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         ENTRY;
  54
  55         LASSERT(dentry != NULL);
  56         LASSERT(!IS_ERR(dentry));
  57         sbi = ll_s2sbi(dentry->d_sb);
  58
  59         lwi = LWI_TIMEOUT(timeout * HZ, NULL, NULL);
  60         for (; !d_mountpoint(dentry) && tries > 0; tries--)
  61                 l_wait_event(sbi->ll_gns_waitq, d_mountpoint(dentry), &lwi);
  62
  63         if (d_mountpoint(dentry)) {
  64                 spin_lock(&sbi->ll_gns_lock);
  65                 sbi->ll_gns_state = LL_GNS_FINISHED;
  66                 spin_unlock(&sbi->ll_gns_lock);
  67                 RETURN(0);
  68         }
  69         RETURN(-ETIME);
  70 }
  71
  72 /*
  73  * sending a signal known to be ignored to cause restarting syscall if GNS mount
  74  * function returns -ERESTARTSYS.
  75  */
  76 static void
  77 ll_gns_send_signal(void)
  78 {
  79         struct task_struct *task = current;
  80         int signal = SIGCONT;
  81
  82         read_lock(&tasklist_lock);
  83         spin_lock_irq(&task->sighand->siglock);
  84         sigaddset(&task->pending.signal, signal);
  85         spin_unlock_irq(&task->sighand->siglock);
  86         read_unlock(&tasklist_lock);
  87         set_tsk_thread_flag(task, TIF_SIGPENDING);
  88 }
  89
  90 /*
  91  * tries to mount the mount object under passed @dentry. In the case of success
  92  * @dentry will become mount point and 0 will be returned. Error code will be
  93  * returned otherwise.
  94  */
  95 int
  96 ll_gns_mount_object(struct dentry *dentry, struct vfsmount *mnt)
  97 {
  98         char *path, *pathpage, *datapage, *argv[4];
  99         struct file *mntinfo_fd = NULL;
 100         int cleanup_phase = 0, rc = 0;
 101         struct ll_sb_info *sbi;
 102         struct dentry *dchild;
 103         ENTRY;
 104
 105         LASSERT(dentry->d_inode != NULL);
 106
 107         if (!S_ISDIR(dentry->d_inode->i_mode))
 108                 RETURN(-EINVAL);
 109
 110         sbi = ll_i2sbi(dentry->d_inode);
 111         LASSERT(sbi != NULL);
 112
 113         spin_lock(&sbi->ll_gns_lock);
 114
 115         if (sbi->ll_gns_state == LL_GNS_DISABLED) {
 116                 spin_unlock(&sbi->ll_gns_lock);
 117                 RETURN(-EINVAL);
 118         }
 119
 120         if (mnt == NULL) {
 121                 CERROR("suid directory found, but no "
 122                        "vfsmount available.\n");
 123                 RETURN(-EINVAL);
 124         }
 125
 126         /*
 127          * another thead is in progress or just finished mounting the
 128          * dentry. Handling that.
 129          */
 130         if (sbi->ll_gns_state == LL_GNS_MOUNTING ||
 131             sbi->ll_gns_state == LL_GNS_FINISHED) {
 132                 /*
 133                  * another thread is trying to mount GNS dentry. We'd like to
 134                  * handling that.
 135                  */
 136                 spin_unlock(&sbi->ll_gns_lock);
 137
 138                 /*
 139                  * check if dentry is mount point already, if so, do not restart
 140                  * syscal.
 141                  */
 142                 if (d_mountpoint(dentry))
 143                         RETURN(0);
 144
 145                 /*
 146                  * causing syscall to restart and find this dentry already
 147                  * mounted.
 148                  */
 149                 ll_gns_send_signal();
 150                 RETURN(-ERESTARTSYS);
 151
 152 #if 0
 153                 wait_for_completion(&sbi->ll_gns_mount_finished);
 154                 if (d_mountpoint(dentry))
 155                         RETURN(0);
 156 #endif
 157         }
 158         LASSERT(sbi->ll_gns_state == LL_GNS_IDLE);
 159
 160         CDEBUG(D_INODE, "mounting dentry %p\n", dentry);
 161
 162         /* mounting started */
 163         sbi->ll_gns_state = LL_GNS_MOUNTING;
 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(dchild));
 177                 GOTO(cleanup, rc = PTR_ERR(dchild));
 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(). So, do not
 192          * handle possible -EAGAIN here.
 193          */
 194         dchild = ll_lookup_one_len(sbi->ll_gns_oname, dentry,
 195                                    strlen(sbi->ll_gns_oname));
 196         up(&sbi->ll_gns_sem);
 197
 198         cleanup_phase = 2;
 199
 200         if (IS_ERR(dchild)) {
 201                 rc = PTR_ERR(dchild);
 202                 CERROR("can't find mount object %*s/%*s err = %d.\n",
 203                        (int)dentry->d_name.len, dentry->d_name.name,
 204                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 205                        rc);
 206                 GOTO(cleanup, rc);
 207         }
 208
 209         /* mount object is not found */
 210         if (!dchild->d_inode)
 211                 GOTO(cleanup, rc = -ENOENT);
 212
 213         /* check if found child is regular file */
 214         if (!S_ISREG(dchild->d_inode->i_mode))
 215                 GOTO(cleanup, rc = -EOPNOTSUPP);
 216
 217         mntget(mnt);
 218
 219         /* ok, mount object if found, opening it. */
 220         mntinfo_fd = dentry_open(dchild, mnt, 0);
 221         if (IS_ERR(mntinfo_fd)) {
 222                 CERROR("can't open mount object %*s/%*s err = %d.\n",
 223                        (int)dentry->d_name.len, dentry->d_name.name,
 224                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 225                        (int)PTR_ERR(mntinfo_fd));
 226                 mntput(mnt);
 227                 GOTO(cleanup, rc = PTR_ERR(mntinfo_fd));
 228         }
 229         cleanup_phase = 3;
 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 = 4;
 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, 0);
 284         if (rc) {
 285                 CERROR("failed to call GNS upcall %s, err = %d\n",
 286                        sbi->ll_gns_upcall, rc);
 287                 GOTO(cleanup, rc);
 288         }
 289
 290         /*
 291          * waiting for dentry become mount point GNS_WAIT_ATTEMPTS times by 1
 292          * second.
 293          */
 294         rc = ll_gns_wait_for_mount(dentry, 1, GNS_WAIT_ATTEMPTS);
 295         complete_all(&sbi->ll_gns_mount_finished);
 296         if (rc == 0) {
 297                 struct dentry *rdentry;
 298                 struct vfsmount *rmnt;
 299
 300                 /* mount is successful */
 301                 LASSERT(sbi->ll_gns_state == LL_GNS_FINISHED);
 302
 303                 rmnt = mntget(mnt);
 304                 rdentry = dget(dentry);
 305
 306                 if (follow_down(&rmnt, &rdentry)) {
 307                         /*
 308                          * registering new mount in GNS mounts list and thus
 309                          * make it accessible from GNS control thread.
 310                          */
 311                         spin_lock(&dcache_lock);
 312                         LASSERT(list_empty(&rmnt->mnt_lustre_list));
 313                         list_add_tail(&rmnt->mnt_lustre_list,
 314                                       &sbi->ll_mnt_list);
 315                         spin_unlock(&dcache_lock);
 316                         rmnt->mnt_last_used = jiffies;
 317                         mntput(rmnt);
 318                         dput(rdentry);
 319                 } else {
 320                         mntput(mnt);
 321                         dput(dentry);
 322                 }
 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 4:
 332                 free_page((unsigned long)datapage);
 333         case 3:
 334                 if (mntinfo_fd != NULL)
 335                         fput(mntinfo_fd);
 336         case 2:
 337                 if (dchild != NULL)
 338                         dput(dchild);
 339         case 1:
 340                 free_page((unsigned long)pathpage);
 341                 complete_all(&sbi->ll_gns_mount_finished);
 342         case 0:
 343                 spin_lock(&sbi->ll_gns_lock);
 344                 sbi->ll_gns_state = LL_GNS_IDLE;
 345                 spin_unlock(&sbi->ll_gns_lock);
 346         }
 347         return rc;
 348 }
 349
 350 /* tries to umount passed @mnt. */
 351 int ll_gns_umount_object(struct vfsmount *mnt)
 352 {
 353         int rc = 0;
 354         ENTRY;
 355
 356         CDEBUG(D_INODE, "unmounting mnt %p\n", mnt);
 357         rc = do_umount(mnt, 0);
 358         if (rc) {
 359                 CDEBUG(D_INODE, "can't umount 0x%p, err = %d\n",
 360                        mnt, rc);
 361         }
 362
 363         RETURN(rc);
 364 }
 365
 366 int ll_gns_check_mounts(struct ll_sb_info *sbi, int flags)
 367 {
 368         struct list_head check_list = LIST_HEAD_INIT(check_list);
 369         struct vfsmount *mnt;
 370         unsigned long pass;
 371         ENTRY;
 372
 373         spin_lock(&dcache_lock);
 374         list_splice_init(&sbi->ll_mnt_list, &check_list);
 375
 376         /*
 377          * walk the list in reverse order, and put them on the front of the sbi
 378          * list each iteration; this avoids list-ordering problems if we race
 379          * with another gns-mounting thread.
 380          */
 381         while (!list_empty(&check_list)) {
 382                 mnt = list_entry(check_list.prev,
 383                                  struct vfsmount,
 384                                  mnt_lustre_list);
 385
 386                 mntget(mnt);
 387
 388                 list_del_init(&mnt->mnt_lustre_list);
 389
 390                 list_add(&mnt->mnt_lustre_list,
 391                          &sbi->ll_mnt_list);
 392
 393                 /* check for timeout if needed */
 394                 pass = jiffies - mnt->mnt_last_used;
 395
 396                 if (flags == LL_GNS_CHECK &&
 397                     pass < sbi->ll_gns_timeout * HZ)
 398                 {
 399                         mntput(mnt);
 400                         continue;
 401                 }
 402                 spin_unlock(&dcache_lock);
 403
 404                 /* umounting @mnt */
 405                 ll_gns_umount_object(mnt);
 406
 407                 mntput(mnt);
 408                 spin_lock(&dcache_lock);
 409         }
 410         spin_unlock(&dcache_lock);
 411         RETURN(0);
 412 }
 413
 414 /*
 415  * GNS timer callback function. It restarts gns timer and wakes up GNS control
 416  * thread to process mounts list.
 417  */
 418 void ll_gns_timer_callback(unsigned long data)
 419 {
 420         struct ll_sb_info *sbi = (void *)data;
 421         ENTRY;
 422
 423         spin_lock(&gns_lock);
 424         if (list_empty(&sbi->ll_gns_sbi_head))
 425                 list_add(&sbi->ll_gns_sbi_head, &gns_sbi_list);
 426         spin_unlock(&gns_lock);
 427
 428         wake_up(&gns_thread.t_ctl_waitq);
 429         mod_timer(&sbi->ll_gns_timer,
 430                   jiffies + sbi->ll_gns_tick * HZ);
 431 }
 432
 433 /* this function checks if something new happened to exist in gns list. */
 434 static int inline ll_gns_check_event(void)
 435 {
 436         int rc;
 437
 438         spin_lock(&gns_lock);
 439         rc = !list_empty(&gns_sbi_list);
 440         spin_unlock(&gns_lock);
 441
 442         return rc;
 443 }
 444
 445 /* should we stop GNS control thread? */
 446 static int inline ll_gns_check_stop(void)
 447 {
 448         mb();
 449         return (gns_thread.t_flags & SVC_STOPPING) ? 1 : 0;
 450 }
 451
 452 /* GNS control thread function. */
 453 static int ll_gns_thread_main(void *arg)
 454 {
 455         struct ll_gns_ctl *ctl = arg;
 456         unsigned long flags;
 457         ENTRY;
 458
 459         {
 460                 char name[sizeof(current->comm)];
 461                 snprintf(name, sizeof(name) - 1, "ll_gns");
 462                 kportal_daemonize(name);
 463         }
 464
 465         SIGNAL_MASK_LOCK(current, flags);
 466         sigfillset(&current->blocked);
 467         RECALC_SIGPENDING;
 468         SIGNAL_MASK_UNLOCK(current, flags);
 469
 470         /*
 471          * letting starting function know, that we are ready and control may be
 472          * returned.
 473          */
 474         gns_thread.t_flags = SVC_RUNNING;
 475         complete(&ctl->gc_starting);
 476
 477         while (!ll_gns_check_stop()) {
 478                 struct l_wait_info lwi = { 0 };
 479
 480                 l_wait_event(gns_thread.t_ctl_waitq,
 481                              (ll_gns_check_event() ||
 482                               ll_gns_check_stop()), &lwi);
 483
 484                 spin_lock(&gns_lock);
 485                 while (!list_empty(&gns_sbi_list)) {
 486                         struct ll_sb_info *sbi;
 487
 488                         sbi = list_entry(gns_sbi_list.prev,
 489                                          struct ll_sb_info,
 490                                          ll_gns_sbi_head);
 491
 492                         list_del_init(&sbi->ll_gns_sbi_head);
 493                         spin_unlock(&gns_lock);
 494                         ll_gns_check_mounts(sbi, LL_GNS_CHECK);
 495                         spin_lock(&gns_lock);
 496                 }
 497                 spin_unlock(&gns_lock);
 498         }
 499
 500         EXIT;
 501         gns_thread.t_flags = SVC_STOPPED;
 502
 503         /* this is SMP-safe way to finish thread. */
 504         complete_and_exit(&ctl->gc_finishing, 0);
 505 }
 506
 507 void ll_gns_add_timer(struct ll_sb_info *sbi)
 508 {
 509         mod_timer(&sbi->ll_gns_timer,
 510                   jiffies + sbi->ll_gns_tick * HZ);
 511 }
 512
 513 void ll_gns_del_timer(struct ll_sb_info *sbi)
 514 {
 515         del_timer(&sbi->ll_gns_timer);
 516 }
 517
 518 /*
 519  * starts GNS control thread and waits for a signal it is up and work may be
 520  * continued.
 521  */
 522 int ll_gns_start_thread(void)
 523 {
 524         int rc;
 525         ENTRY;
 526
 527         LASSERT(gns_thread.t_flags == 0);
 528         init_completion(&gns_ctl.gc_starting);
 529         init_completion(&gns_ctl.gc_finishing);
 530         init_waitqueue_head(&gns_thread.t_ctl_waitq);
 531
 532         rc = kernel_thread(ll_gns_thread_main, &gns_ctl,
 533                            (CLONE_VM | CLONE_FILES));
 534         if (rc < 0) {
 535                 CERROR("cannot start GNS control thread, "
 536                        "err = %d\n", rc);
 537                 RETURN(rc);
 538         }
 539         wait_for_completion(&gns_ctl.gc_starting);
 540         LASSERT(gns_thread.t_flags == SVC_RUNNING);
 541         RETURN(0);
 542 }
 543
 544 /* stops GNS control thread and waits its actual stop. */
 545 void ll_gns_stop_thread(void)
 546 {
 547         ENTRY;
 548         gns_thread.t_flags = SVC_STOPPING;
 549         wake_up(&gns_thread.t_ctl_waitq);
 550         wait_for_completion(&gns_ctl.gc_finishing);
 551         LASSERT(gns_thread.t_flags == SVC_STOPPED);
 552         gns_thread.t_flags = 0;
 553         EXIT;
 554 }