fs/nfs/unlink.c

   1 /*
   2  *  linux/fs/nfs/unlink.c
   3  *
   4  * nfs sillydelete handling
   5  *
   6  */
   7
   8 #include <linux/slab.h>
   9 #include <linux/string.h>
  10 #include <linux/dcache.h>
  11 #include <linux/sunrpc/sched.h>
  12 #include <linux/sunrpc/clnt.h>
  13 #include <linux/nfs_fs.h>
  14 #include <linux/sched.h>
  15 #include <linux/swait.h>
  16 #include <linux/namei.h>
  17 #include <linux/fsnotify.h>
  18
  19 #include "internal.h"
  20 #include "nfs4_fs.h"
  21 #include "iostat.h"
  22 #include "delegation.h"
  23
  24 #include "nfstrace.h"
  25
  26 /**
  27  * nfs_free_unlinkdata - release data from a sillydelete operation.
  28  * @data: pointer to unlink structure.
  29  */
  30 static void
  31 nfs_free_unlinkdata(struct nfs_unlinkdata *data)
  32 {
  33         put_rpccred(data->cred);
  34         kfree(data->args.name.name);
  35         kfree(data);
  36 }
  37
  38 /**
  39  * nfs_async_unlink_done - Sillydelete post-processing
  40  * @task: rpc_task of the sillydelete
  41  *
  42  * Do the directory attribute update.
  43  */
  44 static void nfs_async_unlink_done(struct rpc_task *task, void *calldata)
  45 {
  46         struct nfs_unlinkdata *data = calldata;
  47         struct inode *dir = d_inode(data->dentry->d_parent);
  48
  49         trace_nfs_sillyrename_unlink(data, task->tk_status);
  50         if (!NFS_PROTO(dir)->unlink_done(task, dir))
  51                 rpc_restart_call_prepare(task);
  52 }
  53
  54 #ifdef CONFIG_PREEMPT_RT_BASE
  55 static void nfs_down_anon(struct semaphore *sema)
  56 {
  57         down(sema);
  58 }
  59
  60 static void nfs_up_anon(struct semaphore *sema)
  61 {
  62         up(sema);
  63 }
  64
  65 #else
  66 static void nfs_down_anon(struct rw_semaphore *rwsem)
  67 {
  68         down_read_non_owner(rwsem);
  69 }
  70
  71 static void nfs_up_anon(struct rw_semaphore *rwsem)
  72 {
  73         up_read_non_owner(rwsem);
  74 }
  75 #endif
  76
  77 /**
  78  * nfs_async_unlink_release - Release the sillydelete data.
  79  * @task: rpc_task of the sillydelete
  80  *
  81  * We need to call nfs_put_unlinkdata as a 'tk_release' task since the
  82  * rpc_task would be freed too.
  83  */
  84 static void nfs_async_unlink_release(void *calldata)
  85 {
  86         struct nfs_unlinkdata   *data = calldata;
  87         struct dentry *dentry = data->dentry;
  88         struct super_block *sb = dentry->d_sb;
  89
  90         nfs_up_anon(&NFS_I(d_inode(dentry->d_parent))->rmdir_sem);
  91         d_lookup_done(dentry);
  92         nfs_free_unlinkdata(data);
  93         dput(dentry);
  94         nfs_sb_deactive(sb);
  95 }
  96
  97 static void nfs_unlink_prepare(struct rpc_task *task, void *calldata)
  98 {
  99         struct nfs_unlinkdata *data = calldata;
 100         struct inode *dir = d_inode(data->dentry->d_parent);
 101         NFS_PROTO(dir)->unlink_rpc_prepare(task, data);
 102 }
 103
 104 static const struct rpc_call_ops nfs_unlink_ops = {
 105         .rpc_call_done = nfs_async_unlink_done,
 106         .rpc_release = nfs_async_unlink_release,
 107         .rpc_call_prepare = nfs_unlink_prepare,
 108 };
 109
 110 static void nfs_do_call_unlink(struct nfs_unlinkdata *data)
 111 {
 112         struct rpc_message msg = {
 113                 .rpc_argp = &data->args,
 114                 .rpc_resp = &data->res,
 115                 .rpc_cred = data->cred,
 116         };
 117         struct rpc_task_setup task_setup_data = {
 118                 .rpc_message = &msg,
 119                 .callback_ops = &nfs_unlink_ops,
 120                 .callback_data = data,
 121                 .workqueue = nfsiod_workqueue,
 122                 .flags = RPC_TASK_ASYNC,
 123         };
 124         struct rpc_task *task;
 125         struct inode *dir = d_inode(data->dentry->d_parent);
 126         nfs_sb_active(dir->i_sb);
 127         data->args.fh = NFS_FH(dir);
 128         nfs_fattr_init(data->res.dir_attr);
 129
 130         NFS_PROTO(dir)->unlink_setup(&msg, dir);
 131
 132         task_setup_data.rpc_client = NFS_CLIENT(dir);
 133         task = rpc_run_task(&task_setup_data);
 134         if (!IS_ERR(task))
 135                 rpc_put_task_async(task);
 136 }
 137
 138 static int nfs_call_unlink(struct dentry *dentry, struct nfs_unlinkdata *data)
 139 {
 140         struct inode *dir = d_inode(dentry->d_parent);
 141         struct dentry *alias;
 142
 143         nfs_down_anon(&NFS_I(dir)->rmdir_sem);
 144         alias = d_alloc_parallel(dentry->d_parent, &data->args.name, &data->wq);
 145         if (IS_ERR(alias)) {
 146                 nfs_up_anon(&NFS_I(dir)->rmdir_sem);
 147                 return 0;
 148         }
 149         if (!d_in_lookup(alias)) {
 150                 int ret;
 151                 void *devname_garbage = NULL;
 152
 153                 /*
 154                  * Hey, we raced with lookup... See if we need to transfer
 155                  * the sillyrename information to the aliased dentry.
 156                  */
 157                 spin_lock(&alias->d_lock);
 158                 if (d_really_is_positive(alias) &&
 159                     !(alias->d_flags & DCACHE_NFSFS_RENAMED)) {
 160                         devname_garbage = alias->d_fsdata;
 161                         alias->d_fsdata = data;
 162                         alias->d_flags |= DCACHE_NFSFS_RENAMED;
 163                         ret = 1;
 164                 } else
 165                         ret = 0;
 166                 spin_unlock(&alias->d_lock);
 167                 dput(alias);
 168                 nfs_up_anon(&NFS_I(dir)->rmdir_sem);
 169                 /*
 170                  * If we'd displaced old cached devname, free it.  At that
 171                  * point dentry is definitely not a root, so we won't need
 172                  * that anymore.
 173                  */
 174                 kfree(devname_garbage);
 175                 return ret;
 176         }
 177         data->dentry = alias;
 178         nfs_do_call_unlink(data);
 179         return 1;
 180 }
 181
 182 /**
 183  * nfs_async_unlink - asynchronous unlinking of a file
 184  * @dir: parent directory of dentry
 185  * @dentry: dentry to unlink
 186  */
 187 static int
 188 nfs_async_unlink(struct dentry *dentry, const struct qstr *name)
 189 {
 190         struct nfs_unlinkdata *data;
 191         int status = -ENOMEM;
 192         void *devname_garbage = NULL;
 193
 194         data = kzalloc(sizeof(*data), GFP_KERNEL);
 195         if (data == NULL)
 196                 goto out;
 197         data->args.name.name = kstrdup(name->name, GFP_KERNEL);
 198         if (!data->args.name.name)
 199                 goto out_free;
 200         data->args.name.len = name->len;
 201
 202         data->cred = rpc_lookup_cred();
 203         if (IS_ERR(data->cred)) {
 204                 status = PTR_ERR(data->cred);
 205                 goto out_free_name;
 206         }
 207         data->res.dir_attr = &data->dir_attr;
 208         init_swait_queue_head(&data->wq);
 209
 210         status = -EBUSY;
 211         spin_lock(&dentry->d_lock);
 212         if (dentry->d_flags & DCACHE_NFSFS_RENAMED)
 213                 goto out_unlock;
 214         dentry->d_flags |= DCACHE_NFSFS_RENAMED;
 215         devname_garbage = dentry->d_fsdata;
 216         dentry->d_fsdata = data;
 217         spin_unlock(&dentry->d_lock);
 218         /*
 219          * If we'd displaced old cached devname, free it.  At that
 220          * point dentry is definitely not a root, so we won't need
 221          * that anymore.
 222          */
 223         kfree(devname_garbage);
 224         return 0;
 225 out_unlock:
 226         spin_unlock(&dentry->d_lock);
 227         put_rpccred(data->cred);
 228 out_free_name:
 229         kfree(data->args.name.name);
 230 out_free:
 231         kfree(data);
 232 out:
 233         return status;
 234 }
 235
 236 /**
 237  * nfs_complete_unlink - Initialize completion of the sillydelete
 238  * @dentry: dentry to delete
 239  * @inode: inode
 240  *
 241  * Since we're most likely to be called by dentry_iput(), we
 242  * only use the dentry to find the sillydelete. We then copy the name
 243  * into the qstr.
 244  */
 245 void
 246 nfs_complete_unlink(struct dentry *dentry, struct inode *inode)
 247 {
 248         struct nfs_unlinkdata   *data;
 249
 250         spin_lock(&dentry->d_lock);
 251         dentry->d_flags &= ~DCACHE_NFSFS_RENAMED;
 252         data = dentry->d_fsdata;
 253         dentry->d_fsdata = NULL;
 254         spin_unlock(&dentry->d_lock);
 255
 256         if (NFS_STALE(inode) || !nfs_call_unlink(dentry, data))
 257                 nfs_free_unlinkdata(data);
 258 }
 259
 260 /* Cancel a queued async unlink. Called when a sillyrename run fails. */
 261 static void
 262 nfs_cancel_async_unlink(struct dentry *dentry)
 263 {
 264         spin_lock(&dentry->d_lock);
 265         if (dentry->d_flags & DCACHE_NFSFS_RENAMED) {
 266                 struct nfs_unlinkdata *data = dentry->d_fsdata;
 267
 268                 dentry->d_flags &= ~DCACHE_NFSFS_RENAMED;
 269                 dentry->d_fsdata = NULL;
 270                 spin_unlock(&dentry->d_lock);
 271                 nfs_free_unlinkdata(data);
 272                 return;
 273         }
 274         spin_unlock(&dentry->d_lock);
 275 }
 276
 277 /**
 278  * nfs_async_rename_done - Sillyrename post-processing
 279  * @task: rpc_task of the sillyrename
 280  * @calldata: nfs_renamedata for the sillyrename
 281  *
 282  * Do the directory attribute updates and the d_move
 283  */
 284 static void nfs_async_rename_done(struct rpc_task *task, void *calldata)
 285 {
 286         struct nfs_renamedata *data = calldata;
 287         struct inode *old_dir = data->old_dir;
 288         struct inode *new_dir = data->new_dir;
 289         struct dentry *old_dentry = data->old_dentry;
 290
 291         trace_nfs_sillyrename_rename(old_dir, old_dentry,
 292                         new_dir, data->new_dentry, task->tk_status);
 293         if (!NFS_PROTO(old_dir)->rename_done(task, old_dir, new_dir)) {
 294                 rpc_restart_call_prepare(task);
 295                 return;
 296         }
 297
 298         if (data->complete)
 299                 data->complete(task, data);
 300 }
 301
 302 /**
 303  * nfs_async_rename_release - Release the sillyrename data.
 304  * @calldata: the struct nfs_renamedata to be released
 305  */
 306 static void nfs_async_rename_release(void *calldata)
 307 {
 308         struct nfs_renamedata   *data = calldata;
 309         struct super_block *sb = data->old_dir->i_sb;
 310
 311         if (d_really_is_positive(data->old_dentry))
 312                 nfs_mark_for_revalidate(d_inode(data->old_dentry));
 313
 314         dput(data->old_dentry);
 315         dput(data->new_dentry);
 316         iput(data->old_dir);
 317         iput(data->new_dir);
 318         nfs_sb_deactive(sb);
 319         put_rpccred(data->cred);
 320         kfree(data);
 321 }
 322
 323 static void nfs_rename_prepare(struct rpc_task *task, void *calldata)
 324 {
 325         struct nfs_renamedata *data = calldata;
 326         NFS_PROTO(data->old_dir)->rename_rpc_prepare(task, data);
 327 }
 328
 329 static const struct rpc_call_ops nfs_rename_ops = {
 330         .rpc_call_done = nfs_async_rename_done,
 331         .rpc_release = nfs_async_rename_release,
 332         .rpc_call_prepare = nfs_rename_prepare,
 333 };
 334
 335 /**
 336  * nfs_async_rename - perform an asynchronous rename operation
 337  * @old_dir: directory that currently holds the dentry to be renamed
 338  * @new_dir: target directory for the rename
 339  * @old_dentry: original dentry to be renamed
 340  * @new_dentry: dentry to which the old_dentry should be renamed
 341  *
 342  * It's expected that valid references to the dentries and inodes are held
 343  */
 344 struct rpc_task *
 345 nfs_async_rename(struct inode *old_dir, struct inode *new_dir,
 346                  struct dentry *old_dentry, struct dentry *new_dentry,
 347                  void (*complete)(struct rpc_task *, struct nfs_renamedata *))
 348 {
 349         struct nfs_renamedata *data;
 350         struct rpc_message msg = { };
 351         struct rpc_task_setup task_setup_data = {
 352                 .rpc_message = &msg,
 353                 .callback_ops = &nfs_rename_ops,
 354                 .workqueue = nfsiod_workqueue,
 355                 .rpc_client = NFS_CLIENT(old_dir),
 356                 .flags = RPC_TASK_ASYNC,
 357         };
 358
 359         data = kzalloc(sizeof(*data), GFP_KERNEL);
 360         if (data == NULL)
 361                 return ERR_PTR(-ENOMEM);
 362         task_setup_data.callback_data = data;
 363
 364         data->cred = rpc_lookup_cred();
 365         if (IS_ERR(data->cred)) {
 366                 struct rpc_task *task = ERR_CAST(data->cred);
 367                 kfree(data);
 368                 return task;
 369         }
 370
 371         msg.rpc_argp = &data->args;
 372         msg.rpc_resp = &data->res;
 373         msg.rpc_cred = data->cred;
 374
 375         /* set up nfs_renamedata */
 376         data->old_dir = old_dir;
 377         ihold(old_dir);
 378         data->new_dir = new_dir;
 379         ihold(new_dir);
 380         data->old_dentry = dget(old_dentry);
 381         data->new_dentry = dget(new_dentry);
 382         nfs_fattr_init(&data->old_fattr);
 383         nfs_fattr_init(&data->new_fattr);
 384         data->complete = complete;
 385
 386         /* set up nfs_renameargs */
 387         data->args.old_dir = NFS_FH(old_dir);
 388         data->args.old_name = &old_dentry->d_name;
 389         data->args.new_dir = NFS_FH(new_dir);
 390         data->args.new_name = &new_dentry->d_name;
 391
 392         /* set up nfs_renameres */
 393         data->res.old_fattr = &data->old_fattr;
 394         data->res.new_fattr = &data->new_fattr;
 395
 396         nfs_sb_active(old_dir->i_sb);
 397
 398         NFS_PROTO(data->old_dir)->rename_setup(&msg, old_dir);
 399
 400         return rpc_run_task(&task_setup_data);
 401 }
 402
 403 /*
 404  * Perform tasks needed when a sillyrename is done such as cancelling the
 405  * queued async unlink if it failed.
 406  */
 407 static void
 408 nfs_complete_sillyrename(struct rpc_task *task, struct nfs_renamedata *data)
 409 {
 410         struct dentry *dentry = data->old_dentry;
 411
 412         if (task->tk_status != 0) {
 413                 nfs_cancel_async_unlink(dentry);
 414                 return;
 415         }
 416
 417         /*
 418          * vfs_unlink and the like do not issue this when a file is
 419          * sillyrenamed, so do it here.
 420          */
 421         fsnotify_nameremove(dentry, 0);
 422 }
 423
 424 #define SILLYNAME_PREFIX ".nfs"
 425 #define SILLYNAME_PREFIX_LEN ((unsigned)sizeof(SILLYNAME_PREFIX) - 1)
 426 #define SILLYNAME_FILEID_LEN ((unsigned)sizeof(u64) << 1)
 427 #define SILLYNAME_COUNTER_LEN ((unsigned)sizeof(unsigned int) << 1)
 428 #define SILLYNAME_LEN (SILLYNAME_PREFIX_LEN + \
 429                 SILLYNAME_FILEID_LEN + \
 430                 SILLYNAME_COUNTER_LEN)
 431
 432 /**
 433  * nfs_sillyrename - Perform a silly-rename of a dentry
 434  * @dir: inode of directory that contains dentry
 435  * @dentry: dentry to be sillyrenamed
 436  *
 437  * NFSv2/3 is stateless and the server doesn't know when the client is
 438  * holding a file open. To prevent application problems when a file is
 439  * unlinked while it's still open, the client performs a "silly-rename".
 440  * That is, it renames the file to a hidden file in the same directory,
 441  * and only performs the unlink once the last reference to it is put.
 442  *
 443  * The final cleanup is done during dentry_iput.
 444  *
 445  * (Note: NFSv4 is stateful, and has opens, so in theory an NFSv4 server
 446  * could take responsibility for keeping open files referenced.  The server
 447  * would also need to ensure that opened-but-deleted files were kept over
 448  * reboots.  However, we may not assume a server does so.  (RFC 5661
 449  * does provide an OPEN4_RESULT_PRESERVE_UNLINKED flag that a server can
 450  * use to advertise that it does this; some day we may take advantage of
 451  * it.))
 452  */
 453 int
 454 nfs_sillyrename(struct inode *dir, struct dentry *dentry)
 455 {
 456         static unsigned int sillycounter;
 457         unsigned char silly[SILLYNAME_LEN + 1];
 458         unsigned long long fileid;
 459         struct dentry *sdentry;
 460         struct rpc_task *task;
 461         int            error = -EBUSY;
 462
 463         dfprintk(VFS, "NFS: silly-rename(%pd2, ct=%d)\n",
 464                 dentry, d_count(dentry));
 465         nfs_inc_stats(dir, NFSIOS_SILLYRENAME);
 466
 467         /*
 468          * We don't allow a dentry to be silly-renamed twice.
 469          */
 470         if (dentry->d_flags & DCACHE_NFSFS_RENAMED)
 471                 goto out;
 472
 473         fileid = NFS_FILEID(d_inode(dentry));
 474
 475         /* Return delegation in anticipation of the rename */
 476         NFS_PROTO(d_inode(dentry))->return_delegation(d_inode(dentry));
 477
 478         sdentry = NULL;
 479         do {
 480                 int slen;
 481                 dput(sdentry);
 482                 sillycounter++;
 483                 slen = scnprintf(silly, sizeof(silly),
 484                                 SILLYNAME_PREFIX "%0*llx%0*x",
 485                                 SILLYNAME_FILEID_LEN, fileid,
 486                                 SILLYNAME_COUNTER_LEN, sillycounter);
 487
 488                 dfprintk(VFS, "NFS: trying to rename %pd to %s\n",
 489                                 dentry, silly);
 490
 491                 sdentry = lookup_one_len(silly, dentry->d_parent, slen);
 492                 /*
 493                  * N.B. Better to return EBUSY here ... it could be
 494                  * dangerous to delete the file while it's in use.
 495                  */
 496                 if (IS_ERR(sdentry))
 497                         goto out;
 498         } while (d_inode(sdentry) != NULL); /* need negative lookup */
 499
 500         /* queue unlink first. Can't do this from rpc_release as it
 501          * has to allocate memory
 502          */
 503         error = nfs_async_unlink(dentry, &sdentry->d_name);
 504         if (error)
 505                 goto out_dput;
 506
 507         /* run the rename task, undo unlink if it fails */
 508         task = nfs_async_rename(dir, dir, dentry, sdentry,
 509                                         nfs_complete_sillyrename);
 510         if (IS_ERR(task)) {
 511                 error = -EBUSY;
 512                 nfs_cancel_async_unlink(dentry);
 513                 goto out_dput;
 514         }
 515
 516         /* wait for the RPC task to complete, unless a SIGKILL intervenes */
 517         error = rpc_wait_for_completion_task(task);
 518         if (error == 0)
 519                 error = task->tk_status;
 520         switch (error) {
 521         case 0:
 522                 /* The rename succeeded */
 523                 nfs_set_verifier(dentry, nfs_save_change_attribute(dir));
 524                 d_move(dentry, sdentry);
 525                 break;
 526         case -ERESTARTSYS:
 527                 /* The result of the rename is unknown. Play it safe by
 528                  * forcing a new lookup */
 529                 d_drop(dentry);
 530                 d_drop(sdentry);
 531         }
 532         rpc_put_task(task);
 533 out_dput:
 534         dput(sdentry);
 535 out:
 536         return error;
 537 }