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