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 /**
  24  * nfs_free_unlinkdata - release data from a sillydelete operation.
  25  * @data: pointer to unlink structure.
  26  */
  27 static void
  28 nfs_free_unlinkdata(struct nfs_unlinkdata *data)
  29 {
  30         iput(data->dir);
  31         put_rpccred(data->cred);
  32         kfree(data->args.name.name);
  33         kfree(data);
  34 }
  35
  36 #define NAME_ALLOC_LEN(len)     ((len+16) & ~15)
  37 /**
  38  * nfs_copy_dname - copy dentry name to data structure
  39  * @dentry: pointer to dentry
  40  * @data: nfs_unlinkdata
  41  */
  42 static int nfs_copy_dname(struct dentry *dentry, struct nfs_unlinkdata *data)
  43 {
  44         char            *str;
  45         int             len = dentry->d_name.len;
  46
  47         str = kmemdup(dentry->d_name.name, NAME_ALLOC_LEN(len), GFP_KERNEL);
  48         if (!str)
  49                 return -ENOMEM;
  50         data->args.name.len = len;
  51         data->args.name.name = str;
  52         return 0;
  53 }
  54
  55 static void nfs_free_dname(struct nfs_unlinkdata *data)
  56 {
  57         kfree(data->args.name.name);
  58         data->args.name.name = NULL;
  59         data->args.name.len = 0;
  60 }
  61
  62 static void nfs_dec_sillycount(struct inode *dir)
  63 {
  64         struct nfs_inode *nfsi = NFS_I(dir);
  65         if (atomic_dec_return(&nfsi->silly_count) == 1)
  66                 wake_up(&nfsi->waitqueue);
  67 }
  68
  69 /**
  70  * nfs_async_unlink_done - Sillydelete post-processing
  71  * @task: rpc_task of the sillydelete
  72  *
  73  * Do the directory attribute update.
  74  */
  75 static void nfs_async_unlink_done(struct rpc_task *task, void *calldata)
  76 {
  77         struct nfs_unlinkdata *data = calldata;
  78         struct inode *dir = data->dir;
  79
  80         if (!NFS_PROTO(dir)->unlink_done(task, dir))
  81                 rpc_restart_call_prepare(task);
  82 }
  83
  84 /**
  85  * nfs_async_unlink_release - Release the sillydelete data.
  86  * @task: rpc_task of the sillydelete
  87  *
  88  * We need to call nfs_put_unlinkdata as a 'tk_release' task since the
  89  * rpc_task would be freed too.
  90  */
  91 static void nfs_async_unlink_release(void *calldata)
  92 {
  93         struct nfs_unlinkdata   *data = calldata;
  94         struct super_block *sb = data->dir->i_sb;
  95
  96         nfs_dec_sillycount(data->dir);
  97         nfs_free_unlinkdata(data);
  98         nfs_sb_deactive(sb);
  99 }
 100
 101 static void nfs_unlink_prepare(struct rpc_task *task, void *calldata)
 102 {
 103         struct nfs_unlinkdata *data = calldata;
 104         NFS_PROTO(data->dir)->unlink_rpc_prepare(task, data);
 105 }
 106
 107 static const struct rpc_call_ops nfs_unlink_ops = {
 108         .rpc_call_done = nfs_async_unlink_done,
 109         .rpc_release = nfs_async_unlink_release,
 110         .rpc_call_prepare = nfs_unlink_prepare,
 111 };
 112
 113 static int nfs_do_call_unlink(struct dentry *parent, struct inode *dir, struct nfs_unlinkdata *data)
 114 {
 115         struct rpc_message msg = {
 116                 .rpc_argp = &data->args,
 117                 .rpc_resp = &data->res,
 118                 .rpc_cred = data->cred,
 119         };
 120         struct rpc_task_setup task_setup_data = {
 121                 .rpc_message = &msg,
 122                 .callback_ops = &nfs_unlink_ops,
 123                 .callback_data = data,
 124                 .workqueue = nfsiod_workqueue,
 125                 .flags = RPC_TASK_ASYNC,
 126         };
 127         struct rpc_task *task;
 128         struct dentry *alias;
 129
 130         alias = d_lookup(parent, &data->args.name);
 131         if (alias != NULL) {
 132                 int ret;
 133                 void *devname_garbage = NULL;
 134
 135                 /*
 136                  * Hey, we raced with lookup... See if we need to transfer
 137                  * the sillyrename information to the aliased dentry.
 138                  */
 139                 nfs_free_dname(data);
 140                 ret = nfs_copy_dname(alias, data);
 141                 spin_lock(&alias->d_lock);
 142                 if (ret == 0 && alias->d_inode != NULL &&
 143                     !(alias->d_flags & DCACHE_NFSFS_RENAMED)) {
 144                         devname_garbage = alias->d_fsdata;
 145                         alias->d_fsdata = data;
 146                         alias->d_flags |= DCACHE_NFSFS_RENAMED;
 147                         ret = 1;
 148                 } else
 149                         ret = 0;
 150                 spin_unlock(&alias->d_lock);
 151                 nfs_dec_sillycount(dir);
 152                 dput(alias);
 153                 /*
 154                  * If we'd displaced old cached devname, free it.  At that
 155                  * point dentry is definitely not a root, so we won't need
 156                  * that anymore.
 157                  */
 158                 kfree(devname_garbage);
 159                 return ret;
 160         }
 161         data->dir = igrab(dir);
 162         if (!data->dir) {
 163                 nfs_dec_sillycount(dir);
 164                 return 0;
 165         }
 166         nfs_sb_active(dir->i_sb);
 167         data->args.fh = NFS_FH(dir);
 168         nfs_fattr_init(data->res.dir_attr);
 169
 170         NFS_PROTO(dir)->unlink_setup(&msg, dir);
 171
 172         task_setup_data.rpc_client = NFS_CLIENT(dir);
 173         task = rpc_run_task(&task_setup_data);
 174         if (!IS_ERR(task))
 175                 rpc_put_task_async(task);
 176         return 1;
 177 }
 178
 179 static int nfs_call_unlink(struct dentry *dentry, struct nfs_unlinkdata *data)
 180 {
 181         struct dentry *parent;
 182         struct inode *dir;
 183         int ret = 0;
 184
 185
 186         parent = dget_parent(dentry);
 187         if (parent == NULL)
 188                 goto out_free;
 189         dir = parent->d_inode;
 190         /* Non-exclusive lock protects against concurrent lookup() calls */
 191         spin_lock(&dir->i_lock);
 192         if (atomic_inc_not_zero(&NFS_I(dir)->silly_count) == 0) {
 193                 /* Deferred delete */
 194                 hlist_add_head(&data->list, &NFS_I(dir)->silly_list);
 195                 spin_unlock(&dir->i_lock);
 196                 ret = 1;
 197                 goto out_dput;
 198         }
 199         spin_unlock(&dir->i_lock);
 200         ret = nfs_do_call_unlink(parent, dir, data);
 201 out_dput:
 202         dput(parent);
 203 out_free:
 204         return ret;
 205 }
 206
 207 void nfs_block_sillyrename(struct dentry *dentry)
 208 {
 209         struct nfs_inode *nfsi = NFS_I(dentry->d_inode);
 210
 211         wait_event(nfsi->waitqueue, atomic_cmpxchg(&nfsi->silly_count, 1, 0) == 1);
 212 }
 213
 214 void nfs_unblock_sillyrename(struct dentry *dentry)
 215 {
 216         struct inode *dir = dentry->d_inode;
 217         struct nfs_inode *nfsi = NFS_I(dir);
 218         struct nfs_unlinkdata *data;
 219
 220         atomic_inc(&nfsi->silly_count);
 221         spin_lock(&dir->i_lock);
 222         while (!hlist_empty(&nfsi->silly_list)) {
 223                 if (!atomic_inc_not_zero(&nfsi->silly_count))
 224                         break;
 225                 data = hlist_entry(nfsi->silly_list.first, struct nfs_unlinkdata, list);
 226                 hlist_del(&data->list);
 227                 spin_unlock(&dir->i_lock);
 228                 if (nfs_do_call_unlink(dentry, dir, data) == 0)
 229                         nfs_free_unlinkdata(data);
 230                 spin_lock(&dir->i_lock);
 231         }
 232         spin_unlock(&dir->i_lock);
 233 }
 234
 235 /**
 236  * nfs_async_unlink - asynchronous unlinking of a file
 237  * @dir: parent directory of dentry
 238  * @dentry: dentry to unlink
 239  */
 240 static int
 241 nfs_async_unlink(struct inode *dir, struct dentry *dentry)
 242 {
 243         struct nfs_unlinkdata *data;
 244         int status = -ENOMEM;
 245         void *devname_garbage = NULL;
 246
 247         data = kzalloc(sizeof(*data), GFP_KERNEL);
 248         if (data == NULL)
 249                 goto out;
 250
 251         data->cred = rpc_lookup_cred();
 252         if (IS_ERR(data->cred)) {
 253                 status = PTR_ERR(data->cred);
 254                 goto out_free;
 255         }
 256         data->res.dir_attr = &data->dir_attr;
 257
 258         status = -EBUSY;
 259         spin_lock(&dentry->d_lock);
 260         if (dentry->d_flags & DCACHE_NFSFS_RENAMED)
 261                 goto out_unlock;
 262         dentry->d_flags |= DCACHE_NFSFS_RENAMED;
 263         devname_garbage = dentry->d_fsdata;
 264         dentry->d_fsdata = data;
 265         spin_unlock(&dentry->d_lock);
 266         /*
 267          * If we'd displaced old cached devname, free it.  At that
 268          * point dentry is definitely not a root, so we won't need
 269          * that anymore.
 270          */
 271         if (devname_garbage)
 272                 kfree(devname_garbage);
 273         return 0;
 274 out_unlock:
 275         spin_unlock(&dentry->d_lock);
 276         put_rpccred(data->cred);
 277 out_free:
 278         kfree(data);
 279 out:
 280         return status;
 281 }
 282
 283 /**
 284  * nfs_complete_unlink - Initialize completion of the sillydelete
 285  * @dentry: dentry to delete
 286  * @inode: inode
 287  *
 288  * Since we're most likely to be called by dentry_iput(), we
 289  * only use the dentry to find the sillydelete. We then copy the name
 290  * into the qstr.
 291  */
 292 void
 293 nfs_complete_unlink(struct dentry *dentry, struct inode *inode)
 294 {
 295         struct nfs_unlinkdata   *data = NULL;
 296
 297         spin_lock(&dentry->d_lock);
 298         if (dentry->d_flags & DCACHE_NFSFS_RENAMED) {
 299                 dentry->d_flags &= ~DCACHE_NFSFS_RENAMED;
 300                 data = dentry->d_fsdata;
 301                 dentry->d_fsdata = NULL;
 302         }
 303         spin_unlock(&dentry->d_lock);
 304
 305         if (data != NULL && (NFS_STALE(inode) || !nfs_call_unlink(dentry, data)))
 306                 nfs_free_unlinkdata(data);
 307 }
 308
 309 /* Cancel a queued async unlink. Called when a sillyrename run fails. */
 310 static void
 311 nfs_cancel_async_unlink(struct dentry *dentry)
 312 {
 313         spin_lock(&dentry->d_lock);
 314         if (dentry->d_flags & DCACHE_NFSFS_RENAMED) {
 315                 struct nfs_unlinkdata *data = dentry->d_fsdata;
 316
 317                 dentry->d_flags &= ~DCACHE_NFSFS_RENAMED;
 318                 dentry->d_fsdata = NULL;
 319                 spin_unlock(&dentry->d_lock);
 320                 nfs_free_unlinkdata(data);
 321                 return;
 322         }
 323         spin_unlock(&dentry->d_lock);
 324 }
 325
 326 /**
 327  * nfs_async_rename_done - Sillyrename post-processing
 328  * @task: rpc_task of the sillyrename
 329  * @calldata: nfs_renamedata for the sillyrename
 330  *
 331  * Do the directory attribute updates and the d_move
 332  */
 333 static void nfs_async_rename_done(struct rpc_task *task, void *calldata)
 334 {
 335         struct nfs_renamedata *data = calldata;
 336         struct inode *old_dir = data->old_dir;
 337         struct inode *new_dir = data->new_dir;
 338         struct dentry *old_dentry = data->old_dentry;
 339         struct dentry *new_dentry = data->new_dentry;
 340
 341         if (!NFS_PROTO(old_dir)->rename_done(task, old_dir, new_dir)) {
 342                 rpc_restart_call_prepare(task);
 343                 return;
 344         }
 345
 346         if (task->tk_status != 0) {
 347                 nfs_cancel_async_unlink(old_dentry);
 348                 return;
 349         }
 350
 351         d_drop(old_dentry);
 352         d_drop(new_dentry);
 353 }
 354
 355 /**
 356  * nfs_async_rename_release - Release the sillyrename data.
 357  * @calldata: the struct nfs_renamedata to be released
 358  */
 359 static void nfs_async_rename_release(void *calldata)
 360 {
 361         struct nfs_renamedata   *data = calldata;
 362         struct super_block *sb = data->old_dir->i_sb;
 363
 364         if (data->old_dentry->d_inode)
 365                 nfs_mark_for_revalidate(data->old_dentry->d_inode);
 366
 367         dput(data->old_dentry);
 368         dput(data->new_dentry);
 369         iput(data->old_dir);
 370         iput(data->new_dir);
 371         nfs_sb_deactive(sb);
 372         put_rpccred(data->cred);
 373         kfree(data);
 374 }
 375
 376 static void nfs_rename_prepare(struct rpc_task *task, void *calldata)
 377 {
 378         struct nfs_renamedata *data = calldata;
 379         NFS_PROTO(data->old_dir)->rename_rpc_prepare(task, data);
 380 }
 381
 382 static const struct rpc_call_ops nfs_rename_ops = {
 383         .rpc_call_done = nfs_async_rename_done,
 384         .rpc_release = nfs_async_rename_release,
 385         .rpc_call_prepare = nfs_rename_prepare,
 386 };
 387
 388 /**
 389  * nfs_async_rename - perform an asynchronous rename operation
 390  * @old_dir: directory that currently holds the dentry to be renamed
 391  * @new_dir: target directory for the rename
 392  * @old_dentry: original dentry to be renamed
 393  * @new_dentry: dentry to which the old_dentry should be renamed
 394  *
 395  * It's expected that valid references to the dentries and inodes are held
 396  */
 397 static struct rpc_task *
 398 nfs_async_rename(struct inode *old_dir, struct inode *new_dir,
 399                  struct dentry *old_dentry, struct dentry *new_dentry)
 400 {
 401         struct nfs_renamedata *data;
 402         struct rpc_message msg = { };
 403         struct rpc_task_setup task_setup_data = {
 404                 .rpc_message = &msg,
 405                 .callback_ops = &nfs_rename_ops,
 406                 .workqueue = nfsiod_workqueue,
 407                 .rpc_client = NFS_CLIENT(old_dir),
 408                 .flags = RPC_TASK_ASYNC,
 409         };
 410
 411         data = kzalloc(sizeof(*data), GFP_KERNEL);
 412         if (data == NULL)
 413                 return ERR_PTR(-ENOMEM);
 414         task_setup_data.callback_data = data;
 415
 416         data->cred = rpc_lookup_cred();
 417         if (IS_ERR(data->cred)) {
 418                 struct rpc_task *task = ERR_CAST(data->cred);
 419                 kfree(data);
 420                 return task;
 421         }
 422
 423         msg.rpc_argp = &data->args;
 424         msg.rpc_resp = &data->res;
 425         msg.rpc_cred = data->cred;
 426
 427         /* set up nfs_renamedata */
 428         data->old_dir = old_dir;
 429         ihold(old_dir);
 430         data->new_dir = new_dir;
 431         ihold(new_dir);
 432         data->old_dentry = dget(old_dentry);
 433         data->new_dentry = dget(new_dentry);
 434         nfs_fattr_init(&data->old_fattr);
 435         nfs_fattr_init(&data->new_fattr);
 436
 437         /* set up nfs_renameargs */
 438         data->args.old_dir = NFS_FH(old_dir);
 439         data->args.old_name = &old_dentry->d_name;
 440         data->args.new_dir = NFS_FH(new_dir);
 441         data->args.new_name = &new_dentry->d_name;
 442
 443         /* set up nfs_renameres */
 444         data->res.old_fattr = &data->old_fattr;
 445         data->res.new_fattr = &data->new_fattr;
 446
 447         nfs_sb_active(old_dir->i_sb);
 448
 449         NFS_PROTO(data->old_dir)->rename_setup(&msg, old_dir);
 450
 451         return rpc_run_task(&task_setup_data);
 452 }
 453
 454 /**
 455  * nfs_sillyrename - Perform a silly-rename of a dentry
 456  * @dir: inode of directory that contains dentry
 457  * @dentry: dentry to be sillyrenamed
 458  *
 459  * NFSv2/3 is stateless and the server doesn't know when the client is
 460  * holding a file open. To prevent application problems when a file is
 461  * unlinked while it's still open, the client performs a "silly-rename".
 462  * That is, it renames the file to a hidden file in the same directory,
 463  * and only performs the unlink once the last reference to it is put.
 464  *
 465  * The final cleanup is done during dentry_iput.
 466  *
 467  * (Note: NFSv4 is stateful, and has opens, so in theory an NFSv4 server
 468  * could take responsibility for keeping open files referenced.  The server
 469  * would also need to ensure that opened-but-deleted files were kept over
 470  * reboots.  However, we may not assume a server does so.  (RFC 5661
 471  * does provide an OPEN4_RESULT_PRESERVE_UNLINKED flag that a server can
 472  * use to advertise that it does this; some day we may take advantage of
 473  * it.))
 474  */
 475 int
 476 nfs_sillyrename(struct inode *dir, struct dentry *dentry)
 477 {
 478         static unsigned int sillycounter;
 479         const int      fileidsize  = sizeof(NFS_FILEID(dentry->d_inode))*2;
 480         const int      countersize = sizeof(sillycounter)*2;
 481         const int      slen        = sizeof(".nfs")+fileidsize+countersize-1;
 482         char           silly[slen+1];
 483         struct dentry *sdentry;
 484         struct rpc_task *task;
 485         int            error = -EIO;
 486
 487         dfprintk(VFS, "NFS: silly-rename(%s/%s, ct=%d)\n",
 488                 dentry->d_parent->d_name.name, dentry->d_name.name,
 489                 dentry->d_count);
 490         nfs_inc_stats(dir, NFSIOS_SILLYRENAME);
 491
 492         /*
 493          * We don't allow a dentry to be silly-renamed twice.
 494          */
 495         error = -EBUSY;
 496         if (dentry->d_flags & DCACHE_NFSFS_RENAMED)
 497                 goto out;
 498
 499         sprintf(silly, ".nfs%*.*Lx",
 500                 fileidsize, fileidsize,
 501                 (unsigned long long)NFS_FILEID(dentry->d_inode));
 502
 503         /* Return delegation in anticipation of the rename */
 504         NFS_PROTO(dentry->d_inode)->return_delegation(dentry->d_inode);
 505
 506         sdentry = NULL;
 507         do {
 508                 char *suffix = silly + slen - countersize;
 509
 510                 dput(sdentry);
 511                 sillycounter++;
 512                 sprintf(suffix, "%*.*x", countersize, countersize, sillycounter);
 513
 514                 dfprintk(VFS, "NFS: trying to rename %s to %s\n",
 515                                 dentry->d_name.name, silly);
 516
 517                 sdentry = lookup_one_len(silly, dentry->d_parent, slen);
 518                 /*
 519                  * N.B. Better to return EBUSY here ... it could be
 520                  * dangerous to delete the file while it's in use.
 521                  */
 522                 if (IS_ERR(sdentry))
 523                         goto out;
 524         } while (sdentry->d_inode != NULL); /* need negative lookup */
 525
 526         /* queue unlink first. Can't do this from rpc_release as it
 527          * has to allocate memory
 528          */
 529         error = nfs_async_unlink(dir, dentry);
 530         if (error)
 531                 goto out_dput;
 532
 533         /* populate unlinkdata with the right dname */
 534         error = nfs_copy_dname(sdentry,
 535                                 (struct nfs_unlinkdata *)dentry->d_fsdata);
 536         if (error) {
 537                 nfs_cancel_async_unlink(dentry);
 538                 goto out_dput;
 539         }
 540
 541         /* run the rename task, undo unlink if it fails */
 542         task = nfs_async_rename(dir, dir, dentry, sdentry);
 543         if (IS_ERR(task)) {
 544                 error = -EBUSY;
 545                 nfs_cancel_async_unlink(dentry);
 546                 goto out_dput;
 547         }
 548
 549         /* wait for the RPC task to complete, unless a SIGKILL intervenes */
 550         error = rpc_wait_for_completion_task(task);
 551         if (error == 0)
 552                 error = task->tk_status;
 553         rpc_put_task(task);
 554 out_dput:
 555         dput(sdentry);
 556 out:
 557         return error;
 558 }