fs/notify/mark.c

   1 /*
   2  *  Copyright (C) 2008 Red Hat, Inc., Eric Paris <eparis@redhat.com>
   3  *
   4  *  This program is free software; you can redistribute it and/or modify
   5  *  it under the terms of the GNU General Public License as published by
   6  *  the Free Software Foundation; either version 2, or (at your option)
   7  *  any later version.
   8  *
   9  *  This program is distributed in the hope that it will be useful,
  10  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  *  GNU General Public License for more details.
  13  *
  14  *  You should have received a copy of the GNU General Public License
  15  *  along with this program; see the file COPYING.  If not, write to
  16  *  the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
  17  */
  18
  19 /*
  20  * fsnotify inode mark locking/lifetime/and refcnting
  21  *
  22  * REFCNT:
  23  * The mark->refcnt tells how many "things" in the kernel currently are
  24  * referencing this object.  The object typically will live inside the kernel
  25  * with a refcnt of 2, one for each list it is on (i_list, g_list).  Any task
  26  * which can find this object holding the appropriete locks, can take a reference
  27  * and the object itself is guaranteed to survive until the reference is dropped.
  28  *
  29  * LOCKING:
  30  * There are 3 spinlocks involved with fsnotify inode marks and they MUST
  31  * be taken in order as follows:
  32  *
  33  * mark->lock
  34  * group->mark_lock
  35  * inode->i_lock
  36  *
  37  * mark->lock protects 2 things, mark->group and mark->inode.  You must hold
  38  * that lock to dereference either of these things (they could be NULL even with
  39  * the lock)
  40  *
  41  * group->mark_lock protects the marks_list anchored inside a given group
  42  * and each mark is hooked via the g_list.  It also sorta protects the
  43  * free_g_list, which when used is anchored by a private list on the stack of the
  44  * task which held the group->mark_lock.
  45  *
  46  * inode->i_lock protects the i_fsnotify_marks list anchored inside a
  47  * given inode and each mark is hooked via the i_list. (and sorta the
  48  * free_i_list)
  49  *
  50  *
  51  * LIFETIME:
  52  * Inode marks survive between when they are added to an inode and when their
  53  * refcnt==0.
  54  *
  55  * The inode mark can be cleared for a number of different reasons including:
  56  * - The inode is unlinked for the last time.  (fsnotify_inode_remove)
  57  * - The inode is being evicted from cache. (fsnotify_inode_delete)
  58  * - The fs the inode is on is unmounted.  (fsnotify_inode_delete/fsnotify_unmount_inodes)
  59  * - Something explicitly requests that it be removed.  (fsnotify_destroy_mark)
  60  * - The fsnotify_group associated with the mark is going away and all such marks
  61  *   need to be cleaned up. (fsnotify_clear_marks_by_group)
  62  *
  63  * Worst case we are given an inode and need to clean up all the marks on that
  64  * inode.  We take i_lock and walk the i_fsnotify_marks safely.  For each
  65  * mark on the list we take a reference (so the mark can't disappear under us).
  66  * We remove that mark form the inode's list of marks and we add this mark to a
  67  * private list anchored on the stack using i_free_list;  At this point we no
  68  * longer fear anything finding the mark using the inode's list of marks.
  69  *
  70  * We can safely and locklessly run the private list on the stack of everything
  71  * we just unattached from the original inode.  For each mark on the private list
  72  * we grab the mark-> and can thus dereference mark->group and mark->inode.  If
  73  * we see the group and inode are not NULL we take those locks.  Now holding all
  74  * 3 locks we can completely remove the mark from other tasks finding it in the
  75  * future.  Remember, 10 things might already be referencing this mark, but they
  76  * better be holding a ref.  We drop our reference we took before we unhooked it
  77  * from the inode.  When the ref hits 0 we can free the mark.
  78  *
  79  * Very similarly for freeing by group, except we use free_g_list.
  80  *
  81  * This has the very interesting property of being able to run concurrently with
  82  * any (or all) other directions.
  83  */
  84
  85 #include <linux/fs.h>
  86 #include <linux/init.h>
  87 #include <linux/kernel.h>
  88 #include <linux/kthread.h>
  89 #include <linux/module.h>
  90 #include <linux/mutex.h>
  91 #include <linux/slab.h>
  92 #include <linux/spinlock.h>
  93 #include <linux/srcu.h>
  94
  95 #include <linux/atomic.h>
  96
  97 #include <linux/fsnotify_backend.h>
  98 #include "fsnotify.h"
  99
 100 struct srcu_struct fsnotify_mark_srcu;
 101 static DEFINE_SPINLOCK(destroy_lock);
 102 static LIST_HEAD(destroy_list);
 103 static DECLARE_WAIT_QUEUE_HEAD(destroy_waitq);
 104
 105 void fsnotify_get_mark(struct fsnotify_mark *mark)
 106 {
 107         atomic_inc(&mark->refcnt);
 108 }
 109
 110 void fsnotify_put_mark(struct fsnotify_mark *mark)
 111 {
 112         if (atomic_dec_and_test(&mark->refcnt)) {
 113                 if (mark->group)
 114                         fsnotify_put_group(mark->group);
 115                 mark->free_mark(mark);
 116         }
 117 }
 118
 119 /*
 120  * Any time a mark is getting freed we end up here.
 121  * The caller had better be holding a reference to this mark so we don't actually
 122  * do the final put under the mark->lock
 123  */
 124 void fsnotify_destroy_mark(struct fsnotify_mark *mark)
 125 {
 126         struct fsnotify_group *group;
 127         struct inode *inode = NULL;
 128
 129         spin_lock(&mark->lock);
 130         /* dont get the group from a mark that is not alive yet */
 131         if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) {
 132                 spin_unlock(&mark->lock);
 133                 return;
 134         }
 135         fsnotify_get_group(mark->group);
 136         group = mark->group;
 137         spin_unlock(&mark->lock);
 138
 139         spin_lock(&group->mark_lock);
 140         spin_lock(&mark->lock);
 141
 142         /* something else already called this function on this mark */
 143         if (!(mark->flags & FSNOTIFY_MARK_FLAG_ALIVE)) {
 144                 spin_unlock(&mark->lock);
 145                 spin_unlock(&group->mark_lock);
 146                 goto put_group;
 147         }
 148
 149         mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
 150
 151         if (mark->flags & FSNOTIFY_MARK_FLAG_INODE) {
 152                 inode = mark->i.inode;
 153                 fsnotify_destroy_inode_mark(mark);
 154         } else if (mark->flags & FSNOTIFY_MARK_FLAG_VFSMOUNT)
 155                 fsnotify_destroy_vfsmount_mark(mark);
 156         else
 157                 BUG();
 158
 159         list_del_init(&mark->g_list);
 160
 161         spin_unlock(&mark->lock);
 162         spin_unlock(&group->mark_lock);
 163
 164         spin_lock(&destroy_lock);
 165         list_add(&mark->destroy_list, &destroy_list);
 166         spin_unlock(&destroy_lock);
 167         wake_up(&destroy_waitq);
 168
 169         /*
 170          * Some groups like to know that marks are being freed.  This is a
 171          * callback to the group function to let it know that this mark
 172          * is being freed.
 173          */
 174         if (group->ops->freeing_mark)
 175                 group->ops->freeing_mark(mark, group);
 176
 177         /*
 178          * __fsnotify_update_child_dentry_flags(inode);
 179          *
 180          * I really want to call that, but we can't, we have no idea if the inode
 181          * still exists the second we drop the mark->lock.
 182          *
 183          * The next time an event arrive to this inode from one of it's children
 184          * __fsnotify_parent will see that the inode doesn't care about it's
 185          * children and will update all of these flags then.  So really this
 186          * is just a lazy update (and could be a perf win...)
 187          */
 188
 189         if (inode && (mark->flags & FSNOTIFY_MARK_FLAG_OBJECT_PINNED))
 190                 iput(inode);
 191         /*
 192          * We don't necessarily have a ref on mark from caller so the above iput
 193          * may have already destroyed it.  Don't touch from now on.
 194          */
 195
 196         atomic_dec(&group->num_marks);
 197
 198 put_group:
 199         fsnotify_put_group(group);
 200 }
 201
 202 void fsnotify_set_mark_mask_locked(struct fsnotify_mark *mark, __u32 mask)
 203 {
 204         assert_spin_locked(&mark->lock);
 205
 206         mark->mask = mask;
 207
 208         if (mark->flags & FSNOTIFY_MARK_FLAG_INODE)
 209                 fsnotify_set_inode_mark_mask_locked(mark, mask);
 210 }
 211
 212 void fsnotify_set_mark_ignored_mask_locked(struct fsnotify_mark *mark, __u32 mask)
 213 {
 214         assert_spin_locked(&mark->lock);
 215
 216         mark->ignored_mask = mask;
 217 }
 218
 219 /*
 220  * Attach an initialized mark to a given group and fs object.
 221  * These marks may be used for the fsnotify backend to determine which
 222  * event types should be delivered to which group.
 223  */
 224 int fsnotify_add_mark(struct fsnotify_mark *mark,
 225                       struct fsnotify_group *group, struct inode *inode,
 226                       struct vfsmount *mnt, int allow_dups)
 227 {
 228         int ret = 0;
 229
 230         BUG_ON(inode && mnt);
 231         BUG_ON(!inode && !mnt);
 232
 233         /*
 234          * LOCKING ORDER!!!!
 235          * group->mark_lock
 236          * mark->lock
 237          * inode->i_lock
 238          */
 239         spin_lock(&group->mark_lock);
 240
 241         spin_lock(&mark->lock);
 242         mark->flags |= FSNOTIFY_MARK_FLAG_ALIVE;
 243
 244         fsnotify_get_group(group);
 245         mark->group = group;
 246         list_add(&mark->g_list, &group->marks_list);
 247         atomic_inc(&group->num_marks);
 248         fsnotify_get_mark(mark); /* for i_list and g_list */
 249
 250         if (inode) {
 251                 ret = fsnotify_add_inode_mark(mark, group, inode, allow_dups);
 252                 if (ret)
 253                         goto err;
 254         } else if (mnt) {
 255                 ret = fsnotify_add_vfsmount_mark(mark, group, mnt, allow_dups);
 256                 if (ret)
 257                         goto err;
 258         } else {
 259                 BUG();
 260         }
 261
 262         /* this will pin the object if appropriate */
 263         fsnotify_set_mark_mask_locked(mark, mark->mask);
 264         spin_unlock(&mark->lock);
 265
 266         spin_unlock(&group->mark_lock);
 267
 268         if (inode)
 269                 __fsnotify_update_child_dentry_flags(inode);
 270
 271         return ret;
 272 err:
 273         mark->flags &= ~FSNOTIFY_MARK_FLAG_ALIVE;
 274         list_del_init(&mark->g_list);
 275         fsnotify_put_group(group);
 276         mark->group = NULL;
 277         atomic_dec(&group->num_marks);
 278
 279         spin_unlock(&mark->lock);
 280         spin_unlock(&group->mark_lock);
 281
 282         spin_lock(&destroy_lock);
 283         list_add(&mark->destroy_list, &destroy_list);
 284         spin_unlock(&destroy_lock);
 285         wake_up(&destroy_waitq);
 286
 287         return ret;
 288 }
 289
 290 /*
 291  * clear any marks in a group in which mark->flags & flags is true
 292  */
 293 void fsnotify_clear_marks_by_group_flags(struct fsnotify_group *group,
 294                                          unsigned int flags)
 295 {
 296         struct fsnotify_mark *lmark, *mark;
 297         LIST_HEAD(free_list);
 298
 299         spin_lock(&group->mark_lock);
 300         list_for_each_entry_safe(mark, lmark, &group->marks_list, g_list) {
 301                 if (mark->flags & flags) {
 302                         list_add(&mark->free_g_list, &free_list);
 303                         list_del_init(&mark->g_list);
 304                         fsnotify_get_mark(mark);
 305                 }
 306         }
 307         spin_unlock(&group->mark_lock);
 308
 309         list_for_each_entry_safe(mark, lmark, &free_list, free_g_list) {
 310                 fsnotify_destroy_mark(mark);
 311                 fsnotify_put_mark(mark);
 312         }
 313 }
 314
 315 /*
 316  * Given a group, destroy all of the marks associated with that group.
 317  */
 318 void fsnotify_clear_marks_by_group(struct fsnotify_group *group)
 319 {
 320         fsnotify_clear_marks_by_group_flags(group, (unsigned int)-1);
 321 }
 322
 323 void fsnotify_duplicate_mark(struct fsnotify_mark *new, struct fsnotify_mark *old)
 324 {
 325         assert_spin_locked(&old->lock);
 326         new->i.inode = old->i.inode;
 327         new->m.mnt = old->m.mnt;
 328         if (old->group)
 329                 fsnotify_get_group(old->group);
 330         new->group = old->group;
 331         new->mask = old->mask;
 332         new->free_mark = old->free_mark;
 333 }
 334
 335 /*
 336  * Nothing fancy, just initialize lists and locks and counters.
 337  */
 338 void fsnotify_init_mark(struct fsnotify_mark *mark,
 339                         void (*free_mark)(struct fsnotify_mark *mark))
 340 {
 341         memset(mark, 0, sizeof(*mark));
 342         spin_lock_init(&mark->lock);
 343         atomic_set(&mark->refcnt, 1);
 344         mark->free_mark = free_mark;
 345 }
 346
 347 static int fsnotify_mark_destroy(void *ignored)
 348 {
 349         struct fsnotify_mark *mark, *next;
 350         LIST_HEAD(private_destroy_list);
 351
 352         for (;;) {
 353                 spin_lock(&destroy_lock);
 354                 /* exchange the list head */
 355                 list_replace_init(&destroy_list, &private_destroy_list);
 356                 spin_unlock(&destroy_lock);
 357
 358                 synchronize_srcu(&fsnotify_mark_srcu);
 359
 360                 list_for_each_entry_safe(mark, next, &private_destroy_list, destroy_list) {
 361                         list_del_init(&mark->destroy_list);
 362                         fsnotify_put_mark(mark);
 363                 }
 364
 365                 wait_event_interruptible(destroy_waitq, !list_empty(&destroy_list));
 366         }
 367
 368         return 0;
 369 }
 370
 371 static int __init fsnotify_mark_init(void)
 372 {
 373         struct task_struct *thread;
 374
 375         thread = kthread_run(fsnotify_mark_destroy, NULL,
 376                              "fsnotify_mark");
 377         if (IS_ERR(thread))
 378                 panic("unable to start fsnotify mark destruction thread.");
 379
 380         return 0;
 381 }
 382 device_initcall(fsnotify_mark_init);