Commit | Line | Data |
---|---|---|
124e68e7 SW |
1 | #include "ceph_debug.h" |
2 | ||
3 | #include <linux/sched.h> | |
4 | #include <linux/file.h> | |
5 | #include <linux/namei.h> | |
6 | #include <linux/writeback.h> | |
7 | ||
8 | #include "super.h" | |
9 | #include "mds_client.h" | |
10 | ||
11 | /* | |
12 | * Ceph file operations | |
13 | * | |
14 | * Implement basic open/close functionality, and implement | |
15 | * read/write. | |
16 | * | |
17 | * We implement three modes of file I/O: | |
18 | * - buffered uses the generic_file_aio_{read,write} helpers | |
19 | * | |
20 | * - synchronous is used when there is multi-client read/write | |
21 | * sharing, avoids the page cache, and synchronously waits for an | |
22 | * ack from the OSD. | |
23 | * | |
24 | * - direct io takes the variant of the sync path that references | |
25 | * user pages directly. | |
26 | * | |
27 | * fsync() flushes and waits on dirty pages, but just queues metadata | |
28 | * for writeback: since the MDS can recover size and mtime there is no | |
29 | * need to wait for MDS acknowledgement. | |
30 | */ | |
31 | ||
32 | ||
33 | /* | |
34 | * Prepare an open request. Preallocate ceph_cap to avoid an | |
35 | * inopportune ENOMEM later. | |
36 | */ | |
37 | static struct ceph_mds_request * | |
38 | prepare_open_request(struct super_block *sb, int flags, int create_mode) | |
39 | { | |
40 | struct ceph_client *client = ceph_sb_to_client(sb); | |
41 | struct ceph_mds_client *mdsc = &client->mdsc; | |
42 | struct ceph_mds_request *req; | |
43 | int want_auth = USE_ANY_MDS; | |
44 | int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN; | |
45 | ||
46 | if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC)) | |
47 | want_auth = USE_AUTH_MDS; | |
48 | ||
49 | req = ceph_mdsc_create_request(mdsc, op, want_auth); | |
50 | if (IS_ERR(req)) | |
51 | goto out; | |
52 | req->r_fmode = ceph_flags_to_mode(flags); | |
53 | req->r_args.open.flags = cpu_to_le32(flags); | |
54 | req->r_args.open.mode = cpu_to_le32(create_mode); | |
55 | req->r_args.open.preferred = -1; | |
56 | out: | |
57 | return req; | |
58 | } | |
59 | ||
60 | /* | |
61 | * initialize private struct file data. | |
62 | * if we fail, clean up by dropping fmode reference on the ceph_inode | |
63 | */ | |
64 | static int ceph_init_file(struct inode *inode, struct file *file, int fmode) | |
65 | { | |
66 | struct ceph_file_info *cf; | |
67 | int ret = 0; | |
68 | ||
69 | switch (inode->i_mode & S_IFMT) { | |
70 | case S_IFREG: | |
71 | case S_IFDIR: | |
72 | dout("init_file %p %p 0%o (regular)\n", inode, file, | |
73 | inode->i_mode); | |
74 | cf = kmem_cache_alloc(ceph_file_cachep, GFP_NOFS | __GFP_ZERO); | |
75 | if (cf == NULL) { | |
76 | ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */ | |
77 | return -ENOMEM; | |
78 | } | |
79 | cf->fmode = fmode; | |
80 | cf->next_offset = 2; | |
81 | file->private_data = cf; | |
82 | BUG_ON(inode->i_fop->release != ceph_release); | |
83 | break; | |
84 | ||
85 | case S_IFLNK: | |
86 | dout("init_file %p %p 0%o (symlink)\n", inode, file, | |
87 | inode->i_mode); | |
88 | ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */ | |
89 | break; | |
90 | ||
91 | default: | |
92 | dout("init_file %p %p 0%o (special)\n", inode, file, | |
93 | inode->i_mode); | |
94 | /* | |
95 | * we need to drop the open ref now, since we don't | |
96 | * have .release set to ceph_release. | |
97 | */ | |
98 | ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */ | |
99 | BUG_ON(inode->i_fop->release == ceph_release); | |
100 | ||
101 | /* call the proper open fop */ | |
102 | ret = inode->i_fop->open(inode, file); | |
103 | } | |
104 | return ret; | |
105 | } | |
106 | ||
107 | /* | |
108 | * If the filp already has private_data, that means the file was | |
109 | * already opened by intent during lookup, and we do nothing. | |
110 | * | |
111 | * If we already have the requisite capabilities, we can satisfy | |
112 | * the open request locally (no need to request new caps from the | |
113 | * MDS). We do, however, need to inform the MDS (asynchronously) | |
114 | * if our wanted caps set expands. | |
115 | */ | |
116 | int ceph_open(struct inode *inode, struct file *file) | |
117 | { | |
118 | struct ceph_inode_info *ci = ceph_inode(inode); | |
119 | struct ceph_client *client = ceph_sb_to_client(inode->i_sb); | |
120 | struct ceph_mds_client *mdsc = &client->mdsc; | |
121 | struct ceph_mds_request *req; | |
122 | struct ceph_file_info *cf = file->private_data; | |
123 | struct inode *parent_inode = file->f_dentry->d_parent->d_inode; | |
124 | int err; | |
125 | int flags, fmode, wanted; | |
126 | ||
127 | if (cf) { | |
128 | dout("open file %p is already opened\n", file); | |
129 | return 0; | |
130 | } | |
131 | ||
132 | /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */ | |
133 | flags = file->f_flags & ~(O_CREAT|O_EXCL); | |
134 | if (S_ISDIR(inode->i_mode)) | |
135 | flags = O_DIRECTORY; /* mds likes to know */ | |
136 | ||
137 | dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode, | |
138 | ceph_vinop(inode), file, flags, file->f_flags); | |
139 | fmode = ceph_flags_to_mode(flags); | |
140 | wanted = ceph_caps_for_mode(fmode); | |
141 | ||
142 | /* snapped files are read-only */ | |
143 | if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE)) | |
144 | return -EROFS; | |
145 | ||
146 | /* trivially open snapdir */ | |
147 | if (ceph_snap(inode) == CEPH_SNAPDIR) { | |
148 | spin_lock(&inode->i_lock); | |
149 | __ceph_get_fmode(ci, fmode); | |
150 | spin_unlock(&inode->i_lock); | |
151 | return ceph_init_file(inode, file, fmode); | |
152 | } | |
153 | ||
154 | /* | |
155 | * No need to block if we have any caps. Update wanted set | |
156 | * asynchronously. | |
157 | */ | |
158 | spin_lock(&inode->i_lock); | |
159 | if (__ceph_is_any_real_caps(ci)) { | |
160 | int mds_wanted = __ceph_caps_mds_wanted(ci); | |
161 | int issued = __ceph_caps_issued(ci, NULL); | |
162 | ||
163 | dout("open %p fmode %d want %s issued %s using existing\n", | |
164 | inode, fmode, ceph_cap_string(wanted), | |
165 | ceph_cap_string(issued)); | |
166 | __ceph_get_fmode(ci, fmode); | |
167 | spin_unlock(&inode->i_lock); | |
168 | ||
169 | /* adjust wanted? */ | |
170 | if ((issued & wanted) != wanted && | |
171 | (mds_wanted & wanted) != wanted && | |
172 | ceph_snap(inode) != CEPH_SNAPDIR) | |
173 | ceph_check_caps(ci, 0, NULL); | |
174 | ||
175 | return ceph_init_file(inode, file, fmode); | |
176 | } else if (ceph_snap(inode) != CEPH_NOSNAP && | |
177 | (ci->i_snap_caps & wanted) == wanted) { | |
178 | __ceph_get_fmode(ci, fmode); | |
179 | spin_unlock(&inode->i_lock); | |
180 | return ceph_init_file(inode, file, fmode); | |
181 | } | |
182 | spin_unlock(&inode->i_lock); | |
183 | ||
184 | dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted)); | |
185 | req = prepare_open_request(inode->i_sb, flags, 0); | |
186 | if (IS_ERR(req)) { | |
187 | err = PTR_ERR(req); | |
188 | goto out; | |
189 | } | |
190 | req->r_inode = igrab(inode); | |
191 | req->r_num_caps = 1; | |
192 | err = ceph_mdsc_do_request(mdsc, parent_inode, req); | |
193 | if (!err) | |
194 | err = ceph_init_file(inode, file, req->r_fmode); | |
195 | ceph_mdsc_put_request(req); | |
196 | dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode)); | |
197 | out: | |
198 | return err; | |
199 | } | |
200 | ||
201 | ||
202 | /* | |
203 | * Do a lookup + open with a single request. | |
204 | * | |
205 | * If this succeeds, but some subsequent check in the vfs | |
206 | * may_open() fails, the struct *file gets cleaned up (i.e. | |
207 | * ceph_release gets called). So fear not! | |
208 | */ | |
209 | /* | |
210 | * flags | |
211 | * path_lookup_open -> LOOKUP_OPEN | |
212 | * path_lookup_create -> LOOKUP_OPEN|LOOKUP_CREATE | |
213 | */ | |
214 | struct dentry *ceph_lookup_open(struct inode *dir, struct dentry *dentry, | |
215 | struct nameidata *nd, int mode, | |
216 | int locked_dir) | |
217 | { | |
218 | struct ceph_client *client = ceph_sb_to_client(dir->i_sb); | |
219 | struct ceph_mds_client *mdsc = &client->mdsc; | |
220 | struct file *file = nd->intent.open.file; | |
221 | struct inode *parent_inode = get_dentry_parent_inode(file->f_dentry); | |
222 | struct ceph_mds_request *req; | |
223 | int err; | |
224 | int flags = nd->intent.open.flags - 1; /* silly vfs! */ | |
225 | ||
226 | dout("ceph_lookup_open dentry %p '%.*s' flags %d mode 0%o\n", | |
227 | dentry, dentry->d_name.len, dentry->d_name.name, flags, mode); | |
228 | ||
229 | /* do the open */ | |
230 | req = prepare_open_request(dir->i_sb, flags, mode); | |
231 | if (IS_ERR(req)) | |
232 | return ERR_PTR(PTR_ERR(req)); | |
233 | req->r_dentry = dget(dentry); | |
234 | req->r_num_caps = 2; | |
235 | if (flags & O_CREAT) { | |
236 | req->r_dentry_drop = CEPH_CAP_FILE_SHARED; | |
237 | req->r_dentry_unless = CEPH_CAP_FILE_EXCL; | |
238 | } | |
239 | req->r_locked_dir = dir; /* caller holds dir->i_mutex */ | |
240 | err = ceph_mdsc_do_request(mdsc, parent_inode, req); | |
241 | dentry = ceph_finish_lookup(req, dentry, err); | |
242 | if (!err && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry) | |
243 | err = ceph_handle_notrace_create(dir, dentry); | |
244 | if (!err) | |
245 | err = ceph_init_file(req->r_dentry->d_inode, file, | |
246 | req->r_fmode); | |
247 | ceph_mdsc_put_request(req); | |
248 | dout("ceph_lookup_open result=%p\n", dentry); | |
249 | return dentry; | |
250 | } | |
251 | ||
252 | int ceph_release(struct inode *inode, struct file *file) | |
253 | { | |
254 | struct ceph_inode_info *ci = ceph_inode(inode); | |
255 | struct ceph_file_info *cf = file->private_data; | |
256 | ||
257 | dout("release inode %p file %p\n", inode, file); | |
258 | ceph_put_fmode(ci, cf->fmode); | |
259 | if (cf->last_readdir) | |
260 | ceph_mdsc_put_request(cf->last_readdir); | |
261 | kfree(cf->last_name); | |
262 | kfree(cf->dir_info); | |
263 | dput(cf->dentry); | |
264 | kmem_cache_free(ceph_file_cachep, cf); | |
265 | return 0; | |
266 | } | |
267 | ||
268 | /* | |
269 | * build a vector of user pages | |
270 | */ | |
271 | static struct page **get_direct_page_vector(const char __user *data, | |
272 | int num_pages, | |
273 | loff_t off, size_t len) | |
274 | { | |
275 | struct page **pages; | |
276 | int rc; | |
277 | ||
278 | pages = kmalloc(sizeof(*pages) * num_pages, GFP_NOFS); | |
279 | if (!pages) | |
280 | return ERR_PTR(-ENOMEM); | |
281 | ||
282 | down_read(¤t->mm->mmap_sem); | |
283 | rc = get_user_pages(current, current->mm, (unsigned long)data, | |
284 | num_pages, 0, 0, pages, NULL); | |
285 | up_read(¤t->mm->mmap_sem); | |
286 | if (rc < 0) | |
287 | goto fail; | |
288 | return pages; | |
289 | ||
290 | fail: | |
291 | kfree(pages); | |
292 | return ERR_PTR(rc); | |
293 | } | |
294 | ||
295 | static void put_page_vector(struct page **pages, int num_pages) | |
296 | { | |
297 | int i; | |
298 | ||
299 | for (i = 0; i < num_pages; i++) | |
300 | put_page(pages[i]); | |
301 | kfree(pages); | |
302 | } | |
303 | ||
304 | void ceph_release_page_vector(struct page **pages, int num_pages) | |
305 | { | |
306 | int i; | |
307 | ||
308 | for (i = 0; i < num_pages; i++) | |
309 | __free_pages(pages[i], 0); | |
310 | kfree(pages); | |
311 | } | |
312 | ||
313 | /* | |
314 | * allocate a vector new pages | |
315 | */ | |
316 | static struct page **alloc_page_vector(int num_pages) | |
317 | { | |
318 | struct page **pages; | |
319 | int i; | |
320 | ||
321 | pages = kmalloc(sizeof(*pages) * num_pages, GFP_NOFS); | |
322 | if (!pages) | |
323 | return ERR_PTR(-ENOMEM); | |
324 | for (i = 0; i < num_pages; i++) { | |
325 | pages[i] = alloc_page(GFP_NOFS); | |
326 | if (pages[i] == NULL) { | |
327 | ceph_release_page_vector(pages, i); | |
328 | return ERR_PTR(-ENOMEM); | |
329 | } | |
330 | } | |
331 | return pages; | |
332 | } | |
333 | ||
334 | /* | |
335 | * copy user data into a page vector | |
336 | */ | |
337 | static int copy_user_to_page_vector(struct page **pages, | |
338 | const char __user *data, | |
339 | loff_t off, size_t len) | |
340 | { | |
341 | int i = 0; | |
342 | int po = off & ~PAGE_CACHE_MASK; | |
343 | int left = len; | |
344 | int l, bad; | |
345 | ||
346 | while (left > 0) { | |
347 | l = min_t(int, PAGE_CACHE_SIZE-po, left); | |
348 | bad = copy_from_user(page_address(pages[i]) + po, data, l); | |
349 | if (bad == l) | |
350 | return -EFAULT; | |
351 | data += l - bad; | |
352 | left -= l - bad; | |
353 | if (po) { | |
354 | po += l - bad; | |
355 | if (po == PAGE_CACHE_SIZE) | |
356 | po = 0; | |
357 | } | |
358 | } | |
359 | return len; | |
360 | } | |
361 | ||
362 | /* | |
363 | * copy user data from a page vector into a user pointer | |
364 | */ | |
365 | static int copy_page_vector_to_user(struct page **pages, char __user *data, | |
366 | loff_t off, size_t len) | |
367 | { | |
368 | int i = 0; | |
369 | int po = off & ~PAGE_CACHE_MASK; | |
370 | int left = len; | |
371 | int l, bad; | |
372 | ||
373 | while (left > 0) { | |
374 | l = min_t(int, left, PAGE_CACHE_SIZE-po); | |
375 | bad = copy_to_user(data, page_address(pages[i]) + po, l); | |
376 | if (bad == l) | |
377 | return -EFAULT; | |
378 | data += l - bad; | |
379 | left -= l - bad; | |
380 | if (po) { | |
381 | po += l - bad; | |
382 | if (po == PAGE_CACHE_SIZE) | |
383 | po = 0; | |
384 | } | |
385 | i++; | |
386 | } | |
387 | return len; | |
388 | } | |
389 | ||
390 | /* | |
391 | * Zero an extent within a page vector. Offset is relative to the | |
392 | * start of the first page. | |
393 | */ | |
394 | static void zero_page_vector_range(int off, int len, struct page **pages) | |
395 | { | |
396 | int i = off >> PAGE_CACHE_SHIFT; | |
397 | ||
398 | dout("zero_page_vector_page %u~%u\n", off, len); | |
399 | BUG_ON(len < PAGE_CACHE_SIZE); | |
400 | ||
401 | /* leading partial page? */ | |
402 | if (off & ~PAGE_CACHE_MASK) { | |
403 | dout("zeroing %d %p head from %d\n", i, pages[i], | |
404 | (int)(off & ~PAGE_CACHE_MASK)); | |
405 | zero_user_segment(pages[i], off & ~PAGE_CACHE_MASK, | |
406 | PAGE_CACHE_SIZE); | |
407 | off += PAGE_CACHE_SIZE; | |
408 | off &= PAGE_CACHE_MASK; | |
409 | i++; | |
410 | } | |
411 | while (len >= PAGE_CACHE_SIZE) { | |
412 | dout("zeroing %d %p\n", i, pages[i]); | |
413 | zero_user_segment(pages[i], 0, PAGE_CACHE_SIZE); | |
414 | off += PAGE_CACHE_SIZE; | |
415 | len -= PAGE_CACHE_SIZE; | |
416 | i++; | |
417 | } | |
418 | /* trailing partial page? */ | |
419 | if (len) { | |
420 | dout("zeroing %d %p tail to %d\n", i, pages[i], (int)len); | |
421 | zero_user_segment(pages[i], 0, len); | |
422 | } | |
423 | } | |
424 | ||
425 | ||
426 | /* | |
427 | * Read a range of bytes striped over one or more objects. Iterate over | |
428 | * objects we stripe over. (That's not atomic, but good enough for now.) | |
429 | * | |
430 | * If we get a short result from the OSD, check against i_size; we need to | |
431 | * only return a short read to the caller if we hit EOF. | |
432 | */ | |
433 | static int striped_read(struct inode *inode, | |
434 | u64 off, u64 len, | |
435 | struct page **pages, int num_pages) | |
436 | { | |
437 | struct ceph_client *client = ceph_inode_to_client(inode); | |
438 | struct ceph_inode_info *ci = ceph_inode(inode); | |
439 | u64 pos, this_len; | |
440 | int page_off = off & ~PAGE_CACHE_SIZE; /* first byte's offset in page */ | |
441 | int left, pages_left; | |
442 | int read; | |
443 | struct page **page_pos; | |
444 | int ret; | |
445 | bool hit_stripe, was_short; | |
446 | ||
447 | /* | |
448 | * we may need to do multiple reads. not atomic, unfortunately. | |
449 | */ | |
450 | pos = off; | |
451 | left = len; | |
452 | page_pos = pages; | |
453 | pages_left = num_pages; | |
454 | read = 0; | |
455 | ||
456 | more: | |
457 | this_len = left; | |
458 | ret = ceph_osdc_readpages(&client->osdc, ceph_vino(inode), | |
459 | &ci->i_layout, pos, &this_len, | |
460 | ci->i_truncate_seq, | |
461 | ci->i_truncate_size, | |
462 | page_pos, pages_left); | |
463 | hit_stripe = this_len < left; | |
464 | was_short = ret >= 0 && ret < this_len; | |
465 | if (ret == -ENOENT) | |
466 | ret = 0; | |
467 | dout("striped_read %llu~%u (read %u) got %d%s%s\n", pos, left, read, | |
468 | ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : ""); | |
469 | ||
470 | if (ret > 0) { | |
471 | int didpages = | |
472 | ((pos & ~PAGE_CACHE_MASK) + ret) >> PAGE_CACHE_SHIFT; | |
473 | ||
474 | if (read < pos - off) { | |
475 | dout(" zero gap %llu to %llu\n", off + read, pos); | |
476 | zero_page_vector_range(page_off + read, | |
477 | pos - off - read, pages); | |
478 | } | |
479 | pos += ret; | |
480 | read = pos - off; | |
481 | left -= ret; | |
482 | page_pos += didpages; | |
483 | pages_left -= didpages; | |
484 | ||
485 | /* hit stripe? */ | |
486 | if (left && hit_stripe) | |
487 | goto more; | |
488 | } | |
489 | ||
490 | if (was_short) { | |
491 | /* was original extent fully inside i_size? */ | |
492 | if (pos + left <= inode->i_size) { | |
493 | dout("zero tail\n"); | |
494 | zero_page_vector_range(page_off + read, len - read, | |
495 | pages); | |
496 | goto out; | |
497 | } | |
498 | ||
499 | /* check i_size */ | |
500 | ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE); | |
501 | if (ret < 0) | |
502 | goto out; | |
503 | ||
504 | /* hit EOF? */ | |
505 | if (pos >= inode->i_size) | |
506 | goto out; | |
507 | ||
508 | goto more; | |
509 | } | |
510 | ||
511 | out: | |
512 | if (ret >= 0) | |
513 | ret = read; | |
514 | dout("striped_read returns %d\n", ret); | |
515 | return ret; | |
516 | } | |
517 | ||
518 | /* | |
519 | * Completely synchronous read and write methods. Direct from __user | |
520 | * buffer to osd, or directly to user pages (if O_DIRECT). | |
521 | * | |
522 | * If the read spans object boundary, just do multiple reads. | |
523 | */ | |
524 | static ssize_t ceph_sync_read(struct file *file, char __user *data, | |
525 | unsigned len, loff_t *poff) | |
526 | { | |
527 | struct inode *inode = file->f_dentry->d_inode; | |
528 | struct page **pages; | |
529 | u64 off = *poff; | |
530 | int num_pages = calc_pages_for(off, len); | |
531 | int ret; | |
532 | ||
533 | dout("sync_read on file %p %llu~%u %s\n", file, off, len, | |
534 | (file->f_flags & O_DIRECT) ? "O_DIRECT" : ""); | |
535 | ||
536 | if (file->f_flags & O_DIRECT) { | |
537 | pages = get_direct_page_vector(data, num_pages, off, len); | |
538 | ||
539 | /* | |
540 | * flush any page cache pages in this range. this | |
541 | * will make concurrent normal and O_DIRECT io slow, | |
542 | * but it will at least behave sensibly when they are | |
543 | * in sequence. | |
544 | */ | |
545 | filemap_write_and_wait(inode->i_mapping); | |
546 | } else { | |
547 | pages = alloc_page_vector(num_pages); | |
548 | } | |
549 | if (IS_ERR(pages)) | |
550 | return PTR_ERR(pages); | |
551 | ||
552 | ret = striped_read(inode, off, len, pages, num_pages); | |
553 | ||
554 | if (ret >= 0 && (file->f_flags & O_DIRECT) == 0) | |
555 | ret = copy_page_vector_to_user(pages, data, off, ret); | |
556 | if (ret >= 0) | |
557 | *poff = off + ret; | |
558 | ||
559 | if (file->f_flags & O_DIRECT) | |
560 | put_page_vector(pages, num_pages); | |
561 | else | |
562 | ceph_release_page_vector(pages, num_pages); | |
563 | dout("sync_read result %d\n", ret); | |
564 | return ret; | |
565 | } | |
566 | ||
567 | /* | |
568 | * Write commit callback, called if we requested both an ACK and | |
569 | * ONDISK commit reply from the OSD. | |
570 | */ | |
571 | static void sync_write_commit(struct ceph_osd_request *req, | |
572 | struct ceph_msg *msg) | |
573 | { | |
574 | struct ceph_inode_info *ci = ceph_inode(req->r_inode); | |
575 | ||
576 | dout("sync_write_commit %p tid %llu\n", req, req->r_tid); | |
577 | spin_lock(&ci->i_unsafe_lock); | |
578 | list_del_init(&req->r_unsafe_item); | |
579 | spin_unlock(&ci->i_unsafe_lock); | |
580 | ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR); | |
581 | } | |
582 | ||
583 | /* | |
584 | * Synchronous write, straight from __user pointer or user pages (if | |
585 | * O_DIRECT). | |
586 | * | |
587 | * If write spans object boundary, just do multiple writes. (For a | |
588 | * correct atomic write, we should e.g. take write locks on all | |
589 | * objects, rollback on failure, etc.) | |
590 | */ | |
591 | static ssize_t ceph_sync_write(struct file *file, const char __user *data, | |
592 | size_t left, loff_t *offset) | |
593 | { | |
594 | struct inode *inode = file->f_dentry->d_inode; | |
595 | struct ceph_inode_info *ci = ceph_inode(inode); | |
596 | struct ceph_client *client = ceph_inode_to_client(inode); | |
597 | struct ceph_osd_request *req; | |
598 | struct page **pages; | |
599 | int num_pages; | |
600 | long long unsigned pos; | |
601 | u64 len; | |
602 | int written = 0; | |
603 | int flags; | |
604 | int do_sync = 0; | |
605 | int check_caps = 0; | |
606 | int ret; | |
607 | struct timespec mtime = CURRENT_TIME; | |
608 | ||
609 | if (ceph_snap(file->f_dentry->d_inode) != CEPH_NOSNAP) | |
610 | return -EROFS; | |
611 | ||
612 | dout("sync_write on file %p %lld~%u %s\n", file, *offset, | |
613 | (unsigned)left, (file->f_flags & O_DIRECT) ? "O_DIRECT" : ""); | |
614 | ||
615 | if (file->f_flags & O_APPEND) | |
616 | pos = i_size_read(inode); | |
617 | else | |
618 | pos = *offset; | |
619 | ||
620 | flags = CEPH_OSD_FLAG_ORDERSNAP | | |
621 | CEPH_OSD_FLAG_ONDISK | | |
622 | CEPH_OSD_FLAG_WRITE; | |
623 | if ((file->f_flags & (O_SYNC|O_DIRECT)) == 0) | |
624 | flags |= CEPH_OSD_FLAG_ACK; | |
625 | else | |
626 | do_sync = 1; | |
627 | ||
628 | /* | |
629 | * we may need to do multiple writes here if we span an object | |
630 | * boundary. this isn't atomic, unfortunately. :( | |
631 | */ | |
632 | more: | |
633 | len = left; | |
634 | req = ceph_osdc_new_request(&client->osdc, &ci->i_layout, | |
635 | ceph_vino(inode), pos, &len, | |
636 | CEPH_OSD_OP_WRITE, flags, | |
637 | ci->i_snap_realm->cached_context, | |
638 | do_sync, | |
639 | ci->i_truncate_seq, ci->i_truncate_size, | |
640 | &mtime, false, 2); | |
641 | if (IS_ERR(req)) | |
642 | return PTR_ERR(req); | |
643 | ||
644 | num_pages = calc_pages_for(pos, len); | |
645 | ||
646 | if (file->f_flags & O_DIRECT) { | |
647 | pages = get_direct_page_vector(data, num_pages, pos, len); | |
648 | if (IS_ERR(pages)) { | |
649 | ret = PTR_ERR(pages); | |
650 | goto out; | |
651 | } | |
652 | ||
653 | /* | |
654 | * throw out any page cache pages in this range. this | |
655 | * may block. | |
656 | */ | |
657 | truncate_inode_pages_range(inode->i_mapping, pos, pos+len); | |
658 | } else { | |
659 | pages = alloc_page_vector(num_pages); | |
660 | if (IS_ERR(pages)) { | |
661 | ret = PTR_ERR(pages); | |
662 | goto out; | |
663 | } | |
664 | ret = copy_user_to_page_vector(pages, data, pos, len); | |
665 | if (ret < 0) { | |
666 | ceph_release_page_vector(pages, num_pages); | |
667 | goto out; | |
668 | } | |
669 | ||
670 | if ((file->f_flags & O_SYNC) == 0) { | |
671 | /* get a second commit callback */ | |
672 | req->r_safe_callback = sync_write_commit; | |
673 | req->r_own_pages = 1; | |
674 | } | |
675 | } | |
676 | req->r_pages = pages; | |
677 | req->r_num_pages = num_pages; | |
678 | req->r_inode = inode; | |
679 | ||
680 | ret = ceph_osdc_start_request(&client->osdc, req, false); | |
681 | if (!ret) { | |
682 | if (req->r_safe_callback) { | |
683 | /* | |
684 | * Add to inode unsafe list only after we | |
685 | * start_request so that a tid has been assigned. | |
686 | */ | |
687 | spin_lock(&ci->i_unsafe_lock); | |
688 | list_add(&ci->i_unsafe_writes, &req->r_unsafe_item); | |
689 | spin_unlock(&ci->i_unsafe_lock); | |
690 | ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR); | |
691 | } | |
692 | ret = ceph_osdc_wait_request(&client->osdc, req); | |
693 | } | |
694 | ||
695 | if (file->f_flags & O_DIRECT) | |
696 | put_page_vector(pages, num_pages); | |
697 | else if (file->f_flags & O_SYNC) | |
698 | ceph_release_page_vector(pages, num_pages); | |
699 | ||
700 | out: | |
701 | ceph_osdc_put_request(req); | |
702 | if (ret == 0) { | |
703 | pos += len; | |
704 | written += len; | |
705 | left -= len; | |
706 | if (left) | |
707 | goto more; | |
708 | ||
709 | ret = written; | |
710 | *offset = pos; | |
711 | if (pos > i_size_read(inode)) | |
712 | check_caps = ceph_inode_set_size(inode, pos); | |
713 | if (check_caps) | |
714 | ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, | |
715 | NULL); | |
716 | } | |
717 | return ret; | |
718 | } | |
719 | ||
720 | /* | |
721 | * Wrap generic_file_aio_read with checks for cap bits on the inode. | |
722 | * Atomically grab references, so that those bits are not released | |
723 | * back to the MDS mid-read. | |
724 | * | |
725 | * Hmm, the sync read case isn't actually async... should it be? | |
726 | */ | |
727 | static ssize_t ceph_aio_read(struct kiocb *iocb, const struct iovec *iov, | |
728 | unsigned long nr_segs, loff_t pos) | |
729 | { | |
730 | struct file *filp = iocb->ki_filp; | |
731 | loff_t *ppos = &iocb->ki_pos; | |
732 | size_t len = iov->iov_len; | |
733 | struct inode *inode = filp->f_dentry->d_inode; | |
734 | struct ceph_inode_info *ci = ceph_inode(inode); | |
735 | ssize_t ret; | |
736 | int got = 0; | |
737 | ||
738 | dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n", | |
739 | inode, ceph_vinop(inode), pos, (unsigned)len, inode); | |
740 | __ceph_do_pending_vmtruncate(inode); | |
741 | ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, CEPH_CAP_FILE_CACHE, | |
742 | &got, -1); | |
743 | if (ret < 0) | |
744 | goto out; | |
745 | dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n", | |
746 | inode, ceph_vinop(inode), pos, (unsigned)len, | |
747 | ceph_cap_string(got)); | |
748 | ||
749 | if ((got & CEPH_CAP_FILE_CACHE) == 0 || | |
750 | (iocb->ki_filp->f_flags & O_DIRECT) || | |
751 | (inode->i_sb->s_flags & MS_SYNCHRONOUS)) | |
752 | /* hmm, this isn't really async... */ | |
753 | ret = ceph_sync_read(filp, iov->iov_base, len, ppos); | |
754 | else | |
755 | ret = generic_file_aio_read(iocb, iov, nr_segs, pos); | |
756 | ||
757 | out: | |
758 | dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n", | |
759 | inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret); | |
760 | ceph_put_cap_refs(ci, got); | |
761 | return ret; | |
762 | } | |
763 | ||
764 | /* | |
765 | * Take cap references to avoid releasing caps to MDS mid-write. | |
766 | * | |
767 | * If we are synchronous, and write with an old snap context, the OSD | |
768 | * may return EOLDSNAPC. In that case, retry the write.. _after_ | |
769 | * dropping our cap refs and allowing the pending snap to logically | |
770 | * complete _before_ this write occurs. | |
771 | * | |
772 | * If we are near ENOSPC, write synchronously. | |
773 | */ | |
774 | static ssize_t ceph_aio_write(struct kiocb *iocb, const struct iovec *iov, | |
775 | unsigned long nr_segs, loff_t pos) | |
776 | { | |
777 | struct file *file = iocb->ki_filp; | |
778 | struct inode *inode = file->f_dentry->d_inode; | |
779 | struct ceph_inode_info *ci = ceph_inode(inode); | |
780 | struct ceph_osd_client *osdc = &ceph_client(inode->i_sb)->osdc; | |
781 | loff_t endoff = pos + iov->iov_len; | |
782 | int got = 0; | |
783 | int ret; | |
784 | ||
785 | if (ceph_snap(inode) != CEPH_NOSNAP) | |
786 | return -EROFS; | |
787 | ||
788 | retry_snap: | |
789 | if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) | |
790 | return -ENOSPC; | |
791 | __ceph_do_pending_vmtruncate(inode); | |
792 | dout("aio_write %p %llx.%llx %llu~%u getting caps. i_size %llu\n", | |
793 | inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len, | |
794 | inode->i_size); | |
795 | ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER, | |
796 | &got, endoff); | |
797 | if (ret < 0) | |
798 | goto out; | |
799 | ||
800 | dout("aio_write %p %llx.%llx %llu~%u got cap refs on %s\n", | |
801 | inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len, | |
802 | ceph_cap_string(got)); | |
803 | ||
804 | if ((got & CEPH_CAP_FILE_BUFFER) == 0 || | |
805 | (iocb->ki_filp->f_flags & O_DIRECT) || | |
806 | (inode->i_sb->s_flags & MS_SYNCHRONOUS)) { | |
807 | ret = ceph_sync_write(file, iov->iov_base, iov->iov_len, | |
808 | &iocb->ki_pos); | |
809 | } else { | |
810 | ret = generic_file_aio_write(iocb, iov, nr_segs, pos); | |
811 | ||
812 | if ((ret >= 0 || ret == -EIOCBQUEUED) && | |
813 | ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host) | |
814 | || ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) | |
815 | ret = vfs_fsync_range(file, file->f_path.dentry, | |
816 | pos, pos + ret - 1, 1); | |
817 | } | |
818 | if (ret >= 0) { | |
819 | spin_lock(&inode->i_lock); | |
820 | __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR); | |
821 | spin_unlock(&inode->i_lock); | |
822 | } | |
823 | ||
824 | out: | |
825 | dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n", | |
826 | inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len, | |
827 | ceph_cap_string(got)); | |
828 | ceph_put_cap_refs(ci, got); | |
829 | ||
830 | if (ret == -EOLDSNAPC) { | |
831 | dout("aio_write %p %llx.%llx %llu~%u got EOLDSNAPC, retrying\n", | |
832 | inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len); | |
833 | goto retry_snap; | |
834 | } | |
835 | ||
836 | return ret; | |
837 | } | |
838 | ||
839 | /* | |
840 | * llseek. be sure to verify file size on SEEK_END. | |
841 | */ | |
842 | static loff_t ceph_llseek(struct file *file, loff_t offset, int origin) | |
843 | { | |
844 | struct inode *inode = file->f_mapping->host; | |
845 | int ret; | |
846 | ||
847 | mutex_lock(&inode->i_mutex); | |
848 | __ceph_do_pending_vmtruncate(inode); | |
849 | switch (origin) { | |
850 | case SEEK_END: | |
851 | ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE); | |
852 | if (ret < 0) { | |
853 | offset = ret; | |
854 | goto out; | |
855 | } | |
856 | offset += inode->i_size; | |
857 | break; | |
858 | case SEEK_CUR: | |
859 | /* | |
860 | * Here we special-case the lseek(fd, 0, SEEK_CUR) | |
861 | * position-querying operation. Avoid rewriting the "same" | |
862 | * f_pos value back to the file because a concurrent read(), | |
863 | * write() or lseek() might have altered it | |
864 | */ | |
865 | if (offset == 0) { | |
866 | offset = file->f_pos; | |
867 | goto out; | |
868 | } | |
869 | offset += file->f_pos; | |
870 | break; | |
871 | } | |
872 | ||
873 | if (offset < 0 || offset > inode->i_sb->s_maxbytes) { | |
874 | offset = -EINVAL; | |
875 | goto out; | |
876 | } | |
877 | ||
878 | /* Special lock needed here? */ | |
879 | if (offset != file->f_pos) { | |
880 | file->f_pos = offset; | |
881 | file->f_version = 0; | |
882 | } | |
883 | ||
884 | out: | |
885 | mutex_unlock(&inode->i_mutex); | |
886 | return offset; | |
887 | } | |
888 | ||
889 | const struct file_operations ceph_file_fops = { | |
890 | .open = ceph_open, | |
891 | .release = ceph_release, | |
892 | .llseek = ceph_llseek, | |
893 | .read = do_sync_read, | |
894 | .write = do_sync_write, | |
895 | .aio_read = ceph_aio_read, | |
896 | .aio_write = ceph_aio_write, | |
897 | .mmap = ceph_mmap, | |
898 | .fsync = ceph_fsync, | |
899 | .splice_read = generic_file_splice_read, | |
900 | .splice_write = generic_file_splice_write, | |
901 | .unlocked_ioctl = ceph_ioctl, | |
902 | .compat_ioctl = ceph_ioctl, | |
903 | }; | |
904 |