]>
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); | |
6a18be16 | 55 | req->r_args.open.preferred = cpu_to_le32(-1); |
124e68e7 SW |
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); | |
195d3ce2 SW |
265 | |
266 | /* wake up anyone waiting for caps on this inode */ | |
267 | wake_up(&ci->i_cap_wq); | |
124e68e7 SW |
268 | return 0; |
269 | } | |
270 | ||
271 | /* | |
272 | * build a vector of user pages | |
273 | */ | |
274 | static struct page **get_direct_page_vector(const char __user *data, | |
275 | int num_pages, | |
276 | loff_t off, size_t len) | |
277 | { | |
278 | struct page **pages; | |
279 | int rc; | |
280 | ||
281 | pages = kmalloc(sizeof(*pages) * num_pages, GFP_NOFS); | |
282 | if (!pages) | |
283 | return ERR_PTR(-ENOMEM); | |
284 | ||
285 | down_read(¤t->mm->mmap_sem); | |
286 | rc = get_user_pages(current, current->mm, (unsigned long)data, | |
287 | num_pages, 0, 0, pages, NULL); | |
288 | up_read(¤t->mm->mmap_sem); | |
289 | if (rc < 0) | |
290 | goto fail; | |
291 | return pages; | |
292 | ||
293 | fail: | |
294 | kfree(pages); | |
295 | return ERR_PTR(rc); | |
296 | } | |
297 | ||
298 | static void put_page_vector(struct page **pages, int num_pages) | |
299 | { | |
300 | int i; | |
301 | ||
302 | for (i = 0; i < num_pages; i++) | |
303 | put_page(pages[i]); | |
304 | kfree(pages); | |
305 | } | |
306 | ||
307 | void ceph_release_page_vector(struct page **pages, int num_pages) | |
308 | { | |
309 | int i; | |
310 | ||
311 | for (i = 0; i < num_pages; i++) | |
312 | __free_pages(pages[i], 0); | |
313 | kfree(pages); | |
314 | } | |
315 | ||
316 | /* | |
317 | * allocate a vector new pages | |
318 | */ | |
319 | static struct page **alloc_page_vector(int num_pages) | |
320 | { | |
321 | struct page **pages; | |
322 | int i; | |
323 | ||
324 | pages = kmalloc(sizeof(*pages) * num_pages, GFP_NOFS); | |
325 | if (!pages) | |
326 | return ERR_PTR(-ENOMEM); | |
327 | for (i = 0; i < num_pages; i++) { | |
328 | pages[i] = alloc_page(GFP_NOFS); | |
329 | if (pages[i] == NULL) { | |
330 | ceph_release_page_vector(pages, i); | |
331 | return ERR_PTR(-ENOMEM); | |
332 | } | |
333 | } | |
334 | return pages; | |
335 | } | |
336 | ||
337 | /* | |
338 | * copy user data into a page vector | |
339 | */ | |
340 | static int copy_user_to_page_vector(struct page **pages, | |
341 | const char __user *data, | |
342 | loff_t off, size_t len) | |
343 | { | |
344 | int i = 0; | |
345 | int po = off & ~PAGE_CACHE_MASK; | |
346 | int left = len; | |
347 | int l, bad; | |
348 | ||
349 | while (left > 0) { | |
350 | l = min_t(int, PAGE_CACHE_SIZE-po, left); | |
351 | bad = copy_from_user(page_address(pages[i]) + po, data, l); | |
352 | if (bad == l) | |
353 | return -EFAULT; | |
354 | data += l - bad; | |
355 | left -= l - bad; | |
6a4ef481 YS |
356 | po += l - bad; |
357 | if (po == PAGE_CACHE_SIZE) { | |
358 | po = 0; | |
359 | i++; | |
124e68e7 SW |
360 | } |
361 | } | |
362 | return len; | |
363 | } | |
364 | ||
365 | /* | |
366 | * copy user data from a page vector into a user pointer | |
367 | */ | |
368 | static int copy_page_vector_to_user(struct page **pages, char __user *data, | |
369 | loff_t off, size_t len) | |
370 | { | |
371 | int i = 0; | |
372 | int po = off & ~PAGE_CACHE_MASK; | |
373 | int left = len; | |
374 | int l, bad; | |
375 | ||
376 | while (left > 0) { | |
377 | l = min_t(int, left, PAGE_CACHE_SIZE-po); | |
378 | bad = copy_to_user(data, page_address(pages[i]) + po, l); | |
379 | if (bad == l) | |
380 | return -EFAULT; | |
381 | data += l - bad; | |
382 | left -= l - bad; | |
383 | if (po) { | |
384 | po += l - bad; | |
385 | if (po == PAGE_CACHE_SIZE) | |
386 | po = 0; | |
387 | } | |
388 | i++; | |
389 | } | |
390 | return len; | |
391 | } | |
392 | ||
393 | /* | |
394 | * Zero an extent within a page vector. Offset is relative to the | |
395 | * start of the first page. | |
396 | */ | |
397 | static void zero_page_vector_range(int off, int len, struct page **pages) | |
398 | { | |
399 | int i = off >> PAGE_CACHE_SHIFT; | |
400 | ||
972f0d3a YS |
401 | off &= ~PAGE_CACHE_MASK; |
402 | ||
124e68e7 | 403 | dout("zero_page_vector_page %u~%u\n", off, len); |
124e68e7 SW |
404 | |
405 | /* leading partial page? */ | |
972f0d3a YS |
406 | if (off) { |
407 | int end = min((int)PAGE_CACHE_SIZE, off + len); | |
124e68e7 | 408 | dout("zeroing %d %p head from %d\n", i, pages[i], |
972f0d3a YS |
409 | (int)off); |
410 | zero_user_segment(pages[i], off, end); | |
411 | len -= (end - off); | |
124e68e7 SW |
412 | i++; |
413 | } | |
414 | while (len >= PAGE_CACHE_SIZE) { | |
29065a51 | 415 | dout("zeroing %d %p len=%d\n", i, pages[i], len); |
124e68e7 | 416 | zero_user_segment(pages[i], 0, PAGE_CACHE_SIZE); |
124e68e7 SW |
417 | len -= PAGE_CACHE_SIZE; |
418 | i++; | |
419 | } | |
420 | /* trailing partial page? */ | |
421 | if (len) { | |
422 | dout("zeroing %d %p tail to %d\n", i, pages[i], (int)len); | |
423 | zero_user_segment(pages[i], 0, len); | |
424 | } | |
425 | } | |
426 | ||
427 | ||
428 | /* | |
429 | * Read a range of bytes striped over one or more objects. Iterate over | |
430 | * objects we stripe over. (That's not atomic, but good enough for now.) | |
431 | * | |
432 | * If we get a short result from the OSD, check against i_size; we need to | |
433 | * only return a short read to the caller if we hit EOF. | |
434 | */ | |
435 | static int striped_read(struct inode *inode, | |
436 | u64 off, u64 len, | |
6a026589 SW |
437 | struct page **pages, int num_pages, |
438 | int *checkeof) | |
124e68e7 SW |
439 | { |
440 | struct ceph_client *client = ceph_inode_to_client(inode); | |
441 | struct ceph_inode_info *ci = ceph_inode(inode); | |
442 | u64 pos, this_len; | |
972f0d3a | 443 | int page_off = off & ~PAGE_CACHE_MASK; /* first byte's offset in page */ |
124e68e7 SW |
444 | int left, pages_left; |
445 | int read; | |
446 | struct page **page_pos; | |
447 | int ret; | |
448 | bool hit_stripe, was_short; | |
449 | ||
450 | /* | |
451 | * we may need to do multiple reads. not atomic, unfortunately. | |
452 | */ | |
453 | pos = off; | |
454 | left = len; | |
455 | page_pos = pages; | |
456 | pages_left = num_pages; | |
457 | read = 0; | |
458 | ||
459 | more: | |
460 | this_len = left; | |
461 | ret = ceph_osdc_readpages(&client->osdc, ceph_vino(inode), | |
462 | &ci->i_layout, pos, &this_len, | |
463 | ci->i_truncate_seq, | |
464 | ci->i_truncate_size, | |
465 | page_pos, pages_left); | |
466 | hit_stripe = this_len < left; | |
467 | was_short = ret >= 0 && ret < this_len; | |
468 | if (ret == -ENOENT) | |
469 | ret = 0; | |
470 | dout("striped_read %llu~%u (read %u) got %d%s%s\n", pos, left, read, | |
471 | ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : ""); | |
472 | ||
473 | if (ret > 0) { | |
474 | int didpages = | |
475 | ((pos & ~PAGE_CACHE_MASK) + ret) >> PAGE_CACHE_SHIFT; | |
476 | ||
477 | if (read < pos - off) { | |
478 | dout(" zero gap %llu to %llu\n", off + read, pos); | |
479 | zero_page_vector_range(page_off + read, | |
480 | pos - off - read, pages); | |
481 | } | |
482 | pos += ret; | |
483 | read = pos - off; | |
484 | left -= ret; | |
485 | page_pos += didpages; | |
486 | pages_left -= didpages; | |
487 | ||
488 | /* hit stripe? */ | |
489 | if (left && hit_stripe) | |
490 | goto more; | |
491 | } | |
492 | ||
493 | if (was_short) { | |
494 | /* was original extent fully inside i_size? */ | |
495 | if (pos + left <= inode->i_size) { | |
496 | dout("zero tail\n"); | |
497 | zero_page_vector_range(page_off + read, len - read, | |
498 | pages); | |
972f0d3a | 499 | read = len; |
124e68e7 SW |
500 | goto out; |
501 | } | |
502 | ||
503 | /* check i_size */ | |
6a026589 | 504 | *checkeof = 1; |
124e68e7 SW |
505 | } |
506 | ||
507 | out: | |
508 | if (ret >= 0) | |
509 | ret = read; | |
510 | dout("striped_read returns %d\n", ret); | |
511 | return ret; | |
512 | } | |
513 | ||
514 | /* | |
515 | * Completely synchronous read and write methods. Direct from __user | |
516 | * buffer to osd, or directly to user pages (if O_DIRECT). | |
517 | * | |
518 | * If the read spans object boundary, just do multiple reads. | |
519 | */ | |
520 | static ssize_t ceph_sync_read(struct file *file, char __user *data, | |
6a026589 | 521 | unsigned len, loff_t *poff, int *checkeof) |
124e68e7 SW |
522 | { |
523 | struct inode *inode = file->f_dentry->d_inode; | |
524 | struct page **pages; | |
525 | u64 off = *poff; | |
526 | int num_pages = calc_pages_for(off, len); | |
527 | int ret; | |
528 | ||
529 | dout("sync_read on file %p %llu~%u %s\n", file, off, len, | |
530 | (file->f_flags & O_DIRECT) ? "O_DIRECT" : ""); | |
531 | ||
532 | if (file->f_flags & O_DIRECT) { | |
533 | pages = get_direct_page_vector(data, num_pages, off, len); | |
534 | ||
535 | /* | |
536 | * flush any page cache pages in this range. this | |
537 | * will make concurrent normal and O_DIRECT io slow, | |
538 | * but it will at least behave sensibly when they are | |
539 | * in sequence. | |
540 | */ | |
124e68e7 SW |
541 | } else { |
542 | pages = alloc_page_vector(num_pages); | |
543 | } | |
544 | if (IS_ERR(pages)) | |
545 | return PTR_ERR(pages); | |
546 | ||
29065a51 YS |
547 | ret = filemap_write_and_wait(inode->i_mapping); |
548 | if (ret < 0) | |
549 | goto done; | |
550 | ||
6a026589 | 551 | ret = striped_read(inode, off, len, pages, num_pages, checkeof); |
124e68e7 SW |
552 | |
553 | if (ret >= 0 && (file->f_flags & O_DIRECT) == 0) | |
554 | ret = copy_page_vector_to_user(pages, data, off, ret); | |
555 | if (ret >= 0) | |
556 | *poff = off + ret; | |
557 | ||
29065a51 | 558 | done: |
124e68e7 SW |
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 | ||
29065a51 YS |
620 | ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + left); |
621 | if (ret < 0) | |
622 | return ret; | |
623 | ||
624 | ret = invalidate_inode_pages2_range(inode->i_mapping, | |
625 | pos >> PAGE_CACHE_SHIFT, | |
626 | (pos + left) >> PAGE_CACHE_SHIFT); | |
627 | if (ret < 0) | |
628 | dout("invalidate_inode_pages2_range returned %d\n", ret); | |
629 | ||
124e68e7 SW |
630 | flags = CEPH_OSD_FLAG_ORDERSNAP | |
631 | CEPH_OSD_FLAG_ONDISK | | |
632 | CEPH_OSD_FLAG_WRITE; | |
633 | if ((file->f_flags & (O_SYNC|O_DIRECT)) == 0) | |
634 | flags |= CEPH_OSD_FLAG_ACK; | |
635 | else | |
636 | do_sync = 1; | |
637 | ||
638 | /* | |
639 | * we may need to do multiple writes here if we span an object | |
640 | * boundary. this isn't atomic, unfortunately. :( | |
641 | */ | |
642 | more: | |
643 | len = left; | |
644 | req = ceph_osdc_new_request(&client->osdc, &ci->i_layout, | |
645 | ceph_vino(inode), pos, &len, | |
646 | CEPH_OSD_OP_WRITE, flags, | |
647 | ci->i_snap_realm->cached_context, | |
648 | do_sync, | |
649 | ci->i_truncate_seq, ci->i_truncate_size, | |
650 | &mtime, false, 2); | |
651 | if (IS_ERR(req)) | |
652 | return PTR_ERR(req); | |
653 | ||
654 | num_pages = calc_pages_for(pos, len); | |
655 | ||
656 | if (file->f_flags & O_DIRECT) { | |
657 | pages = get_direct_page_vector(data, num_pages, pos, len); | |
658 | if (IS_ERR(pages)) { | |
659 | ret = PTR_ERR(pages); | |
660 | goto out; | |
661 | } | |
662 | ||
663 | /* | |
664 | * throw out any page cache pages in this range. this | |
665 | * may block. | |
666 | */ | |
667 | truncate_inode_pages_range(inode->i_mapping, pos, pos+len); | |
668 | } else { | |
669 | pages = alloc_page_vector(num_pages); | |
670 | if (IS_ERR(pages)) { | |
671 | ret = PTR_ERR(pages); | |
672 | goto out; | |
673 | } | |
674 | ret = copy_user_to_page_vector(pages, data, pos, len); | |
675 | if (ret < 0) { | |
676 | ceph_release_page_vector(pages, num_pages); | |
677 | goto out; | |
678 | } | |
679 | ||
680 | if ((file->f_flags & O_SYNC) == 0) { | |
681 | /* get a second commit callback */ | |
682 | req->r_safe_callback = sync_write_commit; | |
683 | req->r_own_pages = 1; | |
684 | } | |
685 | } | |
686 | req->r_pages = pages; | |
687 | req->r_num_pages = num_pages; | |
688 | req->r_inode = inode; | |
689 | ||
690 | ret = ceph_osdc_start_request(&client->osdc, req, false); | |
691 | if (!ret) { | |
692 | if (req->r_safe_callback) { | |
693 | /* | |
694 | * Add to inode unsafe list only after we | |
695 | * start_request so that a tid has been assigned. | |
696 | */ | |
697 | spin_lock(&ci->i_unsafe_lock); | |
698 | list_add(&ci->i_unsafe_writes, &req->r_unsafe_item); | |
699 | spin_unlock(&ci->i_unsafe_lock); | |
700 | ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR); | |
701 | } | |
702 | ret = ceph_osdc_wait_request(&client->osdc, req); | |
703 | } | |
704 | ||
705 | if (file->f_flags & O_DIRECT) | |
706 | put_page_vector(pages, num_pages); | |
707 | else if (file->f_flags & O_SYNC) | |
708 | ceph_release_page_vector(pages, num_pages); | |
709 | ||
710 | out: | |
711 | ceph_osdc_put_request(req); | |
712 | if (ret == 0) { | |
713 | pos += len; | |
714 | written += len; | |
715 | left -= len; | |
716 | if (left) | |
717 | goto more; | |
718 | ||
719 | ret = written; | |
720 | *offset = pos; | |
721 | if (pos > i_size_read(inode)) | |
722 | check_caps = ceph_inode_set_size(inode, pos); | |
723 | if (check_caps) | |
724 | ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, | |
725 | NULL); | |
726 | } | |
727 | return ret; | |
728 | } | |
729 | ||
730 | /* | |
731 | * Wrap generic_file_aio_read with checks for cap bits on the inode. | |
732 | * Atomically grab references, so that those bits are not released | |
733 | * back to the MDS mid-read. | |
734 | * | |
735 | * Hmm, the sync read case isn't actually async... should it be? | |
736 | */ | |
737 | static ssize_t ceph_aio_read(struct kiocb *iocb, const struct iovec *iov, | |
738 | unsigned long nr_segs, loff_t pos) | |
739 | { | |
740 | struct file *filp = iocb->ki_filp; | |
741 | loff_t *ppos = &iocb->ki_pos; | |
742 | size_t len = iov->iov_len; | |
743 | struct inode *inode = filp->f_dentry->d_inode; | |
744 | struct ceph_inode_info *ci = ceph_inode(inode); | |
6a026589 | 745 | void *base = iov->iov_base; |
124e68e7 SW |
746 | ssize_t ret; |
747 | int got = 0; | |
6a026589 | 748 | int checkeof = 0, read = 0; |
124e68e7 SW |
749 | |
750 | dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n", | |
751 | inode, ceph_vinop(inode), pos, (unsigned)len, inode); | |
6a026589 | 752 | again: |
124e68e7 SW |
753 | __ceph_do_pending_vmtruncate(inode); |
754 | ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, CEPH_CAP_FILE_CACHE, | |
755 | &got, -1); | |
756 | if (ret < 0) | |
757 | goto out; | |
758 | dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n", | |
759 | inode, ceph_vinop(inode), pos, (unsigned)len, | |
760 | ceph_cap_string(got)); | |
761 | ||
762 | if ((got & CEPH_CAP_FILE_CACHE) == 0 || | |
763 | (iocb->ki_filp->f_flags & O_DIRECT) || | |
764 | (inode->i_sb->s_flags & MS_SYNCHRONOUS)) | |
765 | /* hmm, this isn't really async... */ | |
6a026589 | 766 | ret = ceph_sync_read(filp, base, len, ppos, &checkeof); |
124e68e7 SW |
767 | else |
768 | ret = generic_file_aio_read(iocb, iov, nr_segs, pos); | |
769 | ||
770 | out: | |
771 | dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n", | |
772 | inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret); | |
773 | ceph_put_cap_refs(ci, got); | |
6a026589 SW |
774 | |
775 | if (checkeof && ret >= 0) { | |
776 | int statret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE); | |
777 | ||
778 | /* hit EOF or hole? */ | |
779 | if (statret == 0 && *ppos < inode->i_size) { | |
780 | dout("aio_read sync_read hit hole, reading more\n"); | |
781 | read += ret; | |
782 | base += ret; | |
783 | len -= ret; | |
784 | checkeof = 0; | |
785 | goto again; | |
786 | } | |
787 | } | |
788 | if (ret >= 0) | |
789 | ret += read; | |
790 | ||
124e68e7 SW |
791 | return ret; |
792 | } | |
793 | ||
794 | /* | |
795 | * Take cap references to avoid releasing caps to MDS mid-write. | |
796 | * | |
797 | * If we are synchronous, and write with an old snap context, the OSD | |
798 | * may return EOLDSNAPC. In that case, retry the write.. _after_ | |
799 | * dropping our cap refs and allowing the pending snap to logically | |
800 | * complete _before_ this write occurs. | |
801 | * | |
802 | * If we are near ENOSPC, write synchronously. | |
803 | */ | |
804 | static ssize_t ceph_aio_write(struct kiocb *iocb, const struct iovec *iov, | |
805 | unsigned long nr_segs, loff_t pos) | |
806 | { | |
807 | struct file *file = iocb->ki_filp; | |
808 | struct inode *inode = file->f_dentry->d_inode; | |
809 | struct ceph_inode_info *ci = ceph_inode(inode); | |
810 | struct ceph_osd_client *osdc = &ceph_client(inode->i_sb)->osdc; | |
811 | loff_t endoff = pos + iov->iov_len; | |
812 | int got = 0; | |
88d892a3 | 813 | int ret, err; |
124e68e7 SW |
814 | |
815 | if (ceph_snap(inode) != CEPH_NOSNAP) | |
816 | return -EROFS; | |
817 | ||
818 | retry_snap: | |
819 | if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) | |
820 | return -ENOSPC; | |
821 | __ceph_do_pending_vmtruncate(inode); | |
822 | dout("aio_write %p %llx.%llx %llu~%u getting caps. i_size %llu\n", | |
823 | inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len, | |
824 | inode->i_size); | |
825 | ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER, | |
826 | &got, endoff); | |
827 | if (ret < 0) | |
828 | goto out; | |
829 | ||
830 | dout("aio_write %p %llx.%llx %llu~%u got cap refs on %s\n", | |
831 | inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len, | |
832 | ceph_cap_string(got)); | |
833 | ||
834 | if ((got & CEPH_CAP_FILE_BUFFER) == 0 || | |
835 | (iocb->ki_filp->f_flags & O_DIRECT) || | |
836 | (inode->i_sb->s_flags & MS_SYNCHRONOUS)) { | |
837 | ret = ceph_sync_write(file, iov->iov_base, iov->iov_len, | |
838 | &iocb->ki_pos); | |
839 | } else { | |
840 | ret = generic_file_aio_write(iocb, iov, nr_segs, pos); | |
841 | ||
842 | if ((ret >= 0 || ret == -EIOCBQUEUED) && | |
843 | ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host) | |
88d892a3 YS |
844 | || ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) { |
845 | err = vfs_fsync_range(file, file->f_path.dentry, | |
124e68e7 | 846 | pos, pos + ret - 1, 1); |
88d892a3 YS |
847 | if (err < 0) |
848 | ret = err; | |
849 | } | |
124e68e7 SW |
850 | } |
851 | if (ret >= 0) { | |
852 | spin_lock(&inode->i_lock); | |
853 | __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR); | |
854 | spin_unlock(&inode->i_lock); | |
855 | } | |
856 | ||
857 | out: | |
858 | dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n", | |
859 | inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len, | |
860 | ceph_cap_string(got)); | |
861 | ceph_put_cap_refs(ci, got); | |
862 | ||
863 | if (ret == -EOLDSNAPC) { | |
864 | dout("aio_write %p %llx.%llx %llu~%u got EOLDSNAPC, retrying\n", | |
865 | inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len); | |
866 | goto retry_snap; | |
867 | } | |
868 | ||
869 | return ret; | |
870 | } | |
871 | ||
872 | /* | |
873 | * llseek. be sure to verify file size on SEEK_END. | |
874 | */ | |
875 | static loff_t ceph_llseek(struct file *file, loff_t offset, int origin) | |
876 | { | |
877 | struct inode *inode = file->f_mapping->host; | |
878 | int ret; | |
879 | ||
880 | mutex_lock(&inode->i_mutex); | |
881 | __ceph_do_pending_vmtruncate(inode); | |
882 | switch (origin) { | |
883 | case SEEK_END: | |
884 | ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE); | |
885 | if (ret < 0) { | |
886 | offset = ret; | |
887 | goto out; | |
888 | } | |
889 | offset += inode->i_size; | |
890 | break; | |
891 | case SEEK_CUR: | |
892 | /* | |
893 | * Here we special-case the lseek(fd, 0, SEEK_CUR) | |
894 | * position-querying operation. Avoid rewriting the "same" | |
895 | * f_pos value back to the file because a concurrent read(), | |
896 | * write() or lseek() might have altered it | |
897 | */ | |
898 | if (offset == 0) { | |
899 | offset = file->f_pos; | |
900 | goto out; | |
901 | } | |
902 | offset += file->f_pos; | |
903 | break; | |
904 | } | |
905 | ||
906 | if (offset < 0 || offset > inode->i_sb->s_maxbytes) { | |
907 | offset = -EINVAL; | |
908 | goto out; | |
909 | } | |
910 | ||
911 | /* Special lock needed here? */ | |
912 | if (offset != file->f_pos) { | |
913 | file->f_pos = offset; | |
914 | file->f_version = 0; | |
915 | } | |
916 | ||
917 | out: | |
918 | mutex_unlock(&inode->i_mutex); | |
919 | return offset; | |
920 | } | |
921 | ||
922 | const struct file_operations ceph_file_fops = { | |
923 | .open = ceph_open, | |
924 | .release = ceph_release, | |
925 | .llseek = ceph_llseek, | |
926 | .read = do_sync_read, | |
927 | .write = do_sync_write, | |
928 | .aio_read = ceph_aio_read, | |
929 | .aio_write = ceph_aio_write, | |
930 | .mmap = ceph_mmap, | |
931 | .fsync = ceph_fsync, | |
932 | .splice_read = generic_file_splice_read, | |
933 | .splice_write = generic_file_splice_write, | |
934 | .unlocked_ioctl = ceph_ioctl, | |
935 | .compat_ioctl = ceph_ioctl, | |
936 | }; | |
937 |