]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
3d14c5d2 | 2 | #include <linux/ceph/ceph_debug.h> |
2817b000 SW |
3 | |
4 | #include <linux/spinlock.h> | |
5 | #include <linux/fs_struct.h> | |
6 | #include <linux/namei.h> | |
5a0e3ad6 | 7 | #include <linux/slab.h> |
2817b000 | 8 | #include <linux/sched.h> |
2cdeb1e4 | 9 | #include <linux/xattr.h> |
2817b000 SW |
10 | |
11 | #include "super.h" | |
3d14c5d2 | 12 | #include "mds_client.h" |
2817b000 SW |
13 | |
14 | /* | |
15 | * Directory operations: readdir, lookup, create, link, unlink, | |
16 | * rename, etc. | |
17 | */ | |
18 | ||
19 | /* | |
20 | * Ceph MDS operations are specified in terms of a base ino and | |
21 | * relative path. Thus, the client can specify an operation on a | |
22 | * specific inode (e.g., a getattr due to fstat(2)), or as a path | |
23 | * relative to, say, the root directory. | |
24 | * | |
25 | * Normally, we limit ourselves to strict inode ops (no path component) | |
26 | * or dentry operations (a single path component relative to an ino). The | |
27 | * exception to this is open_root_dentry(), which will open the mount | |
28 | * point by name. | |
29 | */ | |
30 | ||
52dfb8ac | 31 | const struct dentry_operations ceph_dentry_ops; |
2817b000 SW |
32 | |
33 | /* | |
34 | * Initialize ceph dentry state. | |
35 | */ | |
ad5cb123 | 36 | static int ceph_d_init(struct dentry *dentry) |
2817b000 SW |
37 | { |
38 | struct ceph_dentry_info *di; | |
39 | ||
99ec2697 | 40 | di = kmem_cache_zalloc(ceph_dentry_cachep, GFP_KERNEL); |
2817b000 SW |
41 | if (!di) |
42 | return -ENOMEM; /* oh well */ | |
43 | ||
2817b000 SW |
44 | di->dentry = dentry; |
45 | di->lease_session = NULL; | |
9b16f03c | 46 | di->time = jiffies; |
48d0cbd1 | 47 | dentry->d_fsdata = di; |
2817b000 | 48 | ceph_dentry_lru_add(dentry); |
2817b000 SW |
49 | return 0; |
50 | } | |
51 | ||
2817b000 | 52 | /* |
f3c4ebe6 YZ |
53 | * for f_pos for readdir: |
54 | * - hash order: | |
55 | * (0xff << 52) | ((24 bits hash) << 28) | | |
56 | * (the nth entry has hash collision); | |
57 | * - frag+name order; | |
58 | * ((frag value) << 28) | (the nth entry in frag); | |
2817b000 | 59 | */ |
f3c4ebe6 YZ |
60 | #define OFFSET_BITS 28 |
61 | #define OFFSET_MASK ((1 << OFFSET_BITS) - 1) | |
62 | #define HASH_ORDER (0xffull << (OFFSET_BITS + 24)) | |
63 | loff_t ceph_make_fpos(unsigned high, unsigned off, bool hash_order) | |
64 | { | |
65 | loff_t fpos = ((loff_t)high << 28) | (loff_t)off; | |
66 | if (hash_order) | |
67 | fpos |= HASH_ORDER; | |
68 | return fpos; | |
69 | } | |
70 | ||
71 | static bool is_hash_order(loff_t p) | |
72 | { | |
73 | return (p & HASH_ORDER) == HASH_ORDER; | |
74 | } | |
75 | ||
2817b000 SW |
76 | static unsigned fpos_frag(loff_t p) |
77 | { | |
f3c4ebe6 | 78 | return p >> OFFSET_BITS; |
2817b000 | 79 | } |
f3c4ebe6 YZ |
80 | |
81 | static unsigned fpos_hash(loff_t p) | |
82 | { | |
83 | return ceph_frag_value(fpos_frag(p)); | |
84 | } | |
85 | ||
2817b000 SW |
86 | static unsigned fpos_off(loff_t p) |
87 | { | |
f3c4ebe6 | 88 | return p & OFFSET_MASK; |
2817b000 SW |
89 | } |
90 | ||
4d5f5df6 YZ |
91 | static int fpos_cmp(loff_t l, loff_t r) |
92 | { | |
93 | int v = ceph_frag_compare(fpos_frag(l), fpos_frag(r)); | |
94 | if (v) | |
95 | return v; | |
96 | return (int)(fpos_off(l) - fpos_off(r)); | |
97 | } | |
98 | ||
fdd4e158 YZ |
99 | /* |
100 | * make note of the last dentry we read, so we can | |
101 | * continue at the same lexicographical point, | |
102 | * regardless of what dir changes take place on the | |
103 | * server. | |
104 | */ | |
105 | static int note_last_dentry(struct ceph_file_info *fi, const char *name, | |
106 | int len, unsigned next_offset) | |
107 | { | |
108 | char *buf = kmalloc(len+1, GFP_KERNEL); | |
109 | if (!buf) | |
110 | return -ENOMEM; | |
111 | kfree(fi->last_name); | |
112 | fi->last_name = buf; | |
113 | memcpy(fi->last_name, name, len); | |
114 | fi->last_name[len] = 0; | |
115 | fi->next_offset = next_offset; | |
116 | dout("note_last_dentry '%s'\n", fi->last_name); | |
117 | return 0; | |
118 | } | |
119 | ||
c530cd24 YZ |
120 | |
121 | static struct dentry * | |
122 | __dcache_find_get_entry(struct dentry *parent, u64 idx, | |
123 | struct ceph_readdir_cache_control *cache_ctl) | |
124 | { | |
125 | struct inode *dir = d_inode(parent); | |
126 | struct dentry *dentry; | |
127 | unsigned idx_mask = (PAGE_SIZE / sizeof(struct dentry *)) - 1; | |
128 | loff_t ptr_pos = idx * sizeof(struct dentry *); | |
129 | pgoff_t ptr_pgoff = ptr_pos >> PAGE_SHIFT; | |
130 | ||
131 | if (ptr_pos >= i_size_read(dir)) | |
132 | return NULL; | |
133 | ||
134 | if (!cache_ctl->page || ptr_pgoff != page_index(cache_ctl->page)) { | |
135 | ceph_readdir_cache_release(cache_ctl); | |
136 | cache_ctl->page = find_lock_page(&dir->i_data, ptr_pgoff); | |
137 | if (!cache_ctl->page) { | |
138 | dout(" page %lu not found\n", ptr_pgoff); | |
139 | return ERR_PTR(-EAGAIN); | |
140 | } | |
141 | /* reading/filling the cache are serialized by | |
142 | i_mutex, no need to use page lock */ | |
143 | unlock_page(cache_ctl->page); | |
144 | cache_ctl->dentries = kmap(cache_ctl->page); | |
145 | } | |
146 | ||
147 | cache_ctl->index = idx & idx_mask; | |
148 | ||
149 | rcu_read_lock(); | |
150 | spin_lock(&parent->d_lock); | |
151 | /* check i_size again here, because empty directory can be | |
152 | * marked as complete while not holding the i_mutex. */ | |
153 | if (ceph_dir_is_complete_ordered(dir) && ptr_pos < i_size_read(dir)) | |
154 | dentry = cache_ctl->dentries[cache_ctl->index]; | |
155 | else | |
156 | dentry = NULL; | |
157 | spin_unlock(&parent->d_lock); | |
158 | if (dentry && !lockref_get_not_dead(&dentry->d_lockref)) | |
159 | dentry = NULL; | |
160 | rcu_read_unlock(); | |
161 | return dentry ? : ERR_PTR(-EAGAIN); | |
162 | } | |
163 | ||
2817b000 SW |
164 | /* |
165 | * When possible, we try to satisfy a readdir by peeking at the | |
166 | * dcache. We make this work by carefully ordering dentries on | |
946e51f2 | 167 | * d_child when we initially get results back from the MDS, and |
2817b000 SW |
168 | * falling back to a "normal" sync readdir if any dentries in the dir |
169 | * are dropped. | |
170 | * | |
2f276c51 | 171 | * Complete dir indicates that we have all dentries in the dir. It is |
2817b000 SW |
172 | * defined IFF we hold CEPH_CAP_FILE_SHARED (which will be revoked by |
173 | * the MDS if/when the directory is modified). | |
174 | */ | |
a30be7cb | 175 | static int __dcache_readdir(struct file *file, struct dir_context *ctx, |
7389f7b2 | 176 | int shared_gen) |
2817b000 | 177 | { |
77acfa29 | 178 | struct ceph_file_info *fi = file->private_data; |
b583043e | 179 | struct dentry *parent = file->f_path.dentry; |
2b0143b5 | 180 | struct inode *dir = d_inode(parent); |
fdd4e158 | 181 | struct dentry *dentry, *last = NULL; |
2817b000 | 182 | struct ceph_dentry_info *di; |
fdd4e158 | 183 | struct ceph_readdir_cache_control cache_ctl = {}; |
c530cd24 YZ |
184 | u64 idx = 0; |
185 | int err = 0; | |
2817b000 | 186 | |
7389f7b2 | 187 | dout("__dcache_readdir %p v%u at %llx\n", dir, (unsigned)shared_gen, ctx->pos); |
2817b000 | 188 | |
c530cd24 YZ |
189 | /* search start position */ |
190 | if (ctx->pos > 2) { | |
191 | u64 count = div_u64(i_size_read(dir), sizeof(struct dentry *)); | |
192 | while (count > 0) { | |
193 | u64 step = count >> 1; | |
194 | dentry = __dcache_find_get_entry(parent, idx + step, | |
195 | &cache_ctl); | |
196 | if (!dentry) { | |
197 | /* use linar search */ | |
198 | idx = 0; | |
199 | break; | |
200 | } | |
201 | if (IS_ERR(dentry)) { | |
202 | err = PTR_ERR(dentry); | |
203 | goto out; | |
204 | } | |
205 | di = ceph_dentry(dentry); | |
206 | spin_lock(&dentry->d_lock); | |
207 | if (fpos_cmp(di->offset, ctx->pos) < 0) { | |
208 | idx += step + 1; | |
209 | count -= step + 1; | |
210 | } else { | |
211 | count = step; | |
212 | } | |
213 | spin_unlock(&dentry->d_lock); | |
214 | dput(dentry); | |
215 | } | |
2817b000 | 216 | |
c530cd24 | 217 | dout("__dcache_readdir %p cache idx %llu\n", dir, idx); |
2817b000 SW |
218 | } |
219 | ||
fdd4e158 | 220 | |
c530cd24 YZ |
221 | for (;;) { |
222 | bool emit_dentry = false; | |
223 | dentry = __dcache_find_get_entry(parent, idx++, &cache_ctl); | |
224 | if (!dentry) { | |
9cfa1098 | 225 | fi->flags |= CEPH_F_ATEND; |
fdd4e158 YZ |
226 | err = 0; |
227 | break; | |
2817b000 | 228 | } |
c530cd24 YZ |
229 | if (IS_ERR(dentry)) { |
230 | err = PTR_ERR(dentry); | |
231 | goto out; | |
fdd4e158 YZ |
232 | } |
233 | ||
fdd4e158 | 234 | spin_lock(&dentry->d_lock); |
48697e8a YZ |
235 | di = ceph_dentry(dentry); |
236 | if (d_unhashed(dentry) || | |
237 | d_really_is_negative(dentry) || | |
238 | di->lease_shared_gen != shared_gen) { | |
239 | spin_unlock(&dentry->d_lock); | |
240 | dput(dentry); | |
241 | err = -EAGAIN; | |
242 | goto out; | |
243 | } | |
244 | if (fpos_cmp(ctx->pos, di->offset) <= 0) { | |
fdd4e158 YZ |
245 | emit_dentry = true; |
246 | } | |
da502956 | 247 | spin_unlock(&dentry->d_lock); |
2817b000 | 248 | |
fdd4e158 | 249 | if (emit_dentry) { |
f3c4ebe6 | 250 | dout(" %llx dentry %p %pd %p\n", di->offset, |
fdd4e158 YZ |
251 | dentry, dentry, d_inode(dentry)); |
252 | ctx->pos = di->offset; | |
253 | if (!dir_emit(ctx, dentry->d_name.name, | |
254 | dentry->d_name.len, | |
255 | ceph_translate_ino(dentry->d_sb, | |
256 | d_inode(dentry)->i_ino), | |
257 | d_inode(dentry)->i_mode >> 12)) { | |
258 | dput(dentry); | |
259 | err = 0; | |
260 | break; | |
261 | } | |
262 | ctx->pos++; | |
0081bd83 | 263 | |
fdd4e158 YZ |
264 | if (last) |
265 | dput(last); | |
266 | last = dentry; | |
267 | } else { | |
268 | dput(dentry); | |
2817b000 | 269 | } |
fdd4e158 | 270 | } |
c530cd24 | 271 | out: |
fdd4e158 YZ |
272 | ceph_readdir_cache_release(&cache_ctl); |
273 | if (last) { | |
274 | int ret; | |
275 | di = ceph_dentry(last); | |
276 | ret = note_last_dentry(fi, last->d_name.name, last->d_name.len, | |
277 | fpos_off(di->offset) + 1); | |
278 | if (ret < 0) | |
279 | err = ret; | |
2817b000 | 280 | dput(last); |
84583cfb YZ |
281 | /* last_name no longer match cache index */ |
282 | if (fi->readdir_cache_idx >= 0) { | |
283 | fi->readdir_cache_idx = -1; | |
284 | fi->dir_release_count = 0; | |
285 | } | |
fdd4e158 | 286 | } |
2817b000 SW |
287 | return err; |
288 | } | |
289 | ||
f3c4ebe6 YZ |
290 | static bool need_send_readdir(struct ceph_file_info *fi, loff_t pos) |
291 | { | |
292 | if (!fi->last_readdir) | |
293 | return true; | |
294 | if (is_hash_order(pos)) | |
295 | return !ceph_frag_contains_value(fi->frag, fpos_hash(pos)); | |
296 | else | |
297 | return fi->frag != fpos_frag(pos); | |
298 | } | |
299 | ||
77acfa29 | 300 | static int ceph_readdir(struct file *file, struct dir_context *ctx) |
2817b000 | 301 | { |
77acfa29 AV |
302 | struct ceph_file_info *fi = file->private_data; |
303 | struct inode *inode = file_inode(file); | |
2817b000 | 304 | struct ceph_inode_info *ci = ceph_inode(inode); |
3d14c5d2 YS |
305 | struct ceph_fs_client *fsc = ceph_inode_to_client(inode); |
306 | struct ceph_mds_client *mdsc = fsc->mdsc; | |
8974eebd | 307 | int i; |
2817b000 | 308 | int err; |
b50c2de5 | 309 | unsigned frag = -1; |
2817b000 | 310 | struct ceph_mds_reply_info_parsed *rinfo; |
2817b000 | 311 | |
8974eebd | 312 | dout("readdir %p file %p pos %llx\n", inode, file, ctx->pos); |
9cfa1098 | 313 | if (fi->flags & CEPH_F_ATEND) |
2817b000 SW |
314 | return 0; |
315 | ||
316 | /* always start with . and .. */ | |
77acfa29 | 317 | if (ctx->pos == 0) { |
2817b000 | 318 | dout("readdir off 0 -> '.'\n"); |
77acfa29 | 319 | if (!dir_emit(ctx, ".", 1, |
ad1fee96 | 320 | ceph_translate_ino(inode->i_sb, inode->i_ino), |
77acfa29 | 321 | inode->i_mode >> 12)) |
2817b000 | 322 | return 0; |
77acfa29 | 323 | ctx->pos = 1; |
2817b000 | 324 | } |
77acfa29 | 325 | if (ctx->pos == 1) { |
b583043e | 326 | ino_t ino = parent_ino(file->f_path.dentry); |
2817b000 | 327 | dout("readdir off 1 -> '..'\n"); |
77acfa29 | 328 | if (!dir_emit(ctx, "..", 2, |
ad1fee96 | 329 | ceph_translate_ino(inode->i_sb, ino), |
77acfa29 | 330 | inode->i_mode >> 12)) |
2817b000 | 331 | return 0; |
77acfa29 | 332 | ctx->pos = 2; |
2817b000 SW |
333 | } |
334 | ||
335 | /* can we use the dcache? */ | |
be655596 | 336 | spin_lock(&ci->i_ceph_lock); |
fdd4e158 | 337 | if (ceph_test_mount_opt(fsc, DCACHE) && |
3d14c5d2 | 338 | !ceph_test_mount_opt(fsc, NOASYNCREADDIR) && |
a0dff78d | 339 | ceph_snap(inode) != CEPH_SNAPDIR && |
70db4f36 | 340 | __ceph_dir_is_complete_ordered(ci) && |
2817b000 | 341 | __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1)) { |
7389f7b2 | 342 | int shared_gen = atomic_read(&ci->i_shared_gen); |
be655596 | 343 | spin_unlock(&ci->i_ceph_lock); |
a30be7cb | 344 | err = __dcache_readdir(file, ctx, shared_gen); |
efa4c120 | 345 | if (err != -EAGAIN) |
2817b000 | 346 | return err; |
efa4c120 | 347 | } else { |
be655596 | 348 | spin_unlock(&ci->i_ceph_lock); |
2817b000 | 349 | } |
2817b000 SW |
350 | |
351 | /* proceed with a normal readdir */ | |
2817b000 SW |
352 | more: |
353 | /* do we have the correct frag content buffered? */ | |
f3c4ebe6 | 354 | if (need_send_readdir(fi, ctx->pos)) { |
2817b000 SW |
355 | struct ceph_mds_request *req; |
356 | int op = ceph_snap(inode) == CEPH_SNAPDIR ? | |
357 | CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR; | |
358 | ||
359 | /* discard old result, if any */ | |
393f6620 | 360 | if (fi->last_readdir) { |
2817b000 | 361 | ceph_mdsc_put_request(fi->last_readdir); |
393f6620 SW |
362 | fi->last_readdir = NULL; |
363 | } | |
2817b000 | 364 | |
f3c4ebe6 | 365 | if (is_hash_order(ctx->pos)) { |
b50c2de5 YZ |
366 | /* fragtree isn't always accurate. choose frag |
367 | * based on previous reply when possible. */ | |
368 | if (frag == (unsigned)-1) | |
369 | frag = ceph_choose_frag(ci, fpos_hash(ctx->pos), | |
370 | NULL, NULL); | |
f3c4ebe6 YZ |
371 | } else { |
372 | frag = fpos_frag(ctx->pos); | |
373 | } | |
374 | ||
2817b000 SW |
375 | dout("readdir fetching %llx.%llx frag %x offset '%s'\n", |
376 | ceph_vinop(inode), frag, fi->last_name); | |
377 | req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS); | |
378 | if (IS_ERR(req)) | |
379 | return PTR_ERR(req); | |
54008399 YZ |
380 | err = ceph_alloc_readdir_reply_buffer(req, inode); |
381 | if (err) { | |
382 | ceph_mdsc_put_request(req); | |
383 | return err; | |
384 | } | |
2817b000 SW |
385 | /* hints to request -> mds selection code */ |
386 | req->r_direct_mode = USE_AUTH_MDS; | |
5d37ca14 YZ |
387 | if (op == CEPH_MDS_OP_READDIR) { |
388 | req->r_direct_hash = ceph_frag_value(frag); | |
389 | __set_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags); | |
390 | } | |
a149bb9a | 391 | if (fi->last_name) { |
687265e5 | 392 | req->r_path2 = kstrdup(fi->last_name, GFP_KERNEL); |
a149bb9a SK |
393 | if (!req->r_path2) { |
394 | ceph_mdsc_put_request(req); | |
395 | return -ENOMEM; | |
396 | } | |
79162547 YZ |
397 | } else if (is_hash_order(ctx->pos)) { |
398 | req->r_args.readdir.offset_hash = | |
399 | cpu_to_le32(fpos_hash(ctx->pos)); | |
a149bb9a | 400 | } |
79162547 | 401 | |
fdd4e158 YZ |
402 | req->r_dir_release_cnt = fi->dir_release_count; |
403 | req->r_dir_ordered_cnt = fi->dir_ordered_count; | |
404 | req->r_readdir_cache_idx = fi->readdir_cache_idx; | |
2817b000 SW |
405 | req->r_readdir_offset = fi->next_offset; |
406 | req->r_args.readdir.frag = cpu_to_le32(frag); | |
956d39d6 YZ |
407 | req->r_args.readdir.flags = |
408 | cpu_to_le16(CEPH_READDIR_REPLY_BITFLAGS); | |
a149bb9a SK |
409 | |
410 | req->r_inode = inode; | |
411 | ihold(inode); | |
412 | req->r_dentry = dget(file->f_path.dentry); | |
2817b000 SW |
413 | err = ceph_mdsc_do_request(mdsc, NULL, req); |
414 | if (err < 0) { | |
415 | ceph_mdsc_put_request(req); | |
416 | return err; | |
417 | } | |
f3c4ebe6 YZ |
418 | dout("readdir got and parsed readdir result=%d on " |
419 | "frag %x, end=%d, complete=%d, hash_order=%d\n", | |
420 | err, frag, | |
2817b000 | 421 | (int)req->r_reply_info.dir_end, |
f3c4ebe6 YZ |
422 | (int)req->r_reply_info.dir_complete, |
423 | (int)req->r_reply_info.hash_order); | |
2817b000 | 424 | |
81c6aea5 YZ |
425 | rinfo = &req->r_reply_info; |
426 | if (le32_to_cpu(rinfo->dir_dir->frag) != frag) { | |
427 | frag = le32_to_cpu(rinfo->dir_dir->frag); | |
f3c4ebe6 YZ |
428 | if (!rinfo->hash_order) { |
429 | fi->next_offset = req->r_readdir_offset; | |
430 | /* adjust ctx->pos to beginning of frag */ | |
431 | ctx->pos = ceph_make_fpos(frag, | |
432 | fi->next_offset, | |
433 | false); | |
434 | } | |
81c6aea5 | 435 | } |
fdd4e158 | 436 | |
f0494206 | 437 | fi->frag = frag; |
2817b000 SW |
438 | fi->last_readdir = req; |
439 | ||
bc2de10d | 440 | if (test_bit(CEPH_MDS_R_DID_PREPOPULATE, &req->r_req_flags)) { |
fdd4e158 YZ |
441 | fi->readdir_cache_idx = req->r_readdir_cache_idx; |
442 | if (fi->readdir_cache_idx < 0) { | |
443 | /* preclude from marking dir ordered */ | |
444 | fi->dir_ordered_count = 0; | |
8974eebd YZ |
445 | } else if (ceph_frag_is_leftmost(frag) && |
446 | fi->next_offset == 2) { | |
fdd4e158 YZ |
447 | /* note dir version at start of readdir so |
448 | * we can tell if any dentries get dropped */ | |
449 | fi->dir_release_count = req->r_dir_release_cnt; | |
450 | fi->dir_ordered_count = req->r_dir_ordered_cnt; | |
451 | } | |
452 | } else { | |
453 | dout("readdir !did_prepopulate"); | |
454 | /* disable readdir cache */ | |
455 | fi->readdir_cache_idx = -1; | |
456 | /* preclude from marking dir complete */ | |
457 | fi->dir_release_count = 0; | |
458 | } | |
459 | ||
f3c4ebe6 YZ |
460 | /* note next offset and last dentry name */ |
461 | if (rinfo->dir_nr > 0) { | |
2a5beea3 YZ |
462 | struct ceph_mds_reply_dir_entry *rde = |
463 | rinfo->dir_entries + (rinfo->dir_nr-1); | |
f3c4ebe6 YZ |
464 | unsigned next_offset = req->r_reply_info.dir_end ? |
465 | 2 : (fpos_off(rde->offset) + 1); | |
2a5beea3 | 466 | err = note_last_dentry(fi, rde->name, rde->name_len, |
f3c4ebe6 | 467 | next_offset); |
2817b000 SW |
468 | if (err) |
469 | return err; | |
f3c4ebe6 YZ |
470 | } else if (req->r_reply_info.dir_end) { |
471 | fi->next_offset = 2; | |
472 | /* keep last name */ | |
2817b000 SW |
473 | } |
474 | } | |
475 | ||
476 | rinfo = &fi->last_readdir->r_reply_info; | |
8974eebd | 477 | dout("readdir frag %x num %d pos %llx chunk first %llx\n", |
f3c4ebe6 | 478 | fi->frag, rinfo->dir_nr, ctx->pos, |
8974eebd | 479 | rinfo->dir_nr ? rinfo->dir_entries[0].offset : 0LL); |
77acfa29 | 480 | |
8974eebd YZ |
481 | i = 0; |
482 | /* search start position */ | |
483 | if (rinfo->dir_nr > 0) { | |
484 | int step, nr = rinfo->dir_nr; | |
485 | while (nr > 0) { | |
486 | step = nr >> 1; | |
487 | if (rinfo->dir_entries[i + step].offset < ctx->pos) { | |
488 | i += step + 1; | |
489 | nr -= step + 1; | |
490 | } else { | |
491 | nr = step; | |
492 | } | |
493 | } | |
494 | } | |
495 | for (; i < rinfo->dir_nr; i++) { | |
496 | struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i; | |
3105c19c SW |
497 | struct ceph_vino vino; |
498 | ino_t ino; | |
b50c2de5 | 499 | u32 ftype; |
3105c19c | 500 | |
8974eebd YZ |
501 | BUG_ON(rde->offset < ctx->pos); |
502 | ||
503 | ctx->pos = rde->offset; | |
504 | dout("readdir (%d/%d) -> %llx '%.*s' %p\n", | |
505 | i, rinfo->dir_nr, ctx->pos, | |
2a5beea3 | 506 | rde->name_len, rde->name, &rde->inode.in); |
8974eebd | 507 | |
2a5beea3 YZ |
508 | BUG_ON(!rde->inode.in); |
509 | ftype = le32_to_cpu(rde->inode.in->mode) >> 12; | |
510 | vino.ino = le64_to_cpu(rde->inode.in->ino); | |
511 | vino.snap = le64_to_cpu(rde->inode.in->snapid); | |
3105c19c | 512 | ino = ceph_vino_to_ino(vino); |
8974eebd | 513 | |
2a5beea3 YZ |
514 | if (!dir_emit(ctx, rde->name, rde->name_len, |
515 | ceph_translate_ino(inode->i_sb, ino), ftype)) { | |
2817b000 SW |
516 | dout("filldir stopping us...\n"); |
517 | return 0; | |
518 | } | |
77acfa29 | 519 | ctx->pos++; |
2817b000 SW |
520 | } |
521 | ||
b50c2de5 YZ |
522 | ceph_mdsc_put_request(fi->last_readdir); |
523 | fi->last_readdir = NULL; | |
524 | ||
f3c4ebe6 | 525 | if (fi->next_offset > 2) { |
b50c2de5 | 526 | frag = fi->frag; |
2817b000 SW |
527 | goto more; |
528 | } | |
529 | ||
530 | /* more frags? */ | |
f3c4ebe6 | 531 | if (!ceph_frag_is_rightmost(fi->frag)) { |
b50c2de5 | 532 | frag = ceph_frag_next(fi->frag); |
f3c4ebe6 YZ |
533 | if (is_hash_order(ctx->pos)) { |
534 | loff_t new_pos = ceph_make_fpos(ceph_frag_value(frag), | |
535 | fi->next_offset, true); | |
536 | if (new_pos > ctx->pos) | |
537 | ctx->pos = new_pos; | |
538 | /* keep last_name */ | |
539 | } else { | |
540 | ctx->pos = ceph_make_fpos(frag, fi->next_offset, false); | |
541 | kfree(fi->last_name); | |
542 | fi->last_name = NULL; | |
543 | } | |
2817b000 SW |
544 | dout("readdir next frag is %x\n", frag); |
545 | goto more; | |
546 | } | |
9cfa1098 | 547 | fi->flags |= CEPH_F_ATEND; |
2817b000 SW |
548 | |
549 | /* | |
550 | * if dir_release_count still matches the dir, no dentries | |
551 | * were released during the whole readdir, and we should have | |
552 | * the complete dir contents in our cache. | |
553 | */ | |
fdd4e158 YZ |
554 | if (atomic64_read(&ci->i_release_count) == fi->dir_release_count) { |
555 | spin_lock(&ci->i_ceph_lock); | |
556 | if (fi->dir_ordered_count == atomic64_read(&ci->i_ordered_count)) { | |
70db4f36 | 557 | dout(" marking %p complete and ordered\n", inode); |
fdd4e158 YZ |
558 | /* use i_size to track number of entries in |
559 | * readdir cache */ | |
560 | BUG_ON(fi->readdir_cache_idx < 0); | |
561 | i_size_write(inode, fi->readdir_cache_idx * | |
562 | sizeof(struct dentry*)); | |
563 | } else { | |
70db4f36 | 564 | dout(" marking %p complete\n", inode); |
fdd4e158 | 565 | } |
70db4f36 YZ |
566 | __ceph_dir_set_complete(ci, fi->dir_release_count, |
567 | fi->dir_ordered_count); | |
fdd4e158 | 568 | spin_unlock(&ci->i_ceph_lock); |
2817b000 | 569 | } |
2817b000 | 570 | |
77acfa29 | 571 | dout("readdir %p file %p done.\n", inode, file); |
2817b000 SW |
572 | return 0; |
573 | } | |
574 | ||
8974eebd | 575 | static void reset_readdir(struct ceph_file_info *fi) |
2817b000 SW |
576 | { |
577 | if (fi->last_readdir) { | |
578 | ceph_mdsc_put_request(fi->last_readdir); | |
579 | fi->last_readdir = NULL; | |
580 | } | |
581 | kfree(fi->last_name); | |
a1629c3b | 582 | fi->last_name = NULL; |
fdd4e158 YZ |
583 | fi->dir_release_count = 0; |
584 | fi->readdir_cache_idx = -1; | |
a78600e7 | 585 | fi->next_offset = 2; /* compensate for . and .. */ |
9cfa1098 | 586 | fi->flags &= ~CEPH_F_ATEND; |
2817b000 SW |
587 | } |
588 | ||
8974eebd YZ |
589 | /* |
590 | * discard buffered readdir content on seekdir(0), or seek to new frag, | |
591 | * or seek prior to current chunk | |
592 | */ | |
593 | static bool need_reset_readdir(struct ceph_file_info *fi, loff_t new_pos) | |
594 | { | |
595 | struct ceph_mds_reply_info_parsed *rinfo; | |
f3c4ebe6 | 596 | loff_t chunk_offset; |
8974eebd YZ |
597 | if (new_pos == 0) |
598 | return true; | |
f3c4ebe6 YZ |
599 | if (is_hash_order(new_pos)) { |
600 | /* no need to reset last_name for a forward seek when | |
601 | * dentries are sotred in hash order */ | |
0f5aa88a | 602 | } else if (fi->frag != fpos_frag(new_pos)) { |
8974eebd | 603 | return true; |
f3c4ebe6 | 604 | } |
8974eebd YZ |
605 | rinfo = fi->last_readdir ? &fi->last_readdir->r_reply_info : NULL; |
606 | if (!rinfo || !rinfo->dir_nr) | |
607 | return true; | |
f3c4ebe6 YZ |
608 | chunk_offset = rinfo->dir_entries[0].offset; |
609 | return new_pos < chunk_offset || | |
610 | is_hash_order(new_pos) != is_hash_order(chunk_offset); | |
8974eebd YZ |
611 | } |
612 | ||
965c8e59 | 613 | static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence) |
2817b000 SW |
614 | { |
615 | struct ceph_file_info *fi = file->private_data; | |
616 | struct inode *inode = file->f_mapping->host; | |
2817b000 SW |
617 | loff_t retval; |
618 | ||
5955102c | 619 | inode_lock(inode); |
06222e49 | 620 | retval = -EINVAL; |
965c8e59 | 621 | switch (whence) { |
2817b000 SW |
622 | case SEEK_CUR: |
623 | offset += file->f_pos; | |
06222e49 JB |
624 | case SEEK_SET: |
625 | break; | |
fdd4e158 YZ |
626 | case SEEK_END: |
627 | retval = -EOPNOTSUPP; | |
06222e49 JB |
628 | default: |
629 | goto out; | |
2817b000 | 630 | } |
06222e49 | 631 | |
f0494206 | 632 | if (offset >= 0) { |
f3c4ebe6 YZ |
633 | if (need_reset_readdir(fi, offset)) { |
634 | dout("dir_llseek dropping %p content\n", file); | |
635 | reset_readdir(fi); | |
636 | } else if (is_hash_order(offset) && offset > file->f_pos) { | |
637 | /* for hash offset, we don't know if a forward seek | |
638 | * is within same frag */ | |
639 | fi->dir_release_count = 0; | |
640 | fi->readdir_cache_idx = -1; | |
641 | } | |
642 | ||
2817b000 SW |
643 | if (offset != file->f_pos) { |
644 | file->f_pos = offset; | |
645 | file->f_version = 0; | |
9cfa1098 | 646 | fi->flags &= ~CEPH_F_ATEND; |
2817b000 SW |
647 | } |
648 | retval = offset; | |
2817b000 | 649 | } |
06222e49 | 650 | out: |
5955102c | 651 | inode_unlock(inode); |
2817b000 SW |
652 | return retval; |
653 | } | |
654 | ||
655 | /* | |
468640e3 | 656 | * Handle lookups for the hidden .snap directory. |
2817b000 | 657 | */ |
468640e3 SW |
658 | int ceph_handle_snapdir(struct ceph_mds_request *req, |
659 | struct dentry *dentry, int err) | |
2817b000 | 660 | { |
3d14c5d2 | 661 | struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb); |
2b0143b5 | 662 | struct inode *parent = d_inode(dentry->d_parent); /* we hold i_mutex */ |
2817b000 SW |
663 | |
664 | /* .snap dir? */ | |
665 | if (err == -ENOENT && | |
455cec0a | 666 | ceph_snap(parent) == CEPH_NOSNAP && |
6b805185 | 667 | strcmp(dentry->d_name.name, |
3d14c5d2 | 668 | fsc->mount_options->snapdir_name) == 0) { |
2817b000 | 669 | struct inode *inode = ceph_get_snapdir(parent); |
a455589f AV |
670 | dout("ENOENT on snapdir %p '%pd', linking to snapdir %p\n", |
671 | dentry, dentry, inode); | |
9358c6d4 | 672 | BUG_ON(!d_unhashed(dentry)); |
2817b000 SW |
673 | d_add(dentry, inode); |
674 | err = 0; | |
675 | } | |
468640e3 SW |
676 | return err; |
677 | } | |
2817b000 | 678 | |
468640e3 SW |
679 | /* |
680 | * Figure out final result of a lookup/open request. | |
681 | * | |
682 | * Mainly, make sure we return the final req->r_dentry (if it already | |
683 | * existed) in place of the original VFS-provided dentry when they | |
684 | * differ. | |
685 | * | |
686 | * Gracefully handle the case where the MDS replies with -ENOENT and | |
687 | * no trace (which it may do, at its discretion, e.g., if it doesn't | |
688 | * care to issue a lease on the negative dentry). | |
689 | */ | |
690 | struct dentry *ceph_finish_lookup(struct ceph_mds_request *req, | |
691 | struct dentry *dentry, int err) | |
692 | { | |
2817b000 SW |
693 | if (err == -ENOENT) { |
694 | /* no trace? */ | |
695 | err = 0; | |
696 | if (!req->r_reply_info.head->is_dentry) { | |
697 | dout("ENOENT and no trace, dentry %p inode %p\n", | |
2b0143b5 DH |
698 | dentry, d_inode(dentry)); |
699 | if (d_really_is_positive(dentry)) { | |
2817b000 SW |
700 | d_drop(dentry); |
701 | err = -ENOENT; | |
702 | } else { | |
703 | d_add(dentry, NULL); | |
704 | } | |
705 | } | |
706 | } | |
707 | if (err) | |
708 | dentry = ERR_PTR(err); | |
709 | else if (dentry != req->r_dentry) | |
710 | dentry = dget(req->r_dentry); /* we got spliced */ | |
711 | else | |
712 | dentry = NULL; | |
713 | return dentry; | |
714 | } | |
715 | ||
3b33f692 | 716 | static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry) |
1d1de916 SW |
717 | { |
718 | return ceph_ino(inode) == CEPH_INO_ROOT && | |
719 | strncmp(dentry->d_name.name, ".ceph", 5) == 0; | |
720 | } | |
721 | ||
2817b000 SW |
722 | /* |
723 | * Look up a single dir entry. If there is a lookup intent, inform | |
724 | * the MDS so that it gets our 'caps wanted' value in a single op. | |
725 | */ | |
726 | static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry, | |
00cd8dd3 | 727 | unsigned int flags) |
2817b000 | 728 | { |
3d14c5d2 YS |
729 | struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb); |
730 | struct ceph_mds_client *mdsc = fsc->mdsc; | |
2817b000 SW |
731 | struct ceph_mds_request *req; |
732 | int op; | |
315f2408 | 733 | int mask; |
2817b000 SW |
734 | int err; |
735 | ||
a455589f AV |
736 | dout("lookup %p dentry %p '%pd'\n", |
737 | dir, dentry, dentry); | |
2817b000 SW |
738 | |
739 | if (dentry->d_name.len > NAME_MAX) | |
740 | return ERR_PTR(-ENAMETOOLONG); | |
741 | ||
2817b000 | 742 | /* can we conclude ENOENT locally? */ |
2b0143b5 | 743 | if (d_really_is_negative(dentry)) { |
2817b000 SW |
744 | struct ceph_inode_info *ci = ceph_inode(dir); |
745 | struct ceph_dentry_info *di = ceph_dentry(dentry); | |
746 | ||
be655596 | 747 | spin_lock(&ci->i_ceph_lock); |
2817b000 SW |
748 | dout(" dir %p flags are %d\n", dir, ci->i_ceph_flags); |
749 | if (strncmp(dentry->d_name.name, | |
3d14c5d2 | 750 | fsc->mount_options->snapdir_name, |
2817b000 | 751 | dentry->d_name.len) && |
1d1de916 | 752 | !is_root_ceph_dentry(dir, dentry) && |
e2c3de04 | 753 | ceph_test_mount_opt(fsc, DCACHE) && |
2f276c51 | 754 | __ceph_dir_is_complete(ci) && |
2817b000 | 755 | (__ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1))) { |
be655596 | 756 | spin_unlock(&ci->i_ceph_lock); |
2817b000 SW |
757 | dout(" dir %p complete, -ENOENT\n", dir); |
758 | d_add(dentry, NULL); | |
7389f7b2 | 759 | di->lease_shared_gen = atomic_read(&ci->i_shared_gen); |
2817b000 SW |
760 | return NULL; |
761 | } | |
be655596 | 762 | spin_unlock(&ci->i_ceph_lock); |
2817b000 SW |
763 | } |
764 | ||
765 | op = ceph_snap(dir) == CEPH_SNAPDIR ? | |
766 | CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP; | |
767 | req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS); | |
768 | if (IS_ERR(req)) | |
7e34bc52 | 769 | return ERR_CAST(req); |
2817b000 SW |
770 | req->r_dentry = dget(dentry); |
771 | req->r_num_caps = 2; | |
315f2408 YZ |
772 | |
773 | mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED; | |
774 | if (ceph_security_xattr_wanted(dir)) | |
775 | mask |= CEPH_CAP_XATTR_SHARED; | |
776 | req->r_args.getattr.mask = cpu_to_le32(mask); | |
777 | ||
3dd69aab JL |
778 | req->r_parent = dir; |
779 | set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags); | |
2817b000 | 780 | err = ceph_mdsc_do_request(mdsc, NULL, req); |
468640e3 | 781 | err = ceph_handle_snapdir(req, dentry, err); |
2817b000 SW |
782 | dentry = ceph_finish_lookup(req, dentry, err); |
783 | ceph_mdsc_put_request(req); /* will dput(dentry) */ | |
784 | dout("lookup result=%p\n", dentry); | |
785 | return dentry; | |
786 | } | |
787 | ||
788 | /* | |
789 | * If we do a create but get no trace back from the MDS, follow up with | |
790 | * a lookup (the VFS expects us to link up the provided dentry). | |
791 | */ | |
792 | int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry) | |
793 | { | |
00cd8dd3 | 794 | struct dentry *result = ceph_lookup(dir, dentry, 0); |
2817b000 SW |
795 | |
796 | if (result && !IS_ERR(result)) { | |
797 | /* | |
798 | * We created the item, then did a lookup, and found | |
799 | * it was already linked to another inode we already | |
4d41cef2 YZ |
800 | * had in our cache (and thus got spliced). To not |
801 | * confuse VFS (especially when inode is a directory), | |
802 | * we don't link our dentry to that inode, return an | |
803 | * error instead. | |
804 | * | |
805 | * This event should be rare and it happens only when | |
806 | * we talk to old MDS. Recent MDS does not send traceless | |
807 | * reply for request that creates new inode. | |
2817b000 | 808 | */ |
5cba372c | 809 | d_drop(result); |
4d41cef2 | 810 | return -ESTALE; |
2817b000 SW |
811 | } |
812 | return PTR_ERR(result); | |
813 | } | |
814 | ||
815 | static int ceph_mknod(struct inode *dir, struct dentry *dentry, | |
1a67aafb | 816 | umode_t mode, dev_t rdev) |
2817b000 | 817 | { |
3d14c5d2 YS |
818 | struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb); |
819 | struct ceph_mds_client *mdsc = fsc->mdsc; | |
2817b000 | 820 | struct ceph_mds_request *req; |
b1ee94aa | 821 | struct ceph_acls_info acls = {}; |
2817b000 SW |
822 | int err; |
823 | ||
824 | if (ceph_snap(dir) != CEPH_NOSNAP) | |
825 | return -EROFS; | |
826 | ||
6aa007ee LH |
827 | if (ceph_quota_is_max_files_exceeded(dir)) |
828 | return -EDQUOT; | |
829 | ||
b1ee94aa YZ |
830 | err = ceph_pre_init_acls(dir, &mode, &acls); |
831 | if (err < 0) | |
832 | return err; | |
833 | ||
1a67aafb | 834 | dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n", |
2817b000 SW |
835 | dir, dentry, mode, rdev); |
836 | req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS); | |
837 | if (IS_ERR(req)) { | |
b1ee94aa YZ |
838 | err = PTR_ERR(req); |
839 | goto out; | |
2817b000 SW |
840 | } |
841 | req->r_dentry = dget(dentry); | |
842 | req->r_num_caps = 2; | |
3dd69aab JL |
843 | req->r_parent = dir; |
844 | set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags); | |
2817b000 SW |
845 | req->r_args.mknod.mode = cpu_to_le32(mode); |
846 | req->r_args.mknod.rdev = cpu_to_le32(rdev); | |
847 | req->r_dentry_drop = CEPH_CAP_FILE_SHARED; | |
848 | req->r_dentry_unless = CEPH_CAP_FILE_EXCL; | |
b1ee94aa YZ |
849 | if (acls.pagelist) { |
850 | req->r_pagelist = acls.pagelist; | |
851 | acls.pagelist = NULL; | |
852 | } | |
2817b000 SW |
853 | err = ceph_mdsc_do_request(mdsc, dir, req); |
854 | if (!err && !req->r_reply_info.head->is_dentry) | |
855 | err = ceph_handle_notrace_create(dir, dentry); | |
856 | ceph_mdsc_put_request(req); | |
b1ee94aa | 857 | out: |
7221fe4c | 858 | if (!err) |
2b0143b5 | 859 | ceph_init_inode_acls(d_inode(dentry), &acls); |
b20a95a0 | 860 | else |
2817b000 | 861 | d_drop(dentry); |
b1ee94aa | 862 | ceph_release_acls_info(&acls); |
2817b000 SW |
863 | return err; |
864 | } | |
865 | ||
4acdaf27 | 866 | static int ceph_create(struct inode *dir, struct dentry *dentry, umode_t mode, |
ebfc3b49 | 867 | bool excl) |
2817b000 | 868 | { |
2d83bde9 | 869 | return ceph_mknod(dir, dentry, mode, 0); |
2817b000 SW |
870 | } |
871 | ||
872 | static int ceph_symlink(struct inode *dir, struct dentry *dentry, | |
873 | const char *dest) | |
874 | { | |
3d14c5d2 YS |
875 | struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb); |
876 | struct ceph_mds_client *mdsc = fsc->mdsc; | |
2817b000 SW |
877 | struct ceph_mds_request *req; |
878 | int err; | |
879 | ||
880 | if (ceph_snap(dir) != CEPH_NOSNAP) | |
881 | return -EROFS; | |
882 | ||
6aa007ee LH |
883 | if (ceph_quota_is_max_files_exceeded(dir)) |
884 | return -EDQUOT; | |
885 | ||
2817b000 SW |
886 | dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest); |
887 | req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS); | |
888 | if (IS_ERR(req)) { | |
b1ee94aa YZ |
889 | err = PTR_ERR(req); |
890 | goto out; | |
2817b000 | 891 | } |
687265e5 | 892 | req->r_path2 = kstrdup(dest, GFP_KERNEL); |
a149bb9a SK |
893 | if (!req->r_path2) { |
894 | err = -ENOMEM; | |
895 | ceph_mdsc_put_request(req); | |
896 | goto out; | |
897 | } | |
3dd69aab JL |
898 | req->r_parent = dir; |
899 | set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags); | |
a149bb9a SK |
900 | req->r_dentry = dget(dentry); |
901 | req->r_num_caps = 2; | |
2817b000 SW |
902 | req->r_dentry_drop = CEPH_CAP_FILE_SHARED; |
903 | req->r_dentry_unless = CEPH_CAP_FILE_EXCL; | |
904 | err = ceph_mdsc_do_request(mdsc, dir, req); | |
905 | if (!err && !req->r_reply_info.head->is_dentry) | |
906 | err = ceph_handle_notrace_create(dir, dentry); | |
907 | ceph_mdsc_put_request(req); | |
b1ee94aa YZ |
908 | out: |
909 | if (err) | |
2817b000 SW |
910 | d_drop(dentry); |
911 | return err; | |
912 | } | |
913 | ||
18bb1db3 | 914 | static int ceph_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode) |
2817b000 | 915 | { |
3d14c5d2 YS |
916 | struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb); |
917 | struct ceph_mds_client *mdsc = fsc->mdsc; | |
2817b000 | 918 | struct ceph_mds_request *req; |
b1ee94aa | 919 | struct ceph_acls_info acls = {}; |
2817b000 SW |
920 | int err = -EROFS; |
921 | int op; | |
922 | ||
923 | if (ceph_snap(dir) == CEPH_SNAPDIR) { | |
924 | /* mkdir .snap/foo is a MKSNAP */ | |
925 | op = CEPH_MDS_OP_MKSNAP; | |
a455589f AV |
926 | dout("mksnap dir %p snap '%pd' dn %p\n", dir, |
927 | dentry, dentry); | |
2817b000 | 928 | } else if (ceph_snap(dir) == CEPH_NOSNAP) { |
18bb1db3 | 929 | dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode); |
2817b000 SW |
930 | op = CEPH_MDS_OP_MKDIR; |
931 | } else { | |
932 | goto out; | |
933 | } | |
b1ee94aa | 934 | |
6aa007ee LH |
935 | if (ceph_quota_is_max_files_exceeded(dir)) { |
936 | err = -EDQUOT; | |
937 | goto out; | |
938 | } | |
939 | ||
b1ee94aa YZ |
940 | mode |= S_IFDIR; |
941 | err = ceph_pre_init_acls(dir, &mode, &acls); | |
942 | if (err < 0) | |
943 | goto out; | |
944 | ||
2817b000 SW |
945 | req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS); |
946 | if (IS_ERR(req)) { | |
947 | err = PTR_ERR(req); | |
948 | goto out; | |
949 | } | |
950 | ||
951 | req->r_dentry = dget(dentry); | |
952 | req->r_num_caps = 2; | |
3dd69aab JL |
953 | req->r_parent = dir; |
954 | set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags); | |
2817b000 SW |
955 | req->r_args.mkdir.mode = cpu_to_le32(mode); |
956 | req->r_dentry_drop = CEPH_CAP_FILE_SHARED; | |
957 | req->r_dentry_unless = CEPH_CAP_FILE_EXCL; | |
b1ee94aa YZ |
958 | if (acls.pagelist) { |
959 | req->r_pagelist = acls.pagelist; | |
960 | acls.pagelist = NULL; | |
961 | } | |
2817b000 | 962 | err = ceph_mdsc_do_request(mdsc, dir, req); |
275dd19e YZ |
963 | if (!err && |
964 | !req->r_reply_info.head->is_target && | |
965 | !req->r_reply_info.head->is_dentry) | |
2817b000 SW |
966 | err = ceph_handle_notrace_create(dir, dentry); |
967 | ceph_mdsc_put_request(req); | |
968 | out: | |
b20a95a0 | 969 | if (!err) |
2b0143b5 | 970 | ceph_init_inode_acls(d_inode(dentry), &acls); |
b20a95a0 | 971 | else |
2817b000 | 972 | d_drop(dentry); |
b1ee94aa | 973 | ceph_release_acls_info(&acls); |
2817b000 SW |
974 | return err; |
975 | } | |
976 | ||
977 | static int ceph_link(struct dentry *old_dentry, struct inode *dir, | |
978 | struct dentry *dentry) | |
979 | { | |
3d14c5d2 YS |
980 | struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb); |
981 | struct ceph_mds_client *mdsc = fsc->mdsc; | |
2817b000 SW |
982 | struct ceph_mds_request *req; |
983 | int err; | |
984 | ||
985 | if (ceph_snap(dir) != CEPH_NOSNAP) | |
986 | return -EROFS; | |
987 | ||
988 | dout("link in dir %p old_dentry %p dentry %p\n", dir, | |
989 | old_dentry, dentry); | |
990 | req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS); | |
991 | if (IS_ERR(req)) { | |
992 | d_drop(dentry); | |
993 | return PTR_ERR(req); | |
994 | } | |
995 | req->r_dentry = dget(dentry); | |
996 | req->r_num_caps = 2; | |
4b58c9b1 | 997 | req->r_old_dentry = dget(old_dentry); |
3dd69aab JL |
998 | req->r_parent = dir; |
999 | set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags); | |
2817b000 SW |
1000 | req->r_dentry_drop = CEPH_CAP_FILE_SHARED; |
1001 | req->r_dentry_unless = CEPH_CAP_FILE_EXCL; | |
ad88f23f YZ |
1002 | /* release LINK_SHARED on source inode (mds will lock it) */ |
1003 | req->r_old_inode_drop = CEPH_CAP_LINK_SHARED; | |
2817b000 | 1004 | err = ceph_mdsc_do_request(mdsc, dir, req); |
70b666c3 | 1005 | if (err) { |
2817b000 | 1006 | d_drop(dentry); |
70b666c3 | 1007 | } else if (!req->r_reply_info.head->is_dentry) { |
2b0143b5 DH |
1008 | ihold(d_inode(old_dentry)); |
1009 | d_instantiate(dentry, d_inode(old_dentry)); | |
70b666c3 | 1010 | } |
2817b000 SW |
1011 | ceph_mdsc_put_request(req); |
1012 | return err; | |
1013 | } | |
1014 | ||
1015 | /* | |
1016 | * For a soon-to-be unlinked file, drop the AUTH_RDCACHE caps. If it | |
1017 | * looks like the link count will hit 0, drop any other caps (other | |
1018 | * than PIN) we don't specifically want (due to the file still being | |
1019 | * open). | |
1020 | */ | |
1021 | static int drop_caps_for_unlink(struct inode *inode) | |
1022 | { | |
1023 | struct ceph_inode_info *ci = ceph_inode(inode); | |
1024 | int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL; | |
1025 | ||
be655596 | 1026 | spin_lock(&ci->i_ceph_lock); |
2817b000 SW |
1027 | if (inode->i_nlink == 1) { |
1028 | drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN); | |
1029 | ci->i_ceph_flags |= CEPH_I_NODELAY; | |
1030 | } | |
be655596 | 1031 | spin_unlock(&ci->i_ceph_lock); |
2817b000 SW |
1032 | return drop; |
1033 | } | |
1034 | ||
1035 | /* | |
1036 | * rmdir and unlink are differ only by the metadata op code | |
1037 | */ | |
1038 | static int ceph_unlink(struct inode *dir, struct dentry *dentry) | |
1039 | { | |
3d14c5d2 YS |
1040 | struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb); |
1041 | struct ceph_mds_client *mdsc = fsc->mdsc; | |
2b0143b5 | 1042 | struct inode *inode = d_inode(dentry); |
2817b000 SW |
1043 | struct ceph_mds_request *req; |
1044 | int err = -EROFS; | |
1045 | int op; | |
1046 | ||
1047 | if (ceph_snap(dir) == CEPH_SNAPDIR) { | |
1048 | /* rmdir .snap/foo is RMSNAP */ | |
a455589f | 1049 | dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry); |
2817b000 SW |
1050 | op = CEPH_MDS_OP_RMSNAP; |
1051 | } else if (ceph_snap(dir) == CEPH_NOSNAP) { | |
1052 | dout("unlink/rmdir dir %p dn %p inode %p\n", | |
1053 | dir, dentry, inode); | |
e36cb0b8 | 1054 | op = d_is_dir(dentry) ? |
2817b000 SW |
1055 | CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK; |
1056 | } else | |
1057 | goto out; | |
1058 | req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS); | |
1059 | if (IS_ERR(req)) { | |
1060 | err = PTR_ERR(req); | |
1061 | goto out; | |
1062 | } | |
1063 | req->r_dentry = dget(dentry); | |
1064 | req->r_num_caps = 2; | |
3dd69aab JL |
1065 | req->r_parent = dir; |
1066 | set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags); | |
2817b000 SW |
1067 | req->r_dentry_drop = CEPH_CAP_FILE_SHARED; |
1068 | req->r_dentry_unless = CEPH_CAP_FILE_EXCL; | |
1069 | req->r_inode_drop = drop_caps_for_unlink(inode); | |
1070 | err = ceph_mdsc_do_request(mdsc, dir, req); | |
1071 | if (!err && !req->r_reply_info.head->is_dentry) | |
1072 | d_delete(dentry); | |
1073 | ceph_mdsc_put_request(req); | |
1074 | out: | |
1075 | return err; | |
1076 | } | |
1077 | ||
1078 | static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry, | |
1cd66c93 MS |
1079 | struct inode *new_dir, struct dentry *new_dentry, |
1080 | unsigned int flags) | |
2817b000 | 1081 | { |
3d14c5d2 YS |
1082 | struct ceph_fs_client *fsc = ceph_sb_to_client(old_dir->i_sb); |
1083 | struct ceph_mds_client *mdsc = fsc->mdsc; | |
2817b000 | 1084 | struct ceph_mds_request *req; |
0ea611a3 | 1085 | int op = CEPH_MDS_OP_RENAME; |
2817b000 SW |
1086 | int err; |
1087 | ||
1cd66c93 MS |
1088 | if (flags) |
1089 | return -EINVAL; | |
1090 | ||
2817b000 SW |
1091 | if (ceph_snap(old_dir) != ceph_snap(new_dir)) |
1092 | return -EXDEV; | |
0ea611a3 YZ |
1093 | if (ceph_snap(old_dir) != CEPH_NOSNAP) { |
1094 | if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR) | |
1095 | op = CEPH_MDS_OP_RENAMESNAP; | |
1096 | else | |
1097 | return -EROFS; | |
1098 | } | |
68a16a23 LH |
1099 | /* don't allow cross-quota renames */ |
1100 | if ((old_dir != new_dir) && | |
1101 | (!ceph_quota_is_same_realm(old_dir, new_dir))) | |
1102 | return -EXDEV; | |
1103 | ||
2817b000 SW |
1104 | dout("rename dir %p dentry %p to dir %p dentry %p\n", |
1105 | old_dir, old_dentry, new_dir, new_dentry); | |
0ea611a3 | 1106 | req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS); |
2817b000 SW |
1107 | if (IS_ERR(req)) |
1108 | return PTR_ERR(req); | |
180061a5 | 1109 | ihold(old_dir); |
2817b000 SW |
1110 | req->r_dentry = dget(new_dentry); |
1111 | req->r_num_caps = 2; | |
1112 | req->r_old_dentry = dget(old_dentry); | |
180061a5 | 1113 | req->r_old_dentry_dir = old_dir; |
3dd69aab JL |
1114 | req->r_parent = new_dir; |
1115 | set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags); | |
2817b000 SW |
1116 | req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED; |
1117 | req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL; | |
1118 | req->r_dentry_drop = CEPH_CAP_FILE_SHARED; | |
1119 | req->r_dentry_unless = CEPH_CAP_FILE_EXCL; | |
1120 | /* release LINK_RDCACHE on source inode (mds will lock it) */ | |
1121 | req->r_old_inode_drop = CEPH_CAP_LINK_SHARED; | |
2b0143b5 DH |
1122 | if (d_really_is_positive(new_dentry)) |
1123 | req->r_inode_drop = drop_caps_for_unlink(d_inode(new_dentry)); | |
2817b000 SW |
1124 | err = ceph_mdsc_do_request(mdsc, old_dir, req); |
1125 | if (!err && !req->r_reply_info.head->is_dentry) { | |
1126 | /* | |
1127 | * Normally d_move() is done by fill_trace (called by | |
1128 | * do_request, above). If there is no trace, we need | |
1129 | * to do it here. | |
1130 | */ | |
ea1409f9 | 1131 | |
fdd4e158 YZ |
1132 | /* d_move screws up sibling dentries' offsets */ |
1133 | ceph_dir_clear_complete(old_dir); | |
1134 | ceph_dir_clear_complete(new_dir); | |
1135 | ||
2817b000 | 1136 | d_move(old_dentry, new_dentry); |
ea1409f9 SW |
1137 | |
1138 | /* ensure target dentry is invalidated, despite | |
1139 | rehashing bug in vfs_rename_dir */ | |
81a6cf2d | 1140 | ceph_invalidate_dentry_lease(new_dentry); |
2817b000 SW |
1141 | } |
1142 | ceph_mdsc_put_request(req); | |
1143 | return err; | |
1144 | } | |
1145 | ||
81a6cf2d SW |
1146 | /* |
1147 | * Ensure a dentry lease will no longer revalidate. | |
1148 | */ | |
1149 | void ceph_invalidate_dentry_lease(struct dentry *dentry) | |
1150 | { | |
1151 | spin_lock(&dentry->d_lock); | |
9b16f03c | 1152 | ceph_dentry(dentry)->time = jiffies; |
81a6cf2d SW |
1153 | ceph_dentry(dentry)->lease_shared_gen = 0; |
1154 | spin_unlock(&dentry->d_lock); | |
1155 | } | |
2817b000 SW |
1156 | |
1157 | /* | |
1158 | * Check if dentry lease is valid. If not, delete the lease. Try to | |
1159 | * renew if the least is more than half up. | |
1160 | */ | |
14fb9c9e JL |
1161 | static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags, |
1162 | struct inode *dir) | |
2817b000 SW |
1163 | { |
1164 | struct ceph_dentry_info *di; | |
1165 | struct ceph_mds_session *s; | |
1166 | int valid = 0; | |
1167 | u32 gen; | |
1168 | unsigned long ttl; | |
1169 | struct ceph_mds_session *session = NULL; | |
2817b000 SW |
1170 | u32 seq = 0; |
1171 | ||
1172 | spin_lock(&dentry->d_lock); | |
1173 | di = ceph_dentry(dentry); | |
14fb9c9e | 1174 | if (di && di->lease_session) { |
2817b000 | 1175 | s = di->lease_session; |
d8fb02ab | 1176 | spin_lock(&s->s_gen_ttl_lock); |
2817b000 SW |
1177 | gen = s->s_cap_gen; |
1178 | ttl = s->s_cap_ttl; | |
d8fb02ab | 1179 | spin_unlock(&s->s_gen_ttl_lock); |
2817b000 SW |
1180 | |
1181 | if (di->lease_gen == gen && | |
9b16f03c | 1182 | time_before(jiffies, di->time) && |
2817b000 SW |
1183 | time_before(jiffies, ttl)) { |
1184 | valid = 1; | |
1185 | if (di->lease_renew_after && | |
1186 | time_after(jiffies, di->lease_renew_after)) { | |
14fb9c9e JL |
1187 | /* |
1188 | * We should renew. If we're in RCU walk mode | |
1189 | * though, we can't do that so just return | |
1190 | * -ECHILD. | |
1191 | */ | |
1192 | if (flags & LOOKUP_RCU) { | |
1193 | valid = -ECHILD; | |
1194 | } else { | |
1195 | session = ceph_get_mds_session(s); | |
1196 | seq = di->lease_seq; | |
1197 | di->lease_renew_after = 0; | |
1198 | di->lease_renew_from = jiffies; | |
1199 | } | |
2817b000 | 1200 | } |
2817b000 SW |
1201 | } |
1202 | } | |
1203 | spin_unlock(&dentry->d_lock); | |
1204 | ||
1205 | if (session) { | |
1206 | ceph_mdsc_lease_send_msg(session, dir, dentry, | |
1207 | CEPH_MDS_LEASE_RENEW, seq); | |
1208 | ceph_put_mds_session(session); | |
1209 | } | |
1210 | dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid); | |
1211 | return valid; | |
1212 | } | |
1213 | ||
1214 | /* | |
1215 | * Check if directory-wide content lease/cap is valid. | |
1216 | */ | |
1217 | static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry) | |
1218 | { | |
1219 | struct ceph_inode_info *ci = ceph_inode(dir); | |
1220 | struct ceph_dentry_info *di = ceph_dentry(dentry); | |
1221 | int valid = 0; | |
1222 | ||
be655596 | 1223 | spin_lock(&ci->i_ceph_lock); |
7389f7b2 | 1224 | if (atomic_read(&ci->i_shared_gen) == di->lease_shared_gen) |
2817b000 | 1225 | valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1); |
be655596 | 1226 | spin_unlock(&ci->i_ceph_lock); |
2817b000 | 1227 | dout("dir_lease_is_valid dir %p v%u dentry %p v%u = %d\n", |
7389f7b2 YZ |
1228 | dir, (unsigned)atomic_read(&ci->i_shared_gen), |
1229 | dentry, (unsigned)di->lease_shared_gen, valid); | |
2817b000 SW |
1230 | return valid; |
1231 | } | |
1232 | ||
1233 | /* | |
1234 | * Check if cached dentry can be trusted. | |
1235 | */ | |
0b728e19 | 1236 | static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags) |
2817b000 | 1237 | { |
bf1c6aca | 1238 | int valid = 0; |
641235d8 | 1239 | struct dentry *parent; |
34286d66 NP |
1240 | struct inode *dir; |
1241 | ||
f49d1e05 | 1242 | if (flags & LOOKUP_RCU) { |
52953d55 | 1243 | parent = READ_ONCE(dentry->d_parent); |
f49d1e05 JL |
1244 | dir = d_inode_rcu(parent); |
1245 | if (!dir) | |
1246 | return -ECHILD; | |
1247 | } else { | |
1248 | parent = dget_parent(dentry); | |
1249 | dir = d_inode(parent); | |
1250 | } | |
34286d66 | 1251 | |
a455589f | 1252 | dout("d_revalidate %p '%pd' inode %p offset %lld\n", dentry, |
2b0143b5 | 1253 | dentry, d_inode(dentry), ceph_dentry(dentry)->offset); |
2817b000 SW |
1254 | |
1255 | /* always trust cached snapped dentries, snapdir dentry */ | |
1256 | if (ceph_snap(dir) != CEPH_NOSNAP) { | |
a455589f | 1257 | dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry, |
2b0143b5 | 1258 | dentry, d_inode(dentry)); |
bf1c6aca | 1259 | valid = 1; |
2b0143b5 DH |
1260 | } else if (d_really_is_positive(dentry) && |
1261 | ceph_snap(d_inode(dentry)) == CEPH_SNAPDIR) { | |
bf1c6aca | 1262 | valid = 1; |
14fb9c9e JL |
1263 | } else { |
1264 | valid = dentry_lease_is_valid(dentry, flags, dir); | |
1265 | if (valid == -ECHILD) | |
1266 | return valid; | |
1267 | if (valid || dir_lease_is_valid(dir, dentry)) { | |
1268 | if (d_really_is_positive(dentry)) | |
1269 | valid = ceph_is_any_caps(d_inode(dentry)); | |
1270 | else | |
1271 | valid = 1; | |
1272 | } | |
2817b000 | 1273 | } |
2817b000 | 1274 | |
200fd27c YZ |
1275 | if (!valid) { |
1276 | struct ceph_mds_client *mdsc = | |
1277 | ceph_sb_to_client(dir->i_sb)->mdsc; | |
1278 | struct ceph_mds_request *req; | |
1097680d JL |
1279 | int op, err; |
1280 | u32 mask; | |
200fd27c | 1281 | |
f49d1e05 JL |
1282 | if (flags & LOOKUP_RCU) |
1283 | return -ECHILD; | |
1284 | ||
200fd27c | 1285 | op = ceph_snap(dir) == CEPH_SNAPDIR ? |
5eb9f604 | 1286 | CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP; |
200fd27c YZ |
1287 | req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS); |
1288 | if (!IS_ERR(req)) { | |
1289 | req->r_dentry = dget(dentry); | |
5eb9f604 JL |
1290 | req->r_num_caps = 2; |
1291 | req->r_parent = dir; | |
200fd27c YZ |
1292 | |
1293 | mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED; | |
1294 | if (ceph_security_xattr_wanted(dir)) | |
1295 | mask |= CEPH_CAP_XATTR_SHARED; | |
1097680d | 1296 | req->r_args.getattr.mask = cpu_to_le32(mask); |
200fd27c | 1297 | |
200fd27c | 1298 | err = ceph_mdsc_do_request(mdsc, NULL, req); |
c3f4688a JL |
1299 | switch (err) { |
1300 | case 0: | |
1301 | if (d_really_is_positive(dentry) && | |
1302 | d_inode(dentry) == req->r_target_inode) | |
1303 | valid = 1; | |
1304 | break; | |
1305 | case -ENOENT: | |
1306 | if (d_really_is_negative(dentry)) | |
1307 | valid = 1; | |
1308 | /* Fallthrough */ | |
1309 | default: | |
1310 | break; | |
200fd27c YZ |
1311 | } |
1312 | ceph_mdsc_put_request(req); | |
1313 | dout("d_revalidate %p lookup result=%d\n", | |
1314 | dentry, err); | |
1315 | } | |
1316 | } | |
1317 | ||
bf1c6aca | 1318 | dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid"); |
9215aeea | 1319 | if (valid) { |
bf1c6aca | 1320 | ceph_dentry_lru_touch(dentry); |
9215aeea YZ |
1321 | } else { |
1322 | ceph_dir_clear_complete(dir); | |
9215aeea | 1323 | } |
641235d8 | 1324 | |
f49d1e05 JL |
1325 | if (!(flags & LOOKUP_RCU)) |
1326 | dput(parent); | |
bf1c6aca | 1327 | return valid; |
2817b000 SW |
1328 | } |
1329 | ||
1330 | /* | |
147851d2 | 1331 | * Release our ceph_dentry_info. |
2817b000 | 1332 | */ |
147851d2 | 1333 | static void ceph_d_release(struct dentry *dentry) |
2817b000 SW |
1334 | { |
1335 | struct ceph_dentry_info *di = ceph_dentry(dentry); | |
2817b000 | 1336 | |
147851d2 | 1337 | dout("d_release %p\n", dentry); |
3d8eb7a9 | 1338 | ceph_dentry_lru_del(dentry); |
5b484a51 JL |
1339 | |
1340 | spin_lock(&dentry->d_lock); | |
1341 | dentry->d_fsdata = NULL; | |
1342 | spin_unlock(&dentry->d_lock); | |
1343 | ||
3d8eb7a9 SW |
1344 | if (di->lease_session) |
1345 | ceph_put_mds_session(di->lease_session); | |
1346 | kmem_cache_free(ceph_dentry_cachep, di); | |
2817b000 SW |
1347 | } |
1348 | ||
b58dc410 SW |
1349 | /* |
1350 | * When the VFS prunes a dentry from the cache, we need to clear the | |
1351 | * complete flag on the parent directory. | |
1352 | * | |
1353 | * Called under dentry->d_lock. | |
1354 | */ | |
1355 | static void ceph_d_prune(struct dentry *dentry) | |
1356 | { | |
48697e8a YZ |
1357 | struct ceph_inode_info *dir_ci; |
1358 | struct ceph_dentry_info *di; | |
1359 | ||
1360 | dout("ceph_d_prune %pd %p\n", dentry, dentry); | |
b58dc410 SW |
1361 | |
1362 | /* do we have a valid parent? */ | |
8842b3be | 1363 | if (IS_ROOT(dentry)) |
b58dc410 SW |
1364 | return; |
1365 | ||
48697e8a YZ |
1366 | /* we hold d_lock, so d_parent is stable */ |
1367 | dir_ci = ceph_inode(d_inode(dentry->d_parent)); | |
1368 | if (dir_ci->i_vino.snap == CEPH_SNAPDIR) | |
b58dc410 | 1369 | return; |
2817b000 | 1370 | |
48697e8a YZ |
1371 | /* who calls d_delete() should also disable dcache readdir */ |
1372 | if (d_really_is_negative(dentry)) | |
18fc8abd AV |
1373 | return; |
1374 | ||
48697e8a YZ |
1375 | /* d_fsdata does not get cleared until d_release */ |
1376 | if (!d_unhashed(dentry)) { | |
1377 | __ceph_dir_clear_complete(dir_ci); | |
1378 | return; | |
1379 | } | |
1380 | ||
1381 | /* Disable dcache readdir just in case that someone called d_drop() | |
1382 | * or d_invalidate(), but MDS didn't revoke CEPH_CAP_FILE_SHARED | |
1383 | * properly (dcache readdir is still enabled) */ | |
1384 | di = ceph_dentry(dentry); | |
1385 | if (di->offset > 0 && | |
1386 | di->lease_shared_gen == atomic_read(&dir_ci->i_shared_gen)) | |
1387 | __ceph_dir_clear_ordered(dir_ci); | |
b58dc410 | 1388 | } |
2817b000 SW |
1389 | |
1390 | /* | |
1391 | * read() on a dir. This weird interface hack only works if mounted | |
1392 | * with '-o dirstat'. | |
1393 | */ | |
1394 | static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size, | |
1395 | loff_t *ppos) | |
1396 | { | |
1397 | struct ceph_file_info *cf = file->private_data; | |
496ad9aa | 1398 | struct inode *inode = file_inode(file); |
2817b000 SW |
1399 | struct ceph_inode_info *ci = ceph_inode(inode); |
1400 | int left; | |
ae598083 | 1401 | const int bufsize = 1024; |
2817b000 | 1402 | |
3d14c5d2 | 1403 | if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT)) |
2817b000 SW |
1404 | return -EISDIR; |
1405 | ||
1406 | if (!cf->dir_info) { | |
687265e5 | 1407 | cf->dir_info = kmalloc(bufsize, GFP_KERNEL); |
2817b000 SW |
1408 | if (!cf->dir_info) |
1409 | return -ENOMEM; | |
1410 | cf->dir_info_len = | |
ae598083 | 1411 | snprintf(cf->dir_info, bufsize, |
2817b000 SW |
1412 | "entries: %20lld\n" |
1413 | " files: %20lld\n" | |
1414 | " subdirs: %20lld\n" | |
1415 | "rentries: %20lld\n" | |
1416 | " rfiles: %20lld\n" | |
1417 | " rsubdirs: %20lld\n" | |
1418 | "rbytes: %20lld\n" | |
1419 | "rctime: %10ld.%09ld\n", | |
1420 | ci->i_files + ci->i_subdirs, | |
1421 | ci->i_files, | |
1422 | ci->i_subdirs, | |
1423 | ci->i_rfiles + ci->i_rsubdirs, | |
1424 | ci->i_rfiles, | |
1425 | ci->i_rsubdirs, | |
1426 | ci->i_rbytes, | |
1427 | (long)ci->i_rctime.tv_sec, | |
1428 | (long)ci->i_rctime.tv_nsec); | |
1429 | } | |
1430 | ||
1431 | if (*ppos >= cf->dir_info_len) | |
1432 | return 0; | |
1433 | size = min_t(unsigned, size, cf->dir_info_len-*ppos); | |
1434 | left = copy_to_user(buf, cf->dir_info + *ppos, size); | |
1435 | if (left == size) | |
1436 | return -EFAULT; | |
1437 | *ppos += (size - left); | |
1438 | return size - left; | |
1439 | } | |
1440 | ||
2817b000 SW |
1441 | /* |
1442 | * We maintain a private dentry LRU. | |
1443 | * | |
1444 | * FIXME: this needs to be changed to a per-mds lru to be useful. | |
1445 | */ | |
1446 | void ceph_dentry_lru_add(struct dentry *dn) | |
1447 | { | |
1448 | struct ceph_dentry_info *di = ceph_dentry(dn); | |
1449 | struct ceph_mds_client *mdsc; | |
2817b000 | 1450 | |
a455589f | 1451 | dout("dentry_lru_add %p %p '%pd'\n", di, dn, dn); |
3d8eb7a9 SW |
1452 | mdsc = ceph_sb_to_client(dn->d_sb)->mdsc; |
1453 | spin_lock(&mdsc->dentry_lru_lock); | |
1454 | list_add_tail(&di->lru, &mdsc->dentry_lru); | |
1455 | mdsc->num_dentry++; | |
1456 | spin_unlock(&mdsc->dentry_lru_lock); | |
2817b000 SW |
1457 | } |
1458 | ||
1459 | void ceph_dentry_lru_touch(struct dentry *dn) | |
1460 | { | |
1461 | struct ceph_dentry_info *di = ceph_dentry(dn); | |
1462 | struct ceph_mds_client *mdsc; | |
2817b000 | 1463 | |
a455589f AV |
1464 | dout("dentry_lru_touch %p %p '%pd' (offset %lld)\n", di, dn, dn, |
1465 | di->offset); | |
3d8eb7a9 SW |
1466 | mdsc = ceph_sb_to_client(dn->d_sb)->mdsc; |
1467 | spin_lock(&mdsc->dentry_lru_lock); | |
1468 | list_move_tail(&di->lru, &mdsc->dentry_lru); | |
1469 | spin_unlock(&mdsc->dentry_lru_lock); | |
2817b000 SW |
1470 | } |
1471 | ||
1472 | void ceph_dentry_lru_del(struct dentry *dn) | |
1473 | { | |
1474 | struct ceph_dentry_info *di = ceph_dentry(dn); | |
1475 | struct ceph_mds_client *mdsc; | |
1476 | ||
a455589f | 1477 | dout("dentry_lru_del %p %p '%pd'\n", di, dn, dn); |
3d8eb7a9 SW |
1478 | mdsc = ceph_sb_to_client(dn->d_sb)->mdsc; |
1479 | spin_lock(&mdsc->dentry_lru_lock); | |
1480 | list_del_init(&di->lru); | |
1481 | mdsc->num_dentry--; | |
1482 | spin_unlock(&mdsc->dentry_lru_lock); | |
2817b000 SW |
1483 | } |
1484 | ||
6c0f3af7 SW |
1485 | /* |
1486 | * Return name hash for a given dentry. This is dependent on | |
1487 | * the parent directory's hash function. | |
1488 | */ | |
e5f86dc3 | 1489 | unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn) |
6c0f3af7 | 1490 | { |
6c0f3af7 SW |
1491 | struct ceph_inode_info *dci = ceph_inode(dir); |
1492 | ||
1493 | switch (dci->i_dir_layout.dl_dir_hash) { | |
1494 | case 0: /* for backward compat */ | |
1495 | case CEPH_STR_HASH_LINUX: | |
1496 | return dn->d_name.hash; | |
1497 | ||
1498 | default: | |
1499 | return ceph_str_hash(dci->i_dir_layout.dl_dir_hash, | |
1500 | dn->d_name.name, dn->d_name.len); | |
1501 | } | |
1502 | } | |
1503 | ||
2817b000 SW |
1504 | const struct file_operations ceph_dir_fops = { |
1505 | .read = ceph_read_dir, | |
77acfa29 | 1506 | .iterate = ceph_readdir, |
2817b000 SW |
1507 | .llseek = ceph_dir_llseek, |
1508 | .open = ceph_open, | |
1509 | .release = ceph_release, | |
1510 | .unlocked_ioctl = ceph_ioctl, | |
da819c81 | 1511 | .fsync = ceph_fsync, |
2817b000 SW |
1512 | }; |
1513 | ||
38c48b5f YZ |
1514 | const struct file_operations ceph_snapdir_fops = { |
1515 | .iterate = ceph_readdir, | |
1516 | .llseek = ceph_dir_llseek, | |
1517 | .open = ceph_open, | |
1518 | .release = ceph_release, | |
1519 | }; | |
1520 | ||
2817b000 SW |
1521 | const struct inode_operations ceph_dir_iops = { |
1522 | .lookup = ceph_lookup, | |
1523 | .permission = ceph_permission, | |
1524 | .getattr = ceph_getattr, | |
1525 | .setattr = ceph_setattr, | |
2817b000 | 1526 | .listxattr = ceph_listxattr, |
7221fe4c | 1527 | .get_acl = ceph_get_acl, |
72466d0b | 1528 | .set_acl = ceph_set_acl, |
2817b000 SW |
1529 | .mknod = ceph_mknod, |
1530 | .symlink = ceph_symlink, | |
1531 | .mkdir = ceph_mkdir, | |
1532 | .link = ceph_link, | |
1533 | .unlink = ceph_unlink, | |
1534 | .rmdir = ceph_unlink, | |
1535 | .rename = ceph_rename, | |
1536 | .create = ceph_create, | |
2d83bde9 | 1537 | .atomic_open = ceph_atomic_open, |
2817b000 SW |
1538 | }; |
1539 | ||
38c48b5f YZ |
1540 | const struct inode_operations ceph_snapdir_iops = { |
1541 | .lookup = ceph_lookup, | |
1542 | .permission = ceph_permission, | |
1543 | .getattr = ceph_getattr, | |
1544 | .mkdir = ceph_mkdir, | |
1545 | .rmdir = ceph_unlink, | |
0ea611a3 | 1546 | .rename = ceph_rename, |
38c48b5f YZ |
1547 | }; |
1548 | ||
52dfb8ac | 1549 | const struct dentry_operations ceph_dentry_ops = { |
2817b000 | 1550 | .d_revalidate = ceph_d_revalidate, |
147851d2 | 1551 | .d_release = ceph_d_release, |
b58dc410 | 1552 | .d_prune = ceph_d_prune, |
ad5cb123 | 1553 | .d_init = ceph_d_init, |
2817b000 | 1554 | }; |