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                 /*
 111                  * check if another control thread is trying to mount some GNS
 112                  * dentry too. Letting it know that we busy and make
 113                  * ll_lookup_it() ti restart syscall and try again later.
 114                  */
 115                 spin_unlock(&sbi->ll_gns_lock);
 116                 CDEBUG(D_INODE, "GNS is in progress now, throwing "
 117                        "-ERESTARTSYS to restart syscall and let "
 118                        "it finish.\n");
 119                 RETURN(-EBUSY);
 120         }
 121         LASSERT(sbi->ll_gns_state == LL_GNS_IDLE);
 122
 123         /* mounting started */
 124         sbi->ll_gns_state = LL_GNS_MOUNTING;
 125         spin_unlock(&sbi->ll_gns_lock);
 126
 127         /* we need to build an absolute pathname to pass to mount */
 128         pathpage = (char *)__get_free_page(GFP_KERNEL);
 129         if (!pathpage)
 130                 GOTO(cleanup, rc = -ENOMEM);
 131         cleanup_phase = 1;
 132
 133         /* getting @dentry path stored in @pathpage. */
 134         path = d_path(dentry, mnt, pathpage, PAGE_SIZE);
 135         if (IS_ERR(path)) {
 136                 CERROR("can't build mount object path, err %d\n",
 137                        (int)PTR_ERR(dchild));
 138                 GOTO(cleanup, rc = PTR_ERR(dchild));
 139         }
 140
 141         /* synchronizing with possible /proc/fs/...write */
 142         down(&sbi->ll_gns_sem);
 143
 144         /*
 145          * mount object name is taken from sbi, where it is set in mount time or
 146          * via /proc/fs... tunable. It may be ".mntinfo" or so.
 147          */
 148
 149         /*
 150          * recursive lookup with trying to mount SUID bit marked directories on
 151          * the way is not possible here, as lookup_one_len() does not pass nd to
 152          * ->lookup() and this is checked in ll_lookup_it(). So, do not handle
 153          * possible -EBUSY here.
 154          */
 155         dchild = ll_lookup_one_len(sbi->ll_gns_oname, dentry,
 156                                    strlen(sbi->ll_gns_oname));
 157         up(&sbi->ll_gns_sem);
 158
 159         cleanup_phase = 2;
 160
 161         if (IS_ERR(dchild)) {
 162                 rc = PTR_ERR(dchild);
 163                 CERROR("can't find mount object %*s/%*s err = %d.\n",
 164                        (int)dentry->d_name.len, dentry->d_name.name,
 165                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 166                        rc);
 167                 GOTO(cleanup, rc);
 168         }
 169
 170         /* mount object is not found */
 171         if (!dchild->d_inode)
 172                 GOTO(cleanup, rc = -ENOENT);
 173
 174         /* check if found child is regular file */
 175         if (!S_ISREG(dchild->d_inode->i_mode))
 176                 GOTO(cleanup, rc = -EOPNOTSUPP);
 177
 178         mntget(mnt);
 179
 180         /* ok, mount object if found, opening it. */
 181         mntinfo_fd = dentry_open(dchild, mnt, 0);
 182         if (IS_ERR(mntinfo_fd)) {
 183                 CERROR("can't open mount object %*s/%*s err = %d.\n",
 184                        (int)dentry->d_name.len, dentry->d_name.name,
 185                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 186                        (int)PTR_ERR(mntinfo_fd));
 187                 mntput(mnt);
 188                 GOTO(cleanup, rc = PTR_ERR(mntinfo_fd));
 189         }
 190         cleanup_phase = 3;
 191
 192         if (mntinfo_fd->f_dentry->d_inode->i_size > PAGE_SIZE - 1) {
 193                 CERROR("mount object %*s/%*s is too big (%Ld)\n",
 194                        (int)dentry->d_name.len, dentry->d_name.name,
 195                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 196                        mntinfo_fd->f_dentry->d_inode->i_size);
 197                 GOTO(cleanup, rc = -EFBIG);
 198         }
 199
 200         datapage = (char *)__get_free_page(GFP_KERNEL);
 201         if (!datapage)
 202                 GOTO(cleanup, rc = -ENOMEM);
 203
 204         cleanup_phase = 4;
 205
 206         /* read data from mount object. */
 207         rc = kernel_read(mntinfo_fd, 0, datapage, PAGE_SIZE - 1);
 208         if (rc < 0) {
 209                 CERROR("can't read mount object %*s/%*s data, err %d\n",
 210                        (int)dentry->d_name.len, dentry->d_name.name,
 211                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 212                        rc);
 213                 GOTO(cleanup, rc);
 214         }
 215
 216         /* no data in mount object? */
 217         if (rc == 0) {
 218                 CERROR("mount object %*s/%*s is empty?\n",
 219                        (int)dentry->d_name.len, dentry->d_name.name,
 220                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname);
 221                 GOTO(cleanup, rc);
 222         }
 223
 224         datapage[rc] = '\0';
 225         fput(mntinfo_fd);
 226         mntinfo_fd = NULL;
 227         dchild = NULL;
 228
 229         /* synchronizing with possible /proc/fs/...write */
 230         down(&sbi->ll_gns_sem);
 231
 232         /*
 233          * upcall is initialized in mount time or via /proc/fs/... tuneable and
 234          * may be /usr/lib/lustre/gns-upcall.sh
 235          */
 236         argv[0] = sbi->ll_gns_upcall;
 237         argv[1] = datapage;
 238         argv[2] = path;
 239         argv[3] = NULL;
 240
 241         up(&sbi->ll_gns_sem);
 242
 243         rc = USERMODEHELPER(argv[0], argv, NULL);
 244         if (rc) {
 245                 CERROR("failed to call GNS upcall %s, err = %d\n",
 246                        sbi->ll_gns_upcall, rc);
 247                 GOTO(cleanup, rc);
 248         }
 249
 250         /*
 251          * wait for mount completion. This is actually not needed, because
 252          * USERMODEHELPER() returns only when usermode process finishes. But we
 253          * doing this just for case USERMODEHELPER() semantics will be changed
 254          * or usermode upcall program will start mounting in backgound and
 255          * return instantly. --umka
 256          */
 257         rc = ll_gns_wait_for_mount(dentry, 1, GNS_WAIT_ATTEMPTS);
 258         complete_all(&sbi->ll_gns_mount_finished);
 259         if (rc == 0) {
 260                 struct dentry *rdentry;
 261                 struct vfsmount *rmnt;
 262
 263                 /* mount is successful */
 264                 LASSERT(sbi->ll_gns_state == LL_GNS_FINISHED);
 265
 266                 rmnt = mntget(mnt);
 267                 rdentry = dget(dentry);
 268
 269                 if (follow_down(&rmnt, &rdentry)) {
 270                         /*
 271                          * registering new mount in GNS mounts list and thus
 272                          * make it accessible from GNS control thread.
 273                          */
 274                         spin_lock(&dcache_lock);
 275                         LASSERT(list_empty(&rmnt->mnt_lustre_list));
 276                         list_add_tail(&rmnt->mnt_lustre_list,
 277                                       &sbi->ll_mnt_list);
 278                         spin_unlock(&dcache_lock);
 279                         rmnt->mnt_last_used = jiffies;
 280                         mntput(rmnt);
 281                         dput(rdentry);
 282                 } else {
 283                         mntput(mnt);
 284                         dput(dentry);
 285                 }
 286         } else {
 287                 CERROR("usermode upcall %s failed to mount %s, err %d\n",
 288                        sbi->ll_gns_upcall, path, rc);
 289         }
 290
 291         EXIT;
 292 cleanup:
 293         switch (cleanup_phase) {
 294         case 4:
 295                 free_page((unsigned long)datapage);
 296         case 3:
 297                 if (mntinfo_fd != NULL)
 298                         fput(mntinfo_fd);
 299         case 2:
 300                 if (dchild != NULL)
 301                         dput(dchild);
 302         case 1:
 303                 free_page((unsigned long)pathpage);
 304
 305                 /*
 306                  * waking up all waiters after gns state is set to
 307                  * LL_GNS_MOUNTING
 308                  */
 309                 complete_all(&sbi->ll_gns_mount_finished);
 310         case 0:
 311                 spin_lock(&sbi->ll_gns_lock);
 312                 sbi->ll_gns_state = LL_GNS_IDLE;
 313                 spin_unlock(&sbi->ll_gns_lock);
 314         }
 315         return rc;
 316 }
 317
 318 /* tries to umount passed @mnt. */
 319 int ll_gns_umount_object(struct vfsmount *mnt)
 320 {
 321         int rc = 0;
 322         ENTRY;
 323
 324         CDEBUG(D_INODE, "unmounting mnt %p\n", mnt);
 325         rc = do_umount(mnt, 0);
 326         if (rc) {
 327                 CDEBUG(D_INODE, "can't umount 0x%p, err = %d\n",
 328                        mnt, rc);
 329         }
 330
 331         RETURN(rc);
 332 }
 333
 334 int ll_gns_check_mounts(struct ll_sb_info *sbi, int flags)
 335 {
 336         struct list_head check_list = LIST_HEAD_INIT(check_list);
 337         struct vfsmount *mnt;
 338         unsigned long pass;
 339         ENTRY;
 340
 341         spin_lock(&dcache_lock);
 342         list_splice_init(&sbi->ll_mnt_list, &check_list);
 343
 344         /*
 345          * walk the list in reverse order, and put them on the front of the sbi
 346          * list each iteration; this avoids list-ordering problems if we race
 347          * with another gns-mounting thread.
 348          */
 349         while (!list_empty(&check_list)) {
 350                 mnt = list_entry(check_list.prev,
 351                                  struct vfsmount,
 352                                  mnt_lustre_list);
 353
 354                 mntget(mnt);
 355
 356                 list_del_init(&mnt->mnt_lustre_list);
 357
 358                 list_add(&mnt->mnt_lustre_list,
 359                          &sbi->ll_mnt_list);
 360
 361                 /* check for timeout if needed */
 362                 pass = jiffies - mnt->mnt_last_used;
 363
 364                 if (flags == LL_GNS_CHECK &&
 365                     pass < sbi->ll_gns_timeout * HZ)
 366                 {
 367                         mntput(mnt);
 368                         continue;
 369                 }
 370                 spin_unlock(&dcache_lock);
 371
 372                 /* umounting @mnt */
 373                 ll_gns_umount_object(mnt);
 374
 375                 mntput(mnt);
 376                 spin_lock(&dcache_lock);
 377         }
 378         spin_unlock(&dcache_lock);
 379         RETURN(0);
 380 }
 381
 382 /*
 383  * GNS timer callback function. It restarts gns timer and wakes up GNS control
 384  * thread to process mounts list.
 385  */
 386 void ll_gns_timer_callback(unsigned long data)
 387 {
 388         struct ll_sb_info *sbi = (void *)data;
 389         ENTRY;
 390
 391         spin_lock(&gns_lock);
 392         if (list_empty(&sbi->ll_gns_sbi_head))
 393                 list_add(&sbi->ll_gns_sbi_head, &gns_sbi_list);
 394         spin_unlock(&gns_lock);
 395
 396         wake_up(&gns_thread.t_ctl_waitq);
 397         mod_timer(&sbi->ll_gns_timer,
 398                   jiffies + sbi->ll_gns_tick * HZ);
 399 }
 400
 401 /* this function checks if something new happened to exist in gns list. */
 402 static int inline ll_gns_check_event(void)
 403 {
 404         int rc;
 405
 406         spin_lock(&gns_lock);
 407         rc = !list_empty(&gns_sbi_list);
 408         spin_unlock(&gns_lock);
 409
 410         return rc;
 411 }
 412
 413 /* should we stop GNS control thread? */
 414 static int inline ll_gns_check_stop(void)
 415 {
 416         mb();
 417         return (gns_thread.t_flags & SVC_STOPPING) ? 1 : 0;
 418 }
 419
 420 /* GNS control thread function. */
 421 static int ll_gns_thread_main(void *arg)
 422 {
 423         struct ll_gns_ctl *ctl = arg;
 424         unsigned long flags;
 425         ENTRY;
 426
 427         {
 428                 char name[sizeof(current->comm)];
 429                 snprintf(name, sizeof(name) - 1, "ll_gns");
 430                 kportal_daemonize(name);
 431         }
 432
 433         SIGNAL_MASK_LOCK(current, flags);
 434         sigfillset(&current->blocked);
 435         RECALC_SIGPENDING;
 436         SIGNAL_MASK_UNLOCK(current, flags);
 437
 438         /*
 439          * letting starting function know, that we are ready and control may be
 440          * returned.
 441          */
 442         gns_thread.t_flags = SVC_RUNNING;
 443         complete(&ctl->gc_starting);
 444
 445         while (!ll_gns_check_stop()) {
 446                 struct l_wait_info lwi = { 0 };
 447
 448                 l_wait_event(gns_thread.t_ctl_waitq,
 449                              (ll_gns_check_event() ||
 450                               ll_gns_check_stop()), &lwi);
 451
 452                 spin_lock(&gns_lock);
 453                 while (!list_empty(&gns_sbi_list)) {
 454                         struct ll_sb_info *sbi;
 455
 456                         sbi = list_entry(gns_sbi_list.prev,
 457                                          struct ll_sb_info,
 458                                          ll_gns_sbi_head);
 459
 460                         list_del_init(&sbi->ll_gns_sbi_head);
 461                         spin_unlock(&gns_lock);
 462                         ll_gns_check_mounts(sbi, LL_GNS_CHECK);
 463                         spin_lock(&gns_lock);
 464                 }
 465                 spin_unlock(&gns_lock);
 466         }
 467
 468         EXIT;
 469         gns_thread.t_flags = SVC_STOPPED;
 470
 471         /* this is SMP-safe way to finish thread. */
 472         complete_and_exit(&ctl->gc_finishing, 0);
 473 }
 474
 475 void ll_gns_add_timer(struct ll_sb_info *sbi)
 476 {
 477         mod_timer(&sbi->ll_gns_timer,
 478                   jiffies + sbi->ll_gns_tick * HZ);
 479 }
 480
 481 void ll_gns_del_timer(struct ll_sb_info *sbi)
 482 {
 483         del_timer(&sbi->ll_gns_timer);
 484 }
 485
 486 /*
 487  * starts GNS control thread and waits for a signal it is up and work may be
 488  * continued.
 489  */
 490 int ll_gns_start_thread(void)
 491 {
 492         int rc;
 493         ENTRY;
 494
 495         LASSERT(gns_thread.t_flags == 0);
 496         init_completion(&gns_ctl.gc_starting);
 497         init_completion(&gns_ctl.gc_finishing);
 498         init_waitqueue_head(&gns_thread.t_ctl_waitq);
 499
 500         rc = kernel_thread(ll_gns_thread_main, &gns_ctl,
 501                            (CLONE_VM | CLONE_FILES));
 502         if (rc < 0) {
 503                 CERROR("cannot start GNS control thread, "
 504                        "err = %d\n", rc);
 505                 RETURN(rc);
 506         }
 507         wait_for_completion(&gns_ctl.gc_starting);
 508         LASSERT(gns_thread.t_flags == SVC_RUNNING);
 509         RETURN(0);
 510 }
 511
 512 /* stops GNS control thread and waits its actual stop. */
 513 void ll_gns_stop_thread(void)
 514 {
 515         ENTRY;
 516         gns_thread.t_flags = SVC_STOPPING;
 517         wake_up(&gns_thread.t_ctl_waitq);
 518         wait_for_completion(&gns_ctl.gc_finishing);
 519         LASSERT(gns_thread.t_flags == SVC_STOPPED);
 520         gns_thread.t_flags = 0;
 521         EXIT;
 522 }