lustre/llite/llite_gns.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2004, 2005 Cluster File Systems, Inc.
   5  *
   6  * Author: Phil Schwan <phil@clusterfs.com>
   7  * Author: Oleg Drokin <green@clusterfs.com>
   8  * Author: Yury Umanets <yury@clusterfs.com>
   9  * Review: Nikita Danilov <nikita@clusterfs.com>
  10  *
  11  *   This file is part of Lustre, http://www.lustre.org.
  12  *
  13  *   Lustre is free software; you can redistribute it and/or
  14  *   modify it under the terms of version 2 of the GNU General Public
  15  *   License as published by the Free Software Foundation.
  16  *
  17  *   Lustre is distributed in the hope that it will be useful,
  18  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  *   GNU General Public License for more details.
  21  *
  22  *   You should have received a copy of the GNU General Public License
  23  *   along with Lustre; if not, write to the Free Software
  24  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  25  */
  26
  27 #define DEBUG_SUBSYSTEM S_LLITE
  28
  29 #include <linux/fs.h>
  30 #include <linux/version.h>
  31 #include <asm/uaccess.h>
  32 #include <linux/file.h>
  33 #include <linux/kmod.h>
  34
  35 #include <linux/lustre_lite.h>
  36 #include "llite_internal.h"
  37
  38 static struct list_head gns_sbi_list = LIST_HEAD_INIT(gns_sbi_list);
  39 static spinlock_t gns_lock = SPIN_LOCK_UNLOCKED;
  40 static struct ptlrpc_thread gns_thread;
  41 static struct ll_gns_ctl gns_ctl;
  42
  43 /*
  44  * waits until passed dentry gets mountpoint or timeout and attempts are
  45  * exhausted. Returns 1 if dentry became mountpoint and 0 otherwise.
  46  */
  47 static int
  48 ll_gns_wait_for_mount(struct dentry *dentry,
  49                       int timeout, int tries)
  50 {
  51         struct l_wait_info lwi;
  52         struct ll_sb_info *sbi;
  53         int rc = 0;
  54         ENTRY;
  55
  56         LASSERT(dentry != NULL);
  57         LASSERT(!IS_ERR(dentry));
  58         sbi = ll_s2sbi(dentry->d_sb);
  59
  60         lwi = LWI_TIMEOUT(timeout * HZ, NULL, NULL);
  61         for (; !d_mountpoint(dentry) && tries > 0; tries--) {
  62                 l_wait_event(sbi->ll_gns_waitq, d_mountpoint(dentry), &lwi);
  63                 if (signal_pending(current))
  64                         GOTO(out, rc = -EINTR);
  65         }
  66
  67         if (!d_mountpoint(dentry))
  68                 rc = -ETIME;
  69
  70         EXIT;
  71 out:
  72         spin_lock(&sbi->ll_gns_lock);
  73         sbi->ll_gns_state = LL_GNS_FINISHED;
  74         spin_unlock(&sbi->ll_gns_lock);
  75         return rc;
  76 }
  77
  78 /*
  79  * tries to mount the mount object under passed @dentry. In the case of success
  80  * @dentry will become mount point and 0 will be returned. Error code will be
  81  * returned otherwise.
  82  */
  83 int
  84 ll_gns_mount_object(struct dentry *dentry, struct vfsmount *mnt)
  85 {
  86         char *path, *pathpage, *datapage = NULL, *argv[4];
  87         struct file *mntinfo_fd = NULL;
  88         int cleanup_phase = 0, rc = 0;
  89         struct ll_sb_info *sbi;
  90         struct dentry *dchild = NULL;
  91         ENTRY;
  92
  93         LASSERT(dentry->d_inode != NULL);
  94
  95         if (!S_ISDIR(dentry->d_inode->i_mode))
  96                 RETURN(-EINVAL);
  97
  98         sbi = ll_i2sbi(dentry->d_inode);
  99
 100         if (mnt == NULL) {
 101                 CERROR("suid directory found, but no "
 102                        "vfsmount available.\n");
 103                 RETURN(-EINVAL);
 104         }
 105
 106         if (atomic_read(&sbi->ll_gns_enabled) == 0)
 107                 RETURN(-EINVAL);
 108
 109         spin_lock(&sbi->ll_gns_lock);
 110
 111         /*
 112          * another thead is in progress or just finished mounting the
 113          * dentry. Handling that.
 114          */
 115         if (sbi->ll_gns_state != LL_GNS_IDLE) {
 116                 /*
 117                  * another thread is trying to mount GNS dentry. We'd like to
 118                  * handling that.
 119                  */
 120                 spin_unlock(&sbi->ll_gns_lock);
 121
 122         restart:
 123                 /*
 124                  * check if dentry is mount point already, if so, do not restart
 125                  * syscal.
 126                  */
 127                 if (d_mountpoint(dentry))
 128                         RETURN(0);
 129
 130                 spin_lock(&sbi->ll_gns_lock);
 131                 if (sbi->ll_gns_pending_dentry &&
 132                     is_subdir(sbi->ll_gns_pending_dentry, dentry)) {
 133                         spin_unlock(&sbi->ll_gns_lock);
 134                         RETURN(-EAGAIN);
 135                 }
 136                 spin_unlock(&sbi->ll_gns_lock);
 137
 138                 /*
 139                  * waiting for GNS complete and check dentry again, it may be
 140                  * mounted already.
 141                  */
 142                 wait_for_completion(&sbi->ll_gns_mount_finished);
 143                 if (d_mountpoint(dentry))
 144                         RETURN(0);
 145
 146                 /*
 147                  * check for he case when there are few waiters and all they are
 148                  * awakened, but only one will find GNS state LL_GNS_IDLE, and
 149                  * the rest will face with LL_GNS_MOUNTING.  --umka
 150                  */
 151                 spin_lock(&sbi->ll_gns_lock);
 152                 if (sbi->ll_gns_state != LL_GNS_IDLE) {
 153                         spin_unlock(&sbi->ll_gns_lock);
 154                         goto restart;
 155                 }
 156                 spin_unlock(&sbi->ll_gns_lock);
 157         }
 158         LASSERT(sbi->ll_gns_state == LL_GNS_IDLE);
 159         CDEBUG(D_INODE, "mounting dentry %p\n", dentry);
 160
 161         /* mounting started */
 162         sbi->ll_gns_state = LL_GNS_MOUNTING;
 163         sbi->ll_gns_pending_dentry = dentry;
 164         spin_unlock(&sbi->ll_gns_lock);
 165
 166         /* we need to build an absolute pathname to pass to mount */
 167         pathpage = (char *)__get_free_page(GFP_KERNEL);
 168         if (!pathpage)
 169                 GOTO(cleanup, rc = -ENOMEM);
 170         cleanup_phase = 1;
 171
 172         /* getting @dentry path stored in @pathpage. */
 173         path = d_path(dentry, mnt, pathpage, PAGE_SIZE);
 174         if (IS_ERR(path)) {
 175                 CERROR("can't build mount object path, err %d\n",
 176                        (int)PTR_ERR(path));
 177                 GOTO(cleanup, rc = PTR_ERR(path));
 178         }
 179
 180         /* synchronizing with possible /proc/fs/...write */
 181         down(&sbi->ll_gns_sem);
 182
 183         /*
 184          * mount object name is taken from sbi, where it is set in mount time or
 185          * via /proc/fs... tunable. It may be ".mntinfo" or so.
 186          */
 187
 188         /*
 189          * recursive lookup with trying to mount SUID bit marked directories on
 190          * the way is not possible here, as lookup_one_len() does not pass @nd
 191          * to ->lookup() and this is checked in ll_lookup_it().
 192          */
 193         dchild = ll_lookup_one_len(sbi->ll_gns_oname, dentry,
 194                                    strlen(sbi->ll_gns_oname));
 195         up(&sbi->ll_gns_sem);
 196
 197         if (IS_ERR(dchild)) {
 198                 rc = PTR_ERR(dchild);
 199                 CERROR("can't find mount object %*s/%*s err = %d.\n",
 200                        (int)dentry->d_name.len, dentry->d_name.name,
 201                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 202                        rc);
 203                 GOTO(cleanup, rc);
 204         }
 205
 206         /* mount object is not found */
 207         if (!dchild->d_inode) {
 208                 dput(dchild);
 209                 GOTO(cleanup, rc = -ENOENT);
 210         }
 211
 212         /* check if found child is regular file */
 213         if (!S_ISREG(dchild->d_inode->i_mode)) {
 214                 dput(dchild);
 215                 GOTO(cleanup, rc = -EBADF);
 216         }
 217
 218         /* ok, mount object if found, opening it. */
 219         mntinfo_fd = dentry_open(dchild, mntget(mnt), 0);
 220         if (IS_ERR(mntinfo_fd)) {
 221                 CERROR("can't open mount object %*s/%*s err = %d.\n",
 222                        (int)dentry->d_name.len, dentry->d_name.name,
 223                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 224                        (int)PTR_ERR(mntinfo_fd));
 225                 mntput(mnt);
 226                 dput(dchild);
 227                 GOTO(cleanup, rc = PTR_ERR(mntinfo_fd));
 228         }
 229         cleanup_phase = 2;
 230
 231         if (mntinfo_fd->f_dentry->d_inode->i_size > PAGE_SIZE - 1) {
 232                 CERROR("mount object %*s/%*s is too big (%Ld)\n",
 233                        (int)dentry->d_name.len, dentry->d_name.name,
 234                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 235                        mntinfo_fd->f_dentry->d_inode->i_size);
 236                 GOTO(cleanup, rc = -EFBIG);
 237         }
 238
 239         datapage = (char *)__get_free_page(GFP_KERNEL);
 240         if (!datapage)
 241                 GOTO(cleanup, rc = -ENOMEM);
 242
 243         cleanup_phase = 3;
 244
 245         /* read data from mount object. */
 246         rc = kernel_read(mntinfo_fd, 0, datapage, PAGE_SIZE - 1);
 247         if (rc < 0) {
 248                 CERROR("can't read mount object %*s/%*s data, err %d\n",
 249                        (int)dentry->d_name.len, dentry->d_name.name,
 250                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 251                        rc);
 252                 GOTO(cleanup, rc);
 253         }
 254
 255         /* no data in mount object? */
 256         if (rc == 0) {
 257                 CERROR("mount object %*s/%*s is empty?\n",
 258                        (int)dentry->d_name.len, dentry->d_name.name,
 259                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname);
 260                 GOTO(cleanup, rc);
 261         }
 262
 263         datapage[rc] = '\0';
 264         fput(mntinfo_fd);
 265         mntinfo_fd = NULL;
 266         dchild = NULL;
 267
 268         /* synchronizing with possible /proc/fs/...write */
 269         down(&sbi->ll_gns_sem);
 270
 271         /*
 272          * upcall is initialized in mount time or via /proc/fs/... tuneable and
 273          * may be /usr/lib/lustre/gns-upcall.sh
 274          */
 275         argv[0] = sbi->ll_gns_upcall;
 276         argv[1] = datapage;
 277         argv[2] = path;
 278         argv[3] = NULL;
 279
 280         up(&sbi->ll_gns_sem);
 281
 282         /* do not wait for helper complete here. */
 283         rc = call_usermodehelper(argv[0], argv, NULL, 1);
 284         if (rc) {
 285                 CWARN("failed to call GNS upcall %s, err = %d, "
 286                       "checking for mount anyway\n", sbi->ll_gns_upcall, rc);
 287         }
 288
 289         /*
 290          * waiting for dentry become mount point GNS_WAIT_ATTEMPTS times by 1
 291          * second.
 292          */
 293         rc = ll_gns_wait_for_mount(dentry, 1, GNS_WAIT_ATTEMPTS);
 294         LASSERT(sbi->ll_gns_state == LL_GNS_FINISHED);
 295
 296         /* checking for mount point anyway to not loss mounts */
 297         if (d_mountpoint(dentry)) {
 298                 struct dentry *rdentry;
 299                 struct vfsmount *rmnt;
 300
 301                 rmnt = mntget(mnt);
 302                 rdentry = dget(dentry);
 303
 304                 if (follow_down(&rmnt, &rdentry)) {
 305                         /*
 306                          * registering new mount in GNS mounts list and thus
 307                          * make it accessible from GNS control thread.
 308                          */
 309                         spin_lock(&dcache_lock);
 310                         LASSERT(list_empty(&rmnt->mnt_lustre_list));
 311                         list_add_tail(&rmnt->mnt_lustre_list,
 312                                       &sbi->ll_mnt_list);
 313                         spin_unlock(&dcache_lock);
 314                         rmnt->mnt_last_used = jiffies;
 315                         mntput(rmnt);
 316                         dput(rdentry);
 317                 } else {
 318                         mntput(mnt);
 319                         dput(dentry);
 320                 }
 321
 322                 rc = 0;
 323         } else {
 324                 CERROR("usermode upcall %s failed to mount %s, err %d\n",
 325                        sbi->ll_gns_upcall, path, rc);
 326         }
 327
 328         EXIT;
 329 cleanup:
 330         switch (cleanup_phase) {
 331         case 3:
 332                 free_page((unsigned long)datapage);
 333         case 2:
 334                 if (mntinfo_fd != NULL) {
 335                         fput(mntinfo_fd);
 336                         dchild = NULL;
 337                 }
 338         case 1:
 339                 free_page((unsigned long)pathpage);
 340         case 0:
 341                 spin_lock(&sbi->ll_gns_lock);
 342                 sbi->ll_gns_state = LL_GNS_IDLE;
 343                 sbi->ll_gns_pending_dentry = NULL;
 344                 spin_unlock(&sbi->ll_gns_lock);
 345
 346                 /* waking up all waiters after GNS state is LL_GNS_IDLE */
 347                 complete_all(&sbi->ll_gns_mount_finished);
 348                 init_completion(&sbi->ll_gns_mount_finished);
 349         }
 350         return rc;
 351 }
 352
 353 /* tries to umount passed @mnt. */
 354 int ll_gns_umount_object(struct vfsmount *mnt)
 355 {
 356         int rc = 0;
 357         ENTRY;
 358
 359         CDEBUG(D_INODE, "unmounting mnt %p\n", mnt);
 360         rc = do_umount(mnt, 0);
 361         if (rc) {
 362                 CDEBUG(D_INODE, "can't umount 0x%p, err = %d\n",
 363                        mnt, rc);
 364         }
 365
 366         RETURN(rc);
 367 }
 368
 369 int ll_gns_check_mounts(struct ll_sb_info *sbi, int flags)
 370 {
 371         struct list_head check_list = LIST_HEAD_INIT(check_list);
 372         struct vfsmount *mnt;
 373         unsigned long pass;
 374         ENTRY;
 375
 376         spin_lock(&dcache_lock);
 377         list_splice_init(&sbi->ll_mnt_list, &check_list);
 378
 379         /*
 380          * walk the list in reverse order, and put them on the front of the sbi
 381          * list each iteration; this avoids list-ordering problems if we race
 382          * with another gns-mounting thread.
 383          */
 384         while (!list_empty(&check_list)) {
 385                 mnt = list_entry(check_list.prev,
 386                                  struct vfsmount,
 387                                  mnt_lustre_list);
 388
 389                 mntget(mnt);
 390
 391                 list_del_init(&mnt->mnt_lustre_list);
 392
 393                 list_add(&mnt->mnt_lustre_list,
 394                          &sbi->ll_mnt_list);
 395
 396                 /* check for timeout if needed */
 397                 pass = jiffies - mnt->mnt_last_used;
 398
 399                 if (flags == LL_GNS_CHECK &&
 400                     pass < sbi->ll_gns_timeout * HZ)
 401                 {
 402                         mntput(mnt);
 403                         continue;
 404                 }
 405                 spin_unlock(&dcache_lock);
 406
 407                 /* umounting @mnt */
 408                 ll_gns_umount_object(mnt);
 409
 410                 mntput(mnt);
 411                 spin_lock(&dcache_lock);
 412         }
 413         spin_unlock(&dcache_lock);
 414         RETURN(0);
 415 }
 416
 417 /*
 418  * GNS timer callback function. It restarts gns timer and wakes up GNS control
 419  * thread to process mounts list.
 420  */
 421 void ll_gns_timer_callback(unsigned long data)
 422 {
 423         struct ll_sb_info *sbi = (void *)data;
 424         ENTRY;
 425
 426         spin_lock(&gns_lock);
 427         if (list_empty(&sbi->ll_gns_sbi_head))
 428                 list_add(&sbi->ll_gns_sbi_head, &gns_sbi_list);
 429         spin_unlock(&gns_lock);
 430
 431         wake_up(&gns_thread.t_ctl_waitq);
 432         mod_timer(&sbi->ll_gns_timer,
 433                   jiffies + sbi->ll_gns_tick * HZ);
 434 }
 435
 436 /* this function checks if something new happened to exist in gns list. */
 437 static int inline ll_gns_check_event(void)
 438 {
 439         int rc;
 440
 441         spin_lock(&gns_lock);
 442         rc = !list_empty(&gns_sbi_list);
 443         spin_unlock(&gns_lock);
 444
 445         return rc;
 446 }
 447
 448 /* should we stop GNS control thread? */
 449 static int inline ll_gns_check_stop(void)
 450 {
 451         mb();
 452         return (gns_thread.t_flags & SVC_STOPPING) ? 1 : 0;
 453 }
 454
 455 /* GNS control thread function. */
 456 static int ll_gns_thread(void *arg)
 457 {
 458         struct ll_gns_ctl *ctl = arg;
 459         unsigned long flags;
 460         ENTRY;
 461
 462         {
 463                 char name[sizeof(current->comm)];
 464                 snprintf(name, sizeof(name) - 1, "ll_gns");
 465                 kportal_daemonize(name);
 466         }
 467
 468         SIGNAL_MASK_LOCK(current, flags);
 469         sigfillset(&current->blocked);
 470         RECALC_SIGPENDING;
 471         SIGNAL_MASK_UNLOCK(current, flags);
 472
 473         /*
 474          * letting starting function know, that we are ready and control may be
 475          * returned.
 476          */
 477         gns_thread.t_flags = SVC_RUNNING;
 478         complete(&ctl->gc_starting);
 479
 480         while (!ll_gns_check_stop()) {
 481                 struct l_wait_info lwi = { 0 };
 482
 483                 l_wait_event(gns_thread.t_ctl_waitq,
 484                              (ll_gns_check_event() ||
 485                               ll_gns_check_stop()), &lwi);
 486
 487                 spin_lock(&gns_lock);
 488                 while (!list_empty(&gns_sbi_list)) {
 489                         struct ll_sb_info *sbi;
 490
 491                         sbi = list_entry(gns_sbi_list.prev,
 492                                          struct ll_sb_info,
 493                                          ll_gns_sbi_head);
 494
 495                         list_del_init(&sbi->ll_gns_sbi_head);
 496                         spin_unlock(&gns_lock);
 497                         ll_gns_check_mounts(sbi, LL_GNS_CHECK);
 498                         spin_lock(&gns_lock);
 499                 }
 500                 spin_unlock(&gns_lock);
 501         }
 502
 503         EXIT;
 504         gns_thread.t_flags = SVC_STOPPED;
 505
 506         /* this is SMP-safe way to finish thread. */
 507         complete_and_exit(&ctl->gc_finishing, 0);
 508 }
 509
 510 void ll_gns_add_timer(struct ll_sb_info *sbi)
 511 {
 512         mod_timer(&sbi->ll_gns_timer,
 513                   jiffies + sbi->ll_gns_tick * HZ);
 514 }
 515
 516 void ll_gns_del_timer(struct ll_sb_info *sbi)
 517 {
 518         del_timer(&sbi->ll_gns_timer);
 519 }
 520
 521 /*
 522  * starts GNS control thread and waits for a signal it is up and work may be
 523  * continued.
 524  */
 525 int ll_gns_thread_start(void)
 526 {
 527         int rc;
 528         ENTRY;
 529
 530         LASSERT(gns_thread.t_flags == 0);
 531         init_completion(&gns_ctl.gc_starting);
 532         init_completion(&gns_ctl.gc_finishing);
 533         init_waitqueue_head(&gns_thread.t_ctl_waitq);
 534
 535         rc = kernel_thread(ll_gns_thread, &gns_ctl,
 536                            (CLONE_VM | CLONE_FILES));
 537         if (rc < 0) {
 538                 CERROR("cannot start GNS control thread, "
 539                        "err = %d\n", rc);
 540                 RETURN(rc);
 541         }
 542         wait_for_completion(&gns_ctl.gc_starting);
 543         LASSERT(gns_thread.t_flags == SVC_RUNNING);
 544         RETURN(0);
 545 }
 546
 547 /* stops GNS control thread and waits its actual stop. */
 548 void ll_gns_thread_stop(void)
 549 {
 550         ENTRY;
 551         gns_thread.t_flags = SVC_STOPPING;
 552         wake_up(&gns_thread.t_ctl_waitq);
 553         wait_for_completion(&gns_ctl.gc_finishing);
 554         LASSERT(gns_thread.t_flags == SVC_STOPPED);
 555         gns_thread.t_flags = 0;
 556         EXIT;
 557 }