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) {
 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);
 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         datapage[PAGE_SIZE - 1] = '\0';
 213
 214         fput(mntinfo_fd);
 215         mntinfo_fd = NULL;
 216         dchild = NULL;
 217
 218         /* synchronizing with possible /proc/fs/...write */
 219         down(&sbi->ll_gns_sem);
 220
 221         /*
 222          * upcall is initialized in mount time or via /proc/fs/... tuneable and
 223          * may be /usr/lib/lustre/gns-upcall.sh
 224          */
 225         argv[0] = sbi->ll_gns_upcall;
 226         argv[1] = datapage;
 227         argv[2] = path;
 228         argv[3] = NULL;
 229
 230         up(&sbi->ll_gns_sem);
 231
 232         rc = USERMODEHELPER(argv[0], argv, NULL);
 233         if (rc) {
 234                 CERROR("failed to call GNS upcall %s, err = %d\n",
 235                        sbi->ll_gns_upcall, rc);
 236                 GOTO(cleanup, rc);
 237         }
 238
 239         /*
 240          * wait for mount completion. This is actually not need, because
 241          * USERMODEHELPER() returns only when usermode process finishes. But we
 242          * doing this just for case USERMODEHELPER() semantics will be changed
 243          * or usermode upcall program will start mounting in backgound and
 244          * return instantly. --umka
 245          */
 246         rc = ll_gns_wait_for_mount(dentry, 1, GNS_WAIT_ATTEMPTS);
 247         complete_all(&sbi->ll_gns_mount_finished);
 248         if (rc == 0) {
 249                 struct dentry *rdentry;
 250                 struct vfsmount *rmnt;
 251
 252                 /* mount is successful */
 253                 LASSERT(sbi->ll_gns_state == LL_GNS_FINISHED);
 254
 255                 rmnt = mntget(mnt);
 256                 rdentry = dget(dentry);
 257
 258                 if (follow_down(&rmnt, &rdentry)) {
 259                         /*
 260                          * registering new mount in GNS mounts list and thus
 261                          * make it accessible from GNS control thread.
 262                          */
 263                         spin_lock(&dcache_lock);
 264                         LASSERT(list_empty(&rmnt->mnt_lustre_list));
 265                         list_add_tail(&rmnt->mnt_lustre_list,
 266                                       &sbi->ll_mnt_list);
 267                         spin_unlock(&dcache_lock);
 268                         rmnt->mnt_last_used = jiffies;
 269                         mntput(rmnt);
 270                         dput(rdentry);
 271                 } else {
 272                         mntput(mnt);
 273                         dput(dentry);
 274                 }
 275         } else {
 276                 CERROR("usermode upcall %s failed to mount %s, err %d\n",
 277                        sbi->ll_gns_upcall, path, rc);
 278         }
 279
 280         EXIT;
 281 cleanup:
 282         switch (cleanup_phase) {
 283         case 4:
 284                 free_page((unsigned long)datapage);
 285         case 3:
 286                 if (mntinfo_fd != NULL)
 287                         fput(mntinfo_fd);
 288         case 2:
 289                 if (dchild != NULL)
 290                         dput(dchild);
 291         case 1:
 292                 free_page((unsigned long)pathpage);
 293
 294                 /*
 295                  * waking up all waiters after gns state is set to
 296                  * LL_GNS_MOUNTING
 297                  */
 298                 complete_all(&sbi->ll_gns_mount_finished);
 299         case 0:
 300                 spin_lock(&sbi->ll_gns_lock);
 301                 sbi->ll_gns_state = LL_GNS_IDLE;
 302                 spin_unlock(&sbi->ll_gns_lock);
 303         }
 304         return rc;
 305 }
 306
 307 /* tries to umount passed @mnt. */
 308 int ll_gns_umount_object(struct vfsmount *mnt)
 309 {
 310         int rc = 0;
 311         ENTRY;
 312
 313         CDEBUG(D_INODE, "unmounting mnt %p\n", mnt);
 314         rc = do_umount(mnt, 0);
 315         if (rc) {
 316                 CDEBUG(D_INODE, "can't umount 0x%p, err = %d\n",
 317                        mnt, rc);
 318         }
 319
 320         RETURN(rc);
 321 }
 322
 323 int ll_gns_check_mounts(struct ll_sb_info *sbi, int flags)
 324 {
 325         struct list_head check_list = LIST_HEAD_INIT(check_list);
 326         struct vfsmount *mnt;
 327         unsigned long pass;
 328         ENTRY;
 329
 330         spin_lock(&dcache_lock);
 331         list_splice_init(&sbi->ll_mnt_list, &check_list);
 332
 333         /*
 334          * walk the list in reverse order, and put them on the front of the sbi
 335          * list each iteration; this avoids list-ordering problems if we race
 336          * with another gns-mounting thread.
 337          */
 338         while (!list_empty(&check_list)) {
 339                 mnt = list_entry(check_list.prev,
 340                                  struct vfsmount,
 341                                  mnt_lustre_list);
 342
 343                 mntget(mnt);
 344
 345                 list_del_init(&mnt->mnt_lustre_list);
 346
 347                 list_add(&mnt->mnt_lustre_list,
 348                          &sbi->ll_mnt_list);
 349
 350                 /* check for timeout if needed */
 351                 pass = jiffies - mnt->mnt_last_used;
 352
 353                 if (flags == LL_GNS_CHECK &&
 354                     pass < sbi->ll_gns_timeout * HZ)
 355                 {
 356                         mntput(mnt);
 357                         continue;
 358                 }
 359                 spin_unlock(&dcache_lock);
 360
 361                 /* umounting @mnt */
 362                 ll_gns_umount_object(mnt);
 363
 364                 mntput(mnt);
 365                 spin_lock(&dcache_lock);
 366         }
 367         spin_unlock(&dcache_lock);
 368         RETURN(0);
 369 }
 370
 371 /*
 372  * GNS timer callback function. It restarts gns timer and wakes up GNS control
 373  * thread to process mounts list.
 374  */
 375 void ll_gns_timer_callback(unsigned long data)
 376 {
 377         struct ll_sb_info *sbi = (void *)data;
 378         ENTRY;
 379
 380         spin_lock(&gns_lock);
 381         if (list_empty(&sbi->ll_gns_sbi_head))
 382                 list_add(&sbi->ll_gns_sbi_head, &gns_sbi_list);
 383         spin_unlock(&gns_lock);
 384
 385         wake_up(&gns_thread.t_ctl_waitq);
 386         mod_timer(&sbi->ll_gns_timer,
 387                   jiffies + sbi->ll_gns_tick * HZ);
 388 }
 389
 390 /* this function checks if something new happened to exist in gns list. */
 391 static int inline ll_gns_check_event(void)
 392 {
 393         int rc;
 394
 395         spin_lock(&gns_lock);
 396         rc = !list_empty(&gns_sbi_list);
 397         spin_unlock(&gns_lock);
 398
 399         return rc;
 400 }
 401
 402 /* should we stop GNS control thread? */
 403 static int inline ll_gns_check_stop(void)
 404 {
 405         mb();
 406         return (gns_thread.t_flags & SVC_STOPPING) ? 1 : 0;
 407 }
 408
 409 /* GNS control thread function. */
 410 static int ll_gns_thread_main(void *arg)
 411 {
 412         struct ll_gns_ctl *ctl = arg;
 413         unsigned long flags;
 414         ENTRY;
 415
 416         {
 417                 char name[sizeof(current->comm)];
 418                 snprintf(name, sizeof(name) - 1, "ll_gns");
 419                 kportal_daemonize(name);
 420         }
 421
 422         SIGNAL_MASK_LOCK(current, flags);
 423         sigfillset(&current->blocked);
 424         RECALC_SIGPENDING;
 425         SIGNAL_MASK_UNLOCK(current, flags);
 426
 427         /*
 428          * letting starting function know, that we are ready and control may be
 429          * returned.
 430          */
 431         gns_thread.t_flags = SVC_RUNNING;
 432         complete(&ctl->gc_starting);
 433
 434         while (!ll_gns_check_stop()) {
 435                 struct l_wait_info lwi = { 0 };
 436
 437                 l_wait_event(gns_thread.t_ctl_waitq,
 438                              (ll_gns_check_event() ||
 439                               ll_gns_check_stop()), &lwi);
 440
 441                 spin_lock(&gns_lock);
 442                 while (!list_empty(&gns_sbi_list)) {
 443                         struct ll_sb_info *sbi;
 444
 445                         sbi = list_entry(gns_sbi_list.prev,
 446                                          struct ll_sb_info,
 447                                          ll_gns_sbi_head);
 448
 449                         list_del_init(&sbi->ll_gns_sbi_head);
 450                         spin_unlock(&gns_lock);
 451                         ll_gns_check_mounts(sbi, LL_GNS_CHECK);
 452                         spin_lock(&gns_lock);
 453                 }
 454                 spin_unlock(&gns_lock);
 455         }
 456
 457         EXIT;
 458         gns_thread.t_flags = SVC_STOPPED;
 459
 460         /* this is SMP-safe way to finish thread. */
 461         complete_and_exit(&ctl->gc_finishing, 0);
 462 }
 463
 464 void ll_gns_add_timer(struct ll_sb_info *sbi)
 465 {
 466         mod_timer(&sbi->ll_gns_timer,
 467                   jiffies + sbi->ll_gns_tick * HZ);
 468 }
 469
 470 void ll_gns_del_timer(struct ll_sb_info *sbi)
 471 {
 472         del_timer(&sbi->ll_gns_timer);
 473 }
 474
 475 /*
 476  * starts GNS control thread and waits for a signal it is up and work may be
 477  * continued.
 478  */
 479 int ll_gns_start_thread(void)
 480 {
 481         int rc;
 482         ENTRY;
 483
 484         LASSERT(gns_thread.t_flags == 0);
 485         init_completion(&gns_ctl.gc_starting);
 486         init_completion(&gns_ctl.gc_finishing);
 487         init_waitqueue_head(&gns_thread.t_ctl_waitq);
 488
 489         rc = kernel_thread(ll_gns_thread_main, &gns_ctl,
 490                            (CLONE_VM | CLONE_FILES));
 491         if (rc < 0) {
 492                 CERROR("cannot start GNS control thread, "
 493                        "err = %d\n", rc);
 494                 RETURN(rc);
 495         }
 496         wait_for_completion(&gns_ctl.gc_starting);
 497         LASSERT(gns_thread.t_flags == SVC_RUNNING);
 498         RETURN(0);
 499 }
 500
 501 /* stops GNS control thread and waits its actual stop. */
 502 void ll_gns_stop_thread(void)
 503 {
 504         ENTRY;
 505         gns_thread.t_flags = SVC_STOPPING;
 506         wake_up(&gns_thread.t_ctl_waitq);
 507         wait_for_completion(&gns_ctl.gc_finishing);
 508         LASSERT(gns_thread.t_flags == SVC_STOPPED);
 509         gns_thread.t_flags = 0;
 510         EXIT;
 511 }