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                 CWARN("failed to call GNS upcall %s, err = %d, "
 286                       "checking for mount anyway\n", sbi->ll_gns_upcall, rc);
 287         }
 288
 289         /*
 290          * waiting for dentry become mount point GNS_WAIT_ATTEMPTS times by 1
 291          * second.
 292          */
 293         rc = ll_gns_wait_for_mount(dentry, 1, GNS_WAIT_ATTEMPTS);
 294         complete_all(&sbi->ll_gns_mount_finished);
 295         if (rc == 0) {
 296                 struct dentry *rdentry;
 297                 struct vfsmount *rmnt;
 298
 299                 /* mount is successful */
 300                 LASSERT(sbi->ll_gns_state == LL_GNS_FINISHED);
 301
 302                 rmnt = mntget(mnt);
 303                 rdentry = dget(dentry);
 304
 305                 if (follow_down(&rmnt, &rdentry)) {
 306                         /*
 307                          * registering new mount in GNS mounts list and thus
 308                          * make it accessible from GNS control thread.
 309                          */
 310                         spin_lock(&dcache_lock);
 311                         LASSERT(list_empty(&rmnt->mnt_lustre_list));
 312                         list_add_tail(&rmnt->mnt_lustre_list,
 313                                       &sbi->ll_mnt_list);
 314                         spin_unlock(&dcache_lock);
 315                         rmnt->mnt_last_used = jiffies;
 316                         mntput(rmnt);
 317                         dput(rdentry);
 318                 } else {
 319                         mntput(mnt);
 320                         dput(dentry);
 321                 }
 322         } else {
 323                 CERROR("usermode upcall %s failed to mount %s, err %d\n",
 324                        sbi->ll_gns_upcall, path, rc);
 325         }
 326
 327         EXIT;
 328 cleanup:
 329         switch (cleanup_phase) {
 330         case 4:
 331                 free_page((unsigned long)datapage);
 332         case 3:
 333                 if (mntinfo_fd != NULL)
 334                         fput(mntinfo_fd);
 335         case 2:
 336                 if (dchild != NULL)
 337                         dput(dchild);
 338         case 1:
 339                 free_page((unsigned long)pathpage);
 340                 complete_all(&sbi->ll_gns_mount_finished);
 341         case 0:
 342                 spin_lock(&sbi->ll_gns_lock);
 343                 sbi->ll_gns_state = LL_GNS_IDLE;
 344                 spin_unlock(&sbi->ll_gns_lock);
 345         }
 346         return rc;
 347 }
 348
 349 /* tries to umount passed @mnt. */
 350 int ll_gns_umount_object(struct vfsmount *mnt)
 351 {
 352         int rc = 0;
 353         ENTRY;
 354
 355         CDEBUG(D_INODE, "unmounting mnt %p\n", mnt);
 356         rc = do_umount(mnt, 0);
 357         if (rc) {
 358                 CDEBUG(D_INODE, "can't umount 0x%p, err = %d\n",
 359                        mnt, rc);
 360         }
 361
 362         RETURN(rc);
 363 }
 364
 365 int ll_gns_check_mounts(struct ll_sb_info *sbi, int flags)
 366 {
 367         struct list_head check_list = LIST_HEAD_INIT(check_list);
 368         struct vfsmount *mnt;
 369         unsigned long pass;
 370         ENTRY;
 371
 372         spin_lock(&dcache_lock);
 373         list_splice_init(&sbi->ll_mnt_list, &check_list);
 374
 375         /*
 376          * walk the list in reverse order, and put them on the front of the sbi
 377          * list each iteration; this avoids list-ordering problems if we race
 378          * with another gns-mounting thread.
 379          */
 380         while (!list_empty(&check_list)) {
 381                 mnt = list_entry(check_list.prev,
 382                                  struct vfsmount,
 383                                  mnt_lustre_list);
 384
 385                 mntget(mnt);
 386
 387                 list_del_init(&mnt->mnt_lustre_list);
 388
 389                 list_add(&mnt->mnt_lustre_list,
 390                          &sbi->ll_mnt_list);
 391
 392                 /* check for timeout if needed */
 393                 pass = jiffies - mnt->mnt_last_used;
 394
 395                 if (flags == LL_GNS_CHECK &&
 396                     pass < sbi->ll_gns_timeout * HZ)
 397                 {
 398                         mntput(mnt);
 399                         continue;
 400                 }
 401                 spin_unlock(&dcache_lock);
 402
 403                 /* umounting @mnt */
 404                 ll_gns_umount_object(mnt);
 405
 406                 mntput(mnt);
 407                 spin_lock(&dcache_lock);
 408         }
 409         spin_unlock(&dcache_lock);
 410         RETURN(0);
 411 }
 412
 413 /*
 414  * GNS timer callback function. It restarts gns timer and wakes up GNS control
 415  * thread to process mounts list.
 416  */
 417 void ll_gns_timer_callback(unsigned long data)
 418 {
 419         struct ll_sb_info *sbi = (void *)data;
 420         ENTRY;
 421
 422         spin_lock(&gns_lock);
 423         if (list_empty(&sbi->ll_gns_sbi_head))
 424                 list_add(&sbi->ll_gns_sbi_head, &gns_sbi_list);
 425         spin_unlock(&gns_lock);
 426
 427         wake_up(&gns_thread.t_ctl_waitq);
 428         mod_timer(&sbi->ll_gns_timer,
 429                   jiffies + sbi->ll_gns_tick * HZ);
 430 }
 431
 432 /* this function checks if something new happened to exist in gns list. */
 433 static int inline ll_gns_check_event(void)
 434 {
 435         int rc;
 436
 437         spin_lock(&gns_lock);
 438         rc = !list_empty(&gns_sbi_list);
 439         spin_unlock(&gns_lock);
 440
 441         return rc;
 442 }
 443
 444 /* should we stop GNS control thread? */
 445 static int inline ll_gns_check_stop(void)
 446 {
 447         mb();
 448         return (gns_thread.t_flags & SVC_STOPPING) ? 1 : 0;
 449 }
 450
 451 /* GNS control thread function. */
 452 static int ll_gns_thread_main(void *arg)
 453 {
 454         struct ll_gns_ctl *ctl = arg;
 455         unsigned long flags;
 456         ENTRY;
 457
 458         {
 459                 char name[sizeof(current->comm)];
 460                 snprintf(name, sizeof(name) - 1, "ll_gns");
 461                 kportal_daemonize(name);
 462         }
 463
 464         SIGNAL_MASK_LOCK(current, flags);
 465         sigfillset(&current->blocked);
 466         RECALC_SIGPENDING;
 467         SIGNAL_MASK_UNLOCK(current, flags);
 468
 469         /*
 470          * letting starting function know, that we are ready and control may be
 471          * returned.
 472          */
 473         gns_thread.t_flags = SVC_RUNNING;
 474         complete(&ctl->gc_starting);
 475
 476         while (!ll_gns_check_stop()) {
 477                 struct l_wait_info lwi = { 0 };
 478
 479                 l_wait_event(gns_thread.t_ctl_waitq,
 480                              (ll_gns_check_event() ||
 481                               ll_gns_check_stop()), &lwi);
 482
 483                 spin_lock(&gns_lock);
 484                 while (!list_empty(&gns_sbi_list)) {
 485                         struct ll_sb_info *sbi;
 486
 487                         sbi = list_entry(gns_sbi_list.prev,
 488                                          struct ll_sb_info,
 489                                          ll_gns_sbi_head);
 490
 491                         list_del_init(&sbi->ll_gns_sbi_head);
 492                         spin_unlock(&gns_lock);
 493                         ll_gns_check_mounts(sbi, LL_GNS_CHECK);
 494                         spin_lock(&gns_lock);
 495                 }
 496                 spin_unlock(&gns_lock);
 497         }
 498
 499         EXIT;
 500         gns_thread.t_flags = SVC_STOPPED;
 501
 502         /* this is SMP-safe way to finish thread. */
 503         complete_and_exit(&ctl->gc_finishing, 0);
 504 }
 505
 506 void ll_gns_add_timer(struct ll_sb_info *sbi)
 507 {
 508         mod_timer(&sbi->ll_gns_timer,
 509                   jiffies + sbi->ll_gns_tick * HZ);
 510 }
 511
 512 void ll_gns_del_timer(struct ll_sb_info *sbi)
 513 {
 514         del_timer(&sbi->ll_gns_timer);
 515 }
 516
 517 /*
 518  * starts GNS control thread and waits for a signal it is up and work may be
 519  * continued.
 520  */
 521 int ll_gns_start_thread(void)
 522 {
 523         int rc;
 524         ENTRY;
 525
 526         LASSERT(gns_thread.t_flags == 0);
 527         init_completion(&gns_ctl.gc_starting);
 528         init_completion(&gns_ctl.gc_finishing);
 529         init_waitqueue_head(&gns_thread.t_ctl_waitq);
 530
 531         rc = kernel_thread(ll_gns_thread_main, &gns_ctl,
 532                            (CLONE_VM | CLONE_FILES));
 533         if (rc < 0) {
 534                 CERROR("cannot start GNS control thread, "
 535                        "err = %d\n", rc);
 536                 RETURN(rc);
 537         }
 538         wait_for_completion(&gns_ctl.gc_starting);
 539         LASSERT(gns_thread.t_flags == SVC_RUNNING);
 540         RETURN(0);
 541 }
 542
 543 /* stops GNS control thread and waits its actual stop. */
 544 void ll_gns_stop_thread(void)
 545 {
 546         ENTRY;
 547         gns_thread.t_flags = SVC_STOPPING;
 548         wake_up(&gns_thread.t_ctl_waitq);
 549         wait_for_completion(&gns_ctl.gc_finishing);
 550         LASSERT(gns_thread.t_flags == SVC_STOPPED);
 551         gns_thread.t_flags = 0;
 552         EXIT;
 553 }