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  * tries to mount the mount object under passed @dentry. In the case of success
  74  * @dentry will become mount point and 0 will be returned. Error code will be
  75  * returned otherwise.
  76  */
  77 int
  78 ll_gns_mount_object(struct dentry *dentry, struct vfsmount *mnt)
  79 {
  80         struct ll_dentry_data *lld = dentry->d_fsdata;
  81         char *path, *pathpage, *datapage, *argv[4];
  82         struct file *mntinfo_fd = NULL;
  83         int cleanup_phase = 0, rc = 0;
  84         struct ll_sb_info *sbi;
  85         struct dentry *dchild;
  86         ENTRY;
  87
  88         if (mnt == NULL) {
  89                 CERROR("suid directory found, but no "
  90                        "vfsmount available.\n");
  91                 RETURN(-EINVAL);
  92         }
  93
  94         CDEBUG(D_INODE, "mounting dentry %p\n", dentry);
  95
  96         LASSERT(dentry->d_inode != NULL);
  97         LASSERT(S_ISDIR(dentry->d_inode->i_mode));
  98         LASSERT(lld != NULL);
  99
 100         sbi = ll_i2sbi(dentry->d_inode);
 101         LASSERT(sbi != NULL);
 102
 103         /*
 104          * another thead is in progress or just finished mounting the
 105          * dentry. Handling that.
 106          */
 107         spin_lock(&sbi->ll_gns_lock);
 108         if (sbi->ll_gns_state == LL_GNS_MOUNTING ||
 109             sbi->ll_gns_state == LL_GNS_FINISHED) {
 110                 spin_unlock(&sbi->ll_gns_lock);
 111                 CDEBUG(D_INODE, "GNS is in progress now, throwing "
 112                        "-ERESTARTSYS to restart syscall and let "
 113                        "it finish.\n");
 114                 RETURN(-ERESTARTSYS);
 115         }
 116         LASSERT(sbi->ll_gns_state == LL_GNS_IDLE);
 117
 118         /* mounting started */
 119         sbi->ll_gns_state = LL_GNS_MOUNTING;
 120         spin_unlock(&sbi->ll_gns_lock);
 121
 122         /* we need to build an absolute pathname to pass to mount */
 123         pathpage = (char *)__get_free_page(GFP_KERNEL);
 124         if (!pathpage)
 125                 GOTO(cleanup, rc = -ENOMEM);
 126         cleanup_phase = 1;
 127
 128         /* getting @dentry path stored in @pathpage. */
 129         path = d_path(dentry, mnt, pathpage, PAGE_SIZE);
 130         if (IS_ERR(path)) {
 131                 CERROR("can't build mount object path, err %d\n",
 132                        (int)PTR_ERR(dchild));
 133                 GOTO(cleanup, rc = PTR_ERR(dchild));
 134         }
 135
 136         /* synchronizing with possible /proc/fs/...write */
 137         down(&sbi->ll_gns_sem);
 138
 139         /*
 140          * mount object name is taken from sbi, where it is set in mount time or
 141          * via /proc/fs... tunable. It may be ".mntinfo" or so.
 142          */
 143         dchild = lookup_one_len(sbi->ll_gns_oname, dentry,
 144                                 strlen(sbi->ll_gns_oname));
 145         up(&sbi->ll_gns_sem);
 146
 147         cleanup_phase = 2;
 148
 149         if (IS_ERR(dchild)) {
 150                 rc = PTR_ERR(dchild);
 151
 152                 if (rc == -ERESTARTSYS) {
 153                         CDEBUG(D_INODE, "possible endless loop is detected "
 154                                "due to mount object is directory marked by "
 155                                "SUID bit.\n");
 156                         GOTO(cleanup, rc = -ELOOP);
 157                 }
 158
 159                 CERROR("can't find mount object %*s/%*s err = %d.\n",
 160                        (int)dentry->d_name.len, dentry->d_name.name,
 161                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 162                        rc);
 163                 GOTO(cleanup, rc);
 164         }
 165
 166         /* mount object is not found */
 167         if (!dchild->d_inode)
 168                 GOTO(cleanup, rc = -ENOENT);
 169
 170         /* check if found child is regular file */
 171         if (!S_ISREG(dchild->d_inode->i_mode))
 172                 GOTO(cleanup, rc = -EOPNOTSUPP);
 173
 174         mntget(mnt);
 175
 176         /* ok, mount object if found, opening it. */
 177         mntinfo_fd = dentry_open(dchild, mnt, 0);
 178         if (IS_ERR(mntinfo_fd)) {
 179                 CERROR("can't open mount object %*s/%*s err = %d.\n",
 180                        (int)dentry->d_name.len, dentry->d_name.name,
 181                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 182                        (int)PTR_ERR(mntinfo_fd));
 183                 mntput(mnt);
 184                 GOTO(cleanup, rc = PTR_ERR(mntinfo_fd));
 185         }
 186         cleanup_phase = 3;
 187
 188         if (mntinfo_fd->f_dentry->d_inode->i_size > PAGE_SIZE - 1) {
 189                 CERROR("mount object %*s/%*s is too big (%Ld)\n",
 190                        (int)dentry->d_name.len, dentry->d_name.name,
 191                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 192                        mntinfo_fd->f_dentry->d_inode->i_size);
 193                 GOTO(cleanup, rc = -EFBIG);
 194         }
 195
 196         datapage = (char *)__get_free_page(GFP_KERNEL);
 197         if (!datapage)
 198                 GOTO(cleanup, rc = -ENOMEM);
 199
 200         cleanup_phase = 4;
 201
 202         /* read data from mount object. */
 203         rc = kernel_read(mntinfo_fd, 0, datapage, PAGE_SIZE - 1);
 204         if (rc < 0) {
 205                 CERROR("can't read mount object %*s/%*s data, err %d\n",
 206                        (int)dentry->d_name.len, dentry->d_name.name,
 207                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 208                        rc);
 209                 GOTO(cleanup, rc);
 210         }
 211
 212         /* no data in mount object? */
 213         if (rc == 0) {
 214                 CERROR("mount object %*s/%*s is empty?\n",
 215                        (int)dentry->d_name.len, dentry->d_name.name,
 216                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname);
 217                 GOTO(cleanup, rc);
 218         }
 219
 220         datapage[rc] = '\0';
 221         fput(mntinfo_fd);
 222         mntinfo_fd = NULL;
 223         dchild = NULL;
 224
 225         /* synchronizing with possible /proc/fs/...write */
 226         down(&sbi->ll_gns_sem);
 227
 228         /*
 229          * upcall is initialized in mount time or via /proc/fs/... tuneable and
 230          * may be /usr/lib/lustre/gns-upcall.sh
 231          */
 232         argv[0] = sbi->ll_gns_upcall;
 233         argv[1] = datapage;
 234         argv[2] = path;
 235         argv[3] = NULL;
 236
 237         up(&sbi->ll_gns_sem);
 238
 239         rc = USERMODEHELPER(argv[0], argv, NULL);
 240         if (rc) {
 241                 CERROR("failed to call GNS upcall %s, err = %d\n",
 242                        sbi->ll_gns_upcall, rc);
 243                 GOTO(cleanup, rc);
 244         }
 245
 246         /*
 247          * wait for mount completion. This is actually not needed, because
 248          * USERMODEHELPER() returns only when usermode process finishes. But we
 249          * doing this just for case USERMODEHELPER() semantics will be changed
 250          * or usermode upcall program will start mounting in backgound and
 251          * return instantly. --umka
 252          */
 253         rc = ll_gns_wait_for_mount(dentry, 1, GNS_WAIT_ATTEMPTS);
 254         complete_all(&sbi->ll_gns_mount_finished);
 255         if (rc == 0) {
 256                 struct dentry *rdentry;
 257                 struct vfsmount *rmnt;
 258
 259                 /* mount is successful */
 260                 LASSERT(sbi->ll_gns_state == LL_GNS_FINISHED);
 261
 262                 rmnt = mntget(mnt);
 263                 rdentry = dget(dentry);
 264
 265                 if (follow_down(&rmnt, &rdentry)) {
 266                         /*
 267                          * registering new mount in GNS mounts list and thus
 268                          * make it accessible from GNS control thread.
 269                          */
 270                         spin_lock(&dcache_lock);
 271                         LASSERT(list_empty(&rmnt->mnt_lustre_list));
 272                         list_add_tail(&rmnt->mnt_lustre_list,
 273                                       &sbi->ll_mnt_list);
 274                         spin_unlock(&dcache_lock);
 275                         rmnt->mnt_last_used = jiffies;
 276                         mntput(rmnt);
 277                         dput(rdentry);
 278                 } else {
 279                         mntput(mnt);
 280                         dput(dentry);
 281                 }
 282         } else {
 283                 CERROR("usermode upcall %s failed to mount %s, err %d\n",
 284                        sbi->ll_gns_upcall, path, rc);
 285         }
 286
 287         EXIT;
 288 cleanup:
 289         switch (cleanup_phase) {
 290         case 4:
 291                 free_page((unsigned long)datapage);
 292         case 3:
 293                 if (mntinfo_fd != NULL)
 294                         fput(mntinfo_fd);
 295         case 2:
 296                 if (dchild != NULL)
 297                         dput(dchild);
 298         case 1:
 299                 free_page((unsigned long)pathpage);
 300
 301                 /*
 302                  * waking up all waiters after gns state is set to
 303                  * LL_GNS_MOUNTING
 304                  */
 305                 complete_all(&sbi->ll_gns_mount_finished);
 306         case 0:
 307                 spin_lock(&sbi->ll_gns_lock);
 308                 sbi->ll_gns_state = LL_GNS_IDLE;
 309                 spin_unlock(&sbi->ll_gns_lock);
 310         }
 311         return rc;
 312 }
 313
 314 /* tries to umount passed @mnt. */
 315 int ll_gns_umount_object(struct vfsmount *mnt)
 316 {
 317         int rc = 0;
 318         ENTRY;
 319
 320         CDEBUG(D_INODE, "unmounting mnt %p\n", mnt);
 321         rc = do_umount(mnt, 0);
 322         if (rc) {
 323                 CDEBUG(D_INODE, "can't umount 0x%p, err = %d\n",
 324                        mnt, rc);
 325         }
 326
 327         RETURN(rc);
 328 }
 329
 330 int ll_gns_check_mounts(struct ll_sb_info *sbi, int flags)
 331 {
 332         struct list_head check_list = LIST_HEAD_INIT(check_list);
 333         struct vfsmount *mnt;
 334         unsigned long pass;
 335         ENTRY;
 336
 337         spin_lock(&dcache_lock);
 338         list_splice_init(&sbi->ll_mnt_list, &check_list);
 339
 340         /*
 341          * walk the list in reverse order, and put them on the front of the sbi
 342          * list each iteration; this avoids list-ordering problems if we race
 343          * with another gns-mounting thread.
 344          */
 345         while (!list_empty(&check_list)) {
 346                 mnt = list_entry(check_list.prev,
 347                                  struct vfsmount,
 348                                  mnt_lustre_list);
 349
 350                 mntget(mnt);
 351
 352                 list_del_init(&mnt->mnt_lustre_list);
 353
 354                 list_add(&mnt->mnt_lustre_list,
 355                          &sbi->ll_mnt_list);
 356
 357                 /* check for timeout if needed */
 358                 pass = jiffies - mnt->mnt_last_used;
 359
 360                 if (flags == LL_GNS_CHECK &&
 361                     pass < sbi->ll_gns_timeout * HZ)
 362                 {
 363                         mntput(mnt);
 364                         continue;
 365                 }
 366                 spin_unlock(&dcache_lock);
 367
 368                 /* umounting @mnt */
 369                 ll_gns_umount_object(mnt);
 370
 371                 mntput(mnt);
 372                 spin_lock(&dcache_lock);
 373         }
 374         spin_unlock(&dcache_lock);
 375         RETURN(0);
 376 }
 377
 378 /*
 379  * GNS timer callback function. It restarts gns timer and wakes up GNS control
 380  * thread to process mounts list.
 381  */
 382 void ll_gns_timer_callback(unsigned long data)
 383 {
 384         struct ll_sb_info *sbi = (void *)data;
 385         ENTRY;
 386
 387         spin_lock(&gns_lock);
 388         if (list_empty(&sbi->ll_gns_sbi_head))
 389                 list_add(&sbi->ll_gns_sbi_head, &gns_sbi_list);
 390         spin_unlock(&gns_lock);
 391
 392         wake_up(&gns_thread.t_ctl_waitq);
 393         mod_timer(&sbi->ll_gns_timer,
 394                   jiffies + sbi->ll_gns_tick * HZ);
 395 }
 396
 397 /* this function checks if something new happened to exist in gns list. */
 398 static int inline ll_gns_check_event(void)
 399 {
 400         int rc;
 401
 402         spin_lock(&gns_lock);
 403         rc = !list_empty(&gns_sbi_list);
 404         spin_unlock(&gns_lock);
 405
 406         return rc;
 407 }
 408
 409 /* should we stop GNS control thread? */
 410 static int inline ll_gns_check_stop(void)
 411 {
 412         mb();
 413         return (gns_thread.t_flags & SVC_STOPPING) ? 1 : 0;
 414 }
 415
 416 /* GNS control thread function. */
 417 static int ll_gns_thread_main(void *arg)
 418 {
 419         struct ll_gns_ctl *ctl = arg;
 420         unsigned long flags;
 421         ENTRY;
 422
 423         {
 424                 char name[sizeof(current->comm)];
 425                 snprintf(name, sizeof(name) - 1, "ll_gns");
 426                 kportal_daemonize(name);
 427         }
 428
 429         SIGNAL_MASK_LOCK(current, flags);
 430         sigfillset(&current->blocked);
 431         RECALC_SIGPENDING;
 432         SIGNAL_MASK_UNLOCK(current, flags);
 433
 434         /*
 435          * letting starting function know, that we are ready and control may be
 436          * returned.
 437          */
 438         gns_thread.t_flags = SVC_RUNNING;
 439         complete(&ctl->gc_starting);
 440
 441         while (!ll_gns_check_stop()) {
 442                 struct l_wait_info lwi = { 0 };
 443
 444                 l_wait_event(gns_thread.t_ctl_waitq,
 445                              (ll_gns_check_event() ||
 446                               ll_gns_check_stop()), &lwi);
 447
 448                 spin_lock(&gns_lock);
 449                 while (!list_empty(&gns_sbi_list)) {
 450                         struct ll_sb_info *sbi;
 451
 452                         sbi = list_entry(gns_sbi_list.prev,
 453                                          struct ll_sb_info,
 454                                          ll_gns_sbi_head);
 455
 456                         list_del_init(&sbi->ll_gns_sbi_head);
 457                         spin_unlock(&gns_lock);
 458                         ll_gns_check_mounts(sbi, LL_GNS_CHECK);
 459                         spin_lock(&gns_lock);
 460                 }
 461                 spin_unlock(&gns_lock);
 462         }
 463
 464         EXIT;
 465         gns_thread.t_flags = SVC_STOPPED;
 466
 467         /* this is SMP-safe way to finish thread. */
 468         complete_and_exit(&ctl->gc_finishing, 0);
 469 }
 470
 471 void ll_gns_add_timer(struct ll_sb_info *sbi)
 472 {
 473         mod_timer(&sbi->ll_gns_timer,
 474                   jiffies + sbi->ll_gns_tick * HZ);
 475 }
 476
 477 void ll_gns_del_timer(struct ll_sb_info *sbi)
 478 {
 479         del_timer(&sbi->ll_gns_timer);
 480 }
 481
 482 /*
 483  * starts GNS control thread and waits for a signal it is up and work may be
 484  * continued.
 485  */
 486 int ll_gns_start_thread(void)
 487 {
 488         int rc;
 489         ENTRY;
 490
 491         LASSERT(gns_thread.t_flags == 0);
 492         init_completion(&gns_ctl.gc_starting);
 493         init_completion(&gns_ctl.gc_finishing);
 494         init_waitqueue_head(&gns_thread.t_ctl_waitq);
 495
 496         rc = kernel_thread(ll_gns_thread_main, &gns_ctl,
 497                            (CLONE_VM | CLONE_FILES));
 498         if (rc < 0) {
 499                 CERROR("cannot start GNS control thread, "
 500                        "err = %d\n", rc);
 501                 RETURN(rc);
 502         }
 503         wait_for_completion(&gns_ctl.gc_starting);
 504         LASSERT(gns_thread.t_flags == SVC_RUNNING);
 505         RETURN(0);
 506 }
 507
 508 /* stops GNS control thread and waits its actual stop. */
 509 void ll_gns_stop_thread(void)
 510 {
 511         ENTRY;
 512         gns_thread.t_flags = SVC_STOPPING;
 513         wake_up(&gns_thread.t_ctl_waitq);
 514         wait_for_completion(&gns_ctl.gc_finishing);
 515         LASSERT(gns_thread.t_flags == SVC_STOPPED);
 516         gns_thread.t_flags = 0;
 517         EXIT;
 518 }