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 #include "nfstrace.h"
  24
  25 /**
  26  * nfs_free_unlinkdata - release data from a sillydelete operation.
  27  * @data: pointer to unlink structure.
  28  */
  29 static void
  30 nfs_free_unlinkdata(struct nfs_unlinkdata *data)
  31 {
  32         iput(data->dir);
  33         put_rpccred(data->cred);
  34         kfree(data->args.name.name);
  35         kfree(data);
  36 }
  37
  38 #define NAME_ALLOC_LEN(len)     ((len+16) & ~15)
  39 /**
  40  * nfs_copy_dname - copy dentry name to data structure
  41  * @dentry: pointer to dentry
  42  * @data: nfs_unlinkdata
  43  */
  44 static int nfs_copy_dname(struct dentry *dentry, struct nfs_unlinkdata *data)
  45 {
  46         char            *str;
  47         int             len = dentry->d_name.len;
  48
  49         str = kmemdup(dentry->d_name.name, NAME_ALLOC_LEN(len), GFP_KERNEL);
  50         if (!str)
  51                 return -ENOMEM;
  52         data->args.name.len = len;
  53         data->args.name.name = str;
  54         return 0;
  55 }
  56
  57 static void nfs_free_dname(struct nfs_unlinkdata *data)
  58 {
  59         kfree(data->args.name.name);
  60         data->args.name.name = NULL;
  61         data->args.name.len = 0;
  62 }
  63
  64 static void nfs_dec_sillycount(struct inode *dir)
  65 {
  66         struct nfs_inode *nfsi = NFS_I(dir);
  67         if (atomic_dec_return(&nfsi->silly_count) == 1)
  68                 wake_up(&nfsi->waitqueue);
  69 }
  70
  71 /**
  72  * nfs_async_unlink_done - Sillydelete post-processing
  73  * @task: rpc_task of the sillydelete
  74  *
  75  * Do the directory attribute update.
  76  */
  77 static void nfs_async_unlink_done(struct rpc_task *task, void *calldata)
  78 {
  79         struct nfs_unlinkdata *data = calldata;
  80         struct inode *dir = data->dir;
  81
  82         trace_nfs_sillyrename_unlink(data, task->tk_status);
  83         if (!NFS_PROTO(dir)->unlink_done(task, dir))
  84                 rpc_restart_call_prepare(task);
  85 }
  86
  87 /**
  88  * nfs_async_unlink_release - Release the sillydelete data.
  89  * @task: rpc_task of the sillydelete
  90  *
  91  * We need to call nfs_put_unlinkdata as a 'tk_release' task since the
  92  * rpc_task would be freed too.
  93  */
  94 static void nfs_async_unlink_release(void *calldata)
  95 {
  96         struct nfs_unlinkdata   *data = calldata;
  97         struct super_block *sb = data->dir->i_sb;
  98
  99         nfs_dec_sillycount(data->dir);
 100         nfs_free_unlinkdata(data);
 101         nfs_sb_deactive(sb);
 102 }
 103
 104 static void nfs_unlink_prepare(struct rpc_task *task, void *calldata)
 105 {
 106         struct nfs_unlinkdata *data = calldata;
 107         NFS_PROTO(data->dir)->unlink_rpc_prepare(task, data);
 108 }
 109
 110 static const struct rpc_call_ops nfs_unlink_ops = {
 111         .rpc_call_done = nfs_async_unlink_done,
 112         .rpc_release = nfs_async_unlink_release,
 113         .rpc_call_prepare = nfs_unlink_prepare,
 114 };
 115
 116 static int nfs_do_call_unlink(struct dentry *parent, struct inode *dir, struct nfs_unlinkdata *data)
 117 {
 118         struct rpc_message msg = {
 119                 .rpc_argp = &data->args,
 120                 .rpc_resp = &data->res,
 121                 .rpc_cred = data->cred,
 122         };
 123         struct rpc_task_setup task_setup_data = {
 124                 .rpc_message = &msg,
 125                 .callback_ops = &nfs_unlink_ops,
 126                 .callback_data = data,
 127                 .workqueue = nfsiod_workqueue,
 128                 .flags = RPC_TASK_ASYNC,
 129         };
 130         struct rpc_task *task;
 131         struct dentry *alias;
 132
 133         alias = d_lookup(parent, &data->args.name);
 134         if (alias != NULL) {
 135                 int ret;
 136                 void *devname_garbage = NULL;
 137
 138                 /*
 139                  * Hey, we raced with lookup... See if we need to transfer
 140                  * the sillyrename information to the aliased dentry.
 141                  */
 142                 nfs_free_dname(data);
 143                 ret = nfs_copy_dname(alias, data);
 144                 spin_lock(&alias->d_lock);
 145                 if (ret == 0 && alias->d_inode != NULL &&
 146                     !(alias->d_flags & DCACHE_NFSFS_RENAMED)) {
 147                         devname_garbage = alias->d_fsdata;
 148                         alias->d_fsdata = data;
 149                         alias->d_flags |= DCACHE_NFSFS_RENAMED;
 150                         ret = 1;
 151                 } else
 152                         ret = 0;
 153                 spin_unlock(&alias->d_lock);
 154                 nfs_dec_sillycount(dir);
 155                 dput(alias);
 156                 /*
 157                  * If we'd displaced old cached devname, free it.  At that
 158                  * point dentry is definitely not a root, so we won't need
 159                  * that anymore.
 160                  */
 161                 kfree(devname_garbage);
 162                 return ret;
 163         }
 164         data->dir = igrab(dir);
 165         if (!data->dir) {
 166                 nfs_dec_sillycount(dir);
 167                 return 0;
 168         }
 169         nfs_sb_active(dir->i_sb);
 170         data->args.fh = NFS_FH(dir);
 171         nfs_fattr_init(data->res.dir_attr);
 172
 173         NFS_PROTO(dir)->unlink_setup(&msg, dir);
 174
 175         task_setup_data.rpc_client = NFS_CLIENT(dir);
 176         task = rpc_run_task(&task_setup_data);
 177         if (!IS_ERR(task))
 178                 rpc_put_task_async(task);
 179         return 1;
 180 }
 181
 182 static int nfs_call_unlink(struct dentry *dentry, struct nfs_unlinkdata *data)
 183 {
 184         struct dentry *parent;
 185         struct inode *dir;
 186         int ret = 0;
 187
 188
 189         parent = dget_parent(dentry);
 190         if (parent == NULL)
 191                 goto out_free;
 192         dir = parent->d_inode;
 193         /* Non-exclusive lock protects against concurrent lookup() calls */
 194         spin_lock(&dir->i_lock);
 195         if (atomic_inc_not_zero(&NFS_I(dir)->silly_count) == 0) {
 196                 /* Deferred delete */
 197                 hlist_add_head(&data->list, &NFS_I(dir)->silly_list);
 198                 spin_unlock(&dir->i_lock);
 199                 ret = 1;
 200                 goto out_dput;
 201         }
 202         spin_unlock(&dir->i_lock);
 203         ret = nfs_do_call_unlink(parent, dir, data);
 204 out_dput:
 205         dput(parent);
 206 out_free:
 207         return ret;
 208 }
 209
 210 void nfs_wait_on_sillyrename(struct dentry *dentry)
 211 {
 212         struct nfs_inode *nfsi = NFS_I(dentry->d_inode);
 213
 214         wait_event(nfsi->waitqueue, atomic_read(&nfsi->silly_count) <= 1);
 215 }
 216
 217 void nfs_block_sillyrename(struct dentry *dentry)
 218 {
 219         struct nfs_inode *nfsi = NFS_I(dentry->d_inode);
 220
 221         wait_event(nfsi->waitqueue, atomic_cmpxchg(&nfsi->silly_count, 1, 0) == 1);
 222 }
 223
 224 void nfs_unblock_sillyrename(struct dentry *dentry)
 225 {
 226         struct inode *dir = dentry->d_inode;
 227         struct nfs_inode *nfsi = NFS_I(dir);
 228         struct nfs_unlinkdata *data;
 229
 230         atomic_inc(&nfsi->silly_count);
 231         spin_lock(&dir->i_lock);
 232         while (!hlist_empty(&nfsi->silly_list)) {
 233                 if (!atomic_inc_not_zero(&nfsi->silly_count))
 234                         break;
 235                 data = hlist_entry(nfsi->silly_list.first, struct nfs_unlinkdata, list);
 236                 hlist_del(&data->list);
 237                 spin_unlock(&dir->i_lock);
 238                 if (nfs_do_call_unlink(dentry, dir, data) == 0)
 239                         nfs_free_unlinkdata(data);
 240                 spin_lock(&dir->i_lock);
 241         }
 242         spin_unlock(&dir->i_lock);
 243 }
 244
 245 /**
 246  * nfs_async_unlink - asynchronous unlinking of a file
 247  * @dir: parent directory of dentry
 248  * @dentry: dentry to unlink
 249  */
 250 static int
 251 nfs_async_unlink(struct inode *dir, struct dentry *dentry)
 252 {
 253         struct nfs_unlinkdata *data;
 254         int status = -ENOMEM;
 255         void *devname_garbage = NULL;
 256
 257         data = kzalloc(sizeof(*data), GFP_KERNEL);
 258         if (data == NULL)
 259                 goto out;
 260
 261         data->cred = rpc_lookup_cred();
 262         if (IS_ERR(data->cred)) {
 263                 status = PTR_ERR(data->cred);
 264                 goto out_free;
 265         }
 266         data->res.dir_attr = &data->dir_attr;
 267
 268         status = -EBUSY;
 269         spin_lock(&dentry->d_lock);
 270         if (dentry->d_flags & DCACHE_NFSFS_RENAMED)
 271                 goto out_unlock;
 272         dentry->d_flags |= DCACHE_NFSFS_RENAMED;
 273         devname_garbage = dentry->d_fsdata;
 274         dentry->d_fsdata = data;
 275         spin_unlock(&dentry->d_lock);
 276         /*
 277          * If we'd displaced old cached devname, free it.  At that
 278          * point dentry is definitely not a root, so we won't need
 279          * that anymore.
 280          */
 281         kfree(devname_garbage);
 282         return 0;
 283 out_unlock:
 284         spin_unlock(&dentry->d_lock);
 285         put_rpccred(data->cred);
 286 out_free:
 287         kfree(data);
 288 out:
 289         return status;
 290 }
 291
 292 /**
 293  * nfs_complete_unlink - Initialize completion of the sillydelete
 294  * @dentry: dentry to delete
 295  * @inode: inode
 296  *
 297  * Since we're most likely to be called by dentry_iput(), we
 298  * only use the dentry to find the sillydelete. We then copy the name
 299  * into the qstr.
 300  */
 301 void
 302 nfs_complete_unlink(struct dentry *dentry, struct inode *inode)
 303 {
 304         struct nfs_unlinkdata   *data = NULL;
 305
 306         spin_lock(&dentry->d_lock);
 307         if (dentry->d_flags & DCACHE_NFSFS_RENAMED) {
 308                 dentry->d_flags &= ~DCACHE_NFSFS_RENAMED;
 309                 data = dentry->d_fsdata;
 310                 dentry->d_fsdata = NULL;
 311         }
 312         spin_unlock(&dentry->d_lock);
 313
 314         if (data != NULL && (NFS_STALE(inode) || !nfs_call_unlink(dentry, data)))
 315                 nfs_free_unlinkdata(data);
 316 }
 317
 318 /* Cancel a queued async unlink. Called when a sillyrename run fails. */
 319 static void
 320 nfs_cancel_async_unlink(struct dentry *dentry)
 321 {
 322         spin_lock(&dentry->d_lock);
 323         if (dentry->d_flags & DCACHE_NFSFS_RENAMED) {
 324                 struct nfs_unlinkdata *data = dentry->d_fsdata;
 325
 326                 dentry->d_flags &= ~DCACHE_NFSFS_RENAMED;
 327                 dentry->d_fsdata = NULL;
 328                 spin_unlock(&dentry->d_lock);
 329                 nfs_free_unlinkdata(data);
 330                 return;
 331         }
 332         spin_unlock(&dentry->d_lock);
 333 }
 334
 335 /**
 336  * nfs_async_rename_done - Sillyrename post-processing
 337  * @task: rpc_task of the sillyrename
 338  * @calldata: nfs_renamedata for the sillyrename
 339  *
 340  * Do the directory attribute updates and the d_move
 341  */
 342 static void nfs_async_rename_done(struct rpc_task *task, void *calldata)
 343 {
 344         struct nfs_renamedata *data = calldata;
 345         struct inode *old_dir = data->old_dir;
 346         struct inode *new_dir = data->new_dir;
 347         struct dentry *old_dentry = data->old_dentry;
 348
 349         trace_nfs_sillyrename_rename(old_dir, old_dentry,
 350                         new_dir, data->new_dentry, task->tk_status);
 351         if (!NFS_PROTO(old_dir)->rename_done(task, old_dir, new_dir)) {
 352                 rpc_restart_call_prepare(task);
 353                 return;
 354         }
 355
 356         if (task->tk_status != 0)
 357                 nfs_cancel_async_unlink(old_dentry);
 358 }
 359
 360 /**
 361  * nfs_async_rename_release - Release the sillyrename data.
 362  * @calldata: the struct nfs_renamedata to be released
 363  */
 364 static void nfs_async_rename_release(void *calldata)
 365 {
 366         struct nfs_renamedata   *data = calldata;
 367         struct super_block *sb = data->old_dir->i_sb;
 368
 369         if (data->old_dentry->d_inode)
 370                 nfs_mark_for_revalidate(data->old_dentry->d_inode);
 371
 372         dput(data->old_dentry);
 373         dput(data->new_dentry);
 374         iput(data->old_dir);
 375         iput(data->new_dir);
 376         nfs_sb_deactive(sb);
 377         put_rpccred(data->cred);
 378         kfree(data);
 379 }
 380
 381 static void nfs_rename_prepare(struct rpc_task *task, void *calldata)
 382 {
 383         struct nfs_renamedata *data = calldata;
 384         NFS_PROTO(data->old_dir)->rename_rpc_prepare(task, data);
 385 }
 386
 387 static const struct rpc_call_ops nfs_rename_ops = {
 388         .rpc_call_done = nfs_async_rename_done,
 389         .rpc_release = nfs_async_rename_release,
 390         .rpc_call_prepare = nfs_rename_prepare,
 391 };
 392
 393 /**
 394  * nfs_async_rename - perform an asynchronous rename operation
 395  * @old_dir: directory that currently holds the dentry to be renamed
 396  * @new_dir: target directory for the rename
 397  * @old_dentry: original dentry to be renamed
 398  * @new_dentry: dentry to which the old_dentry should be renamed
 399  *
 400  * It's expected that valid references to the dentries and inodes are held
 401  */
 402 static struct rpc_task *
 403 nfs_async_rename(struct inode *old_dir, struct inode *new_dir,
 404                  struct dentry *old_dentry, struct dentry *new_dentry)
 405 {
 406         struct nfs_renamedata *data;
 407         struct rpc_message msg = { };
 408         struct rpc_task_setup task_setup_data = {
 409                 .rpc_message = &msg,
 410                 .callback_ops = &nfs_rename_ops,
 411                 .workqueue = nfsiod_workqueue,
 412                 .rpc_client = NFS_CLIENT(old_dir),
 413                 .flags = RPC_TASK_ASYNC,
 414         };
 415
 416         data = kzalloc(sizeof(*data), GFP_KERNEL);
 417         if (data == NULL)
 418                 return ERR_PTR(-ENOMEM);
 419         task_setup_data.callback_data = data;
 420
 421         data->cred = rpc_lookup_cred();
 422         if (IS_ERR(data->cred)) {
 423                 struct rpc_task *task = ERR_CAST(data->cred);
 424                 kfree(data);
 425                 return task;
 426         }
 427
 428         msg.rpc_argp = &data->args;
 429         msg.rpc_resp = &data->res;
 430         msg.rpc_cred = data->cred;
 431
 432         /* set up nfs_renamedata */
 433         data->old_dir = old_dir;
 434         ihold(old_dir);
 435         data->new_dir = new_dir;
 436         ihold(new_dir);
 437         data->old_dentry = dget(old_dentry);
 438         data->new_dentry = dget(new_dentry);
 439         nfs_fattr_init(&data->old_fattr);
 440         nfs_fattr_init(&data->new_fattr);
 441
 442         /* set up nfs_renameargs */
 443         data->args.old_dir = NFS_FH(old_dir);
 444         data->args.old_name = &old_dentry->d_name;
 445         data->args.new_dir = NFS_FH(new_dir);
 446         data->args.new_name = &new_dentry->d_name;
 447
 448         /* set up nfs_renameres */
 449         data->res.old_fattr = &data->old_fattr;
 450         data->res.new_fattr = &data->new_fattr;
 451
 452         nfs_sb_active(old_dir->i_sb);
 453
 454         NFS_PROTO(data->old_dir)->rename_setup(&msg, old_dir);
 455
 456         return rpc_run_task(&task_setup_data);
 457 }
 458
 459 #define SILLYNAME_PREFIX ".nfs"
 460 #define SILLYNAME_PREFIX_LEN ((unsigned)sizeof(SILLYNAME_PREFIX) - 1)
 461 #define SILLYNAME_FILEID_LEN ((unsigned)sizeof(u64) << 1)
 462 #define SILLYNAME_COUNTER_LEN ((unsigned)sizeof(unsigned int) << 1)
 463 #define SILLYNAME_LEN (SILLYNAME_PREFIX_LEN + \
 464                 SILLYNAME_FILEID_LEN + \
 465                 SILLYNAME_COUNTER_LEN)
 466
 467 /**
 468  * nfs_sillyrename - Perform a silly-rename of a dentry
 469  * @dir: inode of directory that contains dentry
 470  * @dentry: dentry to be sillyrenamed
 471  *
 472  * NFSv2/3 is stateless and the server doesn't know when the client is
 473  * holding a file open. To prevent application problems when a file is
 474  * unlinked while it's still open, the client performs a "silly-rename".
 475  * That is, it renames the file to a hidden file in the same directory,
 476  * and only performs the unlink once the last reference to it is put.
 477  *
 478  * The final cleanup is done during dentry_iput.
 479  *
 480  * (Note: NFSv4 is stateful, and has opens, so in theory an NFSv4 server
 481  * could take responsibility for keeping open files referenced.  The server
 482  * would also need to ensure that opened-but-deleted files were kept over
 483  * reboots.  However, we may not assume a server does so.  (RFC 5661
 484  * does provide an OPEN4_RESULT_PRESERVE_UNLINKED flag that a server can
 485  * use to advertise that it does this; some day we may take advantage of
 486  * it.))
 487  */
 488 int
 489 nfs_sillyrename(struct inode *dir, struct dentry *dentry)
 490 {
 491         static unsigned int sillycounter;
 492         unsigned char silly[SILLYNAME_LEN + 1];
 493         unsigned long long fileid;
 494         struct dentry *sdentry;
 495         struct rpc_task *task;
 496         int            error = -EIO;
 497
 498         dfprintk(VFS, "NFS: silly-rename(%s/%s, ct=%d)\n",
 499                 dentry->d_parent->d_name.name, dentry->d_name.name,
 500                 d_count(dentry));
 501         nfs_inc_stats(dir, NFSIOS_SILLYRENAME);
 502
 503         /*
 504          * We don't allow a dentry to be silly-renamed twice.
 505          */
 506         error = -EBUSY;
 507         if (dentry->d_flags & DCACHE_NFSFS_RENAMED)
 508                 goto out;
 509
 510         fileid = NFS_FILEID(dentry->d_inode);
 511
 512         /* Return delegation in anticipation of the rename */
 513         NFS_PROTO(dentry->d_inode)->return_delegation(dentry->d_inode);
 514
 515         sdentry = NULL;
 516         do {
 517                 int slen;
 518                 dput(sdentry);
 519                 sillycounter++;
 520                 slen = scnprintf(silly, sizeof(silly),
 521                                 SILLYNAME_PREFIX "%0*llx%0*x",
 522                                 SILLYNAME_FILEID_LEN, fileid,
 523                                 SILLYNAME_COUNTER_LEN, sillycounter);
 524
 525                 dfprintk(VFS, "NFS: trying to rename %s to %s\n",
 526                                 dentry->d_name.name, silly);
 527
 528                 sdentry = lookup_one_len(silly, dentry->d_parent, slen);
 529                 /*
 530                  * N.B. Better to return EBUSY here ... it could be
 531                  * dangerous to delete the file while it's in use.
 532                  */
 533                 if (IS_ERR(sdentry))
 534                         goto out;
 535         } while (sdentry->d_inode != NULL); /* need negative lookup */
 536
 537         /* queue unlink first. Can't do this from rpc_release as it
 538          * has to allocate memory
 539          */
 540         error = nfs_async_unlink(dir, dentry);
 541         if (error)
 542                 goto out_dput;
 543
 544         /* populate unlinkdata with the right dname */
 545         error = nfs_copy_dname(sdentry,
 546                                 (struct nfs_unlinkdata *)dentry->d_fsdata);
 547         if (error) {
 548                 nfs_cancel_async_unlink(dentry);
 549                 goto out_dput;
 550         }
 551
 552         /* run the rename task, undo unlink if it fails */
 553         task = nfs_async_rename(dir, dir, dentry, sdentry);
 554         if (IS_ERR(task)) {
 555                 error = -EBUSY;
 556                 nfs_cancel_async_unlink(dentry);
 557                 goto out_dput;
 558         }
 559
 560         /* wait for the RPC task to complete, unless a SIGKILL intervenes */
 561         error = rpc_wait_for_completion_task(task);
 562         if (error == 0)
 563                 error = task->tk_status;
 564         switch (error) {
 565         case 0:
 566                 /* The rename succeeded */
 567                 nfs_set_verifier(dentry, nfs_save_change_attribute(dir));
 568                 d_move(dentry, sdentry);
 569                 break;
 570         case -ERESTARTSYS:
 571                 /* The result of the rename is unknown. Play it safe by
 572                  * forcing a new lookup */
 573                 d_drop(dentry);
 574                 d_drop(sdentry);
 575         }
 576         rpc_put_task(task);
 577 out_dput:
 578         dput(sdentry);
 579 out:
 580         return error;
 581 }