1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* |
3 | * linux/fs/nfs/unlink.c |
4 | * |
5 | * nfs sillydelete handling |
6 | * |
7 | */ |
8 | |
9 | #include <linux/slab.h> |
10 | #include <linux/string.h> |
11 | #include <linux/dcache.h> |
12 | #include <linux/sunrpc/sched.h> |
13 | #include <linux/sunrpc/clnt.h> |
14 | #include <linux/nfs_fs.h> |
15 | #include <linux/sched.h> |
16 | #include <linux/wait.h> |
17 | #include <linux/namei.h> |
18 | #include <linux/fsnotify.h> |
19 | |
20 | #include "internal.h" |
21 | #include "nfs4_fs.h" |
22 | #include "iostat.h" |
23 | #include "delegation.h" |
24 | |
25 | #include "nfstrace.h" |
26 | |
27 | /** |
28 | * nfs_free_unlinkdata - release data from a sillydelete operation. |
29 | * @data: pointer to unlink structure. |
30 | */ |
31 | static void |
32 | nfs_free_unlinkdata(struct nfs_unlinkdata *data) |
33 | { |
34 | put_cred(cred: data->cred); |
35 | kfree(objp: data->args.name.name); |
36 | kfree(objp: data); |
37 | } |
38 | |
39 | /** |
40 | * nfs_async_unlink_done - Sillydelete post-processing |
41 | * @task: rpc_task of the sillydelete |
42 | * @calldata: pointer to nfs_unlinkdata |
43 | * |
44 | * Do the directory attribute update. |
45 | */ |
46 | static void nfs_async_unlink_done(struct rpc_task *task, void *calldata) |
47 | { |
48 | struct nfs_unlinkdata *data = calldata; |
49 | struct inode *dir = d_inode(dentry: data->dentry->d_parent); |
50 | |
51 | trace_nfs_sillyrename_unlink(data, error: task->tk_status); |
52 | if (!NFS_PROTO(inode: dir)->unlink_done(task, dir)) |
53 | rpc_restart_call_prepare(task); |
54 | } |
55 | |
56 | /** |
57 | * nfs_async_unlink_release - Release the sillydelete data. |
58 | * @calldata: struct nfs_unlinkdata to release |
59 | * |
60 | * We need to call nfs_put_unlinkdata as a 'tk_release' task since the |
61 | * rpc_task would be freed too. |
62 | */ |
63 | static void nfs_async_unlink_release(void *calldata) |
64 | { |
65 | struct nfs_unlinkdata *data = calldata; |
66 | struct dentry *dentry = data->dentry; |
67 | struct super_block *sb = dentry->d_sb; |
68 | |
69 | up_read_non_owner(sem: &NFS_I(inode: d_inode(dentry: dentry->d_parent))->rmdir_sem); |
70 | d_lookup_done(dentry); |
71 | nfs_free_unlinkdata(data); |
72 | dput(dentry); |
73 | nfs_sb_deactive(sb); |
74 | } |
75 | |
76 | static void nfs_unlink_prepare(struct rpc_task *task, void *calldata) |
77 | { |
78 | struct nfs_unlinkdata *data = calldata; |
79 | struct inode *dir = d_inode(dentry: data->dentry->d_parent); |
80 | NFS_PROTO(inode: dir)->unlink_rpc_prepare(task, data); |
81 | } |
82 | |
83 | static const struct rpc_call_ops nfs_unlink_ops = { |
84 | .rpc_call_done = nfs_async_unlink_done, |
85 | .rpc_release = nfs_async_unlink_release, |
86 | .rpc_call_prepare = nfs_unlink_prepare, |
87 | }; |
88 | |
89 | static void nfs_do_call_unlink(struct inode *inode, struct nfs_unlinkdata *data) |
90 | { |
91 | struct rpc_message msg = { |
92 | .rpc_argp = &data->args, |
93 | .rpc_resp = &data->res, |
94 | .rpc_cred = data->cred, |
95 | }; |
96 | struct rpc_task_setup task_setup_data = { |
97 | .rpc_message = &msg, |
98 | .callback_ops = &nfs_unlink_ops, |
99 | .callback_data = data, |
100 | .workqueue = nfsiod_workqueue, |
101 | .flags = RPC_TASK_ASYNC | RPC_TASK_CRED_NOREF, |
102 | }; |
103 | struct rpc_task *task; |
104 | struct inode *dir = d_inode(dentry: data->dentry->d_parent); |
105 | |
106 | if (nfs_server_capable(inode, NFS_CAP_MOVEABLE)) |
107 | task_setup_data.flags |= RPC_TASK_MOVEABLE; |
108 | |
109 | nfs_sb_active(sb: dir->i_sb); |
110 | data->args.fh = NFS_FH(inode: dir); |
111 | nfs_fattr_init(fattr: data->res.dir_attr); |
112 | |
113 | NFS_PROTO(inode: dir)->unlink_setup(&msg, data->dentry, inode); |
114 | |
115 | task_setup_data.rpc_client = NFS_CLIENT(inode: dir); |
116 | task = rpc_run_task(&task_setup_data); |
117 | if (!IS_ERR(ptr: task)) |
118 | rpc_put_task_async(task); |
119 | } |
120 | |
121 | static int nfs_call_unlink(struct dentry *dentry, struct inode *inode, struct nfs_unlinkdata *data) |
122 | { |
123 | struct inode *dir = d_inode(dentry: dentry->d_parent); |
124 | struct dentry *alias; |
125 | |
126 | down_read_non_owner(sem: &NFS_I(inode: dir)->rmdir_sem); |
127 | alias = d_alloc_parallel(dentry->d_parent, &data->args.name, &data->wq); |
128 | if (IS_ERR(ptr: alias)) { |
129 | up_read_non_owner(sem: &NFS_I(inode: dir)->rmdir_sem); |
130 | return 0; |
131 | } |
132 | if (!d_in_lookup(dentry: alias)) { |
133 | int ret; |
134 | void *devname_garbage = NULL; |
135 | |
136 | /* |
137 | * Hey, we raced with lookup... See if we need to transfer |
138 | * the sillyrename information to the aliased dentry. |
139 | */ |
140 | spin_lock(lock: &alias->d_lock); |
141 | if (d_really_is_positive(dentry: alias) && |
142 | !nfs_compare_fh(a: NFS_FH(inode), b: NFS_FH(inode: d_inode(dentry: alias))) && |
143 | !(alias->d_flags & DCACHE_NFSFS_RENAMED)) { |
144 | devname_garbage = alias->d_fsdata; |
145 | alias->d_fsdata = data; |
146 | alias->d_flags |= DCACHE_NFSFS_RENAMED; |
147 | ret = 1; |
148 | } else |
149 | ret = 0; |
150 | spin_unlock(lock: &alias->d_lock); |
151 | dput(alias); |
152 | up_read_non_owner(sem: &NFS_I(inode: dir)->rmdir_sem); |
153 | /* |
154 | * If we'd displaced old cached devname, free it. At that |
155 | * point dentry is definitely not a root, so we won't need |
156 | * that anymore. |
157 | */ |
158 | kfree(objp: devname_garbage); |
159 | return ret; |
160 | } |
161 | data->dentry = alias; |
162 | nfs_do_call_unlink(inode, data); |
163 | return 1; |
164 | } |
165 | |
166 | /** |
167 | * nfs_async_unlink - asynchronous unlinking of a file |
168 | * @dentry: parent directory of dentry |
169 | * @name: name of dentry to unlink |
170 | */ |
171 | static int |
172 | nfs_async_unlink(struct dentry *dentry, const struct qstr *name) |
173 | { |
174 | struct nfs_unlinkdata *data; |
175 | int status = -ENOMEM; |
176 | void *devname_garbage = NULL; |
177 | |
178 | data = kzalloc(size: sizeof(*data), GFP_KERNEL); |
179 | if (data == NULL) |
180 | goto out; |
181 | data->args.name.name = kstrdup(s: name->name, GFP_KERNEL); |
182 | if (!data->args.name.name) |
183 | goto out_free; |
184 | data->args.name.len = name->len; |
185 | |
186 | data->cred = get_current_cred(); |
187 | data->res.dir_attr = &data->dir_attr; |
188 | init_waitqueue_head(&data->wq); |
189 | |
190 | status = -EBUSY; |
191 | spin_lock(lock: &dentry->d_lock); |
192 | if (dentry->d_flags & DCACHE_NFSFS_RENAMED) |
193 | goto out_unlock; |
194 | dentry->d_flags |= DCACHE_NFSFS_RENAMED; |
195 | devname_garbage = dentry->d_fsdata; |
196 | dentry->d_fsdata = data; |
197 | spin_unlock(lock: &dentry->d_lock); |
198 | /* |
199 | * If we'd displaced old cached devname, free it. At that |
200 | * point dentry is definitely not a root, so we won't need |
201 | * that anymore. |
202 | */ |
203 | kfree(objp: devname_garbage); |
204 | return 0; |
205 | out_unlock: |
206 | spin_unlock(lock: &dentry->d_lock); |
207 | put_cred(cred: data->cred); |
208 | kfree(objp: data->args.name.name); |
209 | out_free: |
210 | kfree(objp: data); |
211 | out: |
212 | return status; |
213 | } |
214 | |
215 | /** |
216 | * nfs_complete_unlink - Initialize completion of the sillydelete |
217 | * @dentry: dentry to delete |
218 | * @inode: inode |
219 | * |
220 | * Since we're most likely to be called by dentry_iput(), we |
221 | * only use the dentry to find the sillydelete. We then copy the name |
222 | * into the qstr. |
223 | */ |
224 | void |
225 | nfs_complete_unlink(struct dentry *dentry, struct inode *inode) |
226 | { |
227 | struct nfs_unlinkdata *data; |
228 | |
229 | spin_lock(lock: &dentry->d_lock); |
230 | dentry->d_flags &= ~DCACHE_NFSFS_RENAMED; |
231 | data = dentry->d_fsdata; |
232 | dentry->d_fsdata = NULL; |
233 | spin_unlock(lock: &dentry->d_lock); |
234 | |
235 | if (NFS_STALE(inode) || !nfs_call_unlink(dentry, inode, data)) |
236 | nfs_free_unlinkdata(data); |
237 | } |
238 | |
239 | /* Cancel a queued async unlink. Called when a sillyrename run fails. */ |
240 | static void |
241 | nfs_cancel_async_unlink(struct dentry *dentry) |
242 | { |
243 | spin_lock(lock: &dentry->d_lock); |
244 | if (dentry->d_flags & DCACHE_NFSFS_RENAMED) { |
245 | struct nfs_unlinkdata *data = dentry->d_fsdata; |
246 | |
247 | dentry->d_flags &= ~DCACHE_NFSFS_RENAMED; |
248 | dentry->d_fsdata = NULL; |
249 | spin_unlock(lock: &dentry->d_lock); |
250 | nfs_free_unlinkdata(data); |
251 | return; |
252 | } |
253 | spin_unlock(lock: &dentry->d_lock); |
254 | } |
255 | |
256 | /** |
257 | * nfs_async_rename_done - Sillyrename post-processing |
258 | * @task: rpc_task of the sillyrename |
259 | * @calldata: nfs_renamedata for the sillyrename |
260 | * |
261 | * Do the directory attribute updates and the d_move |
262 | */ |
263 | static void nfs_async_rename_done(struct rpc_task *task, void *calldata) |
264 | { |
265 | struct nfs_renamedata *data = calldata; |
266 | struct inode *old_dir = data->old_dir; |
267 | struct inode *new_dir = data->new_dir; |
268 | struct dentry *old_dentry = data->old_dentry; |
269 | |
270 | trace_nfs_sillyrename_rename(old_dir, old_dentry, |
271 | new_dir, new_dentry: data->new_dentry, error: task->tk_status); |
272 | if (!NFS_PROTO(inode: old_dir)->rename_done(task, old_dir, new_dir)) { |
273 | rpc_restart_call_prepare(task); |
274 | return; |
275 | } |
276 | |
277 | if (data->complete) |
278 | data->complete(task, data); |
279 | } |
280 | |
281 | /** |
282 | * nfs_async_rename_release - Release the sillyrename data. |
283 | * @calldata: the struct nfs_renamedata to be released |
284 | */ |
285 | static void nfs_async_rename_release(void *calldata) |
286 | { |
287 | struct nfs_renamedata *data = calldata; |
288 | struct super_block *sb = data->old_dir->i_sb; |
289 | |
290 | if (d_really_is_positive(dentry: data->old_dentry)) |
291 | nfs_mark_for_revalidate(inode: d_inode(dentry: data->old_dentry)); |
292 | |
293 | /* The result of the rename is unknown. Play it safe by |
294 | * forcing a new lookup */ |
295 | if (data->cancelled) { |
296 | spin_lock(lock: &data->old_dir->i_lock); |
297 | nfs_force_lookup_revalidate(dir: data->old_dir); |
298 | spin_unlock(lock: &data->old_dir->i_lock); |
299 | if (data->new_dir != data->old_dir) { |
300 | spin_lock(lock: &data->new_dir->i_lock); |
301 | nfs_force_lookup_revalidate(dir: data->new_dir); |
302 | spin_unlock(lock: &data->new_dir->i_lock); |
303 | } |
304 | } |
305 | |
306 | dput(data->old_dentry); |
307 | dput(data->new_dentry); |
308 | iput(data->old_dir); |
309 | iput(data->new_dir); |
310 | nfs_sb_deactive(sb); |
311 | put_cred(cred: data->cred); |
312 | kfree(objp: data); |
313 | } |
314 | |
315 | static void nfs_rename_prepare(struct rpc_task *task, void *calldata) |
316 | { |
317 | struct nfs_renamedata *data = calldata; |
318 | NFS_PROTO(inode: data->old_dir)->rename_rpc_prepare(task, data); |
319 | } |
320 | |
321 | static const struct rpc_call_ops nfs_rename_ops = { |
322 | .rpc_call_done = nfs_async_rename_done, |
323 | .rpc_release = nfs_async_rename_release, |
324 | .rpc_call_prepare = nfs_rename_prepare, |
325 | }; |
326 | |
327 | /** |
328 | * nfs_async_rename - perform an asynchronous rename operation |
329 | * @old_dir: directory that currently holds the dentry to be renamed |
330 | * @new_dir: target directory for the rename |
331 | * @old_dentry: original dentry to be renamed |
332 | * @new_dentry: dentry to which the old_dentry should be renamed |
333 | * @complete: Function to run on successful completion |
334 | * |
335 | * It's expected that valid references to the dentries and inodes are held |
336 | */ |
337 | struct rpc_task * |
338 | nfs_async_rename(struct inode *old_dir, struct inode *new_dir, |
339 | struct dentry *old_dentry, struct dentry *new_dentry, |
340 | void (*complete)(struct rpc_task *, struct nfs_renamedata *)) |
341 | { |
342 | struct nfs_renamedata *data; |
343 | struct rpc_message msg = { }; |
344 | struct rpc_task_setup task_setup_data = { |
345 | .rpc_message = &msg, |
346 | .callback_ops = &nfs_rename_ops, |
347 | .workqueue = nfsiod_workqueue, |
348 | .rpc_client = NFS_CLIENT(inode: old_dir), |
349 | .flags = RPC_TASK_ASYNC | RPC_TASK_CRED_NOREF, |
350 | }; |
351 | |
352 | if (nfs_server_capable(inode: old_dir, NFS_CAP_MOVEABLE) && |
353 | nfs_server_capable(inode: new_dir, NFS_CAP_MOVEABLE)) |
354 | task_setup_data.flags |= RPC_TASK_MOVEABLE; |
355 | |
356 | data = kzalloc(size: sizeof(*data), GFP_KERNEL); |
357 | if (data == NULL) |
358 | return ERR_PTR(error: -ENOMEM); |
359 | task_setup_data.task = &data->task; |
360 | task_setup_data.callback_data = data; |
361 | |
362 | data->cred = get_current_cred(); |
363 | |
364 | msg.rpc_argp = &data->args; |
365 | msg.rpc_resp = &data->res; |
366 | msg.rpc_cred = data->cred; |
367 | |
368 | /* set up nfs_renamedata */ |
369 | data->old_dir = old_dir; |
370 | ihold(inode: old_dir); |
371 | data->new_dir = new_dir; |
372 | ihold(inode: new_dir); |
373 | data->old_dentry = dget(dentry: old_dentry); |
374 | data->new_dentry = dget(dentry: new_dentry); |
375 | nfs_fattr_init(fattr: &data->old_fattr); |
376 | nfs_fattr_init(fattr: &data->new_fattr); |
377 | data->complete = complete; |
378 | |
379 | /* set up nfs_renameargs */ |
380 | data->args.old_dir = NFS_FH(inode: old_dir); |
381 | data->args.old_name = &old_dentry->d_name; |
382 | data->args.new_dir = NFS_FH(inode: new_dir); |
383 | data->args.new_name = &new_dentry->d_name; |
384 | |
385 | /* set up nfs_renameres */ |
386 | data->res.old_fattr = &data->old_fattr; |
387 | data->res.new_fattr = &data->new_fattr; |
388 | |
389 | nfs_sb_active(sb: old_dir->i_sb); |
390 | |
391 | NFS_PROTO(inode: data->old_dir)->rename_setup(&msg, old_dentry, new_dentry); |
392 | |
393 | return rpc_run_task(&task_setup_data); |
394 | } |
395 | |
396 | /* |
397 | * Perform tasks needed when a sillyrename is done such as cancelling the |
398 | * queued async unlink if it failed. |
399 | */ |
400 | static void |
401 | nfs_complete_sillyrename(struct rpc_task *task, struct nfs_renamedata *data) |
402 | { |
403 | struct dentry *dentry = data->old_dentry; |
404 | |
405 | if (task->tk_status != 0) { |
406 | nfs_cancel_async_unlink(dentry); |
407 | return; |
408 | } |
409 | } |
410 | |
411 | #define SILLYNAME_PREFIX ".nfs" |
412 | #define SILLYNAME_PREFIX_LEN ((unsigned)sizeof(SILLYNAME_PREFIX) - 1) |
413 | #define SILLYNAME_FILEID_LEN ((unsigned)sizeof(u64) << 1) |
414 | #define SILLYNAME_COUNTER_LEN ((unsigned)sizeof(unsigned int) << 1) |
415 | #define SILLYNAME_LEN (SILLYNAME_PREFIX_LEN + \ |
416 | SILLYNAME_FILEID_LEN + \ |
417 | SILLYNAME_COUNTER_LEN) |
418 | |
419 | /** |
420 | * nfs_sillyrename - Perform a silly-rename of a dentry |
421 | * @dir: inode of directory that contains dentry |
422 | * @dentry: dentry to be sillyrenamed |
423 | * |
424 | * NFSv2/3 is stateless and the server doesn't know when the client is |
425 | * holding a file open. To prevent application problems when a file is |
426 | * unlinked while it's still open, the client performs a "silly-rename". |
427 | * That is, it renames the file to a hidden file in the same directory, |
428 | * and only performs the unlink once the last reference to it is put. |
429 | * |
430 | * The final cleanup is done during dentry_iput. |
431 | * |
432 | * (Note: NFSv4 is stateful, and has opens, so in theory an NFSv4 server |
433 | * could take responsibility for keeping open files referenced. The server |
434 | * would also need to ensure that opened-but-deleted files were kept over |
435 | * reboots. However, we may not assume a server does so. (RFC 5661 |
436 | * does provide an OPEN4_RESULT_PRESERVE_UNLINKED flag that a server can |
437 | * use to advertise that it does this; some day we may take advantage of |
438 | * it.)) |
439 | */ |
440 | int |
441 | nfs_sillyrename(struct inode *dir, struct dentry *dentry) |
442 | { |
443 | static unsigned int sillycounter; |
444 | unsigned char silly[SILLYNAME_LEN + 1]; |
445 | unsigned long long fileid; |
446 | struct dentry *sdentry; |
447 | struct inode *inode = d_inode(dentry); |
448 | struct rpc_task *task; |
449 | int error = -EBUSY; |
450 | |
451 | dfprintk(VFS, "NFS: silly-rename(%pd2, ct=%d)\n" , |
452 | dentry, d_count(dentry)); |
453 | nfs_inc_stats(inode: dir, stat: NFSIOS_SILLYRENAME); |
454 | |
455 | /* |
456 | * We don't allow a dentry to be silly-renamed twice. |
457 | */ |
458 | if (dentry->d_flags & DCACHE_NFSFS_RENAMED) |
459 | goto out; |
460 | |
461 | fileid = NFS_FILEID(inode: d_inode(dentry)); |
462 | |
463 | sdentry = NULL; |
464 | do { |
465 | int slen; |
466 | dput(sdentry); |
467 | sillycounter++; |
468 | slen = scnprintf(buf: silly, size: sizeof(silly), |
469 | SILLYNAME_PREFIX "%0*llx%0*x" , |
470 | SILLYNAME_FILEID_LEN, fileid, |
471 | SILLYNAME_COUNTER_LEN, sillycounter); |
472 | |
473 | dfprintk(VFS, "NFS: trying to rename %pd to %s\n" , |
474 | dentry, silly); |
475 | |
476 | sdentry = lookup_one_len(silly, dentry->d_parent, slen); |
477 | /* |
478 | * N.B. Better to return EBUSY here ... it could be |
479 | * dangerous to delete the file while it's in use. |
480 | */ |
481 | if (IS_ERR(ptr: sdentry)) |
482 | goto out; |
483 | } while (d_inode(dentry: sdentry) != NULL); /* need negative lookup */ |
484 | |
485 | ihold(inode); |
486 | |
487 | /* queue unlink first. Can't do this from rpc_release as it |
488 | * has to allocate memory |
489 | */ |
490 | error = nfs_async_unlink(dentry, name: &sdentry->d_name); |
491 | if (error) |
492 | goto out_dput; |
493 | |
494 | /* run the rename task, undo unlink if it fails */ |
495 | task = nfs_async_rename(old_dir: dir, new_dir: dir, old_dentry: dentry, new_dentry: sdentry, |
496 | complete: nfs_complete_sillyrename); |
497 | if (IS_ERR(ptr: task)) { |
498 | error = -EBUSY; |
499 | nfs_cancel_async_unlink(dentry); |
500 | goto out_dput; |
501 | } |
502 | |
503 | /* wait for the RPC task to complete, unless a SIGKILL intervenes */ |
504 | error = rpc_wait_for_completion_task(task); |
505 | if (error == 0) |
506 | error = task->tk_status; |
507 | switch (error) { |
508 | case 0: |
509 | /* The rename succeeded */ |
510 | nfs_set_verifier(dentry, verf: nfs_save_change_attribute(dir)); |
511 | spin_lock(lock: &inode->i_lock); |
512 | NFS_I(inode)->attr_gencount = nfs_inc_attr_generation_counter(); |
513 | nfs_set_cache_invalid(inode, NFS_INO_INVALID_CHANGE | |
514 | NFS_INO_INVALID_CTIME | |
515 | NFS_INO_REVAL_FORCED); |
516 | spin_unlock(lock: &inode->i_lock); |
517 | d_move(dentry, sdentry); |
518 | break; |
519 | case -ERESTARTSYS: |
520 | /* The result of the rename is unknown. Play it safe by |
521 | * forcing a new lookup */ |
522 | d_drop(dentry); |
523 | d_drop(dentry: sdentry); |
524 | } |
525 | rpc_put_task(task); |
526 | out_dput: |
527 | iput(inode); |
528 | dput(sdentry); |
529 | out: |
530 | return error; |
531 | } |
532 | |