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         spin_lock(&dentry->d_lock);
 119         dentry->d_flags |= DCACHE_GNS_MOUNTING;
 120         spin_unlock(&dentry->d_lock);
 121
 122         /* mounting started */
 123         sbi->ll_gns_state = LL_GNS_MOUNTING;
 124         spin_unlock(&sbi->ll_gns_lock);
 125
 126         /* we need to build an absolute pathname to pass to mount */
 127         pathpage = (char *)__get_free_page(GFP_KERNEL);
 128         if (!pathpage)
 129                 GOTO(cleanup, rc = -ENOMEM);
 130         cleanup_phase = 1;
 131
 132         /* getting @dentry path stored in @pathpage. */
 133         path = d_path(dentry, mnt, pathpage, PAGE_SIZE);
 134         if (IS_ERR(path)) {
 135                 CERROR("can't build mount object path, err %d\n",
 136                        (int)PTR_ERR(dchild));
 137                 GOTO(cleanup, rc = PTR_ERR(dchild));
 138         }
 139
 140         /* synchronizing with possible /proc/fs/...write */
 141         down(&sbi->ll_gns_sem);
 142
 143         /*
 144          * mount object name is taken from sbi, where it is set in mount time or
 145          * via /proc/fs... tunable. It may be ".mntinfo" or so.
 146          */
 147
 148         /*
 149          * FIXME: lookup_one_len() requires dentry->d_inode->i_sem to be locked,
 150          * but we can't use ll_lookup_one_len() as this function is called from
 151          * different contol paths and some of them take dentry->d_inode->i_sem
 152          * and others do not.
 153          */
 154         dchild = lookup_one_len(sbi->ll_gns_oname, dentry,
 155                                 strlen(sbi->ll_gns_oname));
 156         up(&sbi->ll_gns_sem);
 157
 158         cleanup_phase = 2;
 159
 160         if (IS_ERR(dchild)) {
 161                 rc = PTR_ERR(dchild);
 162
 163                 if (rc == -ERESTARTSYS) {
 164                         CDEBUG(D_INODE, "possible endless loop is detected "
 165                                "due to mount object is directory marked by "
 166                                "SUID bit.\n");
 167                         GOTO(cleanup, rc = -ELOOP);
 168                 }
 169
 170                 CERROR("can't find mount object %*s/%*s err = %d.\n",
 171                        (int)dentry->d_name.len, dentry->d_name.name,
 172                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 173                        rc);
 174                 GOTO(cleanup, rc);
 175         }
 176
 177         /* mount object is not found */
 178         if (!dchild->d_inode)
 179                 GOTO(cleanup, rc = -ENOENT);
 180
 181         /* check if found child is regular file */
 182         if (!S_ISREG(dchild->d_inode->i_mode))
 183                 GOTO(cleanup, rc = -EOPNOTSUPP);
 184
 185         mntget(mnt);
 186
 187         /* ok, mount object if found, opening it. */
 188         mntinfo_fd = dentry_open(dchild, mnt, 0);
 189         if (IS_ERR(mntinfo_fd)) {
 190                 CERROR("can't open mount object %*s/%*s err = %d.\n",
 191                        (int)dentry->d_name.len, dentry->d_name.name,
 192                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 193                        (int)PTR_ERR(mntinfo_fd));
 194                 mntput(mnt);
 195                 GOTO(cleanup, rc = PTR_ERR(mntinfo_fd));
 196         }
 197         cleanup_phase = 3;
 198
 199         if (mntinfo_fd->f_dentry->d_inode->i_size > PAGE_SIZE) {
 200                 CERROR("mount object %*s/%*s is too big (%Ld)\n",
 201                        (int)dentry->d_name.len, dentry->d_name.name,
 202                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 203                        mntinfo_fd->f_dentry->d_inode->i_size);
 204                 GOTO(cleanup, rc = -EFBIG);
 205         }
 206
 207         datapage = (char *)__get_free_page(GFP_KERNEL);
 208         if (!datapage)
 209                 GOTO(cleanup, rc = -ENOMEM);
 210
 211         cleanup_phase = 4;
 212
 213         /* read data from mount object. */
 214         rc = kernel_read(mntinfo_fd, 0, datapage, PAGE_SIZE);
 215         if (rc < 0) {
 216                 CERROR("can't read mount object %*s/%*s data, err %d\n",
 217                        (int)dentry->d_name.len, dentry->d_name.name,
 218                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 219                        rc);
 220                 GOTO(cleanup, rc);
 221         }
 222
 223         datapage[PAGE_SIZE - 1] = '\0';
 224
 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 need, 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                 spin_lock(&dentry->d_lock);
 287                 dentry->d_flags &= ~DCACHE_GNS_PENDING;
 288                 spin_unlock(&dentry->d_lock);
 289         } else {
 290                 CERROR("usermode upcall %s failed to mount %s, err %d\n",
 291                        sbi->ll_gns_upcall, path, rc);
 292         }
 293
 294         EXIT;
 295 cleanup:
 296         switch (cleanup_phase) {
 297         case 4:
 298                 free_page((unsigned long)datapage);
 299         case 3:
 300                 if (mntinfo_fd != NULL)
 301                         fput(mntinfo_fd);
 302         case 2:
 303                 if (dchild != NULL)
 304                         dput(dchild);
 305         case 1:
 306                 free_page((unsigned long)pathpage);
 307
 308                 /*
 309                  * waking up all waiters after gns state is set to
 310                  * LL_GNS_MOUNTING
 311                  */
 312                 complete_all(&sbi->ll_gns_mount_finished);
 313         case 0:
 314                 spin_lock(&sbi->ll_gns_lock);
 315                 sbi->ll_gns_state = LL_GNS_IDLE;
 316                 spin_unlock(&sbi->ll_gns_lock);
 317
 318                 spin_lock(&dentry->d_lock);
 319                 dentry->d_flags &= ~DCACHE_GNS_MOUNTING;
 320                 spin_unlock(&dentry->d_lock);
 321         }
 322         return rc;
 323 }
 324
 325 /* tries to umount passed @mnt. */
 326 int ll_gns_umount_object(struct vfsmount *mnt)
 327 {
 328         int rc = 0;
 329         ENTRY;
 330
 331         CDEBUG(D_INODE, "unmounting mnt %p\n", mnt);
 332         rc = do_umount(mnt, 0);
 333         if (rc) {
 334                 CDEBUG(D_INODE, "can't umount 0x%p, err = %d\n",
 335                        mnt, rc);
 336         }
 337
 338         RETURN(rc);
 339 }
 340
 341 int ll_gns_check_mounts(struct ll_sb_info *sbi, int flags)
 342 {
 343         struct list_head check_list = LIST_HEAD_INIT(check_list);
 344         struct vfsmount *mnt;
 345         unsigned long pass;
 346         ENTRY;
 347
 348         spin_lock(&dcache_lock);
 349         list_splice_init(&sbi->ll_mnt_list, &check_list);
 350
 351         /*
 352          * walk the list in reverse order, and put them on the front of the sbi
 353          * list each iteration; this avoids list-ordering problems if we race
 354          * with another gns-mounting thread.
 355          */
 356         while (!list_empty(&check_list)) {
 357                 mnt = list_entry(check_list.prev,
 358                                  struct vfsmount,
 359                                  mnt_lustre_list);
 360
 361                 mntget(mnt);
 362
 363                 list_del_init(&mnt->mnt_lustre_list);
 364
 365                 list_add(&mnt->mnt_lustre_list,
 366                          &sbi->ll_mnt_list);
 367
 368                 /* check for timeout if needed */
 369                 pass = jiffies - mnt->mnt_last_used;
 370
 371                 if (flags == LL_GNS_CHECK &&
 372                     pass < sbi->ll_gns_timeout * HZ)
 373                 {
 374                         mntput(mnt);
 375                         continue;
 376                 }
 377                 spin_unlock(&dcache_lock);
 378
 379                 /* umounting @mnt */
 380                 ll_gns_umount_object(mnt);
 381
 382                 mntput(mnt);
 383                 spin_lock(&dcache_lock);
 384         }
 385         spin_unlock(&dcache_lock);
 386         RETURN(0);
 387 }
 388
 389 /*
 390  * GNS timer callback function. It restarts gns timer and wakes up GNS control
 391  * thread to process mounts list.
 392  */
 393 void ll_gns_timer_callback(unsigned long data)
 394 {
 395         struct ll_sb_info *sbi = (void *)data;
 396         ENTRY;
 397
 398         spin_lock(&gns_lock);
 399         if (list_empty(&sbi->ll_gns_sbi_head))
 400                 list_add(&sbi->ll_gns_sbi_head, &gns_sbi_list);
 401         spin_unlock(&gns_lock);
 402
 403         wake_up(&gns_thread.t_ctl_waitq);
 404         mod_timer(&sbi->ll_gns_timer,
 405                   jiffies + sbi->ll_gns_tick * HZ);
 406 }
 407
 408 /* this function checks if something new happened to exist in gns list. */
 409 static int inline ll_gns_check_event(void)
 410 {
 411         int rc;
 412
 413         spin_lock(&gns_lock);
 414         rc = !list_empty(&gns_sbi_list);
 415         spin_unlock(&gns_lock);
 416
 417         return rc;
 418 }
 419
 420 /* should we stop GNS control thread? */
 421 static int inline ll_gns_check_stop(void)
 422 {
 423         mb();
 424         return (gns_thread.t_flags & SVC_STOPPING) ? 1 : 0;
 425 }
 426
 427 /* GNS control thread function. */
 428 static int ll_gns_thread_main(void *arg)
 429 {
 430         struct ll_gns_ctl *ctl = arg;
 431         unsigned long flags;
 432         ENTRY;
 433
 434         {
 435                 char name[sizeof(current->comm)];
 436                 snprintf(name, sizeof(name) - 1, "ll_gns");
 437                 kportal_daemonize(name);
 438         }
 439
 440         SIGNAL_MASK_LOCK(current, flags);
 441         sigfillset(&current->blocked);
 442         RECALC_SIGPENDING;
 443         SIGNAL_MASK_UNLOCK(current, flags);
 444
 445         /*
 446          * letting starting function know, that we are ready and control may be
 447          * returned.
 448          */
 449         gns_thread.t_flags = SVC_RUNNING;
 450         complete(&ctl->gc_starting);
 451
 452         while (!ll_gns_check_stop()) {
 453                 struct l_wait_info lwi = { 0 };
 454
 455                 l_wait_event(gns_thread.t_ctl_waitq,
 456                              (ll_gns_check_event() ||
 457                               ll_gns_check_stop()), &lwi);
 458
 459                 spin_lock(&gns_lock);
 460                 while (!list_empty(&gns_sbi_list)) {
 461                         struct ll_sb_info *sbi;
 462
 463                         sbi = list_entry(gns_sbi_list.prev,
 464                                          struct ll_sb_info,
 465                                          ll_gns_sbi_head);
 466
 467                         list_del_init(&sbi->ll_gns_sbi_head);
 468                         spin_unlock(&gns_lock);
 469                         ll_gns_check_mounts(sbi, LL_GNS_CHECK);
 470                         spin_lock(&gns_lock);
 471                 }
 472                 spin_unlock(&gns_lock);
 473         }
 474
 475         EXIT;
 476         gns_thread.t_flags = SVC_STOPPED;
 477
 478         /* this is SMP-safe way to finish thread. */
 479         complete_and_exit(&ctl->gc_finishing, 0);
 480 }
 481
 482 void ll_gns_add_timer(struct ll_sb_info *sbi)
 483 {
 484         mod_timer(&sbi->ll_gns_timer,
 485                   jiffies + sbi->ll_gns_tick * HZ);
 486 }
 487
 488 void ll_gns_del_timer(struct ll_sb_info *sbi)
 489 {
 490         del_timer(&sbi->ll_gns_timer);
 491 }
 492
 493 /*
 494  * starts GNS control thread and waits for a signal it is up and work may be
 495  * continued.
 496  */
 497 int ll_gns_start_thread(void)
 498 {
 499         int rc;
 500         ENTRY;
 501
 502         LASSERT(gns_thread.t_flags == 0);
 503         init_completion(&gns_ctl.gc_starting);
 504         init_completion(&gns_ctl.gc_finishing);
 505         init_waitqueue_head(&gns_thread.t_ctl_waitq);
 506
 507         rc = kernel_thread(ll_gns_thread_main, &gns_ctl,
 508                            (CLONE_VM | CLONE_FILES));
 509         if (rc < 0) {
 510                 CERROR("cannot start GNS control thread, "
 511                        "err = %d\n", rc);
 512                 RETURN(rc);
 513         }
 514         wait_for_completion(&gns_ctl.gc_starting);
 515         LASSERT(gns_thread.t_flags == SVC_RUNNING);
 516         RETURN(0);
 517 }
 518
 519 /* stops GNS control thread and waits its actual stop. */
 520 void ll_gns_stop_thread(void)
 521 {
 522         ENTRY;
 523         gns_thread.t_flags = SVC_STOPPING;
 524         wake_up(&gns_thread.t_ctl_waitq);
 525         wait_for_completion(&gns_ctl.gc_finishing);
 526         LASSERT(gns_thread.t_flags == SVC_STOPPED);
 527         gns_thread.t_flags = 0;
 528         EXIT;
 529 }