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 Cluster File Systems, Inc.
   5  *   Author: Phil Schwan <phil@clusterfs.com>
   6  *
   7  *   This file is part of Lustre, http://www.lustre.org.
   8  *
   9  *   Lustre is free software; you can redistribute it and/or
  10  *   modify it under the terms of version 2 of the GNU General Public
  11  *   License as published by the Free Software Foundation.
  12  *
  13  *   Lustre is distributed in the hope that it will be useful,
  14  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  *   GNU General Public License for more details.
  17  *
  18  *   You should have received a copy of the GNU General Public License
  19  *   along with Lustre; if not, write to the Free Software
  20  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  */
  22
  23 #define DEBUG_SUBSYSTEM S_LLITE
  24
  25 #include <linux/fs.h>
  26 #include <linux/version.h>
  27 #include <asm/uaccess.h>
  28 #include <linux/file.h>
  29 #include <linux/kmod.h>
  30
  31 #include <linux/lustre_lite.h>
  32 #include "llite_internal.h"
  33
  34 static struct list_head gns_sbi_list = LIST_HEAD_INIT(gns_sbi_list);
  35 static spinlock_t gns_lock = SPIN_LOCK_UNLOCKED;
  36 static struct ptlrpc_thread gns_thread;
  37 static struct ll_gns_ctl gns_ctl;
  38
  39 /*
  40  * waits until passed dentry gets mountpoint or timeout and attempts are
  41  * exhausted. Returns 1 if dentry became mountpoint and 0 otherwise.
  42  */
  43 static int
  44 ll_gns_wait_for_mount(struct dentry *dentry,
  45                       int timeout, int tries)
  46 {
  47         struct l_wait_info lwi;
  48         struct ll_sb_info *sbi;
  49         int rc;
  50         ENTRY;
  51
  52         LASSERT(dentry != NULL);
  53         LASSERT(!IS_ERR(dentry));
  54         sbi = ll_s2sbi(dentry->d_sb);
  55
  56         for (; !d_mountpoint(dentry) && tries > 0; tries--) {
  57                 lwi = LWI_TIMEOUT(timeout * HZ, NULL, NULL);
  58                 l_wait_event(sbi->ll_gns_waitq, d_mountpoint(dentry), &lwi);
  59         }
  60
  61         if ((rc = d_mountpoint(dentry) ? 1 : 0)) {
  62                 spin_lock(&sbi->ll_gns_lock);
  63                 LASSERT(sbi->ll_gns_state == LL_GNS_MOUNTING);
  64                 sbi->ll_gns_state = LL_GNS_FINISHED;
  65                 spin_unlock(&sbi->ll_gns_lock);
  66         }
  67
  68         complete(&sbi->ll_gns_mount_finished);
  69         RETURN(rc);
  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 retuned. Error code will be
  75  * returned otherwise.
  76  */
  77 int ll_gns_mount_object(struct dentry *dentry,
  78                         struct vfsmount *mnt)
  79 {
  80         struct ll_dentry_data *lld = dentry->d_fsdata;
  81         char *p, *path, *pathpage, *argv[4];
  82         struct file *mntinfo_fd = NULL;
  83         struct address_space *mapping;
  84         int cleanup_phase = 0, rc = 0;
  85         struct ll_sb_info *sbi;
  86         struct dentry *dchild;
  87         struct page *datapage;
  88         filler_t *filler;
  89         ENTRY;
  90
  91         if (mnt == NULL) {
  92                 CERROR("suid directory found, but no "
  93                        "vfsmount available.\n");
  94                 RETURN(-EINVAL);
  95         }
  96
  97         CDEBUG(D_INODE, "mounting dentry %p\n", dentry);
  98
  99         LASSERT(dentry->d_inode != NULL);
 100         LASSERT(S_ISDIR(dentry->d_inode->i_mode));
 101         LASSERT(lld != NULL);
 102
 103         sbi = ll_i2sbi(dentry->d_inode);
 104         LASSERT(sbi != NULL);
 105
 106         /* another thead is in progress of mouning some entry */
 107         spin_lock(&sbi->ll_gns_lock);
 108         if (sbi->ll_gns_state == LL_GNS_MOUNTING) {
 109                 spin_unlock(&sbi->ll_gns_lock);
 110
 111                 wait_for_completion(&sbi->ll_gns_mount_finished);
 112                 if (d_mountpoint(dentry))
 113                         RETURN(0);
 114         }
 115
 116         /* another thread mounted it already */
 117         if (sbi->ll_gns_state == LL_GNS_FINISHED) {
 118                 spin_unlock(&sbi->ll_gns_lock);
 119
 120                 /* we lost a race; just return */
 121                 if (d_mountpoint(dentry))
 122                         RETURN(0);
 123         }
 124         LASSERT(sbi->ll_gns_state == LL_GNS_IDLE);
 125
 126         spin_lock(&dentry->d_lock);
 127         dentry->d_flags |= DCACHE_GNS_MOUNTING;
 128         spin_unlock(&dentry->d_lock);
 129
 130         /* mounting started */
 131         sbi->ll_gns_state = LL_GNS_MOUNTING;
 132         spin_unlock(&sbi->ll_gns_lock);
 133
 134         /* we need to build an absolute pathname to pass to mount */
 135         pathpage = (char *)__get_free_page(GFP_KERNEL);
 136         if (!pathpage)
 137                 GOTO(cleanup, rc = -ENOMEM);
 138         cleanup_phase = 1;
 139
 140         /* getting @dentry path stored in @pathpage. */
 141         path = d_path(dentry, mnt, pathpage, PAGE_SIZE);
 142         if (IS_ERR(path)) {
 143                 CERROR("can't build mount object path, err %d\n",
 144                        (int)PTR_ERR(dchild));
 145                 GOTO(cleanup, rc = PTR_ERR(dchild));
 146         }
 147
 148         /* sychronizing with possible /proc/fs/...write */
 149         down(&sbi->ll_gns_sem);
 150
 151         /*
 152          * mount object name is taken from sbi, where it is set in mount time or
 153          * via /proc/fs... tunable. It may be ".mntinfo" or so.
 154          */
 155         dchild = ll_d_lookup(sbi->ll_gns_oname, dentry,
 156                              strlen(sbi->ll_gns_oname));
 157         up(&sbi->ll_gns_sem);
 158
 159         if (!dchild)
 160                 GOTO(cleanup, rc = -ENOENT);
 161
 162         if (IS_ERR(dchild)) {
 163                 CERROR("can't find mount object %*s/%*s err = %d.\n",
 164                        (int)dentry->d_name.len, dentry->d_name.name,
 165                        (int)dchild->d_name.len, dchild->d_name.name,
 166                        (int)PTR_ERR(dchild));
 167                 GOTO(cleanup, rc = PTR_ERR(dchild));
 168         }
 169
 170         mntget(mnt);
 171
 172         /* ok, mount object if found, opening it. */
 173         mntinfo_fd = dentry_open(dchild, mnt, 0);
 174         if (IS_ERR(mntinfo_fd)) {
 175                 CERROR("can't open mount object %*s/%*s err = %d.\n",
 176                        (int)dentry->d_name.len, dentry->d_name.name,
 177                        (int)dchild->d_name.len, dchild->d_name.name,
 178                        (int)PTR_ERR(mntinfo_fd));
 179                 dput(dchild);
 180                 mntput(mnt);
 181                 GOTO(cleanup, rc = PTR_ERR(mntinfo_fd));
 182         }
 183         cleanup_phase = 2;
 184
 185         if (mntinfo_fd->f_dentry->d_inode->i_size > PAGE_SIZE) {
 186                 CERROR("mount object %*s/%*s is too big (%Ld)\n",
 187                        (int)dentry->d_name.len, dentry->d_name.name,
 188                        (int)dchild->d_name.len, dchild->d_name.name,
 189                        mntinfo_fd->f_dentry->d_inode->i_size);
 190                 GOTO(cleanup, rc = -EFBIG);
 191         }
 192
 193         /* read data from mount object. */
 194         mapping = mntinfo_fd->f_dentry->d_inode->i_mapping;
 195         filler = (filler_t *)mapping->a_ops->readpage;
 196         datapage = read_cache_page(mapping, 0, filler,
 197                                    mntinfo_fd);
 198         if (IS_ERR(datapage)) {
 199                 CERROR("can't read data from mount object %*s/%*s\n",
 200                        (int)dentry->d_name.len, dentry->d_name.name,
 201                        (int)dchild->d_name.len, dchild->d_name.name);
 202                 GOTO(cleanup, rc = PTR_ERR(datapage));
 203         }
 204
 205         p = kmap(datapage);
 206         LASSERT(p != NULL);
 207         p[PAGE_SIZE - 1] = '\0';
 208         cleanup_phase = 3;
 209
 210         fput(mntinfo_fd);
 211         mntinfo_fd = NULL;
 212
 213         /* sychronizing with possible /proc/fs/...write */
 214         down(&sbi->ll_gns_sem);
 215
 216         /*
 217          * upcall is initialized in mount time or via /proc/fs/... tuneable and
 218          * may be /usr/lib/lustre/gns-upcall.sh
 219          */
 220         argv[0] = sbi->ll_gns_upcall;
 221         argv[1] = p;
 222         argv[2] = path;
 223         argv[3] = NULL;
 224
 225         up(&sbi->ll_gns_sem);
 226
 227         rc = USERMODEHELPER(argv[0], argv, NULL);
 228         if (rc) {
 229                 CERROR("failed to call GNS upcall %s, err = %d\n",
 230                        sbi->ll_gns_upcall, rc);
 231                 GOTO(cleanup, rc);
 232         }
 233
 234         /*
 235          * wait for mount completion. This is actually not need, because
 236          * USERMODEHELPER() returns only when usermode process finishes. But we
 237          * doing this just for case USERMODEHELPER() semanthics will be changed
 238          * or usermode upcall program will start mounting in backgound and
 239          * return instantly. --umka
 240          */
 241         if (ll_gns_wait_for_mount(dentry, 1, GNS_WAIT_ATTEMPTS)) {
 242                 struct dentry *rdentry;
 243                 struct vfsmount *rmnt;
 244
 245                 /* mount is successful */
 246                 LASSERT(sbi->ll_gns_state == LL_GNS_FINISHED);
 247
 248                 rmnt = mntget(mnt);
 249                 rdentry = dget(dentry);
 250
 251                 if (follow_down(&rmnt, &rdentry)) {
 252                         /*
 253                          * registering new mount in GNS mounts list and thus
 254                          * make it accessible from GNS control thread.
 255                          */
 256                         spin_lock(&dcache_lock);
 257                         LASSERT(list_empty(&rmnt->mnt_lustre_list));
 258                         list_add_tail(&rmnt->mnt_lustre_list,
 259                                       &sbi->ll_mnt_list);
 260                         spin_unlock(&dcache_lock);
 261                         rmnt->mnt_last_used = jiffies;
 262                         mntput(rmnt);
 263                         dput(rdentry);
 264                 } else {
 265                         mntput(mnt);
 266                         dput(dentry);
 267                 }
 268                 spin_lock(&dentry->d_lock);
 269                 dentry->d_flags &= ~DCACHE_GNS_PENDING;
 270                 spin_unlock(&dentry->d_lock);
 271         } else {
 272                 CERROR("usermode upcall %s failed to mount %s\n",
 273                        sbi->ll_gns_upcall, path);
 274                 rc = -ETIME;
 275         }
 276
 277         EXIT;
 278 cleanup:
 279         switch (cleanup_phase) {
 280         case 3:
 281                 kunmap(datapage);
 282                 page_cache_release(datapage);
 283         case 2:
 284                 if (mntinfo_fd != NULL)
 285                         fput(mntinfo_fd);
 286         case 1:
 287                 free_page((unsigned long)pathpage);
 288         case 0:
 289                 spin_lock(&sbi->ll_gns_lock);
 290                 sbi->ll_gns_state = LL_GNS_IDLE;
 291                 spin_unlock(&sbi->ll_gns_lock);
 292
 293                 spin_lock(&dentry->d_lock);
 294                 dentry->d_flags &= ~DCACHE_GNS_MOUNTING;
 295                 spin_unlock(&dentry->d_lock);
 296         }
 297         return rc;
 298 }
 299
 300 /* tries to umount passed @mnt. */
 301 int ll_gns_umount_object(struct vfsmount *mnt)
 302 {
 303         int rc = 0;
 304         ENTRY;
 305
 306         CDEBUG(D_INODE, "unmounting mnt %p\n", mnt);
 307         rc = do_umount(mnt, 0);
 308         if (rc) {
 309                 CDEBUG(D_INODE, "can't umount 0x%p, err = %d\n",
 310                        mnt, rc);
 311         }
 312
 313         RETURN(rc);
 314 }
 315
 316 int ll_gns_check_mounts(struct ll_sb_info *sbi, int flags)
 317 {
 318         struct list_head check_list = LIST_HEAD_INIT(check_list);
 319         struct vfsmount *mnt;
 320         unsigned long pass;
 321         ENTRY;
 322
 323         spin_lock(&dcache_lock);
 324         list_splice_init(&sbi->ll_mnt_list, &check_list);
 325
 326         /*
 327          * walk the list in reverse order, and put them on the front of the sbi
 328          * list each iteration; this avoids list-ordering problems if we race
 329          * with another gns-mounting thread.
 330          */
 331         while (!list_empty(&check_list)) {
 332                 mnt = list_entry(check_list.prev,
 333                                  struct vfsmount,
 334                                  mnt_lustre_list);
 335
 336                 mntget(mnt);
 337
 338                 list_del_init(&mnt->mnt_lustre_list);
 339
 340                 list_add(&mnt->mnt_lustre_list,
 341                          &sbi->ll_mnt_list);
 342
 343                 /* check for timeout if needed */
 344                 pass = jiffies - mnt->mnt_last_used;
 345
 346                 if (flags == LL_GNS_CHECK &&
 347                     pass < sbi->ll_gns_timeout * HZ)
 348                 {
 349                         mntput(mnt);
 350                         continue;
 351                 }
 352                 spin_unlock(&dcache_lock);
 353
 354                 /* umounting @mnt */
 355                 ll_gns_umount_object(mnt);
 356
 357                 mntput(mnt);
 358                 spin_lock(&dcache_lock);
 359         }
 360         spin_unlock(&dcache_lock);
 361         RETURN(0);
 362 }
 363
 364 /*
 365  * GNS timer callback function. It restarts gns timer and wakes up GNS cvontrol
 366  * thread to process mounts list.
 367  */
 368 void ll_gns_timer_callback(unsigned long data)
 369 {
 370         struct ll_sb_info *sbi = (void *)data;
 371         ENTRY;
 372
 373         spin_lock(&gns_lock);
 374         if (list_empty(&sbi->ll_gns_sbi_head))
 375                 list_add(&sbi->ll_gns_sbi_head, &gns_sbi_list);
 376         spin_unlock(&gns_lock);
 377
 378         wake_up(&gns_thread.t_ctl_waitq);
 379         mod_timer(&sbi->ll_gns_timer,
 380                   jiffies + sbi->ll_gns_tick * HZ);
 381 }
 382
 383 /* this function checks if something new happened to exist in gns list. */
 384 static int inline ll_gns_check_event(void)
 385 {
 386         int rc;
 387
 388         spin_lock(&gns_lock);
 389         rc = !list_empty(&gns_sbi_list);
 390         spin_unlock(&gns_lock);
 391
 392         return rc;
 393 }
 394
 395 /* should we staop GNS control thread? */
 396 static int inline ll_gns_check_stop(void)
 397 {
 398         mb();
 399         return (gns_thread.t_flags & SVC_STOPPING) ? 1 : 0;
 400 }
 401
 402 /* GNS control thread function. */
 403 static int ll_gns_thread_main(void *arg)
 404 {
 405         struct ll_gns_ctl *ctl = arg;
 406         unsigned long flags;
 407         ENTRY;
 408
 409         {
 410                 char name[sizeof(current->comm)];
 411                 snprintf(name, sizeof(name) - 1, "ll_gns");
 412                 kportal_daemonize(name);
 413         }
 414
 415         SIGNAL_MASK_LOCK(current, flags);
 416         sigfillset(&current->blocked);
 417         RECALC_SIGPENDING;
 418         SIGNAL_MASK_UNLOCK(current, flags);
 419
 420         /*
 421          * letting starting function know, that we are ready and control may be
 422          * returned.
 423          */
 424         gns_thread.t_flags = SVC_RUNNING;
 425         complete(&ctl->gc_starting);
 426
 427         while (!ll_gns_check_stop()) {
 428                 struct l_wait_info lwi = { 0 };
 429
 430                 l_wait_event(gns_thread.t_ctl_waitq,
 431                              (ll_gns_check_event() ||
 432                               ll_gns_check_stop()), &lwi);
 433
 434                 spin_lock(&gns_lock);
 435                 while (!list_empty(&gns_sbi_list)) {
 436                         struct ll_sb_info *sbi;
 437
 438                         sbi = list_entry(gns_sbi_list.prev,
 439                                          struct ll_sb_info,
 440                                          ll_gns_sbi_head);
 441
 442                         list_del_init(&sbi->ll_gns_sbi_head);
 443                         spin_unlock(&gns_lock);
 444                         ll_gns_check_mounts(sbi, LL_GNS_CHECK);
 445                         spin_lock(&gns_lock);
 446                 }
 447                 spin_unlock(&gns_lock);
 448         }
 449
 450         /*
 451          * letting know stop function know that thread is stoped and it may
 452          * return.
 453          */
 454         EXIT;
 455         gns_thread.t_flags = SVC_STOPPED;
 456
 457         /* this is SMP-safe way to finish thread. */
 458         complete_and_exit(&ctl->gc_finishing, 0);
 459 }
 460
 461 void ll_gns_add_timer(struct ll_sb_info *sbi)
 462 {
 463         mod_timer(&sbi->ll_gns_timer,
 464                   jiffies + sbi->ll_gns_tick * HZ);
 465 }
 466
 467 void ll_gns_del_timer(struct ll_sb_info *sbi)
 468 {
 469         del_timer(&sbi->ll_gns_timer);
 470 }
 471
 472 /*
 473  * starts GNS control thread and waits for a signal it is up and work may be
 474  * continued.
 475  */
 476 int ll_gns_start_thread(void)
 477 {
 478         int rc;
 479         ENTRY;
 480
 481         LASSERT(gns_thread.t_flags == 0);
 482         init_completion(&gns_ctl.gc_starting);
 483         init_completion(&gns_ctl.gc_finishing);
 484         init_waitqueue_head(&gns_thread.t_ctl_waitq);
 485
 486         rc = kernel_thread(ll_gns_thread_main, &gns_ctl,
 487                            (CLONE_VM | CLONE_FILES));
 488         if (rc < 0) {
 489                 CERROR("cannot start GNS control thread, "
 490                        "err = %d\n", rc);
 491                 RETURN(rc);
 492         }
 493         wait_for_completion(&gns_ctl.gc_starting);
 494         LASSERT(gns_thread.t_flags == SVC_RUNNING);
 495         RETURN(0);
 496 }
 497
 498 /* stops GNS control thread and waits its actual stop. */
 499 void ll_gns_stop_thread(void)
 500 {
 501         ENTRY;
 502         gns_thread.t_flags = SVC_STOPPING;
 503         wake_up(&gns_thread.t_ctl_waitq);
 504         wait_for_completion(&gns_ctl.gc_finishing);
 505         LASSERT(gns_thread.t_flags == SVC_STOPPED);
 506         gns_thread.t_flags = 0;
 507         EXIT;
 508 }