc->for_reclaim) | |
251 | return FLUSH_HIGHPRI; | |
252 | if (wbc->for_kupdate) | |
253 | return FLUSH_LOWPRI; | |
254 | return 0; | |
255 | } | |
256 | ||
257 | /* | |
258 | * Write an mmapped page to the server. | |
259 | */ | |
260 | int nfs_writepage(struct page *page, struct writeback_control *wbc) | |
261 | { | |
262 | struct nfs_open_context *ctx; | |
263 | struct inode *inode = page->mapping->host; | |
264 | unsigned long end_index; | |
265 | unsigned offset = PAGE_CACHE_SIZE; | |
266 | loff_t i_size = i_size_read(inode); | |
267 | int inode_referenced = 0; | |
268 | int priority = wb_priority(wbc); | |
269 | int err; | |
270 | ||
271 | /* | |
272 | * Note: We need to ensure that we have a reference to the inode | |
273 | * if we are to do asynchronous writes. If not, waiting | |
274 | * in nfs_wait_on_request() may deadlock with clear_inode(). | |
275 | * | |
276 | * If igrab() fails here, then it is in any case safe to | |
277 | * call nfs_wb_page(), since there will be no pending writes. | |
278 | */ | |
279 | if (igrab(inode) != 0) | |
280 | inode_referenced = 1; | |
281 | end_index = i_size >> PAGE_CACHE_SHIFT; | |
282 | ||
283 | /* Ensure we've flushed out any previous writes */ | |
284 | nfs_wb_page_priority(inode, page, priority); | |
285 | ||
286 | /* easy case */ | |
287 | if (page->index < end_index) | |
288 | goto do_it; | |
289 | /* things got complicated... */ | |
290 | offset = i_size & (PAGE_CACHE_SIZE-1); | |
291 | ||
292 | /* OK, are we completely out? */ | |
293 | err = 0; /* potential race with truncate - ignore */ | |
294 | if (page->index >= end_index+1 || !offset) | |
295 | goto out; | |
296 | do_it: | |
297 | ctx = nfs_find_open_context(inode, FMODE_WRITE); | |
298 | if (ctx == NULL) { | |
299 | err = -EBADF; | |
300 | goto out; | |
301 | } | |
302 | lock_kernel(); | |
303 | if (!IS_SYNC(inode) && inode_referenced) { | |
304 | err = nfs_writepage_async(ctx, inode, page, 0, offset); | |
305 | if (err >= 0) { | |
306 | err = 0; | |
307 | if (wbc->for_reclaim) | |
308 | nfs_flush_inode(inode, 0, 0, FLUSH_STABLE); | |
309 | } | |
310 | } else { | |
311 | err = nfs_writepage_sync(ctx, inode, page, 0, | |
312 | offset, priority); | |
313 | if (err >= 0) { | |
314 | if (err != offset) | |
315 | redirty_page_for_writepage(wbc, page); | |
316 | err = 0; | |
317 | } | |
318 | } | |
319 | unlock_kernel(); | |
320 | put_nfs_open_context(ctx); | |
321 | out: | |
322 | unlock_page(page); | |
323 | if (inode_referenced) | |
324 | iput(inode); | |
325 | return err; | |
326 | } | |
327 | ||
328 | /* | |
329 | * Note: causes nfs_update_request() to block on the assumption | |
330 | * that the writeback is generated due to memory pressure. | |
331 | */ | |
332 | int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc) | |
333 | { | |
334 | struct backing_dev_info *bdi = mapping->backing_dev_info; | |
335 | struct inode *inode = mapping->host; | |
336 | int err; | |
337 | ||
338 | err = generic_writepages(mapping, wbc); | |
339 | if (err) | |
340 | return err; | |
341 | while (test_and_set_bit(BDI_write_congested, &bdi->state) != 0) { | |
342 | if (wbc->nonblocking) | |
343 | return 0; | |
344 | nfs_wait_on_write_congestion(mapping, 0); | |
345 | } | |
346 | err = nfs_flush_inode(inode, 0, 0, wb_priority(wbc)); | |
347 | if (err < 0) | |
348 | goto out; | |
349 | wbc->nr_to_write -= err; | |
350 | if (!wbc->nonblocking && wbc->sync_mode == WB_SYNC_ALL) { | |
351 | err = nfs_wait_on_requests(inode, 0, 0); | |
352 | if (err < 0) | |
353 | goto out; | |
354 | } | |
355 | err = nfs_commit_inode(inode, 0, 0, wb_priority(wbc)); | |
356 | if (err > 0) { | |
357 | wbc->nr_to_write -= err; | |
358 | err = 0; | |
359 | } | |
360 | out: | |
361 | clear_bit(BDI_write_congested, &bdi->state); | |
362 | wake_up_all(&nfs_write_congestion); | |
363 | return err; | |
364 | } | |
365 | ||
366 | /* | |
367 | * Insert a write request into an inode | |
368 | */ | |
369 | static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req) | |
370 | { | |
371 | struct nfs_inode *nfsi = NFS_I(inode); | |
372 | int error; | |
373 | ||
374 | error = radix_tree_insert(&nfsi->nfs_page_tree, req->wb_index, req); | |
375 | BUG_ON(error == -EEXIST); | |
376 | if (error) | |
377 | return error; | |
378 | if (!nfsi->npages) { | |
379 | igrab(inode); | |
380 | nfs_begin_data_update(inode); | |
381 | if (nfs_have_delegation(inode, FMODE_WRITE)) | |
382 | nfsi->change_attr++; | |
383 | } | |
384 | nfsi->npages++; | |
385 | atomic_inc(&req->wb_count); | |
386 | return 0; | |
387 | } | |
388 | ||
389 | /* | |
390 | * Insert a write request into an inode | |
391 | */ | |
392 | static void nfs_inode_remove_request(struct nfs_page *req) | |
393 | { | |
394 | struct inode *inode = req->wb_context->dentry->d_inode; | |
395 | struct nfs_inode *nfsi = NFS_I(inode); | |
396 | ||
397 | BUG_ON (!NFS_WBACK_BUSY(req)); | |
398 | ||
399 | spin_lock(&nfsi->req_lock); | |
400 | radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index); | |
401 | nfsi->npages--; | |
402 | if (!nfsi->npages) { | |
403 | spin_unlock(&nfsi->req_lock); | |
404 | nfs_end_data_update(inode); | |
405 | iput(inode); | |
406 | } else | |
407 | spin_unlock(&nfsi->req_lock); | |
408 | nfs_clear_request(req); | |
409 | nfs_release_request(req); | |
410 | } | |
411 | ||
412 | /* | |
413 | * Find a request | |
414 | */ | |
415 | static inline struct nfs_page * | |
416 | _nfs_find_request(struct inode *inode, unsigned long index) | |
417 | { | |
418 | struct nfs_inode *nfsi = NFS_I(inode); | |
419 | struct nfs_page *req; | |
420 | ||
421 | req = (struct nfs_page*)radix_tree_lookup(&nfsi->nfs_page_tree, index); | |
422 | if (req) | |
423 | atomic_inc(&req->wb_count); | |
424 | return req; | |
425 | } | |
426 | ||
427 | static struct nfs_page * | |
428 | nfs_find_request(struct inode *inode, unsigned long index) | |
429 | { | |
430 | struct nfs_page *req; | |
431 | struct nfs_inode *nfsi = NFS_I(inode); | |
432 | ||
433 | spin_lock(&nfsi->req_lock); | |
434 | req = _nfs_find_request(inode, index); | |
435 | spin_unlock(&nfsi->req_lock); | |
436 | return req; | |
437 | } | |
438 | ||
439 | /* | |
440 | * Add a request to the inode's dirty list. | |
441 | */ | |
442 | static void | |
443 | nfs_mark_request_dirty(struct nfs_page *req) | |
444 | { | |
445 | struct inode *inode = req->wb_context->dentry->d_inode; | |
446 | struct nfs_inode *nfsi = NFS_I(inode); | |
447 | ||
448 | spin_lock(&nfsi->req_lock); | |
449 | nfs_list_add_request(req, &nfsi->dirty); | |
450 | nfsi->ndirty++; | |
451 | spin_unlock(&nfsi->req_lock); | |
452 | inc_page_state(nr_dirty); | |
453 | mark_inode_dirty(inode); | |
454 | } | |
455 | ||
456 | /* | |
457 | * Check if a request is dirty | |
458 | */ | |
459 | static inline int | |
460 | nfs_dirty_request(struct nfs_page *req) | |
461 | { | |
462 | struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode); | |
463 | return !list_empty(&req->wb_list) && req->wb_list_head == &nfsi->dirty; | |
464 | } | |
465 | ||
466 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) | |
467 | /* | |
468 | * Add a request to the inode's commit list. | |
469 | */ | |
470 | static void | |
471 | nfs_mark_request_commit(struct nfs_page *req) | |
472 | { | |
473 | struct inode *inode = req->wb_context->dentry->d_inode; | |
474 | struct nfs_inode *nfsi = NFS_I(inode); | |
475 | ||
476 | spin_lock(&nfsi->req_lock); | |
477 | nfs_list_add_request(req, &nfsi->commit); | |
478 | nfsi->ncommit++; | |
479 | spin_unlock(&nfsi->req_lock); | |
480 | inc_page_state(nr_unstable); | |
481 | mark_inode_dirty(inode); | |
482 | } | |
483 | #endif | |
484 | ||
485 | /* | |
486 | * Wait for a request to complete. | |
487 | * | |
488 | * Interruptible by signals only if mounted with intr flag. | |
489 | */ | |
490 | static int | |
491 | nfs_wait_on_requests(struct inode *inode, unsigned long idx_start, unsigned int npages) | |
492 | { | |
493 | struct nfs_inode *nfsi = NFS_I(inode); | |
494 | struct nfs_page *req; | |
495 | unsigned long idx_end, next; | |
496 | unsigned int res = 0; | |
497 | int error; | |
498 | ||
499 | if (npages == 0) | |
500 | idx_end = ~0; | |
501 | else | |
502 | idx_end = idx_start + npages - 1; | |
503 | ||
504 | spin_lock(&nfsi->req_lock); | |
505 | next = idx_start; | |
506 | while (radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree, (void **)&req, next, 1, NFS_PAGE_TAG_WRITEBACK)) { | |
507 | if (req->wb_index > idx_end) | |
508 | break; | |
509 | ||
510 | next = req->wb_index + 1; | |
511 | BUG_ON(!NFS_WBACK_BUSY(req)); | |
512 | ||
513 | atomic_inc(&req->wb_count); | |
514 | spin_unlock(&nfsi->req_lock); | |
515 | error = nfs_wait_on_request(req); | |
516 | nfs_release_request(req); | |
517 | if (error < 0) | |
518 | return error; | |
519 | spin_lock(&nfsi->req_lock); | |
520 | res++; | |
521 | } | |
522 | spin_unlock(&nfsi->req_lock); | |
523 | return res; | |
524 | } | |
525 | ||
526 | /* | |
527 | * nfs_scan_dirty - Scan an inode for dirty requests | |
528 | * @inode: NFS inode to scan | |
529 | * @dst: destination list | |
530 | * @idx_start: lower bound of page->index to scan. | |
531 | * @npages: idx_start + npages sets the upper bound to scan. | |
532 | * | |
533 | * Moves requests from the inode's dirty page list. | |
534 | * The requests are *not* checked to ensure that they form a contiguous set. | |
535 | */ | |
536 | static int | |
537 | nfs_scan_dirty(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages) | |
538 | { | |
539 | struct nfs_inode *nfsi = NFS_I(inode); | |
540 | int res; | |
541 | res = nfs_scan_list(&nfsi->dirty, dst, idx_start, npages); | |
542 | nfsi->ndirty -= res; | |
543 | sub_page_state(nr_dirty,res); | |
544 | if ((nfsi->ndirty == 0) != list_empty(&nfsi->dirty)) | |
545 | printk(KERN_ERR "NFS: desynchronized value of nfs_i.ndirty.\n"); | |
546 | return res; | |
547 | } | |
548 | ||
549 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) | |
550 | /* | |
551 | * nfs_scan_commit - Scan an inode for commit requests | |
552 | * @inode: NFS inode to scan | |
553 | * @dst: destination list | |
554 | * @idx_start: lower bound of page->index to scan. | |
555 | * @npages: idx_start + npages sets the upper bound to scan. | |
556 | * | |
557 | * Moves requests from the inode's 'commit' request list. | |
558 | * The requests are *not* checked to ensure that they form a contiguous set. | |
559 | */ | |
560 | static int | |
561 | nfs_scan_commit(struct inode *inode, struct list_head *dst, unsigned long idx_start, unsigned int npages) | |
562 | { | |
563 | struct nfs_inode *nfsi = NFS_I(inode); | |
564 | int res; | |
565 | res = nfs_scan_list(&nfsi->commit, dst, idx_start, npages); | |
566 | nfsi->ncommit -= res; | |
567 | if ((nfsi->ncommit == 0) != list_empty(&nfsi->commit)) | |
568 | printk(KERN_ERR "NFS: desynchronized value of nfs_i.ncommit.\n"); | |
569 | return res; | |
570 | } | |
571 | #endif | |
572 | ||
573 | static int nfs_wait_on_write_congestion(struct address_space *mapping, int intr) | |
574 | { | |
575 | struct backing_dev_info *bdi = mapping->backing_dev_info; | |
576 | DEFINE_WAIT(wait); | |
577 | int ret = 0; | |
578 | ||
579 | might_sleep(); | |
580 | ||
581 | if (!bdi_write_congested(bdi)) | |
582 | return 0; | |
583 | if (intr) { | |
584 | struct rpc_clnt *clnt = NFS_CLIENT(mapping->host); | |
585 | sigset_t oldset; | |
586 | ||
587 | rpc_clnt_sigmask(clnt, &oldset); | |
588 | prepare_to_wait(&nfs_write_congestion, &wait, TASK_INTERRUPTIBLE); | |
589 | if (bdi_write_congested(bdi)) { | |
590 | if (signalled()) | |
591 | ret = -ERESTARTSYS; | |
592 | else | |
593 | schedule(); | |
594 | } | |
595 | rpc_clnt_sigunmask(clnt, &oldset); | |
596 | } else { | |
597 | prepare_to_wait(&nfs_write_congestion, &wait, TASK_UNINTERRUPTIBLE); | |
598 | if (bdi_write_congested(bdi)) | |
599 | schedule(); | |
600 | } | |
601 | finish_wait(&nfs_write_congestion, &wait); | |
602 | return ret; | |
603 | } | |
604 | ||
605 | ||
606 | /* | |
607 | * Try to update any existing write request, or create one if there is none. | |
608 | * In order to match, the request's credentials must match those of | |
609 | * the calling process. | |
610 | * | |
611 | * Note: Should always be called with the Page Lock held! | |
612 | */ | |
613 | static struct nfs_page * nfs_update_request(struct nfs_open_context* ctx, | |
614 | struct inode *inode, struct page *page, | |
615 | unsigned int offset, unsigned int bytes) | |
616 | { | |
617 | struct nfs_server *server = NFS_SERVER(inode); | |
618 | struct nfs_inode *nfsi = NFS_I(inode); | |
619 | struct nfs_page *req, *new = NULL; | |
620 | unsigned long rqend, end; | |
621 | ||
622 | end = offset + bytes; | |
623 | ||
624 | if (nfs_wait_on_write_congestion(page->mapping, server->flags & NFS_MOUNT_INTR)) | |
625 | return ERR_PTR(-ERESTARTSYS); | |
626 | for (;;) { | |
627 | /* Loop over all inode entries and see if we find | |
628 | * A request for the page we wish to update | |
629 | */ | |
630 | spin_lock(&nfsi->req_lock); | |
631 | req = _nfs_find_request(inode, page->index); | |
632 | if (req) { | |
633 | if (!nfs_lock_request_dontget(req)) { | |
634 | int error; | |
635 | spin_unlock(&nfsi->req_lock); | |
636 | error = nfs_wait_on_request(req); | |
637 | nfs_release_request(req); | |
638 | if (error < 0) | |
639 | return ERR_PTR(error); | |
640 | continue; | |
641 | } | |
642 | spin_unlock(&nfsi->req_lock); | |
643 | if (new) | |
644 | nfs_release_request(new); | |
645 | break; | |
646 | } | |
647 | ||
648 | if (new) { | |
649 | int error; | |
650 | nfs_lock_request_dontget(new); | |
651 | error = nfs_inode_add_request(inode, new); | |
652 | if (error) { | |
653 | spin_unlock(&nfsi->req_lock); | |
654 | nfs_unlock_request(new); | |
655 | return ERR_PTR(error); | |
656 | } | |
657 | spin_unlock(&nfsi->req_lock); | |
658 | nfs_mark_request_dirty(new); | |
659 | return new; | |
660 | } | |
661 | spin_unlock(&nfsi->req_lock); | |
662 | ||
663 | new = nfs_create_request(ctx, inode, page, offset, bytes); | |
664 | if (IS_ERR(new)) | |
665 | return new; | |
666 | } | |
667 | ||
668 | /* We have a request for our page. | |
669 | * If the creds don't match, or the | |
670 | * page addresses don't match, | |
671 | * tell the caller to wait on the conflicting | |
672 | * request. | |
673 | */ | |
674 | rqend = req->wb_offset + req->wb_bytes; | |
675 | if (req->wb_context != ctx | |
676 | || req->wb_page != page | |
677 | || !nfs_dirty_request(req) | |
678 | || offset > rqend || end < req->wb_offset) { | |
679 | nfs_unlock_request(req); | |
680 | return ERR_PTR(-EBUSY); | |
681 | } | |
682 | ||
683 | /* Okay, the request matches. Update the region */ | |
684 | if (offset < req->wb_offset) { | |
685 | req->wb_offset = offset; | |
686 | req->wb_pgbase = offset; | |
687 | req->wb_bytes = rqend - req->wb_offset; | |
688 | } | |
689 | ||
690 | if (end > rqend) | |
691 | req->wb_bytes = end - req->wb_offset; | |
692 | ||
693 | return req; | |
694 | } | |
695 | ||
696 | int nfs_flush_incompatible(struct file *file, struct page *page) | |
697 | { | |
698 | struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data; | |
699 | struct inode *inode = page->mapping->host; | |
700 | struct nfs_page *req; | |
701 | int status = 0; | |
702 | /* | |
703 | * Look for a request corresponding to this page. If there | |
704 | * is one, and it belongs to another file, we flush it out | |
705 | * before we try to copy anything into the page. Do this | |
706 | * due to the lack of an ACCESS-type call in NFSv2. | |
707 | * Also do the same if we find a request from an existing | |
708 | * dropped page. | |
709 | */ | |
710 | req = nfs_find_request(inode, page->index); | |
711 | if (req) { | |
712 | if (req->wb_page != page || ctx != req->wb_context) | |
713 | status = nfs_wb_page(inode, page); | |
714 | nfs_release_request(req); | |
715 | } | |
716 | return (status < 0) ? status : 0; | |
717 | } | |
718 | ||
719 | /* | |
720 | * Update and possibly write a cached page of an NFS file. | |
721 | * | |
722 | * XXX: Keep an eye on generic_file_read to make sure it doesn't do bad | |
723 | * things with a page scheduled for an RPC call (e.g. invalidate it). | |
724 | */ | |
725 | int nfs_updatepage(struct file *file, struct page *page, | |
726 | unsigned int offset, unsigned int count) | |
727 | { | |
728 | struct nfs_open_context *ctx = (struct nfs_open_context *)file->private_data; | |
729 | struct dentry *dentry = file->f_dentry; | |
730 | struct inode *inode = page->mapping->host; | |
731 | struct nfs_page *req; | |
732 | int status = 0; | |
733 | ||
734 | dprintk("NFS: nfs_updatepage(%s/%s %d@%Ld)\n", | |
735 | dentry->d_parent->d_name.name, dentry->d_name.name, | |
736 | count, (long long)(page_offset(page) +offset)); | |
737 | ||
738 | if (IS_SYNC(inode)) { | |
739 | status = nfs_writepage_sync(ctx, inode, page, offset, count, 0); | |
740 | if (status > 0) { | |
741 | if (offset == 0 && status == PAGE_CACHE_SIZE) | |
742 | SetPageUptodate(page); | |
743 | return 0; | |
744 | } | |
745 | return status; | |
746 | } | |
747 | ||
748 | /* If we're not using byte range locks, and we know the page | |
749 | * is entirely in cache, it may be more efficient to avoid | |
750 | * fragmenting write requests. | |
751 | */ | |
752 | if (PageUptodate(page) && inode->i_flock == NULL && !(file->f_mode & O_SYNC)) { | |
753 | loff_t end_offs = i_size_read(inode) - 1; | |
754 | unsigned long end_index = end_offs >> PAGE_CACHE_SHIFT; | |
755 | ||
756 | count += offset; | |
757 | offset = 0; | |
758 | if (unlikely(end_offs < 0)) { | |
759 | /* Do nothing */ | |
760 | } else if (page->index == end_index) { | |
761 | unsigned int pglen; | |
762 | pglen = (unsigned int)(end_offs & (PAGE_CACHE_SIZE-1)) + 1; | |
763 | if (count < pglen) | |
764 | count = pglen; | |
765 | } else if (page->index < end_index) | |
766 | count = PAGE_CACHE_SIZE; | |
767 | } | |
768 | ||
769 | /* | |
770 | * Try to find an NFS request corresponding to this page | |
771 | * and update it. | |
772 | * If the existing request cannot be updated, we must flush | |
773 | * it out now. | |
774 | */ | |
775 | do { | |
776 | req = nfs_update_request(ctx, inode, page, offset, count); | |
777 | status = (IS_ERR(req)) ? PTR_ERR(req) : 0; | |
778 | if (status != -EBUSY) | |
779 | break; | |
780 | /* Request could not be updated. Flush it out and try again */ | |
781 | status = nfs_wb_page(inode, page); | |
782 | } while (status >= 0); | |
783 | if (status < 0) | |
784 | goto done; | |
785 | ||
786 | status = 0; | |
787 | ||
788 | /* Update file length */ | |
789 | nfs_grow_file(page, offset, count); | |
790 | /* Set the PG_uptodate flag? */ | |
791 | nfs_mark_uptodate(page, req->wb_pgbase, req->wb_bytes); | |
792 | nfs_unlock_request(req); | |
793 | done: | |
794 | dprintk("NFS: nfs_updatepage returns %d (isize %Ld)\n", | |
795 | status, (long long)i_size_read(inode)); | |
796 | if (status < 0) | |
797 | ClearPageUptodate(page); | |
798 | return status; | |
799 | } | |
800 | ||
801 | static void nfs_writepage_release(struct nfs_page *req) | |
802 | { | |
803 | end_page_writeback(req->wb_page); | |
804 | ||
805 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) | |
806 | if (!PageError(req->wb_page)) { | |
807 | if (NFS_NEED_RESCHED(req)) { | |
808 | nfs_mark_request_dirty(req); | |
809 | goto out; | |
810 | } else if (NFS_NEED_COMMIT(req)) { | |
811 | nfs_mark_request_commit(req); | |
812 | goto out; | |
813 | } | |
814 | } | |
815 | nfs_inode_remove_request(req); | |
816 | ||
817 | out: | |
818 | nfs_clear_commit(req); | |
819 | nfs_clear_reschedule(req); | |
820 | #else | |
821 | nfs_inode_remove_request(req); | |
822 | #endif | |
823 | nfs_clear_page_writeback(req); | |
824 | } | |
825 | ||
826 | static inline int flush_task_priority(int how) | |
827 | { | |
828 | switch (how & (FLUSH_HIGHPRI|FLUSH_LOWPRI)) { | |
829 | case FLUSH_HIGHPRI: | |
830 | return RPC_PRIORITY_HIGH; | |
831 | case FLUSH_LOWPRI: | |
832 | return RPC_PRIORITY_LOW; | |
833 | } | |
834 | return RPC_PRIORITY_NORMAL; | |
835 | } | |
836 | ||
837 | /* | |
838 | * Set up the argument/result storage required for the RPC call. | |
839 | */ | |
840 | static void nfs_write_rpcsetup(struct nfs_page *req, | |
841 | struct nfs_write_data *data, | |
842 | unsigned int count, unsigned int offset, | |
843 | int how) | |
844 | { | |
845 | struct rpc_task *task = &data->task; | |
846 | struct inode *inode; | |
847 | ||
848 | /* Set up the RPC argument and reply structs | |
849 | * NB: take care not to mess about with data->commit et al. */ | |
850 | ||
851 | data->req = req; | |
852 | data->inode = inode = req->wb_context->dentry->d_inode; | |
853 | data->cred = req->wb_context->cred; | |
854 | ||
855 | data->args.fh = NFS_FH(inode); | |
856 | data->args.offset = req_offset(req) + offset; | |
857 | data->args.pgbase = req->wb_pgbase + offset; | |
858 | data->args.pages = data->pagevec; | |
859 | data->args.count = count; | |
860 | data->args.context = req->wb_context; | |
861 | ||
862 | data->res.fattr = &data->fattr; | |
863 | data->res.count = count; | |
864 | data->res.verf = &data->verf; | |
865 | ||
866 | NFS_PROTO(inode)->write_setup(data, how); | |
867 | ||
868 | data->task.tk_priority = flush_task_priority(how); | |
869 | data->task.tk_cookie = (unsigned long)inode; | |
870 | data->task.tk_calldata = data; | |
871 | /* Release requests */ | |
872 | data->task.tk_release = nfs_writedata_release; | |
873 | ||
874 | dprintk("NFS: %4d initiated write call (req %s/%Ld, %u bytes @ offset %Lu)\n", | |
875 | task->tk_pid, | |
876 | inode->i_sb->s_id, | |
877 | (long long)NFS_FILEID(inode), | |
878 | count, | |
879 | (unsigned long long)data->args.offset); | |
880 | } | |
881 | ||
882 | static void nfs_execute_write(struct nfs_write_data *data) | |
883 | { | |
884 | struct rpc_clnt *clnt = NFS_CLIENT(data->inode); | |
885 | sigset_t oldset; | |
886 | ||
887 | rpc_clnt_sigmask(clnt, &oldset); | |
888 | lock_kernel(); | |
889 | rpc_execute(&data->task); | |
890 | unlock_kernel(); | |
891 | rpc_clnt_sigunmask(clnt, &oldset); | |
892 | } | |
893 | ||
894 | /* | |
895 | * Generate multiple small requests to write out a single | |
896 | * contiguous dirty area on one page. | |
897 | */ | |
898 | static int nfs_flush_multi(struct list_head *head, struct inode *inode, int how) | |
899 | { | |
900 | struct nfs_page *req = nfs_list_entry(head->next); | |
901 | struct page *page = req->wb_page; | |
902 | struct nfs_write_data *data; | |
903 | unsigned int wsize = NFS_SERVER(inode)->wsize; | |
904 | unsigned int nbytes, offset; | |
905 | int requests = 0; | |
906 | LIST_HEAD(list); | |
907 | ||
908 | nfs_list_remove_request(req); | |
909 | ||
910 | nbytes = req->wb_bytes; | |
911 | for (;;) { | |
912 | data = nfs_writedata_alloc(); | |
913 | if (!data) | |
914 | goto out_bad; | |
915 | list_add(&data->pages, &list); | |
916 | requests++; | |
917 | if (nbytes <= wsize) | |
918 | break; | |
919 | nbytes -= wsize; | |
920 | } | |
921 | atomic_set(&req->wb_complete, requests); | |
922 | ||
923 | ClearPageError(page); | |
924 | SetPageWriteback(page); | |
925 | offset = 0; | |
926 | nbytes = req->wb_bytes; | |
927 | do { | |
928 | data = list_entry(list.next, struct nfs_write_data, pages); | |
929 | list_del_init(&data->pages); | |
930 | ||
931 | data->pagevec[0] = page; | |
932 | data->complete = nfs_writeback_done_partial; | |
933 | ||
934 | if (nbytes > wsize) { | |
935 | nfs_write_rpcsetup(req, data, wsize, offset, how); | |
936 | offset += wsize; | |
937 | nbytes -= wsize; | |
938 | } else { | |
939 | nfs_write_rpcsetup(req, data, nbytes, offset, how); | |
940 | nbytes = 0; | |
941 | } | |
942 | nfs_execute_write(data); | |
943 | } while (nbytes != 0); | |
944 | ||
945 | return 0; | |
946 | ||
947 | out_bad: | |
948 | while (!list_empty(&list)) { | |
949 | data = list_entry(list.next, struct nfs_write_data, pages); | |
950 | list_del(&data->pages); | |
951 | nfs_writedata_free(data); | |
952 | } | |
953 | nfs_mark_request_dirty(req); | |
954 | nfs_clear_page_writeback(req); | |
955 | return -ENOMEM; | |
956 | } | |
957 | ||
958 | /* | |
959 | * Create an RPC task for the given write request and kick it. | |
960 | * The page must have been locked by the caller. | |
961 | * | |
962 | * It may happen that the page we're passed is not marked dirty. | |
963 | * This is the case if nfs_updatepage detects a conflicting request | |
964 | * that has been written but not committed. | |
965 | */ | |
966 | static int nfs_flush_one(struct list_head *head, struct inode *inode, int how) | |
967 | { | |
968 | struct nfs_page *req; | |
969 | struct page **pages; | |
970 | struct nfs_write_data *data; | |
971 | unsigned int count; | |
972 | ||
973 | if (NFS_SERVER(inode)->wsize < PAGE_CACHE_SIZE) | |
974 | return nfs_flush_multi(head, inode, how); | |
975 | ||
976 | data = nfs_writedata_alloc(); | |
977 | if (!data) | |
978 | goto out_bad; | |
979 | ||
980 | pages = data->pagevec; | |
981 | count = 0; | |
982 | while (!list_empty(head)) { | |
983 | req = nfs_list_entry(head->next); | |
984 | nfs_list_remove_request(req); | |
985 | nfs_list_add_request(req, &data->pages); | |
986 | ClearPageError(req->wb_page); | |
987 | SetPageWriteback(req->wb_page); | |
988 | *pages++ = req->wb_page; | |
989 | count += req->wb_bytes; | |
990 | } | |
991 | req = nfs_list_entry(data->pages.next); | |
992 | ||
993 | data->complete = nfs_writeback_done_full; | |
994 | /* Set up the argument struct */ | |
995 | nfs_write_rpcsetup(req, data, count, 0, how); | |
996 | ||
997 | nfs_execute_write(data); | |
998 | return 0; | |
999 | out_bad: | |
1000 | while (!list_empty(head)) { | |
1001 | struct nfs_page *req = nfs_list_entry(head->next); | |
1002 | nfs_list_remove_request(req); | |
1003 | nfs_mark_request_dirty(req); | |
1004 | nfs_clear_page_writeback(req); | |
1005 | } | |
1006 | return -ENOMEM; | |
1007 | } | |
1008 | ||
1009 | static int | |
1010 | nfs_flush_list(struct list_head *head, int wpages, int how) | |
1011 | { | |
1012 | LIST_HEAD(one_request); | |
1013 | struct nfs_page *req; | |
1014 | int error = 0; | |
1015 | unsigned int pages = 0; | |
1016 | ||
1017 | while (!list_empty(head)) { | |
1018 | pages += nfs_coalesce_requests(head, &one_request, wpages); | |
1019 | req = nfs_list_entry(one_request.next); | |
1020 | error = nfs_flush_one(&one_request, req->wb_context->dentry->d_inode, how); | |
1021 | if (error < 0) | |
1022 | break; | |
1023 | } | |
1024 | if (error >= 0) | |
1025 | return pages; | |
1026 | ||
1027 | while (!list_empty(head)) { | |
1028 | req = nfs_list_entry(head->next); | |
1029 | nfs_list_remove_request(req); | |
1030 | nfs_mark_request_dirty(req); | |
1031 | nfs_clear_page_writeback(req); | |
1032 | } | |
1033 | return error; | |
1034 | } | |
1035 | ||
1036 | /* | |
1037 | * Handle a write reply that flushed part of a page. | |
1038 | */ | |
1039 | static void nfs_writeback_done_partial(struct nfs_write_data *data, int status) | |
1040 | { | |
1041 | struct nfs_page *req = data->req; | |
1042 | struct page *page = req->wb_page; | |
1043 | ||
1044 | dprintk("NFS: write (%s/%Ld %d@%Ld)", | |
1045 | req->wb_context->dentry->d_inode->i_sb->s_id, | |
1046 | (long long)NFS_FILEID(req->wb_context->dentry->d_inode), | |
1047 | req->wb_bytes, | |
1048 | (long long)req_offset(req)); | |
1049 | ||
1050 | if (status < 0) { | |
1051 | ClearPageUptodate(page); | |
1052 | SetPageError(page); | |
1053 | req->wb_context->error = status; | |
1054 | dprintk(", error = %d\n", status); | |
1055 | } else { | |
1056 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) | |
1057 | if (data->verf.committed < NFS_FILE_SYNC) { | |
1058 | if (!NFS_NEED_COMMIT(req)) { | |
1059 | nfs_defer_commit(req); | |
1060 | memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf)); | |
1061 | dprintk(" defer commit\n"); | |
1062 | } else if (memcmp(&req->wb_verf, &data->verf, sizeof(req->wb_verf))) { | |
1063 | nfs_defer_reschedule(req); | |
1064 | dprintk(" server reboot detected\n"); | |
1065 | } | |
1066 | } else | |
1067 | #endif | |
1068 | dprintk(" OK\n"); | |
1069 | } | |
1070 | ||
1071 | if (atomic_dec_and_test(&req->wb_complete)) | |
1072 | nfs_writepage_release(req); | |
1073 | } | |
1074 | ||
1075 | /* | |
1076 | * Handle a write reply that flushes a whole page. | |
1077 | * | |
1078 | * FIXME: There is an inherent race with invalidate_inode_pages and | |
1079 | * writebacks since the page->count is kept > 1 for as long | |
1080 | * as the page has a write request pending. | |
1081 | */ | |
1082 | static void nfs_writeback_done_full(struct nfs_write_data *data, int status) | |
1083 | { | |
1084 | struct nfs_page *req; | |
1085 | struct page *page; | |
1086 | ||
1087 | /* Update attributes as result of writeback. */ | |
1088 | while (!list_empty(&data->pages)) { | |
1089 | req = nfs_list_entry(data->pages.next); | |
1090 | nfs_list_remove_request(req); | |
1091 | page = req->wb_page; | |
1092 | ||
1093 | dprintk("NFS: write (%s/%Ld %d@%Ld)", | |
1094 | req->wb_context->dentry->d_inode->i_sb->s_id, | |
1095 | (long long)NFS_FILEID(req->wb_context->dentry->d_inode), | |
1096 | req->wb_bytes, | |
1097 | (long long)req_offset(req)); | |
1098 | ||
1099 | if (status < 0) { | |
1100 | ClearPageUptodate(page); | |
1101 | SetPageError(page); | |
1102 | req->wb_context->error = status; | |
1103 | end_page_writeback(page); | |
1104 | nfs_inode_remove_request(req); | |
1105 | dprintk(", error = %d\n", status); | |
1106 | goto next; | |
1107 | } | |
1108 | end_page_writeback(page); | |
1109 | ||
1110 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) | |
1111 | if (data->args.stable != NFS_UNSTABLE || data->verf.committed == NFS_FILE_SYNC) { | |
1112 | nfs_inode_remove_request(req); | |
1113 | dprintk(" OK\n"); | |
1114 | goto next; | |
1115 | } | |
1116 | memcpy(&req->wb_verf, &data->verf, sizeof(req->wb_verf)); | |
1117 | nfs_mark_request_commit(req); | |
1118 | dprintk(" marked for commit\n"); | |
1119 | #else | |
1120 | nfs_inode_remove_request(req); | |
1121 | #endif | |
1122 | next: | |
1123 | nfs_clear_page_writeback(req); | |
1124 | } | |
1125 | } | |
1126 | ||
1127 | /* | |
1128 | * This function is called when the WRITE call is complete. | |
1129 | */ | |
1130 | void nfs_writeback_done(struct rpc_task *task) | |
1131 | { | |
1132 | struct nfs_write_data *data = (struct nfs_write_data *) task->tk_calldata; | |
1133 | struct nfs_writeargs *argp = &data->args; | |
1134 | struct nfs_writeres *resp = &data->res; | |
1135 | ||
1136 | dprintk("NFS: %4d nfs_writeback_done (status %d)\n", | |
1137 | task->tk_pid, task->tk_status); | |
1138 | ||
1139 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) | |
1140 | if (resp->verf->committed < argp->stable && task->tk_status >= 0) { | |
1141 | /* We tried a write call, but the server did not | |
1142 | * commit data to stable storage even though we | |
1143 | * requested it. | |
1144 | * Note: There is a known bug in Tru64 < 5.0 in which | |
1145 | * the server reports NFS_DATA_SYNC, but performs | |
1146 | * NFS_FILE_SYNC. We therefore implement this checking | |
1147 | * as a dprintk() in order to avoid filling syslog. | |
1148 | */ | |
1149 | static unsigned long complain; | |
1150 | ||
1151 | if (time_before(complain, jiffies)) { | |
1152 | dprintk("NFS: faulty NFS server %s:" | |
1153 | " (committed = %d) != (stable = %d)\n", | |
1154 | NFS_SERVER(data->inode)->hostname, | |
1155 | resp->verf->committed, argp->stable); | |
1156 | complain = jiffies + 300 * HZ; | |
1157 | } | |
1158 | } | |
1159 | #endif | |
1160 | /* Is this a short write? */ | |
1161 | if (task->tk_status >= 0 && resp->count < argp->count) { | |
1162 | static unsigned long complain; | |
1163 | ||
1164 | /* Has the server at least made some progress? */ | |
1165 | if (resp->count != 0) { | |
1166 | /* Was this an NFSv2 write or an NFSv3 stable write? */ | |
1167 | if (resp->verf->committed != NFS_UNSTABLE) { | |
1168 | /* Resend from where the server left off */ | |
1169 | argp->offset += resp->count; | |
1170 | argp->pgbase += resp->count; | |
1171 | argp->count -= resp->count; | |
1172 | } else { | |
1173 | /* Resend as a stable write in order to avoid | |
1174 | * headaches in the case of a server crash. | |
1175 | */ | |
1176 | argp->stable = NFS_FILE_SYNC; | |
1177 | } | |
1178 | rpc_restart_call(task); | |
1179 | return; | |
1180 | } | |
1181 | if (time_before(complain, jiffies)) { | |
1182 | printk(KERN_WARNING | |
1183 | "NFS: Server wrote zero bytes, expected %u.\n", | |
1184 | argp->count); | |
1185 | complain = jiffies + 300 * HZ; | |
1186 | } | |
1187 | /* Can't do anything about it except throw an error. */ | |
1188 | task->tk_status = -EIO; | |
1189 | } | |
1190 | ||
1191 | /* | |
1192 | * Process the nfs_page list | |
1193 | */ | |
1194 | data->complete(data, task->tk_status); | |
1195 | } | |
1196 | ||
1197 | ||
1198 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) | |
1199 | static void nfs_commit_release(struct rpc_task *task) | |
1200 | { | |
1201 | struct nfs_write_data *wdata = (struct nfs_write_data *)task->tk_calldata; | |
1202 | nfs_commit_free(wdata); | |
1203 | } | |
1204 | ||
1205 | /* | |
1206 | * Set up the argument/result storage required for the RPC call. | |
1207 | */ | |
1208 | static void nfs_commit_rpcsetup(struct list_head *head, | |
1209 | struct nfs_write_data *data, int how) | |
1210 | { | |
1211 | struct rpc_task *task = &data->task; | |
1212 | struct nfs_page *first, *last; | |
1213 | struct inode *inode; | |
1214 | loff_t start, end, len; | |
1215 | ||
1216 | /* Set up the RPC argument and reply structs | |
1217 | * NB: take care not to mess about with data->commit et al. */ | |
1218 | ||
1219 | list_splice_init(head, &data->pages); | |
1220 | first = nfs_list_entry(data->pages.next); | |
1221 | last = nfs_list_entry(data->pages.prev); | |
1222 | inode = first->wb_context->dentry->d_inode; | |
1223 | ||
1224 | /* | |
1225 | * Determine the offset range of requests in the COMMIT call. | |
1226 | * We rely on the fact that data->pages is an ordered list... | |
1227 | */ | |
1228 | start = req_offset(first); | |
1229 | end = req_offset(last) + last->wb_bytes; | |
1230 | len = end - start; | |
1231 | /* If 'len' is not a 32-bit quantity, pass '0' in the COMMIT call */ | |
1232 | if (end >= i_size_read(inode) || len < 0 || len > (~((u32)0) >> 1)) | |
1233 | len = 0; | |
1234 | ||
1235 | data->inode = inode; | |
1236 | data->cred = first->wb_context->cred; | |
1237 | ||
1238 | data->args.fh = NFS_FH(data->inode); | |
1239 | data->args.offset = start; | |
1240 | data->args.count = len; | |
1241 | data->res.count = len; | |
1242 | data->res.fattr = &data->fattr; | |
1243 | data->res.verf = &data->verf; | |
1244 | ||
1245 | NFS_PROTO(inode)->commit_setup(data, how); | |
1246 | ||
1247 | data->task.tk_priority = flush_task_priority(how); | |
1248 | data->task.tk_cookie = (unsigned long)inode; | |
1249 | data->task.tk_calldata = data; | |
1250 | /* Release requests */ | |
1251 | data->task.tk_release = nfs_commit_release; | |
1252 | ||
1253 | dprintk("NFS: %4d initiated commit call\n", task->tk_pid); | |
1254 | } | |
1255 | ||
1256 | /* | |
1257 | * Commit dirty pages | |
1258 | */ | |
1259 | static int | |
1260 | nfs_commit_list(struct list_head *head, int how) | |
1261 | { | |
1262 | struct nfs_write_data *data; | |
1263 | struct nfs_page *req; | |
1264 | ||
1265 | data = nfs_commit_alloc(); | |
1266 | ||
1267 | if (!data) | |
1268 | goto out_bad; | |
1269 | ||
1270 | /* Set up the argument struct */ | |
1271 | nfs_commit_rpcsetup(head, data, how); | |
1272 | ||
1273 | nfs_execute_write(data); | |
1274 | return 0; | |
1275 | out_bad: | |
1276 | while (!list_empty(head)) { | |
1277 | req = nfs_list_entry(head->next); | |
1278 | nfs_list_remove_request(req); | |
1279 | nfs_mark_request_commit(req); | |
1280 | nfs_clear_page_writeback(req); | |
1281 | } | |
1282 | return -ENOMEM; | |
1283 | } | |
1284 | ||
1285 | /* | |
1286 | * COMMIT call returned | |
1287 | */ | |
1288 | void | |
1289 | nfs_commit_done(struct rpc_task *task) | |
1290 | { | |
1291 | struct nfs_write_data *data = (struct nfs_write_data *)task->tk_calldata; | |
1292 | struct nfs_page *req; | |
1293 | int res = 0; | |
1294 | ||
1295 | dprintk("NFS: %4d nfs_commit_done (status %d)\n", | |
1296 | task->tk_pid, task->tk_status); | |
1297 | ||
1298 | while (!list_empty(&data->pages)) { | |
1299 | req = nfs_list_entry(data->pages.next); | |
1300 | nfs_list_remove_request(req); | |
1301 | ||
1302 | dprintk("NFS: commit (%s/%Ld %d@%Ld)", | |
1303 | req->wb_context->dentry->d_inode->i_sb->s_id, | |
1304 | (long long)NFS_FILEID(req->wb_context->dentry->d_inode), | |
1305 | req->wb_bytes, | |
1306 | (long long)req_offset(req)); | |
1307 | if (task->tk_status < 0) { | |
1308 | req->wb_context->error = task->tk_status; | |
1309 | nfs_inode_remove_request(req); | |
1310 | dprintk(", error = %d\n", task->tk_status); | |
1311 | goto next; | |
1312 | } | |
1313 | ||
1314 | /* Okay, COMMIT succeeded, apparently. Check the verifier | |
1315 | * returned by the server against all stored verfs. */ | |
1316 | if (!memcmp(req->wb_verf.verifier, data->verf.verifier, sizeof(data->verf.verifier))) { | |
1317 | /* We have a match */ | |
1318 | nfs_inode_remove_request(req); | |
1319 | dprintk(" OK\n"); | |
1320 | goto next; | |
1321 | } | |
1322 | /* We have a mismatch. Write the page again */ | |
1323 | dprintk(" mismatch\n"); | |
1324 | nfs_mark_request_dirty(req); | |
1325 | next: | |
1326 | nfs_clear_page_writeback(req); | |
1327 | res++; | |
1328 | } | |
1329 | sub_page_state(nr_unstable,res); | |
1330 | } | |
1331 | #endif | |
1332 | ||
1333 | static int nfs_flush_inode(struct inode *inode, unsigned long idx_start, | |
1334 | unsigned int npages, int how) | |
1335 | { | |
1336 | struct nfs_inode *nfsi = NFS_I(inode); | |
1337 | LIST_HEAD(head); | |
1338 | int res, | |
1339 | error = 0; | |
1340 | ||
1341 | spin_lock(&nfsi->req_lock); | |
1342 | res = nfs_scan_dirty(inode, &head, idx_start, npages); | |
1343 | spin_unlock(&nfsi->req_lock); | |
1344 | if (res) { | |
1345 | struct nfs_server *server = NFS_SERVER(inode); | |
1346 | ||
1347 | /* For single writes, FLUSH_STABLE is more efficient */ | |
1348 | if (res == nfsi->npages && nfsi->npages <= server->wpages) { | |
1349 | if (res > 1 || nfs_list_entry(head.next)->wb_bytes <= server->wsize) | |
1350 | how |= FLUSH_STABLE; | |
1351 | } | |
1352 | error = nfs_flush_list(&head, server->wpages, how); | |
1353 | } | |
1354 | if (error < 0) | |
1355 | return error; | |
1356 | return res; | |
1357 | } | |
1358 | ||
1359 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) | |
1360 | int nfs_commit_inode(struct inode *inode, unsigned long idx_start, | |
1361 | unsigned int npages, int how) | |
1362 | { | |
1363 | struct nfs_inode *nfsi = NFS_I(inode); | |
1364 | LIST_HEAD(head); | |
1365 | int res, | |
1366 | error = 0; | |
1367 | ||
1368 | spin_lock(&nfsi->req_lock); | |
1369 | res = nfs_scan_commit(inode, &head, idx_start, npages); | |
1370 | if (res) { | |
1371 | res += nfs_scan_commit(inode, &head, 0, 0); | |
1372 | spin_unlock(&nfsi->req_lock); | |
1373 | error = nfs_commit_list(&head, how); | |
1374 | } else | |
1375 | spin_unlock(&nfsi->req_lock); | |
1376 | if (error < 0) | |
1377 | return error; | |
1378 | return res; | |
1379 | } | |
1380 | #endif | |
1381 | ||
1382 | int nfs_sync_inode(struct inode *inode, unsigned long idx_start, | |
1383 | unsigned int npages, int how) | |
1384 | { | |
1385 | int error, | |
1386 | wait; | |
1387 | ||
1388 | wait = how & FLUSH_WAIT; | |
1389 | how &= ~FLUSH_WAIT; | |
1390 | ||
1391 | do { | |
1392 | error = 0; | |
1393 | if (wait) | |
1394 | error = nfs_wait_on_requests(inode, idx_start, npages); | |
1395 | if (error == 0) | |
1396 | error = nfs_flush_inode(inode, idx_start, npages, how); | |
1397 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) | |
1398 | if (error == 0) | |
1399 | error = nfs_commit_inode(inode, idx_start, npages, how); | |
1400 | #endif | |
1401 | } while (error > 0); | |
1402 | return error; | |
1403 | } | |
1404 | ||
1405 | int nfs_init_writepagecache(void) | |
1406 | { | |
1407 | nfs_wdata_cachep = kmem_cache_create("nfs_write_data", | |
1408 | sizeof(struct nfs_write_data), | |
1409 | 0, SLAB_HWCACHE_ALIGN, | |
1410 | NULL, NULL); | |
1411 | if (nfs_wdata_cachep == NULL) | |
1412 | return -ENOMEM; | |
1413 | ||
1414 | nfs_wdata_mempool = mempool_create(MIN_POOL_WRITE, | |
1415 | mempool_alloc_slab, | |
1416 | mempool_free_slab, | |
1417 | nfs_wdata_cachep); | |
1418 | if (nfs_wdata_mempool == NULL) | |
1419 | return -ENOMEM; | |
1420 | ||
1421 | nfs_commit_mempool = mempool_create(MIN_POOL_COMMIT, | |
1422 | mempool_alloc_slab, | |
1423 | mempool_free_slab, | |
1424 | nfs_wdata_cachep); | |
1425 | if (nfs_commit_mempool == NULL) | |
1426 | return -ENOMEM; | |
1427 | ||
1428 | return 0; | |
1429 | } | |
1430 | ||
1431 | void nfs_destroy_writepagecache(void) | |
1432 | { | |
1433 | mempool_destroy(nfs_commit_mempool); | |
1434 | mempool_destroy(nfs_wdata_mempool); | |
1435 | if (kmem_cache_destroy(nfs_wdata_cachep)) | |
1436 | printk(KERN_INFO "nfs_write_data: not all structures were freed\n"); | |
1437 | } | |
1438 |