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