1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2004, 2005 Cluster File Systems, Inc.
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>
11 * This file is part of Lustre, http://www.lustre.org.
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.
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.
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.
27 #define DEBUG_SUBSYSTEM S_LLITE
30 #include <linux/version.h>
31 #include <asm/uaccess.h>
32 #include <linux/file.h>
33 #include <linux/kmod.h>
35 #include <linux/lustre_lite.h>
36 #include "llite_internal.h"
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;
43 #define CONCUR_GNS_RESTART_APPROACH 0
46 * waits until passed dentry gets mountpoint or timeout and attempts are
47 * exhausted. Returns 1 if dentry became mountpoint and 0 otherwise.
50 ll_gns_wait_for_mount(struct dentry *dentry,
51 int timeout, int tries)
53 struct l_wait_info lwi;
54 struct ll_sb_info *sbi;
57 LASSERT(dentry != NULL);
58 LASSERT(!IS_ERR(dentry));
59 sbi = ll_s2sbi(dentry->d_sb);
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);
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);
74 #if (CONCUR_GNS_RESTART_APPROACH == 1)
76 * sending a signal known to be ignored to cause restarting syscall if GNS mount
77 * function returns -ERESTARTSYS.
80 ll_gns_send_signal(void)
82 struct task_struct *task = current;
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);
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
100 ll_gns_mount_object(struct dentry *dentry, struct vfsmount *mnt)
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;
109 LASSERT(dentry->d_inode != NULL);
111 if (!S_ISDIR(dentry->d_inode->i_mode))
114 sbi = ll_i2sbi(dentry->d_inode);
117 CERROR("suid directory found, but no "
118 "vfsmount available.\n");
122 if (atomic_read(&sbi->ll_gns_enabled) == 0)
125 spin_lock(&sbi->ll_gns_lock);
128 * another thead is in progress or just finished mounting the
129 * dentry. Handling that.
131 if (sbi->ll_gns_state == LL_GNS_MOUNTING ||
132 sbi->ll_gns_state == LL_GNS_FINISHED) {
134 * another thread is trying to mount GNS dentry. We'd like to
137 spin_unlock(&sbi->ll_gns_lock);
140 * check if dentry is mount point already, if so, do not restart
143 if (d_mountpoint(dentry))
146 #if (CONCUR_GNS_RESTART_APPROACH == 1)
148 * causing syscall to restart and possibly find this dentry
151 ll_gns_send_signal();
152 RETURN(-ERESTARTSYS);
155 * waiting for GNS complete and check dentry again, it may be
158 wait_for_completion(&sbi->ll_gns_mount_finished);
159 if (d_mountpoint(dentry))
163 LASSERT(sbi->ll_gns_state == LL_GNS_IDLE);
165 CDEBUG(D_INODE, "mounting dentry %p\n", dentry);
167 /* mounting started */
168 sbi->ll_gns_state = LL_GNS_MOUNTING;
169 spin_unlock(&sbi->ll_gns_lock);
171 /* we need to build an absolute pathname to pass to mount */
172 pathpage = (char *)__get_free_page(GFP_KERNEL);
174 GOTO(cleanup, rc = -ENOMEM);
177 /* getting @dentry path stored in @pathpage. */
178 path = d_path(dentry, mnt, pathpage, PAGE_SIZE);
180 CERROR("can't build mount object path, err %d\n",
181 (int)PTR_ERR(dchild));
182 GOTO(cleanup, rc = PTR_ERR(dchild));
185 /* synchronizing with possible /proc/fs/...write */
186 down(&sbi->ll_gns_sem);
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.
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.
199 dchild = ll_lookup_one_len(sbi->ll_gns_oname, dentry,
200 strlen(sbi->ll_gns_oname));
201 up(&sbi->ll_gns_sem);
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,
214 /* mount object is not found */
215 if (!dchild->d_inode)
216 GOTO(cleanup, rc = -ENOENT);
218 /* check if found child is regular file */
219 if (!S_ISREG(dchild->d_inode->i_mode))
220 GOTO(cleanup, rc = -EBADF);
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));
232 GOTO(cleanup, rc = PTR_ERR(mntinfo_fd));
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);
244 datapage = (char *)__get_free_page(GFP_KERNEL);
246 GOTO(cleanup, rc = -ENOMEM);
250 /* read data from mount object. */
251 rc = kernel_read(mntinfo_fd, 0, datapage, PAGE_SIZE - 1);
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,
260 /* no data in mount object? */
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);
273 /* synchronizing with possible /proc/fs/...write */
274 down(&sbi->ll_gns_sem);
277 * upcall is initialized in mount time or via /proc/fs/... tuneable and
278 * may be /usr/lib/lustre/gns-upcall.sh
280 argv[0] = sbi->ll_gns_upcall;
285 up(&sbi->ll_gns_sem);
287 /* do not wait for helper complete here. */
288 rc = call_usermodehelper(argv[0], argv, NULL, 1);
290 CWARN("failed to call GNS upcall %s, err = %d, "
291 "checking for mount anyway\n", sbi->ll_gns_upcall, rc);
295 * waiting for dentry become mount point GNS_WAIT_ATTEMPTS times by 1
298 rc = ll_gns_wait_for_mount(dentry, 1, GNS_WAIT_ATTEMPTS);
300 struct dentry *rdentry;
301 struct vfsmount *rmnt;
303 LASSERT(sbi->ll_gns_state == LL_GNS_FINISHED);
306 rdentry = dget(dentry);
308 if (follow_down(&rmnt, &rdentry)) {
310 * registering new mount in GNS mounts list and thus
311 * make it accessible from GNS control thread.
313 spin_lock(&dcache_lock);
314 LASSERT(list_empty(&rmnt->mnt_lustre_list));
315 list_add_tail(&rmnt->mnt_lustre_list,
317 spin_unlock(&dcache_lock);
318 rmnt->mnt_last_used = jiffies;
326 CERROR("usermode upcall %s failed to mount %s, err %d\n",
327 sbi->ll_gns_upcall, path, rc);
332 switch (cleanup_phase) {
334 free_page((unsigned long)datapage);
336 if (mntinfo_fd != NULL)
342 free_page((unsigned long)pathpage);
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);
352 /* tries to umount passed @mnt. */
353 int ll_gns_umount_object(struct vfsmount *mnt)
358 CDEBUG(D_INODE, "unmounting mnt %p\n", mnt);
359 rc = do_umount(mnt, 0);
361 CDEBUG(D_INODE, "can't umount 0x%p, err = %d\n",
368 int ll_gns_check_mounts(struct ll_sb_info *sbi, int flags)
370 struct list_head check_list = LIST_HEAD_INIT(check_list);
371 struct vfsmount *mnt;
375 spin_lock(&dcache_lock);
376 list_splice_init(&sbi->ll_mnt_list, &check_list);
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.
383 while (!list_empty(&check_list)) {
384 mnt = list_entry(check_list.prev,
390 list_del_init(&mnt->mnt_lustre_list);
392 list_add(&mnt->mnt_lustre_list,
395 /* check for timeout if needed */
396 pass = jiffies - mnt->mnt_last_used;
398 if (flags == LL_GNS_CHECK &&
399 pass < sbi->ll_gns_timeout * HZ)
404 spin_unlock(&dcache_lock);
407 ll_gns_umount_object(mnt);
410 spin_lock(&dcache_lock);
412 spin_unlock(&dcache_lock);
417 * GNS timer callback function. It restarts gns timer and wakes up GNS control
418 * thread to process mounts list.
420 void ll_gns_timer_callback(unsigned long data)
422 struct ll_sb_info *sbi = (void *)data;
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);
430 wake_up(&gns_thread.t_ctl_waitq);
431 mod_timer(&sbi->ll_gns_timer,
432 jiffies + sbi->ll_gns_tick * HZ);
435 /* this function checks if something new happened to exist in gns list. */
436 static int inline ll_gns_check_event(void)
440 spin_lock(&gns_lock);
441 rc = !list_empty(&gns_sbi_list);
442 spin_unlock(&gns_lock);
447 /* should we stop GNS control thread? */
448 static int inline ll_gns_check_stop(void)
451 return (gns_thread.t_flags & SVC_STOPPING) ? 1 : 0;
454 /* GNS control thread function. */
455 static int ll_gns_thread_main(void *arg)
457 struct ll_gns_ctl *ctl = arg;
462 char name[sizeof(current->comm)];
463 snprintf(name, sizeof(name) - 1, "ll_gns");
464 kportal_daemonize(name);
467 SIGNAL_MASK_LOCK(current, flags);
468 sigfillset(¤t->blocked);
470 SIGNAL_MASK_UNLOCK(current, flags);
473 * letting starting function know, that we are ready and control may be
476 gns_thread.t_flags = SVC_RUNNING;
477 complete(&ctl->gc_starting);
479 while (!ll_gns_check_stop()) {
480 struct l_wait_info lwi = { 0 };
482 l_wait_event(gns_thread.t_ctl_waitq,
483 (ll_gns_check_event() ||
484 ll_gns_check_stop()), &lwi);
486 spin_lock(&gns_lock);
487 while (!list_empty(&gns_sbi_list)) {
488 struct ll_sb_info *sbi;
490 sbi = list_entry(gns_sbi_list.prev,
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);
499 spin_unlock(&gns_lock);
503 gns_thread.t_flags = SVC_STOPPED;
505 /* this is SMP-safe way to finish thread. */
506 complete_and_exit(&ctl->gc_finishing, 0);
509 void ll_gns_add_timer(struct ll_sb_info *sbi)
511 mod_timer(&sbi->ll_gns_timer,
512 jiffies + sbi->ll_gns_tick * HZ);
515 void ll_gns_del_timer(struct ll_sb_info *sbi)
517 del_timer(&sbi->ll_gns_timer);
521 * starts GNS control thread and waits for a signal it is up and work may be
524 int ll_gns_start_thread(void)
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);
534 rc = kernel_thread(ll_gns_thread_main, &gns_ctl,
535 (CLONE_VM | CLONE_FILES));
537 CERROR("cannot start GNS control thread, "
541 wait_for_completion(&gns_ctl.gc_starting);
542 LASSERT(gns_thread.t_flags == SVC_RUNNING);
546 /* stops GNS control thread and waits its actual stop. */
547 void ll_gns_stop_thread(void)
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;