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         iput(data->dir);
  34         put_rpccred(data->cred);
  35         kfree(data->args.name.name);
  36         kfree(data);
  37 }
  38
  39 #define NAME_ALLOC_LEN(len)     ((len+16) & ~15)
  40 /**
  41  * nfs_copy_dname - copy dentry name to data structure
  42  * @dentry: pointer to dentry
  43  * @data: nfs_unlinkdata
  44  */
  45 static int nfs_copy_dname(struct dentry *dentry, struct nfs_unlinkdata *data)
  46 {
  47         char            *str;
  48         int             len = dentry->d_name.len;
  49
  50         str = kmemdup(dentry->d_name.name, NAME_ALLOC_LEN(len), GFP_KERNEL);
  51         if (!str)
  52                 return -ENOMEM;
  53         data->args.name.len = len;
  54         data->args.name.name = str;
  55         return 0;
  56 }
  57
  58 static void nfs_free_dname(struct nfs_unlinkdata *data)
  59 {
  60         kfree(data->args.name.name);
  61         data->args.name.name = NULL;
  62         data->args.name.len = 0;
  63 }
  64
  65 static void nfs_dec_sillycount(struct inode *dir)
  66 {
  67         struct nfs_inode *nfsi = NFS_I(dir);
  68         if (atomic_dec_return(&nfsi->silly_count) == 1)
  69                 wake_up(&nfsi->waitqueue);
  70 }
  71
  72 /**
  73  * nfs_async_unlink_done - Sillydelete post-processing
  74  * @task: rpc_task of the sillydelete
  75  *
  76  * Do the directory attribute update.
  77  */
  78 static void nfs_async_unlink_done(struct rpc_task *task, void *calldata)
  79 {
  80         struct nfs_unlinkdata *data = calldata;
  81         struct inode *dir = data->dir;
  82
  83         trace_nfs_sillyrename_unlink(data, task->tk_status);
  84         if (!NFS_PROTO(dir)->unlink_done(task, dir))
  85                 rpc_restart_call_prepare(task);
  86 }
  87
  88 /**
  89  * nfs_async_unlink_release - Release the sillydelete data.
  90  * @task: rpc_task of the sillydelete
  91  *
  92  * We need to call nfs_put_unlinkdata as a 'tk_release' task since the
  93  * rpc_task would be freed too.
  94  */
  95 static void nfs_async_unlink_release(void *calldata)
  96 {
  97         struct nfs_unlinkdata   *data = calldata;
  98         struct super_block *sb = data->dir->i_sb;
  99
 100         nfs_dec_sillycount(data->dir);
 101         nfs_free_unlinkdata(data);
 102         nfs_sb_deactive(sb);
 103 }
 104
 105 static void nfs_unlink_prepare(struct rpc_task *task, void *calldata)
 106 {
 107         struct nfs_unlinkdata *data = calldata;
 108         NFS_PROTO(data->dir)->unlink_rpc_prepare(task, data);
 109 }
 110
 111 static const struct rpc_call_ops nfs_unlink_ops = {
 112         .rpc_call_done = nfs_async_unlink_done,
 113         .rpc_release = nfs_async_unlink_release,
 114         .rpc_call_prepare = nfs_unlink_prepare,
 115 };
 116
 117 static int nfs_do_call_unlink(struct dentry *parent, struct inode *dir, struct nfs_unlinkdata *data)
 118 {
 119         struct rpc_message msg = {
 120                 .rpc_argp = &data->args,
 121                 .rpc_resp = &data->res,
 122                 .rpc_cred = data->cred,
 123         };
 124         struct rpc_task_setup task_setup_data = {
 125                 .rpc_message = &msg,
 126                 .callback_ops = &nfs_unlink_ops,
 127                 .callback_data = data,
 128                 .workqueue = nfsiod_workqueue,
 129                 .flags = RPC_TASK_ASYNC,
 130         };
 131         struct rpc_task *task;
 132         struct dentry *alias;
 133
 134         alias = d_lookup(parent, &data->args.name);
 135         if (alias != NULL) {
 136                 int ret;
 137                 void *devname_garbage = NULL;
 138
 139                 /*
 140                  * Hey, we raced with lookup... See if we need to transfer
 141                  * the sillyrename information to the aliased dentry.
 142                  */
 143                 nfs_free_dname(data);
 144                 ret = nfs_copy_dname(alias, data);
 145                 spin_lock(&alias->d_lock);
 146                 if (ret == 0 && d_really_is_positive(alias) &&
 147                     !(alias->d_flags & DCACHE_NFSFS_RENAMED)) {
 148                         devname_garbage = alias->d_fsdata;
 149                         alias->d_fsdata = data;
 150                         alias->d_flags |= DCACHE_NFSFS_RENAMED;
 151                         ret = 1;
 152                 } else
 153                         ret = 0;
 154                 spin_unlock(&alias->d_lock);
 155                 nfs_dec_sillycount(dir);
 156                 dput(alias);
 157                 /*
 158                  * If we'd displaced old cached devname, free it.  At that
 159                  * point dentry is definitely not a root, so we won't need
 160                  * that anymore.
 161                  */
 162                 kfree(devname_garbage);
 163                 return ret;
 164         }
 165         data->dir = igrab(dir);
 166         if (!data->dir) {
 167                 nfs_dec_sillycount(dir);
 168                 return 0;
 169         }
 170         nfs_sb_active(dir->i_sb);
 171         data->args.fh = NFS_FH(dir);
 172         nfs_fattr_init(data->res.dir_attr);
 173
 174         NFS_PROTO(dir)->unlink_setup(&msg, dir);
 175
 176         task_setup_data.rpc_client = NFS_CLIENT(dir);
 177         task = rpc_run_task(&task_setup_data);
 178         if (!IS_ERR(task))
 179                 rpc_put_task_async(task);
 180         return 1;
 181 }
 182
 183 static int nfs_call_unlink(struct dentry *dentry, struct nfs_unlinkdata *data)
 184 {
 185         struct dentry *parent;
 186         struct inode *dir;
 187         int ret = 0;
 188
 189
 190         parent = dget_parent(dentry);
 191         if (parent == NULL)
 192                 goto out_free;
 193         dir = d_inode(parent);
 194         /* Non-exclusive lock protects against concurrent lookup() calls */
 195         spin_lock(&dir->i_lock);
 196         if (atomic_inc_not_zero(&NFS_I(dir)->silly_count) == 0) {
 197                 /* Deferred delete */
 198                 hlist_add_head(&data->list, &NFS_I(dir)->silly_list);
 199                 spin_unlock(&dir->i_lock);
 200                 ret = 1;
 201                 goto out_dput;
 202         }
 203         spin_unlock(&dir->i_lock);
 204         ret = nfs_do_call_unlink(parent, dir, data);
 205 out_dput:
 206         dput(parent);
 207 out_free:
 208         return ret;
 209 }
 210
 211 void nfs_wait_on_sillyrename(struct dentry *dentry)
 212 {
 213         struct nfs_inode *nfsi = NFS_I(d_inode(dentry));
 214
 215         wait_event(nfsi->waitqueue, atomic_read(&nfsi->silly_count) <= 1);
 216 }
 217
 218 void nfs_block_sillyrename(struct dentry *dentry)
 219 {
 220         struct nfs_inode *nfsi = NFS_I(d_inode(dentry));
 221
 222         wait_event(nfsi->waitqueue, atomic_cmpxchg(&nfsi->silly_count, 1, 0) == 1);
 223 }
 224
 225 void nfs_unblock_sillyrename(struct dentry *dentry)
 226 {
 227         struct inode *dir = d_inode(dentry);
 228         struct nfs_inode *nfsi = NFS_I(dir);
 229         struct nfs_unlinkdata *data;
 230
 231         atomic_inc(&nfsi->silly_count);
 232         spin_lock(&dir->i_lock);
 233         while (!hlist_empty(&nfsi->silly_list)) {
 234                 if (!atomic_inc_not_zero(&nfsi->silly_count))
 235                         break;
 236                 data = hlist_entry(nfsi->silly_list.first, struct nfs_unlinkdata, list);
 237                 hlist_del(&data->list);
 238                 spin_unlock(&dir->i_lock);
 239                 if (nfs_do_call_unlink(dentry, dir, data) == 0)
 240                         nfs_free_unlinkdata(data);
 241                 spin_lock(&dir->i_lock);
 242         }
 243         spin_unlock(&dir->i_lock);
 244 }
 245
 246 /**
 247  * nfs_async_unlink - asynchronous unlinking of a file
 248  * @dir: parent directory of dentry
 249  * @dentry: dentry to unlink
 250  */
 251 static int
 252 nfs_async_unlink(struct inode *dir, struct dentry *dentry)
 253 {
 254         struct nfs_unlinkdata *data;
 255         int status = -ENOMEM;
 256         void *devname_garbage = NULL;
 257
 258         data = kzalloc(sizeof(*data), GFP_KERNEL);
 259         if (data == NULL)
 260                 goto out;
 261
 262         data->cred = rpc_lookup_cred();
 263         if (IS_ERR(data->cred)) {
 264                 status = PTR_ERR(data->cred);
 265                 goto out_free;
 266         }
 267         data->res.dir_attr = &data->dir_attr;
 268
 269         status = -EBUSY;
 270         spin_lock(&dentry->d_lock);
 271         if (dentry->d_flags & DCACHE_NFSFS_RENAMED)
 272                 goto out_unlock;
 273         dentry->d_flags |= DCACHE_NFSFS_RENAMED;
 274         devname_garbage = dentry->d_fsdata;
 275         dentry->d_fsdata = data;
 276         spin_unlock(&dentry->d_lock);
 277         /*
 278          * If we'd displaced old cached devname, free it.  At that
 279          * point dentry is definitely not a root, so we won't need
 280          * that anymore.
 281          */
 282         kfree(devname_garbage);
 283         return 0;
 284 out_unlock:
 285         spin_unlock(&dentry->d_lock);
 286         put_rpccred(data->cred);
 287 out_free:
 288         kfree(data);
 289 out:
 290         return status;
 291 }
 292
 293 /**
 294  * nfs_complete_unlink - Initialize completion of the sillydelete
 295  * @dentry: dentry to delete
 296  * @inode: inode
 297  *
 298  * Since we're most likely to be called by dentry_iput(), we
 299  * only use the dentry to find the sillydelete. We then copy the name
 300  * into the qstr.
 301  */
 302 void
 303 nfs_complete_unlink(struct dentry *dentry, struct inode *inode)
 304 {
 305         struct nfs_unlinkdata   *data = NULL;
 306
 307         spin_lock(&dentry->d_lock);
 308         if (dentry->d_flags & DCACHE_NFSFS_RENAMED) {
 309                 dentry->d_flags &= ~DCACHE_NFSFS_RENAMED;
 310                 data = dentry->d_fsdata;
 311                 dentry->d_fsdata = NULL;
 312         }
 313         spin_unlock(&dentry->d_lock);
 314
 315         if (data != NULL && (NFS_STALE(inode) || !nfs_call_unlink(dentry, data)))
 316                 nfs_free_unlinkdata(data);
 317 }
 318
 319 /* Cancel a queued async unlink. Called when a sillyrename run fails. */
 320 static void
 321 nfs_cancel_async_unlink(struct dentry *dentry)
 322 {
 323         spin_lock(&dentry->d_lock);
 324         if (dentry->d_flags & DCACHE_NFSFS_RENAMED) {
 325                 struct nfs_unlinkdata *data = dentry->d_fsdata;
 326
 327                 dentry->d_flags &= ~DCACHE_NFSFS_RENAMED;
 328                 dentry->d_fsdata = NULL;
 329                 spin_unlock(&dentry->d_lock);
 330                 nfs_free_unlinkdata(data);
 331                 return;
 332         }
 333         spin_unlock(&dentry->d_lock);
 334 }
 335
 336 /**
 337  * nfs_async_rename_done - Sillyrename post-processing
 338  * @task: rpc_task of the sillyrename
 339  * @calldata: nfs_renamedata for the sillyrename
 340  *
 341  * Do the directory attribute updates and the d_move
 342  */
 343 static void nfs_async_rename_done(struct rpc_task *task, void *calldata)
 344 {
 345         struct nfs_renamedata *data = calldata;
 346         struct inode *old_dir = data->old_dir;
 347         struct inode *new_dir = data->new_dir;
 348         struct dentry *old_dentry = data->old_dentry;
 349
 350         trace_nfs_sillyrename_rename(old_dir, old_dentry,
 351                         new_dir, data->new_dentry, task->tk_status);
 352         if (!NFS_PROTO(old_dir)->rename_done(task, old_dir, new_dir)) {
 353                 rpc_restart_call_prepare(task);
 354                 return;
 355         }
 356
 357         if (data->complete)
 358                 data->complete(task, data);
 359 }
 360
 361 /**
 362  * nfs_async_rename_release - Release the sillyrename data.
 363  * @calldata: the struct nfs_renamedata to be released
 364  */
 365 static void nfs_async_rename_release(void *calldata)
 366 {
 367         struct nfs_renamedata   *data = calldata;
 368         struct super_block *sb = data->old_dir->i_sb;
 369
 370         if (d_really_is_positive(data->old_dentry))
 371                 nfs_mark_for_revalidate(d_inode(data->old_dentry));
 372
 373         dput(data->old_dentry);
 374         dput(data->new_dentry);
 375         iput(data->old_dir);
 376         iput(data->new_dir);
 377         nfs_sb_deactive(sb);
 378         put_rpccred(data->cred);
 379         kfree(data);
 380 }
 381
 382 static void nfs_rename_prepare(struct rpc_task *task, void *calldata)
 383 {
 384         struct nfs_renamedata *data = calldata;
 385         NFS_PROTO(data->old_dir)->rename_rpc_prepare(task, data);
 386 }
 387
 388 static const struct rpc_call_ops nfs_rename_ops = {
 389         .rpc_call_done = nfs_async_rename_done,
 390         .rpc_release = nfs_async_rename_release,
 391         .rpc_call_prepare = nfs_rename_prepare,
 392 };
 393
 394 /**
 395  * nfs_async_rename - perform an asynchronous rename operation
 396  * @old_dir: directory that currently holds the dentry to be renamed
 397  * @new_dir: target directory for the rename
 398  * @old_dentry: original dentry to be renamed
 399  * @new_dentry: dentry to which the old_dentry should be renamed
 400  *
 401  * It's expected that valid references to the dentries and inodes are held
 402  */
 403 struct rpc_task *
 404 nfs_async_rename(struct inode *old_dir, struct inode *new_dir,
 405                  struct dentry *old_dentry, struct dentry *new_dentry,
 406                  void (*complete)(struct rpc_task *, struct nfs_renamedata *))
 407 {
 408         struct nfs_renamedata *data;
 409         struct rpc_message msg = { };
 410         struct rpc_task_setup task_setup_data = {
 411                 .rpc_message = &msg,
 412                 .callback_ops = &nfs_rename_ops,
 413                 .workqueue = nfsiod_workqueue,
 414                 .rpc_client = NFS_CLIENT(old_dir),
 415                 .flags = RPC_TASK_ASYNC,
 416         };
 417
 418         data = kzalloc(sizeof(*data), GFP_KERNEL);
 419         if (data == NULL)
 420                 return ERR_PTR(-ENOMEM);
 421         task_setup_data.callback_data = data;
 422
 423         data->cred = rpc_lookup_cred();
 424         if (IS_ERR(data->cred)) {
 425                 struct rpc_task *task = ERR_CAST(data->cred);
 426                 kfree(data);
 427                 return task;
 428         }
 429
 430         msg.rpc_argp = &data->args;
 431         msg.rpc_resp = &data->res;
 432         msg.rpc_cred = data->cred;
 433
 434         /* set up nfs_renamedata */
 435         data->old_dir = old_dir;
 436         ihold(old_dir);
 437         data->new_dir = new_dir;
 438         ihold(new_dir);
 439         data->old_dentry = dget(old_dentry);
 440         data->new_dentry = dget(new_dentry);
 441         nfs_fattr_init(&data->old_fattr);
 442         nfs_fattr_init(&data->new_fattr);
 443         data->complete = complete;
 444
 445         /* set up nfs_renameargs */
 446         data->args.old_dir = NFS_FH(old_dir);
 447         data->args.old_name = &old_dentry->d_name;
 448         data->args.new_dir = NFS_FH(new_dir);
 449         data->args.new_name = &new_dentry->d_name;
 450
 451         /* set up nfs_renameres */
 452         data->res.old_fattr = &data->old_fattr;
 453         data->res.new_fattr = &data->new_fattr;
 454
 455         nfs_sb_active(old_dir->i_sb);
 456
 457         NFS_PROTO(data->old_dir)->rename_setup(&msg, old_dir);
 458
 459         return rpc_run_task(&task_setup_data);
 460 }
 461
 462 /*
 463  * Perform tasks needed when a sillyrename is done such as cancelling the
 464  * queued async unlink if it failed.
 465  */
 466 static void
 467 nfs_complete_sillyrename(struct rpc_task *task, struct nfs_renamedata *data)
 468 {
 469         struct dentry *dentry = data->old_dentry;
 470
 471         if (task->tk_status != 0) {
 472                 nfs_cancel_async_unlink(dentry);
 473                 return;
 474         }
 475
 476         /*
 477          * vfs_unlink and the like do not issue this when a file is
 478          * sillyrenamed, so do it here.
 479          */
 480         fsnotify_nameremove(dentry, 0);
 481 }
 482
 483 #define SILLYNAME_PREFIX ".nfs"
 484 #define SILLYNAME_PREFIX_LEN ((unsigned)sizeof(SILLYNAME_PREFIX) - 1)
 485 #define SILLYNAME_FILEID_LEN ((unsigned)sizeof(u64) << 1)
 486 #define SILLYNAME_COUNTER_LEN ((unsigned)sizeof(unsigned int) << 1)
 487 #define SILLYNAME_LEN (SILLYNAME_PREFIX_LEN + \
 488                 SILLYNAME_FILEID_LEN + \
 489                 SILLYNAME_COUNTER_LEN)
 490
 491 /**
 492  * nfs_sillyrename - Perform a silly-rename of a dentry
 493  * @dir: inode of directory that contains dentry
 494  * @dentry: dentry to be sillyrenamed
 495  *
 496  * NFSv2/3 is stateless and the server doesn't know when the client is
 497  * holding a file open. To prevent application problems when a file is
 498  * unlinked while it's still open, the client performs a "silly-rename".
 499  * That is, it renames the file to a hidden file in the same directory,
 500  * and only performs the unlink once the last reference to it is put.
 501  *
 502  * The final cleanup is done during dentry_iput.
 503  *
 504  * (Note: NFSv4 is stateful, and has opens, so in theory an NFSv4 server
 505  * could take responsibility for keeping open files referenced.  The server
 506  * would also need to ensure that opened-but-deleted files were kept over
 507  * reboots.  However, we may not assume a server does so.  (RFC 5661
 508  * does provide an OPEN4_RESULT_PRESERVE_UNLINKED flag that a server can
 509  * use to advertise that it does this; some day we may take advantage of
 510  * it.))
 511  */
 512 int
 513 nfs_sillyrename(struct inode *dir, struct dentry *dentry)
 514 {
 515         static unsigned int sillycounter;
 516         unsigned char silly[SILLYNAME_LEN + 1];
 517         unsigned long long fileid;
 518         struct dentry *sdentry;
 519         struct rpc_task *task;
 520         int            error = -EBUSY;
 521
 522         dfprintk(VFS, "NFS: silly-rename(%pd2, ct=%d)\n",
 523                 dentry, d_count(dentry));
 524         nfs_inc_stats(dir, NFSIOS_SILLYRENAME);
 525
 526         /*
 527          * We don't allow a dentry to be silly-renamed twice.
 528          */
 529         if (dentry->d_flags & DCACHE_NFSFS_RENAMED)
 530                 goto out;
 531
 532         fileid = NFS_FILEID(d_inode(dentry));
 533
 534         /* Return delegation in anticipation of the rename */
 535         NFS_PROTO(d_inode(dentry))->return_delegation(d_inode(dentry));
 536
 537         sdentry = NULL;
 538         do {
 539                 int slen;
 540                 dput(sdentry);
 541                 sillycounter++;
 542                 slen = scnprintf(silly, sizeof(silly),
 543                                 SILLYNAME_PREFIX "%0*llx%0*x",
 544                                 SILLYNAME_FILEID_LEN, fileid,
 545                                 SILLYNAME_COUNTER_LEN, sillycounter);
 546
 547                 dfprintk(VFS, "NFS: trying to rename %pd to %s\n",
 548                                 dentry, silly);
 549
 550                 sdentry = lookup_one_len(silly, dentry->d_parent, slen);
 551                 /*
 552                  * N.B. Better to return EBUSY here ... it could be
 553                  * dangerous to delete the file while it's in use.
 554                  */
 555                 if (IS_ERR(sdentry))
 556                         goto out;
 557         } while (d_inode(sdentry) != NULL); /* need negative lookup */
 558
 559         /* queue unlink first. Can't do this from rpc_release as it
 560          * has to allocate memory
 561          */
 562         error = nfs_async_unlink(dir, dentry);
 563         if (error)
 564                 goto out_dput;
 565
 566         /* populate unlinkdata with the right dname */
 567         error = nfs_copy_dname(sdentry,
 568                                 (struct nfs_unlinkdata *)dentry->d_fsdata);
 569         if (error) {
 570                 nfs_cancel_async_unlink(dentry);
 571                 goto out_dput;
 572         }
 573
 574         /* run the rename task, undo unlink if it fails */
 575         task = nfs_async_rename(dir, dir, dentry, sdentry,
 576                                         nfs_complete_sillyrename);
 577         if (IS_ERR(task)) {
 578                 error = -EBUSY;
 579                 nfs_cancel_async_unlink(dentry);
 580                 goto out_dput;
 581         }
 582
 583         /* wait for the RPC task to complete, unless a SIGKILL intervenes */
 584         error = rpc_wait_for_completion_task(task);
 585         if (error == 0)
 586                 error = task->tk_status;
 587         switch (error) {
 588         case 0:
 589                 /* The rename succeeded */
 590                 nfs_set_verifier(dentry, nfs_save_change_attribute(dir));
 591                 d_move(dentry, sdentry);
 592                 break;
 593         case -ERESTARTSYS:
 594                 /* The result of the rename is unknown. Play it safe by
 595                  * forcing a new lookup */
 596                 d_drop(dentry);
 597                 d_drop(sdentry);
 598         }
 599         rpc_put_task(task);
 600 out_dput:
 601         dput(sdentry);
 602 out:
 603         return error;
 604 }