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 #define CONCUR_GNS_RESTART_APPROACH 0
  44
  45 /*
  46  * waits until passed dentry gets mountpoint or timeout and attempts are
  47  * exhausted. Returns 1 if dentry became mountpoint and 0 otherwise.
  48  */
  49 static int
  50 ll_gns_wait_for_mount(struct dentry *dentry,
  51                       int timeout, int tries)
  52 {
  53         struct l_wait_info lwi;
  54         struct ll_sb_info *sbi;
  55         ENTRY;
  56
  57         LASSERT(dentry != NULL);
  58         LASSERT(!IS_ERR(dentry));
  59         sbi = ll_s2sbi(dentry->d_sb);
  60
  61         lwi = LWI_TIMEOUT(timeout * HZ, NULL, NULL);
  62         for (; !d_mountpoint(dentry) && tries > 0; tries--)
  63                 l_wait_event(sbi->ll_gns_waitq, d_mountpoint(dentry), &lwi);
  64
  65         if (d_mountpoint(dentry)) {
  66                 spin_lock(&sbi->ll_gns_lock);
  67                 sbi->ll_gns_state = LL_GNS_FINISHED;
  68                 spin_unlock(&sbi->ll_gns_lock);
  69                 RETURN(0);
  70         }
  71         RETURN(-ETIME);
  72 }
  73
  74 #if (CONCUR_GNS_RESTART_APPROACH == 1)
  75 /*
  76  * sending a signal known to be ignored to cause restarting syscall if GNS mount
  77  * function returns -ERESTARTSYS.
  78  */
  79 static void
  80 ll_gns_send_signal(void)
  81 {
  82         struct task_struct *task = current;
  83         int signal = SIGCONT;
  84
  85         read_lock(&tasklist_lock);
  86         spin_lock_irq(&task->sighand->siglock);
  87         sigaddset(&task->pending.signal, signal);
  88         spin_unlock_irq(&task->sighand->siglock);
  89         read_unlock(&tasklist_lock);
  90         set_tsk_thread_flag(task, TIF_SIGPENDING);
  91 }
  92 #endif
  93
  94 /*
  95  * tries to mount the mount object under passed @dentry. In the case of success
  96  * @dentry will become mount point and 0 will be returned. Error code will be
  97  * returned otherwise.
  98  */
  99 int
 100 ll_gns_mount_object(struct dentry *dentry, struct vfsmount *mnt)
 101 {
 102         char *path, *pathpage, *datapage, *argv[4];
 103         struct file *mntinfo_fd = NULL;
 104         int cleanup_phase = 0, rc = 0;
 105         struct ll_sb_info *sbi;
 106         struct dentry *dchild;
 107         ENTRY;
 108
 109         LASSERT(dentry->d_inode != NULL);
 110
 111         if (!S_ISDIR(dentry->d_inode->i_mode))
 112                 RETURN(-EINVAL);
 113
 114         sbi = ll_i2sbi(dentry->d_inode);
 115
 116         if (mnt == NULL) {
 117                 CERROR("suid directory found, but no "
 118                        "vfsmount available.\n");
 119                 RETURN(-EINVAL);
 120         }
 121
 122         if (atomic_read(&sbi->ll_gns_enabled) == 0)
 123                 RETURN(-EINVAL);
 124
 125         spin_lock(&sbi->ll_gns_lock);
 126
 127         /*
 128          * another thead is in progress or just finished mounting the
 129          * dentry. Handling that.
 130          */
 131         if (sbi->ll_gns_state == LL_GNS_MOUNTING ||
 132             sbi->ll_gns_state == LL_GNS_FINISHED) {
 133                 /*
 134                  * another thread is trying to mount GNS dentry. We'd like to
 135                  * handling that.
 136                  */
 137                 spin_unlock(&sbi->ll_gns_lock);
 138
 139                 /*
 140                  * check if dentry is mount point already, if so, do not restart
 141                  * syscal.
 142                  */
 143                 if (d_mountpoint(dentry))
 144                         RETURN(0);
 145
 146 #if (CONCUR_GNS_RESTART_APPROACH == 1)
 147                 /*
 148                  * causing syscall to restart and possibly find this dentry
 149                  * already mounted.
 150                  */
 151                 ll_gns_send_signal();
 152                 RETURN(-ERESTARTSYS);
 153 #else
 154                 /*
 155                  * waiting for GNS complete and check dentry again, it may be
 156                  * mounted already.
 157                  */
 158                 wait_for_completion(&sbi->ll_gns_mount_finished);
 159                 if (d_mountpoint(dentry))
 160                         RETURN(0);
 161 #endif
 162         }
 163         LASSERT(sbi->ll_gns_state == LL_GNS_IDLE);
 164
 165         CDEBUG(D_INODE, "mounting dentry %p\n", dentry);
 166
 167         /* mounting started */
 168         sbi->ll_gns_state = LL_GNS_MOUNTING;
 169         spin_unlock(&sbi->ll_gns_lock);
 170
 171         /* we need to build an absolute pathname to pass to mount */
 172         pathpage = (char *)__get_free_page(GFP_KERNEL);
 173         if (!pathpage)
 174                 GOTO(cleanup, rc = -ENOMEM);
 175         cleanup_phase = 1;
 176
 177         /* getting @dentry path stored in @pathpage. */
 178         path = d_path(dentry, mnt, pathpage, PAGE_SIZE);
 179         if (IS_ERR(path)) {
 180                 CERROR("can't build mount object path, err %d\n",
 181                        (int)PTR_ERR(dchild));
 182                 GOTO(cleanup, rc = PTR_ERR(dchild));
 183         }
 184
 185         /* synchronizing with possible /proc/fs/...write */
 186         down(&sbi->ll_gns_sem);
 187
 188         /*
 189          * mount object name is taken from sbi, where it is set in mount time or
 190          * via /proc/fs... tunable. It may be ".mntinfo" or so.
 191          */
 192
 193         /*
 194          * recursive lookup with trying to mount SUID bit marked directories on
 195          * the way is not possible here, as lookup_one_len() does not pass @nd
 196          * to ->lookup() and this is checked in ll_lookup_it(). So, do not
 197          * handle possible -EAGAIN here.
 198          */
 199         dchild = ll_lookup_one_len(sbi->ll_gns_oname, dentry,
 200                                    strlen(sbi->ll_gns_oname));
 201         up(&sbi->ll_gns_sem);
 202
 203         cleanup_phase = 2;
 204
 205         if (IS_ERR(dchild)) {
 206                 rc = PTR_ERR(dchild);
 207                 CERROR("can't find mount object %*s/%*s err = %d.\n",
 208                        (int)dentry->d_name.len, dentry->d_name.name,
 209                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 210                        rc);
 211                 GOTO(cleanup, rc);
 212         }
 213
 214         /* mount object is not found */
 215         if (!dchild->d_inode)
 216                 GOTO(cleanup, rc = -ENOENT);
 217
 218         /* check if found child is regular file */
 219         if (!S_ISREG(dchild->d_inode->i_mode))
 220                 GOTO(cleanup, rc = -EBADF);
 221
 222         mntget(mnt);
 223
 224         /* ok, mount object if found, opening it. */
 225         mntinfo_fd = dentry_open(dchild, mnt, 0);
 226         if (IS_ERR(mntinfo_fd)) {
 227                 CERROR("can't open mount object %*s/%*s err = %d.\n",
 228                        (int)dentry->d_name.len, dentry->d_name.name,
 229                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 230                        (int)PTR_ERR(mntinfo_fd));
 231                 mntput(mnt);
 232                 GOTO(cleanup, rc = PTR_ERR(mntinfo_fd));
 233         }
 234         cleanup_phase = 3;
 235
 236         if (mntinfo_fd->f_dentry->d_inode->i_size > PAGE_SIZE - 1) {
 237                 CERROR("mount object %*s/%*s is too big (%Ld)\n",
 238                        (int)dentry->d_name.len, dentry->d_name.name,
 239                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 240                        mntinfo_fd->f_dentry->d_inode->i_size);
 241                 GOTO(cleanup, rc = -EFBIG);
 242         }
 243
 244         datapage = (char *)__get_free_page(GFP_KERNEL);
 245         if (!datapage)
 246                 GOTO(cleanup, rc = -ENOMEM);
 247
 248         cleanup_phase = 4;
 249
 250         /* read data from mount object. */
 251         rc = kernel_read(mntinfo_fd, 0, datapage, PAGE_SIZE - 1);
 252         if (rc < 0) {
 253                 CERROR("can't read mount object %*s/%*s data, err %d\n",
 254                        (int)dentry->d_name.len, dentry->d_name.name,
 255                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
 256                        rc);
 257                 GOTO(cleanup, rc);
 258         }
 259
 260         /* no data in mount object? */
 261         if (rc == 0) {
 262                 CERROR("mount object %*s/%*s is empty?\n",
 263                        (int)dentry->d_name.len, dentry->d_name.name,
 264                        strlen(sbi->ll_gns_oname), sbi->ll_gns_oname);
 265                 GOTO(cleanup, rc);
 266         }
 267
 268         datapage[rc] = '\0';
 269         fput(mntinfo_fd);
 270         mntinfo_fd = NULL;
 271         dchild = NULL;
 272
 273         /* synchronizing with possible /proc/fs/...write */
 274         down(&sbi->ll_gns_sem);
 275
 276         /*
 277          * upcall is initialized in mount time or via /proc/fs/... tuneable and
 278          * may be /usr/lib/lustre/gns-upcall.sh
 279          */
 280         argv[0] = sbi->ll_gns_upcall;
 281         argv[1] = datapage;
 282         argv[2] = path;
 283         argv[3] = NULL;
 284
 285         up(&sbi->ll_gns_sem);
 286
 287         /* do not wait for helper complete here. */
 288         rc = call_usermodehelper(argv[0], argv, NULL, 1);
 289         if (rc) {
 290                 CWARN("failed to call GNS upcall %s, err = %d, "
 291                       "checking for mount anyway\n", sbi->ll_gns_upcall, rc);
 292         }
 293
 294         /*
 295          * waiting for dentry become mount point GNS_WAIT_ATTEMPTS times by 1
 296          * second.
 297          */
 298         rc = ll_gns_wait_for_mount(dentry, 1, GNS_WAIT_ATTEMPTS);
 299         if (rc == 0) {
 300                 struct dentry *rdentry;
 301                 struct vfsmount *rmnt;
 302
 303                 LASSERT(sbi->ll_gns_state == LL_GNS_FINISHED);
 304
 305                 rmnt = mntget(mnt);
 306                 rdentry = dget(dentry);
 307
 308                 if (follow_down(&rmnt, &rdentry)) {
 309                         /*
 310                          * registering new mount in GNS mounts list and thus
 311                          * make it accessible from GNS control thread.
 312                          */
 313                         spin_lock(&dcache_lock);
 314                         LASSERT(list_empty(&rmnt->mnt_lustre_list));
 315                         list_add_tail(&rmnt->mnt_lustre_list,
 316                                       &sbi->ll_mnt_list);
 317                         spin_unlock(&dcache_lock);
 318                         rmnt->mnt_last_used = jiffies;
 319                         mntput(rmnt);
 320                         dput(rdentry);
 321                 } else {
 322                         mntput(mnt);
 323                         dput(dentry);
 324                 }
 325         } else {
 326                 CERROR("usermode upcall %s failed to mount %s, err %d\n",
 327                        sbi->ll_gns_upcall, path, rc);
 328         }
 329
 330         EXIT;
 331 cleanup:
 332         switch (cleanup_phase) {
 333         case 4:
 334                 free_page((unsigned long)datapage);
 335         case 3:
 336                 if (mntinfo_fd != NULL)
 337                         fput(mntinfo_fd);
 338         case 2:
 339                 if (dchild != NULL)
 340                         dput(dchild);
 341         case 1:
 342                 free_page((unsigned long)pathpage);
 343         case 0:
 344                 spin_lock(&sbi->ll_gns_lock);
 345                 sbi->ll_gns_state = LL_GNS_IDLE;
 346                 spin_unlock(&sbi->ll_gns_lock);
 347                 complete_all(&sbi->ll_gns_mount_finished);
 348         }
 349         return rc;
 350 }
 351
 352 /* tries to umount passed @mnt. */
 353 int ll_gns_umount_object(struct vfsmount *mnt)
 354 {
 355         int rc = 0;
 356         ENTRY;
 357
 358         CDEBUG(D_INODE, "unmounting mnt %p\n", mnt);
 359         rc = do_umount(mnt, 0);
 360         if (rc) {
 361                 CDEBUG(D_INODE, "can't umount 0x%p, err = %d\n",
 362                        mnt, rc);
 363         }
 364
 365         RETURN(rc);
 366 }
 367
 368 int ll_gns_check_mounts(struct ll_sb_info *sbi, int flags)
 369 {
 370         struct list_head check_list = LIST_HEAD_INIT(check_list);
 371         struct vfsmount *mnt;
 372         unsigned long pass;
 373         ENTRY;
 374
 375         spin_lock(&dcache_lock);
 376         list_splice_init(&sbi->ll_mnt_list, &check_list);
 377
 378         /*
 379          * walk the list in reverse order, and put them on the front of the sbi
 380          * list each iteration; this avoids list-ordering problems if we race
 381          * with another gns-mounting thread.
 382          */
 383         while (!list_empty(&check_list)) {
 384                 mnt = list_entry(check_list.prev,
 385                                  struct vfsmount,
 386                                  mnt_lustre_list);
 387
 388                 mntget(mnt);
 389
 390                 list_del_init(&mnt->mnt_lustre_list);
 391
 392                 list_add(&mnt->mnt_lustre_list,
 393                          &sbi->ll_mnt_list);
 394
 395                 /* check for timeout if needed */
 396                 pass = jiffies - mnt->mnt_last_used;
 397
 398                 if (flags == LL_GNS_CHECK &&
 399                     pass < sbi->ll_gns_timeout * HZ)
 400                 {
 401                         mntput(mnt);
 402                         continue;
 403                 }
 404                 spin_unlock(&dcache_lock);
 405
 406                 /* umounting @mnt */
 407                 ll_gns_umount_object(mnt);
 408
 409                 mntput(mnt);
 410                 spin_lock(&dcache_lock);
 411         }
 412         spin_unlock(&dcache_lock);
 413         RETURN(0);
 414 }
 415
 416 /*
 417  * GNS timer callback function. It restarts gns timer and wakes up GNS control
 418  * thread to process mounts list.
 419  */
 420 void ll_gns_timer_callback(unsigned long data)
 421 {
 422         struct ll_sb_info *sbi = (void *)data;
 423         ENTRY;
 424
 425         spin_lock(&gns_lock);
 426         if (list_empty(&sbi->ll_gns_sbi_head))
 427                 list_add(&sbi->ll_gns_sbi_head, &gns_sbi_list);
 428         spin_unlock(&gns_lock);
 429
 430         wake_up(&gns_thread.t_ctl_waitq);
 431         mod_timer(&sbi->ll_gns_timer,
 432                   jiffies + sbi->ll_gns_tick * HZ);
 433 }
 434
 435 /* this function checks if something new happened to exist in gns list. */
 436 static int inline ll_gns_check_event(void)
 437 {
 438         int rc;
 439
 440         spin_lock(&gns_lock);
 441         rc = !list_empty(&gns_sbi_list);
 442         spin_unlock(&gns_lock);
 443
 444         return rc;
 445 }
 446
 447 /* should we stop GNS control thread? */
 448 static int inline ll_gns_check_stop(void)
 449 {
 450         mb();
 451         return (gns_thread.t_flags & SVC_STOPPING) ? 1 : 0;
 452 }
 453
 454 /* GNS control thread function. */
 455 static int ll_gns_thread_main(void *arg)
 456 {
 457         struct ll_gns_ctl *ctl = arg;
 458         unsigned long flags;
 459         ENTRY;
 460
 461         {
 462                 char name[sizeof(current->comm)];
 463                 snprintf(name, sizeof(name) - 1, "ll_gns");
 464                 kportal_daemonize(name);
 465         }
 466
 467         SIGNAL_MASK_LOCK(current, flags);
 468         sigfillset(&current->blocked);
 469         RECALC_SIGPENDING;
 470         SIGNAL_MASK_UNLOCK(current, flags);
 471
 472         /*
 473          * letting starting function know, that we are ready and control may be
 474          * returned.
 475          */
 476         gns_thread.t_flags = SVC_RUNNING;
 477         complete(&ctl->gc_starting);
 478
 479         while (!ll_gns_check_stop()) {
 480                 struct l_wait_info lwi = { 0 };
 481
 482                 l_wait_event(gns_thread.t_ctl_waitq,
 483                              (ll_gns_check_event() ||
 484                               ll_gns_check_stop()), &lwi);
 485
 486                 spin_lock(&gns_lock);
 487                 while (!list_empty(&gns_sbi_list)) {
 488                         struct ll_sb_info *sbi;
 489
 490                         sbi = list_entry(gns_sbi_list.prev,
 491                                          struct ll_sb_info,
 492                                          ll_gns_sbi_head);
 493
 494                         list_del_init(&sbi->ll_gns_sbi_head);
 495                         spin_unlock(&gns_lock);
 496                         ll_gns_check_mounts(sbi, LL_GNS_CHECK);
 497                         spin_lock(&gns_lock);
 498                 }
 499                 spin_unlock(&gns_lock);
 500         }
 501
 502         EXIT;
 503         gns_thread.t_flags = SVC_STOPPED;
 504
 505         /* this is SMP-safe way to finish thread. */
 506         complete_and_exit(&ctl->gc_finishing, 0);
 507 }
 508
 509 void ll_gns_add_timer(struct ll_sb_info *sbi)
 510 {
 511         mod_timer(&sbi->ll_gns_timer,
 512                   jiffies + sbi->ll_gns_tick * HZ);
 513 }
 514
 515 void ll_gns_del_timer(struct ll_sb_info *sbi)
 516 {
 517         del_timer(&sbi->ll_gns_timer);
 518 }
 519
 520 /*
 521  * starts GNS control thread and waits for a signal it is up and work may be
 522  * continued.
 523  */
 524 int ll_gns_start_thread(void)
 525 {
 526         int rc;
 527         ENTRY;
 528
 529         LASSERT(gns_thread.t_flags == 0);
 530         init_completion(&gns_ctl.gc_starting);
 531         init_completion(&gns_ctl.gc_finishing);
 532         init_waitqueue_head(&gns_thread.t_ctl_waitq);
 533
 534         rc = kernel_thread(ll_gns_thread_main, &gns_ctl,
 535                            (CLONE_VM | CLONE_FILES));
 536         if (rc < 0) {
 537                 CERROR("cannot start GNS control thread, "
 538                        "err = %d\n", rc);
 539                 RETURN(rc);
 540         }
 541         wait_for_completion(&gns_ctl.gc_starting);
 542         LASSERT(gns_thread.t_flags == SVC_RUNNING);
 543         RETURN(0);
 544 }
 545
 546 /* stops GNS control thread and waits its actual stop. */
 547 void ll_gns_stop_thread(void)
 548 {
 549         ENTRY;
 550         gns_thread.t_flags = SVC_STOPPING;
 551         wake_up(&gns_thread.t_ctl_waitq);
 552         wait_for_completion(&gns_ctl.gc_finishing);
 553         LASSERT(gns_thread.t_flags == SVC_STOPPED);
 554         gns_thread.t_flags = 0;
 555         EXIT;
 556 }