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
  18 #include "internal.h"
  19 #include "nfs4_fs.h"
  20 #include "iostat.h"
  21 #include "delegation.h"
  22
  23 struct nfs_unlinkdata {
  24         struct hlist_node list;
  25         struct nfs_removeargs args;
  26         struct nfs_removeres res;
  27         struct inode *dir;
  28         struct rpc_cred *cred;
  29         struct nfs_fattr dir_attr;
  30 };
  31
  32 /**
  33  * nfs_free_unlinkdata - release data from a sillydelete operation.
  34  * @data: pointer to unlink structure.
  35  */
  36 static void
  37 nfs_free_unlinkdata(struct nfs_unlinkdata *data)
  38 {
  39         iput(data->dir);
  40         put_rpccred(data->cred);
  41         kfree(data->args.name.name);
  42         kfree(data);
  43 }
  44
  45 #define NAME_ALLOC_LEN(len)     ((len+16) & ~15)
  46 /**
  47  * nfs_copy_dname - copy dentry name to data structure
  48  * @dentry: pointer to dentry
  49  * @data: nfs_unlinkdata
  50  */
  51 static int nfs_copy_dname(struct dentry *dentry, struct nfs_unlinkdata *data)
  52 {
  53         char            *str;
  54         int             len = dentry->d_name.len;
  55
  56         str = kmemdup(dentry->d_name.name, NAME_ALLOC_LEN(len), GFP_KERNEL);
  57         if (!str)
  58                 return -ENOMEM;
  59         data->args.name.len = len;
  60         data->args.name.name = str;
  61         return 0;
  62 }
  63
  64 static void nfs_free_dname(struct nfs_unlinkdata *data)
  65 {
  66         kfree(data->args.name.name);
  67         data->args.name.name = NULL;
  68         data->args.name.len = 0;
  69 }
  70
  71 static void nfs_dec_sillycount(struct inode *dir)
  72 {
  73         struct nfs_inode *nfsi = NFS_I(dir);
  74         if (atomic_dec_return(&nfsi->silly_count) == 1)
  75                 wake_up(&nfsi->waitqueue);
  76 }
  77
  78 /**
  79  * nfs_async_unlink_done - Sillydelete post-processing
  80  * @task: rpc_task of the sillydelete
  81  *
  82  * Do the directory attribute update.
  83  */
  84 static void nfs_async_unlink_done(struct rpc_task *task, void *calldata)
  85 {
  86         struct nfs_unlinkdata *data = calldata;
  87         struct inode *dir = data->dir;
  88
  89         if (!NFS_PROTO(dir)->unlink_done(task, dir))
  90                 rpc_restart_call_prepare(task);
  91 }
  92
  93 /**
  94  * nfs_async_unlink_release - Release the sillydelete data.
  95  * @task: rpc_task of the sillydelete
  96  *
  97  * We need to call nfs_put_unlinkdata as a 'tk_release' task since the
  98  * rpc_task would be freed too.
  99  */
 100 static void nfs_async_unlink_release(void *calldata)
 101 {
 102         struct nfs_unlinkdata   *data = calldata;
 103         struct super_block *sb = data->dir->i_sb;
 104
 105         nfs_dec_sillycount(data->dir);
 106         nfs_free_unlinkdata(data);
 107         nfs_sb_deactive(sb);
 108 }
 109
 110 #if defined(CONFIG_NFS_V4_1)
 111 void nfs_unlink_prepare(struct rpc_task *task, void *calldata)
 112 {
 113         struct nfs_unlinkdata *data = calldata;
 114         struct nfs_server *server = NFS_SERVER(data->dir);
 115
 116         if (nfs4_setup_sequence(server, &data->args.seq_args,
 117                                 &data->res.seq_res, 1, task))
 118                 return;
 119         rpc_call_start(task);
 120 }
 121 #endif /* CONFIG_NFS_V4_1 */
 122
 123 static const struct rpc_call_ops nfs_unlink_ops = {
 124         .rpc_call_done = nfs_async_unlink_done,
 125         .rpc_release = nfs_async_unlink_release,
 126 #if defined(CONFIG_NFS_V4_1)
 127         .rpc_call_prepare = nfs_unlink_prepare,
 128 #endif /* CONFIG_NFS_V4_1 */
 129 };
 130
 131 static int nfs_do_call_unlink(struct dentry *parent, struct inode *dir, struct nfs_unlinkdata *data)
 132 {
 133         struct rpc_message msg = {
 134                 .rpc_argp = &data->args,
 135                 .rpc_resp = &data->res,
 136                 .rpc_cred = data->cred,
 137         };
 138         struct rpc_task_setup task_setup_data = {
 139                 .rpc_message = &msg,
 140                 .callback_ops = &nfs_unlink_ops,
 141                 .callback_data = data,
 142                 .workqueue = nfsiod_workqueue,
 143                 .flags = RPC_TASK_ASYNC,
 144         };
 145         struct rpc_task *task;
 146         struct dentry *alias;
 147
 148         alias = d_lookup(parent, &data->args.name);
 149         if (alias != NULL) {
 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                 nfs_free_dname(data);
 158                 ret = nfs_copy_dname(alias, data);
 159                 spin_lock(&alias->d_lock);
 160                 if (ret == 0 && alias->d_inode != NULL &&
 161                     !(alias->d_flags & DCACHE_NFSFS_RENAMED)) {
 162                         devname_garbage = alias->d_fsdata;
 163                         alias->d_fsdata = data;
 164                         alias->d_flags |= DCACHE_NFSFS_RENAMED;
 165                         ret = 1;
 166                 } else
 167                         ret = 0;
 168                 spin_unlock(&alias->d_lock);
 169                 nfs_dec_sillycount(dir);
 170                 dput(alias);
 171                 /*
 172                  * If we'd displaced old cached devname, free it.  At that
 173                  * point dentry is definitely not a root, so we won't need
 174                  * that anymore.
 175                  */
 176                 kfree(devname_garbage);
 177                 return ret;
 178         }
 179         data->dir = igrab(dir);
 180         if (!data->dir) {
 181                 nfs_dec_sillycount(dir);
 182                 return 0;
 183         }
 184         nfs_sb_active(dir->i_sb);
 185         data->args.fh = NFS_FH(dir);
 186         nfs_fattr_init(data->res.dir_attr);
 187
 188         NFS_PROTO(dir)->unlink_setup(&msg, dir);
 189
 190         task_setup_data.rpc_client = NFS_CLIENT(dir);
 191         task = rpc_run_task(&task_setup_data);
 192         if (!IS_ERR(task))
 193                 rpc_put_task_async(task);
 194         return 1;
 195 }
 196
 197 static int nfs_call_unlink(struct dentry *dentry, struct nfs_unlinkdata *data)
 198 {
 199         struct dentry *parent;
 200         struct inode *dir;
 201         int ret = 0;
 202
 203
 204         parent = dget_parent(dentry);
 205         if (parent == NULL)
 206                 goto out_free;
 207         dir = parent->d_inode;
 208         /* Non-exclusive lock protects against concurrent lookup() calls */
 209         spin_lock(&dir->i_lock);
 210         if (atomic_inc_not_zero(&NFS_I(dir)->silly_count) == 0) {
 211                 /* Deferred delete */
 212                 hlist_add_head(&data->list, &NFS_I(dir)->silly_list);
 213                 spin_unlock(&dir->i_lock);
 214                 ret = 1;
 215                 goto out_dput;
 216         }
 217         spin_unlock(&dir->i_lock);
 218         ret = nfs_do_call_unlink(parent, dir, data);
 219 out_dput:
 220         dput(parent);
 221 out_free:
 222         return ret;
 223 }
 224
 225 void nfs_block_sillyrename(struct dentry *dentry)
 226 {
 227         struct nfs_inode *nfsi = NFS_I(dentry->d_inode);
 228
 229         wait_event(nfsi->waitqueue, atomic_cmpxchg(&nfsi->silly_count, 1, 0) == 1);
 230 }
 231
 232 void nfs_unblock_sillyrename(struct dentry *dentry)
 233 {
 234         struct inode *dir = dentry->d_inode;
 235         struct nfs_inode *nfsi = NFS_I(dir);
 236         struct nfs_unlinkdata *data;
 237
 238         atomic_inc(&nfsi->silly_count);
 239         spin_lock(&dir->i_lock);
 240         while (!hlist_empty(&nfsi->silly_list)) {
 241                 if (!atomic_inc_not_zero(&nfsi->silly_count))
 242                         break;
 243                 data = hlist_entry(nfsi->silly_list.first, struct nfs_unlinkdata, list);
 244                 hlist_del(&data->list);
 245                 spin_unlock(&dir->i_lock);
 246                 if (nfs_do_call_unlink(dentry, dir, data) == 0)
 247                         nfs_free_unlinkdata(data);
 248                 spin_lock(&dir->i_lock);
 249         }
 250         spin_unlock(&dir->i_lock);
 251 }
 252
 253 /**
 254  * nfs_async_unlink - asynchronous unlinking of a file
 255  * @dir: parent directory of dentry
 256  * @dentry: dentry to unlink
 257  */
 258 static int
 259 nfs_async_unlink(struct inode *dir, struct dentry *dentry)
 260 {
 261         struct nfs_unlinkdata *data;
 262         int status = -ENOMEM;
 263         void *devname_garbage = NULL;
 264
 265         data = kzalloc(sizeof(*data), GFP_KERNEL);
 266         if (data == NULL)
 267                 goto out;
 268
 269         data->cred = rpc_lookup_cred();
 270         if (IS_ERR(data->cred)) {
 271                 status = PTR_ERR(data->cred);
 272                 goto out_free;
 273         }
 274         data->res.dir_attr = &data->dir_attr;
 275
 276         status = -EBUSY;
 277         spin_lock(&dentry->d_lock);
 278         if (dentry->d_flags & DCACHE_NFSFS_RENAMED)
 279                 goto out_unlock;
 280         dentry->d_flags |= DCACHE_NFSFS_RENAMED;
 281         devname_garbage = dentry->d_fsdata;
 282         dentry->d_fsdata = data;
 283         spin_unlock(&dentry->d_lock);
 284         /*
 285          * If we'd displaced old cached devname, free it.  At that
 286          * point dentry is definitely not a root, so we won't need
 287          * that anymore.
 288          */
 289         if (devname_garbage)
 290                 kfree(devname_garbage);
 291         return 0;
 292 out_unlock:
 293         spin_unlock(&dentry->d_lock);
 294         put_rpccred(data->cred);
 295 out_free:
 296         kfree(data);
 297 out:
 298         return status;
 299 }
 300
 301 /**
 302  * nfs_complete_unlink - Initialize completion of the sillydelete
 303  * @dentry: dentry to delete
 304  * @inode: inode
 305  *
 306  * Since we're most likely to be called by dentry_iput(), we
 307  * only use the dentry to find the sillydelete. We then copy the name
 308  * into the qstr.
 309  */
 310 void
 311 nfs_complete_unlink(struct dentry *dentry, struct inode *inode)
 312 {
 313         struct nfs_unlinkdata   *data = NULL;
 314
 315         spin_lock(&dentry->d_lock);
 316         if (dentry->d_flags & DCACHE_NFSFS_RENAMED) {
 317                 dentry->d_flags &= ~DCACHE_NFSFS_RENAMED;
 318                 data = dentry->d_fsdata;
 319                 dentry->d_fsdata = NULL;
 320         }
 321         spin_unlock(&dentry->d_lock);
 322
 323         if (data != NULL && (NFS_STALE(inode) || !nfs_call_unlink(dentry, data)))
 324                 nfs_free_unlinkdata(data);
 325 }
 326
 327 /* Cancel a queued async unlink. Called when a sillyrename run fails. */
 328 static void
 329 nfs_cancel_async_unlink(struct dentry *dentry)
 330 {
 331         spin_lock(&dentry->d_lock);
 332         if (dentry->d_flags & DCACHE_NFSFS_RENAMED) {
 333                 struct nfs_unlinkdata *data = dentry->d_fsdata;
 334
 335                 dentry->d_flags &= ~DCACHE_NFSFS_RENAMED;
 336                 dentry->d_fsdata = NULL;
 337                 spin_unlock(&dentry->d_lock);
 338                 nfs_free_unlinkdata(data);
 339                 return;
 340         }
 341         spin_unlock(&dentry->d_lock);
 342 }
 343
 344 struct nfs_renamedata {
 345         struct nfs_renameargs   args;
 346         struct nfs_renameres    res;
 347         struct rpc_cred         *cred;
 348         struct inode            *old_dir;
 349         struct dentry           *old_dentry;
 350         struct nfs_fattr        old_fattr;
 351         struct inode            *new_dir;
 352         struct dentry           *new_dentry;
 353         struct nfs_fattr        new_fattr;
 354 };
 355
 356 /**
 357  * nfs_async_rename_done - Sillyrename post-processing
 358  * @task: rpc_task of the sillyrename
 359  * @calldata: nfs_renamedata for the sillyrename
 360  *
 361  * Do the directory attribute updates and the d_move
 362  */
 363 static void nfs_async_rename_done(struct rpc_task *task, void *calldata)
 364 {
 365         struct nfs_renamedata *data = calldata;
 366         struct inode *old_dir = data->old_dir;
 367         struct inode *new_dir = data->new_dir;
 368         struct dentry *old_dentry = data->old_dentry;
 369         struct dentry *new_dentry = data->new_dentry;
 370
 371         if (!NFS_PROTO(old_dir)->rename_done(task, old_dir, new_dir)) {
 372                 rpc_restart_call_prepare(task);
 373                 return;
 374         }
 375
 376         if (task->tk_status != 0) {
 377                 nfs_cancel_async_unlink(old_dentry);
 378                 return;
 379         }
 380
 381         d_drop(old_dentry);
 382         d_drop(new_dentry);
 383 }
 384
 385 /**
 386  * nfs_async_rename_release - Release the sillyrename data.
 387  * @calldata: the struct nfs_renamedata to be released
 388  */
 389 static void nfs_async_rename_release(void *calldata)
 390 {
 391         struct nfs_renamedata   *data = calldata;
 392         struct super_block *sb = data->old_dir->i_sb;
 393
 394         if (data->old_dentry->d_inode)
 395                 nfs_mark_for_revalidate(data->old_dentry->d_inode);
 396
 397         dput(data->old_dentry);
 398         dput(data->new_dentry);
 399         iput(data->old_dir);
 400         iput(data->new_dir);
 401         nfs_sb_deactive(sb);
 402         put_rpccred(data->cred);
 403         kfree(data);
 404 }
 405
 406 #if defined(CONFIG_NFS_V4_1)
 407 static void nfs_rename_prepare(struct rpc_task *task, void *calldata)
 408 {
 409         struct nfs_renamedata *data = calldata;
 410         struct nfs_server *server = NFS_SERVER(data->old_dir);
 411
 412         if (nfs4_setup_sequence(server, &data->args.seq_args,
 413                                 &data->res.seq_res, 1, task))
 414                 return;
 415         rpc_call_start(task);
 416 }
 417 #endif /* CONFIG_NFS_V4_1 */
 418
 419 static const struct rpc_call_ops nfs_rename_ops = {
 420         .rpc_call_done = nfs_async_rename_done,
 421         .rpc_release = nfs_async_rename_release,
 422 #if defined(CONFIG_NFS_V4_1)
 423         .rpc_call_prepare = nfs_rename_prepare,
 424 #endif /* CONFIG_NFS_V4_1 */
 425 };
 426
 427 /**
 428  * nfs_async_rename - perform an asynchronous rename operation
 429  * @old_dir: directory that currently holds the dentry to be renamed
 430  * @new_dir: target directory for the rename
 431  * @old_dentry: original dentry to be renamed
 432  * @new_dentry: dentry to which the old_dentry should be renamed
 433  *
 434  * It's expected that valid references to the dentries and inodes are held
 435  */
 436 static struct rpc_task *
 437 nfs_async_rename(struct inode *old_dir, struct inode *new_dir,
 438                  struct dentry *old_dentry, struct dentry *new_dentry)
 439 {
 440         struct nfs_renamedata *data;
 441         struct rpc_message msg = { };
 442         struct rpc_task_setup task_setup_data = {
 443                 .rpc_message = &msg,
 444                 .callback_ops = &nfs_rename_ops,
 445                 .workqueue = nfsiod_workqueue,
 446                 .rpc_client = NFS_CLIENT(old_dir),
 447                 .flags = RPC_TASK_ASYNC,
 448         };
 449
 450         data = kzalloc(sizeof(*data), GFP_KERNEL);
 451         if (data == NULL)
 452                 return ERR_PTR(-ENOMEM);
 453         task_setup_data.callback_data = data;
 454
 455         data->cred = rpc_lookup_cred();
 456         if (IS_ERR(data->cred)) {
 457                 struct rpc_task *task = ERR_CAST(data->cred);
 458                 kfree(data);
 459                 return task;
 460         }
 461
 462         msg.rpc_argp = &data->args;
 463         msg.rpc_resp = &data->res;
 464         msg.rpc_cred = data->cred;
 465
 466         /* set up nfs_renamedata */
 467         data->old_dir = old_dir;
 468         ihold(old_dir);
 469         data->new_dir = new_dir;
 470         ihold(new_dir);
 471         data->old_dentry = dget(old_dentry);
 472         data->new_dentry = dget(new_dentry);
 473         nfs_fattr_init(&data->old_fattr);
 474         nfs_fattr_init(&data->new_fattr);
 475
 476         /* set up nfs_renameargs */
 477         data->args.old_dir = NFS_FH(old_dir);
 478         data->args.old_name = &old_dentry->d_name;
 479         data->args.new_dir = NFS_FH(new_dir);
 480         data->args.new_name = &new_dentry->d_name;
 481
 482         /* set up nfs_renameres */
 483         data->res.old_fattr = &data->old_fattr;
 484         data->res.new_fattr = &data->new_fattr;
 485
 486         nfs_sb_active(old_dir->i_sb);
 487
 488         NFS_PROTO(data->old_dir)->rename_setup(&msg, old_dir);
 489
 490         return rpc_run_task(&task_setup_data);
 491 }
 492
 493 /**
 494  * nfs_sillyrename - Perform a silly-rename of a dentry
 495  * @dir: inode of directory that contains dentry
 496  * @dentry: dentry to be sillyrenamed
 497  *
 498  * NFSv2/3 is stateless and the server doesn't know when the client is
 499  * holding a file open. To prevent application problems when a file is
 500  * unlinked while it's still open, the client performs a "silly-rename".
 501  * That is, it renames the file to a hidden file in the same directory,
 502  * and only performs the unlink once the last reference to it is put.
 503  *
 504  * The final cleanup is done during dentry_iput.
 505  *
 506  * (Note: NFSv4 is stateful, and has opens, so in theory an NFSv4 server
 507  * could take responsibility for keeping open files referenced.  The server
 508  * would also need to ensure that opened-but-deleted files were kept over
 509  * reboots.  However, we may not assume a server does so.  (RFC 5661
 510  * does provide an OPEN4_RESULT_PRESERVE_UNLINKED flag that a server can
 511  * use to advertise that it does this; some day we may take advantage of
 512  * it.))
 513  */
 514 int
 515 nfs_sillyrename(struct inode *dir, struct dentry *dentry)
 516 {
 517         static unsigned int sillycounter;
 518         const int      fileidsize  = sizeof(NFS_FILEID(dentry->d_inode))*2;
 519         const int      countersize = sizeof(sillycounter)*2;
 520         const int      slen        = sizeof(".nfs")+fileidsize+countersize-1;
 521         char           silly[slen+1];
 522         struct dentry *sdentry;
 523         struct rpc_task *task;
 524         int            error = -EIO;
 525
 526         dfprintk(VFS, "NFS: silly-rename(%s/%s, ct=%d)\n",
 527                 dentry->d_parent->d_name.name, dentry->d_name.name,
 528                 dentry->d_count);
 529         nfs_inc_stats(dir, NFSIOS_SILLYRENAME);
 530
 531         /*
 532          * We don't allow a dentry to be silly-renamed twice.
 533          */
 534         error = -EBUSY;
 535         if (dentry->d_flags & DCACHE_NFSFS_RENAMED)
 536                 goto out;
 537
 538         sprintf(silly, ".nfs%*.*Lx",
 539                 fileidsize, fileidsize,
 540                 (unsigned long long)NFS_FILEID(dentry->d_inode));
 541
 542         /* Return delegation in anticipation of the rename */
 543         nfs_inode_return_delegation(dentry->d_inode);
 544
 545         sdentry = NULL;
 546         do {
 547                 char *suffix = silly + slen - countersize;
 548
 549                 dput(sdentry);
 550                 sillycounter++;
 551                 sprintf(suffix, "%*.*x", countersize, countersize, sillycounter);
 552
 553                 dfprintk(VFS, "NFS: trying to rename %s to %s\n",
 554                                 dentry->d_name.name, silly);
 555
 556                 sdentry = lookup_one_len(silly, dentry->d_parent, slen);
 557                 /*
 558                  * N.B. Better to return EBUSY here ... it could be
 559                  * dangerous to delete the file while it's in use.
 560                  */
 561                 if (IS_ERR(sdentry))
 562                         goto out;
 563         } while (sdentry->d_inode != NULL); /* need negative lookup */
 564
 565         /* queue unlink first. Can't do this from rpc_release as it
 566          * has to allocate memory
 567          */
 568         error = nfs_async_unlink(dir, dentry);
 569         if (error)
 570                 goto out_dput;
 571
 572         /* populate unlinkdata with the right dname */
 573         error = nfs_copy_dname(sdentry,
 574                                 (struct nfs_unlinkdata *)dentry->d_fsdata);
 575         if (error) {
 576                 nfs_cancel_async_unlink(dentry);
 577                 goto out_dput;
 578         }
 579
 580         /* run the rename task, undo unlink if it fails */
 581         task = nfs_async_rename(dir, dir, dentry, sdentry);
 582         if (IS_ERR(task)) {
 583                 error = -EBUSY;
 584                 nfs_cancel_async_unlink(dentry);
 585                 goto out_dput;
 586         }
 587
 588         /* wait for the RPC task to complete, unless a SIGKILL intervenes */
 589         error = rpc_wait_for_completion_task(task);
 590         if (error == 0)
 591                 error = task->tk_status;
 592         rpc_put_task(task);
 593 out_dput:
 594         dput(sdentry);
 595 out:
 596         return error;
 597 }