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