Commit | Line | Data |
---|---|---|
3d14c5d2 | 1 | #include <linux/ceph/ceph_debug.h> |
124e68e7 | 2 | |
3d14c5d2 | 3 | #include <linux/module.h> |
124e68e7 | 4 | #include <linux/sched.h> |
5a0e3ad6 | 5 | #include <linux/slab.h> |
124e68e7 SW |
6 | #include <linux/file.h> |
7 | #include <linux/namei.h> | |
8 | #include <linux/writeback.h> | |
9 | ||
10 | #include "super.h" | |
11 | #include "mds_client.h" | |
12 | ||
13 | /* | |
14 | * Ceph file operations | |
15 | * | |
16 | * Implement basic open/close functionality, and implement | |
17 | * read/write. | |
18 | * | |
19 | * We implement three modes of file I/O: | |
20 | * - buffered uses the generic_file_aio_{read,write} helpers | |
21 | * | |
22 | * - synchronous is used when there is multi-client read/write | |
23 | * sharing, avoids the page cache, and synchronously waits for an | |
24 | * ack from the OSD. | |
25 | * | |
26 | * - direct io takes the variant of the sync path that references | |
27 | * user pages directly. | |
28 | * | |
29 | * fsync() flushes and waits on dirty pages, but just queues metadata | |
30 | * for writeback: since the MDS can recover size and mtime there is no | |
31 | * need to wait for MDS acknowledgement. | |
32 | */ | |
33 | ||
34 | ||
35 | /* | |
36 | * Prepare an open request. Preallocate ceph_cap to avoid an | |
37 | * inopportune ENOMEM later. | |
38 | */ | |
39 | static struct ceph_mds_request * | |
40 | prepare_open_request(struct super_block *sb, int flags, int create_mode) | |
41 | { | |
3d14c5d2 YS |
42 | struct ceph_fs_client *fsc = ceph_sb_to_client(sb); |
43 | struct ceph_mds_client *mdsc = fsc->mdsc; | |
124e68e7 SW |
44 | struct ceph_mds_request *req; |
45 | int want_auth = USE_ANY_MDS; | |
46 | int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN; | |
47 | ||
48 | if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC)) | |
49 | want_auth = USE_AUTH_MDS; | |
50 | ||
51 | req = ceph_mdsc_create_request(mdsc, op, want_auth); | |
52 | if (IS_ERR(req)) | |
53 | goto out; | |
54 | req->r_fmode = ceph_flags_to_mode(flags); | |
55 | req->r_args.open.flags = cpu_to_le32(flags); | |
56 | req->r_args.open.mode = cpu_to_le32(create_mode); | |
6a18be16 | 57 | req->r_args.open.preferred = cpu_to_le32(-1); |
124e68e7 SW |
58 | out: |
59 | return req; | |
60 | } | |
61 | ||
62 | /* | |
63 | * initialize private struct file data. | |
64 | * if we fail, clean up by dropping fmode reference on the ceph_inode | |
65 | */ | |
66 | static int ceph_init_file(struct inode *inode, struct file *file, int fmode) | |
67 | { | |
68 | struct ceph_file_info *cf; | |
69 | int ret = 0; | |
70 | ||
71 | switch (inode->i_mode & S_IFMT) { | |
72 | case S_IFREG: | |
73 | case S_IFDIR: | |
74 | dout("init_file %p %p 0%o (regular)\n", inode, file, | |
75 | inode->i_mode); | |
76 | cf = kmem_cache_alloc(ceph_file_cachep, GFP_NOFS | __GFP_ZERO); | |
77 | if (cf == NULL) { | |
78 | ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */ | |
79 | return -ENOMEM; | |
80 | } | |
81 | cf->fmode = fmode; | |
82 | cf->next_offset = 2; | |
83 | file->private_data = cf; | |
84 | BUG_ON(inode->i_fop->release != ceph_release); | |
85 | break; | |
86 | ||
87 | case S_IFLNK: | |
88 | dout("init_file %p %p 0%o (symlink)\n", inode, file, | |
89 | inode->i_mode); | |
90 | ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */ | |
91 | break; | |
92 | ||
93 | default: | |
94 | dout("init_file %p %p 0%o (special)\n", inode, file, | |
95 | inode->i_mode); | |
96 | /* | |
97 | * we need to drop the open ref now, since we don't | |
98 | * have .release set to ceph_release. | |
99 | */ | |
100 | ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */ | |
101 | BUG_ON(inode->i_fop->release == ceph_release); | |
102 | ||
103 | /* call the proper open fop */ | |
104 | ret = inode->i_fop->open(inode, file); | |
105 | } | |
106 | return ret; | |
107 | } | |
108 | ||
109 | /* | |
110 | * If the filp already has private_data, that means the file was | |
111 | * already opened by intent during lookup, and we do nothing. | |
112 | * | |
113 | * If we already have the requisite capabilities, we can satisfy | |
114 | * the open request locally (no need to request new caps from the | |
115 | * MDS). We do, however, need to inform the MDS (asynchronously) | |
116 | * if our wanted caps set expands. | |
117 | */ | |
118 | int ceph_open(struct inode *inode, struct file *file) | |
119 | { | |
120 | struct ceph_inode_info *ci = ceph_inode(inode); | |
3d14c5d2 YS |
121 | struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb); |
122 | struct ceph_mds_client *mdsc = fsc->mdsc; | |
124e68e7 SW |
123 | struct ceph_mds_request *req; |
124 | struct ceph_file_info *cf = file->private_data; | |
125 | struct inode *parent_inode = file->f_dentry->d_parent->d_inode; | |
126 | int err; | |
127 | int flags, fmode, wanted; | |
128 | ||
129 | if (cf) { | |
130 | dout("open file %p is already opened\n", file); | |
131 | return 0; | |
132 | } | |
133 | ||
134 | /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */ | |
135 | flags = file->f_flags & ~(O_CREAT|O_EXCL); | |
136 | if (S_ISDIR(inode->i_mode)) | |
137 | flags = O_DIRECTORY; /* mds likes to know */ | |
138 | ||
139 | dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode, | |
140 | ceph_vinop(inode), file, flags, file->f_flags); | |
141 | fmode = ceph_flags_to_mode(flags); | |
142 | wanted = ceph_caps_for_mode(fmode); | |
143 | ||
144 | /* snapped files are read-only */ | |
145 | if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE)) | |
146 | return -EROFS; | |
147 | ||
148 | /* trivially open snapdir */ | |
149 | if (ceph_snap(inode) == CEPH_SNAPDIR) { | |
150 | spin_lock(&inode->i_lock); | |
151 | __ceph_get_fmode(ci, fmode); | |
152 | spin_unlock(&inode->i_lock); | |
153 | return ceph_init_file(inode, file, fmode); | |
154 | } | |
155 | ||
156 | /* | |
7421ab80 SW |
157 | * No need to block if we have caps on the auth MDS (for |
158 | * write) or any MDS (for read). Update wanted set | |
124e68e7 SW |
159 | * asynchronously. |
160 | */ | |
161 | spin_lock(&inode->i_lock); | |
7421ab80 SW |
162 | if (__ceph_is_any_real_caps(ci) && |
163 | (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) { | |
124e68e7 SW |
164 | int mds_wanted = __ceph_caps_mds_wanted(ci); |
165 | int issued = __ceph_caps_issued(ci, NULL); | |
166 | ||
167 | dout("open %p fmode %d want %s issued %s using existing\n", | |
168 | inode, fmode, ceph_cap_string(wanted), | |
169 | ceph_cap_string(issued)); | |
170 | __ceph_get_fmode(ci, fmode); | |
171 | spin_unlock(&inode->i_lock); | |
172 | ||
173 | /* adjust wanted? */ | |
174 | if ((issued & wanted) != wanted && | |
175 | (mds_wanted & wanted) != wanted && | |
176 | ceph_snap(inode) != CEPH_SNAPDIR) | |
177 | ceph_check_caps(ci, 0, NULL); | |
178 | ||
179 | return ceph_init_file(inode, file, fmode); | |
180 | } else if (ceph_snap(inode) != CEPH_NOSNAP && | |
181 | (ci->i_snap_caps & wanted) == wanted) { | |
182 | __ceph_get_fmode(ci, fmode); | |
183 | spin_unlock(&inode->i_lock); | |
184 | return ceph_init_file(inode, file, fmode); | |
185 | } | |
186 | spin_unlock(&inode->i_lock); | |
187 | ||
188 | dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted)); | |
189 | req = prepare_open_request(inode->i_sb, flags, 0); | |
190 | if (IS_ERR(req)) { | |
191 | err = PTR_ERR(req); | |
192 | goto out; | |
193 | } | |
70b666c3 SW |
194 | req->r_inode = inode; |
195 | ihold(inode); | |
124e68e7 | 196 | req->r_num_caps = 1; |
9bae113a SW |
197 | err = ceph_mdsc_do_request(mdsc, (flags & (O_CREAT|O_TRUNC)) ? |
198 | parent_inode : NULL, req); | |
124e68e7 SW |
199 | if (!err) |
200 | err = ceph_init_file(inode, file, req->r_fmode); | |
201 | ceph_mdsc_put_request(req); | |
202 | dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode)); | |
203 | out: | |
204 | return err; | |
205 | } | |
206 | ||
207 | ||
208 | /* | |
209 | * Do a lookup + open with a single request. | |
210 | * | |
211 | * If this succeeds, but some subsequent check in the vfs | |
212 | * may_open() fails, the struct *file gets cleaned up (i.e. | |
213 | * ceph_release gets called). So fear not! | |
214 | */ | |
215 | /* | |
216 | * flags | |
217 | * path_lookup_open -> LOOKUP_OPEN | |
218 | * path_lookup_create -> LOOKUP_OPEN|LOOKUP_CREATE | |
219 | */ | |
220 | struct dentry *ceph_lookup_open(struct inode *dir, struct dentry *dentry, | |
221 | struct nameidata *nd, int mode, | |
222 | int locked_dir) | |
223 | { | |
3d14c5d2 YS |
224 | struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb); |
225 | struct ceph_mds_client *mdsc = fsc->mdsc; | |
468640e3 | 226 | struct file *file; |
124e68e7 | 227 | struct ceph_mds_request *req; |
468640e3 | 228 | struct dentry *ret; |
124e68e7 SW |
229 | int err; |
230 | int flags = nd->intent.open.flags - 1; /* silly vfs! */ | |
231 | ||
232 | dout("ceph_lookup_open dentry %p '%.*s' flags %d mode 0%o\n", | |
233 | dentry, dentry->d_name.len, dentry->d_name.name, flags, mode); | |
234 | ||
235 | /* do the open */ | |
236 | req = prepare_open_request(dir->i_sb, flags, mode); | |
237 | if (IS_ERR(req)) | |
7e34bc52 | 238 | return ERR_CAST(req); |
124e68e7 SW |
239 | req->r_dentry = dget(dentry); |
240 | req->r_num_caps = 2; | |
241 | if (flags & O_CREAT) { | |
242 | req->r_dentry_drop = CEPH_CAP_FILE_SHARED; | |
243 | req->r_dentry_unless = CEPH_CAP_FILE_EXCL; | |
244 | } | |
245 | req->r_locked_dir = dir; /* caller holds dir->i_mutex */ | |
acda7657 SW |
246 | err = ceph_mdsc_do_request(mdsc, |
247 | (flags & (O_CREAT|O_TRUNC)) ? dir : NULL, | |
248 | req); | |
468640e3 SW |
249 | err = ceph_handle_snapdir(req, dentry, err); |
250 | if (err) | |
251 | goto out; | |
252 | if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry) | |
124e68e7 | 253 | err = ceph_handle_notrace_create(dir, dentry); |
468640e3 SW |
254 | if (err) |
255 | goto out; | |
256 | file = lookup_instantiate_filp(nd, req->r_dentry, ceph_open); | |
257 | if (IS_ERR(file)) | |
258 | err = PTR_ERR(file); | |
259 | out: | |
260 | ret = ceph_finish_lookup(req, dentry, err); | |
124e68e7 | 261 | ceph_mdsc_put_request(req); |
468640e3 SW |
262 | dout("ceph_lookup_open result=%p\n", ret); |
263 | return ret; | |
124e68e7 SW |
264 | } |
265 | ||
266 | int ceph_release(struct inode *inode, struct file *file) | |
267 | { | |
268 | struct ceph_inode_info *ci = ceph_inode(inode); | |
269 | struct ceph_file_info *cf = file->private_data; | |
270 | ||
271 | dout("release inode %p file %p\n", inode, file); | |
272 | ceph_put_fmode(ci, cf->fmode); | |
273 | if (cf->last_readdir) | |
274 | ceph_mdsc_put_request(cf->last_readdir); | |
275 | kfree(cf->last_name); | |
276 | kfree(cf->dir_info); | |
277 | dput(cf->dentry); | |
278 | kmem_cache_free(ceph_file_cachep, cf); | |
195d3ce2 SW |
279 | |
280 | /* wake up anyone waiting for caps on this inode */ | |
03066f23 | 281 | wake_up_all(&ci->i_cap_wq); |
124e68e7 SW |
282 | return 0; |
283 | } | |
284 | ||
124e68e7 SW |
285 | /* |
286 | * Read a range of bytes striped over one or more objects. Iterate over | |
287 | * objects we stripe over. (That's not atomic, but good enough for now.) | |
288 | * | |
289 | * If we get a short result from the OSD, check against i_size; we need to | |
290 | * only return a short read to the caller if we hit EOF. | |
291 | */ | |
292 | static int striped_read(struct inode *inode, | |
293 | u64 off, u64 len, | |
6a026589 | 294 | struct page **pages, int num_pages, |
c3cd6283 | 295 | int *checkeof, bool o_direct, |
ab226e21 | 296 | unsigned long buf_align) |
124e68e7 | 297 | { |
3d14c5d2 | 298 | struct ceph_fs_client *fsc = ceph_inode_to_client(inode); |
124e68e7 SW |
299 | struct ceph_inode_info *ci = ceph_inode(inode); |
300 | u64 pos, this_len; | |
b7495fc2 | 301 | int io_align, page_align; |
124e68e7 SW |
302 | int left, pages_left; |
303 | int read; | |
304 | struct page **page_pos; | |
305 | int ret; | |
306 | bool hit_stripe, was_short; | |
307 | ||
308 | /* | |
309 | * we may need to do multiple reads. not atomic, unfortunately. | |
310 | */ | |
311 | pos = off; | |
312 | left = len; | |
313 | page_pos = pages; | |
314 | pages_left = num_pages; | |
315 | read = 0; | |
b7495fc2 | 316 | io_align = off & ~PAGE_MASK; |
124e68e7 SW |
317 | |
318 | more: | |
c3cd6283 | 319 | if (o_direct) |
ab226e21 | 320 | page_align = (pos - io_align + buf_align) & ~PAGE_MASK; |
b7495fc2 SW |
321 | else |
322 | page_align = pos & ~PAGE_MASK; | |
124e68e7 | 323 | this_len = left; |
3d14c5d2 | 324 | ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode), |
124e68e7 SW |
325 | &ci->i_layout, pos, &this_len, |
326 | ci->i_truncate_seq, | |
327 | ci->i_truncate_size, | |
b7495fc2 | 328 | page_pos, pages_left, page_align); |
124e68e7 SW |
329 | if (ret == -ENOENT) |
330 | ret = 0; | |
0e98728f SW |
331 | hit_stripe = this_len < left; |
332 | was_short = ret >= 0 && ret < this_len; | |
124e68e7 SW |
333 | dout("striped_read %llu~%u (read %u) got %d%s%s\n", pos, left, read, |
334 | ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : ""); | |
335 | ||
336 | if (ret > 0) { | |
773e9b44 | 337 | int didpages = (page_align + ret) >> PAGE_CACHE_SHIFT; |
124e68e7 SW |
338 | |
339 | if (read < pos - off) { | |
340 | dout(" zero gap %llu to %llu\n", off + read, pos); | |
773e9b44 | 341 | ceph_zero_page_vector_range(page_align + read, |
3d14c5d2 | 342 | pos - off - read, pages); |
124e68e7 SW |
343 | } |
344 | pos += ret; | |
345 | read = pos - off; | |
346 | left -= ret; | |
347 | page_pos += didpages; | |
348 | pages_left -= didpages; | |
349 | ||
350 | /* hit stripe? */ | |
351 | if (left && hit_stripe) | |
352 | goto more; | |
353 | } | |
354 | ||
355 | if (was_short) { | |
c3cd6283 SW |
356 | /* did we bounce off eof? */ |
357 | if (pos + left > inode->i_size) | |
358 | *checkeof = 1; | |
359 | ||
360 | /* zero trailing bytes (inside i_size) */ | |
361 | if (left > 0 && pos < inode->i_size) { | |
362 | if (pos + left > inode->i_size) | |
363 | left = inode->i_size - pos; | |
364 | ||
365 | dout("zero tail %d\n", left); | |
773e9b44 | 366 | ceph_zero_page_vector_range(page_align + read, left, |
3d14c5d2 | 367 | pages); |
c3cd6283 | 368 | read += left; |
124e68e7 | 369 | } |
124e68e7 SW |
370 | } |
371 | ||
124e68e7 SW |
372 | if (ret >= 0) |
373 | ret = read; | |
374 | dout("striped_read returns %d\n", ret); | |
375 | return ret; | |
376 | } | |
377 | ||
378 | /* | |
379 | * Completely synchronous read and write methods. Direct from __user | |
380 | * buffer to osd, or directly to user pages (if O_DIRECT). | |
381 | * | |
382 | * If the read spans object boundary, just do multiple reads. | |
383 | */ | |
384 | static ssize_t ceph_sync_read(struct file *file, char __user *data, | |
6a026589 | 385 | unsigned len, loff_t *poff, int *checkeof) |
124e68e7 SW |
386 | { |
387 | struct inode *inode = file->f_dentry->d_inode; | |
388 | struct page **pages; | |
389 | u64 off = *poff; | |
ab226e21 | 390 | int num_pages, ret; |
124e68e7 SW |
391 | |
392 | dout("sync_read on file %p %llu~%u %s\n", file, off, len, | |
393 | (file->f_flags & O_DIRECT) ? "O_DIRECT" : ""); | |
394 | ||
ab226e21 HC |
395 | if (file->f_flags & O_DIRECT) { |
396 | num_pages = calc_pages_for((unsigned long)data, len); | |
b6aa5901 | 397 | pages = ceph_get_direct_page_vector(data, num_pages, true); |
ab226e21 HC |
398 | } else { |
399 | num_pages = calc_pages_for(off, len); | |
34d23762 | 400 | pages = ceph_alloc_page_vector(num_pages, GFP_NOFS); |
ab226e21 | 401 | } |
124e68e7 SW |
402 | if (IS_ERR(pages)) |
403 | return PTR_ERR(pages); | |
404 | ||
e98b6fed SW |
405 | /* |
406 | * flush any page cache pages in this range. this | |
407 | * will make concurrent normal and sync io slow, | |
408 | * but it will at least behave sensibly when they are | |
409 | * in sequence. | |
410 | */ | |
29065a51 YS |
411 | ret = filemap_write_and_wait(inode->i_mapping); |
412 | if (ret < 0) | |
413 | goto done; | |
414 | ||
b7495fc2 | 415 | ret = striped_read(inode, off, len, pages, num_pages, checkeof, |
ab226e21 HC |
416 | file->f_flags & O_DIRECT, |
417 | (unsigned long)data & ~PAGE_MASK); | |
124e68e7 SW |
418 | |
419 | if (ret >= 0 && (file->f_flags & O_DIRECT) == 0) | |
3d14c5d2 | 420 | ret = ceph_copy_page_vector_to_user(pages, data, off, ret); |
124e68e7 SW |
421 | if (ret >= 0) |
422 | *poff = off + ret; | |
423 | ||
29065a51 | 424 | done: |
124e68e7 | 425 | if (file->f_flags & O_DIRECT) |
b6aa5901 | 426 | ceph_put_page_vector(pages, num_pages, true); |
124e68e7 SW |
427 | else |
428 | ceph_release_page_vector(pages, num_pages); | |
429 | dout("sync_read result %d\n", ret); | |
430 | return ret; | |
431 | } | |
432 | ||
433 | /* | |
434 | * Write commit callback, called if we requested both an ACK and | |
435 | * ONDISK commit reply from the OSD. | |
436 | */ | |
437 | static void sync_write_commit(struct ceph_osd_request *req, | |
438 | struct ceph_msg *msg) | |
439 | { | |
440 | struct ceph_inode_info *ci = ceph_inode(req->r_inode); | |
441 | ||
442 | dout("sync_write_commit %p tid %llu\n", req, req->r_tid); | |
443 | spin_lock(&ci->i_unsafe_lock); | |
444 | list_del_init(&req->r_unsafe_item); | |
445 | spin_unlock(&ci->i_unsafe_lock); | |
446 | ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR); | |
447 | } | |
448 | ||
449 | /* | |
450 | * Synchronous write, straight from __user pointer or user pages (if | |
451 | * O_DIRECT). | |
452 | * | |
453 | * If write spans object boundary, just do multiple writes. (For a | |
454 | * correct atomic write, we should e.g. take write locks on all | |
455 | * objects, rollback on failure, etc.) | |
456 | */ | |
457 | static ssize_t ceph_sync_write(struct file *file, const char __user *data, | |
458 | size_t left, loff_t *offset) | |
459 | { | |
460 | struct inode *inode = file->f_dentry->d_inode; | |
461 | struct ceph_inode_info *ci = ceph_inode(inode); | |
3d14c5d2 | 462 | struct ceph_fs_client *fsc = ceph_inode_to_client(inode); |
124e68e7 SW |
463 | struct ceph_osd_request *req; |
464 | struct page **pages; | |
465 | int num_pages; | |
466 | long long unsigned pos; | |
467 | u64 len; | |
468 | int written = 0; | |
469 | int flags; | |
470 | int do_sync = 0; | |
471 | int check_caps = 0; | |
b7495fc2 | 472 | int page_align, io_align; |
ab226e21 | 473 | unsigned long buf_align; |
124e68e7 SW |
474 | int ret; |
475 | struct timespec mtime = CURRENT_TIME; | |
476 | ||
477 | if (ceph_snap(file->f_dentry->d_inode) != CEPH_NOSNAP) | |
478 | return -EROFS; | |
479 | ||
480 | dout("sync_write on file %p %lld~%u %s\n", file, *offset, | |
481 | (unsigned)left, (file->f_flags & O_DIRECT) ? "O_DIRECT" : ""); | |
482 | ||
483 | if (file->f_flags & O_APPEND) | |
484 | pos = i_size_read(inode); | |
485 | else | |
486 | pos = *offset; | |
487 | ||
29065a51 YS |
488 | ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + left); |
489 | if (ret < 0) | |
490 | return ret; | |
491 | ||
492 | ret = invalidate_inode_pages2_range(inode->i_mapping, | |
493 | pos >> PAGE_CACHE_SHIFT, | |
494 | (pos + left) >> PAGE_CACHE_SHIFT); | |
495 | if (ret < 0) | |
496 | dout("invalidate_inode_pages2_range returned %d\n", ret); | |
497 | ||
124e68e7 SW |
498 | flags = CEPH_OSD_FLAG_ORDERSNAP | |
499 | CEPH_OSD_FLAG_ONDISK | | |
500 | CEPH_OSD_FLAG_WRITE; | |
501 | if ((file->f_flags & (O_SYNC|O_DIRECT)) == 0) | |
502 | flags |= CEPH_OSD_FLAG_ACK; | |
503 | else | |
504 | do_sync = 1; | |
505 | ||
506 | /* | |
507 | * we may need to do multiple writes here if we span an object | |
508 | * boundary. this isn't atomic, unfortunately. :( | |
509 | */ | |
510 | more: | |
d7f124f1 SW |
511 | io_align = pos & ~PAGE_MASK; |
512 | buf_align = (unsigned long)data & ~PAGE_MASK; | |
124e68e7 | 513 | len = left; |
ab226e21 | 514 | if (file->f_flags & O_DIRECT) { |
b7495fc2 SW |
515 | /* write from beginning of first page, regardless of |
516 | io alignment */ | |
ab226e21 HC |
517 | page_align = (pos - io_align + buf_align) & ~PAGE_MASK; |
518 | num_pages = calc_pages_for((unsigned long)data, len); | |
519 | } else { | |
b7495fc2 | 520 | page_align = pos & ~PAGE_MASK; |
ab226e21 HC |
521 | num_pages = calc_pages_for(pos, len); |
522 | } | |
3d14c5d2 | 523 | req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, |
124e68e7 SW |
524 | ceph_vino(inode), pos, &len, |
525 | CEPH_OSD_OP_WRITE, flags, | |
526 | ci->i_snap_realm->cached_context, | |
527 | do_sync, | |
528 | ci->i_truncate_seq, ci->i_truncate_size, | |
b7495fc2 | 529 | &mtime, false, 2, page_align); |
a79832f2 SW |
530 | if (!req) |
531 | return -ENOMEM; | |
124e68e7 | 532 | |
124e68e7 | 533 | if (file->f_flags & O_DIRECT) { |
b6aa5901 | 534 | pages = ceph_get_direct_page_vector(data, num_pages, false); |
124e68e7 SW |
535 | if (IS_ERR(pages)) { |
536 | ret = PTR_ERR(pages); | |
537 | goto out; | |
538 | } | |
539 | ||
540 | /* | |
541 | * throw out any page cache pages in this range. this | |
542 | * may block. | |
543 | */ | |
213c99ee | 544 | truncate_inode_pages_range(inode->i_mapping, pos, |
5c6a2cdb | 545 | (pos+len) | (PAGE_CACHE_SIZE-1)); |
124e68e7 | 546 | } else { |
34d23762 | 547 | pages = ceph_alloc_page_vector(num_pages, GFP_NOFS); |
124e68e7 SW |
548 | if (IS_ERR(pages)) { |
549 | ret = PTR_ERR(pages); | |
550 | goto out; | |
551 | } | |
3d14c5d2 | 552 | ret = ceph_copy_user_to_page_vector(pages, data, pos, len); |
124e68e7 SW |
553 | if (ret < 0) { |
554 | ceph_release_page_vector(pages, num_pages); | |
555 | goto out; | |
556 | } | |
557 | ||
558 | if ((file->f_flags & O_SYNC) == 0) { | |
559 | /* get a second commit callback */ | |
560 | req->r_safe_callback = sync_write_commit; | |
561 | req->r_own_pages = 1; | |
562 | } | |
563 | } | |
564 | req->r_pages = pages; | |
565 | req->r_num_pages = num_pages; | |
566 | req->r_inode = inode; | |
567 | ||
3d14c5d2 | 568 | ret = ceph_osdc_start_request(&fsc->client->osdc, req, false); |
124e68e7 SW |
569 | if (!ret) { |
570 | if (req->r_safe_callback) { | |
571 | /* | |
572 | * Add to inode unsafe list only after we | |
573 | * start_request so that a tid has been assigned. | |
574 | */ | |
575 | spin_lock(&ci->i_unsafe_lock); | |
49bcb932 HC |
576 | list_add_tail(&req->r_unsafe_item, |
577 | &ci->i_unsafe_writes); | |
124e68e7 SW |
578 | spin_unlock(&ci->i_unsafe_lock); |
579 | ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR); | |
580 | } | |
78a25565 | 581 | |
3d14c5d2 | 582 | ret = ceph_osdc_wait_request(&fsc->client->osdc, req); |
78a25565 HC |
583 | if (ret < 0 && req->r_safe_callback) { |
584 | spin_lock(&ci->i_unsafe_lock); | |
585 | list_del_init(&req->r_unsafe_item); | |
586 | spin_unlock(&ci->i_unsafe_lock); | |
587 | ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR); | |
588 | } | |
124e68e7 SW |
589 | } |
590 | ||
591 | if (file->f_flags & O_DIRECT) | |
b6aa5901 | 592 | ceph_put_page_vector(pages, num_pages, false); |
124e68e7 SW |
593 | else if (file->f_flags & O_SYNC) |
594 | ceph_release_page_vector(pages, num_pages); | |
595 | ||
596 | out: | |
597 | ceph_osdc_put_request(req); | |
598 | if (ret == 0) { | |
599 | pos += len; | |
600 | written += len; | |
601 | left -= len; | |
d7f124f1 | 602 | data += written; |
124e68e7 SW |
603 | if (left) |
604 | goto more; | |
605 | ||
606 | ret = written; | |
607 | *offset = pos; | |
608 | if (pos > i_size_read(inode)) | |
609 | check_caps = ceph_inode_set_size(inode, pos); | |
610 | if (check_caps) | |
611 | ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, | |
612 | NULL); | |
613 | } | |
614 | return ret; | |
615 | } | |
616 | ||
617 | /* | |
618 | * Wrap generic_file_aio_read with checks for cap bits on the inode. | |
619 | * Atomically grab references, so that those bits are not released | |
620 | * back to the MDS mid-read. | |
621 | * | |
622 | * Hmm, the sync read case isn't actually async... should it be? | |
623 | */ | |
624 | static ssize_t ceph_aio_read(struct kiocb *iocb, const struct iovec *iov, | |
625 | unsigned long nr_segs, loff_t pos) | |
626 | { | |
627 | struct file *filp = iocb->ki_filp; | |
2962507c | 628 | struct ceph_file_info *fi = filp->private_data; |
124e68e7 SW |
629 | loff_t *ppos = &iocb->ki_pos; |
630 | size_t len = iov->iov_len; | |
631 | struct inode *inode = filp->f_dentry->d_inode; | |
632 | struct ceph_inode_info *ci = ceph_inode(inode); | |
cd84db6e | 633 | void __user *base = iov->iov_base; |
124e68e7 | 634 | ssize_t ret; |
2962507c | 635 | int want, got = 0; |
6a026589 | 636 | int checkeof = 0, read = 0; |
124e68e7 SW |
637 | |
638 | dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n", | |
639 | inode, ceph_vinop(inode), pos, (unsigned)len, inode); | |
6a026589 | 640 | again: |
124e68e7 | 641 | __ceph_do_pending_vmtruncate(inode); |
2962507c SW |
642 | if (fi->fmode & CEPH_FILE_MODE_LAZY) |
643 | want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO; | |
644 | else | |
645 | want = CEPH_CAP_FILE_CACHE; | |
646 | ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, &got, -1); | |
124e68e7 SW |
647 | if (ret < 0) |
648 | goto out; | |
649 | dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n", | |
650 | inode, ceph_vinop(inode), pos, (unsigned)len, | |
651 | ceph_cap_string(got)); | |
652 | ||
2962507c | 653 | if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 || |
124e68e7 | 654 | (iocb->ki_filp->f_flags & O_DIRECT) || |
4918b6d1 SW |
655 | (inode->i_sb->s_flags & MS_SYNCHRONOUS) || |
656 | (fi->flags & CEPH_F_SYNC)) | |
124e68e7 | 657 | /* hmm, this isn't really async... */ |
6a026589 | 658 | ret = ceph_sync_read(filp, base, len, ppos, &checkeof); |
124e68e7 SW |
659 | else |
660 | ret = generic_file_aio_read(iocb, iov, nr_segs, pos); | |
661 | ||
662 | out: | |
663 | dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n", | |
664 | inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret); | |
665 | ceph_put_cap_refs(ci, got); | |
6a026589 SW |
666 | |
667 | if (checkeof && ret >= 0) { | |
668 | int statret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE); | |
669 | ||
670 | /* hit EOF or hole? */ | |
671 | if (statret == 0 && *ppos < inode->i_size) { | |
c3cd6283 | 672 | dout("aio_read sync_read hit hole, ppos %lld < size %lld, reading more\n", *ppos, inode->i_size); |
6a026589 SW |
673 | read += ret; |
674 | base += ret; | |
675 | len -= ret; | |
676 | checkeof = 0; | |
677 | goto again; | |
678 | } | |
679 | } | |
680 | if (ret >= 0) | |
681 | ret += read; | |
682 | ||
124e68e7 SW |
683 | return ret; |
684 | } | |
685 | ||
686 | /* | |
687 | * Take cap references to avoid releasing caps to MDS mid-write. | |
688 | * | |
689 | * If we are synchronous, and write with an old snap context, the OSD | |
690 | * may return EOLDSNAPC. In that case, retry the write.. _after_ | |
691 | * dropping our cap refs and allowing the pending snap to logically | |
692 | * complete _before_ this write occurs. | |
693 | * | |
694 | * If we are near ENOSPC, write synchronously. | |
695 | */ | |
696 | static ssize_t ceph_aio_write(struct kiocb *iocb, const struct iovec *iov, | |
697 | unsigned long nr_segs, loff_t pos) | |
698 | { | |
699 | struct file *file = iocb->ki_filp; | |
33caad32 | 700 | struct ceph_file_info *fi = file->private_data; |
124e68e7 SW |
701 | struct inode *inode = file->f_dentry->d_inode; |
702 | struct ceph_inode_info *ci = ceph_inode(inode); | |
3d14c5d2 YS |
703 | struct ceph_osd_client *osdc = |
704 | &ceph_sb_to_client(inode->i_sb)->client->osdc; | |
124e68e7 | 705 | loff_t endoff = pos + iov->iov_len; |
33caad32 | 706 | int want, got = 0; |
88d892a3 | 707 | int ret, err; |
124e68e7 SW |
708 | |
709 | if (ceph_snap(inode) != CEPH_NOSNAP) | |
710 | return -EROFS; | |
711 | ||
712 | retry_snap: | |
713 | if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) | |
714 | return -ENOSPC; | |
715 | __ceph_do_pending_vmtruncate(inode); | |
716 | dout("aio_write %p %llx.%llx %llu~%u getting caps. i_size %llu\n", | |
717 | inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len, | |
718 | inode->i_size); | |
33caad32 SW |
719 | if (fi->fmode & CEPH_FILE_MODE_LAZY) |
720 | want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO; | |
721 | else | |
722 | want = CEPH_CAP_FILE_BUFFER; | |
723 | ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, endoff); | |
124e68e7 | 724 | if (ret < 0) |
d8de9ab6 | 725 | goto out_put; |
124e68e7 SW |
726 | |
727 | dout("aio_write %p %llx.%llx %llu~%u got cap refs on %s\n", | |
728 | inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len, | |
729 | ceph_cap_string(got)); | |
730 | ||
33caad32 | 731 | if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 || |
124e68e7 | 732 | (iocb->ki_filp->f_flags & O_DIRECT) || |
4918b6d1 SW |
733 | (inode->i_sb->s_flags & MS_SYNCHRONOUS) || |
734 | (fi->flags & CEPH_F_SYNC)) { | |
124e68e7 SW |
735 | ret = ceph_sync_write(file, iov->iov_base, iov->iov_len, |
736 | &iocb->ki_pos); | |
737 | } else { | |
d8de9ab6 SW |
738 | /* |
739 | * buffered write; drop Fw early to avoid slow | |
740 | * revocation if we get stuck on balance_dirty_pages | |
741 | */ | |
742 | int dirty; | |
743 | ||
744 | spin_lock(&inode->i_lock); | |
745 | dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR); | |
746 | spin_unlock(&inode->i_lock); | |
747 | ceph_put_cap_refs(ci, got); | |
124e68e7 | 748 | |
d8de9ab6 | 749 | ret = generic_file_aio_write(iocb, iov, nr_segs, pos); |
124e68e7 SW |
750 | if ((ret >= 0 || ret == -EIOCBQUEUED) && |
751 | ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host) | |
88d892a3 | 752 | || ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) { |
8018ab05 | 753 | err = vfs_fsync_range(file, pos, pos + ret - 1, 1); |
88d892a3 YS |
754 | if (err < 0) |
755 | ret = err; | |
756 | } | |
d8de9ab6 SW |
757 | |
758 | if (dirty) | |
759 | __mark_inode_dirty(inode, dirty); | |
760 | goto out; | |
124e68e7 | 761 | } |
d8de9ab6 | 762 | |
124e68e7 | 763 | if (ret >= 0) { |
fca65b4a | 764 | int dirty; |
124e68e7 | 765 | spin_lock(&inode->i_lock); |
fca65b4a | 766 | dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR); |
124e68e7 | 767 | spin_unlock(&inode->i_lock); |
fca65b4a SW |
768 | if (dirty) |
769 | __mark_inode_dirty(inode, dirty); | |
124e68e7 SW |
770 | } |
771 | ||
d8de9ab6 | 772 | out_put: |
124e68e7 SW |
773 | dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n", |
774 | inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len, | |
775 | ceph_cap_string(got)); | |
776 | ceph_put_cap_refs(ci, got); | |
777 | ||
d8de9ab6 | 778 | out: |
124e68e7 SW |
779 | if (ret == -EOLDSNAPC) { |
780 | dout("aio_write %p %llx.%llx %llu~%u got EOLDSNAPC, retrying\n", | |
781 | inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len); | |
782 | goto retry_snap; | |
783 | } | |
784 | ||
785 | return ret; | |
786 | } | |
787 | ||
788 | /* | |
789 | * llseek. be sure to verify file size on SEEK_END. | |
790 | */ | |
791 | static loff_t ceph_llseek(struct file *file, loff_t offset, int origin) | |
792 | { | |
793 | struct inode *inode = file->f_mapping->host; | |
794 | int ret; | |
795 | ||
796 | mutex_lock(&inode->i_mutex); | |
797 | __ceph_do_pending_vmtruncate(inode); | |
798 | switch (origin) { | |
799 | case SEEK_END: | |
800 | ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE); | |
801 | if (ret < 0) { | |
802 | offset = ret; | |
803 | goto out; | |
804 | } | |
805 | offset += inode->i_size; | |
806 | break; | |
807 | case SEEK_CUR: | |
808 | /* | |
809 | * Here we special-case the lseek(fd, 0, SEEK_CUR) | |
810 | * position-querying operation. Avoid rewriting the "same" | |
811 | * f_pos value back to the file because a concurrent read(), | |
812 | * write() or lseek() might have altered it | |
813 | */ | |
814 | if (offset == 0) { | |
815 | offset = file->f_pos; | |
816 | goto out; | |
817 | } | |
818 | offset += file->f_pos; | |
819 | break; | |
820 | } | |
821 | ||
822 | if (offset < 0 || offset > inode->i_sb->s_maxbytes) { | |
823 | offset = -EINVAL; | |
824 | goto out; | |
825 | } | |
826 | ||
827 | /* Special lock needed here? */ | |
828 | if (offset != file->f_pos) { | |
829 | file->f_pos = offset; | |
830 | file->f_version = 0; | |
831 | } | |
832 | ||
833 | out: | |
834 | mutex_unlock(&inode->i_mutex); | |
835 | return offset; | |
836 | } | |
837 | ||
838 | const struct file_operations ceph_file_fops = { | |
839 | .open = ceph_open, | |
840 | .release = ceph_release, | |
841 | .llseek = ceph_llseek, | |
842 | .read = do_sync_read, | |
843 | .write = do_sync_write, | |
844 | .aio_read = ceph_aio_read, | |
845 | .aio_write = ceph_aio_write, | |
846 | .mmap = ceph_mmap, | |
847 | .fsync = ceph_fsync, | |
40819f6f GF |
848 | .lock = ceph_lock, |
849 | .flock = ceph_flock, | |
124e68e7 SW |
850 | .splice_read = generic_file_splice_read, |
851 | .splice_write = generic_file_splice_write, | |
852 | .unlocked_ioctl = ceph_ioctl, | |
853 | .compat_ioctl = ceph_ioctl, | |
854 | }; | |
855 |