Merge tag 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mst/vhost
[linux-2.6-microblaze.git] / fs / nfs / unlink.c
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(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  *
43  * Do the directory attribute update.
44  */
45 static void nfs_async_unlink_done(struct rpc_task *task, void *calldata)
46 {
47         struct nfs_unlinkdata *data = calldata;
48         struct inode *dir = d_inode(data->dentry->d_parent);
49
50         trace_nfs_sillyrename_unlink(data, task->tk_status);
51         if (!NFS_PROTO(dir)->unlink_done(task, dir))
52                 rpc_restart_call_prepare(task);
53 }
54
55 /**
56  * nfs_async_unlink_release - Release the sillydelete data.
57  * @task: rpc_task of the sillydelete
58  *
59  * We need to call nfs_put_unlinkdata as a 'tk_release' task since the
60  * rpc_task would be freed too.
61  */
62 static void nfs_async_unlink_release(void *calldata)
63 {
64         struct nfs_unlinkdata   *data = calldata;
65         struct dentry *dentry = data->dentry;
66         struct super_block *sb = dentry->d_sb;
67
68         up_read_non_owner(&NFS_I(d_inode(dentry->d_parent))->rmdir_sem);
69         d_lookup_done(dentry);
70         nfs_free_unlinkdata(data);
71         dput(dentry);
72         nfs_sb_deactive(sb);
73 }
74
75 static void nfs_unlink_prepare(struct rpc_task *task, void *calldata)
76 {
77         struct nfs_unlinkdata *data = calldata;
78         struct inode *dir = d_inode(data->dentry->d_parent);
79         NFS_PROTO(dir)->unlink_rpc_prepare(task, data);
80 }
81
82 static const struct rpc_call_ops nfs_unlink_ops = {
83         .rpc_call_done = nfs_async_unlink_done,
84         .rpc_release = nfs_async_unlink_release,
85         .rpc_call_prepare = nfs_unlink_prepare,
86 };
87
88 static void nfs_do_call_unlink(struct inode *inode, struct nfs_unlinkdata *data)
89 {
90         struct rpc_message msg = {
91                 .rpc_argp = &data->args,
92                 .rpc_resp = &data->res,
93                 .rpc_cred = data->cred,
94         };
95         struct rpc_task_setup task_setup_data = {
96                 .rpc_message = &msg,
97                 .callback_ops = &nfs_unlink_ops,
98                 .callback_data = data,
99                 .workqueue = nfsiod_workqueue,
100                 .flags = RPC_TASK_ASYNC,
101         };
102         struct rpc_task *task;
103         struct inode *dir = d_inode(data->dentry->d_parent);
104         nfs_sb_active(dir->i_sb);
105         data->args.fh = NFS_FH(dir);
106         nfs_fattr_init(data->res.dir_attr);
107
108         NFS_PROTO(dir)->unlink_setup(&msg, data->dentry, inode);
109
110         task_setup_data.rpc_client = NFS_CLIENT(dir);
111         task = rpc_run_task(&task_setup_data);
112         if (!IS_ERR(task))
113                 rpc_put_task_async(task);
114 }
115
116 static int nfs_call_unlink(struct dentry *dentry, struct inode *inode, struct nfs_unlinkdata *data)
117 {
118         struct inode *dir = d_inode(dentry->d_parent);
119         struct dentry *alias;
120
121         down_read_non_owner(&NFS_I(dir)->rmdir_sem);
122         alias = d_alloc_parallel(dentry->d_parent, &data->args.name, &data->wq);
123         if (IS_ERR(alias)) {
124                 up_read_non_owner(&NFS_I(dir)->rmdir_sem);
125                 return 0;
126         }
127         if (!d_in_lookup(alias)) {
128                 int ret;
129                 void *devname_garbage = NULL;
130
131                 /*
132                  * Hey, we raced with lookup... See if we need to transfer
133                  * the sillyrename information to the aliased dentry.
134                  */
135                 spin_lock(&alias->d_lock);
136                 if (d_really_is_positive(alias) &&
137                     !(alias->d_flags & DCACHE_NFSFS_RENAMED)) {
138                         devname_garbage = alias->d_fsdata;
139                         alias->d_fsdata = data;
140                         alias->d_flags |= DCACHE_NFSFS_RENAMED;
141                         ret = 1;
142                 } else
143                         ret = 0;
144                 spin_unlock(&alias->d_lock);
145                 dput(alias);
146                 up_read_non_owner(&NFS_I(dir)->rmdir_sem);
147                 /*
148                  * If we'd displaced old cached devname, free it.  At that
149                  * point dentry is definitely not a root, so we won't need
150                  * that anymore.
151                  */
152                 kfree(devname_garbage);
153                 return ret;
154         }
155         data->dentry = alias;
156         nfs_do_call_unlink(inode, data);
157         return 1;
158 }
159
160 /**
161  * nfs_async_unlink - asynchronous unlinking of a file
162  * @dir: parent directory of dentry
163  * @dentry: dentry to unlink
164  */
165 static int
166 nfs_async_unlink(struct dentry *dentry, const struct qstr *name)
167 {
168         struct nfs_unlinkdata *data;
169         int status = -ENOMEM;
170         void *devname_garbage = NULL;
171
172         data = kzalloc(sizeof(*data), GFP_KERNEL);
173         if (data == NULL)
174                 goto out;
175         data->args.name.name = kstrdup(name->name, GFP_KERNEL);
176         if (!data->args.name.name)
177                 goto out_free;
178         data->args.name.len = name->len;
179
180         data->cred = get_current_cred();
181         data->res.dir_attr = &data->dir_attr;
182         init_waitqueue_head(&data->wq);
183
184         status = -EBUSY;
185         spin_lock(&dentry->d_lock);
186         if (dentry->d_flags & DCACHE_NFSFS_RENAMED)
187                 goto out_unlock;
188         dentry->d_flags |= DCACHE_NFSFS_RENAMED;
189         devname_garbage = dentry->d_fsdata;
190         dentry->d_fsdata = data;
191         spin_unlock(&dentry->d_lock);
192         /*
193          * If we'd displaced old cached devname, free it.  At that
194          * point dentry is definitely not a root, so we won't need
195          * that anymore.
196          */
197         kfree(devname_garbage);
198         return 0;
199 out_unlock:
200         spin_unlock(&dentry->d_lock);
201         put_cred(data->cred);
202         kfree(data->args.name.name);
203 out_free:
204         kfree(data);
205 out:
206         return status;
207 }
208
209 /**
210  * nfs_complete_unlink - Initialize completion of the sillydelete
211  * @dentry: dentry to delete
212  * @inode: inode
213  *
214  * Since we're most likely to be called by dentry_iput(), we
215  * only use the dentry to find the sillydelete. We then copy the name
216  * into the qstr.
217  */
218 void
219 nfs_complete_unlink(struct dentry *dentry, struct inode *inode)
220 {
221         struct nfs_unlinkdata   *data;
222
223         spin_lock(&dentry->d_lock);
224         dentry->d_flags &= ~DCACHE_NFSFS_RENAMED;
225         data = dentry->d_fsdata;
226         dentry->d_fsdata = NULL;
227         spin_unlock(&dentry->d_lock);
228
229         if (NFS_STALE(inode) || !nfs_call_unlink(dentry, inode, data))
230                 nfs_free_unlinkdata(data);
231 }
232
233 /* Cancel a queued async unlink. Called when a sillyrename run fails. */
234 static void
235 nfs_cancel_async_unlink(struct dentry *dentry)
236 {
237         spin_lock(&dentry->d_lock);
238         if (dentry->d_flags & DCACHE_NFSFS_RENAMED) {
239                 struct nfs_unlinkdata *data = dentry->d_fsdata;
240
241                 dentry->d_flags &= ~DCACHE_NFSFS_RENAMED;
242                 dentry->d_fsdata = NULL;
243                 spin_unlock(&dentry->d_lock);
244                 nfs_free_unlinkdata(data);
245                 return;
246         }
247         spin_unlock(&dentry->d_lock);
248 }
249
250 /**
251  * nfs_async_rename_done - Sillyrename post-processing
252  * @task: rpc_task of the sillyrename
253  * @calldata: nfs_renamedata for the sillyrename
254  *
255  * Do the directory attribute updates and the d_move
256  */
257 static void nfs_async_rename_done(struct rpc_task *task, void *calldata)
258 {
259         struct nfs_renamedata *data = calldata;
260         struct inode *old_dir = data->old_dir;
261         struct inode *new_dir = data->new_dir;
262         struct dentry *old_dentry = data->old_dentry;
263
264         trace_nfs_sillyrename_rename(old_dir, old_dentry,
265                         new_dir, data->new_dentry, task->tk_status);
266         if (!NFS_PROTO(old_dir)->rename_done(task, old_dir, new_dir)) {
267                 rpc_restart_call_prepare(task);
268                 return;
269         }
270
271         if (data->complete)
272                 data->complete(task, data);
273 }
274
275 /**
276  * nfs_async_rename_release - Release the sillyrename data.
277  * @calldata: the struct nfs_renamedata to be released
278  */
279 static void nfs_async_rename_release(void *calldata)
280 {
281         struct nfs_renamedata   *data = calldata;
282         struct super_block *sb = data->old_dir->i_sb;
283
284         if (d_really_is_positive(data->old_dentry))
285                 nfs_mark_for_revalidate(d_inode(data->old_dentry));
286
287         /* The result of the rename is unknown. Play it safe by
288          * forcing a new lookup */
289         if (data->cancelled) {
290                 spin_lock(&data->old_dir->i_lock);
291                 nfs_force_lookup_revalidate(data->old_dir);
292                 spin_unlock(&data->old_dir->i_lock);
293                 if (data->new_dir != data->old_dir) {
294                         spin_lock(&data->new_dir->i_lock);
295                         nfs_force_lookup_revalidate(data->new_dir);
296                         spin_unlock(&data->new_dir->i_lock);
297                 }
298         }
299
300         dput(data->old_dentry);
301         dput(data->new_dentry);
302         iput(data->old_dir);
303         iput(data->new_dir);
304         nfs_sb_deactive(sb);
305         put_cred(data->cred);
306         kfree(data);
307 }
308
309 static void nfs_rename_prepare(struct rpc_task *task, void *calldata)
310 {
311         struct nfs_renamedata *data = calldata;
312         NFS_PROTO(data->old_dir)->rename_rpc_prepare(task, data);
313 }
314
315 static const struct rpc_call_ops nfs_rename_ops = {
316         .rpc_call_done = nfs_async_rename_done,
317         .rpc_release = nfs_async_rename_release,
318         .rpc_call_prepare = nfs_rename_prepare,
319 };
320
321 /**
322  * nfs_async_rename - perform an asynchronous rename operation
323  * @old_dir: directory that currently holds the dentry to be renamed
324  * @new_dir: target directory for the rename
325  * @old_dentry: original dentry to be renamed
326  * @new_dentry: dentry to which the old_dentry should be renamed
327  *
328  * It's expected that valid references to the dentries and inodes are held
329  */
330 struct rpc_task *
331 nfs_async_rename(struct inode *old_dir, struct inode *new_dir,
332                  struct dentry *old_dentry, struct dentry *new_dentry,
333                  void (*complete)(struct rpc_task *, struct nfs_renamedata *))
334 {
335         struct nfs_renamedata *data;
336         struct rpc_message msg = { };
337         struct rpc_task_setup task_setup_data = {
338                 .rpc_message = &msg,
339                 .callback_ops = &nfs_rename_ops,
340                 .workqueue = nfsiod_workqueue,
341                 .rpc_client = NFS_CLIENT(old_dir),
342                 .flags = RPC_TASK_ASYNC,
343         };
344
345         data = kzalloc(sizeof(*data), GFP_KERNEL);
346         if (data == NULL)
347                 return ERR_PTR(-ENOMEM);
348         task_setup_data.callback_data = data;
349
350         data->cred = get_current_cred();
351
352         msg.rpc_argp = &data->args;
353         msg.rpc_resp = &data->res;
354         msg.rpc_cred = data->cred;
355
356         /* set up nfs_renamedata */
357         data->old_dir = old_dir;
358         ihold(old_dir);
359         data->new_dir = new_dir;
360         ihold(new_dir);
361         data->old_dentry = dget(old_dentry);
362         data->new_dentry = dget(new_dentry);
363         nfs_fattr_init(&data->old_fattr);
364         nfs_fattr_init(&data->new_fattr);
365         data->complete = complete;
366
367         /* set up nfs_renameargs */
368         data->args.old_dir = NFS_FH(old_dir);
369         data->args.old_name = &old_dentry->d_name;
370         data->args.new_dir = NFS_FH(new_dir);
371         data->args.new_name = &new_dentry->d_name;
372
373         /* set up nfs_renameres */
374         data->res.old_fattr = &data->old_fattr;
375         data->res.new_fattr = &data->new_fattr;
376
377         nfs_sb_active(old_dir->i_sb);
378
379         NFS_PROTO(data->old_dir)->rename_setup(&msg, old_dentry, new_dentry);
380
381         return rpc_run_task(&task_setup_data);
382 }
383
384 /*
385  * Perform tasks needed when a sillyrename is done such as cancelling the
386  * queued async unlink if it failed.
387  */
388 static void
389 nfs_complete_sillyrename(struct rpc_task *task, struct nfs_renamedata *data)
390 {
391         struct dentry *dentry = data->old_dentry;
392
393         if (task->tk_status != 0) {
394                 nfs_cancel_async_unlink(dentry);
395                 return;
396         }
397
398         /*
399          * vfs_unlink and the like do not issue this when a file is
400          * sillyrenamed, so do it here.
401          */
402         fsnotify_nameremove(dentry, 0);
403 }
404
405 #define SILLYNAME_PREFIX ".nfs"
406 #define SILLYNAME_PREFIX_LEN ((unsigned)sizeof(SILLYNAME_PREFIX) - 1)
407 #define SILLYNAME_FILEID_LEN ((unsigned)sizeof(u64) << 1)
408 #define SILLYNAME_COUNTER_LEN ((unsigned)sizeof(unsigned int) << 1)
409 #define SILLYNAME_LEN (SILLYNAME_PREFIX_LEN + \
410                 SILLYNAME_FILEID_LEN + \
411                 SILLYNAME_COUNTER_LEN)
412
413 /**
414  * nfs_sillyrename - Perform a silly-rename of a dentry
415  * @dir: inode of directory that contains dentry
416  * @dentry: dentry to be sillyrenamed
417  *
418  * NFSv2/3 is stateless and the server doesn't know when the client is
419  * holding a file open. To prevent application problems when a file is
420  * unlinked while it's still open, the client performs a "silly-rename".
421  * That is, it renames the file to a hidden file in the same directory,
422  * and only performs the unlink once the last reference to it is put.
423  *
424  * The final cleanup is done during dentry_iput.
425  *
426  * (Note: NFSv4 is stateful, and has opens, so in theory an NFSv4 server
427  * could take responsibility for keeping open files referenced.  The server
428  * would also need to ensure that opened-but-deleted files were kept over
429  * reboots.  However, we may not assume a server does so.  (RFC 5661
430  * does provide an OPEN4_RESULT_PRESERVE_UNLINKED flag that a server can
431  * use to advertise that it does this; some day we may take advantage of
432  * it.))
433  */
434 int
435 nfs_sillyrename(struct inode *dir, struct dentry *dentry)
436 {
437         static unsigned int sillycounter;
438         unsigned char silly[SILLYNAME_LEN + 1];
439         unsigned long long fileid;
440         struct dentry *sdentry;
441         struct inode *inode = d_inode(dentry);
442         struct rpc_task *task;
443         int            error = -EBUSY;
444
445         dfprintk(VFS, "NFS: silly-rename(%pd2, ct=%d)\n",
446                 dentry, d_count(dentry));
447         nfs_inc_stats(dir, NFSIOS_SILLYRENAME);
448
449         /*
450          * We don't allow a dentry to be silly-renamed twice.
451          */
452         if (dentry->d_flags & DCACHE_NFSFS_RENAMED)
453                 goto out;
454
455         fileid = NFS_FILEID(d_inode(dentry));
456
457         sdentry = NULL;
458         do {
459                 int slen;
460                 dput(sdentry);
461                 sillycounter++;
462                 slen = scnprintf(silly, sizeof(silly),
463                                 SILLYNAME_PREFIX "%0*llx%0*x",
464                                 SILLYNAME_FILEID_LEN, fileid,
465                                 SILLYNAME_COUNTER_LEN, sillycounter);
466
467                 dfprintk(VFS, "NFS: trying to rename %pd to %s\n",
468                                 dentry, silly);
469
470                 sdentry = lookup_one_len(silly, dentry->d_parent, slen);
471                 /*
472                  * N.B. Better to return EBUSY here ... it could be
473                  * dangerous to delete the file while it's in use.
474                  */
475                 if (IS_ERR(sdentry))
476                         goto out;
477         } while (d_inode(sdentry) != NULL); /* need negative lookup */
478
479         ihold(inode);
480
481         /* queue unlink first. Can't do this from rpc_release as it
482          * has to allocate memory
483          */
484         error = nfs_async_unlink(dentry, &sdentry->d_name);
485         if (error)
486                 goto out_dput;
487
488         /* run the rename task, undo unlink if it fails */
489         task = nfs_async_rename(dir, dir, dentry, sdentry,
490                                         nfs_complete_sillyrename);
491         if (IS_ERR(task)) {
492                 error = -EBUSY;
493                 nfs_cancel_async_unlink(dentry);
494                 goto out_dput;
495         }
496
497         /* wait for the RPC task to complete, unless a SIGKILL intervenes */
498         error = rpc_wait_for_completion_task(task);
499         if (error == 0)
500                 error = task->tk_status;
501         switch (error) {
502         case 0:
503                 /* The rename succeeded */
504                 nfs_set_verifier(dentry, nfs_save_change_attribute(dir));
505                 spin_lock(&inode->i_lock);
506                 NFS_I(inode)->attr_gencount = nfs_inc_attr_generation_counter();
507                 NFS_I(inode)->cache_validity |= NFS_INO_INVALID_CHANGE
508                         | NFS_INO_INVALID_CTIME
509                         | NFS_INO_REVAL_FORCED;
510                 spin_unlock(&inode->i_lock);
511                 d_move(dentry, sdentry);
512                 break;
513         case -ERESTARTSYS:
514                 /* The result of the rename is unknown. Play it safe by
515                  * forcing a new lookup */
516                 d_drop(dentry);
517                 d_drop(sdentry);
518         }
519         rpc_put_task(task);
520 out_dput:
521         iput(inode);
522         dput(sdentry);
523 out:
524         return error;
525 }